Cancer driver mutation prediction through Bayesian integration of multi-omic data.

Zixing Wang,Kenna Shaw,Gordon B Mills,Kwok-Shing Ng,Funda Meric-Bernstam,Kenneth L Scott,Fang Wang,Tenghui Chen,Tae-Beom Kim,Ken Chen,Joel S Bader

doi:10.1371/journal.pone.0196939

Abstract

Identification of cancer driver mutations is critical for advancing cancer research and personalized medicine. Due to inter-tumor genetic heterogeneity, many driver mutations occur at low frequencies, which make it challenging to distinguish them from passenger mutations. Here, we show that a novel Bayesian hierarchical modeling approach, named rDriver can achieve enhanced prediction accuracy by identifying mutations that not only have high functional impact scores but also are associated with systemic variation in gene expression levels. In examining 3,080 tumor samples from 8 cancer types in The Cancer Genome Atlas, rDriver predicted 1,389 driver mutations. Compared with existing tools, rDriver identified more low frequency mutations associated with lineage specific functional properties, timing of occurrence and patient survival. Evaluation of rDriver predictions using engineered cell-line models resulted in a positive predictive value of 0.94 in PIK3CA genes. Our study highlights the importance of integrating multi-omic data in predicting cancer driver mutations and provides a statistically rigorous solution for cancer target discovery and development.

Highlights

Large-scale cancer genome projects, such as The Cancer Genome Altas (TCGA), and the International Cancer genome Consortium (ICGC) have systematically catalogued hundreds of thousands of somatic mutations in a wide variety of adult and pediatric cancers
We developed a rDriver approach, which predicts driver mutations based on genome-wide mRNA/protein expression levels, and the functional impact scores (FISs) of individual mutations (Fig 1 and Methods)
A p-value is further computed for Cancer driver mutation prediction through Bayesian integration of multi-omic data each mutation based on empirical distribution of the driver scores (Methods), reflecting the probability that the observed amount of association was not produced by chance

Summary

Introduction

Large-scale cancer genome projects, such as The Cancer Genome Altas (TCGA), and the International Cancer genome Consortium (ICGC) have systematically catalogued hundreds of thousands of somatic mutations in a wide variety of adult and pediatric cancers. The functional significance of the majority of these mutations remains unknown. As implicated in previous studies, only a small fraction of genetic alterations are expected to be driver mutations that functionally drive the malignancy of tumor cells and the rest are likely passenger mutations conferring no selective advantage [1]. Distinguishing driver mutations from passenger mutations remains one of the most pressing challenges in ongoing cancer genomic research [2].

Methods

Results

Conclusion