characterization of altered microrna expression …...iv acknowledgments first and foremost, i would...

i

Characterization of Altered MicroRNA Expression in Cervical Cancer

by

Christine How

A thesis submitted in conformity with the requirements for the degree of Doctor of Philosophy

Graduate Department of Medical Biophysics University of Toronto

© Copyright by Christine How 2013

ii

Characterization of Altered MicroRNA

Expression in Cervical Cancer

Christine How

Doctor of Philosophy

Graduate Department of Medical Biophysics University of Toronto

2013

Abstract

Cervical cancer is the third most common cancer among women worldwide, and the

fourth leading cause of cancer mortality. Despite significant declines in the incidence and

mortality rates of cervical cancer in Canada, it remains the 4th most common cancer in women

aged 20-29 years. In order to gain novel insights into cervical cancer tumourigenesis and clinical

outcome, we investigated and characterized the alterations in microRNA (miRNA) expression in

this disease. Firstly, we performed global miRNA expression profiling of cervical cancer cell

lines (n=3), and patient specimens (n=79). From this analysis, we identified miR-196b to be

significantly down-regulated in cervical cancer, and characterized its role in regulating the

HOXB7~VEGF axis. The global miRNA expression data also led to the development of a

candidate 9-miRNA signature that was prognostic for disease-free survival in patients with

cervical cancer, although we were unable to validate this signature in an independent cohort.

This report describes important considerations concerning the development and validation of

microRNA signatures for cervical cancer.

Our investigations also led us to a comparison of three methods for measuring miRNA

abundance: the TaqMan Low Density Array, the NanoString nCounter assay, and single-well

quantitative real-time PCR. Our findings demonstrated limited concordance between the TLDA

and NanoString platforms, although each platform correlated well with PCR, which is considered

the gold standard for nucleic acid quantification. Furthermore, we examined biases created by

amplification protocols for microarray studies. Our analysis demonstrated that performing a

iii

correction using the LTR-method (linear transformation of replicates) could help mitigate, but

not completely eliminate such biases.

Overall, this report presents insights into the role of miRNAs in cervical cancer, as well

as an evaluation of technical considerations concerning miRNA and mRNA expression profiling

studies.

iv

Acknowledgments

First and foremost, I would like to acknowledge my PhD supervisor, Dr. Fei-Fei Liu, for

giving me the opportunity to pursue my graduate studies in her lab and collaborate with world-

class scientists and clinicians. You have been an invaluable teacher and role model, and I am

truly thankful for your mentorship and support during my PhD studies. My achievements and

accolades could not have been possible without your continuous encouragement and guidance.

Special thanks to my supervisory committee, Dr. Anthony Fyles and Dr. Mark Minden, for

providing me with keen advice and valuable feedback. You have always encouraged and

challenged me to think more broadly, which I will continue to strive to do in the future.

It has been a pleasure to work with all members of the Liu Lab, and there are a few

current and past members whom I would like to acknowledge in particular: Angela Hui, for

providing guidance and sharing your wealth of knowledge, you have played an instrumental role

in conceptualizing and developing my projects. Willa Shi, for providing encouragement and

advice, and assisting with specimen preparation and immunohistochemistry. Nehad Alajez, for

your advice and assistance with experimental design and data analysis. Jeff Bruce, for always

being willing and enthusiastic in assisting with my projects. Michelle Lenarduzzi, for being a

great lab bench neighbor, conference roommate, and friend. Winnie Yue, for assisting with

animal work and always offering encouragement. Other colleagues I would like to thank for their

contributions to this thesis include: Melania Pintilie, Trudey Nicklee, Blaise Clarke, Rui Yan,

and Paul Boutros.

Finally, I owe my sincere gratitude to my family and friends for always supporting me

throughout the highs and lows of graduate school. To my mother (Mary) and father (Gary), thank

you for your unconditional love, support and patience. I am humbled and grateful for everything

you have done for me, and I would not be the person I am today without you. I hope I have made

you proud. To my sister (Maxine) and brother (Andrew), thank you for keeping me motivated

and reminding me to have fun. To my dearest friends, thank you for always trying to put a smile

on my face, and for believing in me even when I didn’t believe in myself.

v

Table of Contents

Abstract ........................................................................................................................................... ii

Acknowledgments.......................................................................................................................... iv

Table of Contents .............................................................................................................................v

List of Tables ...................................................................................................................................x

List of Figures ................................................................................................................................ xi

List of Abbreviations ................................................................................................................... xiii

CHAPTER 1: INTRODUCTION ....................................................................................................1

1.1 MicroRNAs ..........................................................................................................................2

1.1.1 Overview ..................................................................................................................2

1.1.2 History......................................................................................................................2

1.1.3 Biogenesis and Processing .......................................................................................2

1.1.4 RNA-induced Silencing Complex (RISC) Assembly and Function ........................6

1.1.5 Target Prediction Databases ....................................................................................7

1.1.6 MicroRNAs in Cancer .............................................................................................7

1.1.7 MicroRNA Expression Profiling .............................................................................9

1.1.8 Platforms for MicroRNA Expression Profiling .....................................................10

1.1.9 microRNAs as Biomarkers ....................................................................................12

1.2 Cervical Cancer ..................................................................................................................13

1.2.1 Background ............................................................................................................13

1.2.2 Human Papillomavirus ...........................................................................................14

1.2.3 Treatment ...............................................................................................................16

1.2.4 Gene Expression Signatures for Cervical Cancer ..................................................16

1.2.5 MicroRNA Expression in Cervical Cancer ............................................................17

1.3 Homeobox Genes ...............................................................................................................18

vi

1.3.1 Overview ................................................................................................................18

1.3.2 Role in Development .............................................................................................20

1.3.3 Role in Cancer ........................................................................................................20

1.4 Vascular Endothelial Growth Factor .................................................................................21

1.4.1 Overview ................................................................................................................21

1.4.2 Receptors................................................................................................................21

1.4.3 Isoforms .................................................................................................................22

1.4.4 VEGF in Cervical Cancer ......................................................................................22

1.5 Research Objectives ...........................................................................................................24

CHAPTER 2: MICRORNA-196B REGULATES THE HOMEOBOX-B7/VASCULAR ENDOTHELIAL GROWTH FACTOR AXIS IN CERVICAL CANCER ..............................25

2.1 Chapter Abstract ................................................................................................................26

2.2 Introduction ........................................................................................................................26

2.3 Materials and Methods .......................................................................................................27

2.3.1 Cell Lines and Transfections .................................................................................27

2.3.2 miRNA Expression Profiling of Cell Lines ...........................................................27

2.3.3 miRNA Expression Profiling of Patient Tissues ...................................................28

2.3.4 Quantitative Real-time PCR Analysis of miRNAs and mRNAs ...........................29

2.3.5 Cell Viability, Proliferation and Colony-Formation Assays..................................31

2.3.6 Cell Migration and Invasion Assays ......................................................................31

2.3.7 Cell Cycle Analysis................................................................................................31

2.3.8 In vivo Experiments ...............................................................................................32

2.3.9 Tri-modal Approach for miRNA Target Identification .........................................32

2.3.10 Luciferase Reporter Assay .....................................................................................33

2.3.11 Immunoblotting......................................................................................................33

2.3.12 Enzyme-linked Immunosorbent Assay (ELISA) ...................................................33

vii

2.3.13 5-aza-2’-deoxycytidine Treatment .........................................................................33

2.3.14 Statistical Analysis .................................................................................................34

2.4 Results ................................................................................................................................34

2.4.1 miR-196b was significantly down-regulated in primary cervical cancer tissues and cell lines, and was strongly associated with DFS ...........................................34

2.4.2 miR-196b over-expression significantly reduced cell viability, clonogenicity, proliferation and invasion ......................................................................................36

2.4.3 miR-196b over-expression suppressed tumor angiogenesis and tumor cell proliferation in vivo ................................................................................................36

2.4.4 miR-196b directly targets HOXB7 ........................................................................39

2.4.5 VEGF is a relevant downstream target of HOXB7 ...............................................41

2.4.6 Knockdown of HOXB7 or VEGF recapitulated the biological effects observed following mir-196b over-expression ......................................................................41

2.5 Discussion ..........................................................................................................................45

CHAPTER 3: CHALLENGES ASSOCIATED WITH DEVELOPING A PROGNOSTIC MICRORNA SIGNATURE FOR HUMAN CERVICAL CARCINOMA ..............................56


3.2 Introduction ........................................................................................................................57


3.3.1 Clinical specimens .................................................................................................58

3.3.2 Sample processing .................................................................................................59

3.3.3 Data normalization .................................................................................................60

3.3.4 Survival analysis ....................................................................................................60

3.4 Results ................................................................................................................................61

3.4.1 The majority of significantly differentially-expressed miRNAs were downregulated ........................................................................................................61

3.4.2 Patient miRNA expression was associated with disease-free survival ..................63

3.4.3 Validation of miRNA signature .............................................................................63

3.4.4 Independent corroboration of the Hu et al. 2-miRNA signature ...........................66

viii

3.5 Discussion ..........................................................................................................................68

CHAPTER 4: IDENTIFICATION AND CORRECTION OF AMPLIFICATION PROTOCOL BIAS IN MICROARRAY STUDIES .................................................................73


4.2 Introduction ........................................................................................................................75


4.3.1 Samples ..................................................................................................................76

4.3.2 Sample processing .................................................................................................77

4.3.3 Microarray hybridization .......................................................................................77

4.3.4 Pre-processing ........................................................................................................78

4.3.5 Unsupervised machine-learning.............................................................................78

4.3.6 Correlation analyses ...............................................................................................79

4.3.7 Statistical analysis of microarray data ...................................................................79

4.3.8 Gene subset division ..............................................................................................80

4.3.9 Functional analysis.................................................................................................80

4.3.10 Sequence analysis ..................................................................................................80

4.3.11 Visualization ..........................................................................................................81

4.4 Results ................................................................................................................................81

4.4.1 Experimental approach ..........................................................................................81

4.4.2 Nugen and express data are not directly comparable .............................................82

4.4.3 Protocol introduces systematic bias .......................................................................85

4.4.4 Protocol noise is larger than biological signal .......................................................85

4.4.5 Protocol difference biases specific biological functions ........................................88

4.4.6 Sequence characteristics do not explain protocol bias ...........................................88

4.4.7 Batch effects adjustment partially mitigates protocol bias ....................................91

4.4.8 Replication analysis ...............................................................................................91

ix

4.5 Discussion ..........................................................................................................................93

CHAPTER 5: SYSTEMATIC COMPARISON OF PLATFORMS FOR MICRORNA PROFILING ............................................................................................................................103

5.1 Chapter Abstract ..............................................................................................................104

5.2 Introduction ......................................................................................................................104

5.3 Materials and Methods .....................................................................................................106

5.3.1 Samples and sample processing ...........................................................................106

5.3.2 Data description ...................................................................................................107

5.3.3 Data pre-processing .............................................................................................109

5.3.4 Unsupervised machine-learning...........................................................................109

5.3.5 Statistical analysis ................................................................................................109

5.3.6 Visualization ........................................................................................................110

5.4 Results ..............................................................................................................................110

5.4.1 Experimental design.............................................................................................110

5.4.2 Biological differences are detectable by platforms ..............................................111

5.4.3 Data from NanoString and TLDA are significantly different ..............................111

5.4.4 PCR is more similar to TLDA than NanoString ..................................................112

5.4.5 Fold-changes are poorly correlated between NanoString vs. TLDA, but differentially-expressed miRNAs are highly correlated across platforms ...........113

5.5 Discussion ........................................................................................................................120

CHAPTER 6: DISCUSSION .......................................................................................................123

6.1 Research Summary ..........................................................................................................124

6.2 Future Directions .............................................................................................................124

6.3 Conclusions ......................................................................................................................126

References ....................................................................................................................................127

x

List of Tables

Table 2.1 Clinical parameters of 79 cervical cancer patients. .............................................. 29

Table 2.2 Primer Sequences used for qRT-PCR ................................................................... 30

Table 3.1 Clinical parameters of patients in the training and validation cohorts .................. 62

Table S3.1 Significantly differentially-expressed miRNAs in cervical cancer ...................... 71

Table S4.1 Sample descriptions (Data: cervix data) ............................................................. 966

Table 5.1 Data descriptions for breast cancer datasets ....................................................... 108

Table 5.2 Data descriptions for cervical cancer datasets .................................................... 108

Table 5.3 Data descriptions for NPC datasets ..................................................................... 108

xi

List of Figures

Figure 1.1 MicroRNA biogenesis and processing .................................................................... 4

Figure 1.2 Map of HPV genome ............................................................................................ 15

Figure 1.3 Human HOX gene clusters ................................................................................... 19

Figure 1.4 VEGF family members and receptors ................................................................... 23

Figure 2.1 miR-196b down-regulation in cervical cancer ...................................................... 35

Figure 2.2 In vitro effects of miR-196b over-expression ....................................................... 37

Figure 2.3 Identification of HOXB7 as an mRNA target of miR-196b ................................. 40

Figure 2.4 Identification of VEGF as a target of the HOXB7 transcription factor ................ 42

Figure 2.5 Downstream effects of siHOXB7 and siVEGF .................................................... 43

Figure S2.1 In vitro effects of 5-aza-DCT treatment and miR-196b over-expression ............. 48

Figure S2.2 Effect of miR-196b on in vivo tumor growth........................................................ 50

Figure S2.3 Effect of miR-196b on in vivo angiogenesis and apoptosis .................................. 51

Figure S2.4 Tri-modal strategy for target identification ......................................................... 522

Figure S2.5 Transcript levels of putative miR-196b targets ................................................... 533

Figure S2.6 In vitro effects of treatment with siHOXB7, siVEGF, or pre-miR-196b ............. 54

Figure 3.1 Kaplan-Meier analysis of DFS according to 9-miRNA signature ........................ 64

Figure 3.2 Application of 9-miRNA signature to validation cohort ...................................... 65

Figure 3.3 Kaplan-Meier analysis of DFS according to Hu et al. 2-miRNA signature ......... 67

Figure S3.1 Application of TCGA Small RNA-Seq data to cervical cancer signatures .......... 72

xii

Figure 4.1 Experimental design............................................................................................ 833

Figure 4.2 Correlations between technical replicates ........................................................... 844

Figure 4.3 Specific genes are affected by protocol bias ....................................................... 866

Figure 4.4 Differentially expressed genes between biological and technical replicates ........ 87

Figure 4.5 Differing characteristics of significantly differentially expressed genes ............ 899

Figure 4.6 Application of LTR algorithm to remove platform specific biases ...................... 92

Figure S4.1 Distributions of Selected Genes (Dataset: Cervix data) ....................................... 97

Figure S4.2 Overlap of significant GO terms by Protocol or sample type ............................... 98

Figure S4.3 Correlations between technical replicates (Data: kidney data) ........................... 999

Figure S4.4 Distributions of mRNA abundances (Data: kidney data) ................................... 100

Figure S4.5 Correlation comparisons between cervix and kidney data. ................................ 101

Figure S4.6 Distribution of Y-chromosome filter process for cervix data ............................. 102

Figure 5.1 Experimental design............................................................................................ 114

Figure 5.2 Correlations between technical replicates (TLDA vs. NanoString) ................. 1155

Figure 5.3 Distributions and correlations of miRNA abundances across platforms .......... 1166

Figure 5.4 Correlations between technical replicates with PCR .......................................... 117

Figure 5.5 Sample-wise and gene-wise correlations between PCR vs. NanoString and PCR

vs. TLDA............................................................................................................. 118

Figure 5.6 Correlations of tumour/normal fold-changes between NanoString and TLDA .. 119

xiii

List of Abbreviations

aCGH Array comparative genomic hybridization AGO Argonaute AGO2 Argonaute 2 ALL Acute lymphoblastic leukemia AML Acute myeloid leukemia AML Acute myeloid leukemia ANKHD1 Ankyrin repeat and KH domain containing 1 ANOVA Analysis of variance ARI Adjusted rand index ATP Adenosine triophosphate AU Adenine-uracil BCL2 B-cell Lymphoma 2 BCL6 B-cell Lymphoma 6 bp Base pair CD31 Cluster of differentiation 31 cDNA Complementary DNA CLL Chronic lymphocytic leukemia COX-2 Cyclooxygenase-2 Ct Threshold cycle CT Computed tomography CTDSP2 CTD (carboxy-terminal domain, RNA polymerase II, polypeptide A) small

phosphatase 2 CRT Chemo-radiotherapy DCE-MRI Dynamic contrast-enhanced magnetic resonance imaging DFS Disease-free survival DGCR8 DiGeorge Syndrome Critical Region 8 DNA Deoxyribonucleic acid eIF4E Eukaryotic translation initiation factor 4E eIF4G Eukaryotic translation initiation factor 4 gamma eIF6 Eukaryotic translation initiation factor 6 ELISA Enzyme-linked immunosorbent assay ER Estrogen receptor FACS Fluorescence-activated cell sorting FBS Fetal bovine serum FDR False-discovery rate FFPE Formalin-fixed paraffin-embedded FGFR1 Fibroblast growth factor receptor 1 FIGO International Federation of Gynecologists and Obstetricians FLT1 Fms-related tyrosine kinase 1 FLT4 Fms-related tyrosine kinase 4

xiv

FOXO1 Forkhead box O1 GAPDH Glyceraldehyde-3-phosphate dehydrogenase GEO Gene Expression Omnibus GO Gene ontology GTP Guanosine triphosphate HDAC Histone deacetylase HE Hematoxylin and eosin HER2 Human epidermal growth factor receptor 2 HOXA Homeobox-A HOXA7 Homeobox-A7 HOXA9 Homeobox-A9 HOXB Homeobox-B HOXB7 Homeobox-B7 HPV Human Papillomavirus Ig Immunoglobulin IGFR1 Insulin-like growth factor receptor 1 IMRT Intensity modulated radiotherapy KDR Kinase insert domain receptor KRT8 Keratin 8 LTR Linear transformation of replicates MAQC MicroArray Quality Control MEIS1 Myeloid ecotropic viral integration site 1 homolog MEM Minimal essential medium miRNA microRNA MLL Mixed lineage leukemia MRI Magnetic resonance imaging mRNA Messenger RNA MTS 3-(4,5-dimethylthiazol-2-yl)-5-(3-carboxymethoxyphenyl)-2-(4-sulfophenyl)-

2H-tetrazolium NC Negative Control NPC Nasopharyngeal carcinoma nt Nucleotide OCT Optimal cutting temperature OS Overall survival Pap Papanicolaou PCR Polymerase chain reaction PlGF Placenta growth factor pRb Retinoblastoma protein Pre-miRNA Precursor microRNA

xv

Pri-miRNA Primary microRNA PTEN Phosphatase and tensin homolog PUM2 Pumilio homolog 2 qRT-PCR Quantitative real-time PCR RIN RNA integrity number RISC RNA-induced silencing complex RMA Robust multi-array average RNA Ribonucleic acid RNAP II RNA Polymerase II RNAP III RNA Polymerase III RT Radiotherapy SCC Squamous cell carcinoma SCJ Squamocolumnar junction SCID Severe combined immunodeficiency SEM Standard error of the mean siRNA Small interfering RNA SLC9A6 solute carrier family 9, subfamily A (NHE6, cation proton antiporter 6), member

6 SMC3 Structural maintenance of chromosomes 3 SMG7 smg-7 homolog, nonsense mediated mRNA decay factor (C. elegans) snoRNA Small nucleolar RNA SR140 U2 snRNP-associated SURP domain containing TCGA The Cancer Genome Atlas TLDA TaqMan Low Density Array TRBP Transactivating response RNA-binding protein TUNEL Terminal deoxynucleotidyl transferase-mediated dUTP nick end labeling TZ Transformation zone URR Upstream regulatory region UTR Untranslated region VEGF Vascular endothelial growth factor VEGFR Vascular endothelial growth factor receptor XRN1 5’-3’ exoribonuclease 1

1

CHAPTER 1: INTRODUCTION

2

1.1 MicroRNAs

1.1.1 Overview

MicroRNAs (miRNAs) are a class of short, non-coding RNAs that regulate gene

expression in a sequence-specific manner (1, 2). These single-stranded RNAs of ~20-22

nucleotides in length are endogenous in plants and animals, and show cross-species evolutionary

conservation (3). As of February 2013, there were 21,262 miRNAs catalogued in the miRBase

database, including 2042 human miRNAs (miRBase Release 19, www.mirbase.org). According

to bioinformatic analyses, up to 1/3 of all human genes are predicted to be regulated by miRNAs

(4), rendering them as one of the largest classes of gene regulators in the mammalian system (5).

1.1.2 History

miRNAs were first described in Caenorhabditis elegans (C elegans) in 1993 (6).

Ambros had previously observed that worms with mutations in the lin-4 and lin-14 genes did not

develop normally (7). Upon further characterization of these two genes, Lee et al. discovered

that the lin-4 gene was transcribed into a 22-nucleotide RNA (lin-4) that did not encode a protein

(6). Furthermore, lin-4 was found to be complementary to a repeated sequence element in the 3’

untranslated region (UTR) of lin-14 messenger RNA (mRNA), and could regulate protein

translation of lin-14 by binding to its 3’UTR.

The existence of miRNAs in higher organisms was first suggested in 2000 when the

second miRNA, let-7, was identified in C elegans (8). Let-7 had several known homologues in

higher organisms, including humans, and was reported to regulate lin-41 and lin-57/hbl-1 in C

elegans (6, 9, 10). Interestingly, the term “microRNA” was first used to describe this novel class

of small, non-coding RNAs in 2001, in three papers published simultaneously in volume

294(5543) of Science (11-13).

1.1.3 Biogenesis and Processing

The biogenesis and processing of miRNAs occurs via a complex pathway outlined in

Figure 1.1. They are initially transcribed inside the nucleus, usually by RNA polymerase II

(RNAP II) (14), to form large RNA precursors called primary miRNAs (pri-miRNAs) that can

be longer than 1 kb (14, 15). Some miRNAs are transcribed by RNA polymerase III (RNAP III),

3

including those surrounded by Alu elements (16). Pri-miRNAs contain at least one hairpin

structure, as well as a 3’ polyadenylated tail, and a 5’ 7-methylguanylate cap, similar to mRNA

transcripts (17).

While still inside the nucleus, the hairpin structure of the pri-miRNA is recognized and

bound by DiGeorge Syndrome Critical Region 8 (DGCR8), which recruits Drosha, an RNase III

endonuclease, to form the microprocessor complex (18, 19). Guided by DGCR8, Drosha cleaves

the pri-miRNA precisely 11 base pairs (bp) away from the base of the hairpin stem (20). This

cleavage yields a ~60-70 nt stem-loop precursor miRNA (pre-miRNA) with a 2-nt overhang at

the 3’ end, which is characteristic of products that have been cleaved by RNase III

endonucleases, and important for downstream steps in the processing pathway (21). Exportin-5

recognizes the pre-miRNA based on its stem-loop structure and 3’ overhang (22, 23), and binds

to the pre-miRNA with high affinity immediately after cleavage by Drosha (24). After binding,

Exportin-5 transports the pre-miRNA to the cytoplasm using Ran-GTP (25, 26).

In the cytoplasm, the pre-miRNA is further cleaved by Dicer, another RNase III

endonuclease (27). Dicer, aided by transactivating response RNA-binding protein (TRBP),

cleaves off the pre-miRNA terminal loop, yielding a ~22-nt double-stranded miRNA duplex

(28). Although Dicer alone is sufficient for this step, TRBP stabilizes Dicer and improves its

processing efficiency (29). Only one strand of the miRNA duplex, referred to as the guide

strand, is loaded onto the RNA-induced silencing complex (RISC), whereas the other strand,

referred to as the passenger strand, is subsequently degraded. The guide strand was traditionally

believed to be the strand with the less stable base pair at the 5’ end (30); however, it has been

recently reported that either strand could be incorporated into the RISC in some instances (31).

4

Figure 1.1 MicroRNA biogenesis and processing

miRNAs are initially transcribed inside the nucleus as large RNA precursors (pri-miRs), and then cleaved by the microprocessor complex into ~70 nt hairpin structures (pre-miRs). The pre-miRs

5

are exported to the cytoplasm where they are further cleaved by the RNase III enzyme Dicer into ~22 nt miRNA duplexes. The less stable of the two strands in the duplex, referred to as the “guide” strand, is incorporated into the RNA-induced silencing complex (RISC), while the opposite strand, referred to as the “passenger” strand, is destroyed. The miRNA-RISC complex can bind to target mRNAs with perfect complementarity, leading to mRNA degradation; or with imperfect complementarity, blocking protein translation. miRNA, microRNA; RNAP II, RNA polymerase II; pri-miRNA, primary microRNA; pre-miRNA, precursor microRNA; DGCR8, DiGeorge Syndrome Critical Region 8; TRBP, transactivating response RNA-binding protein; RISC, RNA-induced silencing complex.

6

1.1.4 RNA-induced Silencing Complex (RISC) Assembly and Function

The RISC is composed of the guide strand, Dicer, TRBP and Argonaute2 (AGO2) (32).

TRBP remains bound to the mature miRNA duplex after cleavage by Dicer and recruits it to the

AGO proteins (33). In humans, there are four Argonaute (AGO) proteins but only one member

of this family, AGO2, harbours endonuclease activity and is the catalytic unit of the RISC (34-

36). The RISC machinery binds via the guide strand to complementary target site(s) in the

3’UTR of the target mRNA (37), or in some cases, located in the 5’UTR (38), or even coding

regions (39). The seed region, a 6 to 8 nt sequence at the 5’ end of the guide strand, determines

the specificity of this binding (2, 40). Seed regions show a high degree of complementarity with

their corresponding binding sites in target mRNAs, and are highly conserved across species (41,

42).

miRNAs regulate their target mRNAs by one of two methods: destabilization of the

target mRNA, or translational repression. Destabilization of the target mRNA occurs when the

seed region of the guide strand has perfect complementarity with binding site(s) in the target

mRNA. AGO2 in the RISC initiates the mRNA degradation process by cleaving the target

mRNA via its endonuclease domain (43). The cleaved target mRNA is subsequently degraded

by 5’-3’ exoribonuclease 1 (XRN1) and the exosome (44).

When the seed region of the guide strand has partial complementarity with the binding

site(s) in the target mRNA, translational repression of the target mRNA occurs, either by

inhibition of translational initiation or repression of translational elongation. Inhibition of

translational initiation is a 5’ cap-dependent process (45). Normally, translational initiation

begins when the 5’ cap of the mRNA is recognized by eukaryotic translation initiation factor 4E

(eIF4E), a component of the translational initiation complex (46). It has been reported that

AGO2 from the RISC and eIF4E are similar in sequence, which allows AGO2 to displace eIF4E,

thereby inhibiting translational initiation (47). In another study, Chendrimada et al. showed that

miRNAs can also inhibit translational initiation by blocking ribsosome assembly (48). The

authors demonstrated that RISC associates with the 60S large ribosomal subunit and eIF6, which

is a ribosome inhibitory protein known to prevent the 40S and 60S ribosomal subunits from

assembling together to form the 80S ribosome. The evidence that translational repression can

occur during elongation came from polyribosome experiments with miRNAs, which showed that

7

active translation of the target was occurring even though the protein product could not be

detected (49-51). Unlike repression of translational initiation, the repression of translational

elongation is a 5’ cap-independent process (51).

1.1.5 Target Prediction Databases

With their ability to bind to target mRNAs with either perfect or partial complementarity,

each miRNA can potentially bind to numerous target mRNAs; conversely, each mRNA can

potentially be regulated by numerous miRNAs (52). Hence, one of the major challenges in

miRNA biology is to accurately identify mRNA targets that are truly regulated by any single

miRNA. Multiple publically-available in silico databases have been generated to assist in

identifying targets of miRNAs, using algorithms that incorporate criteria such as seed sequence

alignment (40), thermodynamic stability (53), conservation in different species (54), or binding

site accessibility (55). Some algorithms also consider the contribution of multiple binding sites

and their proximity to each other within the 3’UTR of the target mRNA (56, 57), or enrichment

of AU base-pairs in the 30 nt region flanking the binding site (58).

Currently there are over ten miRNA target prediction databases, each utilizing a unique

algorithm for determining putative target mRNAs, including: Targetscan (58), RNA22 (59),

PicTar (60), Diana-microT (61), miRDB (62), GenMir++ (63), miRanda (64, 65), TarBase (66),

PITA (55), and miRBase (4). These target prediction databases often provide hundreds of

candidate target mRNAs for any single miRNA; however, the algorithms used have been shown

to have high false discovery and high false negative rates (40, 67). Therefore, functional

experiments are necessary to validate putative target mRNAs that were determined using in

silico prediction databases. We used a tri-modal approach for miRNA target identification (68),

which combines: i) in silico target prediction; ii) clinical gene expression data from patients; and

iii) in vitro functional experiments. This tri-modal approach is described in further detail in

Chapter 2.

1.1.6 MicroRNAs in Cancer

It is not surprising that alterations in miRNA expression have been linked to most human

cancers, given the ability of miRNAs to regulate the expression of genes involved in important

cellular activities, including cell proliferation and apoptosis (69). Over-expression of oncogenic

miRNAs can contribute to cancer progression by down-regulating target genes that function as

8

tumour supressors. For example, miR-21, which is frequently over-expressed in many types of

solid tumours, targets the tumour suppressor phosphatase and tensin homolog (PTEN) (70), and

pro-apoptotic genes (71, 72). Conversely, under-expression of tumour suppressor-like miRNAs

can result in up-regulation of oncogenes. In chronic lymphocytic leukemia (CLL), miR-15a and

miR-16-1 are often down-regulated, and have been reported to target B Cell Lymphoma 2

(BCL2) (73), a well-known oncogene (74).

There are multiple mechanisms that can lead to dysregulation of miRNA expression,

including: epigenetic alterations (75-77), chromosomal aberrations (73, 78), and deregulated

expression of miRNA processing enzymes (79-81). Altered miRNA expression has been

attributed to epigenetic events for a number of human malignancies, which is unsurprising

considering bioinformatic analyses have demonstrated that 13 and 27% of human miRNA genes

are located within 3 and 10 kb of a CpG island, respectively (76). Aberrant CpG methylation can

lead to silencing of tumour supressor-like miRNAs, such as miR-376a and miR-376c in

melanoma (77). The authors of this study showed that both of these miRNAs were

downregulated in melanoma and directly targeted insulin-like growth factor receptor 1 (IGFR1),

which may contribute to IGFR1 overexpression in melanoma and thereby promote

tumourigenesis and metastasis. Another study reported that miR-127 was epigenetically silenced

in bladder cancer, and expression could be reactivated in vitro after treatment with 5-aza-2’-

deoxycytidine and 4-phenylbutyric acid, which inhibit DNA methylation and histone

deacetylase, respectively (82). Reactivation of miR-127 expression was accompanied by a

corresponding downregulation of its target B-cell CLL/lymphoma 6 (BCL6), which is a known

proto-oncogene.

Approximately half of miRNA genes are located in fragile chromosomal sites, which are

prone to chromosomal aberrations such as amplifications, deletions or translocations (83). For

example, the oncogenic miR-17~92 cluster is located in a region (13q31.3) that is frequently

amplified in several human cancers such as Burkitt’s lymphomas, follicular lymphomas, diffuse

large B-cell lymphomas, and lung cancer (84). As expected, overexpression of the miR-17~92

cluster, which encodes six miRNAs (miR-17, miR-18a, miR-19a, miR-20a, miR-19b-1, and

miR-92-1), is frequently observed in a number of human cancers including lymphomas and lung

cancer (85, 86). Conversely, the aforementioned tumour suppressor-like miR-15a and miR-16-1

9

are located in a region (13q14.3) that has been reported to be deleted in several hematopoietic

and solid malignancies (73, 74).

A number of studies have reported dysregulation of miRNA processing enzymes, either

up- or down-regulation, in various cancers, including colorectal (87, 88), skin (89, 90), renal

(91), and ovarian (92). In colorectal cancer, down-regulation of Dicer was significantly

associated with higher FIGO stage, tumour grade and nodal metastasis, and low Dicer expression

was correlated with shortened overall survival, independent of other prognostic factors (87).

Merritt et al. demonstrated that Dicer and Drosha were down-regulated in over 50% of the 111

ovarian cancer patients in their study (92). Furthermore, patients with low expression of both

Dicer and Drosha had worse overall survival, again independent of other prognostic factors.

