HelixADMET: a robust and endpoint extensible ADMET system incorporating self-supervised knowledge transfer.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in pharmacology

Background:

• Accurate absorption, distribution, metabolism, excretion, and toxicity (ADMET) predictions are crucial for early-stage drug candidate screening.
• Existing ADMET systems often exhibit weak extrapolation abilities, particularly for molecules with unobserved scaffolds due to limited labeled data.
• Current systems typically offer fixed endpoints, lacking the flexibility to meet diverse drug research and development requirements.

Purpose of the Study:

• To develop a robust and endpoint-extensible ADMET system named HelixADMET (H-ADMET).
• To address the limitations of weak extrapolation and lack of customization in existing ADMET prediction tools.

Main Methods:

• H-ADMET utilizes a self-supervised learning approach to generate a robust pre-trained model.
• The model is subsequently fine-tuned using a multi-task and multi-stage framework.
• This framework facilitates knowledge transfer across various ADMET endpoints, auxiliary tasks, and self-supervised tasks.

Main Results:

• H-ADMET demonstrated an overall improvement of 4% compared to existing ADMET systems on comparable endpoints.
• The pre-trained model within H-ADMET can be fine-tuned to create new, customized ADMET endpoints.
• These capabilities effectively address various demands in drug research and development.

Conclusions:

• H-ADMET offers a significant advancement in ADMET prediction accuracy and flexibility.
• The system's extensible nature and customizable endpoints empower researchers in drug discovery.
• H-ADMET provides a valuable tool for screening drug candidates and meeting evolving research needs.