Chemistry agent proposes ring-shaped molecules for drug design, Google DeepMind says
In a demonstration with two lung cancer drugs, Google DeepMind's Antigravity tool, given chemistry skills, searched thousands of compounds and suggested ways to make new molecules. The company says every design must still be made and tested.
An AI agent, which carries out a series of steps on its own, searched thousands of compounds and proposed new ring-shaped molecules, with possible ways to make them, in about an hour, Google DeepMind says.
In a post dated 3 November 2026, the company says a drug design workflow of this kind typically takes multiple days for a single round. It has added Cheminformatics Skills, tools for studying chemical compounds with computers, to its Google Antigravity tool.
How it works: the language model, the part that reads and writes text, plans the steps, searches databases and does the reasoning. Separate fixed chemistry tools, built on an open-source chemistry toolkit and linked to public compound databases, do the molecular calculations.
The company's example uses two real lung cancer drugs, crizotinib and lorlatinib. Lorlatinib resembles crizotinib but is closed into a ring at one end, which the post says leads to higher potency, among other effects. Potency is how strongly a molecule acts on its target.
The idea is called rigidification. The post explains that a floppy molecule can twist into many shapes, but must hold one shape to bind its target. Adding a bond in the right place holds the shape, which the post says can boost potency a great deal.
In the demonstration, the agent retrieved thousands of compounds within minutes, filtered out irrelevant molecules such as buffers and solvents, and recorded where each came from in a report.
It then compared pairs of molecules that differ by one small change. The post says the agent suggested this analysis independently. The analysis showed that closing molecules into rings generally increased potency, as the authors expected.
The agent also built many models to predict potency and checked them on compounds held back from training. Asked for new designs, it recommended reactions and synthesis routes, gave a written rationale for each design and predicted each design's potency.
Its report noted that the ring-shaped molecules were largely predicted to reach the brain, which the company says matters because this lung cancer often spreads there.
These results are Google DeepMind's own demonstration. The company states that every proposed design must still be made and tested in experiments.
It also says sparse data, noise between labs' tests and narrow ranges of measurements limit how precise predictions can be. It adds that computational designs are "testable hypotheses, not finished drugs".
The company says it is offering 10,000 scientists free access for 6 months, subject to eligibility, capacity and selection. The skills are available on Google Antigravity: users download the Antigravity App and install the Science plugin.
Sources
- Agentic cheminformatics with google antigravity · deepmind.google