cs 5854: projects - undergraduate...

Focus Projects PL 2.0 Expression to Signals ToxCast PanCancer Orient XTalk 2.0 NLP Feedback Support

CS 5854: Projects

T. M. Murali

January 26, 28, February 2, 2016

T. M. Murali January 26, 28, February 2, 2016 CS 5854: Projects


Continuum of Models in Systems Biology

From Building with a scaffold: emerging strategies for high- to low-level cellular modeling, Ideker

and Lauffenburger, Trends in Biotechnology Volume 21, Issue 6 , June 2003, Pages 255-262.



Goals of the Course

I Emphasise a data-driven approach to systems biology.

I Integrate massive quantities of different types of data

I Stress methods that can prioritise experiments.

I Learn techniques from clustering, data mining, and graph theory andapply them to solve specific biological questions.



Student PresentationsI Start on Tuesday, February 9.I Each presentation is 75 minutes long (including 20–30 minutes of

questions).I Forms groups of 2 ≤ k ≤ 3 students.I A group of k students will make k presentations.I Each group of k students should select one–two papers, one for each

class they will present.I Single student can lead/give each individual presentation, but read,

understand, and discuss papers as a group.I Use the tag 2016-spring-csb-papers on CiteULike.I Read (quickly) the papers before you select them!I Use the current schedule only as a guideline for how long I think each

paper will take to present. It is not the final schedule!I Send me the names of the groups and your top four paper choices by

Tuesday, January 26.I I will post the schedule by Thursday, January 28.


http://www.citeulike.org/user/tmmurali/tag/2016-spring-csb-papers


Class Projects that Resulted in Papers1. VIRGO: Computational Prediction of Gene Functions, Naveed Massjouni,

Corban Rivera, and T. M. Murali, Nucleic Acids Research, 2006.2. Network Legos: Building Blocks of Cellular Wiring Diagrams, T. M. Murali

and Corban G. Rivera, RECOMB 2007, JCB 2008.3. Computational Prediction of Interactions between Host and Pathogen

Proteins, Matthew Dyer, T. M. Murali, and Bruno Sobral, ISMB 2007.4. Divergence of Gene Expression Profiles in Tandemly Arrayed Genes in

Human and Mouse, Valia Shoja, T. M. Murali, and Liqing Zhang,Comparative and Functional Genomics, 2007.

5. Network-Based Prediction and Analysis of HIV Dependency Factors, T. M.Murali, Matthew D. Dyer, David Badger, Brett M. Tyler, and Michael G.Katze, PLoS Computational Biology, 2011.

6. Top-Down Network Analysis to Drive Bottom-Up Modeling of PhysiologicalProcesses, Christopher L. Poirel, Richard R. Rodrigues, Katherine C. Chen,John J. Tyson, and T. M. Murali, JCB, 2013.

7. “Pathways on Demand: Automatic Reconstruction of Human SignalingNetworks,” Anna Ritz, Christopher L. Poirel, Allison N. Tegge, NicholasSharp, Allison Powell, Kelsey Simmons, Shiv D. Kale, and T. M. Murali, npj:Systems Biology and Applications, 2016.



List of Projects

1. Develop PathLinker 2.0

2. Predict signal transduction pathways from gene expression data.

3. Compute chemical response networks

4. Analyze PanCancer data

5. Predict orientation of interactions in interaction networks.

6. Develop XTalk 2.0, discover links between insulin signaling andinflammation in diabetes

7. Use NLP to build gold standard for XTalk

8. Compute feedback loops in yeast regulatory networks



Overview

Focus of the Course

Projects

Develop PathLinker 2.0

Predict signal transduction pathways from gene expression data

Compute chemical response networks

Analyze PanCancer data

Predict orientation of interactions in networks

Develop XTalk 2.0

Use NLP to build gold standard for XTalk

Compute feedback loops in yeast regulatory networks

Support



Baron and Kneissel. WNT signaling in bone homeostasis and disease: from human mutations to treatments. Nature Medicine,2013.



