I have worked on a tablet small enough to hold between two fingers, and I have worked on a cell therapy batch that filled a single bag with exactly one patient’s name on the label. Both times the badge said something like validation, or process development, or manufacturing sciences and technology. Almost everything else was different. The math was different, the failure modes were different, the distance between “we made it” and “someone received it” was different by orders of magnitude. What stayed constant across six modalities and about two decades was not a technique. It was a way of asking questions before trusting an answer.
People early in this field often ask whether they should specialize immediately, small molecule or biologics or cell and gene therapy, and commit. I understand the instinct. The tools genuinely do not transfer directly; a resin lifetime study and a viral vector potency assay have almost nothing in common on paper. But the instinct underneath both of them is the same instinct, and once you have it, moving between modalities is a lot less frightening than it looks from the outside.
Six modalities, one instinct
Small molecule process validation, at Novartis, taught me to think in populations of batches, not single events. Medical devices, at Medtronic, taught me that validation takes a completely different shape when what you are proving is mechanical, and a person’s hands are part of the system. RNAi lipid nanoparticles, at Alnylam, in the years around ONPATTRO’s 2018 approval and GIVLAARI’s the following year, taught me what happens when the delivery vehicle is as novel as the drug substance. Oligonucleotide validation, at Wave, taught me that “it is just chemistry” is exactly the sentence to distrust. Autologous cell therapy process development, at SQZ, taught me what it means when the batch and the patient are the same event. Commercial-stage gene therapy manufacturing sciences and technology, at Sarepta, taught me to sit with an assay that may never fully resolve, and ship product anyway, honestly. RNA editing, where I work now, is teaching me that the questions repeat even when the biology is new again.
What never changes
Underneath all six, four things never changed. I have come to think of them as the actual job, with the modality as the costume.
Risk thinking, not the form but the habit. Before you build a control, ask what happens if this specific thing fails, who it fails on, and whether you would find out in time to matter. A blocked pore in a sterilizing filter and an underdosed viral vector are unrelated failures, but the question that finds both of them is the same question, asked early and asked honestly.
Data integrity. Every modality I have touched eventually comes down to whether you can trust the number in front of you: who generated it, under what conditions, and whether the record lets you reconstruct that chain later. A potency result and a tablet dissolution result are trusted or distrusted by the same logic, contemporaneous, attributable, original, accurate, no matter how exotic the assay producing them looks.
Process understanding before you lock the process. This is the step schedule pressure skips first, and the one that costs the most later. A process you understand can be defended when it drifts, because you know which parameters actually matter. A process you locked before you understood it is a process you will be afraid to touch for years, because you genuinely do not know what happens if you do.
Statistics as honesty, not decoration. A sample size, a confidence interval, an acceptance criterion tied to real process capability instead of picked because the last few engineering runs happened to clear it, none of that is paperwork. It is the difference between a claim that survives a second question and one that only survives if nobody asks it. That distinction does not change across modalities. Only the units do.
A worked version of that distinction, using illustrative figures: a process characterization data set from twenty engineering runs shows a critical quality attribute with a mean of 98.2 percent and a standard deviation of 0.6 percent, comfortably inside a proposed 95 to 102 percent acceptance range. An acceptance criterion picked because “the last five runs cleared 97 to 99 percent” describes those five runs, not the process. An acceptance criterion derived from the full twenty-run capability estimate, with a stated confidence level and a stated sample size behind it, describes the process and survives a second reviewer rebuilding it from the same raw data. The arithmetic is not the hard part. The discipline of deriving the number from the whole data set instead of the convenient run is.
Before you trust a control, ask this
The four things that never change compress into a sequence, not just a list, when a control from a prior modality shows up asking to be trusted again.
