1. Canadian Bioinformatics Workshops

2. Module #: Title of Module2

3. Pathway and Network Analysis
Module 7 – Part III
Pathway and Network Analysis
Lincoln Stein
Bioinformatics for Cancer Genomics
May 27-31, 2013

4. Classes of Gene Set Analysis
DAVID
GSEA
Reactome FI network
PARADIGM
Khatri et al. PLOS Comp Bio. 8:1 2012

5. Limitations of Gene Set Enrichment Analysis
Many possible gene sets – diseases, molecular function, biological process, cellular compartment, pathways...
Gene sets are heavily overlapping; need to sort through lists of enriched gene sets!
“Bags of genes” obscure regulatory relationships among them.

6. Pathway Databases Advantages: Disadvantages: Usually curated.
Biochemical view of biological processes.
Cause and effect captured.
Human-interpretable visualizations.
Disadvantages:
Sparse coverage of genome.
Different databases disagree on boundaries of pathways.

7. KEGG

8. Reactome

9. Reactome Hand-curated pathways in human.
Rigorous curation standards – every reaction traceable to primary literature.
Automatically-projected pathways to non-human species.
22 species; 1112 human pathways; 5078 proteins.
Features:
Google-map style reaction diagrams with overlays;
Find pathways containing your gene list;
Calculate gene overrepresentation in pathways;
Find corresponding pathways in other species.
Open access.

10. Pathway Commons

11. Pathway Colorization Main feature offered by all pathway databases.
Upload a gene list
Database calculates an enrichment score on each pathway and displays ranked list.
Browse into pathways of interest; download colorized pictures.

12. Example from Reactome

13. Example from Reactome

15. Networks
Pathways capture only the “well understood” portion of biology.
Networks cover less well understood relationships:
Genetic interactions
Physical interaction
Coexpression
GO term sharing
Adjacency in pathways

21. Biological Networks are Scale Free
Properties:
The degree (# connections) of nodes follows a power law. A node of degree k+1 is exponentially less likely to occur than a node of degree k.
The local clustering coefficient (tendency of nodes to interconnect) is independent of the degree of the node.
Nature Reviews Genetics 5, (February 2004) | doi: /nrg1272

22. Biological Networks are Scale Free
Implications:
A small number of genes have a large number of connections (chokepoints).
A large number of genes have a small number of connections (leaves).
Genes cluster (functional groups).
The cluster sizes are also scale-free (many small clusters, few large clusters).
Nature Reviews Genetics 5, (February 2004) | doi: /nrg1272

23. Network Databases Can be built automatically or via curation.
Popular sources of curated networks:
BioGRID – Curated interactions from literature; 529,000 genes, 167,000 interactions.
InTact – Curated interactions from literature; 60,000 genes, 203,000 interactions.
MINT – Curated interactions from literature; 31,000 genes, 83,000 interactions.

24. Uncurated Interaction Sources
Text mining approaches
Computationally extract gene relationships from text, such as PubMed abstracts.
Much faster than hand curation.
Not perfect:
Problems recognizing gene names. Is hedgehog a gene or a species?
Natural language processing is difficult.
Popular resources:
iHOP
PubGene

25. Uncurated Interaction Sources
Experimental techniques
Yeast 2 hybrid protein interactions.
Protein complex pulldowns/mass spec.
Genetic screens, such as synthetic lethals, enhancer/suppressor screens.
NOT perfect
Y2H interactions have taken proteins out of natural context; physical interaction != biological interaction.
Protein complex pulldowns plagued by “sticky” proteins such as actin.
Genetic screens highly sensitive to genetic background (“network effects”).

26. Integrative Approaches
Combine multiple sources of evidence to increase accuracy.
Simple example:
“Party hubs” are Y2H interactions that have been filtered for those partners that share the same temporal-spatial location.
Complex example:
Combine multiple sources of curated and uncurated evidence.

