Robust and scalable inference of cancer progression pathways using Conjunctive Bayesian Networks

Abstract

Cancer is an evolutionary disorder driven by stepwise accumulation of selectively advantageous mutations forming mutational pathways, characterization of which is essential for diagnosis, prognosis and treatment of cancer. Conjunctive Bayesian networks (CBN) are probabilistic graphical models that have enabled the inference of these pathways of cancer progression from genomic data. Previously, we showed that the CBN model can be used to estimate the predictability of cancer evolution as it is able to reflect the underlying cancer fitness landscapes directly from genotypic data. However, the reliability of the inferred pathway probability distributions has not yet been ascertained, which motivates the need for a robust inferential framework. To fill this gap, in this study I have introduced the robust-CBN model (R-CBN). By analyzing synthetic, simulated and real data, I have rigorously compared R-CBN with previous CBN models including CT-CBN, H-CBN, and B-CBN, and the results indicate a superior robustness of the R-CBN model in various settings. Furthermore, I have devised a dynamic programming approximation algorithm, which renders the model amenable to scalability. Thus, R-CBN has the potential to be broadly utilized as a reliable framework to infer cancer-driving evolutionary trajectories, and to distill mechanistic insights from cross-sectional cancer genomic data.