Google DeepMind’s new cheminformatics walkthrough shows how an AI agent can connect scientific records, established chemistry software and predictive analysis. Its importance lies in the structure of that collaboration—and in the distance between a computationally attractive proposal and an experimentally validated discovery
A research question in chemistry rarely arrives as a clean dataset. Relevant evidence may be scattered across papers, tables, patents and databases, with different representations of the same compound and measurements collected under different conditions. Before a scientist can assess a promising idea, somebody must organise that material and decide which comparisons are meaningful. An AI system that helps with this preparatory work could be useful even without inventing a new chemical theory. Its value would come from making the route between a question and an inspectable analysis shorter and easier to repeat.
On 3 November 2026, Google DeepMind published a walkthrough of new Cheminformatics Skills for Antigravity. It describes an agent gathering compound information, conducting analysis and proposing candidate structures and synthetic routes in a workflow the company says took about an hour. This is a developer-presented demonstration, not an independently reproduced productivity trial or evidence that a new medicine has been validated. The interesting technical claim is a division of labour: the language model coordinates the work, while established chemistry tools carry out calculations.
That distinction offers a useful way to judge agentic science. A fluent explanation can suggest an experiment, but a scientific workflow must also expose its inputs, operations and outputs. The more work an agent can perform, the more important it becomes to preserve those connections. Otherwise speed can simply move uncertainty from a visible manual process into an automated sequence that is harder to inspect. The test is whether a researcher gains stronger control over the analysis as well as a quicker answer.
Domain tools constrain what language alone cannot
The RDKit project supplies an established open-source toolkit for cheminformatics, including molecular representations, analysis and related computational operations. Connecting such software to a language model changes what a request can produce. Instead of merely describing how a calculation might be performed, an agent can arrange for software to perform it on specified structures. This does not make every step correct. It creates an opportunity to inspect an operation and its result separately from the model’s verbal account.
The distinction matters because a molecular description is precise in ways ordinary language often is not. A small change in connectivity or stereochemistry can refer to a different compound. A pipeline may successfully execute on a structure while that structure is not the one the source intended. Deterministic computation therefore constrains one class of error without resolving the entire chain of interpretation. A useful system needs checks at the boundary between extraction and calculation, and a record of any normalisation or exclusion applied before analysis begins.
Google’s public science-skills repository makes the broader idea of domain-specific agent guidance concrete. Such resources can be inspected as part of understanding how an agent is instructed to use scientific tools and databases. Their availability should be distinguished from independent validation of an end-to-end workflow. A publicly accessible instruction or connector can support reproducibility, but reproducing a scientific result also requires the relevant data, software environment, settings and interpretation. These are different layers of openness, each useful in its own way.
A database entry is not a universal measurement
One of the resources connected in the demonstration is ChEMBL, which describes itself as a manually curated database bringing together chemical, bioactivity and genomic information. This is important infrastructure rather than a replacement for critical reading. A compound can have several reported measurements, reflecting different targets, assay designs and experimental conditions. A clean-looking table assembled from reliable records may still combine values that should not be treated as interchangeable. Retrieval quality and comparison quality are separate questions.
An illustrative workflow might gather activity values for a family of compounds and rank them from strongest to weakest. Before interpreting that ranking, a chemist would want to know whether the measurements concern the same biological system, use compatible units and represent comparable endpoints. Missing values also require care: absence of a reported result is not evidence that a property is absent. An agent can help organise those questions, but it should not silently settle them through a convenient default that disappears from the final report.
This gives provenance a practical role. A useful result should let a researcher return from a plotted point to the underlying record and, where possible, its experimental source. Decisions to remove duplicates, exclude assays or convert measurements should remain visible. That record is not administrative decoration. It is what allows another person to discover whether an attractive relationship reflects chemistry, a measurement convention or a data-processing mistake. Automation is most valuable when it makes those distinctions easier to examine.
