---

Version of record: 10.5281/zenodo.21799849, published 5 August 2026. That identifier is the citable address for this paper and it resolves at https://doi.org/10.5281/zenodo.21799849. 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_2_BASE_EDITOR_SPECIFICITY.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 Earlier on 6 August 2026 it was given a banner reading "superseded revision, do not cite any figure in it".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 carries five defects, corrected here on 6 and 7 August 2026, after the deposit. (This header said four until 7 August 2026. A fifth correction was added to the manuscript that day and is described at the end of this note; the count is corrected here rather than left standing.) Two are the same count twice: Table 2a's protein-changing column read 3, 16, 23, 55 and now reads 5, 22, 34, 71, and section 6.5 read "16 of which change a protein" and now reads 22. The deposited paper therefore prints 16 protein-changing sites for the SpG route in Table 2a and 22 for the same route two subsections later, which is an internal contradiction rather than a divergence from the archive. The corrected column reconciles exactly with the coding-exon column, which never changed. The editable-site columns were always right, so the 6.2-fold PAM-relaxation cost, which is what that section exists to report, does not move.

The third correction is a new section 3.4.8, and half of it is a defect in a published data file. ABE_CONSEQUENCE_RECHECK.csv, as deposited in 10.5281/zenodo.21799234, truncates its recomputed_aa field at exactly 1,500 characters. Fifty-eight of its 668 call-carrying rows end mid-token, and 1,706 of 8,843 consequence calls, 19.3 percent, are missing from a published file. The truncated rows are not marked and do not look truncated. Every consequence figure in the deposited paper is therefore a statement about the 7,137 calls the table still prints, and the deposited paper does not say so. This copy says so, in section 3.4.8, in limitation 6.1.11 and in its data availability section. What the table printed for the missing 1,706 is unrecoverable; the producing script is not in the project tree. The same section reports a fourth wrong consequence call, at KCNQ2 chr20:63,424,178, where p.Arg416Gly, p.Arg406Gly and p.Arg393Gly are really p.Ser416Gly, p.Ser406Gly and p.Ser393Gly. The call is missense under both readings and nothing downstream of the residue label changes.

The fourth correction is structural, it moves no number, and it is the one a reader can see without opening a data file. The deposited version prints, as a section heading of a published scientific paper, the line ## Insert into PUBLISH_2_BASE_EDITOR_SPECIFICITY.md as section 3.4.7, immediately after 3.4.6. That is a build instruction to whoever assembled the manuscript. It was pasted in with the section it introduces, never removed, and deposited; its presence in the deposited PDF was confirmed by downloading that PDF from the Zenodo API on 6 August 2026. The instruction was also never followed, so the deposited version prints section 3.4 in the order 3.4.1, 3.4.2, 3.4.3, 3.4.4, 3.4.6, 3.4.5, and then prints section 3.4.7 twenty-five pages later, after the Discussion has ended. In this copy sections 3.4.5 through 3.4.8 are in numerical order and the instruction line is removed. Not one word of any section was rewritten, and the removed line is quoted verbatim in a dated note at the position the displaced sections occupied.

The fifth correction was added on 7 August 2026, it is the only favourable one, and that is the reason to be most sceptical of it. New section 3.4.9 answers the question section 3.4.8 raised and stopped at: this paper is deposited under a permanent identifier and rests on a file that cannot be repaired, so what does the truncation actually cost? The truncation is confined to recomputed_aa, the per-transcript audit column, and every count in this paper is a site-level count read from recomputed_worst and recomputed_region, which are separate and untruncated. Every consequence figure was regenerated from those columns and 102 of 102 reproduce, with the two truncated rows at which a destroyed call could in principle have changed a count adjudicated one by one and shown that it cannot have. Nothing is repaired: the deposited file is republished in version 2 byte for byte unchanged, 1,706 consequence calls stay destroyed, and a reader auditing any single call still has nothing to audit. The counts survive and the audit trail does not, and only the first half of that is good news.

A sixth change in this copy is not a correction and is recorded so it is not mistaken for one. The four junction-straddling census tables that section 3.4.8 rests on were named as absent from the archive when this copy said the archive did not hold them; they are staged for version 2 and this copy now says so, with the struck sentence left visible. Version 2 has not been uploaded, so a reader holding version 1 of 10.5281/zenodo.21799234 will not find them.

One further defect in the deposited data, recorded here because this paper describes it. The paper says its lead guide has 22 protein-changing predicted off-targets and its Table 3 lists 22. The summary file in the data deposit, MS_TABLE3_LEAD_PROTEIN_CHANGING.csv, still holds the 16 rows of the superseded count, because the paper was corrected and the file was not regenerated. That defect is in the deposited summary table, not in this text. (This paragraph previously ended by calling ABE_CONSEQUENCE_RECHECK.csv complete at 22,717 rows. The row count is right and the word complete was wrong, and it was corrected here on 6 August 2026: the file has 22,717 rows and is missing 19.3 percent of its consequence calls, as set out above. A sentence that states a true row count and infers completeness from it is exactly the failure mode this project keeps finding, so it is corrected in place rather than deleted.)

No conclusion, risk tier, editor route, sequencing-panel entry or recommendation changes. Two of the four corrections make the negative result larger rather than smaller.

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.

---

Editability-scored off-target counting, and the specificity cost of PAM relaxation, in adenine base editor design: SCN5A p.Arg104Gln as a worked example

Ethan Bradley

Independent researcher, no institutional affiliation

ORCID: 0009-0008-8925-7975

Correspondence: the author

Preprint. Version 1, 2026-08-04. Corrected 6 August 2026 in three places: a fourth wrong consequence call in my own recheck table and a 1,500-character truncation in that table as deposited (section 3.4.8, carried into 6.1 and 7); a stale protein-changing column in Table 2a; and a stale protein-changing count in section 6.5. Corrected again 7 August 2026 with a new section 3.4.9, which establishes what that truncation actually costs: every consequence figure in this paper was regenerated from the broken file's untruncated columns and 102 of 102 reproduce, while the audit trail for 1,706 destroyed calls stays permanently gone. The corrections are collected in "Corrections, 6 and 7 August 2026" before section 7. The record deposited at 10.5281/zenodo.21799850 on 5 August 2026 carries none of them, and it is the record that states this paper's figures come from a file it never says is defective. No conclusion, no recommendation, no risk tier, no editor route and no sequencing-panel entry changes.

Keywords: adenine base editing, off-target prediction, PAM relaxation, SpG, Cas9-NG, SpRY, guide RNA design, SCN5A, Brugada syndrome, dominant negative

Abstract

Background. Base-editor off-targets are conventionally ranked by mismatch count, but a bound adenine base editor with no adenine in its window does nothing. SCN5A base editing already works in arrhythmia models (PMID 42193469, 37965733); the open question at a new locus is how specificity is counted.

Methods. With the Brugada variant SCN5A p.Arg104Gln as the worked example, I scanned 3,099,750,718 bp of GRCh38 on both strands with no seed heuristic, scoring sites on PAM compatibility and an editable adenine rather than sequence match, then rescanned under 1-nucleotide bulge tolerance (Cas-OFFinder-bulge v1.2).

Results. Editability scoring filters 85.1 percent of an SpG route's 4,388 matches but only 7.5 percent of a near-PAMless SpRY route's, so similarity counting makes the two appear 5.6-fold apart when the editable gap is 34.7-fold. Varying only the PAM rule on an identical 20-mer gives 136, 655, 1,598 and 4,064 editable sites for NGG, NG, purine-restricted and unrestricted PAMs, a 6.2-fold cost. Specificity claims hold only with an alignment model attached: the lead guide has nothing editable at 0 mismatches under either model, and nothing at 1 mismatch ungapped but five alignments at four loci with bulges, none protein-changing. Bulge tolerance also moves its nearest protein-changing off-target from 3 mismatches to 2, and penalises a relaxed-PAM editor twice: for accepting more PAMs, and because each dockable position is searched under every bulged geometry.

Conclusion. Editability, not similarity, is the correct scoring axis, and bulge tolerance raises the cost of PAM relaxation.

A key to the terms used in this paper

I am the person who carries this variant. I am not a bioinformatician, and I have written this paper so that a reader who is also not one can follow it. Every technical term is glossed at first use. This section collects the recurring ones.

A pharmacy analogy carries the central methodological point of this paper, so I will state it here and return to it in the Results. A guide matching a genomic sequence is like a prescription whose patient name matches. That alone does not mean a drug gets dispensed. The label also has to scan, which is the PAM, and there has to be something in the bottle to dispense, which is an adenine in the window. Counting name matches and calling them dispensing errors would badly overstate how many errors the system can actually make.

1. Introduction

1.0 What this paper contributes, and what it does not claim

I want to be direct about where the contribution lies, because the obvious reading of this paper is the wrong one.

Base editing of SCN5A is already an established approach. A 2026 review of gene therapy for cardiovascular and cerebrovascular disease names SCN5A base editing in its abstract as one of the strategies that have demonstrated capacity to re-establish electrophysiological stability in arrhythmia models, listed alongside KCNQ1/KCNH2 suppression-and-replacement and structural protein restoration via PKP2 and TMEM43 (PMID 42193469). The underlying demonstration is Qi M et al., Circulation 2024 (PMID 37965733), described in section 1.4. Base editing has also been used at this gene in the opposite direction, to generate SCN5A point mutations in human pluripotent stem cells for disease modelling rather than to repair one (PMID 33102492), and the general approach of CRISPR modelling and correction in cardiovascular disease has been reviewed (PMID 35679361).

So "I designed a guide for a new variant" is not a contribution, and I am not claiming it as one. If the premise were speculative, novelty would be the thing to argue. It is not speculative, which is more useful: a reader does not need convincing that a base editor can be pointed at this gene.

The contributions are methodological, and neither requires this variant.

1. Editability-scored off-target counting. Off-targets for base editors are conventionally ranked by mismatch count. For a nuclease that is defensible, because a bound nuclease cuts. For a base editor it is not, because a bound editor with no adenine in its editing window does nothing. Scoring on editability rather than similarity removes 85.1 percent of the lead guide's 4,388 genomic sequence matches, and it removes them unevenly between routes, which is the part that matters for any comparison. Any base-editor design that ranks off-targets by mismatch count alone is making this error.
2. A controlled measurement of the specificity cost of PAM relaxation. Holding the 20-mer identical and varying only the PAM rule separates the cost of the editor from the cost of guide choice, giving a 6.2-fold penalty from NG to unrestricted. This is cheap to run at any locus and nobody needs this variant to use it.

R104Q is the worked example, not the claim. I use it because it is the variant I carry, because it forces the interesting case (standard SpCas9 cannot reach it, so engineered editors and their relaxed PAMs are mandatory rather than optional), and because it lets both methodological points be demonstrated on a real therapeutic target rather than a synthetic one.

There is one respect in which the target itself is new, and I state it as a gap rather than an achievement: as far as this search reaches, there is no published base-editing design for any Brugada or loss-of-function SCN5A variant, and no prior guide design, off-target analysis or editor comparison at this locus. Section 1.5 argues that the reason is structural rather than accidental, and section 5.3 develops it.

Bounds of the prior-art search. Two independent searches were run, one against PubMed through NCBI E-utilities and one against Europe PMC across five query families. Neither covers patents, company development pipelines, conference abstracts, or non-English literature. A query for "R104Q OR Arg104Gln" returns 72 hits in Europe PMC, none concerning SCN5A: they are R104Q substitutions in RNF114, in Charcot-Marie-Tooth genes, and in cytochrome P450 enzymes. A bounded search finding nothing is evidence about the search, not proof of absence, and I do not claim priority on that basis. I claim only that I could not find it, and I have said where I looked.

1.1 The variant used as the worked example

SCN5A encodes Nav1.5, the voltage-gated sodium channel that carries the depolarising current starting each heartbeat in ventricular muscle. Loss-of-function variants in SCN5A cause Brugada syndrome, in which the electrocardiogram shows a characteristic pattern in the right precordial leads and the carrier is at raised risk of ventricular arrhythmia and sudden death.

The variant treated here is NM_000335.5:c.311G>A, p.Arg104Gln, at mRNA position 520, which changes codon 104 from CGG (arginine) to CAG (glutamine). In genomic coordinates it is NC_000003.12:g.38630392C>T on GRCh38, ClinVar VariationID 67780. It was first reported in 2001 in a series of seven Brugada patients from Israel, in which two novel SCN5A variants were found and low penetrance was proposed as the reason so few family members were symptomatic (PMID 11960580). Position 104 lies in the N-terminal domain of the channel, and N-terminally mutated Nav1.5 constructs including R104Q have been characterised electrophysiologically (PMID 23805106).

