Google Pushes Forward with New AI for Science Tools
Google is launching a series of new tools to help scientists leverage AI technology as a force multiplier to accelerate the pursuit of scientific knowledge and discovery. In addition to the collection of tools under the Gemini for Science banner, it is also launching Science Skills, which centralizes data from more than 30 life science databases, as well as a major upgrade to Gemini 3 Deep Think.
Ever since its DeepMind arm launched AlphaFold back in 2018, Google has been intimately involved in utilizing AI for science. The company was an early backer in the Department of Energy’s Genesis Mission project to pursue AI for science and engineering in the National Labs.
This week at its Google I/O developer conference, the tech giant brought together several of its existing AI for science projects into a product called Gemini for Science. The new offering includes three primary tools, including Literature Insights for AI-assisted document review, Co-Scientist for hypothesis generation, and Computational Discovery, an AI-powered research engine.
Literature Insights enables researchers to perform targeted searches and analyze large bodies of scientific literature. The software, which is powered by Google’s AI-powered research assistant and note-taking app called NotebookLM, allows researchers to interact and “chat” with the data in the literature to uncover nunaces and patterns. Literature Insights also generates high-quality tables, reports, and infographics with embedded links back to the texts.
Hypothesis Generation, meanwhile, is a new offering built on Co-Scientist, the AI assistant developed by Google and DeepMind. The software picks up where Literature Insights leaves off by emulating the scientific method and enabling researchers to create research direciton and define hypothesis. It acts as a research partner and uses the concept of an “idea tournament,” where AI agents compete to generate, debate and evaluate hypotheses. Researchers can also chat with Hypothesis Generation agents to identify the best candidates for further evaluation.
Once a promising hypothesis is in hand, the researcher can fire up Computational Discovery to automate the execution of experiments to prove or disprove the hypothesis. Computational Discovery is built with DeepMind’s AlphaEvolve and ERA (Empirical Research Assistance) software, and utilizes AI agents to generate and score thousands of code variations in parallel, thereby enabling scientists to test novel modeling approaches that would otherwise take months to test manually.
Google researchers published a pair of papers in the journal Nature that showcase the advances made in ERA and Co-Scientist. In the ERA Nature paper, Google researchers say the tool, which is based on an LLM and Tree Search, “discovered 40 novel methods for single-cell data analysis that outperformed the top human-developed methods on a public leaderboard. In epidemiology, the tool generated 14 models that outperformed the CDC ensemble and all other individual models for forecasting COVID-19 hospitalizations. In the Co-Scientist Nature paper, the Googlers document how “the tool helped identify new drug repurposing candidates and synergistic combination therapies for acute myeloid leukemia, which were validated through in vitro experiments.”
Gemini for Science collectively will help drive scientific discovery from time-consuming manual processes to AI-powered automation, write Pushmeet Kohli, the Chief Scientist at Google Cloud and a Vice President in Google DeepMind, and Yossi Matias, a Vice President in Google and General Manager of Google Research, in a May 19 blog post.
“Today science faces a paradox: our collective knowledge is growing so fast that it’s becoming harder for individual scientists to see the full picture,” the Google executives write. “Scientific breakthroughs often rely upon making creative connections between data, but the time required to do this manually can take weeks or even months. AI can help eliminate this bottleneck and serve as a force multiplier for scientific work by handling complex tasks. This allows researchers to focus on identifying and tackling the most impactful scientific problems and directions that would drive progress.”
Rutgers mathematician Lisa Carborne was surprised Gemini 3 Deep Think mode identified errors in her paper
Kohli and Matias say these tools are already being adopted in pilot projects. It’s working with more than 100 institutions on a variety of projects, including with Stanford University on liver fibrosis, Imperial College London on antimicrobial resistance and The Crick Institute. The company is also working with scientific conferences like ICML, STOC and NeurIPS to develop tools for agentic peer review and scientific validation, such as its Paper Assistant Tool (PAT) and ScholarPeer.
Google Science Skills
Google is also launching Science Skills, a collection of data and tools for aiding the scientific process in areas like bioinformatics and genomic analysis.