Comparing small changes can sharpen a hypothesis
The walkthrough uses matched molecular pair analysis as one route from collected records to a more focused question. The established mmpdb project illustrates this family of methods: it identifies related structures sharing a common part and differing in a defined transformation, then associates transformations with observed property changes. The appeal is interpretability. Instead of comparing wholly unrelated compounds, a researcher can ask what tends to change when a particular structural modification is made within the available examples.
Such a comparison can help generate a hypothesis, but it does not turn observational data into an unrestricted chemical rule. The surrounding structure and experimental context still matter. A transformation associated with a useful change in one series may not behave similarly elsewhere. Where several properties matter, an apparent gain can also arrive with a disadvantage that the chosen dataset does not measure. The output should therefore be read as evidence about a set of comparisons, with uncertainty about how far it extends.
This is where an agent could improve the scientist’s working environment. It can organise candidate comparisons, prepare visualisations and help trace why a particular proposal was considered. The scientific benefit would be strongest if the researcher can question the selection and revise it without losing the earlier record. A system that supplies a polished narrative but hides the pair definitions or excluded examples makes scrutiny harder. A collaborative tool should make alternative explanations accessible, not merely accelerate the production of the preferred explanation.
A prediction is only as informative as its test
The next stage of an agentic workflow may train a model to estimate properties of untested structures. Here a numerical score can appear authoritative even when the evaluation answers an easier question than the scientist intended. If closely related examples appear on both sides of a training–test division, success may mainly show interpolation within a familiar series. A harder evaluation might separate structural families or respect the order in which information became available. The right choice depends on the proposed use, and no single split provides a complete account of generalisation.
For an application intended to prioritise genuinely unfamiliar structures, the evaluation should resemble that challenge. It should also distinguish the model’s uncertainty from the ordinary variability of measurements. Trying many feature sets and models creates another responsibility: the test data should not quietly become part of repeated optimisation. These are general evaluation considerations rather than a claim that the published walkthrough has failed them. They identify what a team would need to inspect before trusting an automated modelling sequence for consequential research decisions.
Google’s demonstration describes generating candidate macrocycles and proposed synthesis plans, while explicitly stating that designs require synthesis and experimental validation. That boundary is essential. A proposal can be coherent, computationally attractive and still fail in practice. Making a compound, measuring the intended property and investigating other relevant properties are separate achievements. Faster hypothesis generation can increase the value of experimental capacity, but it cannot substitute for it.
The next benchmark is a useful scientific handover
Antigravity’s science use-case page describes installation through a Science plugin and points to the public customisations repository. That establishes a route for eligible users to explore the tooling, rather than a universal guarantee of access to every model, resource or service. A research team would still need to consider its data permissions, execution environment and institutional requirements. The practical unit of adoption is the complete workflow, including the people responsible for reviewing its products.
A strong assessment of this approach would compare more than elapsed time in a selected demonstration. It could examine the accuracy of extracted structures, the proportion of usable comparisons, reproducibility across runs and the effort required for a chemist to correct mistakes. It could ask whether another researcher can reproduce the analysis from the saved artefacts and whether proposed experiments prove informative. Those measures would help distinguish genuine assistance from a transfer of work into checking, where the apparent time saving disappears during review.
The November walkthrough is therefore best understood as a concrete example of connecting language-based coordination to established scientific infrastructure. It is a direction worth testing because many research bottlenecks lie between tools and records rather than inside a single calculation. Its promise will be realised when the handover to the scientist becomes more reliable: sources recoverable, assumptions explicit, computations repeatable and proposals clearly labelled as proposals. An agent that achieves that can make research easier to pursue without blurring the line between a persuasive computational story and an experimentally supported result.
Image: Chemistry lab 4 by Santhosh Notagar99, CC BY 4.0, via Wikimedia Commons. Illustrative laboratory photograph; it does not depict Google’s demonstration or experimental validation of its proposals.