A note on adjacency that I record because it shaped my expectations. R104W, a different substitution at the same residue, sits at the immediately adjacent base, chr3:38630393, ClinVar VariationID 67778. R104W is classified Pathogenic and has published dominant-negative and trafficking data. R104Q is Conflicting and has far less. One base apart, and one has an answer key while the other does not.

1.2 The functional baseline, and the one thing that is not settled

The quantitative anchor for everything downstream is O'Neill MJ et al., Genetics in Medicine 2022 (PMID 35305865), Supplementary Table 1, which I read from the open preprint doi:10.1101/2021.09.22.461398 at page 29. That study used Sleeping Beauty genomic integration, so wild-type expression is not diluted by the variant construct and a variant with no dominant-negative effect reads 100 percent.

The variant protein alone carries essentially no current. Co-expressed with wild type it removes 31.7 points from what the wild-type allele would deliver alone. That is the signature of a dominant-negative effect: the variant protein interferes with its healthy partner rather than merely being absent.

This is the load-bearing uncertainty of the paper and I am placing it in the introduction rather than burying it in the limitations. That measurement was made in a heterologous expression system. Whether the dominant-negative effect operates the same way in human heart muscle is unresolved. If R104Q behaves in a human heart as simple haploinsufficiency, meaning one allele simply stops contributing and does not interfere, then correcting the mutant allele buys a 2.18-fold gain rather than the 3.20-fold I calculate below. The case weakens without collapsing. An allele-resolved surface-expression experiment would settle it, and I have requested one from a laboratory with the relevant assay. Until that result exists, every rescue figure in this paper is reported under both mechanisms.

1.3 What a corrective therapy would be compared against

I want to be precise about the comparator, because the intuitive answer is wrong.

Guideline indications for an implantable cardioverter-defibrillator in Brugada syndrome centre on prior cardiac arrest, documented sustained ventricular arrhythmia, or arrhythmic syncope with a spontaneous type 1 pattern. Most carriers are asymptomatic, do not meet those criteria, and therefore have no protection at all. A defibrillator also does not treat the disease: it terminates an arrhythmia after it has started and does nothing to the sodium current. And it carries measured harm. A systematic review and meta-analysis of 63 studies and 4,916 patients with inherited arrhythmia syndromes found inappropriate shocks in 20 percent of patients, at a crude rate of 4.7 percent per year (PMID 26385533).

The comparator for a disease-modifying therapy is therefore no treatment, not a functioning device. This raises what an off-target burden must be weighed against. It does not lower the bar for measuring that burden, which is the substance of this paper.

1.4 Why an adenine base editor, and what the established precedent is

Codon 104 needs its middle A changed back to a G. An adenine base editor performs exactly that reaction and no other. The repair required and the chemistry available are the same operation, which is not true of most disease variants.

The precedent is unusually close, and it is the demonstration behind the review statement quoted in section 1.0. Qi M et al., Circulation 2024 (PMID 37965733) built a knock-in mouse carrying the SCN5A T1307M variant, split an adenine base editor across two AAV9 vectors, gave a single intraperitoneal injection at postnatal day 14, and corrected up to 99.20 percent of Scn5a transcripts. Correction above 60 percent eliminated the QT-prolongation phenotype, and no off-target DNA or RNA editing was detected in treated hearts. Same gene, same organ, same editor class, and a phenotype rescue rather than an efficiency measurement. Section 1.5 states the one way it does not transfer.

Two further results frame the approach. Levy JM et al., Nature Biomedical Engineering 2020 (PMID 31937940) delivered split base editors by dual AAV and reported 20 percent editing in mouse heart in a general tissue survey. Koblan LW et al., Nature 2021 (PMID 33408413) corrected the progeria point mutation in LMNA in vivo and extended lifespan, establishing that a dominant-negative single-base change is a tractable base-editing target, which is the structural situation here.

1.5 Why the established successes are all the easier class of variant

T1307M is a gain-of-function variant causing long QT syndrome type 3. R104Q is loss of function causing Brugada syndrome. That difference is not cosmetic, and I think it explains the pattern in the literature rather than being incidental to it.

Long QT syndrome type 3 has a cheap, quantitative, non-invasive animal endpoint: the QT interval on a surface electrocardiogram. Brugada syndrome does not. Its diagnostic ST-segment signature depends on a transmural voltage gradient across the right ventricular wall. Mouse hearts beat near 500 beats per minute with a substantially different complement of repolarising currents and do not reproduce that gradient. A mouse carrying a Brugada variant therefore offers no equivalent readout, and a correction experiment in one would demonstrate editing efficiency without demonstrating phenotype rescue.

So the absence noted in section 1.0, that I could find no base-editing design for any Brugada or loss-of-function SCN5A variant, is unlikely to be an oversight the field simply has not got round to. The gain-of-function variants are the ones with a cheap endpoint, and that is a strong reason for them to have been done first. Designing for the harder class means accepting up front that the phenotype question cannot be answered the way the precedent answered it, and saying so rather than implying the precedent transfers whole.

This is a limitation of the whole approach for this disease, not of this design. Section 5.3 develops what it implies for anyone attempting the correction, and section 6 lists it among the limitations.

2. Methods

Everything in this section is reproducible from public data by anyone, with no institutional access and no licensed software. Every input is named with an accession, a checksum, or a coordinate.

2.1 Sequence retrieval and the coordinate assertions

Reference sequence was retrieved from NCBI E-utilities efetch, accession NC_000003.12 (GRCh38.p14), region 38630362 to 38630423, 1-based inclusive, forward strand. The 62-base window reads:

AGGACATACAAGGCGTTGGTGGCACTGAACCGGAAGATGGTCTTGCCTTTATTCAGTACGAT

Two assertions guard every coordinate operation, and any reproduction of this work should include them.

1. The reference allele at forward position 38630392 must be C, matching the ClinVar record.
2. Codon 104 must read CGG in wild type and CAG after a forward C-to-T substitution at 38630392.

I state plainly that these assertions caught two real errors of mine, because a design that fails silently is the dangerous kind.

The first error was a convention mismatch. UCSC REST returns 0-based half-open coordinates; NCBI efetch returns 1-based inclusive. The same request to both returns strings offset by one base. My first sequence gave G where ClinVar requires C, and CCG where the codon must read CGG. Both assertions failed loudly, and I rebuilt on the NCBI sequence.

The second error survived the first fix and is subtler. c.311 is the middle base of codon 104, and SCN5A is on the minus strand. The codon is therefore forward 38630391 to 38630393 read as reverse complement, not 38630392 to 38630394. Reading forward 38630392 to 38630394 returns CGG, which looks like a passing check and is the wrong codon. Only the combination of the correct span and the reverse complement satisfies both assertions. I verified this again while preparing this manuscript: forward 38630391-38630393 is CCG, whose reverse complement is CGG (arginine), and the C-to-T substitution makes it CAG (glutamine).

2.2 Strand determination

SCN5A is Ensembl ENSG00000183873, chr3:38548057-38649743, strand -1. The transcript notation c.311G>A and the genomic notation g.38630392C>T are not contradictory: a G on the messenger RNA pairs with a C on the forward strand, which is the signature of a minus-strand gene. The consequence is concrete:

A base editor acts on whichever strand its guide binds, so every guide here must be designed on the reverse strand. Anyone working from the cDNA number alone will design a forward-strand guide and get nothing. I flag this because the cDNA notation is the form the variant is almost always quoted in.

2.3 Guide enumeration

The mutant allele was constructed by substituting T at forward position 38630392 in memory. For each of the 20 possible registers, the protospacer was taken as the 20 bases ending at forward coordinate 38630392 + (p - 1) read as reverse complement, where p is the position the target adenine occupies in the protospacer, and the PAM as the three bases immediately 3' of the protospacer on the same strand. A register was retained as a candidate if p fell within the editing window, defined as positions 3 to 9 inclusive. Bystanders were counted as adenines within positions 3 to 9 other than the target. The same register was extracted from the unmodified reference to obtain the healthy-allele protospacer and its editable-adenine count.

PAM rules applied: NGG for SpCas9, NG for SpG and Cas9-NG, NRN for purine-restricted SpRY, and unrestricted for SpRY (PMID 32217751 introduced both SpG and SpRY).

All 20 registers were enumerated, not sampled. Seven fall in the window and constitute the candidate set reported in Table 1.

2.4 Genome-wide off-target scan

Assembly. NCBI GRCh38 no-alt analysis set, GCA_000001405.15_GRCh38_no_alt_analysis_set.fna.gz, 872,949,833 bytes, md5 a08035b6a6e31780e96a34008ff21bd6, gzip integrity verified. The file contains 195 sequences. I scanned 194 of them, totalling 3,099,750,718 bp: the 22 autosomes, X, Y, mitochondrial DNA, and all unplaced and unlocalised scaffolds. I excluded only chrEBV, which is Epstein-Barr virus included as a sequencing decoy and is not human DNA. The no-alt set excludes alternate haplotype contigs, which are duplicate representations of a few variable regions; including them would double-count the same locus.

Search. Every one of the 3.1 billion positions was compared against every guide at all 20 positions, on both strands. There is no seed requirement, no index, and no heuristic filter, so a distant match cannot be silently missed. The eight candidate guide-and-editor pairs collapse to 7 unique 20-mers, because the SpG guide at position 5 and the SpRY guide at position 5 are the same sequence used by two editors. Each unique sequence was searched as itself and as its reverse complement, giving 14 searches. Mismatch budget 0 through 6. Runtime 669 seconds on 10 cores.

Editability scoring. A hit was scored editable only if both conditions held: the three bases immediately adjacent satisfy the PAM rule of the editor in question, at the correct offset and on the correct strand; and at least one adenine sits within protospacer positions 3 to 9. A hit failing either condition cannot be edited even if the guide binds. Naive sequence-match counts and editable counts are reported separately throughout, and I do not quote the naive count as a risk figure.

Annotation. NCBI RefSeq GCF_000001405.40_GRCh38.p14_genomic.gtf.gz, 56,947,363 bytes, 67,512 gene records, 1,849,964 coding-exon records. RefSeq accessions were mapped to chr names using the matching assembly report so annotation and sequence share one coordinate system. Amino-acid consequences were computed from the GTF frame field and genomic sequence.

Cardiac expression. GTEx v8, tissues Heart_Left_Ventricle and Heart_Atrial_Appendage, plus skeletal muscle and two arteries for context. A gene was flagged cardiac-expressed at a threshold of 1 TPM in either heart tissue. I requested 330 genes and resolved 295; the 35 unresolved are mostly LOC-prefixed and predicted-model genes absent from the GTEx v8 reference.

2.5 Pipeline validation: five gates

I do not trust a pipeline that has not recovered something whose answer I already know.

1. Coordinate gate. The retrieved 62-base window matched the expected string exactly, and the forward base at 38630392 is C.
2. Codon gate. Codon 104 reads CGG wild-type and CAG mutant, as described in 2.1. This gate caught the middle-base error.
3. Array gate. The window read back from the parsed chromosome array is byte-identical to what NCBI returned, confirming array indexing is aligned to 1-based genomic coordinates.
4. On-target recovery gate. All 7 unique guides found their own site on the minus strand with exactly the expected PAM. An important subtlety: GRCh38 is the wild-type reference and these guides target the mutant allele, so at the target locus each guide matches the reference at 1 mismatch, not 0. Patching C-to-T at 38630392 in memory to build the mutant allele, every guide then matches at 0 mismatches. A pipeline reporting 0 mismatches against unmodified GRCh38 would have been wrong.
5. PAM offset gate, both strands. Plus-strand and minus-strand hits were independently re-extracted from the chromosome array; protospacer and PAM matched in every case. The on-target gate exercises only the minus strand, so without this a plus-strand offset error would have gone unnoticed.

Independently, the consequence engine reproduces p.Arg104Gln at codon 104 for the intended edit given only the GTF frame field and genomic sequence, with no hard-coded knowledge of the variant.

