The mRNA vaccine's 'booster' was not the contents but the container — the ionizable lipid hits TLR4 and sets off innate immunity

Vaccines normally contain an adjuvant, an agent added to boost the immune response. Antigen on its own does not make the immune system take things seriously, so you add a substance that tells innate immunity “a foreign body has arrived here.” Aluminium salts, and molecules that mimic bacterial components, are what get used.

And yet mRNA vaccines contain no adjuvant. The ingredient list holds nothing but mRNA and lipids. Even so, they raise strong antibody and cell-mediated immunity. Which means that this vaccine must be hiding the work of an adjuvant somewhere inside itself.

Where?

The obvious answer is the mRNA. RNA is, by nature, a molecule that gets sensed as foreign. But the story runs the other way: the mRNA in current vaccines carries a modification called N1-methylpseudouridine that deliberately makes it harder to sense, and on top of that it is purified to remove the double-stranded RNA formed as a by-product. A molecule engineered not to be sensed makes no sense as the lead actor among boosters.

What I am taking up today is a 2025 npj Vaccines paper that interrogated the remaining suspect — the lipid nanoparticle (LNP) itself. Using a human monocyte cell line, the authors asked whether “empty LNPs” carrying no mRNA could switch innate immunity on all by themselves.

Let me give the conclusion first. They could. And adding mRNA did not increase the response. What held the key was one of the four lipid components — the ionizable lipid — together with TLR4, a receptor whose supposed job is to watch for bacterial toxins.

Let me say this firmly up front: this paper is not an argument that mRNA vaccines are dangerous. Innate immunity coming up is itself the mechanism by which a vaccine works. What this paper is aiming at is to pin that mechanism down and connect it to designs that keep the efficacy while reducing reactogenicity. I set aside a section for this in the middle of the article.

One more thing: this article is the companion piece to “What happens when you count the lipid nanoparticles that carry mRNA one particle at a time”, published the same day. That one is about how most LNPs carry no cargo; this one is about how those empty particles raise immunity. Read together, they get a notch more interesting.

Journal information

• Paper title: Ionizable lipid nanoparticles of mRNA vaccines elicit NF-κB and IRF responses through toll-like receptor 4
• Type of paper: original research article (vaccinology and innate immunity)
• Journal: npj Vaccines 2025, volume 10, article number 73 (received 2 May 2024, accepted 24 March 2025). An open-access journal in the Nature family, published in partnership with the Sealy Institute for Vaccine Sciences at the University of Texas Medical Branch at Galveston (UTMB)
• DOI / link: 10.1038/s41541-025-01124-x
• License: open access (CC BY 4.0). There is also a note invoking the provision on works prepared by US government employees in the course of their duties (Title 17, Section 105 of the United States Code)
• Impact Factor: roughly 7 or thereabouts (it moves year to year; among specialist vaccinology journals this puts it near the top)
• Authors: eight in all — Amanda E. Zelkoski (first author), Zhongyan Lu, Gauthaman Sukumar, Clifton Dalgard, Hooda Said, Mohamad-Gabriel Alameh, Edward Mitre and Allison M. W. Malloy (corresponding author)
• Affiliations: the departments of Pediatrics, of Anatomy, Physiology and Genetics, and of Microbiology and Immunology at the Uniformed Services University of the Health Sciences (USU, Bethesda, Maryland); the Henry M. Jackson Foundation; the Department of Pathology at the Children’s Hospital of Philadelphia (CHOP); and the Department of Pathology at the University of Pennsylvania together with the Penn Institute for RNA Innovation
• Funding: a RAMP grant from the Department of Pediatrics, plus funds from the US Department of Defense and the Defense Health Agency (Defense Health Program grants HU00012020067 and HU00012120094, and Immunization Healthcare Branch grant HU00012120104). The paper states explicitly that “the funders had no role in the study design, in the collection, analysis or interpretation of the data, in the writing of the report, or in the decision to submit”
• Conflict-of-interest disclosure: the authors declare no commercial or financial conflicts of interest
• Data availability: the RNA sequencing data are deposited in the NCBI Sequence Read Archive as PRJNA1196969. Other data are available on request

What kind of researcher is the corresponding author?

The corresponding author, Allison M. W. Malloy, is associate professor of pediatrics at the F. Edward Hébert School of Medicine of the Uniformed Services University, where she heads the Laboratory of Infectious Diseases and Host Defense. She is at the same time a clinician in pediatric infectious diseases at the Walter Reed National Military Medical Center. She has been on the faculty of that school of medicine since 2013; before that she was a clinical research fellow at the Vaccine Research Center of the National Institute of Allergy and Infectious Diseases (NIAID) at the National Institutes of Health (NIH).

