Google DeepMind announced its Cheminformatics Skills for Google Antigravity on November 3, 2026, and said they were available through the Science plugin. The workflow pairs a language model’s coordination and scientific reasoning with deterministic chemistry tools that perform molecular calculations. Google DeepMind describes the skills and access route.
The company also offered up to 10,000 scientists the highest Antigravity usage tier free for six months. The offer has eligibility, capacity, and selection conditions, with onboarding on a rolling basis.
How the model and chemistry tools divide the work
Google DeepMind describes the model as handling workflow orchestration, database navigation, and scientific reasoning. Deterministic tools handle molecular calculations. The skills are built around RDKit and connect to the curated chemistry resources ChEMBL and PubChem.
That division matters: the model coordinates the steps, while specialized tools carry out the calculations. The announced workflow includes literature searches, compound extraction and curation, matched molecular pair analysis (MMPA), potency modeling, and proposed molecular designs with suggested synthesis routes.
From literature search to proposed molecules
In its demonstration, Google DeepMind searched for compounds related to crizotinib and lorlatinib, filtered out irrelevant results such as buffers and solvents, and recorded sources. The company says the workflow retrieved thousands of compounds and displayed 568 curated compounds in a dashboard.
The analysis then used MMPA to compare related molecules, trained potency models, and proposed macrocycles and possible synthesis routes. In the company’s example, crizotinib is described as having five major rotatable bonds; lorlatinib, after cyclization, has zero effective rotatable bonds. Google DeepMind says cyclization generally increased potency in the molecular pairs it analyzed, a finding limited to those pairs and that analysis.
How to interpret the model evaluation
Google DeepMind says its models were evaluated on held-out data across four experimental readouts. It describes using chronological splits when dates allow and scaffold splits as another way to estimate generalizability. A scaffold split separates compounds by their core chemical structures.
The choice of split should match the prediction task. If closely related compounds appear in both training and test sets, performance may mainly reflect interpolation among familiar structures. A researcher aiming to predict results for unfamiliar chemical families needs an evaluation designed to test that broader generalization. The IDHUS Institute has explained this as a general methodological distinction; it is not an independent assessment of Google’s models.
Proposed designs still need lab validation
Google DeepMind says sparse data, noise between laboratories’ assays, and narrow measurement ranges limit predictive precision. The demonstration’s proposed macrocycles and synthesis routes are computational designs: the company says they must still be synthesized and experimentally validated.
A proposed reaction scheme is one of the workflow’s outputs; establishing that a molecule can be made and testing its properties require experimental work.
