AI for Bio · Model

Co-Scientist

Google DeepMind's Gemini-based multi-agent system that generates and ranks research hypotheses

Text, LLMs & agentsBig-tech labUnited Kingdom, United StatesProduct only, no public artifact中文

Overview

NameCo-Scientist · aliases: AI co-scientist, Google Co-Scientist
Full nameAI co-scientist
Ambiguous names (not auto-linked)Co-Scientist, co-scientist
FamilyNot a family entry. A single system; the 2025 preprint and 2026 journal version are versions within one record
OrganizationGoogle DeepMind, Google Research
Organization typeBig-tech lab
CountryUnited Kingdom, United States
Model classAgent (hypothesis generation)
DomainText, LLMs & agents
First release · paper revision
Basis: arXiv preprint v1 (official blog the same period) source
Other milestones
  • First public evidence: · arXiv preprint v1 source
  • Preprint v1: · arXiv preprint v1 source
  • Journal online: · Nature publication (official blog the same day) source
  • Official announcement: † · Google Research blog source
Latest update · paper revision
Basis: arXiv v2
arXiv v2 posted (scope of revision not inspected), after the 2026-05-19 Nature paper and opening to individual researchers through Gemini for Science source
The most recent sourced event, not a claim that nothing changed since.
StatusProduct only, no public artifact · Closed; offered through an experimental tool and a cloud preview
OpennessCode: Closed
Weights: Closed
Closed; individual researchers apply via Gemini for Science and an enterprise version is in Google Cloud preview
Benchmarks reported0 counted (unconfirmed 6)
Wet-lab validation1 recorded campaign
Unverified fields† wetlab_validation[0], release_milestones.official_announcement, training_data.a_public
As of2026-11-17

What it does

A Gemini-based multi-agent system in which generation, reflection, ranking, evolution and meta-review agents iterate on hypotheses, ranking them in an Elo tournament of pairwise comparisons and simulated debates, and output research proposals; checks call web search and public databases such as ChEMBL and UniProt. Unlike Kosmos or Biomni, it outputs ranked hypotheses rather than running data analysis.

Benchmarks reported

Every row is one pairing from the survey's relation file; a score belongs to the benchmark page's ranking, where it is shown with metric, eval set, source type and date. 0 of 6 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
14 PubMedQAUnconfirmed (not counted)—Unconfirmed sourceNo report found in official materials †disputed · evidence
18 LAB-Bench / LABBench2Unconfirmed (not counted)—Unconfirmed sourceNo report found in official materials †disputed · evidence
23 BioASQUnconfirmed (not counted)—Unconfirmed sourceNo report by it was found in the official material. †disputed · evidence
29 BixBenchUnconfirmed (not counted)—Unconfirmed sourceNo report found in official materials †disputed · evidence
41 BioML-benchUnconfirmed (not counted)—Unconfirmed sourceNo report on this benchmark found in official materials. †disputed · evidence
42 BioMysteryBenchUnconfirmed (not counted)—Unconfirmed sourceNo report on this benchmark found in official materials. †disputed · evidence

Other evaluations (outside the 44)

  • Site text and agent benchmarks (none disclosed) Self-reported: Checked LAB-Bench, BixBench, BioASQ, PubMedQA, BioML-bench, BioMysteryBench: not reported in the paper or official blog
  • Elo vs GPQA correctness (Public benchmark not among the 44) Self-reported: Uses GPQA Diamond to check whether system Elo tracks answer correctness; the Elo is self-assessed with no independent ground truth
  • Expert evaluation (user or partner report) Self-reported: A small expert panel rated novelty and impact; the developers describe the sample as small
  • Elo tournament (Own task set) Self-reported: Self-play ranking improves with test-time compute; no independent ground truth

Developers report: a multi-agent Gemini system that generates and ranks hypotheses, with Elo tracking quality and some hypotheses confirmed in partner labs.

Independent evidence: Press reports describe a UK university team trying it on a bacterial-evolution question, and 2025 press criticism noted a lack of independent evaluation; no systematic third-party study was found.

Training data

Group in Figure 7: Public data + wet lab for validation only.

ClassValueDetail
a. Public databasesNot disclosedBase Gemini training data undisclosed; ChEMBL, UniProt and similar are inference-time tools source
b. Private or licensed dataNot disclosed
c. Own wet-lab data used for trainingNoThe system is not trained on wet-lab data
d. Wet lab used only for validationYesSome hypotheses were tested in partner labs; the developers give no aggregate counts source

Wet-lab validation

Every rate carries its numerator and denominator. Claims that could not be verified are shown as unverified, without numbers; different paper versions are separate entries.

Partner-lab case studies (paper/blog)
Targets — · Designs tested — · Hits — · Success — · Best affinity —

Denominator basis: No aggregate counts; individual cases such as liver-fibrosis targets, AML drug repurposing and cellular senescence, each run by a partner lab
Qualitative cases only; must not be converted into hit rates or quantitative validation
Executor: Run by external partner labs; not independent replication

Self-reported · · †unverified · source

Background

  • Organization: Google DeepMind was founded in London in 2010, acquired by Google in 2014 and merged with Google Brain in 2023-04; this system was built with Google Research, Google Cloud and Google Labs.
  • Funding: A wholly owned internal research unit of Alphabet with no separate funding.
  • Products and partnerships: Enterprise preview partners include Daiichi Sankyo, Bayer Crop Science and US national laboratories; academic cases include Stanford, MIT, Edinburgh, Cambridge and Calico; the developers cite researchers at over 100 institutions.
  • Related models: Kosmos / Edison Analysis, Biomni, Robin, MedGemma / Med-Gemini, VERID

Reading notes

  1. It has no standard site benchmark score and cannot be compared directly with Kosmos or Biomni.
  2. Partner-lab results are single cases without stated controls or replicates and cannot be read as efficacy or hit rates.
  3. The Elo-GPQA correlation is self-assessed and comes from the 2025 preprint.
  4. The 2026 claim of planning experiments has no detail in official text, so it is best read as a hypothesis-generation system.

Caveats

  1. Wet lab was done by partner labs with no aggregate counts; treated as unverified.
  2. Closed, so external evaluation is not possible.
  3. The scope of the arXiv v2 revision was not inspected.

Sources

  1. Nature paper (journal)
  2. arXiv preprint (preprint)
  3. DeepMind official blog (2026-05-19) (company blog)
  4. Google Research 2025 blog (company blog)
  5. Gemini for Science launch blog (company blog)
  6. Nature paper registry record (journal)

Corrections

2 corrections
FieldWasNowBasisFound
latest_update2026-05-19 (Nature publication)2026-06-29(arXiv v2)source2026-11-16
latest_update.event_kindjournal-publicationpaper-revisionsource2026-11-16

Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-17.