2.6 Rescue arithmetic

Let b be the untreated current as a percentage of a heart with two working alleles. Under the dominant-negative measurement, b = 68.3 / 2.184 = 31.3, because the O'Neill figure is expressed relative to one wild-type allele and their own measured two-allele reference reads 218.4 percent of that, not 200. Under the alternative haploinsufficiency mechanism, b = 100 / 2.184 = 45.8. A corrected cell delivers 100. For a mosaic fraction f of corrected cells, predicted tissue current is 100f + b(1 - f), and fold gain is that quantity divided by b. Per-cell gain is 100 / b, giving 3.20-fold under the dominant-negative mechanism and 2.18-fold under haploinsufficiency. This is linear arithmetic on a measured starting point, not a model of cardiac electrophysiology, and section 6 states what it therefore cannot capture.

2.7 Literature retrieval, and its bounds

Citations were retrieved and verified through NCBI E-utilities esearch, esummary and efetch against PubMed. Every PMID cited in this paper was resolved to its author list, journal, year, volume, pages and DOI, and those are recorded in MS_REFERENCES.csv.

This literature search is bounded and I do not call it exhaustive. It covers PubMed only. It does not cover conference abstracts, patents, preprint servers other than those PubMed indexes, or non-English-language literature. Publisher websites were unreachable from my environment, which was blocked by a network allowlist and returned proxy errors rather than server errors; that is evidence about my retrieval, not about the literature. Specifically, I could not verify current author guidelines for any candidate journal, and every such requirement is marked UNVERIFIED in the accompanying venue assessment.

3. Results

3.1 Standard SpCas9 cannot reach this site. This is the first result and it is a negative.

Of the 20 possible protospacer registers, seven place the target adenine inside the editing window. Enumerating all seven (Table 1), none has an NGG PAM. Standard SpCas9 therefore has no usable guide at this locus at any editing position. The site is reachable only with an engineered editor accepting a relaxed PAM.

This matters because it forces the design onto engineered editors whose relaxed docking requirement is itself the main source of off-target exposure. Everything in section 3.4 is a consequence of this negative.

3.2 One guide is clean by construction

Table 1. The seven guide candidates that place the target adenine in the editing window. All are on the reverse, mRNA-sense strand. Full table with genomic spans in MS_TABLE1_GUIDES.csv.

Six of the seven carry no bystander adenine. Exactly one has a PAM that SpG can read: the guide at position 5, PAM CGC, whose middle base is G. Its genomic span including PAM is chr3:38630374-38630396.

The lead guide is therefore the guide at position 5. It requires only the NG relaxation rather than full PAM-lessness, and it sits at the position with the best editing geometry with no other adenine in its window.

3.3 What happens to the healthy copy, and why the answer is unusually clean

mutant allele    5'-TTCCAGTTCAGTGCCACCAA-3'
healthy allele   5'-TTCCGGTTCAGTGCCACCAA-3'
                        ^ one base different

The guide differs from the healthy allele at a single position, so it will bind both copies of the gene. But the editing window on the healthy allele contains no adenine at all. The editor docks and finds nothing it can act on, because an adenine base editor cannot alter a G.

Allele selectivity here does not depend on the guide discriminating between the two copies. It depends on the enzyme having no substrate on the healthy copy. That is a stronger position than sequence-based discrimination, and it is the reason this variant is a better base-editing target than an allele-specific silencing target.

3.4 The predicted off-target profile

Across all searches, 2,051,639 sequence matches exist at up to 6 mismatches. The annotated set stops at 4 mismatches and contains 24,562 rows, of which 7 are the on-target sites of the 7 unique guides and 24,555 are off-target.

Naive sequence-match counts, pooled across all guides:

Zero matches at 0 mismatches is expected and correct: the guides match the mutant allele, which is not present in the reference.

3.4.1 First methodological result: scoring editability rather than similarity removes 85 percent of the apparent risk

For the lead guide, 4,388 off-target sequence matches reduce to 655 editable sites, so 85.1 percent of matches cannot be edited at all, lacking either an acceptable PAM or an adenine in positions 3 to 9. For the SpRY route, 24,555 matches reduce only to 22,710, filtering just 7.5 percent, because with no PAM requirement left there is almost nothing to filter on.

The editability filter is what makes the SpG number small, and it is precisely the filter SpRY discards. A worked example makes the point concrete. The closest sequence matches in the entire scan are at chr2:166,058,637-166,058,640, one mismatch from three of the SpRY guides. That looks alarming. It is not editable. The site reads TCTTCCGGTTCAGTGCCACC: it carries G where the guide carries A, because it is a paralogous sodium-channel gene retaining the ancestral wild-type sequence. There is no adenine in the window, so there is nothing to convert. A ranking by sequence similarity alone would place this at the top of the risk list.

3.4.2 Second methodological result: relaxing the PAM costs 6.2-fold with the guide held identical

Holding the 20-mer identical and varying only the PAM rule isolates the cost of relaxation from the cost of guide choice.

Table 2a. Controlled comparison. Identical lead 20-mer TTCCAGTTCAGTGCCACCAA, PAM rule varied.

(Corrected 6 August 2026. The protein-changing column read 3, 16, 23 and 55. Those are pre-recheck counts: this table alone was never recomputed from ABE_CONSEQUENCE_RECHECK.csv when Table 2b, section 3.4.4 and Table 3 were, so the paper printed 16 protein-changing sites for the SpG route here and 22 for the same route two subsections later. The corrected column is the recheck's missense calls on the same row sets, and it reconciles with the unchanged coding-exon column exactly — 5 + 1, 22 + 4, 34 + 6 and 71 + 8 synonymous give 6, 26, 40 and 79. The editable columns are unaffected and were correct. The correction makes the protein-changing load larger on every route, so it is unfavourable to this paper's own recommendation in absolute terms and marginally favourable in relative terms — the NG-to-unrestricted ratio on this column moves from 55/16 = 3.44 to 71/22 = 3.23, while the editable-load figure of 6.2-fold, which is what this section is about, does not move at all. The record deposited on 5 August 2026 prints the retired column.)

Relaxing the docking collar from NG to unrestricted multiplies the editable off-target load 6.2-fold on a sequence that has not changed at all. The NGG row is included for calibration only; as section 3.1 established, no NGG guide exists at this site, so that row is not an available option.

3.4.3 Route level: SpRY carries 34.7 times the load for no benefit

The comparison that decides the recommendation is not editor against editor but route against route, because only one of the seven guides has an on-target PAM SpG can read. The other six have on-target PAMs CCT, GCC, ACG, AAC, CAA, CCA, none satisfying NG. So the SpG route is one guide and that is its entire exposure, while the SpRY route is all seven.

Table 2b. Route-level comparison.

Protein-changing rows in this table are recomputed from ABE_CONSEQUENCE_RECHECK.csv for the reason given in section 3.4.6, so they are higher on both routes than an earlier version of this analysis reported. The editable counts are unaffected, because the correction was to consequence calling and not to the scan. Both routes are counted the same way, so the comparison the recommendation rests on is unchanged.

SpRY's purpose is to provide guide options where no NG PAM exists. Here an NG PAM does exist, at the position with the best editing geometry and with no bystander adenine. The SpRY guides buy nothing at this site and cost between 24 and 45 times the burden depending on which column is read.

3.4.4 The lead guide's own profile, enumerated

Every count in this subsection is stated with its alignment model attached, because the two models answer different questions and one of them is blind to insertions and deletions.

• Nothing editable at 0 mismatches, under either alignment model.
• At 1 mismatch, nothing editable under ungapped alignment, and five alignments at four distinct loci under 1-nucleotide bulge tolerance. All four loci were characterised completely and none is protein-changing: three intronic, in TNKS1BP1 (17.17 TPM left ventricle), MON2 (2.95) and MYO5B (0.18), and one in coding sequence but synonymous, TGM2 p.Pro375Pro. TGM2 is the one that needed checking, because at 230.76 TPM in left ventricle it runs 6.6 times SCN5A's own 35.11, so a protein-changing edit there would have been serious. I verified the synonymous call by translating full-length NM_004613.4 end to end rather than trusting my own annotator, and the edited base sits 25 bp from the nearest splice boundary, so splice disruption is not in play either. Five alignments describe four loci because one docking window at chr12:62,533,866 admits two equivalent bulge placements at equal mismatch cost.
• At 2 mismatches, 3 editable sites under ungapped alignment, and I enumerated all three: chr1:114,999,332 (intergenic), chr14:96,276,087 (intergenic), and chr21:15,190,061 (an exon of the long non-coding RNA LOC124905047). None is protein-coding. Under bulge tolerance the same distance carries 4 protein-changing loci, in TXNIP, SAMD15, SLC68A1 and HNRNPAB, so bulge tolerance puts protein-changing risk closer to this guide than the ungapped scan could see.
• 51 editable at up to 3 mismatches, 655 at up to 4, ungapped.
• 26 fall in coding exons: 22 missense and 4 synonymous.
• Of the 22 protein-changing sites, 7 are in genes expressed at 1 TPM or above in heart, and 4 of those reach that level in left ventricle specifically, which is the column Table 3 reports.

The consequence counts in this subsection and in Table 3 are recomputed from ABE_CONSEQUENCE_RECHECK.csv and not from the original scan table, for the reason given in section 3.4.6: the original consequence calls were shifted by one residue in 133 rows. The missense count for this guide is 22, not the 16 an earlier version of this analysis reported.

Table 3. All 22 protein-changing predicted off-targets of the lead guide, ungapped alignment. Consequences are given on the lowest-numbered RefSeq transcript carrying them; full detail including every overlapping transcript is in ABE_CONSEQUENCE_RECHECK.csv.

A correction to my own earlier figure. An earlier version of the off-target figure labelled the highest-expressed cardiac hit GFM1, because the site at chr3:158,672,417 falls within the annotated span of both GFM1 and LXN. The coding consequence is in LXN (p.Tyr20His, NM_020169.4). Figure 1 and Table 3 carry the corrected label. The counts are unaffected; only the gene name attached to that one row changes.

MSH6 deserves naming rather than a footnote. It is a DNA mismatch-repair gene, and inherited loss of function causes Lynch syndrome, a cancer-predisposition condition. It is the hit on the lead guide I would not wave through on paper. The mitigating detail is specific and testable: it sits at 3 mismatches, at guide positions 6, 7 and 12, with the mismatch at position 12 inside the PAM-proximal seed region where mismatches are least tolerated; the substitution p.Val263Ala is conservative rather than truncating; and cardiac expression is low at 2.82 TPM in left ventricle. It is a site to assay empirically, and it is listed explicitly for that purpose, not a reason to abandon the guide.

MSH6 is no longer the nearest protein-changing off-target, and I state that plainly because an earlier version of this analysis said it was. Under ungapped alignment it remains the nearest, at 3 mismatches. Once 1-nucleotide bulges are allowed, four protein-changing loci appear at only 2 mismatches (TXNIP, SAMD15, SLC68A1, HNRNPAB), and 41 appear at 3 mismatches against the 2 the ungapped scan found, among them CACNA1D, MPL, CHD5, RPTOR and the cardiac myopalladin gene MYPN. MSH6 is still a real hit and still the one I would assay first on consequence grounds. It is simply not the closest.

One accessibility question about MSH6 that I did not resolve. Its four editable protein-changing rows read closed in purified cardiomyocyte data, 0 of 3 experiments with the nearest peak 5,925 bp away, and open in bulk left-ventricle ATAC, 6 of 15 experiments. The cardiomyocyte reading is the one relevant to the delivery target, and it is the favourable one, which is exactly why I am not resolving the disagreement in that direction: the cardiomyocyte tier is 3 experiments against 15, and choosing the tier that gives the preferred answer is the move that produces a wrong result. The two data types disagree, and MSH6 stays flagged on sequence grounds.

3.4.5 Two cardiac genes are reachable only by the SpRY route

Table 5 (MS_TABLE5_CARDIAC_GENE_HITS.csv) enumerates all 24 editable hits in named cardiac ion-channel and sarcomeric genes. The two coding hits are:

• TRIM63, 116.2 TPM in left ventricle, the highest-expressed cardiac gene in the whole hit set. Missense, 3 mismatches, reached by the SpRY guides at positions 8 and 9. TRIM63 is a cardiac protein-turnover gene associated with hypertrophic cardiomyopathy.
• KCNJ8, 40.2 TPM, missense, 4 mismatches, reached by the SpRY guide at position 3. KCNJ8 encodes a cardiac potassium channel subunit and has itself been linked to early-repolarisation and Brugada phenotypes. Editing it while attempting to treat Brugada syndrome would be a poor outcome.

Neither coding hit is reachable by the lead SpG guide. The lead guide's only hits in this gene set are two non-coding sites in CACNA1C, both at 4 mismatches, one intronic and one in an untranslated or non-coding exon. Intronic hits are lower concern than coding ones but not zero, since they can affect splicing.

