The Standards Bodies: Who Actually Builds Biopharma's Shared Vocabulary
📍 Where we are: Part VIII · Ontologies in Industry Today. The earlier parts validated, governed, and FAIR-measured the whole model on one antibody batch. Now we leave the bench of our own example and ask the empirical question the earlier chapters quietly assumed away.
For twenty-three chapters we leaned on a comforting phrase: the shared standard. We anchored bp:DS-001 to an IDMP substance identifier, aligned our classes to BFO, borrowed mid-level scaffolding from IOF Core, and validated with SHACL. Each time, we spoke as if some authoritative vocabulary simply existed, waiting to be imported. It is time to be honest about who writes those vocabularies — because the answer reshapes how a real plant should plan its semantics.
The short version is that almost none of it comes from a software vendor. In biopharma, the engines of shared meaning are pre-competitive consortia: alliances where rival companies pool effort on the parts that confer no competitive advantage — the words, the file formats, the identifiers — precisely so they can compete on the science instead. Each consortium owns a slice of the process. None owns the whole. The result is a patchwork that is converging, but slowly.
Imagine a dozen rival restaurants that all need the same thing: an agreed list of ingredient names, so a supplier's "caster sugar" means the same thing in every kitchen. No single restaurant should own that list — the others would never trust it — so they form a club to maintain it together. Biopharma has not one such club but a handful, each curating a different aisle of the pantry: one for lab instruments, one for the drug substance, one for the recipe, one for the shipping label. The food is fine. The trouble is that the aisles were stocked by different clubs that do not always agree on where the shelves end.
What this chapter covers
This chapter is a map of the consortia, not the vocabularies themselves — those get their own chapter next. We hold one distinction firm throughout: a true formal ontology is an OWL/RDF artifact a reasoner can run over, whereas a structured information model is an XML or object schema that organizes data without formal logical semantics. The two are easy to conflate and must not be. We tour the Allotrope Foundation and the Pistoia Alliance as the leading formal-ontology consortia; MESA International, the OPC Foundation, and PROFIBUS & PROFINET International as the structured-model bodies behind ISA-88 and ISA-95; ISPE and BioPhorum as the maturity-framework authors; and GS1 together with the OAGi/NIIMBL collaboration for identification and emerging biomanufacturing ontologies. Every adoption claim carries a maturity tag in bold parentheses, because a published spec and a deployed one are years apart.
Each pre-competitive consortium owns one aisle of biopharma's shared vocabulary, color-coded by how mature its work really is, and no single body owns the whole.
Original diagram by the authors, created with AI assistance.
Read the figure's colors as the maturity tags made visual: a green card is (production), amber is (piloted), violet is (proposed). Every adoption claim below carries the same tag in bold parentheses, so the map and the prose are one argument seen two ways.
The formal-ontology consortia
Two consortia do the genuine OWL-and-BFO work this book has assumed.
The Allotrope Foundation is the most industrially adopted formal-ontology effort in the lab-data space (production). It was formed on 4 June 2012 as an independent legal entity, spun off from the IQ Consortium — the International Consortium for Innovation and Quality in Pharmaceutical Development [1]. Allotrope publishes a three-part framework: the AFO (Allotrope Foundation Ontology), an OWL controlled vocabulary for analytical and lab data that draws on BFO-style upper distinctions but mints its own terms under purl.allotrope.org (the af-r:/af-p: namespaces this book aligns its SEC result and assay to); the ADM (Allotrope Data Models); and the ADF (Allotrope Data Format), an HDF5-based binary data container whose v1.0 shipped in October 2015 [2]. Its weight comes from its members: nine Foundation Members — Amgen, BASF, Bayer, Boehringer Ingelheim, Dow, Genentech, GSK, Johnson & Johnson, and Merck & Co. — with instrument vendors such as Agilent, Bruker, SCIEX, Shimadzu, Benchling, and BIOVIA seated in a separate Partner Network tier [1]. When an earlier chapter spoke of importing a controlled term for an SEC result like our 98.611 % monomer, AFO is the kind of vocabulary it had in mind. What AFO buys a real release lab is breadth: not just the SEC monomer and aggregate terms behind that 98.611 % result, but a controlled word for every attribute on the drug-substance Certificate of Analysis — CE-SDS purity, the HILIC N-glycan map (afucosylation, high-mannose, galactosylation), charge variants by icIEF or cation-exchange, host-cell protein and residual DNA, and the cell-based potency assay. That analytical spread — a dozen orthogonal methods, each on a different instrument from a different vendor — is exactly why a single shared lab-data vocabulary is worth a foundation's effort, and exactly why no one company could impose one.
