Semantic data infrastructure is the part of the stack that decides whether identity survives

by Marina VasquezHead of Semantic Systems, HMND

Everyone keeps buying more surfaces to stare at. The scorecard grows, the portal gets shinier, and the sample still arrives in another system wearing a different name, with the history shaved off at the boundary.

Biopharma has spent years talking about platforms, interoperability, and federated data as if those words were destination signs. The current momentum is real. Federated research environments are now being built around shared services, common security rules, and explicit authorization for users and environments, which is exactly the kind of disciplined plumbing serious science needs. The Pistoia Alliance’s Pharma General Ontology work points in the same direction, pushing shared semantics so pharmaceutical concepts can mean the same thing across domains. FAIR guidance keeps repeating the same hard truth in machine language: persistent identifiers, rich metadata, formal vocabularies, qualified references, provenance. Good. The industry finally has enough vocabulary to admit the crime scene.

The failure mode is embarrassingly familiar. A discovery team registers an assay run in one system, a translational group copies the result into a warehouse, and manufacturing later receives a material record that sounds close enough to be the same thing. Close enough is how identity dies. The sample becomes a row, the study becomes a folder, the material becomes a string in a payload, and every team thinks integration happened because the export completed without error. It did not. The meaning moved out of the object and nobody put it back.

Meaning has to survive the handoff

Semantic data infrastructure is the thing that decides whether an entity remains itself after it crosses a system boundary. That means assay, sample, study, and material need durable identifiers, versioned definitions, and explicit relationships that survive mapping into LIMS, ELN, CTMS, CDMS, ERP, MES, and cloud analysis workspaces. Federated research platforms already show the architectural shape of this problem. Common services can move data. Common security can gate access. Shared semantics decides whether two records are actually talking about the same biological event, the same specimen, or the same batch lineage.

A warehouse alone cannot do that. A warehouse stores what was sent. A semantic layer stores what it means, when that meaning changed, and which upstream system is authoritative for which property. That distinction matters the moment a clinical site reconsents a participant, a bioanalytical lab reruns an assay, or a manufacturing lot is split and renumbered. The entity changed. The software has to know that, and so do the humans.

FAIR gets flattened into ceremony when organizations stop at findability and treat interoperability as a checkbox. The sources say otherwise. Interoperability requires formal, shared language and vocabularies, plus qualified links among metadata and provenance that can be traced back through the chain. In practice, that means an ontology service or terminology service at the center of the stack, not as decoration but as the authority for controlled concepts, synonyms, mappings, and version histories. It means identifiers that are globally unique, resolvable, and persistent. It means event sourcing or immutable audit trails for state changes so downstream systems can answer a simple question without guessing: what was this thing at the time it was used?

The architecture choice is where the argument becomes real

Discovery, clinical, and manufacturing teams do not need more dashboards to admire. They need agreement about identity at the point where systems disagree. That agreement usually lands in a few concrete design choices.

A canonical entity model has to exist, even if each domain keeps its own local shape. Not a fake universal table, just a semantic contract for core objects and their state transitions. An assay result, for example, should carry the assay definition, protocol version, instrument context, sample identifier, and provenance of transformation. A sample should carry collection time, source subject or material lineage, processing state, and custody chain. A material should carry composition, lot, derivation, and the exact timestamp when a status changed. Without those fields, every downstream reconstruction is a guess dressed as analytics.

Security boundaries matter too. Federated systems work only when authorization follows the data and the environment, with explicit trust rules for where sensitive data can be hosted and who can access it. In pharma, that boundary is where the clinical privacy regime, the quality system, and the manufacturing record all collide. A sane architecture keeps raw identifiers, operational metadata, and regulated content in the right zones, then exposes only the semantic links needed for cross domain work. The trick is to share meaning without leaking control.

A proper implementation also needs machine readable terminology services, not just a document full of approved terms. Controlled vocabularies need versioning. Mappings need stewardship. Provenance needs to be queryable, not buried in a PDF or some integration ticket nobody owns anymore. That is the whole game. If a scientist can ask why a sample appears under two names, or why an assay changed interpretation after a protocol update, the system is doing meaning work. If the answer requires a meeting, the architecture failed earlier than the meeting did.

The point

The market has momentum, and that is useful, but momentum alone will not rescue sloppy identity. More platforms will not fix a sample that loses its lineage between discovery and GMP. More dashboards will not tell a study when its meaning changed. Semantic data infrastructure does that work, because it decides whether the thing the system stores is still the thing the organization thinks it has.

If this is the kind of handoff problem sitting on your desk, hello@example.com is where the conversation starts.