A data-driven approach to modeling drug synergy and antagonism

Drug combination therapies are essential for improving treatment efficacy and reducing resistance in complex diseases. However, identifying synergistic or antagonistic interactions experimentally is resource-intensive and often infeasible at large scale. Conventional synergy assessment depends on dose–response matrices, but their application is largely retrospective and limited by experimental throughput. Machine Learning (ML) offers a data-driven approach to predict drug synergy from chemical structures and bioactivity profiles. By leveraging molecular fingerprints and observed interaction labels, ML models can capture subtle structure–response relationships that are not easily identifiable through traditional approaches. 

In this study, three ML algorithms, i.e., Support Vector Machine (SVM), k-Nearest Neighbors (k-NN), and Random Forest (RF), were applied to a subset of the DrugCombDB dataset comprising experimental drug-combination results generated in lung, colon, breast, and prostate cancer cell lines. The dataset includes chemical descriptors for paired drugs along with in vitro drug-combination measurements. Across all four cancer types, the tuned classification models demonstrated comparable performance, achieving accuracies between 0.72 and 0.83 across the prediction of both synergistic and antagonistic classes. For regression, ZIP synergy scores served as continuous response variables, and Random Forest achieved the strongest performance (R² = 0.56; MSE = 39.40). However, the overall predictive accuracy remains limited, indicating the need for additional optimization. 

These findings illustrate the feasibility of predicting drug interaction effects using molecular fingerprints; nonetheless, improvements in feature representation, incorporation of biological context, and expansion to larger, more diverse datasets are likely required to enhance model generalizability and real-world applicability.