Her research has run on one consistent axis: where the immune response to respiratory viruses works on the protective side, and where it works to make the disease worse. She has worked for a long time in particular on respiratory syncytial virus (RSV) and immunity in infants, building mouse models to study how mucosal immunity to RSV develops in the airway. According to public information, she has been involved in work showing that raising innate immune activation in neonatal mice, together with increasing exposure to antigen, produces lung-resident memory T cells and protects against RSV, and in prospective pediatric cohort studies following the development of mucosal and circulating immunity to respiratory viruses. Her collaborators include NIAID (the Vaccine Research Center, the Laboratory of Infectious Diseases and the Mucosal Immunobiology Section), the Children’s Hospital of Philadelphia and the Walter Reed Army Institute of Research.

🧬 Dr. Exotaro: Once you have this background in view, the paper’s aim comes into focus. A clinician who has spent her career on how you can reliably raise immunity in infants and the elderly with as little stimulation as possible turned around and asked what the adjuvant in this vaccine actually is — that is the kind of paper this is. The motive in front is not pure curiosity so much as wanting to know the mechanism in order to design with it.

There is one more person: a co-author without whom this study could not have been done. Mohamad-Gabriel Alameh (Children’s Hospital of Philadelphia, University of Pennsylvania, Penn Institute for RNA Innovation). According to the author-contributions section, he supplied the LNPs used in the experiments and contributed to the study design. In Immunity 2021, volume 54, page 2877, he published a paper titled “Lipid nanoparticles enhance the efficacy of mRNA and protein subunit vaccines by inducing robust T follicular helper cell and humoral responses”, and the present paper is its sequel. The University of Pennsylvania is the home ground of mRNA vaccine technology, and the shape of the thing is that that home ground is itself driving the research that says the real booster is the container. First author Zelkoski conceived and designed the study with Malloy, performed all the experiments and wrote the manuscript.

What was it that nobody knew until now?

The puzzle of “it works although there is no adjuvant”

The immunogenicity of mRNA vaccines has been demonstrated, but the paper opens by naming two problems. The duration is limited, and the cause of reactogenicity is not understood. Neither can be improved while it remains unknown by what mechanism immunity is being raised.

There are several candidate sensing routes. mRNA (single-stranded RNA) can be sensed by TLR7 and TLR8, and the double-stranded RNA formed as a by-product by TLR3, RIG-I and MDA5. But as noted above, the mRNA in current vaccines is built, through modification and purification, to avoid exactly these.

👦 Student: If it is built not to be sensed, why does immunity come up at all?

🧬 Dr. Exotaro: That is exactly the puzzle. Which is why the gaze turned to the container. And there were in fact clues. Several reports, in cultured cells and in mice, had found that innate immunity comes up with LNPs alone, carrying no mRNA. Give empty LNPs to human monocyte-derived dendritic cells and out come cytokines: IL-1, IL-6, IL-12, interferon α, interferon γ. The present paper cites a review and writes that “61% elicited an immune response” (though what the cited source is counting is published studies of lipid-based nanoparticles, not 61% of LNP formulations and not 61% of vaccine recipients. Here too the meaning changes once you go back to the source).

A further clue, going deeper, came from mouse work by co-author Alameh and colleagues. The adjuvant activity of empty ionizable LNPs depended in part on MyD88 — and since MyD88 is an adapter molecule that works downstream of TLRs and the IL-1 receptor, that suggests a TLR is involved. And in the same study, removing the ionizable lipid reduced cytokine production.

Why “empty LNPs stimulate” was not enough on its own

Up to this point, “empty LNPs stimulate immunity” was more or less established. What was missing was the identification of the route: which receptor it goes through, which transcription factor it moves, and which component is responsible.

And the earlier work disagreed with itself. There are reports that in mouse models there was no evidence that TLRs were involved. To sort out the contradiction, somebody had to break the receptors and the adapters one at a time, in human cells, and compare.

What do you have to measure to get an answer?

At the centre of the innate immune switch sit two families of transcription factor.

• NF-κB (nuclear factor-kappa B) — governs the production of inflammatory cytokines and chemokines
• IRF (interferon regulatory factor) — governs the production of type I interferon

When a TLR is stimulated, these two families are driven by way of an adapter — either MyD88 or TRIF.

What the authors used was THP-1-Dual, a human monocyte-like cell line. It is a cell carrying two reporters: when NF-κB moves, secreted alkaline phosphatase (SEAP) appears in the culture medium, and when IRF moves, a secreted luciferase appears. Take the medium, measure colour and light, and how far the switch has been thrown becomes a number.

👦 Student: Why monocytes?

