Mutational Signatures for Survival Prediction

Motivation - Cancer remains one of the deadliest diseases worldwide and while advancements have been made in cancer treatment, cancer's heterogeneous nature makes it challenging to find a good treatment. Survival prediction for cancer patients can aid in choosing a treatment plan. Various machine learning methods have been employed to predict the survival of cancer patients, but they offer little insight into why a patient's survival is likely or not. Mutational signatures can offer an explanation on what a patient's cancer originates from, and can be linked to certain outside factors such as UV radiation. Even though mutational signatures have been employed in other problems, like predicting DNA repair pathway deficiencies, they have not been used in survival prediction. Integrating the survival problem with the extraction of mutational signatures could allow for extracting signatures that are particularly indicative of a patient's survival, providing a better prediction and more insight into why a patient's survival is predicted that way.

Results - We propose Multi-Task Auto-Encoder Cox (MTAE-Cox), which combines a non-negative auto-encoder for signature extraction with a Cox model for survival prediction and optimizes these in a multi-task manner. Our method jointly optimizes the auto-encoder's reconstruction error and the Cox loss, integrating the survival prediction problem into the signature extraction. MTAE-Cox is applied to four cancers of the TCGA dataset (GBM, HNSC, OV, SKCM) and its prediction performance is compared to Cox models using Gene Expression, Mutational Catalog, and exposures to COSMIC signatures. MTAE-Cox outperforms the generally applied gene expression (median C-index of 0.579 over 0.561 for gene expression) for GBM and outperforms Cox using non-integrated signatures derived by NMF for three of the four cancers. MTAE-Cox can extract biologically relevant signatures that are similar to COSMIC signatures that are known to be common in the specific type of cancer, for example SBS3 for ovarian cancer.