Interaction effect between mutational events on predicting efficacious immunotherapy in nonsquamous NSCLC: Results from 1,083 patients.

Abstract

e21100 Background: Despite that the biomarkers of single mutational events predicting efficacy of immune checkpoint inhibitors (ICIs) have been intensively studied in non-squamous NSCLC, the interaction effects between mutations on prediction were scarcely investigated due to the limited sample size. Deeper understanding of these inter-relations could optimize the prediction model and provide insights for basic research on anti-tumor immunity. Methods: In total, 1083 patients with non-squamous NSCLC were included. 288 patients with both PFS and OS data receiving monotherapy from the MSKCC, DFCI, and SNCC cohorts were included in the training set-1. 304 and 294 patients from the POPLAR/OAK cohort treated with atezolizumab or docetaxel were included in the training set-2 and the control set respectively. Validation set-1 and validation set-2 respectively consists of 104 patients receiving monotherapy with only OS data and 93 patients treated with combination therapy from the MSKCC cohort. Risk models using single mutations (uni-model) and both single mutations and interactions (inter-model) were developed via multivariable Cox proportional hazards regression model. Results: By cross-validation in training set-1/2, five single mutational events, including mutations of EGFR, STK11, PTPRD, members in PI3K pathway, and members in HRR pathway, and five interactions ( EGFR* PI3K pathway, TP53* KRAS, TP53* ERBB4, cell cycle pathway*HRR pathway, and PI3K pathway*chromatin remodeling pathway) were identified as predictive biomarkers. The score of inter-model exhibited higher correlation with immunotherapeutic PFS in the training sets, compared to the uni-model ( Dxy: 0.173 vs. 0.139). Uni-score low was moderately associated with longer OS in the validation set-1 (HR = 0.62, 95% CI 0.35-1.08, P = 0.092), and PFS in the validation set-2 (HR = 0.32, 95% CI 0.18-0.60, P < 0.001). Inter-score low showed better predictive utility in the validation set-1 (HR = 0.53, 95% CI 0.30-0.94, P = 0.030), and the validation set-2 (HR = 0.23, 95% CI 0.10-0.50, P < 0.001). On the contrary, inter-score low was associated with worse PFS and OS in the control set, indicating that inter-score low might be a negative prognostic biomarker, but positive predictive biomarker to efficacious immunotherapy. Conclusions: Our study suggests that the predictive biomarkers might not be similarly predictive in every single subpopulation, and they might be merely valid under certain circumstances, which is termed by “interaction effect”. Given this, the future researches probing into immunotherapeutic predictor should focus on the mutual influence between distinct biomarkers, for instance, adding an interaction term in the Cox proportional hazards regression model.