Batch Genealogy for Biologics: Tracing a Lot Back Through Bulk and Cell Bank

Why Biologic Genealogy Is a Graph, Not a Chain

Ask a quality leader to describe batch genealogy and you will usually get a picture of a tree. Cell bank at the top, finished lots at the bottom, branches in between. That picture is closer than a straight line, but it is still wrong in a way that matters, because a tree only branches downward. Every node in a tree has exactly one parent.

Biologic manufacturing does not work that way. Material converges as often as it diverges. Fractions from a chromatography column are pooled. Several drug substance lots are combined into one formulated bulk. Buffers made from multiple raw material lots feed a single step. A finished lot can draw from more than one upstream lot, and a single bulk lot can become many finished lots. The correct picture is a directed acyclic graph with fan-out and fan-in at multiple levels, and the difference is not academic. It determines whether your system can answer a recall question in an afternoon or in three weeks.

What “one lot” means at each stage

Part of the confusion comes from the word itself. Under US drug CGMP definitions, a batch is a specific quantity of a drug or other material intended to have uniform character and quality within specified limits and produced according to a single manufacturing order. A lot is a batch or a specific identified portion of a batch, and a lot number is any distinctive combination of letters, numbers, or symbols from which the complete history of the manufacture, processing, packing, holding, and distribution of a batch or lot can be determined.¹

Read that last clause carefully. The regulation does not require a chain. It requires that the complete history be determinable from the identifier. That is a graph traversal requirement written in 1970s language, and it is the standard your system is actually held to.

In practice, the objects that carry a lot identity in a biologics process are not the same kind of thing at each stage. A working cell bank is an inventory of vials, and the meaningful unit is often the individual vial rather than the bank. A drug substance lot is commonly defined by a single production bioreactor cycle carried through recovery and purification. A formulated bulk lot is a pooling decision. A filled lot is a filling run, which may be split into sublots by filling session, by container size, or by market presentation. Each of those definitions is legitimate. The problem is that they are different shapes, and a single “parent lot” field cannot hold them.

The failure mode: a single parent field

The most common genealogy defect we see is structural rather than procedural. The system has a field on the lot record called something like source lot or parent batch. It holds one value. When a step pools three inputs, someone enters the largest contributor, or the first one, or the one that was in the header of the paper record. The other two are still written down somewhere, usually in the batch record body, but they are no longer in the field the query runs against.

Nothing fails. The record looks complete. Release proceeds. Then two years later an adventitious agent finding on a media lot triggers a forward search, and the search returns a partial list that looks authoritative. The people running the investigation do not know it is partial. That is the specific way this goes wrong, and it goes wrong quietly.

Where Pooling and Splitting Actually Happen

Naming the specific points where the graph branches is more useful than describing the process again. Here is where it happens in a typical mammalian cell culture process, and what each point does to the genealogy record.

Seed expansion: split, then converge

A single vial from the working cell bank is thawed and expanded through shake flasks, wave bags, and one or more seed bioreactors before inoculating the production vessel. The expansion may run in parallel vessels that are later combined. If your system records only “WCB vial X to production bioreactor run Y,” you have lost the intermediate vessels, and with them the ability to answer a question about a contamination event in one seed vessel that fed two production runs.

The cell bank tier itself is a branch point that spans years. Under ICH Q5D, the industry standard is a two-tiered system in which a master cell bank gives rise to one or more working cell banks, each a homogeneous pool distributed into many containers, with documented history and passage limits.²³ A single MCB can therefore sit above a decade of commercial production across multiple WCBs, multiple sites, and multiple product presentations. If a characterization finding on the MCB ever emerges, the forward search from that one node is the largest traversal your system will ever be asked to do.

Harvest and capture: the first real pooling

Clarified harvest is loaded onto the capture column over multiple cycles. Elution fractions are pooled according to a defined collection criterion. Two harvests may be combined before capture, or one harvest may be split across two columns to fit within campaign scheduling. Each of those decisions is made on the floor, recorded in the batch record, and frequently never lands in a structured field.

Drug substance to formulated bulk: pooling by design

This is the point most often modeled incorrectly. Formulation frequently combines more than one drug substance lot to reach a target batch size, to blend within a potency range, or to use up remaining material before an expiry. The formulated bulk is then a genuine many-to-one node. Every finished lot downstream of it inherits exposure to every DS lot in the pool, in proportion to the contribution.