Science Skills comes prebundled with insights from more than 30 life science databases, as well as access to tools like UniProt (Univeral Protein Resource), the free bioinformatics database InterPro, and DeepMind’s own AlphaFold Database and AlphaGenome API.
“We have oceans of data and amazing predictive models,” says Saz Basu, a Google DeepMind research scientist. “But moving from an initial observation to a physical experiment means navigating fragmented manual and sometimes slow process.”
In a demo, Basu showed how Science Skills can be used to glean insights about the underlying mechanism behind a rare genetic disease caused by mutations in the AK2 gene. Working from Google Antigravity, the company’s IDE for AI-assisted development, Basu is able to bring the power of AI models to bear on AK2 gene data in the Alphagenome database to come up with a hypothesis. Basu then connects the model with a protein database to see if there’s the possibility of a causal connection that could be fleshed out in a wet lab experiment.
“It used to take me an entire afternoon to do a workflow like this,” Basu said. “And now we go from observations to hypothesis and designing the blueprints needed to decipher the mechanism in the real world in a matter of minutes.”
(Source: Google)
Google Gemini 3 Deep Think
Finally, Google is announcing an update to Gemini 3 Deep Think, its AI-powered reasoning “mode” for science, research, and engineering. In a blog post, the Deep Think team says it worked closely with scientists and researchers to enable the model to tackle tough research challenges, “where problems often lack clear guardrails or a single correct solution and data is often messy or incomplete.”
One user of the new mode was Lisa Carborne, a mathematician at Rutgers University who works on the mathematical structures required by the high-energy physics community. In a video posted to Google’s site, Carborne says she was surprised when, after running a draft of a paper through Gemini 3 Deep Think, the model told her some of her math was off.
“It gave three separate, irrefutable reasons why our mathematical arguments around one particular statement were incompatible,” Carborne says. “This was pretty destabilizing because the paper had already been peer reviewed.”
The mathematician was surprised to discover that Gemini 3 Deep Think had done the tough work of a highly skilled math expert, like herself and her peer reviewers. The AI model had a different perspective on the work, which she wasn’t expecting.
“It took me a while to understand, because it was really outside of my thought process, and the model’s reasoning was completely correct,” the professor says. “The paper’s at the forefront of research in the subject, and so there’s very little context or training data that the model could have been trained on. So it seemed as if it did the work of a highly trained mathematician.”
According to Google the Deep Think mode is pushing the bounds on what AI can do. The AI mode scored 48.4% on the Humanity’s Last Exam benchmark, which is designed to push modern frontier models to their limits. It scored 84.6% on ARC-AGI-2, got an Elo rating of 3455 on the Codeforces benchmark and reached gold-medal level performance on the International Math Olympiad 2025
Google AI Ultra subscribers now have access to the updated Deep Think mode. Google is also making it available to scientists, engineers and enterprises via the Gemini API. If you would like to get access, you can send a request to Google here.
Editor’s note: A version of this story first ran on HPCwire
Related
Yann LeCun’s AMI Secures $1B Seed to Develop AI World Models
Turing Award–winning AI researcher Yann LeCun has spent years arguing that large language models, at…
Deloitte’s State of AI 2026: Why Enterprise Execution Is Falling Behind Adoption
Deloitte’s latest State of the AI report shows that AI adoption continues to accelerate rapidly,…
Broadcom Announces VMware Cloud Foundation 9.1
PALO ALTO, Calif., May 5, 2026 — Broadcom Inc., a global technology leader that designs,…
Anthropic Unveils ‘Project Glasswing’ as Claude Mythos Targets Software Vulnerabilities
April 9, 2026 — Anthropic has announced Project Glasswing, a new initiative that brings together Amazon…
Google Unveils Gemini Enterprise Agent Platform, Expands Vertex AI into Full Agent Stack
At Google Cloud Next 2026 in Las Vegas this week, the company announced major changes…
OpenAI Shutters Sora, Shifts Business Strategy Ahead of IPO
Back in 2022, OpenAI set off a chain reaction in the tech world when it…
