Predicting In Vivo Compound Brain Penetration Using Multi-task Graph Neural Networks.

Area of Science:

• Computational chemistry
• Pharmacology
• Machine learning in drug discovery

Background:

• Estimating compound brain penetration is crucial in drug discovery for efficacy and safety.
• In vivo studies measuring brain-to-blood concentration ratios (Kp) are standard but resource-intensive.

Purpose of the Study:

• To develop machine learning models for predicting in vivo compound brain penetration (LogKp) from chemical structures.
• To evaluate the utility of in vitro data as auxiliary tasks in multi-task learning for improved predictions.

Main Methods:

• Development of multi-task graph neural network (MT-GNN) models incorporating in vitro data.
• Comparison of MT-GNN performance against single-task (ST) models trained only on in vivo data.
• Prospective validation of the best-performing model on unseen compounds.

Main Results:

• MT-GNN models significantly outperformed ST models in predicting LogKp.
• The best MT-GNN achieved a R-squared of 0.42 and MAE of 0.39 on prospective data.
• Binary classification of brain-penetrant vs. non-penetrant compounds yielded a Matthew's correlation coefficient of 0.66.

Conclusions:

• Incorporating in vitro assay data as auxiliary tasks in MT-GNNs improves in vivo brain penetration predictions.
• This approach enhances prospective compound prioritization in drug discovery pipelines.
• The developed models offer a valuable tool for efficient assessment of brain penetration potential.