Discovering novel cancer bio-markers in acquired lapatinib resistance using Bayesian methods

TL;DR

This paper introduces a Bayesian methodology to identify novel cancer biomarkers associated with acquired resistance to lapatinib in breast cancer, focusing on signaling pathway disruptions and causal network modeling.

Contribution

It develops a Bayesian framework, PathTurbEr, for discovering driver biomarkers in signaling pathways using high-throughput gene expression data and causal Bayesian networks.

Findings

Identified 22 dysregulated signaling pathways in lapatinib resistance.

Developed a robust Bayesian network modeling approach for pathway analysis.

Revealed key pathways like PI3K-AKT and TGF-beta involved in resistance.

Abstract

Genes/Proteins do not work alone within our body, rather as a group they perform certain activities indicated as pathways. Signalling transduction pathways (STPs) are some of the important pathways that transmit biological signals from protein-to-protein controlling several cellular activities. However, many diseases such as cancer target some of these signalling pathways for their growth and malignance, but demystifying their underlying mechanisms are a very complicated tasks. In this study, we use a fully Bayesian approach to develop methodologies in discovering novel driver bio-markers in aberrant STPs given two-conditional high-throughput gene expression data. This project, namely PathTurbEr (Pathway Perturbation Driver), is applied on a global gene expression dataset derived from the lapatinib (an EGFR/HER dual inhibitor) sensitive and resistant samples from breast cancer cell lines (SKBR3). Differential expression analysis revealed 512 differentially expressed genes (DEGs) and their signalling pathway enrichment analysis revealed 22 singalling pathways as aberrated including PI3K-AKT, Hippo, Chemokine, and TGF-beta singalling pathway as highly dysregulated in lapatinib resistance. Next, we model the aberrant activities in TGF-beta STP as a causal Bayesian network (BN) from given observational datasets using three Markov Chain Monte Carlo (MCMC) sampling methods, i.e. Neighbourhood sampler (NS) and Hit-and-Run (HAR) sampler, which has already proven to have more robust inference with lower chances of getting stuck at local optima and faster convergence compared to other state-of-art methods. Next, we examined the structural features of the optimal BN as a statistical process that generates the global structure using, $p_1$-model, a special class of Exponential Random Graph Models (ERGMs) and MCMC methods for their hyper-parameter sampling....