Counting mRNA-Carrying Lipid Nanoparticles One at a Time — The Average Load Was 2.8 mRNA Molecules, and 8 Out of 10 Particles Were Running 'Empty'

A lipid nanoparticle (LNP) is a carrier that wraps fragile mRNA in a bead of lipid and delivers it all the way inside a cell. mRNA vaccines brought it into practical use all at once, and it is now used in therapies that deliver gene-editing tools to the liver as well.

Here is a plain question. How many mRNA molecules are inside a single one of those particles?

Almost no one could answer on the spot. What appears on a formulation’s quality spec is a number like “encapsulation efficiency (EE) 94%.” That means “94% of the mRNA we put in ended up inside the particles,” and it is a fraction seen from the mRNA’s side. It says not a word about the questions seen from the particle’s side: how many copies per particle, and how many particles are empty-handed.

What we take up this time is a 2022 Nature Communications paper that answered that question one particle at a time. Let me give the conclusion first. In LNPs made with a standard laboratory formulation, the average per particle among the particles that carry mRNA was 2.80 copies, and the mode was 2 copies. And of all the particles, 77% were not carrying a single molecule of mRNA.

Two caveats up front. First, this is a paper from 2022. I take it up anyway because, in my own view, the number that came out of it — “most of them are running empty” — became a starting point for thinking about this field thereafter, and because the thinking behind the measurement goes straight to the heart of our own work on extracellular vesicles (EVs). Second, what was measured was a commercially available dye-labeled mRNA together with a “textbook” laboratory formulation that uses DLin-MC3-DMA as the ionizable lipid. This is not a paper that measured a COVID-19 vaccine itself. Please do not confuse the two.

Publication Details

• Paper title: Payload distribution and capacity of mRNA lipid nanoparticles
• Paper type: Original research article (measurement-technology development plus formulation physical chemistry)
• Journal: Nature Communications 2022, Volume 13, Article number 5561 (received January 13, 2022; accepted September 5; published online September 23)
• DOI / Link: 10.1038/s41467-022-33157-4
• License: Open access (CC BY 4.0). Anyone can read the full text
• Impact Factor: roughly around 15 (it varies by year; it is one of the very top general science journals)
• Authors: Sixuan Li and Yizong Hu (these two are co-first authors), Andrew Li, Jinghan Lin, Kuangwen Hsieh, Zachary Schneiderman, Pengfei Zhang, Yining Zhu, Chenhu Qiu, Efrosini Kokkoli, Tza-Huei Wang, and Hai-Quan Mao — twelve in all
• Affiliations: All at Johns Hopkins University (Baltimore, Maryland, USA). Department of Mechanical Engineering; Institute for NanoBioTechnology (INBT); Department of Biomedical Engineering in the School of Medicine; Translational Tissue Engineering Center; Department of Chemical and Biomolecular Engineering; Department of Materials Science and Engineering
• Corresponding authors: Yizong Hu, Tza-Huei Wang, and Hai-Quan Mao — three in all (Wang and Mao are explicitly noted as having “contributed equally as co-corresponding authors”)
• Funding: U01AI155313 (Mao) and R01AI137272 (Wang), from the US National Institute of Allergy and Infectious Diseases (NIAID)
• Conflict-of-interest disclosure: Four of the authors — Li, Hu, Wang, and Mao — have a patent application pending on the technology described in this paper (filed through the Johns Hopkins University Office of Technology Ventures). The other authors declare none
• Peer review: Under Nature Communications policy the peer review file is public, and Xucheng Hou is disclosed by name as one of the reviewers

What Kind of Researchers Are the Corresponding Authors?

Hai-Quan Mao leads the Institute for NanoBioTechnology (INBT) at Johns Hopkins University and holds professorships in both the Department of Materials Science and Engineering and the Department of Biomedical Engineering in the School of Medicine. His path is a little unusual: he took a bachelor’s degree in chemistry (1988) and a PhD in polymer chemistry (1993) at Wuhan University in China, served as a lecturer there, and then did a postdoc at Johns Hopkins University from 1995 to 1998. From 1999 to 2003 he was a co-principal investigator at Johns Hopkins Singapore; from 2001 to 2003 he held a concurrent assistant professorship at the National University of Singapore; and in 2003 he joined the faculty of the Whiting School of Engineering at Johns Hopkins University. He became associate director of INBT in 2016 and is now its director. Royal Society of Chemistry Fellow (2014), American Institute for Medical and Biological Engineering (AIMBE) Fellow (2018), and National Academy of Inventors (NAI) Fellow (2023).

