# No antisense oligonucleotide has demonstrated target engagement in a human cardiomyocyte, and an unprotected ADAR-recruiting oligonucleotide against SCN5A R104Q would make the current worse

**Ethan Bradley**

Independent researcher, no institutional affiliation

ORCID: [0009-0008-8925-7975](https://orcid.org/0009-0008-8925-7975)

## Abstract

Two independent barriers close the oligonucleotide route to the SCN5A Brugada variant p.Arg104Gln, and
both are negative. The first is delivery: across 1,410 unique records from 26 PubMed queries, no
antisense oligonucleotide has demonstrated target engagement inside a human cardiomyocyte. Four records tagged as human cardiac engagement fail on reading: two are adenoviral
transfer into explanted cells (PMID 11864915, 12479247), one a review reporting animal data
(PMID 33472516), one patisiran acting on liver (PMID 38437698). The strongest conjugate results stop
at mouse and nonhuman primate, and the first human trial of that platform reports target reduction only
from skeletal muscle biopsy (PMID 41707138). There is no cardiac
equivalent of the liver-targeting N-acetylgalactosamine ligand, and the one chemistry with repeated
cardiac engagement, arginine-rich peptide conjugation, carries a dose-limiting renal toxicity in which
arginine content drives both cardiac uptake and nephrotoxicity (PMID 32782413). The
second barrier is bystander editing. Of the editable adenosines an ADAR-recruiting oligonucleotide
would expose, 13 of 15 produce damaging amino acid changes and nine of those products are already
catalogued ClinVar variants of uncertain significance, so an unprotected guide is predicted to make the
current worse: 34.1 percent of wild type falls to 17.6 percent at 50 percent editing. Across 3,517,306 annotated cardiac cells, ADAR1 is detected in 3.5
percent of cardiac muscle cells against 16.2 percent of endothelial cells, while SCN5A runs the
opposite way at 29 to 31 percent against about 2: the enzyme is scarcest in the cells carrying the
target.

---

## A key to the terms used here

- **SCN5A** is the gene for the main sodium channel in heart muscle; **Nav1.5** is the protein it
  makes. **p.Arg104Gln**, or **R104Q**, means arginine at protein position 104 has been replaced by
  glutamine.
- **Messenger RNA** is the working copy the cell makes from a gene and reads to build protein. Editing
  the messenger RNA changes the product without altering the gene itself, so it has to be redone
  continuously, the way a maintenance action with a fixed interval has to be repeated rather than
  signed off once.
- An **antisense oligonucleotide**, or **ASO**, is a short synthetic strand of genetic letters designed
  to stick to one specific messenger RNA by matching its sequence.
- **ADAR** is an enzyme already present in human cells that changes the RNA letter A into I, which the
  cell then reads as G. An **ADAR-recruiting oligonucleotide** is an ASO designed to pair with the
  target and invite ADAR to make one specific letter change. **ADAR1** and **ADAR2** are the two
  relevant forms.
- **Target engagement** means the drug was shown to act on its intended target inside the intended
  cell, not merely to be present in the tissue. This distinction carries most of the weight in this
  paper. Stock on the shelf is not a dose given to a patient.
- A **cardiomyocyte** is a heart muscle cell, the cell that would need to be reached here.
- The **editing window** is the stretch of transcript exposed to the enzyme when the oligonucleotide is
  paired at its target. A **bystander** edit is an unintended change at another letter inside that
  window.
- **GalNAc**, N-acetylgalactosamine, is a sugar tag that liver cells actively take up, which is why
  liver-directed oligonucleotide drugs work. No equivalent tag exists for heart.
- A **conjugate** is an oligonucleotide chemically attached to something that helps it enter cells, such
  as an antibody or a peptide. **Nephrotoxicity** means kidney damage.
- **TPM**, transcripts per million, is a unit of how much a gene is expressed in a tissue.
- **Variant of uncertain significance**, or **VUS**, is a ClinVar classification meaning the evidence
  does not yet establish whether a variant is harmful or harmless: not known to be safe, not known to be
  dangerous.
- **kcal/mol** measures binding strength here; more negative means tighter binding.

