---

Version of record: 10.5281/zenodo.21799866, published 5 August 2026. That identifier is the citable address for this paper and it resolves at https://doi.org/10.5281/zenodo.21799866. 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_8_MECHANISM_ELIMINATION.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 Earlier on 6 August 2026 it was given a banner reading "superseded revision, do not cite any figure in it".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.

This copy is ahead of the version of record, and the addition is dated. On 6 August 2026, after the deposit, Tano et al. 2026 (PMID 41582807) was folded in at six places: a requirement sharpened in the Part 4 discussion, a bullet in the evidence list, a new subsection on what Tano adds and the one thing it takes away, a trafficking paragraph that constrains the class rather than this variant, a limitation, and reference 9. The record at the identifier above contains none of it. No conclusion changes.

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.

---

Neither local strain nor exposed hydrophobic surface explains pathogenicity at SCN5A Arg104, and three further candidate mechanisms are ruled out

Ethan Bradley

Independent researcher, no institutional affiliation

ORCID: 0009-0008-8925-7975

Abstract

Five candidate structural mechanisms for loss of function at SCN5A Arg104 were tested and four are refuted, in each case by a control that a plausible-looking positive result failed. Local energy redistribution on substitution is real but non-specific: it tracks side-chain volume (Spearman +0.77) and produces its largest effect in the study on a benign control, V125L. Exposed apolar surface separates R104W from R104Q decisively (p = 1.7e-7), which I first read as support for a sticky-surface mechanism, until a cross-position control substituting tryptophan at seven N-terminal-domain positions put the benign control R34C first, gaining 2.3 times more apolar area than position 104, with the dominant-negative and benign split running the wrong way (Mann-Whitney p = 1.000). Two protocols reached that refutation and disagreed on absolute magnitude while agreeing on the ordering. A linear degradation or endoplasmic-reticulum retention motif is ruled out at Arg104 by sequence scan, while the pathogenic control R121W does lie in one. The N-terminal domain, residues 1 to 130 of NP_932173.1, contains no N-X-S/T glycosylation sequon, so no glycosylation-mediated mechanism is available to it. What the structure does establish is that Arg104 anchors an invariant buried acidic pocket whose salt bridge to Asp84 is conserved across all nine human Nav paralogues. The observable that separates R104W from benign substitutions has not been identified, and the remaining candidates are ones a single-chain calculation cannot see.

A key to the terms used here

• SCN5A is the gene for the heart's main sodium channel; Nav1.5 is the protein. Arg104, or R104, is the arginine at protein position 104. R104Q replaces it with glutamine, R104W with tryptophan. R34C and V125L are used as benign controls, variants not expected to cause disease.
• The N-terminal domain is the first stretch of the protein, residues 1 to 130 here.
• A mechanism in this paper means a specific physical story for how a variant breaks the channel. Each is tested separately and most are eliminated.
• Local strain is stress introduced into the surrounding structure when one amino acid is swapped for another that does not fit as well.
• Hydrophobic means water-repelling. Proteins normally bury their hydrophobic parts inside. Exposed apolar surface is water-repelling surface left on the outside, which can make a protein sticky and get it flagged for disposal. SASA is the measured amount of that surface, in square ångströms.
• An ångström, Å, is a ten-billionth of a metre. A REU is Rosetta's internal energy unit.
• Repacking means letting the computer rearrange nearby side chains after a substitution; minimisation then relaxes the structure to relieve leftover stress. A rotamer is one of the positions a side chain can adopt.
• N-glycosylation is the cell's attachment of a sugar chain to a protein at a specific three-letter pattern called a sequon (N-X-S/T). No sequon means no attachment site, so that mechanism is unavailable.
• A degron is a short sequence that marks a protein for destruction; a retention motif holds it in the endoplasmic reticulum, the cell's assembly and inspection area. A protein failing inspection is destroyed rather than shipped, which is why these motifs matter here.
• A salt bridge is an attraction between a positively and a negatively charged side chain, such as Arg104 to Asp84 (aspartate at position 84). A hydrogen-bond donor is an atom offering a hydrogen to such a contact.
• Paralogues are related genes in the same organism arising from duplication; the nine human Nav alpha subunits are paralogues of each other. Conservation is how unchanged a position stays across them.
• Cryo-EM yields a 3D protein model; 8VYJ is the one used here, at 3.6 Å resolution, fine enough to place the backbone but not to fix every side chain.
• Cliff's δ measures how completely two groups separate. A confidence interval that excludes zero means the effect is distinguishable from no effect. Spearman ρ measures rank agreement.
• A control is a case whose answer is known in advance. The controls do the real work in this paper: each mechanism that looked supported was eliminated by a variant known to be harmless scoring as high.

