Google's new chemistry agent shrunk a drug discovery step from days to an hour

Key points

• Google DeepMind says its Antigravity agent, running new Cheminformatics Skills, completed a drug design loop on real lung cancer compounds in about an hour, a task it says normally takes multiple days.
• The demo rebuilt the known path from crizotinib (five rotatable bonds) to lorlatinib (zero rotatable bonds), two approved non-small cell lung cancer drugs, using matched molecular pair analysis.
• The skills are built on RDKit, the open source chemistry toolkit developed over two decades, with connectors to the public ChEMBL and PubChem databases.
• The agent proposed new rigidified macrocycle candidates with synthesis routes and potency predictions from models trained during the same session.
• No clinical results, no new approved drug, and no pricing or availability for the Skills were disclosed.

Google has been quietly pushing its AI into one of the slowest, most expensive corners of science: finding new medicines. Its latest move is an agent, software that carries out multi-step tasks on its own, that behaves like a junior computational chemist.

The company published a technical walkthrough showing Antigravity equipped with new Cheminformatics Skills completing a real lung cancer case in roughly an hour. It's a task Google says normally takes a working chemist several days.

What did the agent actually do?

It rebuilt, in miniature, a famous success story in cancer drug design. Crizotinib and lorlatinib are both approved treatments for non-small cell lung cancer. Lorlatinib is the newer one, and chemists made it more potent by locking a floppy molecule into a single rigid shape, a technique called rigidification.

The agent pulled thousands of related compounds from the scientific literature, filtered out solvents and buffers, and lined the molecules up against their reported potency. It then ran matched molecular pair analysis, a standard medicinal chemistry method that compares pairs of molecules differing by one small change, confirming what the published record already shows: cyclising the molecule cuts rotatable bonds and tends to raise potency.

After that it trained several small machine learning models to predict potency and benchmarked them, then proposed new macrocycle designs with synthesis routes and predicted potency scores. A human chemist stayed in the loop through an interactive viewer, clicking molecules and plotting data.

Is this a new drug?

No. This is a reconstruction of a known problem with a known answer, which is the honest way to test a tool like this. Nothing in the post suggests a new approved medicine or a validated hit against a fresh target.

That matters. Drug discovery has a long graveyard of software that looked brilliant on retrospective cases and fell flat on new ones. The useful question is whether chemists at pharma companies and university labs find the hour-long loop reliable on problems where the answer isn't already in the textbook.

How does this fit with the rest of Google's AI push?

We first covered Antigravity on 15 October, when the agent hunted for antibiotics inside protein 'dark matter' and produced a lead now headed to the lab bench. The chemistry work also sits next to AlphaFold, the protein structure system from Google DeepMind that the agent can hand its designs off to.

It also lands in a tighter spending climate. We reported last week that Microsoft and Google trimmed their AI spending forecasts, which puts pressure on these science demos to turn into paying customers. Pharma is one of the few buyers that can actually afford the compute bill.

What should a working chemist or patient take from this?

For patients, nothing changes today. The drugs in the demo are already on pharmacy shelves, and any agent-designed candidate would still face a decade-long path through safety and efficacy trials.

For computational and bench chemists, the practical read is narrower and more interesting. An RDKit-grounded agent with ChEMBL and PubChem access that writes up its own reasoning is worth stress-testing on a project where you already know the answer. That's the only honest way to find out whether the hour holds up.

The thing I'd watch: Google's own drug discovery archive here runs 16 stories deep over the past 90 days across this beat, and none of them have yet named a pharma partner putting this into a live pipeline. That's the data point that would actually change the story.