When you hear “exosome therapy,” the picture that comes to mind is probably one of gathering the sacs that cells have released and administering them to a patient. Yet the subject of today’s review is not the sacs. It is about rewriting the genes of the cells that make the sacs—a story about reworking the factory’s blueprint rather than the product. The target is insulin resistance (IR). Skeletal muscle, fat, and liver stop answering insulin’s orders, and this becomes the foundation of type 2 diabetes and fatty liver. The troublesome part is that it produces no symptoms for a long time. While the pancreatic β cells (the cells that make insulin) compensate by increasing their secretion, the blood glucose on a health check stays normal—and the β cells, having gone on raising their voices for years, eventually wear out.
What we are reading today is a paper published in the Journal of Cellular and Molecular Medicine by a team at the 940th Hospital of the Joint Logistics Support Force of the Chinese People’s Liberation Army (Lanzhou, Gansu Province). Into the exosomes released by mesenchymal stem cells (MSCs), they would raise the microRNAs (miRNAs) miR-21 and miR-3075, load Sirtuin-3 (SIRT3), the enzyme that protects mitochondria, and erase miR-29b-3p—four moves aimed at restarting insulin signaling and rebuilding mitochondria at the same time. This is “a shift from symptomatic treatment to precise disease-modifying therapy,” the authors write.
—But there are three premises. This is a review (a narrative review), and it contains not a single experiment the authors themselves ran. Across the 14 pages of main text, not one p-value, percentage, effect size, dose, or n is printed (the only figures that even look like numbers are “30–150 nm,” “~22 nucleotides,” “4-week HFD,” and “As of 2025”). And not one clinical trial in which MSC exosomes were given to humans as a treatment for insulin resistance is cited. What is more, this is not an empty lot. Metformin reduced all-cause mortality by 36% over 10.7 years of follow-up (95% CI 9–55, p=0.011; UK Prospective Diabetes Study Group, Lancet 1998;352(9131):854–865), and a dietary weight-loss programme led 46% into remission at 12 months (4% in the control group; Lean MEJ, et al., Lancet 2018;391(10120):541–551). Both of these are numbers this article has brought in from outside the review; neither appears even once in its text.
Let me settle two pieces of terminology. A “gene-modified exosome” does not mean that the DNA of the vesicle itself has been rewritten. Exosomes have no nucleus; what gets rewritten is the parent cell, the MSC. Second, this article writes “exosome” in line with the paper, and writes “vesicle” when referring to the particles as a whole regardless of origin—but MISEV2023 recommends that particles not demonstrated to be of endosomal origin be called “small extracellular vesicles” (sEVs) (Welsh JA, et al., Journal of Extracellular Vesicles 2024;13(2):e12404). This distinction will matter later.
Publication Details
- Paper title: Therapeutic Potential of Gene-Modified MSC-Exos in Insulin Resistance: Mitochondrial Crosstalk and miRNA Regulation
- Paper type: A review (narrative review). None of the authors’ own experiments or primary data are included. It is not a systematic review either. One figure, four tables, 125 references
- Authors: Qinghua Jia (first author; manuscript writing; funding acquisition), Bing Li, Yuan Xing, Yuanyuan Zhang, Baodong Gao, Xiangying Li, and Xiaoqin Ha (corresponding author)—seven in all
- Affiliation: Department of Clinical Laboratory, the 940th Hospital of the Joint Logistics Support Force of the Chinese People’s Liberation Army (Lanzhou, Gansu Province) / Gansu Provincial Key Laboratory of Stem Cells and Gene Drugs. All of them belong to both
- Journal: Journal of Cellular and Molecular Medicine (published by Wiley), 2026, Volume 30, Issue 17, Article No. e71342
- DOI / Link: 10.1111/jcmm.71342
- Peer review and publication dates: Received May 25, 2026; accepted August 20 (87 days)
- Open access: Yes (CC BY 4.0)
- Impact Factor: 4.7 (2025 Journal Citation Reports, Q2). The 2024 edition gave 4.2, and the 2022 edition 5.3
- About the journal: A fully open-access journal indexed in MEDLINE, Web of Science, Scopus, and DOAJ, and neither “predatory” nor a de-indexed title. That said, of the 9,506 papers it has in PubMed, 73 have been retracted and 5 carry expressions of concern, and 73% of the retractions are concentrated in 2024 or later (machine tally as of September 9, 2026). It is a mid-tier journal in the middle of putting out the fires from a period when it expanded its publication volume rapidly. This paper itself is not among them
