Factor Models for Cancer Signatures

TL;DR

This paper introduces a new statistical method adapted from finance risk models to extract more stable and computationally efficient cancer signatures from genome data, discovering three novel signatures.

Contribution

It applies risk modeling techniques to cancer genomics, enabling faster and more reliable extraction of mutational signatures with the identification of new cancer-specific signatures.

Findings

Lower variability in signatures from filtered data

Tenfold reduction in computational cost

Discovery of three novel cancer signatures

Abstract

We present a novel method for extracting cancer signatures by applying statistical risk models (http://ssrn.com/abstract=2732453) from quantitative finance to cancer genome data. Using 1389 whole genome sequenced samples from 14 cancers, we identify an "overall" mode of somatic mutational noise. We give a prescription for factoring out this noise and source code for fixing the number of signatures. We apply nonnegative matrix factorization (NMF) to genome data aggregated by cancer subtype and filtered using our method. The resultant signatures have substantially lower variability than those from unfiltered data. Also, the computational cost of signature extraction is cut by about a factor of 10. We find 3 novel cancer signatures, including a liver cancer dominant signature (96% contribution) and a renal cell carcinoma signature (70% contribution). Our method accelerates finding new cancer signatures and improves their overall stability. Reciprocally, the methods for extracting cancer signatures could have interesting applications in quantitative finance.