There may be no word whose meaning diverges more between biology and everyday speech than “aging.” What cell biology calls cellular senescence is the phenomenon that a single cell stops dividing—and still does not die. That normal human cells divide only a few dozen times in a culture dish was shown 60 years ago (Hayflick L, Exp Cell Res 1965;37:614–636), and what was counting the divisions turned out to be the telomeres, which shorten with every division (Harley CB, Futcher AB, Greider CW, Nature 1990;345(6274):458–460). Even on reaching that ceiling, the cell does not die. Not death, but retirement. Let us begin with that metaphor.
What we are reading today is not about cells growing old in a patient’s body, but about cells being run inside a device in order to make the raw material for a drug. An extracellular vesicle (EV) is a sac of thousands of kinds of molecules, and it cannot be written as a structural formula. So instead, we have declared “this is an EV preparation” on the basis of particle size, surface markers, and particle number. We have been defining a product’s identity by the outside of the box rather than by its contents. Today’s paper ran human bone marrow–derived mesenchymal stem cells (BMSCs) in a hollow fiber bioreactor (HFB) for about 40 days and, with mass spectrometry, opened up the contents of the small extracellular vesicles (sEVs) that came out four times a week.
On the cell side, things look to have tilted in the direction of senescence. And yet particle size, markers, yield—none of the three items that verify a vesicle’s identity moved at all, the paper reports. Meanwhile, in the contents they opened up, 1,618 proteins were “differentially altered” between the early phase (Phase 1) and the late phase (Phase 3). This is not “1,618 molecules were swapped out”; it is the number of proteins whose between-phase difference in amount crossed a statistical threshold.
Let me state the scale of this up front. Everything that appears here is only what happened inside a culture device. And the most important caveat of all—this study never once showed that the sEVs harvested later had lost efficacy. What was measured is composition, not potency. Neither the number of donors nor the number of devices is stated. Note too that “the three items did not move” is not the same as “they were constant.” Particle concentration fluctuated between 3.9 × 10⁸ and 8.2 × 10⁹ particles/mL, and the paper itself calls this a “reproducible oscillatory pattern.” What was preserved is the order of magnitude, not the same number.
Publication Details
- Paper title: Small extracellular vesicles reflect senescence progression in human bone marrow–derived mesenchymal stem cells during hollow fiber bioreactor culture
- Paper type: Original research article (Research Paper). It consists of culture experiments only, and includes neither animal experiments nor clinical trials. There is not a single experiment measuring the function of the sEVs either
- Authors: Ali Dinari (first author and co-corresponding author), Armin M. Ebrahimi, Bartosz Leszczyński, Kamil Wawrowicz, Masoud Rezaei, Maciej Słotwiński, Zenon Rajfur, and Ewa L. Stępień (corresponding author)—eight in all. No note of equal contribution
- Affiliation: Centered on the Department of Medical Physics in the Faculty of Physics at Jagiellonian University in Kraków, Poland
- Journal: Nanotheranostics 2026, Volume 10, pages 142–157. A fully open-access journal from Ivyspring International Publisher
- DOI / Link: 10.7150/ntno.132088
- Peer review and publication dates: Received January 25, 2026; accepted March 26; published July 19
- Open access: Yes (CC BY 4.0)
- Impact Factor: What follows is a check made outside the paper—no official Clarivate impact factor has been assigned. The “8.25” the journal displays is a third-party tally that the publisher itself flags as an “Unofficial Impact Score,” and the Scopus-derived CiteScore is 8.4 (we could not pin down which year it covers)
- About the journal: A legitimate title indexed in MEDLINE, PubMed, and Scopus, and a small one, publishing on the order of a few dozen papers a year. This too is a check made outside the paper: just under 300 papers are indexed in PubMed, and we found no retractions or expressions of concern. Not a dubious journal, but not a top journal either
- Funding: Polish National Science Centre (NCN) 2022/45/B/NZ7/01430. The paper’s Funding section gives only the number and does not write the project title. Tracing the funder’s public records, the project title is “Modification of extracellular vesicles towards future drug delivery systems”—a budget awarded under a heading that is neither “stem cell senescence” nor “quality control” (the project title is information from outside the paper)
- Conflicts of interest: The authors declare no competing interests
