Gene Regulatory Networks II Computa.onal Genomics Seyoung Kim

Transcription

1 Gene Regulatory Networks II Computa.onal Genomics Seyoung Kim

2 Goal: Discover Structure and Func;on of Complex systems in the Cell Identify the different regulators and their target genes that are involved in the system. Represent the relationship between regulators and their target genes as a network Nodes: entities (regulators, target genes) Edges: regulatory relationship High- level goal: Use high throughput data to discover paaerns of combinatorial regula.on and to understand how the ac.vity of genes involved in related biological processes is coordinated and interconnected.

3 Overview Bayesian networks (network with directed edges): Module networks and their extensions Module network (Segal et al., Nature Gene.cs 2003): Gene module s ac.vity is determined by their expression levels of regulator genes Geronemo (Lee et al., PNAS 2006): Gene module s ac.vity is determined by their expression levels of regulator gene and SNPs Lirnet (Lee et al., PLoS Gene.cs 2009): incorporates prior knowledge CONEXIC (Akavia et al., Cell 2010): cancer data analysis for copy number varia.on and gene expression data Gaussian graphical models (network with undirected edges) and their extensions

4 Probabilis;c Graphical Models p(x 3 X 2 ) X 3 p(x 2 X 1 ) X 2 p(x 1 ) X 1 X 6 p(x 4 X 1 ) p(x 5 X 4 ) X 4 X 5 p(x 6 X 2, X 5 ) The joint distribu.on on (X 1, X 2,, X N ) factors according to the parent- of rela.ons defined by the edges E : p(x 1, X 2, X 3, X 4, X 5, X 6 ) = p(x 1 ) p(x 2 X 1 )p(x 3 X 2 ) p(x 4 X 1 )p(x 5 X 4 )p(x 6 X 2, X 5 )

5 Learning Bayesian Networks Density es.ma.on Model data distribu.on in popula.on Learn both the graph structure and the associated probability density Probabilis.c inference: Predic.on Classifica.on Data p(x 3 X 2 ) X 3 p(x 2 X 1 ) X 2 p(x 1 ) X 1 X 6 p(x 4 X 1 ) p(x 5 X 4 ) p(x 6 X 2, X 5 ) X 4 X 5 p(x 1, X 2, X 3, X 4, X 5, X 6 ) = p(x 1 ) p(x 2 X 1 )p(x 3 X 2 ) p(x 4 X 1 )p(x 5 X 4 )p(x 6 X 2, X 5 )

6 The Module Network Idea Unlike methods that represent individual genes, module- based methods explicitly model modular structure. This helps in: Reducing dependence on possibly noisy measurements for individual genes, by combining informa.on among genes in the same module Elevated sta.s.cal significance

7 The Module Network Idea Bayesian Network CPD 1 Module Network Share parameters and dependencies between variables with similar behavior CPD 1 MSFT Module I MSFT CPD 2 CPD 3 MOT CPD 4 MOT CPD 2 DELL INTL DELL INTL Module II CPD 5 AMAT HPQ CPD 6 AMAT HPQ CPD 3 Module III Slides from the presenta.on by Segal et al. UAI03

8 Learning Module Network Module Network Model defini.on Learning the model Experimental results

9 Module Network Components Module Assignment Func.on A( ) A(MSFT)=M I AMAT DELL MSFT MOT HPQ INTL A(MOT)=A(DELL)=A(INTL) =M II A(AMAT)= A(HPQ)=M III Module I MSFT MOT DELL INTL Module II Module III AMAT HPQ Slides from the presenta.on by Segal et al. UAI03

10 Module Network Components Module Assignment Func.on Set of parents for each module Pa(M I )= Pa(M II )={MSFT} Pa(M III )={DELL, INTL} Module I MSFT MOT DELL INTL Module II Module III AMAT HPQ Slides from the presenta.on by Segal et al. UAI03

11 Module Network Components Module Assignment Func.on Set of parents for each module Condi.onal probability density (CPD) template for each module Module I MSFT MOT DELL INTL Module II Module III AMAT HPQ Slides from the presenta.on by Segal et al. UAI03

12 Ground Bayesian Network A module network induces a ground BN over X A module network defines a coherent probabilty distribu.on over X if the ground BN is acyclic Module I MSFT MSFT DELL MOT INTL Module II DELL MOT INTL AMAT HPQ Ground Bayesian Network Module III AMAT HPQ Slides from the presenta.on by Segal et al. UAI03