🧬 Dr. Exotaro: Monocytes are front-line innate immune cells circulating in the blood, and they respond quickly to signals from pathogens and from injured cells. They are also the cells that gather at an injection site and present antigen to hand things on to T cells. As a stand-in for the first cell to meet a foreign body, it is a reasonable choice. But this is an immortalized cell line derived from a leukemia patient, not a cell from a living human being. I come back and criticize this point properly later on.

What did this paper find?

The four containers they prepared

The comparison is cleanly designed. They prepared four kinds of LNP.

All of them contain cholesterol, DSPC (the helper lipid) and a PEG lipid. The PEG lipid in LNP-ALC315 is ALC-0159; in LNP-SM102 it is PEG-DMG. As controls they used R848 (a TLR7/8 agonist) and MPLA (a TLR4 agonist), plus TNFα (50 pg/mL) and interferon α (5 ng/mL). Every LNP was given at 1 µg/mL (as mRNA equivalent).

① Empty LNPs alone threw the switch

NF-κB was detectable at 24 hours, rose at 48 hours and then stayed roughly flat out to 120 hours.

• LNP-ALC315: 2-fold over the unstimulated control at 24 hours, and 4-fold from 48 hours on
• LNP-1: 6- to 7-fold at 48 and 72 hours (the strongest)
• LNP-SM102: about the same as LNP-ALC315

IRF was more modest than NF-κB. LNP-ALC315 gave 3-fold at 48 hours and 1.7-fold at 72 hours. LNP-1 was about the same as LNP-ALC315, and the result was that LNP-SM102 was significantly stronger. R848 showed the strongest IRF response of all.

For every LNP the response was dose-dependent, plateauing at the higher doses (IRF plateaued at 2 µg/mL).

② The cause was the ionizable lipid

This is the core of the paper. When they gave the LNP with the ionizable lipid alone removed, both the NF-κB response and the IRF response disappeared.

👦 Student: Take out just one of the four lipid components and it goes to zero?

🧬 Dr. Exotaro: Yes. Cholesterol, DSPC and the PEG lipid are all still there, and no response comes up. It is an extremely legible result: the ionizable lipid is the source of the stimulus. But — and I have to say this honestly — the ionizable lipid they removed is the principal component, the largest share of the lipid in an LNP (though this paper does not report the molar ratios of its own formulations — only the component names). Take it out and what remains is not “the same particle minus one component” but a different particle. And in fact this control had a diameter of 80.26 nm, larger than the 54.07 nm of LNP-ALC315. I take this up again in the section on critical appraisal.

③ The route was TLR4 and MyD88

Next, identify the receptor. The authors started by looking at gene expression comprehensively.

They sequenced the RNA of THP-1 cells stimulated with LNP-1 for 24 hours, four samples per condition, and 10,585 of the 57,773 genes detected were called differentially expressed (false discovery rate 0.01, log2 fold change above 1 or below −1).

Among the NF-κB subunits, NFKB2 (log2 fold change 2.35) and RELB (2.74) rose strongly, and NFKB1 also rose significantly at 0.67 (just short of the cut-off). Among the IRFs, IRF7 (3.07), IRF4 (1.96), IRF9 (1.46) and IRF3 (0.19) rose. This is consistent with the reporter experiments.

And gene set enrichment analysis (GSEA) picked up an increase in the Toll-like receptor signaling pathway from the KEGG database. Looking at the genes inside that pathway, the list runs CD14 (4.07), CD86 (2.96), FOS (2.15), JUN (1.93), CD40 (1.26), STAT1 (1.18), MYD88 (0.97), TICAM1 (0.77, which encodes TRIF) and TLR4 (0.60). It was confirmed at the protein level too: by flow cytometry, CD14 and CD86 had risen about 2-fold and CD40 about 1.3-fold.

Here the authors get to the main business. They use THP-1 cells with TLR4 knocked out.

The result: for every ionizable LNP, both the NF-κB response and the IRF response fell markedly. But they do not go all the way to zero. The authors put it carefully: TLR4 is the main receptor, but other routes may also contribute.

One careful touch: they measured the LNPs for endotoxin contamination and confirmed that it was below the limit of detection. Endotoxin (bacterial LPS) is the native ligand of TLR4, so contamination would have ruined the result.

They broke the adapters too. In the MyD88 knockout, the NF-κB and IRF responses fell for all three LNPs. In the TRIF knockout, by contrast, the NF-κB response to LNP-1 actually rose slightly, and LNP-ALC315 and LNP-SM102 fell but not as far as in the MyD88 knockout.

