---

Version of record: 10.5281/zenodo.21799862, published 5 August 2026. That identifier is the citable address for this paper and it resolves at https://doi.org/10.5281/zenodo.21799862. It is a version identifier; Zenodo minted a second one that resolves to all versions, and the version identifier is the one to cite.

What this file is. This project's authoritative copy of the manuscript, SUBMIT_THESE/papers/PUBLISH_6_UPREGULATION_CEILING.md, which is the file the deposited PDF was built from. Synced 7 August 2026 by scripts/sync-manuscripts.mjs, which copies the source byte for byte and prepends this note. Nothing in the manuscript below has been rewritten for the website.

What was here before, because nothing on this site is deleted quietly. Until 6 August 2026 this file was a copy taken before the corrections of 4 August 2026 evening and never resynced, so it served the retired 34.1 percent rescaling of the measured current. Earlier on 6 August 2026 it was given a banner reading "superseded revision, do not cite any figure in it". That banner was an accurate description of a stale file and a poor thing to serve on a research site, so the stale file has been replaced with the authoritative text rather than annotated. The retired rescaling divided the measured 68.3 percent by two; O'Neill 2022's own two-allele control reads 218.4 percent of a single allele rather than 200, so the divisor is 2.184, the baseline is 31.3 percent and the comparator for simple loss of one allele is 45.8 percent.

Warning, and it points at the published record rather than at this page. The version deposited on 5 August 2026 contains seven figures that are wrong. They were corrected here on 6 August 2026, after the deposit. The deposited paper carried the corrected baseline of 31.3 percent throughout and paired it with the retired comparator of 50 percent throughout. The worst of the results, 36.0 percent for the ceiling and 51.2, 68.2 and 85.2 percent in the boost table, are the retired 34.1 baseline multiplied out, sitting in a table whose own column header reads 31.3. A reader checking the arithmetic from the printed anchor got a different answer from the printed result.

A second correction, and this one is worse than the arithmetic above. The deposited version's data availability statement said this paper's derived tables were in the shared archive and named three of them: the per-sample junction and transcript classifications, the transcript biotype table, and the headroom calculations. The archive at 10.5281/zenodo.21799234 contained none of the three. It held no GTEx junction file, no transcript classification, no biotype table and no headroom table, and its own README had no section for this paper. That is a statement a reader cannot check except by downloading the archive and finding nothing there. It was found in an audit of all eleven papers' data availability statements on 6 August 2026, which found the same class of defect in six of them.

Three of the four tables were not lost, and the fourth had the same defect as this paper. SCN5A_SPLICING_MEASUREMENT.csv, jx_classification.csv and ensembl_tx_biotypes.csv were saved by the original pipeline on 4 August 2026 and simply never copied into the deposit; all three are staged for version 2. The fourth, the headroom table, existed on disk carrying the retired 34.1 and 50 anchors that the first correction above withdraws, so it was recomputed from 31.3, 45.8 and 100 and staged as UPREGULATION_HEADROOM_CORRECTED.csv. The retired file is neither deleted nor deposited. Every quantity printed in the paper was then recomputed from the deposited files by p6_verify_and_headroom.py, and 116 of 119 reproduce at the precision printed. The three that do not are recorded in the paper rather than quietly adjusted, and the largest of them is a misattributed Mann-Whitney p-value that understated this paper's own result.

No conclusion in this paper changes. The measured non-productive fraction of 0.0045 percent and the generous transcript-level upper bound of 5.38 percent are junction and transcript measurements, and the baseline question never touched them. The 1.46-fold requirement is unchanged, and it is worth saying why rather than leaving it looking untouched: the required boost is the ratio of the one-allele level to the measured heterozygous level, so the scale factor cancels out of it. It was right for the wrong reason before and is right for the right reason now.

If a figure on this page disagrees with the same figure at the identifier above, this page is the corrected one. The full divergence, and what a version-2 deposit would have to include, is recorded in SUBMIT_THESE/ZENODO_DIVERGENCE_20260806.md. (This paragraph used to open by asserting that no version 2 had been deposited and nothing had been uploaded. That was true when written on 6 August 2026 and is not a claim a generated page can keep true, because it would turn false the moment anything is deposited and nothing here would notice. The sentence is removed rather than updated: to find out what is deposited, resolve the identifier, which is the only source that cannot go stale.)

None of this is peer reviewed, and none of it has been through a wet lab. No cell has been edited and no current has been recorded for this variant by this project. Every therapeutic statement in the manuscript below is a prediction.

