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

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Drug discovery

Background:

• The RET V804M gatekeeper mutation confers clinical resistance to multiple kinase inhibitors.
• This mutation significantly reduces the efficacy of targeted cancer therapies.

Purpose of the Study:

• To develop an interpretable Quantitative Structure-Activity Relationship (QSAR) model for predicting inhibitory potency against the RET V804M variant.
• To facilitate rapid in silico prediction of drug efficacy for RET V804M inhibitors.

Main Methods:

• A curated dataset of RET V804M inhibitors was used to build the QSAR model.
• 140 RDKit descriptors were selected after preprocessing and filtering.
• Gradient Boosting Regression was employed and optimized via extensive grid search and 5-fold cross-validation.

Main Results:

• The Gradient Boosting Regression model achieved a Pearson correlation coefficient (r) of 0.737 and an RMSE of 0.564.
• A leverage-residual Williams plot confirmed that most compounds were within the model's applicability domain.
• Feature importance analysis identified electronegativity, hydrophobic surface distribution, molecular flexibility, and polarizability as key drivers of inhibition.

Conclusions:

• The developed QSAR model accurately predicts RET V804M inhibitor potency.
• The model is interpretable, highlighting key molecular features influencing inhibition.
• An interactive web application was created to enable practical use for predicting pIC50 values of novel RET V804M inhibitors.