IPDsim: An interpretable model to assess individual clinical antagonism in combination therapies for cancers.

e13646 Background: Combination therapies, integrating chemotherapies, biologics, and radiotherapies, have shown promise in treating cancer and the potential to cure cancer. However, the success of these treatments in clinical trials is limited, with synergy rare and antagonism common. Most combination therapies, despite offering group-level benefits, may not outperform individual treatments for specific patients. Current models for assessing clinical efficacy of combination therapies overlook variability in correlations across different combinations, which impedes application in clinical assessments and hinders their use in combinations design and approval decisions. Methods: We initially presented extensive and systematical evidence from cell trials, Patient-Derived Xenografts (PDX) mouse models, and mathematical theoretical models to emphasize the importance of the individual patient perspective in combination therapy and state that individual clinical antagonism is common among combination therapies even they are favorable and marketed. Subsequently, we developed a interpretable model (IPDsim) based on collaborative filtering algorithm to simulate counterfactual individual clinical effect from PDX models and clinical trials, assessing and comparing different similarity model based on PDX models. And then, we systematically collect combination therapies in clinical trials to treat cancers, and map single-agent clinical trials to assess clinical antagonism for those combination therapies for both individual and population perspectives. Furthermore, we integrated our findings with combination properties, sequential therapies, and biomarkers to facilitate the design of combination therapies and inform drug approval decisions. Results: IPDsim model significantly enhances prediction accuracy, raising the AUC from 0.68 to 0.98 compared to current independent action model, and achieving an R 2 of 0.91 for predicting individual antagonistic probability. Over one-third of cancer combination therapies are population favorable yet exhibit individual antagonism in systematically collected clinical trials, with over 50% of patients potentially experiencing inferior effects compared to monotherapies. Distance differences significantly show that closer proximity leads to greater individual benefits(P < 0.05), with all distances differences > 0. Conclusions: We have developed the IPDsim model, an interpretable and efficient tool for predicting clinical individual antagonism, informed by systematic evidence of the prevalence of individual antagonism. The use of biomarkers, the combination of segregated drugs with lower distance differences, and new technologies may reduce individual antagonism, enhance the likelihood of market success, and confer greater benefits to individual patients.