Google DeepMind Adds Cheminformatics Skills to Antigravity for Drug Discovery Work

In a blog post published on November 3, DeepMind said chemists still spend too much time cleaning and reconciling data. Part of the problem is that the same molecule can be recorded in many forms. It might appear as an IUPAC name in a journal paper, a Markush structure in a patent or a structure-data file in a supplier catalogue. Getting any of these wrong has real costs: the post notes that changing a single atom or bond describes a different compound.

DeepMind says the skills split the work. The language model handles workflow orchestration, database navigation and scientific reasoning, while deterministic chemistry software performs the molecular calculations. The skills are built on RDKit, a widely used open-source cheminformatics toolkit, and include connectors to the public ChEMBL and PubChem databases. DeepMind also lists tools for manipulating and validating reactions, drawing curved-arrow electron-pushing diagrams and finding similar reactions in public databases.

Availability and Access

The skills are available now through the Antigravity app and its Science plugin, according to the post. Enterprise users can sign in with Google Cloud credentials. DeepMind is offering its highest usage tier free for six months to 10,000 scientists, who will be onboarded on a rolling basis subject to eligibility, capacity and selection. The post does not say whether the skills themselves are open source.

A Lung Cancer Drug as the Test Case

To show how the skills work, DeepMind used rigidification, a standard medicinal chemistry technique. Molecules with many rotatable bonds can twist into many shapes, and locking a molecule into its active shape can increase its potency. The example compares two tyrosine kinase inhibitors used against certain non-small cell lung cancers. Crizotinib, a first-generation drug, has five major rotatable bonds. Lorlatinib, a third-generation drug, is cyclised so that its effective count of rotatable bonds falls to zero.

In the walkthrough, the agent pulled thousands of compounds from the literature, removed irrelevant entries such as buffers and solvents, and recorded where each came from, leaving 568 curated compounds in an interactive viewer. It then ran a matched molecular pair analysis, which found that cyclisation generally increased potency. It built potency models and validated them on held-out data, and it proposed new macrocycle designs with synthesis routes and predicted potencies. DeepMind says the full workflow took about an hour, compared with the multiple days a single iteration typically takes.

The company stresses that the agent's outputs are hypotheses to be tested. It says sparse data, assay noise and narrow dynamic ranges limit how precise the models can be, and that every proposed design needs experimental validation. Possible next steps named in the post include linking the workflow to AlphaFold and to molecular dynamics simulations.

Why It Matters

The release moves Antigravity, which Google launched as a coding environment, into domain-specific scientific work. DeepMind is pairing a general-purpose model with established, deterministic scientific software rather than asking the model to do the chemistry itself. For pharmaceutical and chemicals companies, the claimed shortening of design cycles from days to hours is the main selling point. The free-access offer is also a way for Google to win over working scientists while AI labs compete for research and enterprise users.