Multitask Learning for Membrane Permeability Prediction across Assays with Prediction Reliability Assessment

Area of Science:

• Computational chemistry
• Pharmacokinetics
• Machine learning

Background:

• Membrane permeability is crucial for drug exposure and distribution.
• Existing in silico models are limited by uneven experimental data availability across different assay systems.
• Predictive models need to be robust and reliable for effective drug development.

Purpose of the Study:

• To develop a multitask graph convolutional neural network (MT-GCN) for predicting apparent permeability (Papp) across multiple experimental assays.
• To benchmark MT-GCN against other machine learning models like Random Forest (RF) and single-task GCN (ST-GCN).
• To introduce an applicability domain framework for quantifying prediction confidence.

Main Methods:

• Curated five apparent permeability (Papp) datasets: Caco-2, MDCK, RRCK, LLC-PK1, and PAMPA.
• Developed and trained a multitask graph convolutional neural network (MT-GCN).
• Implemented an applicability domain framework using ensemble dispersion, learned-space similarity, and local-consistency metrics to generate reliability scores.

Main Results:

• MT-GCN achieved the best performance for Caco-2 and MDCK assays (R² values of 0.503 and 0.500).
• MT-GCN demonstrated superior data efficiency, outperforming RF and ST-GCN in data-limited scenarios.
• The applicability domain framework provided reliable scores consistent with observed errors, enabling accuracy control.

Conclusions:

• Integrating multitask learning with applicability domain assessment enhances the precision and practicality of permeability predictions across assays.
• The developed MT-GCN model improves predictive performance and decision confidence, especially for data-limited endpoints.
• This approach offers a more robust tool for in silico drug permeability prediction in drug discovery.