## Methods

**Literature search, and its bounds.** I queried PubMed through NCBI E-utilities with 26 distinct
queries combining the molecule class (antisense oligonucleotide, ASO, siRNA, ADAR-recruiting
oligonucleotide, conjugate) with the tissue and evidence terms (cardiomyocyte, cardiac, myocardium,
heart, target engagement, knockdown, biopsy) and with the named delivery platforms. That returned 1,410
unique records. I read every record whose title, abstract or full text claimed cardiac evidence, and
traced each claim to whether the measurement was made in human cardiac tissue, in animal cardiac
tissue, or in another tissue entirely. Positive controls were run on every query before any zero was
believed, because a zero from a malformed query is indistinguishable from a zero in the literature.

This search is bounded and I do not claim it is exhaustive. It covers PubMed-indexed literature and the
trial registry entries I pulled. It does not reach conference abstracts, patent filings, unpublished
industry programmes, or any trial whose cardiac measurements exist but are unreported. A single
published human cardiac target-engagement measurement would overturn the delivery finding, and my claim
is only that I looked in the places named and did not find one.

**Bystander analysis.** The editing window is the stretch of transcript an ADAR-recruiting
oligonucleotide would expose to the enzyme when duplexed at the target site. Within that window I
enumerated every adenosine, translated the consequence of an A-to-I change at each in the reading frame
of RefSeq NM_000335.5, and classified each as silent, tolerated, or damaging by the resulting amino acid
substitution. Relative editing propensity was ranked using the published Eggington sequence-preference
scale, which is ordinal: a rank of 8 of 16 does not mean half the rate, and I did not convert ranks to
rates anywhere.

**Rescue arithmetic.** Predicted current as a function of editing efficiency was computed from the
published co-expression measurements in O'Neill et al. 2022 (PMID 35305865, Supplementary Table 1),
taking the heterozygous R104Q baseline as 34.1 percent of wild type, unaffected as 100, and simple loss
of one allele as 50. The model is linear in functional protein and linear in the dominant-negative
penalty, which is an assumption stated in the limitations rather than a measurement.

**Specificity.** Duplex thermodynamics were computed with ViennaRNA. Transcriptome-wide competitive
binding was assessed both by mismatch-capped scanning, which is bounded by that cap, and by an independent comparison
against 120,000 random transcript windows, which carries no mismatch cap.

Tool versions and exact retrieval dates are not recorded in my working notes for every step, and I
report that gap rather than reconstructing dates after the fact.

## Lead finding, stated first because it is negative

**No antisense oligonucleotide has ever been shown to engage a target inside a human heart
muscle cell.** I searched 1,410 records across 26 separate PubMed queries and read every record
that claimed cardiac evidence. Four came back tagged as human cardiac target engagement. All four
fail on reading:

- PMID 11864915 and PMID 12479247 are adenovirus gene transfer into cardiomyocytes already
  removed from the patient and sitting in a dish. That is not a delivered drug.
- PMID 33472516 is a review whose cardiac dystrophin numbers come from animals.
- PMID 38437698 is patisiran in ATTR cardiac amyloidosis. Patisiran works on **liver** cells to
  lower a protein circulating in the blood that then stops depositing in the heart. The heart
  never sees the drug in a therapeutically meaningful way. Real benefit, wrong organ.

The best real evidence stops in animals. PMID 35944903 achieved dystrophin at 77 percent of
wild-type in **mouse** heart using an antibody-oligonucleotide conjugate that grabs the
transferrin receptor. PMID 37224533 got greater than 75 percent target messenger RNA reduction in
striated muscle including cardiac in **mice and monkeys**. PMID 40207629 showed the same with a
small peptide instead of an antibody, in **nonhuman primates**. These are genuinely strong
results and the platform is real.

Then it reaches humans and the heart drops out of the readout. PMID 41707138 is the first
published human trial of this delivery platform, del-desiran in myotonic dystrophy. It reduced
the target messenger RNA by 46, 44 and 37 percent across three dose groups versus 0.9 percent on
placebo. Every one of those numbers comes from a **skeletal muscle biopsy**. There is no cardiac
tissue measurement in the paper, in the companion molecular-pathology paper (PMID 41821312), or
anywhere in the trial registry entries I pulled.


