Assay2Mol: large language model-based drug design using BioAssay context

TL;DR

Assay2Mol introduces a large language model–driven workflow that transforms unstructured PubChem BioAssay descriptions into context-rich prompts for de novo molecule generation, bypassing the need for protein structures. By retrieving and summarizing relevant BioAssays, the approach grounds molecule design in experimental context and counterscreen information, improving docking performance and drug-likeness while enabling safety-oriented optimization (e.g., hERG). Across CrossDocked benchmarks, Assay2Mol outperforms several structure-based baselines and demonstrates robust performance even in zero-shot scenarios, provided BioAssay relevance is leveraged. The work highlights both the promise and caveats of context-aware, text-guided chemistry, including data quality, open-access limitations, and the need for wet-lab validation to confirm in silico findings.

Abstract

Scientific databases aggregate vast amounts of quantitative data alongside descriptive text. In biochemistry, molecule screening assays evaluate candidate molecules' functional responses against disease targets. Unstructured text that describes the biological mechanisms through which these targets operate, experimental screening protocols, and other attributes of assays offer rich information for drug discovery campaigns but has been untapped because of that unstructured format. We present Assay2Mol, a large language model-based workflow that can capitalize on the vast existing biochemical screening assays for early-stage drug discovery. Assay2Mol retrieves existing assay records involving targets similar to the new target and generates candidate molecules using in-context learning with the retrieved assay screening data. Assay2Mol outperforms recent machine learning approaches that generate candidate ligand molecules for target protein structures, while also promoting more synthesizable molecule generation.

Figures (10)

• Figure 1: The Assay2Mol workflow. A chemist provides a target description, which is used to retrieve BioAssays from the pre-embedded vector database. After filtering for relevance, the BioAssays are summarized by an LLM. The BioAssay ID is then used to retrieve experimental tables. The final molecule generation prompt is formed by combining the description, summarization, and selected test molecules with associated test outcomes, enabling the LLM to generate relevant active molecules. Icons are from Flaticon.com and svgrepo.com
• Figure 2: Docked binding poses of generated molecules without (left) and with (right) BioAssay context. With BioAssay context, ChatGPT 4o generates a molecule with three hydrogen bonds to the GRK4 pocket residue MET-267, improving the docking score.
• Figure 3: Change in predicted hERG score and docking score between initial and optimized molecules for three proteins. The up arrow indicates the docking score decreases and the down arrow indicates it increases. The length of the arrow (top-left) serves as a scale bar, representing an increase of 1 score unit (kcal/mol) from Vina Dock.
• Figure 4: Distribution of docking scores of different relevance level groups.
• Figure 5: Average docking score under different $max\_assay\_num$