Wnt Signaling in a Pathway Database

Destruction Complex

www.netpath.org/netslim



Signaling Pathways as Directed Graphs

Destruction Complex

FzdWnt

DirectedRegulatoryInteractions

GSK

β-Catenin

p

BidirectedPhysical

InteractionsFzd

Wnt

CTNNB1GSK



Signaling Pathways as Directed Graphs

Destruction Complex

Receptors

Transcription Factors andTranscriptional Regulators



Reconstructing Signaling Pathways

Human protein-protein interactome




Curated pathway is a subnetwork of the interactome




Question: Can we reconstruct the curated pathway given onlyreceptors and transcriptional regulators?




Proposed pathway reconstruction



Evaluating Reconstructed Pathways

Curated Pathway



Evaluating Reconstructed Pathways

Curated Pathway and Proposed Reconstruction



Output

ProposedReconstruction

Input

Receptors

TranscriptionalRegulators (TRs)

Human Interactome

PathLinker

I Developed PathLinker to reconstruct proteins and interactions

I Systematically evaluated PathLinker and other algorithms onhuman signaling pathways from the NetPath and KEGG databases

“Pathways on Demand: Automatic Reconstruction of Human Signaling Pathways,” Ritz et al., Systems Biology andApplications, a Nature partner journal, to appear, 2016.



Output

ProposedReconstruction

Evaluation

Curated Pathway

Input

Receptors

TranscriptionalRegulators (TRs)

Human Interactome

PathLinkerNetPathPathways

I Developed PathLinker to reconstruct proteins and interactions

I Systematically evaluated PathLinker and other algorithms onhuman signaling pathways from the NetPath and KEGG databases

“Pathways on Demand: Automatic Reconstruction of Human Signaling Pathways,” Ritz et al., Systems Biology andApplications, a Nature partner journal, to appear, 2016.



How Does PathLinker work?

PathLinker

RWR

IPA

0.4

0.6

0.8

Precision

Recall0.2 0.4

IPA (prec. 0.72)

PathLinker(prec. 0.65)

RWR(prec. 0.61)

(a)

(b)

Wnt network in GraphSpaceT. M. Murali January 26, 28, February 2, 2016 CS 5854: Projects

http://graphspace.org/graphs/[email protected]/Wnt-pathlinker-top200paths-labeled?layout=pathlinker-labeled-layout&[email protected]


PathLinker 2.0

I Current algorithm does not learn fromthe structure of the pathway.

I Goal: develop a (machine-learning)algorithm that can use information onthe edges in a pathway to predict newedges.

I Use cross-validation to testperformance: develop meaningful waysof deleting edges.



Evaluating Reconstructions with Cross Validation

Curated pathway




Edges removed for cross validation




Proposed reconstruction




Curated pathway and proposed reconstruction




Cross validation edges and proposed reconstruction




Precision and recall



Project Details

I Paper: Pathways on Demand: Automatic Reconstruction of HumanSignaling Pathways, Ritz et al., Systems Biology and Applications, aNature partner journal, to appear, 2016

I PathLinker code on GitHub

I Update network dataset used in PathLinker paper.

I Develop machine learning method that can utilize partial informationabout edges in a pathway.

I Find several meaningful alternatives methods to compare youralgorithm with.

I Create computational analyses that are different from the PathLinkerpaper.

I Do literature analysis of predicted interactions.


http://bioinformatics.cs.vt.edu/~murali/papers/2016-sysbio-appl-pathlinker-with-supplement.pdf


https://github.com/Murali-group/PathLinker


Overview

Focus of the Course

Projects






Develop XTalk 2.0



Support



Dissect Cellular Responses to Signals

Space of Disse

Hepatocytes

LSECs and Kupffer cells

PEM

Collagen

Endothelial/Kupffer cells

Hepatocyte

Collaboration with Padma Rajagopalan (Chem Eng)



Project DetailsI Papers:

I Designing a Multi-cellular Organotypic 3D Liver Model with a Detachable,Nanoscale Polymeric Space of Disse, Larkin et al., Tissue Eng., 2013.