### There is one genuinely encouraging precedent, and it is not a conjugate

An **unconjugated** antisense oligonucleotide given under the skin worked in the heart of a mouse
carrying a cardiomyopathy variant. PMID 34462437 targeted phospholamban in PLN-R14del mice:
protein aggregates prevented, cardiac dysfunction prevented, survival up three-fold, plus a
reversal in a second unrelated model and improved left-ventricular contractility in rats after
myocardial infarction. PMID 35269571 halted advanced disease. PMID 40905134 confirmed cardiac
target engagement by dose-response. This is the closest structural analogue to the R104Q problem
that exists: a single-nucleotide cardiomyopathy variant, an oligonucleotide, the heart, and it
worked without any targeting ligand.

That programme has now reached humans. AZD4063 (NCT07241104, AstraZeneca, Phase 1, recruiting) is
a first-in-human subcutaneous agent in PLN-R14del dilated cardiomyopathy. A second, ATR 1072
(NCT07738107, Atrium Therapeutics, Phase 1/2, not yet recruiting), delivers a small interfering
RNA against PRKAG2 for a genetic cardiomyopathy by intravenous infusion every six weeks. Neither
has posted cardiac pharmacodynamic results. **The question of whether an oligonucleotide can work
inside a human cardiomyocyte is being actively answered right now, and the answer is not in yet.**

### The hardest number in the delivery literature

PMID 24549299 dosed mice for eight weeks and then measured drug, splicing and protein in each
tissue. In heart, the oligonucleotide half-life was about 65 days, longer than in skeletal
muscle, liver or kidney at about 35 days. And in the same heart tissue: **high oligonucleotide
levels but low splicing correction and low protein**. The drug got in, persisted longer than
anywhere else, and did not work.

That is the single most dangerous fact for this route, and it is worse than a delivery failure.
Drug measured in cardiac tissue is not evidence of drug doing anything in cardiac tissue.
Whatever fraction sits in the wrong compartment inside the cell, or bound to the wrong thing, is
invisible to a tissue-concentration assay. PMID 20407428 independently found heart had the lowest
oligonucleotide levels of any muscle group examined, with the longest half-life, about 46 days.
The heart is both the hardest striated muscle to reach and the one that most misleads you about
whether you reached it.

### For the exact molecule class on this route, delivery is liver-only

The route needs an ADAR-recruiting oligonucleotide. Of every record I classified, **not one
demonstrated ADAR-recruiting oligonucleotide editing in heart tissue of any species**. The single
nonhuman-primate demonstration of this exact chemistry (PMID 35256816) reached up to 50 percent
editing with no detectable bystander editing of the control transcript. In **liver**, using an
N-acetylgalactosamine sugar that a liver-specific receptor grabs.

Every clinical-stage programme of this class is liver-directed and treats the same liver disease:
WVE-006 (NCT06186492 completed, NCT06405633 active), KRRO-110 (NCT06677307, terminated), AIR-001
(NCT07431112, recruiting), all in alpha-1 antitrypsin deficiency. The reason is not coincidence.
GalNAc works because hepatocytes display a receptor that vacuums up anything carrying that sugar.
**There is no cardiomyocyte equivalent of GalNAc.** The closest thing is the transferrin receptor,
which is not cardiac-specific, and its human data stop at skeletal muscle.

### The one demonstrated cardiac oligonucleotide route carries a dose-limiting toxicity, and it is driven by the same property that gets the drug into heart

Peptide conjugation is the only chemistry in this literature with repeated demonstrated target
engagement in cardiac tissue. Peptide-conjugated phosphorodiamidate morpholino oligomers restored
dystrophin in mouse heart with functional improvement (PMID 18784278), sustained it over months
(PMID 18545222), and prevented cardiomyopathy durably (PMID 19815563). Those are the strongest cardiac
delivery results in the whole set, and every one of them is in rodent.

The problem is what makes them work. In the delivery review by Roberts, Langer and Wood (*Nat Rev
Drug Discov* 2020, PMID 32782413), the cell-penetrating peptides used for these conjugates are
arginine-rich, and **arginine content correlates with both cardiac uptake and nephrotoxicity**. The
property being optimised for delivery to heart is the property that damages kidney, so the two cannot
be separated by tuning the same variable: raising arginine content to reach more cardiomyocytes raises
renal exposure in step. That is a dose-limiting toxicity rather than a formulation problem, and it is
the reason this route has not advanced to human cardiac dosing despite the strength of the animal
data. Any design that reaches for peptide conjugation to solve the delivery barrier inherits it.