There are two pillars to his research. One is a method called flash nanocomplexation, or FNC, which made the process of mixing nucleic acids or proteins with a polymer to form nanoparticles scalable. The other is the nanofiber-hydrogel composite (NHC), an injectable scaffold material in which fragments of electrospun polycaprolactone (PCL) nanofiber are woven by covalent bonds into a hyaluronic acid gel network.

This is the point that matters to me, so let me put it first. The Mao laboratory has in fact been pursuing research that uses that NHC for spinal cord injury. Biomaterials 2020, volume 245, 119978 (first author Xiaowei Li, last author Mao) injected NHC into a rat spinal cord contusion model, and Cells 2022, volume 11, issue 7, 1137 (first author Agnes Haggerty, with Mao and Martin Oudega as co-authors; open access) went on to inject bone marrow-derived mesenchymal stromal cells (MSCs) combined with the NHC into the injured spinal cord. In other words, the corresponding author of today’s paper is someone running the physical chemistry of mRNA-LNPs and the regenerative medicine of MSCs for spinal cord injury in the same laboratory. I will come back to this in the second half.

Tza-Huei (Jeff) Wang is the Louis M. Sardella Professor in the same university’s Department of Mechanical Engineering (an endowed chair established in 2021), holds a joint appointment in Biomedical Engineering, and directs the BioMEMS and Single Molecule Dynamics Lab. He took a PhD in mechanical engineering at the University of California, Los Angeles (UCLA) in 2002, joined Johns Hopkins University the same year, and became a full professor in July 2013. AIMBE Fellow (2017, elected for work on nanobiosensors and single-molecule detection in the analysis of genetic and epigenetic disease). More than 150 papers and more than 20 patents.

The central technology of this paper is itself an invention of the Wang laboratory. CICS (cylindrical illumination confocal spectroscopy), used throughout the text, is a method that same laboratory published in Biophysical Journal 2008, volume 95, page 2964. Their usual theme runs toward early cancer detection by reading methylation in liquid biopsy (trace DNA in blood) with digital high-resolution melting, and the one point it shares with this LNP analysis is “counting a very small number of molecules accurately, one at a time.”

Yizong Hu, co-first author and also a corresponding author, is, according to the author contributions section, the person who conceived this study. He took a bachelor’s degree in polymer materials at Tsinghua University (2016) and a PhD in biomedical engineering at Johns Hopkins University (2022). According to public profiles, he has since been a postdoc at the Massachusetts Institute of Technology (MIT), studying non-viral gene delivery.

What Wasn’t Understood Before?

The Question That “Encapsulation Efficiency 94%” Does Not Answer

Encapsulation efficiency comes up without fail whenever LNP quality is discussed. The measurement is simple: using a dye such as RiboGreen that binds nucleic acid and fluoresces, you compare the fluorescence “before the particles are broken open with detergent” and “after.” Before, only the mRNA that is outside fluoresces; after, all of it fluoresces. The difference gives you “what percentage was inside the particles.”

👦 Student: If 94% of it is getting in, that feels like plenty to me.

🧬 Dr. Exotaro: No — the question is a different one. This is a roll call of the cargo. It says, “of the 100 parcels we shipped, 94 were loaded onto trucks.” But what we want to know is about the trucks: how many trucks are running, how many parcels are on each one, and how many are driving around empty. A roll call will never tell us.

This is not word play. Once you know the payload per particle, you know the number of particles. Once you know the number of particles, you can start arguing about how many particles a single cell has to take up for the drug to work. And if many particles are empty, it means we are giving patients lipid that contributes nothing to the effect. Since it is already known that the lipids themselves can stimulate the immune system and burden the liver, this connects directly to safety as well.

Why No One Could Measure It

Every existing measurement method fell one step short.

Cryo-electron microscopy (cryo-TEM) can look at the particles directly. But at physiological pH 7.4 the ionizable lipid loses its charge and forms a hydrophobic amorphous core at the center of the particle, so particles look equally black and packed whether or not mRNA is inside. This paper too imaged empty LNPs made without mRNA, and they cannot be told apart.

Small-angle neutron scattering, NMR, and nanoparticle tracking analysis (NTA) hit the same wall. They can give you particle size and average structure, but they cannot judge loaded versus unloaded one particle at a time.

So the authors turned to fluorescence. An ordinary microscope, however, is not enough. What was needed was an instrument that could pass the particles through one at a time without missing essentially any particle, and record which colors lit up together (the paper’s own wording is “~100% mass detection efficiency”).

CICS — Sending Particles Through a Capillary “Ticket Gate” One at a Time

The idea behind CICS goes like this. Sample is pushed through a fused-silica capillary about as thin as a human hair, with an inner diameter of 10 µm, under pressure (argon gas at 42 psi). The flow velocity is 1 mm/s. Three lasers (488 nm, 552 nm, 640 nm) are combined into one by a beam combiner and then spread out sideways by a cylindrical lens and applied as a sheet of light.

