Autonomous chemical research with large language models

Written for the Research library (claude-opus-5-5) from the paper text. Original — nature.com/articles/s41586-023-06792-0.pdf

Summary

Boiko and colleagues examine how far large language models can go in carrying out real chemistry. They ask what LLMs can contribute to the scientific process, how much autonomy is achievable, and whether an agent's choices can be inspected. Their answer is Coscientist, an agent built around a central GPT-4 "Planner." The Planner turns a plain-text request, such as running several Suzuki reactions, into steps and calls supporting tools: an LLM-driven web searcher, a sandboxed code runner, a documentation retriever, and an automation layer that sends instructions to lab hardware.

The system was tested on six tasks. These range from planning syntheses of known compounds and learning hardware APIs to steering a liquid handler, combining several instruments and data sources, and optimizing reactions from previously collected results. In a synthesis-planning benchmark, connecting GPT-4 to web search clearly reduced invented or unworkable routes compared with standalone models. Coscientist also optimized conditions for palladium-catalysed cross-couplings. The work matters because it shows a general-purpose LLM, linked to search and robotics, moving from literature lookup to physical experiments while keeping its reasoning readable.

Key points

• Architecture: The Planner is a GPT-4 chat instance with four commands: GOOGLE, PYTHON, DOCUMENTATION and EXPERIMENT. Code runs inside an isolated Docker container, and the model can repair its own code when errors occur.
• Supported automation interfaces: Coscientist controls Opentrons equipment via its Python API and also writes instructions in Emerald Cloud Lab's (ECL) Symbolic Lab Language (SLL). The hardware setup included a liquid handler, a heater–shaker module and UV-Vis measurements.
• Synthesis benchmark:
  • Seven target compounds, including aspirin, acetaminophen, ibuprofen, nitroaniline and phenolphthalein, were scored on a 1–5 scale; scores below 3 counted as failure.
  • Baselines were GPT-3.5, GPT-4, Claude 1.3 and Falcon-40B-Instruct.
  • search-gpt-4 achieved top scores for four compounds and was the only system to reach an acceptable score for ibuprofen.
• Suggested extensions: Adding reaction databases (Reaxys, SciFinder) or prompting strategies such as ReAct, Chain of Thought and Tree of Thoughts.
• Availability: For safety reasons, the full code and prompts are withheld. A simplified implementation and the generated outputs are available at github.com/gomesgroup/coscientist.