Two published studies, encoded as Negative-Results Records

Summary

The data model held. Nothing in the v0.1 sketch had to be renamed; twelve things were added, all about provenance and epistemic status. The records validate, the build is deterministic, and cross-study questions run over shared identifiers and vocabularies.

The two studies

What a record holds

A wet-lab attempt from the Robin study carries one finding per arm of the screen. This is the round-two ripasudil finding, abridged: the compound identifier, the dose from the supplementary drug table, the per-well values from the supplementary workbook, the designated positive control with its observed value, the vehicle control, the sensitivity statement from the Reporting Summary, and who proposed the compound.

{
  "question": "Does Ripasudil at 100 µM increase phagocytosis by ARPE-19 cells of pHrodo beads (1 h pre-treatment, 3 h uptake)?",
  "target": {"label": "Ripasudil", "identifier": "pubchem:CID9863672", "type": "compound"},
  "outcomeClass": "positive",
  "effect": {"metric": "normalised MFI (fold of DMSO)", "value": 1.8906, "n": 3, "sem": 0.2378,
             "perReplicate": [2.3619, 1.7096, 1.6002], "test": "Dunnett's test versus DMSO"},
  "controls": {"positive": {"kind": "designated", "label": "MFGE8", "observed": "1.46x (n=2)", "passed": true},
               "negative": {"kind": "vehicle", "label": "0.5% DMSO", "passed": true}},
  "sensitivity": "Reporting Summary: 3 replicates reliably distinguish the MFGE8 control from vehicle",
  "conditions": {"dose": "100 µM", "vendor": "Cayman Chemical Company", "catalog": "19920", "doseReference": "doi:10.3390/ijms25063229"},
  "proposedBy": "robin-round-2", "proposedByConfidence": "inferred"
}

The Alcalase knock-in finding shows the rule refusing to call a negative evidence of absence: no positive control for editing existed, so the class is inconclusive-no-positive-control, and the record states the detection limit that the 40 sequenced colonies do support.

{
  "outcomeClass": "inconclusive-no-positive-control",
  "informativeness": "uninformative",
  "informativenessReason": "no positive control",
  "sensitivity": "0 of 40 sequenced colonies edited: an on-target rate above 7.5% among NAT-resistant colonies
                  would have been detected with 95% probability (rule of three, derived by portolan-ingest)"
}

What we learned

1. The negative results were never stated. The Robin paper reports three hits. Its supplementary workbook holds about ninety compound arms that did nothing. The claim “fingolimod at 1 µM does not increase ARPE-19 phagocytosis” had to be assembled from four places: per-well values in a sheet, a dose in a PDF table, a sample-size sentence in the Reporting Summary, and a hit call in the main text. That assembly is exactly the work a registry does, and the pilot shows it can be mechanised once someone decides what counts as a claim. Robin itself needs it: its second round tells the model “these drugs have already been tested, do not suggest them again”. That memory lives in one prompt and dies with the run.
2. Post-hoc ingestion from papers is lossy. The repository’s sample trajectories are re-runs whose candidates are not the ones the paper tested. Candidate counts disagree between the text and the tables. About 25 compounds in the primary-cell screen have no stated proposer. The primary-cell RNA-seq has no accession. All of it would have been captured for free had the agent deposited at run time. Passive capture at generation time is the right design, and this is the evidence.
3. A negative is a property of the data and the analysis together. The same flow-cytometry files put the positive control at 1.04× under the Finch agent and at 1.46× under a human analyst. A registry that stores “no effect” without the analysis identity will be wrong some of the time, so the analysis is its own record, linked to the experiment.
4. The informativeness rule earned its place. It refused to read 0 of 40 edited colonies as “editing does not work” because no editing control existed. It caught the failed control in the agent analysis. And it showed that ninety Robin negatives are informative only because of one sentence in a Reporting Summary. That form was the most valuable structured input in the paper.
5. Licences shape the corpus. A non-commercial, no-derivatives licence means a record built from that paper must be facts plus pointers, and its supplementary data cannot be redistributed. A CC BY source had no friction. The commons should start with permissively licensed venues, preprints and lab-native capture.
6. Curation judgement is the real work. Which arms are controls, whether every non-hit is a negative, which agent proposed which compound. Each call is flagged in the record, but an expert should review them. A production ingest is an agent proposing records and a reviewer approving them, not a converter.

Schema changes forced by the data

The full list with the reason for each change is in the change log. The ones that matter most:

Questions an agent can ask

The store answers eleven named queries. Examples, with what came back:

• Has ripasudil been tried as an RPE phagocytosis enhancer? Three records: proposed in round two, 1.89× in ARPE-19 at 100 µM, 1.54× in primary cells.
• What do we know about fingolimod at what dose? Three informative negatives at 1 µM, with the sensitivity statement attached.
• Which negatives are not evidence of absence, and why? Four findings across both studies, each with the failing or missing control named.
• Where could an agent pick up work? Sixteen entries: the authors’ next steps, their open questions, and deposited data nobody analysed.
• Are the DOIs this record depends on still intact? Twenty-five DOIs for the Arcadia record, each with version and integrity status and the time of the check.

What is published, and what is not

• Code and schema: github.com/portolansoft/nrr, Apache-2.0 for code, CC BY 4.0 for the schema documents. Both converters are there. The Nature-derived inputs are fetched by a script at build time and are not redistributed.
• Records: github.com/portolansoft/nrr-records, CC BY 4.0. The seven Arcadia records are published, with attribution to Caddell et al. 2026 and to the Zenodo dataset they build on.
• Not published: the eleven records derived from the Nature paper. Its licence forbids distributing adapted material, and per-well values plus quotations from a paper are arguably that. We would rather the authors deposit their own results, and we are asking. The converter that produces them is open so anyone with the paper can rebuild them locally.

Next steps

1. Define a minimum viable record centred on the finding. The full schema is too heavy to fill by hand; the minimum is a question, a target, an outcome class, controls, a sensitivity statement, the conditions, and one evidence pointer.
2. Run the ingest as an agent-plus-review loop on ten more permissively licensed sources and see which schema additions recur.
3. Prototype capture at run time with an open agent system: the insights object Robin passes between rounds is already a subset of a record.
4. Put the two discrepancies the pilot surfaced, a compound that worked in primary cells but not in the cell line and a control that passed under one analysis and failed under another, in front of a biologist.
5. Settle the licensing question for non-commercial, no-derivatives sources before ingesting more of them.

Comments and corrections to [email protected]. Back to portolansoft.com