Methods blog · Bayesian & Network Inference

PathTurbEr

PathTurbEr: finding the driver genes of perturbed signalling pathways with causal Bayesian networks

Causal Bayesian networks learned with Neighbourhood, Hit-and-Run and Metropolis–Hastings samplers, then analysed with the p₁-model, pinpoint perturbation drivers of TGF-β signalling in lapatinib-resistant breast cancer.

#Bayesian networks#MCMC#ERGM#p1-model#biomarker discovery#TGF-beta#lapatinib resistance

Cancers hijack signal transduction pathways (STPs) to grow and resist therapy, but finding which genes drive the aberration is hard. Differential expression tells us what changed, not what is causing the change. PathTurbEr (Pathway Perturbation Driver), published in Briefings in Bioinformatics with Salem Alyami, combines causal network learning and random-graph modelling to find drivers inside a perturbed pathway.

Pipeline at a glance

Step 1: find the perturbed pathways

Using expression data from lapatinib-sensitive and lapatinib-resistant SKBR3 breast cancer cells (lapatinib is a dual EGFR/HER2 inhibitor), differential expression analysis found 512 differentially expressed genes. Pathway enrichment revealed 13 highly perturbed signalling pathways in resistance, including PI3K–AKT, chemokine, Hippo and TGF-β signalling.

Step 2: learn the causal structure of the perturbed pathway

The aberrant TGF-β pathway was modelled as a causal Bayesian network, with posterior

\[\Pr(G \mid D) \;\propto\; \Pr(D \mid G)\,\Pr(G),\]

sampled with three MCMC samplers: the Neighbourhood sampler (NS) and Hit-and-Run (HAR) sampler, which yield robust inference with less risk of local optima and faster convergence, and Metropolis–Hastings (MH) as a baseline. (See our posts on the Neighbourhood sampler and BNMCMC.)

Step 3: ask which nodes shape the network

An optimal network is still just a graph; drivers are the nodes that shape its global structure. We treated the learned BN as the output of a statistical generative process, the p₁-model, a special class of exponential random graph model (ERGM) with node-level expansiveness and attractiveness parameters, and sampled its hyper-parameters by MCMC. Genes whose parameters stand out are the ones that drive the aberration within the perturbed pathway.

Results

80genes in the TGF-β pathway model
34 / 34 / 23driver genes (NS / HAR / MH networks)
13perturbed signalling pathways

Functional-relevance and disease-relevance analyses showed that the drivers are significantly associated with breast cancer progression and resistance. That makes them candidate biomarkers and co-targets to pair with EGFR/HER2 inhibition.

Code

PathTurbEr is open source: github.com/Akmazad/PathTurbEr.

Read the paper

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

    AKM Azad, Salem A. Alyami

    Briefings in Bioinformatics 22 (5), bbab137 · 2021

    Q1First author OA 5

BibTeX