Who owns validation, by modality
The job title is usually the same across all six, validation, process development, or manufacturing sciences and technology, but what that role actually does, and who else has to sign off before a lot releases, shifts more than people expect moving in.
| Modality | Who typically drives validation | Who else signs off before release | What is different about the sign-off |
|---|---|---|---|
| Small molecule tablets | Validation / manufacturing sciences | Quality Assurance | Sign-off leans on trended data across a large batch population |
| Medical devices | Process validation, alongside design assurance | Quality Assurance, and design engineering for anything touching design intent | Separate design verification and design validation sign-offs already happened before process validation starts |
| RNAi lipid nanoparticles | Process development / drug product | Quality Assurance, plus a drug product SME for particle attributes | The delivery particle, not just the drug substance, needs its own sign-off authority |
| Oligonucleotides | Process development / analytical | Quality Assurance, plus an analytical SME who can speak to the full species ladder | Release sign-off requires someone who can defend individual near-identical species, not an aggregate purity number |
| Autologous cell therapy | Manufacturing sciences and technology, in real time with manufacturing | Quality Assurance, often reviewing in real time rather than after the fact | There is no batch to hold while a slow investigation runs; disposition timelines compress hard |
| Commercial gene therapy | Manufacturing sciences and technology | Quality Assurance, plus whoever owns the potency triangulation strategy | Release rests on a panel of complementary assays, not a single validated method |
The pattern underneath the differences: population-scale modalities let sign-off wait for fully trended data, and patient-scale modalities compress that timeline until sign-off is closer to a live, in-the-moment decision than a paperwork review after the fact.
Six modalities at a glance
The narrative below goes deep on each modality in order. This table is the fast version, for anyone who wants the shape of the argument before the stories behind it.
| Modality | Where I learned it | The shift it forced | Signature failure mode | What carried forward |
|---|---|---|---|---|
| Small molecule tablets | Novartis | Think in populations of batches, not single events | A process locked before it was understood, defended only by convention | Continued process verification as a habit, not a phase gate |
| Medical devices | Medtronic | Validation splits into three separate questions: design verification, design validation, and process validation | A part that is perfectly reproducible to print and still capable of a use error | Never trusting reproducibility alone as proof of safety |
| RNAi lipid nanoparticles | Alnylam | The delivery vehicle is validated as its own critical system, not a packaging afterthought | Cold chain treated as logistics instead of process | Reading distribution qualification data the way I read a batch record |
| Oligonucleotides | Wave | ”It is just chemistry” undercounts the failure modes a stepwise synthesis actually produces | An aggregate purity number standing in for a ladder of near-identical species | Naming every closely related species instead of averaging past it |
| Autologous cell therapy | SQZ | The batch is a patient; population-scale statistics stop applying cleanly | A chain-of-identity gap invisible until the wrong result reaches the wrong record | Earning confidence before the first real batch, not across a thousand you will never run |
| Commercial gene therapy | Sarepta | Ship product on an assay you know is imperfect, and say so | One clean-looking number quietly retired because it disagreed with the others | Triangulating potency across complementary measures instead of waiting for a perfect one |
The regulatory framework underneath each modality
The four questions that never change sit on top of six genuinely different legal frameworks, and knowing which framework actually governs a given modality prevents a specific, recurring mistake: assuming a rule that applies to one modality applies to all of them because the underlying science feels similar.
