Decoding the Blood-Brain Barrier: Innovative and Scalable Open-Source Machine Learning Model for Drug Permeability

Area of Science:

• Computational chemistry
• Pharmacology
• Drug discovery

Background:

• The blood-brain barrier (BBB) is crucial for drug development, but its assessment is resource-intensive.
• Existing in vitro and in vivo methods for measuring BBB transport are costly and time-consuming.
• Developing efficient methods to predict BBB permeability is essential for accelerating drug discovery.

Purpose of the Study:

• To develop and validate reliable machine learning models for assessing blood-brain barrier (BBB) permeability.
• To introduce a novel approach using classification labels as additional descriptors in regression tasks.
• To provide publicly available models for scientists to use in their research.

Main Methods:

• Utilized extensive datasets to train machine learning models for BBB permeability prediction.
• Developed both classification and regression models to assess BBB properties.
• Implemented a novel strategy of incorporating classification labels into regression models.

Main Results:

• The best BBB classification model achieved an ROC-AUCcv of 0.963.
• The top BBB regression model attained R2 = 0.954, Q2 = 0.728, and RMSEcv = 0.321.
• Models demonstrated strong generalization, with validation metrics closely matching cross-validation results.

Conclusions:

• The novel approach of using classification labels as descriptors significantly enhanced regression model performance.
• The developed regression model is more robust due to an expanded, diverse training dataset.
• The study provides publicly accessible machine learning models for BBB permeability assessment, aiding scientific research.