Google Antigravity Cut the Analysis of Thousands of Molecules From Several Days to One Hour

Drug discovery has an uncomfortable engineering reality: a substantial share of the time goes not into flashes of insight, but into collecting and cleaning data. Chemists have to extract molecular structures from papers and images, reconcile names, discard solvents, link compounds to experimental results, and then prepare everything for analysis. A single error in an atom or bond means the system is already working with a different molecule.

Google DeepMind demonstrated how the Antigravity agent, equipped with a set of Cheminformatics Skills, can complete this cycle in about an hour. In the demonstration, it collected data on thousands of compounds associated with treatments for non-small cell lung cancer, compared them, built activity models, and proposed new molecules along with possible synthesis routes. Previously, a similar iteration took several days and required the involvement of a computational chemistry specialist.

The agent manages the process; the tools do the calculations

The main idea is not that a language model was taught to “guess chemistry.” It was assigned the role of coordinator: finding sources, choosing the next stage of analysis, invoking the required software, and assembling the results into a report. Calculations and structural validation are performed by specialized tools—primarily the open-source RDKit library—while the data comes from curated resources such as ChEMBL and PubChem.

This division of labor is useful far beyond chemistry. A language model is good at connecting the stages of a long process and explaining them to a person, but poorly suited to operations where one incorrect character changes the result. Molecular calculations therefore remain with deterministic code that can be rerun and verified.

As an example, the team chose molecular “rigidification.” A flexible molecule can adopt many shapes, but it needs a particular configuration to bind to its target. An additional chemical bond restricts movement and can sometimes increase activity. This is how lorlatinib, a more rigid macrocyclic drug, differs from the more flexible crizotinib.

Antigravity found related compounds in the literature, cleaned the dataset, and displayed them in an interactive interface. A chart made it possible to compare cellular activity with the number of rotatable bonds: in this dataset, more rigid macrocycles more often showed high activity. This does not establish a universal rule, but it gives researchers a clear hypothesis for the next test.

From data discovery to new candidates

After preparing the dataset, the agent analyzed pairs of molecules that differed by a small modification. This approach helps estimate what specifically happened after a molecule was closed into a ring, without conflating the effect with a dozen other differences. The system then built several activity models using different features and architectures.

Importantly, the models were not evaluated only on randomly held-out examples. The team used a split based on chemical scaffolds so that the test set differed more substantially from the training set. When compound publication dates are available, the skills also recommend training the model on older data and testing it on newer data. Both approaches better reflect the real task: predicting the properties of an unfamiliar molecule, rather than recognizing a close relative of a previously seen compound.

At the final stage, the agent proposed several new macrocycles, explained the chemical rationale, selected possible reactions, and drew electron-transfer schemes. For each option, it used the previously trained models to estimate expected activity. The results can be exported in formats compatible with industry-standard chemistry software or passed into further calculations, including molecular dynamics and protein-structure analysis.

Where the acceleration ends

This remains a demonstration by Google itself, not an independent comparison with laboratory workflows. The material does not include a published series of physical experiments confirming that the proposed compounds can actually be synthesized and work better. Sparse data, differences between laboratory protocols, and a narrow range of measurements limit the accuracy of any model.

The practical lesson for developers of agentic systems is straightforward: value comes not from a model holding an autonomous conversation with itself, but from connecting it to verifiable tools, sources, and checkpoints for specialists. For a product team, this is a useful blueprint for a vertical agent: take an expensive, multistep process, leave the calculations to specialized software, and give the model responsibility for navigation and explanation. But the success metric should not be an attractive report; it should be a reduction in the time required to reach a testable experiment. The claimed result here is about one hour instead of several days; the final answer still comes from the laboratory.