Enterprise Knowledge Graphs at Big Pharma

📍 Where we are: Part VIII · Ontologies in Industry Today. The previous chapter walked the vendor landscape — the engines you can buy. This one asks the harder question: with those engines in hand, what have the largest drug makers actually built, and where did they put it?

A vendor demo and a production system are different animals. The triple stores, the reasoners, and the SHACL validators of the last chapter all work — but a working engine tells you nothing about where a company decided to bolt it in. This chapter inventories the enterprise knowledge graphs — large, cross-domain graphs that unify a company's data under a shared semantic model — that big pharma has put into production. The case studies are real, peer-reviewed or vendor-documented, and impressive at petabyte scale. They also share a pattern so consistent it becomes the chapter's real finding: the graphs cluster almost entirely in research, FAIR-data cataloging, and regulatory master data — and, on the public evidence, they have not yet reached the GMP (Good Manufacturing Practice — the regulated, inspected production environment every licensed medicine must be made under) manufacturing floor where this book's running batch is actually made.

The simple version

Imagine a global company that has finally built one giant, searchable index of everything it knows — every experiment, every dataset, every product record — and it works beautifully. Then you notice the index covers the research labs and the head-office filing cabinets but stops at the factory door. The machines that physically make the medicine are still keeping their own private notebooks. That gap, not the index, is the story.

What this chapter covers

We survey eight enterprise knowledge graphs — six in production and two piloted — across nine named pharmaceutical companies (the Methods Hub pilot is a two-company Merck-and-GSK effort), laid out as a single comparison table and then read closely. We distinguish a true formal ontology — an RDF/OWL (RDF, the Resource Description Framework graph data format, plus OWL, the Web Ontology Language logic layer on top of it) model a reasoner (the software that draws logical conclusions from the model) can act on — from a contextualization program, which maps and links data without committing to a logic-bearing model; the difference is not pedantry, because only the former carries the guarantees earlier parts of this book were built to deliver. We trace where these graphs live: R&D discovery graphs, FAIR-data catalogs, and regulatory master data. We then name the pattern — that GMP-production deployments on the manufacturing floor are absent from the surveyed public evidence — and the coverage gaps in that evidence, chiefly contract manufacturers and Asian producers. Throughout, the maturity of each claim is marked in bold parentheses, because the distance between "shipped" and "announced" is exactly where this book refuses to oversell.

Every deployed pharma knowledge graph in this survey clusters in R and D, FAIR catalogs, and master data, while the GMP manufacturing floor where the running batch is made stays a rose, ungraphed frontier. Original diagram by the authors, created with AI assistance.

The inventory

The table below is the chapter in miniature. Read the Status and Technology columns together: where the technology is RDF/OWL, the company built a formal ontology; where it is a mapping or "contextualization" layer, it did not. One reading rule carries the whole table: scan down the Technology column, and wherever you see RDF/OWL/SPARQL the company committed to a logic-bearing model a reasoner can act on; wherever you see property graph, Foundry Ontology, or contextualized mapping, it built a knowledge graph in the engineering sense — powerful, but without the description-logic guarantees Parts III–VI of this book were written to deliver. (The platforms chapter defines these tooling terms — triplestore, property-graph database, reasoner — and what it means for a reasoner to act on a model.)

Company	System	Status	Technology	What it does
Roche	EDIS / Dataset Portal	(production)	DCAT, Dublin Core, PROV-O, SKOS, FOAF, PAV, JSON-LD over Ontotext GraphDB	FAIR catalog of ~20,000 datasets
Boehringer Ingelheim	Enterprise Knowledge Graph	(production)	RDF/OWL/SPARQL on metaphactory; also a Stardog R&D layer	Federates omics, IT, documents, trial data
AstraZeneca	BIKG	(production, R&D)	Property graph, ~14M nodes / 136M edges, 55 sources	Powers Mantis-ML 2.0 target identification
Novartis	data42	(production)	Palantir Foundry Ontology (earlier AWS Neptune)	Unifies ~20 PB of R&D data
Novo Nordisk	OBDM	(production, R&D)	RDF/OWL reusing AFO, OBI, ChEBI, BFO; SKOS/SSSOM/ROBOT	Inferencing KG over research data
Johnson & Johnson	IDMP-O product master	(production master data)	Pistoia IDMP Ontology + Accurids	Regulatory product master for EMA PMS
Sanofi	Modulus / Connected Smart Factories	(piloted)	Contextualized data mapping across MES/LIMS/SAP	Links factory data — not a formal ontology
Merck (MSD) & GSK	Pistoia Methods Hub	(piloted)	ADF/AFO + a novel RDF graph model	Machine-readable HPLC-UV method transfer