That proportion matters. If DS lot A contributed 80 percent of the pool and DS lot B contributed 20 percent, a specification problem traced to lot B has a different exposure profile than one traced to lot A. A system that stores only the fact of the link, without the quantity, can tell you that a relationship exists but cannot help you rank what to quarantine first.

Fill and finish: splitting into many

One formulated bulk becomes many filled units, and those units are commonly divided into sublots by filling session, by line, by stopper or vial lot, or by the sterilization load they went through. Those sublots are then combined again into market-specific finished lots with different labeling, different secondary packaging, and different serialization. A single finished pack in a hospital pharmacy can therefore be four or five graph hops from the drug substance, with a fan-in at each hop.

Rework, reprocessing, and retains

Three smaller branch points get forgotten regularly. Reprocessing (running material through a purification step a second time) gives a lot a second parentage that includes its own earlier state. Rework of a finished lot into new packaging creates a node whose parent is another finished lot. And retain samples, reference standards, and stability pull samples are all objects drawn from a lot that later become the physical evidence in an investigation. If your genealogy does not include them, the investigation team has to reconstruct which retain corresponds to which sublot by hand.

Forward and Backward: The Two Questions Under Time Pressure

All of this becomes concrete in one scenario. A problem is found. It might be an out-of-specification stability result, a field complaint, a sterility failure, a supplier notification about a raw material, or a positive adventitious agent test. Whatever the trigger, the same two questions get asked immediately, and they are asked by people who need an answer today.

The backward question

Backward traversal starts from a failing lot and enumerates every input. Every upstream lot at every level, transitively, plus every raw material, component, and consumable that touched each step, plus the equipment used, plus the analytical results that supported each release decision along the way.

Most organizations do this reasonably well, for a simple reason: the batch record is organized backward. It is a document about how this lot was made. Someone can read it and build the list. The weakness is not completeness, it is speed and transitivity. Getting one level back is easy. Getting four levels back, through a formulated bulk that pooled three DS lots, each of which used a different media lot and a different resin cycle, means opening a dozen documents and copying numbers into a spreadsheet.

The forward question

Forward traversal starts from an implicated input and enumerates everything derived from it. Every DS lot made with that media lot. Every finished lot containing any of those DS lots. Every market and every customer those finished lots went to.

This is the question that decides how much product you have to quarantine, and it is the question almost nobody can answer quickly. Nothing in a batch record system is indexed this way. The batch record for lot A does not know that lot B exists. Answering forward means scanning every batch record made in a window, which is why the honest internal answer is often “give us two weeks.”

Two weeks is not an answer. It is a decision to quarantine everything, because you cannot defend a narrower boundary. The 2009 vesivirus event at a single enzyme therapy facility is the reference case for what happens when upstream biology fails: one contaminated bioreactor among six led to a shortage that ran for a median of 1.3 years for one patient cohort, with involuntary dose reductions and product switches.⁴ The scale of an event like that is set by biology and capacity, not by data. But the scope of what has to be held is set by data, and that part is under your control.

The first day of an investigation

What the Data Model Has to Store

If both traversals have to be fast, the data model follows directly. It is a small number of ideas, none of them exotic, and most organizations already have the raw information. What is usually missing is the structure that makes it queryable.

Store edges, not paths

The single most important design decision is to store each consumption event as its own record linking exactly two nodes, rather than storing a lot’s ancestry as a summarized path or a denormalized string. Edges compose. Paths do not. If every edge is stored separately, both directions are the same query run with the direction reversed, and adding a new stage to the process does not require anyone to redesign anything.

An edge should carry, at minimum: the consuming node, the consumed node, the process step at which consumption occurred, the quantity or proportion, the timestamp, the person or system that recorded it, and a pointer to the executed record that is the source of truth. That last field is what makes the graph auditable rather than merely convenient.

Quantities turn a list into a ranking

A genealogy that returns 140 affected finished lots is technically correct and operationally useless. A genealogy that returns 140 lots ranked by the proportion of implicated material each contains lets quality set a defensible boundary: full recall above a threshold, enhanced testing in a middle band, monitoring below. That distinction is worth more than any other single improvement, and it depends entirely on storing quantity on the edge.

