Predicting the Absorption Potential of Chemical Compounds Through a Deep Learning Approach

Area of Science:

• Pharmacology
• Computational Chemistry
• Drug Discovery

Background:

• The Caco-2 cell line assay is a standard in-vitro method for predicting oral drug absorption.
• In-silico models enhance high-throughput screening but often rely on dataset-specific handcrafted features, leading to overfitting.
• Deep Neural Networks (DNNs) can generate robust, high-level features from raw data, improving model generalization.

Purpose of the Study:

• To develop and validate a DNN-based binary classifier for predicting Caco-2 cellular permeability.
• To demonstrate the superiority of DNN-generated high-level features over handcrafted features for permeability prediction.

Main Methods:

• A DNN model was constructed using in-vitro Caco-2 apparent permeability data for 663 chemical compounds.
• Two hundred nine molecular descriptors were utilized as input for DNN feature generation.
• Dropout regularization and Rectified Linear Unit (ReLU) activation were employed to address overfitting and vanishing gradient problems, respectively.

Main Results:

• The DNN model successfully classified Caco-2 permeability.
• DNN-generated high-level features showed greater robustness compared to handcrafted features.
• The model demonstrated effective prediction of cellular permeability for structurally diverse compounds.

Conclusions:

• DNNs offer a powerful approach for developing generalized in-silico models for drug permeability prediction.
• The proposed DNN-based classifier provides a more reliable method for assessing drug absorption potential.
• This approach can significantly enhance the efficiency and accuracy of drug discovery pipelines.