- Funding: Lanzhou Science and Technology Plan Project (2023-ZD-166) and Gansu Provincial Health Industry Scientific Research Plan Project (GSWSKY2023-06). Both are at the municipal or provincial level, and the one who obtained them was the first author, not the corresponding author
- Conflicts of interest: A single sentence only—“the authors declare no conflict of interest.” No mention of patents
- Use of generative AI: Not stated in the paper
About corresponding author Xiaoqin Ha: All the paper records is the assigned role (project administration, manuscript review and editing); no position or title is stated. What follows from here is checking done outside the review, against PubMed. Narrowing by affiliation, three pillars can be traced—tissue repair with gene-modified MSCs, gene delivery using attenuated Salmonella, and clinical laboratory biomarkers for type 2 diabetes (Ha X, et al., Journal of International Medical Research 2021;49(11):3000605211048293). Representative of the first is an experimental study that treated myocardial ischemia with bone marrow MSCs made to overexpress hepatocyte growth factor (not a human clinical trial; Duan HF, et al., Molecular Therapy 2003;8(3):467–474), cited 172 times. “MSCs with a gene put into them” has been a theme for more than 20 years. The first author, too, has work on the miR-17-92 cluster in chronic myeloid leukemia (Jia Q, et al., Biochemical and Biophysical Research Communications 2017;487(4):868–874), so there is hands-on experience in miRNA biology. However, the same search turned up no primary research with “MSC-derived exosomes” as its subject (because the reference list abbreviates author names with “et al.,” whether there is self-citation cannot be settled from this review either). This is not material for declaring that they lack experience, but a record of having run this technology with their own hands cannot be read off this paper, at least.
What Wasn’t Understood Before?
When insulin binds to its receptor, the receptor itself is phosphorylated; next the tyrosine residues of the insulin receptor substrate (IRS) are phosphorylated, and PI3K (phosphoinositide 3-kinase) and Akt start moving in turn. Then the sugar transporter GLUT4 is carried from storage vesicles inside the cell (a different thing from exosomes—sacs that work only within the cell) to the cell membrane, where it becomes the entrance for sugar. IRS1→PI3K→Akt→GLUT4—this is the chain the paper uses as its backbone. In resistance, tyrosine phosphorylation decreases and serine phosphorylation becomes excessive in its place. The staff member is not broken, but a sticky note of a different color has been put on them and the message cannot be passed along. The culprits are stress kinases such as JNK and PKC; the background is chronic inflammation and lipotoxicity. What matters is that the sticky note can be peeled off. That is exactly why exercise and weight loss improve resistance.
How is insulin resistance measured? The convenient method is HOMA-IR (homeostasis model assessment of insulin resistance)—an index calculated from fasting blood glucose and insulin. The reference method is the hyperinsulinemic-euglycemic clamp—insulin is infused while blood glucose is held constant, and the amount of glucose required is read off. Neither appears even once in this review.
The other axis is mitochondria. Run a power plant and electrons leak, throwing off sparks called reactive oxygen species (ROS). So a maintenance crew is needed. Fusion and fission cordon off the bad parts, new power plants are built (biogenesis, with PGC-1α/NRF1/TFAM as the command center), and what cannot be repaired is bagged up and degraded. This is mitophagy. In resistant skeletal muscle, oxidative capacity falls, the lipid intermediates that cannot be fully burned activate PKC and JNK, and those extra serine sticky notes get applied. Mitophagy stalls, biogenesis drops, ATP runs dry, and that in turn breaks signaling—the state the paper calls a “self-amplifying loop.” The review writes this as a causal vicious cycle. But it is not a description that demonstrates which comes first.
👦 Student: Does it really matter that much which comes first? It’s a vicious cycle either way, isn’t it?
🧬 Dr. Exotaro: Try replacing the power plant with a restaurant kitchen. In a restaurant whose sales are falling, the stove and the oven in the kitchen are both worn out. Did sales fall because the kitchen is old, or was the kitchen never renewed because sales fell? One photograph cannot tell you. But the treatment strategy is decided by the direction of the arrow. If the mitochondria are the “result,” then making the kitchen brand new will not bring sales back. This paper is built on top of that bet.