1.1.7 MicroRNA Expression Profiling

miRNA expression profiling of human cancers has been used to distinguish between

different types of cancers (86), as well as categorize cancer sub-types (93, 94), including poorly-

differentiated tumours (95). Due to their small sizes, miRNAs are extremely stable and can be

readily extracted from cell lines and various types of clinical specimens, including frozen and

formalin-fixed paraffin embedded (FFPE) tissues, blood, serum, plasma, urine, and saliva (96-

102). The ability to utilize FFPE specimens for miRNA expression profiling is a powerful tool

for biomarker discovery, since FFPE processing is a universally standard procedure, and FFPE

specimens are often linked to clinical databases with extensive patient follow-up. In contrast,

there are many technical limitations associated with mRNA expression profiling using FFPE

specimens, since mRNAs are significantly longer than miRNAs; thus are easily degraded during

formalin fixation (103, 104).

Several reports have demonstrated highly concordant data when the same platform is

used to measure the expression of a panel of miRNAs in tissue-matched frozen and FFPE tissue

specimens; this has been documented for kidney, breast, prostate, lung, skin, glioblastoma and

melanoma specimens (96, 105-112). However, some of these studies have also shown poorly

correlated patterns of under- or over-expression for select miRNAs in the frozen vs. FFPE

specimens (105, 108, 109, 111, 112). For example, Leite et al. analyzed a panel of 14 miRNAs in

five prostate cancer specimens (5 frozen, 5 FFPE) and reported that four of the 14 miRNAs

demonstrated opposing patterns of tumour-normal fold changes in tissue-matched sample pairs

10

(112). Furthermore, Weng et al. reported that miRNAs with significant tumour-normal fold

changes were poorly correlated between tissue-matched frozen vs. FFPE renal carcinoma

samples; when the lists of dysregulated miRNAs were compared between matched samples, the

correlation peaked at 55% for the top 20 miRNAs (108).

Most methods of miRNA expression profiling utilize total RNA that contains small RNA

(<200 nt), and do not require separation or enrichment of small RNAs in the sample. Methods

such as TRIzol and the Ambion mirVana miRNA Isolation Kit use phenol:chloroform extraction

followed by alcohol precipitation. Other methods use column-based RNA extraction, including

the Norgen Total RNA Purification Kit and the Ambion RecoverAll Total Nucleic Acid Isolation

Kit. Furthermore, methods have been developed and optimized for extracting high-quality RNA

from FFPE specimens, such as the Ambion RecoverAll Total Nucleic Acid Isolation Kit, which

was shown to recover the most amplifiable RNA in a study that compared ten RNA extraction

methods for FFPE samples (113).

1.1.8 Platforms for MicroRNA Expression Profiling

Various platforms for global miRNA expression profiling have been developed,

including the Applied Biosystems TaqMan Low Density Array Human MicroRNA Array

(TLDA), NanoString nCounter Human miRNA Expression Assay (NanoString), and deep

sequencing. Individual quantitative real-time PCR (qRT-PCR), which many consider to be the

gold standard for nucleic acid quantification, is most often used to examine the expression of a

panel of selected miRNAs, since it is difficult to analyze a large number of miRNAs

simultaneously using this method. Due to their small sizes, miRNAs are not compatible with

standard PCR primers that require a minimum template size of 40 nt. miRNA-specific stem-loop

primers elongate the template RNA during the reverse transcription step, and are specific enough

to discriminate between mature miRNAs that differ by a single nucleotide (114). During the PCR

stage, the reverse transcription product is subjected to 40 cycles where TaqMan probes hybridize

with their target miRNA sequences. The raw data consists of the threshold cycle (Ct), which is

defined as the cycle number at which the fluorescence signal emitted passes the threshold.

Hence, there is an inverse relationship between Ct values and the amount of template RNA in the

11

sample; higher Ct values indicate lower levels of the template miRNA, whereas lower Ct values

indicate higher levels of the template miRNA.

The TLDA method simultaneously measures the expression of 377 miRNAs and four

endogenous controls using a microfluidic card. This method utilizes the same stem-loop primers

as individual qRT-PCR, in a multiplex reverse transcription step. The reverse transcription

product is then loaded onto the microfluidic card with 384 wells, which are each pre-loaded with

PCR primers specific for one target sequence (there is one well per miRNA, and four wells for

the U6 control). The same TaqMan chemistry as individual qRT-PCR is used to measure the

level of each miRNA, and the raw data is also outputted as Ct.

NanoString is a method that provides a direct count of miRNAs using optical

quantification of fluorescently-tagged miRNA molecules. First, the sample is prepared by

ligating the miRNA to a miRtag sequence. The miRtagged miRNA is then hybridized to a probe

pair consisting of a reporter probe, which carries the fluorescent signal/barcode, and a capture

probe, which allows the complex to be immobilized in the next step. Excess probes are removed

during the purification step and then probe-target miRNA complexes are immobilized. In the

final step, the colour barcode for each target miRNA is counted by the nCounter Digital

Analyzer. The raw data consists of a direct count of the number of target miRNA molecules, and

there is no bias due to amplification steps or enzymatic reactions.

Deep sequencing has been noted for its increased specificity and sensitivity, as well as its

ability to detect both known and novel miRNAs (115, 116). The first step of this method is the

reverse transcription of small RNAs to a cDNA library. Adapters are then ligated to the cDNAs,

which allows the library either to be affixed to beads for emulsion PCR (e.g. Applied Biosystems

SOLiD, Roche GS FLX), or a surface for solid-phase PCR (e.g. Illumina RNA-Seq). Next,

millions of individual cDNAs from the library are sequenced in parallel, with the number of

sequence reads for each small RNA species determined digitally. For each miRNA, the relative

quantification is provided, which is the number of sequence reads for a given miRNA relative to

the total reads in the sample.

12

1.1.9 microRNAs as Biomarkers

Biomarkers are used in the clinic to guide treatment decisions. A prognostic biomarker

indicates the likely clinical outcome of a patient with a particular disease, regardless of

treatment, whereas a predictive biomarker indicates the likely response of a patient to a particular

treatment. Established biomarkers include estrogen receptor (ER) and human epidermal growth

factor receptor 2 (HER2), which are both used for breast cancer patients. The absence or

presence of ER is used to determine whether or not a patient will respond to treatment with

tamoxifen, which targets the ER in breast tissue. HER2 is used to predict sensitivity to

trastuzumab, a monoclonal antibody against HER2. Patients with HER2 positive tumours, about

15% of all primary breast cancers, benefit significantly from trastuzumab therapy (117).

Despite the large quantity of reports describing prognostic factors for cancer, very few of

these prognostic factors have been developed as biomarkers for clinical use. Simon et al.

described the key steps required for the development and validation of therapeutically relevant

biomarkers for the clinic (118). Firstly, a biomarker should be developed using a training cohort

in which patients are adequately homogenous and are receiving the same treatment. Next, the

biomarker should be internally validated, and its prognostic value should be compared to

standard prognostic factors for the given disease. If the biomarker performs better than existing

standard prognostic factors, it should be translated to a platform that can be broadly used in the

clinic. The biomarker assay should be standardized and highly reproducible. Finally, the

biomarker should be independently validated with an external cohort of patients. The steps

required to develop and validate clinical biomarkers require robust experimental design and can

be very costly, which explains why so few prognostic factors have been developed into clinical

biomarkers.

Given their notable stability in clinical specimens, miRNAs could potentially be used as

biomarkers for cancer. Multiple reports to date have demonstrated a correlation between miRNA

expression and clinical outcome in cancer patients. The first reported miRNA signature was a

13-miRNA set associated with prognosis and progression in patients with CLL (119). Since

then, a number of prognostic miRNA signatures have been published for various cancers,

including cervical (120), colorectal (121), esophageal (122), gastric (123), hepatocellular (124),

lung (125-127), nasopharyngeal (128), prostate (129), and acute myeloid leukemia (AML) (130).

13

These studies highlight the prognostic value and potential translational impact of miRNA

expression profiling in cancer.

1.2 Cervical Cancer

1.2.1 Background

The cervix is covered by both mucin-secreting columnar epithelium, which line the

endocervix, and non-keratinising stratified squamous epithelium, which line the ectocervix.

These two types of epithelia meet at the squamocolumnar junction (SCJ). The original SCJ is the

location at which the endocervix and ectocervix meet at birth, while the new SCJ consists of SCJ

that is newly formed by the continuous remodeling process that occurs throughout the life of a

female. The transformation zone (TZ), which is the border between the old and new SCJ, is the

most common area where cervical cancer develops. The TZ undergoes metaplasia under normal

physiological conditions, which makes cells in the TZ more susceptible to malignant

transformation. Squamous cell carcinoma (SCC) and adenocarcinoma are the two most common

histological types, comprising 80-90% and 10-15% of all cervical cancers, respectively. SCC’s

arise from the squamous epithelial cells of the ectocervix, whereas adenocarcinomas arise from

the columnar epithelial cells of the endocervix. The International Federation of Gynecology and

Obstetrics (FIGO) system is used for staging cervical cancer, which is primarily based on clinical

examination.

Globally, cervical cancer is the third most common cancer amongst women, with 530,000

new cases and 275,000 deaths each year, making it the fourth leading cause of cancer mortality

(131). The vast majority of these cases and deaths (greater than 85%) occur in low income

countries in South America, Africa and South-Central Asia. In high income countries, such as

Canada, the United States and Australia, cervical cancer incidence and mortality rates have

fortunately declined significantly over the past few decades, primarily due to widespread

Papanicolaou (Pap) test screening programs that allow for early detection of pre-cancerous

lesions (132, 133). Although the incidence and mortality rates of cervical cancer in Canada have

been declining by 1.4% and 2.9% per year, respectively; it still remains as the 4th most common

cancer in women aged 20-29 years in this country (134).

14

1.2.2 Human Papillomavirus

Infection with human papillomavirus (HPV) has been identified as a necessary cause of

cervical cancer (135), with 99.7% of all cervical cancers being HPV-positive (136). The

connection between HPV and cervical cancer was first postulated in 1974 due to the

morphologic changes in the cancer lesions which were suggestive of a viral infection (137).

About a decade later, the causal relationship was established when HPV16 DNA was identified

and cloned in cervical cancer lesions by Professor Harold zur Hausen (138). Due to the rigour of

his science, and the impact of this discovery, which subsequently led to the development of the

HPV vaccine (139, 140), Professor zur Hausen was awarded the Nobel Prize in Physiology and

Medicine in 2008 (http://www.nobelprize.org/nobel_prizes/medicine/laureates/2008/). To date,

there are 15 high-risk HPV subtypes (HPV16, 18, 31, 33, 35, 39, 45, 51, 52, 56, 58, 59, 68, 73

and 82) that are considered carcinogenic (141), with HPV16 and 18 accounting for 70% of all

HPV-positive cases of cervical cancer (135).

HPV is a small, non-enveloped DNA virus with a double-stranded, circular genome of

approximately 8 kb (Figure 1.2). The HPV genome encodes six non-structural regulatory

proteins (E1, E2, E4, E5, E6 and E7) from the early region, and two structural viral capsid

proteins (L1 and L2) from the late region (142). The upstream regulatory region (URR) contains

the viral origin of replication, and encodes cis-acting elements that regulate viral gene

transcription and synthesis of viral DNA. E1 and E2 are involved in viral DNA replication; E1

encodes an ATP-dependent helicase, and E2 encodes a transcription factor. E4 facilitates viral

assembly and release, while E5 enhances cell immortalization.

The E6 and E7 viral oncoproteins mediate cervical cancer development by inactivating

the p53 and retinoblastoma protein (pRb) tumour suppressor proteins, respectively. P53 induces

cell cycle arrest upon recognizing DNA damage, and can initiate apoptosis if the damage is

irreparable. E6 ubiquitinates p53, which results in its degradation through the proteasomal

pathway. pRb binds to and inhibits E2F transcription factor, which prevents the cell from

replicating damaged DNA. However, when E7 competes for pRb, E2F is free to transactivate its

target genes, which facilitates the transition from G1 to S phase, thereby promoting cell cycle

progression.

15

Figure 1.2 Map of HPV genome The HPV genome, indicated by the circle, is a 7,857 bp double-stranded DNA genome. The coloured boxes outside the circle represent the open reading frames that encode the viral proteins. The upstream regulatory region (URR) contains the viral origin of replication, and encodes cis-acting elements that regulate viral gene transcription and synthesis of viral DNA.

1,000

5,000

3,000

2,000 4,000

6,000 7,857/1

7,000

HPV

E6

E7

E1

E4

E2

E5

L2

L1

16

1.2.3 Treatment

Cervical cancer patients with non-bulky, early-stage disease can be effectively treated

with surgery alone, achieving 5-year overall survival (OS) rates of ~80% (143). Unfortunately,

patients with high-risk, early-stage cervical cancer, such as those with lymph node metastases,

parametrial invasion, or deeply invasive lesions, fare much worse with surgery, with 5-year OS

rates ranging from 50-70% (144-148).

Radiotherapy (RT) has been used to treat cervical cancer patients with advanced-stage

disease since the 1930’s (149). Since then, many improvements have been made to RT, and

chemotherapy has been added to the standard treatment regimen, based on a recommendation

made by the National Cancer Institute in 1999. This decision was made after five major

randomized clinical trials had provided consistent evidence that RT combined with concurrent

chemotherapy was superior to RT alone, in terms of patient survival (150-154). Hence, current

standard therapy for advanced stages of cervical cancer consists of concurrent chemo-RT (CRT)

using cisplatin. Typically, external-beam RT is applied to the primary cervical tumour and pelvic

lymph nodes (45 to 50 (Gy) total, in 1.8-to-2.0 Gy daily fractions with 18-to-25-MV photons),

combined with weekly doses of cisplatin (40 mg/m2 total, 5 doses).

1.2.4 Gene Expression Signatures for Cervical Cancer

Prognostic gene signature development for cervical cancer has been slow compared to

other types of cancers, likely due to the low incidence of this disease in high income countries

with research capabilities. To date, there have been four gene signatures reported for cervical

cancer, but none of them have been validated in an independent patient cohort. The earliest

reported gene signature was a 58-gene (mRNA) set associated with recurrence in cervical cancer

patients treated with CRT (155). The second was a 7-gene (mRNA) set to predict pelvic relapse,

derived from a group of patients participating in a Phase II clinical trial evaluating the benefit of

a COX-2 inhibitor Celecoxib in addition to CRT using cisplatin (156). A third signature

consisted of a 7-gene set used to identify patients who were likely to respond to RT alone; hence

could be spared from the toxicity induced by the addition of chemotherapy (157). Another

unvalidated signature was a 63-gene set associated with metastasis of advanced tumours

following treatment with RT (158). The clinical relevance of these gene signatures remains

unknown, since none have been validated in an independent cohort. Furthermore, all of these

17

studies used either frozen tissues (156, 158) or fresh tissues that were immediately stored in

RNAlater stabilization solution upon tissue collection (155, 157); neither of which are routine

practice or cost-efficient compared to FFPE tissues.

1.2.5 MicroRNA Expression in Cervical Cancer

The first report to examine miRNA expression in cervical cancer used direct sequencing

to compare global miRNA expression in six cervical cancer cell lines and five normal cervix

tissue samples, and identified six differentially expressed miRNAs (one upregulated: miR-21;

five downregulated: miR-143, miR-196b, let-7b, let-7c, miR-23b) (159). When the expression

of two of these dysregulated miRNAs, miR-21 and miR-143, were examined in 29 matched pairs

of cervical cancer and normal tissues, it was observed that the expression variation in the cell

lines was reproduced in the tissue samples. The second published study identified 70 miRNAs

that were differentially expressed in ten early-stage invasive squamous cell carcinomas (SCC’s)

compared to ten normal cervical epithelial samples (160). Furthermore, up regulation of miR-

127 was found to be significantly associated with lymph node metastases and was validated in an

independent group of 31 patient samples.

Using age-matched cervical cancer and normal cervix tissue samples, Wang et al.

reported significant down regulation of miR-126, miR-143 and miR-145, along with up

regulation of miR-15b, miR-16, miR-146a and miR-155 (161). Functional experiments in

cervical cancer cell lines demonstrated that miR-146a promoted cell proliferation, whereas miR-

143 and miR-145 inhibited cell growth. A study by Pereira et al. examined miRNA expression

in four SCC’s, five high-grade squamous intraepithelial lesions, nine low-grade squamous

epithelial lesions, and 19 normal cervix tissues, but could not derive a miRNA signature for

cervical cancer due to high variability in miRNA expression, particularly among the normal

cervix samples (162). Hu et al. reported a 2-miRNA signature (miR-200a and miR-9) that was

associated with overall survival (OS) (120). This was the only study that had an adequate

number of patient samples to develop a prognostic miRNA signature (training set: n=60;

validation set: n=40).

A recent report described five dysregulated miRNAs that were directly linked with

frequent chromosomal alterations in cervical cancer, namely miR-9 (1q23.2), miR-15b

(3q25.32), miR-28-5p (3q27.3), miR-100 and miR-125b (both 11q24.1) (163). Functional

18

experiments performed in vitro demonstrated that overexpression of miR-9, which is linked to a

chromosomal gain of 1q, promoted cell viability, anchorage-independent growth, and cell

migration, supporting the oncogenic role of miR-9. Another recent study identified nine down

regulated (miR-211, miR-145, miR-223, miR-150, miR-142-5p, miR-328, miR-195, miR-199b,

and miR-142-3p), and two up regulated (miR-182 and miR-183) miRNAs in cervical cancer cell

lines (164). The most up regulated miRNA, miR-182, was characterized further and shown to be

significantly upregulated in cervical cancer tissues, correlating with advanced stage of disease.

In vitro experiments demonstrated that miR-182 was involved in apoptosis and associated with

forkhead box O1 (FOXO1) regulation, and in vivo experiments confirmed that inhibiting miR-

182 could reduce tumour growth.

1.3 Homeobox Genes

1.3.1 Overview

The Homeobox genes, which were first reported in 1984, are a family of regulatory genes

that encode transcription factors that play important roles in directing the formation and

patterning of body structures during embryonic development (165, 166). Homeobox genes

contain a 180 bp homeobox sequence encoding a highly conserved 60-residue helix-turn-helix

protein domain, called the homeodomain (166). The homeodomain is a DNA-binding motif that

recognizes and binds to DNA elements in a sequence-specific manner (167). Homeobox genes

are contained in genomic clusters, and are expressed in a spatial and temporal manner (168).

Human class I human homeobox (HOX) genes consist of 39 genes organized in four

clusters (HOXA, HOXB, HOXC and HOXD) located on different chromosomes, 7p, 17q, 12q,

and 2q, respectively (Figure 1.3). Within each cluster, the HOX genes are assigned to 13 paralog

groups, with each cluster containing 9 to 11 HOX genes.

19

Figure 1.3 Human HOX gene clusters Schematic representation of the 39 human HOX genes assigned to 13 paralog groups in four HOX clusters (A-D) located on different chromosomes (7p, 17q, 12q, 2q). During embryonic development, the HOX genes are expressed sequentially from the 3’ to 5’, with the early genes (shown in orange/yellow) being expressed in anterior regions, and late genes (shown in blue) being expressed in posterior regions.

1 2 3 4 5 6 7 8 9 10 11 12

HOX A

HOX B

HOX C

HOX D

3’ 5’

20

1.3.2 Role in Development

During embryogenesis, HOX genes are fundamentally important for proper development

along the anterior-posterior axis (169). The HOX genes are expressed sequentially from the 3’ to

the 5’ end of each cluster, with 3’ end genes, such as HOXB1, expressed early in development in

anterior regions, and 5’ end genes, such as HOXB13, expressed later in development in the

posterior regions (170). HOX genes have also been shown to play a role in directing patterns

development in limbs (171, 172), lungs (173-175), and genitals (176, 177).

1.3.3 Role in Cancer

The expression of HOX genes in normal adult tissues indicates a role for these genes

beyond embryonic development (178). Takahasi et al. reported that HOX genes were expressed

in adult organs in a highly specific manner, and deregulated HOX gene expression was

associated with tumour development (178). Given that HOX genes have been shown to play

important roles in normal hematopoiesis (179-181), it is thus not surprising that aberrant HOX

gene expression has been linked to leukemias (182, 183). Ferrando et al. reported consistent up

regulation of HOXA9, HOXA10 and HOXC6 in acute lymphoblastic leukemia (ALL) patients

with chromosomal translocations involving the mixed lineage leukemia (MLL) gene (184). In a

microarray study of 6817 genes tested in patients with acute myeloid leukemia (AML), HOXA9

was identified as a diagnostic marker, and was the only gene that was strongly correlated with

clinical outcome (185). Furthermore, inhibition of HOXA9 was shown to reduce proliferation

and induce apoptosis in AML, especially in samples with MLL translocations (186).

Aberrant HOX gene expression has also been reported in solid tumours. Three

mechanisms of HOX gene deregulation in cancer have been described by Abate-Shen (187). In

the first mechanism, which is responsible for the majority of cases of HOX gene deregulation,

HOX genes that are normally only expressed during embryogenesis are re-expressed in tumour

cells. Secondly, HOX genes can be aberrantly expressed in tumour cells derived from tissues in

which the gene is not normally expressed. The third and final mechanism involves down

regulation of HOX genes in tumour cells derived from tissues in which the gene is normally

expressed.

21

In prostate cancer, HOXC4, HOXC5, HOXC6 and HOXC8 were found to be

overexpressed in tumours and nodal metastases (188). Interestingly, HOXC8 overexpression has

been linked with loss of differentiation in prostate cancer (189), as well as increased cell invasion

in vitro (190). HOXA5 under expression in breast cancer was associated with loss of p53

transcript and protein levels (191). In renal carcinoma, HOXB5 and HOXB9 have been reported

to be down regulated, whereas HOXB2 was up regulated compared to normal kidney tissues

(192).

1.4 Vascular Endothelial Growth Factor

1.4.1 Overview

The Vascular endothelial growth factor (VEGF) family proteins are secreted platelet-

derived growth factors (193). There are five members of the VEGF family: VEGF-A, VEGF-B,

VEGF-C, VEGF-D and placenta growth factor (PlGF) (Figure 1.4) (194). VEGF-A was the first

to be discovered over 20 years ago, and is the most well-characterized family member (195). It

was identified as a key regulator of signaling pathways involved in vasculogenesis and

angiogenesis of endothelial cells (196). VEGF-A also induces chemotaxis in various types of

cells (197-199), and stimulates vasodilation mediated via nitric oxide release from endothelial

cells (200). VEGF-B shares high sequence homology with VEGF-A, and was initially described

to be an angiogenic factor (201). However, more recent studies have shown that VEGF-B can

both positively and negatively regulate angiogenesis, depending on the conditions (202). The

role of VEGF-B remains unclear at this time; VEGF-C and VEGF-D appear both to be involved

in angiogenesis and lymphangiogenesis (203, 204).

1.4.2 Receptors

The VEGF family members bind to and activate three VEGF receptors (VEGFR’s):

VEGFR-1 (FLT1), VEGFR-2 (KDR), and VEGFR-3 (FLT4) (205, 206). Each VEGF ligand has

distinct binding specificities for each of the VEGFRs, which contributes to their diverse

functions (Figure 1.4). The VEGFRs are receptor tyrosine kinases with an extracellular portion

composed of seven immunoglobulin (Ig)-like domains, a single membrane-spanning region, and

an intracellular portion with a conserved intracellular tyrosine kinase domain containing a kinase

insert (207, 208). FLT1 has the highest affinity for VEGF and can also bind VEGF-B and PlGF.

This receptor is expressed on monocytes, macrophages and hematopoietic stem cells. VEGFR-2

22

is the most widely expressed receptor and also has a high affinity for VEGF. It is mainly

responsible for mediating mitogenesis and angiogenesis. VEGFR-3 is mainly expressed on the

lymphatic endothelium, and is involved in lymphangiogenesis.

1.4.3 Isoforms

VEGF-A, herein referred to as VEGF, undergoes alternative splicing to yield different

isoforms which are denoted as VEGFxxx, where xxx represents the number of residues present in

the protein, excluding the signaling peptide. To date, eight VEGF isoforms have been identified

in humans, ranging in length from 121 to 206 residues (209). The most common and well-

studied isoforms are VEGF121, VEGF165 and VEGF189, which are expressed in most VEGF-

producing cells (210). The five minor isoforms (VEGF111, VEGF145, VEGF148, VEGF183 and

VEGF206) are less common, and their functions are not clearly understood (211-216).

1.4.4 VEGF in Cervical Cancer

VEGF levels in patients undergoing radiotherapy for cervical cancer, as measured in

serum by ELISA (217, 218), have been shown to be a prognostic marker of disease-free survival

(DFS), whereby high pre-treatment VEGF levels are associated with worse DFS. Bevacizumab

is a humanized anti-VEGF monoclonal antibody (A.4.6.1) that recognizes all biologically active

isoforms of VEGF and prevents them from binding to their receptors. This drug has been shown

to be effective in treating several malignancies by blocking angiogenesis (219). Thus far,

bevacizumab has been approved by the U.S. Food and Drug Administration (FDA) for treatment

of metastatic colorectal cancer, non-small cell lung cancer, glioblastoma, and metastatic kidney

cancer. Recently, the National Institutes of Health announced the results of a multi-centred

randomized Phase III clinical trial (GOG240), where Bevacizumab was evaluated in patients

with advanced, recurrent or persistent cervical cancer who could not be cured with standard

treatment (220). The preliminary data from this trial reported that the addition of Bevacizumab

to chemotherapy with cisplatin plus paclitaxel significantly improved OS by a median of 3.7

months.

23

Figure 1.4 VEGF family members and receptors The five VEGF family members (VEGF-A, VEGF-B, VEGF-C, VEGF-D, and PlGF) and the three principal VEGF receptors (VEGFR-1, VEGFR-2, VEGFR-3). VEGF-A binds to both VEGFR-1 and VEGFR-2. VEGF-B and PlGF bind to only VEGFR-1. VEGF-C and VEGF-D bind to both VEGFR-2 and VEGFR-3.

VEGFR‐1 VEGFR‐2 VEGFR‐3

S‐S S‐S

Cytosol

VEGF‐A VEGF‐D VEGF‐C PlGF VEGF‐B

24

1.5 Research Objectives

Novel insights are required, to develop new therapeutic approaches that could improve

clinical outcome for patients with advanced stages of cervical cancer. Altered miRNA expression

has been shown to play a significant role in many human malignancies (69), including cervical

cancer (120, 159, 161, 162), and target genes that contribute to tumor progression have been

described for several miRNAs (68, 221, 222). The overarching objective of this thesis is to

investigate and characterize the role of altered miRNA expression in cervical cancer, and to

provide analyses of the currently available methods for miRNA expression profiling.

Firstly, global miRNA expression profiling will be performed using cervical cancer cell

lines and patient specimens, to identify differentially-expressed miRNAs in cervical cancer.

From the list of differentially-expressed miRNAs, we will select miRNA(s) of interest and

identify target genes that contribute to tumour progression; we will characterize the selected

miRNA(s) and their target genes(s) in vitro and in vivo, using various assays that assess colony

formation, cell migration and invasion, cell viability, and tumour formation (Chapter 2). The

global miRNA expression data from the patient specimens (n = 79) will also be used to develop a

prognostic miRNA signature, which we will attempt to validate with an independent cohort of

patients (n = 87) (Chapter 3).

Analyses of methods for genome-wide miRNA and mRNA expression studies will be

conducted. We will examine the effect of amplification protocols used for microarray studies,

and determine if biases can be corrected computationally (Chapter 4). Furthermore, two

platforms for measuring global miRNA expression will be evaluated: the TaqMan Low Density

Array and the NanoString nCounter assay; we will also compare these two platforms with qRT-

PCR, which is considered to be the gold-standard for nucleic acid quantification (Chapter 5).

Overall, this thesis will provide insights into the important role of miRNAs in cervical

cancer, as well as an evaluation of technical considerations associated with genome-wide

expression profiling studies. The findings from this thesis will potentially guide future studies

and help improve therapeutic approaches for cervical cancer patients.

25

CHAPTER 2: MICRORNA-196B REGULATES THE HOMEOBOX-B7/VASCULAR ENDOTHELIAL GROWTH

FACTOR AXIS IN CERVICAL CANCER

The data in this chapter have been published in PLoS ONE:

How C, Hui ABY, Alajez NM, Shi W, Boutros PC, Clarke BA, Yan R, Pintilie M, Fyles A, Hedley DW, Hill RP, Milosevic M, Liu FF. (2013) MicroRNA-196b Regulates the HomeoboxB7-Vascular Endothelial Growth Factor Axis in Cervical Cancer. PLoS ONE 8(7): e67846.

26

2.1 Chapter Abstract

The down-regulation of microRNA-196b (miR-196b) has been reported, but its

contribution to cervical cancer progression remains to be investigated. In this study, we first

demonstrated that miR-196b down-regulation was significantly associated with worse disease-

free survival (DFS) for cervical cancer patients treated with combined chemo-radiation.

Secondly, using a tri-modal approach for target identification, we determined that homeobox-B7

(HOXB7) was a bona fide target for miR-196b, and in turn, vascular endothelial growth factor

(VEGF) was a downstream transcript regulated by HOXB7. Reconstitution of miR-196b

expression by transient transfection resulted in reduced cell growth, clonogenicity, migration and

invasion in vitro, as well as reduced tumor angiogenesis and tumor cell proliferation in vivo.

Concordantly, siRNA knockdown of HOXB7 or VEGF phenocopied the biological effects of

miR-196b over-expression. Our findings have demonstrated that the miR-196b/HOXB7/VEGF

pathway plays an important role in cervical cancer progression; hence targeting this pathway

could be a promising therapeutic strategy for the future management of this disease.

2.2 Introduction

Worldwide, cervical cancer is the third most frequently diagnosed malignancy and the

fourth leading cause of cancer mortality in women, with an estimated 530,000 new cases and

275,000 deaths each year (131). Although cervical cancer incidence and mortality rates have

declined over the past thirty years in the United States (223), the 5-year survival rate has

remained below 40% for patients diagnosed with Stage III or IV disease (224). Novel insights

are required to better understand the mechanisms that contribute to disease progression, in order

to design improved therapies for patients with locally advanced cervical cancer.

Micro-RNAs (miRNAs) are short, non-coding RNAs that regulate gene expression post-

transcriptionally (1, 2), and aberrant miRNA expression has been shown to be important in many

human malignancies (69). Gene targets that contribute to tumor progression have been described

for several miRNAs (68, 221, 222); however, the biological function of the majority of miRNAs

still remains unknown. One of the major challenges to miRNA target identification is the ability

of miRNAs to bind mRNA targets with imperfect complementarity; hence a single miRNA can

potentially regulate several hundreds or thousands of genes (52). Unfortunately, the currently

27

available in silico miRNA target prediction algorithms have high false-discovery and false-

negative rates (40, 67); thereby mandating experimental validation of miRNA targets.

Down-regulation of miR-196b in cervical cancer has been previously reported (159), but

its role in tumor progression in this disease has not been previously investigated. Herein, we

report down-regulation of miR-196b in primary human cervical cancer tissues and cell lines.

Furthermore, we identified the HOXB7 transcription factor as a novel, direct and specific target

of miR-196b, which in turn, regulates VEGF in cervical cancer. Most importantly, miR-196b

down-regulation was associated with worse DFS in patients treated with chemo-radiation,

highlighting the biological importance of miR-196b in cervical cancer progression.

2.3 Materials and Methods

2.3.1 Cell Lines and Transfections

Human cervical cancer cell lines (ME-180, SiHa, and HT-3) were obtained from

American Type Culture Collection (ATTC), and grown in α-MEM supplemented with 10%

FBS at 37°C, 5% CO2. All cells were authenticated every six months at the Centre for Applied

Genomics (Hospital for Sick Children, Toronto, Canada) using the AmpF/STR Identifier PCR

Amplification Kit (Applied Biosystems), and determined to be free from Mycoplasma

contamination using the MycoAlert Mycoplasma Detection Kit (Lonza). ME-180 and SiHa cells

were transfected using the LipofectAMINE 2000 (Invitrogen) forward transfection protocol,

according to the manufacturer’s instructions. Pre-miR Negative Control #1 (NC), Pre-miR-196b

(Ambion), All Stars Negative Control (siNEG), siHOXB7 and siVEGF (Qiagen) were all

transfected at a final concentration of 30 nmol/L.