The Pistoia Alliance is the central vehicle for pre-competitive pharma ontologies more broadly (piloted to production, by component). It was incorporated in 2008 by representatives of AstraZeneca, GSK, Novartis, and Pfizer who had met at a conference in Pistoia, Italy; it now reports more than 200 members [4]. Its portfolio is a useful index of what the industry is trying to standardize, even where the work is unfinished:
| Project | What it standardizes | Maturity |
|---|---|---|
| IDMP Ontology (IDMP-O) | Substance and product identification; v1.0 released 24 January 2024, open-source, co-developed by 11 pharma companies [3] | (production) by component |
| Unified Data Model (UDM) | Open chemical-reaction exchange; v6.0 January 2021, MIT license | (production) |
| Methods Hub | Machine-readable analytical-method transfer | (piloted) |
| Pharma General Ontology (PGO) | A shared upper-level reference targeting the "FAIR silos" problem — data FAIR within a company but not across companies | (proposed) |
| CMC Process Ontology | An ISA-88-aligned process vocabulary; v1.0 announced for around mid-2026 [5] | (proposed) |
IDMP-O is exactly what sits behind our bp:DS-001 in a real plant: the formal ontology a manufacturer uses to express the regulated substance identity. PGO is worth flagging for its candor — it exists because FAIR within one company has proven insufficient across companies, the precise gap that Part VII's FAIR scoring would expose the moment our model met a partner's.
One layer of IDMP is worth making explicit, because the running example carries it. IDMP is not one identifier but a family, and its first cut is substance versus product: ISO 11238 governs the substance identity (in the US, the FDA's UNII / GSRS code), while the Medicinal Product Identifier (MPID) and Pharmaceutical Product Identifier (PhPID) name the marketed product. That split is exactly why the drug substance bp:DS-001 and the drug product bp:DP-001 cannot share one regulated name. In instances.ttl, bp:IDMP-DS-001 accordingly carries both a bp:uniiCode (the substance side) and a bp:mpid (the product side) — a single bp:SubstanceIdentifier node holding the two ISO-11238/IDMP coordinates the regulator actually reasons over.
One consortium the rest of this chapter under-credits, because it predates the biopharma-specific bodies, is the OBO Foundry — the life-sciences alliance (RO, OBI, ChEBI, PR, GO, NCBITaxon, IAO, PATO) whose coordinated, BFO-grounded, PURL-published ontologies are the maturest pre-competitive semantic commons in any industry, and the explicit template IOF was modeled on. It is also the consortium our own model reuses most: align.ttl binds the antibody to GO's IgG immunoglobulin complex, the release sample to OBI's specimen, the HCP ELISA to OBI's enzyme-linked immunosorbent assay, the host cell to NCBITaxon's Cricetulus griseus, and antibody–target binding to RO's molecularly interacts with (RO_0002436). When the OAGi/NIIMBL effort says it aligns to "BFO and IOF Core," the governance discipline it is buying into — versioned PURLs, the Foundry principles, terms that interlock rather than overlap — is OBO's, ported to manufacturing; the Vocabularies in Use chapter walks these OBO/RO/OBI/AFO edges term by term.
