Google's Gemini for Science Beats CDC's COVID Forecasting Team in Nature

For most of AI's history in science, the big wins came from narrow specialists: a model trained to fold proteins, another to predict weather, another to read genomic sequences. At Google I/O 2026, Google made a public bet that this era is ending. Gemini for Science is a new suite of agentic tools designed to cover the entire arc of the scientific method, from reading literature to generating hypotheses to running thousands of computational experiments in parallel.

The launch was backed by something few AI-for-science announcements carry: same-day peer-reviewed validation in Nature. Two papers, one on Co-Scientist and one on Empirical Research Assistance (ERA), were published on May 19, 2026, establishing benchmarks that include outperforming the U.S. Centers for Disease Control and Prevention's own COVID-19 hospitalization forecasting ensemble. That is a meaningful signal. Google is not just shipping demos.

The three tools, explained

Gemini for Science experimental tools on Google Labs include three primary prototypes designed to handle distinct tasks across the scientific workflow.

• Hypothesis Generation (built with Co-Scientist): The system simulates the scientific method through a multi-agent "idea tournament" that generates, debates, evaluates, and verifies hypotheses with supporting citations. Think of it as a panel of AI reviewers that argue over your research ideas and surface the most defensible ones, all grounded in the published literature.
• Computational Discovery (built with AlphaEvolve and ERA): This agentic research engine is capable of generating and testing thousands of code variations in parallel, allowing scientists to explore novel modeling approaches for areas such as solar forecasting and epidemiology much faster than traditional methods.
• Literature Insights (built with NotebookLM): This tool searches scientific literature and structures results into tables with custom, searchable attributes for side-by-side analysis. Researchers can use chat to uncover nuances grounded in their curated corpus, and create high-fidelity artifacts such as reports, slide decks, infographics, and audio and video overviews.

The numbers that matter

The ERA paper is the most quantitatively compelling piece of the launch. Research papers on ERA and Co-Scientist have been published in Nature. The ERA results span six benchmark domains. In epidemiology, ERA generated 14 COVID-19 hospitalization forecasting models that beat the CDC's CovidHub Ensemble, which aggregates the predictions of every professional forecasting team in the country. In bioinformatics, it produced 40 novel methods for single-cell data analysis that surpassed the top human-developed methods on a public leaderboard.

Google's Co-Scientist paper, published in Nature, documented a practical result: Stanford University School of Medicine researchers used it to identify Vorinostat, an FDA-approved anti-cancer drug, as a candidate for liver fibrosis treatment. In hepatic organoid lab tests, the compound reduced TGFβ-induced chromatin structural changes by 91%.

ERA has helped accelerate discoveries from neuroscience to cosmology, with results including predicting hospital admissions for respiratory illnesses and forecasting seasonal runoff across California's river basins.

How each tool actually works

Co-Scientist, which powers Hypothesis Generation, is a multi-agent system. That means it is not a single model answering a question, but a network of specialized agents with different roles. The working cycle involves three stages: generating, debating, and evolving hypotheses, completed through the collaboration of multiple specialized agents including Generation, Reflection, Ranking, Evolution, Meta-review, Proximity, and Supervisor. As the computational volume expands during testing, the quality of the hypotheses continues to improve, meaning you get better results the longer you let it run.

Computational Discovery is built on two systems. AlphaEvolve is an evolutionary algorithm engine that mutates and scores code automatically. ERA (Empirical Research Assistance) is the research coding layer that translates scientific tasks into runnable experiments. The engine generates thousands of code variants in parallel, and each variant is automatically scored, compressing the complex modeling path that originally took humans months to explore into the scope of machine search.

The fourth component, Science Skills, is the data plumbing underneath all three tools. It is a specialized bundle that integrates insights from over 30 major life science databases and tools including UniProt, AlphaFold Database, AlphaGenome API, and InterPro. Using these skills on agentic platforms like Google Antigravity allows researchers to perform complex and often manual workflows like structural bioinformatics and genomic analyses in minutes rather than hours.

Who is already using it

Named enterprise organizations in private preview include BASF, which is using AlphaEvolve to optimize supply chains; Klarna, which is using AlphaEvolve for machine learning models; and Daiichi Sankyo, Bayer Crop Science, and the U.S. National Labs through the U.S. Department of Energy's Genesis Mission, which are using Co-Scientist.

To ensure the integrity of AI-generated insights, Google built a trusted tester community ranging from PhD students to industry researchers to Nobel laureates, to stress test their systems against complex real-world challenges. The company has also created pilots with conferences including ICML, STOC, and NeurIPS for agentic peer review and scientific validation tools, including Paper Assistant Tool and ScholarPeer.

What the scientific community is saying

Reactions from early testers have been striking. Gary Peltz, a Stanford geneticist, compared using the AI Co-Scientist to "consulting the oracle of Delphi" in a Nature Medicine article. That is a strong endorsement from someone who has actually used it on real research problems.

But not everyone is convinced. Skeptics warn about reproducibility and overreliance. The current toolset leans heavily toward bioscience applications, and how well Co-Scientist and ERA transfer to fields like materials science, climate modeling, or theoretical physics has not yet been demonstrated publicly. Whether the tools perform at scale outside controlled benchmarks, and whether they integrate cleanly into the heterogeneous software environments of most research institutions, remains an open question.

A strategic shift, not just a product launch

The broader context here is important. Google believes that a new era of discovery won't come from narrow, specialized models, but from general agents that empower researchers across every scientific field. This is a direct philosophical pivot from the AlphaFold era, where breakthroughs came from training one model to do one thing extraordinarily well.

Gemini for Science is not incompatible with specialized tools; agentic systems can be designed to call on such tools when they might be useful, and no agentic system can predict the structure that a protein will fold into without AlphaFold's help. But the framing has shifted. The goal is now a general research collaborator, not a domain-specific oracle.

Google has been careful about the language it uses here. The name is AI Co-Scientist, not AI Scientist. Gemini for Science packages multiple specialized agents into a coordinated research environment, representing a network of AI tools designed to collaborate with scientists rather than replace them. Whether that framing holds as these systems become more capable is a question the field will be watching closely.

How to get access

Researchers can register interest for the experimental tools at labs.google/science; the underlying Science Skills data layer is available now on GitHub and inside Google Antigravity. Science Skills connects agentic platforms like Google Antigravity to over 30 major life science databases and tools. Pricing for the experimental tools has not been announced; enterprise access through Google Cloud is available via a separate private preview application.

The tools are most immediately useful if you work in life sciences, computational biology, epidemiology, or any field where you spend significant time reading papers, writing experimental code, or trying to connect findings across a fragmented literature. AlphaFold has already helped over 3 million researchers tackle problems like malaria vaccines and plastic-eating enzymes, and Google is betting that a general agentic layer built on top of that foundation can reach even further.