The FAIR catalogs: Roche and Novartis

Roche's EDIS / Roche Dataset Portal is the cleanest example of a knowledge graph built for FAIR — the principle that data be Findable, Accessible, Interoperable, and Reusable. It rests on a stack of public vocabularies — DCAT, Dublin Core, PROV-O, SKOS, FOAF, PAV, served as JSON-LD through a FAIR Data Point — sitting over Ontotext GraphDB, and it catalogs roughly 20,000 datasets (production) [1]. Roche reports an internal FAIR-maturity score of 4.75 out of 5 and reuse of the underlying terminology stack across more than 100 applications [1]. Both are self-reported figures — read them as the company's own assessment, not an independent audit. What is unambiguous is the architecture: this is an ontology in the service of a catalog, the same FAIR pattern Part VII argued for, running at the scale of a global research organization. Worth naming for the modeler: every vocabulary in Roche's stack is a metadata vocabulary about information artifacts — records and labels about the science rather than the science itself, the kind of entity that can be copied without losing its identity. DCAT describes datasets, PROV-O the activities that produced them, PAV the authorship and versioning provenance, SKOS the concept schemes that label them. None reaches into the domain entities — the molecule, the batch, the bioreactor — the way the companion bioproc.ttl does. That is the signature of a FAIR catalog: it is an ontology of the records, not of the science, which is precisely why it stops cleanly at the dataset and never has to model the GMP process underneath.

Novartis's data42 is the same idea at brute scale: roughly 20 petabytes of R&D data and on the order of 2 million patient-years, unified through the Palantir Foundry Ontology, with an earlier iteration on AWS Neptune and SageMaker, and with FAIR described as a stated "mantra" (production) [4]. Tellingly, Novartis was hiring an Ontology Designer to define object and action types and the "semantic contracts" between them [4] — a sign that the discipline this book teaches has become a salaried role, not a research curiosity. One caution on the noun: Foundry's "Ontology" is a structured object model — the preceding platforms chapter unpacks exactly this distinction — and whether it constitutes a formal OWL ontology in the BFO sense is a question the public record does not settle. The petabyte figure is real; the word "ontology" is doing softer work than it does elsewhere in this table.

There is an open-versus-proprietary axis hiding in the Technology column worth reading on its own. Roche's GraphDB, Boehringer's metaphactory-and-Stardog, and Novo Nordisk's standards-based stack rest on open, W3C-standard surfaces — RDF, SPARQL, SHACL, OWL — so the graph is portable by construction: a SPARQL query and a Turtle dump move to Apache Jena Fuseki or any other compliant store without a rewrite, the same portability the open-source knowledge-graph chapter in Book 3 builds on with RDFLib and Jena. Novartis's Foundry Ontology and AstraZeneca's property graph sit at the other end: their object models, query dialects, and "semantic contracts" are vendor-specific, so the data is reachable but the model is not trivially exportable — the classic lock-in the open standards exist to avoid. The trade is real both ways, not a morality tale: a proprietary platform buys integration speed and a managed runtime that a raw triplestore makes you assemble yourself, while the open stack buys exit rights and tool-independence. For a reader following the book's own choices, the point is that the recommended architecture is the portable one, and that portability is exactly what lets the modest companion graph — built with open tooling — be a faithful small-scale specimen of the petabyte deployments above, rather than a different kind of object.