| Modality | US marketing pathway | Primary US regulation | International counterpart |
|---|---|---|---|
| Small molecule tablets | New Drug Application or Abbreviated New Drug Application, reviewed by CDER | 21 CFR 210 and 211, current good manufacturing practice for finished pharmaceuticals | EU GMP Volume 4 |
| Medical devices | 510(k), De Novo, or Premarket Approval depending on device class, reviewed by CDRH | The Quality Management System Regulation, effective February 2, 2026, which replaced most of 21 CFR 820 by incorporating ISO 13485:2016 directly into US law | EU Medical Device Regulation 2017/745, and ISO 13485 internationally |
| RNAi lipid nanoparticles and oligonucleotides | New Drug Application, reviewed by CDER, the same pathway as small molecule tablets | 21 CFR 210 and 211 | EU Marketing Authorisation Application under Directive 2001/83/EC |
| Autologous cell therapy and commercial gene therapy | Biologics License Application under Section 351(a) of the Public Health Service Act, reviewed by CBER | 21 CFR 600 series, plus 21 CFR Part 1271 Subpart C for donor eligibility and identification of human cells and tissues | EU Marketing Authorisation for an Advanced Therapy Medicinal Product under Regulation (EC) No 1394/2007 |
| RNA editing | Investigational New Drug application during development; the eventual marketing pathway depends on whether the final product is classified as a chemically synthesized drug substance or a biological product | 21 CFR 312, the same IND framework every modality uses during clinical development | Varies by eventual product classification |
The row most people get wrong on first exposure is RNAi and oligonucleotides. A chemically synthesized nucleic acid drug, however biologic-like its complexity or its lipid nanoparticle delivery vehicle looks, is regulated as a drug, not a biologic, because it is manufactured by chemical synthesis rather than derived from a living system. ONPATTRO (patisiran) and GIVLAARI (givosiran) were both approved as New Drug Applications, NDA 210922 and NDA 212194 respectively, not as Biologics License Applications, even though both moved through Alnylam’s organization alongside biologics-trained colleagues who, understandably, sometimes assumed otherwise on first contact with the program. The device row carries its own version of the same lesson from the other direction: as of February 2026, the Quality Management System Regulation folds ISO 13485:2016 directly into 21 CFR 820 by reference, closing a gap that used to leave a US-only regulation and an internationally recognized standard pointing at the same expectations without being, formally, the same document.
Where the ground shifts first
Small molecule validation at Novartis is where I learned continued process verification, the third stage of FDA’s process validation lifecycle (the day-to-day version of that stage is walked in full in continued process verification at Stage 3), built on the idea that a handful of qualification batches proves far less than years of routine production trended honestly. You learn to see a process the way an actuary sees a population, not the way a chemist sees one reaction. That habit turned out to be the most portable thing I own.
Medical devices at Medtronic taught me something small molecule work never had to: that “validated” splits into two different questions when the product is a mechanical, often electromechanical, system a person operates, rather than a chemical reaction settling out in a vessel. Design controls under the device quality system regulation, 21 CFR 820.30, separate design verification, did we build the device to its own specification, from design validation, did we build the device that actually meets the user’s need under real conditions of use, sometimes stressed, sometimes distracted, sometimes used in a way the design history file never quite anticipated. Process validation layers a third, related question on top of both: does the manufacturing process reproducibly make that verified, validated design, run after run. The trap is collapsing all three into a single process-validation habit, confirming the line makes identical parts to print and calling the device proven, without ever closing the loop on whether a clinician’s hands, under pressure, get the intended result from that part. Once you have sat with that gap, a device that is perfectly reproducible and still capable of a use error, you do not fully stop thinking about the end user, in any modality that follows.
The night RNAi grew up
I was at Alnylam through the stretch of years when RNA interference stopped being a Nobel-winning idea from a lab bench and became an approved medicine. ONPATTRO, patisiran, was approved in August 2018, the first RNAi therapeutic the FDA ever cleared, delivered inside a lipid nanoparticle because naked siRNA has almost no chance of reaching liver cells intact on its own. GIVLAARI followed about a year later, built on a different delivery chemistry entirely, a GalNAc conjugate that lets the drug find liver cells directly, no nanoparticle required. Two approved medicines, two philosophies of delivery, inside about fifteen months. I am not sure I fully appreciated at the time how rare that is.
The validation lesson underneath both was the same lesson: when the delivery vehicle is as novel as the drug substance, you are not validating one thing, you are validating an interaction. A lipid nanoparticle’s encapsulation efficiency and particle size distribution are not background characteristics you check once and move past. They are as central to whether the medicine works as the sequence itself, because a poorly formed particle does not deliver its payload no matter how correct that sequence is.