🧬 Dr. Exotaro: So the main route runs from TLR4 to MyD88, and how far TRIF is involved varies by formulation. And one more thing: in A549 cells, a human lung epithelial cell line that does not express TLR4, none of the LNPs produced any response.

④ It happens at the cell surface — but serum is needed

There was something odd about the timing of the response. Signal transduction normally reaches phosphorylation of the transcription factor in about an hour, yet the LNP response peaks at 48 to 72 hours. TLR agonists come up strongly within 24 hours; the LNPs are slow.

The authors wondered whether the particles had to be taken up first. Following LNPs labelled with the fluorescent dye DiO, uptake is visible within 2 hours and is maximal at 48 hours. So they tried four endocytosis inhibitors: dynasore, cytochalasin D, chloroquine and methyl-β-cyclodextrin.

The result was unexpected. Chloroquine alone lowered the IRF response, but because dynasore, which blocks the same clathrin-dependent pathway, produced no change, the chloroquine effect was judged most likely to be an off-target action.

⚠️ A word here about the paper’s own wording as well. The paper introduces chloroquine as an inhibitor of clathrin-mediated endocytosis, or CME, but chloroquine is properly a lysosomotropic agent, whose main action is to stop the acidification of the endosome. It is not a drug that blocks CME itself. The authors’ line of argument — that it disagrees with dynasore, so it is probably off-target — holds up on its own terms, but the way the drug’s classification is written is inaccurate. Taken as a whole, the inhibitors had no significant effect on the LNP response.

The decider is localization. DiO-labelled LNPs and TLR4 did not colocalize at Rab5a, the early endosome marker. Which suggests that the encounter with TLR4 happens at the cell surface. TLR4 is known to use MyD88 at the surface and TRIF after endocytosis, so this also fits the finding that MyD88 dominates.

The authors do state explicitly, though, that “signaling from the endosome cannot be ruled out entirely”. The inhibition was not complete enough.

Another finding was the need for serum. Switch to serum-free medium and both the NF-κB response and the IRF response fell significantly.

👦 Student: Why would serum be needed?

🧬 Dr. Exotaro: The paper offers two possibilities. One is the protein corona — serum proteins forming a layer on the LNP surface, and that layer helping the particle latch onto the receptor. The other is LPS-binding protein (LBP), whose native role is to hand LPS over to TLR4. This paper does not decide between the two. But the implication matters, because the composition of a protein corona changes with the surrounding environment, so the corona that forms in the body after an intramuscular injection and the corona that forms after an intravenous one must differ. Serum in a test tube and the environment inside a human muscle differ too. The paper flags this explicitly as future work.

⑤ And adding mRNA did not increase the response

The last finding is the most unexpected one in the paper.

The authors had expected that LNPs loaded with mRNA would give the stronger response. But when they measured it, the response to mRNA-loaded LNPs was about the same as, or actually lower than, the matching empty LNPs. Surface CD86 expression was about the same as well. Since R848 does produce a properly strong response, the TLR7/8 route itself is alive. Even so, mRNA added nothing on top.

The authors’ conclusion runs like this: ionizable LNPs are capable of fully activating innate immunity in a monocyte model without mRNA, and adding mRNA did not increase the response. Put another way: the principal adjuvant in the BNT162b2 vaccine is the LNP.

⚠️ Reading this result as “vaccines are dangerous” is a mistake

I am setting aside a section here. Results of this kind are easy to misquote, so let me sort out what the paper itself does and does not say.

First, activation of innate immunity is a necessary condition for a vaccine to work. An adjuvant is precisely the ingredient you put in to aim at this. TLR agonists are in fact used as adjuvants in clinical trials and in approved vaccines. “LNPs raise innate immunity” is not the identification of a defect; it is an explanation of why mRNA vaccines work without an adjuvant. In the authors’ own words, the goal is to gain insight that will improve the immunogenicity and/or reactogenicity of future vaccines.

Second, this is an experiment in a cultured cell line. There are no animal experiments and no human data. Nor is it a study that measured the severity of reactogenicity. It is not a paper that settles “this is the cause of reactogenicity”; it is a paper that says “this is the mechanism of innate immune activation, so there is room to tune it”.

Third, the LNPs that were measured are not the commercial vaccines themselves. As the methods section says, they were built “following” the formulations — that is, rebuilt in the laboratory. (Please do not take the particle size figures as the grounds for this. As I discuss below, the 90.2 nm and 75.5 nm that the paper cites as “the values for the commercial products” were in fact values for LNPs that another study made itself. The grounds are the methods description instead.)

Fourth, the paper honestly lays out the earlier work that contradicts it. It states explicitly that there are reports in which knocking out TLR4 in mice did not abolish the vaccine response, and reports that MDA5 is the real key, and it also writes that TLR signaling in humans and in mice do not match completely.