### The enzyme this route depends on is scarce in exactly the cells that carry the target

An ADAR-recruiting oligonucleotide does no work by itself. It has to recruit an ADAR enzyme already
present in the cell, so the route needs the enzyme and the target transcript in the same cell. In
heart they are largely in different cells.

Across **3,517,306 annotated cardiac cells** in 11 cell types, ADAR1 is detected in **3.5 percent of
cardiac muscle cells** against **16.2 percent of endothelial cells**, while SCN5A runs the opposite
way: **29 to 31 percent of cardiac myocytes** against about **2 percent of endothelial cells**. The
cells that carry the transcript needing repair are the cells with least of the enzyme that would
repair it, and the cells with the most enzyme barely express the target at all. Expressed as a ratio
of detection rates, ADAR1 to SCN5A is 0.11 in cardiac muscle cells and 9.35 in endothelial cells, an
83-fold difference in the wrong direction.

This is a sharper statement of the same problem than the bulk-tissue measurement gives. Bulk left
ventricle reports ADAR at 21.48 TPM, rank 53 of 54 tissues, which already says heart is enzyme-poor.
The single-cell reading says the shortfall is concentrated precisely in the cell type that matters,
and bulk tissue understates it because endothelial and other non-myocyte cells contribute most of the
ADAR signal a bulk assay measures. A route whose efficiency scales with local enzyme availability is
therefore worse off than the bulk number suggests.

## The bystander problem: I can solve it, and solving it breaks something else

### What is exposed

Sequence contexts are ranked on the Eggington 2011 scale (PMID 21587236), which quantified editing
at 406 sites in one long double-stranded RNA and ranks all 16 three-letter contexts worst to best;
the target's own CAG context is rank 8 of 16, the exact median.

Within 25 nucleotides of the target there are **15 editable letters, and 13 of them change the
protein if edited**. Widening to 50 nucleotides: 27 letters, 23 damaging. Twelve of these sit in
sequence contexts the enzyme likes **more** than the target's own. The worst is 5 letters away,
in an identical CAG context, meaning the enzyme finds it exactly as attractive as the intended
site; editing it gives S106G.

Two findings here appear not to have been drawn together before.

**Nine of the bystander products are already catalogued human variants.** These are not
hypothetical proteins. S106G is ClinVar VCV000850004, T101A is VCV001798889, I102V is
VCV000967224, Y112C is VCV001332491, and five more. All are currently classified uncertain
significance, so none is known to be harmful, but none is known to be safe either. An editing
drug that produced one of these would be manufacturing a variant of uncertain significance on
purpose, in a heart, in a person who already has an arrhythmia syndrome.

**D84G is an editable letter.** It sits 60 nucleotides upstream, outside every footprint I
considered, so it is not a design problem today. It matters because D84 is the carboxylate that
the rescue hypothesis under test here is built around gripping. If a future design reached that
far, editing there would destroy the target of the parallel chemistry route.

### The protection scheme works, on paper, and the paper is weaker support than it looks

Schneider 2014 (PMID 24744243) is the sole published basis for the protection scheme. It is usually
cited as showing that a G placed opposite an unwanted A blocks editing there while permitting
near-quantitative editing 5 nucleotides away. I read the full text (open access,
doi 10.1093/nar/gku272) to check the geometry, and found three things, one better than expected
and two worse.

**Better than expected: the blocking distance.** The paper protected the GFP Tyr66 codon while
editing the adjacent Ser67 codon, and states that the strategy succeeded even when the targeted
base and the off-site base were separated by only **one intervening nucleotide**. My footprints
contain no damaging bystander closer than 5 nucleotides to the target, so every site I need to
block sits comfortably inside the distance range where blocking has actually been demonstrated.
That part of the design is on firmer ground than the 5-nucleotide figure implies.

**Worse, first: the enzyme was not endogenous ADAR.** Schneider used SNAP-ADAR1 and SNAP-ADAR2,
engineered fusions in which the enzyme's own double-stranded-RNA-binding domains were **replaced**
by a SNAP-tag, and the guide RNA was **covalently attached** to the enzyme through a
benzylguanine linker. That is a one-to-one enzyme-guide complex delivered as a unit. This route
depends on the opposite arrangement: a bare chemically modified oligonucleotide that must recruit
whatever endogenous ADAR happens to diffuse past, in a heart where ADAR1 sits at 21.48 TPM,
rank 53 of 54 tissues. The word "recruit" does not appear anywhere in Schneider 2014. Every
protection number in my tables inherits an untested assumption: that a blocking G behaves the
same way when the enzyme is not tethered to the guide.

