Advancing ADMET prediction through multiscale fragment-aware pretraining with MSformer-ADMET

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• Absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties are crucial for drug candidate success.
• Early prediction of ADMET is vital to reduce costs and accelerate drug development.
• Existing deep learning models often lack interpretability and generalization.

Purpose of the Study:

• To adapt and specialize the MSformer framework for predicting ADMET properties.
• To evaluate the performance of the adapted model, MSformer-ADMET, across diverse ADMET tasks.
• To investigate the interpretability of MSformer-ADMET for understanding structure-property relationships.

Main Methods:

• Developed MSformer-ADMET, a fragmentation-based deep learning model.
• Fine-tuned MSformer-ADMET on 22 ADMET tasks from the Therapeutics Data Commons (TDC).
• Compared MSformer-ADMET against traditional smiles-based and graph-based models.
• Performed interpretability analyses using attention distributions and fragment mappings.

Main Results:

• MSformer-ADMET achieved superior performance on a wide range of ADMET endpoints.
• The model consistently outperformed conventional molecular representation methods.
• Interpretability analyses identified key structural fragments associated with specific ADMET properties.
• Demonstrated effective performance in both classification and regression ADMET tasks.

Conclusions:

• MSformer-ADMET is a highly effective and broadly applicable model for ADMET prediction.
• The fragmentation-based approach enhances generalization and interpretability in molecular property prediction.
• MSformer-ADMET offers transparent insights into structure-property relationships, aiding drug discovery.