2.3.2 miRNA Expression Profiling of Cell Lines

Total RNA was isolated from SiHa, ME-180 and HT3 cervical cancer cell lines using the

mirVana miRNA Isolation Kit (Ambion) according to the manufacturer’s instructions.

FirstChoice® Total RNA: Human Normal Cervix Tissue (Ambion) from 3 different tissue

donors was utilized as normal comparators. Expression levels of 377 miRNAs and 3 snoRNAs

(controls) were assayed in the cervical cancer cell lines and normal cervix tissues using the

TaqMan® Low Density Array (TLDA) Human MicroRNA Panel (Applied Biosystems), with the

Applied Biosystems 7900HT Real-Time PCR System, as we have previously described (96).

28

2.3.3 miRNA Expression Profiling of Patient Tissues

Flash-frozen punch biopsies were obtained from patients with locally advanced cervical

cancer who were planned to receive primary treatment with standard chemo-radiation, consisting

of external-beam radiotherapy to the primary cervical tumor and pelvic lymph nodes (45 to 50

Gy total, in 1.8-to-2-Gy daily fractions with 18-to-25-MV photons), combined with weekly

doses of cisplatin (40 mg/m2 total, 5 doses). FIGO (International Federation of Gynecologists

and Obstetricians) staging was determined using a combination of: pretreatment evaluation under

anesthesia, computed tomography (CT) scans of the abdomen and pelvis, chest x-ray, and

magnetic resonance imaging (MRI) of the pelvis. MRI was also used to determine lymph node

status; pelvic and para-aortic lymph nodes were classified as positive for metastatic disease if the

MRI short-axis dimension was >1 cm and equivocal if it was 8 to 10 mm. After biopsy, the

specimens were placed in optimal cutting temperature (OCT) storage medium for

histopathologic examination, then flash-frozen in liquid nitrogen. H&E-stained tissue sections

were cut from the OCT-embedded material and evaluated by a gynecologic pathologist (B

Clarke). Total cell content (stroma and tumor cells) was estimated for all tissue samples using a

light microscope, and only samples containing >70% tumor cells were considered for further

analysis (n=79). The clinical characteristics of these 79 patients are provided in Table 2.1. The

median follow-up time for this cohort was 3 years. Flash-frozen normal cervix tissues obtained

from 11 patients who underwent hysterectomy for benign causes served as normal comparators.

Two sections of 50-micron thickness were cut from the OCT-embedded flash-frozen

tissues and placed in a nuclease-free microtube. Total RNA was isolated using the Norgen Total

RNA Purification Kit (Norgen Biotek), according to the manufacturer’s instructions. Global

miRNA expression was measured in the cervical cancer and normal cervix tissues with the

TaqMan® Low Density Array (TLDA) Human MicroRNA A Array v2.0 (Applied Biosystems)

using the Applied Biosystems 7900HT Real-Time PCR System, as already described (96).

29

Table 2.1 Clinical parameters of 79 cervical cancer patients.

Category Subcategory N = 79 Percent Age, years Median 48

Range 26-84

Tumour size ≤5 cm 48 61%

>5 cm 31 39%

FIGO stage IB 24 30% IIA 2 3% IIB 35 44% IIIA 0 0%

IIIB 18 23%

Histology Squamous 74 94% Adenocarcinoma 4 5%

Adenosquamous 1 1%

Pelvic node involvement Positive 25 32% Equivocal 15 19% Negative 39 49%

2.3.4 Quantitative Real-time PCR Analysis of miRNAs and mRNAs

Total RNA was isolated from the cell lines using the Total RNA Purification Kit (Norgen

Biotek), according to the manufacturer’s instructions. The expression of miR-196b was

measured by quantitative real-time polymerase chain reaction (qRT-PCR) using the standard

TaqMan MicroRNA Assay (Applied Biosystems). Briefly, RNA was first reverse transcribed

using the TaqMan MicroRNA Reverse Transcription Kit and a stem-loop primer specific for

miR-196b (Applied Biosystems) (114). The 2-ΔΔCt method was used to calculate relative levels

of miR-196b expression, using RNU44 as a reference gene (225).

The RT products were amplified with miR-196b-specific primers, as we previously

described (96). The expression levels of previously-described (c-myc, BCL2, HOXA9, MEIS1),

and candidate mRNA targets for miR-196b (ANKHD1, CTDSP2, FGFR1, HDAC, HOXA7,

HOXB7, KRT8, PUM2, SLC9A6, SMC3, SMG7, SR140), and HOXB7 (VEGF, Ku70, Ku80,

DNA-PK, FGF2, MMP2, WNT5a, PDGFA, THBS2) were also measured by qRT-PCR. One

30

microgram of total RNA was reverse-transcribed using SuperScript II Reverse Transcriptase

(Invitrogen) according to the manufacturer’s instructions. Quantitative RT-PCR was performed

using SYBR Green PCR Master Mix (Applied Biosystems) and PCR primers (Table 2.2)

designed using Primer 3 Input. The 2-ΔΔCt method was used to calculate relative levels of gene

expression, with GAPDH as a reference gene (225).

Table 2.2 Primer Sequences used for qRT-PCR

Gene Forward Primer Reverse Primer

HOXB7 5' GTTGTTACTAGTCCAGCTCT GGGAACTGAATC 3'

5' GTTGTTAAGCTTCCACAGTCCACAAAGACAGC 3'

KRT8 5' CAAAGGCCAGAGGGCTTC 3' 5' GACGTTCATCAGCTCCTGGT 3' HDAC 5' CCTGGACCACCAGTTCTCAC 3' 5' CCTGTTGTTGCTTGATGTGC 3' PUM2 5' GGGAGCTTCTCACCATTCAA 3' 5' CCCTTCCTAAATCGAGAATCC 3' SMG7 5' GATTTACCATGCCGTGTGAA 3' 5' TTCTGTATCGAGCAATGTCTCC 3' SR140 5' CGGACGGAAGATTTTCGTAT 3' 5' CCCTCTGTTCTTCCTTAAGTGC 3' ANKHD1 5' CGAAGTGTCCGAGGTTGAAT 3' 5' TGCTTCTAGTCGTGCCTGTG3' SMC3 5' AAAGAAACAGAGGGCAAACG 3' 5' CATCAAGTTTGGCACGAGTC 3' CTDSP2 5' AAGCAAGGCCTGGTCTCC 3' 5' GCAATGGTGTTTGCTTCCTC 3' FGFR1 5' CGATGTGCAGAGCATCAACT 3' 5' AGGGGAGAGCATCTGAAACA 3' HOXA7 5' CAGTGACCTCGCCAAAGG 3' 5' AGGTCCTGAAGACCGCATC 3' SLC9A6 5' CATTTGCCTTGGCCATTC 3' 5' AATCAACACCAACCCTGATATG 3'Ku70 5' CGACAGGTGTTTGCTGAGAA 3' 5' CCTGGTTGGATTTTGCTTTC 3' Ku80 5' AATCCAGGTGCAAAACGAAT 3' 5' GGGGTTGTCTTCATTGGTGA 3' DNA-PK 5' TGCAAGGTTATAAACGCAAAA 3' 5' ACCTCAGGAACTGTGTCAAGG 3' VEGF 5' GCTACTGCCATCCAATCGAG 3' 5' ATCCGCATAATCTGCATGGT 3' HOXA9 5' CCACGCTTGACACTCACACT 3' 5' GGGTTATTGGGATCGATGG 3' BCL-2 5' GAGGATTGTGGCCTTCTTTG 3' 5' CATCCCAGCCTCCGTTATC 3' MEIS1 5' TGATGGCTTGGACAACAGTG 3' 5' AGGGTGTGTTAGATGCTGGA 3' MYC 5' TAGTGGAAAACCAGCAGCCT 3' 5' GTTCACCATGTCTCCTCCCA 3' FGF2 5' CTGGCTTCTAAATGTGTTACGG 3' 5' CCCAGGTCCTGTTTTGGAT 3' MMP2 5' TGGGCAACAAATATGAGAGAG 3' 5' CGGCATCCAGGTTATCGGGG 3' WNT5a 5' ATTCTTGGTGGTCGCTAGGT 3' 5' TGTACTGCATGTGGTCCTGA 3' PDGFA 5' CATGTTCTGGCCGAGGAAG 3' 5' AGTCTATCTCCAGGAGTCGC 3' THBS2 5' ACTCGCAGTGGAAGAACGTC 3' 5' GAAGCAAACCCCTGAAGTGA 3'

31

2.3.5 Cell Viability, Proliferation and Colony-Formation Assays

The viability of transfected cells was assessed by the Trypan blue exclusion assay. ME-

180 and SiHa cells were transfected in triplicate with 30 nmol/L of pre-miR-196b, NC,

siHOXB7, siVEGF, or siNEG and incubated at 37°C, 5% CO2. At 48 and 72 hours post-

transfection, cells were trypsinized, stained with Trypan blue and counted using a

hemocytometer. Cell proliferation was examined using the CellTiter 96 Non-Radioactive Cell

Proliferation Assay (MTS Assay) (Promega BioSciences), according to the manufacturer’s

instructions. For colony formation assays, cells were transfected with 30 nmol/L of pre-miR-

196b, Pre-miR Negative Control #1, siHOXB7, siVEGF, or siNEG and incubated at 37°C, 5%

CO2. At 48 hours post-transfection, cells were re-seeded at low density in 6-well plates in

triplicate. Cells were incubated at 37°C, 5% CO2 for 10-12 days, then fixed and stained with

0.1% crystal violet in 50% methanol. The number of colonies containing at least 50 cells was

counted, and the surviving fraction was calculated by comparison with cells transfected with

Negative Control.

2.3.6 Cell Migration and Invasion Assays

BD BioCoat Matrigel Invasion Chambers and Control Inserts (BD Biosciences) were

used to assay migration and invasion of transfected cells. Chambers contained a polyethylene

terephthalate membrane with 8 µm pores. ME-180 and SiHa cells were transfected with 30

nmol/L of pre-miR-196b, NC, siHOXB7, siVEGF, or siNEG and incubated at 37°C, 5% CO2.

At 24 hours after transfection, 1.5 x 105 cells were re-seeded inside each chamber with medium

containing low serum (1% FBS). The chambers were placed in a 24-well plate, with high serum

(20% FBS) medium in each lower chamber to serve as a chemo-attractant. Cells were incubated

at 37°C, 5% CO2 for 48 hours, then the membranes were washed, stained, and mounted onto

slides. A light microscope was used to count the number of migrating or invading cells. Relative

migration was calculated by comparison with cells transfected with the negative control. Percent

invasion was calculated as the number of cells that invaded through the Matrigel insert, divided

by the number of cells that migrated through the control insert.

2.3.7 Cell Cycle Analysis

Cell cycle analysis was performed on ME-180 and SiHa cells after transfection with 30

nmol/L of pre-miR-196b or NC, to measure the fraction of cells in the sub-G1 phase of the cell

32

cycle. Cells were harvested and washed twice in FACS buffer (PBS/0.5% BSA), re-suspended in

1 mL of FACS buffer, then fixed in 1 mL of ice-cold 70% ethanol. After 1h of incubation on ice,

cells were washed again and re-suspended in 500 uL of FACS buffer containing 40 µg/mL

RNase A (Sigma) and 50 µg/mL propidium iodide, then incubated in the dark at room

temperature for 30 minutes. Cells were analyzed in the BD FACScalibur (Becton Dickinson)

using the FL-2A and FL-2W channels. The flow cytometry data were analyzed using FlowJo 7.5

software (Tree Star).

2.3.8 In vivo Experiments

Six- to 8-week-old severe combined immunodeficient (SCID) female mice were utilized

for xenograft experiments, according to guidelines of the Animal Care Committee, Ontario

Cancer Institute, University Health Network. Cells were transfected with pre-miR-196b or NC

and incubated at 37°C, 5% CO2. At 48 hours post-transfection, cells were harvested and 5 x 105

viable cells were diluted in 100 µL of growth medium. Cells were injected intramuscularly into

the left gastrocnemius muscle of female SCID mice. Tumor plus leg diameter was measured

twice a week and mice were euthanized when 15 mm was attained. Tumors were removed at 25

days after implantation and immediately fixed in 10% buffered formalin for 24 h, placed in 70%

ethanol for 24 h, embedded in paraffin, and sectioned (5 µM) for immunostaining. In addition to

hematoxylin and eosin (H&E) staining, cluster of differentiation 31 (CD31) was used for

assessing tumor angiogenesis, Ki-67 for tumor cell proliferation, and terminal deoxynucleotidyl

transferase-mediated dUTP nick end labeling (TUNEL) for apoptosis.

2.3.9 Tri-modal Approach for miRNA Target Identification

Candidate mRNA targets of miR-196b were determined using a previously-described tri-

modal approach (68) combining: i) all predicted targets of miR-196b from five in silico miRNA

target prediction databases (TargetScan, PicTar, GenMir++, miRBase, miRDB); ii) mRNAs up-

regulated at least 2-fold in cervical cancer tissues compared to normal cervix tissues, using two

independent publicly-available microarray datasets (226, 227); and iii) mRNAs down-regulated

at least 0.5-fold at both 24 and 72 hours after transfection with 30 nmol/L of pre-miR-196b,

where transcript levels were measured using the Whole Human Genome 4 x 44K One-Color

Array (Agilent).

33

2.3.10 Luciferase Reporter Assay

Wildtype or mutant fragments of the 3’-untranslated region (UTR) of HOXB7 containing

the predicted binding site (position 220-226) for miR-196b were individually amplified by

AmpliTaq Gold DNA Polymerase (Applied Biosystems) using the primers listed in Table S1.

The PCR products were purified, then cloned downstream of the firefly luciferase gene in the

pMIR-REPORT vector (Ambion) at the SpeI and HindIII restriction sites, to produce the pMIR-

HOXB7 or pMIR-HOXB7-mut vector. ME-180 and SiHa cells were co-transfected with 100

nmol/L of pre-miR-196b or NC, and 100 ng of the reporter vector of interest. As a reference

control, 50 ng of pRL-SV vector (Promega) containing the Renilla luciferase gene was also

transfected with each condition. Firefly and renilla luciferase activities were measured at 24

hours post-transfection using the Dual-Glo Luciferase Assay System (Promega) according to the

manufacturer’s instructions.

2.3.11 Immunoblotting

Cells were transfected with either 100 nmol/L of pre-miR-196b or NC, and total protein

extracts were harvested on ice after 48 and 72 hours. Proteins of interest were probed with rabbit

anti-HOXB7 (1:500 dilution; Invitrogen) or mouse anti-GAPDH (1:10,000 dilution; Sigma), and

detected with IRDye fluorescent secondary antibodies (1:20,000 dilution, LI-COR). GAPDH

was used as a loading control. Immunoblots were scanned and quantified using the Odyssey

Infrared Imaging System (LI-COR).

2.3.12 Enzyme-linked Immunosorbent Assay (ELISA)

Cells were transfected with 30 nmol/L of siHOXB7 or siNEG, and the level of secreted

VEGF was measured at 48 and 72 hours post-transfection, using the Human VEGF DuoSet

ELISA (R&D Systems) according to manufacturer’s instructions.

2.3.13 5-aza-2’-deoxycytidine Treatment

ME-180 and SiHa and were seeded in 24-well plates and incubated overnight at 37°C,

5% CO2. Media containing 2 µM 5-aza-2’-deoxycytidine (5-aza-DCT) (Sigma-Aldrich) was

added to cells, 1 and 3 days after seeding, respectively. Cells were harvested 4 days after seeding

and total RNA was isolated for qRT-PCR analysis of miR-196b expression.

34

2.3.14 Statistical Analysis

All experiments have been performed at least three independent times, and the data are

presented as the mean ± standard error of the mean (SEM). The Student’s t-test function

(unpaired, two-tailed) in Microsoft Excel (Microsoft, Redmond, WA) was used to compare two

treatment groups. Graphs were plotted using GraphPad Prism software (GraphPad software, San

Diego, CA). The Kaplan-Meier method was used for univariate analysis, and the log-rank test

was used to examine associations between miR-19b expression and DFS, where the expression

was dichotomized at the median value. The relationship between tumor size and miR-196b

expression was investigated using Pearson’s correlation. Correlations between miR-196b

expression with either FIGO stage or nodal status were analyzed using one-way ANOVA.

2.4 Results

2.4.1 miR-196b was significantly down-regulated in primary cervical cancer tissues and cell lines, and was strongly associated with DFS

Expression of miR-196b was significantly reduced by almost 4-fold in 79 primary

cervical cancer tissues compared to 11 normal cervix tissues (P < 0.001; Fig. 2.1A). Importantly,

patients with lower than median miR-196b expression level at the time of diagnosis experienced

worse DFS compared to those with higher miR-196b expression (P = 0.02; hazard ratio = 0.39;

Fig. 2.1B). miR-196b expression was not significantly correlated with tumor size (P = 0.12),

FIGO stage (P = 0.14), or nodal status (P = 0.60). Global miRNA expression profiling conducted

on three cervical cancer cell lines (ME-180, SiHa and HT-3) also confirmed the down-regulation

of miR-196b in cervical cancer. From the 55 miRNAs that were deregulated at least 2-fold in all

three cell lines compared to 3 normal cervix tissues, miR-196b was amongst the most

significantly down-regulated miRNAs (Fig. 2.1C), consistent with a previously published

miRNA expression profiling study (159).

To explore whether miR-196b down-regulation was epigenetically determined, in

addition to chromosomal loss [201], ME-180 and SiHa cells were treated with the demethylating

agent 5-aza-2’-deoxycytidine (5-aza-DCT). This treatment resulted in only a minimal increase in

miR-196b expression (Fig. S2.1A left), indicating that promoter methylation was unlikely to be a

major mechanism for miR-196b under-expression, whereas the levels of miR-375, which is

known to be epigenetically regulated, increased significantly (Fig. S2.1A right). Furthermore,

35

examination of publically-available microarray datasets of gene expression in primary cervical

cancer tissues did not reveal any significant alterations in the expression of DICER, drosha,

DGCR8, Exportin-5, or any subunits of RNA Polymerase II, which are all involved in miRNA

biogenesis and processing (data not shown).

Figure 2.1 miR-196b down-regulation in cervical cancer

A) miR-196b expression levels were measured using qRT-PCR in 79 primary cervical cancer samples, compared to 11 normal cervix epithelial tissue controls. B) Kaplan-Meier analysis of DFS in patients with cervical cancer. Red, patients with higher than median miR-196b expression level (n = 39); blue, patients with lower than median miR-196b expression (n = 39); one patient was removed from survival analysis due to missing survival information. C) Basal levels of miR-196b in three cervical cancer cell lines (SiHa, ME-180, and HT-3), as compared to normal cervix epithelial tissues, assayed by qRT-PCR. ** P < 0.01.

36

2.4.2 miR-196b over-expression significantly reduced cell viability, clonogenicity, proliferation and invasion

To assess the biological significance of miR-196b down-regulation, cells were

transfected with 30 nmol/L NC or pre-miR-196b. Up-regulation of miR-196b expression was

sustained for up to 72 hours after transfection (Fig. S2.1B). Transfection with pre-miR-196b led

to significantly decreased cell viability compared to controls at 48 and 72 hours post-transfection

(ME-180: 25% at 48h; 41% at 72h, SiHa: 29% at 48h; 54% at 72h) (Fig. 2.2A). In addition, miR-

196b over expression resulted in significant reductions in clonogenicity (ME-180: 57%

compared to NC, SiHa: 64% compared to NC) (Fig. 2.2B), proliferation (ME-180: 36% at 48h,

35% at 72h, SiHa: 22% at 48h; 21% at 72h) (Fig. 2.2C), and migration (32% compared to NC),

plus invasion (32% vs. 63% for NC) (Fig. 2.2D). Cells treated with pre-miR-196b demonstrated

a small, but not statistically significant, increase in the percentage of cells in the sub G1

population, accompanied by a small decrease in the G0-G1 population (Fig. S2.1C).

2.4.3 miR-196b over-expression suppressed tumor angiogenesis and tumor cell proliferation in vivo

Cells transfected with pre-miR-196b did not show a statistically significant difference in

tumor growth in mice, compared to NC-treated cells (Fig. S2.2). At 25 days post-implantation

however, CD31 immunostaining was nonetheless observed to be reduced to 69% in pre-miR-

196b-treated tumors compared to control tumors (Fig. S2.3, top). Furthermore, this was

associated with a modest yet significant reduction in Ki-67 expression (46% vs. 53% for cells

treated with NC) (Fig. S2.3, middle), and a minor, but not statistically significant increase in

TUNEL staining (Fig. S2.3, bottom). Hence, our data suggested that pre-miR-196b contributed

to reduced angiogenesis and tumor cell proliferation.

37

Figure 2.2 In vitro effects of miR-196b over-expression

A) Relative viability of ME-180 and SiHa cells were assessed at 48 and 72 hours post-transfection with pre-miR-196b (30 nmol/L), compared to Negative Control pre-miR (30 nmol/L), using the Trypan Blue assay. B) Clonogenicity of ME-180 and SiHa cells were assessed by transfection with 30 nmol/L of pre-miR 196b or Negative Control pre-miR. At 48

38

hours post-transfection, cells were harvested then counted and re-seeded at low density in 6-well plates. After 10 days of incubation, cells were fixed and stained and the number of colonies (>50 cells) were counted. C) Relative proliferation of ME-180 and SiHa cells were examined at 24, 48, and 72 hours post-transfection with pre-miR-196b (30 nmol/L), compared to Negative Control pre-miR (30 nmol/L), using the MTS assay. D) Representative images (left) and histograms (right) depicting migratory ability (top) and invasiveness (bottom) of ME-180 cells that were transfected with 30 nmol/L of pre-miR 196b or Negative Control pre-miR, harvested at 24 hours post-transfection, then counted and re-seeded in the invasion chambers. Cell migration (top), and invasion (bottom), assessed at 48 hours after seeding in transwell chambers. All data represent the mean ± SEM from 3 independent experiments. NC, pre-miR Negative Control; * P < 0.05; ** P < 0.01.

39

2.4.4 miR-196b directly targets HOXB7

The down-regulation of miR-196b in cervical cancer tissues and cell lines, and the

significant phenotypic effects of miR-196b over-expression both in vitro and in vivo, indicated

that miR-196b appears to be an important mediator of cervical cancer progression. Thus, a tri-

modal strategy (68) was utilized to identify potential mRNA targets that could account for these

phenotypic changes (Fig. S2.4). This method identified 15 overlapping candidate targets

(ANKHD1, CALM3, CLK2, CTDSP2, FGFR1, HDAC, HOXA7, HOXB7, KRT8, PCCB,

PUM2, SLC9A6, SMC3, SMG7, SR140). For target validation, ME-180 cells were transfected

with pre-miR-196b or NC, and transcript levels at 24 hours post-transfection were measured with

qRT-PCR for 12 candidate targets (suitable primers could not be designed for the other 3

transcripts), plus 4 previously-described targets of miR-196b (c-myc, BCL2, HOXA9, MEIS1).

None of the previously-described targets were significantly altered after miR-196b over-

expression (Fig S2.5A). Only 5 of the 12 tested candidate targets were significantly down-

regulated after miR-196b over-expression (Fig. S2.5B); HOXB7 was selected for further

functional evaluation since it was the candidate target that demonstrated the greatest level of

down-regulation (40% compared to controls). Furthermore, HOXB7 is a member of the Hox

gene cluster, a family of genes which have been reported to be dysregulated in various

malignancies (228), and the miR-196 family is known to target the mammalian Hox genes (229).

A luciferase binding assay confirmed that miR-196b directly and specifically interacted

with the 3’-UTR of HOXB7. In comparison to control cells transfected with pMIR-REPORT,

cells transfected with pMIR-HOXB7 showed reduced luciferase activity (ME-180: 68%, SiHa:

71%) when co-transfected with pre-miR-196b (Fig. 2.3A). This inhibitory effect was completely

abrogated with pMIR-HOXB7-mut, which contained a mutation in the miR-196b binding site.

In addition, transfection with pre-miR-196b resulted in significantly reduced HOXB7 mRNA

transcript (ME-180: 49% at 48h, 62% at 72h; SiHa: 55% at 48h, 69% at 72h) (Fig. 2.3B), and

protein (74% at 48h; 80% at 72h) (Fig. 2.3C) levels at 48 and 72 hours post-transfection. There

appeared to be a greater fold change in HOXB7 at the mRNA level compared to protein, which

is not surprising since miRNAs interact directly with mRNA transcripts and not proteins. In

addition, protein translation can be affected by a number of mechanisms that occur upstream,

such as nuclear export of RNA transcripts and recruitment of ribosomal subunits.

40

Figure 2.3 Identification of HOXB7 as an mRNA target of miR-196b

A) Relative luciferase activity of ME-180 or SiHa cells at 24 hours after co-transfection with pMIR-REPORT, pMIR-HOXB7 or pMIR-HOXB7-mut vectors and pre-miR-196b or NC (30 nmol/L). B) Relative HOXB7 mRNA expression levels in ME-180 or SiHa cells after transfection (48, and 72 hrs) with pre-miR-196b or NC (30 nmol/L), as measured by qRT-PCR. Expression levels were normalized to GAPDH expression. C) Representative Western blot image (top), and relative quantification of HOXB7 protein levels (bottom) after transfection (48, and 72 hrs) with pre-miR-196b or NC (30nmol/L). All data represented the mean ± SEM from 3 independent experiments. OD, optical density; NC, pre-miR Negative Control; *P < 0.05; **P < 0.01.

41

2.4.5 VEGF is a relevant downstream target of HOXB7

Since HOXB7 is a transcription factor, it was necessary to determine which gene(s)

regulated by HOXB7 could be relevant in this context for cervical cancer. Hence, cells were

transfected with 30 nmol/L of siHOXB7 or siNEG, and mRNA levels of known HOXB7 targets

such as Ku70, Ku80, DNA-PK, FGF2, MMP2, WNT5a, PDGFA, thromobospondin 2 (THBS2)

and VEGF (230-233) were measured at 48 hours post-transfection. VEGF demonstrated the

greatest level of down-regulation (43%) following HOXB7 knockdown (Fig. S2.6A).

Concordantly, significant reduction in VEGF mRNA (ME-180: 63% at 48h, 33% at 72h; SiHa:

69% at 48h, 41% at 72h) (Fig. 2.4A) and secreted protein (ME-180: 82% at 48h, 77% at 72h;

SiHa: 84% at 48h, 75% at 72h) (Fig. 2.4B) levels were observed at both 48 and 72 hours post-

transfection with siHOXB7, confirming that VEGF was indeed a relevant downstream HOXB7

target in this disease. Furthermore, other pro-angiogenic known targets of HOXB7 (FGF2,

MMP2, WNT5a, PDGFA, THBS2) were not significantly altered following HOXB7 knockdown

(Fig. S2.6A) or miR-196b overexpression (Fig. S2.6E), suggesting that HOXB7 mediates

angiogenesis via VEGF.

2.4.6 Knockdown of HOXB7 or VEGF recapitulated the biological effects observed following mir-196b over-expression

To further corroborate this newly-described pathway of miR-196b targeting HOXB7

which in turn regulated VEGF, ME-180 and SiHa cells were treated with siHOXB7 or siVEGF

to determine whether these interventions could recapitulate the effects of miR-196b over-

expression. Knockdown of HOXB7 and VEGF transcript and protein levels were sustained for

up to 72 hours after siRNA transfection (Fig. S2.6B, S2.6C, and S2.6D). We observed that cell

viability was reduced significantly after transfection with either siHOXB7 (ME-180: 77% at 48h,

66% at 72h; SiHa: 80% at 48h, 70% at 72h) (Fig. 2.5A, left) or siVEGF (ME-180: 78% at 48h,

60% at 72h; SiHa: 77% at 48h, 68% at 72h) (Fig. 2.5A, right), similar to the effects observed

after transfection with pre-miR-196b (ME-180: 25% at 48h, 41% at 72h; SiHa: 29% at 48h, 54%

at 72h) (Fig. 2.2A). Clonogenicity decreased significantly following either siHOXB7 (ME-180:

69%; SiHa: 71%) (Fig. 2.5B, left) or siVEGF (ME-180: 78%; SiHa: 75%) (Fig. 2.5B, right)

transfection, as previously observed for pre-miR-196b transfection (ME-180: 57% compared to

NC; SiHa: 64% compared to NC) (Fig. 2.2B). Cell migration (Fig. 2.5C, top) was also reduced

42

significantly following transfection with either siHOXB7 (60%) or siVEGF (62%), similar to the

results observed following pre-miR-196b transfection (32% vs. NC) (Fig. 2.2D, top). Finally,

cell invasion also significantly decreased after transfection with either siHOXB7 (52% vs. 74%

for control cells) or siVEGF (42% vs. 74% for control cells) (Fig. 2.5C, bottom), all

recapitulating the effects observed after transfection with pre-miR-196b (32% vs. 63% for

control cells) (Fig. 2.2D, bottom).

Figure 2.4 Identification of VEGF as a target of the HOXB7 transcription factor

A) Relative VEGF expression at the mRNA level after transfection of ME-180 or SiHa cells with siHOXB7 or siNEG (30 nmol/L), as determined by qRT-PCR. Expression levels were normalized to GAPDH expression. B) Relative levels of secreted VEGF protein after transfection of ME-180 or SiHa cells with siHOXB7 or siNEG (30 nmol/L), as determined by ELISA. All data represented the mean ± SEM from 3 independent experiments. siNEG, All Stars Negative Control; *P < 0.05; **P < 0.01.

43

Figure 2.5 Downstream effects of siHOXB7 and siVEGF

ME-180 and SiHa cells were transfected with siHOXB7, siVEGF or siNEG (30 nmol/L): A) Relative cell viability assessed using the Trypan blue assay at 48 and 72 hours post-transfection with siHOXB7 (left) or siVEGF (right); B) Cells were harvested at 48 hours post-transfection, then counted and re-seeded at low density in 6-well plates; after 10 days of incubation, cells were

44

fixed and stained and colonies (>50 cells) were counted. Histograms depict the relative number of surviving colonies post-transfection with siHOXB7 (left), and siVEGF (right). C) Representative images (left) and histograms (right) depict migratory ability (top), and invasiveness (bottom) after transfection. All data represented the mean ± SEM from 3 independent experiments. siNEG, All Stars Negative Control; *P < 0.05.

45

2.5 Discussion

In this study, we identified that miR-196b was significantly down-regulated in both

cervical cancer cell lines and primary tissues, which promoted tumour cell proliferation,

migration, invasion, and angiogenesis, mediated through VEGF regulation by HOXB7.

MicroRNA-196b was one of the six deregulated miRNAs first reported by Lui et al. for human

cervical cancer (159). Its biological role was first described for endometriosis (234); since then,

miR-196b has been reported to be deregulated in various human malignancies aside from

cervical cancer, including over-expression in acute lymphoblastic leukemia (ALL) (235) and

colon cancer (236); as well as under-expression in glioblastoma (237) and B-cell lineage ALL

(238). These discordant observations highlight the fact that miR-196b can function as either an

oncogene or a tumor suppressor, as described for many miRNAs. In glioblastoma, miR-196b

levels have been positively correlated with overall survival (237). In contrast, we report for the

first time that lower levels of miR-196b were associated with worse DFS for cervical cancer, by

promoting cellular proliferation, clonogenicity, migration and invasion in vitro, as well as tumor

cell proliferation and angiogenesis in vivo.

The mechanisms for miR-196b down-regulation are complex. Using array comparative

genomic hybridization (aCGH) profiling of cervical cancer samples, the chromosomal location

of miR-196b (7p15.2) was noted to be a region with a high level of homozygous loss in

squamous cell cervical carcinoma (239); hence this might be one explanation for down-

regulation of miR-196b in this disease. miR-196b has also been reported to be epigenetically

regulated in gastric cancer (240), which was not corroborated based on our data (Figure S2.2).