- Data availability: Stated as deposited in the public repository RODBUK. The DOI is printed broken by a space at the end of a line, but joined up it reads
10.57903/UJ/GIM176. In a check outside the paper, however, this DOI did not resolve, and a search on RODBUK itself failed to turn up the dataset in question. No deposition in the field-standard PRIDE / ProteomeXchange is mentioned
About corresponding author Ewa L. Stępień (everything below rests on checks made outside the paper): A full professor (prof. dr hab.) at Jagiellonian University, based in the Department of Medical Physics and holding a joint appointment at the Center for Theranostics. The pillars of her record are a lineage that ran from diabetic vascular pathology to EV biomarkers (Theranostics 2018;8:3874–3890), and total-body PET and positronium imaging (Nat Rev Phys 2019;1(9):527–529). As far as we could check, she has almost no prior track record in the senescence biology of mesenchymal stem cells or in bioreactor engineering. Nor is she included in the author lists of MISEV2018 or MISEV2023, the international guidelines for EV research: she is on the side that complies with them, not the side that writes them. First author and co-corresponding author Ali Dinari likewise came to Jagiellonian University in 2025 by way of Tarbiat Modares University in Iran and the Gwangju Institute of Science and Technology in South Korea, and his specialty is enzyme-immobilizing nanomaterials and magnetic nanoparticles—neither stem cell biology nor senescence biology. The hollow fiber bioreactor first appears in this laboratory in a pre-peer-review preprint from the end of 2024 (whose first author is Kamil Wawrowicz, the person this paper explicitly credits with “HFB methodology”). In other words, this is a paper that brought an established EV analysis technique to bear on a new subject.
What Wasn’t Understood Before?
A retired cell does not sleep quietly. It keeps on secreting inflammatory cytokines and proteases. This is called the senescence-associated secretory phenotype (SASP) (Coppé J-P, et al., PLoS Biol 2008;6(12):e301). Mesenchymal stem cells (MSCs) have this ceiling too, and the more you expand them in culture, the flatter their shape becomes and the weaker their surface markers grow. What is more, this does not happen abruptly at late passages: it is a downhill slope that advances continuously along with passaging (the operation of detaching expanded cells and replanting them in a new vessel) (Wagner W, et al., PLoS ONE 2008;3(5):e2213).
👦 Student: If all it did was stop dividing, then as long as it stays quiet there’s no harm, is there?
🧬 Dr. Exotaro: Think of a senescent cell as a car parked with the handbrake pulled on. It will never drive again, yet the engine is running and the radio is still playing. That radio is the SASP. That vesicles released by senescent cells make the surrounding cells senescent has also been shown in cultured cells (Borghesan M, et al., Cell Rep 2019;27(13):3956–3971.e6). The premise that a retiree is harmless simply does not hold.
The trouble is that those aging cells are themselves the factory that makes the drug. Let me get one thing out of the way first. There is not a single approved extracellular vesicle drug, at least in the United States, and no release specification laid down by a regulator exists either (in 2019 the US FDA issued a public safety notification stating that “there are no FDA-approved exosome products.” This is information from outside the paper). Today’s paper is not a study dealing with release specifications or pass/fail judgments either. The three-item set at the top of this article is no more than a convention of the field, and if you ask whether we should simply measure function instead, there is no consensus there either—which assay can reliably measure function is unresolved, as an international workshop report on large-scale EV production states explicitly (Paolini L, et al., J Extracell Biol 2022;1(10):e63). The only move left is to “manage the state of the cells that made them.” That is why ISEV (the International Society for Extracellular Vesicles) spells out, among its reporting items, that you should state the passage number and the total conditioning time (how long the medium was left on the cells) (Shekari F, et al., J Extracell Biol 2023;2(10):e115).
Let me describe the device as well. A hollow fiber is a thin tube whose wall is a semipermeable membrane; in a cartridge packed with several hundred to several thousand of them, the medium circulates through the inside of the tubes (the lumen), and the cells live in a small chamber on the outside of the tubes. The membrane’s molecular weight cut-off (MWCO) is 20 kDa—glucose (180 Da) passes through, but vesicles tens to 200 nm in diameter cannot. So the EVs become concentrated in that small chamber. The price is that you cannot look inside the device.
👦 Student: If this is about cells growing old, hasn’t it been studied over and over already?