The discovery graphs: AstraZeneca and Boehringer Ingelheim

AstraZeneca's BIKG (Biological Insights Knowledge Graph) is a discovery-side graph of roughly 14 million nodes and 136 million edges drawn from 55 data sources, and it powers Mantis-ML 2.0 for disease-gene target identification (production, R&D) [3]. One caution on the numbers: the 136-million-edge figure is the count as of that 2024 paper, and edge counts drift across AstraZeneca's publications — a widely repeated "146 million" figure is not the value in this source. Treat 136M as as-of-paper, not a canonical constant. Note too that BIKG is a property graph, not an RDF/OWL deployment; it is a knowledge graph in the engineering sense, optimized for traversal and machine learning rather than for description-logic inference. A labeled property graph has no standardized model theory — its edges are not IRIs and it carries no T-Box or entailment regime — so the distinction is not just performance tuning: there is no logical layer for a reasoner to act on, whereas an RDF/OWL graph carries open-world DL semantics by construction.

Boehringer Ingelheim runs an Enterprise Knowledge Graph Platform on metaphactory, using RDF/OWL with SPARQL federation across omics data, IT systems, documents, and the secondary use of clinical-trial data (production) [2]; it also operates a separate Stardog-based semantic layer for R&D. This is a genuine formal-ontology deployment, and the closest of the discovery graphs to the architecture this book recommends. The often-cited drug-discovery and supply-chain use cases, however, are best read as aspirational direction rather than shipped capability — the platform is real; some of the destinations it is pointed at are not yet arrivals.

What the graph is for: it is the ground truth a model stands on

BIKG is the clearest case in the table of a knowledge graph built to feed a model, and that role is worth reading carefully, because it is where the survey touches the companion ML book. Mantis-ML 2.0 does not learn from raw tables; it learns from the graph's typed edges — gene-to-disease, gene-to-pathway, target-to-tractability — and that typing is what makes the graph a defensible substrate rather than a feature dump. A label property graph supplies graph features (a node's degree, its shortest path to a known drug target, an embedding of its neighbourhood) that a downstream classifier consumes; the ontology, where one is present, is what guarantees an edge labelled "associated-with" means the same thing wherever it appears, so a learned feature is not silently averaging two different relations. This is the structure-versus-substance split the AI chapter draws, seen from the enterprise side: the graph supplies verified substance, the model supplies fluent structure over it.

Two cautions the ML book makes precise apply directly here, and they are the reason a graph helps an ML team rather than merely housing its data. First, the validation paradox: a graph that a reasoner has entailed — every owl:TransitiveProperty lineage edge closed, every owl:disjointWith contradiction caught — is in one sense more trustworthy than any learned model, because its conclusions are proven, not estimated, and a SHACL gate can certify the training set is complete and well-typed before a single epoch runs. The book's own SHACL release gate is exactly that certifier: shapes that refuse a non-conformant graph are, repurposed, the check that refuses a hollow or mislabelled training graph — every lot with its bp:derivedFrom parent, every result with its value and unit — before a model learns the gaps as if they were signal. Second, learning over graph instances re-opens the leakage trap the ML book names: rows drawn from one batch, one cell line, or one densely-connected neighbourhood are not independent, so an honest held-out score needs a grouped, leave-one-batch-out split (a GroupKFold keyed on the batch or lineage node), never a row-wise shuffle that scatters near-twins across the train/test line. A 136-million-edge graph makes that trap easy to fall into and easy to fix: the group key is already a node in the graph. The discipline this book teaches — identity, typed relations, validated completeness — is, read in this light, the prerequisite for trustworthy ML over the graph, not a parallel concern; AstraZeneca's pairing of BIKG with Mantis-ML is the production proof that the two travel together.

The research ontology done right: Novo Nordisk

Novo Nordisk's OBDM (ontology-based data management) is the cleanest peer-reviewed account of an industrial pharma knowledge graph, published in 2025 (production, R&D) [5]. It is an RDF/OWL graph that deliberately reuses public ontologies — AFO, OBI, ChEBI, and BFO, the upper ontology this book built on — and is tooled with SKOS, SSSOM, and ROBOT. Its three components map almost exactly to the architecture earlier parts described: a data model, a set of controlled vocabularies, and an inferencing knowledge graph. For a reader who has followed this book, OBDM reads less like news and more like confirmation — evidence that the recommended pattern survives contact with a real, regulated enterprise rather than living only in textbooks.