👦 Student: Why a sheet, instead of an ordinary point?

🧬 Dr. Exotaro: To block the full width of the ticket gate. If you focus the laser to a point, only the particles passing through the center of the capillary shine brightly, and those passing near the edge come out dim. Then you cannot tell a “dim particle” from a “particle with less inside.” Cover the whole cross-section with a sheet of light and a particle shines just as brightly wherever it passes. The paper describes this as nearly 100% mass detection efficiency. Everyone goes through the gate — and everyone goes through under the same conditions.

The detectors are three avalanche photodiodes (APDs) that count single photons. Recording in 0.1-millisecond bins, they collect more than 50,000 events in 20 minutes per formulation (3,000 to 5,000 events per minute).

And three fluorescent labels are used for different jobs.

Using the combination of simultaneous flashes among these three colors, three species are distinguished.

• mRNA-loaded LNP = YOYO-1 (−), TMR (+), Cy5 (+)
• Empty LNP = TMR (+) only
• Unencapsulated naked mRNA = YOYO-1 (+), Cy5 (+), TMR (−)

How They Counted “One Copy, Two Copies, or Three”

This is the most ingenious part of the paper. Cy5 brightness is roughly set by “how many mRNA molecules are there,” but it is also swayed by the variation in the number of Cy5 molecules per copy, the Poisson-like fluctuation of photon emission and detection, and drift in laser power and flow rate. So you cannot declare particle by particle, “this one is bright, therefore three copies.”

Instead, the authors used a different move: they solved the distribution (deconvolution).

1. First, measure the brightness distribution of a single naked mRNA molecule (the paper writes this as D(RNA,1))
2. Multiply the brightness of that distribution by n and you get D(RNA,n), the distribution for a particle holding exactly n copies
3. Adjust the weights w(n) so that a weighted sum of D(RNA,1), D(RNA,2), … reproduces the measured LNP distribution
4. The optimal set of weights is itself the answer to “what percentage carries 1 copy, what percentage carries 2, and so on”

The optimization used simulated annealing in MATLAB, with a penalty coefficient of α = 0.1 imposed so that the sum of the weights would not deviate by more than 1% from the measured total particle count. The upper limit of payload considered was set at 6 times the estimated mean payload (the ratio of the geometric means of the fluorescence of mRNA-loaded LNPs and of naked mRNA).

👦 Student: You say “count them one particle at a time,” but in the end it is an estimate of a distribution, isn’t it?

🧬 Dr. Exotaro: Sharp. The loaded-or-not call is made one particle at a time; the how-many is an estimate for the population. This is a distinction worth understanding precisely, and I will touch on it again in the critical reading later.

What Did This Paper Find?

① Measured Values for the Standard Formulation — 77% Were Running Empty

What they measured was the benchmark formulation often used in the literature: the ionizable lipid DLin-MC3-DMA, the helper lipid DSPC, cholesterol, and the PEG lipid DMG-PEG2000 at a molar ratio of 50:10:38.5:1.5. The mRNA was a Cy5-labeled product 1929 nucleotides (nt) long, at a final concentration of 20 µg/mL and an N/P ratio of 6 (the molar ratio of the ionizable lipid’s amine groups to the mRNA’s phosphate groups). The lipid-ethanol solution at 1 mL/min and the aqueous mRNA solution at 3 mL/min are run into a T-junction and mixed, then dialyzed through a membrane with a 3,500 molecular weight cut-off at 4 ℃ for 12 hours.

The decisive thing is that they measured both states: still acidic at pH 4.0, and after the return to neutral at pH 7.4.

⚠️ A word of caution about terminology is needed. Table 2 of the paper calls the former “Before dialysis at pH 4.0 — that is, the initial LNP,” but reading the Methods section shows that the pH 4.0 sample was also dialyzed for 12 hours — against acetate buffer, in order to remove the ethanol. So strictly speaking this is not a comparison of “before and after dialysis” but of something “dialyzed while still acidic” versus something “dialyzed all the way to neutrality.” Because the paper’s own labels conflict with its Methods, I will follow the paper’s naming below, but please read it with this point in mind. The protagonist of the mechanism is always deprotonation caused by the shift in pH, not whether or not dialysis was performed.

At pH 4.0, the paper’s so-called before-dialysis sample, three populations were visible.

• Lipophilic mRNA complexes, 34% (TMR+ Cy5+; they carry a fair amount of helper lipid)
• Non-lipophilic mRNA complexes, 25% (TMR− Cy5+; almost no helper lipid — strongly charged clusters of mRNA and ionizable lipid)
• Empty LNPs, 41%

At pH 7.4, after dialysis, the picture changes completely.

