Quantum-Informed Molecular Representation Learning Enhancing ADMET Property Prediction.

Area of Science:

• Computational Chemistry
• Machine Learning
• Drug Discovery

Background:

• Molecular representation learning is crucial for predicting Absorption, Distribution, Metabolism, Excretion, Toxicity, and Disposal (ADMET) properties.
• Existing pretraining tasks for molecular representations have limitations.
• Graph transformers show promise for molecular property prediction.

Purpose of the Study:

• To develop and evaluate novel pretraining tasks for graph transformers to enhance molecular representation learning.
• To improve the accuracy of ADMET property prediction using enhanced molecular representations.
• To identify more effective training targets beyond traditional methods.

Main Methods:

• Investigated various pretraining tasks, incorporating data from 2D molecular descriptors to quantum chemistry simulations.
• Implemented supervised pretraining tasks and multitask learning with a shared encoder.
• Utilized the Therapeutics Data Commons dataset for evaluating 22 ADMET tasks.

Main Results:

• The proposed pretraining strategy significantly outperforms conventional methods.
• Achieved state-of-the-art performance in 7 out of 22 ADMET prediction tasks.
• Demonstrated the effectiveness of integrating diverse data sources into pretraining.

Conclusions:

• Novel pretraining tasks enhance molecular representation learning for graph transformers.
• The approach offers a scalable solution for improving ADMET property prediction.
• This work represents a significant advancement in leveraging data for drug discovery and development.