A Comprehensive Machine Learning Framework for Predicting Blood-Brain Barrier Permeability Enhanced by Ant Colony...

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Neuroscience

Background:

• The blood-brain barrier (BBB) limits drug delivery to the central nervous system (CNS).
• Accurate in silico prediction of BBB permeability is crucial for efficient drug discovery.
• Developing robust predictive models is essential to overcome BBB challenges.

Purpose of the Study:

• To develop and optimize a machine learning framework for predicting BBB permeability.
• To enhance a Deep Neural Network (DNN) model using Ant Colony Optimization (ACO).
• To provide an accurate tool for virtual screening in CNS drug discovery.

Main Methods:

• Curated a dataset of 1957 compounds with 212 molecular descriptors.
• Employed a scaffold-based split for dataset partitioning.
• Benchmarked baseline models including Random Forest and a standard DNN.
• Utilized ACO for systematic hyperparameter optimization of the DNN.

Main Results:

• Random Forest achieved a baseline AUC of 0.913.
• The ACO algorithm identified an optimized DNN architecture.
• The ACO-Optimized DNN achieved a state-of-the-art AUC-ROC of 0.921.
• The optimized model outperformed all baseline classifiers.

Conclusions:

• Metaheuristic optimization (ACO) offers advantages over traditional hyperparameter tuning.
• The ACO-Optimized DNN demonstrates stability and strong generalization capabilities.
• The developed methodology is a generalizable template for complex model optimization in drug discovery.