The Pulse
Google Antigravity Compresses Cheminformatics Work into an Hour
Google DeepMind says Google Antigravity completed a lung-cancer cheminformatics workflow in an hour, compared with the multiple days a typical iteration takes. The demonstration covers compound searches, potency analysis and proposed macroc
AI.info Team · Written with AI · fact-checked ·
A drug-discovery workflow that Google DeepMind says typically takes multiple days for one iteration took an hour in a demonstration using Google Antigravity. Published November 3, 2026, the walkthrough applies new Cheminformatics Skills to lung-cancer treatments, moving from literature searches to potency analysis and proposed macrocycles. The company presents the system as a way to reduce data-wrangling work, not to replace laboratory testing.
From literature search to 568 compounds
In the demonstration, the agent searches literature for crizotinib, lorlatinib and related compounds, extracts molecules and associates them with reported data. It retrieves thousands of compounds within minutes, filters out items such as buffers and solvents, and records the sources in a report. The resulting curated dataset contains 568 compounds, which researchers can explore in an interactive viewer with compound metadata and customizable plots.
The example focuses on rigidification: adding bonds to restrict a molecule’s movement. Google describes crizotinib as having five major rotatable bonds, while lorlatinib has been cyclized to reduce its effective number to zero. The viewer lets researchers compare potency with rotatable-bond counts and distinguish macrocycles from other compounds. In the reported dataset, the highest-potency compounds are the most rigid.
Language-model orchestration, chemistry-tool calculations
Google’s design divides tasks between the language model and chemistry software. The model handles workflow orchestration, database navigation and scientific reasoning; deterministic tools perform molecular calculations. The skills are grounded in RDKit and connect to ChEMBL and PubChem, resources the post identifies as curated chemistry databases.
The agent then performs matched molecular pair analysis, comparing compounds that differ by whether they are cyclized. Google says the analysis found that cyclization reduced rotatable bonds while generally increasing potency. The system also builds potency models using different feature sets and architectures, validates them on held-out data, and benchmarks them across four experimental readouts. Its guidance favors chronological splits when dates are available, with scaffold splits as another way to estimate generalizability.
Proposed molecules remain hypotheses
After analysis, the agent proposes new macrocycles, possible reactions and synthesis routes. It can provide electron-pushing diagrams, check reactions against public databases and predict potency using models built earlier in the workflow. Google says the designs can be prepared for further analysis with AlphaFold or molecular-dynamics simulations, and that results can be exported to chemistry software.
“Computational designs are testable hypotheses, not finished drugs,” write Zoi Tsangalidou, Ruijie (Kelly) Zhang, Sukhdeep Singh and Michael V. LeVine. The post identifies sparse data, differences between laboratory assays and narrow dynamic ranges as limits on predictive precision. It says every proposed design still needs to be synthesized and validated experimentally.
Access through Antigravity’s Science plugin
Google says the Cheminformatics Skills are available through the Science plugin in the Antigravity app, including for Enterprise users who sign in with Google Cloud credentials. The company is also offering 10,000 scientists its highest Antigravity usage tier free for six months, with rolling onboarding subject to eligibility, capacity and selection. Applications are open, according to the post.