And then there is cold chain, which I had spent my entire small molecule career treating as a logistics problem, something shipping handled after the real work was done. Lipid nanoparticles taught me that cold chain is part of the process, not downstream of it. A temperature excursion is not only a risk to the product, it can be a physical change to the particle itself, one a release test performed before shipping will never catch, because the excursion has not happened yet. The first time I understood that a refrigerated truck failure was, functionally, a process deviation and not a shipping incident, I started reading distribution qualification data the way I used to read a bioreactor batch record. Cold chain qualification for cryoshippers and dry-vapor shippers walks that same instinct through in detail: a shipping lane is part of the process, not downstream of it.
When it is just chemistry is the trap
Oligonucleotide validation at Wave sharpened something RNAi had only hinted at. Synthetic oligonucleotides are made by solid-phase chemical synthesis, which sounds, on paper, like the most controllable process on this entire list, closer to small molecule manufacturing than anything else here. That is exactly the trap. “It is just chemistry” is the sentence that gets people to under-invest in characterization, because a stepwise synthesis feels countable and predictable in a way a living cell never does.
It is not that simple. A solid-phase synthesis builds the sequence one coupling cycle at a time, and an incomplete coupling at any cycle does not just disappear, it becomes a shortmer, a sequence one or more bases short of the target, chemically similar enough to co-elute or nearly co-elute with the intended product. Run enough cycles and the failure modes stack into a ladder: n-1 and n-2 truncations, depurination products, double-coupling longmers, a family of closely related species an analytical method has to resolve individually, not average into a single purity number. The invariant is that every rung of that ladder needs its own identity and its own control rationale, because a sequence one base short is not chemically identical to the intended drug substance and can carry a different activity or a different risk profile entirely. The trap is treating oligonucleotide purity the way a small molecule assay treats potency, one aggregate percentage, when what actually matters is which specific rung is present and at what level. Stereochemistry at the phosphorus backbone, where synthesis can produce a mixture of stereoisomers at each linkage unless the chemistry specifically controls for it, layers a second axis of near-identical-but-not-identical species on top of the first. Chemical does not mean simple.
The other thing that modality taught me, alongside newer high-concentration biologics generally, is the viscosity wall. Push a formulation toward the concentration a small-volume subcutaneous injection demands, and viscosity stops being a minor product attribute and starts constraining what a patient’s own hands and a small-gauge needle can actually deliver. That is a validation question as much as a formulation one. The process has to reliably hit a target that a syringe, not just a specification sheet, has to tolerate. Formulation development patterns across advanced therapies walks that viscosity ceiling through in more detail, across modalities beyond oligonucleotides.
When the batch is a patient
Nothing in the first four modalities prepared me for the first time I sat with an autologous cell therapy batch record and understood, fully, that the batch was a person. Not a sample that represented a population of patients. One patient’s starting material, collected on one day, processed once, with no second chance and no larger lot to fall back on if something went sideways. At SQZ, working on autologous cell therapy process development, that fact reorganized how I thought about nearly every control in the process.
In a small molecule plant, a batch that fails release gets investigated, and in the worst case discarded, and the loss is measured in dollars and schedule. An autologous batch that fails is not something you discard and remake next week. It may be the only material that patient’s biology can provide on that timeline, and a manufacturing failure is not an inconvenience, it is a person who does not get their treatment when they needed it. I have watched people who spent years in traditional biologics manufacturing go quiet the first time they truly understood what “batch failure” means in this context. It changes how carefully you write a deviation procedure. It changes what “acceptable risk” means when there is no next lot to fall back on.
