Antigravity chemistry skills: setup, evidence and access
Google DeepMind’s November 3 cheminformatics walkthrough shows Antigravity organizing chemical literature, comparing compounds and proposing designs through its Science plugin. Google reports completing the demonstrated workflow in about an hour. The outputs remain research hypotheses requiring synthesis and experimental validation. ByteForward has not reproduced the demonstration.
What Google demonstrated
The DeepMind walkthrough combines a language model’s coordination with deterministic chemistry tools, including RDKit and database connections to ChEMBL and PubChem. It compares related molecules, builds predictive models and suggests candidate designs. Google identifies sparse data, variation between laboratories and limited measurement ranges as constraints. The reported speed does not establish clinical benefit or a general success rate.
How to enable the Science plugin
Google’s Build with Google documentation describes Science as a curated bundle of scientific skills. Bundles can be selected during onboarding or managed afterward in settings. That packaging matters: adding specialized instructions and tools is a different change from replacing the underlying model.
The public Science Skills repository gives the specific path: Customizations, Build with Google, Science, Add. Its instructions recommend the latest Antigravity version. Each skill packages instructions, helper scripts and optional references. Researchers can inspect these materials rather than treating a successful-looking answer as the entire record of the work.
Setup can involve more than clicking Add. The repository uses uv for dependency management and says the agent requests approval to install it on first use. Some skills require an API key, while others work without one but gain higher limits with a key. Those requirements depend on the selected skill. The repository also says third-party data licenses and terms remain applicable, and labels the repository as not an official Google product.
Our Antigravity Marketplace guide explains the broader distinction between installing a bundle and configuring the services it uses. For a lab, that suggests checking dependencies, connected accounts and permitted data before assigning a real project.
What to retain from an evaluation
A useful trial should produce an inspectable record. Keep the source documents, compound identifiers, data-cleaning decisions and generated scripts together. Record which measurements were observed and which values were predicted. Ask for the evaluation split and the assumptions behind it, then have an appropriate domain expert examine the results. These are recommended review steps, not tests we performed.
A polished visualization helps a researcher ask better questions, but it cannot resolve a mistaken structure or an incompatible measurement. Teams should decide in advance what evidence would justify moving from a computational suggestion to an experiment. That decision should remain visible in the workflow rather than being hidden inside a final summary.
Free access depends on selection
Google’s science access page links an application for 10,000 scientists. The November 3 announcement specifies six months on the highest usage tier, with rolling onboarding subject to eligibility, capacity and selection. Applying is not a guarantee of acceptance. Check the program’s current requirements before planning work around the offer.
