Predicting the impact of mutations using pathway-

guided integrative genomicsNetwork Biology SIG, ISMB 2012

Josh Stuart, UC Santa Cruz

July 12, 2012

Overview of pathway-guided approach

Integrate many data sources to gain accurate view of how genes are functioning in pathways

Predict the functional consequences of mutations by quantifying the effect on the surrounding pathway

Use pathway signatures to implicate mutations in novel genes to (re-)focus targeting

Identify critical “Achilles Heels” in the pathways that distinguish a particular sub-type

Flood of Data Analysis Challenges

Expression

DNA Methylation

StructuralVariation

ExomeSequences

Copy NumberAlterations

Multiple, PossiblyConflicting Signals

This is What itDoes to You

Genomics, Functional Genomics, Metabolomics, Epigenomics =

Analysis of disease samples like automotive repair

(or detective work or other sleuthing)Patient Sample 1 Patient Sample 2

Patient Sample 3 Patient Sample N

…

Sleuths use as much knowledge as possible.

Much Cell Machinery Known:Gene circuitry now available.

Curated and/or CollectedReactome

KEGGBiocartaNCI-PID

Pathway Commons…

Integration key to interpret gene function

Expression not always an indicator of activityDownstream effects often provide clues

TF

Inference:TF is OFF

(high expressionbut inactive)

TF

Inference:TF is ON

(low-expressionbut active )

TF

Inference:TF is ON(expression

reflectsactivity)

Expression of 3 transcription factors:

high high low

Integration key to interpret gene function

Need multiple data modalities to get it right.

TF

Expression -> TF ON

BUT, targets are amplified

Lowers our beliefin active TF becauseexplained away bycis evidence.

Copy Number -> TF OFF

Probabilistic Graphical Models:A Language for Integrative Genomics

Generalize HMMs, Kalman Filters, Regression, Boolean Nets, etc.Language of probability ties together multiple aspects of gene

function & regulationEnable data-driven discovery of biological mechanismsFoundation: J. Pearl, D. Heckerman, E. Horvitz, G. Cooper, R.

Schacter, D. Koller, N. Friedman, M. Jordan, …Bioinformatics: D. Pe’er, A. Hartemink, E. Segal, E Schadt…

Nir Friedman, Science (2004) - Review

Integration Approach: Detailed models of gene expression and interaction

MDM2

TP53

Integration Approach: Detailed models of expression and interaction

MDM2

TP53

Two Parts:

1. Gene Level Model(central dogma)

2. Interaction Model(regulation)

PARDIGM Gene Model to Integrate Data

Vaske et al. 2010. Bioinformatics

1. Central Dogma-LikeGene Model of Activity

2. Interactions that connect to specific pointsin gene regulation map

Charlie VaskeSteve Benz

Integrated Pathway Analysis for Cancer Integrated dataset for downstream analysis

Inferred activities reflect neighborhood of influence around a gene.

Can boost signal for survival analysis and mutation impact

Multimodal Data

CNV

mRNA

meth

Pathway Modelof Cancer

Cohort Inferred Activities

…

TCGA Ovarian CancerInferred Pathway Activities

Patient Samples (247)Pa

thw

ay C

once

pts

(867

)

TCGA Network. 2011. Nature(lead by Paul Spellman)

Ovarian: FOXM1 pathway alteredin majority of serous ovarian tumors

FOXM1 Transcription Network

Patient Samples (247)

Path

way

Con

cept

s (8

67)

TCGA Network. 2011. Nature(lead by Paul Spellman)

FOXM1 central to cross-talk betweenDNA repair and cell proliferation

in Ovarian Cancer

TCGA Network. 2011. Nature (lead by Paul Spellman)

Pathway signatures of mutations to reveal therapeutic candidates

Mutated genes are the focus of many targeted approaches.

Some patients with “right” mutation don’t respond. Why?

Many cancers have one of several “novel” mutations. Can these be targeted with current approaches?

Pathway-motivated approaches:Identify gain-of-function from loss-of-function.