• mRNA-loaded LNPs, 23%
• Empty LNPs, 77%
• Unencapsulated naked mRNA accounted for only 4% of all events

Note that there are two different denominators here. The 23% and the 77% are fractions whose denominator is the events in which TMR lit up — that is, the lipid particles, while the 4% is a fraction whose denominator is all of the detected events (the legend to Figure 2 of the paper says so). That is why the first two sum to 100%, and the naked mRNA is not included among them.

And then the payload. The number-average for mRNA-loaded LNPs was 2.80 ± 0.41 copies, the mode was 2, and roughly three-quarters carried 1 to 3 copies (n = 6, independent formulation experiments). At pH 4.0 the lipophilic complexes held 3.43 ± 0.38 copies, the non-lipophilic complexes 1.34 ± 0.20 copies, and the population as a whole 2.51 ± 0.24 copies.

The supporting measurements are all there too. Particle size (z-average by dynamic light scattering) was 106.3 ± 13.0 nm at pH 4.0 and 120.5 ± 6.0 nm at pH 7.4. Zeta potential went from +45.1 ± 0.9 mV to −6.3 ± 1.3 mV. Encapsulation efficiency was 94.2% ± 3.6% by RiboGreen, and 85.6% ± 5.1% when derived from the CICS counts.

The authors also showed that three-color discrimination really was necessary. With two-color discrimination, which does not use YOYO-1, up to 10.4% of events are misclassified as mRNA-loaded LNPs. The extra trouble was worth taking.

② Why Empty Particles Increase — During Dialysis, Particles “Split and Stick Together”

Empty LNPs were 41% at pH 4.0, but 77% at pH 7.4. They went up. The authors read an assembly mechanism into this.

The key is a prior finding that the PEG lipid sets the ceiling on particle size. PEG occupies a certain area at the interface, so if the amount of the other lipids is the same, the higher the PEG fraction, the more the particle has no choice but to raise its surface-area-to-volume ratio — that is, to get smaller. In fact, raising DMG-PEG from 0.25% to 3% shrank the mean particle size at pH 7.4 from 210 nm to 100 nm. At pH 4.0, by contrast, changing the PEG content changes neither particle size nor payload. In other words, all of the change happens during dialysis.

👦 Student: Isn’t dialysis just the step for pulling the ethanol out?

🧬 Dr. Exotaro: That is the pitfall. When dialysis raises the pH from 4.0 to 7.4, the ionizable lipid loses its positive charge. A particle that at pH 4.0 was stabilized by charge repulsion plus PEG now has to be stabilized at pH 7.4 by PEG alone. The means of support changes, so the particle rearranges its shape. On top of that, it has to move from the bilayer vesicle of pH 4.0 to the pH 7.4 structure in which a single layer wraps a hydrophobic core. The same amount of lipid now needs more surface area. So the particles presumably split — that is the paper’s reading (for the storylines of Figures 4 and 6, the paper itself writes explicitly that they are “hypothesized”).

The storyline the paper presents runs as follows. Everything below is a hypothesis the paper advanced as an inference from population statistics, not something observed at the moment of splitting or the moment of fusion.

When PEG is high, at 1.5% or more: empty LNPs are thought to have increased by splitting (from the observed facts that the mean TMR signal falls and the number concentration rises). Some of the lipophilic complexes that started with a high payload are also interpreted as having split and lost payload. Those with medium to low payload keep their payload. When PEG is very high (3.0% or more), splitting dominates: the mean payload falls and the particle count rises. The non-lipophilic complexes, which had no helper lipid, are meanwhile inferred to have fused with empty LNPs and taken up helper lipid from them (a back-calculation from the fact that the payload distribution becomes relatively uniform after dialysis, and that a population which had no TMR acquires TMR).

When PEG is low, at 1.0% or less: the size ceiling loosens, so the lipophilic complexes can fuse with one another — or so the paper reasons. As a result, the payload rose considerably. The TMR signal rose consistently more than the mRNA payload did here — evidence that the lipid is being received from empty LNPs.

The contrast in the real data is easy to read. At PEG 1.5%, empty went 41% → 77%; at PEG 0.5%, empty went 33% → 81%. In both cases the empty fraction rose after dialysis.

③ Lower the N/P Ratio and the Payload Rises at the Same Particle Size

The next variable to turn is the N/P ratio. The mRNA concentration is held constant while all the lipids are raised or lowered in proportion to the ionizable lipid. Since PEG’s fraction of the total lipid does not change, the ceiling on particle size does not change either.

And yet at the same particle size, lowering the N/P ratio increased the payload per particle. The logic is one of stock ratios. A lower N/P ratio means relatively less lipid per unit of mRNA, which means less lipid left over to become empty LNPs. In fact, the fraction of empty LNPs at pH 4.0 showed a strong positive correlation with the N/P ratio (50% at N/P = 12, 15% at N/P = 2).

