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

Caco-2 monolayer assays are widely used to assess drug absorption, yet existing predictive approaches mainly rely on predefined features and 2D representations, potentially overlooking key 3D structural information. This study evaluates models leveraging 3D molecular representations (3D-NN), lower-dimensional approaches (1D, 2D), and predefined features through four tasks: benchmarking on merged literature data sets, scaffold-based splitting, comparison with published models, and external testing on transporter-mediated compounds. Our findings demonstrated that 3D-NN can independently extract more meaningful features for the permeability task, achieving superior generalizability across scaffolds and performing competitively with task-specific literature models. Additionally, its embeddings showed clear separation between low- and high-permeability compounds, highlighting promise for future applications as larger data sets emerge. However, analyses on transporter-mediated permeability were inconclusive due to current data limitations, indicating the need for more representative data sets to fully evaluate model performance in this aspect.