3.4.6 Bulge-aware rescan, and an error it exposed in my own consequence table

A position-by-position comparison can only count substitutions. It is structurally blind to a bulge, one extra unpaired nucleotide on one side of the pairing: a bulged site throws every downstream position out of register, so a well-matched bulged site looks like ten or more mismatches and falls outside any sane cutoff. It is not down-ranked, it is invisible. I therefore rescanned the identical 3,099,750,718 bp under Cas-OFFinder-bulge v1.2 semantics, reading that tool's source rather than inventing my own rules: up to one DNA bulge or one RNA bulge per alignment, never both, with mismatches counted separately from the bulge. For a 20-nucleotide guide that is 38 geometries, 304 patterns across 8 guides, and the editing window is anchored to the PAM rather than to guide index so bulged and unbulged sites are scored by the same physical rule.

Bulge tolerance adds risk, and I report it as a problem rather than a reassurance. The raw count is 52,994 bulge-added editable sites for the lead guide against 655 ungapped, a factor of 80.9, and that ratio should not be quoted, because "one bulge plus four mismatches" permits five deviations where "four mismatches" permits four. Matched so each column allows the same total deviations, the honest figure is 1.7 to 8.8 times more editable sites: 5 against 3 at two deviations, 174 against 48 at three, 3,369 against 604 at four, and 49,446 against 5,597 at five.

A second, independent argument for SpG over SpRY, on a corrected ratio. Every one of the seven SpRY guides carries more bulge-added editable load than the lead: 117,069 to 337,986 against 52,994, a range of 2.21 to 6.38 times, and in coding sequence 3,771 and above against 1,167, a range of 3.2 to 5.3 times. An earlier version of this comparison reported 6.8-fold by mixing the ungapped baseline into the bulged totals; the corrected figures are bulge-added only. The recommendation survives on a budget-matched comparison and is strengthened by holding across two alignment models whose failure modes differ, since the ungapped model can only miscount substitutions and the bulged model can only miscount insertions and deletions.

The mechanism generalises past this variant. A relaxed-PAM editor is penalised twice by bulge tolerance: once directly, because accepting nearly any PAM makes more genomic positions dockable at all, and again indirectly, because each of those extra dockable positions is itself searched under every bulged geometry. The two penalties multiply rather than add. The usual reasoning is that a relaxed-PAM editor costs specificity in proportion to how many more PAMs it accepts, and under bulge-aware scoring that understates the cost. The practical rule for anyone designing at a site with no canonical NGG PAM in range, which is the situation this variant presents: prefer the most PAM-restrictive editor that can still reach the target base, and if a relaxed-PAM editor is unavoidable, score it with bulge tolerance rather than position-by-position comparison.

An error in my own published consequence table, found by building the annotator for this rescan. The independent caller I wrote for the bulged and tail hits reproduces the patient's own variant from genomic coordinates alone, chr3:38,630,392 landing at cumulative CDS position 310 of NM_000335.5, codon 104 position 2, CGG to CAG, p.Arg104Gln. It was validated against a known answer, where it returned p.Val263Ala for the MSH6 off-target. It then disagreed with ABE_OFFTARGET_SCAN.csv on 133 rows substantively: 105 called synonymous that are in fact missense, 15 the reverse, 12 that carried no call at all and now do, 8 of those 12 protein-changing and 4 synonymous, and 1 called noncoding that is missense. A further 1,183 rows differ only in label vocabulary and I do not count those as errors. The cause was a one-residue offset. I resolved three disagreements by translating the full coding sequence from NCBI end to end, checking for an ATG start and no internal stops before deciding which call was right, and MTREX is the decisive case because both analyses used the same transcript NM_015360.5: the old call was p.Ala1018Ala synonymous, the truth is p.Leu1019Pro missense. Every consequence figure in this paper is recomputed from ABE_CONSEQUENCE_RECHECK.csv. No number is carried forward from the original table.

That recheck table has since been found wrong in a fourth place, and incomplete in a way that was invisible when this section was written. Section 3.4.8, added 6 August 2026, counts how many calls are exposed to the failure mode described here rather than how many were caught, and reports a truncation in the deposited copy of the file this paragraph directs the reader to. Neither finding moves a count in this paper.

3.4.7 A splice-site result the coding-only annotation could not have found, and the general lesson in it

Everything up to this point counted off-targets by what they do to a protein's amino acid sequence. That choice quietly decided which off-targets could be found at all, and this section is the correction.

A note on two terms first. A splice site is the junction where the cell cuts a non-coding stretch out of a gene's message before building the protein. Two bases on the intron side of each junction are effectively invariant across human genes: a junction the cell cuts into almost always begins GT and the one it cuts out of almost always ends AG. In pharmacy terms these are the tamper-evident seals on a blister pack. The pill inside can be perfect and the dose can be right, and if the seal is broken the pack does not get dispensed. Break the AG and the cell can no longer find that junction, so it usually splices to the next one it can find and the entire intervening block is dropped from every copy of the message. That is a much larger change than swapping one amino acid.

Why my earlier annotation could not see this class. Section 6.1 point 6 recorded that the 5-and-6 mismatch tail was annotated only where a hit overlaps a coding block. A hit sitting 1 or 2 bases from a splice junction lies in the intron, on the non-coding side. It was therefore outside the annotated set by construction. It was not scored as low risk and it was not ranked below other hits. It was invisible. I want that stated in those words, because "annotated only where it overlaps coding sequence" reads like a completeness caveat and it was actually a blind spot with a specific shape.

This is the second blind spot of the same shape in this analysis. Section 3.4.6 describes the first: a position-by-position comparison is structurally incapable of seeing a bulged alignment, so allowing one unpaired base promoted four protein-changing loci to 2 mismatches that the ungapped scan could not represent. Both blind spots were created by a filter that looked like a reasonable scoping decision, and in both cases the filter removed exactly the class of event the analysis was asking about. The general lesson for off-target prediction is that a filter chosen for tractability has to be tested against the question being asked, not just documented. Two independent instances in one project is enough to state that as a methods point rather than an anecdote.

What the rescan covered. I rescanned the same 3,099,750,718 base pairs of GRCh38 as the original scan, which is the no-alt analysis set across 194 sequences with the Epstein-Barr virus decoy contig excluded, both strands, ungapped, keeping every editable hit at 5 and 6 mismatches regardless of whether it touches a coding block. That returns 1,924,387 editable tail hits across the eight guides, 219,802 at 5 mismatches and 1,704,585 at 6. The 6-mismatch figure reproduces the count in section 6.1 exactly, which is the check that the rescan and the original scan agree on the part they share. The lead guide contributes 49,376 of those hits, of which 856 place an editable adenine within 2 bases of a splice junction.

An adenine base editor cannot damage the two junction seals equally, and the asymmetry is the result. The editor converts adenine to guanine and nothing else. The AG that ends a cut-out junction has an adenine in its first position, so that position is directly editable and an edit changes AG to GG, which the cell no longer recognises. The GT that begins a junction has no adenine on the sense strand at all; the editable base is the complementary thymine, reached only from the opposite strand, and editing it yields GC rather than destroying the seal. GC is a real minor-class junction that the cell can still use, if less efficiently. So an adenine base editor can destroy an acceptor and can only weaken a donor. Across all eight guides that gives 1,184 distinct genomic adenines that would destroy an invariant AG against 256 that would weaken a GT, and in the 139-gene cardiac panel 20 against 2. Counted as guide-by-adenine records rather than distinct positions the same data gives 18,422 and 2,883, because one genomic adenine can be reached by several guides; the distinct-position count is the one to quote. The two classes are reported separately in GAP_A_CARDIAC_SPLICE_DINUCLEOTIDE.csv and should not be pooled.

The lead guide has two, and one of them matters.

OBSCN is the finding. Obscurin is a giant sarcomeric scaffolding protein that binds titin, and its variants are associated with cardiomyopathy. At 35.49 TPM in heart left ventricle it is the second highest of 54 GTEx tissues (ENSG00000154358.20, GTEx v8), which puts it level with SCN5A itself at 35.11 TPM in the same tissue. Heart is where this gene lives. The edited adenine is position 1 of the acceptor dinucleotide for an exon that is 282 base pairs long and therefore an exact multiple of three, so skipping it removes 94 amino acids in frame rather than shifting the reading frame: residues approximately 4617 to 4710 of the 6,620-residue NM_052843.4 product, and 5574 to 5667 of the 8,925-residue NM_001386125.1 product. All four affected transcripts are curated NM_ records, not computational models.

Why OBSCN outranks every missense hit in Table 3 despite being further away in mismatch distance. Table 3's worst entry swaps one amino acid for another. This hit can remove a 94-residue block from every copy of a structural protein's message in the tissue being treated. Consequence class and mismatch distance are different axes, and a reader should not have to infer the ranking: on consequence class OBSCN is the most severe predicted off-target of the lead guide, and on mismatch distance it is among the least likely. MSH6 at 3 mismatches remains the site I would assay first on probability grounds. OBSCN at 6 mismatches is the site whose consequence, if it were edited, would be worst.

JPH2 is reported and explicitly down-weighted. Junctophilin-2 is a genuine cardiac gene at 49.98 TPM in left ventricle, and the edited base does sit on an invariant acceptor, but only in XM_006723833.5, a computationally predicted transcript model. No curated NM_ transcript of JPH2 has a junction there. The consequence is conditional on that model being real, which I cannot establish from sequence, so it goes on the sequencing panel and not into the risk narrative.

What this does not do. It does not retire the lead guide. Both sites sit at 6 mismatches, editing probability falls steeply with mismatch count, and neither reads open in purified cardiomyocyte accessibility data (OBSCN 0 of 3 experiments, JPH2 0 of 3; OBSCN is open in 2 of 12 bulk left-ventricle DNase experiments). What it does is convert an unassessed tail into two named sites with a sequencing plan. Both have been added to AMPLICON_PANEL.csv as tier T4_CARDIAC_SPLICE_mm6, taking the panel from 16 sites to 18, with primers designed under the same constraints as the original 16 and each edited base at least 89 bases from either primer. A named tail risk with a plan to measure it is a stronger position than an unexamined tail.

One caveat about the underlying table that anyone reusing it must know. The original scan stored a single editable-adenine coordinate per hit rather than the full set. An adenine base editor edits every adenine in its window, and these windows average 1.74 editable adenines each. Recomputing the complete set for every candidate window is what surfaced the JPH2 site, which the stored single coordinate missed. ABE_TAIL_ANNOTATION.csv and the scan tables list one editable position per hit, not all of them, and any reuse for consequence prediction has to recompute the window from the reference sequence.

Prior art, and it bounds this contribution. Using a base editor to disrupt a splice site is an established deliberate technique, not a discovery here. Adenine base editing of an intron acceptor to inactivate a gene was reported in 2018, a dedicated design tool for splice-disrupting guides exists (SpliceR), and a head-to-head screen in primary cells found that targeting donors gives more reliable knockouts than targeting acceptors (Kluesner et al., Nat Commun 2021, PMID 33893286). The asymmetry described above is therefore consistent with published mechanism rather than new mechanism: the field already knows which seal an adenine editor can break.

What is new here is the direction of the question. That literature asks how to disrupt a splice site deliberately, and chooses the editor and the seal accordingly. This analysis asks the inverse: given a guide chosen for an entirely different purpose, how much unintended splice-disrupting capacity does it carry across the genome, and is that capacity symmetric between the two seal types. I have not found that quantified for an off-target set, and the answer bears on any therapeutic base-editing design rather than on knockout screens. The 4.62-fold excess of acceptor-destroying over donor-weakening positions means an off-target analysis that treats splice sites as one category will misestimate the risk, and one that ignores intronic hits entirely, as my own earlier version did, will not see it at all.

3.4.8 A fourth wrong consequence call in my own corrected table, and a truncation in the deposited copy of it

Added 6 August 2026. Two defects in tables this paper published. Neither changes a count, a consequence class, a risk tier, an editor route, a sequencing-panel entry or the recommendation. They are reported in the Results, and again in section 6.1, because they are errors in this paper's own data rather than caveats about someone else's.

Section 3.4.6 recorded three wrong consequence calls whose shared cause is that the edited codon straddles an exon-exon junction: three consecutive bases read off the genome spell AGG, which is arginine, where the spliced codon spells serine. Those three — ALPK3 p.Arg48Gly and CCDC180 p.Arg317Gly and p.Arg320Gly — were found by verifying corrected calls against full-length reference protein. What was never done was to count how many calls are exposed to that failure mode, as opposed to how many were caught. That census has now been run at three nested levels, and the exposure is small, bounded, and contains one error nobody had seen.

