Abstract
Low-cost multi-omics sequencing is expected to become clinical routine and transform precision oncology. Viable computational methods that can facilitate tailored intervention while tolerating sequencing biases are in high demand. Here we propose a class of transparent and interpretable computational methods called integral genomic signature (iGenSig) analyses, that address the challenges of cross-dataset modeling through leveraging information redundancies within high-dimensional genomic features, averaging feature weights to prevent overweighing, and extracting unbiased genomic information from large tumor cohorts. Using genomic dataset of chemical perturbations, we develop a battery of iGenSig models for predicting cancer drug responses, and validate the models using independent cell-line and clinical datasets. The iGenSig models for five drugs demonstrate predictive values in six clinical studies, among which the Erlotinib and 5-FU models significantly predict therapeutic responses in three studies, offering clinically relevant insights into their inverse predictive signature pathways. Together, iGenSig provides a computational framework to facilitate tailored cancer therapy based on multi-omics data.