The Deal Crossed a Boundary. The Data Model Has to Cross It Too

by Sofia ReyesTools and Context Columnist, HMND

Novartis’s agreement with Abogen puts worldwide rights to the early stage mRNA encoded T cell engager ABO2203 across a company boundary, with $575 million upfront and as much as $7.2 billion in potential milestones. The science may travel quickly. Its evidence model will travel more slowly, carrying every ambiguity about subjects, samples, protocol versions, corrections, and ownership with it.

That is where the serious work begins.

A partnered asset arrives with a story attached. The story usually contains promising results, a few tables, a data room, and a confident explanation of how the pieces fit. The receiving organisation then discovers that the subject identifier in one system is a visit identifier in another, that a laboratory result was transformed before export, and that the protocol amendment exists as a PDF whose effective date depends on someone remembering an email.

Clinical development has a talent for making context look optional.

The handoff is a systems problem

Abogen’s ABO2203 is being developed as an mRNA encoded CD19 by CD3 T cell engager for autoimmune disease. Its early evidence will have to survive movement from a biotech’s systems into a multinational sponsor’s clinical, statistical, regulatory, safety, and manufacturing processes. Each system will ask a reasonable question in its own dialect.

The clinical trial management system asks whether the subject attended the visit. The electronic data capture system asks what the investigator entered. The laboratory system asks which specimen produced the result. The safety system asks whether an event was serious, expected, and reportable. The statistical environment asks which frozen dataset supported the analysis. A data lake asks everyone to stop arguing and put it all in one place.

The lake does not resolve the argument.

Source system lineage has to survive the transfer. A value needs a traceable route from the original record through extraction, transformation, mapping, review, and final analysis. “Imported from vendor” is not lineage. It is a shrug with a timestamp.

A receiving team should be able to answer which source produced a value, which transformation changed its representation, which person or service performed the change, and which version entered the analysis dataset. The answer needs to remain available after the person who built the transfer has moved to another programme.

That documentation belongs inside the platform. If the transfer cannot replay the evidence path, the handoff has produced a new dataset, not a trustworthy continuation of the old one.

Subject identity is where clean data goes to hide

Subject identity deserves more suspicion than it receives. A participant may have a screening number, a randomisation number, a treatment number, a specimen accession number, and a safety case identifier. Those keys can refer to the same human while obeying different rules about visibility, retention, and correction.

The mapping belongs behind a controlled boundary. It needs explicit stewardship, access rules, and an audit history that can distinguish a legitimate reconciliation from a convenient overwrite. A model that joins records on a similar string can produce a beautiful cohort with one extra person and one missing dose. The chart will not complain.

Protocol identity creates a second fault line. The protocol is a versioned object with amendments, inclusion criteria, visit windows, endpoint definitions, treatment assignments, and statistical consequences. A subject’s data acquires meaning under the protocol version active at the relevant time. A visit outside the original window may be a deviation, an amended allowance, or a data entry error. The row itself cannot decide.

Clinical data models such as CDISC standards can give the handoff a shared shape, including structures for tabulation and analysis. They do not provide the missing facts. A standardised field can still hold a value with an unclear source, a stale definition, or the wrong subject.

The schema provides a map, while the evidence provides the territory and the witness.

Attribution is a decision, not a column

Partnered development also raises a question that dashboards tend to bury: who is allowed to say what a data point means?

The originating laboratory may own the assay record. The investigator may own the clinical observation. The sponsor may own the analysis dataset. A vendor may operate the system that stores the data. Those roles overlap without becoming interchangeable.

Attribution therefore needs a governed record of responsibility. A corrected haemoglobin value should retain the original value, reason for change, authorisation, time, and downstream impact. A derived endpoint should identify its input records and calculation version. A query resolution should state who answered it and under which protocol interpretation.

Access controls should follow those responsibilities. A scientist may need row level access to deidentified subjects. A safety physician may need identifiable case context. A statistical programmer may need frozen analysis data while lacking permission to alter source records. A data steward may resolve a mapping issue without changing the clinical observation itself.

Viewing a record does not grant permission to interpret it, and permission to correct it does not grant permission to approve the change.