The exposure first, because it is the negative and it is the larger of the two results. Across the consequence table reconstructed to its full size, 67 of 8,843 transcript-level protein calls sit on a codon that straddles an exon-exon junction. That is 0.758 percent, at four genomic adenines out of the 577 in this scan that land in a coding sequence, in four genes: ALPK3, CCDC180, KCNQ2 and UBR4. Exon structure alone predicts 4.80 exposed adenines against the 4 observed, so this analysis was neither lucky nor unlucky and there is no large hidden population of suspect calls. The test uses no sequence at all — a codon is exposed when its three genomic positions are not consecutive in the direction of transcription, which is a property of the exon coordinates — so exposure is decided exhaustively and cannot be biased by which reference base sits anywhere. A separate level of the census crossed every editable adenine in the scan against every RefSeq coding transcript containing it, 5,099 adenine-and-transcript pairs at 577 distinct adenines in 3,160 transcripts, to answer the question the other levels cannot: whether an exposed adenine was missed because no call was ever made on it. None was. The same four adenines are the whole population.

The error is KCNQ2, and it is the fourth wrong call of this type. At chr20:63,424,178 the table publishes p.Arg416Gly across seven transcript records, p.Arg406Gly across four, and p.Arg393Gly across one — three residue numbers across twelve transcript records. The reference residue is serine at all three positions, so the correct calls are p.Ser416Gly, p.Ser406Gly and p.Ser393Gly. The signature is identical to the three already known: the spliced codon reads AGC or AGT depending on the transcript, the genome-contiguous reading spells AGG. Residue 406 of NM_004518.6 and residue 416 of NM_172107.4 were checked against NCBI's own translated proteins and both read S. Residue 393 (XM_017027843.2) is computed rather than NCBI-verified; it is the same genomic adenine and the same spliced codon as the other eleven, differing only in that transcript's 5' structure, which is what moves the residue number.

Three residue labels change and nothing else does. The call is missense under both readings in all twelve records, so no consequence count in this paper moves. The site is not on the lead route: it is a minus-strand match at 4 mismatches with 1 seed mismatch, PAM TGC, reachable only by SpRY (spry_route_editable true, spg_route_editable false, pam_ok_SpCas9_NGG false), tier TIER5_coding_mm4. Heart left-ventricle expression is 0.033 TPM and cardiac_expressed is false — this is the neonatal-epilepsy and developmental-encephalopathy potassium channel, effectively absent from the tissue any therapy here would target. KCNQ2 is a serious gene in its own domain and a predicted missense off-target in it is worth naming correctly, which is why this correction is in the Results. It is not a cardiac risk finding, and correcting the residue changes no risk tier, no mismatch count, no editor route and no entry on the amplicon panel.

The fourth exposed gene was called correctly, and it is the case that shows exposure is not the same thing as error. UBR4 contributes 50 of the 67 exposed records at a single adenine, chr1:19,105,189, spanning 35 distinct residue numbers from 4132 to 4213 because fifty transcript models differ in 5' structure over one shared genomic base. All fifty are synonymous under both readings. Thirteen are still printed in the deposited file and all thirteen read serine, matching the spliced codon and NCBI residue 4168 of NM_020765.3. The other thirty-seven cannot be adjudicated at all, for the reason in the next subsection.

So the count of demonstrably wrong consequence calls attributable to junction straddling is 17 transcript-level records at three adenines — one ALPK3, four CCDC180, twelve KCNQ2. Five of the seventeen were already known, being the three verified rows expanded to their transcript records; twelve are new. In gene-and-site terms the arithmetic is three known wrong calls plus one new. Four adenines are exposed and wrong calls occur at three of them, because UBR4's produced none.

All four codons split the same way, with codon bases one and two at the end of one exon and base three at the start of the next, and all four junctions are canonical GT-AG donor-acceptor pairs, checked directly in the reference sequence, so none of this is an annotation artefact of a non-canonical or single-transcript exon boundary.

One observation about which cases failed, offered as an observation rather than a rule, because n is four. The three wrong calls all place the edited adenine at codon position one, where the base carried across the junction is one the caller had to read but never had to edit. The one correct call, UBR4, places it at codon position three, which is itself the base across the junction. The caller appears to fail when it reads across a junction and to survive when it edits across one. All fifty UBR4 records, printed and destroyed alike, sit at codon position three, which is a reason to expect the thirty-seven unadjudicable ones were also right and is not evidence that they were.

A data-integrity defect in the file this paper directs every consequence figure to, and it is the larger finding. ABE_CONSEQUENCE_RECHECK.csv, deposited in this paper's data archive, silently truncates its recomputed_aa field at exactly 1,500 characters. Fifty-eight of its 668 call-carrying rows sit at exactly 1,500 characters and each ends mid-token — the first stops at …NRXN3:XM_047431956.1: with no residue and no consequence class — and no row exceeds it. The next longest field below the cap is 1,469 characters, so a pile-up at a round number is the signature of a fixed-width slice rather than of rows that happened to be long. The file prints 7,137 of the 8,843 transcript-level calls it should carry; 1,706, or 19.3 percent, are gone, and 28 genes lose at least one call. Only that column is affected: every other column of that file, and every column of the full-CDS verification table, tops out at a value with no rows piled at it. This defect is why the first version of the census reported "30 exposed calls of 7,496" — it had parsed the surviving tokens and taken their number for the size of the table. That figure is withdrawn and replaced by the 67 of 8,843 above.

What the truncation costs, at its real size. What is recoverable is which calls exist, not what the table said about them. The caller's universe is deterministic — one call per row, RefSeq coding transcript and codon number, over every editable adenine of the row's protospacer that lands in a coding sequence — and that rule was tested rather than assumed: on the 610 rows the cap never touched it predicts 5,032 calls and the file carries exactly those 5,032, with none missing and none extra. Applied to all 668 rows it gives 8,843. The honest limit is that the largest untruncated row it is validated against carries 38 calls, while the capped rows run to 153 reconstructed calls each, so above 38 it is an extrapolation of a rule that never failed below it rather than a verified reproduction. The published residue letters for the 1,706 destroyed calls are unrecoverable, which is exactly why 37 exposed UBR4 records are counted above as unadjudicable rather than as correct. Any statement of the form "n published calls were wrong" in this paper is a statement about the 7,137 calls the deposited file still prints, not about the 8,843 it should carry. The producing script is not in my tree and there is no second copy of the CSV, so the clean repair is to regenerate the table without the cap and redeposit it; until that is done, "this cannot be counted" is the correct answer for those 37 records, and I record it as an open defect rather than as a closed one.

A second, smaller cap of the same class, named so the next reader does not rediscover it as a surprise: the older ABE_OFFTARGET_SCAN.csv caps its transcripts column at 4 entries per row, with 671 rows sitting at exactly 4, and its aa_consequences column at 6 entries, with 345 rows at exactly 6. That table is superseded for consequence calls by the recheck table, as section 7 already states, and no count in this paper depends on it for them.

Method and validation. Annotation is UCSC hg38 ncbiRefSeq.txt.gz, retrieved 6 August 2026, md5 c4e88a1a1e6a526fa7c5155521a513fe, 197,654 transcript records of which 145,446 have a coding sequence; reference bases come from the UCSC hg38 sequence endpoint. Validation was against material the census did not itself produce. For all 133 rows of the full-CDS verification table the computed codon number, both codons, the reference residue and the protein length equal the independently recorded values, zero failures in 133. The reference residue agrees with the published table on every non-straddling call it still prints, 7,466 of 7,466. Run against the superseded pre-recheck consequence column the census independently rediscovers the one-residue offset of section 3.4.6, reproducing 894 displaced calls at 133 distinct adenines. Six residues pulled from NCBI's own GenBank translations agree with the spliced reading. What is measured: the exon coordinates, the reference bases, the four canonical GT-AG junctions, the six NCBI residues, and every field length, row count and exposure count above. What is inferred: the genome-contiguous codon column, which reconstructs what a genome-contiguous reader produces rather than observing the code that produced the original calls, and the existence of the 1,706 destroyed calls. What is not established at all: what the published table said about those 1,706, and in particular about the 37 exposed UBR4 records among them.

Where the census tables are. The four census tables — the per-row exposure determination, the 8,843 reconstructed calls flagged for whether the published file still prints them, the 5,099 adenine-transcript pairs, and the record of what the truncated file actually contains — together with the full method and limitations, are in the project record as JUNCTION_STRADDLING_CENSUS.md and its CSVs. They are not in the version-1 data archive 10.5281/zenodo.21799234, which was built before the census was run, and adding them is listed as an open item against the next version of that deposit. (Corrected 7 August 2026. The struck sentence was true of version 1 and is no longer the whole truth, so it is left visible rather than rewritten.) The four census tables are JUNCTION_STRADDLING_PER_ROW.csv, JUNCTION_STRADDLING_RECONSTRUCTED_CALLS.csv, JUNCTION_STRADDLING_ALL_ADENINES.csv and JUNCTION_STRADDLING_RECHECK_CALLS.csv, with the method and limitations in JUNCTION_STRADDLING_CENSUS.md. They are not in version 1 of the data archive 10.5281/zenodo.21799234, which was built before the census was run, and they are in version 2. If you are holding version 1, they are not there.

3.4.9 The figures in this paper were regenerated from the broken file, and they all reproduce — which is a narrower result than it sounds

Added 7 August 2026. This section exists because section 3.4.8 reports a defect in a file this paper sends every reader to, and reporting a defect is not the same as establishing what it costs. This establishes what it costs. The answer has two halves and the second is the one to carry away: the counts survive, and the audit trail does not.

The escalation this answers, stated as it was put to me. A paper published under a permanent identifier depends on a file that cannot be repaired, and version 1 of the paper does not say so. That is the same defect class as paper 10's false data availability statement in this project's own set, and it is worse in kind, because paper 10's missing tables could be regenerated and this file cannot be. The question is therefore not whether the file can be fixed — it cannot — but whether anything published on top of it can be checked.

The structural fact that decides it. The truncation is confined to one column, and it is not the column the figures come from. recomputed_aa is the per-transcript detail column: the thing a reader uses to audit a single call. Every count in this paper is a site-level count, and a site's consequence class lives in recomputed_worst and its region in recomputed_region, which are separate columns of the same file. Measured across all twelve columns of ABE_CONSEQUENCE_RECHECK.csv, only recomputed_aa shows the pile-up at a round width that is the signature of a fixed-width slice — recomputed_worst tops out at 13 characters with 5 rows there, recomputed_region at 15 with 13,112 rows there. So the file is broken in its audit trail and intact in its summary, and this paper's numbers are computed from the summary.

What was done. p2_regen_consequence_figures.py, deposited in version 2 of the data archive, regenerates every consequence figure in this paper from those untruncated columns joined to ABE_OFFTARGET_SCAN.csv, and compares each against the number printed here. 102 quantities were compared. 102 reproduce. None differs. The full print-out is P2_VERIFICATION_OUTPUT.txt and the figure-by-figure table is P2_CONSEQUENCE_FIGURE_REGENERATION.csv, both in that archive.

The one way the truncation could still have moved a count, adjudicated rather than argued. A site's class is the worst class among its calls, so a destroyed call could in principle hide a class outranking the one the file records. Protein-changing here means missense or start-lost, which is the top class these counts use, so the only rows at risk are truncated rows recorded as synonymous. There are exactly two, and they are one locus reached by two overlapping SpRY guides: chr11:68,591,644 in PPP6R3, 4 mismatches, tier TIER6_other_mm4, cardiac_expressed false, not reachable by the lead SpG guide at all. All 148 reconstructed calls across those two rows sit at one genomic adenine, at codon position 3, on the plus strand, with none straddling an exon junction, so every call reads the same three bases and the same substitution. The 70 that survive are all p.Asp→Asp. The 78 destroyed ones cannot be anything else. The per-row adjudication is P2_TRUNCATION_ADJUDICATION.csv, one row for each of the 58 truncated rows.

A supporting check that makes the whole approach testable rather than assumed. On the 610 call-carrying rows the cap never touched, recomputed_worst equals the worst class among that row's own printed calls in 610 of 610 cases, zero disagreements. Wherever that column can be checked against the detail it summarises, it is faithful.

