The Unspoken Truth About Cell Therapy Product Understanding

By Arnaud Deladeriere, Ph.D., Cell&Gene Consulting Inc.

Here's a sticky premise, an elephant in the room, if you will: cell therapy developers often lack full product understanding, even up to the point where the patient is receiving a cell infusion.

Participants in a recent gathering of the Cell&Gene Foundry concluded that cell therapies are still moving toward the clinic with too little product understanding, too late in the process; however, the group did not treat it as a simple and universal conclusion. As a whole, the cell therapy field knows far more than it did when the modality emerged. Regulators have grown more comfortable as certain technologies standardized, and the trajectory is better than it was. But the basic premise, in several important contexts, remains true.

The clearest examples are allogeneic and regenerative medicine products, where too little product understanding can still arrive at the point of translation. When a team does not know what to measure, its ability to characterize both product and process erodes, and what should become an early-phase manufacturing process is assembled on partial information.

Autologous CAR-T presents a different version of the same problem: the starting material is variable from the outset, and applying uniform manufacturing concepts to a non-uniform input produces an output that remains, in the Foundry participants' framing, a question mark shaped by whatever entered the process at the front end.

That observation carried a note of retrospective candor. The CAR-T field moved quickly enough that it missed the opportunity to capture data it now recognizes as essential: the relationships between apheresis characteristics, manufacturing timelines, cell exhaustion, dose, and the persistence of the infused cells. Those relationships were not predicted early, and the parameters that would have illuminated them were not collected.

Survey work on what should be captured, from apheresis through follow-up, has since exposed how unsettled the field remains on characterizing the starting material and relating manufacturing data to therapy and patient outcome. The knowledge is being generated now, later than it might have been.

A Case For Assaying Early, Even Before It's Required

Developers tend to discover the attributes they should be measuring only once it is too late to act on them. A well-designed program works backward from the end game, asking what will be needed at commercialization and reasoning about it from day one, anchored in mechanism of action. If a developer understands how the product works, what it does, and what it should not do, most of what matters can be anticipated, and engineered products make this concrete: a construct designed for a defined function tells the developer what the cells are meant to do and therefore what to watch for.

Predictable problems that were not considered mark poor development; unpredictable problems that surface late are part of the nature of the work, and the response is a plan built to confirm that nothing sits outside the original thinking and to manage risk when something does. The cost of learning late is highest where the omission was foreseeable.

In early development, it's helpful to produce a functional readout earlier, particularly for pluripotent-derived products with long differentiation protocols. A qualified potency assay is not required for early phase, but some measure of function is needed to develop the process and define the design space at all, ideally an early yes-or-no signal before the cells reach a final format that makes measurement difficult, such as a scaffold.

Alongside function, basic characterization of what else is present in the product, and whether the same populations appear consistently, was treated as underused. Adaptive manufacturing work reinforced the point, using metabolomic profiling of source material and in-process product together with inline sensors to predict where a cell population is heading while the process runs. The underlying claim, that manufacturing genuinely matters and that something steerable is happening during it, framed characterization as a live control problem rather than a retrospective audit.

Unburdening Potency With Surrogate Markers

The Foundry reframed the question of whether the field asks potency assays to do something they were never designed for. The concern is less that potency is the wrong concept and more that it is asked to carry too much weight and is made too complex in the process. Potency is essential but does not stand alone. It sits at the end of a chain running from mechanism of action through defined characteristics and release criteria to expected function, and only then, with patient variability layered on, to clinical benefit.

The productive path identified is surrogacy done rigorously. If a developer can document that a set of markers reliably accompanies a demonstrated function, and validate that connection, the markers can serve as a faster, more practical assay that still carries meaning. An engineered MSC product designed to home to inflammation and secrete IL-10 allows the secreted cytokine to be measured as a correlate of an immunomodulatory function documented in vitro and in vivo.

The pluripotent field can drift toward potency assays that attempt to mimic the clinical outcome itself, which becomes theoretical and fragile, especially when a product matures in vivo over weeks before it secretes anything measurable.

The Foundry participants agreed it's better to identify a biological activity the cells must exhibit, measure it, and link it to trial outcomes, rather than overcomplicate the assay in place of understanding the biology. In vivo approaches sharpen the difficulty: the administered product is the vector rather than the cell, control is held at arm's length, and potency shifts toward attributes such as transduction, with an additional in vivo step beyond the developer's control. Efficacy is difficult to harness there; safety, even more so.