The regulatory master: Johnson & Johnson

Johnson & Johnson's product master is built on the Pistoia IDMP Ontology plus the Accurids platform, aimed at the EMA's Product Management Service obligations (production master data) [6]. Two honesty notes. First, this is regulatory master data — the canonical, governed record of what a product is — not process execution; it is the formal substrate behind the ISO IDMP substance identity (our bp:IDMP-DS-001, a bp:SubstanceIdentifier) that the graph attaches to the drug-substance lot bp:DS-001 — a bp:DrugSubstance material individual — via bp:hasSubstanceIdentifier, not a model of the batch that produced it. In our own running example this is the difference between two nodes, not one: the material drug-substance lot bp:DS-001 (a BFO independent continuant — a thing that simply is, as opposed to an event that happens) and the information artifact that names it for the regulator, bp:IDMP-DS-001, which align.ttl maps up to iof:InformationContentEntity (a BFO generically dependent continuant carrying the FDA UNII / GSRS code). The two are tied by bp:hasSubstanceIdentifier, and the identifier points back at the substance with bp:isAbout, a sub-property of IAO's is about (obo:IAO_0000136). A regulatory master like J&J's lives entirely on the information-artifact side of that edge; it governs what the product is without ever modeling the occurrent — the batch run — that produced a given lot. Second, because the EMA deadline it serves falls in mid-2026, "production" here is partly built-for and rolling-out rather than fully shipped; read it as a system going live against a deadline, not one with years of operation behind it.

This book does not just survey that pattern — it tests it. Competency question CQ-16, exercised in the publication chapter, checks that the IDMP substance identity attaches to the same bp:DS-001 node the release gate validated rather than a duplicate, with an explicit guard against an owl:sameAs over-merge that would silently fuse the manufacturer's view and the regulator's into a copy-of-a-copy. The J&J master is the industrial instance of a guarantee the running example makes machine-checkable.

The two that touch the floor — barely

Two efforts come closest to manufacturing, and both are instructive precisely because of how they stop short. Sanofi's Modulus, within its Connected Smart Factories program, is a "contextualized data" effort that maps and links data across MES (the Manufacturing Execution System, the plant-floor software that runs and records batches), LIMS (the Laboratory Information Management System holding QC results), SAP (the enterprise resource-planning system of record), and paper records, with sites in France and Singapore rolling out around 2026 (piloted) [7]. Be exact about what it is: a data-mapping and contextualization program, not a formal RDF/OWL ontology, and the tidy "ISA-95-style" framing sometimes attached to it is an outside inference, not Sanofi's own claim.

The Pistoia Methods Hub pilot, run with Merck (MSD) and GSK, validated machine-readable transfer of an HPLC-UV analytical method between sites using ADF/AFO plus a novel RDF graph model — 55 standardized parameters, with about 1.35% relative retention-time reproducibility across the transfer, peer-reviewed in 2025 (piloted) [8]. This is the single case that lands nearest the regulated QC bench, and it remains a pilot, not a plant-wide system. The identification of "Merck" here as MSD rather than the unrelated Merck KGaA rests partly on conference-program inference, so it is offered with that caveat attached.

To see why this case matters, picture the same machinery on our own release panel. A monoclonal-antibody lot is dispositioned on a fixed battery of methods — SEC for monomer and HMW aggregate, cation-exchange (CEX) for charge variants, an HCP ELISA, residual-DNA qPCR, bacterial-endotoxin (LAL), sterility, and a potency bioassay, all present as classes in the companion model (what each of these methods actually measures is Book 1's analytical chapter) — every one of which must be transferred, with documented inter-site equivalence, before a second site may release product. A machine-readable method (the 55 standardized HPLC-UV parameters here) lets that equivalence be checked by query rather than by re-typing a paper SOP, and lets an out-of-trend retention-time shift be flagged before it becomes an investigation. That is the narrow, high-value wedge by which semantics reaches the regulated bench — one method at a time, not the whole floor at once, and the book's per-CQ PASS/FAIL acceptance tests are its analogue of method-transfer acceptance criteria.

The pattern: everywhere but the floor

Lay the inventory out and one shape is unmistakable. The production graphs cluster in two places: R&D and FAIR-data cataloging — Roche, Novartis, AstraZeneca, Boehringer Ingelheim, Novo Nordisk — and regulatory master data, Johnson & Johnson. The two efforts that reach toward manufacturing — Sanofi's Modulus and the Methods Hub — are a contextualization program and a pilot, respectively, and both are still rolling out. The discipline has proven itself on discovery and on the regulatory record. It has not, in public, proven itself on the GMP line.