Few studies have identified mRNA targets of miR-196b, aside from c-myc (238), and the

Hox gene cluster (229). The Hox gene family consists of a set of 39 genes which encode

transcription factors that direct the basic structure and orientation of an organism during

embryonic development (241), regulating many crucial processes such as differentiation,

apoptosis, motility, angiogenesis and receptor signaling (228). Aberrant Hox gene expression has

been reported to mediate oncogenesis in many human cancers, including hepatocellular (242),

ovarian (243), as well as acute myeloid leukemia (AML) (244). There are at least three

mechanisms that have been described which can lead to Hox gene deregulation: a) over-

expression of Hox genes in a specific tissue type; b) epigenetic deregulation, whereby Hox genes

46

are silenced in a tissue when they should normally be expressed; and c) temporo-spatial

deregulation, whereby Hox gene expression in a tumor arising in a specific tissue temporo-

spatially differs from that in normal tissue (187). In this current study, we provide evidence for

an alternate mechanism for HOXB7 deregulation, via miR-196b, in cervical cancer.

Our current study also demonstrated that VEGF was a downstream transcriptional target

of HOXB7 in cervical cancer, which has also been reported for breast cancer (231), as well as

multiple myeloma (233). VEGF is a known key mediator of angiogenesis (245); a major

hallmark of human cancers (246). In contrast, other pro-angiogenic factors that are known targets

of HOXB7 (FGF2, MMP2, WNT5a and PDGFA) (231, 233) were not significantly altered

following HOXB7 knockdown or miR-196b overexpression. Although VEGF was initially

regarded to be an endothelial-specific ligand, reports have shown that VEGF can promote cancer

cell proliferation (247, 248), migration and invasion (249, 250). Interestingly, serum VEGF

levels have been shown to be a prognostic marker for DFS in cervical cancer patients, whereby

high pre-treatment VEGF levels were associated with worse survival (217, 218). In addition,

alterations in the serum concentration of VEGF have been used to measure treatment response in

cervical cancer patients (251). Serum expression of VEGF might indeed be a superior read-out of

angiogenic activity in cervix tumors, compared to tissue expression of VEGF. Unfortunately,

sera samples were not available from the patients in this current study.

A humanized anti-VEGF monoclonal antibody (A.4.6.1) that recognizes all biologically

active isoforms of VEGF and prevents their binding to VEGF receptors (VEGFR-1 and VEGFR-

2), Bevacizumab, has been shown to be effective in treating several human malignancies by

blocking angiogenesis (252). Thus far, Bevacizumab has been approved by the FDA for treating

metastatic colorectal cancer, non-small cell lung cancer, glioblastoma, and metastatic kidney

cancer. A multi-centre randomized Phase III clinical trial (GOG240) has recently been

completed, in which Bevacizumab in combination with standard treatment was evaluated in

cervical cancer patients with advanced (stage IVB), persistent, or recurrent cervical cancer. A

National Cancer Institute (NCI) press release in February 2013 announced that the Phase III trial

demonstrated that the addition of Bevacizumab significantly improved median survival by 3.7

months.

47

In conclusion, we report for the first time that miR-196b is a novel tumor suppressor in

cervical cancer, by regulating the transcription factor HOXB7, which in turn, induced VEGF

expression. The resulting phenotype of miR-196b down-regulation included increased cell

growth, clonogenicity, migration and invasion, as well as increased tumor cell proliferation and

vascularity. Furthermore, patients with lower miR-196b expression experienced a worse 5-year

DFS. Hence, this novel axis of miR-196b~HOXB7~VEGF might well provide the biological

rationale for the potential efficacy of an anti-angiogenic therapeutic strategy for cervical cancer.

48

Figure S2.1 In vitro effects of 5-aza-DCT treatment and miR-196b over-expression

A) qRT-PCR analysis of miR-196b (left) and miR-375 (right) levels in ME-180 and SiHa cells after treatment with DMSO or 2 µM 5-aza-2’-deoxycytidine (5-aza-DCT). Expression levels were normalized to RNU44 expression, relative to cells treated with DMSO. B) qRT-PCR analysis of miR-196b levels in ME-180 and SiHa cells after treatment with NC or pre-miR-196b (30 nmol/L). Expression levels were normalized to RNU44 expression, relative to cells treated with NC. C) Cell cycle analysis performed on ME-180 (left) and SiHa cells (right) using flow cytometry after treatment with pre-miR-196b or NC (30 nmol/L). The data represented the mean

49

± SEM from 3 independent experiments. NC, pre-miR Negative Control; **P < 0.01; P = ns (not significant).

50

Figure S2.2 Effect of miR-196b on in vivo tumor growth

Tumor-plus-leg diameter measurements of ME-180 tumors in SCID mice after intramuscular injection of cells transfected with pre-miR-196b or Negative Control pre-miR (60nmol/L). The plotted data represent the mean ± SEM from 9 mice in each group.

51

Figure S2.3 Effect of miR-196b on in vivo angiogenesis and apoptosis

Tumours removed at 25 days post-implantation were immunostained for CD31, Ki-67, and TUNEL expression; representative photomicrographs are shown for NC vs. miR-196b for CD31 (top), Ki-67 (middle), and TUNEL (bottom) immuno-expression. The corresponding histograms represented the mean ± SEM scoring obtained from 6 representative regions, from 2 independent tumors. NC, pre-miR Negative Control; *P < 0.05; P = ns (not significant).

52

Figure S2.4 Tri-modal strategy for target identification

A tri-modal strategy to elucidate targets of miR-196b in cervical cancer used a combination of: i) All predicted targets of miR-196b from five in silico miRNA target prediction databases (in silico); ii) mRNA transcripts up-regulated at least 2-fold in primary cervical cancer samples compared to normal cervix tissues [Cervical cancer (Up)]; and iii) mRNA transcripts down-regulated at least 0.5-fold at both 24 and 72 hours after transfection with 30 nmol/L of pre-miR-196b (Exp. Det.).

53

Figure S2.5 Transcript levels of putative miR-196b targets

qRT-PCR analysis of: A) previously described; and B) candidate targets of miR-196b. ME-180 cells were transfected with NC or pre-miR-196b (30 nmol/L) and transcript levels of candidate targets were measured at 24 hours post-transfection. Expression levels were normalized to GAPDH expression, relative to cells transfected with pre-miR Negative Control. The data represent the mean ± SEM from 3 independent experiments. *P < 0.05.

54

Figure S2.6 In vitro effects of treatment with siHOXB7, siVEGF, or pre-miR-196b

55

A) qRT-PCR analysis of candidate targets of HOXB7. Cells were transfected with siNEG or siHOXB7 (30 nmol/L) and transcript levels of candidate targets were measured at 24 hours post-transfection. B) qRT-PCR analysis of HOXB7 (left) and VEGF (right) transcript levels after treatment with siHOXB7, siVEGF, or siNEG (30 nmol/L). C) Western blot analysis of HOXB7 protein levels after treatment with siNEG or siHOXB7 (30 nmol/L). D) VEGF protein levels as measured by ELISA, after treatment with siNEG or siVEGF (30 nmol/L). E) qRT-PCR analysis of candidate targets of HOXB7. Cells were transfected with NC or pre-miR-196b (30 nmol/L) and transcript levels of candidate targets were measured at 24 hours post-transfection. The data represent the mean ± SEM from 3 independent experiments. NC, pre-miR Negative Control; siNEG, All Stars Negative Control; *P < 0.05; **P < 0.01.

56

CHAPTER 3: CHALLENGES ASSOCIATED WITH DEVELOPING A PROGNOSTIC MICRORNA

SIGNATURE FOR HUMAN CERVICAL CARCINOMA

The data in this chapter have been submitted for publication and is currently under review: How C, Pintilie M, Bruce J, Hui ABY, Clarke B, Wong P, Yin S, Yan R, Waggott D, Boutros PC, Fyles A, Hedley DW, Hill RP, Milosevic M, Liu FF.

57


Purpose: Cervical cancer remains the third most frequently diagnosed and fourth leading cause

of cancer death in women worldwide. We sought to develop a micro-RNA signature for cervical

cancer that was prognostic for disease-free survival, which could potentially allow tailoring of

treatment for patients.

Experimental Design: The TaqMan Low Density Array (Applied Biosystems) was utilized to

measure the expression of 377 unique human micro-RNAs in 79 frozen cervical cancer

specimens with five years median clinical follow-up (range, 0.64-10.6 years). LASSO regression

was applied to the normalized micro-RNA expression data, to select a subset of micro-RNAs

associated with disease-free survival. For our validation cohort (n=87), which were formalin-

fixed paraffin-embedded samples, three different methods were used to measure micro-RNA

expression: TaqMan Low Density Array, NanoString nCounter Human miRNA Expression

Assay, and individual single-well quantitative real-time PCR using TaqMan Micro-RNA Assays.

Results: A candidate prognostic 9-micro-RNA signature set for disease-free survival was

identified in the frozen samples. Three different approaches however, to validate this signature in

an independent cohort of 87 patients with FFPE specimens, were unsuccessful.

Conclusions: There are several challenges and considerations associated with developing a

prognostic micro-RNA signature for cervical cancer, namely: tumour heterogeneity, lack of

concordance between frozen and FFPE specimens, and platform selection for global micro-RNA

expression profiling in this disease.

3.2 Introduction

Micro-RNAs (miRNAs) are a class of small, non-coding RNAs that play important roles

in regulating target genes by binding to complementary sequences in mRNA transcripts (4).

Deregulation of miRNA expression has been reported for numerous solid and hematological

malignancies, which is not surprising given their involvement in multiple critical biological

processes, including development, proliferation, and apoptosis (2, 253). miRNAs have been

described to be extremely stable, and can be readily extracted from cell lines and various types of

clinical specimens, including frozen and formalin-fixed paraffin embedded (FFPE) tissues,

58

blood, serum, plasma, urine, and saliva (96, 98-102). miRNA expression profiling of solid and

hematological human malignancies has identified disease-specific miRNA signatures associated

with diagnosis, progression, staging, prognosis and response to therapy (254). Numerous

prognostic miRNA signatures have since been described for various malignancies, including:

lung (125-127), colorectal (121), gastric (123), esophageal (122), hepatocellular (124), prostate

(129), nasopharyngeal (128), and cervical cancers (120).

Cervical cancer is the third most frequently diagnosed cancer, and the fourth leading

cause of cancer mortality in women worldwide, with an estimated 530,000 new cases and

175,000 deaths each year (131). Although cervical cancer incidence and mortality have declined

over the past 30 years in the United States (223), the 5-year survival rate remains less than 40%

for patients diagnosed with Stage III disease and above (224). A two-miRNA signature that

could predict overall survival in cervical cancer patients was reported by Hu et al. (120), and

remains the only reported miRNA signature for cervical cancer to date. However, we were

unable to independently corroborate this signature with our own cohort of patients. Given these

limitations and the poor survival for patients with advanced stages of cervical cancer, we sought

to develop an independent miRNA signature that could predict disease-free survival (DFS) for

cervical cancer patients; and potentially allow tailoring of treatment according to risk. Herein, we

describe the challenges and considerations associated with developing such a prognostic miRNA

signature for cervical cancer.


3.3.1 Clinical specimens

Pre-treatment cancer samples were collected from patients with cervical cancer who were

undergoing curative chemo-radiation, consisting of external-beam radiotherapy to the primary

cervical tumour and pelvic lymph nodes (45 to 50 Gy total, in 1.8-to-2-Gy daily fractions using

18 or 25MV photons), combined with weekly cisplatin (40 mg/m2 total, 5 doses). Patients were

staged using the FIGO (International Federation of Gynecologists and Obstetricians) system,

with additional clinical information gathered using computed tomography (CT) scans of the

abdomen and pelvis and magnetic resonance imaging (MRI) of the pelvis to assess local and

59

lymphatic disease. Pelvic and para-aortic lymph nodes were classified as positive for metastatic

disease if the MRI short-axis dimension was >1 cm, and equivocal if it was 8 to 10 mm.

The training cohort comprised of flash-frozen punch biopsies obtained from 79 patients

treated at the Princess Margaret Cancer Centre between 2000 to 2007, inclusively. The biopsy

specimens were placed in a storage medium (optimal cutting temperature (OCT) compound) for

histopathologic examination, then flash-frozen in liquid nitrogen. H&E-stained tissue sections

were cut from the OCT-embedded material, and evaluated by a gynecological pathologist (B.

Clarke). The total cell content (stroma and tumour cells) was estimated for all tissue samples

using a light microscope, and only samples containing at least 70% tumour cells were considered

for further analysis. Flash-frozen normal cervix tissues obtained from 11 patients who underwent

total hysterectomy for benign causes served as the normal comparators.

The validation cohort comprised of diagnostic FFPE blocks collected from 87 cervical

patients treated between 1999 and 2007, inclusively. All samples contained at least 70%

malignant epithelial cells, as determined by a gynecologic pathologist (B. Clarke), or were

macro-dissected prior to RNA purification. FFPE normal cervix tissues obtained from 9 patients

who underwent hysterectomy for benign causes served as the normal comparators.

3.3.2 Sample processing

For the training cohort specimens, two sections of 50-µm thickness were cut from the

OCT-embedded flash-frozen tissues and placed in a nuclease-free microtube. Total RNA was

isolated using the Norgen Total RNA Purification Kit (Norgen Biotek), according to the

manufacturer’s instructions. Global miRNA expression was measured in both the cervical

cancer and normal cervix tissues with the TaqMan® Low Density Array (TLDA) Human

MicroRNA A Array v2.0 (Applied Biosystems) using the Applied Biosystems 7900HT Real-

Time PCR System, as previously described (96).

For the validation cohort specimens, ten sections of 5-µm thickness were cut from the

FFPE tissues and placed in a nuclease-free microtube. Total RNA was isolated using the Norgen

Total RNA Purification Kit (Norgen Biotek), according to the manufacturer’s instructions. We

measured the expression of the 9 miRNAs in our prognostic signature using three methods: 1)

Applied Biosystems TLDA Human MicroRNA A Array v2.0; 2) NanoString nCounter Human

60

miRNA Expression Assay v1.6.0; and 3) individual single-well qRT-PCR using Applied

Biosystems TaqMan MicroRNA Assays.

3.3.3 Data normalization

To normalize the miRNA expression data from TLDA, the raw miRNA abundances were

loaded into the R statistical environment (v2.15.2). Three control genes were utilized for

normalization: RNU44, RNU48, and U6. Normalized miRNA abundances were calculated as

− log2(2−(CT− C

c)), where CT represents the abundance of a given miRNA, and CC represents the

mean of the control gene abundances.

To normalize the miRNA expression data from NanoString, the R package

‘NanoStringNorm’ (255) was utilized with the following settings:

1) Probe level correction.

2) Code Count Correction = "geo.mean" (geometric mean)

3) Background Correction = "mean.2sd", i.e. mean +/- 2 standard deviations (Background is

calculated based on negative controls, the calculated background is subtracted from each sample)

4) Sample Content Correction = "top.geo.mean" (The option 'top.geo.mean' is a method which

ranks miRNAs based on the sum of all samples and then takes the geometric mean of the top 75)

5) log2 transformation

3.3.4 Survival analysis

The normalized data were filtered to remove miRNAs with low expression (Ct > 35) in

over 20% of samples. LASSO regression was applied to the normalized TLDA miRNA

expression data for the training cohort (256), to select a subset of miRNAs associated with DFS.

The parameter was obtained through cross-validation (0.087942). The DFS estimates were

calculated based on the Kaplan-Meier method. The hazard ratios were obtained from the

unadjusted Cox regression method; miRNAs significant for DFS were selected, leading to a 9-

miRNA signature prognostic for DFS. A risk score was calculated using the coefficients obtained

in LASSO regression and the normalized miRNA expression levels. The risk scores were

dichotomized at the median, and the cohort was divided into low and high risk groups. Kaplan-

Meier survival analysis was used to illustrate the difference in DFS between the high and low

risk groups. The p-value associated with the two curves was determined using the log-rank test.

61

All statistical analyses were performed with R statistical environment (v2.15.2) using the

“Survival” package for survival analysis, and “glmnet” for LASSO analysis. Significance was

defined as p-values below 0.05.

Publically-available cervical cancer Illumina Small RNA-Seq data from The Cancer

Genome Atlas (TCGA) Data Portal was utilized as another independent validation cohort (n =

48). Level three Small RNA-Seq (isoform_expression) data and clinical information for the 48

cervical cancer patients were downloaded from the Broad Firehose (stddata run 2013_06_06).

Reads-per-million (RPM) data were log2 transformed and z-score standardized before the 9-

miRNA signature equation was applied to calculate risk scores. Patients were dichotomized by

the previously established cut-point (median risk score in training cohort) and compared using a

log-rank test.

3.4 Results

3.4.1 The majority of significantly differentially-expressed miRNAs were

downregulated

The clinical characteristics of the patients in the training (n = 79) and validation (n = 87)

cohorts are provided in Table 3.1. Analysis of tumour-normal (T/N) fold changes revealed that

the majority of significantly differentially-expressed miRNAs (P < 0.01) were downregulated in

our cervical cancer samples (Table S3.1). Of the 29 significantly differentially-expressed

miRNAs, only 2 were upregulated (miR-21 and miR-187).

62

Table 3.1 Clinical parameters of patients in the training and validation cohorts

Training cohort

(n = 79)

Validation cohort

(n = 87)

Age (years)

Median 48 48

Range 26-84 19-83

Tumour size

≤ 5 cm 48 (61%) 43 (52%)

> 5 cm 31 (39%) 39 (48%)

FIGO stage

IB 24 (30%) 22 (25%)

IIA 2 (3%) 5 (6%)

IIB 35 (44%) 31 (36%)

IIIA 0 2 (2%)

IIIB 18 (23%) 27 (31%)

Pelvic or para-aortic node involvement

Positive 25 (32%) 29 (33%)

Equivocal 15 (19%) 18 (21%)

Negative 39 (49%) 40 (46%)

Overall survival

Deaths 24 (31%) 26 (30%)

Disease-free survival

Relapses or deaths 28 (35%) 33 (38%)

Follow-up (years)

Median 6.0 5.3

Range 0.7-10.6 1.0-10.5

63

3.4.2 Patient miRNA expression was associated with disease-free survival

Using LASSO regression on the normalized TLDA data from the training cohort, we

derived the following model to calculate the risk score for each patient using the expression

values of nine miRNAs:

Risk Score = (0.197109 x Elet-7c) – (0.07048 x EmiR-21) + (0.045797 x EmiR-222) – (0.56469 x EmiR-

451) + (0.171838 x EmiR-455-5p) – (0.01725 x EmiR-134) + (0.506956 x EmiR-148a) + (0.203466 x EmiR-

218) + (0.22355 x EmiR-500)

Where EX represents the normalized expression level of the given miRNA.

In this prediction model, a higher risk score would predict for worse DFS. The risk score

was calculated for each patient in the training set, and the median risk score (-0.05373) was used

to dichotomize the low vs. high risk groups. Our 9-miRNA signature was significantly predictive

of DFS for the 79 patients in the training cohort, with a hazard ratio of 9.26 and log-rank p-value

of 6.9 x 10-7 (Fig 3.1).

3.4.3 Validation of miRNA signature

In our attempts to validate the 9-miRNA signature, we used three different methods to

measure miRNA expression in the validation cohort: 1) TLDA (n = 87); 2) NanoString (n = 87);

and 3) qRT-PCR (n = 68; 19 samples omitted due to insufficient RNA). A risk-score was

calculated for each of the patients by applying the miRNA TLDA expression values to our

prediction model. The patients were divided into the high vs. low risk groups based on their risk

score, using the same cut-off point as defined in the training cohort (Fig 3.2A). This analysis was

repeated for the NanoString and the qRT-PCR datasets (Fig 3.2B-C). Regardless of the method

used for measuring miRNA expression, the 9-miRNA signature was not significant when applied

to the patients in the validation cohort. We also utilized publically-available cervical cancer

Illumina Small RNA-Seq data derived from The Cancer Genome Atlas (TCGA) Data Portal as

another independent validation cohort (n = 48). Interestingly, this validation attempt approached

statistical significance (p = 0.05251) (Fig S3.1A).

64

Figure 3.1 Kaplan-Meier analysis of DFS according to 9-miRNA signature

A risk score was calculated for each patient in the training cohort (n = 79) using our 9-miRNA signature for DFS in cervical cancer. The median risk score was used to divide patients into the high vs. low risk groups. HR; hazard ratio, DFS; disease-free survival, CI; 95% confidence interval.

0 2 4 6 8 10

0.0

0.2

0.4

0.6

0.8

1.0

Time to death or relapse (years)

DF

S

||

|||| | | |

|| | | ||| | | ||| || | | | || || | || | | |

| | || ||

| | |

| | || | |

Low Risk, n=40, 5y DFS=89%High Risk, n=39, 5y DFS=44%

Log-rank p-value=6.9e-07High Risk vs. Low Risk , HR= 9.26 , 95% CI: 3.2 - 26.78

65

Figure 3.2 Application of 9-miRNA signature to validation cohort

Kaplan-Meier analysis of DFS. A risk score was calculated for each patient in the validation cohort, by applying our 9-miRNA signature for DFS to the miRNA expression data generated using A) TLDA, B) NanoString, and C) individual qRT-PCR. The same cut-off point from the training set was used. HR; hazard ratio, DFS; disease-free survival, CI; 95% confidence interval.

66

3.4.4 Independent corroboration of the Hu et al. 2-miRNA signature

We attempted to perform an independent corroboration of the only published prognostic

miRNA signature to date for cervical cancer by Hu, et al. (S = 17.9 − 0.284 × EmiR-9 − 0.376 ×

EmiR-200a, where S represents the risk score for each patient, and EmiR-9 and EmiR-200a

represent the normalized expression levels of miR-9 and miR-200a in each patient, respectively)

(120). This 2-miRNA signature was first applied to the miRNA TLDA expression values from

our training frozen samples (n = 79) to calculate a risk-score for each patient. Although the

authors used 0 as the cut-off point with their training set, we were not able to use this value for

corroboration because the risk scores were all above 0. We thus divided our patients into high

vs. low risk groups, with 1/3 in the former and 2/3 in the latter categories, which reflected the Hu

et al. population when 0 was used as their cut-off. Based on the Kaplan-Meier analysis, the 2-

miRNA signature was not significant when applied to the patients in our frozen cohort utilizing

the TLDA platform (Fig 3.3A). This analysis was repeated for the FFPE TLDA, and FFPE

NanoString datasets; neither of which could corroborate the 2-miRNA signature (Fig 3.3B-C).

In a final attempt to corroborate the Hu et al. signature, we utilized Small RNA-Seq data derived

from the TCGA Data Portal as yet another independent cohort (n = 48). Using the same

analytical methods, again, the 2-miRNA signature could not be corroborated (Fig S3.1B),

although at least the trend was in the correct direction, in contrast to the other 3 datasets.

An attempt was also made to cross-validate our own 9-miRNA signature with the

miRNA expression data from the Hu, et al. study; unfortunately, access to their raw data was not

provided; hence no further analysis was possible.

67

Figure 3.3 Kaplan-Meier analysis of DFS according to Hu et al. 2-miRNA signature

Using the Hu et al. 2-miRNA signature, a risk score was calculated for each patient from: A) TLDA frozen cohort (n = 79), B) TLDA FFPE cohort (n = 87), and C) NanoString FFPE cohort (n = 87). HR; hazard ratio, DFS; disease-free survival, CI; 95% confidence interval.

68

3.5 Discussion

In the clinical management of cancer patients, prognostic evaluation is essential to guide

appropriate treatment decisions. Unfortunately, in cervical cancer, there remains significant

differences in patient survival despite being assigned to the same clinical stage; underscoring the

gaps in the current system, as well as the need to develop more useful prognostic biomarkers.

We developed a candidate 9-miRNA signature that was prognostic for DFS in patients with

cervical cancer. According to our analysis, our validation cohort (n=87) was large enough to

detect a signature with a HR of 2.5 with 83% power. However, we could not validate our

signature (HR = 9.26) in our validation cohort, despite trying three separate techniques.

Furthermore, attempts to corroborate the only published miRNA signature to date for cervical

cancer (120) in three independent patient cohorts were also unsuccessful.

We believe that there are three key reasons as to why we failed to validate our candidate

9-miRNA prognostic signature for cervical cancer: i) intra-tumour heterogeneity; ii) miRNA

expression data from frozen and FFPE samples could not be directly compared in this disease;

and iii) the current platforms for global miRNA expression profiling are not sufficiently robust.

Firstly, it is well-established that human tumours are intrinsically heterogeneous. Two

types of tumour heterogeneity exist: a) inter-tumour heterogeneity, with differences between

tumours arising from different patients; and b) intra-tumour heterogeneity, with differences

between distinct sub-populations of cancer cells within a single individual’s tumour. In cervical

cancer, intra-tumour heterogeneity has been described on various levels. Many studies have

demonstrated significant variations in interstitial fluid pressure (257), blood perfusion (258), and

oxygen tension (259) in different regions within an individual tumour. At the chromosomal level,

intra-tumour heterogeneity has been described with respect to specific genetic mutations and

chromosomal abnormalities such as gains and deletions, reflecting the polyclonal derivation of

cervical cancer (260-264). At the mRNA transcript level, Bachtiary et al. evaluated intra-

tumoural heterogeneity across 11 cervical cancer patients, and demonstrated that multiple

biopsies from distinct areas of an individual tumour were necessary to reduce sampling bias,

with genes displaying low intra-tumour heterogeneity requiring two to three biopsies, and genes

with high intra-tumour heterogeneity requiring more than six biopsies per tumour (265).

69

To date, intra-tumoural heterogeneity of miRNA expression has only been reported in

breast cancer thus far (266). However, natural inter-patient variability has been shown to exist

among normal cervix samples, which complicates miRNA expression profiling studies (162).

Given that intra-tumour heterogeneity exists in cervical cancer at the macroscopic, chromosomal

and transcript levels, it would be a logical extension to assume that this would also apply to

miRNAs. In our study, we only utilized one biopsy from each patient in the training and

validation cohorts, which is therefore probably insufficient to obtain a representative measure of

miRNA expression for the patient’s entire tumour.

A second reason for the lack of signature validation is the lack of concordance with

respect to the tissue preservation method used for the two patient cohorts; specifically, the

training and validation sets consisted of frozen and FFPE samples, respectively. This experience

would appear to contradict several previous studies that have reported high concordance in the

miRNA expression data from tissue-matched frozen and FFPE samples derived from various

types of human tissues, including: breast (96), lung (107), kidney (108), skin (106), glioblastoma

(109), melanoma (111, 267), prostate (105, 112), and lymph nodes (268). However, with the

exception of our own single study that used the TLDA platform (96), the remaining reports

utilized other miRNA profiling technologies, such as microarray (106, 107, 111, 267, 268), deep

sequencing (107, 108), custom PCR array (105), and qRT-PCR using stem-loop (112) or locked-

nucleic-acid primers (109). There has only been one report to date that evaluated miRNA

expression data from tissue-matched frozen and FFPE cervix samples, which only analyzed 3

cervix specimens, in addition to 3 breast and 2 gall bladder samples (269). This report by

Doleshal et al. analyzed 3 miRNAs (miR-24, miR-103, miR-191) using qRT-PCR, and only

calculated the ΔCt values between frozen and matched FFPE samples without performing any

correlation tests. Furthermore, several of these published reports have demonstrated that

although there were high correlations (> 0.5) between tissue-matched frozen and FFPE samples

for the overall panel of miRNAs tested, there were specific individual miRNAs that were poorly

correlated between the matched samples (108, 109, 111), and sometimes even miRNAs that

demonstrated opposing patterns of under- or over-expression in tissue-matched sample pairs

(105, 112). The reasons behind these discrepancies, as to why some miRNAs correlate between

frozen and FFPE samples; yet others do not, remain unclear.

70

Lastly, we believe that the current methods used for global miRNA expression profiling

are not sufficiently robust. We used the TLDA and NanoString platforms for global miRNA

expression profiling, which are both widely-used for miRNA expression profiling. However, in

our own recent analyses, the correlation in miRNA expression levels between these two

platforms, even using the same FFPE RNA samples, had only a ρ of 0.65 (manuscript in

preparation), underscoring the challenges in the technologies to be able to reliably identify

clinically useful biomarkers in this disease. This might also provide one technical explanation for

the recent review describing the difficulties in validating miRNAs for human malignancies

(270). More recent reports have demonstrated the advantages of deep sequencing, including

increased sensitivity and specificity with few false-positive calls (115, 116). Given that the

majority of differentially-expressed miRNAs in cervical cancer are down-regulated, deep

sequencing would likely be a more suitable and robust platform to identify and validate a

prognostic miRNA signature in this disease.

Our validation attempt using Small RNA-Seq data from the TCGA data portal was

promising, with our 9-miR signature predicting worse outcome for high-risk patients than low-

risk patients (p = 0.053). Interestingly, the datasets from the TCGA cohort and our training

cohort were both generated from frozen tissues. With the addition of more cervical cancer

samples to the TCGA data portal with survival information and miRNA expression data, we

could potentially obtain a statistically significant result in a future validation attempt.

In conclusion, a prognostic miRNA signature for cervical cancer could not be validated,

due to intra-tumoural heterogeneity, incompatibilities between miRNA expression data from

frozen and FFPE samples, and insufficiently robust technical platforms for global miRNA

expression profiling. We propose that this could potentially be resolved in the near future, with

the advancement of technologies for miRNA expression profiling such as deep-sequencing. At

present, the TCGA Data Portal contains deep sequencing miRNA expression data with clinical

annotation for only 48 cervical cancer patients; when this dataset is expanded to include more

samples, it could potentially be utilized as an important resource to identify and corroborate

potential miRNA signature sets. Our observations provide an important cautionary tale for future

miRNA signature studies for cervical cancer, which can also be potentially applicable to miRNA

profiling studies involving other types of human malignancies.

71

Table S3.1 Significantly differentially-expressed miRNAs in cervical cancer

miRNA Fold Change (log2) P-value

miR-149 -4.30 2.09 x 10-9

let-7c -3.42 5.69 x 10-7

miR-218 -3.79 5.14 x 10-6

miR-139-5p -2.75 1.44 x 10-5

miR-203 -3.43 1.44 x 10-5

miR-125b -2.80 1.44 x 10-5

miR-376c -3.08 4.77 x 10-5

miR-361-5p -3.76 1.15 x 10-4

miR-320a -1.07 1.33 x 10-4

miR-324-3p -1.85 1.67 x 10-4

miR-196b -1.59 2.07 x 10-4

let-7b -1.48 4.24 x 10-4

let-7a -1.84 4.24 x 10-4

miR-100 -2.61 4.24 x 10-4

let-7e -1.29 1.16 x 10-3

miR-193b -1.14 1.23 x 10-3

miR-99a -2.86 1.32 x 10-3

miR-328 -1.84 1.35 x 10-3

miR-195 -1.82 1.35 x 10-3

miR-21 1.89 1.35 x 10-3

miR-24 -1.10 1.43 x 10-3

miR-574-3p -1.57 2.40 x 10-3

miR-375 -5.72 2.43 x 10-3

miR-148a -2.22 2.43 x 10-3

miR-29a -1.51 2.70 x 10-3

let-7g -0.91 6.46 x 10-3

miR-210 -1.70 6.80 x 10-3

miR-455-5p -2.26 8.50 x 10-3

miR-187 2.85 8.70 x 10-3

72

Figure S3.1 Application of TCGA Small RNA-Seq data to cervical cancer signatures

miRNA expression data from the TCGA miRNASeq cohort (n = 48) was used to test: A) our 9-miRNA signature, and B) the Hu et al. 2-miRNA signature. HR, hazard ratio; CI, 95% confidence interval.

A

Hu et al. signatureB

9-miRNA signature

73

CHAPTER 4: IDENTIFICATION AND CORRECTION OF AMPLIFICATION PROTOCOL BIAS IN MICROARRAY

STUDIES

The data in this chapter have been submitted for publication and is currently under review: Yan R*, How C*, Saha S, Haider S, Hui ABY, Waggott D, Chong LC, Harding NJ, Kasprzyk A, Clarke BA, Fyles A, Reich HN, Liu FF, and Boutros PC. *These authors contributed equally to this work.

74


Background: Microarray platforms allow the simultaneous measurement of thousands of

transcript abundances, but their clinical use has been limited because surgical specimens often

yield low quantities of highly-degraded RNA. To resolve this issue, several novel amplification

protocols have been developed. We evaluated the bias introduced by alternative amplification

protocols by systematically comparing mRNA abundance profiles generated using standard and

low-input/low-quality amplification protocols from matched samples.

