Google DeepMind Puts Cheminformatics Skills Into Antigravity, Compressing a Multi-Day Drug Design Loop Into an Hour

Google DeepMind has released a suite of Cheminformatics Skills for Google Antigravity, its agentic development surface, pitching the combination as a "science workbench" for medicinal chemists. In a technical walkthrough published November 3, the team says an agent searched thousands of compounds drawn from the literature, ran matched molecular pair analysis, trained potency models and proposed new macrocyclic designs with synthetic routes — a loop the authors describe as normally taking multiple days — in roughly an hour. The post is a demonstration, not a validated result: DeepMind is explicit that every design it generated remains an untested hypothesis.

What the skills actually do

The pitch rests on a division of labor that will be familiar to anyone who has built a tool-using agent. According to the DeepMind team, the language model handles orchestration, database navigation and scientific reasoning, while deterministic chemistry tools do the molecular math. Those tools are grounded in RDKit, the open source cheminformatics toolkit the post describes as having been built by a global community over two decades, plus connectors into ChEMBL and PubChem.

That split matters more in chemistry than in most agent domains. DeepMind frames the core difficulty as representational: the same molecule can appear as an IUPAC name in a paper, a Markush structure in a patent, or an SDF record in a vendor catalog, and the team notes there is effectively no error tolerance — change one atom or bond and you have a different compound, potentially an inactive one. Handing structure manipulation to deterministic libraries rather than to a generative model is the obvious mitigation, though the post does not quantify how often the agent's extraction and curation steps still produce wrong structures.

The skills also encode methodological guidance. DeepMind says the agent is steered toward chronological splits when training potency models — training on older compounds to predict newer ones — and falls back to scaffold splits to estimate generalization. Embedding that kind of domain opinion into the skill definition is arguably the most transferable idea in the announcement: it is prompt-level methodology enforcement, aimed at stopping a capable model from producing a statistically flattering but scientifically meaningless result.

The walkthrough: rigidification, from crizotinib to lorlatinib

The demo targets a textbook medicinal chemistry move. DeepMind uses the example of crizotinib, a first-generation small-molecule tyrosine kinase inhibitor for certain non-small cell lung carcinomas, which the post says has five major rotatable bonds. Lorlatinib, a third-generation TKI, is described as a structurally similar compound cyclized at one end to bring the effective number of rotatable bonds to zero, yielding higher potency among other effects. The underlying logic — paying the entropic cost of locking a molecule into its binding conformation up front — is well established, which is precisely why it works as a demo. The agent is being asked to rediscover a known answer.

In the described run, the agent pulled thousands of compounds from literature, filtered out buffers and solvents, and documented provenance in a report. DeepMind says the report flagged that the retrieved macrocycles were largely predicted to be brain-penetrant, a relevant property given that the target cancer frequently metastasizes to the brain. An interactive viewer then let the user plot potency against rotatable bonds, colored by macrocycle status, surfacing the expected relationship between rigidity and potency.

From there the agent ran matched molecular pair analysis across cyclized and non-cyclized pairs, benchmarked several model architectures and feature sets across four experimental read-outs using a scaffold split, and finally proposed new macrocycles with stated medicinal chemistry rationale, predicted potencies from its own trained models, and suggested synthesis routes. DeepMind says the skills let the agent validate reactions, generate curved-arrow electron-pushing diagrams, and look up similar reactions in public databases, and that outputs export into standard chemistry software. The post also notes designs can be prepared for downstream work with AlphaFold or molecular dynamics.

The caveats DeepMind states, and the ones it doesn't

To its credit, the post includes a frank limitations paragraph. DeepMind cites sparse data, inter-lab assay noise and narrow dynamic ranges as constraints on predictive precision, and states plainly that computational designs are testable hypotheses rather than finished drugs, and that every proposal must be synthesized and validated experimentally.

Several things are absent. No accuracy or error figures are reported for the model benchmarking step — the post describes statistical and visual outputs for the scientist to assess, but publishes none. No comparison is offered against a human chemist or an existing automated pipeline running the same task, so the "multiple days to one hour" framing is an internal estimate rather than a measured baseline. The post does not disclose which underlying model powers the agent, how many runs were attempted before the documented one, or how often the agent produced invalid structures or unworkable routes. No wet-lab results are reported for the proposed macrocycles. And because the chosen case study has a known successful outcome in lorlatinib, the exercise cannot distinguish genuine design capability from reconstruction of a pattern already present in the training data and the retrieved literature.

For teams evaluating this, the practical question is narrower than drug design: how much of the hour saved is data wrangling versus reasoning? DeepMind's own framing suggests the bulk of the value is in curation, deduplication and provenance tracking — unglamorous work that agents are genuinely suited to, and where errors are at least auditable against source documents.

Distribution, and the pattern behind it

The Cheminformatics Skills are available now through the Antigravity app with the Science plugin installed, DeepMind says, and to enterprise users who sign in with Google Cloud credentials via the "Use Business Account" option. The company is also offering the highest Antigravity usage tier free for six months to 10,000 scientists, onboarded on a rolling basis and subject to eligibility, capacity and selection; applications are open. Eligibility criteria, the tier's normal price and what happens after six months were not disclosed.

This follows a related DeepMind post from October 2026 on using Antigravity to search for antimicrobial peptides, which suggests a deliberate strategy: rather than shipping one monolithic science model, Google is packaging domain tooling as installable skills on a general agent surface and seeding adoption with free capacity. That is a distribution play as much as a research one, and it puts Antigravity in competition with the entrenched commercial cheminformatics stacks that already sit in pharma IT. Whether chemists move depends less on demo speed than on reproducibility, audit trails and whether proposed compounds survive contact with a synthesis lab.

What to watch

• Whether DeepMind or external groups publish wet-lab results for agent-proposed designs, and the hit rate when they do.
• Any quantitative evaluation of structure extraction accuracy from papers, patents and catalogs — the step with the least error tolerance and no reported metrics.
• Whether the free usage-tier program publishes eligibility criteria and uptake numbers, and what pricing looks like after the six-month window.
• How enterprise pharma handles data governance when curated internal compound sets meet a cloud-hosted agent; the post addresses access via Google Cloud credentials but not data handling.
• Whether the skills pattern extends to further domains beyond cheminformatics and antimicrobial peptides, which would signal a platform strategy rather than a series of one-off demos.