Google DeepMind gives its Antigravity agent cheminformatics skills, turning days of drug-design work into about an hour

The design choice that matters most is the division of labour. The language model orchestrates the workflow, navigates databases and reasons about the science, while deterministic chemistry tools handle the molecular calculations. The skills are grounded in RDKit, the open-source cheminformatics toolkit built over two decades by a global community, with connectors to curated data from ChEMBL and PubChem[1].

DeepMind explains why that matters: the same molecule can show up as an IUPAC name in a paper, a Markush structure in a patent or a structure-data file in a vendor catalog, and changing a single atom or bond describes a different compound altogether[1].

The walkthrough uses a textbook medicinal chemistry move, rigidification. Crizotinib, a first-generation kinase inhibitor for certain non-small cell lung cancers, has five major rotatable bonds. Lorlatinib, a third-generation successor, is cyclized so that its effective number of rotatable bonds drops to zero, which boosts potency[1].

Asked to search the literature for these drugs and their analogues, the agent retrieved thousands of compounds within minutes, filtered out irrelevant molecules such as buffers and solvents, and documented its sources in a report. It flagged that the macrocycles were largely predicted to be brain-penetrant, which matters because this cancer often spreads to the brain. An interactive viewer let scientists plot potency against rotatable bonds and inspect individual molecules[1].

The agent then ran matched molecular pair analysis, which it had also suggested on its own, confirming that cyclization generally reduces rotatable bonds and increases potency. It trained and benchmarked a range of potency models, using chronological or scaffold splits to better test how well they generalize. Finally, it designed new macrocycles, proposed reactions and synthesis routes with electron-pushing diagrams, and scored each design with the models it had just trained. Designs can be passed on to AlphaFold or molecular dynamics, and results exported to standard chemistry software[1].

The skills are available now through the Science plugin in the Antigravity app, and Enterprise users can sign in with Google Cloud credentials. DeepMind is offering 10,000 scientists the highest Antigravity usage tier free for six months, subject to eligibility, capacity and selection[1]. The skills live in DeepMind's open science-skills repository, which also covers genomics, structural biology and literature search, draws on AlphaGenome, the AlphaFold database, UniProt and more than 30 other databases and tools, and can be installed outside Antigravity with npx[2].

Science is becoming a key battleground for agent platforms, and DeepMind's bet is that grounding an agent in trusted, deterministic tools beats asking a model to do chemistry in its head. The caveats are DeepMind's own: sparse data, assay noise between labs and narrow dynamic ranges limit even the best models, and the designs are hypotheses, not drugs. Nothing in the post was synthesized or tested in a lab, and the one-hour figure is DeepMind's estimate rather than an independent benchmark[1].