Google DeepMind has launched a suite of Cheminformatics Skills for Google Antigravity, its agentic development surface. According to Google DeepMind, the skills turn Antigravity into a collaborative science workbench with full access to local tooling, built for computational chemists and bench chemists alike. The pitch: stop burning hours on data wrangling, spend them on hypotheses instead.
The interesting part isn’t that an AI touches chemistry. It’s how the work gets divided.
5 things worth knowing
1. The model plans, the tools calculate
Google DeepMind describes a clear division of labour. The language model handles workflow orchestration, database navigation and scientific reasoning. Deterministic chemistry tools run the actual molecular calculations. Picture a lab lead who decides which tests to run while calibrated instruments produce the numbers.
2. RDKit is the foundation
The skills are grounded in RDKit, the open source cheminformatics toolkit built over two decades by a global community of contributors. Smart call. Chemists already trust it, so nobody’s being asked to accept a black box for the core math.
3. Connectors reach into ChEMBL and PubChem
Two major curated public chemistry resources are wired in. An agent that can look things up has less reason to guess. Google also says the skills encode domain best practices directly into the agent’s workflow, so the standard habits of a working chemist are baked in rather than typed into every prompt.
4. It targets chemistry’s representation mess
The same molecule shows up as an IUPAC name (the formal naming system) in a journal article, a Markush structure (a generic template covering many compounds) in a patent, and a structure-data file in a vendor catalog. Any serious system has to unify those sources and handle the many ways errors creep in.
And the tolerance here is basically zero. Alter one atom or bond and you’ve described a different compound, one that could render a promising inhibitor or catalyst inert.
5. The demo tackles a real drug design technique
Google’s technical walkthrough covers rigidification, a classic medicinal chemistry move for boosting potency. Drug molecules often link substructures through rotatable bonds, which lets them twist into many 3D shapes. Floppy, in other words. To bind its target the molecule has to lock into one active shape, and giving up that freedom of motion carries an entropic penalty that weakens binding affinity.
The fix is to add bonds that limit the twisting, a strategy chemists call paying the entropic cost in advance. The case study uses two lung cancer drugs:
- Crizotinib: a first-generation tyrosine kinase inhibitor used against certain non-small cell lung carcinomas, with five major rotatable bonds
- Lorlatinib: a third-generation inhibitor that resembles crizotinib but is cyclized at its end, cutting the effective rotatable bond count to zero and landing higher potency among other effects
Why it matters
Science is turning into the proving ground for AI agents, and chemistry is an unforgiving place to run that test. A model that confidently describes the wrong compound isn’t useful, it’s expensive. By handing calculations to deterministic tools and keeping the model on orchestration and reasoning, Google is admitting something useful out loud: the language model is the worst possible place to put the math.
That pattern shows up everywhere agents meet technical domains. The wins come from pairing a good planner with trustworthy tools, not from asking one model to do everything.
What the article doesn’t say
Google DeepMind’s post is a technical walkthrough, not a full product page. It doesn’t spell out pricing for the skills, a general availability date, or whether access is limited to specific Antigravity tiers. The rigidification example is also a retrospective one: crizotinib and lorlatinib are known drugs with a known relationship, which is a reasonable way to show a workflow but not proof the agent can find a novel rigidification nobody has tried.
Still, this is the most concrete move yet toward agents doing day-to-day scientific work rather than summarizing it. Full technical details, including the complete walkthrough, are available at the original source.