An integral genomic signature approach for tailored cancer targeted therapy using genome-wide sequencing data

Abstract

With the advent of low-cost sequencing, transcriptome and genome sequencing is expected to become clinical routine and transform precision oncology within next decade. However, viable genome-wide modeling methods that can facilitate rational selection of patients for tailored intervention while tolerating sequencing biases are far lacking. Here we propose an integral genomic signature (iGenSig) analysis as a new class of transparent, interpretable, and resilient methods for precision oncology based on multiple types of genome-wide sequencing data. We postulate that the redundant high-dimensional genomic features, which are typically eliminated during multi-omics modeling, may help overcome the sequencing biases. We thus conceive a novel method that models the therapeutic response using the high-dimensional transcriptional and mutational features predictive of tumor response, which we termed as an integral genomic signature (iGenSig), and then algorithmically resolve the feature redundancy tailored for each patient subject. Using genomic dataset of chemical perturbations, we developed the iGenSig models for predicting targeted therapy responses, and applied selected models to independent datasets for cancer cell lines, patient-derived xenografts, and patient subjects. iGenSig models exhibit outstanding cross-dataset performance compared to artificial intelligence methods, with exceptional resilience against simulated errors in genomic features. In particular, the iGenSig model for the EGFR inhibitor Erlotinib significantly predicted the responses of patient-derived xenografts and patients from a clinical trial, biological interpretation of which led to new insights into the predictive signature pathways with clinical relevance. Together, iGenSig will provide a computational infrastructure to empower tailored cancer intervention based on genome-wide sequencing data.