Compare novel signatures

Sam NgISMB

Oral Poster

PARADIGM-Shift Predicting the Impact of Mutations On Genetic Pathways

FG

Inference using all neighbors

FG

Inference using downstream neighbors

FG

Inference using upstream neighbors

SHIFT

Sam Ng, ECCB 2012

High Inferred Activity

Low Inferred Activity

mutatedgene

FG

PARADIGM-Shift Overview

Sam Ng, ECCB 2012

FG FG

1.IdentifyLocal

Neighbor-hood

PARADIGM-Shift Overview

Sam Ng, ECCB 2012

FG FG

1.IdentifyLocal

Neighbor-hood

FG

FG

2a.Regulators

Run

2b.Targets

Run

PARADIGM-Shift Overview

Sam Ng, ECCB 2012

FG FG

1.IdentifyLocal

Neighbor-hood

FG

FG

2a.Regulators

Run

2b.Targets

Run

P-ShiftScore

3.Calculate

Difference

FGFG -

FG

(LOF)

PARADIGM-Shift Overview

Sam Ng, ECCB 2012

P-Shift Predicts RB1 Loss-of-Function in GBM

Shift Score

PARADIGM downstream

PARADIGM upstream

Expression

Mutation

RB1

Sam Ng, ECCB 2012

RB1 Network (GBM)

Neighbor Gene Key

Activity

Expression

RB1 Mutation

Focus Gene Key

P-Shift

Expression

Mutation

T-Run

R-Run

Sam Ng, ECCB 2012

RB1 Discrepancy Scores distinguish mutated vs non-mutated samples

Signal Score (t-statistic) = -5.78

Sam Ng

RB1 discrepancy distinction is significant

Given the same network topology, how likely would we call a gain/loss of function Background model: permute gene labels in our dataset Compare observed signal score to signal scores (SS) obtained

from background modelObserved SS

Background SS

Sam Ng

TP53 Network

Sam Ng

Gain-of-Function (LUSC)

NFE2L2

P-Shift Score

PARADIGM downstream

PARADIGM upstream

Expression

Mutation

Sam Ng

NFE2L2 Network (LUSC)

Sam Ng

Neighbor Gene Key

Activity

Expression

RB1 Mutation

Focus Gene Key

P-Shift

Expression

Mutation

T-Run

R-Run

NFE2L2

Discrepancy scores are sensitive

Signal Score (t-statistic) = -5.78 Signal Score (t-statistic) = 4.985

RB1 NFE2L2

Signal Score (t-statistic) = -10.94

TP53

Observed SS

Background SS

Sam Ng

Expect passenger mutations to lack shifts

Is the discrepancy specific?Negative control: calculate scores for “passenger” mutations

Passengers:insignificant by MutSig (p > 0.10) well-represented in our pathways

Discrepancy of these “neutral” mutations should be close to what’s expected by chance (from permuted)

Sam Ng

Discrepancies of Passenger Mutationsare NOT distinctive

Sam Ng

PARADIGM-Shift gives orthogonal view of the importance of mutations in

LUSC

Enables probing into infrequent eventsCan detect non-coding mutation impact (pseudo FPs) Can detect presence of pathway compensation for

those seemingly functional mutations (pseudo FPs)Extend beyond mutationsLimited to genes w/ pathway representation

NFE2L2 (29)

CDKN2A (n=30)

Pat

hway

Dis

crep

ancy

LUS

C

MET (n=7) (gefitinib resistance)

HIF3A (n=7)TBC1D4 (n=9) (AKT signaling)

MAP2K6 (n=5)

EIF4G1 (n=20)

GLI2 (n=10) (SHH signaling)

AR (n=8)

Sam Ng

PARADIGM in TCGA patient BRCA tumors

Christina Yau,Buck Inst

Christina Yau, Buck Inst.

PATHMARK: Identify Pathway-based “markers” that underlie sub-types

Identify sub-pathways that distinguish patients sub-types (e.g. mutant vs. non-mutant, response to drug, etc)

Predict mutation impact on pathway “neighborhood”

Identify master control points for drug targeting.

Predict outcomes with quantitative simulations.