13 Module Graph Nodes correspond to modules M i M j if at least one variable in M i is a parent of M j M I M II M III Module graph Module I MSFT Theorem: The ground BN is acyclic if the module graph is acyclic Module II DELL MOT INTL Acyclicity checked efficiently using the module graph Module III AMAT HPQ Slides from the presenta.on by Segal et al. UAI03

14 Learning Module Network Module Network Model defini.on Learning the model Experimental results

15 Learning Overview Given data D, find assignment func.on A and structure S that maximize the Bayesian score Marginal data likelihood Marginal likelihood Assignment / structure prior Data likelihood Parameter prior Slides from the presenta.on by Segal et al. UAI03

16 Bayesian Score Decomposi;on Bayesian score decomposes by modules Module j parents Module j variables

17 Bayesian Score Decomposi;on Bayesian score decomposes by modules Module I MSFT MOT DELL INTL Module II Module III AMAT HPQ Slides from the presenta.on by Segal et al. UAI03

18 Likelihood Func;on θ MI Score M2 (MSFT, X 2 : D) = Score(DELL,MSFT : D) + Score(MOT,MSFT : D) + Score(INTL,MSFT : D) θ MII MSFT Module I MSFT MOT DELL INTL θ MIII DELL,INTL Module II Module score decomposes by variables in the module Module III Instance 1 Instance 2 Instance 3 AMAT HPQ Slides from the presenta.on by Segal et al. UAI03

19 Algorithm Overview Find assignment func.on A and structure S that maximize the Bayesian score Find initial assignment A Assignment function A Improve assignments Improve structure Dependency structure S Slides from the presenta.on by Segal et al. UAI03

20 Learning Dependency Structure Heuris.c search with operators Add/delete parent for module Module I MSFT MOT DELL INTL Module II Module III AMAT HPQ Slides from the presenta.on by Segal et al. UAI03

21 Learning Dependency Structure Heuris.c search with operators Add/delete parent for module Handle acyclicity Can be checked efficiently on the module graph Module I MSFT X MOT M I M II M III DELL INTL X INTL Module I AMAT HPQ Module II Module III Slides from the presenta.on by Segal et al. UAI03

22 Learning Dependency Structure Heuris.c search with operators Add/delete parent for module Handle acyclicity Can be checked efficiently on the module graph M I M II M III Module I DELL MSFT MOT INTL X INTL Module I INTL Module III Module II AMAT HPQ Module III Slides from the presenta.on by Segal et al. UAI03

23 Learning Dependency Structure Heuris.c search with operators Add/delete parent for module Handle acyclicity Can be checked efficiently on the module graph Module I MSFT Efficient computa.on Ajer applying operator for module M j, only update score of operators for module M j DELL Module II MOT INTL AMAT HPQ Module III Slides from the presenta.on by Segal et al. UAI03

24 Algorithm Overview Find assignment func.on A and structure S that maximize the Bayesian score Find initial assignment A Improve assignments Assignment function A Dependency structure S Improve structure Slides from the presenta.on by Segal et al. UAI03

25 Learning Assignment Func;on DELL Module I MSFT MOT INTL Module II Module III AMAT HPQ Slides from the presenta.on by Segal et al. UAI03

26 Learning Assignment Func;on A(DELL)=M I Score: 0.7 DELL Module I MSFT MOT INTL Module II Module III AMAT HPQ Slides from the presenta.on by Segal et al. UAI03

27 Learning Assignment Func;on A(DELL)=M I Score: 0.7 Module I MSFT A(DELL)=M II Score: 0.9 DELL MOT INTL Module II Module III AMAT HPQ Slides from the presenta.on by Segal et al. UAI03

28 Learning Assignment Func;on A(DELL)=M I Score: 0.7 Module I MSFT A(DELL)=M II Score: 0.9 MOT INTL A(DELL)=M III Score: cyclic! Module II DELL AMAT HPQ Module III Slides from the presenta.on by Segal et al. UAI03

29 Learning Module Network Module Network Model defini.on Learning the model Experimental results

30 Learning Module Network from Gene Expression Data

31 Determine Combinatorial Control From Gene Expression data Under Mul;ple Condi;ons

32 Resul;ng Module Row: genes Columns: array (condi.on) Segal et al Nature Gene.cs 2003

33 Experimental Design Hypothesis: Regulator X ac;vates process Y Experiment: Knock out X and repeat experiment false false true HAP4 true Ypl230W X?