Edges that are not material transfers

The category people miss is the shared resource. A chromatography column packed with a particular resin lot and used for cycles across two campaigns links those campaigns even though no material was transferred from one product lot to another. The same is true of a shared hold vessel, a shared buffer preparation, a shared filter housing, and a single autoclave load.

These belong in the graph as a distinct edge type. They are not bill-of-materials relationships and they should not be modeled as such, but a contamination or carryover investigation traverses them exactly the way it traverses a material edge. Systems that only understand material consumption will not find them, and the investigator will have to reconstruct them from equipment logbooks.

As-of queries and immutability

Genealogy data changes. A transcription correction lands. A deviation reclassifies a sublot. A CDMO sends a revised record six weeks after the fact. If your system overwrites, then the answer you gave last month is unreproducible today, and you cannot demonstrate what you knew when you made a quarantine decision.

The fix is to make edge records append-only with effective dating, so the graph can be queried as of any past date. This is the same principle behind audit trail expectations under data integrity guidance, applied to the relationship rather than to the result.⁸⁹

Raw Materials, Resins, and Single-Use Components

Everything above concerns product material. The other half of the graph is everything that touched the product without becoming it, and this is where genealogy models are thinnest.

Attach material lots to steps, not to batches

US CGMP requires records showing the identity and quantity of each component and container used, and requires batch production records to identify each batch of component or in-process material used.¹⁰¹¹ Most systems satisfy this by listing components against the batch. That is compliant and it is also insufficient for fast investigation, because it loses the step.

A biologics batch may consume twenty or more distinct material lots: basal media, feeds, supplements, buffer salts, acids and bases, WFI, protein A ligand, filters, tubing sets, bags, vials, stoppers. Knowing that media lot M was used somewhere in batch B is much less useful than knowing it was used in the N-1 seed vessel on a specific date. When the supplier notification arrives saying that media lot M has a problem introduced at a specific manufacturing step, the step-level linkage lets you decide whether the exposure is even relevant.

Buffers deserve special attention because they are themselves compound objects. A buffer prep is made from several raw material lots and WFI, gets its own preparation record and its own identifier, and is then consumed by one or more process steps, often across multiple batches. That is a node in the graph, not an attribute of a batch. Organizations that treat buffer preps as line items lose an entire layer of the ancestry.

The resin problem

Chromatography resin is the clearest case of a link that standard genealogy models miss completely. Resin is expensive and durable. It is packed into a column, used for a validated number of cycles, cleaned and regenerated between cycles and between campaigns, and eventually replaced. During its life it may process many lots of the same product, and increasingly it processes lots of different products.

Published work on multiproduct resin reuse describes exactly this: resins dedicated to a single product are often used for only a fraction of their potential lifetime, and extending them to multiple products reduces resin waste and cost while improving flexibility when raw materials are scarce.¹² Doing it responsibly requires a documented strategy covering risk assessment, small-scale feasibility, cleaning with sodium hydroxide to inactivate and degrade carryover protein, and blank buffer elution testing to demonstrate that carryover falls to an acceptable level.¹²¹³ Acceptance limits for product carryover are more stringent for multiproduct reuse than for single-product reuse, and the between-campaign cleaning has to consistently meet them.¹³

All of that is process validation. The genealogy consequence is separate and rarely handled: the resin creates a directed link from every lot processed on that column to every lot processed after it. If your model only records material consumption, that link does not exist in your data. The identity you need is not “resin lot R” alone. It is the combination of resin lot, column packing event, and cycle number, because the packing event is when the resin becomes a distinct physical object with its own use history and its own cleaning record.

Single-use components and the supplier chain behind them

Single-use bags, tubing sets, filters, and connectors carry the same problem in a different shape. Each has a supplier lot. Each attaches to a specific step. And each sits on top of a supply chain the manufacturer does not directly control, because a single-use assembly is built from films, connectors, and filters made by other companies.

Industry work on change notification practices for single-use systems addresses this directly. A change that originates further back in the supply chain has to keep the alphanumeric identifier of the originator of the change, so the notification remains traceable to its source rather than arriving as an anonymous statement from the assembler.¹⁴ That is the same identifier discipline problem discussed later in this article, appearing on the supplier side.

The practical implication is that your genealogy should be able to answer a supplier’s question, not just your own. When a film supplier notifies a change affecting a resin grade used in a bag family over a date range, you need to move from that date range to the assemblies you received, to the steps they were used on, to the product lots those steps produced. That is three joins across three systems, and it is worth testing before you need it.