🧬 Dr. Exotaro: If you mean “comparisons across passages,” yes. There is a report that EVs from BMSCs passaged in flask culture out to passage 8, unlike those at passage 2, failed to raise Arg1 or IL-10 in a culture-based assay (Liu Y, et al., J Extracell Vesicles 2024;13(8):e12472). But that is comparing snapshots of separately established cultures. Records that kept a single manufacturing run going without stopping it and followed how the cargo shifted on the inside were almost nonexistent. That is the film today’s paper set out to shoot.
What Did This Paper Find?
Splitting the Forty Days into Three by Medium and Operating Conditions
BMSCs passaged three times were expanded, and about 6.5 × 10⁷ cells were injected into the cell-side space of a cartridge (FiberCell Systems, Cat#: C2025D—a laboratory scale with 450 cm² of membrane area and a 3.2 mL cell-side space; these two numbers are not printed in the paper but are the manufacturer’s published values). Medium containing sEVs was harvested four times a week, 3 to 3.2 mL each time, for a total of about 70 mL over the whole period (divide one by the other and you get roughly 23 harvests, but the number of harvests itself is not printed in the paper). The three phases are not divisions carved out of the data but divisions of medium and operating conditions: (i) cell adaptation, days 1–7; (ii) serum removal, days 8–21; and (iii) transition to xeno-free (containing no animal-derived components) conditions, days 22–40. Note that the number of operating days is inconsistent between 39 days (main text) and 40 days (abstract and phase definitions). For a paper selling a time-lapse, the clock is not set right.
The Methods list three monitored items—glucose consumption, viability, and sEV recovery—but the only one described as “continuously monitored” is glucose consumption. And that is an indicator on the medium side, not of the cells themselves. Because MSCs are adapted to a hypoxic niche and depend heavily on glycolysis, the amount of sugar consumed serves as a surrogate indicator of whether metabolism is turning over—that is the paper’s logic. The seeding itself succeeded, and the authors describe the cells as adhering uniformly to the fiber surfaces and rapidly forming a confluent layer. On that basis the main text writes that “throughout the 39-day culture period, glucose showed a steady-state consumption profile,” and concludes that the stepwise transition to xeno-free conditions did not disturb metabolic stability either.
Yet the main text attaches neither statistical support nor a criterion for the judgment to this “steady state.” The one quantity recorded continuously during the run is being called “steady” with no grounds shown. And the authors themselves write that this metabolic flatness and the cells’ function are separate things—“despite sustained glucose utilization, functional viability declined progressively with extended culture duration.” Even when a metabolic indicator is flat, it does not follow that the cells inside have stayed the same. The paper says so itself.
The Three Identity-Checking Items Picked Up Nothing
The sEVs were recovered by differential centrifugation followed by ultracentrifugation at 150,000 × g for 90 minutes. Particle number by nanoparticle tracking analysis (NTA) was 3.9 × 10⁸ to 8.2 × 10⁹ particles/mL, averaging roughly 2.5 × 10⁹. Particle size was “a consistently narrow distribution below 200 nm”—but neither the mean diameter nor the mode diameter (the most frequent particle size) nor any numerical value for the distribution appears in the main text or the figure legends (supplementary materials were not checked for this article). For markers, Western blot detected CD9 (about 25 kDa), HSP70 (about 70 kDa), and ARF6 (about 20 kDa) (the material was a “representative sEV fraction,” and there is no quantitative comparison between phases). Total protein was 50 to 500 µg/mL.
👦 Student: If the size, the number, and the markers are all the same, isn’t it the same product?
🧬 Dr. Exotaro: That is exactly the premise this paper overturned. Here’s an analogy—the same factory kept shipping cardboard boxes of the same size for 40 days. The boxes carry the same logo (CD9). Count them and check the logo, and you can say the boxes are the same.—And yet when they were opened, the composition of the contents was quite different between early and late. A logo is the manufacturer’s name, not an inventory of the contents. Particle counting only looks at scattered light; in principle it cannot see the contents. Incidentally, the paper’s “was stable” means “stayed within the same order of magnitude.” The spread between 3.9 × 10⁸ and 8.2 × 10⁹ is about 21-fold—that multiple is not printed in the paper, so the division is ours—and “stable” is a word used with that spread included.