The structured-model bodies
The next tier produces information models, not formal ontologies — and the distinction earns its keep here, because you cannot run a reasoner over an XML schema the way you can over OWL.
The backbone is the ISA standards family: ISA-88 for batch recipe structure and ISA-95 (also published as IEC/ISO 62264) for connecting the manufacturing execution system to the enterprise, with object models for equipment, material, and personnel (production) [6]. The two standards split the labor cleanly: ISA-88 governs a recipe's internal structure — phases, operations, the order in which steps run — while ISA-95 governs the conversation between the recipe-runner on the floor and the business systems above it, supplying shared object models for equipment, material, and personnel. Keep that division in mind, because it is the seam the rest of this section runs along: B2MML and OPC 10030 carry the ISA-95 side, while the recipe structure ISA-88 describes is exactly what the still-proposed CMC Process Ontology is trying to give formal words. ISA-88's procedural hierarchy (procedure → unit procedure → operation → phase) is what gives a CHO fed-batch run its recipe spine: the seed-train expansion, the production-bioreactor operation with its scheduled bolus or continuous feeds, and harvest each map to that structure — which is why the Pistoia CMC Process Ontology choosing an ISA-88 alignment is the right backbone and not an arbitrary one. These standards describe their objects in prose and tables; turning them into something machines exchange falls to implementations.
MESA International supplies that implementation as B2MML (Business to Manufacturing Markup Language), a royalty-free, W3C-XSD rendering of the ISA-95 family. Its latest is B2MML Version 7 (V0700), announced 24 November 2020, aligning with the 2018 ISA-95 editions and adding a first B2MML-JSON specification [6]. B2MML is widely described as the de facto MES exchange layer in pharma — but that reputation rests largely on system-integrator assertion rather than published adoption data, and should be read as such (production, vendor/integrator-reported).
Where B2MML carries transactional data, the OPC Foundation adds real-time semantics through the OPC UA for ISA-95 Common Object Model (OPC 10030), Release 1.00 dated 6 November 2013 — a companion specification that lets live equipment speak the same object vocabulary B2MML moves between systems (production) [7].
It helps to see what each layer actually moves off the floor for one batch. OPC UA (and OPC 10030) carries the live signal — the production bioreactor's pH, dissolved oxygen, temperature, agitation, and feed-pump setpoints and process values, second by second. B2MML carries the transactional record after the fact — the MES batch record: material genealogy, equipment usage, the as-executed recipe, and the operator and QA dispositions. AFO/ASM carries the analytical result — the discrete SEC, CE-SDS, glycan, and charge-variant numbers off the QC instruments, including the 98.611 % monomer our release gate read. Three standards, three data shapes, one batch — which is why the running example needs a separate loader (opcua_to_rdf.py, b2mml_to_rdf.py, asm_to_rdf.py) to pull each into the graph. The unit a value wears survives the crossing because OPC UA's EUInformation and B2MML's UnitOfMeasure field both carry a UCUM code, which the loaders deposit as qudt:ucumCode / qudt:hasUnit — so a setpoint arrives at the graph already bound to a QUDT quantity kind. The XML model transports the datum; the ontology types it so a reasoner and a SHACL gate can check it. The loader boundary itself is owned by from the wire to the graph.