Why eliminate mechanisms at all

A folding-stability calculation says nothing about SCN5A Arg104: a companion analysis found that folding free energy misses three of four known loss-of-function variants in this domain. The natural next move is to ask what observable does separate them. This paper reports that search, and it is mostly a record of hypotheses that did not survive their controls. I report it because the controls are the useful part. Each refuted hypothesis here produced an encouraging first result, and each was killed by asking whether a variant known to be benign scored as high.

Part 1. Local strain and exposed hydrophobic surface

1. What was computed

Nine models per frame on all 24 relaxed ensemble members (12 Rosetta density-guided, 12 restrained-OpenMM) × 3 independent repacker seeds = 1008 models, 0 failures. Position 104 substitutions Q, W, K plus A/F/Y as a volume-and-aromaticity series; benign controls R34C and V125L at their own positions.

Each model was scored in two tiers:

The control that makes this interpretable. A wild-type pose keeps its input rotamer through IncludeCurrent; a mutant cannot. Comparing mutant-to-WT therefore charges every mutant for the repacker's inability to rediscover the input conformation. So every variant is compared against a self-mutation control, MutateResidue(104, "ARG") on the wild-type pose, then the identical repack/minimise protocol. Arg→Arg is chemically null; everything it registers is method noise. All effect sizes below are frame- and seed-matched differences against that control, with 95% paired-bootstrap CIs (20 000 resamples) and Cliff's δ.

Metrics: per-residue ref2015 energy for all 8 Å-shell residues (target excluded, its own energy changes trivially with residue type); Rosetta HBondSet donor counts and an independent geometric donor count at the Asp84/Asp82 carboxylates; BuriedUnsatHbondFilter (vsasa burial); FreeSASA side-chain and apolar SASA; compute_residue_packing_scores (packstat) for cavity formation; Cα displacement over the 82-92 loop and 100-110 strand; DSSP secondary structure over B78-115.

2. Result 1: local energy redistribution is real, and it is not specific

Excess local |ΔE| over the 8 Å shell, relative to the matched self-mutation control (OpenMM prep, n = 36 models each):

R104W does rank first. But V125L, a benign control with no ClinVar entry, correctly predicted neutral by the global method, ranks fourth of eight and has the single largest effect size in the table (δ = +0.79, and δ = +1.00 in the Rosetta prep). A metric that fires this hard on a benign control cannot be the missing discriminator.

Hypothesis A's exact predicted signature appears in the wrong variant. The prediction was "significant local redistribution with near-zero total energy." V125L delivers precisely that: local |ΔE| = +3.99 REU (CI excludes zero) while total energy change is +0.27 REU (CI -0.39 to +1.09, includes zero). R104W's total is also indistinguishable from control (+1.75, CI -0.47 to +4.06), but so is a benign variant's. The signature is not diagnostic.

Ranking test, explicitly. Local disruption at position 104 anti-correlates with global ΔΔG (Spearman ρ = -0.77, n = 6, p = 0.072) and correlates with side-chain volume at exactly the same magnitude (ρ = +0.77). The local ranking is W > Y > F > A > Q > K. Clinically, R104W is Pathogenic/Likely pathogenic and R104Q is Conflicting; R104Y and R104F have no ClinVar entries and no functional data, so the local ranking places two uncharacterised variants above the one variant of clinical interest. Volume, not pathogenicity, is what this axis measures.

3. Result 2: the fixed-backbone strain signal is repacker noise

Tier 1 (no relaxation) appears to show a large R104W effect: local |ΔE| = 26.4 REU, total ΔE = +38.0 REU, roughly 3× any other variant. That signal has a standard deviation equal to its own mean (SD 26.1 REU across 36 models; coefficient of variation 0.99, the highest of any variant, control included). Cartesian relaxation removes it: R104W's local |ΔE| falls from 26.4 to 12.4 REU and its total ΔE from +38.0 to +2.1 REU.

