A genomic signature for accurate classification and prediction of clinical outcomes in cancer patients treated with immune checkpoint blockade immunotherapy

Abstract

Abstract Tumor mutational burden (TMB) is associated with clinical response to immunotherapy, but application has been limited to a subset of cancer patients. We hypothesized that advanced machine-learning and proper modeling could identify mutations that classify patients most likely to derive clinical benefits. Training data: Two sets of public whole-exome sequencing (WES) data for metastatic melanoma. Validation data: One set of public non-small cell lung cancer (NSCLC) data. Least Absolute Shrinkage and Selection Operator (LASSO) machine-learning and proper modeling were used to identify a set of mutations (biomarker) with maximum predictive accuracy (measured by AUROC). Kaplan–Meier and log-rank methods were used to test prediction of overall survival. The initial model considered 2139 mutations. After pruning, 161 mutations (11%) were retained. An optimal threshold of 0.41 divided patients into high-weight (HW) or low-weight (LW) TMB groups. Classification for HW-TMB was 100% (AUROC = 1.0) on melanoma learning/testing data; HW-TMB was a prognostic marker for longer overall survival. In validation data, HW-TMB was associated with survival (p = 0.0057) and predicted 6-month clinical benefit (AUROC = 0.83) in NSCLC. In conclusion, we developed and validated a 161-mutation genomic signature with “outstanding” 100% accuracy to classify melanoma patients by likelihood of response to immunotherapy. This biomarker can be adapted for clinical practice to improve cancer treatment and care.