Insight from contrast

Ted GoldsteinSam Ng

PathMark: Differential Subnetworks from a “SuperPathway”

Pathway Activities

Pathway Activities

Ted Goldstein Sam Ng

PathMark: Differential Subnetworks from a “SuperPathway”

SuperPathway Activities

SuperPathway Activities

PathwaySignature

Ted Goldstein Sam Ng

Traditional methods treat each gene as a separate feature

Use features reflecting overall pathway activity

Smaller number of features are now fed to predictorsPredictor

PIL: Pathway-informed Learning

Artem Sokolov

Traditional methods treat each gene as a separate feature

Use features reflecting overall pathway activity

Smaller number of features are now fed to predictorsPredictor

PIL: Pathway-informed Learning

Artem Sokolov

ISMB poster

Basal vs. Luminal

Recursive feature elimination: we train an SVM, drop the least important half of features and recurse

The number of times each feature survived the elimination across 100 random splits of data

Artem Sokolov

Methotrexate Sensitivity

Non sub-type specific drug

Artem Sokolov

Pathway involvingthe target of the drug.

One large highly-connectedcomponent (size and connectivity

significant according to permutationanalysis)

Higher activity in ER-

Lower activity in ER-

Triple Negative Breast Pathway MarkersIdentified from 50 Cell Lines

980 pathway concepts1048 interactions

HIF1A/ARNT

Characterized byseveral “hubs’

IL23/JAK2/TYK2

Myc/Max

P53tetramer

ER

FOXA1

Sam Ng, Ted Goldstein

Identify master controllers usingSPIA (signaling pathway impact analysis)

Impactfactor:

Google PageRank for NetworksDetermines affect of a given pathway on each nodeCalculates perturbation factor for each node in the networkTakes into account regulatory logic of interactions.

Yulia Newton (NetBio SIG Poster)

Google’s PageRank-Like

Slight Trick: Run SPIA in reverse

Reverse edges in Super PathwayHigh scoring genes now those at the “top” of

the pathway

Yulia Newton

PageRank findshighly referenced

Reverse to findHighly referencing

Master Controller Analysis on Breast Cell Lines

BasalLuminal

Yulia Newton

Master regulators predict response to drugs:PLK3 predicted as a target for basal breast

UpDown

• DNA damage network is upregulated in basal breast cancers

• Basal breast cancers are sensitive to PLK inhibitors

Basal

Claudin-low

Luminal

GSK-PLKi

Heiser et al. 2011 PNAS

Ng, Goldstein

• HDAC Network is down-regulated in basal breast cancer cell lines

• Basal/CL breast cancers are resistant to HDAC inhibitors

VORINOSTATHDAC inhibitor

HDAC inhibitors predicted for luminal breast

Heiser et al. 2011 PNAS Ng, Goldstein

Connect genomic alterations to downstream expression/activity

• What circuitry connects mutations to transcriptional changes?– Mutations general (epi-) genomic perturbation– Expression activity

• Mutation/perturbation and expression/activity treated as heat diffusing on a network– HotNet, Vandin F, Upfal E, B.J. Raphael, 2008.– HotNet used in ovarian to implicate Notch pathway

• Find subnetworks that link genetic to mRNA and protein-level changes.

?

Evan PaullISMB

Oral Poster

HotLinkGenomic Perturbations(Mutations, Methylation, Focal Copy Number)

Gene Activity(Expression, RPPA, PARADIGM)

?

Evan Paull

HotLinkGenomic Perturbations(Mutations, Methylation, Focal Copy Number)

Gene Activity(Expression, RPPA, PARADIGM)

1. Add heat 2. Diffuse heat 3. Cut out linkers

Evan Paull

HotLinkGenomic Perturbations(Mutations, Methylation, Focal Copy Number)

Gene Activity(Expression, RPPA, PARADIGM)

Linking Sub-Pathway

Evan Paull

HotLink “Double Diffusion” Overview

HotLink Causal Paths

Significance of Interlinking Network

Double diffusion better than single

Basal-LumAHotLink

Basal LumA

TCGA Interlinking Network

Basal-LumA HotLink MapMAP, PI3K, AKT

TP53, RB1

MYC, FOXM1

Basal LumA

AKT/PI3KMAPK8, MAPK14 (p38-alpha)identified as mediators.

TP53, RB1

Basal LumA

MYC Neighborhood

Basal LumA

Double feedback loop involving TP53, RB1, CDK4, FOXM1, and MYC

AKT signaling

Basal LumA

Defining Pathway Signatures for Mutations and Sub-Types

Build a signature for every mutation and tumor/clinical event.