I A genome-wide analysis in Saccharomyces cerevisiae demonstrates the influence ofchromatin modifiers on transcription, Steinfeld, Shamir, and Kupiec, NatureGenetics, 2007. Read the technique.

I Pathways on Demand: Automatic Reconstruction of Human Signaling Pathways,Ritz et al., Systems Biology and Applications, a Nature partner journal, in press,2016

I Use liver gene expression data from Padma Rajagopalan’s group.I Find other appropriate datasets (by the middle of February):

1. Gene expression measurements after a signal.I ideally for signals on liver tissue.I ideally, accompanied by a proteomic dataset.

2. Prefer human or mammalian datasets.3. Create a high-quality dataset of molecular interactions in appropriate

organism.4. If proteomic data is available, use it to measure performance.

I Use software for existing algorithms (PathLinker, eQED, PCSF, etc).I Perform literature based validation of predicted pathways.


http://online.liebertpub.com/doi/full/10.1089/ten.tec.2012.0700

http://online.liebertpub.com/doi/full/10.1089/ten.tec.2012.0700

http://www.nature.com/ng/journal/v39/n3/full/ng1965.html

http://www.nature.com/ng/journal/v39/n3/full/ng1965.html



Overview

Focus of the Course

Projects






Develop XTalk 2.0



Support



ToxCast → Chemical Response Networks

I Efforts such as the EPA’s ToxCast seek to quickly and efficientlyscreen thousands of chemicals for potential human and environmentaleffects.

I Information is partial: for each chemical, these efforts study a fixedset of proteins and pathways.

I Assays are independent: Do not consider the complex network ofinteractions among assayed proteins and pathways.

I We seek to connect the responding proteins in the context of theunderlying network of regulatory and physical interactions.

I Toxicant response network: network of regulatory, signaling andphysical interactions that connects proteins that are perturbed as aresult of toxicant exposure.

I Toxicant response networks may revealI important intermediate proteins that have not have been tested andI physiological processes that have not been previously implicated in

connection with the chemical.



ToxCast → Chemical Response Networks

I Efforts such as the EPA’s ToxCast seek to quickly and efficientlyscreen thousands of chemicals for potential human and environmentaleffects.

I Information is partial: for each chemical, these efforts study a fixedset of proteins and pathways.

I Assays are independent: Do not consider the complex network ofinteractions among assayed proteins and pathways.

I We seek to connect the responding proteins in the context of theunderlying network of regulatory and physical interactions.

I Toxicant response network: network of regulatory, signaling andphysical interactions that connects proteins that are perturbed as aresult of toxicant exposure.

I Toxicant response networks may revealI important intermediate proteins that have not have been tested andI physiological processes that have not been previously implicated in

connection with the chemical.T. M. Murali January 26, 28, February 2, 2016 CS 5854: Projects


Bisphenol-A Response Network

Responsive Protein

Non-responsive Protein

Physical Interaction

Regulatory Interaction

ALK1 Pathway

BMPPathway

RXR/VDRPathway

AktPathway

p38 Pathway

Activation of AP1

Down-regulation of TGFbeta Receptor

Signaling

PYK2Pathway

Posterior LiteraturePathway Probability Support

ALK1 Pathway (PID) 0.998BMP Pathway (PID) 0.996RXR VDR Pathway (PID) 0.981 XAKT Pathway (Biocarta) 0.881 XP38 γδ Pathway (PID) 0.808Activation of AP1 Family of TFs (Reactome) 0.711 XDownregulation of TGFβR Signaling (Reactome) 0.669PYK2 Pathway (Biocarta) 0.654 X



