Google Antigravity's chemistry skills put traceable tools at the center

Google DeepMind's November 3 account of Cheminformatics Skills describes an Antigravity workflow connecting an agent to RDKit and resources including ChEMBL and PubChem. The model coordinates searches and analysis while specialist tools perform molecular calculations. DeepMind demonstrates literature curation, compound comparison and proposed designs, presenting the process as support for scientists rather than a finished drug-discovery result.

A division of work between model and tools

The important implementation detail is the division of responsibility. A conversational system can help organize a sequence of tasks, but the chemistry work is grounded in named tools and data sources. The demonstration includes reports and an interactive viewer linking compounds to associated information. That gives a reader something more specific to investigate than an undifferentiated claim that an AI system understands science.

NEXTECH analysis: provenance is the useful interface

The valuable interface here is the chain of evidence. A scientist needs to distinguish a retrieved observation, a calculated property and a new hypothesis, then inspect where each came from. We would judge an agent-assisted workflow by how easily a colleague can challenge or reproduce a step. Fluent explanations are useful, but they should make that inspection easier instead of standing in for it.

A proposed molecule remains a hypothesis

DeepMind explicitly says proposed designs require synthesis and experimental validation. We have not reproduced its workflow, evaluated the predictions or assessed any medical application. Our view is that the story concerns scientific work organization: reducing effort around a hypothesis while preserving the ability to examine it. It should not be read as evidence that an agent has produced an approved treatment, established a compound's safety or replaced experimental expertise.