This is a known confound of fixed-backbone repacking, and it is quantitatively confirmed: the apparent fixed-backbone strain of the bulkiest substitution is dominated by the repacker's failure to place a large side chain, not by physical strain. Any local-strain claim built on fixed-backbone repacking alone would have been an error of method, not a finding. Reported here rather than debugged away.

The self-mutation controls put a floor on everything: chemically-null Arg→Arg registers local |ΔE| = 6.37 ± 4.61 REU (OpenMM) and 0.72 ± 1.34 REU (Rosetta). The Rosetta ensemble's ~9× smaller noise floor is the conformational degeneracy already documented for that prep (SD 0.005 Å in side-chain distances); it makes the Rosetta prep look more precise without making it more informative, which is why OpenMM is the primary prep here.

4. Result 3: the acidic pocket is dismantled by every substitution

H-bond donors within 3.5 Å of the Asp84 carboxylate (Rosetta HBondSet, tier 2):

Every substitution except lysine drops Asp84 to essentially zero donors (all CIs exclude zero, δ ≈ -0.76 to -1.00, both preps). R104W and R104Q are indistinguishable on this metric (paired W-Q difference +0.03 donors, p = 0.66). Only R104K retains partial coordination, 0.78-0.97 donors, which is consistent with the prior finding that lysine cannot reach Asp84 but can still contribute at the pocket periphery.

This reproduces the project's prior numbers (WT ~2.5 → R104Q ~1.0 by the older geometric count; the absolute values differ with detection method, the direction and magnitude do not) and confirms the microenvironment note's framing. But it also settles a question: loss of Asp84 coordination cannot explain the R104W/R104Q clinical difference, because both lose it identically.

Buried-unsatisfied-polar counts inverted from expectation, and the reason is instructive: removing arginine exposes Asp84 to solvent rather than leaving it buried and unsatisfied. Shell-summed unsatisfied polars go down in every mutant (Δ = -0.3 to -2.6). The "orphaned buried charge" picture overstates the defect, as the microenvironment note already concluded from a different direction, because the carboxylate gains solvent access when its partner leaves.

5. Result 4: the backbone barely moves, and secondary structure never changes

R104W's loop displacement exceeds control (+0.05 Å, CI +0.01 to +0.10) and exceeds R104Q (+0.03 Å, p = 0.019). These are hundredths of an ångström, far below the ~3.6 Å resolution of the parent structure and below any threshold at which a backbone claim is meaningful. Statistical significance here reflects 36 matched replicates, not physical magnitude.

DSSP over B78-115 is identical to the matched wild-type in 100% of Rosetta models (all 396) and 83-94% of OpenMM models. The residual OpenMM variation is method noise, not mutation effect: the chemically-null Arg→Arg control scores 0.861, below R104W's 0.944, and the lowest score of any variant belongs to the benign control V125L (0.833). The 101-106 β-strand and the 82-89 loop survive every substitution intact. No variant unzips the sheet. This is the expected result given the prior finding that R104 occupies the non-H-bonded cross-strand position, so its side chain is not load-bearing for sheet register, and it independently confirms that framing.

Packstat detects no cavity: shell-averaged change is ≤0.05 for every variant in both preps, including the R104A "cavity-creating" case.

6. Result 5: what R104W actually does differently

One metric separates R104W from R104Q decisively, and it is not strain. Exposed apolar side-chain surface, normalised to each residue's Gly-X-Gly extended-reference maximum (removing the trivial size dependence):

R104W and R104Q move in opposite directions, and this is the largest, most reproducible separation in the whole analysis: paired W-Q difference = +0.150 relative units, p = 1.7 × 10⁻⁷, consistent in both preps. R104W adds ~28 Ų of exposed apolar surface at a position that is 11.6% solvent-accessible in the assembled channel; R104Q removes ~13 Ų. Both benign controls move slightly negative (V125L -1.1 Ų) or, in R34C's case, positive at an already solvent-exposed position, R34 measures 52.4% RSA in this carve-out (V125 is 5.3%, R104 15.1%), so R34C adds apolar area to an already-wet surface rather than creating a buried hydrophobic patch. R34 sits three residues from a carve-out chain terminus, so its exposure is an upper bound.

This looked like hypothesis B, and I initially reported it as support for it. That was wrong, and the control that shows why is in §6a. The observation itself stands: tryptophan at 104 presents an apolar face where arginine presented a polar one, without destabilising the fold. What it does not do is explain pathogenicity.