There is a hole in the counting as well. What NTA returns is the number of particles per 1 mL of medium, not per cell. And for the number of cells inside the cartridge, no value after the time of seeding is reported (the Methods state that the recovered cells were “counted,” but that number never appears in the main text). Even so, the conclusions state that both “production efficiency” and “vesicle biogenesis” were preserved. With no data normalized to cell number, the maintenance of production is being asserted from measurements per unit volume. “Yield was stable” is equally consistent with “the cell population shrank and the remaining cells secreted more.”
Opening the Box—5,071 → 4,415 → 3,337
Three biological replicates per phase, nine samples in all, were measured on an Orbitrap Astral mass spectrometer in data-independent acquisition (DIA) mode and analyzed with DIA-NN v2.1.0. The database was the human UniProtKB reference proteome (20,663 sequences), with a false discovery rate (FDR) of 1%. One point of terminology here. The “number of proteins quantified” is not a property of the specimen itself; half of it is a property of the measurement—how many kinds that instrument could dig up on that day. With that said, the numbers quantified were 5,071 in Phase 1, 4,415 in Phase 2, and 3,337 in Phase 3, and the numbers differentially altered were 1,618 for Phase 3 vs Phase 1, 579 for Phase 2 vs Phase 1, and 550 for Phase 3 vs Phase 2.
👦 Student: If the proteins dropped from 5,071 to 3,337, doesn’t that mean the vesicles themselves decreased?
🧬 Dr. Exotaro: The authors raise the same question and reject it. Because particle number and protein amount were both maintained, they argue, it is not that “production fell” but that “the cargo was selectively restricted.” But here a two-way choice has been set up. There is a third possibility—the Phase 1 medium contains 10% fetal bovine serum (FBS), and the design replaces it partway through with xeno-free medium at a one-to-one ratio. On top of that, the database is human only. A mere decrease in bovine-derived proteins could move the number of human proteins detected in the same direction. How much that contributed cannot be estimated from the paper’s description, and there is no sign that it was examined.
The change had structure. Hierarchical clustering splits into “a group enriched toward Phase 3” and “a group that becomes depleted.” Analysis with STRING, which gathers the connections between proteins, gave 71 clusters on the upregulated side and 148 on the downregulated side (five or more proteins per cluster), and in the UpSet plot that tallies which comparison each set is unique to, the largest set was again the downregulated side of Phase 3 vs Phase 1 at 560, followed by the upregulated side at 402—the direction of “loading less” dominates over “loading something new.” Enriched in the late phase were extracellular matrix (ECM) construction, adhesion signaling, vesicle biogenesis, and the endolysosomal pathway. Depleted were RNA splicing, ribosome biogenesis, transcription, and the TCA cycle. The main text, however, merely lists these pathway names and attaches not a single per-pathway enrichment statistic—no FDR, no enrichment ratio, no number of molecules involved. Only 21 upregulated and 6 downregulated proteins moved consistently across all three comparisons, and on the upregulated side ECM-related proteins dominated—fibulins (FBLN1, FBLN4), tenascin (TNC), collagens (COL4A2 and others) (the nuclear envelope protein LMNB1 and the endoplasmic reticulum stress–related NUCB1 and DNAJC3 are included too). What came to stand out was patching putty and sealing tape—the number of boxes did not change, yet the character of the contents shifted. It was not a straight line either: there is a group transiently induced in Phase 2 (MAT2A, GSTA1/2, XDH, and others), and the authors conclude that senescence is “not a linear decline but multi-stage adaptive remodeling.” That is a conclusion a two-point comparison could not produce, but when you are dealing with thousands of proteins, non-monotonic patterns arise in large numbers from chance alone. There is no mention of permutation testing or of stability assessment.