The paper’s explanation is that this changes whom a particle meets during dialysis. If the N/P ratio is high, the surroundings are nothing but empty LNPs, so an mRNA-carrying particle reaches the ceiling size by fusing with empty LNPs — that is, by receiving only lipid and no extra mRNA. If the N/P ratio is low, the surroundings are mRNA complexes, so particles that carry mRNA fuse with each other. Neither of these was observed; both are storylines derived from the distributions. The number of parcels loaded by the time the same ceiling size is reached comes out different.

However, the empty fraction at pH 7.4 did not fall when the N/P ratio was lowered. At a low N/P ratio the absolute number of empty LNPs falls too, but because the payload rises, the number of mRNA-loaded LNPs falls as well. The two cancel out and the fraction does not move — that is the explanation.

④ Concentration Does Not Matter, Size Does — “Payload Is a Weight, Not a Count”

When the mRNA concentration was varied 20-fold, from 5 to 100 µg/mL, with the lipids scaled in proportion, neither particle size nor the payload distribution changed much (only the empty fraction differed, being lowest at the lowest concentration of 5 µg/mL and nearly the same from 20 to 100 µg/mL).

This runs against common sense in the polymer world. For complexes of pDNA with polyethylenimine, it was known that raising the nucleic acid concentration leaves particles dynamically frozen at a high payload. The authors interpret this as meaning that cationic lipids move faster than polymeric cations, so they can access mRNA well enough and are less prone to bridging multiple mRNA molecules.

What did matter, conversely, was the length of the mRNA. Replacing the 1929 nt with 996 nt — roughly half — left the size ceiling unchanged yet made the mean payload increase significantly, and the mode doubled at both pH 4.0 and pH 7.4. The empty fraction also fell at both N/P = 3 and N/P = 6.

🧬 Dr. Exotaro: The sentence that comes out of this is the most beautiful conclusion in the paper. “The loading capacity of an LNP that has a given size ceiling can be expressed not as a number of mRNA molecules, but as a certain fixed mass.” Since the volume of the cargo bed is fixed, small parcels can be loaded in greater numbers. That there are reports of studies loading siRNA (about 20 nt) in which no empty LNPs existed connects through the very same logic.

⑤ Adding Empty Particles Lowered Expression in the Liver

Finally, they check whether any of this means anything inside the body. Alongside the N/P = 3 and N/P = 6 formulations (the paper describes both as being about 75% empty, which differs slightly from the benchmark value of 77% ± 8% in Table 2, though both figures are the paper’s own), they prepared an “N/P = 3+3” in which LNPs made at N/P = 3 were mixed afterwards with empty LNPs to bring the total N/P ratio to 6 (86% empty). This is a group in which the number of mRNA-loaded LNPs is unchanged and only the empty particles are roughly doubled.

Loaded with luciferase mRNA, 0.5 mg mRNA/kg was given into the tail vein of Balb/cJ mice (female, 8 weeks old); 12 hours later D-luciferin was injected intraperitoneally, imaging was done with IVIS, and the organs were removed and quantified.

• Liver: expression fell in the group to which empty LNPs had been added (N/P = 3+3) (P = 0.0298, unpaired two-tailed t-test, comparing N/P = 3 with N/P = 3+3). The amount distributed to the liver itself did not change
• Spleen: only the two groups derived from N/P = 3 (3 and 3+3) showed meaningful expression, and N/P = 6 showed almost no expression (P = 0.0280)

LNPs are said to collect in the liver through interaction with apolipoprotein E (ApoE) in the blood. The authors raise the possibility that empty LNPs competed with mRNA-loaded LNPs for ApoE, while closing cautiously with “further investigation is needed.”

How Will the Future Change? (The Path to the Clinic)

Manufacturing QC Can Add “Metrics Seen From the Particle’s Side”

This is the most realistic use. Because CICS measures in flow, it needs only a tiny amount of sample, and it can get through 50,000 events or more in 20 minutes per formulation (events include other species such as naked mRNA). The paper also demonstrated applying it to the assessment of storage stability (Supplementary Figure 12).

What is used today to judge an LNP lot is particle size, polydispersity index, zeta potential, and encapsulation efficiency. What happens if the empty-particle fraction and the per-particle payload distribution are added to that list? Between two lots with the same 94% encapsulation efficiency, one at 77% empty and one at 50% empty, the total amount of lipid delivered differs when the same amount of mRNA is given. In other words, the lipid burden needed to produce the same effect becomes something you can manage as a metric.