A reconciliation this paper did not carry, and a reader would have hit it. Counting well-formed calls in ABE_CONSEQUENCE_RECHECK.csv gives 7,496, and section 3.4.8 says the file prints 7,137. Both are right. The difference is 359 duplicate calls, a second call on a (transcript, codon) pair that already has one, and the cause is that a protospacer can put two editable adenines inside the same codon of the same transcript. Each adenine gets its own call, so the pair shares a key and differs in consequence — which is why every one of the 359 disagrees with its partner rather than repeating it: 208 are missense-against-synonymous and 151 are missense-against-missense at different residues. LIPK codon 239 is the clearest case, where one adenine gives p.Gln239Arg and the other p.Gln239Gln. This paper's counting rule is one call per row, transcript and codon, so 7,137 is the distinct-key count and 7,496 is the raw call count, and neither is an error.

What this does not establish, stated at the same volume as what it does.

• Not one of the 1,706 destroyed calls is recovered. Nothing recovers them. What the published table said about each is gone, and the regeneration does not touch that.
• The deposited file is republished in version 2 byte for byte unchanged, because it cannot be repaired: the producing script is on neither machine, there is no second copy, and there is no version control on this project's science tree. A replacement table would be a new computation with its own validation burden and it has not been built.
• The 37 exposed UBR4 records stay unadjudicable on the one question that needs the missing text.
• This is a check of the arithmetic, not of the biology. It establishes that the counts printed here are the counts the file supports. It does not establish that the file's consequence calls are right; that is what section 3.4.6's recheck and section 3.4.8's census address, and both of those found errors.
• A reader auditing an individual per-transcript residue label for one of the 1,706 has nothing to audit and will continue to have nothing. That is the permanent cost, and it is why this section does not close section 3.4.8's open defect. It bounds it.

[Figure 1 is supplied as a separate file, FIG1_OFFTARGET.png. Upload it with the submission.]

Figure 1. Predicted off-target load of the SCN5A R104Q base-editing guides. Complete both-strand comparison across the 3,099,750,718 bp of GRCh38 scanned, at all 20 protospacer positions with no seed heuristic, every candidate site scored on PAM compatibility and the presence of an editable adenine. Complete within that scanned span and mismatch budget; the excluded contigs, the unannotated 5-and-6 mismatch tail, and bulged alignments are stated in section 6.1. (a) Cost of PAM relaxation with guide identity held constant: the identical 20-mer scored under four PAM rules. The NGG row is calibration only, since no NGG guide exists at this site. (b) Naive sequence-match counts (open circles) against editable counts (filled). The SpG route's 4,388 matches reduce to 655; the SpRY route's 24,555 reduce only to 22,710. (c) Cumulative editable sites against mismatch budget, ungapped alignment. The lead guide has no editable site at 1 mismatch under that model, marked by the triangle at the axis floor; under bulge tolerance it has five alignments at four loci there, none protein-changing. (d) The lead guide's protein-changing hits in genes expressed at 1 TPM or above in heart left ventricle, ordered by expression; MSH6 at 3 mismatches is the closest under ungapped alignment, which is the model this panel uses. Gene symbols name the transcript carrying the coding consequence. Counts are off-target sites only, excluding the on-target locus. Cardiac expression is GTEx v8 bulk left-ventricle median.

3.5 Predicted rescue, reported under both mechanisms

Correcting the mutant allele in a single cell takes that cell from the untreated baseline to 100 percent. Because editing is mosaic, tissue-level current depends on the corrected fraction. I report both candidate mechanisms side by side because the mechanism is unresolved (section 1.2).

Table 4. Predicted sodium current as a percentage of a heart with two working alleles. Full table in MS_TABLE4_RESCUE_MODEL.csv.

Two observations. First, the efficiency anchor matters more than the mechanism at low correction rates: at 20 percent editing, the two mechanisms predict 45.0 and 56.6 percent of normal, and neither is far above what a simple single-allele loss would give. Second, at the 60 percent that sufficed in mice, the dominant-negative mechanism predicts 72.5 percent of normal and haploinsufficiency predicts 78.3. Both mechanisms support the same conclusion, that high correction fractions are the ones worth targeting, and they differ in how much the therapy is worth rather than in whether it works.

I record that my first version of this analysis cited only the weaker efficiency anchor, Levy 2020 at 20 percent cardiac editing. That figure is drawn from a general tissue survey covering brain, liver, retina, heart and skeletal muscle, and it predicts a benefit barely above single-allele loss. The stronger and more relevant anchor, Qi 2024, was surfaced by a parallel analysis of a different therapeutic route rather than by my own literature search, and finding it materially raised this route's standing.

4. Recommendation

Use the SpG guide at target position 5, TTCCAGTTCAGTGCCACCAA with PAM CGC, delivered with SpG or Cas9-NG. Do not use the SpRY route unless SpG fails experimentally.

The reasoning is not merely that SpG scores better on the tables. It is that SpRY's advantage was never specificity. SpRY exists to provide targeting options where no NG PAM is available. At this locus an NG PAM is available, at the best editing position, with no bystander adenine in its window. The SpRY guides therefore buy nothing and cost roughly 35 times the editable off-target load, including coding hits in two cardiac genes, one of which is itself Brugada-associated.

The empirical programme this design implies, in order:

1. Targeted deep sequencing of the on-target site in a cell model carrying R104Q, to measure editing efficiency and confirm the absence of bystander editing.
2. Targeted deep sequencing of the ranked sites in ABE_OFFTARGET_SCAN.csv, prioritising the 3 sites at 2 mismatches and the MSH6 and ATP4A sites at 3.
3. An unbiased empirical off-target assay, GUIDE-seq or CIRCLE-seq, which is the only way to detect the classes of event no sequence scan can predict.
4. An allele-resolved surface-expression measurement to settle the mechanism question of section 1.2.

5. Discussion

5.1 The methodological results, which are the contribution

Neither of the two results in this section is about R104Q, and neither requires it.

Editability, not similarity, is the right scoring axis for base editors. The field's default off-target ranking is by mismatch count. For a nuclease that is defensible, because a bound nuclease cuts. For a base editor it is not, because a bound editor with no adenine in its window does nothing. In this dataset that distinction removes 85.1 percent of the apparent risk for the lead guide, and it correctly demotes the single closest sequence match in the genome, a paralogous sodium-channel site at 1 mismatch that carries G where the guide carries A.

The uneven filtering is the part that matters, and it is easy to miss. Overstating risk uniformly would be tolerable, because it would preserve the ordering between options. This does not: the SpG route's naive count is inflated 6.7-fold while the SpRY route's is inflated only 1.08-fold, because with no PAM requirement left there is almost nothing to filter on. Two routes compared on naive counts would appear 5.6 times apart when the editable gap is 34.7 times. A similarity-based comparison therefore does not merely exaggerate risk, it exaggerates it in the direction that favours the worse editor. Any base-editor design that ranks off-targets by mismatch count alone inherits that bias.

PAM relaxation has a measurable, separable cost. By holding the 20-mer identical and varying only the PAM rule, the 6.2-fold penalty from NG to unrestricted is attributable to the editor and not confounded with guide choice. As far as my bounded search reaches, this comparison has not been reported in this controlled form, and it is cheap to run at any locus: it requires one guide and one genome scan, not a panel. The practical corollary is a design rule that does not mention this gene: use the least relaxed editor that can reach the target, and treat each step of PAM relaxation as costing roughly a factor of two to three in editable off-target load.

Why the worked example was worth using. A synthetic target would have demonstrated both points, but it would not have produced the case that makes the second one bite. Standard SpCas9 cannot reach this site at all, so relaxed-PAM editors are mandatory rather than a convenience, and the cost measured in section 3.4.2 is a cost that has to be paid rather than avoided. That is the situation in which a designer actually needs the number.

5.2 What the design half establishes, and what it does not

The design is a prediction, not a therapy and not an efficacy claim. It establishes that the variant is chemically correctable by the most developed editor class, with a guide whose allele selectivity rests on substrate absence rather than on binding discrimination; that standard SpCas9 cannot reach the site; and that among the available engineered options one guide dominates on every measured axis.

It does not establish that editing will occur at a useful rate at this site, which is not predictable from sequence. It does not establish that the predicted off-target profile matches the empirical one. It does not establish that the dominant-negative mechanism operates in human heart. And it does not establish anything about whether any specific person should receive such a therapy, which is a clinical judgement belonging to physicians with access to the full patient picture.

I am also not claiming that pointing a base editor at SCN5A is new. Section 1.0 states the opposite: it is established, and the review literature already lists it as such (PMID 42193469). What I could not find is a design for a loss-of-function SCN5A variant, and section 5.3 argues that absence is structural.

5.3 The harder class of variant, and why designing for it is itself a contribution

Every published cardiac base-editing success I could find corrects a gain-of-function variant with a measurable electrocardiogram readout in mice. Qi 2024 corrected SCN5A T1307M and eliminated the QT-prolongation phenotype. That is the closest possible precedent short of the same variant, and I want to be exact about the one way it does not transfer.

T1307M causes LQT3, whose animal endpoint is the QT interval on a surface electrocardiogram: cheap, quantitative, repeatable, non-invasive, and measurable in a conscious mouse. R104Q causes Brugada syndrome, whose diagnostic ST-segment signature depends on a transmural voltage gradient across the right ventricular wall. Mouse hearts, beating near 500 beats per minute with a different complement of repolarising currents, do not reproduce that gradient. A Brugada mouse therefore offers no equivalent phenotype readout.

This is why I read the absence of prior loss-of-function designs as structural rather than accidental. The class with the cheap endpoint got done first, which is what one would expect. The corollary is a specific and uncomfortable prediction for anyone attempting this correction: on-target editing efficiency will be measurable in a mouse, and the thing one actually wants to know, whether arrhythmia risk falls, will not. That is a harder experimental position than the precedent occupied, and no improvement to the guide changes it.

The available routes are all partial. Human induced pluripotent stem cell cardiomyocytes give a directly measurable sodium current density, which tests the rescue arithmetic of section 3.5 at the cellular level, but they have no transmural gradient either, so they cannot produce the diagnostic signature. Larger animals with more human-like repolarisation could, at a cost and ethical burden that a single-variant design does not currently justify. The honest position is that the cellular endpoint is reachable and the clinical one is not, and that a design for this class of variant should say so at the outset rather than let the reader assume the precedent covers it.

Stating that plainly, and designing for the harder class anyway with the endpoint limitation on the record, is the part of this paper I would defend most strongly to a reviewer. It is also the part that generalises to the other loss-of-function channelopathies, where the same asymmetry applies.

5.4 The comparator, restated because it is easy to get wrong

I stated in section 1.3 that the comparator is no treatment rather than a functioning defibrillator. That cuts in both directions and I want both directions on the record.

It raises the tolerable risk, because most carriers have no protection at all and a device would not treat the disease in any case. It does not lower the standard of evidence. An off-target edit in MSH6 or TRIM63 would be a permanent change in a person who may never have had an arrhythmia. The measurement has to be honest in both directions, and the reason this paper spends most of its length on the off-target profile rather than on the rescue arithmetic is that the off-target profile is the part that could cause harm.

Sections 3.4.7 and 3.4.8 used to sit here, and an editorial instruction used to sit here with them

Moved 6 August 2026 23:35 CDT. Nothing was deleted except one line that was never part of the paper, and that line is quoted below in full. A reader working from the version of record — the PDF deposited at 10.5281/zenodo.21799850 on 5 August 2026 — will find sections 3.4.7 and 3.4.8 printed at this point, after section 5.4 and before section 6, and will find them introduced by a heading reading:

## Insert into PUBLISH_2_BASE_EDITOR_SPECIFICITY.md as section 3.4.7, immediately after 3.4.6

That heading was an instruction to whoever assembled this paper, not text of the paper. It was pasted in with the section it describes and never removed, and it was deposited. It is removed here rather than struck through, because striking it through would leave a sentence in the manuscript that addresses a file-editing operation and not a reader — the only case in this document where deletion is the honest option. It is quoted above so that nothing is lost.

The instruction was followed, 6 August 2026, and it was correct. Section 3.4.7 now sits immediately after 3.4.6, and section 3.4.8 immediately after 3.4.7, inside section 3.4 where the numbering says they belong. Section 3.4.5 was also out of order — it was printed after 3.4.6 — and is now printed before it. Not one word of any of the four sections was changed by the move, and Figure 1 stays where it was, at the end of section 3.4 just before 3.5, because its caption reports the bulge-aware results that section 3.4.6 introduces.