---

SCN5A lacks the non-productive mRNA reserve that antisense upregulation therapy would need in human heart

Ethan Bradley

Independent researcher, no institutional affiliation

ORCID: 0009-0008-8925-7975

Abstract

Raising total SCN5A output by blocking non-productive mRNA splicing is not a viable strategy in human heart, because the material this strategy needs to redirect is essentially absent from cardiac tissue. Splice-junction read counting across 827 GTEx v10 heart samples (371 left ventricle, 456 atrial appendage) puts the non-productive fraction of SCN5A mRNA at 0.0045 percent (69 of 1,545,656 reads, 95 percent Wilson confidence interval 0.0035 to 0.0056 percent). The same method applied to SCN1A in 491 brain cortex and frontal cortex samples, the tissue and gene in which this drug class already works clinically (zorevunersen, an antisense oligonucleotide approved for Dravet syndrome, N Engl J Med 2026;394(10):969-982, PMID 41780062), gives 1.388 percent. The gap is 308-fold, and the separation between the two per-sample distributions is overwhelming by a one-sided Mann-Whitney test (junction method p = 4.0e-219 with tie correction, 1.3e-168 without; transcript method p = 1.6e-160). Even under a generous transcript-level estimate of 5.38 percent non-productive SCN5A mRNA, redirecting all of it with perfect efficiency raises output only 1.057-fold. The R104Q variant, whose heterozygous current is 31.3 percent of normal (O'Neill et al. 2022, PMID 35305865), needs 1.46-fold to reach the 45.8 percent level that simple loss of one allele would give. A frameshifting cassette exon with the structural signature this drug class targets does exist in the SCN5A annotation, but its splice junctions are undetectable anywhere in the GTEx v10 junction file. The strategy has no substrate in cardiac tissue as currently characterized.

A key to the terms used here

• SCN5A is the gene for the heart's main sodium channel; Nav1.5 is the protein. R104Q is the variant used as the worked example, arginine 104 replaced by glutamine.
• Messenger RNA, or mRNA, is the working copy a cell makes from a gene and reads to build protein. Transcription is making it; expression is how much gets made.
• Splicing is the cutting and rejoining that turns a raw transcript into a finished messenger RNA. Exons are the kept pieces, introns the removed ones, and a junction is a joint between two kept pieces.
• A non-productive transcript is one the cell makes and then discards without building protein from it: waste. Upregulation therapy aims to rescue that waste stream into working product, so the amount of waste sets the ceiling on how much extra protein is available. If almost nothing is being wasted, there is almost nothing to recover.
• A poison exon is an exon whose inclusion makes a transcript non-productive on purpose. Blocking it with a drug converts waste into product, which is the mechanism this paper tests for SCN5A. It has worked for other genes, which is why the question was worth asking.
• NMD, nonsense-mediated decay, is the cell's process for destroying transcripts carrying a premature stop signal.
• GTEx is a public atlas of gene expression across human tissues. TPM, transcripts per million, is its unit of abundance. Long-read sequencing reads whole transcripts end to end rather than in fragments, which is what would be needed to settle the question directly.
• Headroom here means how much more functional protein the mechanism could deliver at best.
• Saturation means a mechanism reaching its maximum, beyond which more drug adds nothing.

Why raising output looked like the only mechanism-independent option

SCN5A encodes the cardiac sodium channel Nav1.5. R104Q (NM_000335.5:c.311G>A, p.Arg104Gln) is a Brugada syndrome variant, and whether it acts through simple loss of function or through active interference with the co-expressed wild-type channel is not established. That question gates most candidate interventions: a folding corrector only helps if the mutant protein is retained inside the cell, an interaction-blocking drug only helps if the mutant subunit actively suppresses the healthy one. Raising total transcript output is different. It helps regardless of which mechanism is operating, because more working mRNA means more working channel either way. That property made it worth testing before the mechanism question is resolved, and it is the reason a negative result here closes something rather than merely narrowing it.

The clinical motivation is straightforward. Most Brugada carriers are asymptomatic and do not meet implant criteria for a defibrillator, so they currently have no disease-modifying protection at all, and a defibrillator in any case terminates an arrhythmia rather than preventing one. A precedent for the mechanism exists: zorevunersen, an antisense oligonucleotide that increases productive SCN1A transcript by blocking non-productive splicing, treats Dravet syndrome, a disease caused primarily by SCN1A haploinsufficiency, and reached publication in the New England Journal of Medicine in 2026. The question this study asks is whether SCN5A in heart has the same kind of non-productive reserve that makes that drug class work in SCN1A and brain.