**Worse, second: they protected one site. I need seven to ten.** Schneider blocked a single
off-site adenosine. My fully protected 30-mer requires seven simultaneous G mismatches and the
40-mer requires ten. The paper contains exactly one observation on multiple simultaneous
mismatches, and it is discouraging: with two mismatches present, an A/C at the target plus an
A/G protection, **SNAP-ADAR2 did not accept the doubly mismatched substrate well** and target
editing yield stayed low, for both A/C+A/C and A/C+A/G combinations. SNAP-ADAR1 handled it and
gave efficient, selective editing.

There is one genuinely favourable inference in that. The enzyme that tolerated the double mismatch
was ADAR1, and heart is overwhelmingly ADAR1-dominant: 21.48 TPM against 3.18 TPM for ADAR2. If
mismatch tolerance is an intrinsic property of the two enzymes, the heart has the more forgiving
one. But this is an inference from one experiment with an engineered fusion, extrapolated from two
mismatches to seven, in the tissue with almost the least ADAR1 in the body. It is a reason to test
the idea, not a reason to believe it.

### Why that is not a solution

Each protective G is a deliberate mismatch, and mismatches cost binding energy. For a 30-mer,
blocking all seven damaging bystanders moves the binding energy from -56.0 to -32.4 kcal/mol, a
penalty of 23.6. For a 40-mer, ten blocks cost 29.0.

I then tested what that weakening does to specificity, and this is the result that decides the
task. I scanned all **669,547 Ensembl 116 transcripts, 1,478,394,455 nucleotides**, every window
position, and separately compared each design against 120,000 randomly drawn transcriptome
windows:

| design | protective blocks | grip on target | random windows binding at least as tightly |
|---|---|---|---|
| 20-mer, unblocked | 0 | -32.1 | 0 of 120,000 |
| 30-mer, unblocked | 0 | -56.0 | 0 of 120,000 |
| 30-mer, worst site only | 2 | -47.1 | 0 of 120,000 |
| 20-mer, fully blocked | 3 | -22.9 | 0.04 percent |
| 25-mer, fully blocked | 5 | -27.6 | 0.34 percent |
| 40-mer, fully blocked | 10 | -47.7 | 0.37 percent |
| 30-mer, fully blocked | 7 | -32.4 | **0.87 percent** |

A fully blocked 30-mer binds roughly **1 in 115 random transcriptome positions** at least as
tightly as it binds its own intended target. Scaled across the transcriptome that is on the order
of ten million competitive sites. The oligo stops being a targeted molecule.

**The tension, stated plainly.** A longer oligo does give better transcriptome
specificity when unblocked, and that is real: the bare 30-mer had zero competitive random
windows. But a longer oligo also exposes more damaging local letters, 3 at 20 nucleotides rising
to 13 at 50. Blocking those letters is the only way to make the local problem safe, and blocking
is precisely what destroys the specificity that made the longer oligo attractive. The two
constraints are not independent; they are coupled through binding energy, and the coupling runs
the wrong way.

**No length optimises both. I tested 33 length-and-protection combinations and zero passed both
measured criteria.** Every unblocked design leaves damaging bystanders exposed. Every fully
blocked design binds competitively across the transcriptome. The intermediate option, blocking
only the worst sites, gives a 30-mer with clean transcriptome behaviour that still leaves five
damaging letters unprotected, including T101A and N109D. That is the closest thing to a viable
design and it is not viable.

### An off-target in the brain sodium channel, which an earlier filter had hidden

The bare 20-mer has an off-target in **SCN1A**, the brain sodium channel and the Dravet syndrome
gene, at the same single-mismatch count as its own target and binding **6.5 kcal/mol more
tightly** (-38.6 versus -32.1). The window is nearly identical because the two channels are
paralogues; the protein sequences read ...NKGKAIFRFSATSA... in SCN1A against ...NKGKTIFRFSA... in
SCN5A. Two editable letters sit in that window, which would give I99V and S103G in SCN1A.

