Improved response prediction to immune checkpoint inhibition by combining TMB and WGS-based genomic features in NSCLC.

Abstract

e21077 Background: The appearance of immune checkpoint inhibitors (ICIs) has revolutionized treatment strategies in advanced NSCLC. The clinical efficacy and indications for ICIs are currently estimated using the immunohistochemistry (IHC) profile of programmed death-ligand 1 (PD-L1) protein. Recent findings indicate that tumor mutational burden (TMB) is a unique feature that may improve response prediction to ICIs across multiple cancer types [1]. Nevertheless, only a minority of NSCLC patients will benefit from ICIs. Therefore, unraveling the genomics of NSCLC and the subsequent discovery of novel predictive biomarkers remains a high unmet need. Whole-genome sequencing (WGS) bounded with AI may identify biomarkers and composed signatures which will better than TMB or PD-L1 predict response to ICIs. Methods: To develop the PD1Dx classifier with the highest predictive value for ICI response, we combined WGS data analyses with machine learning approaches. Consistently, we divided WGS data of 57 NSCLCs derived from the Hartwig Medical Foundation (HMF) database into training (n = 39) and hold-out (n = 18) sets. Patients treated with pembrolizumab or nivolumab alone were labeled based solely on clinical metadata provided by the repository, reaching a total of 22 responders and 35 non-responders. Results: Recursive Feature Elimination using SHAP importance (ShapRFECV) selected 15 top predictive genomic features, based on which the final random forest model was built. The random forest model achieved the best performance, with an accuracy of 0.72, a recall of 0.73, and a precision of 0.80. Using the same learning method and dataset, we built a model based on a single TMB feature whose performance determined by accuracy, recall and precision was 0.67, 0.64, and 0.78, respectively. Apart from TMB, we also identified other predictive features, among the 15 top of which were, e.g., features related to transcriptional strand bias and C > A/T single nucleotide variants correlating with mismatch repair. Conclusions: Our results confirm the predictive value of features other than TMB alone in predicting ICI responses in NSCLC. Further investigation is required to validate biomarkers of sensitivity and resistance to ICIs. Krieger T, Pearson I, Bell J, Doherty J, Robbins P. Targeted literature review on use of tumor mutational burden status and programmed cell death ligand 1 expression to predict outcomes of checkpoint inhibitor treatment. Diagn Pathol 2020;15:6.