Gemini for Science Launches With Peer-Reviewed Benchmarks: ERA Beat CDC Forecasting Model

Google DeepMind launched Gemini for Science on Tuesday at Google I/O 2026, unveiling a suite of agentic AI tools built to automate the most labor-intensive phases of scientific research — and backed 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), published in the journal 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. Researchers can register interest for the experimental tools today; the underlying Science Skills data layer is available now on GitHub and inside Google Antigravity.

Three Tools, Three Stages of Discovery

Gemini for Science bundles three experimental prototypes targeting distinct phases of the scientific method. All three are available through a gradual access program beginning today.

Hypothesis Generation, built on Co-Scientist, addresses one of the most acute bottlenecks in modern research: no scientist can read the millions of papers published annually. The tool simulates scientific discourse by running a multi-agent "idea tournament" in which AI agents generate, debate, and score competing hypotheses for novelty and feasibility. Every claim is backed by verified, clickable citations, letting researchers trace outputs to primary sources rather than accepting the model's word. Google's Co-Scientist paper, published today 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%.

Computational Discovery, built with AlphaEvolve and ERA, functions as what Google calls an "agentic research engine." Rather than requiring researchers to write and iterate code for each computational experiment, the tool auto-generates and scores thousands of code variations in parallel. The ERA paper, also published today in Nature, reported results across six benchmark tests spanning genomics, epidemiology, geospatial analysis, neuroscience, time-series forecasting, and numerical analysis. ERA generated 14 COVID-19 hospitalization forecasting models that outperformed the CDC's official CovidHub Ensemble — the gold-standard aggregate of all professional forecasting teams. In bioinformatics, ERA produced 40 novel methods for single-cell data analysis that surpassed the top human-developed methods on a public leaderboard. Tasks that would take human teams months to navigate manually can be explored in parallel within the system.

Literature Insights, built with Google NotebookLM, handles the other end of the research process: making sense of what already exists. The tool searches scientific literature and structures results into tables with custom, searchable attributes for side-by-side analysis. Researchers can converse with their curated corpus to uncover nuances across papers, identify research gaps, and generate artifacts including reports, slide decks, and audio overviews.

Science Skills: 30-Database Integration Layer Available Today

Underpinning all three experimental tools is a fourth component called Science Skills, which connects agentic platforms — including Google's own Antigravity — to more than 30 major life science databases and tools including UniProt, the AlphaFold Database, the AlphaGenome API, and InterPro. Science Skills is the most immediately accessible part of the launch: available today on GitHub and directly within Antigravity. In internal testing, Google reported that a complex structural bioinformatics analysis that normally takes hours was completed in minutes, producing new insights about potential disease mechanisms caused by mutations in the AK2 gene.

Enterprise Partners Already in Private Preview

Google confirmed that enterprise partners in private preview include pharmaceutical developer Daiichi Sankyo, agricultural sciences company Bayer Crop Science, chemical giant BASF — which is using AlphaEvolve to accelerate supply chain decision-making across its global network — and financial technology company Klarna. The U.S. Department of Energy's Genesis Mission is also using Co-Scientist as part of a formal collaboration with Google DeepMind.

100 Institutions, Including Stanford and Imperial College London

The tools are being validated collaboratively with more than 100 research institutions. Stanford University School of Medicine contributed the liver fibrosis study cited in the Co-Scientist Nature paper. Imperial College London's Fleming Initiative has tested Co-Scientist on antimicrobial resistance research, with Professor José Penadés reporting that the system proposed the same hypothesis his team had reached through a decade of painstaking laboratory work, but in a fraction of the time. The Francis Crick Institute has maintained a multi-year machine learning and biology partnership with Google DeepMind. Google has also created dedicated pilots with academic conferences ICML, STOC, and NeurIPS to test agentic peer review tools including the Paper Assistant Tool and ScholarPeer.

What Separates This From Prior AI-for-Science Claims

When Google announced the earlier standalone Co-Scientist prototype in early 2025, Sara Beery, a computer vision researcher at MIT, told TechCrunch she was unconvinced, describing the system as "not likely to be seriously used" and questioning whether there was genuine demand from scientists for hypothesis-generation tools. The critique was fair at the time: results were preliminary and empirical benchmarks were thin. The simultaneous publication of two Nature papers today directly addresses that gap, placing ERA and Co-Scientist in the same peer-reviewed record as earlier AlphaFold research that has since been used by over three million researchers worldwide.

Whether the tools perform at scale outside controlled benchmarks — and whether they integrate cleanly into the heterogeneous software environments of most research institutions — remains to be established in practice. The current toolset leans most heavily toward bioscience applications; how well Co-Scientist and ERA transfer to fields like materials science, climate modeling, or theoretical physics will depend on tuning work that has not yet been demonstrated publicly.

Pushmeet Kohli, Chief Scientist at Google Cloud and Vice President at Google DeepMind, and Yossi Matias, Vice President at Google and General Manager of Google Research, authored the official launch post, framing the initiative as an effort to close what they called a central paradox of modern science: collective knowledge grows faster than any individual researcher can absorb, making AI synthesis not a luxury but a structural requirement for discovery at scale.

Researchers can register interest at labs.google/science. Enterprise teams can apply for prioritized access through Google Cloud.