ChemCrow
GPT-4-driven chemistry agent combining chain-of-thought reasoning with 18 expert-designed chemistry tools (retrosynthesis, molecule property lookup, web/literature search, reaction execution).
| Affiliation | EPFL / University of Rochester (Bran et al.) |
| First introduced | 2023-04 (arXiv preprint); Nature Machine Intelligence 2024 |
| Lifecycle stages | Experiment design, Analysis |
| Autonomy level | Semi-autonomous (closed-loop with checkpoints) |
| Domain focus | Chemistry (organic synthesis, reaction planning) |
| Availability | Open source |
Approach
GPT-4 driven agent that combines chain-of-thought reasoning with 18 expert-designed chemistry tools. The agent plans, calls tools, and integrates results to complete chemistry tasks.
Validation
Demonstrated on synthesis planning tasks (e.g., insect repellent, organocatalyst, novel chromophore). Evaluated by expert chemists and an LLM grader against an unaugmented GPT-4 baseline.
Notable results
Autonomously planned and executed syntheses. Outperformed GPT-4 baseline on expert and automated chemistry evaluations per the Gao et al. Cell perspective citation.
Primary paper
Bran et al., “Augmenting large language models with chemistry tools,” Nat. Mach. Intell. 2024.
Other references
- Gao et al., Cell Perspective 2024 (cites ChemCrow as a Level 1 agent example)