DECIPHERING CANCER MECHANISMS BY INTEGRATIVE NETWORK ANALYSIS Research Seminar Duke-NUS Medical School June 2014 Sriganesh Srihari Institute for Molecular Bioscience, The University of Queensland, QLD, Australia
Cancer: A large class of diseases affecting different organs of the body Even if it affects the same organ site E.g. Breast cancer – five • “intrinsic” subtypes (identified from gene-expression; Perou et al . 2000) • At least ten subtypes from genomic and expression data (Curtis et al ., 2012) Sources: National Cancer Institute USA; Science Sriganesh Srihari, Institute for Molecular Bioscience, The University of Queensland 2
Cancer: The origin is in the genome… Hampton et al., Genome Research 2009 Sriganesh Srihari, Institute for Molecular Bioscience, The University of Queensland 5
Leading up to pathways and processes Sources: Nature Reviews, Science Sriganesh Srihari, Institute for Molecular Bioscience, The University of Queensland 6
Understanding dysregulation in pathways: Usually studied as individual genes #Up-regulated = 8 #Down-regulated = 8 So, is this pathway up- or down-regulated? • Mean / maximum / voting of genes? “Top” portion is down, but “bottom” portion is up-regulated. • Where do you draw the boundaries? By studying genes individually, we are missing their aggregate effect. • Not taking into account the structure or topology of the pathway. Sriganesh Srihari, Institute for Molecular Bioscience, The University of Queensland 7
Understanding dysregulation in pathways: Studying complexes Proteins seldom perform their functions in isolation, but instead form stable functional complexes. By looking at complexes, • Aggregate the effect of individual genes; • Factor in topological structure. We obtain an aggregate or “systems biology” view of underlying mechanisms. Sriganesh Srihari, Institute for Molecular Bioscience, The University of Queensland 8
Complexes in Pathways and Processes Affected in Cancer Sources: Nature Reviews, Science 9
Complexes in Diseases Sriganesh Srihari, Institute for Molecular Bioscience, The University of Queensland 10
ANALYSING COMPLEXES IN CANCER Sriganesh Srihari, Institute for Molecular Bioscience, The University of Queensland 11
Human complex databases: CORUM (Mewes et al. , NAR 2004) Havugimana et al. ( Cell , 2012) Coverage ~30 - 40% Sriganesh Srihari, Institute for Molecular Bioscience, The University of Queensland 12
Complex Identification from Protein Interactions A typical pipeline (Spirin & Merny, PNAS 2003) 3. Validation Validate against bona fide complexes. Study roles of novel complexes. 1. PPI network Low- and high- throughput experiments. Assembled as a network after filtering noise. 2. Complexes Human complex databases: Clustering the PPI network CORUM (Mewes et al., NAR 2004) to predict complexes. Havugimana et al. ( Cell , 2012) Coverage ~30 - 40% Sriganesh Srihari, Institute for Molecular Bioscience, The University of Queensland 13
Complex Prediction from Protein Interaction Networks Computational methods in the literature (Srihari & Leong, 2013) There are several methods for identifying complexes from PPI networks. Experiments on yeast suggest ~75% coverage. Srihari S and Leong HW, J Bioinf Comp Biol 2013. 14
Can we use this pipeline ? “Cancer PPI network” Complexes in cancer Validate against known, study roles of novel complexes in cancer Sriganesh Srihari, Institute for Molecular Bioscience, The University of Queensland 15
Can we use this pipeline ? ‘Cancer PPI network’ Complexes in cancer Validate against known, study roles of novel complexes in cancer Sure! But, • how do we gather such a “cancer PPI network” ? • PPI networks do not have what constitutes the network? • contextual information which complexes are affected in cancer? • not all complexes are involved in cancer. • Can we predict ‘cancer PPI network’? Sriganesh Srihari, Institute for Molecular Bioscience, The University of Queensland 16
Identify complexes that are “dysfunctional” in cancer Track complexes for behavioural differences across conditions by integrating diverse information , The key! Mutated genes coding for dysfunctional proteins within complexes Changes in expression of coding genes Mutations in genes or Changes in protein composition or abundance chromosomal locii mRNA expression Possible to predict Normal tumour Sriganesh Srihari, Institute for Molecular Bioscience, The University of Queensland 17