Methods: RNA extracted from 12 cervix cancers and 2 unmatched normal cervix samples was

divided into two aliquots. One was amplified using a low-input protocol and the other using

standard techniques; both samples were analyzed on Affymetrix microarrays. Raw array data

were preprocessed using the Robust Multi-array Average (RMA) algorithm, followed by linear-

modeling and correlation analyses to identify genes showing different patterns between the two

protocols. Agglomerative hierarchical clustering was used to visualize the distribution of the

mRNA abundances across the protocols. Significant genes were then extracted for pathway and

sequence analyses to determine whether differentially expressed genes are biased in gene

structure or function. Lastly, to generalize these conclusions we repeated this experimental

approach in two normal kidney samples and one Michigan reference RNA sample (RNA

control).

Results: Data generated using differed dramatically between the two amplification protocols,

with systematic biases affecting thousands of genes. Altered genes were biased towards specific

biological pathways, but did not exhibit unique sequence characteristics. Batch correction

protocols reduced, but did not eliminate, protocol bias.

Conclusions: Each amplification protocol creates a unique, systematic bias in microarray data.

As a result, experiments performed using different protocols cannot be directly merged. Batch-

correction methods partially mitigate these effects, but it remains essential to ensure consistency

in the amplification protocol when comparing microarray data.

75

4.2 Introduction

mRNA microarrays are the most widely used high-throughput technology for functional

genomics. Modern microarrays measure the abundances of tens of thousands of transcripts in

parallel, based on fluorescent nucleotide labeling and hybridization between complementary

sequences (271). Despite the emergence of sequencing-based methods, they remain useful in

both basic (272, 273) and translational (274, 275) research, in part because of their cost-

effectiveness and favourable signal-to-noise ratios for low-abundance transcripts.

Importantly for clinical applications, the error-characteristics of microarrays have been

extensively characterized. Comprehensive empirical studies have led to the development and

assessment of sophisticated analysis approaches. For example, the statistical power of a study

can be increased by direct integration of intensity-level data from samples measured using two

different microarray technologies (276). Despite differing surface chemistry, signal-detection,

and hybridization methods, cross-platform comparisons such as the MAQC studies have shown

that most platforms are inherently reliable (277, 278). Indeed, biomarkers developed by different

teams and using different microarray platforms show remarkable concordance (279). Hence,

expression microarrays currently remain the best-studied and most established “–omic”

technology.

Surprisingly, despite this extensive characterization and consequent wide adoption of

microarrays in discovery research, the uptake of microarrays as clinical diagnostic tools remains

relatively limited. Outside of a small number of high-profile projects (280), most successful

transcriptomic biomarkers are implemented using medium-throughput techniques such as real-

time PCR (281-283), even when the initial discovery used microarrays. This occurs for several

reasons. First, clinical samples are usually archived after undergoing formalin-fixation followed

by embedding in paraffin (FFPE), a procedure used because it maintains morphological features

of the tissue for histological analysis. However it is very challenging to extract high-quality RNA

from FFPE tissue samples because paraffin embedding does not preserve nucleic acids well

(103) and formalin fixation cross-links nucleic acids and proteins (284, 285). Second, clinical

specimens are often small in size, thus yield a small quantity of RNA. For example, two 10 µm

sections of a breast cancer biopsy yielded only 150 ng of total RNA, an amount insufficient for

standard microarray labeling procedures (286). Finally, independent of preservation method and

76

specimen size, nucleic acids isolated from clinical specimens are often degraded relative to those

from model systems. After tumour tissue is surgically excised, it is usually sent for pathological

assessment prior to freezing or fixation. This lag between surgery and tissue preservation

significantly influences RNA quality, and affects mRNA abundance patterns (287, 288).

Similarly, thawing of frozen tissues prior to and during RNA isolation contributes to RNA

degradation and changes isoform abundances (289).

Several approaches have been taken to mitigate these challenges and to enable routine

microarray profiling of clinical specimens. Unfortunately, to date there have been no systematic

comparisons of the accuracy and bias of these protocols, nor a template for how data created

using optimized protocols can be merged with data generated from traditional approaches. Here,

we take the first step towards resolving these challenges by evaluating the bias introduced by the

NuGEN WT-Ovation Pico protocol for the robust linear amplification of limited amounts of

significantly degraded RNA. This protocol works with transcripts that have lost their poly-A tail

because it combines random- and poly-A priming (290). We considered three issues. First, we

determined if the use of a low-input/low-quality RNA amplification protocol altered

transcriptional profiles either globally, or for specific genes. Second, we investigated whether

these changes altered biological conclusions drawn from the dataset. Third, we examined if there

was a bias towards specific classes of genes, either in terms of their sequence characteristics (e.g.

length, GC-content) or biological function.


4.3.1 Samples

Twelve flash-frozen punch biopsies were obtained from patients with locally advanced

cervix cancer. After biopsy, the specimens were placed in optimal cutting temperature (OCT)

storage medium for histopathologic examination, then flash-frozen in liquid nitrogen. H&E-

stained tissue sections were cut from the OCT-embedded material and evaluated by a

gynecologic oncology pathologist (B Clarke). Total cell content (stroma and tumour cells) were

estimated for all tissue samples using a light microscope, to ensure that all samples contained

>70% malignant epithelial cells. Flash-frozen normal cervix tissues obtained from two patients

who underwent hysterectomy for benign causes served as normal comparators.

77

Kidney biopsies were performed at University of Health Network Toronto using an

ultrasound-guided automated 18-gauge needle system. A section of the third clinical core

obtained for kidney biopsy is routinely stored in RNAlater (Life Technologies, Carlsbad CA) at

4C. Once determined that this tissue is not required for clinical diagnosis, this core segment is

available for research purposes as per institutional informed consent process. The RNA mixture

control sample is a Stratagene Universal human reference RNA sample (Catalog Number:

740000-41). All tissue acquisition protocols (for both cervix and kidney samples) were approved

by the Research Ethics Board of the University Health Network.

4.3.2 Sample processing

Two sections of 50-µm thickness were cut from the OCT-embedded flash-frozen cervix


RNA Purification Kit (Norgen Biotek), according to the manufacturer’s instructions, and

aliquoted for further analysis. RNA quantity and quality were determined using the NanoDrop

ND-1000 Spectrophotometer (NanoDrop Technologies) and the Agilent 2100 Bioanalyser RNA

6000 Pico LabChip kit (Agilent Technologies), respectively.

The kidney biopsy core was micro-dissected using a stereomicroscope to separate the

glomerular and tubulo-interstitial compartments as previously described [33]. Following tissue

disruption and homogenization total RNA was isolated using the RNeasy kit (Qiagen, Valencia

CA), and eluted in 30 µL of RNase-free water. The sample obtained from the tubulo-interstitial

compartment was split into two aliquots. RNA quantity and quality were evaluated as described

above for cervix samples.

4.3.3 Microarray hybridization

Each aliquot of total RNA from the 12 cervix cancer and 2 normal cervix tissue samples

was reverse-transcribed and linearly amplified using one of two methods: 1) 50 ng of total RNA

was reverse transcribed into cDNA and linearly amplified using the NuGEN WT Ovation Pico

kit, according to the manufacturer’s protocol. The cDNA was fragmented and biotin labelled

using the NuGEN FL-Ovation cDNA Biotin Module v2.; 2) 200 ng of total RNA was reverse

transcribed into cDNA with the Affymetrix GeneChip One-Cycle cDNA Synthesis Kit and

linearly amplified using the Affymetrix 3’ IVT Express kit, according to the manufacturer’s

protocol. For both reverse transcription/amplification methods, the quality and quantity of the

78

amplified, labelled cDNA product was confirmed using the Agilent 2100 Bioanalyzer, and then

hybridized to the Affymetrix GeneChip Human Genome U133 Plus 2.0 arrays, according to the

manufacturer’s protocol.

4.3.4 Pre-processing

Quantitative microarray data (CEL files) were first loaded into the R statistical

environment (v2.15.2) using the affy package (v1.36.0) of the BioConductor library. Samples

were measured in three different batches. All RNA samples were subjected to quality

assessment: RNA Integrity Numbers (RINs) ranged from 2.2 – 8.5 (Table S4.1); no arrays were

excluded. RINs were rounded to the nearest integer for clustering analyses. These data were then

pre-processed with Robust Multi-array Average (RMA) algorithm (291). To remove unexpressed

genes, we applied a Y chromosome filter (i.e., signal intensity < 5, Figure S4.6), as described

previously (265).

For some analyses we considered data produced by the two protocols separately and for

others we merged them as a complete dataset. We therefore used the Robust Multi-array Average

(RMA) pre-processing algorithm and the Y chromosome filtering method three times, once on

the entire dataset and once for each of the two protocol-specific datasets. ProbeSet mappings

were updated to modern Entrez Gene ID annotations as implemented in R package

hgu133plus2hsentrezgcdf (v15.0.0) (292), and these were annotated with Entrez Gene

information from NCBI as of 2012-12-05. Raw and pre-processed data are available in the GEO

repository (GSE42764).

4.3.5 Unsupervised machine-learning

Unsupervised hierarchical clustering was used to identify the strongest trends within the

dataset. The agglomerative hierarchical clustering using complete linkage and Pearson's

correlation were used as a similarity metric chosen for clustering, as implemented in the cluster

package (v1.14.3) for the R statistical environment (v2.15.2). The randomness of clustering

patterns was assessed using the Adjusted Rand Index (ARI), implemented in the mclust package

(v3.5).

79

4.3.6 Correlation analyses

To assess if Nugen- and Express-processing yielded identical ordering of samples of

genes, Spearman's rank-order correlation coefficient was applied between cervix samples

processed with the Nugen protocol and those processed with the Express protocol. In the

replication analysis (i.e., 2 normal kidney samples and 1 RNA reference), we performed

Pearson's correlation instead of Spearman's rank-order correlation because Pearson’s correlation

is able to capture the degree of concordances or discordances, while Spearman’s rank-order

correlation is likely to overlook the differences between small and large concordances or

discordances for small sample size.

4.3.7 Statistical analysis of microarray data

To address the effects of the Nugen and Express protocols, we performed three separate

linear-modeling analyses, all using the limma (v3.14.2) package for the R statistical environment

(v2.15.2).

First, we pre-processed all data together and fitted a model to explicitly test the effect of

protocol on each gene. The model used was: Yi = α0,i + α1,iXi, where Y is a binary indicator value

representing protocol (Nugen vs. Express) and Xi is the expression of the 14 replicate Nugen and

Express arrays for the i-th gene. A Linear Transformation of Replicates (LTR) was then used in

an attempt to remove platform-and batch-specific biases (293). Normalized data from a subset of

arrays were fitted into the LTR model, which was then used to adjust the remainder of the data.

The LTR analysis was performed with the LTR (v1.0.0) package for R (v2.15.2).

Second, we pre-processed and modeled the Nugen and Express datasets separately to

determine if they yielded similar conclusions about biological effects. The model used was

again: Yi = α0,i + α1,iXi, but here Y is a binary indicator value representing the biological effect

(tumour vs. normal) and Xi is the expression of the fourteen replicate Nugen and Express arrays

for the i-th gene. This model was fitted twice, once to the Nugen data and once to the Express

data.

After model-fitting, an empirical Bayes moderation of the standard error was applied

(294), followed by model-based t-tests to determine if each α1,i was significantly different from

zero. A false-discovery rate (FDR) adjustment was used to control for multiple testing (295).

80

4.3.8 Gene subset division

For downstream analyses we selected three subsets of genes. First, we identified genes

whose mRNA abundance measurements were affected by protocol as those with padjusted < 10-4 in

both protocols. This threshold ensured that the selected genes were statistically significant and

the number was reasonable for downstream analyses. Second, we identified genes that showed

tumour/normal differential expression in data generated using the Express protocol as those with

padjusted < 0.05 in the Express-only model. Third, we identified genes that showed tumour/normal

differential expression in data generated using the Nugen protocol as those with padjusted < 0.05 in

the Nugen-only model.

4.3.9 Functional analysis

Pathway analysis was performed using GOMiner (v2009-09) (296). GOMiner compares

a list of differentially expressed genes to the complete set of genes tested and assesses

enrichment of Gene Ontology (GO) annotations. We used a false discovery rate threshold of 0.1,

generated by 1000 randomizations, and tested all human databases, look-up options, GO

evidence codes and ontologies. This analysis was performed on all three of the gene sets

described above.

4.3.10 Sequence analysis

Sequence annotations (gene length, GC content, untranslated sequence length and

number of exons) for the genes affected in each of the Express and Nugen samples, differentially

expressed genes in the protocol and the total genes on the array were retrieved from Ensembl

BioMart v61 (Human genome assembly GRCh37). The GC content was calculated using the

formula:

G+CA+T+G+C

× 100 (1)

The multiple transcripts for each gene were further processed to compute mean

untranslated (UTR) sequence length using the formula:

UTRgene=1n(∑i=1

n

5 'UTRi)+ 1n(∑i=1

n

3 'UTRi) (2)

81

Where n is the total number of transcripts for a given gene. 5’UTR and 3’UTR represents five-

prime and three-prime untranslated sequence length respectively.

The number of exons across multiple transcripts was averaged to compute the number of

exons for the corresponding gene. The GC content of all genes on the array were compared to the

GC content of genes that were differentially expressed in the protocols and to the genes that were

affected in each of the protocols using two sample unpaired t-test with Welch's correction for

heteroscedasticity. Similarly, the gene length, UTR length and the number of exons between the

above-mentioned groups of gene sets were compared using the unpaired Wilcoxon rank-sum

test. All analyses were performed in the R statistical environment (v2.15.2).

4.3.11 Visualization

Visualization employed the lattice and latticeExtra packages (v0.20-11 and v0.6-24

respectively) for the R statistical environment (v2.15.2). P-value sensitivity plots were produced

by plotting all genes altered at 100 padjusted cut-offs spanning from padjusted = 1 to padjusted = 10-12.

Venn diagrams were created using the VennDiagram package (v1.5.2) (297). All clustering

analyses employed agglomerative hierarchical clustering using complete linkage and Pearson’s

correlation as the similarity metric as implemented in the cluster package (v1.14.3).

4.4 Results

4.4.1 Experimental approach

Our aim was to evaluate the effects of low-input/low-quality RNA amplification

methodologies and to determine if their use introduced any biases (systematic or stochastic) into

microarray data. To achieve this, we selected 12 cervix cancers and two unmatched normal

cervix samples, which were flash-frozen in liquid nitrogen after biopsy. Total RNA was

extracted from each sample and divided into two aliquots. One was reverse transcribed and

amplified using the NuGEN WT-Ovation Pico protocol (henceforth simply “Nugen”), while the

other was processed using the standard Affymetrix 3’ IVT Express protocol (henceforth

“Express”). Each aliquot was hybridized to an Affymetrix HG-U133 Plus 2.0 microarray. We

selected this platform because of its widespread use (2,854 GEO experiments encompassing

77,923 samples at writing). Lastly, to generalize these conclusions we repeated this analysis with

82

two normal kidney samples and one RNA mixture control sample. Figure 4.1 summarizes the

experimental design and analysis workflow.

4.4.2 Nugen and express data are not directly comparable

We selected cervix samples with a broad range of RNA quality, with RNA Integrity

Numbers (RINs) ranging from 2.2 to 8.5 (Table S4.1). Following pre-processing, we first used

unsupervised machine-learning to identify the strongest trends within the dataset (Figure 4.2A).

Samples processed with the Nugen protocol (purple annotation bars) clustered separately from

those processed with the Express protocol (yellow annotation bars). To test the significance of

this observation we calculated the Adjusted Rand Index (ARI), which ranges from -1 to +1, with

higher values indicating non-random associations. Despite using identical samples, the ARI for

protocol bias was 1.0, compared to ARIs of -0.02 for the biological differences between tumour

and normal (red/white annotation bars), 0.26 for different microarray hybridization batches

(grey-scale annotation bars) and -0.07 for RNA quality (white-to-green annotation bars). This

technical bias was not a function of low-expression (Figure S4.1A) or low-variability genes

(Figure S4.1B).

These results showed that the two protocols produced distinctly different data. However,

these differences might not be sufficient to change biological conclusions. For example, if

different protocols created a global scaling artifact that survived data pre-processing, it would

lead to rank-order comparable data that could be evaluated using non-parametric statistical

methods. We therefore assessed if the two protocols yielded identical ordering of samples for

each gene, using Spearman’s correlation coefficient. In total 96% of genes (14,289/14,859) were

positively correlated between the two protocols (ρ > 0), suggesting a broad similarity between

the two methods (Figure 4.2B). However, only 17.6% of genes (2,620) were strongly correlated

(ρ > 0.90). In fact, 4% of genes (570) showed opposite trends between the two protocols (ρ < 0;

Figure 4.2C), and these genes were only weakly biased towards lower mRNA abundances

(Figure 4.2D). Taken together, these data demonstrated that Nugen & Express data were not

directly comparable, even in a rank-order manner.

83

Figure 4.1 Experimental design

Raw array data were preprocessed by Robust Multi-array Average (RMA) algorithm. Three normalized datasets were generated: Nugen, Express and Protocol (all samples). Agglomerative hierarchical clustering was used to visualize the distribution of the mRNA abundances across the protocols. Linear models were then applied to the three datasets to examine the technical/biological effects. Downstream analyses were performed on significant genes to reveal the effects of technical and biological variables. Significant genes were then extracted for pathway and sequence analyses to determine whether differentially expressed genes are biased in gene structure. The LTR algorithm was used to remove technical biases. Lastly, to generalize these conclusions we repeated this experimental approach in two normal kidney samples and one Michigan reference RNA sample (RNA control). Experiments performed in cervix data and kidney data were indicated as 1 and 2 respectively.

84

Figure 4.2 Correlations between technical replicates

Following pre-processing, we studied the relationship between Nugen-Express effects and other aspects of the microarray. A) On this sample vs. sample heatmap, the white-to-blue colour bar reflects the range of Pearson's correlations. The four-column annotation bars indicate array protocol, biological effects (tumour vs. normal), hybridization batch and RIN. The strong separation by protocol highlights the magnitude of technical artifacts. B) A histogram of gene-wise Spearman’s rho (ρ) between Nugen and Express samples was skewed to the right, suggesting most ProbeSets show similar trends. C) However a cumulative plot shows that only a small fraction of genes were highly correlated. D) Correlation between Nugen and Express protocols was not strongly associated with signal intensity (a proxy for mRNA abundance).

85

4.4.3 Protocol introduces systematic bias

We next sought to determine if the RNA amplification protocol creates a systematic bias

that preferentially affects specific genes, or a stochastic one that altered all genes (or a random

subset thereof). This is critical because if a reproducible and limited set of genes is affected by

protocol, these can be excluded from comparative analyses and data from the two protocols can

be safely integrated.

We performed gene-wise linear-modeling to assess protocol effects. Stochastic effects

would be expected to yield few reproducible differences between matched samples. Instead, we

identified large numbers of genes altered in a reproducible manner by protocol (Figure 4.3A).

The dashed line indicates a 1% false-discovery rate; at this stringent threshold, 72% of the array

(10,703 genes) was altered by protocol. These effects were very large in magnitude: 86 genes

exhibit at least a 10-fold difference between protocols. Both their magnitude/direction (Figure

4.3B) and statistical significance (Figure 4.3C) were independent of expression level. This

systematic bias was only weakly related to the poor correlations outlined in Figure 4.2B: ~24%

(620/2,620) of the highly correlated genes (ρ > 0.9) differed by at least two-fold between

protocols. Interestingly, negatively correlated genes showed generally smaller magnitude

differences (Figure 4.3D). Thus amplification protocols systematically bias the majority of the

transcriptome in a gene-specific manner.

4.4.4 Protocol noise is larger than biological signal

To put the magnitude of protocol-specific changes in context, we compared this effect to

the tumour/normal differential expression between normal and malignant cervix tissues. We

compared linear models that were fitted separately on the Nugen and Express data to the

technical effects described above. Independent of the statistical threshold, an order of magnitude

more genes were altered by protocol than by cancer, with the green curve for the protocol effect

shifted far to the right of the red and blue curves in Figure 4.4A. Interestingly, protocol changes

led to artifactual up- and down-regulation in equal proportions, whereas an excess of down-

regulated genes were observed in cancer (Figure 4.4B), consistent with the reduction of

transcriptomic complexity in malignancies (298).

86

Figure 4.3 Specific genes are affected by protocol bias

This figure reveals how statistically significantly differentially-expressed genes were altered by protocols. Fold change between Nugen and Express is in log2 scale, describing the changes of mRNA abundances between Nugen and Express. P-values were calculated with FDR-adjusted model-based t-tests. A) Large numbers of genes were altered in a reproducible manner by protocol changes. These effects were independent of expression level, as high magnitude (B) and highly significant protocol biases (C) were evident in genes at all signal intensities. The degree of this systematic bias was independent of the previously observed correlation differences (D).

87

Figure 4.4 Differentially expressed genes between biological and technical replicates

Following a general-linear model, we identified a number of genes at various p-value thresholds. This figure shows how the genes were affected by protocol. A) The x-axis is p-value threshold ranging from 1 to 10-12. The y-axis is all genes that are differentially expressed. More genes were altered by technical differences (green curve) than by biological differences (blue and red curves) between samples. B) This figure shows the proportion of up-regulated genes at different p-value thresholds.

88

4.4.5 Protocol difference biases specific biological functions

Our data demonstrated that changing the amplification protocols introduce systematic

artifacts that affect specific genes. To determine if these changes will bias pathway analyses we

subjected the set of genes showing differential signal intensity to gene ontology (GO) enrichment

analysis. We again put the protocol effects in context of true biology driven by tumour-normal

differences. A significant fraction of pathways affected by protocol bias were also altered

between normal and tumour cells, and again protocol bias dramatically overwhelms biological

signal (Figure S4.2).

4.4.6 Sequence characteristics do not explain protocol bias

Because only some genes were affected by different amplification protocols, and because

these genes were affected in different ways, we hypothesized that sequence-specific

characteristics were involved. We considered four gene features: GC-content, number of exons,

gene length and untranslated region (UTR) length. We compared genes showing protocol

differences and those showing tumour-normal differences to the overall gene cohort. Genes

affected by protocol were not unusual in their GC content (PP vs A = 1.83 x 10-1; Figure 4.5A),

number of exons (PP vs A = 3.81 x 10-1; Figure 4.5B), length (PP vs A = 7.80 x 10-1; Figure 4.5C) or

UTR length (PP vs A = 3.83 x 10-1; Figure 4.5D), indicating that the protocol bias was not be

explained by obvious sequence characteristics. However, we observed a significantly increased

number of exons in tumour-associated genes (PTE vs A = 4.88 x 10-2, PTN vs A = 2.00 x 10-2; Figure

4.5B), which to our knowledge has not previously been described. Interestingly, genes identified

as tumour-normal differentially expressed by the Nugen protocol had a higher GC-content than

those identified using the Express protocol (red curve, Figure 4.5A), suggesting that the Nugen

protocol may be more sensitive at amplifying genes with high GC-content. Traditionally, there

have been limitations in the amplification of sequences with high GC-content due to higher

affinity bonds between guanine and cytosine nucleotides, compared to those found between

adenine and thymine (299). There are clearly many other sequence-based factors that might

influence protocol bias, but the most common ones do not appear to play a role.

89

Figure 4.5 Differing characteristics of significantly differentially expressed genes

We were interested to determine whether differentially expressed genes between protocols differed on any structural characteristics of a gene. A) This figure shows the distribution of GC content in significantly differentially expressed genes between the different protocols, and Nugen and Express tumour and normal samples. Unpaired t-tests calculated p-values of the contrasts between distributions and found a significant difference between all genes (black curve) and Express genes (blue curve). We then studied the distribution of the number of exons (B), gene-length (C) and UTR lengths (D) of significantly differentially expressed genes in protocol, Nugen, and Express. The number of exons in all genes (black curve) versus the Express genes (blue curve) and Nugen genes (red curve) differed significantly. Comparisons for these features

90

(number of exons, gene-length and UTR length) were conducted using unpaired Wilcoxon rank sum test.

91

4.4.7 Batch effects adjustment partially mitigates protocol bias

Lastly, we wished to determine whether the technical artifacts created by protocol

difference could be mitigated. LTR (Linear Transformation of Replicates) is a linear-modeling-

based method that learns the relationship between different microarray batches from replicate

hybridizations (293). Normalized data from half of the arrays were used to train an LTR model,

which was used to adjust the remainder of the dataset. LTR removal of noise diminished the

effect of protocol bias (Figure 4.6), reducing the ARI for protocol effects from 1.0 to -0.04.

However a significant number of genes remained poorly correlated between the two protocols,

with 11% (1,643) having negative correlations.

4.4.8 Replication analysis

To generalize our findings, we applied the same approach to normal kidney samples and

an RNA reference sample. Clustering confirms a strong separation by protocol (Figure S3A,

purple/yellow annotation, ARI = 1). The RNA mixture control sample was very highly-

correlated between Nugen and Express (ρ = 0.9, p < 2.2 x 10-16, Figure S4.4A), while the two

kidney samples were strongly correlated (ρ = 0.71 and 0.78, p < 2.2 x 10-16, Figure S4.4B and C).

We then examined the correlation of each ProbeSet on the microarray. We performed

Pearson's correlations between samples processed with the Nugen protocol and those processed

with the Express protocol. Most ProbeSets were positively correlated between the two protocols

(77% of ProbeSets with ρ > 0, Figure S4.3B), and 36% (6,840/18,988) were strongly correlated

(ρ > 0.9, Figure S4.3C). Moreover, 16% (1,098 /6,840) of the most strongly correlated genes in

the kidney dataset (ρ > 0.9) were also strongly positively correlated (ρ > 0.9) in the cervix dataset

(Figure 4.5). This degree of concordance was independent of expression level (Figure S4.3D).

These results thus suggest that amplification bias is a fundamental property of the preparation

procedure with only modest sample-specific effects.

92

Figure 4.6 Application of LTR algorithm to remove platform specific biases

Following an LTR-adjustment, we visualized the Pearson's correlations between testing samples (A) before and (B) after LTR-adjustment for the validation cohort. On each of sample vs. sample heatmaps, the white-to-blue colour bar reflects the range of Pearson's correlations. The four-column annotation bars indicate protocol, biological effects, batch and RIN, respectively. The strong separation by protocol in Figure 4.6A highlights the magnitude of technical artifacts, and its disappearance in Figure 4.6B suggested that the LTR algorithm removed protocol bias.

93

4.5 Discussion

mRNA microarrays have become an important high-throughput technology in mRNA

expression analysis. Despite their extensive characterization, their use as clinical diagnostic tools

remains limited due to the challenges of amplifying low amounts of significantly degraded RNA

in clinical samples. Recent studies have focused on protocols that address these challenges to

enable routine microarray profiling of clinical samples (300, 301). Furthermore, the emergence

of large consortia such as the TCGA and ICGC make it increasingly common to integrate data

generated at multiple centres using divergent protocols. It is therefore critical to assess the

sources of error so that they can be reported, controlled for, and adjusted statistically. However,

to date, very few efforts have been made towards evaluating RNA amplification protocols for

expression profiling.

Previous evaluations examined the reproducibility of different protocols by investigating

correlations of RNA amplification methods using 8 colon cancer biopsy samples isolated using

laser capture microdissection (LCM) (302), or picogram amounts of input RNA from two cell

lines (303). Some researchers evaluated RNA integrity and purity (303, 304), and suggested a

minimal amount of input RNA (250 pg) for amplification (303). However, none of these studies

were systematically comparing accuracy and bias of these protocols, but rather focused on the

agreement of different approaches rather than the integration of data generated in diverse labs

and from multiple protocols. Our work fills this gap by evaluating the performances of a low-

input/low-quality protocol, the NuGEN WT Ovation Pico kit, with a standard protocol, the

Affymetrix 3’ IVT Express kit. We investigated 17 unique samples: 2 normal cervix samples, 12

cervix tumour samples, 2 normal kidney samples and 1 RNA mixture control sample. The three

key findings from this work are: 1) Nugen and Express data are not directly comparable; 2)

technical biases can be partially removed; 3) a template for the evaluation and analysis of

different sample-handling protocols is proposed. We discuss these in turn.

The most prominent finding of this evaluation was the inconsistency of the two

amplification protocols, given that the two amplification protocols both used linear

amplification. For example, only 17.6% of genes were highly correlated (ρ > 0.9) between the

two protocols, and over 24% of these highly correlated genes showed a difference between the

two protocols of over 2-fold. The bias was not a simple global scaling artifact and was

94

independent of expression level. These results implied that comparing studies generated with

different protocols or mixing protocols within a study would affect differentially expressed

genes, creating artificial disparities and elevating false-positive rates. These findings were

consistent with previous reports that showed microarray data being affected by the RNA

amplification protocol used, and that data generated with the same microarray platform using

different protocols were not directly comparable (305, 306).

Moreover, we found that the technical differences (whether arrays were run with Nugen

or Express reverse transcription/amplification protocols) had a greater effect on differential

expression of genes than biological differences between samples (whether they were tumour or

normal). These discrepancies in the two protocols artificially affected specific biological

pathways and could affect the apparent biological significance of analyses. For these reasons, it

is important to utilize only one amplification methodology within a study. However, if it is

unavoidable to use multiple methodologies, there are steps that can be taken to mitigate the

technical bias. Firstly, it is imperative to use the same type of amplification, either linear or

exponential. Technical bias between two linear-amplification protocols (Nugen and Express) was

greater in magnitude than true biological differences. Hence, it is expected that the technical

artifacts would be significantly greater with mixed types of amplification protocols (linear and

exponential). Secondly, sophisticated algorithms are needed to remove technical artifacts before

interpreting the biological significance of the dataset. In this study, we applied a linear-

modeling-based method, LTR, to remove technical artifacts. The ARI was decreased from 1.00

(highly non-random clustering) to -0.04 (random association) after LTR-adjustment. However,

even after LTR-adjustment, about 11% of genes had negative correlations, suggesting technical

biases were only partially removed. There is a clear need for more sophisticated batch-effect

removal algorithms.

Lastly, this work provides a template for the evaluation and analysis of different sample-

handling protocols and can be extended to other areas. Before conducting any analysis on the

integrated multi-protocol dataset, two generalized steps are suggested: First, data normalization

remains essential, with robust multi-array average (RMA) remaining a typical method for

normalizing Affymetrix data. Second, data cleaning is needed to remove technical bias from the

normalized data. Explicit batch-effect reduction approaches can be used, with some such as LTR

(293) specializing in taking advantage of technical replicates, while others such as ComBat (307)

95

employing empirical Bayes models to handle unknown error structures. Unsupervised machine-

learning methods are valuable for validating whether and how much technical bias is removed,

particularly when combined with metrics of clustering-randomness such as ARI, which is

infrequently used in bioinformatics. Lastly, linear model analysis can be utilized to fit the

integrated data with batch-effect terms to explicitly model technical artifacts.

Many techniques and tools are available to measure mRNA, miRNA and DNA

abundances, and this template demonstrates how to evaluate the protocol performance. Protocol-

specific changes are expected to increase in magnitude as next-generation sequencing

techniques, which require larger amounts of input material, become more widely employed,

making careful consideration of this issue particularly important.

96

Table S4.1 Sample descriptions (Data: cervix data)

Legends: SampleID Unique ID for sample used in Nugen or Express Amplification Protocol Nugen or Express RIN RNA Integrity Number BatchNumber Three different batches, Batch 1, Batch 2, and Batch 3 Type Sample type: Tumour vs. Normal

Sample ID Amplification

Protocol RIN BatchNumber Type

1 Express 5.4 Batch2 Tumour 1 Nugen 5.4 Batch3 Tumour 14 Express 6.5 Batch2 Tumour 14 Nugen 6.5 Batch3 Tumour 18 Express 6.0 Batch2 Tumour 18 Nugen 6.0 Batch3 Tumour 27 Express 7.1 Batch1 Tumour 27 Nugen 7.1 Batch1 Tumour 38 Express 7.9 Batch1 Tumour 38 Nugen 7.9 Batch1 Tumour 44 Express 8.5 Batch1 Tumour 44 Nugen 8.5 Batch1 Tumour 45 Express 6.7 Batch2 Tumour 45 Nugen 6.7 Batch3 Tumour 50 Express 7.0 Batch1 Tumour 50 Nugen 7.0 Batch1 Tumour 57 Express 2.2 Batch1 Tumour 57 Nugen 2.2 Batch1 Tumour 58 Express 4.9 Batch1 Tumour 58 Nugen 4.9 Batch1 Tumour 67 Express 7.4 Batch2 Tumour 67 Nugen 7.4 Batch3 Tumour 86 Express 7.4 Batch2 Tumour 86 Nugen 7.4 Batch3 Tumour

3856 Express 7.1 Batch2 Normal 3856 Nugen 7.1 Batch3 Normal 3926 Express 6.0 Batch2 Normal 3926 Nugen 6.0 Batch3 Normal

97

Figure S4.1 Distributions of Selected Genes (Dataset: Cervix data)

Following pre-processing, we visualized the genes that passed the expression-filtering selection (mRNA abundances > 6, Figure S4.1A) and variance-filtering selection (variance > 1, Figure S4.1B). On each heatmap, rows represent samples and columns represent genes. The white-to-blue colour bar at the bottom indicates scaled signal intensity of a gene. The four-column row-annotation displays four aspects: protocols, biological effects, batch numbers, and RNA Integrity numbers (RIN). The first column indicates the protocol (Nugen in purple; Express in yellow), the second indicates whether a sample is derived from a tumour (red) or a normal (white), the third indicates batch numbers, and the fourth is the discretized RINs. In this case, we see that the samples appear to cluster primarily by protocol. Most Nugen (purple) samples were clustered together, as well as Express (yellow) samples.

