Supervised Learning-Driven Prediction of Small-Molecule Modulator Bioactivity Against Protein-Protein Interactions

<title>Abstract</title> <p> Small molecules can modulate protein-protein interactions (PPIs). The structural complexity of PPIs and the evaluation of small-molecule bioactivity remain major barriers in drug discovery. To address this, we proposed a supervised learning-driven framework to computationally predict the bioactivity of small molecules, often quantified by the half-maximal inhibitory concentration (IC ₅₀ ). We used multiple supervised regression algorithms (RF, GB, SVR, LSTM) across three chemical descriptor sets (RDKit, PubChem, PaDEL) on a comprehensive dataset of 3,451 small molecules with known bioactivity against 176 distinct biological targets, including 47 PPIs, 91 single proteins, and 38 cell lines. Random forest with RDKit descriptors achieved best performance, with an R² of 0.75 and RMSE of 0.78 in cross-validation and an R² of 0.74 and RMSE of 0.79 in blind-set evaluation, closely matched by PaDEL descriptors (R² of 0.75 in both sets); PubChem descriptors consistently underperformed across all four regressors (R² of 0.47–0.66) in cross-validation. Performance of the model decreased on a multitarget external validation set of 1,528 structurally diverse compounds, with an R² of 0.10 and an RMSE of 1.25. This reflects the well-known challenge of extending QSAR models beyond their training domain. Median absolute percentage error remained low, falling in the range of 16% to 25% across cross-validation and blind evaluation, while symmetric mean absolute percentage error was elevated across all models and descriptor sets, ranging from 76% to 93%. We also found that RDKit chemical descriptors were highly influential, as indicated by SHAP-based feature importance analysis, in predicting bioactivity for key PPIs, including p53-MDM2, menin-MLL, IL-15–IL-15Rα, and YAP1-TEF-1. The proposed framework establishes a generic, scalable, and data-driven strategy for predicting the biological activity of different PPI modulators, making it a valuable resource for computational chemists and drug discovery researchers. </p>