HDAC3_VS_assistant: cheminformatics-driven discovery of histone deacetylase 3 inhibitors

Histone deacetylase 3 (HDAC3) inhibitors keep significant therapeutic promise for treating oncological, neurodegenerative, and inflammatory diseases. In this work, we developed robust QSAR regression models for HDAC3 inhibitory activity and acute toxicity (LD₅₀, intravenous administration in mice). A total of 1751 compounds were curated for HDAC3 activity, and 15,068 for toxicity. The models employed molecular descriptors such as Morgan fingerprints, MACCS-166 keys, and Klekota-Roth, PubChem fingerprints integrated with machine learning algorithms including random forest, gradient boosting regressor, and support vector machine. The HDAC3 QSAR models achieved Q²_test values of up to 0.76 and RMSE values as low as 0.58, while toxicity models attained Q²_test values of 0.63 and RMSE values down to 0.41, with applicability domain (AD) coverage exceeding 68%. Internal validation by fivefold cross-validation (Q²cv = 0.70 for HDAC3 and 0.60 for toxicity) and y-randomization confirmed model reliability. Shapley additive explanation (SHAP) was also used to explain the influence of modeling features on model prediction results. The most predictive QSAR models are integrated into the developed HDAC3_VS_assistant application, which is freely available at https://hdac3-vs-assistant-v2.streamlit.app/ . Virtual screening conducted using the HDAC3_VS_assistant web application allowed us to reveal a number of potential inhibitors, and the nature of their bonds with the active HDAC3 site was additionally investigated by molecular docking.