Google DeepMind's Science Skills Cuts Biology Research From Hours to Minutes

Google Antigravity just got a serious upgrade for life scientists. Google DeepMind has released Science Skills, a curated bundle of agent skills that connects Antigravity , Google's agentic coding and analysis platform , directly to more than 30 major scientific databases and tools. The pitch is simple but significant: workflows that used to take a researcher hours of manual database-hopping can now be completed in minutes by an AI agent.

The Science Skills bundle is part of the broader Gemini for Science initiative, which also includes three experimental tools: Literature Insights (built with NotebookLM), Hypothesis Generation (built with Co-Scientist), and Computational Discovery (built with AlphaEvolve and ERA). Science Skills is the most immediately practical piece of that suite , it's available right now, free, on GitHub.

What's actually in the box

Science Skills integrates more than 30 major life science databases and tools, including the AlphaFold Database for protein structure predictions, AlphaGenome, UniProt, and InterPro. The bundle is organized into four functional categories:

• Genomics, Transcriptomics, and Regulatory Biology (11 skills): Tools for querying genetic variants, gene expression, and transcriptional regulation across databases like gnomAD, Ensembl, ClinVar, ENCODE, GTEx, dbSNP, JASPAR, and UCSC.
• Proteomics and Structural Biology (10 skills): Fetch and analyze 3D protein structures via the AlphaFold Database, run structural similarity searches with Foldseek, align sequences with Clustal Omega, visualize molecules in PyMOL, and query UniProt, PDB, InterPro, Human Protein Atlas, and STRING.
• Cheminformatics and Clinical Translation (5 skills): Query ChEMBL for drug targets, ClinicalTrials.gov, OpenFDA, OpenTargets, and PubChem , bridging basic biochemistry to drug discovery pipelines.
• Scientific Literature, Pathways, and Ontologies (8 skills): Search PubMed, arXiv, bioRxiv, Europe PMC, OpenAlex, Reactome, QuickGO, and EMBL-EBI's ontology service.

There's also a Workflow Automation skill worth calling out: workflow-skill-creator lets you "handhold" the agent through a custom workflow once, and it will distill that interaction into a reusable skill you can run repeatedly or share with others. That's a meaningful escape hatch for the inevitable coverage gaps.

The grounding problem this actually solves

The core problem Science Skills addresses isn't knowledge , it's grounding. A general-purpose agent like Antigravity already has broad scientific knowledge from training, but when it tries to access a specific database or cite a primary source, it often stumbles: wrong API calls, hallucinated URLs, or generic web searches that return outdated results.

The technical report shows that the bundle improves the reliability and efficiency of scientific workflow tasks: more tasks succeed and the number of tokens and model calls necessary to complete tasks is reduced. With skills, smaller Flash-category models attain Pro-level reliability. That's a meaningful result , it means you don't need to burn the more expensive model to get trustworthy outputs on biology tasks.

The numbers back this up concretely. Science Skills launched with a technical report showing quantitative benchmark results across internal and external evaluations. On the team's internal 67-task capability benchmark, Gemini 3 Flash without skills succeeded on 49% of tasks. With Science Skills, that jumped to 93% , matching and slightly exceeding Gemini 3.1 Pro without skills (67%). Token consumption also dropped by a factor of 2.04× for Flash and 1.62× for Pro. On the external BioReason benchmark (a multi-step variant-effect reasoning dataset with 2,106 tasks), accuracy on coding variant prediction improved from 41.4% to 60.9%, and on non-SNV variants from 46.6% to 81.6%.

The agent traces in the technical report make the mechanism intuitive. Without skills, the agent for a simple ChEMBL query fires off 11 web searches, makes multiple failed API calls, and eventually gets the right answer. With the chembl-database skill, it reads the SKILL.md instruction file, runs the right script directly against the ChEMBL API, and finishes in 6 steps. The skill essentially pre-teaches the agent how to use each database correctly , the right endpoints, the right scripts, the right output format.

How it works under the hood

Skills are implemented using the Agent Skills standard , a cross-vendor standard for customizing agents through a layered set of instructions, code, and data. Each skill directory contains a SKILL.md file with YAML frontmatter and detailed markdown instructions, a scripts/ folder with helper Python scripts, and optional reference documentation. The agent reads the SKILL.md to understand what the skill does and how to invoke its scripts, then calls those scripts directly rather than improvising API access from scratch.

The bundle uses the uv package manager to handle dependencies. The first time you trigger a Science Skill, the agent will ask for approval and install uv, then proceed to respond to your query. Restarting Antigravity after this first installation is recommended. Some skills require API keys: AlphaGenome and OpenAlex require an API key to function. Others like ClinVar work without one but unlock higher rate limits with a key.

The quality testing framework is also worth noting. The team validated each skill at three levels: unit tests (does the agent correctly invoke each individual capability?), workflow tests (can it orchestrate multiple skills across a multi-turn task?), and capability tests (67 harder, skill-agnostic tasks that measure end-to-end improvement). All unit and workflow tests pass before a skill ships.

Getting started

Installation is a single command. If you're using the GitHub repo directly with any npx-compatible agent:

npx skill add google-deepmind/science-skills/

For Antigravity users specifically, new users can check the "Science" box during the "Build with Google" setup step. Existing users can go to Settings → Customizations → Build with Google Plugins → Customize → Download the Science plugin. The bundle is free and open-source under Apache 2.0.

What it's good at , and where it falls short

Science Skills allows researchers to run workflows including structural bioinformatics, genomic analysis, protein analysis, and structural searches in minutes rather than hours. The real-world validation is compelling: Google's research teams used Science Skills in early testing to carry out an analysis that usually takes hours in minutes, leading to insights about possible mechanisms for a rare genetic disease caused by mutations in the AK2 gene.

The limitations are honestly documented in the technical report. Coverage is intentionally narrow , the initial focus is biology and life sciences, and scientific work spans far more territory than 34 skills can cover. Long-running workflows (those spanning many hours with complex multi-step dependencies) are also not yet well-supported; the current skills target workflows of "dozens of steps" lasting a few hours. And reproducibility across different user environments remains a challenge, which is why the team standardized on uv , a tradeoff that adds setup friction but improves reliability.

Why this matters beyond the headline

The premise is not subtle: Google believes the same agentic stack that powers software development agents can be pointed at the scientific method itself. Literature review, hypothesis generation, experimental design, and algorithm discovery are all, in some sense, search-and-synthesis problems.

AlphaFold has already helped over 3 million researchers tackle challenges from malaria vaccines to plastic-eating enzymes, and AlphaGenome is helping scientists identify the drivers of disease. Science Skills is the infrastructure layer that makes those models , and the 30+ databases around them , accessible to an agent that can chain them together autonomously. The real unlock isn't any single database connection; it's the ability to compose them. A researcher can now ask a multi-hop question , "find variants in the conserved region of FOXP2, predict their functional consequences, and pull the relevant literature" , and get a grounded, cited answer in a single session rather than a day of manual work.

The simultaneous publication of two Nature papers on ERA and Co-Scientist directly addresses earlier skepticism about the scientific rigor of these tools, placing them in the same peer-reviewed record as earlier AlphaFold research. Science Skills ships with a technical report and a full benchmark suite , a signal that Google is treating this as infrastructure to be validated, not just a demo to be shipped.