Chain of identity, tracking one patient’s cells from collection through processing and back to that same patient without ever risking a mix-up, is not a paperwork nicety here. It is the control everything else in the process assumes is already working. You can validate potency, sterility, and every release specification perfectly and still have built nothing safe if the identity chain has a gap anywhere in it. The regulatory root of that expectation sits in 21 CFR Part 1271, Subpart C, current good tissue practice’s donor eligibility and identification requirements for human cells and tissues; a BLA-licensed autologous product carries an analogous chain-of-identity expectation under its own approved procedures and general current good manufacturing practice, but the underlying discipline, never lose track of which physical unit belongs to which person, traces back to the same source.
What I would want anyone moving into cell and gene therapy to sit with before they start is that patient-scale manufacturing does not scale the way population-scale manufacturing does. You cannot average your way to confidence across a thousand batches you will never run. You have to earn that confidence before the first real patient batch, because by the time real batches are running, there is no room left to learn from failure the way a larger, more forgiving process would allow. A working log for documenting that identity chain at every handoff, collection through infusion, is at chain of identity verification log.
The modality I have not worked: allogeneic cell therapy, and where “the batch is a patient” breaks
I want to name one modality honestly rather than blur it into the six above: allogeneic, “off the shelf,” cell therapy. I have not worked a program there myself, and a reader moving into cell and gene therapy deserves to know that plainly rather than have me write around the gap. But the field splits cleanly into autologous and allogeneic, and anyone entering it needs to know the validation posture flips again depending on which side of that line they land on, so it earns a section even without the firsthand story the other six get.
Autologous taught me that the batch is a patient, one person’s starting material, processed once, with no second chance. Allogeneic inverts that. A single donor’s cells, expanded, often genetically engineered, and banked once, become the starting material for a program that may run for a decade and serve hundreds or thousands of unrelated patients from that same banked source. The batch stops being a patient. In one sense this looks like a return to population-scale thinking, a lot tested and released the way a biologics lot is. In practice it is not that either, because it swaps the chain-of-identity risk for a different, longer-horizon one: a characterization gap discovered in the master cell bank three years into commercial distribution does not just affect batches made since the gap was found. It can call into question every lot already distributed from that bank, across every patient who ever received one.
Donor eligibility is the clearest place the posture flips. For a strictly autologous product, the donor and the recipient are the same person, and 21 CFR Part 1271.90(a) waives most of Subpart C’s eligibility determinations on that basis; the identity chain described in the section above, not a communicable-disease screen, is the control that actually matters. Allogeneic starting material carries none of that waiver. Full donor eligibility screening and testing under 21 CFR Part 1271 Subpart C, relevant communicable disease agents including HIV-1, HIV-2, hepatitis B, hepatitis C, and syphilis among others, applies because the risk now crosses to people who share nothing biologically with the donor.
The master cell bank and the working cell bank derived from it become the single most consequential validation activity in an allogeneic program, characterized under ICH Q5D and against the long-standing FDA guidance that preceded it, “Points to Consider in the Characterization of Cell Lines Used to Produce Biologicals” (1993). Identity shifts from confirming one patient’s chain of custody to confirming, commonly by short tandem repeat profiling, that a working cell bank vial pulled for a lot manufactured in year six is still demonstrably derived from the same master bank qualified in year one. Genetic and, for a cell type capable of extended or indefinite expansion such as an induced pluripotent stem cell line, karyotypic stability across passage number becomes its own tracked attribute, because a bank that drifts genetically between early and late passage is not the same starting material the original characterization qualified. For an iPSC-derived product specifically, tumorigenicity risk is a critical quality attribute with no real analog in a typical autologous product, assessed through a risk-based testing strategy that extends the same identity-purity-safety-stability framework ICH Q5D and the 1993 FDA guidance already establish, rather than through a single named compliance test.
