Deep Transferable Compound Representation across Domains and Tasks for Low Data Drug Discovery

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• cheminformatics

Background:

• Small molecule drug discovery faces challenges in identifying candidates with optimal pharmacological activity, ADME properties, and low toxicity.
• Machine learning (ML) significantly contributes to drug discovery but often requires large datasets, which are scarce in the lead optimization phase.
• Existing methods overlook differences in compound distribution and target groups between source and target assays.

Purpose of the Study:

• To develop a novel ML approach for drug discovery that overcomes data limitations in lead optimization.
• To leverage knowledge from auxiliary assays (source data) to improve predictions for target assays (new compounds).
• To address the domain shift issue between source and target assays with varying compound distributions.

Main Methods:

• Utilized Graph Convolutional Networks (GCN) for feature extraction from molecular structures.
• Employed Adversarial Domain Adaptation (ADA) to bridge the gap between source and target assay data distributions.
• Integrated GCN with ADA to create a novel architecture for knowledge transfer.

Main Results:

• The proposed GCN-ADA approach demonstrated effectiveness in transferring knowledge from source to target datasets.
• Evaluated on Tox21, ToxCast, SIDER, HIV, and BACE datasets, showing improved predictive performance.
• Successfully addressed the challenge of low data availability in lead optimization for small molecule drug discovery.

Conclusions:

• The developed GCN-ADA model offers a robust solution for data-scarce scenarios in drug discovery.
• This approach enhances the prediction of compound properties by effectively utilizing related auxiliary data.
• The findings pave the way for more efficient and accurate small molecule drug development.