🧬 Dr. Exotaro: As a physician I handle this sort of story carefully. What this paper showed is that part of the mechanism by which the vaccine works has been understood. And once it is understood, you can aim at designs that produce the same effect with less stimulation. That is the direction the authors are heading.

How will the future change? (The road to the clinic)

The route of “design the lipid to tune the immunity”

This is the application the paper shows most clearly. Change the structure of the ionizable lipid and you can change the strength of the innate immune stimulus.

There are three grounds for it. One, remove the ionizable lipid and the response disappeared. Two, the size and the pattern of the response differed across the three LNPs — LNP-SM102 gave the strongest IRF, LNP-1 the strongest NF-κB, and LNP-ALC315 about the same NF-κB as LNP-SM102 and about the same IRF as LNP-1 (the paper’s own description). Different structure, different response. Three, earlier work is closing in on the structural elements. Among the papers cited, there is a report that the structure of the amine head group of the ionizable lipid governs the immune response through binding to TLR4 and CD1d (Nature Biomedical Engineering 2024, volume 8, page 1483), and a report that differences in the head group of cationic lipids modulate TLR4 signaling.

🧬 Dr. Exotaro: The authors liken this to the way the structure of LPS differs between bacterial species, changing the strength of TLR4 signaling. The natural world holds a continuous range, from strong TLR4 ligands to weak TLR4 ligands. If so, the thinking goes, lipids too ought to let you choose where in that range to aim.

The formulation strategy that follows runs in two directions. If you want to raise the adjuvant activity — in the elderly or in infants, say, where the immune response is hard to get going, or in cancer vaccines — you choose a strong ionizable lipid. If you want to hold the reactogenicity down, you choose a weak one. Lipids that are at present chosen on a more or less single axis, “delivery efficiency”, would become selectable on an immunological axis as well.

”Tinkering with the lipid beats tinkering with the mRNA”

The change of direction set out in the paper’s last sentence matters too. The authors write that since the LNP rather than the RNA is the main driver of innate immune activation, effort should be turned toward understanding how to make LNPs that can target particular signaling pathways, rather than toward modifying the mRNA further.

Optimization on the mRNA side — modified nucleosides, purification, UTRs, codons — has already come a long way. What this paper is saying is that the headroom that remains is on the container’s side.

What comes into view when you lay the two papers on top of each other

Let me lay this over the other piece published the same day — a 2022 Nature Communications paper finding that 77% of LNPs carried no cargo.

That one measured that particles loaded with mRNA are only 23% of the total. This one measured that empty particles alone bring innate immunity fully up, and that adding mRNA does not increase it. Lay them straightforwardly on top of each other and the picture you get is that most of the adjuvant activity in a single administered dose of vaccine is carried by particles that hold not one strand of mRNA.

👦 Student: Is that a conclusion you are allowed to draw from these two papers?

🧬 Dr. Exotaro: Half allowed. The logic looks natural enough, but there are two cautions. One, the ionizable lipid in the paper that measured the empty fraction was DLin-MC3-DMA, not the ALC-0315 and SM-102 used here. How the empty fraction differs between formulations has not been measured. Two, with the way the present paper matched doses — “1 µg/mL as mRNA equivalent” — you cannot follow from the text exactly how the total lipid was matched between the empty LNPs and the mRNA-loaded LNPs. So “suggested” is as far as the correct wording goes.

Even so, I think it is a sharp question. A formulation improvement that reduces the empty fraction might raise delivery efficiency and at the same time lower the adjuvant activity. The two metrics can be a trade-off. You only get to notice that possibility by reading the two papers side by side.

The distance to the clinic

Let me write this plainly. This paper has no direct bridge to the clinic. It is an experiment in a cell line. Reactogenicity in animals, tolerability in humans, and the relationship to how long immunity lasts are all untested. Even so, it means something that the knob to be adjusted has been identified. If you do not know where the knob is, you cannot turn it.

How to read this study critically (its limits, and ways to raise its quality)

What deserves credit

Let me be fair. The comparison is correctly designed. Same cells, same dose, same time points, with three real LNP formulations laid alongside “the one with the ionizable lipid removed”. They verify by breaking genes — using cell lines with TLR4, MyD88 and TRIF individually knocked out is far stronger evidence than relying on inhibitors. They measured and excluded endotoxin contamination, which is obligatory work when you are arguing about TLR4 and which no small number of papers skip. The order — form a hypothesis from a comprehensive analysis, test it by breaking individual genes, then back it up at the protein level — is clean as well. They do not hide the negative results either: that NF-κB to LNP-1 actually rose in the TRIF knockout, that breaking TLR4 does not abolish the response completely, that endocytosis inhibition was not decisive, that there is contradictory earlier work — they write all of it. And they have released the sequence data.

