Systemic immunometabolic profiling classifies cisplatin sensitivity states using interpretable machine learning

Abstract

Cisplatin resistance limits the effectiveness of platinum-based chemotherapy for lung adenocarcinoma, yet practical systemic diagnostics for cisplatin sensitivity are lacking. We developed ImmunoMetabolic Profiling Analysis and Classification Tool (IMPACT), an interpretable machine learning pipeline that selects the best performing model and reduces it to a minimal, mechanistically informative feature set via recursive feature elimination. In a syngeneic orthotopic model, we quantified 25 serum amino acids and 16 immune cell populations across bone marrow, spleen, lung, and mediastinal lymph nodes to capture systemic immunometabolic states. IMPACT classified cisplatin-sensitive versus cisplatin-resistant tumors with high accuracy (AUC = 0.950), driven primarily by bone marrow MDSCs and serum glutamine. Using the same framework, we also classified Cancer (CS + CR) versus no cancer controls with high accuracy (AUC = 0.955), with lung MDSCs and phosphoserine among the top features.