Table of Contents

E. coli reads per million across the eleven Achs sequencing lots, log scale, splitting by brand

Source documents. Achs et al., "Systematic analysis of COVID-19 mRNA vaccines using four orthogonal approaches demonstrates no excessive DNA impurities", npj Vaccines 10:259, 13 December 2025, DOI 10.1038/s41541-025-01304-9, with the sequencing deposit PRJNA1303864. The published method critique is McKernan, Speicher and Rose, "Systematic methodological flaws in DNA contamination assessment of mRNA vaccines: a critical analysis of Achs et al. (2025)", Oncotarget 2026;17:368-376. Every number in the reanalysis section below came off the public deposit with open tools, minimap2 and samtools, with predictions written into the script headers before each run. Any reader can check all of it.

Why this article exists

On 13 December 2025, npj Vaccines published a paper whose title is its conclusion: no excessive DNA impurities. It has since been cited to regulators as reassurance that residual DNA in Comirnaty and Spikevax is within limits, the fragments are small, and nothing derives from host genomes.

The methodological critique of that paper already exists and is peer-reviewed. Kevin McKernan, David Speicher and Jessica Rose published it in Oncotarget in August 2026, after the authors ignored the same points on the preprint server in September 2025 and a critique preprint in October 2025. That critique stands on its own. I do not repeat it here beyond a summary.

What I did was the boring, systematic version of the simplest test. Kevin's team had already used the deposit's coverage plots to expose the KAN-versus-spike qPCR problem, and as he put it, they got distracted by the rest of the technical pile before the Oncotarget paper went out. So I downloaded all fifteen runs Achs et al. deposited at NCBI under PRJNA1303864, re-mapped them, and checked every conclusion in the paper against the reads, lot by lot. This article is that pass.

Credit where it is due, and here it is due almost everywhere. The critique architecture, the qPCR and fluorometry and capillary electrophoresis arguments, the fragment-length platform argument, and the RNA:DNA hybrid mechanism belong to McKernan, Speicher, Rose, Rixey, Buckhaults, Kämmerer, Steger, König and Fleming. Kevin has publicly credited me for the starting observation on those coverage plots, and the full reanalysis package went to him before this article went up. The one thing I claim here is the read-level reanalysis of the deposit itself: the E. coli screen, the topology work, the fragment-length contrast, the mixture scan. Nothing else is mine.

What the paper claims, in its own words

Two sentences carry the headline claims:

"Remaining reads in very low relative amounts either mapped to Escherichia coli K12 genome (GenBank accession NC_000913.3; non-pathogenic strain widely used in microbiology research and biotechnologies), bacteriophage discovered in microbiome study (PV093624.1...) or low complexity reads ... these reads can at least partially be a result of reagent contamination."

"83.52-98.08% of the DNA sequences were derived from plasmids used as templates during in vitro transcription of mRNA, rather than from exogenous sources or host genomes. These findings confirm that the detected DNA is a known and expected by-product of the manufacturing process, not an evidence of unintended contamination."

Read those two together. E. coli K12 is the production host of the DNA template plasmids. Reads mapping to it are host-genome reads by definition. The paper files them under reagent contamination and then concludes no host genomes. Both halves cannot be right, and the deposit decides which one is.

The method critique, summarised

The full version is the Oncotarget paper and Kevin's article series. The eight points:

#The flawThe effect
1qPCR amplicons sized 63 bp (KAN) to 233 bp (SPIKE) against a fragment population their own data put at a 130-201 bp medianThe 233 bp spike target is mostly unamplifiable; the spike:backbone 4-6 Ct offset seen in four independent labs (McKernan, Speicher, Buckhaults, Fleming) disappears into a size artifact
2Copy-to-mass conversion applied full-plasmid molecular weight to a fragmented single-target countThe only coherent readings are ~0.07 ng/dose (fragment mass) or 12.2 ng/dose (fragmentation-corrected, over the 10 ng limit). The published 8 ng sits between the two
3No fragmentation correction anywhere in the qPCR chainEven the 63 bp KAN amplicon needs a 1.5x correction at their own median fragment length; none was applied
4Fluorometry switched to a DNA-purification front end no prior lab used, then the methods were edited post-hoc (ladder changed, no recovery number moved)Recovery never validated by size, in a regime where kits lose sub-200 bp fragments preferentially
5Capillary electrophoresis with electrokinetic injection, no matrix controls, standards in clean bufferAn LNP matrix that cannot inject reads as "below detection" regardless of true content
6The CE instrument's smear floor is 50 pg/uL; the 10 ng per 300 uL dose limit equals 33.3 pg/uLThe instrument is blind across the entire band where compliance is decided
7Fragment sizing done on Illumina after 95C x 10 min, two SPRI cleanups and 10 PCR cycles, in a lab that owns an ONT instrumentThe measured "130-201 bp" medians describe the library prep, not the vials; ONT work from the same field shows kilobase fragments
8Qubit trusted to titrate DNA input for their own Illumina runs, dismissed as unreliable when other labs reported over-limit DNA with itBoth statements appear in the same paper

Jessica Rose's Lander-Waterman treatment gets to the same place as points 2 and 3 by a different road: detected KAN copies represent roughly 37 percent of the true KAN-bearing fragment population at conservative boundary assumptions, which pushes the corrected mass to around 25 ng per dose. Different boundary assumptions, same direction. Every coherent correction lands above the 10 ng limit.

