Machine learning-enhanced QSAR modeling for predicting drug efficacy against the RET V804M kinase domain mutation

The RET V804M gatekeeper mutation is a clinically significant resistance mechanism that reduces the efficacy of several kinase inhibitors. To support rapid insilico prediction of inhibitory potency against this variant, we built an interpretable QSAR model using a curated dataset of experimentally reported RET V804M inhibitors. Following descriptor preprocessing and filtering, 140 RDKit descriptors were retained for model development. Among the algorithms evaluated, Gradient Boosting Regression provided the most reliable performance. Its parameters were optimized using an extensive grid search involving 1728 combinations with 5-fold cross-validation. The resulting model achieved a Pearson correlation coefficient (r) of 0.737 and a Root Mean Squared Error (RMSE) of 0.564, indicating good agreement between the predicted and experimental activities. A leverage-residual Williams plot confirmed that most of the test compounds reside within the applicability domain (AD) of the model. Evaluation of newly reported RET inhibitors further showed that reliable predictions are achieved for compounds contained within this established chemical space. To ensure interpretability, model behaviour was examined using tree-based feature importances, which identified that electronegativity patterns, hydrophobic surface distribution, and molecular flexibility and polarizability are major drivers of RET V804M inhibition. To facilitate practical use, the fine-tuned Gradient Boosting Regressor has been deployed as an interactive web application ( https://ret-biopredictor.streamlit.app/ ), enabling users to predict pIC₅₀ values for candidate RET V804M inhibitors.