A Comparative Study of QSPR Methods on a Unique Multitask PAMPA Data Set.

Area of Science:

• Computational chemistry
• Drug discovery
• Pharmacokinetics

Background:

• Predicting drug permeability across biological membranes is crucial for drug development.
• The parallel artificial membrane permeability assay (PAMPA) is a common in vitro method to assess passive permeability.
• Developing accurate predictive models for membrane permeability is an ongoing challenge.

Purpose of the Study:

• To create a comprehensive dataset of drug permeability across six different artificial membranes.
• To systematically evaluate various molecular descriptors and machine learning models for predicting passive membrane permeability.
• To analyze the trade-off between predictive accuracy and model interpretability.

Main Methods:

• Generated a dataset of 143 drug and drug candidate molecules.
• Performed in vitro parallel artificial membrane permeability assays (PAMPA) using six distinct model membranes.
• Assessed predictive performance using models from linear regression to deep learning (transformer architecture).

Main Results:

• Expert-designed physicochemical descriptors demonstrated superior performance compared to deep learning representations in this study.
• The study highlights challenges in model interpretability with advanced machine learning approaches.
• Novel insights into membrane-specific permeability profiles were gained.

Conclusions:

• Physicochemical descriptors are more suitable for permeability prediction with limited sample sizes.
• The developed dataset and findings offer valuable resources for drug discovery and pharmacokinetic modeling.
• Balancing predictive power and interpretability remains a key consideration in computational drug design.