Machine learning based dynamic consensus model for predicting blood-brain barrier permeability.

Area of Science:

• Computational chemistry
• Pharmacology
• Bioinformatics

Background:

• The blood-brain barrier (BBB) protects the brain but complicates drug delivery.
• Existing in silico models for BBB permeability prediction suffer from low reliability and high false positive rates due to small, imbalanced datasets.

Purpose of the Study:

• To develop and compare machine learning and deep learning models for accurate prediction of BBB permeability.
• To address the challenge of class imbalance in datasets used for BBB permeability prediction.

Main Methods:

• Curated a dataset of 8153 compounds (BBB permeable and non-permeable).
• Generated molecular descriptors and fingerprints for feature engineering.
• Applied machine learning (XGBoost, Random Forest, Extra-tree) and deep learning (deep neural network) models.
• Utilized three balancing techniques to mitigate class imbalance.
• Developed a dynamic consensus model for enhanced prediction confidence.

Main Results:

• The deep neural network model trained on balanced MACCS fingerprints achieved the highest performance.
• Achieved 97.8% accuracy and a 0.98 ROC-AUC score.
• The dynamic consensus model demonstrated higher confidence scores on a benchmark dataset.

Conclusions:

• Deep learning, particularly a deep neural network with MACCS fingerprints and balancing techniques, offers a highly accurate method for predicting BBB permeability.
• The developed models and consensus approach provide a reliable tool for in silico BBB permeability assessment, aiding drug discovery and development.