Designing a new drug candidate usually means days of fiddly work before any chemistry begins: pulling molecules out of papers and patents, cleaning the data, training a model and then proposing something to make. Google DeepMind's new Cheminformatics Skills for Antigravity aim at that drudgery, letting an agent do the data wrangling so a scientist can spend more time on the hypotheses that matter.
The design choice at the centre is a division of labour. The language model plans the workflow, navigates databases and reasons about the chemistry, while the exact molecular calculations are handed to deterministic tools built on RDKit, the open-source toolkit the field has used for two decades. Connectors reach curated data in ChEMBL and PubChem. That split matters because chemistry has no margin for error: change one atom or bond and you have described a different compound, so letting a probabilistic model do arithmetic it can fumble would be the wrong way round. The skills install through Antigravity's Science plugin, and we review Antigravity itself.
DeepMind's walkthrough uses a real lung cancer case. The agent searches the literature, pulls out thousands of compounds, filters out buffers and solvents, then runs matched molecular pair analysis to isolate the effect of small changes and trains potency models validated on held-out compounds. It ends by proposing new rigidified macrocycles, molecules locked into a single shape to bind their target more tightly, complete with candidate synthesis routes and electron-pushing diagrams. DeepMind says the whole cycle took about an hour rather than the usual multiple days, and that 10,000 scientists can get the top tier free for six months.
The honesty of the write-up is its strongest feature. DeepMind states plainly that sparse data, noise between labs and narrow measurement ranges limit even the best predictive models, and that every computational design is a testable hypothesis, not a finished drug. Each proposed molecule still has to be synthesised and checked in a lab, which is where the real time and cost sit.
Two caveats temper the headline. The one-hour figure comes from a single showcase case chosen to work well, so a medicinal chemist should read it as a best case, not a typical one. And the deeper question of whether agent-proposed molecules actually bind better in the lab is not something a blog walkthrough can answer. The split between a reasoning model and deterministic tools is a sound pattern that other scientific-agent work could copy; the drug-design payoff is still a promise waiting on wet-lab results.