Elamur Research
☰
home / papers / 2025 / Accelerating scientific discovery with Co-Scientist

Accelerating scientific discovery with Co-Scientist

Written for the Research library (claude-opus-5-5) from the paper text. Original — arxiv.org/pdf/2502.18864

Summary

Modern research needs deep expertise in one field and also broad knowledge that spans several fields. One scientist can rarely have both. The authors present Co-Scientist, a multi-agent system built on Gemini. It takes a research goal written in plain language and proposes original, testable hypotheses with experimental plans. A Supervisor agent sends tasks through an asynchronous queue to specialised agents. These agents draft ideas, critique them, compare them head-to-head in tournaments and evolve the strongest ones. Web search and specialised models help ground the output, and scientists can steer the process through chat. The system was tested automatically on 15 open problems chosen by experts. Hypothesis quality kept improving as more inference-time compute was added, and Co-Scientist outperformed other agentic and reasoning baselines. The authors also checked its proposals in laboratory experiments across three biomedical tasks of increasing difficulty. This matters because it shows that an LLM-based system can help produce new findings confirmed in the lab, rather than only summarising existing literature.

Key points

  • Architecture: the Generation, Reflection, Ranking, Evolution, Proximity and Meta-review agents run a cycle of generating, debating and evolving ideas, coordinated by a Supervisor. The design does not depend on one particular model, so the backbone LLM can be replaced without retraining the framework.
  • Acute myeloid leukemia (AML): the system suggested drugs to repurpose, both alone and in combination. In vitro, they selectively killed cancer cells at clinically relevant concentrations.
  • Liver fibrosis: the system proposed new epigenetic targets. Tests in human hepatic organoids showed reduced fibrosis and hepatocyte regeneration, and one of the effective compounds is an FDA-approved drug.
  • Antimicrobial resistance: the system independently arrived at an unpublished finding about cf-PICIs. It proposed that these genetic elements use the tails of many different phages to spread to more bacterial host species.
  • Limitations: the system only uses open-access literature, so it misses paywalled work and negative results. It can hallucinate, and the lab validation so far is preliminary.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: AI & machine learning

Cited by

102 papers of our library cite it (of the 140673 whose bibliographies we read); 3 of them have a page here, 99 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.1038/s41586-026-10644-y · arXiv 2502.18864

@misc{gottweis2025accelerating,
  title = {{Accelerating scientific discovery with Co-Scientist}},
  author = {Gottweis, Juraj and Weng, Wei-Hung and Daryin, Alexander and Tu, Tao and Sirkovic, Petar and Myaskovsky, Artiom and Glowaty, Grzegorz and Weissenberger, Felix and Orlandi, Alessio and Popovici, Dan and Palepu, Anil and Rong, Keran and Tanno, Ryutaro and Saab, Khaled and Zhang, Fan and Blum, Jacob and Carroll, Andrew and Kulkarni, Kavita and Tomasev, Nenad and Zverinski, Dina and Rendulic, Ivor and Vedadi, Elahe and Hasler, Florian and Rimanic, Luka and Boia, Marina and Budiselic, Ivan and Feinstein, Ben and Bellaiche, Mathias and Sheffer, Tom and Freyberg, Jan and Ratcliff, Jeremy and Bertolli, Ottavia and Chou, Katherine and Hassidim, Avinatan and Gokturk, Burak and Vahdat, Amin and Guan, Yuan and Dhillon, Vikram and Vaishnav, Eeshit Dhaval and Lee, Byron and Costa, Tiago R D and Penadés, José R and Peltz, Gary and Matias, Yossi and Manyika, James and Hassabis, Demis and Xu, Yunhan and Kohli, Pushmeet and Pawlosky, Annalisa and Karthikesalingam, Alan and Natarajan, Vivek},
  year = {2025},
  doi = {10.1038/s41586-026-10644-y},
  eprint = {2502.18864},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.1038/s41586-026-10644-y}
}

References

8 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. 1.​ Guan, Y. et al. AI-assisted drug re-purposing for human liver fibrosis. Adv. Sci. 12, e08751 (2025). 2.​ Penadés, J. R. et al. AI mirrors experimental science to uncover a novel mechanism of gene transfer crucial to bacterial evolution. Cell 188, 6654–6665 (2025). 3.​ Jinek, M. et al. A programm… find our card
  2. Preprint at https://arxiv.org/abs/2403.08295 (2024). 12.​ Leslie, D. et al. ‘Frontier AI,’ power, and the public interest: who benefits, who decides? ↗
  3. Harvard Data Sci. Rev. https://doi.org/10.1162/99608f92.4a42495c (2024). 13.​ Chen, L. et al. Are more LLM calls all you need? Towards the scaling properties of compound AI systems. Adv. Neural Inform. Process. Syst. 37, 45767–45790 (2024). 14.​ Comanici, G. et al. Gemini 2.5: pushing the frontier w… ↗ our card
  4. (2019). 18.​ Silver, D. et al. Mastering the game of Go with deep neural networks and tree search. find
  5. Nature 529, 484–489 (2016). 19.​ Ringel, M. S., Scannell, J. W., Baedeker, M. & Schulze, U. Breaking Eroom’s law. Nat. Rev. find
  6. Drug Discov. 19, 833–834 (2020). 20.​ Pushpakom, S. et al. Drug repurposing: progress, challenges and recommendations. Nat. find
  7. Rev. Drug Discov. 18, 41–58 (2019). 21.​ He, L. et al. Chimeric infective particles expand species boundaries in phage-inducible chromosomal island mobilization. Cell 188, 6636–6653 (2025). 22.​ Döhner, H., Weisdorf, D. J. & Bloomfield, C. D. Acute myeloid leukemia. N. Engl. J. Med. 373, 1136–1152 (… find
  8. Commun. 12, 6138 (2021). 27.​ Guan, Y. et al. Live-cell imaging of human liver fibrosis using hepatic micro-organoids. JCI Insight 10, e187099 (2024). find