98

Figure S4.2 Overlap of significant GO terms by Protocol or sample type

Overlap of significant GO terms in a three-way interaction between biological and technical cohorts. We estimated the number of GO terms affected by the three groups at padjusted < 0.05. More GO terms were significant in Nugen than Express, and half of the GO terms that were significant in Express were also differentially expressed in Nugen (approximately 50%).

99

Figure S4.3 Correlations between technical replicates (Data: kidney data)

Following pre-processing, we revealed intra-sample relationships to study how close the samples are. We calculated pair-wise Pearson's correlations of the two kidney samples and then applied agglomerative hierarchical clustering on the correlations. A) On this heatmap, rows and columns represent samples. The white-to-blue colour bar at the bottom reflects the range of correlations [0, 1]. The one-column row-annotation displays the protocol (Nugen in purple and Express in yellow). B) Pearson’s correlation and adjusted p-values were calculated for the expression level of each gene between Nugen and Express for 2 kidney samples and one RNA mixture control sample. C) Cumulative rho tells us what percentage of ProbeSets has a specific correlation. D) Genes with a positive rho between two protocols and with high signal intensity are in the top right corner of the plot.

100

Figure S4.4 Distributions of mRNA abundances (Data: kidney data)

We examined the distributions of mRNA abundances for: (A) The RNA mixture control sample and two kidney samples (B and C) between Nugen (x-axis) and Express (y-axis). Pairwise Spearman’s rank-order correlation was calculated on each ProbeSet between Nugen and Express.

101

Figure S4.5 Correlation comparisons between cervix and kidney data.

To examine whether highly correlated genes in one dataset are also consistent in another, we compared the correlations of each ProbeSet on the microarray between cervix data and kidney data. X-axis is the Pearson’s correlations from kidney data and Y-axis is the Spearman’s correlations from cervix data. Highly correlated genes in both datasets can be found on the right top corner.

102

Figure S4.6 Distribution of Y-chromosome filter process for cervix data

After RMA normalization, we performed Y-chromosome filter process to remove the noise genes. As data is from Cervix samples, we examined the distribution of genes on chromosome Y (red curve) and genes not on chromosome Y (blue curve) for differentially expressed genes in A) protocol, B) Nugen, and C) Express. Genes on chromosome Y showed low mRNA expression (intensity < 5). This suggested that mRNA with low mRNA expression (<5) was mostly noise, as the mRNA abundance was same as the genes from Y chromosome. Hence, we performed Y-chromosome filter process to remove those genes with mRNA abundance for all cervix samples lower than 5.

103

CHAPTER 5: SYSTEMATIC COMPARISON OF PLATFORMS FOR MICRO-RNA PROFILING

The data in this chapter have been prepared as a manuscript for submission: How C*, Yan R*, Waggott D, Bruce J, Yin S, Hui ABY, Chong LC, Harding NJ, Liu FF, and Boutros PC. *These authors contributed equally to this work.

104


Background: MicroRNAs (miRNAs) are small, non-coding RNAs that are involved in crucial

biological processes, including development, proliferation and apoptosis. Altered miRNA

expression has been described for numerous diseases, and various methods for miRNA

expression profiling have been used for biomarker studies. Here, we evaluated two platforms for

miRNA expression profiling, the Applied Biosystems TaqMan Low Density Array Human

MicroRNA Array (TLDA) and the NanoString nCounter miRNA Expression Assay

(NanoString), and both methods were also compared to a gold standard method, single-well

quantitative real-time PCR (PCR).

Methods: Total RNA was extracted from 79 cervical cancer flash-frozen punch biopsies, 87

cervical cancer formalin-fixed paraffin-embedded (FFPE) biopsies, 144 nasopharyngeal

carcinoma FFPE biopsies, and 40 breast cancer FFPE biopsies. The miRNA abundances were

then measured by three different platforms, TLDA, NanoString and PCR. These data were

normalized and a series of correlation analyses were then conducted on the common sample- and

gene-pairs, to reveal whether genes were biased by the technical differences.

Results: Data generated using TLDA and NanoString were weakly, but positively correlated,

suggesting that TLDA and NanoString resulted in distinct expression profiles. PCR was found to

be comparable to TLDA and NanoString, but gene-wise comparisons were more similar between

PCR vs. TLDA than PCR vs. Nanostring.

Conclusions: Experiments performed using different platforms for measuring miRNA

expression cannot be directly merged, suggesting that it is essential to keep platform consistency

when comparing miRNA data.

5.2 Introduction

Micro-RNAs (miRNAs) are small (~22 nucleotides), non-coding RNAs that regulate

gene expression by binding to mRNA transcripts in a sequence-specific manner (2). Since

miRNAs were initially reported in C. elegans in 1993 (6), 2042 human miRNAs have been

identified thus far (miRBase Release 19) and the number continues to steadily increase. Altered

105

miRNA expression has been shown to be important in many human cancers (69), which is

unsurprising given that miRNAs are involved in crucial biological processes, including

development, proliferation and apoptosis (2, 253).

Due to their small sizes and considerable stability, miRNAs can be easily extracted from

both frozen and formalin-fixed paraffin-embedded (FFPE) samples (96, 97). MicroRNA

expression profiling of clinical samples has been shown to be particularly useful for

distinguishing between various types of cancers (86), and classifying cancer subtypes (93, 94),

including poorly-differentiated tumours (95). Furthermore, miRNA expression profiling has led

to the identification of disease-specific miRNA signatures associated with diagnosis (126, 308,

309) and prognosis (119, 127, 310) for various solid and haematological malignancies.

The Applied Biosystems TaqMan Low Density Array Human MicroRNA Array (TLDA)

and NanoString nCounter miRNA Expression Assay (NanoString) are two commonly used

methods for miRNA expression profiling. We sought to compare these two miRNA platforms in

cervical, breast and nasopharyngeal (NPC) cancers. Both platforms were also compared to

single-well quantitative real-time PCR (PCR), which has long been considered the gold standard

for nucleic acid quantification due to its specificity and sensitivity (311). For PCR quantification,

we used the Applied Biosystems TaqMan MicroRNA Assays, which specifically measure mature

miRNAs using stem-loop reverse transcription followed by TaqMan PCR analysis. The high

specificity and sensitivity of this method allows it to distinguish between miRNAs that differ by

a single nucleotide (114).

The TLDA method uses multiplex target-specific reverse transcription, where a panel of

up to 378 miRNAs and controls are reverse transcribed in a single reaction, using a mixture of

miRNA-specific stem-loop primers. After the RT step, the cDNA is loaded into a 384-well

microfluidic card containing TaqMan MicroRNA Assays specific for the panel of miRNAs and

controls, and the quantity of each miRNA is measured using TaqMan chemistry. A more recently

developed method, NanoString, provides a direct count of miRNAs, by fluorescently tagging

individual miRNA molecules and then optically scanning and counting the tagged miRNA

molecules. This method is a direct measurement of miRNA levels, without bias from

amplification or enzymatic reactions (312).

106

In our analyses of platforms for miRNA expression profiling, we conducted a cross-

platform comparison between TLDA, NanoString and PCR, using FFPE and frozen samples

from three cancer types (Breast cancer, Cervical cancer and NPC). We aimed to determine if the

choice of platform had a significant impact on the detection of differentially-expressed miRNAs.

Drawing comparisons between protocol differences and biological differences, we sought to

investigate both the biological and technical effects, and the potential influence these could have

on miRNA expression profiling studies.


5.3.1 Samples and sample processing

Flash-frozen punch biopsies were obtained from 79 patients with cervical cancer prior to

treatment with standard chemo-radiotherapy between 2000 and 2007, inclusive. After removal

from the patient, the specimens were placed in a storage medium (optimal cutting temperature

compound) for histopathologic examination, then flash-frozen in liquid nitrogen. H&E-stained

tissue sections were cut from the optimal cutting temperature-embedded material to ensure

samples contained at least 70% tumor cells. Normal cervix tissues obtained from eleven patients

who underwent hysterectomy for benign causes were also flash frozen in OCT and served as

normal comparators. Two sections of 50-micron thickness were cut from the frozen cervix


RNA Purification Kit (Norgen Biotek), according to the manufacturer’s instructions for frozen

tissue samples.

Diagnostic formalin-fixed paraffin-embedded (FFPE) blocks were collected from 87

human cervical cancer tumours between 1997 and 2000, inclusive. All samples contained at

least 70% malignant epithelial cells, or were macrodissected prior to RNA purification. Nine

normal cervix FFPE tissues obtained from patients who underwent hysterectomy for benign

causes served as normal comparators. Ten sections of 5-micron thickness were cut from the

FFPE cervix tissues and places in a nuclease-free microtube. Total RNA was isolated using the

Norgen Total RNA Purification Kit, according to the manufacturer’s instructions for FFPE tissue

samples.

107

FFPE primary biopsies were collected from 144 NPC patients diagnosed from 1993 to

2000. A representative section from each biopsy was stained with hematoxylin and eosin stain,

and reviewed to determine regions with malignant epithelium for macrodissection. Total RNA

was isolated from the macrodissected samples, using the RecoverAll Total Nucleic Acid

Isolation Kit for FFPE (Ambion Inc.) according to the manufacturer’s instructions for

enrichment for small RNA species.

Thirty-four FFPE lumpectomy blocks and six reduction mammoplasty specimens were

macrodissected and RNA was isolated as described previously described (96).

5.3.2 Data description

This paper studied three types of cancer patients: breast cancer, cervical cancer and NPC.

Detailed data descriptions are listed in Tables 5.1, 5.2, and 5.3. Herein, we refer to the miRNAs

measured by the platforms as “genes”.

Breast cancer FFPE samples were collected and gene expression was measured using

three platforms: NanoString nCounterTM Human miRNA Expression Assay (NanoString)

(v1.6.0), Applied Biosystems TaqMan Low Density Array Human MicroRNA Array (TLDA)

(v1.0), and single-well quantitative real-time PCR (PCR) using Applied Biosystems TaqMan

MicroRNA Assays.

Cervical cancer data included FFPE and Frozen samples. Gene expression was measured

by two platforms: NanoString nCounterTM Human miRNA Expression Assay (NanoString)

(v1.7.0) and TLDA Human MicroRNA A Array (v2.0).

NPC data included FFPE samples and gene expression was measured by two platforms:

NanoString nCounterTM Human miRNA Expression Assay (NanoString) (v1.7.0) and TLDA

Human MicroRNA Array (v1.0).

108

Table 5.1 Data descriptions for breast cancer datasets

FFPE

NanoString 742 genes * 16 samples (4 normals, 12 tumours)

TLDA 365 genes * 40 samples (6 normals, 34 tumours)

PCR 20 genes * 40 samples (7 normals, 33 tumours)

Table 5.2 Data descriptions for cervical cancer datasets

FFPE Frozen


753 genes * 12 samples (0 normals, 12 tumours)


378 genes * 90 samples (11 normals, 79 tumours)

Table 5.3 Data descriptions for NPC datasets

FFPE



109

5.3.3 Data pre-processing

Prior to any further analyses, data pre-processing was performed to remove the technical

assay bias (i.e., data normalization) and extract the common genes and samples between

platforms (i.e., gene and sample mapping).

We used standard ∆CT methods to normalize gene abundances measured from TLDA and

PCR. We subtracted the gene abundance (CT) from the mean of the control gene abundance (CC)

and then calculated the normalized gene abundance as: 2 . The reason for using

the log2 -space was to be consistent with the amplification processes in PCR and TLDA.

To normalize gene abundances from NanoString, we applied the R package

'NanoStringNorm' (v1.1.11) (255). The gene abundances were adjusted based on the sum of all

the samples, which was able to normalize the majority of the technical variations (i.e., NS1). We

further normalized gene abundances using the mean of all samples (i.e., NS2). To evaluate the

platform performance, we extracted the common samples and genes from every pair of

platforms, since not all the genes and samples were measured in all platforms.

5.3.4 Unsupervised machine-learning

An unsupervised machine-learning hierarchical clustering analysis was applied to the

normalized mapped gene abundances, to understand how gene abundances were distributed

across the platforms. The divisive hierarchical clustering algorithm DIANA was used to cluster

all the genes, as implemented in the cluster package (v1.14.3) for the R statistical environment

(v2.15.1), and heatmaps were visualized using the lattice (v0.20.10) and latticeExtra (v0.6-24)

packages. Pearson's correlation was used as the distance metric.

5.3.5 Statistical analysis

Correlation analyses were first performed to measure the impact of platform on gene

expression. Pair-wise comparisons between platforms were calculated using Spearman's rank

correlation (rho, ρ). We measured three different correlations to understand different aspects of

the platform effects:

1. Overall correlations: The correlations of gene expression per sample and gene

between two platforms were calculated. This gives an overall evaluation of two

110

platforms regardless of the number of genes and samples.

2. Sample-wise correlations: The correlations of gene expression per sample between

two platforms were examined. The correlations depend on the number of genes per

sample involved.

3. Gene-wise correlations: Correlations of gene expression per gene between two

platforms were examined. The correlations depend on the number of samples per

gene involved.

Next, to study the protocol effects on differentially-expressed genes, we calculated the

FDR-adjusted p-value and fold change between tumour and normal for each pair of platforms.

Fold change was calculated as: mean(tumour) / mean(normal) in log2-space. A Spearman’s

correlation was also calculated based on the fold change and p-value, to measure the correlations

between differentially-expressed genes.

5.3.6 Visualization

For visualization, we employed the lattice and latticeExtra packages (v0.19-33 and v0.6-

33 respectively) for the R statistical environment (v2.13.2). Venn diagrams were created using

the VennDiagram package (v1.1.1) (297). All clustering analyses employed agglomerative

hierarchical clustering using complete linkage, and Pearson’s correlation as the distance metric,

as implemented in the cluster package (v1.14.1). Heatmaps were created using the lattice (v0.19-

33) and latticeExtra (v0.6-19) packages.

5.4 Results

5.4.1 Experimental design

The experimental design for this study is outlined in Figure 5.1. Briefly, tissue samples

were collected from patients with: breast cancer (FFPE), cervical cancer (FFPE and frozen), or

NPC (FFPE). The gene abundances were measured using: NanoString nCounterTM Human

miRNA Expression Assay (NS) (v1.6.0 for breast cancer; v1.7.0 for cervical cancer and NPC),

Applied Biosystems TaqMan Low Density Array Human MicroRNA Array (TLDA) (TLDA

microRNA array v1.0 for breast cancer and NPC; TLDA microRNA A Array v2.0 for cervical

111

cancer), and qRT-PCR using Applied Biosystems TaqMan MicroRNA Assays (breast cancer

only). The data were pre-processed to find the common samples and genes across the three

platforms. The filtered genes and samples were subjected to a series of statistical analyses to

reveal the platform relationships. Survival analysis was performed on the cervical cancer data, to

understand the impact of platform on the biologically-significant genes. Lastly, these

biologically-significant genes were subjected to sequence and pattern analysis, to determine the

effect of sequence characteristics.

5.4.2 Biological differences are detectable by platforms

It is important to assess data quality before any analysis. Thus, after normalization,

miRNA abundances for every pair of samples across protocols were subjected to unsupervised

machine-learning. As expected, all platforms were able to distinguish biological differences. For

example, almost all tumour and normal samples were tightly clustered to each other, regardless

of the protocol (Figure 5.2A and B), suggesting that the platforms were sensitive to biological

differences between cancer and normal samples. We noticed that about 5% to 10% of normalized

miRNA abundances from NanoString in breast cancer samples were exceptionally high; for

better visualization, these miRNAs were removed from these heatmaps.

5.4.3 Data from NanoString and TLDA are significantly different

To identify the strongest trends in the miRNA abundances between TLDA and

NanoString, we first employed an unbiased machine-learning approach (313). The correlations

of the miRNA abundance profiles of breast and cervical samples were calculated between each

pair of arrays, generating a correlation-matrix. The strongest trend within these two matrices was

the separation of samples processed with the TLDA platform (yellow annotation bars) vs. those

processed with the NanoString platform (blue annotation bars) (Figure 5.2A for breast samples;

Figure 5.2B for cervical FFPE samples). The effects of platform were stronger than the

biological differences between cancer and normal samples (red/white annotation bars). The

strong technical effects were also observed in cervical frozen samples (Figure 5.2C) and NPC

samples (Figure 5.2D). In these two latter cases, cancer samples clustered perfectly by protocol

with the absence of normal samples. To test the significance of this observation, we calculated

the Adjusted Rand Index (ARI), which ranges from -1 to +1, with higher values indicating non-

random associations. The higher the ARI, the more similar two clustering patterns are. An ARI

112

of 1.0 indicates perfect partitioning and a non-positive ARI (ARI ≤0) indicates independent or

random partitioning. The ARI for platform bias was 1.0, indicating that the TLDA and

NanoString datasets contained significantly different expression profiles.

To quantify these similarities or dissimilarities, we calculated pair-wise Spearman's

correlations on the miRNA abundances between protocols. The correlations between NanoString

and TLDA were: 0.32 for breast FFPE samples (Fig 5.3A), 0.32 for cervical FFPE samples (Fig

5.3B), -0.02 for NPC FFPE samples (Fig 5.3C), and 0.42 for cervical frozen samples (Fig 5.3F),

again suggesting the weak correlation between miRNA abundances from these two platforms.

The exceptionally low correlation (ρ = -0.02) for the NPC samples (Fig 5.3C) could be due to the

smaller number of samples and miRNAs that were common between the two platforms, with

only 194 miRNAs and 12 samples (= 2328 sample pairs), which is 56% the sample pairs

compared to the cervical frozen samples (345 miRNAs and 12 samples in common), and 7% of

sample pairs compared to the cervical FFPE samples (345 miRNAs and 96 samples in common).

5.4.4 PCR is more similar to TLDA than NanoString

We next compared the miRNA abundances between PCR and NanoString (Figure 5.4A

and B), and between PCR and TLDA (Figure 5.4C and D) for the breast samples. Platform

effects between PCR (purple annotation bars) and NanoString (blue annotation bars) were not as

strong as the biological differences between cancer and normal samples (red/white annotation

bars) (Figure 5.4B). Furthermore, the differences between PCR (purple annotation bars) and

TLDA (yellow annotation bars) were weaker compared to the biological differences (red/white

annotation bars) (Figure 5.4D). These results suggest that the datasets produced from either

NanoString or TLDA, were more similar to the dataset produced from PCR than they were to

each other.

In contrast to the low correlations observed between NanoString and TLDA for all

sample types, the correlations on the miRNA abundances between PCR and NanoString (ρ = 0.6;

Figure 5.3D), and between PCR and TLDA (ρ = 0.7; Figure 5.3E) in the breast FFPE samples

were significantly higher. Sample-wise correlations were high for both PCR vs. NanoString (ρ =

0.75, p = 4.32 x 10-44; Figure 5.5A), and PCR vs. TLDA (ρ = 0.70, p = 6.69 x 10-58; Figure 5.5B).

Despite this, gene-wise correlations were not comparable, with almost 50% of genes being

113

negatively correlated between PCR and NanoString (ρ < 0; Figure 5.5C), whereas over 90% of

genes were positively correlated between PCR and TLDA (ρ > 0; Figure 5.5D).

5.4.5 Fold-changes are poorly correlated between NanoString vs. TLDA, but differentially-expressed miRNAs are highly correlated across platforms

We investigated the relationship of tumour/normal fold-changes between NanoString and

TLDA data. This analysis was performed using the breast FFPE and cervical FFPE datasets; the

cervical frozen and NPC FFPE datasets were excluded due to their small numbers of samples.

Fold-changes were positively- but lowly-correlated between NanoString and TLDA in the breast

FFPE samples (ρ = 0.33; Fig 5.6A) and cervical FFPE samples (ρ = 0.37; Fig 5.6B). However,

miRNAs that were significantly differentially-expressed between tumour and normal (fold-

change ≥ 2.0 in both platforms; upper right section of plots in Fig 5.6C and D), were generally

significantly highly-correlated between the two platforms (ρ > 0.5 and PRho ≤ 0.05). Altogether,

these results show that although tumour/normal fold-changes were lowly-correlated between the

NanoString and TLDA datasets, significantly differentially-expressed miRNAs were

significantly highly-correlated between the two platforms.

114

Figure 5.1 Experimental design

FFPE tissue samples were collected from three types of cancer: breast (1), cervical (2) and NPC (4). Frozen tissue samples were collected from cervical cancer (3). Total RNA was extracted from each sample and gene abundances were measured using three different platforms: a) NanoString nCounterTM Human miRNA Expression Assay (NanoString, or NS), b) Applied Biosystems TaqMan Low Density Array Human MicroRNA Array (TLDA), and c) qRT-PCR using Applied Biosystems TaqMan MicroRNA Assays (PCR). These data were normalized and common samples and genes were extracted across the three platforms. The common genes and samples were subjected to unsupervised machine-learning to visualize the distribution of the gene abundances across the platforms. A series of correlation analyses were then conducted to investigate the technical and biological effects. To determine how the technical biases affected the biologically-significant genes, statistical analyses were performed and Spearman's correlations were calculated based on the tumour/normal fold changes. Lastly, survival analysis was conducted to understand the effect of platform on the prognostic genes. Datasets involved in the analysis are indicated with 1-4.

115

Figure 5.2 Correlations between technical replicates (TLDA vs. NanoString)

Following pre-processing, the relationships between platform effects (TLDA vs. Nanostring) and biological effects (tumour vs. normal) were examined for the: A) breast FFPE dataset, B) cervical FFPE dataset, C) cervical frozen dataset, and D) NPC FFPE dataset. On each sample vs. sample heatmap, the white-to-blue colour bar reflects the range of Pearson’s correlations. The two-column annotation bars indicate platform (TLDA in yellow, NanoString in blue) and biological effects (tumour in red, normal in white). The strong separation by platform highlights the magnitude of technical artifacts.

116

Figure 5.3 Distributions and correlations of miRNA abundances across platforms

Following pre-processing, we compared the distributions of miRNA abundances and calculated the correlations between the NanoString and TLDA datasets for A) breast FFPE samples, B) cervical FFPE samples, C) NPC FFPE samples, and D) cervical frozen samples. We also examined the distributions and correlations of miRNA abundances in the PCR dataset in comparison with the D) NanoString and E) TLDA datasets for breast FFPE samples. P-values were calculated using FDR-adjusted model-based t-tests.

117

Figure 5.4 Correlations between technical replicates with PCR

Following pre-processing, we studied the relationships between NanoString-PCR (A and B), and TLDA-PCR effects (C and D), and biological effects (tumour vs. normal) for the breast FFPE dataset. On each sample vs. sample heatmap, the white-to-blue colour bar reflects the range of Pearson's correlations. The two-column annotation bars indicate platform (TLDA in yellow, NanoString in blue, and PCR in purple), and biological effects (tumour in red, normal in white). The strong separation by platform highlights the magnitude of technical artifacts.

118

Figure 5.5 Sample-wise and gene-wise correlations between PCR vs. NanoString and PCR

vs. TLDA

Spearman’s rank correlations were calculated for miRNA abundances in the breast cancer dataset, for every pair of replicate samples between A) PCR vs. NanoString, and B) PCR vs. TLDA, and every pair of genes between C) PCR vs. NanoString and D) PCR vs. TLDA. The dashed line indicates the overall correlation between platforms.

119

Figure 5.6 Correlations of tumour/normal fold-changes between NanoString and TLDA

Tumour/normal fold-changes (shown in log2-scale) were plotted for the NanoString (NS) and TLDA datasets, for all miRNAs that were common to both platforms, for A) breast FFPE, and B) cervical FFPE samples. P-values were calculated using FDR-adjusted model-based t-tests. Only significantly differentially-expressed miRNAs, those with statistically significant fold-changes (Pboth ≤ 0.05, PTLDA ≤ 0.05, or PNS ≤ 0.05), were re-plotted for C) breast FFPE, and D) cervical FFPE samples, to show the correlations between NS and TLDA datasets. The yellow-to-purple colour bar at the bottom represents the range of correlations (rho).

120

5.5 Discussion

The use of global miRNA profiling platforms for biomarker development continues to

increase in popularity, largely due to the fact that miRNAs are extremely stable in many types of

clinical specimens, and have been implicated in nearly all human diseases. With various

platforms available for the high-throughput measurement of miRNA abundance, it is important

to consider the platform-specific effects associated with the detection of differentially-expressed

miRNAs. The performance of technologies used for miRNA detection are not as well

characterized as those used for DNA, mRNA or protein, with respect to their use in clinical

research. Many methods for mRNA data normalization and analysis have been adapted for

miRNA datasets, however, the accuracy of any methods that require a normal distribution of data

may not be as robust, because most global miRNA platforms can only detect several hundred

different miRNAs, compared to global mRNA platforms that can detect tens of thousands of

different gene transcripts (314).

Previous studies comparing global miRNA profiling techonologies have mainly focused

on various microarray and next-generation sequencing technologies (315-318), however, there

are no published reports to date that directly compare NanoString with TLDA for miRNA

expression profiling. Our study fills this gap by evaluating the performance of the NanoString

nCounter miRNA Expression Assay and the TaqMan Low Density Array Human MicroRNA

Array across three types cancers (breast, cervical and NPC) and two types of specimens (FFPE

and frozen tissues). We also compared both of these platforms with PCR, which is widely

regarded as the gold standard and the method of choice for validation. The four key findings of

our study were: i) miRNA abundances were poorly correlated between NanoString and TLDA,

ii) PCR was more similar to TLDA than NanoString, iii) Tumour/normal fold-changes were

poorly correlated between NanoString and TLDA, and iv) Significantly differentially-expressed

miRNAs were highly-correlated between Nanostring and TLDA.

Consistent with other studies that have reported low cross-platform concordance between

various global miRNA profiling platforms (315, 316, 318), we demonstrated poor correlations

between the NanoString and TLDA datasets, with respect to miRNA abundances (ρ = 0.32 for

breast FFPE samples ; ρ = 0.32 for cervical FFPE samples; ρ = -0.02 for NPC FFPE samples,

and ρ = 0.42 for cervical frozen samples) and tumour/normal fold changes (ρ = 0.33 for breast

121

FFPE samples; ρ = 0.37 for cervix FFPE samples). The poor correlation between NanoString and

TLDA could potentially be explained by the fundamental differences between the chemistries of

these two methods. The TLDA method relies on target-specific reverse transcription and the

hybridization of stem-loop primers, while NanoString provides a direct count of miRNAs using

optical quantification of fluorescently-tagged miRNA molecules. The possibility of bias from

amplification steps or enzymatic reactions is present for data generated using TLDA, but not

NanoString.

Despite low cross-platform correlations between NanoString and TLDA, we

demonstrated that PCR was highly comparable to both NanoString (ρ = 0.6) and TLDA (ρ = 0.7)

in terms of miRNA abundances, which was consistent with other published reports that found

high concordance between PCR and various global miRNA profiling platforms (315, 319).

However, gene-wise correlations of replicate samples revealed that almost 50% of genes were

negatively correlated between PCR and NanoString, whereas over 90% of genes were positively

correlated between PCR and TLDA. The similarity between PCR and TLDA are not surprising,

since TLDA is a PCR-based method and utilizes the same stem-loop reverse transcription

primers and TaqMan chemistry.

Although tumour/normal fold-changes were poorly-correlated between the NanoString

and TLDA datasets (ρ = 0.33 for breast FFPE samples; ρ = 0.37 for cervix FFPE samples),

miRNAs that were significantly differentially-expressed between tumour and normal were highly

correlated between the two platforms. This indicates that biologically-significant miRNAs are

generally concordant between the two platforms, and are likely to be identified with either

NanoString or TLDA.

In conclusion, our study demonstrated that despite poor cross-platform correlation

between NanoString and TLDA for global miRNA quantification, both platforms performed well

in comparison with PCR, although TLDA was more similar to PCR. To our knowledge, this is

the first reported study describing a systematic comparison of the Nanostring and TLDA

platforms for global miRNA expression profiling. When selecting a platform for global miRNA

profiling, many consideration must be taken into account, including: cost, available facilities,

amount of RNA required, and familiarity with platform-specific data normalization and analysis

122

methods. For future studies, it would be useful to develop methods for integrating data across

platforms.

123

CHAPTER 6: DISCUSSION

124

6.1 Research Summary

The focus of this thesis was to examine the role of miRNAs in cervical cancer, and

develop new insights into their contributions to tumourigenesis and clinical outcome. During our

investigations, several methods were utilized for measuring miRNA expression, and our focus

was expanded to include an analysis of various technologies for quantifying miRNA abundance.

MicroRNA-196b was demonstrated to be significantly down regulated in cervical cancer

cell lines and patient samples, and low miR-196b expression in patients was shown to be

associated with poorer DFS. Furthermore, we observed that miR-196b directly targeted the

HOXB7 transcription factor, which in turn, regulated the expression of VEGF. Our miRNA

profiling studies in patients also led to the development of a candidate 9-miRNA signature that

was prognostic for DFS in cervical cancer patients, independent of tumour size and nodal status.

However, although three different methods were utilized to measure these nine miRNAs in an

independent cohort of patients, we were unable to validate this signature. Furthermore, our

attempts to independently corroborate the only published miRNA signature to date for cervical

cancer (120) were similarly unsuccessful. These results highlight the considerations that must be

taken into account with regards to the evaluation of miRNAs as prognostic biomarkers.

Our analyses of the TLDA and Nanostring methods for global miRNA expression

profiling demonstrated an overall lack of concordance between these two platforms, although

they individually correlated well with PCR, the gold-standard for nucleic acid quantification. On

a positive note, biologically-relevant miRNAs were found to be highly correlated between the

two platforms. In addition, our investigations identified a protocol bias in microarray studies that

was created by the choice of the amplification protocol; the LTR-method was able to help

remove, but was unable to completely eliminate this bias.

6.2 Future Directions

Our characterization of the role of miR-196b in the HOXB7~VEGF axis was very

promising, given the recent announcement of the preliminary data derived from this Phase III

clinical trial for Bevacizumab (220). This drug is currently approved for use in metastatic

colorectal cancer, non-small cell lung cancer, glioblastoma, and metastatic kidney cancer, and

has been shown to impact tumour progression by blocking angiogenesis (219). If Bevacizumab

125

were to be added to the treatment regimen for cervical cancer patients, further investigations

would be warranted to determine if miR-196b levels could possibly be used as a biomarker to

predict response to Bevacizumab.

Additional attempts to validate our 9-miRNA signature should be made. One option is to

utilize an alternate RNA extraction method for the FFPE samples that make up the validation

cohort. After our experiments were completed, a study that compared ten RNA extraction

methods for FFPE samples demonstrated that the Ambion RecoverAll Total Nucleic Acid

Isolation Kit recovered the most amplifiable RNA (113), which would be well-suited for PCR-

based RNA expression profiling methods including the TLDA. Another option is to use a

validation cohort consisting of frozen samples, given the lack of consistency in the tissue

preservation methods (frozen vs. FFPE) between our training and validation cohorts. One

strategy is to collect frozen cervical cancer specimens to create a new validation cohort, and

measure the expression of the nine miRNAs in our signature using TLDA or individual real-time

PCR. Another strategy is to utilize the frozen cervical cancer samples from the TCGA data portal

as another independent validation set, after the number of frozen cervical cancer samples with

miRNASeq data and clinical information is expanded; as of June 2013, only 48 samples had

survival information.

