Meta-Modeling with Drug Discovery Stack Regressor for Drug Discovery: An Explainable AI Perspective

Drug discovery faces persistent challenges, including the need to handle heterogeneous datasets, extended timelines, and difficulties in accurately predicting drug-target interactions. These issues hinder the timely development of therapeutic interventions, especially during public health crises such as COVID-19. This study integrates ensemble machine learning with explainable artificial intelligence (XAI) to enhance predictive accuracy and transparency.

The dataset of 104 COVID-19-targeting compounds was used to train three regression models: Random Forest, Support Vector Regression, and Multi-Layer Perceptron. Ensemble strategies-Voting and Stacking Regressors-were implemented. SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) were employed to identify feature importance at global and local levels.

The Drug Discovery Stack Regressor achieved the best performance, with a mean squared error (MSE) of 0.18 and R² of 0.88. SHAP and LIME analyses identified EffectiveRotorCount3D and YStericQuadrupole3D as the most influential descriptors. These features correspond to molecular flexibility and steric effects relevant to drug activity.

Combining ensemble modeling with explainability improves both prediction robustness and interpretability. The integration of SHAP and LIME enables chemically meaningful insights into compound behavior, supporting informed molecular design and increasing model transparency. This dual-layer approach enhances confidence in AI-driven decision-making in the drug discovery process.

This study highlights that explainable ensemble models can improve the reliability, interpretability, and applicability of AI in drug discovery. The framework is scalable for broader datasets and offers actionable insights for rational therapeutic development and regulatory alignment.