Google Antigravity adds chemistry skills for research workflows

A workbench for moving between chemistry tools

Google DeepMind introduced Cheminformatics Skills for Antigravity on November 3. The workflow combines language-model orchestration with deterministic chemistry tools, including RDKit, and resources such as ChEMBL and PubChem. The announcement demonstrates literature gathering, compound analysis and computational design proposals.

Google says the skills are available through Antigravity’s Science plugin. It also describes a program offering selected scientists six months of its highest usage tier, subject to eligibility and capacity. The demonstration produces hypotheses for research; it does not establish that proposed molecules are effective medicines. Source: Agentic cheminformatics with Google Antigravity

The integration is more interesting than the stopwatch

Our analysis: the useful architectural choice is to give different jobs to different components. An agent can organize the workflow and interpret the researcher’s request, while specialized software performs calculations that should be reproducible. That division makes it easier to inspect where an answer came from.

A scientist still needs to know which structures, measurements and assumptions entered the calculation. If a molecule was extracted incorrectly from a source, a perfectly executed downstream tool can produce a precise answer about the wrong compound.

That is why a polished dashboard is not sufficient evidence. A valuable workbench should make it easy to move backward from a plotted point to the underlying record and its source. The ability to inspect provenance is part of the scientific utility, not merely an administrative feature.

Data preparation creates consequential choices

Combining records from several sources requires decisions about identity, units, missing values and duplicate entries. Those choices can alter the apparent pattern in a dataset. An agent that makes them silently may produce an attractive analysis that is difficult to audit.

Our recommendation is to request an explicit transformation log. It should identify which records were excluded, which values were standardized and which ambiguities remain unresolved. A researcher can then review the consequential choices before using the prepared dataset to compare candidates.

The log is also valuable when a result looks surprising. Instead of rerunning an entire conversation, the team can inspect the specific stage at which the dataset changed. That turns an agent-assisted workflow into a more reproducible computational process.

Prediction is a way to prioritize experiments

A computational proposal can help decide what to investigate next without being a proven result. The practical question is whether the proposal makes a better experiment possible: a clearer comparison, a more informative measurement or a more efficient use of laboratory resources.

Predicted performance should therefore remain visibly separate from measured performance. Mixing those values in a table without clear labels can make a generated candidate appear better established than it is. The same distinction should survive exports and presentations.

A useful review asks what observation would disprove the proposal. If no plausible experiment could change the conclusion, the output is closer to a persuasive narrative than a scientific hypothesis. Agents should make uncertainty easier to examine, not hide it behind confident language.

How a research team could evaluate the workflow

Start with a bounded retrospective task whose source records and expected checks are already understood. Compare the prepared dataset with an independently reviewed reference. Record extraction errors and unresolved cases, rather than only the time taken to generate a report.

Then examine whether a second researcher can reproduce the analysis from saved inputs, scripts and tool versions. If reproduction requires guessing what happened in an earlier chat, the workflow needs stronger records before it becomes dependable.

Finally, assess the human work that remains. Faster initial preparation may still be worthwhile even when careful review is necessary, but the net benefit should include that review time.

Antigravity’s chemistry skills illustrate a concrete direction for scientific agents: connecting established tools into an inspectable workflow. The strongest outcome would be better-supported decisions about the next experiment, with a clear boundary between computational suggestion and experimental evidence.