Bisphenol-A Response Network

Posterior LiteraturePathway Probability Support

ALK1 Pathway (PID) 0.998BMP Pathway (PID) 0.996RXR VDR Pathway (PID) 0.981 XAKT Pathway (Biocarta) 0.881 XP38 γδ Pathway (PID) 0.808Activation of AP1 Family of TFs (Reactome) 0.711 XDownregulation of TGFβR Signaling (Reactome) 0.669PYK2 Pathway (Biocarta) 0.654 X



Project Details

I Papers:I Pathways on Demand: Automatic Reconstruction of Human Signaling Pathways,

Ritz et al., Systems Biology and Applications, a Nature partner journal, to appear,2016

I Profiling 976 ToxCast Chemicals across 331 Enzymatic and Receptor SignalingAssays, Sipes et al., Chem Res Toxicol, 17; 26(6), 878–895, 2013.

I Phenotypic screening of the ToxCast chemical library to classify toxic andtherapeutic mechanisms, Kleinstreuer et al., Nature Biotechnology 32, 583–591,2014.

I Find other papers on ToxCast and Tox21 datasets and understandhow these data are accessible through the EPA and NIEHS websites.

I Apply PathLinker to each chemical to compute its “responsenetwork.”

I Integrate relevant gene expression dataset for each chemical, ifavailable.

I Demonstrate increased knowledge in response networks whencompared to the original ToxCast data.



http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3685188

http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3685188

http://www.nature.com/nbt/journal/v32/n6/full/nbt.2914.html

http://www.nature.com/nbt/journal/v32/n6/full/nbt.2914.html


Overview

Focus of the Course

Projects






Develop XTalk 2.0



Support



PanCancer + PathLinker

I Use PathLinker to analyze PanCancer data.I Papers:

I Pan-cancer network analysis identifies combinations of rare somaticmutations across pathways and protein complexes, Leiserson et al., NatGenet, 47, 2. 2015, 106–114, 2015.

I Discovering causal pathways linking genomic events to transcriptionalstates using Tied Diffusion Through Interacting Events (TieDIE)

I Pathway and network analysis of cancer genomes (Review)

I Most of these papers analyze overlap of the genes in the computednetworks with known physiological processes.

I Can we use PathLinker to discover mechanisms by which genomicevents control transcriptional changes?


http://www.citeulike.org/user/tmmurali/article/13475571




http://www.nature.com/nmeth/journal/v12/n7/full/nmeth.3440.html


Overview

Focus of the Course

Projects






Develop XTalk 2.0



Support



Baron and Kneissel. WNT signaling in bone homeostasis and disease: from human mutations to treatments. Nature Medicine,2013.



Wnt Signaling in a Pathway Database

Destruction Complex

www.netpath.org/netslim



Orienting Interactions

I Interactions in pathways are often directed because they arecorrespond to regulation.

I Interactions in PPI networks are undirected because they are physical.I Human protein interactome:

I 12K nodes and 110K edgesI 87K physical interactions

BIND, DIP, InnateDB, IntAct, MINT, MatrixDB, Reactome, NetPath, KEGG, SPIKE

I 33K signaling interactionsNetPath, KEGG, SPIKE

I Develop methods to assign direction to edges in PPI networks andquantitatively evaluate your results.



Approach for Orienting Interactions

Human protein-protein interactome




Pathway is a subnetwork of the interactome




Question: Can we assign edge orientations given onlyreceptors and transcriptional regulators?




Proposed orientations



Project Details

I Papers:I Discovering pathways by orienting edges in protein interaction

networks, Gitter et al., Nucleic Acids Research, 2011I Network orientation via shortest paths, Silverbush and Sharan,

Bioinformatics, 2014

I Software: PathLinker (Murali), OrientEdges (Gitter), Shortest(Sharan)

I Implement OrientEdges and Shortest yourself in Python within thePathLinker package.

