Google bundles its AI-for-science work under Gemini for Science

Google packaged its research-side AI work into Gemini for Science, with three experimental tools in Google Labs and a Science Skills bundle on Antigravity. Hypothesis Generation, built on Co-Scientist, runs what the team calls a multi-agent idea tournament, where parallel agents generate, debate, and score hypotheses against verified citations. Computational Discovery uses AlphaEvolve together with a system called ERA to generate and evaluate thousands of code variations in parallel for problems like solar power forecasting and epidemiology. Literature Insights, built on NotebookLM, pulls scientific literature into searchable tables and flags missing pieces in a researcher's coverage.

The Science Skills bundle is the more concrete piece. It wires more than thirty life-science databases, including UniProt, AlphaFold Database, AlphaGenome API, and InterPro, into Google Antigravity, so an agent can run end-to-end structural bioinformatics or genomic queries without the researcher hand-stitching each API. Listed partners include Stanford on liver fibrosis, Imperial College London on antimicrobial resistance, and the Crick Institute, with private-preview enterprise users like BASF and Klarna already running on AlphaEvolve.

Two of the underlying systems landed in Nature alongside the announcement, one for ERA and one for Co-Scientist. That puts the marketing on a different footing from most AI launches: there is a peer-reviewed result to argue from, not just a benchmark slide.

Why it matters

For working scientists, this is the first time Google's various science models are presented as a single workflow rather than scattered demos. The Nature papers also raise the bar on what counts as a credible AI-for-science announcement. If you evaluate vendors for a research lab, the question to ask now is "where is the paper" rather than "what does the demo do".