Integration Analysis of Bayesian and Machine Learning for Heterogeneity, Biomarkers, and Optimal Combination Regimens of Pucotenlimab in Solid Tumors

The efficacy of PD-1 inhibitor pucotenlimab (HX008) in solid tumors exhibits heterogeneity. This study integrated data from 6 clinical trials (covering gastric/gastroesophageal junction cancer, triple-negative breast cancer, melanoma, and dMMR/MSI-H solid tumors) using Bayesian meta-analysis, machine learning (optimal XGBoost AUC = 0.86), and network meta-analysis to construct an integrated "efficacy-prediction-safety" framework. Bayesian analysis showed pucotenlimab significantly improved outcomes versus control (ORR OR = 4.82, 95% CrI: 3.65-6.38; PFS HR = 0.41, 0.32-0.52; OS HR = 0.37, 0.26-0.51). Subgroups revealed TNBC patients with gemcitabine/cisplatin achieved highest ORR (80.6%, 62.5%-92.6%), while mucosal melanoma showed lowest response (8.7%, 1.1%-28.0%). Combination therapy demonstrated superior efficacy to monotherapy (ORR OR: 5.91 vs. 2.35). Machine learning identified 4 efficacy biomarkers (KMT2D mutation, post-treatment NLR decrease, PD-L1 CPS ≥ 1, high eotaxin) and 3 irAE risk factors (baseline NLR ≥ 4, irinotecan combination, high VEGF). Network analysis recommended regimens: gemcitabine/cisplatin for TNBC (SUCRA = 95.7%), oxaliplatin/capecitabine for G/GEJ cancer (ORR = 60.0% vs. irinotecan 27.6%, HR = 0.45). The integrated model classified high-benefit (≥ 3 points; ORR 78.2%) and low-benefit (≤ 0 points; ORR 28.3%) groups, plus high-risk (≤ -2 points; grade ≥ 3 irAEs 41.2%) and low-risk (≥ 1 point; irAEs 3.5%) groups, validated by decision curve analysis. This defines precise application scenarios and provides an extensible analytical paradigm.