I Evaluate by comparing to gold standard dataset of interactions withexperimentally known orientation.






https://github.com/Murali-group/PathLinker

http://sb.cs.cmu.edu/OrientEdges/software.html

http://www.cs.tau.ac.il/~roded/shortest.zip


Overview

Focus of the Course

Projects






Develop XTalk 2.0



Support



Signaling pathway crosstalk

Cell

NF2Mst1/2

Hippo Receptor

Lats1/2

YAP TAZ

YAP TAZp

General Assumption: pathways that share nodes crosstalk




Nucleus

SMAD2

TGFBR1/2

SMAD2

SMAD3

SMAD3p

SMAD2SMAD3p

Cell cycleApoptosis





Cell

SMAD2/3NF2

Mst1/2

Hippo Receptor

Lats1/2

E2F4/5

SMAD4

TGFBR1/2

YAP TAZSMAD2/3

X

YAPTAZp




Importance of Crosstalk

I Crosstalk is a naturally occuring phenomenon

I Crosstalk has been implicated in many diseases, including cancer, hostpathogen defense, and regulation of cell survival

I Surprisingly, no computational methods exist to predict which pairs ofpathways may crosstalk

XTalk: a method to identify which pathway pairs may crosstalk.



XTalk: average length of the k shortest paths

Inputs:

A Pathway A

B Pathway B

RA Receptors in A

TB TFs in B

k number of paths

G background networkPathway AReceptor

Pathway BTranscription Factor

Notation:

d(r , t, k) the length of the kth shortest path from r to t

χ(A,B, k) the XTalk statistic

χ(A,B, k) =1

k |TB |∑

t∈TB

k∑l=1

d(RA, t, l),



XTalk: average length of the k shortest paths

Inputs:

A Pathway A

B Pathway B

RA Receptors in A

TB TFs in B

k number of paths

G background networkNotation:

d(r , t, k) the length of the kth shortest path from r to t

χ(A,B, k) the XTalk statistic

χ(A,B, k) =1

k |TB |∑

t∈TB

k∑l=1

d(RA, t, l),



Performance of XTalk

0.0 0.2 0.4 0.6 0.8 1.0FPR

0.0

0.2

0.4

0.6

0.8

1.0TPR

XTalk BPLN NIC

AUC:0.65

AUC:0.58

AUC:0.50



XTalk network from Hippo to TGF-β Pathway

SMAD6,SMAD7

SP1

TGFBR2,TGFBR1

RHOA

SMAD2,SMAD3

SMAD3

TGFBR2

EP300,CREBBPSMAD1

RAF1

SMURF1,SMURF2

TGFBR1

BMPR1A,BMPR1B,BMPR2

RBL1

TFDP2,TFDP1

E2F5,E2F4

SMAD4

WWTR1,YAP1

YAP1

SMAD5,SMAD1,SMAD9

BMPR1A,BMPR1B,ACVR1

SMAD7

TFDP1

PPP2CA,PPP2R1A,PPP2R1B,PPP2CB

MAP2K2,MAP2K1

MAPK3,MAPK1

MAP2K1

MAP2K2

ESR2,ESR1

JUN,SP1,FOS

LEF1,TCF7,TCF7L2,TCF7L1

MYC

GSK3B

DVL2,DVL3,DVL1

FZD1,FZD2,FZD3,FZD4,FZD5FZD6,FZD7,FZD8,FZD9,FZD10

+p

+p

http://graphspace.org/graphs/[email protected]/kegg-hippo-tgfbetaCrosstalk Network on GraphSpaceT. M. Murali January 26, 28, February 2, 2016 CS 5854: Projects


XTalk 2.0

I Simple: improve the AUC of XTalk!

I Just like PathLinker, XTalk does not learn anything from thestructure of the pathways.

I Can we use machine learning techniques to improve XTalk?

I XTalk can use edges from any pathway in its crosstalk networks.Can we constrain the types of labels that can appear in a path?

