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

Area of Science:

• Computational chemistry
• Pharmacology
• Machine learning

Background:

• Blood-brain barrier permeability (BBBP) is crucial for central nervous system drug discovery.
• Existing machine learning (ML) models for BBBP lack interpretability.
• Understanding physicochemical drivers of BBBP is essential for rational drug design.

Purpose of the Study:

• To develop an interpretable multimodal ML framework for BBBP prediction.
• To integrate molecular fingerprints and image features for enhanced predictive accuracy.
• To elucidate key physicochemical features governing blood-brain barrier passage.

Main Methods:

• A multimodal ML framework combining molecular fingerprints (Morgan, MACCS, RDK) and image features.
• A stacking ensemble model for classification (BBB-permeable vs. nonpermeable).
• A deep learning framework with Transformer and CNN for regression (logBB value prediction).
• Principal Component Analysis (PCA) and SHAP analysis for feature interpretability.

Main Results:

• The multimodal framework achieved competitive predictive stability and generalization.
• Feature interpretability analysis identified key molecular descriptors influencing BBBP.
• Attention maps revealed token-level relationships in molecular representations.
• The model offers enhanced transparency and mechanistic insight into BBBP.

Conclusions:

• The proposed interpretable multimodal ML framework advances BBBP prediction accuracy and transparency.
• This approach provides mechanistic insights into the physicochemical basis of blood-brain barrier permeability.
• The framework serves as a foundation for developing more transparent and physics-informed drug discovery tools.