Google Antigravity Cheminformatics Skills: What Launched and How to Use Them

Google DeepMind introduced Cheminformatics Skills for Antigravity on November 3. The update helps researchers connect chemical literature, molecular datasets and analysis tools in an agent workflow. It targets the preparation and analysis around drug discovery, with researchers still responsible for checking results and validating proposed compounds experimentally.

What the chemistry skills do

In Google's demonstration, the agent searches thousands of compounds, compares related molecular structures and proposes new candidates and synthetic routes in roughly an hour. That is a company demonstration, not an independently established time saving for every research project.

The workflow separates tasks. The language model coordinates research and analysis, while RDKit performs deterministic molecular calculations and databases such as ChEMBL and PubChem supply chemical records. The agent can also prepare predictive models and evaluate them with chronological or scaffold-based data splits.

That separation is useful. Asking a model for a plausible molecule and asking it to run a reproducible calculation are different operations. A researcher needs to know which operation produced an answer before deciding how much weight to give it.

How to install the Science plugin

The skills are available through Antigravity's Science plugin. The official installation instructions say to install the latest Antigravity app, open Customizations, select Science under Build with Google, and click Add.

The repository describes skills as instructions, scripts and supporting resources. They give the agent a repeatable way to use specialist tools rather than requiring a researcher to explain each tool from scratch. Some workflows need additional credentials: AlphaGenome requires an API key, for example. Other services can work without a key but offer higher limits with one.

The first use can request permission to install the uv dependency manager. The documentation recommends restarting Antigravity after that setup. It also warns that plugin updates overwrite the installation directory, so researchers should keep their own modifications elsewhere.

The repository also says its code is not an official Google product. Access to the files should not be read as a guarantee of production support.

These details matter more than a polished demo when deciding whether a workflow can be handed to a lab colleague. Installation, credentials and reproducible environments are part of the product experience.

What access is available

The Antigravity for Science page describes integrations with more than 30 life-science databases and tools, including UniProt, PubChem and ChEMBL. The launch announcement also offers an access program giving 10,000 scientists the highest usage tier free for six months. Enrollment is subject to eligibility, capacity and selection, rather than an unrestricted free tier for everyone.

The chemistry announcement does not establish a separate per-analysis API price. Researchers should distinguish access to the plugin from the usage limits and terms of the Antigravity account running it, as well as any external services it calls.

What still needs a scientist

Google explicitly notes sparse data, noisy assays and narrow measurement ranges as limits on predictions. Proposed compounds still need synthesis and experimental testing. Generating a route does not establish that it will work in a particular lab.

The useful question is whether the workflow leaves enough evidence to inspect. Which records were included? Were incompatible measurements combined? Did the evaluation separate genuinely unfamiliar structures from close relatives of the training examples?

Those checks can change a result without changing the underlying AI model. An impressive prediction built on a weak dataset remains a weak prediction, even if the agent assembled it quickly.

For research teams, the promising shift is less time spent moving between papers, database queries and scripts. If that works reliably, the bottleneck moves toward choosing which hypotheses deserve an expensive experiment. The strongest evidence will be workflows that survive that handoff, not simply workflows that produce more candidates.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn research workflows and product launches into clear stories that developers and founders understand and share.

Book a 30-minute call.