Order in the deposited version 1: 3.4.1, 3.4.2, 3.4.3, 3.4.4, 3.4.6, 3.4.5, Figure 1, 3.5, 4, 5, 5.1–5.4, [the editorial heading], 3.4.7, then section 6. Order here: 3.4.1 through 3.4.8 in numerical order, then Figure 1, then 3.5, 4, 5, 5.1–5.4, then section 6.

This changes page layout throughout and is therefore an editorial decision on a published paper, not a typographical fix. It is recorded as such in SUBMIT_THESE/V2_STAGING/05_paper_2/VERSION_NOTE.md and it is reversible: no text was rewritten, so restoring the deposited order is a matter of moving the same blocks back. Nothing has been uploaded.

A correction to the rescue baseline, made 4 August 2026

An earlier version of this analysis rescaled the measured heterozygous current to a two-allele heart by dividing by two, giving 34.1 percent. That assumed perfect additivity. O'Neill 2022 measured the two-allele case directly: their WT+WT cells read 218.4 ± 7.7 percent of a single wild-type allele (n = 199), not 200. Dividing by 2.184 rather than 2 gives 31.3 percent, and the haploinsufficiency comparator moves from 50.0 to 45.8 percent by the same argument. Every rescue figure in this document uses the measured control. The direction of the correction is unfavourable to the untreated baseline and favourable to the per-cell gain, which rises from 2.93-fold to 3.20-fold.

6. Limitations, and what would falsify this

I have used the phrase "predicted off-target profile" throughout deliberately. The following are the specific reasons that word is load-bearing.

6.1 Limitations of the off-target prediction

1. Mismatch count is not binding affinity. I report position-resolved mismatches and a separate seed-region count, which captures the established fact that PAM-proximal mismatches are least tolerated. I applied no trained activity-prediction model. A systematic benchmark of 14 in silico off-target prediction tools against 3,827 deep-sequenced sites found correlation with measured indel frequency to be weak to moderate, and recall of roughly 77 percent among the top 500 predicted sites and 83 percent among the top 1,250, leaving a substantial fraction of true off-targets undetected (PMID 42503868). That benchmark concerns Cas9 nuclease tools rather than base editors, so it is an indication of the general difficulty rather than a direct measurement of my pipeline's performance. My method is a complete enumeration of sequence matches with an editability filter, which is a different object from a trained ranker, and it inherits the same fundamental limitation: it identifies candidates, it does not predict magnitude.
2. Bulges were modelled, to one nucleotide only. The rescan in section 3.4.6 allows up to one DNA bulge or one RNA bulge per alignment, never both in the same alignment, and caps bulged alignments at 4 mismatches. Larger bulges and two-bulge combinations were not searched, so this is a bounded search and not an exhaustive one. My deduplication rule treats a bulged alignment sharing a docking window with an editable ungapped hit as redundant; counting every alignment separately, as Cas-OFFinder-bulge does by default, inflates every bulge number here by roughly the 7 percent of bulged rows that are redundant.
3. This is a reference-genome scan, not my genome. Every count is against GRCh38. My own variants could create editable sites absent from the reference or destroy ones present in it. This is directly falsifiable by re-running the pipeline against my own sequence data, and tools now exist specifically for variant-aware off-target identification (PMID 40337925).
4. Alternate haplotypes were excluded by design. Correct for counting, but a guide could match a haplotype not represented in the primary assembly.
5. Annotation is RefSeq only. A site called intergenic here could be exonic under another annotation. Many hits fall in XM_-prefixed predicted transcripts, which are computational models rather than experimentally confirmed transcripts.
6. The 5-and-6 mismatch tail is annotated only where it overlaps coding sequence. The tail contains most of the 2,051,639 raw matches. Roughly 79 percent of tail hits pass the editability filter, which removes far less than I expected, and what collapses the tail is coding-sequence overlap instead: of 1,704,585 editable hits at 6 mismatches only 29,887, or 1.8 percent, fall in coding sequence. Because I annotated only the editable hits overlapping a coding block, a 5-or-6-mismatch hit that disrupts a cardiac promoter or a splice site would not have been seen, and the protein-changing tail counts do not include that class at all. For the lead guide the tail contributes 894 protein-changing loci, 128 at 5 mismatches and 766 at 6, none in a cardiac disease gene at 5 mismatches. Individually improbable, collectively not empty.

6a. I did not resolve the MSH6 accessibility disagreement. Its editable protein-changing rows read closed in purified cardiomyocyte data, 0 of 3 experiments, and open in bulk left-ventricle ATAC, 6 of 15. The cardiomyocyte reading is the relevant one for a cardiomyocyte-directed therapy and it is also the favourable one, and I decline to resolve the disagreement in that direction, because the cardiomyocyte tier is 3 experiments against 15 and picking the tier that returns the preferred answer is how a wrong result gets published. What would settle it: a primary purified-cardiomyocyte ATAC or DNase experiment at adequate depth covering chr2:47,798,7xx. 7. Expression evidence is partial and bulk. I resolved 295 of 330 genes in GTEx v8. A hit in an unresolved gene has no expression evidence here, which is not evidence of no expression. GTEx medians are bulk adult post-mortem tissue, not cardiomyocyte-specific and not developmental. One row in Table 3, CAVIN2, has no resolved expression value at all. 8. Guide-independent deaminase activity is outside any sequence scan. Base editors produce off-target editing that does not depend on the guide at all. No sequence method can predict it. 9. RNA off-target editing is outside the scope of this scan. Adenine base editors edit RNA as well as DNA. Qi 2024 detected none in treated mouse hearts, which is encouraging but is a measurement of their editor in their system. 10. Four residue labels in my corrected consequence table were wrong, and the exposure to that error is now counted rather than assumed. (Added 6 August 2026.) A codon that straddles an exon-exon junction is read incorrectly by a genome-contiguous reader, which is the cause of the three errors in section 3.4.6 and of the fourth, KCNQ2, in section 3.4.8. 67 of 8,843 transcript-level calls, 0.758 percent, are exposed to it, at four adenines in four genes; 17 transcript-level records at three of those adenines carry a wrong reference residue, 13 carry a right one, and 37 cannot be judged. The consequence class is unchanged in every one of them, so no count, tier or route in this paper moves — the defect is in residue labels only. What remains a genuine limitation is that the correction was derived by re-deriving the reference residue and not by re-deriving whether an editor would edit any of these adenines at a measurable rate, which is unmeasured here as everywhere in this paper. One of the three KCNQ2 residue numbers, 393 in XM_017027843.2, is computed from the exon model rather than checked against an NCBI translation. 11. The deposited ABE_CONSEQUENCE_RECHECK.csv is truncated, and 19.3 percent of its consequence calls are not in it. (Added 6 August 2026.) Its recomputed_aa field is capped at exactly 1,500 characters; 58 of 668 call-carrying rows sit at the cap ending mid-token; the file prints 7,137 of the 8,843 calls it should carry. This is a defect in a file that has been published, not a caveat about a method. Every consequence figure in this paper is drawn from that file, so every such figure is a statement about the 7,137 calls it prints. The reconstruction of which calls are missing is validated 5,032 times without a miss on the rows the cap never touched, but the published text for the missing 1,706 is gone and is not recoverable from anything. The producing script is not in my tree. Until the table is regenerated without the cap and redeposited, a reader reusing that file for any purpose other than reproducing this paper's numbers must know that a fifth of it is absent, and that the absence is invisible in the file itself because the truncated rows still look like well-formed rows. (Bounded, not closed, 7 August 2026. The sentence "every consequence figure in this paper is a statement about the 7,137 calls it prints" is now checked rather than assumed: section 3.4.9 regenerates every one of those figures from the file's untruncated recomputed_worst and recomputed_region columns and all 102 reproduce, and the two truncated rows at which a destroyed call could have moved a count are adjudicated individually and cannot have. This limitation is not withdrawn. What it costs is now known to be the audit trail rather than the arithmetic, and the audit trail is still gone. The file is republished unchanged in version 2 of the deposit because it cannot be repaired.)

Only an empirical assay can confirm any of this. GUIDE-seq, CIRCLE-seq, or targeted deep sequencing of the ranked sites in ABE_OFFTARGET_SCAN.csv.

6.2 Limitations of the rescue prediction

1. The mechanism is unresolved, as stated in section 1.2 and carried through Table 4 as two parallel columns rather than one number.
2. The arithmetic is linear and the heart is not. Table 4 assumes tissue-level sodium current is the corrected-cell fraction weighted average of corrected and uncorrected cell currents. Real cardiac conduction is a threshold phenomenon in a coupled syncytium, where a patchwork of restored and unrestored cells may behave better or worse than the average of the two. The mouse result that 60 percent correction eliminated the phenotype (PMID 37965733) is consistent with a threshold rather than with linearity. Table 4 should be read as an ordering, not a prediction of any specific current density.
3. The efficiency anchors are mouse, at research doses, in research vectors. Human cardiac delivery at comparable efficiency is unproven.
4. The editor and guide exceed single-AAV capacity, so a split-vector approach is required, adding reconstitution efficiency as a further variable between the anchor figures and any human result.
5. Editing efficiency at a specific site is not predictable from sequence. There is no efficiency number in this paper that was measured at this locus, and none of the figures in Table 4 is an efficiency prediction. They are the arithmetic of a rescue conditional on an efficiency that has not been measured.

6.3 Limitations of the literature and prior-art search

The search is bounded and I do not call it exhaustive. Two independent searches were run, one against PubMed through NCBI E-utilities and one against Europe PMC across five query families. Neither covers patents, company development pipelines, conference abstracts, non-indexed preprints, or non-English literature. Publisher websites were unreachable from my environment because of a network allowlist, returning proxy errors rather than server errors.

This bounds the prior-art claim specifically. When I write in section 1.0 that I could find no base-editing design for a loss-of-function SCN5A variant, that is a statement about two bounded searches, not proof that none exists. The most likely place for such work to be invisible to both is a patent filing or an unpublished industry programme, and neither search reaches those. I claim only that I looked in the two places named and did not find it. A failed retrieval is evidence about the retrieval.

Two further caveats on the same point. My own PubMed search for the closest precedent did not surface it; Qi 2024 reached me through a parallel analysis of a different therapeutic route, which is direct evidence that my search strategy has misses. And a keyword search cannot find work that uses vocabulary I did not think of, so the counts throughout are lower bounds.

6.4 What would falsify the central claims

Each of these is a specific, runnable experiment.

1. An empirical off-target assay showing measurable editing at the MSH6 or ATP4A sites, or at any site this scan ranked as low-risk. This would falsify the specificity claim and, depending on magnitude, the recommendation.
2. SpG failing to edit the on-target site efficiently in cells. This would force the SpRY route and turn its off-target burden from something to avoid into something to manage.
3. A trained activity model promoting a site this method ranked as distant. This would falsify the ranking without falsifying the enumeration. The bulge-aware half of this test has now been run, and it did promote sites: see section 3.4.6. 3a. Measured editing at the four two-mismatch bulged loci showing none detectable. TXNIP, SAMD15, SLC68A1 and HNRNPAB are the sites bulge tolerance brought closer than the ungapped scan could see. No detectable editing at them would mean bulged binding does not translate into bulged editing here, and would return the nearest protein-changing risk to 3 mismatches. 3b. Measured editing at chr20:38,139,629 changing TGM2 expression despite the synonymous codon. That would mean the reassurance about the four one-mismatch bulged loci rests on the wrong mechanism, since it currently rests on protein sequence and splice-boundary distance alone. 3c. A full-CDS translation of any corrected transcript showing my residue numbering off by one in the other direction. I verified 3 of the 133 substantive consequence corrections against full-length translated coding sequence; the remaining 130 rest on the same validated code path and were not individually checked against NCBI. Superseded 6 August 2026, and the superseding result is the reason section 3.4.8 exists. All 133 were subsequently verified against full-length reference protein, every reconstructed coding sequence matching the official translation exactly before being used to judge anything, and three of the 133 were wrong — ALPK3 p.Arg48Gly and CCDC180 p.Arg317Gly and p.Arg320Gly, all three from junction-straddling codons, all three unchanged in consequence class. The falsification test named here has therefore been run and it returned three failures, not zero. What has not been run is the equivalent verification for the 1,706 consequence calls the deposited table truncated away, and that test remains open in the sense that it cannot be run at all until the table is regenerated.
4. My own genome containing a variant that creates a close editable site absent from the reference. Directly checkable.
5. An allele-resolved surface-expression experiment showing simple haploinsufficiency. This would replace the 3.20-fold per-cell figure with 2.18-fold and weaken the therapeutic case without eliminating it.
6. Bystander editing detected at the on-target locus despite zero adenines in the window. This would indicate the editing window is wider than the positions 3 to 9 assumed here, which would invalidate the bystander analysis for all seven guides.