Alternately, we could try to generate a new prognostic miRNA signature for cervical

cancer, using a more robust method for global miRNA expression profiling. Reports have

demonstrated the increased sensitivity and specificity of deep sequencing for measuring miRNA

expression (115, 116). Since we have shown that the majority of significantly differentially-

expressed miRNAs in cervical cancer are down-regulated, deep sequencing should be able to

more accurately detect more lowly-expressed miRNAs in our samples that were beyond the limit

of detection for TLDA.

Another potential avenue would be to perform a global miRNA expression profiling

experiment using multiple biopsies from each individual patient’s tumour, similar to the

Bachtiary et al. study (265), to explore intra-tumour heterogeneity of miRNA expression in

cervical cancer. It would also be interesting to develop a novel signature using these data, since

this should provide a more representative miRNA profile of the tumour, in its entirety.

126

In addition, we will perform co-analyses of global miRNA and mRNA expression data

that we have generated for our 79 cervical cancer and 11 normal cervix frozen samples, similar

to studies that have been reported for glioma (320), ovarian cancer (321), primary human

lymphoblastoid cell lines (322), and the NCI-60 panel of human cancer cell lines (323). Global

miRNA and mRNA levels have been measured using the TLDA Human MicroRNA A Array

(v2.0) and the Affymetrix GeneChip Human Genome U133 Plus 2.0, respectively. We will

examine the correlation of genome-wide expression profiles between miRNAs and their

predicted target mRNAs, and determine if predicted miRNA-mRNA pairings are negatively

correlated in expression.

6.3 Conclusions

This thesis has provided insights into the role of microRNAs in cervical cancer, and

provided a thorough analysis of the issues and considerations pertaining to prognostic miRNA

signature development and validation. miR-196b was identified and characterized to be

important in cervical cancer development and progression. These findings are significant

contributions to the body of knowledge that will help us to realize the promise of personalized

cancer medicine. In addition, various platforms for miRNA expression profiling were

investigated and compared to each other, and amplification methods for Affymetrix microarrays

were evaluated. In both cases, the low concordance in data produced by different protocols was

assessed and methods to mitigate biases were described.

127

References

1. Ambros V. The functions of animal microRNAs. Nature 2004;431: 350-5.

2. Bartel DP. MicroRNAs: genomics, biogenesis, mechanism, and function. Cell 2004;116: 281-97.

3. Ibanez-Ventoso C, Vora M, Driscoll M. Sequence relationships among C. elegans, D. melanogaster and human microRNAs highlight the extensive conservation of microRNAs in biology. PLoS One 2008;3: e2818.

4. Griffiths-Jones S, Grocock RJ, van Dongen S, Bateman A, Enright AJ. miRBase: microRNA sequences, targets and gene nomenclature. Nucleic Acids Res 2006;34: D140-4.

5. Stefani G, Slack FJ. Small non-coding RNAs in animal development. Nat Rev Mol Cell Biol 2008;9: 219-30.

6. Lee RC, Feinbaum RL, Ambros V. The C. elegans heterochronic gene lin-4 encodes small RNAs with antisense complementarity to lin-14. Cell 1993;75: 843-54.

7. Ambros V. A hierarchy of regulatory genes controls a larva-to-adult developmental switch in C. elegans. Cell 1989;57: 49-57.

8. Reinhart BJ, Slack FJ, Basson M, et al. The 21-nucleotide let-7 RNA regulates developmental timing in Caenorhabditis elegans. Nature 2000;403: 901-6.

9. Lin SY, Johnson SM, Abraham M, et al. The C elegans hunchback homolog, hbl-1, controls temporal patterning and is a probable microRNA target. Dev Cell 2003;4: 639-50.

10. Abrahante JE, Daul AL, Li M, et al. The Caenorhabditis elegans hunchback-like gene lin-57/hbl-1 controls developmental time and is regulated by microRNAs. Dev Cell 2003;4: 625-37.

11. Lee RC, Ambros V. An extensive class of small RNAs in Caenorhabditis elegans. Science 2001;294: 862-4.

12. Lau NC, Lim LP, Weinstein EG, Bartel DP. An abundant class of tiny RNAs with probable regulatory roles in Caenorhabditis elegans. Science 2001;294: 858-62.

13. Lagos-Quintana M, Rauhut R, Lendeckel W, Tuschl T. Identification of novel genes coding for small expressed RNAs. Science 2001;294: 853-8.

14. Lee Y, Kim M, Han J, et al. MicroRNA genes are transcribed by RNA polymerase II. EMBO J 2004;23: 4051-60.

15. Lee Y, Jeon K, Lee JT, Kim S, Kim VN. MicroRNA maturation: stepwise processing and subcellular localization. EMBO J 2002;21: 4663-70.

128

16. Borchert GM, Lanier W, Davidson BL. RNA polymerase III transcribes human microRNAs. Nat Struct Mol Biol 2006;13: 1097-101.

17. Cai X, Hagedorn CH, Cullen BR. Human microRNAs are processed from capped, polyadenylated transcripts that can also function as mRNAs. RNA 2004;10: 1957-66.

18. Lee Y, Ahn C, Han J, et al. The nuclear RNase III Drosha initiates microRNA processing. Nature 2003;425: 415-9.

19. Landthaler M, Yalcin A, Tuschl T. The human DiGeorge syndrome critical region gene 8 and Its D. melanogaster homolog are required for miRNA biogenesis. Curr Biol 2004;14: 2162-7.

20. Han J, Lee Y, Yeom KH, et al. Molecular basis for the recognition of primary microRNAs by the Drosha-DGCR8 complex. Cell 2006;125: 887-901.

21. Basyuk E, Suavet F, Doglio A, Bordonne R, Bertrand E. Human let-7 stem-loop precursors harbor features of RNase III cleavage products. Nucleic Acids Res 2003;31: 6593-7.

22. Zeng Y, Cullen BR. Structural requirements for pre-microRNA binding and nuclear export by Exportin 5. Nucleic Acids Res 2004;32: 4776-85.

23. Gwizdek C, Ossareh-Nazari B, Brownawell AM, et al. Exportin-5 mediates nuclear export of minihelix-containing RNAs. J Biol Chem 2003;278: 5505-8.

24. Bohnsack MT, Czaplinski K, Gorlich D. Exportin 5 is a RanGTP-dependent dsRNA-binding protein that mediates nuclear export of pre-miRNAs. RNA 2004;10: 185-91.

25. Yi R, Qin Y, Macara IG, Cullen BR. Exportin-5 mediates the nuclear export of pre-microRNAs and short hairpin RNAs. Genes Dev 2003;17: 3011-6.

26. Lund E, Guttinger S, Calado A, Dahlberg JE, Kutay U. Nuclear export of microRNA precursors. Science 2004;303: 95-8.

27. Winter J, Jung S, Keller S, Gregory RI, Diederichs S. Many roads to maturity: microRNA biogenesis pathways and their regulation. Nat Cell Biol 2009;11: 228-34.

28. Bernstein E, Caudy AA, Hammond SM, Hannon GJ. Role for a bidentate ribonuclease in the initiation step of RNA interference. Nature 2001;409: 363-6.

29. Haase AD, Jaskiewicz L, Zhang H, et al. TRBP, a regulator of cellular PKR and HIV-1 virus expression, interacts with Dicer and functions in RNA silencing. EMBO Rep 2005;6: 961-7.

30. Khvorova A, Reynolds A, Jayasena SD. Functional siRNAs and miRNAs exhibit strand bias. Cell 2003;115: 209-16.

31. Marco A, Macpherson JI, Ronshaugen M, Griffiths-Jones S. MicroRNAs from the same precursor have different targeting properties. Silence;3: 8.

129

32. Gregory RI, Chendrimada TP, Cooch N, Shiekhattar R. Human RISC couples microRNA biogenesis and posttranscriptional gene silencing. Cell 2005;123: 631-40.

33. Chendrimada TP, Gregory RI, Kumaraswamy E, et al. TRBP recruits the Dicer complex to Ago2 for microRNA processing and gene silencing. Nature 2005;436: 740-4.

34. Rivas FV, Tolia NH, Song JJ, et al. Purified Argonaute2 and an siRNA form recombinant human RISC. Nat Struct Mol Biol 2005;12: 340-9.

35. Meister G, Landthaler M, Patkaniowska A, Dorsett Y, Teng G, Tuschl T. Human Argonaute2 mediates RNA cleavage targeted by miRNAs and siRNAs. Mol Cell 2004;15: 185-97.

36. Liu J, Carmell MA, Rivas FV, et al. Argonaute2 is the catalytic engine of mammalian RNAi. Science 2004;305: 1437-41.

37. Lai EC. Micro RNAs are complementary to 3' UTR sequence motifs that mediate negative post-transcriptional regulation. Nat Genet 2002;30: 363-4.

38. Lytle JR, Yario TA, Steitz JA. Target mRNAs are repressed as efficiently by microRNA-binding sites in the 5' UTR as in the 3' UTR. Proc Natl Acad Sci U S A 2007;104: 9667-72.

39. Duursma AM, Kedde M, Schrier M, le Sage C, Agami R. miR-148 targets human DNMT3b protein coding region. RNA 2008;14: 872-7.

40. Lewis BP, Shih IH, Jones-Rhoades MW, Bartel DP, Burge CB. Prediction of mammalian microRNA targets. Cell 2003;115: 787-98.

41. Lim LP, Lau NC, Weinstein EG, et al. The microRNAs of Caenorhabditis elegans. Genes Dev 2003;17: 991-1008.

42. Stark A, Brennecke J, Russell RB, Cohen SM. Identification of Drosophila MicroRNA targets. PLoS Biol 2003;1: E60.

43. Bagga S, Bracht J, Hunter S, et al. Regulation by let-7 and lin-4 miRNAs results in target mRNA degradation. Cell 2005;122: 553-63.

44. Orban TI, Izaurralde E. Decay of mRNAs targeted by RISC requires XRN1, the Ski complex, and the exosome. RNA 2005;11: 459-69.

45. Humphreys DT, Westman BJ, Martin DI, Preiss T. MicroRNAs control translation initiation by inhibiting eukaryotic initiation factor 4E/cap and poly(A) tail function. Proc Natl Acad Sci U S A 2005;102: 16961-6.

46. Gingras AC, Raught B, Sonenberg N. eIF4 initiation factors: effectors of mRNA recruitment to ribosomes and regulators of translation. Annu Rev Biochem 1999;68: 913-63.

47. Kiriakidou M, Tan GS, Lamprinaki S, De Planell-Saguer M, Nelson PT, Mourelatos Z. An mRNA m7G cap binding-like motif within human Ago2 represses translation. Cell 2007;129: 1141-51.

130

48. Chendrimada TP, Finn KJ, Ji X, et al. MicroRNA silencing through RISC recruitment of eIF6. Nature 2007;447: 823-8.

49. Maroney PA, Yu Y, Fisher J, Nilsen TW. Evidence that microRNAs are associated with translating messenger RNAs in human cells. Nat Struct Mol Biol 2006;13: 1102-7.

50. Nottrott S, Simard MJ, Richter JD. Human let-7a miRNA blocks protein production on actively translating polyribosomes. Nat Struct Mol Biol 2006;13: 1108-14.

51. Petersen CP, Bordeleau ME, Pelletier J, Sharp PA. Short RNAs repress translation after initiation in mammalian cells. Mol Cell 2006;21: 533-42.

52. Bentwich I, Avniel A, Karov Y, et al. Identification of hundreds of conserved and nonconserved human microRNAs. Nat Genet 2005;37: 766-70.

53. Rehmsmeier M, Steffen P, Hochsmann M, Giegerich R. Fast and effective prediction of microRNA/target duplexes. RNA 2004;10: 1507-17.

54. Friedman RC, Farh KK, Burge CB, Bartel DP. Most mammalian mRNAs are conserved targets of microRNAs. Genome Res 2009;19: 92-105.

55. Kertesz M, Iovino N, Unnerstall U, Gaul U, Segal E. The role of site accessibility in microRNA target recognition. Nat Genet 2007;39: 1278-84.

56. Saetrom P, Heale BS, Snove O, Jr., Aagaard L, Alluin J, Rossi JJ. Distance constraints between microRNA target sites dictate efficacy and cooperativity. Nucleic Acids Res 2007;35: 2333-42.

57. Doench JG, Sharp PA. Specificity of microRNA target selection in translational repression. Genes Dev 2004;18: 504-11.

58. Lewis BP, Burge CB, Bartel DP. Conserved seed pairing, often flanked by adenosines, indicates that thousands of human genes are microRNA targets. Cell 2005;120: 15-20.

59. Miranda KC, Huynh T, Tay Y, et al. A pattern-based method for the identification of MicroRNA binding sites and their corresponding heteroduplexes. Cell 2006;126: 1203-17.

60. Krek A, Grun D, Poy MN, et al. Combinatorial microRNA target predictions. Nat Genet 2005;37: 495-500.

61. Maragkakis M, Alexiou P, Papadopoulos GL, et al. Accurate microRNA target prediction correlates with protein repression levels. BMC Bioinformatics 2009;10: 295.

62. Wang X. miRDB: a microRNA target prediction and functional annotation database with a wiki interface. RNA 2008;14: 1012-7.

63. Huang JC, Babak T, Corson TW, et al. Using expression profiling data to identify human microRNA targets. Nat Methods 2007;4: 1045-9.

131

64. Enright AJ, John B, Gaul U, Tuschl T, Sander C, Marks DS. MicroRNA targets in Drosophila. Genome Biol 2003;5: R1.

65. John B, Enright AJ, Aravin A, Tuschl T, Sander C, Marks DS. Human MicroRNA targets. PLoS Biol 2004;2: e363.

66. Sethupathy P, Corda B, Hatzigeorgiou AG. TarBase: A comprehensive database of experimentally supported animal microRNA targets. RNA 2006;12: 192-7.

67. Bentwich I. Prediction and validation of microRNAs and their targets. FEBS Lett 2005;579: 5904-10.

68. Alajez NM, Lenarduzzi M, Ito E, et al. MiR-218 suppresses nasopharyngeal cancer progression through downregulation of survivin and the SLIT2-ROBO1 pathway. Cancer Res 2011;71: 2381-91.

69. Garzon R, Calin GA, Croce CM. MicroRNAs in Cancer. Annu Rev Med 2009;60: 167-79.

70. Meng F, Henson R, Wehbe-Janek H, Ghoshal K, Jacob ST, Patel T. MicroRNA-21 regulates expression of the PTEN tumor suppressor gene in human hepatocellular cancer. Gastroenterology 2007;133: 647-58.

71. Chan JA, Krichevsky AM, Kosik KS. MicroRNA-21 is an antiapoptotic factor in human glioblastoma cells. Cancer Res 2005;65: 6029-33.

72. Si ML, Zhu S, Wu H, Lu Z, Wu F, Mo YY. miR-21-mediated tumor growth. Oncogene 2007;26: 2799-803.

73. Calin GA, Dumitru CD, Shimizu M, et al. Frequent deletions and down-regulation of micro- RNA genes miR15 and miR16 at 13q14 in chronic lymphocytic leukemia. Proc Natl Acad Sci U S A 2002;99: 15524-9.

74. Cimmino A, Calin GA, Fabbri M, et al. miR-15 and miR-16 induce apoptosis by targeting BCL2. Proc Natl Acad Sci U S A 2005;102: 13944-9.

75. Baer C, Claus R, Plass C. Genome-wide epigenetic regulation of miRNAs in cancer. Cancer Res 2013;73: 473-7.

76. Choudhry H, Catto JW. Epigenetic regulation of microRNA expression in cancer. Methods Mol Biol 2011;676: 165-84.

77. Zehavi L, Avraham R, Barzilai A, et al. Silencing of a large microRNA cluster on human chromosome 14q32 in melanoma: biological effects of mir-376a and mir-376c on insulin growth factor 1 receptor. Mol Cancer 2012;11: 44.

78. Calin GA, Croce CM. MicroRNAs and chromosomal abnormalities in cancer cells. Oncogene 2006;25: 6202-10.

132

79. Kobel M, Gilks CB, Huntsman DG. Dicer and Drosha in ovarian cancer. N Engl J Med 2009;360: 1150-1; author reply 1.

80. Yan M, Huang HY, Wang T, et al. Dysregulated expression of dicer and drosha in breast cancer. Pathol Oncol Res;18: 343-8.

81. Kumar MS, Lu J, Mercer KL, Golub TR, Jacks T. Impaired microRNA processing enhances cellular transformation and tumorigenesis. Nat Genet 2007;39: 673-7.

82. Saito Y, Liang G, Egger G, et al. Specific activation of microRNA-127 with downregulation of the proto-oncogene BCL6 by chromatin-modifying drugs in human cancer cells. Cancer Cell 2006;9: 435-43.

83. Calin GA, Sevignani C, Dumitru CD, et al. Human microRNA genes are frequently located at fragile sites and genomic regions involved in cancers. Proc Natl Acad Sci U S A 2004;101: 2999-3004.

84. Ota A, Tagawa H, Karnan S, et al. Identification and characterization of a novel gene, C13orf25, as a target for 13q31-q32 amplification in malignant lymphoma. Cancer Res 2004;64: 3087-95.

85. He L, Thomson JM, Hemann MT, et al. A microRNA polycistron as a potential human oncogene. Nature 2005;435: 828-33.

86. Volinia S, Calin GA, Liu CG, et al. A microRNA expression signature of human solid tumors defines cancer gene targets. Proc Natl Acad Sci U S A 2006;103: 2257-61.

87. Faggad A, Kasajima A, Weichert W, et al. Down-regulation of the microRNA processing enzyme Dicer is a prognostic factor in human colorectal cancer. Histopathology 2012;61: 552-61.

88. Papachristou DJ, Korpetinou A, Giannopoulou E, et al. Expression of the ribonucleases Drosha, Dicer, and Ago2 in colorectal carcinomas. Virchows Arch 2011;459: 431-40.

89. Sand M, Gambichler T, Skrygan M, et al. Expression levels of the microRNA processing enzymes Drosha and dicer in epithelial skin cancer. Cancer Invest 2010;28: 649-53.

90. Sand M, Skrygan M, Georgas D, et al. Expression levels of the microRNA maturing microprocessor complex component DGCR8 and the RNA-induced silencing complex (RISC) components argonaute-1, argonaute-2, PACT, TARBP1, and TARBP2 in epithelial skin cancer. Mol Carcinog 2012;51: 916-22.

91. Horikawa Y, Wood CG, Yang H, et al. Single nucleotide polymorphisms of microRNA machinery genes modify the risk of renal cell carcinoma. Clin Cancer Res 2008;14: 7956-62.

92. Merritt WM, Lin YG, Han LY, et al. Dicer, Drosha, and outcomes in patients with ovarian cancer. N Engl J Med 2008;359: 2641-50.

93. Blenkiron C, Goldstein LD, Thorne NP, et al. MicroRNA expression profiling of human breast cancer identifies new markers of tumor subtype. Genome Biol 2007;8: R214.

133

94. Mattie MD, Benz CC, Bowers J, et al. Optimized high-throughput microRNA expression profiling provides novel biomarker assessment of clinical prostate and breast cancer biopsies. Mol Cancer 2006;5: 24.

95. Lu J, Getz G, Miska EA, et al. MicroRNA expression profiles classify human cancers. Nature 2005;435: 834-8.

96. Hui AB, Shi W, Boutros PC, et al. Robust global micro-RNA profiling with formalin-fixed paraffin-embedded breast cancer tissues. Lab Invest 2009;89: 597-606.

97. Xi Y, Nakajima G, Gavin E, et al. Systematic analysis of microRNA expression of RNA extracted from fresh frozen and formalin-fixed paraffin-embedded samples. RNA 2007;13: 1668-74.

98. Mitchell PS, Parkin RK, Kroh EM, et al. Circulating microRNAs as stable blood-based markers for cancer detection. Proc Natl Acad Sci U S A 2008;105: 10513-8.

99. Cortez MA, Calin GA. MicroRNA identification in plasma and serum: a new tool to diagnose and monitor diseases. Expert Opin Biol Ther 2009;9: 703-11.

100. Gilad S, Meiri E, Yogev Y, et al. Serum microRNAs are promising novel biomarkers. PLoS One 2008;3: e3148.

101. Hanke M, Hoefig K, Merz H, et al. A robust methodology to study urine microRNA as tumor marker: microRNA-126 and microRNA-182 are related to urinary bladder cancer. Urol Oncol;28: 655-61.

102. Park NJ, Zhou H, Elashoff D, et al. Salivary microRNA: discovery, characterization, and clinical utility for oral cancer detection. Clin Cancer Res 2009;15: 5473-7.

103. Bresters D, Schipper ME, Reesink HW, Boeser-Nunnink BD, Cuypers HT. The duration of fixation influences the yield of HCV cDNA-PCR products from formalin-fixed, paraffin-embedded liver tissue. J Virol Methods 1994;48: 267-72.

104. Macabeo-Ong M, Ginzinger DG, Dekker N, et al. Effect of duration of fixation on quantitative reverse transcription polymerase chain reaction analyses. Mod Pathol 2002;15: 979-87.

105. Nonn L, Vaishnav A, Gallagher L, Gann PH. mRNA and micro-RNA expression analysis in laser-capture microdissected prostate biopsies: valuable tool for risk assessment and prevention trials. Exp Mol Pathol 2009;88: 45-51.

106. Lovendorf MB, Zibert JR, Hagedorn PH, et al. Comparison of microRNA expression using different preservation methods of matched psoriatic skin samples. Exp Dermatol 2012;21: 299-301.

107. Kolbert CP, Feddersen RM, Rakhshan F, et al. Multi-platform analysis of microRNA expression measurements in RNA from fresh frozen and FFPE tissues. PLoS One 2013;8: e52517.

134

108. Weng L, Wu X, Gao H, et al. MicroRNA profiling of clear cell renal cell carcinoma by whole-genome small RNA deep sequencing of paired frozen and formalin-fixed, paraffin-embedded tissue specimens. J Pathol 2010;222: 41-51.

109. de Biase D, Visani M, Morandi L, et al. miRNAs expression analysis in paired fresh/frozen and dissected formalin fixed and paraffin embedded glioblastoma using real-time pCR. PLoS One 2012;7: e35596.

110. Liu H, Kohane IS. Tissue and process specific microRNA-mRNA co-expression in mammalian development and malignancy. PLoS One 2009;4: e5436.

111. Glud M, Klausen M, Gniadecki R, et al. MicroRNA expression in melanocytic nevi: the usefulness of formalin-fixed, paraffin-embedded material for miRNA microarray profiling. J Invest Dermatol 2009;129: 1219-24.

112. Leite KR, Canavez JM, Reis ST, et al. miRNA analysis of prostate cancer by quantitative real time PCR: comparison between formalin-fixed paraffin embedded and fresh-frozen tissue. Urol Oncol 2011;29: 533-7.

113. Okello JB, Zurek J, Devault AM, et al. Comparison of methods in the recovery of nucleic acids from archival formalin-fixed paraffin-embedded autopsy tissues. Anal Biochem 2010;400: 110-7.

114. Chen C, Ridzon DA, Broomer AJ, et al. Real-time quantification of microRNAs by stem-loop RT-PCR. Nucleic Acids Res 2005;33: e179.

115. Glazov EA, Cottee PA, Barris WC, Moore RJ, Dalrymple BP, Tizard ML. A microRNA catalog of the developing chicken embryo identified by a deep sequencing approach. Genome Res 2008;18: 957-64.

116. Kelly AD, Hill KE, Correll M, et al. Next-generation sequencing and microarray-based interrogation of microRNAs from formalin-fixed, paraffin-embedded tissue: Preliminary assessment of cross-platform concordance. Genomics 2013;102: 8-14.

117. Mass RD, Press MF, Anderson S, et al. Evaluation of clinical outcomes according to HER2 detection by fluorescence in situ hybridization in women with metastatic breast cancer treated with trastuzumab. Clin Breast Cancer 2005;6: 240-6.

118. Simon R. Roadmap for developing and validating therapeutically relevant genomic classifiers. J Clin Oncol 2005;23: 7332-41.

119. Calin GA, Ferracin M, Cimmino A, et al. A MicroRNA signature associated with prognosis and progression in chronic lymphocytic leukemia. N Engl J Med 2005;353: 1793-801.

120. Hu X, Schwarz JK, Lewis JS, Jr., et al. A microRNA expression signature for cervical cancer prognosis. Cancer Res 2010;70: 1441-8.

121. Schetter AJ, Leung SY, Sohn JJ, et al. MicroRNA expression profiles associated with prognosis and therapeutic outcome in colon adenocarcinoma. JAMA 2008;299: 425-36.

135

122. Guo Y, Chen Z, Zhang L, et al. Distinctive microRNA profiles relating to patient survival in esophageal squamous cell carcinoma. Cancer Res 2008;68: 26-33.

123. Li X, Zhang Y, Ding J, Wu K, Fan D. Survival prediction of gastric cancer by a seven-microRNA signature. Gut 2010;59: 579-85.

124. Budhu A, Jia HL, Forgues M, et al. Identification of metastasis-related microRNAs in hepatocellular carcinoma. Hepatology 2008;47: 897-907.

125. Raponi M, Dossey L, Jatkoe T, et al. MicroRNA classifiers for predicting prognosis of squamous cell lung cancer. Cancer Res 2009;69: 5776-83.

126. Yanaihara N, Caplen N, Bowman E, et al. Unique microRNA molecular profiles in lung cancer diagnosis and prognosis. Cancer Cell 2006;9: 189-98.

127. Yu SL, Chen HY, Chang GC, et al. MicroRNA signature predicts survival and relapse in lung cancer. Cancer Cell 2008;13: 48-57.

128. Liu N, Chen NY, Cui RX, et al. Prognostic value of a microRNA signature in nasopharyngeal carcinoma: a microRNA expression analysis. Lancet Oncol 2012;13: 633-41.

129. Tong AW, Fulgham P, Jay C, et al. MicroRNA profile analysis of human prostate cancers. Cancer Gene Ther 2009;16: 206-16.

130. Marcucci G, Maharry K, Radmacher MD, et al. Prognostic significance of, and gene and microRNA expression signatures associated with, CEBPA mutations in cytogenetically normal acute myeloid leukemia with high-risk molecular features: a Cancer and Leukemia Group B Study. J Clin Oncol 2008;26: 5078-87.

131. Jemal A, Bray F, Center MM, Ferlay J, Ward E, Forman D. Global cancer statistics. CA Cancer J Clin 2011;61: 69-90.

132. Duarte-Franco E, Franco EL. Cancer of the Uterine Cervix. BMC Womens Health 2004;4 Suppl 1: S13.

133. Franco EL, Duarte-Franco E, Ferenczy A. Cervical cancer: epidemiology, prevention and the role of human papillomavirus infection. CMAJ 2001;164: 1017-25.

134. Statistics CCSsSCoC. Canadian Cancer Statistics 2012. Toronto, ON: Canadian Cancer Society; 2012.

135. Walboomers JM, Jacobs MV, Manos MM, et al. Human papillomavirus is a necessary cause of invasive cervical cancer worldwide. J Pathol 1999;189: 12-9.

136. Ault KA. Epidemiology and natural history of human papillomavirus infections in the female genital tract. Infect Dis Obstet Gynecol 2006;2006 Suppl: 40470.

137. zur Hausen H, Meinhof W, Scheiber W, Bornkamm GW. Attempts to detect virus-secific DNA in human tumors. I. Nucleic acid hybridizations with complementary RNA of human wart virus. Int J Cancer 1974;13: 650-6.

136

138. Durst M, Gissmann L, Ikenberg H, zur Hausen H. A papillomavirus DNA from a cervical carcinoma and its prevalence in cancer biopsy samples from different geographic regions. Proc Natl Acad Sci U S A 1983;80: 3812-5.

139. Garland SM, Hernandez-Avila M, Wheeler CM, et al. Quadrivalent vaccine against human papillomavirus to prevent anogenital diseases. N Engl J Med 2007;356: 1928-43.

140. Giuliano AR, Palefsky JM, Goldstone S, et al. Efficacy of quadrivalent HPV vaccine against HPV Infection and disease in males. N Engl J Med 2011;364: 401-11.

141. Munoz N, Bosch FX, de Sanjose S, et al. Epidemiologic classification of human papillomavirus types associated with cervical cancer. N Engl J Med 2003;348: 518-27.

142. Scheurer ME, Tortolero-Luna G, Adler-Storthz K. Human papillomavirus infection: biology, epidemiology, and prevention. Int J Gynecol Cancer 2005;15: 727-46.

143. Landoni F, Maneo A, Colombo A, et al. Randomised study of radical surgery versus radiotherapy for stage Ib-IIa cervical cancer. Lancet 1997;350: 535-40.

144. Inoue T, Okumura M. Prognostic significance of parametrial extension in patients with cervical carcinoma Stages IB, IIA, and IIB. A study of 628 cases treated by radical hysterectomy and lymphadenectomy with or without postoperative irradiation. Cancer 1984;54: 1714-9.

145. Alvarez RD, Soong SJ, Kinney WK, et al. Identification of prognostic factors and risk groups in patients found to have nodal metastasis at the time of radical hysterectomy for early-stage squamous carcinoma of the cervix. Gynecol Oncol 1989;35: 130-5.

146. Fuller AF, Jr., Elliott N, Kosloff C, Hoskins WJ, Lewis JL, Jr. Determinants of increased risk for recurrence in patients undergoing radical hysterectomy for stage IB and IIA carcinoma of the cervix. Gynecol Oncol 1989;33: 34-9.

147. Lin HH, Cheng WF, Chan KW, Chang DY, Chen CK, Huang SC. Risk factors for recurrence in patients with stage IB, IIA, and IIB cervical carcinoma after radical hysterectomy and postoperative pelvic irradiation. Obstet Gynecol 1996;88: 274-9.

148. Zreik TG, Chambers JT, Chambers SK. Parametrial involvement, regardless of nodal status: a poor prognostic factor for cervical cancer. Obstet Gynecol 1996;87: 741-6.

149. Meigs JV, Dresser R. Carcinoma of the Cervix Treated by the Roentgen Ray and Radium. Ann Surg 1937;106: 653-67.

150. Rose PG, Bundy BN, Watkins EB, et al. Concurrent cisplatin-based radiotherapy and chemotherapy for locally advanced cervical cancer. N Engl J Med 1999;340: 1144-53.

151. Whitney CW, Sause W, Bundy BN, et al. Randomized comparison of fluorouracil plus cisplatin versus hydroxyurea as an adjunct to radiation therapy in stage IIB-IVA carcinoma of the cervix with negative para-aortic lymph nodes: a Gynecologic Oncology Group and Southwest Oncology Group study. J Clin Oncol 1999;17: 1339-48.

137

152. Peters WA, 3rd, Liu PY, Barrett RJ, 2nd, et al. Concurrent chemotherapy and pelvic radiation therapy compared with pelvic radiation therapy alone as adjuvant therapy after radical surgery in high-risk early-stage cancer of the cervix. J Clin Oncol 2000;18: 1606-13.

153. Morris M, Eifel PJ, Lu J, et al. Pelvic radiation with concurrent chemotherapy compared with pelvic and para-aortic radiation for high-risk cervical cancer. N Engl J Med 1999;340: 1137-43.

154. Keys HM, Bundy BN, Stehman FB, et al. Cisplatin, radiation, and adjuvant hysterectomy compared with radiation and adjuvant hysterectomy for bulky stage IB cervical carcinoma. N Engl J Med 1999;340: 1154-61.

155. Klopp AH, Jhingran A, Ramdas L, et al. Gene expression changes in cervical squamous cell carcinoma after initiation of chemoradiation and correlation with clinical outcome. Int J Radiat Oncol Biol Phys 2008;71: 226-36.

156. Weidhaas JB, Li SX, Winter K, et al. Changes in gene expression predicting local control in cervical cancer: results from Radiation Therapy Oncology Group 0128. Clin Cancer Res 2009;15: 4199-206.

157. Rajkumar T, Vijayalakshmi N, Sabitha K, et al. A 7 gene expression score predicts for radiation response in cancer cervix. BMC Cancer 2009;9: 365.

158. Harima Y, Ikeda K, Utsunomiya K, et al. Identification of genes associated with progression and metastasis of advanced cervical cancers after radiotherapy by cDNA microarray analysis. Int J Radiat Oncol Biol Phys 2009;75: 1232-9.