What Was Happening on the Cell Side—And the Asymmetry of the Clock
Cells recovered from the HFB and replated in 2D were flat and irregular in shape with low density, their F-actin disordered and their nuclei enlarged. Senescence-associated β-galactosidase (SA-β-gal) staining turned the cytoplasm densely blue, while young controls barely stained at all. This stain is not looking at an enzyme specific to senescence; it looks at a trick whereby, because lysosomes enlarge with senescence, the activity can be detected even at pH 6.0, where it would normally fall off (Lee BY, et al., Aging Cell 2006;5(2):187–195). Since serum starvation and high-density culture also turn it positive (Severino J, et al., Exp Cell Res 2000;257(1):162–171), staining at the two densities of 20% (5 × 10³ cells/well) and 80% (2.5 × 10⁴ cells/well) is a proper countermeasure. By flow cytometry, the 2D condition met the surface marker requirements of the International Society for Cellular Therapy’s minimal criteria (Dominici M, et al., Cytotherapy 2006;8(4):315–317), whereas under the HFB condition the positivity rates and fluorescence intensity distributions of CD105 and CD73 declined. For viability, multiple harvest time points were assessed in the form of replating the recovered cells, and the main text writes that cells from the earlier time points reattached and proliferated well, while the late-phase cells fell off in both attachment efficiency and proliferative capacity. The figure legends, however, state neither which operating day each time point corresponds to nor the number of replicates.
There is an asymmetry in the design here. What was followed longitudinally is the vesicle side. On the cell side, multiple time points exist only as far as survival, reattachment, and proliferation after replating; the SA-β-gal staining, confocal microscopy, and flow cytometry used to judge senescence were all measured on cells recovered after the run had ended. There are no longitudinal senescence-marker data matched to the sEV harvest time points.
👦 Student: Then how did they confirm that “the vesicles reflected the progression of senescence in the cells”?
🧬 Dr. Exotaro: The honest answer is that they never matched them up numerically. There is no correlation coefficient and no concordance rate, and the parent cells’ proteome was not measured either, so there are not two numbers standing on the same ground. What is being compared is a quantitative proteome on the vesicle side against descriptions of morphology and staining on the cell side. The paper’s “reflected it well” is the authors’ reading that both sides can be explained by the same story.
How Will the Future Change? (The Path to the Clinic)
Let me state exactly where we stand. This is a culture experiment on human-derived cells; there is no administration to animals and there are no patients. Nor is it a study that verified, in a culture run on a laboratory-scale device, whether a clinical dose can be made. To write “a step forward toward clinical application” would be contrary to fact. This is not an advance; it is an inspection.
Even so, the question connects directly to the manufacturing floor. Over the past decade, regulatory thinking has shifted from “measure the final product” to “manage the whole way it is made.” The following is information from outside the paper, but the draft guidance on potency (the actual power of a product to work) assurance for cell and gene therapy products that the US FDA issued in December 2023 moved its center of gravity to a potency assurance strategy bundling manufacturing design, process control, in-process testing, and release testing (we could not confirm whether a final version has been issued). Japanese-language guidance likewise states explicitly that the origin of the producer cells, the culture conditions, and the passage number must be managed as risk factors (Terai S, et al., Regen Ther 2025;29:43–50).
👦 Student: So should we just stop the culture at day 21?
🧬 Dr. Exotaro: It isn’t that simple. This paper never once showed that the late-phase vesicles had lost efficacy. If anything, a figure legend writes that the yield in the late phase was “slightly higher, with smaller variability.” That said, there is no room for optimism either. EVs from MSCs made senescent with a drug failed to protect cartilage in mouse osteoarthritis and, if anything, raised inflammatory markers (Boulestreau J, et al., J Nanobiotechnol 2024;22:255). Vesicles from old cells may not be “a diluted drug” but “a different drug”—that is the safer way to think about it.
Let me also spell out the limits of extrapolation. The guidance just mentioned recommends xeno-free medium containing no animal-derived components, yet this study began with 10% FBS and transitioned partway through, and that transition is itself what defines the phases. The cargo changes observed include a medium transition that would not exist if you manufactured according to the guidance. Umbilical cord– and adipose-derived MSCs differ from bone marrow–derived ones in both the rate of senescence and EV secretion. Even so, the paper’s closing reaches all the way to design guidance for “scalable stem cell culture platforms” in general, with not one sentence acknowledging how narrow the conditions it examined were.
How to Critically Read This Research (Limitations and Paths to Improve It)
The virtues first, concretely. That longitudinal sampling was done on a single run in an actual manufacturing device. That the phase divisions correspond to operating conditions rather than to the data. That the cell side and the vesicle side were both taken and measured from one and the same run (they did not get as far as matching them up numerically, but the design of lining them up within the same run is sound). The unglamorous good practices are observed too—fixed measurement conditions, printed tool versions, and naming where the raw data were deposited. And above all, that the finding which weakens their own sales pitch—that the three items do not move while the cargo changes—was not hidden but laid out in the abstract.
