Discovery of mutations predictive of survival benefit from immunotherapy in first-line NSCLC: A retrospective machine learning study of IMpower150 liquid biopsy data.

Predictive biomarker identification in cancer treatment has traditionally relied on pre-defined analyses, limiting discoveries to expected biomarkers and potentially overlooking novel ones predictive of therapy response. In this work, we develop a novel machine-learning approach capable of exploring full landscape of mutations and combinations and identify potentially new predictive biomarkers for chemoimmunotherapy. Utilizing the liquid biopsy dataset from 313 non-small cell lung cancer (NSCLC) patients in the Phase 3 Impower150 trial (NCT02366143), we developed the HRdiffRF algorithm with a novel hazard ratio-splitting criterion. Predictive mutations and combinations were identified for overall survival (OS) improvement with atezolizumab plus bevacizumab plus carboplatin and paclitaxel (ABCP) compared to bevacizumab plus carboplatin and paclitaxel (BCP). Our analysis confirms the predictive role of KRAS mutations and reveals the predictive value of PTPRD and SMARCA4 mutations in chemoimmunotherapy efficacy. Unlike other KRAS wild-type NSCLC patients, NSCLC patients with KRAS wild-type status and mutations in FAT1, ERBB2, or PTPRD may benefit from chemoimmunotherapy, while NTRK3 and GNAS mutations could negatively impact survival. Patients harboring concurrent KRAS and KEAP1 mutations may not benefit from chemoimmunotherapy. These findings highlight the complex genetic factors influencing treatment response for chemoimmunotherapy in NSCLC. In summary, the proposed machine-learning tool identified potential predictive biomarkers for first-line chemoimmunotherapy in NSCLC and can be readily applied to other tumor types and studies. It can also be extended to explore predictive biomarkers beyond mutations.