This controversy has gone nearly 40 years without being settled. “Whether changes in mitochondrial content, respiratory function and substrate oxidation are central to the development of insulin resistance, or occur as a response, remains unclear” (Genders AJ, et al., International Journal of Molecular Sciences 2020;21(18):6948). The review writes from the causal side without touching this fork in the road.
Why the sacs rather than the cells? The premise is that “the substance of MSCs’ therapeutic effect is not the cells turning into tissue, but their secretions.” Almost no risk of tumorigenicity (the property of forming tumors), low immunogenicity (the ease of being flagged by the immune system), easy storage and standardization, and the possibility of becoming an “off-the-shelf” preparation—these are the advantages the review lines up by citing other papers. Which advantage has been demonstrated and to what extent, however, cannot be verified from this review. Moreover, “easy to standardize” is the very item the same paper lists in Chapter 7 as a manufacturing bottleneck. The biogenesis side is solid. The membrane of an endosome—formed when the cell membrane indents—buds inward; then the multivesicular body (MVB), where those vesicles have accumulated, fuses with the cell membrane and releases its contents. These are exosomes. But there are also, separately, vesicles that pinch off directly from the cell membrane, and there is still no way to tell with certainty, after the fact, where collected particles came from. The MISEV2023 recommendation rests on this one point.
What Did This Paper Find?
Everything below consists of claims the authors gathered from other people’s papers. The evidence, in the review’s own words, is “mainly derived from short-term high-fat-diet or streptozotocin-induced rodent models, with relatively small sample sizes.” Streptozotocin is a method of destroying pancreatic β cells with a drug; it is not a model of insulin resistance itself. The five rows of Table 1 are two studies of type 2 diabetes in rats, two of high-fat-diet obese mice, and one of cultured human adipocytes. Human clinical trials: zero—and the effect column is a string of arrows. There is not a single measured value of insulin sensitivity itself.
There Are Two Places to Rewrite—The Factory, or the Lunchbox Already Shipped
Table 4 divides the modification methods into three. Two rewrite the cargo, and one presents a targeting ligand on the surface—“surface display modification” (expressing a fusion protein with Lamp2b or CD63 in the parent cell to aim at the pancreas or muscle). But there are two places where the hand goes in. Indirect modification (engineering of the parent cell) puts a gene into the MSC with a lentivirus or a plasmid and waits for the vesicles that emerge to inherit the cargo. Direct modification (manipulation of the vesicles after isolation) opens pores in already-harvested vesicles with electrical pulses and pushes nucleic acid in. In lunchbox-factory terms, the former is “adding ‘triple the fried chicken’ to the recipe,” the latter “making a hole in the lid of the finished lunchbox and injecting the filling.” The table is honest about the drawbacks too, writing that viral methods are prone to causing insertional mutagenesis and can leave virus contaminating the final product, that post-isolation electroporation can destroy membrane integrity and induce aggregation, and that chemical transfection makes it difficult to separate the reagent-vesicle complexes.
👦 Student: Injecting into a lunchbox that’s already been shipped sounds better, though—you could put in exactly the amount you aimed for.
🧬 Dr. Exotaro: As an intuition, that’s right. But in practice, open a hole in the lid and the lunchbox breaks, and the lunchboxes stick to each other into a lump. There’s another pitfall, too: the nucleic acid you pushed in aggregates and merely “looks as if it were loaded.” That’s why this field has chosen the roundabout route—“rewriting the factory’s blueprint”—as the mainstream.
What to Write In—Four Modifications Whose Evidence Is Not of Equal Strength
The abstract lists the four side by side, but the strength of the evidence differs. The most coherent is the proposal to erase miR-29b-3p. It is enriched in exosomes derived from senescent bone marrow MSCs and, by targeting SIRT1—the enzyme that removes acetyl groups from metabolism-related proteins to switch their function—worsens metabolism. The foundation is primary research reporting that inhibiting it made the harmful effects disappear and restored insulin sensitivity and glucose uptake (the review itself does not write the animal species; Su T, et al., ACS Nano 2019;13(2):2450–2462).
