Apoptosis centric Bayesian network perturbation analysis of signaling pathways in colorectal cancer for synergistic drug targets discovery

In this work we first investigate clustering techniques to identify different molecular subtypes in colorectal cancer based on RPPA protein expression data for 461 tumor samples across 60 proteins. Principal component analysis (PCA) is used to examine variances and determine small number of clusters without losing accuracy. Heatmap and hierarchical clustering with various distance measures are used to visually validate choice of number of clusters. Partitioning around Medoids (PAM) clustering algorithm is used to partition the tumor protein expression data into three clusters. The clusters are analyzed for their protein expressions data to gain insight into their molecular subtypes. Our analysis shows that colorectal tumors can be classified primarily along the degree of Epithelial-mesenchymal transition in cancer progression.

After partitioning tumor samples into three clusters, we use Bayesian network (BN) learned from RPPA data to make class specific inference regarding network structure and pathway perturbations. We use bootstrap sampling technique to learn 500 class specific causal Bayesian networks from protein phosphorylation data and use model averaging to build representative network with high statistical significance. The network structure from each bootstrap sample was learned with a hill-climbing (hc) search and a BDe (Bayesian Dirichlet equivalent) posterior density network score. The averaged Bayesian network was created using a statistically motivated algorithm described in [2] and implemented as the default behavior of averaged.network() function available within R package bnlearn [3]. Bayesian estimator was used to learn parameters of local distribution. Bayesian estimates of network parameters are smoother than the maximum likelihood estimates, making inference more robust and easier. Authors in [1] showed that Bayesian estimates of network parameters produce networks that are close to the true networks.

We used the learned Bayesian networks to discover influential proteins whose synergistic perturbation causes favorable apoptosis conditions. In particular we use conditional probability query on various protein perturbations to elicit high apoptotic signals Caspase-3, Caspase-9 and Caspase-8. We believe that cajoling cancer cells to sensitize at least one of many parallel apoptosis pathways represents a more robust approach to arresting cancer than inhibiting multitude of growth signals whose complex regulation and crosstalk in still unknown. Approximate inference algorithm based on Markov Chain Monte Carlo (MCMC) scheme was employed to determine probability of high or low expression of proteins.

Our research shows that JAK2/STAT5 and TGF-β/SMAD4 pathways are frequently deregulated in colon tumors that have not yet undergone Epithelial-mesenchymal transition (EMT) and hence are good targets for drugs. Tumors with high level of N-Cadherin (CDH2) expression imply that they have undergone EMT transition and are difficult to tame despite having favorable level of apoptotic proteins (Bax, Bak, Bid, Bcl-xL) and other onco-proteins (mTOR, β-Catenin, STAT5, BRAF).