6a. The control that refutes hypothesis B: a cross-position test

The §6 comparison holds position fixed and varies the substitution. That design cannot detect a metric which is simply large wherever a big hydrophobic residue is introduced. The missing control is the transpose: substitute tryptophan at seven different NTD positions and ask whether apolar gain tracks phenotype. Same tier-2 protocol, same self-mutation controls, 24 OpenMM models per position:

The benign control ranks first, gaining 2.3× more apolar area than position 104, and position 104 is third of seven. Splitting the phenotyped positions into dominant-negative (87, 104, 121) versus benign (34, 125) gives means of 21.0 vs 32.6 Ų, the wrong direction, Mann-Whitney p = 1.000. R121W, a demonstrated dominant-negative, gains little.

Hypothesis B is therefore refuted as an explanation of pathogenicity. A second parallel analysis in this work (reported below) reached the same verdict via a different protocol, Bio.PDB ShrakeRupley, 10 Å shell, repack-plus-minimise, and reported position 34 first and position 121 last on the same test. Two protocols, same refutation. My §6 numbers differ from theirs in absolute magnitude (+30.0 vs +13.7 Ų at position 104) because the shell definition, SASA implementation and reference state differ; the ordering, which is what the claim rests on, agrees.

What survives. Two narrower statements, both still useful. (i) R104W and R104Q change the chemical character of the position-104 surface in opposite directions (p = 1.7 × 10⁻⁷), so whatever distinguishes them, it is not a shared surface property, they are candidates for mechanistically distinct variants at the same residue. (ii) A folding-ΔΔG method is blind to surface chemistry by construction, so the negative result is not evidence that R104W is structurally uneventful, only that it is thermodynamically uneventful. Neither statement licenses a hydrophobic-association mechanism.

7. Which hypothesis do the data support?

Not A. Local perturbation is measurable but non-specific: it tracks side-chain volume (ρ = +0.77), fires with the largest effect size in the study on a benign control (V125L, δ = +0.79 to +1.00), and produces a ranking (W > Y > F > A > Q > K) that places two uncharacterised variants above the variant of clinical interest. R104W and R104Q are statistically indistinguishable on the Asp84 coordination loss that would have been hypothesis A's mechanism. Local strain is not the thing global ΔΔG was averaging away.

Not B either. Exposed apolar surface is the one axis on which R104W and R104Q separate decisively (§6), and I initially read that as support for hypothesis B. The cross-position control (§6a) removes it: benign position 34 gains 2.3× more apolar area on tryptophan substitution than position 104 does, dominant-negative position 121 gains little, and the DN-versus-benign split runs in the wrong direction (p = 1.000). The metric measures "a large hydrophobic residue was introduced at a partly-exposed site", not pathogenicity. Both hypothesis A and hypothesis B are refuted by the same class of control, a benign variant that scores at or above the pathogenic one.

Not C. Nothing here suggests the ΔΔG calculations were wrong. Both benign controls behave correctly, the two preps agree on every conclusion that clears its own noise floor, self-mutation controls are near-null where they should be, and the one metric that looked dramatic (tier-1 strain) failed its own reproducibility check and was discarded rather than reported. The negative ΔΔG result stands as a true statement about folding stability, the mechanistic explanation is that folding stability is the wrong observable, not that it was measured badly.

Why the global ΔΔG method missed R104W, stated precisely and within what the data support: the perturbation R104W causes is not a stability perturbation at all, not globally, and not locally. It leaves fold stability, secondary structure, backbone geometry, packing quality and carboxylate coordination in the same state a benign substitution does. Restricting the radius of a stability calculation therefore cannot recover it, and neither can the one surface observable tested here. The observable that separates R104W from benign substitutions has not yet been identified in this work. That is a weaker and more honest conclusion than either hypothesis A or B would have given, and it redirects effort: the remaining candidates are properties a monomer calculation cannot see at all, α-α interface geometry, glycosylation-state processing, or CaM-region effects, none of which is measurable in the assay this analysis used.

