A Machine Learning Approach for Predicting Synergism of Mixtures of Compounds in Treating Cancers

Combination therapy is a common strategy in cancer treatment to improve efficacy, minimize side effects, and overcome drug resistance. However, the vast number of possible drug pairings makes experimental screening both time-consuming and costly. Drug interactions can result in synergy and antagonism, emphasizing the need for efficient predictive models. In this study, we develop a pairwise-compound quantitative structure-activity relationship (pQSAR) framework based on an ensemble stacking machine learning strategy to enhance prediction of empirically defined combination-effect classes using molecular and PK-related descriptors. Our approach leverages molecular descriptors along with key pharmacokinetic properties, the estimated accumulated drug concentration over 24 h (AUC), bioavailability (Fb) and volume of distribution at steady state (Vdss) to train the model on pairwise drug combinations. Building upon this foundation, we develop an ensemble stacking classification model that integrates multiple ML algorithms, providing a more accurate and robust prediction of drug synergy. Using the ALMANAC drug dataset across 8 cell lines from 4 different cancer types, our model achieves an accuracy exceeding 0.88 in predicting the effects of drug pairs. These results indicate that the proposed framework can capture predictive patterns associated with experimentally defined drug-combination effect classes in the evaluated dataset.