HLA matching and immune rejection risk sit on top of all of this as a clinical and formulation question an autologous product, immunologically the patient’s own tissue, never has to answer. An allogeneic program has to decide, and defend, how much matching the clinical strategy actually requires, from none, for a cell type engineered to evade or tolerate mismatch, to a defined matching algorithm, and that decision reaches back into formulation and distribution planning in ways a strictly autologous program never encounters.
| Dimension | Autologous | Allogeneic |
|---|---|---|
| What the batch is | One patient’s own cells, one use | A manufacturing lot from a banked source, many unrelated patients |
| Donor eligibility scope | Largely waived under 21 CFR 1271.90(a), donor and recipient are the same person | Full Subpart C screening and testing for relevant communicable disease agents |
| The central validation artifact | The chain-of-identity record from collection to infusion | The master and working cell bank characterization package under ICH Q5D |
| Identity testing answers | Is this physical unit still this one patient’s material | Is this bank vial still demonstrably the same donor-derived line, years later |
| Attribute with no real counterpart on the other side | None beyond identity itself; there is no banked source to drift | Genetic and karyotypic stability across passage; tumorigenicity risk for an iPSC-derived line |
| Consequence of a found gap | Contained to one patient’s batch | Can reach back across every lot ever distributed from the affected bank |
I have not sat in the room for this one the way I have for the other six, and I would rather say that directly than imply otherwise. But the instinct that this whole piece is actually about, ask what fails and who it fails on before you trust a control, would not change if I did. Only the object it points at would: from one patient’s physical identity to one bank’s genetic and biological identity standing behind years of patients who never met each other. A ready-to-use protocol built to that posture, master and working cell bank qualification matched to this same characterization logic, is at Master Cell Bank / Working Cell Bank Qualification Protocol.
The question that never fully closes
Commercial-stage gene therapy manufacturing sciences and technology, at Sarepta, introduced a different kind of discomfort: living for years with a potency assay you know is imperfect, and shipping product anyway, because imperfect and validated is still the honest state of the science.
FDA’s thinking on potency has moved twice since the original 2011 potency-testing guidance for cellular and gene therapy products. A December 2023 draft, Potency Assurance for Cellular and Gene Therapy Products, reframed the whole exercise as a strategy built from several pieces working together, manufacturing process design, process control, material control, in-process testing, and lot-release assays, rather than one assay carrying the full burden of proof. An August 2026 draft supplement, Potency Assessment of Active Immunotherapy Products, extends that same assurance-strategy thinking beyond the classic viral-vector or cell-therapy frame to peptide- and protein-based, vectored, personalized, and cell-based active immunotherapy products. The agency keeps rewriting this guidance because the underlying problem does not go away: a viral vector’s potency has to reflect the biological activity that actually matters in the patient, not just a proxy that is easy to measure and loosely correlated with what you really care about. Cell-based potency assays are, by nature, higher variability than a chromatography peak, and at commercial scale you do not get to wait for a perfect assay before releasing product to patients waiting for a one-time, potentially transformative treatment.
What I learned to do, and what I would tell anyone facing this for the first time, is stop treating assay imperfection as something to hide and start treating it as something to manage openly. The invariant underneath commercial gene therapy potency is triangulation: several complementary measures that each capture a piece of the same biological claim, a transduction or expression readout alongside a functional bioassay alongside a molecular measure, so no single imperfect number carries the whole burden of proof. The trap is waiting for one assay to become perfect before you trust it, which in practice means waiting forever, or worse, picking whichever number looks cleanest and quietly retiring the others because they disagree with it. A single number everyone has agreed to stop questioning is not more honest than three that visibly do not always agree. Statistics as honesty again, wearing a different modality’s clothes. A worksheet for building and defending that panel is at potency assurance triangulation matrix.
The newest frontier, the same old questions
These days I work on RNA editing drug product, a modality young enough that some of the analytical methods I now rely on did not exist in a validated form five years ago. ADAR-mediated editing and the platforms built around it raise a genuinely new set of questions: how do you demonstrate editing specificity with real confidence, how do you characterize an outcome that is, by design, a change to a molecule rather than the simple presence or absence of one.