Amplicon survival under random fragmentation, e to the minus A over L

Four things the discussion never touches: the 10 ng limit descends from naked-DNA kinetics and was never constructed for LNP-encapsulated DNA; Moderna's own patents teach against qPCR for this measurement; Moderna's own patents state residual DNA has the potential to be oncogenic; and Georgiou et al. 2006 documented PicoGreen underestimating fragmented DNA by up to 70 percent, which the fluorometry numbers inherit unexamined.

The 5lovak Academy of 5cience (McKernan's coinage)

The section title is Kevin's joke, not mine, and it is doing serious work. The conflict-of-interest record below is his, assembled with receipts collected mostly by a researcher posting as Miroslav. I summarise it because the method critique and the deposit analysis cannot be read without it.

The work was commissioned and paid for. Kevin documents approximately €350,000 paid for the study, funded through the Government of the Slovak Republic via the Ministry of Health. The commissioned institute is part of the Slovak Academy of Sciences, and its Biomedical Research Center entered a formal agreement with Sensible Biotech, a commercial mRNA company, on 28 August 2024, with shared laboratory infrastructure. Slovak government and EU investment in the institute's vaccine-capable infrastructure runs to roughly €25 million since 2014. The receipts are public: the institute's annual reports and the Slovak public-procurement register (crz.gov.sk) entries are linked from Kevin's article.

The problem is the dual role. The communicating author of the paper, who declared no conflicts of interest, is also, per Kevin's documentation, the institute director whose signature is on the collaboration agreement with the mRNA company. The conflict was raised in preprint commentary in September 2025, four months before publication. It is absent from the published paper.

None of the COI material is needed to refute the paper's conclusions. The methods and the deposit do that on their own. What the COI record explains is why every methodological error in the paper points the same way: underestimation. A paper with eight independent biases that all err toward the sponsor's preferred answer is not a paper with eight unlucky accidents.

What the deposit itself shows

I re-mapped eleven libraries, SRR34932925/926/927/928/929/930/931/932/933/937/938, and extended the fragment-length comparison to all fifteen runs. What I ran: vector-unmapped reads re-mapped with minimap2 -ax sr against E. coli K12 NC_000913.3 (4,641,652 bp) and phage PV093624.1, primary alignments only, MAPQ >= 20; topology from full-depth proper-pair counts with a mid-reference control; depth with samtools depth -a -q 20 -d 0. Predictions went into the script headers before each run, including the ones that failed.

1. The host genome they said was not there

In four of five Comirnaty lots (925, 930, 931, 932), 1.4-2.4 percent of ALL sequencing reads map across 96 percent of the complete E. coli K12 genome at 56-95x mean depth. That is 13,700 to 23,700 reads per million. All six Spikevax lots sit at 8-472 rpm. Every Comirnaty lot is above every Spikevax lot. The spread across brands is 3,100x.

Both brands were processed in the same lab with the same reagents and the same pipeline, and the vector-mapped loads overlap between brands (Comirnaty 106-127M vs Spikevax 80-115M reads). A shared-reagent kitome cannot produce a brand split. The phage PV093624.1 runs the other way (28-67 rpm in Comirnaty, 151-762 rpm in Spikevax): that is what an actual reagent component looks like, and it is uncorrelated with the E. coli signal.

Depth per 100 kb bin along the whole E. coli genome for three representative lots

Coverage is flat along the entire 4.64 Mb genome in 100 kb bins. That is a genomic DNA population, not a sparse kit contaminant. The fifth Comirnaty lot, 937, sits seven to twelve times below its siblings but still above every Spikevax lot. This is manufacturing carryover, co-purified host genomic DNA from plasmid propagation, consistent in direction with Kevin's ONT finding of E. coli gDNA in Pfizer bivalent vials. "Rather than from ... host genomes" is refuted by the paper's own reads.

2. Whole circular plasmid, not only small fragments

Cross-junction read pairs, one read in the first 800 bp of the reference and the mate in the last 800 bp, proper pairs, MAPQ >= 20, full depth, mid-reference control, are nonzero in every lot. The circular molecule fraction is ~1e-3 in all Comirnaty lots (1.4-2.7e-3) and the pv-vector Spikevax lots (1.0-9.7e-3), versus 7e-6 to 1.5e-5 in the two original-vector Spikevax lots.