8. What would falsify this

1. A solvent-exposure measurement showing R104W does not add apolar surface in the assembled channel. The 28 Ų figure is measured on a multi-segment carve-out (NTD 49-140 plus chains 12-37, 166-192, 229-242, 422-433) that reproduces the local burial environment. A calculation on the complete 1395-residue channel, or on a cryo-EM structure of R104W, could revise it. If R104W's apolar surface is occluded in the full assembly, hypothesis B loses its structural basis here.
2. A co-immunoprecipitation or FRET titration showing R104Q associates with wild-type Naᵥ1.5 as avidly as R104W does. Hypothesis B predicts R104W > R104Q in α-α association propensity, since R104Q reduces exposed apolar surface. Equal association would break the link between apolar exposure and the dominant-negative effect. Note the Iamshanova caveat: this must not be done in an overexpression system.
3. A local-strain metric that separates V125L from R104W. If some better-chosen local observable cleanly ranks benign below pathogenic at these positions, hypothesis A returns. The specific failure documented here is that energy redistribution, carboxylate coordination, backbone displacement, packstat and buried-unsat all fail to do so.
4. A phenotype pattern that rescues apolar exposure. The cross-position control (§6a) kills it on two counts: benign position 34 ranks first and dominant-negative position 121 ranks fifth of seven. If R34W were found to be pathogenic and R121W's dominant-negative status were overturned, Wang 2020 already disputes the latter, the control would weaken. Both would have to go the same way, which is unlikely but is the specific evidence that would reopen hypothesis B.

9. Honest limitations

• A repacked-and-minimised model is not a physical ensemble. Every energy here is ref2015 on a locally-relaxed model, with the confound the task named. The self-mutation control quantifies it rather than removing it: the OpenMM noise floor is 6.4 REU of local |ΔE| for a chemically null mutation, which is why R104Q's +2.70 REU excess is reported as above-noise-but-small and R104K's +1.72 is reported as not resolved.
• No sampling of alternative backbone conformations. Cartesian minimisation relaxes the shell around the input backbone; it does not explore whether R104W would prefer a different local fold. A short MD run per variant would test that and was not performed.
• The apolar-exposure result is a static-structure measurement. It says a tryptophan at 104 presents apolar surface in this conformational ensemble. It does not measure association free energy, does not model a second Naᵥ1.5 protomer, and cannot distinguish genuine interface formation from non-specific stickiness, a distinction Iamshanova et al. show is experimentally fraught in this exact protein.
• Gly-X-Gly normalisation is a convention. Absolute Ų values are reported alongside the normalised fractions so any other reference state can be substituted.
• The carve-out is not the full channel. R104 reads 15.1% RSA in the ensemble carve-out versus 11.6% in the complete 1395-residue chain A, the DI-region segments retained (166-192, 229-242, 422-433) reproduce most but not all of the body-donated burial. Apolar exposure for R104W is therefore a slight over-estimate in absolute terms; the W-vs-Q direction, which is the claim, is unaffected because both are measured in the same environment.
• R104F/R104Y/R104A have no clinical or functional data. They are included as a chemical series to separate volume from chemistry, and they succeed at that. They are not answer-key members and are not treated as such.
• Statistical significance ≠ physical significance, and this is load-bearing for the backbone result. 36 matched replicates make a 0.03 Å Cα difference formally significant. It is not structurally meaningful at 3.6 Å parent resolution and is not claimed to be.
• n = 6 for the position-104 rank correlations. ρ = -0.77 has p = 0.072, i.e. not significant at the conventional threshold. The claim rests on the benign-control failure (V125L) and the volume correlation, which are independent of that correlation's significance.
• Single energy function. All local energies are ref2015. The prior project finding that ThermoMPNN and RaSP disagree on R104W's sign is a warning that method choice matters; a second local-energy model was not run.

Part 2. A second test of the hydrophobic hypothesis, by a different protocol

Why this was worth testing

Andrew Glazer, who measured R104Q's dominant-negative effect directly, described the mechanism as probably involving "multi-Nav1.5 complexes that are formed and deleterious interactions between wildtype and misfolded mutant Nav1.5's." That puts misfolding upstream of the interaction.

My own data contradict that at this residue. R104W is the least destabilising of all nineteen substitutions at position 104, and it is the one classified Pathogenic/Likely pathogenic with nine concordant ClinVar submitters. R104Q is clearly destabilising and is only Conflicting.

One reconciliation would have explained everything: what if tryptophan does not destabilise the fold, but instead exposes a hydrophobic surface patch that promotes subunit-to-subunit association? A dominant-negative effect with no misfolding required. The idea was independently plausible because Iamshanova et al. (Sci Rep 2026, PMID 42082654) document that overexpressed Nav1.5 is predisposed to hydrophobic stickiness that mimics genuine interaction.