The limitations the authors themselves acknowledge are astonishingly few. There is no “Limitations” section. The hedges one can pick up are five: (1) asserting irreversible senescence would require additional functional analysis; (2) senescence is heterogeneous and asynchronous, not a uniform terminal arrest; (3) the markers show not “loss” but “shrinkage” of a high-expressing subpopulation; (4) RNA cargo was not measured; and (5) proportionality between particle number and protein amount does not rule out qualitative remodeling. All of them are sound, and the conclusions retain the qualifier “an adaptive, early senescence-like phenotype.” The problem is that far stronger phrasings run alongside them in the same paper—“a controlled senescence program,” “establishes,” “robust.” Weak phrasing and strong phrasing coexist, uncalibrated.
First, senescence and the change of medium are confounded. The three phases are, just as they stand, three divisions of medium condition: “10% FBS,” “serum removal,” and “xeno-free transition.” Phase 3 vs Phase 1, which produced the largest difference (1,618 proteins), is a comparison between the two points whose media differ most. And the paper does not write the final serum concentration in Phase 3 (the description goes no further than “a one-to-one mixture of DMEM and xeno-free medium”). There is no mention of EV-depleted FBS either, and the database is human only. The cells’ response to the change of medium, and the contribution of serum-derived components themselves, have not been separated out from the effect of senescence.
And this confounding compounds with the absence of controls. To say “HFB culture promotes senescence,” you need a comparison against a non-HFB condition. For the cell phenotype a 2D control was placed, and that part is properly done. What cannot be confirmed from the main text is a control in which cells of the same passage were cultured in parallel in 2D for the same 40 days and sEVs were harvested, a static 3D control, and a cell-free cartridge control. What is supported goes only as far as “after 40 days of HFB culture the cells showed a senescence-like phenotype, and sEV composition changed between phases”—not the attribution that “the HFB caused it.” The phrase “during hollow fiber bioreactor culture” in the paper’s title merely states the condition observed, and on that point it is accurate. It is the abstract and the main text that overstep, using causal verbs: “prolonged HFB culture promotes” and “induces a stress-associated senescence program.” These are not sentences you can write without a control.
Second, the “viability” in Figure 1E is placed in a form that makes it easy to misread what survival rate it is. What Figure 1E shows is viability after the cells recovered from the cartridge were replated in 2D—not the survival rate inside the cartridge during the run. Confuse the two and the story becomes “they were harvesting vesicles from a culture that had all but died off,” but that is not what the paper showed. Even so, the main text explicitly states that attachment efficiency and proliferative capacity fell in the late-phase recovered cells, and between that sentence and the conclusions’ phrasing that “the culture supported cell survival” there is a mismatch in strength. As for contamination by membrane fragments of dead cells or by apoptotic bodies, the paper states, from the NTA size distribution, that it “excluded significant contamination by larger microvesicles or cell debris.” Whether one size distribution is enough is another matter, but you cannot write that they “did not examine it.” There is one further ambiguity: the Methods describe the object of the viability assessment both as “the cells remaining in the ECS” and as “the medium recovered from the ECS,” leaving it undetermined whether what is being measured is adherent cells or detached cells.
So is this cell death rather than senescence? It is highly likely that senescence itself was genuinely occurring. The point at issue is the single question of how far the proteome changes can be attributed to senescence.
Third, the judgment of senescence stands almost on the single leg of SA-β-gal plus morphology and staining. The HFB group went through a recovery process of one hour of Accutase treatment at 37 °C and vigorous flushing before being replated in 2D, and how far the 2D control received the same treatment is not written in the paper (the control cells for the flow analysis were also detached from flasks). In addition, on the cell side, neither p16 (CDKN2A) nor p21 (CDKN1A) nor p53, nor lamin B1, γH2AX, or Ki-67, nor telomere length, nor population doubling number, nor SASP itself was actually measured—not one of them (Gorgoulis V, et al., Cell 2019;179(4):813–827 states explicitly that senescence cannot be defined by a single marker). Not even a numerical value for the SA-β-gal positivity rate appears in the main text; the Methods say it was “quantified,” yet the results give only adjectives. In the main text, the primary evidence supporting the senescence in the title goes no further than images and adjectives. —Incidentally, lamin B1 (LMNB1) was detected not in the cells but on the side of the vesicle proteome.