Stitching the Record Across MES, LIMS, ERP, and Paper

None of this lives in one place. Being honest about that is the difference between a genealogy program that works and one that stalls in a system selection exercise.

The seams are where the graph breaks

Genealogy rarely fails in the middle of a system. It fails at the joins. Six seams account for most of it:

• Cell bank vial to production. The vial inventory is often a validated spreadsheet or a small standalone application. The link from a specific vial to a specific production run is frequently recorded only on paper.
• Buffer and solution prep to use. Prep records sit in a separate logbook or a separate module, and the consumption link is recorded by transcription.
• Equipment and column history to batch. Cycle counters, packing records, and cleaning records live with engineering rather than with production.
• Drug substance to drug product. If these are made at different sites, the handoff is a shipment, and the receiving site often assigns its own identifier.
• Internal to CDMO. The most common seam, and the hardest, discussed below.
• Finished lot to distribution. Serialization and warehouse systems that were designed for supply chain security rather than for quality investigation.

Paper at the interface is a data problem, not a compliance problem

A paper record at a seam is fully compliant. It satisfies the batch record requirements. What it does not do is participate in a query. The distinction to hold onto is that this is not a data integrity failure, it is a data availability failure, and the two need different remedies. Data integrity guidance from both FDA and MHRA is concerned with whether records are attributable, legible, contemporaneous, original, and accurate, and with whether audit trails are reviewed.⁸⁹ A perfectly ALCOA-compliant paper logbook still cannot be traversed in ten minutes.

The practical answer is usually not to replace the system that holds the paper record. It is to extract the linkage into a structured index while leaving the executed record where it is. The index holds identifiers and edges. The executed record remains the source of truth, and every edge points back to it. That is a much smaller project than an MES expansion, and it delivers most of the investigation speed.

Identifier Discipline and the Shared Lot Numbering Convention

Here is the part that surprises people: the numbering convention matters more than the systems. You can stitch a graph across four mediocre systems if the identifiers are clean. You cannot stitch one across four excellent systems if a lot is called three different things at three boundaries.

Six rules that do most of the work

The CDMO boundary

Most biologics involve at least one external party, and often several: a CDMO for drug substance, another for fill and finish, a testing laboratory, a packaging site. Each of those parties runs its own systems and its own numbering. This is the seam where genealogy most often breaks, and it breaks in a specific way. The quality agreement covers what the parties will do and what records they will provide, and the records arrive as PDFs of executed batch records. Those PDFs are compliant. They are also not data.

US regulations are unusually direct about the obligation here. Where two or more establishments participate in manufacturing a biological product, the records of each must plainly show the degree of its responsibility, and each participating manufacturer must furnish to the manufacturer who prepares the product in final form a copy of all records relating to its operations insofar as they concern the safety, purity, and potency of the lots involved.¹⁵ FDA’s guidance on contract manufacturing arrangements makes the same point from the quality systems side: the owner remains responsible for ensuring the drug is made in compliance with CGMP, and the quality agreement should define responsibilities clearly.¹⁶

What that means in practice is that the genealogy data transfer belongs in the quality agreement as a named deliverable, with the same specificity as any other. Not “the contract manufacturer will provide batch documentation” but a defined set of fields, a defined format, a defined identifier convention, a defined transfer frequency, and a defined obligation to notify when a record is corrected after the fact.

That last point is worth dwelling on. Biologics records under US rules must be retained for no less than five years after the manufacturing records are complete, or six months after the latest expiration date for the product, whichever is later.¹⁵ For advanced therapies in the EU, traceability data must be kept for a minimum of 30 years after the expiry date.⁵⁶ Contracts with contract manufacturers do not typically run that long. If the retention and access obligations are not written to survive termination, you can lose the ability to answer a question you are still legally required to answer.

Regulatory Expectations, and a Drill to Test Yourself

The regulatory basis for all of this is older and plainer than most people expect. It does not use the word genealogy, and it does not prescribe a system. It prescribes an outcome.

What the requirements actually say

For biological products, records must be made concurrently with performance of each step in manufacture and distribution, in such a manner that at any time the successive steps in the manufacture and distribution of any lot may be traced by an inspector.¹⁵ That phrase, “at any time,” is doing real work. It is not satisfied by an assurance that the information could be assembled.

