When an assay loses its identity, everything downstream starts lying

by Marina VasquezHead of Semantic Systems, HMND

Pharma keeps buying interoperability as if it were furniture, then acts surprised when the room still does not hold meaning. Assay, sample, study, and material identities have to survive the trip across ELN, LIMS, MES, data lake, and analytics stack, or research and manufacturing end up debating four different entities with the same label and a shared vocabulary of disappointment.

The pressure is real. Regulated product data is being pushed toward more explicit semantic structure through standards work around IDMP, while laboratory and research systems are being asked to speak in terms that survive exchange instead of merely export. CDISC, OBI, BAO, LOINC, and related ontologies exist because a flat field called sample_id has never been enough to carry provenance, context, and permitted meaning from one system boundary to another.

The part everyone pretends is solved

A study in a discovery platform can call something a specimen, a biospecimen, a matrix, or a sample, then a manufacturing system receives a batch record that points to the same physical thing with different timing, different status, and a different reason for existence. The record moves. The identity often does not.

That is where teams start arguing in meetings that should have been spare, technical, and almost boring. The assay owner says the result came from the correct material. The manufacturing lead says the material is a different lot. The data engineer shows a join that worked exactly as written, which is the kind of success that creates a larger failure.

Semantic infrastructure is the refusal to let those arguments remain vague. A shared ontology gives the nouns shape. Identity resolution gives them continuity. Together they decide whether a study references the same assay object in every system or only a familiar looking string that has been copied through three interfaces and one spreadsheet exile.

Where the failure shows up

I have seen this failure in the ordinary places, which is where the damage is honest. A lab uploads assay results into one repository, manufacturing reconciles the release data in another, and someone in the middle decides a dashboard can bridge the gap. The dashboard looks civilized. The identifiers underneath have already split apart.

Once that split happens, compliance reports get patched by hand, QC trends blur, and a supposedly searchable data asset becomes a museum of near matches. The problem presents as integration pain because that is a respectable label. The real issue is that the systems no longer agree on meaning, and the disagreement has a checksum.

The useful engineering move is usually unglamorous. You define the entity that should exist. You assign a durable identifier. You map synonyms, parent child relations, provenance, and allowed transformations in an ontology that people are forced to maintain, not admire. OBI and BAO help with assay and experimental structure. IDMP helps with medicinal product identity. CDISC and LOINC help keep clinical and laboratory terms from drifting into private dialects.

A warehouse full of copied fields does not become semantic just because someone drew a box around it and called it a graph. GraphDB, Protégé, RDF, OWL, SPARQL, whatever the stack, none of it matters if the organization refuses to own the meaning of a sample once it leaves the originating system. FAIR as a slide title will not rescue a pipeline that cannot tell whether a material is a parent, a derivative, or a late arriving correction.

The hard part is social, which is annoying in the way only architecture can be. Somebody has to accept that the identifier is a shared contract across research and manufacturing, and that contract survives only if every boundary treats it as authoritative. Otherwise each team keeps its own local truth, and everyone spends the quarter reconciling stories that should have been the same story from the start.

The question worth asking in the room

When a sample leaves discovery and enters a regulated workflow, does the identity travel with it intact, or does it get translated into a different creature by each system that touches it? If the answer depends on who exported the CSV, the infrastructure is decorative and the ontology is ornamental.

If this kind of meaning problem is sitting inside your stack, hello@example.com is where the conversation starts.