Interpretable Machine Learning Frameworks for QSAR Modeling: A Comparative Study of Classical Algorithms and Transformer-CNN Architectures

Abstract

Phosphodiesterase type 5A (PDE5A) inhibitors represent an important class of bioactive molecules with therapeutic relevance across cardiovascular, neurological, and oncological disorders. In this work, we develop an interpretable Quantitative Structure–Activity Relationship (QSAR) framework that integrates classical machine learning (ML) approaches with a hybrid deep learning architecture (Transformer–Convolutional Neural Network, Transformer–CNN) to model enzyme inhibition using a curated and diverse data set from ChEMBL. Multiple structural descriptors and fingerprint-based representations were evaluated using established ML algorithms, with fingerprint-driven models, particularly those employing FCFP4 features, achieving the highest predictive performance with Matthews Correlation Coefficients up to 0.783. Comparative analysis revealed that optimized classical ML models slightly outperform the Transformer–CNN approach for this data set. Model interpretability was explored through Shapley Additive Explanations (SHAP) analysis, similarity maps, and atom-wise relevance scores, providing mechanistic insights and highlighting the importance of molecular symmetry and topological features in determining inhibitory potency. To support practical deployment, the final models were implemented in a freely accessible web application built with Flask and Streamlit for rapid prediction of novel compounds’ inhibitory potential. Overall, this work provides a transparent and reproducible QSAR workflow, clarifies the relative strengths of classical and deep learning approaches, and offers a practical tool to assist in the design of next-generation PDE5A inhibitors.