Demonstrating that specificity turns out to be the same triangulation instinct gene therapy taught me, wearing yet another costume. On-target editing efficiency at the intended adenosine, read by sequencing, is necessary and nowhere near sufficient on its own. Bystander editing at adenosines near the intended site, and off-target editing elsewhere across the transcriptome, both need their own characterization, because a therapy that edits its target well but also edits sites nobody specified is not a specific therapy, it is a lucky one so far. Sequencing tells you where editing happened; it does not by itself tell you what that editing did inside the cell, so a functional or phenotypic readout downstream of the edit still has to confirm the biological claim that sequencing data alone cannot make. The trap here is the mirror image of the oligonucleotide trap: treating a single on-target editing percentage as the whole story, when the real question a validation package has to answer is what happened everywhere else in the transcriptome, not only at the one site chosen on purpose.
New questions, familiar posture. I still ask the same four questions I asked at Novartis in a room full of tablet presses. What fails, and would I know in time. Can I trust this number back to its source. Do I understand this process well enough to defend it when it drifts, before I lock it. Is this statistic honest, or built backward from the answer I wanted. The molecule changes every few years. The posture has not changed once.
Where people trip moving between modalities
The same handful of mistakes shows up whenever someone, myself included in earlier years, moves too fast between modalities without re-deriving the basics.
- Carrying a control forward because it worked last time, not because the new modality shares the failure mode it was built to catch. A disposition procedure written for a batch you can discard and remake does not survive contact with a batch that is a patient.
- Treating a novel delivery vehicle as packaging rather than a co-equal critical system. A lipid nanoparticle’s particle size distribution is not background data; it is as central to whether the medicine works as the payload sequence.
- Waiting for a single assay to become perfect before trusting anything it says, instead of triangulating several imperfect, complementary measures that each capture a piece of the same biological claim.
- Under-investing in characterization because a process sounds simple on paper. Chemical synthesis, mechanical assembly, and enzymatic editing all invite this trap in their own way; “simple to describe” and “simple to control” are not the same claim.
- Losing sight of cold chain, or any distribution step, as part of the manufacturing process rather than a logistics function that happens after the real work is done.
Common inspection findings, by modality
A generic list of common GMP findings, missing signatures, undocumented deviations, applies to all six modalities equally and says nothing useful about any of them specifically. The findings worth naming are the ones that only show up because of what makes each modality different, the same distinctions the sections above walk through in depth.
| Modality | A finding pattern specific to this modality | Why it is specific to this modality, not generic GMP |
|---|---|---|
| Small molecule tablets | Process validation trending built on a batch population that quietly excludes discarded or reworked batches, hiding a signal the full trend would show | A population-scale modality’s real risk is false confidence from an incomplete population, not a missing-record problem |
| Medical devices | A design history file that cannot show design verification and design validation as two separate, distinct records with their own evidence | The habit of “we tested it and it worked” collapses a three-question structure, design verification, design validation, process validation, that the regulation keeps explicitly separate |
| RNAi lipid nanoparticles and oligonucleotides | A specification treating an aggregate purity percentage as sufficient, with individual near-identical species, shortmers, longmers, stereoisomers, not separately identified and controlled | A small-molecule-style single purity number applied to a stepwise synthesis that actually produces a ladder of chemically distinct, not merely trace, related species |
| Autologous cell therapy | A gap in the chain-of-identity record between collection and infusion, invisible until an audit specifically traces one physical unit forward from start to finish | Batch averaging, the tool that hides small errors in a population-scale process, does not exist here; a single-unit error is the whole batch |
| Commercial gene therapy | A potency release strategy resting on one assay with no triangulating orthogonal measure, or a disagreeing assay quietly dropped from the panel with no documented rationale for dropping it | The multi-assay triangulation described above is the current regulatory expectation for this modality specifically; a single clean-looking number is not automatically defensible the way it might be for a simpler product |
| RNA editing | Sequencing data reported for on-target editing efficiency with no corresponding bystander or off-target characterization anywhere in the same package | A validation package that answers “did it edit the intended site” without answering “what else did it edit” is answering an easier question than the one that matters |
What this looks like as an interview question
“You are experienced in biologics but have never worked in cell therapy. What would worry you most moving into that role?” The strong answer names the batch-is-a-patient shift explicitly: population-scale statistical thinking stops applying cleanly, and every control built around “there is a next lot if this one fails” has to be re-examined, starting with deviation and disposition procedures. A weak answer treats the move as the same job with new vocabulary.