An earlier version of this transcriptome scan did not report SCN1A at all. I checked why: that
scan retained only hits carrying an editable A at the edit-site position, and every one of its
1,966 rows has that flag set. The SCN1A window has a G there, matching the oligo's C as an
ordinary pair, so it was filtered out before counting. This is the same class of error as the
seed-and-extend failure already documented on this route: a filter that looks reasonable removes
exactly the hits that matter. The 30-mer does not have this problem.

## Rescue arithmetic, redone for RNA editing

DNA editing fixes a cell once. RNA editing fixes a fraction of messages continuously and must be
re-dosed, so what matters is the steady-state fraction repaired.

Anchors, from O'Neill 2022 (PMID 35305865, Supplementary Table 1, open preprint
10.1101/2021.09.22.461398 p.29, Sleeping Beauty genomic integration so wild-type is not diluted
and the no-effect baseline is 100): heterozygous R104Q at 68.3 plus or minus 6.1 percent of a
single wild-type allele, n=34. Rescaled to a heart where two working alleles equal 100:
unaffected 100, simple loss of one allele 50, measured R104Q 34.1.

With every bystander perfectly blocked:

| target repaired | dominant-negative | simple loss of one allele |
|---|---|---|
| 0 percent | 34.1 | 50.0 |
| 25 percent | 50.6 | 62.5 |
| 50 percent | 67.1 | 75.0 |
| 75 percent | 83.5 | 87.5 |
| 100 percent | 100.0 | 100.0 |

**24.1 percent** of mutant messages must be repaired to reach 50, what simple loss of one allele
would give. **51.9 percent** to reach 68.3. Under the loss-of-one-allele mechanism the baseline is
already 50, so any editing is gain, and 36.6 percent reaches 68.3.

Now the coupling that DNA editing does not have. The worst bystander shares the target's exact
sequence context, so the honest assumption is that it is edited at the same rate as the target.
Under that assumption:

| target repaired | unprotected, bystander damaging | unprotected, bystander merely non-functional |
|---|---|---|
| 10 percent | 33.4 | 35.2 |
| 25 percent | 30.0 | 35.0 |
| 50 percent | 17.6 | 29.6 |
| 90 percent | -19.3 | 9.5 |

**Every unprotected curve peaks at or below the untreated baseline of 34.1 and then falls.** The
best the damaging-bystander case ever achieves is 34.1, at zero editing. Even assuming the
bystander product is merely non-functional rather than actively poisonous, the curve tops out at
35.4 and never approaches 50. Under simple loss of one allele the unprotected curve starts at 50
and only declines. **An unprotected ADAR oligo at this site is predicted to make the sodium
current worse under all three mechanistic assumptions.** The harder it works, the worse the
outcome.

This inverts how bystander editing is usually discussed. It is normally a tolerability footnote.
Here it is the sign of the treatment effect.

How much bystander editing can be tolerated: at 50 percent on-target editing, no more than
**17.3 percent** of transcripts may carry a damaging bystander edit and still hold 50 percent
current. At 75 percent on-target, **13.2 percent** to hold 68.3. These are demanding numbers for
a site whose worst neighbour is 5 nucleotides away in an identical context.

## Does the route survive

**Conditionally, and not on the strength of anything I found.** Both barriers came back negative in
their strong form, and the piece usually treated as already solved is not: the protection scheme
rests on a single paper using a tethered engineered enzyme rather than the endogenous recruitment
this route requires.

The bystander constraint is solvable in principle by Schneider blocking, and every dangerous site
is inside the distance range where blocking has been demonstrated, but only ever for one site at a
time and with a tethered enzyme. No length and protection combination I tested, 33 of them, passes
both the local and the transcriptome-wide test, because the blocks that fix the local problem
destroy the global one. The specific claim that a longer oligo buys specificity
is true only for unblocked oligos, and unblocked oligos are predicted to be net harmful.

Delivery has never been demonstrated for this molecule class in a heart of any species, and has
never been demonstrated for any oligonucleotide inside a human cardiomyocyte.

## The single most likely thing to kill this route

**Delivery, and specifically the finding in PMID 24549299 that cardiac drug levels do not predict
cardiac drug activity.** Not the bystanders, because those are a design problem with a known
mechanism of attack even if no current design threads it. Not the enzyme shortage, though that is
real and already documented on this route. Delivery, because it is the one failure mode where the
usual evidence would not tell you that you had failed. A trial could measure oligonucleotide in a
cardiac biopsy, see plenty of it, and still be looking at a drug doing nothing, exactly as
happened in mouse heart with an eight-week dosing schedule.