The implementation is concrete too. It writes that antagonists (artificial nucleic acids such as antagomirs, which bind the target miRNA and stop it working), siRNA (short double-stranded RNA that destroys mRNA of a targeted sequence), or CRISPR interference would suppress expression from the MSC stage onward, reducing the amount loaded into the exosomes in the first place. Still, since miR-29b-3p is enriched in “senescent” MSCs, wouldn’t it be enough—before touching any genes—simply to choose young donors and set an upper limit on passage number (how many times the cells have been re-plated)? The review writes down that observation without connecting it to the manufacturing chapter.
miR-3075 is precarious. It is enriched in hepatocyte-derived exosomes from “early obese” mice given a 4-week high-fat diet, and targets fatty acid 2-hydroxylase (FA2H) (Ji Y, et al., Nature Metabolism 2021;3(9):1163–1174). The review itself writes that direct evidence that this is present at high concentration in MSC-derived exosomes is limited, and the original study writes that in chronic obesity this compensation is lost and hepatocyte exosomes instead promote resistance. Fixing in the direction of increase a molecule whose sign flips with disease stage is a bet.
SIRT3 is of a different stripe. It is a deacetylase inside the mitochondria that assists the respiratory chain, reduces ROS, and promotes PGC-1α-mediated biogenesis. The idea is to send in the maintenance crew itself—but the data showing that “modified exosomes loaded with SIRT3 worked” come from a myocardial ischemia-reperfusion model (Yang J, et al., Journal of Nanobiotechnology 2025;23(1):439), and the review does not even write the animal species.
And then miR-21. The review treats it as “the hero that protects β cells” (Chen J, et al., Stem Cell Research & Therapy 2020;11(1):97), but the explanation is strange. The paper writes that “miR-21 suppresses pro-apoptotic pathways such as PTEN/PI3K/AKT, p38, and MAPK.” PI3K/Akt is the very thing this review has spent pages calling “the pathway of survival and glucose uptake.” What miR-21 suppresses is PTEN, and if PTEN decreases, PI3K/Akt is if anything activated. On top of that, in cancer research miR-21 is a textbook cancer-promoting miRNA (oncomiR). One of the references the review cites is a study in which miR-21 promotes the proliferation of glioma cells through the same pathway (Chai C, et al., CNS Neuroscience & Therapeutics 2018;24:369–380). Therapeutic effect on one side, carcinogenic risk on the other. The same pathway, the same story. Note also that for the last molecule Table 2 lists, miR-210, the table’s evidence concerns miR-210-5p (target SIDT2), while another section cites miR-210-3p, which is on the side that suppresses GLUT4—and both the main text and the table write only “miR-210,” never distinguishing the two. Since both point in the direction of lowering sensitivity, it cannot be said that the effects are opposite; but as long as two different molecules with different sequences are handled under one name, which one should be increased cannot be decided. In Table 2 as well, two of the four molecules are hepatocyte-derived and adipose-tissue-macrophage-derived—the table’s title says “exosomes from various sources,” but the section heading says “key miRNAs of MSC exosomes.” Readers have to sort out for themselves which is demonstrated for MSC-derived material and which is knowledge from other cell sources.
Counting the Letters—How Many Are Inside One Sac
miRNAs are short RNAs of about 22 bases that bind to the 3′ untranslated region (3′UTR) of messenger RNA (mRNA) and either cause its degradation or suppress its translation. They tighten the production-volume line of the recipe a little—not a switch but a dial.
👦 Student: Because there are letters inside the sac, the gene expression of the cell that receives it changes. That was the explanation, wasn’t it?
🧬 Dr. Exotaro: In 2014, a research team counted both the number of sacs and the number of letters. The answer was the same for every source—plasma, semen, dendritic cells, and others. Even for the most abundant miRNA, it came to an average of 0.00825 ± 0.02 molecules per exosome (the mean ± SD the original reported across six exosome sources). You open more than 100 sacs before you may or may not find one letter. That is the order of magnitude we are talking about.
The source is Chevillet JR, et al., Proceedings of the National Academy of Sciences of the United States of America 2014;111(41):14888–14893. This is not proof that “exosomes don’t work.” The original itself raises the possibilities that only some “winning” vesicles carry a large load, that the numbers may be sufficient anyway, and that what is working may be membrane lipids or surface proteins rather than miRNA. The problem is that a review with “miRNA regulation” in its title never once puts a number on how many copies could be loaded per vesicle. Unless how many molecules there are before and after modification is written down, “amplified” cannot be checked by arithmetic. The review also uses the literature for the competing hypothesis that “MSC exosomes work through protein-based mechanisms” (Toh WS, et al., Biochemical Society Transactions 2018;46:843–853) only to list surface markers.