The careful claim is this: named, GxP-production ontology deployments on the manufacturing floor are not found in the surveyed public evidence (GxP being the umbrella of "Good Practice" regulations that govern regulated drug work — GMP for manufacturing among them — under which a release-relevant system must be validated). That is deliberately not the same as saying they do not exist. Manufacturing IT is routinely confidential and competitively sensitive; a plant running a quiet OWL-backed batch-genealogy graph would have every reason not to publish it. Absence from the public record is weak evidence of absence in fact — but it is the only evidence we honestly have, and the consistency of the gap is itself striking.

There is also a structural reason the gap is so clean, and it is not only secrecy. The moment a graph touches a release-relevant fact — a CPP (critical process parameter — a setting that must be controlled to keep the product in spec) trajectory, an in-process result, a disposition — it falls under GxP computerized-system validation (GAMP 5, 21 CFR Part 11, EU GMP Annex 11) and the ALCOA+ data-integrity expectations: every value attributable, legible, contemporaneous, original, accurate, plus complete, consistent, enduring and available, and carried with a complete audit trail [9]. A discovery or FAIR-catalog graph that helps a scientist find a dataset carries none of that weight; a reasoner that silently infers that bp:DP-004 is out of specification on HMW aggregate is, in regulatory terms, a system contributing to a batch decision — and must be validated, change-controlled, and auditable before it may. That asymmetry, as much as confidentiality, is why the same companies deploy graphs freely in research and stop at the validated line. The point cuts the other way too: the wire-to-the-graph loaders show the floor is graphable from OPC UA, the historian, and B2MML — the obstacle is regulatory, not technical — and the SHACL release gate is exactly the validated, auditable mechanism a regulator would expect to find holding such a graph to spec.

There is an ontological reason the boundary is hard, too, not just a commercial one. The graphs that have shipped — FAIR catalogs, discovery graphs, regulatory masters — are overwhelmingly graphs of continuants and information artifacts: datasets, substances, targets, product records, the things that simply are. The GMP floor is the opposite shape. Its core entities are occurrents — events that happen in time: the batch run, the chromatography step, the moment bp:DP-004 goes out of spec — and process-bound realizables: the disposition a column has to bind, the role a QP (Qualified Person — the named individual legally responsible for certifying each batch's release) signature plays at release. A DCAT catalog needs almost none of BFO's occurrent or realizable machinery; a batch-genealogy graph is built from little else. Crossing the validation boundary is therefore not a port of the same model to a new server — it is adding the half of BFO the deployed graphs never had to use.

Three zones, then: a crowded R&D / FAIR quadrant (Roche, Novartis, AstraZeneca, Boehringer Ingelheim, Novo Nordisk), a governed regulatory master-data quadrant of one (Johnson & Johnson), and a near-empty GMP-floor quadrant that two efforts only graze. The hero figure draws exactly this map; the rest of the chapter is the reading of it.

The unsolved part: bridging the R&D graph to the validated line

There is a second silence in the evidence, and it is worth naming as plainly as the first. The published case studies skew heavily toward Western big-pharma innovators, and two populations are conspicuously under-represented. Contract manufacturers (CDMOs) — exactly where multi-client data-interoperability pain is sharpest — barely appear. The sharpest of that pain is not the last cold-chain mile but the tech-transfer package itself: when a process moves from innovator to CDMO, the handoff is a process description, a control strategy with its CPPs and acceptance criteria, the in-process and release method transfers, and a comparability protocol — followed by a per-batch certificate of analysis and disposition for every lot the contract site makes. Each of those is a place where two companies' identifiers, units, and method definitions must reconcile, and each is today reconciled mostly by document and email. A federated graph that let bp:DS-001's full lineage and its acceptance criteria travel intact across that corporate boundary is exactly the unbuilt bridge — and exactly why CDMOs, not innovators, would gain the most from it. And Asian manufacturers are nearly absent from the public-evidence base: Samsung Biologics surfaces once in connection with digital twins, while Celltrion, Lotte Biologics, WuXi, and NMPA-region players are largely missing from the surveyed literature. This book invents no details about them; it only flags the hole. The unfinished work is therefore twofold: bridging the proven R&D knowledge graph across the GMP validation boundary onto the production floor, and federating graphs across the CDMO and partner network so that a batch's full digital thread survives the handoffs between companies. By digital thread we mean the unbroken, queryable chain of every fact about one product — discovery target, cell bank, process parameters, in-process results, release disposition — connected as one graph rather than scattered across systems; it is exactly the single connected graph the running example assembles for WCB-CHO-001 → DS-001 → DP-001, and what today's company-internal graphs each hold only a slice of.