A Direction for Optimization — “Reduce the Empty Particles” — Has Come Into View

This paper laid out which knob moves what. The PEG content sets the ceiling on particle size, the mass ratio of lipid to mRNA sets the initial tug-of-war between the mRNA complexes that form first and the empty LNPs, and splitting and fusion during dialysis finish the job. Concentration does not matter. If the cargo is small, empty particles decrease.

👦 Student: So shouldn’t we just make LNPs with zero empty particles?

🧬 Dr. Exotaro: The interesting part is that it is not that simple. If you lower the N/P ratio in order to reduce empty particles, the payload per particle rises and the number of particles falls. In this paper’s animal experiment, N/P = 6 gave almost no expression in the spleen, while N/P = 3 did express. A small number of heavily loaded particles and a large number of lightly loaded ones behave differently in the body, even at the same total dose of mRNA. The optimum changes with where you want to deliver. The empty fraction should be read not as “an evil to be minimized” but as one of the degrees of freedom in design.

That Said, the Distance to the Clinic Is Frankly Large

There are no clinical data in this paper. The mouse experiment, too, is a single dose, a single 12-hour time point, liver and spleen only, and a reporter gene. Repeat dosing, immune response, expression of a therapeutic protein, and pharmacokinetics in humans are all out of scope. And as I write in the next section, the formulation measured is not the formulation of a current vaccine.

How to Critically Read This Research (Limitations and Paths to Improve It)

What Deserves Credit

Let me be fair. The measurement is designed correctly. The idea of covering the cross-section with a sheet of light to equalize detection efficiency, the design that distinguishes three species from three colors lighting up together, and the notion of measuring the distribution of a single naked mRNA molecule and using it as the basis set all point toward counting without fudging. By measuring both before and after dialysis, a static description became an account of mechanism. That they quantified for themselves that two colors could misclassify up to 10.4% of events is honest too. They corroborated it with an orthogonal method — observing LNPs set in resin by super-resolution confocal microscopy and confirming by a separate route that many are empty. And the code is available on request and the source data are attached. The attitude toward reproducibility is good.

Points to Read With Care

① This is not a measurement of a current mRNA vaccine. This is the point most easily misunderstood, so I put it first. The ionizable lipid used was DLin-MC3-DMA (the lipid brought into practical use in the siRNA drug patisiran), not ALC-0315 (BNT162b2) or SM-102 (mRNA-1273) from the COVID-19 vaccines. The mRNA, too, was a commercially available dye-labeled mRNA in which 25% of the uridines were replaced with a Cy5-labeled form, a different thing from the N1-methylpseudouridine-modified mRNA used in the vaccines. The dialysis was also a laboratory-scale 12-hour batch, unlike industrial tangential flow filtration. Summarizing this as “80% of the vaccine was empty particles” is a misreading of this paper.

② No guarantee is shown that the fluorescence is linear. The deconvolution depends entirely on the assumption that “the brightness of a particle holding n copies is n times the single-copy distribution.” But dyes self-quench when crowded, and energy transfer occurs when they are close together. Whether Cy5 really shines n times as brightly when multiple mRNA molecules are packed into a lipid core is not verified in the paper. If quenching occurs, the payload is underestimated. A control with a known amount of mRNA encapsulated, or an experiment probing for quenching, would have made this assumption far stronger.

③ The units of the particle number concentration do not add up. Table 2 gives the number concentration of mRNA-loaded LNPs as 1.29 × 10¹⁶ ± 2.22 × 10¹⁵ mL⁻¹. But estimating from the amount of mRNA put in: mRNA at 20 µg/mL with 1929 nt (molecular weight roughly 6.4 × 10⁵) comes to about 1.9 × 10¹³ molecules/mL, and at 85.6% encapsulation efficiency with 2.80 copies per particle that is on the order of 6 × 10¹² particles/mL. The printed value is about 2,000 times larger. Even if the unit were meant to be L⁻¹, a discrepancy of about 2.2-fold would remain, so we cannot establish that this is simply a misprinted unit either. The details of the calculation are in the supplementary material (Supplementary Discussion 2–5) and cannot be followed in the main text, so I will not assert anything; but if you read only the main text and quote the figure, you will be off by orders of magnitude. Note that the ratio of empty to mRNA-loaded (4.88 × 10¹⁶ versus 1.29 × 10¹⁶ ≈ 79% versus 21%) is consistent with 77% / 23%, so read as a ratio, there is no problem.

④ There is a self-contradictory sentence in the main text. Immediately after arguing that three-color discrimination is necessary, the paper writes: “the two-color discrimination method was therefore used for the calculation of the payload distribution and capacity, in order to obtain higher accuracy.” From the context it is natural to read this as a slip for “three-color discrimination was used,” but the main text does not let us settle which method produced the principal results, which hurts in a paper whose selling point is counting. Similarly, the laser wavelength is given as 647 nm in the Results and 640 nm in the Methods (from the product name, 640 nm is the correct one). The legend of Figure 7 also points to the liver result as “panel i” when it should be panel l. None of these overturns the conclusions, but they are flaws that proofreading should have caught.