Fourth, the experimental unit and the statistics remain unclear. All it says is “three biological replicates per phase”; whether those are three independent cartridges, or three samples taken from the same cartridge on different days, is not written. If the latter, they are pseudoreplicates, and a t-test can answer no more than “is there a difference within this one device?” Which samples were chosen out of the roughly 70 mL harvested is also unstated. The description of the thresholds is inconsistent between the Methods (p<0.05 and |log₂FC|>1), the Results (q<0.05), and the figure legends (adjusted p), and moreover neither the multiple-comparison correction method nor whether an FC threshold was used alongside q is written. The principal component analysis (PCA) that is said to have been performed has no results appearing in the main text either. The handling of missing values amounts to the single sentence “only proteins consistently identified were included,” with no definition of whether “consistently” means 3/3 or 2/3, no number excluded, and nothing on whether imputation was used. This filter has room to act asymmetrically across phases—because proteins prone to being missing in Phase 3, where fewer were detected, could be systematically dropped. Normalization is only said to have been “performed” (row-wise z-scoring for the heatmap is stated explicitly), but which method was used for normalization between samples is not written.
Fifth, the characterization does not fully support the claims. The presentation of the positive markers is itself relegated to a supplementary figure, and neither the main text nor the figure legends contain negative controls for non-EV markers (Calnexin, GM130, albumin, apolipoproteins), electron micrographs, or purification by a different principle (orthogonal purification) such as a density gradient or size exclusion chromatography (supplementary materials were not checked for this article). Even so, the Methods and the main text write that the detection of CD9, HSP70, and ARF6 “verified purity.” That positive markers show up can be evidence that EVs are there, but it is not evidence that nothing else is in there. A different kind of evidence, assessing the co-isolation of lipoproteins and aggregates, is required.
Sixth, several of the abstract’s headline claims are stronger than the grounds in the main text. “Particle size was stable”—there are neither time-course data nor numbers in the main text. “Coordinated changes in oxidative stress–related processes”—the main text does describe proteins involved in redox buffering and antioxidant defense (GSTA1/2, XDH, and others), so they are not empty-handed here. But not a single enrichment statistic is attached to those pathway names. There are no numbers to support “coordinated.” On the other hand, the Results’ “CD105 and CD73 declined” and the conclusions’ “retention of core markers” are not a contradiction—the paper writes that “positivity rate and fluorescence intensity fell while remaining detectable” and that this is “shrinkage of the high-expressing population rather than complete loss,” which is compatible with the conclusion that lineage identity was preserved (though the degree of retention is not shown quantitatively). And the abstract writes that vesicle analysis “establishes” sEVs as “a convergent, non-invasive readout of BMSC senescence,” going so far as to write that they can be monitored sensitively and are robust. Yet this paper contains no sensitivity, no specificity, no limit of detection, and no direct comparison with a cell-based assay. There are no numbers to support “establishes”—and that is the biggest structural hole. To keep things distinct: (1) performance as an indicator for reading senescence in the parent cells and (2) validity as an indicator of sEV efficacy (potency) are separate problems; (1) is what this paper claims, and (2) is what it does not claim. Neither has been verified quantitatively.
So how could the quality have been raised? Let me start with what can be fixed with zero additional experiments. Settle whether it is p or q and state the correction method, and 1,618 becomes an interpretable number for the first time. Put the already-performed PCA into a figure. And state STRING’s background set explicitly, and check whether the enrichment survives when the proteins detectable in this experiment are taken as the background (which background was used is not written in the paper). EV proteomes are skewed toward ECM, membrane trafficking, and ribosome-related proteins to begin with, so if the whole proteome is taken as the background, those are automatically prone to come out “enriched.” On top of that, calibrate the strength of the claims—“establishes” to “suggests,” “controlled senescence program” to “senescence-associated phenotype.” And deposit the raw data in ProteomeXchange / PRIDE with a PXD number attached, release the analysis scripts, and put the RODBUK deposition into a state where it can actually be traced. As things stand, the information for a third party to reproduce 1,618 independently is not in place.
