Tumor microbiome-based deep learning diagnostic and prognostic models for predicting immunotherapy efficacy and progression-free survival in lung cancer

Immune checkpoint inhibitors have become an important treatment for advanced lung cancer, but patient responses exhibit substantial interindividual variability. The tumor-associated microbiome is closely related to immunotherapy efficacy. Tumor tissue samples from lung cancer patients receiving immunotherapy were analyzed using 16 S-5R rRNA sequencing (Cohort 1). Two core models were constructed: a deep learning-based bacterial diagnostic model (BDM) incorporating hidden layers and a Squeeze-and-Excitation block to predict immunotherapy response status, and a Cox proportional hazards (CoxPH) prognostic model to predict progression-free survival (PFS). Furthermore, four core bacterial species were selected using SHAP values, and a probe-based multiplex qPCR detection system (BDM-mqPCR) was developed (Cohort 2). The BDM achieved a sensitivity of 96.55% and 87.50%, specificity of 96.15% and 100.00%, and AUC of 0.989 and 1.000 in the training and validation set. The CoxPH model showed high concordance between predicted PFS curves and actual outcomes in both the training and validation sets, effectively distinguishing high-risk from low-risk patients, further providing the probability distribution of their PFS to help physicians decide whether to prepare subsequent treatment in advance or intensify monitoring. BDM-mqPCR constructed using the four SHAP-selected species achieved AUCs of 0.828 (training) and 0.869 (validation), which were lower than those of the BDM but significantly superior to Biomarker model, offering the advantages of rapidity, cost-effectiveness, and target specificity. BDM model and the CoxPH model achieved a complete assessment from whether a response occurs to how long the response lasts, providing a microbiome-driven precision tool for individualized decision-making in lung cancer immunotherapy.