34 Biological Experiments Valida;on Ypl230w Were the differen.ally expressed genes predicted as targets? Rank modules by enrichment for diff. expressed genes # Module Significance 39 Protein folding 7/23, 1e Cell differentiation 6/41, 2e Glycolysis and foldin g 5/37, 4e Mitochondrial and protein fate 5/37, 4e - 2 Ppt1 # Module Significance 14 Ribosomal and phosphate metabolism 8/32, 9e 3 11 Amino acid and purine metabolism 11/53, 1e 2 15 mrna, rrna and trna processing 9/43, 2e 2 39 Protein f olding 6/23, 2e 2 30 Cell cycle 7/30, 2e 2 Kin82 All regulators regulate predicted modules # Module Significance 3 Ener gy and osmotic stress I 8/31, 1e 4 2 Energy, osmolarity & camp signaling 9/64, 6e 3 15 mrna, rrna and trna processing 6/43, 2e 2 Segal et al., Nature Genetics, 2003

35 Extensions of Module Network: Geronemo (Lee et al., PNAS 2006) Both gene expression levels and SNPs can be regulators for gene expression levels of other genes Purple regulator: expression regulator Blue regulator: SNP regulator

36 Extensions of Module Network: Lirnet (Lee et al., PLoS Gene;cs 2009) Incorporate prior knowledge (regulatory features) on gene- expression regulators and SNP regulators Regulatory features (on the lej) and learned regulatory priors (purple bar)

37 Extensions of Module Network: CONEXIC (Akavia et al., Cell 2010) Extends module networks to handle cancer copy number varia.on and gene expression data to find cancer- causing (driver) muta.on A driver muta.on should occur in mul.ple tumors more ojen than would be expected by chance

38 Extensions of Module Network: CONEXIC (Akavia et al., Cell 2010) A driver muta.on may be associated (correlated) with the expression of a group of genes that form a module A driver may be over- expressed due to amplifica.on of the DNA encoding it or the ac.on of other factors. The target genes correlate with driver gene expression rather than driver copy number

39 Overview Bayesian networks (network with directed edges): Module networks and their extensions Module network (Segal et al., Nature Gene.cs 2003): Gene module s ac.vity is determined by their expression levels of regulator genes Geronemo (Lee et al., PNAS 2006): Gene module s ac.vity is determined by their expression levels of regulator gene and SNPs Lirnet (Lee et al., PLoS Gene.cs 2009): incorporates prior knowledge CONEXIC (Akavia et al., Cell 2010): cancer data analysis for copy number varia.on and gene expression data Gaussian graphical models (network with undirected edges) and their extensions

40 Gaussian Graphical Models The gene expressions for K genes Y={y 1,, y K } are Gaussian distributed: Y ~ N(0 K,Θ 1 ) 0 K : vector of K zeros Θ: K by K inverse covariance matrix Then, the inverse covariance matrix Θ encodes a Gaussian graphical model Non- zero elements in Θ correspond to edges

41 Gaussian Graphical Models Non- zero elements in Θ correspond to edges Gaussian graphical model encoded by Θ y 1 y 2 y 3 y 4 y y 1 y 3 y 2 y 4 y 5

42 Gaussian Graphical Models Non- zero elements in Θ correspond to edges Gaussian graphical model encoded by Θ y 1 y 2 y 3 y 4 y y 1 y 3 y 2 y 4 y 5 Nonzero/zero paaern of the y 5 s column matches the neighbors of the node y 5

43 Probabilis;c Graphical Models Sta.s.cs and Mechanics are independent of each other condi.onal on Algebra

44 Learning a Sparse Gaussian Graphical Models Minimize nega.ve log likelihood of data with L 1 penalty argmin logdet Θ tr(sθ) λ Θ 1 where - tr(a) is the trace of matrix A - S is a K by K sample covariance - Θ 1 is an L 1 regulariza.on The op.miza.on problem is convex! Many sojware packages exist (e.g., BIG&QUIC, Hsieh et al., NIPS 2013; FastGGM, Wang et al., Plos Comp Bio 2016)

45 Extensions of Gaussian Graphical Models Network with hubs (Tan et al., JMLR 2014) Network with block structures (Tan et al., UAI 2009) Learned from Glioblastoma expression data (Hubs as pink nodes)

46 Summary Modeling gene networks with Bayesian networks Probabilis.c model for learning modules of variables and their structural dependencies Module networks have improved performance over Bayesian networks Sta.s.cal robustness Interpretability Reconstruc.on of many known regulatory modules and predic.on of targets for unknown regulators Modeling gene networks with undirected networks Gaussian graphical models are extremely popular: fast learning methods are available (more efficient than Bayesian network learning)