Measuring SASP cytokines (IL-6, IL-8, IL-1β, MCP-1) in the stored medium would close the hole of explaining things by SASP without measuring SASP, and would yield longitudinal data as well. Re-searching the mass spectrometry data against a concatenated human-plus-bovine database would allow the contribution of serum-derived components to be examined with no additional experiments. Quantitative Western blots across all phases and all replicates are cheap too, and lining up the negative markers mentioned above would make “the markers were stable” a quantitative claim for the first time. Taking electron micrographs for each phase, and tabulating mode diameter, mean diameter, and D10/D50/D90 by phase from the raw NTA data, would put numbers on “particle size was stable” as well. If the design is to be changed: hold the medium xeno-free and constant for the whole period, culture cells from the same donor at the same passage in parallel in flasks for the same 40 days, make the cell side longitudinal and add SA-β-gal positivity rate, p16, p21, and population doubling number, with a minimum of 3 donors × 2–3 independent runs.
And if the discussion is to be carried through to where this paper makes no claims, the next experiments are required. Administer the Phase 1 and Phase 3 sEVs matched by particle number and compare function. The sharpest of them would be a “senescence transmission assay,” giving late-phase sEVs to young MSCs and seeing whether SA-β-gal or proliferative capacity moves. That sEVs from senescent cells transmit senescence cell-non-autonomously has already been shown by studies using cultured cells (Borghesan M, et al., Cell Rep 2019;27(13):3956–3971.e6). The paper states that the sEV pellets were stored at −80 °C until analysis. If the remaining amount and their condition permit, these are experiments the authors themselves could run starting now.
Dr. Exotaro’s Perspective
I myself have continued to work on applying MSC-derived sEVs to the treatment of spinal cord injury (SCI) and neurological disease. Make them, measure them, put them into animals. What this paper deals with is that “make them, measure them” part, and this time it did not make for comfortable reading.
When preparing sEVs for animal experiments, I have been making an assumption: that a preparation harvested at the end of one manufacturing run is the same product as one harvested at the beginning. I do not have data that verify that assumption. This paper is saying, “that assumption is not safe.”
In a study we reported earlier, in a rodent model of severe spinal cord injury, the intravenously administered MSCs themselves never reached the injury site, yet the sEVs the MSCs released did migrate to the injury site and associated specifically with the M2 macrophages that work on tissue repair (Nakazaki, M., et al., Journal of Extracellular Vesicles 2021;10(11):e12137). The crux is that the same total amount worked when it was delivered divided over 3 days. In a study reported in Neurotherapeutics, an identical total amount delivered over 6 days gave greater motor recovery when infused continuously by osmotic pump than when split into daily injections (Nakazaki, M., et al., Neurotherapeutics 2026;23:e00892). How you deliver it decides the outcome. So where did those 3 days’ worth, those 6 days’ worth of sEVs come from? They were harvested through a single manufacturing run.
More uncomfortable still is the list of characteristics I wrote myself. The sEV requirements I recorded in my papers were particle size (70–150 nm, mean about 130 nm), positivity for the tetraspanins (CD63, CD9) and Alix, negativity for Calnexin, and particle number and protein amount. There is not one line in that list that records a time axis. And what did not move over about 40 days in the present paper is precisely that kind of item. What was moving was the proteome, the part we do not normally look at. In routine quality checks, this change is in principle uncatchable. That is not negligence; it follows from how the items were designed. The field still has no consensus on a potency assay for EV preparations, and this paper does not provide one either. To put it without rounding off: on the cell side, viability after replating and senescence findings from staining and morphology were measured. What was not measured is the potency of the sEVs those cells released. The correct summary is not “they stop working” but “a change that does not show up where we are looking was genuinely occurring.”
What can be done starting tomorrow is unglamorous. Record the harvest date in the lot record. Decide on a functional assay and run it along the time axis. Earlier in this series I wrote, “Is your control really the right control?” Today is its companion piece—is your lot really the same lot? It is a question I myself have yet to answer. This study on its own rewrites no standard, but it has placed before us, in actual measurements, the possibility that something may be moving in the place where the items we routinely measure fall silent. As one of those who sits in the maker’s chair, I will keep this paper close at hand. Not as an answer, but as an invoice.
Note: this article was written with the assistance of AI.