I Papers:I XTalk: a path-based approach for identifying crosstalk between

signaling pathways, Tegge, Sharp, and Murali, Bioinformatics, 32,242–251

I Engineering Label-Constrained Shortest-Path Algorithms, Barrett et al.,Algorithmic Aspects in Information and Management, LNCS 5034,27–37, 2008





http://i11www.iti.uni-karlsruhe.de/extra/publications/bbhkmw-elcsp-08.pdf




Overview

Focus of the Course

Projects






Develop XTalk 2.0



Support



XTalk Gold Standard

I Selected 17 signaling pathways in KEGG.I For each pair of pathways,

I Performed PubMed query andI Read papers for evidence of crosstalkI Recorded sentence providing evidence, if any.

I Goal: use natural language processing techniques to train system thatreads articles or their abstracts to automatically determine whether apair of pathways crosstalk.

I Requires some experience with NLP tools.

Crosstalk database


http://crosstalk.herokuapp.com/


Overview

Focus of the Course

Projects






Develop XTalk 2.0



Support



Destruction Complex



Feedback Loops

I These computed networks contain no loops or feedback cycles!

I Feedback cycles are fundamental components of regulatory circuits.

I How difficult is to compute all the cycles of a length three in a graph?Length four, five, six? The large number of possible cycles may makecost prohibitive.

I Computing all cycles also may not be very interesting.

I Goal: Given a pair of query nodes s and t, compute the k shortestcycles containing s and t.

I Develop algorithm.

I Apply it to yeast or human regulatory and signaling networks.

I Develop methods to evaluate results.

I Use literature to provide additional support.



Feedback Loops



I How difficult is to compute all the cycles of a length three in a graph?

Length four, five, six? The large number of possible cycles may makecost prohibitive.









Feedback Loops



I How difficult is to compute all the cycles of a length three in a graph?Length four, five, six?

The large number of possible cycles may makecost prohibitive.









Feedback Loops



I How difficult is to compute all the cycles of a length three in a graph?Length four, five, six? The large number of possible cycles may makecost prohibitive.









Overview

Focus of the Course

Projects






Develop XTalk 2.0



Support



Hardware Support for Projects

I 40-processor, 20-node cluster in the Department of Computer Sciencededicated to bioinformatics (baobab.cs.vt.edu).

I Obtain accounts on bioinformatics.cs.vt.edu from Rob Hunter(rhunter at vt dot edu).



Software Support for Projects

I My software page:http://bioinformatics.cs.vt.edu/~murali/software

I My Github page: http://github.com/Murali-group

I Biorithm software suite: http:

//bioinformatics.cs.vt.edu/~murali/software/biorithm-docs

I PathLinker, PCSF, ResponseNet, etc.: Code or pipelines implemented inPython and other languages.


http://bioinformatics.cs.vt.edu/~murali/software

http://github.com/Murali-group

http://bioinformatics.cs.vt.edu/~murali/software/biorithm-docs

http://bioinformatics.cs.vt.edu/~murali/software/biorithm-docs


Ground Rules for Projects

I 1 hour meetings with each group every 2 weeks or 4 weeks.

I Maintain Google docs describing your project and your progress.

I Mid-term project review presentations on Tuesday, March 15 in class.

I Project descriptions (motivation, background, related and previousresearch, approach, data, any preliminary results) due on Tuesday,March 15.

I Final project presentations possibly on Tuesday, May 3 andWednesday, May 4. Also possible to set aside a special day just forfinal presentations.

I Final project reports due on 5pm, Friday, May 6.



List of Projects

1. Develop PathLinker 2.0

2. Predict signal transduction pathways from gene expression data.

3. Compute chemical response networks

4. Analyze PanCancer data

5. Predict orientation of interactions in interaction networks.

6. Develop XTalk 2.0, discover links between insulin signaling andinflammation in diabetes

7. Use NLP to build gold standard for XTalk

8. Compute feedback loops in yeast regulatory networks


cs 5854: projects - undergraduate...

Documents