If I had to name the experiment that decides this route, it is not an editing experiment. It is:
does AZD4063 or ATR 1072 produce measurable target engagement in human cardiac tissue. That
result is roughly one to three years out and is being generated by somebody else.

## Limitations, and what would falsify this

1. **The entire protection scheme rests on one paper that used a different enzyme system.**
   Schneider 2014 used SNAP-ADAR fusions with the guide RNA covalently tethered to the enzyme and
   the enzyme's own RNA-binding domains removed. This route needs endogenous ADAR recruited by an
   untethered oligonucleotide. It protected one site; my designs need seven to ten. Its only
   multiple-mismatch observation showed ADAR2 rejecting a doubly mismatched substrate, though
   ADAR1, the enzyme that dominates heart, tolerated it. **Falsified by:** a G-protected
   multi-mismatch oligo tested against endogenous ADAR1. This is the single most informative bench
   experiment available on this route and is far cheaper than anything on the delivery side.
2. **The Eggington ranking is an ordinal scale, not a rate.** Rank 8 of 16 does not mean half
   speed. I used it to rank relative risk, which is what it supports, and I did not convert ranks
   to rates anywhere in the rescue model. Where I needed a bystander rate I assumed equality with
   the target for the site that shares the target's exact context, which is a mechanistic argument
   rather than a measurement. **Falsified by:** measured editing rates at positions 520 and 525 in
   the same reporter.
3. **The rescue model is linear in functional protein and linear in the dominant-negative
   penalty.** The 15.9-point gap between simple loss of one allele at 50 and the measured 34.1 is
   attributed entirely to dominant-negative interference and assumed proportional. If interference
   saturates, unprotected editing is less catastrophic than I show. **Falsified by:** an allelic
   titration series measuring current against mutant fraction.
4. **Whether R104Q is dominant-negative in a human heart is unresolved.** Settling it requires a
   measurement I cannot make myself, in a cardiomyocyte model rather than a heterologous expression
   system. I report both mechanisms throughout, simple loss of one allele and dominant-negative
   interference, and neither changes the sign of the unprotected result.
5. **My transcriptome scan capped reported hits at 8 mismatches.** For designs whose intended
   duplex carries more than 8 mismatches, the 40-mer with 11, that cap sits below the intended
   duplex, so that hit count is a lower bound and I did not use it as a pass. This is
   why I ran the independent 120,000-random-window comparison, which has no mismatch cap and
   showed the 40-mer binding 0.37 percent of random windows competitively.
6. **The specificity test is thermodynamic, not cellular.** ViennaRNA duplex energy ignores
   target structure, RNA-binding proteins, and whether ADAR is even present at a given transcript.
   A site that binds tightly is not automatically edited. **Falsified by:** transcriptome-wide
   editing measurement after transfection of a blocked oligo.
7. **Chemical modification is out of scope here.** It changes binding energies and could shift the
   trade-off numbers in either direction. My protection penalties are computed on unmodified RNA, so
   a modified-backbone design might recover some of the binding energy that protection costs.
8. **The peptide-conjugate toxicity is a reported correlation, not my measurement.** Roberts, Langer
   and Wood (PMID 32782413) state that arginine content tracks both cardiac uptake and
   nephrotoxicity; I did not measure either, and I did not find a study that separates the two
   properties. **Falsified by:** a peptide conjugate achieving cardiac target engagement at an
   arginine content low enough to avoid renal toxicity, which would mean the two properties are
   separable after all and this barrier is a formulation problem rather than a dose-limiting one.
9. **The single-cell enzyme measurement is detection rate, not molecule count.** Percentage of cells
   with at least one read is not concentration, and droplet single-cell data undercounts low-expressed
   transcripts, so the absolute ADAR1 figures are lower bounds and the myocyte-versus-endothelial
   contrast could be inflated by differences in capture efficiency between cell types. The comparison
   holds only if SCN5A and ADAR1 are undercounted similarly, which I did not verify. **Falsified by:**
   a protein-level or targeted-RNA measurement of ADAR1 in purified human cardiomyocytes showing enzyme
   abundance sufficient for the editing rates this route needs.
10. **My literature search is bounded, not exhaustive.** 26 PubMed queries, 1,410 unique records, every one classified and every human cardiac claim read individually, plus 6 registry queries.
   I could not reach conference abstracts, company pipeline disclosures, or the two 2026 records
   without abstracts (PMID 41707145 is an editorial). A cardiac pharmacodynamic result presented
   at a meeting and not indexed would not appear. **Falsified by:** any published measurement of
   oligonucleotide target engagement in human cardiac tissue.
11. **One registry field name failure worth recording.** My first ClinicalTrials.gov queries
   returned HTTP 400 because `Phases` and `Conditions` are not valid field names, `Phase` and
   `Condition` are. A 400 is loud, but the same class of error returning an empty list instead
   would have looked like absence of trials. I bisected every field before trusting any count, and
   ran positive controls on every PubMed query before believing any zero.

