Google demonstrated cheminformatics in Antigravity. Here is how to check molecule parsing, retain rejected records and distinguish calculations from hypotheses.
On November 3, Google DeepMind demonstrated a cheminformatics workflow in Antigravity: the agent organizes the work while specialist tools, including RDKit, perform molecular calculations. The walkthrough is by Zoi Tsangalidou, Ruijie (Kelly) Zhang, Sukhdeep Singh and Michael V. LeVine. It is a developer's research demonstration, best assessed through intermediate data and calculations that can be inspected.
Google's Science page directs users to install the plugin through the in-app Customizations marketplace. The science-skills repository describes instructions, helper scripts and dependencies managed with uv; some skills require API keys. Data-source licenses need separate review from the code license. Environment preparation and data access remain part of the work after plugin installation.
A concrete first check is whether every molecule was read successfully. The RDKit documentation shows that reading can return None instead of a molecule object, and recommends checking records before using them. If an agent silently skips those rows, its report describes a changed sample. Ask for input, accepted and rejected counts, alongside identifiers for rejected records.
For reviewing the analysis, keep a provenance table containing the dataset source and version, molecule identifier, parsing status, script used and saved calculation output. Failed rows should remain visible for investigation. This is NODVIEW's proposed review method; it does not establish chemical correctness for every conclusion or imply that we ran the plugin.
Label calculated values separately from newly proposed structures. Google's authors explicitly say computational designs require synthesis and experimental validation. A polished visualization, a successfully parsed structure or a model prediction does not replace that step. NODVIEW reviewed the original sources and RDKit documentation, without running chemistry experiments or evaluating drug efficacy.