“How do you evaluate whether a technique or control from one modality will transfer to another?” The strong answer starts from the failure mode, not the technique: name what the control was actually built to catch, then check whether that failure mode is even physically possible in the new modality, rather than assuming a control transfers because it worked before. A weak answer assumes prior experience transfers directly without re-deriving it.
“Tell me about a time a potency or identity assay was not perfect, and how you handled it.” The strong answer describes triangulating multiple complementary measures rather than waiting for, or pretending to have, confidence in a single imperfect number, and being transparent about the limitation rather than hiding it. A weak answer implies the assay was eventually perfected, or avoids the question of imperfection entirely.
“Is an RNAi lipid nanoparticle therapeutic regulated as a drug or a biologic, and why does it matter operationally?” The strong answer states plainly that it is a drug, reviewed by CDER under a New Drug Application, because it is chemically synthesized rather than derived from a living system, citing ONPATTRO and GIVLAARI as both approved via the NDA pathway despite their biologic-like manufacturing complexity. It goes further to state why this matters beyond trivia: the applicable current good manufacturing practice regulation, the specification philosophy, and the organizational reviewers who sign off on release all trace back to that classification, so getting it wrong early shapes the wrong validation package for months before anyone catches it. A weak answer assumes the classification from the delivery vehicle’s complexity rather than the actual legal definition, or treats the question as an administrative detail with no downstream consequence.
“How would validation change if you moved from an autologous program into an allogeneic cell therapy program?” The strong answer names the shift directly: the batch stops being one patient’s physical identity and becomes a bank’s genetic and biological identity standing behind years of distributed lots to unrelated patients. Donor eligibility screening under 21 CFR Part 1271 Subpart C applies in full rather than being largely waived, and the master and working cell bank characterization under ICH Q5D, identity by short tandem repeat profiling, and, for an iPSC-derived line, karyotypic stability and tumorigenicity risk, replace the chain-of-identity log as the central validation artifact. A weak answer assumes allogeneic is simply the autologous discipline applied to a bigger batch, with no change to what is actually being controlled.
A checklist built from this framework, for anyone about to move a validation program, or their own career, into a new modality, is at cross-modality validation transition checklist.
What I would tell someone just starting
If you are early in this field, wondering whether to specialize immediately, my honest answer is: not yet, and maybe not ever completely. Go deep enough in one modality to know what real process understanding feels like, because you will recognize its absence instantly in every modality after that. Then let yourself move, if the chance comes, because the discomfort of not knowing a new modality’s specific failure modes wears off faster than people expect, usually within one real investigation, and what does not wear off is the instinct underneath it.
Ask what fails and who it fails on before you ask what the regulation requires. The regulation is almost always trying to codify that same question, written down after someone already learned it the hard way. Trust a number only as far as you can trace it back. Understand a process before you lock it, even when the schedule is screaming at you not to take the time, especially then. And when a statistic becomes an acceptance criterion, ask honestly whether it would survive someone else building it from the same data with no idea what answer you wanted them to reach. That same honesty is what separates a filing’s validation summary a reviewer learns to trust from one they learn to discount; see writing CMC validation summaries for regulatory filings for how that plays out on the page.
Six modalities in, that is still the whole job. Everything else is vocabulary.