Prediction of Permeability and Efflux Using Multitask Learning.

Area of Science:

• Computational Chemistry
• Pharmacology
• Machine Learning

Background:

• * In silico prediction of cell membrane permeability is vital for drug discovery, impacting efficacy, bioavailability, and pharmacokinetics.
• * Existing public datasets for permeability prediction are limited in size and consistency.
• * Caco-2 and MDCK cell lines are standard experimental models for drug permeability and efflux assays.

Purpose of the Study:

• * To investigate the efficacy of multitask graph neural networks (MTL) for predicting cell membrane permeability endpoints.
• * To benchmark MTL models against single-task approaches using a large, harmonized dataset.
• * To evaluate the impact of molecular features on prediction accuracy.

Main Methods:

• * Development and training of multitask graph neural networks on a proprietary dataset (>10K compounds).
• * Benchmarking MTL models against single-task models.
• * Evaluation of models on an external public dataset to assess applicability domain.
• * Incorporation of molecular features (pKa, LogD) into MTL models.

Main Results:

• * Multitask learning models demonstrated higher accuracy than single-task models by leveraging shared information across endpoints.
• * Augmenting MTL models with molecular features, specifically pKa and LogD, significantly improved prediction accuracy for both permeability and efflux.
• * The study provides benchmarking results and guidelines for optimal validation strategies in multitask learning.

Conclusions:

• * Multitask graph neural networks offer a powerful approach for accurate in silico prediction of cell membrane permeability.
• * Integrating physicochemical properties like pKa and LogD enhances the predictive performance of these models.
• * The findings support the use of MTL for robust drug permeability and efflux prediction in drug discovery pipelines.