159. Lui WO, Pourmand N, Patterson BK, Fire A. Patterns of known and novel small RNAs in human cervical cancer. Cancer Res 2007;67: 6031-43.

160. Lee JW, Choi CH, Choi JJ, et al. Altered MicroRNA expression in cervical carcinomas. Clin Cancer Res 2008;14: 2535-42.

161. Wang X, Tang S, Le SY, et al. Aberrant expression of oncogenic and tumor-suppressive microRNAs in cervical cancer is required for cancer cell growth. PLoS One 2008;3: e2557.

162. Pereira PM, Marques JP, Soares AR, Carreto L, Santos MA. MicroRNA expression variability in human cervical tissues. PLoS One 2010;5: e11780.

163. Wilting SM, Snijders PJ, Verlaat W, et al. Altered microRNA expression associated with chromosomal changes contributes to cervical carcinogenesis. Oncogene 2013;32: 106-16.

164. Tang T, Wong HK, Gu W, et al. MicroRNA-182 plays an onco-miRNA role in cervical cancer. Gynecol Oncol.

165. Scott MP, Weiner AJ. Structural relationships among genes that control development: sequence homology between the Antennapedia, Ultrabithorax, and fushi tarazu loci of Drosophila. Proc Natl Acad Sci U S A 1984;81: 4115-9.

138

166. McGinnis W, Levine MS, Hafen E, Kuroiwa A, Gehring WJ. A conserved DNA sequence in homoeotic genes of the Drosophila Antennapedia and bithorax complexes. Nature 1984;308: 428-33.

167. Corsetti MT, Briata P, Sanseverino L, et al. Differential DNA binding properties of three human homeodomain proteins. Nucleic Acids Res 1992;20: 4465-72.

168. Krumlauf R. Hox genes in vertebrate development. Cell 1994;78: 191-201.

169. Favier B, Dolle P. Developmental functions of mammalian Hox genes. Mol Hum Reprod 1997;3: 115-31.

170. Grier DG, Thompson A, Kwasniewska A, McGonigle GJ, Halliday HL, Lappin TR. The pathophysiology of HOX genes and their role in cancer. J Pathol 2005;205: 154-71.

171. Zakany J, Gerard M, Favier B, Potter SS, Duboule D. Functional equivalence and rescue among group 11 Hox gene products in vertebral patterning. Dev Biol 1996;176: 325-8.

172. Rijli FM, Chambon P. Genetic interactions of Hox genes in limb development: learning from compound mutants. Curr Opin Genet Dev 1997;7: 481-7.

173. Volpe MV, Pham L, Lessin M, et al. Expression of Hoxb-5 during human lung development and in congenital lung malformations. Birth Defects Res A Clin Mol Teratol 2003;67: 550-6.

174. Bogue CW, Gross I, Vasavada H, Dynia DW, Wilson CM, Jacobs HC. Identification of Hox genes in newborn lung and effects of gestational age and retinoic acid on their expression. Am J Physiol 1994;266: L448-54.

175. Golpon HA, Geraci MW, Moore MD, et al. HOX genes in human lung: altered expression in primary pulmonary hypertension and emphysema. Am J Pathol 2001;158: 955-66.

176. Warot X, Fromental-Ramain C, Fraulob V, Chambon P, Dolle P. Gene dosage-dependent effects of the Hoxa-13 and Hoxd-13 mutations on morphogenesis of the terminal parts of the digestive and urogenital tracts. Development 1997;124: 4781-91.

177. Podlasek CA, Clemens JQ, Bushman W. Hoxa-13 gene mutation results in abnormal seminal vesicle and prostate development. J Urol 1999;161: 1655-61.

178. Takahashi Y, Hamada J, Murakawa K, et al. Expression profiles of 39 HOX genes in normal human adult organs and anaplastic thyroid cancer cell lines by quantitative real-time RT-PCR system. Exp Cell Res 2004;293: 144-53.

179. Shen WF, Detmer K, Mathews CH, et al. Modulation of homeobox gene expression alters the phenotype of human hematopoietic cell lines. EMBO J 1992;11: 983-9.

180. Beslu N, Krosl J, Laurin M, Mayotte N, Humphries KR, Sauvageau G. Molecular interactions involved in HOXB4-induced activation of HSC self-renewal. Blood 2004;104: 2307-14.

139

181. Giampaolo A, Sterpetti P, Bulgarini D, et al. Key functional role and lineage-specific expression of selected HOXB genes in purified hematopoietic progenitor differentiation. Blood 1994;84: 3637-47.

182. Alharbi RA, Pettengell R, Pandha HS, Morgan R. The role of HOX genes in normal hematopoiesis and acute leukemia. Leukemia 2012.

183. Lawrence HJ, Largman C. Homeobox genes in normal hematopoiesis and leukemia. Blood 1992;80: 2445-53.

184. Armstrong SA, Staunton JE, Silverman LB, et al. MLL translocations specify a distinct gene expression profile that distinguishes a unique leukemia. Nat Genet 2002;30: 41-7.

185. Golub TR, Slonim DK, Tamayo P, et al. Molecular classification of cancer: class discovery and class prediction by gene expression monitoring. Science 1999;286: 531-7.

186. Faber J, Krivtsov AV, Stubbs MC, et al. HOXA9 is required for survival in human MLL-rearranged acute leukemias. Blood 2009;113: 2375-85.

187. Abate-Shen C. Deregulated homeobox gene expression in cancer: cause or consequence? Nat Rev Cancer 2002;2: 777-85.

188. Miller GJ, Miller HL, van Bokhoven A, et al. Aberrant HOXC expression accompanies the malignant phenotype in human prostate. Cancer Res 2003;63: 5879-88.

189. Waltregny D, Alami Y, Clausse N, de Leval J, Castronovo V. Overexpression of the homeobox gene HOXC8 in human prostate cancer correlates with loss of tumor differentiation. Prostate 2002;50: 162-9.

190. Axlund SD, Lambert JR, Nordeen SK. HOXC8 inhibits androgen receptor signaling in human prostate cancer cells by inhibiting SRC-3 recruitment to direct androgen target genes. Mol Cancer Res;8: 1643-55.

191. Raman V, Martensen SA, Reisman D, et al. Compromised HOXA5 function can limit p53 expression in human breast tumours. Nature 2000;405: 974-8.

192. Cillo C, Barba P, Freschi G, Bucciarelli G, Magli MC, Boncinelli E. HOX gene expression in normal and neoplastic human kidney. Int J Cancer 1992;51: 892-7.

193. Ferrara N, Gerber HP, LeCouter J. The biology of VEGF and its receptors. Nat Med 2003;9: 669-76.

194. Park JE, Keller GA, Ferrara N. The vascular endothelial growth factor (VEGF) isoforms: differential deposition into the subepithelial extracellular matrix and bioactivity of extracellular matrix-bound VEGF. Mol Biol Cell 1993;4: 1317-26.

195. Ferrara N, Henzel WJ. Pituitary follicular cells secrete a novel heparin-binding growth factor specific for vascular endothelial cells. Biochem Biophys Res Commun 1989;161: 851-8.

140

196. Ferrara N. Role of vascular endothelial growth factor in the regulation of angiogenesis. Kidney Int 1999;56: 794-814.

197. Hayashibara T, Yamada Y, Miyanishi T, et al. Vascular endothelial growth factor and cellular chemotaxis: a possible autocrine pathway in adult T-cell leukemia cell invasion. Clin Cancer Res 2001;7: 2719-26.

198. Mayr-Wohlfart U, Waltenberger J, Hausser H, et al. Vascular endothelial growth factor stimulates chemotactic migration of primary human osteoblasts. Bone 2002;30: 472-7.

199. Waltenberger J, Lange J, Kranz A. Vascular endothelial growth factor-A-induced chemotaxis of monocytes is attenuated in patients with diabetes mellitus: A potential predictor for the individual capacity to develop collaterals. Circulation 2000;102: 185-90.

200. Ku DD, Zaleski JK, Liu S, Brock TA. Vascular endothelial growth factor induces EDRF-dependent relaxation in coronary arteries. Am J Physiol 1993;265: H586-92.

201. Olofsson B, Pajusola K, Kaipainen A, et al. Vascular endothelial growth factor B, a novel growth factor for endothelial cells. Proc Natl Acad Sci U S A 1996;93: 2576-81.

202. Cao Y. Positive and negative modulation of angiogenesis by VEGFR1 ligands. Sci Signal 2009;2: re1.

203. Kukk E, Lymboussaki A, Taira S, et al. VEGF-C receptor binding and pattern of expression with VEGFR-3 suggests a role in lymphatic vascular development. Development 1996;122: 3829-37.

204. Achen MG, Jeltsch M, Kukk E, et al. Vascular endothelial growth factor D (VEGF-D) is a ligand for the tyrosine kinases VEGF receptor 2 (Flk1) and VEGF receptor 3 (Flt4). Proc Natl Acad Sci U S A 1998;95: 548-53.

205. Terman BI, Dougher-Vermazen M, Carrion ME, et al. Identification of the KDR tyrosine kinase as a receptor for vascular endothelial cell growth factor. Biochem Biophys Res Commun 1992;187: 1579-86.

206. Shibuya M, Yamaguchi S, Yamane A, et al. Nucleotide sequence and expression of a novel human receptor-type tyrosine kinase gene (flt) closely related to the fms family. Oncogene 1990;5: 519-24.

207. Pajusola K, Aprelikova O, Korhonen J, et al. FLT4 receptor tyrosine kinase contains seven immunoglobulin-like loops and is expressed in multiple human tissues and cell lines. Cancer Res 1992;52: 5738-43.

208. Matthews W, Jordan CT, Gavin M, Jenkins NA, Copeland NG, Lemischka IR. A receptor tyrosine kinase cDNA isolated from a population of enriched primitive hematopoietic cells and exhibiting close genetic linkage to c-kit. Proc Natl Acad Sci U S A 1991;88: 9026-30.

209. Hilmi C, Guyot M, Pages G. VEGF spliced variants: possible role of anti-angiogenesis therapy. J Nucleic Acids 2012;2012: 162692.

141

210. Tischer E, Mitchell R, Hartman T, et al. The human gene for vascular endothelial growth factor. Multiple protein forms are encoded through alternative exon splicing. J Biol Chem 1991;266: 11947-54.

211. Houck KA, Ferrara N, Winer J, Cachianes G, Li B, Leung DW. The vascular endothelial growth factor family: identification of a fourth molecular species and characterization of alternative splicing of RNA. Mol Endocrinol 1991;5: 1806-14.

212. Lei J, Jiang A, Pei D. Identification and characterization of a new splicing variant of vascular endothelial growth factor: VEGF183. Biochim Biophys Acta 1998;1443: 400-6.

213. Poltorak Z, Cohen T, Sivan R, et al. VEGF145, a secreted vascular endothelial growth factor isoform that binds to extracellular matrix. J Biol Chem 1997;272: 7151-8.

214. Jingjing L, Xue Y, Agarwal N, Roque RS. Human Muller cells express VEGF183, a novel spliced variant of vascular endothelial growth factor. Invest Ophthalmol Vis Sci 1999;40: 752-9.

215. Whittle C, Gillespie K, Harrison R, Mathieson PW, Harper SJ. Heterogeneous vascular endothelial growth factor (VEGF) isoform mRNA and receptor mRNA expression in human glomeruli, and the identification of VEGF148 mRNA, a novel truncated splice variant. Clin Sci (Lond) 1999;97: 303-12.

216. Mineur P, Colige AC, Deroanne CF, et al. Newly identified biologically active and proteolysis-resistant VEGF-A isoform VEGF111 is induced by genotoxic agents. J Cell Biol 2007;179: 1261-73.

217. Bachtiary B, Selzer E, Knocke TH, Potter R, Obermair A. Serum VEGF levels in patients undergoing primary radiotherapy for cervical cancer: impact on progression-free survival. Cancer Lett 2002;179: 197-203.

218. Zusterzeel PL, Span PN, Dijksterhuis MG, Thomas CM, Sweep FC, Massuger LF. Serum vascular endothelial growth factor: a prognostic factor in cervical cancer. J Cancer Res Clin Oncol 2009;135: 283-90.

219. Shih T, Lindley C. Bevacizumab: an angiogenesis inhibitor for the treatment of solid malignancies. Clin Ther 2006;28: 1779-802.

220. Tewari KS, Sill M, Long HJ, et al. Incorporation of bevacizumab in the treatment of recurrent and metastatic cervical cancer: A phase III randomized trial of the Gynecologic Oncology Group. In: 2013 ASCO Annual Meeting; 2013 June 2, 2013; Chicago, IL; 2013.

221. Ma L, Teruya-Feldstein J, Weinberg RA. Tumour invasion and metastasis initiated by microRNA-10b in breast cancer. Nature 2007;449: 682-8.

222. Hui AB, Bruce JP, Alajez NM, et al. Significance of dysregulated metadherin and microRNA-375 in head and neck cancer. Clin Cancer Res 2011;17: 7539-50.

142

223. Ries LAG EM, Kosary CL, Hankey BF, Miller BA, Clegg L, Mariotto A, Fay MP, Feuer EJ, Edwards BK (eds). SEER Cancer Statistics Review, 1975-2000. 2003 [cited; Available from: http://seer.cancer.gov/csr/1975_2000/

224. Cancer Facts and Figures 2012. Atlanta, GA: American Cancer Society; 2012.

225. Livak KJ, Schmittgen TD. Analysis of relative gene expression data using real-time quantitative PCR and the 2(-Delta Delta C(T)) Method. Methods 2001;25: 402-8.

226. Scotto L, Narayan G, Nandula SV, et al. Identification of copy number gain and overexpressed genes on chromosome arm 20q by an integrative genomic approach in cervical cancer: potential role in progression. Genes Chromosomes Cancer 2008;47: 755-65.

227. Perez-Plasencia C, Vazquez-Ortiz G, Lopez-Romero R, Pina-Sanchez P, Moreno J, Salcedo M. Genome wide expression analysis in HPV16 cervical cancer: identification of altered metabolic pathways. Infect Agent Cancer 2007;2: 16.

228. Shah N, Sukumar S. The Hox genes and their roles in oncogenesis. Nat Rev Cancer 2010;10: 361-71.

229. Yekta S, Shih IH, Bartel DP. MicroRNA-directed cleavage of HOXB8 mRNA. Science 2004;304: 594-6.

230. Rubin E, Wu X, Zhu T, et al. A role for the HOXB7 homeodomain protein in DNA repair. Cancer Res 2007;67: 1527-35.

231. Care A, Felicetti F, Meccia E, et al. HOXB7: a key factor for tumor-associated angiogenic switch. Cancer Res 2001;61: 6532-9.

232. Care A, Silvani A, Meccia E, et al. HOXB7 constitutively activates basic fibroblast growth factor in melanomas. Mol Cell Biol 1996;16: 4842-51.

233. Storti P, Donofrio G, Colla S, et al. HOXB7 expression by myeloma cells regulates their pro-angiogenic properties in multiple myeloma patients. Leukemia 2011;25: 527-37.

234. Ohlsson Teague EM, Van der Hoek KH, Van der Hoek MB, et al. MicroRNA-regulated pathways associated with endometriosis. Mol Endocrinol 2009;23: 265-75.

235. Schotte D, Chau JC, Sylvester G, et al. Identification of new microRNA genes and aberrant microRNA profiles in childhood acute lymphoblastic leukemia. Leukemia 2009;23: 313-22.

236. Wang YX, Zhang XY, Zhang BF, Yang CQ, Chen XM, Gao HJ. Initial study of microRNA expression profiles of colonic cancer without lymph node metastasis. J Dig Dis 2010;11: 50-4.

237. Lakomy R, Sana J, Hankeova S, et al. MiR-195, miR-196b, miR-181c, miR-21 expression levels and MGMT methylation status are associated with clinical outcome in glioblastoma patients. Cancer Sci 2011;102: 2186-90.

143

238. Bhatia S, Kaul D, Varma N. Potential tumor suppressive function of miR-196b in B-cell lineage acute lymphoblastic leukemia. Mol Cell Biochem 2010;340: 97-106.

239. Choi YW, Bae SM, Kim YW, et al. Gene expression profiles in squamous cell cervical carcinoma using array-based comparative genomic hybridization analysis. Int J Gynecol Cancer 2007;17: 687-96.

240. Tsai KW, Hu LY, Wu CW, et al. Epigenetic regulation of miR-196b expression in gastric cancer. Genes Chromosomes Cancer 2010;49: 969-80.

241. Mallo M, Wellik DM, Deschamps J. Hox genes and regional patterning of the vertebrate body plan. Dev Biol 2010;344: 7-15.

242. Cillo C, Schiavo G, Cantile M, et al. The HOX gene network in hepatocellular carcinoma. Int J Cancer 2011;129: 2577-87.

243. Kelly ZL, Michael A, Butler-Manuel S, Pandha HS, Morgan RG. HOX genes in ovarian cancer. J Ovarian Res 2011;4: 16.

244. Rice KL, Licht JD. HOX deregulation in acute myeloid leukemia. J Clin Invest 2007;117: 865-8.

245. Leung DW, Cachianes G, Kuang WJ, Goeddel DV, Ferrara N. Vascular endothelial growth factor is a secreted angiogenic mitogen. Science 1989;246: 1306-9.

246. Hanahan D, Weinberg RA. Hallmarks of cancer: the next generation. Cell 2011;144: 646-74.

247. Liang Y, Brekken RA, Hyder SM. Vascular endothelial growth factor induces proliferation of breast cancer cells and inhibits the anti-proliferative activity of anti-hormones. Endocr Relat Cancer 2006;13: 905-19.

248. Buchler P, Reber HA, Buchler MW, Friess H, Hines OJ. VEGF-RII influences the prognosis of pancreatic cancer. Ann Surg 2002;236: 738-49; discussion 49.

249. Fan F, Wey JS, McCarty MF, et al. Expression and function of vascular endothelial growth factor receptor-1 on human colorectal cancer cells. Oncogene 2005;24: 2647-53.

250. Wey JS, Fan F, Gray MJ, et al. Vascular endothelial growth factor receptor-1 promotes migration and invasion in pancreatic carcinoma cell lines. Cancer 2005;104: 427-38.

251. Dirix LY, Vermeulen PB, Pawinski A, et al. Elevated levels of the angiogenic cytokines basic fibroblast growth factor and vascular endothelial growth factor in sera of cancer patients. Br J Cancer 1997;76: 238-43.

252. Ferrara N, Hillan KJ, Gerber HP, Novotny W. Discovery and development of bevacizumab, an anti-VEGF antibody for treating cancer. Nat Rev Drug Discov 2004;3: 391-400.

253. Harfe BD. MicroRNAs in vertebrate development. Curr Opin Genet Dev 2005;15: 410-5.

144

254. Calin GA, Croce CM. MicroRNA signatures in human cancers. Nat Rev Cancer 2006;6: 857-66.

255. Waggott D, Chu K, Yin S, Wouters BG, Liu FF, Boutros PC. NanoStringNorm: an extensible R package for the pre-processing of NanoString mRNA and miRNA data. Bioinformatics 2012;28: 1546-8.

256. Tibshirani R. The lasso method for variable selection in the Cox model. Stat Med 1997;16: 385-95.

257. Fyles A, Milosevic M, Hedley D, et al. Tumor hypoxia has independent predictor impact only in patients with node-negative cervix cancer. J Clin Oncol 2002;20: 680-7.

258. Lyng H, Vorren AO, Sundfor K, et al. Intra- and intertumor heterogeneity in blood perfusion of human cervical cancer before treatment and after radiotherapy. Int J Cancer 2001;96: 182-90.

259. Wong RK, Fyles A, Milosevic M, Pintilie M, Hill RP. Heterogeneity of polarographic oxygen tension measurements in cervix cancer: an evaluation of within and between tumor variability, probe position, and track depth. Int J Radiat Oncol Biol Phys 1997;39: 405-12.

260. Lyng H, Beigi M, Svendsrud DH, et al. Intratumor chromosomal heterogeneity in advanced carcinomas of the uterine cervix. Int J Cancer 2004;111: 358-66.

261. Guo Z, Wu F, Asplund A, et al. Analysis of intratumoral heterogeneity of chromosome 3p deletions and genetic evidence of polyclonal origin of cervical squamous carcinoma. Mod Pathol 2001;14: 54-61.

262. Chang AR, Grignon DJ, Keeney MM, Koster JL, Kirk ME. DNA content in cervical carcinoma: a flow cytometric assessment of DNA heterogeneity. Int J Gynecol Pathol 1994;13: 330-6.

263. Nguyen HN, Sevin BU, Averette HE, Ramos R, Ganjei P, Perras J. Evidence of tumor heterogeneity in cervical cancers and lymph node metastases as determined by flow cytometry. Cancer 1993;71: 2543-50.

264. Koopman LA, Corver WE, van der Slik AR, Giphart MJ, Fleuren GJ. Multiple genetic alterations cause frequent and heterogeneous human histocompatibility leukocyte antigen class I loss in cervical cancer. J Exp Med 2000;191: 961-76.

265. Bachtiary B, Boutros PC, Pintilie M, et al. Gene expression profiling in cervical cancer: an exploration of intratumor heterogeneity. Clin Cancer Res 2006;12: 5632-40.

266. Raychaudhuri M, Schuster T, Buchner T, et al. Intratumoral heterogeneity of microRNA expression in breast cancer. J Mol Diagn 2012;14: 376-84.

267. Liu A, Tetzlaff MT, Vanbelle P, et al. MicroRNA expression profiling outperforms mRNA expression profiling in formalin-fixed paraffin-embedded tissues. Int J Clin Exp Pathol 2009;2: 519-27.

145

268. Zhang X, Chen J, Radcliffe T, Lebrun DP, Tron VA, Feilotter H. An array-based analysis of microRNA expression comparing matched frozen and formalin-fixed paraffin-embedded human tissue samples. J Mol Diagn 2008;10: 513-9.

269. Doleshal M, Magotra AA, Choudhury B, Cannon BD, Labourier E, Szafranska AE. Evaluation and validation of total RNA extraction methods for microRNA expression analyses in formalin-fixed, paraffin-embedded tissues. J Mol Diagn 2008;10: 203-11.

270. Nair VS, Maeda LS, Ioannidis JP. Clinical outcome prediction by microRNAs in human cancer: a systematic review. J Natl Cancer Inst 2012;104: 528-40.

271. Tan PK, Downey TJ, Spitznagel EL, Jr., et al. Evaluation of gene expression measurements from commercial microarray platforms. Nucleic Acids Res 2003;31: 5676-84.

272. Badis G, Berger MF, Philippakis AA, et al. Diversity and complexity in DNA recognition by transcription factors. Science 2009;324: 1720-3.

273. Laurie CC, Laurie CA, Rice K, et al. Detectable clonal mosaicism from birth to old age and its relationship to cancer. Nat Genet 2012;44: 642-50.

274. Kang HJ, Voleti B, Hajszan T, et al. Decreased expression of synapse-related genes and loss of synapses in major depressive disorder. Nat Med 2012;18: 1413-7.

275. Irizarry RA, Warren D, Spencer F, et al. Multiple-laboratory comparison of microarray platforms. Nat Methods 2005;2: 345-50.

276. van Hooff SR, Leusink FK, Roepman P, et al. Validation of a gene expression signature for assessment of lymph node metastasis in oral squamous cell carcinoma. J Clin Oncol 2012;30: 4104-10.

277. Turnbull AK, Kitchen RR, Larionov AA, Renshaw L, Dixon JM, Sims AH. Direct integration of intensity-level data from Affymetrix and Illumina microarrays improves statistical power for robust reanalysis. BMC Med Genomics 2012;5: 35.

278. Canales RD, Luo Y, Willey JC, et al. Evaluation of DNA microarray results with quantitative gene expression platforms. Nat Biotechnol 2006;24: 1115-22.

279. Shi L, Campbell G, Jones WD, et al. The MicroArray Quality Control (MAQC)-II study of common practices for the development and validation of microarray-based predictive models. Nat Biotechnol 2010;28: 827-38.

280. Rutgers E, Piccart-Gebhart MJ, Bogaerts J, et al. The EORTC 10041/BIG 03-04 MINDACT trial is feasible: results of the pilot phase. Eur J Cancer 2011;47: 2742-9.

281. Kratz JR, He J, Van Den Eeden SK, et al. A practical molecular assay to predict survival in resected non-squamous, non-small-cell lung cancer: development and international validation studies. Lancet 2012;379: 823-32.

282. Cobleigh MA, Tabesh B, Bitterman P, et al. Tumor gene expression and prognosis in breast cancer patients with 10 or more positive lymph nodes. Clin Cancer Res 2005;11: 8623-31.

146

283. Paik S, Shak S, Tang G, et al. A multigene assay to predict recurrence of tamoxifen-treated, node-negative breast cancer. N Engl J Med 2004;351: 2817-26.

284. Finke J, Fritzen R, Ternes P, Lange W, Dolken G. An improved strategy and a useful housekeeping gene for RNA analysis from formalin-fixed, paraffin-embedded tissues by PCR. Biotechniques 1993;14: 448-53.

285. Park YN, Abe K, Li H, Hsuih T, Thung SN, Zhang DY. Detection of hepatitis C virus RNA using ligation-dependent polymerase chain reaction in formalin-fixed, paraffin-embedded liver tissues. Am J Pathol 1996;149: 1485-91.

286. Budczies J, Weichert W, Noske A, et al. Genome-wide gene expression profiling of formalin-fixed paraffin-embedded breast cancer core biopsies using microarrays. J Histochem Cytochem 2011;59: 146-57.

287. Huang J, Qi R, Quackenbush J, Dauway E, Lazaridis E, Yeatman T. Effects of ischemia on gene expression. J Surg Res 2001;99: 222-7.

288. Spruessel A, Steimann G, Jung M, et al. Tissue ischemia time affects gene and protein expression patterns within minutes following surgical tumor excision. Biotechniques 2004;36: 1030-7.

289. Botling J, Edlund K, Segersten U, et al. Impact of thawing on RNA integrity and gene expression analysis in fresh frozen tissue. Diagn Mol Pathol 2009;18: 44-52.

290. Turner L, Heath JD, Kurn N. Gene expression profiling of RNA extracted from FFPE tissues: NuGEN technologies' whole-transcriptome amplification system. Methods Mol Biol 2011;724: 269-80.

291. Irizarry RA, Bolstad BM, Collin F, Cope LM, Hobbs B, Speed TP. Summaries of Affymetrix GeneChip probe level data. Nucleic Acids Res 2003;31: e15.

292. Dai M, Wang P, Boyd AD, et al. Evolving gene/transcript definitions significantly alter the interpretation of GeneChip data. Nucleic Acids Res 2005;33: e175.

293. Boutros PC. LTR: Linear Cross-Platform Integration of Microarray Data. Cancer Inform 2010;9: 197-208.

294. Smyth GK. Linear models and empirical bayes methods for assessing differential expression in microarray experiments. Stat Appl Genet Mol Biol 2004;3: Article3.

295. Storey JD, Tibshirani R. Statistical significance for genomewide studies. Proc Natl Acad Sci U S A 2003;100: 9440-5.

296. Zeeberg BR, Feng W, Wang G, et al. GoMiner: a resource for biological interpretation of genomic and proteomic data. Genome Biol 2003;4: R28.

297. Chen H, Boutros PC. VennDiagram: a package for the generation of highly-customizable Venn and Euler diagrams in R. BMC Bioinformatics 2011;12: 35.

147

298. Martinez R, Esteller M. The DNA methylome of glioblastoma multiforme. Neurobiol Dis 2010;39: 40-6.

299. Arezi B, Xing W, Sorge JA, Hogrefe HH. Amplification efficiency of thermostable DNA polymerases. Anal Biochem 2003;321: 226-35.

300. Tillinghast GW. Microarrays in the clinic. Nat Biotechnol 2010;28: 810-2.

301. Tang W, Hu Z, Muallem H, Gulley ML. Clinical implementation of RNA signatures for pharmacogenomic decision-making. Pharmgenomics Pers Med 2011;4: 95-107.

302. Caretti E, Devarajan K, Coudry R, et al. Comparison of RNA amplification methods and chip platforms for microarray analysis of samples processed by laser capture microdissection. J Cell Biochem 2008;103: 556-63.

303. Clement-Ziza M, Gentien D, Lyonnet S, Thiery JP, Besmond C, Decraene C. Evaluation of methods for amplification of picogram amounts of total RNA for whole genome expression profiling. BMC Genomics 2009;10: 246.

304. Vermeulen J, Derveaux S, Lefever S, et al. RNA pre-amplification enables large-scale RT-qPCR gene-expression studies on limiting sample amounts. BMC Res Notes 2009;2: 235.

305. Wilson CL, Pepper SD, Hey Y, Miller CJ. Amplification protocols introduce systematic but reproducible errors into gene expression studies. Biotechniques 2004;36: 498-506.

306. Puskas LG, Zvara A, Hackler L, Jr., Van Hummelen P. RNA amplification results in reproducible microarray data with slight ratio bias. Biotechniques 2002;32: 1330-4, 6, 8, 40.

307. Johnson WE, Li C, Rabinovic A. Adjusting batch effects in microarray expression data using empirical Bayes methods. Biostatistics 2007;8: 118-27.

308. Bianchi F, Nicassio F, Marzi M, et al. A serum circulating miRNA diagnostic test to identify asymptomatic high-risk individuals with early stage lung cancer. EMBO Mol Med 2011;3: 495-503.

309. Iorio MV, Ferracin M, Liu CG, et al. MicroRNA gene expression deregulation in human breast cancer. Cancer Res 2005;65: 7065-70.

310. Volinia S, Galasso M, Sana ME, et al. Breast cancer signatures for invasiveness and prognosis defined by deep sequencing of microRNA. Proc Natl Acad Sci U S A 2012;109: 3024-9.

311. Ginzinger DG. Gene quantification using real-time quantitative PCR: an emerging technology hits the mainstream. Exp Hematol 2002;30: 503-12.

312. Geiss GK, Bumgarner RE, Birditt B, et al. Direct multiplexed measurement of gene expression with color-coded probe pairs. Nat Biotechnol 2008;26: 317-25.

313. Boutros PC, Okey AB. Unsupervised pattern recognition: an introduction to the whys and wherefores of clustering microarray data. Brief Bioinform 2005;6: 331-43.

148

314. Wang B, Wang XF, Howell P, et al. A personalized microRNA microarray normalization method using a logistic regression model. Bioinformatics 2010;26: 228-34.

315. Sato F, Tsuchiya S, Terasawa K, Tsujimoto G. Intra-platform repeatability and inter-platform comparability of microRNA microarray technology. PLoS One 2009;4: e5540.

316. Git A, Dvinge H, Salmon-Divon M, et al. Systematic comparison of microarray profiling, real-time PCR, and next-generation sequencing technologies for measuring differential microRNA expression. RNA 2010;16: 991-1006.

317. Sah S, McCall MN, Eveleigh D, Wilson M, Irizarry RA. Performance evaluation of commercial miRNA expression array platforms. BMC Res Notes 2010;3: 80.

318. Wang B, Howel P, Bruheim S, et al. Systematic evaluation of three microRNA profiling platforms: microarray, beads array, and quantitative real-time PCR array. PLoS One 2011;6: e17167.

319. Ach RA, Wang H, Curry B. Measuring microRNAs: comparisons of microarray and quantitative PCR measurements, and of different total RNA prep methods. BMC Biotechnol 2008;8: 69.

320. Liu T, Papagiannakopoulos T, Puskar K, et al. Detection of a microRNA signal in an in vivo expression set of mRNAs. PLoS One 2007;2: e804.

321. Creighton CJ, Hernandez-Herrera A, Jacobsen A, et al. Integrated analyses of microRNAs demonstrate their widespread influence on gene expression in high-grade serous ovarian carcinoma. PLoS One 2012;7: e34546.

322. Wang L, Oberg AL, Asmann YW, et al. Genome-wide transcriptional profiling reveals microRNA-correlated genes and biological processes in human lymphoblastoid cell lines. PLoS One 2009;4: e5878.

323. Wang YP, Li KB. Correlation of expression profiles between microRNAs and mRNA targets using NCI-60 data. BMC Genomics 2009;10: 218.

characterization of altered microrna expression …...iv acknowledgments first and foremost, i would...

Documents