Google DeepMind says Antigravity’s Cheminformatics Skills let researchers inspect retrieved compounds and their metadata, customize scatter plots, and export results to chemistry software. The workflow also proposes molecular designs, but those still need synthesis and experimental testing. Google DeepMind’s description of the skills explains the features and their limits.
What researchers can inspect and export
Google DeepMind describes an interactive viewer for browsing individual compounds and their associated metadata, creating customizable scatter plots, and exporting results in formats used by established chemistry software. That gives researchers ways to examine the records and visualizations behind an agent’s summary, then carry the results into other chemistry tools.
A useful chemistry workflow makes the compounds behind an agent’s summary inspectable and lets researchers carry the analysis into tools they already use.
How the skills coordinate cheminformatics work
The language model coordinates the workflow, navigates databases, and handles scientific reasoning. Deterministic chemistry tools perform the molecular calculations. Google DeepMind names RDKit as a foundation and describes connections to the chemical resources ChEMBL and PubChem.
The demonstration retrieves chemical information from scientific literature, filters and curates compounds, and supports potency modeling and proposed molecule designs with possible synthesis routes. It also uses matched molecular pair analysis (MMPA), which compares related molecules that differ by a defined structural change.
In one example, Google DeepMind describes crizotinib as having five major rotatable bonds and lorlatinib as having zero effective rotatable bonds after cyclization. Those are differently labeled bond counts. The company reports that cyclization generally increased potency among the molecular pairs it analyzed.
Google DeepMind announced the skills on November 3, 2026, and said they were available through the Antigravity Science plugin.
What the model evaluation and timing claims mean
Google DeepMind says it benchmarked multiple model architectures and feature sets across four experimental readouts. It describes using scaffold splits to gauge performance on compounds outside the training distribution, and says chronological splits may be used when compound dates are available.
The company’s blog describes the demonstrated workflow as taking about an hour and says a conventional iteration typically takes multiple days. In the demonstration video, presenter Michael says an agent takes less than half an hour for “this work,” compared with a week for his earlier process. The descriptions use different or unclear task boundaries, so they are not one standardized timing result.
Why proposed molecules still need experiments
The proposed macrocycles and possible synthesis routes are computational hypotheses, not validated medicines. Google DeepMind says the molecules must be synthesized and experimentally tested. It also identifies sparse data, differences between laboratories’ assays, and narrow measurement ranges as limits on predictive precision.
