News
Google DeepMind's Antigravity Agent Designs Lung Cancer Macrocycles in an Hour
Google DeepMind has unveiled Cheminformatics Skills for Antigravity. In a demo built around non-small cell lung cancer drugs, the agent proposed new macrocycles in about an hour, though the designs still need synthesis and experimental testing.
On November 3, 2026, Google DeepMind published a description of Cheminformatics Skills, a set of capabilities for the Google Antigravity agent that combines literature search, compound analysis, predictive models and synthesis route suggestions. In a demo based on non-small cell lung cancer drugs, the agent went from raw data to new molecule designs in about an hour.
What was shown
The authors of the DeepMind post describe Antigravity with the new skills as a "collaborative science workbench". The language model handles orchestration, database navigation and reasoning, while deterministic chemistry tools carry out the calculations on molecules.
According to DeepMind, advanced cheminformatics systems have to unify different ways of writing down molecules, including IUPAC names, Markush structures and SDF files. This matters because changing a single atom or bond produces a different compound.
The crizotinib example
The starting point was crizotinib and lorlatinib. Crizotinib is a first-generation tyrosine kinase inhibitor used in certain forms of non-small cell lung cancer, and it has five major rotatable bonds. Lorlatinib is a third-generation inhibitor that has been cyclized, cutting the effective number of rotatable bonds to zero, which according to the description leads to higher potency, among other effects.
The agent searched the literature, retrieved thousands of compounds, filtered out substances such as buffers and solvents, and prepared a report citing its sources. The report noted that the macrocycles are largely predicted to be brain-penetrant, which is important because this type of cancer frequently metastasizes to the brain.
An interactive viewer showed that the most potent compounds are also the most rigid. Matched molecular pair analysis, which compares molecules that differ by only a small fragment, showed that cyclization reduces the number of rotatable bonds and generally increases potency.
From models to synthesis
The agent built many models that predict potency. According to the description, it typically uses chronological splits when dates are available, and in the benchmark shown it used a split by molecular scaffolds. The models were tested across four experimental read-outs.
It then proposed new macrocycles together with precursors, cyclization reactions, mechanism diagrams, synthesis routes and predicted potency values. For the proposed cyclization reaction, it produced electron-pushing diagrams illustrating the mechanism.
Caveats and access
DeepMind notes that sparse data, inter-lab assay noise and narrow dynamic ranges limit the precision of the predictions. The designs are hypotheses to be tested, not finished drugs, and each one must be synthesized and validated experimentally. The hour therefore covers computational design only.
The skills are available now once you download the Antigravity app and install the Science plugin. Enterprise users can log in with a Google Cloud account. The program for 10,000 scientists is onboarding participants on a rolling basis, and applications are already open.
