Google Antigravity cheminformatics skills: an agent that designs drug candidates in an hour, and what that claim leaves out

On November 3, 2026, Google DeepMind published a post titled "Agentic cheminformatics with Google Antigravity." It describes a set of Cheminformatics Skills for Antigravity, Google's agent workspace, that help chemists curate data, train predictive models and propose molecules. The authors say the agent completed a workflow that "typically takes multiple days" in about an hour. That is a striking claim, and it is worth separating the work the agent did from the claim about what the work means.

Searches for Google Antigravity cheminformatics skills turn up the same demo everywhere, because there is one demo. This piece reads it closely.

What the demo actually did

The test case is the lung cancer drug crizotinib and its descendant lorlatinib. Crizotinib has five major rotatable bonds. Lorlatinib is cyclized at its end, so it has effectively zero. The question the agent worked on is whether stiffening a molecule this way tends to raise potency.

According to the DeepMind post, the agent did several things in sequence:

• Searched academic and patent literature for macrocyclic and acyclic analogues of the two drugs, extracted the molecules and linked them to reported data, filtering out non-drug entries such as buffers and solvents.
• Retrieved thousands of compounds, of which an interactive viewer displayed 568 curated ones.
• Ran a matched molecular pair analysis, comparing pairs that differ only in whether they are cyclized. The result: cyclization reduces rotatable bonds while generally increasing potency, and the highest-potency compounds were the most rigid.
• Built many potency models with varied features and architectures, validated on a held-out set using a scaffold split.
• Proposed new macrocycles with synthesis routes and predicted potency. One example is a 12-membered cyclic ether labeled LOR-DES-3.

Nothing in that list is exotic to a medicinal chemist. The finding that rigid analogues are often more potent is a known design principle, and the post presents it as a demonstration of the workflow, not as a new result.

The one-hour claim

The speed figure is the headline, and it is the least verifiable part. It comes from the authors describing their own run. Remio, a note-taking vendor whose blog covered the launch, notes that the comparison with a multi-day workflow is "a company account, rather than an independently established productivity benchmark," and points out that speed measured only while the software runs may leave out the expert time spent reviewing and correcting the output.

That critique is fair. A chemist still has to decide whether the extracted compounds are right, whether the assay data are comparable across labs and whether the model's potency predictions deserve trust. The DeepMind authors say as much: sparse data, assay noise between labs and narrow dynamic ranges limit predictive precision.

What would change the verdict

The post reports no experimental results. The designs are predictions from models the agent trained on a few hundred public data points, and the authors say they must be validated experimentally. Three things would raise our confidence:

• A synthesized compound from the agent's list, measured against the predicted potency.
• A comparison with a human chemist given the same data and a time limit, scored blind.
• Independent evidence that tool-using chemistry agents beat the underlying model. Remio cites a finding that the tool-augmented ChemAgent "did not consistently outperform its underlying language models," which is the benchmark question this release has not yet answered.

Earlier systems set the context. Remio names ChemCrow, published in May 2024 with 18 integrated tools, and Coscientist, published in December 2023, as precedents for chemistry agents.

Access and incentives

Google's Antigravity science page says the Science plugin installs in one click, connects the agent to more than 30 life-science databases including UniProt, PubChem and ChEMBL, and offers access to DeepMind tools such as AlphaFold database queries. The company is giving 10,000 scientists its highest usage tier of Antigravity and its Gemini 4 Argon model free for six months, subject to eligibility, capacity and selection. That is a developer-adoption program as much as a research program, and it is a reasonable thing to say plainly.

How this compares with other labs' science pushes

Anthropic has made a parallel move, covered in our report on Claude's drug development partnerships. The pattern across labs is the same: agents that read literature, run standard tools and propose candidates, with validation left to partners. The differences are in how much experimental evidence each lab publishes.

Reading the numbers in the post

Several figures in the DeepMind write-up sound more precise than they are. "Thousands of compounds" were retrieved, but the viewer showed 568 after curation, so most of the raw yield did not survive filtering. The models were validated on a held-out set using a scaffold split, which tests whether predictions generalize to new core structures, a harder test than a random split. That is good practice. It is still a small public dataset assembled from papers, where assay conditions differ between labs.

The proposed design, LOR-DES-3, a 12-membered cyclic ether derived from a precursor labeled LOR-PRE-3, comes with synthesis routes, mechanism diagrams and a predicted potency. The post notes these designs can be passed to AlphaFold or molecular dynamics for further checks. That is a pipeline of predictions feeding predictions, which is why the call for experimental validation is the sentence that matters most.

Our take

Judge this as a workflow tool. Mining literature, building a viewer and fitting models quickly is useful, and the authors are honest that outputs are hypotheses. But "designs a drug in an hour" is not what the post shows. It shows a plausible, fast, unvalidated first pass on a well-studied family. The next signal to watch is a wet-lab result from any Antigravity user, not another demo.