Interpretable Multimodal Deep Ensemble Framework Dissecting Bloodbrain Barrier Permeability with Molecular Features

Blood-brain barrier permeability (BBBP) prediction plays a critical role in the drug discovery process, particularly for compounds targeting the central nervous system. While machine learning (ML) has significantly advanced the prediction of BBBP, there remains an urgent need for interpretable ML models that can reveal the physicochemical principles governing BBB permeability. In this study, we propose a multimodal ML framework that integrates molecular fingerprints (Morgan, MACCS, RDK) and image features to improve BBBP prediction. The classification task (BBB-permeable vs nonpermeable) is addressed with a stacking ensemble model combining multiple base classifiers. The proposed framework demonstrates competitive predictive stability, generalization ability, and feature interpretability compared with recent approaches, under comparable evaluation settings. Beyond predictive performance, our framework incorporates Principal Component Analysis (PCA) and Shapley Additive Explanations (SHAP) analysis to highlight key fingerprint features contributing to predictions. The regression task (logBB value prediction) is tackled by a multi-input deep learning framework, incorporating a Transformer encoder for fingerprint processing, a convolutional neural network (CNN) for image feature extraction, and a Multi-Head Attention fusion mechanism to enhance feature interactions. Attention maps derived from the multimodal features reveal token-level relationships within molecular representations. This work provides an interpretable framework for modeling BBBP with enhanced transparency and mechanistic insight and lays the foundation for future studies incorporating transparent descriptors and physics-informed features.