Google Antigravity gets Cheminformatics Skills; agent runs drug design workflow in an hour, DeepMind says

Google DeepMind has introduced a suite of Cheminformatics Skills for Google Antigravity. According to the company, an agent using the skills can search thousands of compounds, run matched molecular pair analysis and propose novel rigidified macrocycles with synthetic routes in about an hour. DeepMind says a single iteration of this workflow typically takes multiple days. The announcement was published November 3, 2026 by Zoi Tsangalidou, Ruijie (Kelly) Zhang, Sukhdeep Singh and Michael V. LeVine.

How the skills are built

DeepMind describes a division of labour. The underlying language model handles workflow orchestration, database navigation and scientific reasoning. Deterministic chemistry tools execute the molecular calculations. The skills are grounded in RDKit, the open source cheminformatics toolkit, and include connectors to curated resources from ChEMBL and PubChem. They encode domain best practices directly into the agent's workflow.

The stated motivation is data wrangling. The same molecule can appear as an IUPAC name in a paper, a Markush structure in a patent or a structure-data file (SDF) in a catalog. DeepMind notes that altering a single atom or bond describes a different compound, so the task has no margin for error.

The case study: rigidifying a kinase inhibitor

The walkthrough uses rigidification, a medicinal chemistry technique that adds bonds to limit a molecule's conformations and reduce the entropic penalty of binding. The example pair is crizotinib, a first-generation tyrosine kinase inhibitor (TKI) for certain forms of non-small cell lung carcinoma, and lorlatinib, a third-generation TKI. Crizotinib has five major rotatable bonds. Lorlatinib is cyclized at its end, which DeepMind says reduces the effective number of rotatable bonds to zero and leads to higher potency among other effects.

The reported agent workflow:

• Literature retrieval: Within minutes, the agent retrieves thousands of compounds, filters out irrelevant molecules such as buffers and solvents, and documents sources in a report.
• Initial analysis: The report notes that the macrocycles are largely predicted to be brain-penetrant. DeepMind flags this as important because the target lung cancer frequently metastasizes to the brain.
• Interactive viewer: The agent launches a viewer showing curated compounds, associated data and customizable scatter plots. Plotting potency against rotatable bonds, colored by macrocycle status, shows that macrocycles like lorlatinib tend to have fewer rotatable bonds and higher potency.
• Matched molecular pair analysis (MMPA): The agent had independently suggested this analysis. It compared many pairs of molecules differing only by whether they are cyclized. DeepMind reports that cyclization reduces rotatable bonds while generally increasing potency.
• Predictive modeling: The agent built a large number of potency models with varying feature sets and model architectures and validated them on a held-out set. The skills guide the agent toward practices such as chronological splits, training on older compounds.

The post concludes with the agent proposing novel rigidified macrocycles with synthetic routes, per the summary at its top.

What this means

The design separates probabilistic reasoning from exact chemistry. The language model plans and interprets, while RDKit-based tools perform the molecular operations. That addresses the failure mode DeepMind identifies, where a single altered atom or bond yields a different compound. Encoding practices such as chronological splits into the skills means the agent's modeling choices follow documented domain conventions rather than ad hoc defaults.

The timing figures (about an hour versus multiple days) come from DeepMind's own demonstration on one lung cancer case study, not from an independent evaluation. The post does not give benchmark scores or pricing for the skills in the text reviewed. Whether results on crizotinib and lorlatinib carry over to other targets is not addressed there.