Methods

Gene and tissue definitions. SCN5A, Ensembl ENSG00000183873, chr3:38548057-38649743, minus strand, GRCh38. SCN1A, Ensembl ENSG00000144285, used as the positive control gene because zorevunersen already demonstrates efficacy against it. Heart tissues: GTEx "Heart - Left Ventricle" and "Heart - Atrial Appendage". Brain tissues: GTEx "Brain - Cortex" and "Brain - Frontal Cortex (BA9)".

Sequencing data. GTEx v10, quantified against GENCODE v39. Two files were streamed in full and filtered to the two genes:

A third file, the GTEx v9 long-read isoform quantification (FLAIR pipeline, GENCODE v26 annotation, 92 samples total), was also examined and is reported separately because of a depth problem described below. No access date is recorded for these three files in the underlying records; the pipeline and reference versions above are as labelled in the file names themselves.

Sample filtering. Samples were required to pass a minimum sequencing-depth quality control flag recorded per sample in the derived counts table; the numeric cutoff for that flag is not separately stated in the records available. After filtering: junction analysis, 827 heart samples (371 left ventricle, 456 atrial appendage) and 491 brain samples (262 cortex, 229 frontal cortex); transcript analysis, 879 heart samples (420 left ventricle, 459 atrial appendage) and 528 brain samples (267 cortex, 261 frontal cortex).

Transcript annotation. Transcript biotypes and exon coordinates were retrieved from the Ensembl REST API, GRCh38, accessed 2026-08-04, for both genes. SCN5A has 21 annotated transcripts, of which 7 are non-productive by biotype: two nonsense-mediated-decay transcripts, ENST00000713730 (8,616 nt, longer than MANE Select at 8,528 nt) and ENST00000713731 (2,873 nt); four retained-intron transcripts, ENST00000491944 (1,382 nt), ENST00000476683 (774 nt), ENST00000718273 (687 nt), and ENST00000718272 (600 nt); and one protein_coding_CDS_not_defined transcript, ENST00000464652 (591 nt). SCN1A has 18 of 30 annotated transcripts classified non-productive by the same scheme. These are annotation counts, not abundance measurements, and are treated as such throughout.

Junction classification and counting. Every intron implied by every annotated transcript of both genes was derived from the Ensembl exon models, and each resulting junction was labelled by which transcripts contain it. A junction was classified non-productive-specific when every transcript containing it is non-productive. Reads spanning each classified junction were then summed directly from the STAR junction file, with no isoform-deconvolution step.

Transcript-level estimation. RSEM transcripts-per-million values were summed across the transcripts classified non-productive for each gene and divided by total gene TPM, per sample. This method requires software to assign short reads (76 to 150 bases in this dataset) to one of several highly similar isoforms, and its output was cross-checked against the junction counts rather than trusted on its own.

Coordinate convention check. GTEx junction identifiers were confirmed empirically, not assumed, to represent 1-based inclusive intron coordinates: SCN1A's MANE intron 1 was computed independently from Ensembl exon boundaries as chr2:165992423-165994145, which matched the GTEx junction identifier string chr2:165992423-165994145:- exactly.

Statistics. Wilson 95 percent confidence intervals were computed for the pooled non-productive fractions. A one-sided Mann-Whitney U test compared per-sample SCN5A-heart fractions against SCN1A-brain fractions.

Electrophysiology baseline. R104Q functional data are from O'Neill et al., 2022 (PMID 35305865, Supplementary Table 1, read from the open preprint 10.1101/2021.09.22.461398, page 29), using a Sleeping Beauty genomic-integration system in which wild-type current is undiluted: heterozygous R104Q current 68.3 ± 6.1 percent of a single wild-type allele (n = 34), homozygous 0.4 ± 0.2 percent (n = 22). The same table's wild-type-plus-wild-type additivity control reads 218.4 percent of a single wild-type allele, not 200, so the scale factor from O'Neill's single-allele units to a heart in which two working alleles equal 100 percent is 218.4/100 = 2.184. Translating on that factor gives an unaffected value of 218.4/2.184 = 100, a simple one-allele value of 100/2.184 = 45.8, and a measured R104Q value of 68.3/2.184 = 31.3 percent.