The distinction matters because a clinical record is not merely a row in a warehouse. It may become a safety signal, an endpoint, a submission argument, or a manufacturing decision. Every downstream use increases the cost of a casual upstream edit.

Automation arrives after the boring machinery

Teams will eventually ask what an analytics or model layer can do. It can flag missing visits, detect unusual laboratory patterns, suggest duplicate subjects, propose mappings between coding systems, and identify a result whose provenance chain breaks halfway through the pipeline.

Useful work, provided the system has somewhere trustworthy to write its findings.

The model should receive governed data products with defined schemas, quality states, provenance links, and permitted actions. Its output should be an observation or recommendation with confidence, evidence references, model version, and a human decision record. It can open a query. It should not silently rewrite a source record. It can identify a likely duplicate. It should not merge subjects because two identifiers look sociable.

Validation must match the decision. A model that ranks records for review needs evidence about ranking performance and review burden. A model that influences eligibility, safety assessment, or an endpoint needs a much higher standard, with controlled changes, reproducible inputs, and documented release decisions. The word “assistive” does not dissolve accountability.

How does it compare with a governed clinical data platform? An autonomous correction agent is a shortcut through the part of the building where the fire exits are.

A dashboard can lie politely

Picture a study team meeting at 8:30 in the morning. The dashboard shows clean enrolment, complete visits, and a tidy response curve. Someone asks why the number of evaluable subjects differs from the statistical report.

The dashboard drew its population from a treatment table loaded after a protocol amendment. The report used a frozen analysis set linked to randomisation records before the amendment. Both displays are internally consistent. Neither display explains the disagreement.

Clean presentation can conceal broken provenance because aggregation removes the scars. A missing source timestamp becomes a completed date. A changed unit becomes a normalised value. A manually repaired identifier becomes a successful join. The screen rewards coherence, even when coherence came from an undocumented decision.

Validation needs to inspect the path, not merely the output. Reconciliation should compare source counts, subject mappings, visit states, protocol versions, transformation logs, and exception queues. Quality checks should fail loudly when a source version disappears or when a derivation uses records from incompatible freezes.

The decision to correct data also needs a named authority. The decision to freeze data needs one too. A sponsor can define the analysis cut, while the originating party may remain responsible for source corrections. Both parties need an impact process for changes discovered after a freeze, including whether the change affects safety reporting, the analysis, or a regulatory submission.

That governance is slower than refreshing a dashboard. It is also cheaper than explaining an unexplained result after the evidence has acquired legal and scientific consequences.

The asset moved first

The Novartis and Abogen agreement shows how quickly a scientific asset can cross organisational boundaries while its evidence remains early and conditional. The deal includes options on additional RNA programmes, which makes the handoff problem larger than one compound. Every future programme will bring its own source systems, coding choices, protocol history, and local definitions of “complete.”

The temptation is to build a universal layer that makes all of this look uniform. Uniformity is a poor substitute for meaning. A shared model can expose differences, preserve them, and make their consequences visible. It cannot erase the need to decide which source has authority.

Senior engineering teams know this from other domains. A platform is only as dependable as its boundaries, contracts, observability, and change control. Clinical data adds a human subject, a regulated record, and a scientific claim to every boundary. The data product therefore needs more than an API and a polished catalogue.

It needs an identity service that respects privacy. It needs protocol and terminology registries. It needs immutable source records, versioned transformations, role based access, audit trails, validation evidence, reconciliation workflows, and explicit freeze states. It needs a user experience that shows uncertainty and provenance beside the result instead of hiding both behind a green status badge.

Only then should a model touch the evidence layer. It may help triage, reconcile, classify, and explain. Its permitted actions should be narrower than its apparent intelligence, because the cost of a plausible mistake rises with every downstream system that accepts it.

What happens when the context is wrong? The system can still produce a confident answer, a complete graph, and a dashboard that passes a casual inspection. The failure appears later, when someone asks which subject, under which protocol version, from which source, with whose correction, supported which decision.

That question is the product.

If this handoff problem is sitting on your desk, hello@example.com is a sensible place to start the conversation.