The “Mitochondrial Crosstalk” Signboard, and an Address That Is Hard to Reach
The introduction and Table 3 put the phenomenon in which MSCs transfer intact, functional mitochondria to damaged cells on the first line of the therapeutic mechanism, and the abstract writes that it “can rescue mitophagy and respiration.” But the sizes do not fit. The review itself defines exosomes as “30–150 nm,” while the diameter of a mitochondrion is generally 500–1,000 nm (this is a figure from outside, not written in the review). The truck bed is an order of magnitude smaller than the load.
👦 Student: Could the authors really not have noticed something that obvious?
🧬 Dr. Exotaro: They noticed. Section 5.3 of the same paper says this—“direct transfer of intact mitochondria by exosomes remains mechanistically debated, and much of the evidence supports miRNA/protein-mediated regulation rather than delivery of whole organelles.” The routes the review lists for handing over a whole organelle are three: tunneling nanotubes (TNTs, the arrangement in which cells connect by a thin tube and pass contents across), gap junctions, and extracellular vesicles. The one that got the caveat attached is the exosome line. There are reports of vesicles carrying mitochondrial fragments or mitochondrial DNA (mtDNA), so you cannot say it is “absolutely impossible” either. The problem is that elsewhere in the same paper there is a table presenting it as if it were established fact. And the references attached to the first and second rows of that table all look, from their titles, to be mechanistic overviews.
The core evidence is beside the point as well. The grounds on which the review concludes that it “validates the therapeutic potential of targeting mitochondrial rescue” come from a study of pancreatic α cells, which put out glucagon—the hormone on the side that raises blood glucose—and moreover what was administered was MSCs as cells, not exosomes (a mouse model of type 2 diabetes; Song J, et al., Stem Cells Translational Medicine 2024;13:776–790). The review writes for itself that “biodistribution analyses show a strong tendency toward systemic clearance via the liver and spleen, whereas bioavailability and retention in skeletal muscle and pancreatic islets are limited.” It does not write that they are zero—but most glucose disposal happens in skeletal muscle, and the main battlefield is an organ that is hard to reach. The single figure (Figure 1) is a conceptual diagram of the therapeutic mechanism (Schematic illustration), not a figure showing that simultaneous delivery has been demonstrated. The confession in the text is far more cautious than the figure.
How Will the Future Change? (The Path to the Clinic)
The evidence for therapeutic effect that this review cites is, in the review’s own words, “mostly short-term rodent models,” with cultured cells for the rest. Not one clinical trial registration number (NCT number) appears. The authors themselves write in the conclusion that it “remains at an early preclinical stage.” The only sentence that touches the current clinical situation is “as of 2025 … no large-scale trials targeting systemic insulin resistance exist”—a paper accepted in August 2026 saying “as of 2025,” with a review in the field of plastic surgery as its source.
The numbers from here on are also from outside the review. As of September 9, 2026, there are only two clinical trials of MSC exosomes in the world targeting type 2 diabetes—a Phase I/II at Peking University First Hospital (NCT07693036, recruiting, 66 patients planned, intravenous infusion once weekly for 8 weeks, with change in HbA1c (hemoglobin A1c; the average blood glucose over the past 1–2 months) at 8 weeks as the primary endpoint) and a Phase II at the University of Indonesia (NCT07144241, not yet recruiting; a design in which the control group also receives exosomes). Human trials of gene-modified exosomes in metabolic disease: zero. By contrast, MSCs as “cells” do have a double-blind, placebo-controlled Phase II: in the group given three intravenous infusions of umbilical cord-derived MSCs at 4-week intervals, HbA1c fell from 9.02 ± 1.27% to 7.52 ± 1.07% (placebo group 8.89 ± 1.11% → 8.19 ± 1.02%, between-group p = 0.0081; Zang L, et al., Stem Cell Research & Therapy 2022;13(1):180). Even so, the rate at which the primary endpoint was achieved was 20.00% in the MSC group (4.55% for placebo), and the original states explicitly that “there was no improvement in β-cell function in either group.” This was a trial of cells, and it did not test β-cell protection by the exosomes the review speaks of. And yet—putting in whole cells did not move β-cell function, and β-cell protection by exosomes has still not been tested in humans. In December 2019, the US FDA, responding to serious adverse events from unapproved exosome products, also gave notice that “there are no FDA-approved exosome products at this time.”
👦 Student: In the end, what would this treatment have to achieve to count as a “win”?
