Google adds chemistry tools to Antigravity, with lab tests still ahead

A workbench for molecular information

Google DeepMind introduced Cheminformatics Skills for its Antigravity application on November 3. Cheminformatics means using computers to organise and analyse information about molecules. The announcement describes an agent gathering published compounds, comparing structures, building prediction models and proposing possible new designs.

Google reports completing its worked example in about an hour. This is a developer demonstration, not an independently reproduced speed comparison. The proposed molecules remain hypotheses that need to be made and tested.

The calculation tools matter

The workflow assigns coordination and database navigation to a language model, while established chemistry software performs molecular calculations. Google names RDKit as a foundation.

RDKit’s documentation describes an open-source toolkit with two- and three-dimensional molecular operations and descriptors for machine learning. Its Python guide shows how software represents individual atoms, bonds and ring systems. It also provides structure checks that can flag invalid representations.

Those functions answer a different question from whether a compound will work in a laboratory. A valid molecular representation lets software analyse the intended structure. It is not, by itself, a measurement of biological activity. Keeping that distinction visible helps readers understand what the agent has actually produced.

A prediction needs an experimental destination

Google’s example compares related compounds and explores closing flexible structures into rings. The resulting designs include suggested synthesis routes and predicted potency. The company itself notes that limited data and differences between laboratories constrain prediction.

An informative handoff would preserve the source records, analysis code, model assumptions and proposed experimental checks. That makes a suggestion inspectable rather than merely convincing. It also lets scientists distinguish a retrieved measurement from a value generated by a model.

The announcement says the skills are available through Antigravity’s Science plugin. Availability is a product fact; useful scientific performance remains something to evaluate for each dataset and task. The immediate story is assistance with analysis and hypothesis preparation, not a shortcut around experimental evidence.