Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 46 of the 46 references it cites were found in a registry or an index.

Autonomous chemical research with large language models

Open access
Boiko, Daniil A., MacKnight, Robert, Kline, Ben, Gomes, Gabe · 2023 · Nature

Abstract

Transformer-based large language models are making significant strides in various fields, such as natural language processing 1–5 , biology 6,7 , chemistry 8–10 and computer programming 11,12 . Here, we show the development and capabilities of Coscientist, an artificial intelligence system driven by GPT-4 that autonomously designs, plans and performs complex experiments by incorporating large language models empowered by tools such as internet and documentation search, code execution and experimental automation. Coscientist showcases its potential for accelerating research across six diverse tasks, including the successful reaction optimization of palladium-catalysed cross-couplings, while exhibiting advanced capabilities for (semi-)autonomous experimental design and execution. Our findings demonstrate the versatility, efficacy and explainability of artificial intelligence systems like Coscientist in advancing research.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1038/s41586-023-06792-0

Cites

46 of 46 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 The system’s architecture. a, Coscientist is composed of multiple modules that exchange messages. Boxes with blue background represent LLM modules, the Planner module is shown in green, and the input prompt is in red. White boxes represent modules that do not use LLMs. b, Types of experiments performed to demonstrate the capabilities when using individual modules or their combinations. c, Image of the experimental setup with a liquid handler. UV-Vis, ultraviolet visible.
Fig. 1 The system’s architecture. a, Coscientist is composed of multiple modules that exchange messages. Boxes with blue background represent LLM modules, the Planner module is shown in green, and the input prompt is in red. White boxes represent modules that do not use LLMs. b, Types of experiments performed to demonstrate the capabilities when using individual modules or their combinations. c, Image of the experimental setup with a liquid handler. UV-Vis, ultraviolet visible.CC BY · PubMed Central
Fig. 2 Coscientist’s capabilities in chemical synthesis planning tasks. a, Comparison of various LLMs on compound synthesis benchmarks. Error bars represent s.d. values. b, Two examples of generated syntheses of nitroaniline. c, Two example of generated syntheses of ibuprofen. UV, ultraviolet.
Fig. 2 Coscientist’s capabilities in chemical synthesis planning tasks. a, Comparison of various LLMs on compound synthesis benchmarks. Error bars represent s.d. values. b, Two examples of generated syntheses of nitroaniline. c, Two example of generated syntheses of ibuprofen. UV, ultraviolet.CC BY · PubMed Central
Fig. 3 Overview of documentation search. a, Prompt-to-code through ada embedding and distance-based vector search. b, Example of code for using OT-2’s heater–shaker module. c, Prompt-to-function/prompt-to-SLL (to symbolic laboratory language) through supplementation of documentation. d, Example of valid ECL SLL code for performing high-performance liquid chromatography (HPLC) experiments.
Fig. 3 Overview of documentation search. a, Prompt-to-code through ada embedding and distance-based vector search. b, Example of code for using OT-2’s heater–shaker module. c, Prompt-to-function/prompt-to-SLL (to symbolic laboratory language) through supplementation of documentation. d, Example of valid ECL SLL code for performing high-performance liquid chromatography (HPLC) experiments.CC BY · PubMed Central
Fig. 4 Robotic liquid handler control capabilities and integration with analytical tools. a, Overview of Coscientist’s configuration. b, Drawing a red cross. c, Colouring every other row. d, Drawing a yellow rectangle. e, Drawing a blue diagonal.
Fig. 4 Robotic liquid handler control capabilities and integration with analytical tools. a, Overview of Coscientist’s configuration. b, Drawing a red cross. c, Colouring every other row. d, Drawing a yellow rectangle. e, Drawing a blue diagonal.CC BY · PubMed Central
Fig. 5 Cross-coupling Suzuki and Sonogashira reaction experiments designed and performed by Coscientist. a, Overview of Coscientist’s configuration. b, Available compounds (DMF, dimethylformamide; DiPP, 2,6-diisopropylphenyl). c, Liquid handler setup. d, Solving the synthesis problem. e, Comparison of reagent selection performance with a large dataset. f, Comparison of reagent choices across multiple runs. g, Overview of justifications made when selecting various aryl halides. h, Frequency of visited URLs. i, Total ion current (TIC) chromatogram of the Suzuki reaction mixture (top panel) and the pure standard, mass spectra at 9.53 min (middle panel) representing the expected reaction product and mass spectra of the pure standard (bottom panel). j, TIC chromatogram of the Sonogashira reaction mixture (top panel) and the pure standard, mass spectra at 12.92 min (middle panel) representing the expected reaction product and mass spectra of the pure standard (bottom panel). Rel., relative.
Fig. 5 Cross-coupling Suzuki and Sonogashira reaction experiments designed and performed by Coscientist. a, Overview of Coscientist’s configuration. b, Available compounds (DMF, dimethylformamide; DiPP, 2,6-diisopropylphenyl). c, Liquid handler setup. d, Solving the synthesis problem. e, Comparison of reagent selection performance with a large dataset. f, Comparison of reagent choices across multiple runs. g, Overview of justifications made when selecting various aryl halides. h, Frequency of visited URLs. i, Total ion current (TIC) chromatogram of the Suzuki reaction mixture (top panel) and the pure standard, mass spectra at 9.53 min (middle panel) representing the expected reaction product and mass spectra of the pure standard (bottom panel). j, TIC chromatogram of the Sonogashira reaction mixture (top panel) and the pure standard, mass spectra at 12.92 min (middle panel) representing the expected reaction product and mass spectra of the pure standard (bottom panel). Rel., relative.CC BY · PubMed Central
Fig. 6 Results of the optimization experiments. a, A general reaction scheme from the flow synthesis dataset analysed in c and d. b, The mathematical expression used to calculate normalized advantage values. c, Comparison of the three approaches (GPT-4 with prior information, GPT-4 without prior information and GPT-3.5 without prior information) used to perform the optimization process. d, Derivatives of the NMA and normalized advantage values evaluated in c, left and centre panels. e, Reaction from the C–N cross-coupling dataset analysed in f and g. f, Comparison of two approaches using compound names and SMILES string as compound representations. g, Coscientist can reason about electronic properties of the compounds, even when those are represented as SMILES strings. DMSO, dimethyl sulfoxide.
Fig. 6 Results of the optimization experiments. a, A general reaction scheme from the flow synthesis dataset analysed in c and d. b, The mathematical expression used to calculate normalized advantage values. c, Comparison of the three approaches (GPT-4 with prior information, GPT-4 without prior information and GPT-3.5 without prior information) used to perform the optimization process. d, Derivatives of the NMA and normalized advantage values evaluated in c, left and centre panels. e, Reaction from the C–N cross-coupling dataset analysed in f and g. f, Comparison of two approaches using compound names and SMILES string as compound representations. g, Coscientist can reason about electronic properties of the compounds, even when those are represented as SMILES strings. DMSO, dimethyl sulfoxide.CC BY · PubMed Central
Extended Data Fig. 1 Using UV-Vis and liquid handler to solve food colouring identification problem. Guiding prompt in the third message is shown in bold. In the first message the user prompt is provided, then code for sample preparation is generated, resulting data is provided as NumPy array, which is then analysed to give the final answer.
Extended Data Fig. 1 Using UV-Vis and liquid handler to solve food colouring identification problem. Guiding prompt in the third message is shown in bold. In the first message the user prompt is provided, then code for sample preparation is generated, resulting data is provided as NumPy array, which is then analysed to give the final answer.CC BY · PubMed Central
Extended Data Fig. 2 Code, generated by Coscientist. The generated code can be split into the following steps: defining metadata for the method, loading labware modules, setting up the liquid handler, performing required reagent transfers, setting up the heater-shaker module, running the reaction, and turning the module off.
Extended Data Fig. 2 Code, generated by Coscientist. The generated code can be split into the following steps: defining metadata for the method, loading labware modules, setting up the liquid handler, performing required reagent transfers, setting up the heater-shaker module, running the reaction, and turning the module off.CC BY · PubMed Central
Extended Data Fig. 3 Additional results on comparison with Bayesian optimization. a, GPT-4 models compared with Bayesian optimization performed starting with different number of initial samples. b, Compound-by-compound comparison of differences between advantages.
Extended Data Fig. 3 Additional results on comparison with Bayesian optimization. a, GPT-4 models compared with Bayesian optimization performed starting with different number of initial samples. b, Compound-by-compound comparison of differences between advantages.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