A prediction was recorded before any measurement: that the ring would bury rather than expose, filling the cavity the arginine vacated, and that the hypothesis would fail on that basis. That prediction was wrong. The hypothesis failed anyway, for a different and better reason.

Method

Twelve OpenMM-relaxed frames of 8VYJ chain A in assembled-channel context (chains A-E present; the NTD is chain B, residues 49-140). At each target position: PyRosetta MutateResidue, repack the 10 Å CA shell with RestrictToRepacking and IncludeCurrent, then minimise side chains and backbone in that shell (LBFGS, two repack/minimise passes). Solvent-accessible area from Bio.PDB ShrakeRupley at atom level, hydrophobic defined as carbon plus sulphur, side chain defined as excluding N, CA, C and O.

Validation before use. ShrakeRupley reproduced this work's independently measured accessibilities on the same structure: R104 13.4% against 11.6%, D82 8.0% against 8.5%, across 171 residues with a median of 45.5% and a wide spread. This check exists because a hand-rolled accessibility routine written earlier in the project returned near-zero for every residue, a consequence of unfiltered hydrogens where the PDB element column is blank.

The clash that nearly produced a false positive

The first pass used repacking only, without minimisation, and appeared to support the hypothesis strongly: aromatics exposed far more hydrophobic carbon than wild type or the polar substitutions, surviving size normalisation, with W versus WT at +15.80 Ų (paired Wilcoxon p = 0.0005).

Then the check for that clash:

A repulsive term of 197 is a severe steric clash, not a comfortably surface-exposed ring. Side-chain repacking alone cannot accommodate an aromatic at position 104; the rotamer is shoved outward and left clashing, and the "exposed area" was largely that clash. Those numbers were discarded.

This is a reusable protocol warning, though it needs care in how it is stated. Running the same test at seven positions shows that after full repack plus minimisation the residual strain from tryptophan is modest everywhere, and lowest of all at position 104 (excess fa_rep +0.70, against +4.21 at position 125, +3.30 at 84, +2.97 at 124). So the +197 was a packing failure of the fixed-backbone protocol, not a physical impossibility: once the backbone can relax, tryptophan is accommodated comfortably at 104. The warning is about protocol, not about the position.

That distinction cuts against an appealing but wrong inference. It would be tempting to read the clash as evidence that R104W is structurally disruptive after all, and therefore that ThermoMPNN's near-zero score is a packing blindness. The control says otherwise: position 104 accommodates tryptophan better than five of the six comparison positions. ThermoMPNN's low score for R104W looks correct, not naive.

After repack plus minimisation, fa_rep returns to baseline: Trp 1.58, Phe 1.46, Tyr 1.24 against wild type 0.84. Only these clash-verified poses were measured.

What the corrected measurement shows

The effect survives the clash correction. It also survives the obvious size objection, and the cleanest framing does not need normalisation at all: total exposed side-chain area is nearly identical across wild type, Gln and Lys (34.2, 37.5 and 34.0 Ų). What changes is the chemical character of the exposed face.

Position 104 is partially exposed in every case. Wild type and the polar substitutions present a polar face to solvent; the aromatics present a carbon face of similar total size.

This refutes my own earlier explanation for R104W's low ΔΔG. I had attributed it to aromatic bulk refilling the vacated cavity (introduces-aromatic r = -0.639, p = 0.0057). A cavity-filling ring would be less exposed than the arginine. The tryptophan is more exposed: side-chain relative accessibility 31.9% against 16.6%. That explanation should be retired or heavily qualified wherever it appears.

The control that kills the hypothesis

The same protocol at seven real N-terminal positions, ranked by carbon area gained on substitution to tryptophan, against what is known about each position:

Two failures, either of which is sufficient:

The benign control ranks first. Position 34 gains three times more exposed hydrophobic area than position 104. If exposed hydrophobic surface drove pathogenicity, position 34 would be the most dangerous site in the domain. It is the one position known to tolerate substitution in healthy people.

A demonstrated dominant-negative variant gains nothing. R121W exposes essentially no additional carbon. The mechanism is not general even across the arginine-to-tryptophan variants it was invented to explain.

Position 104 is mid-pack, fourth of seven.