Points to read with care

① A great deal is loaded onto a single cell line. Every major conclusion comes out of THP-1, an immortalized cell line derived from a patient with acute monocytic leukemia. A549, used as a supplement, is a cell line derived from a lung cancer too. There are no data from human primary monocytes or monocyte-derived dendritic cells, and no animal experiments. In a cell line, both receptor expression levels and signaling thresholds are shifted from those in a living cell. It is the trade you make for being able to use knockouts, but it falls one step short of a generalization about human innate immune cells.

② The control without ionizable lipid is a different particle altogether. I repeat myself, but this is the largest confounder. The ionizable lipid is the principal component, the largest share of the lipid in an LNP (and because this paper does not report the molar ratios, how much was removed cannot be worked out from the paper). The particle with it removed had a diameter of 80.26 nm (the LNP-ALC315 control, 54.07 nm) and a zeta potential of −8.548 (the control, −4.447; the paper gives these without units, conventionally mV), so both the size and the surface charge differ. Uptake, stability and the protein corona must all change with them. So “the response disappeared” does not show that “the ionizable lipid is the ligand”; it goes only as far as showing that “a particle has to contain the ionizable lipid to get a response”. There is independent evidence supporting the authors (the earlier work in which the amine head structure governs TLR4 binding), but this paper on its own does not show that the lipid molecule itself bound TLR4. Had there been an experiment lowering the amount in steps, or swapping in a non-ionizable lipid of similar shape, the claim would have been much stronger.

③ The description of particle properties is weak. Diameter, PDI and zeta potential are all single values per condition, and they are written out to digits that carry no meaning, as in LNP-SM102’s 51.590 nm or LNP-ALC315’s zeta potential of −4.447 (the zeta potentials carry no units). Replicate measurements and SDs are not reported (whether they were not performed or merely not reported cannot be told apart from the paper). On top of that, the diameters of LNP-1 and mRNA-LNP-1 are not reported. And there is a comparison here that falls apart once you go back to the source. The paper’s introduction cites the diameters of the “vaccine LNPs” of BNT162b2 and mRNA-1273 as 90.2 ± 7.8 nm and 75.5 ± 0.4 nm (the citation is Zhang L et al., npj Vaccines 2023;8:156). But read that source and what was measured is not the commercial vaccines themselves. They are values for LNPs loaded with firefly luciferase mRNA that the cited study made itself, with lipid molar ratios of 50:10:38.5:1.5. The cited source itself states explicitly that “the composition of these LNPs is not the same as the Pfizer/BioNTech or Moderna vaccines”.

Which means that setting this study’s 54.07 nm beside the cited 90.2 nm is a comparison between two reconstructions, and proves nothing about a difference from the commercial product. The misattribution here is on this paper’s side, and I myself took it at face value at first. The fact itself that “what was measured was a reconstruction made with reference to the formulation, not the commercial vaccine itself” still stands, unchanged, from the methods description (built “following” the formulation) — it is only that the grounds for it are not the diameter figures.

④ The modification status of the mRNA they used is not stated in the methods. This bears on the centre of the claim, so let me point it out concretely. The methods section does say that the mRNA encodes the SARS-CoV-2 Omicron spike (S2P), that it was capped co-transcriptionally with CleanCap, that it was purified by cellulose chromatography, that removal of double-stranded RNA was confirmed by dot blot, and that endotoxin was measured (less than 0.05 EU/mL). But whether the N1-methylpseudouridine modification was performed is not stated. Since the paper’s introduction takes “the mRNA in current vaccines is modified” as its starting point, whether their own mRNA was modified is essential information. If it was unmodified, then even an mRNA that ought to be the more stimulatory added nothing on top, which favours the authors; if it was modified, it can also be read as “it was quiet thanks to the modification”. The reader is left unable to choose between the interpretations.

⑤ The comparison with the mRNA-loaded particles is limited. mRNA-loaded LNPs were made for only two of them, LNP-1 and LNP-ALC315, and there is no mRNA-LNP-SM102. The comparison is at a single time point, 72 hours. The reach of the conclusion that “mRNA adds nothing on top” stops inside that range. The authors themselves, moreover, offer as an explanation for the discrepancy with the earlier study that used GMP-manufactured BNT162b2 (the report that MDA5 is the key) the fact that their own mRNA was a cellulose-purified product. In other words, the authors too accept that the result can change with how the mRNA is made.

