Scientific data fabric fails where meaning is vague

by Marina VasquezHead of Semantic Systems, HMND

R&D teams keep buying the promise of data fabric as if the hard part were moving bytes fast enough, when the real break point sits one layer deeper: what a sample, assay, patient, or batch means across systems, and who owns that meaning when the record crosses a boundary. The organizations that get this wrong end up with elegant access to inconsistent nouns, which is a very expensive way to stay confused.

The current momentum is real

The momentum around data fabric is not imaginary. Recent industry and standards writing keeps circling the same mechanism: connect distributed sources through metadata, APIs, governance, semantic services, and often knowledge graph style enrichment so users can query across systems without centralizing everything first.

RDA materials describe data fabric as machinery for reproducible data science, with actionable policies, self documenting steps, and interfaces that keep scientists out of infrastructure weeds unless they are unlucky enough to be paged into them. Gartner and several vendors describe the same shape in more commercial language, but the useful core survives the polish: metadata has to stay active, a shared semantic layer has to do real work, and access has to function across locations without flattening the differences that matter.

Identity is where the project lives or dies

A federated query layer can only answer the questions you think you asked. If one lab calls a vial a sample and another system treats the same thing as a specimen, a derivative, or a child entity with its own lifecycle, the query will return numbers with the confidence of a liar in a clean shirt.

This is why interoperability work keeps failing in pharma. Teams announce API readiness, then leave ownership vague, then discover that the same patient, assay, or batch exists in three identifiers with three stewardship models and one silent reconciliation job nobody trusts. The warehouse fills. The disagreement remains.

FHIR helps because it gives clinical systems a common resource model and a defined exchange grammar, and HL7 frames it as a standard for exchanging health care information across systems. FHIR still leaves semantic drift sitting in the room. The EHR, LIMS, ELN, MES, and CDISC sidecars can all move data and still disagree about what an entity is unless someone defines identity across the boundary.

The scene in the lab is usually small

It is 7:40 on a Tuesday in a translational review room, and somebody is trying to line up an assay result from discovery with a patient cohort from clinical ops and a release event from manufacturing. The dashboard is pretty. The definitions are not.

One team says the sample is the original draw. Another means the aliquot used in sequencing. A third means the batch of material that survived processing and only exists because the plant needed a release label. The federated query runs anyway, because software rarely stops you from being wrong.

That is the part people miss when they call this an integration issue. Integration is the transport. Meaning is the decision.

What changes when the meaning model is right

Once ontology and identity are treated as first class design objects, portfolio decisions change because leaders can see which programs share material, which assays truly replicate across sites, and where a manufacturing constraint will invalidate a translational claim before the claim gets dressed up as insight. The same semantic layer that reconciles specimen lineage also sharpens operational handoffs between lab, clinical, and plant, because a batch failure can be traced back to the exact material state that created it instead of to three incompatible export files.

That changes translational analysis in a very direct way. A scientist can ask whether the biomarker seen in discovery survives the move into patient material without pretending that every source system already agrees on the entity being measured. A manufacturing engineer can ask whether a release deviation touches the same material lineage that the clinical team is using in its analysis. The query stops being an argument with spreadsheets and becomes a test of shared definition.

The failure mode is familiar

A company builds APIs, adds a catalog, buys a semantic product, and declares victory while ownership stays blurred. Then every group invents its own rules for identity resolution, and the fabric becomes a polite wrapper around contradictions. Gartner’s framing of data fabric depends on metadata that can actively guide access and organization, and that guidance collapses when stewardship is undefined.

Marina will call that a meaning failure, and she would be right. Daniel would call it an API problem and still be half wrong.

The practical fix is boring in the best possible way. Name the entity. Name the steward. Name the allowed transformations. Then let the federated query layer do the work it was actually built for, which is cross system reasoning over agreed semantics, not ceremonial access to incompatible records.

If this handoff problem is sitting in your stack, write hello@example.com. We build the systems side of that mess.