The comparator is 45.8, not 50, and the two cannot be mixed. Fifty is what the single allele would be worth if two wild-type alleles produced exactly twice the current of one; O'Neill's own control says they produce 218.4 percent of one allele. Halving 68.3 to get 34.1 makes the same assumption and is retired for the same reason. Pairing the corrected 31.3 baseline with the retired 50 comparator, or the retired 34.1 baseline with 45.8, gives arithmetic that does not close. Every derived figure below is computed from 31.3, 45.8 and 100.

The junction count is direct, and it is close to zero

The gap between the two pooled fractions is 308-fold, a figure computed from the rounded fractions 1.388 and 0.0045; from the raw read counts it is 310.9-fold. A one-sided Mann-Whitney test comparing the per-sample junction fractions, 827 heart samples against 491 brain samples, gives p = 4.0e-219 with tie correction and p = 1.3e-168 without it. The same test on the transcript-level fractions of the next section, 879 heart against 528 brain, gives p = 1.6e-160. With most SCN5A junction values exactly zero, no asymptotic p-value at these extremities should be read as a probability; what they establish is that the separation is not marginal under any convention. The distribution matters as much as the pooled number: the 90th percentile for SCN5A in both heart tissues is still 0.000 percent, and the single highest value across all 827 heart samples is 0.935 percent. There is no hidden subpopulation of high-splicing-waste hearts obscured by an average.

Sensitivity is not the explanation for the low number. All 27 annotated SCN5A MANE junctions were quantified at 19 to 107 reads per sample, with median per-sample junction depth of 1,173 to 1,947 reads depending on tissue and junction. The sequencing depth was adequate to detect a non-productive signal had one existed; none was found.

The transcript-level estimate is less reliable, and here is the reason

Pooled across heart, the transcript-level method gives roughly 4.9 percent non-productive SCN5A mRNA against roughly 40.6 percent for SCN1A in brain, an 8.3-fold gap in the same direction as the junction result. But this method has a specific limitation for SCN5A: the RSEM reference used by GTEx v10 quantifies only 3 of the 7 transcripts classified non-productive (ENST00000464652, ENST00000476683, ENST00000491944). Both large nonsense-mediated-decay transcripts, including the one carrying the poison-exon candidate described below, are absent from that reference and could not have been counted regardless of their abundance. The three transcripts that do carry the transcript-level signal have median heart abundances of 0.47, 0.32, and 0.07 TPM and attract only the same 69 junction reads across 827 samples that the direct method already counted. They are short, 591 to 1,382 nucleotides, and mostly overlap the coding transcript, which is exactly the setting in which isoform-assignment software parks ambiguous reads. The junction method does not share this limitation, because a read spanning a splice seam is evidence for that seam regardless of which transcripts are annotated. For that reason the junction fraction, 0.0045 percent, is treated as the measurement, and the transcript-level fraction, generously rounded to 5.38 percent, is treated as an upper bound rather than an estimate to be believed on its own terms.

A poison exon exists in the genome but heart does not use it

The long nonsense-mediated-decay transcript ENST00000713730 contains a 100-nucleotide cassette exon at chr3:38612802-38612901 that MANE Select splices past entirely, using one continuous intron at chr3:38609965-38613742. A 100-nucleotide insertion shifts the reading frame and destroys the protein, which is the same structural pattern zorevunersen blocks in SCN1A, and the exon shares a splice site with MANE Select, which would make it an attractive antisense target if it were used.

It is not detectably used in human heart. Neither inclusion junction, chr3:38612902-38613974 nor chr3:38609965-38612801, appears anywhere in the GTEx v10 junction file spanning 392,955 junctions and 19,788 samples. The MANE junction that skips the exon is present at a median of 19 reads per heart sample. In direct contrast, SCN1A's non-productive signal in brain concentrates in two specific junctions, chr2:166007294-166009718 and chr2:166002754-166007229, carrying 832 and 724 reads across brain cortex samples, where the same kind of poison-exon splicing is demonstrably active. The SCN5A element is annotated but silent in the tissue that matters.

Long-read sequencing could not settle this

GTEx v9 long-read data (16 cardiac samples, full-length isoforms, no isoform-assignment guesswork) gave SCN5A non-productive fractions of 20 percent in left ventricle and 50 percent in atrial appendage, numbers that would have supported the strategy. They are not usable. Depth on SCN5A in this dataset is 1 to 10 long reads per sample, median 5.5, and technical replicates of the same sample disagree completely: sample GTEX-WY7C-1126 gave fractions of 0.20, 0.83, and 0.20 from 5, 6, and 5 reads respectively. No sample reaches 50 reads. The reference annotation used for this quantification is GENCODE v26, which predates the annotation of both ENST00000713730 and ENST00000713731 and so could not have counted the poison-exon transcript regardless of depth. This dataset is reported here for completeness and is not treated as evidence for or against the conclusion.