⑥ There is a place where the text and their own table disagree. On the IRFs, the authors write that “among the IRFs that rose, only IRF4 and IRF7 met the criteria for a differentially expressed gene”. Yet the table in the same figure lists IRF9 at a log2 fold change of 1.46 and an adjusted p-value of 3.3 × 10⁻⁵⁸. That clearly meets the authors’ own criteria (log2 fold change above 1, false discovery rate 0.01). It is not an error that changes the conclusion, but a reader quoting the numbers needs to look at the table.

⑦ The statistics are described inconsistently. The statistical analysis section mentions nothing but one-way ANOVA (Dunnett or Tukey), yet the figure legends also use a t-test. The legend to Figure 4 reads “significance was assessed by t test (A) or one-way ANOVA with Dunnett’s multiple comparisons (A, B) or Student’s t test (C)”, which assigns two different tests to panel A. The reader cannot tell which one a given value came from. On top of that, in the legends to Figures 2 and 4, the asterisks for the fourth significance level are printed as three (***P ≤ 0.001, ***P ≤ 0.0001). A proofreading flaw.

⑧ What is inside n = 3 is never defined. The methods say that “biological and technical replicates were performed for each condition”, and the figure legends say n = 3. Whether that 3 means independent biological replicates or technical replicates from the same culture cannot be made out. The data are also shown as standard error of the mean (SEM) rather than standard deviation, which makes the scatter look small.

⑨ A cell-death confounder remains. The authors write it honestly: viability fell under every condition, and at 48 hours LNP-1 caused the strongest cell death, which corresponded to the strongest NF-κB activation. Since the substances released by dying cells (DAMPs) can hit TLR4, what is being watched — stimulation by the LNP, or secondary stimulation by the cell death the LNP caused — cannot in principle be fully separated. The authors argue for a route distinct from cell death on the grounds that LNP-SM102 caused comparable cell death yet barely moved NF-κB at this time point, and on the earlier work in which the lipid structure binds TLR4. It is persuasive, but it is not settled.

⑩ The reporters are a cumulative readout. Both SEAP and luciferase are secreted enzymes that accumulate in the culture medium. What is being measured is not the instantaneous signal strength but something close to the sum of all production up to that point. The observation that “the LNP response peaks at 48 to 72 hours, later than the TLR agonists” is interesting, but it is also a result in which gradual uptake, cell proliferation and enzyme accumulation are mixed together. Had there been an experiment pinning the time course down with a direct readout such as phosphorylation, the argument about kinetics would have been far firmer.

⑪ Cytokine proteins were not measured. The outputs of this paper are reporter enzymes, surface markers and transcripts. How much IL-6, IL-1β or interferon was actually released was not measured. As a foundation for arguing about reactogenicity or tolerability, there is a hole right here.

⑫ LNP-1 is a proprietary formulation. LNP-1, which gave the strongest NF-κB response and was also chosen as the model for the transcriptome analysis, has a composition that is not public outside the patent text. A third party would find it hard to reproduce from the paper’s description alone, and it is not the formulation of an approved vaccine either. Making a proprietary formulation the pillar of an argument about the generality of a mechanism is a weakness, whatever the unavoidable circumstances.

⑬ A small thing, but there is an error in an affiliation. The seventh affiliation reads “Uniformed Services University of Health Sciences, Bethesda, PA”, but Bethesda is in Maryland (MD). The other affiliations give MD.

What would have made this a higher-quality study

• Reproduce it in human primary cells. Confirm at least the main conclusions — dependence on the ionizable lipid, dependence on TLR4, no addition from mRNA — in peripheral blood monocytes or monocyte-derived dendritic cells. Make up for the loss of knockouts with neutralizing antibodies or inhibitors
• Do not “remove” the ionizable lipid; reduce it, or swap it out. Compare with the particle properties matched, either through a stepwise dose–response or by swapping in a non-ionizable lipid of similar shape. With the present control, the contribution of the molecule and the contribution of the particle cannot be separated
• Report the particle properties with replicate measurements, for every formulation. Match the significant figures to reality. Discuss the differences from the commercial products explicitly
• State the modification status of the mRNA. Where possible, put modified and unmodified mRNA side by side and compare with a GMP-manufactured product too. That would let the cause of the discrepancy with earlier work be identified
• Measure the cytokines as proteins. Reporters and transcripts alone do not connect to an argument about reactogenicity or tolerability
• Follow it in animals as far as reactogenicity readouts. The largest claim — that changing the lipid structure lets you tune the stimulus — ultimately has to be shown in both tolerability and immunogenicity in a living body
• Check the numbers, the statistics and the figures and tables mechanically. The IRF9 contradiction, the duplicated test attributions, the asterisk typo and the wrong state name could all be crushed with a pre-submission checklist

Dr. Exotaro’s perspective

A container is not just a box for carrying the contents

Let me be honest with you. Vaccinology is not my field. What I work on is treating neurological disease, spinal cord injury (SCI) above all, using mesenchymal stem cells (MSCs) and the extracellular vesicles (EVs) they release. Nor have I any experience handling TLR4 knockout cell lines.