Why it matters

These eight systems are the real-world instances of the digital thread this book has been building toward — living proof that the engine runs at petabyte scale inside the most demanding companies in the industry. They are also proof of the harder fact: the most regulated, highest-stakes mile of the journey — the GMP manufacturing floor where BATCH-2026-001 is physically made and bp:DP-004 fails its release spec on HMW aggregate (2.41% against a 2.0% limit) while its monomer stays in spec — is precisely where the graphs have not yet reached. That is the quiet irony of this survey: the very thing the public record is missing — an OWL-backed graph of a GMP batch, its released lots, its CQA (critical quality attribute) panel, and the SHACL gate that holds them to spec — is exactly what the preceding chapters built around BATCH-2026-001. The book is, in effect, a worked specimen of the deployment the industry has not yet published. The pattern is not a failure of the technology; it is a map of the frontier. Knowing exactly where the deployed art stops is what lets the next reader aim at the part that is still open.

Key terms

• GMP — Good Manufacturing Practice, the legally binding, inspected quality regime a facility must follow to make a marketed medicine.
• GxP — the umbrella of "Good x Practice" regulations (GMP for manufacturing, and the like) under which a release-relevant computerized system must be validated.
• Enterprise knowledge graph — a large, cross-domain graph that unifies a company's data under a shared semantic model, queryable as one connected whole.
• FAIR catalog — a knowledge graph whose primary job is to make datasets Findable, Accessible, Interoperable, and Reusable, typically via public metadata vocabularies such as DCAT and PROV-O.
• Contextualization program — a data effort that maps and links records across systems without committing to a formal, logic-bearing RDF/OWL model; not the same as an ontology.
• Regulatory master data — the canonical, governed record of what a product is (identifiers, ingredients, strengths), as opposed to a record of how a given batch was made.
• BIKG — AstraZeneca's Biological Insights Knowledge Graph, a discovery-side property graph powering computational target identification.
• OBDM — Novo Nordisk's ontology-based data management approach: a data model, controlled vocabularies, and an inferencing RDF/OWL knowledge graph reusing public ontologies.
• IDMP Ontology — the Pistoia Alliance's OWL rendering of the ISO IDMP family of identification standards for medicinal products and their substances (ISO 11238/11239/11240/11615/11616), used to build regulatory product masters; the companion's bp:SubstanceIdentifier keys on the substance standard, ISO 11238.
• Argument from absence — the careful claim that something is "not found in surveyed public evidence," which is weaker than, and must not be stated as, "does not exist."
• Graph features — quantities derived from a node's position in a graph (its degree, shortest paths, neighbourhood embedding) that a machine-learning model consumes as inputs; what makes a knowledge graph a substrate for ML rather than only a query target.
• Validation paradox — that an ontology-reasoned, SHACL-validated graph yields proven conclusions while a learned model yields estimated ones, so the graph is the trustworthy ground truth a model is checked against, not the other way round.
• Grouped / leave-one-batch-out cross-validation — an honest held-out evaluation for learning over graph instances, in which every row from one batch or lineage goes wholly to train or to test, never both, so the score is not inflated by correlated near-duplicate rows.
• Vendor lock-in — the cost of a model expressed in a proprietary platform's object types and query dialect rather than in open W3C surfaces (RDF, SPARQL, SHACL, OWL), which makes the data reachable but the model not freely portable.

Where this leads

Johnson & Johnson's product master pointed at a target this book has so far only glanced at: the regulator. The next chapter, Regulatory Semantics: IDMP, SPL, KASA, and the Structured Submission, follows the data the whole way to the agency — how the identifiers, structured product labels, and submission formats that health authorities mandate are themselves becoming semantic, and what it means when the record a company files is meant to be read by a machine as much as by a reviewer.