What the ceiling means for R104Q

A drug in this class cannot exceed the theoretical maximum of redirecting every non-productive transcript into a working one, a boost of 1/(1 minus the non-productive fraction).

Applying a range of boost factors to the 31.3 percent baseline, under the assumption that suppression scales proportionally with allele output:

The strategy needs 1.46-fold to move R104Q current to the 45.8 percent level that simple loss of one allele would give. That figure does not depend on the rounding, because the required boost is the ratio of the one-allele level to the measured heterozygous level and the scale factor cancels out of it: (100/2.184)/(68.3/2.184) = 100/68.3 = 1.464, and from the rounded anchors 45.8/31.3 = 1.463. The measurement caps the achievable boost at 1.057-fold under the most generous reading of the data and at essentially 1.0000-fold under the reliable one. What is needed is a 46 percent increase in output; what is available at the generous upper bound is 5.7 percent, a shortfall of roughly eightfold, and that is using an efficiency assumption, perfect redirection of every non-productive transcript, that no real drug achieves. Under the reliable reading the available increase is 0.0045 percent and the shortfall is four orders of magnitude.

The saturation risk that outlives this negative

The headroom table above assumes that proportional changes in both alleles produce proportional changes in surviving current. That response law is untested for the proposed intervention. A fixed absolute suppressive amount alone does not imply that raising both alleles lowers current: for a positive unsuppressed contribution A and constant decrement C, I(λ)=λA−C increases with λ. Suppression proportional to mutant abundance likewise scales with both alleles under a constant-parameter homogeneous model. Worsening would require an additional assumption, such as superlinear suppression, unequal changes in WT and mutant availability, altered activity, or changing coupling. None is established by the fixed-amount premise. Therefore the perturbation path and current response need experimental qualification; neither worsening nor safe, effective rescue follows from the mechanism label alone.

What this search covered, and what it did not reach

This was a bounded search of specific named resources, not an exhaustive one. Reached and used in full: the GTEx v10 junction file, the GTEx v10 RSEM transcript file, the GTEx v9 long-read file, GTEx sample annotations, and the Ensembl REST API for transcript biotypes and exon structures. Not reached: recount3 (host duffel.rail.bio was not accessible from the environment used) and a generic Google Cloud Storage endpoint; no mirror was attempted for either. recount3 would have added an independent junction quantification from cohorts outside GTEx, which is the single most useful check not performed here. ENCODE heart RNA-seq was reachable but was not pulled, because GTEx already supplied roughly 900 heart samples at far greater depth than the long-read alternative and ENCODE's cardiac sample count is considerably smaller. Intron retention was not quantified directly from base-level coverage; the junction method captures splice-seam evidence only, and a direct coverage analysis of the four retained-intron biotypes specifically was not performed.

What would falsify this

1. Tissue and developmental stage. GTEx is post-mortem bulk tissue from adults. If the poison exon or other non-productive splicing is used in fetal or diseased heart, or concentrated in a cell type diluted out in bulk tissue, this measurement would miss it. Single-nucleus or fetal cardiac RNA-seq showing meaningful poison-exon inclusion would overturn the conclusion.
2. Turnover masking flux. Nonsense-mediated-decay transcripts are degraded quickly, so a low steady-state level can hide a high rate of production. The decisive experiment is cardiomyocyte RNA-seq with nonsense-mediated decay inhibited, for example by UPF1 knockdown; a sharp rise in SCN5A poison-exon inclusion under that condition would mean this measurement understates the available substrate. This is a wet-lab experiment and was not performed here. Retained-intron transcripts are not degraded by this pathway and were also near zero, and the same turnover caveat applies to SCN1A in brain, where the corresponding signal is nonetheless 308-fold higher.
3. Annotation dependence. The non-productive set used here comes entirely from Ensembl biotype calls. An unannotated poison exon would not be counted by this method, though the junction file does contain unannotated junctions and a targeted search of those was not carried out.
4. Structural blindness of the junction method. A retained intron produces no junction read at all, so this method cannot see intron retention directly and can only undercount terminal splicing events. The transcript-level estimate partially covers this gap and it, too, came out low.
5. Baseline tissue. The 31.3 percent R104Q current figure comes from a heterologous cell line, not human heart. Whether the underlying dominant-negative mechanism, if one exists, operates the same way in cardiac tissue is unresolved and is not addressed by this measurement.

