Reliable Prediction of Caco-2 Permeability by Supervised Recursive Machine Learning Approaches.

Area of Science:

• Pharmacokinetics and Drug Metabolism
• Computational Chemistry
• In Silico Drug Discovery

Background:

• Caco-2 cell permeability assays exhibit high variability due to cell line heterogeneity and protocol differences.
• This variability limits the development of accurate predictive models for drug absorption.
• A need exists for robust computational methods to predict intestinal permeability.

Purpose of the Study:

• To develop a Quantitative Structure-Property Relationship (QSPR) model for predicting Caco-2 cell permeability.
• To establish a reliable and automated platform for early-stage drug discovery by assessing intestinal permeability.
• To overcome the limitations of traditional in vitro Caco-2 assays.

Main Methods:

• A QSPR approach was implemented on the KNIME analytical platform using a dataset of over 4900 diverse molecules.
• Random forest supervised recursive algorithms were employed for data cleaning and feature selection.
• A conditional consensus model combining regional and global random forest regression was developed.

Main Results:

• The developed QSPR models achieved Root Mean Square Error (RMSE) values between 0.43-0.51 across all validation sets.
• Blind prediction of 32 International Council for the Harmonization (ICH) recommended drugs demonstrated the model's applicability as a surrogate for in vitro Caco-2 assays.
• The model was validated for preliminary Biopharmaceutics Classification System (BCS)/Biomembrane Drug Disposition Classification System (BDDCS) class estimation.

Conclusions:

• The developed QSPR model and KNIME workflow provide a reliable, automated tool for predicting intestinal permeability.
• This platform can significantly aid in identifying promising drug candidates with high permeability during early drug discovery stages.
• The freely available KNIME workflow facilitates the prediction of new drug candidates.