⑤ The label “before dialysis” conflicts with the Methods. Let me raise again, as a limitation, the point touched on above. Table 2 calls the pH 4.0 sample “before dialysis (= the initial LNP),” but the Methods section states explicitly that the pH 4.0 sample was also dialyzed for 12 hours (against acetate buffer, to remove the ethanol). What is being compared is not whether dialysis was done, but the pH it was dialyzed into. The paper’s claim — that the particles rearrange once the pH rises and deprotonation occurs — is not shaken by this correction, but taking the phrase “before dialysis” at face value leads you to think they measured particles with the ethanol still in them. The table and the Methods should have used the same words.

⑥ The “mechanism” is largely a hypothesis. The splitting-and-fusion storyline of Figures 4 and 6 is, exactly as the paper itself writes with hypothesized and presumably, an inference from population statistics — proportions, geometric means of TMR, and number concentrations. The moment at which an individual particle splits or fuses was not observed with time resolution. There is supporting consistency with earlier FRET and cryo-EM reports, but the gap between saying it must be so and saying it is so remains.

⑦ n is small, which leaves the negative conclusions weak. Most conditions are n = 3 (only the benchmark is n = 6). To make a claim that there is no difference — “concentration does not affect payload,” “mRNA size does not affect the ceiling size” — at n = 3, one would want an equivalence design and a statement of statistical power.

⑧ The gap between the two encapsulation-efficiency values is left unfilled. For the roughly 9-point difference between RiboGreen’s 94.2% and CICS’s 85.6%, the paper only sets them side by side and never discusses the reason. Which is closer to the truth, and whether the difference comes from detection efficiency or from dye access, bears on the credibility of the counting itself.

⑨ The interpretation of the animal experiment is confounded. The N/P = 3+3 group is a group to which empty LNPs were added. That means the total amount of lipid administered also went up. If the drop in liver expression is to be explained by competition for ApoE, a control with matched total lipid is needed — for instance a group given the same amount of an unrelated liposome in place of the empty LNPs. The authors are right to avoid asserting, but the reading that “empty particles get in the way” cannot be established from these data alone. On top of that, n = 3 to 4, a single 12-hour time point, females only, and a single strain.

⑩ What is being measured is the state after dilution and after standing for more than an hour. CICS dilutes the sample, incubates it with YOYO-1 in the dark for more than an hour, and then takes 20 minutes to measure it. Since an LNP is not a rigid structure at equilibrium, whether its state shifts during the sample preparation for the measurement is something one would really want checked (demonstrating a storage-stability measurement is a good sign, but it is not a validation of the preparation itself).

⑪ The link between payload and function is missing. Even knowing “how many copies per particle,” how that connects to endosomal escape or to the amount of translation has not been measured. For a particle carrying 2 copies and one carrying 8, how many times greater is the protein production when a single particle enters a cell? Only once that is filled in does the empty fraction become complete as a design metric.

How This Could Have Been a Higher-Quality Study

• Test for quenching. Confirm linearity with standard particles that encapsulate a known copy number of mRNA, or with mRNA of differing Cy5 labeling density. That would firm up the foundation of the deconvolution
• Run the same measurement on the current vaccine formulations, ALC-0315 and SM-102. This is the biggest extrapolation target of the present study, and the number readers most want
• Cross-check numbers and units mechanically. The unit of the particle number concentration, the two-color/three-color description, the laser wavelength, and the figure cross-references could all have been caught with a pre-submission checklist
• Observe the mechanism directly. Capture splitting and fusion in real time with stopped-flow plus time-resolved FRET, or with a microfluidic channel that can follow the dialysis process
• Match the total lipid dose in the animal experiment. If you want to claim an effect of empty LNPs, a control group with matched lipid is required. Likewise a time course after dosing, and multiple organs
• Link payload to expression at the single-cell level. Make up particles sorted by payload, and measure uptake number and protein production at the same time

Dr. Exotaro’s Perspective

A Corresponding Author Who Studies the Same Disease I Do

Let me be honest. The formulation physical chemistry of mRNA-LNPs is not my own specialty. What I work on is the treatment of neurological disease, spinal cord injury above all, using mesenchymal stem cells (MSCs) and the extracellular vesicles (EVs) that come out of them.