Batch production and control records must include complete information relating to the production and control of each batch, including specific identification of each batch of component or in-process material used.¹¹ Component and container records must show identity and quantity, the supplier’s lot number, and the number assigned by the receiving firm.¹⁰ Distribution records must contain the name and strength of the product, the lot or control number, the name and address of the consignee, and the date and quantity shipped.⁷ The label of a biological product must bear the lot number.¹⁷

In the EU, Annex 2 applies specific expectations to biological active substances and medicinal products, including the seed lot and cell bank system and the maintenance of traceability from starting material through to the finished product batch.¹⁸ WHO GMP for biological products sets parallel expectations.¹⁹ ICH Q5D governs the derivation and characterization of the cell substrate itself, including documented origin and passage history for the two-tiered cell bank system.²³

Where cell and gene therapy raises the bar

For advanced therapies, the expectations go further and get more specific about direction. EU rules require a traceability system from the donor through to the recipient and back, with information archived for a minimum of 30 years after the expiry date.⁵⁶ That bidirectional phrasing is exactly the two-traversal problem described earlier, written into regulation.

Chain of identity and chain of custody are the terms of art in that space, and they are distinct: chain of identity establishes that a given material belongs to a given patient, while chain of custody records every transfer and handling event across the material’s life. We have written about how those work for autologous cell therapies elsewhere and will not repeat it here. The point worth carrying back to conventional biologics is that the discipline autologous therapy developers are forced into by regulation is the same discipline that makes a monoclonal antibody investigation fast. They just do not get to skip it.

Where this is heading

Two developments are pushing conventional biologics toward the same rigor. The first is process intensification. Perfusion culture and connected downstream processing blur the discrete batch boundary that biologics genealogy has always relied on. ICH Q13 addresses continuous manufacturing of drug substances and drug products and treats batch definition, material traceability, and diversion as core elements of the control strategy.²⁰ Work on residence time distribution shows how material tracking becomes a modeling exercise rather than a records exercise when flow is continuous.²¹ Biologics processes are moving in that direction, and organizations that cannot answer genealogy questions on discrete batches will not manage it on continuous ones.

The second is external. Traceability expectations across biological medicinal products have been sharpening for years, driven by patient safety and by supply chain concerns, and the literature on traceability for biological products has been explicit that current practice varies widely.²² The direction of travel is toward more granularity, not less.

A drill you can run next month

The fastest way to find out where you stand is to stop assessing and start testing. Pick a date. Have quality select a finished lot at random. Give the team four hours and a stopwatch. Score it like this.

Run it once and the gaps rank themselves. In our experience the failures cluster predictably: forward traversal, the cell bank seam, equipment and resin history, and anything on the far side of a contract manufacturing boundary. Those four account for most of the delay in most investigations, and none of them require a new system to fix. They require identifiers, edges, and a place to put them.

Conclusion

Batch genealogy for a biologic is not a longer version of the small molecule problem. It is a different shape. Pooling and splitting make it a many-to-many graph, and a graph does not fit in a parent-child field no matter how carefully people fill that field in. The consequence only appears under pressure, when a problem is found and the organization has hours rather than weeks to determine what else is affected in both directions. Backward traversal is usually adequate because batch records are organized that way. Forward traversal is usually not, because nothing was built to run in that direction, and the fallback is to quarantine broadly because a narrower boundary cannot be defended.

The fixes are less dramatic than they sound. Store consumption events as edges with quantities and record references rather than storing paths. Attach material lots to process steps rather than to whole batches. Treat buffer preps, cell bank vials, column packing events, and filling sublots as real objects with real identifiers. Model the shared-resource links, particularly reused chromatography resin, which creates a connection between product lots that no bill of materials will ever show you. Then get the identifiers right at every boundary, and write the genealogy data transfer into the quality agreement with the same specificity you would apply to a specification, because the partner boundary is where the graph most reliably breaks. A shared numbering convention that everyone honors is worth more than any single system on the list.

Sakara Digital works with pharma and biotech organizations building traceability that holds up during an investigation rather than only during an audit of the documentation. If you are trying to work out whether your genealogy would survive the four-hour drill, or where to place a structured index across MES, LIMS, ERP, and the seams in between, we are happy to have that conversation.

For Further Reading