7 more claims about this paper

Create a free account to see every claim other papers make about this one.

See all claims — free account

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 37 references

Research works printed but not linked: 3

  1. Brown, T. et al. in Advances in Neural Information Processing Systems Vol. 33 (eds Larochelle, H. et al.) 1877–1901 (Curran Associates, 2020).
  2. Data Mining. Mining of Massive Datasets (Cambridge Univ., 2011).
  3. Running experiments. Emerald Cloud Lab https://www.emeraldcloudlab.com/guides/runningexperiments (2023).

Cited by

See all 1,050

Metadata

Authors
Boiko, D. A., MacKnight, R., Kline, B., & Gomes, G.
Affiliations
Carnegie Mellon University (US)
Corresponding author
Gomes, Gabe
Venue
Nature
Published
2023-12-20
Volume
624
Issue
7992
Pages
570-578
ISSN
  • 0028-0836 (ISSN-L)
  • 1476-4687
OpenAlex
W4389991792
Cited by
1,050
References
37
Citations by year
2026: 458 · 2025: 401 · 2024: 186 · 2023: 5
FWCI
77.41
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
Funders
  • National Science Foundation (2221346, 2202693)
  • Carnegie Mellon University
Topics
  • Physical Sciences › Materials Science › Materials Chemistry › Machine Learning in Materials Science
  • Physical Sciences › Computer Science › Artificial Intelligence › Artificial Intelligence Applications
  • Life Sciences › Pharmacology, Toxicology and Pharmaceutics › Drug Discovery › Computational Drug Discovery Methods
Keywords
large language models, GPT-4, experimental automation, reaction condition optimization, palladium-catalyzed cross-coupling, program execution
MeSH terms
Artificial Intelligence, Automation, Catalysis, Documentation, Electric Power Supplies

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Autonomous chemical research with large language models. https://app.cochat.ai/papers/10.1038/s41586-023-06792-0
The paper
Boiko, D. A., MacKnight, R., Kline, B., & Gomes, G. (2023). Autonomous chemical research with large language models. Nature. https://doi.org/10.1038/s41586-023-06792-0

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.