Comirnaty coverage is flat across the whole plasmid and the T7 promoter is depressed, not peaked (T7:spike depth 0.58-0.77 in all five lots). The Spikevax enriched lots show the promoter at or above spike (1.03-1.54). Comirnaty spike:kanR sits at 0.43-0.64 across all nine Comirnaty runs in the corrected fifteen-run table, which reproduces the phase-2 numbers cross-platform.

Circular fraction and T7-to-spike ratio per lot

So the Comirnaty residual DNA is whole plasmid, with a circular fraction two to three orders of magnitude above the original-vector lots, plus host genomic DNA. It is the debris of an under-digested E. coli plasmid prep. The original-vector Spikevax lots are linear, cassette-bounded molecules. The paper's "small fragments" describe its own 95C x 10 min library prep, not the vials, and "small fragments only" is false for the paper's own Comirnaty data.

3. The RNA:DNA hybrid footprint, in their own fragment lengths

Within-lot contrast, same library, same size selection: median spike-region fragment length minus median kanR-region fragment length. Every lot with continuous backbone coverage is positive, +6 to +26 nt. All nine Comirnaty runs sit at +6..+20 (sign test p = 0.004). The spike-rich pv lots 926 and 928 sit at +26 and +25. The three lots whose coverage collapses at the cassette boundary are the three negative ones (929/933 at -17, 938 at -21).

Spike-region minus kanR-region median fragment length per lot

Longer, heavier spike-region DNA exactly where spike is enriched is what the hybrid-protection mechanism predicts: RNA:DNA-hybrid-protected spike DNA takes fewer DNase cuts. The mechanism and the prediction are McKernan, Rixey and Rose's, and BioNTech's own scientists published the hundredfold digestion-resistance of RNA:DNA hybrids (Lenk et al. 2024) a year before Achs. What I added is the measurement inside the Achs libraries. The 4-6 Ct spike enrichment that four independent labs measured, and that the Achs amplicon sizing erases, is sitting in the deposit's own fragment lengths.

4. The mixture scan, and why the brand split is not an artifact

A minor-allele scan (>= 3 percent minor allele, MAPQ >= 20, baseQ >= 25, mpileup depth cap set above true depth) returns zero MIX rows in all five Comirnaty lots and both original-vector lots, and 29-68 MIX rows in every pv-vector lot, at the same shared spike positions with lot-varying fractions (926 near 50/50, 928/929/933 at 10-18 percent). A real second DNA population per pv lot, coherent within each lot.

The zero in every non-pv lot matters beyond the mixture finding: it bounds cross-lot contamination in the sequencing pipeline below 3 percent. That closes the standing objection that the E. coli brand split is index hopping or library carryover. It cannot be, because there is no cross-lot signal where a mixed-up pipeline would show one.

Standing caveats

Every one of these stays in.

  • All fifteen runs are downstream of the authors' 95C x 10 min step. Fragment lengths reported here are lower bounds; visibility corrections computed from them are upper bounds. Their prep degrades the very evidence their conclusions rest on.
  • Per Kevin's T4-ligase rule, spike:kanR and ecoli:plasmid read ratios are sequencing ratios, not in-vial mass ratios. Hybrid-protected spike is suppressed in these libraries, so sequencing under-states spike enrichment. That bias is shared by both brands and cannot manufacture a 3,100x brand split.
  • Strain identity is resolved to K12/MG1655 level only. Production strains are K12 derivatives; strain-level attribution is not resolved by this screen.
  • I am not converting the host-genome read share into ng/dose. The only anchors available are the paper's fluorometry and qPCR, and both are the contested methods. The anchor argument has to be won before the conversion means anything.

The one-paragraph version

The method critique already stood on its own. The deposit adds this: the paper's own sequencing data contain near-complete E. coli genomic DNA in four of five Comirnaty lots at 56-95x depth, whole circular plasmid in every lot, and a fragment-length signature consistent with RNA:DNA hybrid protection, while the reported qPCR arithmetic converts a 12.2 ng over-limit figure into a reported 8 ng. The paper concludes the opposite of what its own data show, and it is cited to regulators as reassurance.

Every claim in the reanalysis section is checkable from PRJNA1303864 with public tools. The scripts, the registered predictions and the per-lot tables went to Kevin before publication. The deposit is the authors' own.

Sources

Kevin's article series on the Achs paper:

Open to corrections

Every number here can be reproduced from PRJNA1303864 and every claim carries its source. If I have something wrong, tell me and I will correct it, and the correction gets the same prominence as the original claim. If the authors want to respond point by point, I will publish their response here, unedited. Standing offer, no expiry. General policy on the Methodology page.