Why the split of labor is the real design choice
A molecule tolerates no approximation. Change one atom or one bond and, as DeepMind puts it, you have described a different compound altogether. So the skills do not ask the language model to do chemistry arithmetic. The model handles workflow orchestration, database navigation and scientific reasoning. The calculations (property prediction, reaction handling, structure manipulation) go to deterministic tools built on RDKit, the open-source cheminformatics toolkit, with connectors to the curated ChEMBL and PubChem databases. That separation is the part of this announcement most likely to outlast the demo. The model proposes and explains; deterministic tools perform the molecular calculations and structure operations. Experimental validation remains the chemist's job.
The hour, read carefully
The demonstration uses a medicinal-chemistry technique called rigidification. DeepMind's comparison starts with crizotinib, a non-small cell lung cancer drug with five major rotatable bonds, and lorlatinib, another NSCLC drug whose structure is cyclized and has effectively zero. The agent uses this structural contrast to investigate how reducing molecular flexibility relates to potency in the demonstration. DeepMind says its matched-pair analysis found that cyclization generally reduced rotatable bonds and increased potency. DeepMind also says it validated its potency models on a held-out set using scaffold and chronological splits, across four experimental readouts. It then proposed new rigidified macrocycles, predicted each design's potency with the models it had just trained, and supplied synthesis routes and reaction-mechanism diagrams.
One hour is DeepMind's figure for its own walkthrough, set against its own "multiple days" baseline. There is no independent timing, no comparison against a named human team on the same task, and no error rate for the proposed routes. Read it as a demonstration of how much of the data work can be compressed, not as a measured speedup.
What an hour does not buy
DeepMind states the boundary itself, and it is worth quoting in substance: sparse data, assay noise between labs and narrow dynamic ranges limit how precise the predictions can be; computational designs are testable hypotheses, not finished drugs; every proposed design must be synthesized and validated experimentally. The expensive, slow part of drug discovery (making the molecule, testing it, finding out the prediction was wrong) is exactly the part this release does not touch. What it shortens is the time a chemist spends wrangling data before that part begins.
The skills are available now: install the Science plugin in the Antigravity app, or, as an Enterprise user, sign in with a business account and Google Cloud credentials. Google is also offering the highest usage tier free for six months to 10,000 scientists, onboarded on a rolling basis and subject to eligibility, capacity and selection (terms as stated on November 3, 2026).
Sources
- [1] Google DeepMind — "Agentic Cheminformatics with Google Antigravity," November 3, 2026 (primary source, read first-hand in full at Gate 1: Cheminformatics Skills in the Antigravity app via the Science plugin, with an Enterprise path via "Use Business Account" and Google Cloud credentials; literature extraction over thousands of compounds, discarding buffers and solvents, with sources documented and an interactive viewer; matched molecular pair analysis; potency models validated on a held-out set with scaffold and chronological splits across four experimental readouts; crizotinib with five major rotatable bonds and lorlatinib cyclized to effectively zero, both non-small cell lung cancer treatments; the model orchestrating and reasoning while deterministic RDKit-grounded tools, with ChEMBL and PubChem connectors, perform the molecular calculations; new rigidified macrocycles proposed with predicted potency, synthesis routes and electron-pushing diagrams consistent with documented procedures; results exportable to industry-standard chemistry software; "in just an hour" set against the workflow "typically takes multiple days for a single iteration" — both DeepMind's own framing; the stated limits — sparse data, inter-lab assay noise and narrow dynamic ranges limit predictive precision, and computational designs are testable hypotheses, not finished drugs, each requiring synthesis and experimental validation; the offer of the highest Antigravity usage tier free for six months to 10,000 scientists, onboarded on a rolling basis, subject to eligibility, capacity and selection)Read source
- [2] Google Antigravity — product page (linked from the primary source; Science plugin install path as stated in the post)Read source