6.5 The negative result, stated plainly

No guide at this locus is clean in an absolute sense. The lead guide has 655 predicted editable off-target sites in the genome at up to 4 mismatches, 22 of which change a protein, one of them in a cancer-predisposition gene. (Corrected 6 August 2026: this sentence read "16 of which change a protein". 16 is the retired count from before the consequence recheck of section 3.4.6, and section 3.4.4 of this same paper already states that the missense count is 22 and not 16. The paper contradicted itself, and the number in the section headed "the negative result, stated plainly" was the wrong one. The correction makes the negative result larger, not smaller. The record deposited on 5 August 2026 prints 16 here.) Standard SpCas9, the best-characterised and most specific option, cannot reach the site at all. "The lead guide survives" means "best of the available options, with a specific and testable mitigation list", not "safe".

The honest headline is that this site is workable with SpG and would be difficult to defend with SpRY.

Corrections, 6 and 7 August 2026

Three corrections were made to this paper on 6 August 2026. Four, as of 6 August 2026 23:35. Five, as of 7 August 2026. The fifth is new section 3.4.9 and it is the only one that was written in response to an escalation from outside this paper, namely that a deposited paper resting on an unrepairable file without saying so is the same defect class as paper 10's false data availability statement. The fourth is structural rather than numerical and is described in its own row below and, at greater length, at the point in the document where the displaced sections used to sit. Two of them are the same defect twice: a count that the consequence recheck of section 3.4.6 had already superseded, left standing in a table and in a sentence that were never revisited when the rest of the paper was. The third is a new finding. None changes a conclusion or the recommendation. Every superseded figure is left visible at its site with its replacement rather than removed.

The direction of the first three is unfavourable to this paper: more predicted protein-changing off-targets on every route, one more wrong call in its own table, and a published data file that is a fifth incomplete. That is the reason to state them at the front of the correction rather than at the back. The fourth is unfavourable in a different currency: it says that a paper this project deposited under a permanent identifier carried a build instruction in its section headings for a day and nobody noticed.

The fifth is favourable and is stated last for that reason. It is the only one that makes this paper look better than it did, so it is the one to be most sceptical of. Two things bound it. It checks the arithmetic and not the biology: it establishes that the figures printed here are the figures the file supports, and says nothing about whether the file's consequence calls are right — sections 3.4.6 and 3.4.8 address that and both found errors. And it repairs nothing: the deposited ABE_CONSEQUENCE_RECHECK.csv is republished in version 2 byte for byte unchanged, 1,706 consequence calls are still gone, and a reader auditing any one of them still has nothing to audit.

None of the five moves a consequence class, a risk tier, an editor route, an amplicon-panel entry, the 6.2-fold PAM-relaxation cost, the 34.7-fold route ratio, the rescue arithmetic of Table 4, or the recommendation in section 4. The fourth moves no number at all. It does move page layout throughout, which is why it is an editorial decision on a published record rather than a typographical fix, and why it is recorded in SUBMIT_THESE/V2_STAGING/05_paper_2/VERSION_NOTE.md as well as here.

7. Data availability

All inputs are public and every derived file accompanies this paper.

Public inputs.

• Reference sequence: NCBI efetch, NC_000003.12 (GRCh38.p14), region 38630362-38630423.
• Assembly for the scan: NCBI GCA_000001405.15_GRCh38_no_alt_analysis_set.fna.gz, 872,949,833 bytes, md5 a08035b6a6e31780e96a34008ff21bd6.
• Annotation: NCBI RefSeq GCF_000001405.40_GRCh38.p14_genomic.gtf.gz, 56,947,363 bytes.
• Expression: GTEx v8, Heart_Left_Ventricle and Heart_Atrial_Appendage.
• Variant records: ClinVar VariationID 67780 (R104Q) and 67778 (R104W). Gene record: Ensembl ENSG00000183873.
• Functional baseline: PMID 35305865 Supplementary Table 1, read from the open preprint doi:10.1101/2021.09.22.461398 page 29.

All derived tables are deposited as a single archive with a permanent identifier. The identifier is recorded in DATA_DOI.txt alongside this manuscript and should be cited as the data source.

Derived files in that archive.

The archive also carries p2_regen_consequence_figures.py, which produces the last two of those from the other files in the archive and needs nothing but the Python standard library; P2_VERIFICATION_OUTPUT.txt, its print-out; P2_REGENERATION_NOTE.md, which states what is unrepairable and what was regenerated; and JUNCTION_STRADDLING_CENSUS.md, the census write-up with its method and limitations. All of these are in version 2 of the archive and none is in version 1.

Three reconciliations, so that a discrepancy does not look like an error. (The third was added 6 August 2026 and, unlike the first two, it is a defect rather than a reconciliation. It is placed here because this is where a reader decides which file to open.) First, the consequence columns of ABE_OFFTARGET_SCAN.csv disagree with this paper: they are the uncorrected calls, wrong on 133 rows, and are retained in the archive only so the correction can be audited row by row against ABE_CONSEQUENCE_RECHECK.csv. Read consequences from the recheck file. Second, summing the spry_route_editable column of ABE_OFFTARGET_SCAN.csv gives 22,712, while this paper reports 22,710 editable off-target sites for the SpRY route. The difference is exactly 2, and those two rows are on-target: the SpRY guides at positions 3 and 4 each carry a known bystander adenine inside the editing window, at window positions 8 and 9 respectively. Filter on is_on_target == False to reproduce every count in this paper. The SpG figure of 655 is identical either way, because the lead guide has no bystander adenine in its window. This is a useful independent confirmation: the genome scan rediscovered from sequence alone the bystanders that the guide enumeration had recorded separately.

Third, and it is a defect rather than a discrepancy: ABE_CONSEQUENCE_RECHECK.csv as deposited is truncated. Its recomputed_aa field is cut at exactly 1,500 characters, 58 of the 668 call-carrying rows end mid-token, and the file prints 7,137 of the 8,843 transcript-level consequence calls it should carry. The truncated rows are not marked and do not look truncated. Anyone reusing this file for consequence prediction must read section 3.4.8 first, and anyone counting calls in it should not report their number as the size of the table — that is the exact mistake the first version of the census made. The full measurement, the reconstruction of which calls are missing, and the limits of that reconstruction are in section 3.4.8. The census tables themselves are not in this archive, which was built before the census was run. (Corrected 7 August 2026: they are in version 2 of the archive and not in version 1. The four named above, plus JUNCTION_STRADDLING_CENSUS.md.)

Fourth, and this one is genuinely a reconciliation: counting calls in ABE_CONSEQUENCE_RECHECK.csv gives 7,496, and this paper says 7,137. (Added 7 August 2026.) Both are right. The difference is 359 duplicate calls — a second call on a (transcript, codon) pair that already has one — and the cause is that a protospacer can put two editable adenines inside the same codon of the same transcript, each of which gets its own call. That is why all 359 disagree with their partner rather than repeating it: 208 are missense-against-synonymous and 151 are missense-against-missense at different residues. This paper counts one call per row, transcript and codon, so 7,137 is the distinct-key count and 7,496 is the raw one. Neither is an error and nothing in the file says so, which is why it is written here.

Fifth, and it is the reason to trust the numbers in this paper despite the defect above: P2_CONSEQUENCE_FIGURE_REGENERATION.csv regenerates every consequence figure in it and all 102 reproduce. (Added 7 August 2026.) The truncation is confined to recomputed_aa; this paper's counts come from recomputed_worst and recomputed_region, which are untruncated. Section 3.4.9 states the method, the adjudication and — at the same length — what this does not establish.

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.

8. Competing interests, and the role of the author

I carry the variant used as the worked example in this paper. I have no financial interest in any editor, vector or reagent named. No funding was received. I performed the design, the scan, the analysis and the writing.

I disclose this because it can be read two ways, as a conflict of interest or as a reason for closer scrutiny of the design half of the paper. The methodological results in section 5.1 do not depend on which variant was used and can be checked without reference to me. For the design half, I note that n-of-1 base editing for rare disease is now an established framing in the literature rather than an unusual one (PMID 40664722), which is context rather than a defence.

Nothing in this paper is a treatment recommendation for any person, including me. Clinical decisions belong to the physicians who manage the patient.

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.

9. References

1. Levy-Nissenbaum E, Eldar M, Wang Q, et al. Genetic analysis of Brugada syndrome in Israel: two novel mutations and possible genetic heterogeneity. Genet Test 2001;5(4):331-4. doi:10.1089/109065701753617480. PMID: 11960580
2. Gütter C, Benndorf K, Zimmer T. Characterization of N-terminally mutated cardiac Na(+) channels associated with long QT syndrome 3 and Brugada syndrome. Front Physiol 2013;4:153. doi:10.3389/fphys.2013.00153. PMID: 23805106
3. Olde Nordkamp LR, Postema PG, Knops RE, et al. Implantable cardioverter-defibrillator harm in young patients with inherited arrhythmia syndromes: A systematic review and meta-analysis of inappropriate shocks and complications. Heart Rhythm 2016;13(2):443-54. doi:10.1016/j.hrthm.2015.09.010. PMID: 26385533
4. Qi T, Wu F, Xie Y, et al. Base Editing Mediated Generation of Point Mutations Into Human Pluripotent Stem Cells for Modeling Disease. Front Cell Dev Biol 2020;8:590581. doi:10.3389/fcell.2020.590581. PMID: 33102492
5. Levy JM, Yeh WH, Pendse N, et al. Cytosine and adenine base editing of the brain, liver, retina, heart and skeletal muscle of mice via adeno-associated viruses. Nat Biomed Eng 2020;4(1):97-110. doi:10.1038/s41551-019-0501-5. PMID: 31937940
6. Walton RT, Christie KA, Whittaker MN, et al. Unconstrained genome targeting with near-PAMless engineered CRISPR-Cas9 variants. Science 2020;368(6488):290-296. doi:10.1126/science.aba8853. PMID: 32217751
7. Koblan LW, Erdos MR, Wilson C, et al. In vivo base editing rescues Hutchinson-Gilford progeria syndrome in mice. Nature 2021;589(7843):608-614. doi:10.1038/s41586-020-03086-7. PMID: 33408413
8. Liu N, Olson EN. CRISPR modeling and correction of cardiovascular disease. Circ Res 2022;130(12):1827-1850. doi:10.1161/CIRCRESAHA.122.320496. PMID: 35679361
9. O'Neill MJ, Muhammad A, Li B, et al. Dominant negative effects of SCN5A missense variants. Genet Med 2022;24(6):1238-1248. doi:10.1016/j.gim.2022.02.010. PMID: 35305865
10. Qi M, Ma S, Liu J, et al. In Vivo Base Editing of Scn5a Rescues Type 3 Long QT Syndrome in Mice. Circulation 2024;149(4):317-329. doi:10.1161/CIRCULATIONAHA.123.065624. PMID: 37965733
11. Kingwell K. Building on the momentum of N-of-1 base editor therapies for rare diseases. Nat Rev Drug Discov 2025;24(8):582-583. doi:10.1038/d41573-025-00119-6. PMID: 40664722
12. Mekonnen AM, Seong K, Kim H, et al. Variant-aware Cas-OFFinder: web-based in silico variant-aware potential off-target site identification for genome editing applications. Nucleic Acids Res 2025;53(W1):W118-W124. doi:10.1093/nar/gkaf389. PMID: 40337925
13. Liu Z, Liu R, Ying Y, et al. Gene Therapy for Cardiovascular and Cerebrovascular Disease: Mechanisms, Translational Barriers, and the Road Ahead. Biomedicines 2026;14(5):1142. doi:10.3390/biomedicines14051142. PMID: 42193469
14. Kaufmann MM, Hackenberg M, Jobson Pargeter W, et al. Systematic Benchmarking of CRISPR-Cas9 Off-Target Prediction Tools Reveals Limitations and Implications for Preclinical Assessment. Hum Gene Ther 2026 (online ahead of print). doi:10.1177/10430342261466407. PMID: 42503868

All fourteen references were resolved through NCBI E-utilities against PubMed, and author lists, titles, journals, years, volumes, pages and DOIs are recorded exactly as retrieved in MS_REFERENCES.csv. None was typed from memory.

Back to all ten papers