Exploring the Effectiveness of Three-Dimensional Molecular Representations in Caco-2 Permeability Prediction

Area of Science:

• Pharmacokinetics and Drug Metabolism
• Computational Chemistry
• Biophysical Chemistry

Background:

• Caco-2 monolayer assays are standard for drug absorption prediction.
• Current methods often rely on 2D representations and predefined features, potentially missing crucial 3D structural insights.
• There is a need for advanced predictive models that incorporate 3D molecular information.

Purpose of the Study:

• To evaluate the performance of 3D neural networks (3D-NN) against traditional 1D, 2D, and feature-based models for predicting drug permeability.
• To assess the generalizability of 3D-NN across different chemical scaffolds and compare them with existing literature models.
• To explore the utility of 3D-NN embeddings for distinguishing between low- and high-permeability compounds.

Main Methods:

• Benchmarking various models (3D-NN, 1D, 2D, predefined features) on merged literature datasets.
• Utilizing scaffold-based splitting for robust evaluation of model generalizability.
• Comparing model performance against established literature models and external test sets, including transporter-mediated compounds.

Main Results:

• 3D-NN models independently extracted more relevant features for permeability prediction, demonstrating superior generalizability across scaffolds.
• 3D-NN performance was competitive with specialized literature models.
• Embeddings from 3D-NN showed clear separation between low- and high-permeability compounds, indicating potential for future applications.
• Analysis of transporter-mediated permeability was inconclusive due to data limitations.

Conclusions:

• 3D molecular representations offer significant advantages for predicting drug permeability using Caco-2 assays.
• 3D-NN models show promise for improving drug absorption prediction, especially with larger datasets.
• Further research with more comprehensive datasets is required to fully evaluate 3D-NN performance in predicting transporter-mediated permeability.