Google's science agent proposes new drug candidates in an hour. Every one still has to be made in a lab

Facts are attributed and linked to their sources. The analysis and opinions are the author's own.

What it is

On 3 November Google DeepMind published Cheminformatics Skills for Google Antigravity, its agent workbench. The design splits the work: the language model handles orchestration, database navigation and scientific reasoning, while deterministic chemistry tools, built on the open-source RDKit toolkit with connectors to ChEMBL and PubChem, do the molecular calculations.

The case study is a well-known example of rigidification, using two lung-cancer drugs, crizotinib and lorlatinib. DeepMind says the agent retrieved thousands of compounds from the literature within minutes, filtered out irrelevant entries such as buffers and solvents, documented its sources, ran matched molecular pair analysis, built and validated potency models on a held-out set, and proposed new ring-shaped molecules with synthesis routes, in about an hour against a workflow that typically takes several days. The skills are available now through a Science plugin, and Google is offering 10,000 scientists six months of its highest Antigravity usage tier free, subject to eligibility and selection.

Why it matters

The most important sentence in DeepMind's post is its caveat: "computational designs are testable hypotheses, not finished drugs", and "every proposed design must be synthesized and validated experimentally". It also names the limits: sparse data, assay noise between labs and narrow dynamic ranges.

The useful pattern is not an agent that knows chemistry. It is an agent that hands the arithmetic to tools that cannot improvise.

That division of labour travels well beyond chemistry. In finance, the equivalent is an agent that gathers filings, cleans the data and proposes a model, while the calculations run in tested, deterministic code and a person signs off before anything reaches a decision. DeepMind's choice of validation method is a detail worth copying too: the agent often uses chronological splits when dates are available, training on older compounds to predict newer ones, which is how credit and fraud models should be validated as well.

What I'd watch

• Whether any agent-proposed molecules are synthesised and their measured potency published.
• How the free usage programme is allocated, and whether academic groups in emerging markets get places.
• Whether Google publishes error rates for the agent's literature extraction, where one wrong atom changes the molecule.