CDK4/6抑制剂联合内分泌治疗HR+/HER2-乳腺癌的精准生物标志物：跨多项新辅助试验的机制模型与混合预测模型的外部验证

Background: Clinical response to CDK4/6 inhibitors (CDK4/6i) and endocrine therapy (ET) in HR+/HER2- breast cancer is heterogeneous, and established biomarkers to predict individual response prior to treatment are lacking. We previously developed a mechanistic mathematical model of the combined mechanisms of action to provide patient-specific response scores (Schmiester et al. Clin Cancer Res 2024). To further enhance predictive accuracy, we developed an extended hybrid model (Mechanistic Boosting) that leverages machine learning (ML) to learn and correct the errors of the mechanistic simulations by incorporating additional molecular data. Methods: The core mechanistic model describes the dynamics of protein-protein and drug-protein interactions. It generates a response score based on the baseline expression of six genes: CCND1, CCNE1, ESR1, RB1, MYC, and CDKN1A . The extended hybrid model integrates this mechanistic framework with an ML component trained on the residuals between mechanistic predictions and observed clinical outcomes using 764 additional genes. Both models were validated using data from the CORALLEEN (n=50), NEOPALANA (n=27) and NEOLETRIB (n=85) trials, assessing response via Ki67 levels and the PAM50 risk of relapse (ROR) score. Results: In the primary validation (including the CORALLEEN cohort), the mechanistic model significantly predicted high residual Ki67 levels (>10%) with an AUC of 0.80 and high PAM50 ROR with an AUC of 0.78. In the NEOPALANA and NEOLETRIB validation, the hybrid model demonstrated superior predictive power over the purely mechanistic model by capturing residual variance and non-modeled biological factors. The hybrid approach more accurately stratified patients into response groups, particularly identifying resistant cases where the expression of six genes alone was insufficient. Both models were drug-specific and showed no association with outcomes in patients treated with chemotherapy, confirming their utility as precision tools for ET+CDK4/6i. Conclusions: The integration of mechanistic modeling with machine learning represents a significant advancement in precision oncology. Validation within external cohorts, using Ki67 and ROR, demonstrates that while mechanistic models provide a robust biological foundation, the hybrid model offers enhanced accuracy for clinical decision-making in HR+/HER2- breast cancer patients.