BiGranMolNet: A deep learning method for predicting blood-brain barrier permeability based on Bi-Granularity...

Area of Science:

• Computational chemistry
• Drug discovery
• Neuroscience

Background:

• The blood-brain barrier (BBB) restricts the passage of most central nervous system (CNS) drugs.
• Accurate prediction of BBB permeability is essential for developing effective CNS therapeutics.

Purpose of the Study:

• To develop a computational model, BiGranMolNet, for predicting blood-brain barrier (BBB) permeability.
• To enhance the efficiency of central nervous system (CNS) drug development through improved permeability prediction.

Main Methods:

• Utilized graph convolutional networks (GCNs) to build a Bi-Granularity Molecular Graph Network (BiGranMolNet).
• Integrated multi-source datasets for comprehensive regression (2148 samples) and classification (16,904 samples) benchmarks.
• Employed atom-level and motif-level molecular graph representations, fused via cross-attention, with a weighted loss function for class imbalance.

Main Results:

• BiGranMolNet demonstrated stable and competitive performance in 5-fold cross-validation.
• The model achieved robust results in structure-aware evaluations for both regression and classification tasks.
• Validated the efficacy of dual-granularity feature integration and weighted loss function.

Conclusions:

• BiGranMolNet offers a reliable computational tool for predicting BBB permeability by integrating atomic and motif-level structural information.
• The framework supports early-stage CNS drug screening by identifying compounds with high brain exposure potential.
• Provides structure-aware insights for optimizing drug candidates during CNS drug development.