ChemXTree: A Feature-Enhanced Graph Neural Network-Neural Decision Tree Framework for ADMET Prediction

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Deep learning (DL) accelerates drug discovery by predicting molecular properties.
• Efficiently using high-dimensional molecular data remains a challenge.

Purpose of the Study:

• Introduce ChemXTree, a novel graph-based model for improved molecular property prediction.
• Address the challenge of utilizing rich, high-dimensional information in drug discovery.

Main Methods:

• Developed ChemXTree, a graph-based model incorporating a Gate Modulation Feature Unit (GMFU).
• Integrated a neural decision tree (NDT) in the output layer for enhanced feature processing.
• Evaluated performance on MoleculeNet and eight additional drug databases.

Main Results:

• ChemXTree demonstrated superior performance, matching or exceeding state-of-the-art models.
• Visualization confirmed improved separation between substrates and nonsubstrates in the latent space.
• Achieved significant improvements in predictive accuracy for drug discovery tasks.

Conclusions:

• ChemXTree offers a promising approach for drug discovery by combining advanced feature extraction and neural decision trees.
• The model shows potential for optimizing molecular properties and advancing predictive accuracy.
• Opens new avenues for research in machine learning-driven drug development.