## What is settled and what is not

Settled: the bystander enumeration is now complete and traceable to the reference sequence, with
protein consequences and ClinVar status for every site. The protection scheme is designed and
costed. The trade-off question has a definite answer, which is that no length optimises both. The
rescue arithmetic now includes the bystander coupling term, which changes the sign of the result.
The delivery question has a definite answer for this molecule class.

One correction to the received position rather than a new finding: the protection scheme is widely
treated as established, and it is not. It is one paper, one protected site, a tethered engineered
enzyme, and a single discouraging observation about simultaneous mismatches.

Not settled: whether the dominant-negative mechanism operates in a human heart. Whether any
oligonucleotide engages a target in a human cardiomyocyte. Whether a chemically modified oligo can
recover the binding energy that protection costs, which is the one route by which the trade-off
verdict here could be overturned.


## Where the single-cell and expression numbers come from

The cardiac single-cell figures are detection rates across 3,517,306 annotated cells in 11 cardiac
cell types, from the cardiac cell atlas data used throughout this project, tabulated in
`ADAR_CARDIOMYOCYTE_EXPRESSION.csv`: for each cell type, the percentage of cells with at least one
read for ADAR1 and for SCN5A, alongside cell counts. Bulk tissue values are GTEx v8 medians, ADAR at
21.48 TPM in left ventricle and 3.18 TPM for ADAR2, tabulated in `gtex_adar_expression.csv` across
all 54 tissues so the rank can be checked rather than taken on trust.

## Data availability

All primary evidence is public. The literature findings rest on the PubMed records cited by PMID
throughout, retrieved through NCBI E-utilities; the search was 26 queries returning 1,410 unique
records, and the queries and dates are as recorded in the methods above. Trial registry entries were
read at ClinicalTrials.gov. The rescue arithmetic uses published co-expression current measurements
from O'Neill et al., Genet Med 2022, PMID 35305865, Supplementary Table 1. Sequence context for the
bystander analysis is RefSeq NM_000335.5. 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.
They comprise the per-adenosine bystander consequence calls, the delivery-evidence classification of
every record read, the cardiac single-cell expression table, and the editing-versus-current curves.

## Competing interests

I am a heterozygous carrier of SCN5A p.Arg104Gln, the variant analysed here, and I have a clinical
diagnosis of Brugada syndrome. No funding was received.

Nothing in this paper is clinical guidance or a treatment recommendation for any person, including me.
It reports why a therapeutic route does not currently work.

## References

Cited by PubMed identifier throughout. Every record below was retrieved from PubMed and read; the
relevant finding from each is stated at the point of citation in the text rather than summarised here.

1. PMID 11864915
2. PMID 12479247
3. PMID 33472516
4. PMID 38437698
5. PMID 35944903
6. PMID 37224533
7. PMID 41707138
8. PMID 35305865
9. PMID 40207629
10. PMID 41821312
11. PMID 34462437
12. PMID 35269571
13. PMID 40905134
14. PMID 24549299
15. PMID 20407428
16. PMID 35256816
17. PMID 21587236
18. PMID 24744243
19. PMID 41707145
20. PMID 32782413 (Roberts TC, Langer R, Wood MJA. Advances in oligonucleotide drug delivery. *Nat Rev Drug Discov* 2020)
21. PMID 18784278 (peptide-conjugated morpholino, cardiac dystrophin restoration in mdx mice)
22. PMID 18545222 (sustained dystrophin expression from peptide-conjugated morpholino oligomers)
23. PMID 19815563 (long-term improvement in mdx cardiomyopathy after peptide-conjugated morpholino therapy)