For R&D teams evaluating an AI chemistry assistant, require its final tables to preserve whether a measurement is exact, above a limit or below a limit. A value converted to nanomolar units can still describe a bound. Sorting the numeric part alone can turn incomplete evidence into an unjustified compound ranking, affecting which candidates a team spends time investigating.
Google DeepMind’s 3 November 2026 cheminformatics announcement makes this a timely review question. It describes Antigravity coordinating chemistry tools and data sources to support discovery work. Its reported acceleration is a vendor demonstration, not independent evidence of a transferable business productivity gain. The announcement also says proposed designs require experimental validation.
This guide examines a narrower, inspectable boundary in a public ChEMBL agent skill. The original helper retains the measurement relation. The risk explored here arises when downstream analysis drops that relation and treats the added number as an exact result. The examples are fictional, and the local check is not an Antigravity trial or a chemistry validation.
Inspect the actual skill and the transformation it performs
Google’s public ChEMBL Database Query skill contains a real SKILL.md workflow, rather than merely a prompt described as a skill. It directs the agent to a utility script, requires output files and documents a --normalize option for bioactivity values. The links here pin the inspected repository revision.
The corresponding Python helper has a function named _normalize_activity. It reads standard_value and standard_units, applies a conversion factor and adds normalized_value_nM and a conversion note to the record. For missing values or unsupported units, it returns a null converted value with an explanation.
That function does not decide whether two experiments are scientifically comparable. It does not turn a “greater than” measurement into an exact one. Keeping these responsibilities separate helps reviewers identify where a proposed workflow needs additional interpretation rules.
Antigravity’s current skill documentation describes instruction bundles with optional scripts and resources. Reading a public bundle establishes what those files say; it does not establish that an installed assistant follows them correctly. Our guide to script-backed agent skills covers that broader distinction between a helper and its surrounding workflow.
Keep the number and its relation together
ChEMBL records include a measurement relation as well as a value and units. Its assay and activity documentation distinguishes assay types, target-assignment confidence and data-validity flags. Its criteria for calculating pChEMBL, a logarithmic activity field, include an equality relation, positive values and specified validity conditions. A converted concentration is therefore a narrower statement than eligibility for that field.
For the following fictional example, assume the measurements concern the same endpoint under comparable conditions. IC50 describes the concentration associated with half-maximal inhibition in an assay. Within this deliberately simplified comparison, a lower value indicates that less concentration was needed for that endpoint. Here, “exact” means recorded with an equality relation, not free from experimental uncertainty. No actual molecules, targets or experimental results are represented.
| Fictional record | Input measurement | Numeric conversion to nM | Meaning to retain |
|---|---|---|---|
| A | Greater than 100 nM | 100 | Above 100 nM |
| B | Equal to 200 nM | 200 | Measured at 200 nM |
| C | Less than 0.1 µM | 100 | Below 100 nM |
| D | Equal to 100 nM | 100 | Measured at 100 nM |
| E | Value 100; units missing | Null | Conversion unavailable |
Records A, C and D have the same numeric conversion, but they do not report the same result. Displaying all three as “100 nM” removes information that changes the interpretation.
Record A also cannot be assigned a definite position relative to B from these values alone. A value of 150 nM and a value of 500 nM would both satisfy A’s reported bound. One would fall below B’s 200 nM result; the other would fall above it. A precise A-versus-B order would add evidence the records do not contain.
Record C supports a below-100 nM statement, under the example’s assumptions. It does not support an exact numerical distance from D. Missing units in E require a different response again: identify the missing information rather than infer the unit from nearby rows.
What the isolated check actually established
For this article, we extracted only the conversion-factor dictionary and _normalize_activity function from the pinned helper into a local Python 3.12.14 check. We passed in the five invented records above. The numeric outputs were 100, 200, 100, 100 and null, respectively. The original relation field remained present and unchanged in every result.
This was an isolated function check. It did not import the complete skill package, query ChEMBL, install the Science plugin, run an agent or reproduce Google’s scientific demonstration. It establishes the behavior of that extracted transformation for these inputs.
The result is also why the proposed control belongs downstream. The helper preserved the relation. A later assistant, dataframe projection, chart or export could still select only the numeric field. That would be a change in meaning introduced after a successful conversion.
Ask to see the final decision table, including its column definitions. A green execution status or a correct conversion note cannot show whether the explanation preserved the underlying evidence.
Research uses explicit selection rules, not just common units
A June 2026 study by Schiebroek, Landrum and Riniker compared bioactivity-data curation and modeling approaches using ChEMBL34. Its inspected methods specify equality-only measurements among several selection criteria, including a single protein and controls concerning target variants. This is a research dataset definition, not a universal requirement that every scientific application discard bounded observations.
That distinction matters. A workflow intended to fit an exact-valued regression dataset may exclude qualified records under an explicit protocol. A literature evidence table may need to retain those same records and display their bounds. A specialist method may handle bounded measurements directly. The assistant should not silently choose among those purposes.
Practitioner work also makes selection visible. In his July 2026 RDKit post, Greg Landrum develops queries for identifying lead-optimization datasets, with explicit assay, target-confidence and document criteria. It is an exploratory practitioner account, not validation of Google’s agent. Its useful lesson here is that selecting a dataset is a documented analytical decision.
Common units alone do not settle differences in endpoint, assay conditions or target assignment. Our example assumes those differences away to isolate the measurement-bound problem. A real research team must review them separately before making a scientific comparison.
Add a bound-preservation check to one pilot workflow
The following is our proposed acceptance approach for a team building or evaluating an AI research workflow. It is not an observed product capability.
Have the scientific owner define what the output is for: an evidence inventory, a compound-prioritization discussion or a modeling dataset. Record how that purpose treats qualified measurements and missing information. Avoid a generic instruction such as “clean and rank the data.”
Then give the workflow a controlled input containing an exact value, an upper bound, a lower bound and a missing unit. Set the expected interpretation before running the assistant. Include the A-versus-B case above so that a numerical sort cannot pass merely by producing a neat table.
Review three outputs separately:
- The retained data: original value, unit, relation, endpoint and source identifiers remain available alongside derived fields.
- The displayed result: tables and charts distinguish exact values from bounds and explain exclusions or unavailable conversions.
- The written recommendation: unresolved comparisons stay unresolved, and any ordering states the assumptions that support it.
If the workflow exports data to another tool, inspect the exported file too. A correct narrative does not repair a CSV that has lost its relation column. Likewise, an intact source JSON does not repair a chart that presents every converted number as an exact observation.
Start with one scientific question and one controlled set of records. If the assistant forces the ambiguous pair into a precise order, revise the transformation or instructions and repeat that case. Keep the experimental prioritization decision with the research team, supported by the measurements and limitations the workflow actually preserved.