Correction, 6 August 2026

The 31.3 percent baseline replaced an earlier 34.1 across this project on 4 August 2026. This paper took the new baseline but kept the comparator that belonged to the old one, 50 percent, and several of its derived figures were still the products of 34.1. That mixture is not a rounding difference; it produces numbers that are wrong. All derived figures were recomputed from 31.3, 45.8 and 100 on 6 August 2026.

The 1.46-fold requirement is unchanged, which is a coincidence worth stating rather than hiding: the ratio of the one-allele level to the measured heterozygous level is 100/68.3 whether or not the scale factor 2.184 is applied, because it cancels. Everything else moved. The three figures in the R104Q column of the boost table were the clearest failures — 51.2, 68.2 and 85.2 are 34.1 multiplied by the boost, not 31.3, so the table's own header disagreed with its own contents.

Nothing in the conclusion changes. The measured non-productive fraction, 0.0045 percent, and the generous upper bound, 5.38 percent, are junction and transcript measurements and were not affected. The gap between the boost this strategy could deliver and the boost R104Q would need is smaller than it looked in one respect — the target is 45.8 rather than 50 — and remains roughly eightfold at the generous bound and four orders of magnitude at the reliable one.

This section is not in the record deposited at 10.5281/zenodo.21799863 on 5 August 2026, which carries the pre-correction arithmetic.

Correction, 6 August 2026, second: the data availability statement was false, and checking it found three more things

Version 1 of this paper said its derived tables were deposited in the shared data archive. They were not. The statement named three of them — the per-sample junction and transcript classifications, the transcript biotype table, and the headroom calculations — and the archive at 10.5281/zenodo.21799234 contained none of the three. It held no GTEx junction file, no transcript classification, no biotype table and no headroom table, and its own README had no section for this paper. That is a statement a reader cannot check except by downloading the archive and finding nothing there, and it is the most serious defect in this paper, worse than the arithmetic corrected above. It was found in an audit of all eleven papers' data availability statements on 6 August 2026, which found the same class of defect in six of them.

Three of the four tables were not lost. They were saved by the original pipeline on 4 August 2026 and simply never copied into the deposit; they are SCN5A_SPLICING_MEASUREMENT.csv, jx_classification.csv and ensembl_tx_biotypes.csv, and all three are deposited now. The fourth, the headroom table, existed on disk in a form that carried the retired 34.1 and 50 anchors this paper's first correction section withdraws, so it was recomputed from 31.3, 45.8 and 100 and deposited as UPREGULATION_HEADROOM_CORRECTED.csv. The retired file is not deleted and not deposited.

Every quantity printed in this paper was then recomputed from the deposited files by p6_verify_and_headroom.py, and 116 of 119 reproduce at the precision printed. Three do not, and all three are recorded here rather than quietly adjusted.

Nothing in the conclusion changes. The misattributed p-value is the only one of the three that touches an argument, and it moves in the direction of a larger separation, not a smaller one.

One documentation gap was found and is not a defect. The sensitivity paragraph's "median per-sample junction depth of 1,173 to 1,947 reads" mixes two sample sets: 1,947 is the left-ventricle median after the depth quality-control filter, and 1,173 is the heart-pooled median before it, across 913 samples rather than the 827 the section is about. Both endpoints reproduce. The point being made — that depth was adequate — is unaffected.

Four claims in this paper are not checkable against any deposited file, because the per-junction read matrix and the per-transcript abundances were aggregated away before saving. They are named in the data availability statement below rather than left for a reader to discover.

Correction, 7 August 2026, third: this paper claims mechanism-independence for a class it did not measure, and then retracts the claim in its own discussion

Lead with what does not change, because it is almost everything. The title is correctly scoped and stays. The abstract is correctly scoped and stays. The measurement stands: the non-productive fraction of SCN5A mRNA in 827 GTEx heart samples is 0.0045 percent, the positive-control figure in brain is 1.388 percent, the shortfall is 308-fold, and the ceiling on splice redirection is 1.057-fold against a 1.463-fold requirement. Splice redirection is closed and this correction does not reopen it. What is withdrawn is a framing claim in the introduction, and the reason it is withdrawn is that this paper already contradicts it further down.

The sentence. The section headed "Why raising output looked like the only mechanism-independent option" reads:

"Raising total transcript output is different. It helps regardless of which mechanism is operating, because more working mRNA means more working channel either way. That property made it worth testing before the mechanism question is resolved, and it is the reason a negative result here closes something rather than merely narrowing it."

