Computational Disease Subtyping based on Joint Analysis of Clinical and Genomic Data

ABSTRACT Objective To investigate application of non-negative tensor decomposition for disease subtype discovery based on joint analysis of clinical and genomic data. Data and Methods Somatic mutation profiles including 11,996 genes of 503 breast cancer patients from the Cancer Genome Atlas (TCGA) along with 11 clinical variables and markers of these patients were used to construct a binary third-order tensor. CANDECOMP/PARAFAC method was applied to decompose the constructed tensor into rank-one component tensors. Definitions of breast cancer verotypes were constructed from the patient, gene and clinical vectors corresponding to each component tensor. Patient membership proportions in the identified verotypes were utilized in a Cox proportional hazards model to predict their survival. Results Qualitative evaluation of the verotypes obtained by tensor factorization indicates that they correspond to clinically meaningful breast cancer subtypes. While some components correspond to the known HER2- or ER-positive breast cancer subtypes, other components correspond to a variant of triple negative subtype and a cohort of patients with high mutation load of tumor suppressor genes. Quantitative evaluation indicates that the Cox model utilizing computationally discovered breast cancer verotypes is more accurate (AUC=0.5796) at predicting patient survival than the Cox models utilizing random patient membership proportions in cancer subtypes (AUC=0.4056) as well as patient membership proportions in genotypes (AUC=0.4731) and phenotypes (AUC=0.5047) obtained by non-negative factorization of the somatic mutation and clinical matrices. Conclusion Non-negative factorization of a binary tensor constructed from clinical and genomic data enables high-throughput discovery of breast cancer verotypes that are effective at predicting patient survival.