That unit-preserving path is also why these information-model bodies, not just the formal-ontology ones, quietly underwrite the machine-learning frontier. A model is only as trustworthy as the ground truth it stands on, and the consortia are who govern that ground truth. The AFO/ASM analytical numbers and the B2MML/OPC-10030 process record arrive at the graph already typed and unit-bearing — a feed rate of 36.5 degC against its qudt:hasUnit, a 98.611 % monomer against iof:CaptureStep — so when a soft sensor learns titer from a Raman spectrum or a release classifier learns pass/fail from a panel, its features are governed quantities, not bare floats a historian emitted with no units. That is precisely the difference the ML book calls leak-free, contextualized fuel: a PLS or gradient-boosted model trained on bp:-typed, IOF-classed instances inherits the consortia's identity discipline for free, where a model trained on a flat export inherits nothing. Two consequences follow that are easy to miss. First, because every record carries its batch identity (the same bp:BATCH-2026-001 lineage the loaders preserve), the honest validation is grouped — leave-one-batch-out cross-validation over whole batches, never a row-wise shuffle that lets two near-identical spectra from one run straddle the train/test line, the leakage taxonomy the ML book treats as its cardinal sin; the consortia's job is to make that batch key survive every standard crossing so the split can be grouped. Second, a graph built on these governed vocabularies is the ground truth that constrains a fluent model: this book's own ontologies-as-ground-truth chapter shows a GraphRAG answer traversing the typed bp:derivedFrom edges and citing them, and the same SHACL gate that refuses a non-conformant release refuses a non-conformant retrieval — so the standard a consortium publishes is, read in this light, the schema that keeps a language model from inventing a lineage the plant never ran. The vocabulary the consortia stock is the substance; the model only ever supplies the structure.
Newer, and aimed more squarely at modular plants, is the Module Type Package (MTP), standardized as VDI/VDE/NAMUR 2658 and built on OPC UA plus AutomationML to enable "plug-and-produce" modular automation. Its governance moved to PROFIBUS & PROFINET International (PI) in November 2021, jointly with NAMUR and ZVEI; MTP V2.0.0 was released in the fall of 2024, and the work is being internationalized as IEC 63280 (piloted) [8]. Pharma and biotech are repeatedly cited as the leading-interest sector for MTP — a reasonable signal of intent, not yet of broad production use.
The maturity-framework bodies
A third kind of consortium standardizes not vocabulary but how far along you are in using it. These are frameworks, and their maturity tag is (production, as a framework) — meaning the framework itself is real and used, not that any given plant has reached its top rung.
ISPE publishes the Pharma 4.0 work: Baseline Guide Vol 8 (Pharma 4.0, 1st edition), released December 2023, which extends ICH Q10 and incorporates 35 use cases [9]. It contains a Maturity Model and Self-Assessment built on the acatech Industrie 4.0 Maturity Index — a six-stage model, so the common "five-level" shorthand is imprecise — across four operating-model dimensions: Resources; Organization and Processes; Information Systems; and Culture [9].
BioPhorum maintains the Digital Plant Maturity Model (DPMM), a five-level model running from paper-based operations to a self-optimizing autonomous plant. DPMM V3 was published in October 2023, adding split QC and QMS dimensions and a Process Development dimension, and renaming a security dimension "Cybersecurity" [10]. A caution on a figure you will see quoted: the frequently cited claim that roughly 80% of members use the DPMM traces to the earlier, pre-V3 tool — treat it as legacy, self-reported evidence, not a current measurement (production, as a framework) [10].
Identification, and the newest entrant
For identification, the converged backbone is GS1: the GTIN, serial number, and GLN encoded in a 2D DataMatrix, with EPCIS for event exchange — the standards that now underpin pharmaceutical serialization (production) [11]. US DSCSA manufacturer enforcement began 27 May 2025; the EU's Falsified Medicines Directive has been live since 9 February 2019. One nuance matters for accuracy: DSCSA mandates interoperable electronic traceability but does not name GS1 as the sole permissible format — the industry chose GS1 EPCIS to meet the requirement [11].
These identifier standards are also where this consortium map touches data integrity, the frame that decides whether a regulator trusts any of it. A serialized SGTIN, an IDMP substance code, or an AFO-typed CoA result is only as defensible as the record that carries it: it has to be attributable, legible, contemporaneous, and the rest of the ALCOA+ attributes, under Part 11 and Annex 11 and the shift from prescriptive CSV to risk-based CSA. That is the quiet payoff of the unit-and-identity discipline above. When a bp: instance binds an SGTIN to a lot, a UNII to bp:DS-001, and a UCUM-coded value to a release result, the ontology is not just naming things — it is making the audit trail machine-checkable, so a SHACL gate can assert that every released vial carries its serial and every CoA value its unit and signer before the graph is trusted. A standard with no governed record behind it is a label; a standard whose identifiers land in a validated, attributable graph is evidence — which is the difference between passing an inspection and failing one.