Two defects, and they compound.

First, the paper contradicts itself, and the contradiction is internal rather than a divergence from anything. The section headed "The saturation risk that outlives this negative" says the opposite in its own words: the proportionality assumption behind mechanism-independence "is untested", and if the mutant subunit interferes with a fixed absolute amount of channel complex rather than a fixed fraction, "raising expression from both alleles delivers more interfering protein alongside more working protein, and the intervention could lower current rather than raise it". A strategy that could lower current under one branch of the mechanism fork does not help regardless of which mechanism is operating. The introduction asserts mechanism-independence as the property that made the question worth asking, and the discussion withdraws it. Both statements are in the deposited record at 10.5281/zenodo.21799863. This is the same defect class already found in paper 2 of this series: an internal contradiction that a diff against the deposit cannot surface, because the deposit reproduces it faithfully.

7 September 2026 interpretation correction: The preceding quotation records the earlier manuscript's internal contradiction, not a valid derivation of worsening. The fixed-absolute-suppression premise does not imply decreasing current under proportional upregulation; the current discussion above now states that limitation. This repairs that argument only. It neither validates the transcript-to-rescue ceiling nor establishes the response of an actual WT/R104Q preparation.

Second, the scope claim is wider than the measurement, and it licences a wider closure than the paper earned. "Raising total transcript output" is a level of intervention. Blocking non-productive splicing is one method at that level, and it is the only one measured here. A non-productive splice fraction is a precondition for splice redirection and for nothing else. Transcriptional activation, Wnt or beta-catenin pathway inhibition, EZH2 inhibition, histone-deacetylase inhibition, microRNA antagonism and mRNA stabilisation do not require one to exist, so none of them is touched by a 308-fold shortfall in one. The clause "it is the reason a negative result here closes something rather than merely narrowing it" is therefore true of splice redirection and false of the class the sentence names.

Why this matters rather than being a wording preference, and it is measurable. This project's own route enumeration filed this paper's closure under a heading reading "Transcriptional upregulation (make more of the healthy copy)", and every later reader in that project, including its author, read the class as closed. It was not. The project's literature corpus holds a measurement in the right cell type: in cardiomyocytes derived from Brugada-syndrome patients, the small-molecule Wnt inhibitor Wnt-C59 gave 2.1-fold Na_v1.5 protein (p = 0.0005), and short hairpin RNA knockdown of beta-catenin gave 4.0-fold protein and 4.9-fold peak I_Na, replicated in a second patient line (PMID 37226398). Against the 1.463-fold requirement stated above, both are above the bar. A closure this paper did not make was read out of it for two days.

What that measurement does not do, stated so this correction cannot be read as an opening. It does not open anything. The deliverable agent, the small molecule, was never measured on current at all, and the 4.9-fold came from a knockdown that is a laboratory tool rather than a therapy. Neither patient line carries R104Q; one carries a different missense variant and the other a frameshift. The intervention is allele-agnostic, so it raises the variant message alongside the healthy one and runs into this paper's own saturation risk unresolved. Wnt inhibition is an organism-wide developmental intervention offered against a comparator of no treatment. And the evidence is one 2023 paper from one laboratory. It is recorded as conditional and it is not a route this paper endorses.

What is corrected, exactly.

One thing this correction deliberately does not do. It does not change the paper's conclusion, its title, its abstract or any number in it, and it does not claim the measurement was misreported. The measurement is right, the ceiling is right, and the closure it supports is right at the width the measurement supports. What was wrong was a sentence claiming a property the paper had not tested and later contradicted, and the cost of that sentence was paid by a downstream document rather than by this one.

This section is not in the record deposited at 10.5281/zenodo.21799863 on 5 August 2026, which carries the framing sentence uncorrected.

Data availability

Public data used: GTEx v10 junction and transcript quantification files and GTEx v9 long-read quantification files, available from the GTEx Portal; Ensembl REST API annotations for ENSG00000183873 and ENSG00000144285, GRCh38, accessed 2026-08-04; O'Neill et al. 2022 electrophysiology data, PMID 35305865, Supplementary Table 1, and the corresponding preprint, doi 10.1101/2021.09.22.461398.

