Google DeepMind announced chemistry skills for Antigravity on November 3, 2026, aimed at chemists and drug-research teams gathering evidence and designing candidate molecules. Available through its Science plugin, the tools connect computational research tasks rather than deliver a finished medicine.
The immediate question for a laboratory is which work to delegate. A source-linked compound list can be checked against existing records; a prediction about a new molecule needs a different kind of evidence.
What the tools can take on
Cheminformatics means using computers to organize and analyze chemical information. In Google’s description, the language model coordinates the work while chemistry software performs molecular calculations. RDKit, the underlying open-source toolkit, already supports molecular operations and numerical features for machine learning.
Match names to records. The documented PubChem connector can resolve chemical names to database identifiers, retrieve properties and find similar structures. Some returned properties are calculated values, not laboratory measurements.
Gather experimental data. The ChEMBL connector supports searches for molecules, biological targets and activity measurements, as well as structure-file exports and conversion of measurement units.
These are documented capabilities, not guarantees that an agent chose the correct compound or comparison. For an initial evaluation, that makes a small, familiar public dataset more useful than an open-ended request to discover a drug.
Matching units is not matching experiments
The ChEMBL connector can convert activity measurements to nanomolar units, giving concentrations a common scale. But ChEMBL’s own guidance distinguishes tests of target binding, biological effects and toxicity. It also records confidence in target assignments and flags potentially problematic data.
The practical implication is that two numbers can share a unit without answering the same scientific question. Before ranking compounds, a chemist should check what was measured, which target was involved and whether flagged records belong in the comparison. Correct conversion cannot make unlike experiments interchangeable.
What Google’s one-hour demonstration establishes
Google reports a one-hour workflow that gathered thousands of compounds, compared related molecules, built models to predict potency, or strength of effect and proposed ring-shaped molecules with possible synthesis routes. The example drew on research involving the cancer drugs crizotinib and lorlatinib.
That is a company-reported demonstration, not an independently established saving for other laboratories. Google acknowledges limits from sparse data and differences between experiments, and says proposed designs still need to be made and tested.
Even laboratory results are not the same as patient benefit: the FDA’s explanation of clinical research distinguishes preclinical findings from studies in people. For the next research stage, see how laboratory feedback can revise AI-generated molecular designs.
Access and a sensible first task
The Science plugin is available through Antigravity’s in-app Customizations marketplace. Google lists a $0-per-month individual tier with weekly limits; quota consumption depends on the work performed, and Pro or Ultra users can opt into paid credit overages. Those terms do not establish the cost of a chemistry workflow.
Separately, Google offers six months at its highest usage tier to 10,000 scientists, with rolling onboarding subject to eligibility, capacity and selection. It is an application-based offer, not unlimited free access for every research team.
A sensible pilot would ask for a compound table built from known public records, with identifiers, source links and the original measurement context preserved. Have a chemist check it before moving on to predictions.
Record correction effort, data quality and actual usage alongside completion time. That would let a team decide whether the tool helps prepare trustworthy research inputs before committing resources to candidate design.