The newest entrant points back toward formal ontologies. In June 2024, NIIMBL contributed the ontologies from its Big Data Program to the Open Applications Group (OAGi) to jointly develop open-source biopharma manufacturing ontologies, aligned to BFO and IOF Core (proposed) [12]. That effort has a published counterpart: the IOF biopharma domain ontology itself. Audited directly against its February 2026 release (Release_202602), it defines 171 Released classes — 44 unit operations, 17 QbD parameters, plus equipment, materials, and recipe terms — the most complete formal vocabulary of the bioprocess interior that exists today, and the one this book's running example now binds to at the class level [13]: the batch class (bp:Batch → iof:MaterialArtifact) and its unit-operation classes (bp:CaptureChromatography → iof:CaptureStep, bp:ViralFiltration → iof:ViralFiltration, and the rest) bind their process steps to it, the individual bp:BATCH-2026-001 only transitively through its type. Of those 171, the campaign's align.ttl actually consumes 20 — part of the 27 distinct IOF classes (7 Core + 20 biopharma) it reuses overall — the same audit that the upper-spine and the-ontologies-in-use chapters draw on, so the three stay numerically consistent rather than each quoting its own count. This is the work most directly continuous with the model this book has built — and yet, published and Released as it is, it still wants the one thing this part keeps circling back to: a plant in production that actually depends on it. The vocabulary has arrived faster than the adoption, which is the clearest marker of how young cross-industry biomanufacturing semantics still are.
The unsolved part: a published spec is not an adopted one
Pre-competitive consensus is slow and partial, and the map above shows why. A different body owns each aisle — Allotrope the lab, Pistoia the substance, MESA and PI the recipe and the plant floor, GS1 the package — and their slices overlap in some places and leave gaps in others. Worse, the dates tell a recurring story: a specification is typically published years ahead of broad adoption, so the existence of a standard says little about whether the plant down the road actually speaks it. Where this chapter could find no published deployment data, that is reported as not found in public evidence rather than as evidence of absence. The consortia have done the genuinely hard part — producing governed, shared vocabulary — but no consortium can mandate its use. Getting an entire industry to converge on what already exists is the unfinished, and largely ungovernable, work.
Why it matters
Every term the earlier chapters anchored to "a shared standard" is, in reality, some consortium's deliberate and governed labor. Knowing who owns which slice is not trivia — it is the difference between a plant importing a vocabulary and reinventing one. It tells you where to reuse AFO instead of writing your own analytical terms, where IDMP-O already encodes the substance identity behind bp:DS-001, and — just as importantly — where, honestly, there is not yet a standard to import and you are building on open ground. The through-line of this book is turning records into knowledge that can be reasoned over under pressure; that knowledge is only as shareable as the vocabularies these bodies maintain. And it locates this book's own model honestly on that map: the bp: vocabulary is a fourth-rung, site-specific namespace that consumes AFO's analytical terms, IDMP-O's substance identity, and IOF biopharma's unit-operation classes, yet is owned by no consortium and mandated by no regulator — which is precisely the position a real plant's local ontology occupies. The consortia stock the aisles; the plant still has to do its own shopping.
