Google DeepMind Unveils Gemini for Science: A Suite of AI Tools for Scientific Discovery

Google DeepMind has announced Gemini for Science — a set of experimental AI tools and applications aimed at expanding the scale and accuracy of scientific research. The approach is based on the idea that a new era of discovery will be driven not by narrowly specialized models, but by universal agents capable of assisting scientists across diverse fields. As part of Gemini for Science, three main prototypes are available on the Google Labs platform. Hypothesis Generation, built on Co-Scientist, helps generate and evaluate hypotheses using a multi-round 'tournament of ideas' system and verifies claims with source references. Computational Discovery, based on AlphaEvolve and ERA, enables parallel generation and evaluation of thousands of code variants for computational experiments, reducing testing time from months to minutes; application examples include solar activity forecasting and epidemiology. Literature Insights, powered by Google NotebookLM, structures scientific literature into tables with customizable attributes and allows creating reports, slides, infographics, and audio and video summaries. Access to the experiments is being rolled out gradually; registration is available at labs.google/science. These capabilities are also provided to enterprise customers via Google Cloud: for example, BASF uses AlphaEvolve to optimize supply chains, Klarna to improve machine learning models, and Daiichi Sankyo, Bayer Crop Science, and U.S. national laboratories (as part of the U.S. Department of Energy's Genesis mission) apply Co-Scientist to accelerate research. Scientific papers on ERA and Co-Scientist have been published in Nature. Additionally, the Science Skills module has been launched — a specialized toolkit aggregating data from over 30 major life sciences databases, including UniProt, AlphaFold Database, AlphaGenome API, and InterPro. Using Science Skills on agent platforms such as Google Antigravity, researchers can perform complex workflows (structural bioinformatics, genomic analysis) in minutes instead of hours. Google collaborates with over 100 institutions, including Stanford University (liver fibrosis), Imperial College London (antimicrobial resistance), and the Francis Crick Institute, to validate the systems. Pilot projects have also been established with conferences ICML, STOC, and NeurIPS for agent-based peer review and validation using the Paper Assistant Tool (PAT) and ScholarPeer.