(Why) Does this help? A case study mapping gene expression from Normal and Tumour pancreatic conditions onto PPIs and complexes Integrate the following two kinds of data: PPI network (BioGrid v3.1.93; Stark et al . 2011) 5824 proteins 29600 interactions [high-quality post filtering] Gene expression 39 matched normal-tumour samples from pancreatic adenocarcinoma patients (Badea et al. 2008) Extract complexes from the PPI network using clique-merging (Srihari et al. , 2013) Mapping gene co-expression [-1,+1] Sriganesh Srihari, Institute for Molecular Bioscience, The University of Queensland 18
(Why) Does this help? A case study mapping gene expression from Normal and tumour pancreatic conditions onto PPIs and complexes Significant loss in correlations of PPIs – Significant loss in correlations for “accelerators” as well as “brakes”. complexes. (KS test: 23.11 > K 0.05 = 1.36) (KS test: 1.69 > K 0.05 = 1.36) The aggregate effect of using protein pairs or complexes! 19
Affected Protein Interactions A case s tudy “mapping” gene expression from Normal and tumour pancreatic conditions onto PPIs and complexes RHOXF2(PEPP2) Involved in carcinogenesis in gastric and pancreatic cell lines. (Shibata-Minoshima et al., 2012) NFKB (Exp mean 4.57, 4.63) KRAS SMN1--TMSB4X TGFB1 RBPMS-- RHOXF2 “Jumps” in correlation of PPIs from normal to tumour Sriganesh Srihari, Institute for Molecular Bioscience, The University of Queensland 20
CONTOUR: An Enhanced Pipeline to Detect Dysfunctional Complexes in Cancer Integrating PPI, Gene Expression and Mutation datasets Normal Normal 1 1 2 2 + s t Identify and match Generic PPI network complexes between conditions tumour tumour Gene expression, mutation profiles Conditional PPI networks Dysfunctional complexes Srihari S & Ragan MA, Bioinformatics 2013 Sriganesh Srihari, Institute for Molecular Bioscience, The University of Queensland 21
APPLICATION TO CANCERS Sriganesh Srihari, Institute for Molecular Bioscience, The University of Queensland 22
Cancer Conditions Studied Normal vs Pancreatic adenocarcinoma (PDAC) PDAC – constitutes 95% of pancreatic tumours 39 pairs of matched gene expression samples from Badea et al . (2008) from normal and PDAC tissues Mutation profiles of 1169 genes from Jones et al. ( Science , 2008) BRCA1 tumours ~ basal-like BRCA1 tumours vs BRCA2 tumours • Aggressive Germline defects in BRCA1 and BRCA2 genes Mostly triple-negative (ER/PR/HER2-) • Deficient in homologous recombination-based DSB repair Profiling of familial breast tumours (kConfab consortium) • Expression data from Waddell et al. (2010) BRCA2 tumours ~ luminal or more heterogeneous (luminal and HER2+) Less aggressive (at least luminal-A) • Badea et al. Hepatogastroenterology 2008 Jones et al. Science 2008 Waddell et al. Breast Cancer Treat Res 2010 Sriganesh Srihari, Institute for Molecular Bioscience, The University of Queensland 23
Differences between conditional PPI networks (KS test: 23.11 > K 0.05 = 1.36) (KS test: 22.85 > K 0.05 = 1.36) Top enriched GO terms for interactions showing ≥ |1.0| change : Cell cycle, chromatin organization, DNA repair and RNA splicing. Pancreatic: KRAS, TGF , RAD21, STAT1, STAT3, P53, SMAD4. Breast: BRCA1, BRCA2, TP53, BRE, BRCC3. 24
Differential PPI Network: Normal vs PDAC High cut-off: |1.0| Proteins: 558 Interactions: 519 (very sparse) Interactions PDAC vis-à-vis Normal Red: Weakened Green: Strengthened EP300 PLK1 ANXA2 PELP1 Top enriched GO terms for interactions showing ≥ |1.0| change : Cell cycle, chromatin organization, DNA repair and RNA splicing. 25
Differential PPI Network: Normal vs PDAC 26
Changes in correlation of complexes Normal vs PDAC and BRCA1 vs BRCA2 tumours CORUM (KS test: 1.69 > K 0.05 = 1.36) (KS test: 5.48 > K 0.05 = 1.36) Sriganesh Srihari, Institute for Molecular Bioscience, The University of Queensland 27
Changes in correlation of complexes: Normal vs PDAC and BRCA1 vs BRCA2 Correlation of complexes Correlation of complexes Condition #Complexes Max Avg Condition #Complexes Max Avg Category Normal 1.206 0.292 BRCA1 0.863 0.218 Our 256 277 PDAC 0.757 0.154 BRCA2 0.479 0.027 Correlation of complexes Correlation of complexes Condition Max Avg Condition #Complexes Max Avg Category #Complexes Normal 1.037 0.216 BRCA1 0.702 0.188 CORUM 189 441 PDAC 0.448 0.113 BRCA2 0.512 0.059 Overall loss in correlation of complexes in PDAC vis-à-vis Normal and BRCA2 vis-à-vis BRCA1. 28
Recommend
More recommend