Correlate every signature to each other.Reveals common molecular similarities

between different divisions of patient subgroups

Mutations in novel genes may “phenocopy” mutations in known genes

Ted Goldstein

Mutation Association to PathwaysWhat pathway activities is a mutation’s presence

associated?Can we classify mutations based on these

associations?Mutations

PA

RA

DIG

M S

igna

ture

s

Ted Goldstein

Mutation Association to Pathways

What pathway activities is a mutation’s presence associated?

Can we classify mutations based on these associations?

(Note: CRC figure below; soon for BRCA)

Mutations

PA

RA

DIG

M S

igna

ture

s

APC and other Wnt

Ted Goldstein

Mutation Association to PathwaysWhat pathway activities is a mutation’s presence

associated?Can we classify mutations based on these

associations?Mutations

PA

RA

DIG

M S

igna

ture

s

Ted Goldstein

Mutation Association to PathwaysWhat pathway activities is a mutation’s presence

associated?Can we classify mutations based on these

associations?

(Note: CRC figure below; soon for BRCA)

Mutations

PA

RA

DIG

M S

igna

ture

s

TGFB Pathway mutations

Ted Goldstein

Mutation Association to PathwaysWhat pathway activities is a mutation’s presence

associated?Can we classify mutations based on these

associations?

(Note: CRC figure below; soon for BRCA)

Mutations

PA

RA

DIG

M S

igna

ture

s

PIK3CA, RTK pathway, KRAS

Ted Goldstein

Mutation Association to PathwaysWhat pathway activities is a mutation’s presence

associated?Can we classify mutations based on these

associations?

(Note: CRC figure below; soon for BRCA)

Mutations

PA

RA

DIG

M S

igna

ture

s

Evidence forAHNAK2 actingPI3KCA-like?

Ted Goldstein

Summary• Modeling information flow on known pathways

gives view of gene activity.• Patient stratification into pathway-based subtypes• Sub-networks provide pathway-based signatures of

sub-types and mutations.• Loss- and gain-of-function predicted from pathway

neighbors for even rare mutations.• Identify interlinking genes associated with

mutations to implicate additional targets even in LOF cases• E.g. Target MYC-related pathways in certain

TP53-deficient cells?• Current work: Use pathways to now simulate the

affects of specific knock-downs.

James DurbinChris Szeto

UCSC Integrative Genomics Group

Sam Ng

Ted Golstein

Dan Carlin Evan Paull

Artem Sokolov

Chris Wong

Marcos Woehrmann

Yulia Newton

Please See Posters!

Acknowledgments

UCSC Cancer Genomics• Kyle Ellrott• Brian Craft• Chris Wilks• Amie Radenbaugh• Mia Grifford• Sofie Salama• Steve Benz

UCSC Genome Browser Staff• Mark Diekins• Melissa Cline• Jorge Garcia• Erich Weiler

Buck Institute for Aging• Christina Yau• Sean Mooney• Janita Thusberg

Collaborators• Joe Gray, LBL• Laura Heiser, LBL• Eric Collisson, UCSF• Nuria Lopez-Bigas, UPF• Abel Gonzalez, UPF

Funding Agencies• NCI/NIH• SU2C• NHGRI• AACR• UCSF Comprehensive Cancer

Center• QB3

Jing Zhu

David Haussler

Chris Benz,

Broad Institute• Gaddy Getz• Mike Noble• Daniel DeCara

UCSC Cancer Browsergenome-cancer.ucsc.edu

Jing Zhu

supplemental

325 Genes, 2213 Interactions

“Backbone” of 43 genes, 90 connections

“Backbone” of 43 genes, 90 connectionsMaster regulators: Cell-Cell Signaling, AKT1, Cyclin D1

“Backbone” of 43 genes, 90 connectionsMajor PARADIGM hubs included: MYC, FOXM1, FOXA1, HIF1A/ARNT

“Backbone” of 43 genes, 90 connectionsSignaling through beta-catenin explains MYC activity in basals:- deletions in CDKN2A de-repress CTNNB1 in basals or- lower expression of Cyclin D1 de-repress CTNNB1

Top 20 basal master controllers

Yulia Newton

RNAi vs. Master Controller (after recurring runs)

Yulia Newton

RPS6KA3- p90 S6 kinase

High-scoring afterIterative runs.

Basals differentiallySensitive to RNAi

Inhibitors available.Bas

al v

s Lu

min

al R

NA

i Gro

wth

Master Controller Score

RPS6KA3

AKT1

RAF1

PDPK1

AKT2