A merged molecular representation deep learning method for blood-brain barrier permeability prediction

Area of Science:

• Computational chemistry
• Pharmacology
• Artificial intelligence in drug discovery

Background:

• Blood-brain barrier (BBB) permeability is critical for developing drugs targeting the central nervous system.
• Experimental methods for BBB permeability assessment are costly and time-consuming, hindering rapid drug discovery.
• High-throughput screening is essential for predicting compound permeability efficiently.

Purpose of the Study:

• To develop an advanced computational tool for predicting blood-brain barrier (BBB) permeability.
• To accelerate the drug discovery process by providing a fast and accurate prediction method.
• To introduce Deep-B3, a deep learning-based multi-model framework for BBB permeability prediction.

Main Methods:

• Developed Deep-B3, a deep learning framework integrating multiple models.
• Encoded compounds using molecular descriptors, fingerprints, molecular graphs, and Simplified Molecular Input Line Entry System (SMILES) notations.
• Utilized pre-trained models to extract latent features from molecular graphs and SMILES, representing data as tabular, image, and text.

Main Results:

• Deep-B3 demonstrated superior performance compared to state-of-the-art models on an independent dataset.
• The multi-modal approach effectively captured complex relationships influencing BBB permeability.
• Validation confirmed the model's accuracy and reliability for predicting compound BBB penetration.

Conclusions:

• Deep-B3 offers a powerful and efficient computational tool for predicting blood-brain barrier permeability.
• The framework has the potential to significantly aid in the central nervous system drug development pipeline.
• A publicly accessible web server, source code, and dataset are available to facilitate research and application.