The derived tables are deposited in the data archive whose identifier is recorded in DATA_DOI.txt alongside this manuscript, and they are these seven files, named individually so that a reader can check this statement against the archive rather than take it on trust:

• SCN5A_SPLICING_MEASUREMENT.csv — the per-sample junction and transcript classifications, 2,942 rows, one per sample per gene per method, carrying the non-productive signal, the total, the fraction, the unit and the depth quality-control flag. Every sample count, median, 90th percentile and maximum in the two tables above is computed from this file.
• SCN5A_SPLICING_SUMMARY.csv — the four headline group statistics.
• jx_classification.csv — every annotated junction of both genes with the transcripts containing it, whether it is in MANE Select, and whether it is non-productive-specific.
• ensembl_tx_biotypes.csv — the transcript biotype table: 21 SCN5A and 30 SCN1A transcripts with biotype, length, coordinates and the non-productive classification.
• UPREGULATION_HEADROOM_CORRECTED.csv — the headroom calculations, both the ceiling table and the boost table, computed from 31.3, 45.8 and 100.
• p6_verify_and_headroom.py — the script that writes the headroom table and checks the other four files against every quantity printed in this paper.
• P6_VERIFICATION_OUTPUT.txt — that script's output. 116 of 119 checked quantities reproduce at the precision this paper prints them to; the three that do not are named in the correction section below.

Four claims in this paper cannot be checked against any deposited file, and this statement says so rather than implying otherwise. The per-junction read matrix was not retained, so the range of 19 to 107 reads per sample across the 27 MANE junctions, the median of 19 reads on the exon-skipping junction, the 832 and 724 reads on the two SCN1A junctions, and the claim that neither SCN5A poison-exon inclusion junction appears anywhere in the GTEx v10 junction file all rest on a streamed file that was not stored. What the deposit does support is that both inclusion junctions exist in the annotation and are correctly classified. Per-transcript abundance was aggregated before saving, so the three median values of 0.47, 0.32 and 0.07 TPM are likewise not recoverable from the deposit. Reproducing any of these would require downloading the GTEx v10 junction and transcript files again.

Archive versioning. The concept DOI 10.5281/zenodo.21799233 always resolves to the current version of the data archive and is the identifier to follow for access. The version current at the time of this revision is version 2, 10.5281/zenodo.21840036. Version DOIs cited elsewhere in this manuscript name the specific version read and are deliberately not rewritten.

Competing interests

I am a heterozygous carrier of the SCN5A R104Q variant examined in this study.

Use of AI tools

This work was carried out with AI coding and research assistants (Anthropic Claude, via Claude Code). That use is disclosed here rather than left to inference.

Analysis code. The great majority of the analysis code in this project -- parsers, genome scans, regeneration scripts and verification scripts -- was written by an AI assistant working to my specification. I set what each script had to compute, chose the thresholds and the decision rules, and checked the output against the claims it is used to support.

Manuscript text. The prose of this manuscript was drafted by an AI assistant. I directed the drafting and revised the result, and I am responsible for every claim it makes.

Scientific decisions. The questions asked, the thresholds set, what was allowed to count as a refutation, and what was published were mine.

Verification, which does not depend on any of the above. Where a claim in this manuscript is regenerable from deposited inputs, the script that regenerates it and that script's own output are in the data deposit. Reproduction does not require trusting any account of who wrote what.

What no AI system did. No AI system generated, altered or selected any experimental measurement; this project contains no wet-lab data of any kind. All primary literature cited was retrieved from PubMed, PMC and publisher sources. Every reference in this manuscript has been machine-resolved against its own record, including a check that each PMID's first author and year match the author and year printed beside it in the text.

References

1. O'Neill MJ, et al. 2022. PMID: 35305865. Supplementary Table 1, also available as preprint doi 10.1101/2021.09.22.461398.
2. Zorevunersen in children and adolescents with Dravet syndrome, phase 1-2a studies MONARCH and ADMIRAL. N Engl J Med. 2026;394(10):969-982. doi: 10.1056/NEJMoa2506295. PMID: 41780062.
3. Correspondence on zorevunersen in Dravet syndrome. N Engl J Med. 2026. PMID: 42308493.
4. GTEx Consortium. GTEx Analysis v10: splice-junction counts (STAR v2.7.10a) and transcript-level quantification (RSEM v1.3.3), quantified against GENCODE v39. GTEx Portal.
5. GTEx Consortium. GTEx v9 long-read RNA-seq quantification (FLAIR, GENCODE v26).
6. Ensembl. GRCh38 annotations for ENSG00000183873 (SCN5A) and ENSG00000144285 (SCN1A), accessed 2026-08-04.

Back to all ten papers