DeepMind puts a chemistry agent inside its dev platform

Google DeepMind has added a set of Cheminformatics Skills to Antigravity, its agentic development platform, and shows them working on a real drug-discovery problem. The idea is a clear division of labor: the language model plans the work and reasons about chemistry, while deterministic tools built on RDKit, with connectors to the ChEMBL and PubChem databases, do the actual molecular calculations. The team says a workflow that used to take several days now runs in about an hour.

The case study looks at kinase inhibitors for lung cancer and a medicinal-chemistry trick called rigidification. A flexible drug like crizotinib has five rotatable bonds and can fold many ways; a later drug, lorlatinib, is cyclized so it has effectively none, which raises its potency by "paying the entropic cost in advance." From a prompt, the agent pulled thousands of compounds from the literature and patents, filtered out things like buffers and solvents, and settled on a curated set of 568. It then flagged that most of these molecules were predicted to reach the brain, which matters because this cancer often spreads there, built models to predict potency, and proposed new ring-shaped molecules complete with synthesis routes and reaction diagrams.

DeepMind is careful to say the output is a set of testable hypotheses, not finished drugs, and that accuracy is limited by sparse and noisy experimental data. The skills are available now through the Antigravity app's Science plugin, and the company is offering 10,000 scientists its top usage tier free for six months.

Why it matters

If you work in computational chemistry, this is worth a direct test: the agent does the tedious data gathering and model building, but it still hands you hypotheses you have to validate in the lab. The free six-month tier makes trying it cheap, so the question is whether its designs hold up against your own assays.