Methods blog · Bayesian & Network Inference

p₁-model cross-talk

Predicting the signalling cross-talks behind lapatinib resistance with a Bayesian p₁-model

A fully Bayesian network model compares resistant and parental breast cancer cells to find gene pairs that newly interact across pathways, linking EGFR/ErbB to Notch, Wnt, GPCR, Hedgehog, IGF1R and TGF-β signalling.

#Bayesian modelling#ERGM#p1-model#acquired drug resistance#breast cancer#lapatinib#pathway cross-talk

Targeted cancer drugs often work brilliantly at first, then stop working. In HER2/ErbB2-positive breast cancer, the EGFR/HER2 inhibitor lapatinib shrinks tumours, but cells acquire resistance. Mutations in the drug target explain only part of this. The rest comes from signalling cross-talk: other pathways rewire to compensate for the inhibited one. This paper, featured as an Editors’ pick in BMC Systems Biology, asked which cross-talks appear when cells become resistant.

The modelling idea

We treat each condition, parental (sensitive) SKBR3 cells and lapatinib-resistant SKBR3-R cells, as an undirected gene network inferred from expression data, and ask for every gene pair \((i,j)\): what is the posterior probability that they interact in this condition?

To answer that in a fully Bayesian way we used the p₁-model, an exponential random graph model (ERGM) originally proposed by Holland and Leinhardt. In its undirected form, the probability of an edge between nodes \(i\) and \(j\) depends on a global density parameter and node-specific “expansiveness” terms:

\[\Pr(x_{ij}=1 \mid \theta, \alpha) \;=\; \frac{\exp(\theta + \alpha_i + \alpha_j)}{1 + \exp(\theta + \alpha_i + \alpha_j)},\]

so hub-like genes (large \(\alpha_i\)) are allowed to be densely connected without inflating the evidence for any single edge. Placing priors on \(\theta\) and the \(\alpha\)s and sampling the joint posterior with MCMC yields, for every pair, a posterior probability of interaction in each condition, rather than a single point estimate.

Defining a drug-resistant cross-talk

A gene pair is a candidate resistance cross-talk if:

  1. one gene belongs to the EGFR/ErbB signalling pathway (the drug’s target) and the other to a different signalling pathway (from Reactome, KEGG or WikiPathways); and
  2. the pair is inferred to interact in resistant cells but not in parental cells.

What we found

104 / 188 / 299candidate cross-talks (Reactome / KEGG / WikiPathways)
168distinct pairs with both genes up-regulated in resistance
6partner pathways recurring with EGFR/ErbB

The 168 pairs in which both partners were up-regulated in resistant cells are the prime candidates. They connect EGFR/ErbB signalling with six pathways: Notch, Wnt, GPCR, Hedgehog, insulin receptor/IGF1R and TGF-β receptor signalling. A literature survey supported a role in drug resistance for many of them.

Crucially, the same partner pathways reappeared when we analysed an independent lapatinib-resistance study in BT474 cells. That replication made us more confident that the signal is biological rather than an artefact of one dataset.

The interpretation

Compensatory pathways can up-regulate EGFR/ErbB pathway genes through cross-talk, counteracting the inhibitor. The up-regulated members of those compensatory pathways are therefore candidate co-targets: inhibiting them alongside lapatinib might delay or prevent resistance.

This paper was the first in a line of work. Signal rewiring and V-structures extended the model, the cross-talk categorisation formalised what counts as a cross-talk, and PathTurbEr moved to causal Bayesian networks.

Read the paper

  1. Prediction of signaling cross-talks contributing to acquired drug resistance in breast cancer cells by Bayesian statistical modeling

    AKM Azad, Alfons Lawen, Jonathan Macgregor Keith

    BMC Systems Biology 9 (1), 2 · 2015

    First author OA Editor's pick, BMC Systems Biology 25

BibTeX