Designing new medicines usually requires curating data from figures, charts and literature, training predictive models, and proposing synthesizable candidate molecules — a workflow that often takes days per iteration. Google Antigravity’s new Cheminformatics Skills aim to speed up this process and let scientists spend more time on hypotheses than on data wrangling.
What the new skills provide
The Cheminformatics Skills combine a language model’s workflow orchestration with deterministic chemical tools. The foundation is RDKit, augmented with connectors to curated resources such as ChEMBL and PubChem. Using these components, an agent can extract molecular structures from documents, associate them with reported data, and prepare the compounds for downstream analysis and visualization.
The system is designed to unify the many ways chemistry is represented — IUPAC names, Markush structures, SDF files, images — reducing the human overhead and error-proneness of data handling. That is critical because changing a single atom or bond converts a description into a different compound, which can invalidate follow-up work.
Demonstration: macrocycles and lung cancer therapies
In a technical walkthrough, the agent analyzed real-world lung cancer treatments, including crizotinib and lorlatinib. The session included these steps:
- Searching thousands of compounds and extracting associated data.
- Launching an interactive viewer to inspect curated compounds, their metadata, and customizable scatter plots (for example potency versus number of rotatable bonds, colored by macrocycle status).
- Running matched molecular pair analysis (MMPA) to isolate the effects of specific structural changes by comparing molecule pairs differing only by a small modification.
- Building and validating many potency models on a held-out dataset, using varying feature sets and model architectures with guidance on best practices (for example chronological splits where possible, or scaffold splits to estimate generalizability).
- Designing novel, rigidified macrocycles and proposing multiple reactions and synthesis routes for them.
The chemical rationale highlighted in the demonstration is familiar in medicinal chemistry: many drug molecules contain multiple rotatable bonds that allow many conformations in solution. Binding often requires a single active conformation; the entropic cost of adopting that conformation reduces potency. Rigidification — for example via cyclization — reduces conformational freedom and can increase binding affinity. Crizotinib has five major rotatable bonds; lorlatinib is a third-generation tyrosine kinase inhibitor (TKI) with a cyclized portion that effectively reduces the number of rotatable bonds to zero, contributing to its higher potency.
Results produced by the agent
The MMPA results showed that cyclization tends to reduce the number of rotatable bonds and generally increase potency, consistent with expectations. The agent built many predictive models and validated them on held-out data; the Cheminformatics Skills provide domain-appropriate guidance for constructing molecular machine learning workflows.
For designed macrocycles, the agent suggested several plausible synthetic routes and prepared output suitable for further analysis with tools such as AlphaFold or physics-based methods like molecular dynamics. Results can be exported in various formats for import into industry-standard chemistry software.
Limits and caveats
The tools speed up generation of stronger design hypotheses but do not replace experimental validation. Sparse data, inter-lab assay noise, and narrow dynamic ranges limit the predictive precision of models; computational designs remain testable hypotheses, not finished drugs.
Availability and support
The Cheminformatics Skills are available now on Google Antigravity. Users can download the Antigravity App and install the Science plugin to access the capabilities. Enterprise users can select “Use Business Account” and sign in with Google Cloud credentials. Google is offering 10,000 scientists the highest usage tier of Antigravity free for six months (onboarded on a rolling basis), subject to eligibility, capacity and selection; applications are open.
Acknowledgements
Contributors include Filip Radlinski, Artem Sobolev, Victoria Johnston, Hannah Kiely-Collins, Zachary Wu, Jue Wang, Jack Dunger, Adriana Fernandez Lara, Ngân Vũ, Doug Fritz, Yunhan Xu, Sebastian Nowozin, and Pushmeet Kohli.
The Cheminformatics Skills are intended to reduce the burden of data handling and accelerate researchers’ ability to make meaningful scientific decisions, while preserving the need for experimental synthesis and validation.
