Assay2Mol: Large Language Model-based Drug Design Using BioAssay Context

Area of Science:

• Biochemistry
• Computational Drug Discovery
• Bioinformatics

Background:

• Scientific databases contain vast quantitative and text data crucial for drug discovery.
• Biochemical screening assays evaluate molecule responses against disease targets.
• Unstructured text in assay data is largely untapped for drug discovery.

Purpose of the Study:

• To present Assay2Mol, a large language model-based workflow.
• To utilize existing biochemical screening assays for early-stage drug discovery.
• To overcome limitations of unstructured text in assay data.

Main Methods:

• Assay2Mol retrieves existing assay records for targets similar to a new target.
• It employs in-context learning with retrieved screening data.
• Generates candidate molecules based on assay information.

Main Results:

• Assay2Mol outperforms recent machine learning approaches for generating candidate ligand molecules.
• The workflow promotes the generation of more synthesizable molecules.
• Effectively capitalizes on existing biochemical screening assay data.

Conclusions:

• Assay2Mol offers a novel approach to early-stage drug discovery.
• Leveraging unstructured assay data with LLMs can accelerate the identification of drug candidates.
• The method enhances molecule generation for drug discovery campaigns.