ADMET property prediction via multi-task graph learning under adaptive auxiliary task selection.

Area of Science:

• Computational Chemistry
• Medicinal Chemistry
• Drug Discovery

Background:

• Accurate prediction of absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties is crucial for lead optimization in drug discovery.
• Classical single-task learning (STL) effectively predicts individual ADMET endpoints but requires extensive labeled data.
• Multi-task learning (MTL) can predict multiple ADMET endpoints with less data but faces challenges in task synergy and identifying key molecular substructures.

Purpose of the Study:

• To develop a novel multi-task graph learning framework (MTGL-ADMET) for predicting multiple ADMET properties of drug-like small molecules.
• To address limitations of existing STL and MTL methods, particularly in ensuring task synergy and interpretability.
• To provide a transparent method for identifying crucial molecular substructures influencing ADMET properties.

Main Methods:

• Elaboration of a multi-task graph learning framework (MTGL-ADMET) based on the "one primary, multiple auxiliaries" paradigm.
• Integration of status theory and maximum flow algorithms for effective auxiliary task selection.
• Development of a primary-task-centric MTL model with integrated modules for enhanced prediction and interpretability.

Main Results:

• MTGL-ADMET significantly outperforms existing single-task learning (STL) and multi-task learning (MTL) methods in predicting multiple ADMET endpoints.
• The framework provides a transparent analysis of crucial molecular substructures impacting ADMET properties.
• Demonstrated improved accuracy and efficiency in predicting ADMET profiles of drug-like small molecules.

Conclusions:

• MTGL-ADMET offers a superior approach for predicting multiple ADMET properties compared to conventional methods.
• The framework enhances the interpretability of ADMET predictions by highlighting key molecular substructures.
• This work is expected to accelerate lead compound identification and optimization in the drug discovery pipeline.