On November 3, Google DeepMind introduced a set of cheminformatics skills for Google Antigravity. The agent can use local tools to process chemical structures, query scientific databases and prepare datasets for analysis. The package is available through the Science plugin in the Antigravity app. The public Science Skills repository describes a wider collection of instructions, scripts and connections to research resources.
From published literature to a dataset
The approach builds on RDKit, an open-source toolkit for chemical structures, and connects the agent to resources such as ChEMBL and PubChem. When investigating a compound class, a researcher can ask the agent to find structures, associate them with reported measurements and produce a table for further inspection. Separate skills support representing and checking reactions. Using specialized tools for structures makes the workflow less dependent on whether a language model can reproduce every detail correctly from prose.
DeepMind's demonstration looked for compounds related to crizotinib and lorlatinib. The agent retrieved candidates from the literature, removed items such as solvents and associated the remaining structures with their data. An interactive viewer lets scientists inspect individual molecules and compare properties such as potency and rotatable bond count. The example is a developer-reported workflow on one specific topic. Its results depend on the inputs and on the way the dataset was curated.
Why the data split matters
The demonstration includes comparisons between related molecules with a specific structural difference and models for predicting potency. DeepMind describes evaluating the models on held-out data. When dates are available, a chronological split can train on older records and test on more recent ones. A scaffold split asks a different question. It tests how the model behaves on molecular families less similar to those it has seen in training.
A useful evaluation for research teams is to trace a few selected compounds back to the original data tables. Check identities, units and whether the reported measurements come from comparable experiments. If close analogues are present on both sides of a split, an aggregate score may look stronger than performance on genuinely new chemistry. Researchers still own those checks when an agent has done the preparation.
What is available now
DeepMind says the skills are available through the Science plugin. The company also announced a program offering 10,000 scientists access to its highest usage tier for six months, subject to eligibility, capacity and selection. That program is separate from the availability of the plugin itself.
The useful output is a set of traceable hypotheses and candidate structures for follow-up work. Sparse data, variation between laboratories and differences in measurement methods can all limit predictions. DeepMind explicitly says that proposed compounds still require synthesis and experimental validation. The account of where the data came from and how it was checked is therefore part of the value of the workflow, alongside the time saved on manual preparation.