27. Example: Reactome FI Network
Curated Human Data – Version 35.
5078 proteins reactions
3870 complexes pathways
Only ~25% of genome!
Goal: add a “corona” of uncurated interaction data around scaffold of curated pathway data.

28. Expanding Reactome’s Coverage
Curated Pathways
Uncurated Information
human PPI
PPI inferred from fly,
worm & yeast
PPI from text mining
Gene co-expression
GO annotation on
biological processes
Protein domain domain interactions
GeneWays
CellMap
TRED
Naïve Bayes Classifier
Annotated Functional Interactions
Predicted Functional Interactions
Wu et al. (2010) Genome Biology

29. Integrated Functional Interaction (FI) Network
2929
Integrated Functional Interaction (FI) Network
10,956 proteins (9,542 genes).
209,988 FIs.
~50% coverage of genome.
False (+) rate < 1%
False (-) rate ~80%
5% of network shown here

30. Active Network Extraction+
Machine Learning
Uncurated
Interaction
Evidence
Curated
Pathway Dbs
Reactome Functional Interaction Network
(~11,000 proteins; 200,000 interactions)
Extract and Cluster Altered Genes
Disease “modules” (10-30)

31. Clustering of TCGA Breast Cancer Mutations
Cadherin signaling
Signaling by Tyrosine
Kinase receptors
NOTCH and wnt signaling
Focal adhesion
ECM-Receptor interaction
Neuroactive ligand-receptor
interaction
Mucin cluster
Cell adhesion
molecules
Ubiquitin-mediated
proteolysis
Metabolism of proteins
Signaling by Rho GTPases
DNA repair
Cell cycle
Axon guidance
M phase
G2/M Transition
Calcium signaling

32. 256 Pancreatic Cancer Mutations
Patient Samples
Genes

33. Pancreatic Mutation Modules
Module 0: MAPK, Hedgehog, TGFβ signaling
Module 4: ECM, focal adhesion, integrin signaling
Module 5: Wnt & Cadherin singaling
Module 3: Translation
Module 2: B-cell receptor, ERBB, FGFR, EGFR signaling
Module 9: Axon guidance
Module 10: muscle contraction
Module 1: Heterotrimeric G-protein signaling
Module 7: Axon guidance
Module 6: Ca2+ signaling
Module 8: MHC class II antigen presentation

34. Modules After Hierarchical Clustering
Patient Samples
Modules

35. Network-Based Clustering Algorithms
Reactome FI network (Wu & Stein, Genome Biol (12):R112)
Expression or SNV analysis
Online analysis via Cytoscape Plugin (lab)
HotNet (Vandin et al. J Comput Biol Mar;18(3):507-22).
Local installation with Python & MatLab
Cytoscape visualization
WGCNA (Langfelder et al BMC Bioinformatics 9: 559.)
Expression analysis
Local installation as R package.

36. Classification of Tumors via Molecular Phenotype
Test
Classify
Through the research and discoveries we are making eventually we will start to deliver tests that will allow physicians to selectively treat cancer patients with the best drugs.
In this room we are using a very sophisticated machine machine to look at differences in genes so that we can identify those genes that correlate to response
We are able to examine about 8000 genotypes/day or roughly 2million/year
Our Company has also taken a lead position in the emerging field of pharmacogenomics and personalized medicines. We believe pharmocogenomics in the clinic offer several advantages to drug development, to improved medicines for patients. Today it is becoming commonplace to examine a single gene on a tumor biopsy to select treatment. We believe that in the future, expanding that approach to scan large sets of genes will improve our ability to find the right drug for the right patient. Through our internal genomics effort, as well as through our alliances with Millennium, the Whitehead Genomics Center, Impath, Exelixis and major clinical centers, we are integrating this approach extensively into our oncology pipeline.
Proteomics
Transcriptomics
Genomics