🧬 Dr. Exotaro: Defining that was, in truth, this review’s job. Added on top of standard of care, it would have to move HbA1c, or liver fat content, or a clamp-derived index; the effect would have to last more than a few weeks; and anti-drug antibodies (ADAs—antibodies the body makes against an administered drug) would have to not become a problem when it is given for years. The authors’ own proposal is sound—“narrow it to prediabetes, or to early type 2 diabetes with mitochondrial dysfunction.” But which index, and at what threshold, that patient population would be selected by is not written. Since they themselves admit in another section that “pharmacodynamic biomarkers to identify responders in advance are not established,” the one step that would bring the proposal down into trial inclusion criteria is still blank.
How to Critically Read This Research (Limitations and Paths to Improve It)
Before the criticism, the virtues. The choice of subject—there is value in pointing the engineering modification of extracellular vesicles (EVs) at systemic insulin resistance itself. The fact that a single mechanistic axis runs through it is kind to the reader too. That at the abstract stage they named GMP manufacturing scalability, the long-term safety of modified products, and advanced therapy medicinal product (ATMP) regulation as “walls not yet cleared” is a candor that does not close the story with “it is promising.” And the fact that they embedded the inconvenient facts in the main text—“direct transfer of intact mitochondria is mechanistically debated,” “many of the reported therapeutic effects are based on correlation,” “delivery to skeletal muscle and pancreatic islets is limited”—all three of these sentences are material that works against their own claim.
The limitations the authors themselves acknowledge are comprehensive as well: (1) most preclinical findings come from short-term rodent models with small n; (2) MSC source, isolation method, and dosing regimen are heterogeneous; (3) therapeutic effect in humans is unproven; (4) the anti-inflammatory findings come from acute models; (5) the evidence for mitochondrial improvement is largely indirect; (6) most of the SIRT3 data come from acute injury models; (7) the modified constructs have never been tested in large animals; (8) pharmacodynamic biomarkers are not established; (9) comparative data against standard drugs are lacking; (10) the gene-modification process itself generates variability in transduction efficiency and loading, making it hard to produce preparations of uniform specification; and (11) there is no validated assay for measuring efficacy in the IR setting.
On the other hand, there are important problems the paper does not touch on.
Within the same paper, the strength of assertion is not consistent. Section 7.4 writes that “much of the reported therapeutic effect relies on evidence based on correlation rather than causation.” Yet the conclusion says that “at the preclinical stage, it has been proved that these vesicles … reactivate mitochondrial function through horizontal transfer.” What is horizontally transferred here is bioactive cargo, which is a different matter from the transfer of intact mitochondria that Section 5.3 qualified—so this is not a self-contradiction. Even so, a story the text admitted “stops at correlation” becomes “proved” in the conclusion. Readers who stop at the abstract or the conclusion take home a stronger conviction than the main text supports. The strategies listed change places as well: the abstract’s four become, in the conclusion, “enrichment of miR-21 or miR-146a and overexpression of SIRT3”—both are examples, but they are not the same four.
Not a single piece of data quantitatively compared against standard of care is laid out. What the introduction names as existing treatments are metformin and thiazolidinediones (Exendin-4 and apelin also appear, but not as objects of comparison—as cargo to be loaded into modified MSCs). SGLT2 inhibitors, tirzepatide, and bariatric surgery do not appear as a single word in the full text, and lifestyle intervention is not discussed as a comparator either (the word “exercise” appears only in the titles of references). On top of that, existing drugs are called “single-target” and “limited to symptom management,” and the conclusion asserts that “standard interventions cannot correct this complexity.” At the very least, “single-target” is contrary to fact. Metformin acts on multiple pathways—mitochondrial complex I, AMPK, and the gut microbiota (Rena G, et al., Diabetologia 2017;60(9):1577–1585)—and pioglitazone is a direct insulin sensitizer acting through PPARγ. Moreover, the basis for “existing drugs cannot reach mitochondrial impairment” was a paper discussing the pathophysiology of gestational diabetes. One number from outside the review. In 3,234 people with impaired glucose tolerance, lifestyle modification reduced the onset of diabetes by 58% and metformin by 31% (Diabetes Prevention Program Research Group, N Engl J Med 2002;346(6):393–403)—numbers for precisely the population this review names as its indication. It was the authors themselves who listed the absence of comparative data as a challenge.
