Generative Adversarial Networks in Precision Oncology

Abstract

Precision medicine strives to deliver improved care based on genetic patient information. Towards this end, it is crucial to find effective data representations on which to perform matching and inference operations. We develop and evaluate a generative adversarial neural network (GAN) approach to representation learning with the goal of patient-centric literature retrieval and treatment recommendation in precision oncology. Several large-scale corpora including the COSMIC Cancer Gene Census, COSMIC Mutation Data, Genomic Data Commons (GDC) and 26M MEDLINE abstracts are used to train GANs for synthesizing genetic mutation patterns that likely correspond to patient properties such as their demographics or cancer type. The introduction of GANs into the literature retrieval and treatment recommendation process results in significant improvements in performance by increasing the recall of a range of methods at stable precision. Finally, we propose a method to discover novel gene-gene interaction hypotheses to guide future research.