I chose this paper anyway, because what is written here lands right in the centre of our own field.

Put the lesson of this paper in one line and it reads: a container is not just a box for carrying the contents; the container itself has biological activity. And our field is handling exactly that one line.

When we explain the therapeutic effect of MSC-EVs, we usually talk about the contents: this miRNA suppresses inflammation, this protein helps axons extend. But an EV is a particle made of a lipid bilayer, with tetraspanins and integrins embedded in the membrane and a layer of protein covering the surface. There is no way that membrane and that surface are doing nothing to the cell that receives them.

The main effects hoped for from MSC-EVs in particular — immunomodulation and anti-inflammatory action — are, by their very nature, not necessarily things that happen only after the contents have been delivered to the cytoplasm. They may begin the moment the particle touches a receptor on the cell surface. Which is to say that in our field, the biological activity of the container may not be a side effect at all but the very main effect we are hoping for.

This paper showed exactly that, cleanly, in the simple system of a synthetic particle and with knockout cells. Take out one component and the response disappeared. Break the receptor and the response fell. When we put the same question to EVs, could we answer it this cleanly? Taking “just that one component” out of an EV is not easy.

TLR4 is one of the leading players on the ground in spinal cord injury

This is the most concrete point for me, and the one that made me sit up straightest.

TLR4 is a receptor that sits at the centre of the pathology after spinal cord injury. Injured tissue releases endogenous molecules from broken cells (DAMPs), and these drive an inflammatory response through TLR4 on microglia and macrophages. That response works in the direction of spreading the secondary injury, and at the same time in the direction needed for clearing the debris and repairing — it is an awkward pathway, not simply a villain, but one whose meaning changes with its strength and its timing.

Now lay this paper’s result on top of that. Ionizable lipid LNPs activate innate immunity through TLR4.

🧬 Dr. Exotaro: If I were to try to deliver mRNA or siRNA to an injured spinal cord with LNPs, what would I be doing? I would be pouring particles that hit TLR4 into a tissue where inflammation is already running on TLR4 signaling.

This is an inference I drew from the paper, not something the paper showed. The paper does not touch the central nervous system at all, and it is a story about a cultured cell line. Between local and systemic administration, both the protein corona and the cell types reached are completely different. So this gives me no grounds at all for saying that LNPs cannot be used in spinal cord injury. There will be situations where a measured innate immune stimulus at the injury site is in fact beneficial.

Even so, you must not choose a courier for the single reason that its delivery efficiency is good — that one point I felt I should certainly take home. The immunological properties of the courier multiply with the state of the tissue it is delivered to. Whether you are delivering into inflamed tissue or quiet tissue changes how the same courier should be judged.

So does the case for using EVs as the courier get stronger?

Let me write this carefully, so as not to give in to the temptation.

“EVs come from a living body, so they are immunologically quiet and safer than LNPs” — it is tempting to say that. And indeed EVs contain no synthetic ionizable lipid. But the inference is dangerous. There is no evidence that EVs are immunologically inert. The surface of an EV carries a diversity of molecules, and both contents and surface change with the state of the donor cell. With allogeneic EVs, the molecules presented can become a problem too. The absence of synthetic lipid is not the same thing as the absence of immune stimulation.

What this paper actually prompted in me is the opposite piece of work: that when you evaluate an EV as a courier, you should measure its immunological properties from the start. Not only delivery efficiency and an analysis of the contents, but which receptor it touches and which transcription factor it moves — at the same standard this paper worked to.

On the writing

One last thing. What I trusted most in this paper is that they write the things inconvenient to their own conclusion into the main text.

They write that breaking TLR4 did not abolish the response completely. They write that endocytosis inhibition did not settle the matter, and that signaling from the endosome cannot be ruled out. They write that there was a formulation whose response actually rose when TRIF was broken. They raise, themselves, the possibility that cell death is mixed in. They write that there are reports in which knocking out TLR4 in mice did not abolish the response, and reports that MDA5 is the real key, and that TLRs in humans and in mice do not match completely. And they attach their own reservation: “further investigation is needed for definitive proof of IRF activation”.

Do not mix what you measured with what you inferred. And introduce the earlier work that counts against your own hypothesis in the main text, not in a footnote. I think that is the attitude befitting a paper that claims a mechanism.

Note: this article was written with the assistance of AI.