Even so, one of the reasons I chose this paper was that the laboratory of the corresponding author, Mao, is actually doing regenerative medicine for spinal cord injury. As noted above: the 2020 Biomaterials work injecting a nanofiber-hydrogel composite into rat spinal cord contusion, and then the 2022 Cells treatment of spinal cord injury that combined MSCs with that composite. Research that counts nanoparticles one by one and research that delivers MSCs to an injured spinal cord live under the same roof. I do not think that is a coincidence. Both are the two ends of the same question: “how much do we actually know about what is inside the carrier?”

The Heaviest Lesson — We Talk Too Much in Averages

What this paper thrust at me was neither the method nor the numbers, but the way the question is framed.

Recall the numbers we write in EV papers. By nanoparticle tracking analysis, “so many particles per mL.” In a separate tube, RNA is extracted and we get “so many ng of RNA,” or “so many copies of this miRNA.” Then we divide and say “so many copies per particle.”

This is precisely the “bulk-average payload” that this paper showed tells you nothing. An average made by dividing a population’s total by the population’s count. It does not answer “which particle holds how many.” (To be clear — this is not the same thing as encapsulation efficiency. Encapsulation efficiency is a fraction on the mRNA’s side, the proportion of the mRNA put in that got encapsulated, and not a per-particle count obtained by division. What this paper demolished was the habit of talking in averages.)

The EV field already had its warning. Chevillet et al., in PNAS 2014, volume 111, number 41, page 14888, analyzed the miRNA content of exosomes quantitatively and stoichiometrically and reported that most individual exosomes do not carry a biologically meaningful number of miRNA molecules. For a particular miRNA whose per-particle average comes out below 0.01 copies, what “an average of 0.01 copies” means is not “every particle holds 0.01 of one,” but “the great majority of particles are empty-handed, and only a very small fraction hold one or more” (since the mean alone does not fix the shape of the distribution, one cannot declare that “one particle in 100 holds exactly one”). Structurally, this is the same story as 77% of LNPs running empty.

Let me write this down as something I myself need to reflect on. When I attribute the therapeutic effect of MSC-EVs to a particular miRNA, I have not measured what percentage of the particles actually carry that miRNA. What the authors of this paper did for synthetic particles, we have not yet managed to do for particles of biological origin.

But the Numbers Do Not Transfer — Only the Question Does

Let me write this part carefully. LNPs and EVs are different things. An LNP is a synthetic particle that is designed and assembled; its composition is known and its size distribution can be controlled. EVs are made by living things, so they differ particle by particle, with membrane proteins such as tetraspanins embedded in the membrane and a layer of protein covering the surface. There is no basis anywhere for applying the figure “77% empty” to EVs.

What transfers is the question and the methodology. And in fact the methodology side is unexpectedly close. CICS is a technique that “flows particles single file through a capillary, classifies them by which colors light up together, and estimates numbers using a known single-molecule distribution as the basis set.” Single-EV analysis, now spreading through the EV field, and high-sensitivity flow cytometry using microfluidic channels lie in the same direction from the same idea.

👦 Student: So we should just do the same thing for EVs, then.

🧬 Dr. Exotaro: We should, and I want to. But let me add one caution to myself that I learned from this paper. What let this paper make its strongest claim was that it measured both before and after dialysis. Had they measured only the finished product, it would have ended with the bare fact that “80% are empty” and never reached “why it turns out that way.” The same holds for EVs: if we only ever measure the finished product recovered from culture supernatant, we will never know why the distribution came out the way it did. The stage at which the cell loads the EV, the stage of release, the stage of recovery and purification — we need to measure, step by step, where what gets decided. Right now we are measuring almost nothing but the finished product.

The Distance to the Destination Is Still Unclosed

Finally, as a task of my own. What this paper’s animal experiment showed was delivery to the liver and the spleen. Those are the organs an intravenously injected nanoparticle reaches most easily, and it rides a known mechanism in ApoE. What I want to deliver to, by contrast, is injured neural tissue on the far side of the blood-brain barrier and the blood-spinal cord barrier.

Being able to count a particle’s contents accurately and having that particle arrive are two different problems. Even so, in order of operations, what cannot be counted cannot be optimized. Now that a number like “2 copies per particle” exists, the discussion of “how many particles, then?” and “how much lipid does that make us carry?” has at last become a calculation rather than a guess.

And I trust the way this paper is written. They write “hypothesized” and “presumably” for their own mechanistic account, and “further investigation is needed” for their interpretation of the animal experiment. They do not mix what was measured with what was surmised. It is unglamorous, but I think it is the right attitude for a paper that sells counting.

One more note: this article is a companion piece to “The mRNA Vaccine’s ‘Booster’ Was the Container, Not the Contents,” published the same day. This one is the story that “most LNPs are empty,” and that one is the story that “it is precisely those empty particles that set off innate immunity.” Read side by side, they get a notch more interesting.

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