In the real world
Stack the map by what is actually deployed, not what is published, and it sorts into three honest bands. One detail to get right before reading the GS1 row: serialization attaches to the unit, not the lot. In instances.ttl, the running example's serialized vial bp:VIAL-DP-001-000042 is the thing that rides the serialization layer — it carries the SGTIN (bp:hasIdentityKey bp:SGTIN-DP-001-000042, a bp:GS1ItemKey cross-linked by skos:exactMatch to a GS1 SGTIN URN), the per-unit identifier a DataMatrix symbol would print, derived from lot bp:DP-001. A GTIN names a trade-item class; the SGTIN = GTIN + serial names the individual saleable vial, aggregated vial → carton → case → pallet — and it is that per-unit serial, not the lot, that DSCSA and FMD make traceable.
| Tier | What is genuinely load-bearing in a plant today | The caveat |
|---|---|---|
| In production | GS1 + EPCIS serialization (now legally mandated — US DSCSA from 27 May 2025, EU FMD since 9 February 2019) [11]; the Allotrope Foundation Ontology for lab data [1] | DSCSA mandates interoperable traceability, not GS1 by name — the industry chose it |
| Production on softer evidence | B2MML as the MES exchange layer; OPC 10030 for live ISA-95 objects [6] | B2MML's reputation rests on integrator assertion, not published adoption data |
| Framework-real, top-rung aspirational | ISPE Pharma 4.0 and BioPhorum DPMM as maturity scaffolds [10] | The DPMM "roughly 80% of members" figure is legacy, pre-V3, self-reported; MTP's pharma lead is intent, not broad production use [8] |
| Released but unadopted | IOF biopharma (171 Released classes) — the work most continuous with this book, which the running example's batch and unit-operation classes bind to [13] | Published and Released, yet no plant in production depends on it |
The shape repeats down every row: vocabulary arrives years ahead of adoption, and a standard with a stable URL is not a standard with a production dependency.
Key terms
- Pre-competitive consortium — an alliance of rival companies that pools effort on shared, non-differentiating assets (vocabularies, formats, identifiers) so members can compete elsewhere.
- Formal ontology vs. structured information model — the former is an OWL/RDF artifact a reasoner can operate over; the latter, such as an XML schema, organizes data without formal logical semantics.
- Allotrope Foundation — the consortium behind AFO, ADM, and the ADF data container; the most industrially adopted formal-ontology effort for lab data (production).
- Pistoia Alliance — the central pre-competitive pharma consortium; home of IDMP-O, UDM, Methods Hub, PGO, and the CMC Process Ontology.
- B2MML — MESA International's royalty-free XML and JSON implementation of the ISA-95 information model; widely asserted, on integrator evidence, as the de facto pharma MES exchange layer.
- OPC 10030 — the OPC UA companion specification mapping the ISA-95 common object model to real-time equipment data.
- Module Type Package (MTP) — VDI/VDE/NAMUR 2658, an OPC UA and AutomationML package for plug-and-produce modular automation, now governed by PI and internationalized as IEC 63280.
- Digital Plant Maturity Model (DPMM) — BioPhorum's five-level framework rating a plant from paper-based to self-optimizing autonomous operation.
- GS1 and EPCIS — the identification and event-exchange standards (GTIN, GLN, DataMatrix, EPCIS) the industry chose to satisfy the DSCSA and FMD serialization mandates.
- Ground truth (for AI) — the verified, governed graph a model is anchored to; the consortia's vocabularies are what make that graph trustworthy, so a model trained or grounded on them inherits identity and unit discipline a flat export lacks.
- Leave-one-batch-out / grouped cross-validation — the honest way to validate a model over instances that carry a batch key: whole batches go entirely to train or test, never split row-wise, so the score answers "how will this do on the next batch?"
- ALCOA+ — the data-integrity attributes (Attributable, Legible, Contemporaneous, Original, Accurate, plus Complete, Consistent, Enduring, Available) under Part 11 and Annex 11 that make a standardized identifier or result regulator-trustworthy.
Where this leads
We have mapped the builders; next we open their products. The following chapter, The Vocabularies in Use: From AFO to IDMP, descends from the consortia to the artifacts themselves — what AFO actually says about an analytical result, how IDMP-O structures a substance, and how these vocabularies would attach to the very batch, pool, and drug product we have modeled throughout this book.