Citation matching has broken down. Within the range checked, there are a number of places where the statement in the text and the title of the reference attached to it do not mesh. The heaviest is that the original title of the reference cited immediately after the proposal to “raise the expression of protective miR-21 and SIRT3” is “miR-21 antagomir improves insulin resistance and lipid metabolism disorder in type 2 diabetic rats” (Wang Y, et al., Annals of Palliative Medicine 2020;9:394–404)—a study in which reducing miR-21 was what improved things. The three references attached to the backbone section are likewise a study of the adipokine CTRP4, one of the polyphenol ferulic acid, and one in which macrophage-derived miR-210-3p suppresses GLUT4 and “breaks” sensitivity; not one is about MSC exosomes, and the third points the opposite way. Checked against Table 1, of the four references in the text only one matches in both number and content. One matches only in number with the content swapped, and the remaining two are a rat study of papaya extract (Roy JR, et al., Nutrients 2022;14:4181) and a rat study of MSC “cells” overexpressing apelin (Gao LR, et al., Stem Cell Research & Therapy 2018;9(1):339). This is not fabrication; it is disorder in citation management, verifiable by cross-checking.
And the sentence that “the modified versions consistently outperformed the unmodified ones” comes with no comparison attached to show it. Three of the four examples given are the very studies the same review lined up in Table 1 as representative examples of unmodified MSC exosomes. The remaining one is the miR-29b-3p study, which is a story about sensitivity recovering when the cargo is reduced. There is one citation said to compare directly against an unmodified product—a comparison against vesicles whose surfaces were decorated with PEG (a polymer that keeps them in the bloodstream longer) and an aptamer (Xia L, et al., International Journal of Nanomedicine 2024;19:9575–9596). That is surface modification, not gene modification. Nowhere is there a place where modification method, control, and effect size can be set side by side and compared. The cheapest way to measure what the modification adds is to line up vesicles with no gene put into them and dose them at the same particle number. Unless that is done, whether the modification worked cannot be separated out.
Safety specific to gene modification is handled in generic language. That Table 4 lists insertional mutagenesis and viral contamination deserves credit, but there is nothing beyond it. Karyotyping of the producer cell bank (checking for chromosomal abnormalities), integration site analysis (checking where in the genome the introduced gene landed), a passage limit, a specification for the upper limit of residual host cell DNA, replication-competent lentivirus (RCL) testing—not one of these is listed. They refer to ATMP regulation, yet the specific specification tests number zero. There is also no discussion of the point that artificial nucleic acids that mimic miRNA function (miRNA mimics) can be detected as double-stranded RNA by innate immune sensors (TLR7/8, RIG-I). In fact, a Phase I trial that tried systemic administration of a miR-34a mimic was discontinued for immune-related adverse events (Hong DS, et al., British Journal of Cancer 2020;122:1630–1637). The only off-target risk named is for SIRT3; miR-21 is not included.
👦 Student: “Erasing” sounds safer than “loading,” at least.
🧬 Dr. Exotaro: There is an oversight in the opposite direction there too. The review itself cites a paper saying that “pancreatic β cells secrete the miR-29 family on exosomes to regulate glucose homeostasis” (Li J, et al., Journal of Extracellular Vesicles 2021;10:e12055). The manipulation being proposed is to reduce the cargo on the MSC side; it does not erase miR-29 throughout the body. Even so, what happens in the recipient tissues when vesicles with this cargo removed are given for years has never once been examined. Even for the modification that looks most coherent, the order in which you confirm things cannot be skipped.
The handling of terminology is also a concern. The context in which the review writes that “the MISEV2023 criteria are insufficient” is that “they do not define efficacy assessment methods for the IR setting.” But MISEV is a minimal standard for the identification, characterization, and reporting of EVs; it is not a document that defines how to measure efficacy disease by disease. So they have judged it “insufficient” on grounds outside its remit. The review itself explains exosomes by both size (30–150 nm) and origin—“formed via the endosomal pathway”—but does not show which of the studies lined up in its tables demonstrated origin. In fact, a study that the original authors carefully called “small extracellular vesicle” has been rewritten as “Exos” inside the table (Yap SK, et al., Pharmaceutics 2022;14:649). Items such as negative markers and the particle-count-to-protein-amount ratio are not raised either, and the MISEV2023 source itself (Welsh JA, et al., Journal of Extracellular Vesicles 2024;13(2):e12404) never appears once in the list of 125 references. Whether the individual cited studies conform to MISEV cannot be judged from this review.
