Google Antigravity gets Cheminformatics Skills for drug design

Google DeepMind has published a blog post presenting a suite of Cheminformatics Skills for Google Antigravity. The post says designing new medicines means curating complex data from charts, figures and literature, training predictive models and proposing candidate molecules to synthesize, and that one iteration of this workflow typically takes multiple days. In the demo, an agent equipped with the skills searches through thousands of compounds, runs matched molecular pair analysis and proposes novel rigidified macrocycles with synthetic routes in just an hour. That timing is the post's own claim for the demo.

The pitch starts from a data problem. The same molecule can show up as an IUPAC name in a paper, a Markush structure in a patent or a structure-data file (SDF) in a catalog, and changing a single atom or bond describes a different compound. DeepMind says it built the skills so scientists can focus on hypotheses rather than data wrangling, turning Antigravity into a collaborative science workbench with full access to local tooling, for both computational chemists and experimental bench chemists. The work is split: the language model handles workflow orchestration, database navigation and scientific reasoning, while deterministic chemistry tools execute the molecular calculations. The skills are grounded in RDKit, the open source cheminformatics toolkit developed over two decades by a global community of contributors, with connectors to curated resources from ChEMBL and PubChem. They encode domain best practices directly into the agent's workflow.

The walkthrough covers rigidification, a classic medicinal chemistry technique for improving potency. Drug molecules often link substructures through rotatable bonds, which lets them twist into many 3D conformations. To bind its target, a floppy molecule must lock into one active shape, and losing that freedom carries an entropic penalty that weakens binding affinity. Chemists often add new bonds to rigidify the molecule, a strategy the post says is often called "paying the entropic cost in advance", and which can dramatically boost potency. The case study is crizotinib, a first-generation small-molecule tyrosine kinase inhibitor used for certain forms of non-small cell lung carcinoma, which has five major rotatable bonds. Lorlatinib, a third-generation TKI that resembles crizotinib, is cyclized at its end, which reduces the effective number of rotatable bonds to zero and leads to higher potency among other effects.

The agent first searches the literature for these drugs and their analogues, extracts molecules from the documents, associates them with the reported data and prepares the compounds for downstream analysis. It then launches an interactive viewer with the curated compounds, their 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.

Next comes matched molecular pair analysis (MMPA), which isolates the effect of a small, discrete change by comparing pairs of molecules that differ only by that change. The post says the agent had independently suggested this analysis. It compared many pairs that differ by whether or not they are cyclized, and the results show that cyclization reduces the number of rotatable bonds while generally increasing potency, as expected. The agent also built a large number of potency models with varying feature sets and architectures and validated them on a held-out set. The skills guide how it builds such models: it will often use chronological splits, training on older compounds to predict newer ones, when dates are available, and can fall back to scaffold splits to better estimate generalizability.

Finally, the researchers ask the agent to design new, improved macrocycles from what it learned. It recommends multiple reactions and synthesis routes involving different compounds, and can prepare those compounds for further analysis with AlphaFold or physics-based techniques such as molecular dynamics simulation. Results can be exported in various formats for loading into many industry standard chemistry software packages.

The post closes with a limit. Sparse data, inter-lab assay noise and narrow dynamic ranges limit the predictive precision of even the most powerful models, and computational designs are testable hypotheses, not finished drugs. A drug discovery scientist can generate more strongly-supported design hypotheses faster, DeepMind says, but every proposed design must be synthesized and validated experimentally.

Availability: the skills are available on Google Antigravity today. Users download the Antigravity App and install the Science plugin; Enterprise users can log in with Google Cloud credentials by choosing "Use Business Account". DeepMind is also offering 10,000 scientists the highest usage tier in Antigravity free of charge for 6 months, onboarded on a rolling basis and subject to eligibility, capacity and selection. Applications are open.

Key facts

• Google DeepMind added Cheminformatics Skills to Google Antigravity, built on RDKit with connectors to ChEMBL and PubChem.
• In the demo on crizotinib and lorlatinib, the agent searches thousands of compounds, runs matched molecular pair analysis and proposes rigidified macrocycles with synthetic routes in an hour, against multiple days per iteration without the tool (the post's own claim).
• Work is split: the language model orchestrates, navigates databases and reasons, while deterministic chemistry tools run the molecular calculations.
• The skills are available now via the Antigravity App and its Science plugin; 10,000 scientists are offered the top usage tier free for 6 months, subject to eligibility, capacity and selection.
• DeepMind says computational designs are testable hypotheses, not finished drugs, and every proposed design must be synthesized and validated experimentally.

Why it matters

Drug design depends on pulling data out of papers, patents and catalogs, where one molecule can appear as a name, a Markush structure or an SDF file, and a single changed atom means a different compound. DeepMind says a single iteration of the workflow typically takes multiple days, and the demo compresses it to about an hour. The notable design choice is the split of labour: the language model orchestrates and reasons, while deterministic RDKit-based tools do the molecular calculations, so the arithmetic of chemistry is not left to the model.

Who it affects

The post names two groups: computational chemists and experimental bench chemists, who get a shared workbench in Antigravity. Drug discovery scientists are the target users. Enterprise users can sign in with a Google Cloud account, and 10,000 scientists are offered free top-tier access for 6 months.

How to use it

Download the Antigravity App and install the Science plugin to start using the skills. Enterprise users can choose "Use Business Account" and log in with Google Cloud credentials. The agent can search literature, curate compounds, open an interactive viewer with scatter plots, run matched molecular pair analysis, build potency models, and propose macrocycles with synthesis routes. Results can be exported in various formats for industry standard chemistry software. The free highest-tier offer for 10,000 scientists lasts 6 months, is onboarded on a rolling basis, and is subject to eligibility, capacity and selection; applications are open.

How solid is it

This is a first-party DeepMind blog post describing a single walkthrough on two well-known drugs. The one-hour figure is the post's own claim for the demo; no independent measurement or comparison methodology is described. No accuracy or benchmark figures are given for the potency models, and the post gives only approximate counts: thousands of compounds, many pairs, a large number of models. The post does not name which language model powers the agent. The proposed macrocycles are not reported as synthesized or tested.

Risks and caveats

DeepMind itself says sparse data, inter-lab assay noise and narrow dynamic ranges limit the predictive precision of even the most powerful models. Computational designs are testable hypotheses, not finished drugs, and every proposed design must be synthesized and validated experimentally. The cyclization finding is a tendency: potency generally, not always, increases. The post does not say the agent replaces chemists; it frames the tool as a collaborative workbench. The free-tier offer is conditional on eligibility, capacity and selection.

“paying the entropic cost in advance”

— Google DeepMind blog post, on the rigidification strategy in medicinal chemistry