Decision model for durable clinical benefit from front- or late-line immunotherapy alone or with chemotherapy in large population of non-small cell lung cancer.

Abstract

e20533 Background: Predictive biomarkers and models of immune checkpoint inhibitors (ICI) have been extensively studied in NSCLC. However, evidence for many biomarkers remains inconclusive, and the opaqueness of machine learning models hinders their practicality. We aimed to provide compelling evidence for biomarkers and develop a transparent decision-tree model. Methods: We consolidated data from ICI-treated patients with NSCLC (n = 3,288) across real-world multicentre studies, public cohorts, and the Choice-01 trial (NCT03607539). Over 50 features were gathered to examine the predictive capacity of durable clinical benefits (DCB) from mono-ICIs and ICIs plus chemotherapy. Noteworthy biomarkers were identified to establish a decision tree model. We also explored the tumour microenvironment and peripheral PD-1+ CD8+ T-cell receptor profiles. Results: Multivariate logistic regression analysis identified tumour histology, PD-L1 expression, tumour mutational burden, line, and regimen of ICI treatment as significant factors. Mutation subtypes of EGFR, KRAS, KEAP1, STK11, and disruptive TP53 mutations demonstrated significant association with DCB. The decision-tree (DT10) model, using the aforementioned five clinicopathological markers and five genomic mutations, predicted DCB with superior performance in the training set (N = 286, AUC = 0.82), and consistently outperformed the logistic regression, random forest, and support vector machine models, as well as FDA-approved biomarkers of PD-L1 and TMB in the internal and three external test sets (Table). We visualized the decision routes for DCB from ICI-based treatment based on the DT10 model, which was interpretable for clinical use. Furthermore, DT10 predicted-DCB patients manifested significantly longer progression-free survival and overall survival than -NDB patients across the train and test tests (HR < 1, p < 0.05). Additionally, the DT10 predicted-DCB group presented an enriched inflamed tumour immune phenotype (67%) and higher peripheral TCR diversity, whereas the -NDB group showed an enriched desert immune phenotype (86%) and higher peripheral TCR clonality. Conclusions: Based on the large population, we provided persuasive evidence for the predictive reliability of biomarkers in NSCLC. The development of the DT10 model introducing an easily interpretable, visually engaging, and clinically relevant model, which assisted the decision making in varied clinical scenarios. [Table: see text]