Can We Quickly Learn to "Translate" Bioactive Molecules with Transformer Models?

Area of Science:

• Medicinal Chemistry
• Machine Learning
• Computational Drug Design

Background:

• Exploring vast chemical spaces for druglike molecules is a major challenge in drug design.
• Combinatorial explosion of molecular modifications limits efficient drug discovery.
• Machine learning (ML) offers potential solutions for navigating complex chemical landscapes.

Purpose of the Study:

• To apply transformer models, initially developed for machine translation, to facilitate meaningful exploration of chemical space in drug design.
• To enable ML models to learn context-dependent molecular transformations relevant to medicinal chemistry.
• To demonstrate the utility of transformer models in generating novel bioactive molecules.

Main Methods:

• Transformer models were trained on pairs of bioactive molecules from the ChEMBL database.
• The models learned to perform medicinal-chemistry-meaningful transformations, including novel ones.
• Retrospective analysis involved testing models on ligands targeting COX2, DRD2, and HERG proteins.

Main Results:

• Trained transformer models successfully generated molecules identical or highly similar to known active ligands.
• Model performance was validated even when trained on limited data for specific protein targets.
• The models demonstrated the ability to learn and apply unseen molecular transformations.

Conclusions:

• Transformer models can effectively "translate" known active molecules into novel candidates for specific protein targets.
• This approach significantly aids human experts in hit expansion and accelerates the drug design process.
• The study highlights the potential of adapting natural language processing techniques for molecular discovery.