Reliable Apparent Permeability Prediction Through Benchmarking, Stability, Uncertainty Quantification, and...

Area of Science:

• Computational chemistry and cheminformatics
• Pharmacokinetics and drug discovery
• Machine learning in drug development

Background:

• Apparent permeability (P app) is crucial for predicting drug absorption and pharmacokinetic modeling.
• In vitro experiments for P app determination are time-consuming and expensive.
• Accurate machine learning (ML) models can streamline drug absorption prediction.

Purpose of the Study:

• To compare different feature representations for predicting log P app values using ML.
• To identify key molecular descriptors influencing drug permeability.
• To quantify prediction uncertainty using a model-agnostic method.

Main Methods:

• Four distinct molecular feature representations were evaluated.
• Three regression models were employed to predict log P app.
• SHAP values were analyzed for descriptor influence interpretation.
• Jackknife+ method was used for confidence interval estimation.

Main Results:

• The best ML model achieved an R 2 of 0.68 and RMSE of 0.42 on test data.
• Octanol-water partition coefficient and basic atom count significantly influenced predictions.
• Specific molecular substructures were identified as positively or negatively impacting permeability.
• Confidence intervals demonstrated high reliability, covering >95% of original and predicted log P app values.

Conclusions:

• The developed ML model offers a robust and informed approach to P app prediction.
• This method can significantly reduce reliance on in vitro permeability assays.
• The findings contribute to more efficient drug discovery and development pipelines.