source: sciencedaily ai: ai lets chemists design molecules by simply describing them

level: research

creating complex molecules usually demands years of experience and many strategic decisions. a team at epfl developed synthegy, a framework that combines large language models with traditional search algorithms. instead of generating structures directly, the ai acts as an evaluator. chemists provide a target molecule and a simple instruction, such as forming a ring early or avoiding protecting groups. standard retrosynthesis software then proposes many possible routes. each route is converted to text and reviewed by the language model, which scores how well it matches the instruction and explains its reasoning.

synthegy also tackles reaction mechanisms by breaking them into basic electron movements and exploring possibilities. the language model evaluates each step and steers the search toward chemically sensible pathways. researchers can add extra details like reaction conditions or hypotheses as text, refining the analysis. in a double-blind study with 36 chemists, the system's assessments agreed with human evaluations 71.2% of the time on average. larger models performed best, while smaller ones showed limited abilities. the framework can flag unnecessary protecting steps, judge reaction feasibility, and prioritize efficient solutions.

this approach positions language models as guides that interpret and refine computational results, rather than replacing human decision-making. chemists describe their goals in plain language and receive solutions that reflect their strategy. the work could speed up drug discovery, improve reaction design, and make advanced tools more accessible. the connection between synthesis planning and mechanisms is bridged through a unified natural language interface, allowing faster iteration on complex synthetic ideas.

why it matters: it shows how language models can make expert chemical reasoning more accessible, potentially accelerating drug discovery and materials design.


source: sciencedaily ai: ai lets chemists design molecules by simply describing them