A task that DeepMind says usually takes medicinal chemists several days, from gathering compound data to proposing new molecules, took its agent about an hour. Google DeepMind on 3 November released a set of Antigravity cheminformatics tools, which it calls Cheminformatics Skills, and showed the agent working through a real lung cancer drug problem from literature search to synthesis plans.
The release turns Antigravity, Google’s agent workspace, into what DeepMind calls “a collaborative science workbench”. The interesting design choice is the split of work: the language model plans the workflow, navigates databases and reasons about the science, while established, deterministic chemistry software does the molecular calculations. DeepMind is just as clear about the limit, writing that “computational designs are testable hypotheses, not finished drugs.”
In brief
- Google DeepMind has released Cheminformatics Skills for its Antigravity agent, built on RDKit with ChEMBL and PubChem connectors.
- In a lung cancer walkthrough, DeepMind says the agent went from literature search to new candidate designs in about an hour, against days for a typical iteration.
- None of the designs has been made or tested; DeepMind says “every proposed design must be synthesized and validated experimentally.”
How the Antigravity cheminformatics skills work
According to DeepMind’s blog post, the skills are grounded in RDKit, the open-source cheminformatics toolkit built up by a global community over two decades, with connectors to the curated ChEMBL and PubChem databases. The skill definitions are published in the google-deepmind/science-skills repository on GitHub, which includes ChEMBL and PubChem database skills.
DeepMind argues the main bottleneck is data wrangling. The same molecule can appear as a chemical name in a paper, a drawn structure in a patent or a data file in a supplier catalogue, and a single wrong atom describes a different compound. The skills encode best practice for pulling these sources together so that the agent’s output can be checked.
The skills are available now through the Science plugin in the Antigravity app, and enterprise users can sign in with Google Cloud credentials. DeepMind is also offering 10,000 scientists the highest Antigravity usage tier free for six months, onboarded on a rolling basis and subject to eligibility, capacity and selection.
A lung cancer drug pair as the test case
The walkthrough uses two real drugs for certain forms of non-small cell lung cancer. Crizotinib, a first-generation kinase inhibitor, has five major rotatable bonds, which make it “floppy” in solution. Lorlatinib, a third-generation drug, is cyclised into a ring, cutting its effective rotatable bonds to zero, which DeepMind links to higher potency.
The agent searched the literature for the two drugs and their analogues, retrieved thousands of compounds within minutes, filtered out buffers and solvents, and documented where each came from. It flagged that the ring-shaped macrocycles were largely predicted to be brain-penetrant, which matters because this cancer often spreads to the brain.
It then ran matched molecular pair analysis, comparing molecules that differ only in whether they are cyclised. The result matched textbook expectations: cyclising cut rotatable bonds and generally raised potency. The agent also trained and benchmarked a range of potency models, using scaffold splits to test how well they generalise to unfamiliar structures, before proposing new rigidified macrocycles with suggested synthesis routes and electron-pushing diagrams.
This follows DeepMind’s earlier demonstration of Antigravity finding an antimicrobial peptide candidate, which also still needed lab testing. Google is meanwhile closing Jules, its earlier coding agent.
The value is in the plumbing, not the molecules
The headline speed is the claim most readers will notice, but the more durable idea is the division of labour. Language models are known to slip on exact chemical structures, and in chemistry a near miss is simply a different molecule. By handing calculations to RDKit and lookups to ChEMBL and PubChem, DeepMind is using the model as a coordinator rather than an oracle. That makes each step traceable to a tool or a database record, which is what a sceptical chemist, or later a regulator reviewing how a candidate was chosen, would want to see.
It is also notable that the showcase rediscovers a known result. Rigidifying crizotinib into something like lorlatinib is a classic medicinal chemistry story, so the demonstration shows the agent can reproduce sound reasoning; it does not show it can find something human chemists missed. The new macrocycles it proposed are untested, and DeepMind says sparse data, assay noise between labs and narrow ranges limit how precise any potency model can be.
That puts this alongside other recent tools such as UCLA’s MAESTRO, which ranked thousands of compounds to pair with a cancer drug, and Pythia Labs’ AI antibody design work: useful for narrowing the search, but dependent on wet-lab results. The free usage offer to 10,000 scientists is the real test. If academic groups, including UK labs, win places and use it on their own targets and publish what survives synthesis, the hour-long workflow will mean something. If not, it stays a polished demo.
What to watch
DeepMind has not said how many scientists have applied for the free tier or when the first cohort will be onboarded. The next useful evidence will be any published case where an Antigravity-proposed molecule has been made and tested in a lab.
Sources
- Google DeepMind: Agentic cheminformatics with Google Antigravity (3 November 2026)
- GitHub: google-deepmind/science-skills
Related stories
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