37. Low risk – reduce treatment High risk – treat aggresively
Risk Stratification
Don’t Treat
TEST
Low risk – reduce treatment
Treat
10-20% progress
Through the research and discoveries we are making eventually we will start to deliver tests that will allow physicians to selectively treat cancer patients with the best drugs.
In this room we are using a very sophisticated machine machine to look at differences in genes so that we can identify those genes that correlate to response
We are able to examine about 8000 genotypes/day or roughly 2million/year
Our Company has also taken a lead position in the emerging field of pharmacogenomics and personalized medicines. We believe pharmocogenomics in the clinic offer several advantages to drug development, to improved medicines for patients. Today it is becoming commonplace to examine a single gene on a tumor biopsy to select treatment. We believe that in the future, expanding that approach to scan large sets of genes will improve our ability to find the right drug for the right patient. Through our internal genomics effort, as well as through our alliances with Millennium, the Whitehead Genomics Center, Impath, Exelixis and major clinical centers, we are integrating this approach extensively into our oncology pipeline.
High risk – treat aggresively
No Relapse
Relapse

38. Challenges in Biomarker Discovery
Overtraining
22,000 genes; any given cancer may show alterations in 1000s of them; patients cohorts are in 100s.
Can find a set of gene alterations that nicely predicts survival in a single cohort by chance.
Field is littered with biomarkers that didn’t replicate in independent cohorts.
Disease Heterogeneity
If there are many subtypes of disease then need even larger cohorts.
Tumor Heterogeneity
A single primary tumor may carry high-risk and low-risk subclones simultaneously.

39. Using Network Architecture to Accelerate Biomarker Selection
Expression Analysis of tumours from multiple patients
Principal component analysis on modules
Disease Module Map
Correlate principal components with clinical parameters
Guanming Wu
Genome Biol Dec 10;13(12):R112

40. Samples Used Built the network using Nejm: van de Vijver et al 2002
295 Samples, ~12,000 genes
Event: death
Validated with GSE4922: Ivshina et al. Cancer Res. 2006
249 Samples, ~13,000 genes
Event: recurrence or death

41. PC Analysis Identifies Module 2 as Explaining Much of Variation in Survival

42. Same Signature Predicts Survival in Independent Data Set

43. And Three More Data Sets as Well…

44. Module 2: Kinetochore + Aurora B Signaling

45. Integration of Multiple Data Sets
Experimental samples can be interrogated many ways:
RNA expression
Genome/exome sequencing
Copy number changes/loss of heterozygosity
shRNA knockdown screens
Integrate multiple functional data types using network/pathway relationships?

46. PARADIGM
Vaske, Benz et al. Bioinformatics 26:i

47. Factor graph: directed graph connecting genes; each gene is activated, inactivated, or unchanged in a single patient.
Vaske, Benz et al. Bioinformatics 26:i

48. Vaske, Benz et al. Bioinformatics 26:i237 2010

49. PARADIGM: The Bad News Distributed in source code form only
Requires several third-party math/graph libraries (all open source).
Tedious to compile!
Scant documentation.
No repositories of formatted pathway data.
No examples of converting experimental data into input files.
Good news: we are working on a web service implementation for a Reactome-based implementation.

50. Take Home Messages
Pathway/network analysis can provide context to altered gene lists.
Pathway/network analysis differs greatly in complexity , power, and usability:
SIMPLE: Pathway diagram colorization
MODERATE: Reactome FI network extraction
COMPLEX: PARADIGM
This type of analysis is work-in-progress, but promises ability to integrate data across many dimensions.

51. URLs KEGG – www.genome.jp/kegg Biocarta – www.biocarta.com
WikiPathways –
Reactome –
NCI/PID – pid.nci.nih.gov
Ingenuity –
Pathway Commons –
PARADIGM --

52. URLs BioGrid – www.thebiogrid.org InTact – www.ebi.ac.uk/intact
MINT – mint.bio.uniroma2.it
iHOP –
PubGene –

53. We are on a Coffee Break & Networking Session