Are We Learning From Data Or Defending Decisions?

A question the Foundry has circled before returned here: is analytical data collected to learn from or mainly to defend decisions already made? The answer was that it is legitimately both, provided the program is aiming at the final product from the outset.

Learning should be concentrated early, when the developer knows least and the design space remains broad, and then narrow as product and process come under control, tolerating the occasional detour but avoiding those that consume time and denature the effort.

This surfaced a useful disagreement. One position held that later-stage data should become more precise and tightly bounded, with manufacturing data distinguished from clinical usage data, and batch-to-batch consistency delivered by release criteria that absorb most starting-material variability. The counter-position held that the field is still failing to collect data that matters, particularly for commercial products, where the correlation between the manufacturing data of individual infused batches and long-term outcome remains largely absent. The contention was not whether release criteria create consistency but whether current criteria capture the full picture of what a product is, and whether the field is closing off knowledge it will later wish it had.

The exchange over how much resolution a release specification needs, framed as the difference between an image of millions of pixels and one of a few thousand that still conveys the information, captured the tension between the effort a measurement costs and the relevance of what it returns.

The clinical and regulatory reality complicated any clean answer. Batch release was described as less binary than it appears. Threshold-based specifications, rather than fixed single values, gave qualified persons room to release more batches while accepting more variation between them, and noncompliant batches are sometimes released with the treating physician holding final responsibility for administration. These are the conditions under which analytical strategy operates, some distance from the idealized version.

Two Different Jobs: Understanding The Product And Releasing The Batch

The closing question asked whether better analytics could become a practical lever for access, scale, and adoption. The answer was carefully qualified. Better analytics help when they are the right analytics, and the improvement lies as much in design, execution, interpretation, and fitness for purpose as in the measurement itself, which is where more capable assessment tools, including AI, have a role. But more is not automatically better.

Powerful new measurement technologies that are not validated as release assays can complicate matters rather than advance them, because not every site will have the instruments, and a standard that cannot be implemented everywhere risks hindering the data collection it was meant to enable. Standardization was endorsed only where a developer already knows where they are and does not intend further tailoring; premature standardization forecloses opportunities while leeway still has value.

The distinction the group kept returning to was between analytics used to understand and characterize a product and analytics needed to release a batch into a trial. The first can be rich, exploratory, and data-hungry, and it is where federated learning and cross-referencing across institutions could accelerate the field if the infrastructure to share data existed.

The second should be simple. Once the product is understood and the process defined, a release panel needs only to link to safety, identity, purity, dose, and biological activity, and that simplicity is what supports adoption, particularly in decentralized models where access to advanced instruments cannot be assumed. Testing becomes a bottleneck mainly when analytics and product are poorly matched, or when a timeline cannot be met, rather than as an inherent feature of the work.

Asked what one change would matter most over the next three years, the answers converged on the infrastructure that does not yet exist: better ways to share data, so the field can learn from its accumulated experience rather than each developer rebuilding that knowledge alone.

Closing Perspective

This Foundry session began with a question about product understanding, but the discussion widened because analytics are no longer just a CMC discipline. They are becoming part of how the field decides what to build, how to manufacture it, how to release it, how to compare it, how to transfer it, and how to make it available across real clinical systems.

A potency assay is not just a regulatory requirement. Characterization is not just a description of what is in the vial. Analytical data is not just a defense of decisions already taken. Used well, these tools create the feedback loop between biology, process, product, and patient. The field does not need more complexity for its own sake. It needs earlier functional insight, clearer separation between exploratory characterization and release control, better ways to share data, and analytical strategies that support confidence across sites, products, and patients.

In cell therapy, the product is still too often understood in motion. The next phase of the field will depend on whether developers can learn what they are making early enough for that knowledge to change the outcome.

Key Takeaways

About The Author:

Arnaud Deladeriere, Ph.D., is principal consultant at Cell&Gene Consulting Inc. Previously, he was head of MSAT and Manufacturing at Triumvira Immunologics, and before that, manufacturing manager at C3i. He received his Ph.D. in biochemistry from the University of Cambridge.