Google’s chemistry agent proposes new lung-cancer drug candidates in an hour

A scientist looking for a better cancer drug might begin with a paper, find a chemical structure in a patent, reconcile two naming conventions and then discover that the assay numbers cannot be compared directly. On November 3, Google DeepMind unveiled a set of Cheminformatics Skills for its Antigravity environment intended to pull much of that administrative and computational work into a single agent-assisted process. In a demonstration involving the lung-cancer drugs crizotinib and lorlatinib, an agent searched thousands of related compounds, curated a smaller dataset, analysed the effect of structural changes and generated candidate molecules and proposed synthesis routes in roughly an hour.[1]

The outcome is a computational design exercise. Google did not report synthesising the proposed new molecules, measuring their actual biological activity or establishing that any candidate is a safer or more effective medicine. The useful output is a set of hypotheses and supporting data for chemists to examine.[1]

How chemistry tools support the agent’s workflow

Google describes a division of labour: the language model coordinates searches and analysis, while specialised software executes chemistry operations. The skills connect to RDKit, a widely used open-source cheminformatics toolkit, and resources including ChEMBL and PubChem. They are designed to reconcile chemical structures across different representations and attach data provenance to the molecules found. That matters because a misread atom or bond changes the identity of the molecule, while an incorrectly attached assay number can make a potency comparison meaningless.[1]

The example used a familiar design idea known as rigidification. Crizotinib contains several bonds that allow its structure to twist. Lorlatinib is a more constrained macrocyclic inhibitor. In some drug series, reducing flexibility helps a candidate bind its intended target with a smaller entropic penalty. But rigidification alone is not a promise of greater effectiveness: solubility, metabolism, toxicity, target selectivity and access to the brain can all change. The company presented the crizotinib–lorlatinib series as an established case worth exploring, not proof that every rigid macrocycle improves a treatment.[1]

After retrieving thousands of candidate records, the agent removed unrelated items and presented an interactive chemical-space view. Google's illustrated curated dataset contained 568 compounds. The next step was matched molecular pair analysis: comparing close relatives that differ by a particular chemical change rather than drawing sweeping conclusions from dissimilar drugs. It also trained predictive potency models with held-out testing and described using scaffold-based splits to probe whether predictions survive a change of molecular family. Those are meaningful methodological safeguards, although the announcement does not disclose a complete benchmark with uncertainty intervals and an external replication.[1]

The last step is still experimental

The agent went on to suggest new macrocycles, mechanisms and plausible synthesis routes. It could export candidate structures for more specialised modelling, including molecular dynamics or structure prediction. An output that includes a chemically intelligible reaction diagram is useful to a medicinal chemist, but does not verify that a reaction will work in a laboratory at acceptable yield, or that the product will reach a tumour in a patient.[1]

Google itself identifies sparse data, differences between laboratories and narrow assay ranges as constraints on potency prediction. A compound may look strong on one biochemical measurement while failing in cells, animals or clinical trials. The most defensible reading of the one-hour demonstration is that an agent made a connected hypothesis-generation workflow faster in a vendor-selected example. The site does not provide the human-hours baseline, machine cost, comparative blind study or failed-design rate needed to calculate a universal speedup.[1]

The skills are available in Antigravity through its Science plugin and through an enterprise route using Google Cloud credentials. Google is also offering six months of its highest usage tier to 10,000 scientists, with rolling onboarding subject to eligibility, capacity and selection. For a laboratory evaluating the workflow, the useful test is whether the exported evidence and proposed experiments help chemists choose which candidates justify synthesis and further study.[1]

Report an error or suggest a correction