Integrating (deep) machine learning and cheminformatics for predicting human intestinal absorption of small molecules

Area of Science:

• Pharmacology
• Computational Chemistry
• Drug Discovery

Background:

• The oral route is the most preferred for drug administration, making oral bioavailability a critical factor in pharmaceutical development.
• Human intestinal absorption (HIA) is a key determinant of oral bioavailability, necessitating accurate prediction methods.
• Developing robust models for HIA prediction can significantly streamline the drug discovery and development process.

Purpose of the Study:

• To develop and evaluate machine learning and deep learning models for predicting drug permeability at HIA.
• To establish a reliable computational tool for assessing oral bioavailability potential early in drug development.
• To curate a comprehensive dataset for training and validating predictive models of HIA.

Main Methods:

• A dataset of 2648 compounds was curated for training and validation.
• Five machine learning algorithms (including Random Forest and LightGBM) were trained using engineered molecular descriptors.
• Two deep learning models, Graph Convolutional Neural Network (GCNN) and Graph Attention Network (GAT), were developed using automated feature extraction.

Main Results:

• Random Forest achieved 87.71% accuracy on the test set and 81.43% on the external validation set.
• LightGBM achieved 86.04% accuracy on the test set and 77.30% on the external validation set.
• GCNN and GAT models demonstrated accuracies of 77.69% and 78.58% on the test set, respectively, and 79.29% and 79.42% on the external validation set.

Conclusions:

• Machine learning models, particularly Random Forest and LightGBM, show high accuracy in predicting drug permeability at HIA.
• Deep learning models (GCNN, GAT) also provide valuable predictive capabilities for HIA.
• The developed models and dataset, available on GitHub, can aid in the screening of oral drugs to enhance drug discovery efficiency.