So how could the quality have been raised? Ten items, in order of least effort. Redo the full cross-check of citations (no additional experiments are needed). Disclose the search strategy and the period covered, and declare that “this is a narrative review.” Rebuild Table 1 as a quantitative evidence table—add columns for dose (particle number or protein amount), number of administrations, species and n, the method used to measure insulin sensitivity, and type of control. Add a “level of target validation” column to Table 2—was binding confirmed with a 3′UTR reporter, was the phenotype reproduced, or is it only a correlation of expression? With that one column, the ordering becomes visible: the well-evidenced miR-29b-3p first, and miR-21 only after target validation and an assessment of carcinogenic risk.
Continuing. Deliberately search for negative studies (the words “publication bias” and “negative result” do not appear in the full text). Put in one comparison table against standard of care, and define a target product profile. Estimate the stoichiometry and the required amounts to the right order of magnitude on the back of an envelope—how many copies ride on one vesicle, and how many particles are needed for one dose in one person. Rewrite the safety section to the level of ATMP specification testing. Rewrite the mitochondria chapter starting from the question of size—add one line saying “an organelle an order of magnitude larger does not fit as-is into a 30–150 nm vesicle,” and the caveat in Section 5.3 arrives as a caveat whose reason you can see. Finally, write the skeleton of a Phase I trial in one paragraph—“in prediabetic people with high liver fat, added on top of standard of care. Primary endpoint: change in MRI-PDFF (liver fat content measured by MRI) at 12 weeks; secondary endpoints: clamp-derived M value, HOMA-IR, and the positivity rate for anti-drug antibodies. Include repeat dosing and tumor surveillance.” Turning the weakness they wrote down themselves—“it ends up collecting in the liver and spleen”—into the rationale for the first choice of indication: that inversion was something only the authors could have written.
Dr. Exotaro’s Perspective
To be honest, metabolic disease is not my own field of expertise. What I have long worked on is how to use MSC-derived small extracellular vesicles (sEVs) in neurological diseases such as spinal cord injury (SCI), and I am not involved in diabetes care either. Even so, there were three points I could not read as somebody else’s problem.
First, the way it is administered. In our research, the MSCs themselves, given intravenously, did not reach the injury site, yet the sEVs the MSCs released did migrate to the injury site and associated specifically with M2 macrophages. And—even at the same total amount, a single large dose did not work; only when it was divided over 3 days was motor function recovery equivalent to MSC cells obtained (Nakazaki, M., et al., Journal of Extracellular Vesicles 2021;10(11):e12137—a rodent model of severe spinal cord injury). What decided the outcome was not “how much was given” but “in what temporal shape the exposure was delivered.” This review does touch on the heterogeneity of dosing regimens and on the risks of repeat administration. But while dealing with a chronic disease in which repeat dosing over years is the premise, there is no description comparing at what interval and how many times to dose.
Second, the observation that the unmodified vesicle is itself the drug. What I used was natural MSC-sEVs, with no gene modification and no drug loading. In the study loading the IL-10 gene onto extracellular vesicles that I read earlier in this series, “ordinary vesicles” with no gene put into them also plainly calmed inflammation, and in primary human macrophages they were equivalent to the IL-10-loaded vesicles except on some readouts. So every time I read that “gene modification amplifies the effect,” I want to check—did you line up unmodified vesicles at the same particle number and compare?
Third, nobody is counting “what is in one dose.” EV papers that can answer for copy number per vesicle and particle number per dose are, my own included, almost nonexistent. Even so, if we are going to bring an intravenous biologic into a chronic disease, we cannot run away from the work of counting what is loaded and how much of it there is. Without a potency specification, you cannot make a product.
I will draw a line too. My target is the injured central nervous system, and I take advantage of a state in which the barrier has been breached and the entrance is open. In metabolic organs with an intact barrier, “reaching” means something different. And I have never once done any of it—enhancing miR-21, silencing miR-29b-3p, or loading SIRT3. On the long-term safety of modified products, I stand in the same place as the reader.
Even so, I will not end by running it down. The proposal to “erase” miR-29b-3p is the most coherent part. If senescent MSCs are concentrating harmful cargo, cutting it off makes sense, and it connects directly to manufacturing specifications you could write tomorrow: a criterion for donor age and an upper limit on passage number. What remains is whether they can write it in as actual dimensions and specifications, not skip validation in large animals, and push it to the point where it can be measured added on top of the best existing treatment. This paper lines up a repair plan. While still not having climbed onto the roof.
Note: this article was written with the assistance of AI.