机制建模平台预测患者应答以指导抗体药物偶联物的临床开发

To address the critical translational gap in predicting patient responses to novel antibody-drug conjugates in preclinical or early clinical stages, we developed a mechanistic modeling platform integrating drug mechanism and patient genomic profiles to predict response rates before mature clinical data becomes available. Methods: We developed PatientMatrix, a systems pharmacology modeling-machine learning hybrid platform, and built a specialized platform for antibody-drug conjugates (PatientMatrix-ADC). The platform comprises three components: a mechanistic ADC model employing systems pharmacology to convert drug properties into patient response predictions, a machine learning model, and a combination model integrating both outputs. This platform enables prediction of ADC monotherapy or combination with PD(L)1 inhibitor. We applied this framework to sacituzumab govitecan, a TROP2-targeted ADC. The mechanistic model captures three critical properties: antigen target, payload sensitivity, and resistance mechanisms. The platform was trained on harmonized patient genomic data and clinical trials including ASCENT, then applied to TCGA reference cohorts to predict patient responses. Results: PatientMatrix-ADC accurately predicted sacituzumab govitecan plus pembrolizumab responses in the EVOKE-02 trial (metastatic NSCLC), achieving accuracy across PD-L1 subgroups: 77.6% predicted vs 75.0% observed (PD-L1 ≥50%), 50.6% predicted vs 44.0% observed (PD-L1 <50%), and 58.1% predicted vs 54.0% observed (ITT population). The model demonstrated qualitative concordance beyond fitted indications, with first-line predictions of 29.0% (endometrial) and 54.3% (urothelial) appropriately trending higher than second-line reports (22.2% and 27.4%, respectively), as expected for treatment-naive patients. Indication prioritization across TCGA identified several candidate tumor indications missed by TROP2 target expression-based approaches. Esophageal cancer emerged as a potential opportunity despite low TROP2 expression, driven by favorable SN38 sensitivity scores. Critically, PatientMatrix-ADC predicted target expression-stratified responses (TROP2 high, medium, and low) learned from ASCENT clinical trials. This enabled the model to estimate TROP2-stratified patient responses in ASCENT-03. Conclusions: The PatientMatrix modeling platform uses genomic, clinical, and mechanistic data to predict patient response rates to novel therapeutics in preclinical and early clinical stage, and guide indication selection and patient stratification for agents in clinical development. Case studies with sacituzumab govitecan clinical trials demonstrated the utility of the approach as well as broader ability to predict response in adjacent clinical settings or indications.