A second control, which sharpens the negative result

Because the clash raised the question of whether position 104 is unusually intolerant of aromatic substitution, the residual strain after full relaxation was measured at every control position:

Position 104 accommodates tryptophan better than four of the six comparison positions. There is no hidden steric penalty. This independently supports the negative result rather than undermining it: R104W really is a structurally benign substitution at this site, by physical packing as well as by the learned predictor, and it is nonetheless the variant with nine concordant pathogenic submissions.

Re-implementation check

This question was worked twice, without shared code: once as described above (side-chain carbon area, 12 OpenMM frames, 8 Å patch definition), and once by a second AI agent given the same question and no access to the first run's implementation, using its own protocol, its own patch definition, and full apolar surface rather than side-chain carbon only. The two runs agree on the refutation.

This is a re-implementation check and not an independent replication, and the distinction is not cosmetic. Both runs were produced by agents of the same model family. Their agreement demonstrates that the refutation does not depend on one implementation of the patch definition. It does not demonstrate independence from a shared systematic error, because a common architecture can make the same wrong assumption twice for the same reason. A genuinely independent replication would require a different investigator and a different method, and none has been performed. (Section retitled 9 August 2026; it read "Independent replication" as published. No number changes and the refutation stands. Extended 13 August 2026: the 9 August retitle changed the section heading and nothing else, so five other places in this paper still asserted independence as fact — the abstract's "two independent protocols", the Part 2 heading "an independent test", two references to "the independent run", and a results-table column header reading "Independent run". All five now read "second" rather than "independent". This is the same defect shape the paper documents elsewhere: a correction applied at one site while the claim it corrects survives everywhere else. The published version 2 record carries the unpropagated form.)

Spearman ρ = 0.750 (p = 0.052) between the two rankings. Both put the benign control at rank 1 and a demonstrated dominant-negative variant near the bottom. The two failure modes that kill the hypothesis are reproduced by both runs. (This read "reproduced independently" until 13 August 2026; both runs were produced by agents of the same model family, so agreement shows the refutation does not depend on one implementation, not that it survives a shared systematic error. See the re-implementation note in Part 2.)

The one substantial disagreement is instructive rather than troubling. At position 124 the native residue is alanine with a single side-chain carbon, so substituting tryptophan adds nine. A metric counting side-chain carbons registers that differently from one measuring total apolar surface. Neither is wrong; they answer different questions, and the phenotype mismatch survives either.

The second run also extended the analysis to the free N-terminal domain as well as the assembled channel, which matters because pre-assembly is when inter-subunit association would have to initiate. Position 104's rank improves there (2 of 7 rather than 4 of 7), which is the reading least favourable to refutation, and it was reported rather than buried. Even so, the benign control still ranks first in that context too (+80.9 Ų against position 104's +33.2), so the hypothesis fails in both contexts.

Verdict

Hypothesis refuted. Not by the size objection and not by the clash, both of which were controlled for. It fails the phenotype test, which is the one that matters.

The structural observation stands on its own terms: aromatic substitution at position 104 converts a polar solvent-facing surface into a partly hydrophobic one, and cannot be accommodated without backbone relaxation. Neither of those facts explains pathogenicity.

Two things worth keeping. The magnitude was never large enough to carry the claim: roughly 17 Ų is a sub-hotspot patch, where a real protein-protein interface buries 600 to 1000 Ų per side and a single hot-spot residue contributes 50 to 200 Ų. And frame-to-frame variation was 50 to 60% of the mean, so the exposure is strongly conformation-dependent rather than a fixed structural feature.

What would actually test the association hypothesis

Solvent accessibility on a monomer cannot answer it. The question requires the α-α interaction interface itself, and per Iamshanova et al. (Sci Rep 2026, PMID 42082654) that interface cannot be mapped in heterologous overexpression systems, because overexpressed Nav1.5 is predisposed to hydrophobic stickiness that mimics genuine interaction. Native-tissue crosslinking or a structure of the complex would be needed.

Tano et al. 2026 (PMID 41582807) sharpen that requirement in a way worth stating here rather than only in Part 4. Their proximity ligation and co-immunoprecipitation signals are positive for a variant, p.G833R, that has no dominant-negative and no coupled-gating effect. So an association assay run on this protein can return a clean positive with no functional consequence whatever, which is the same failure mode Iamshanova describe from the direction of overexpression artefact, reached instead from the direction of a real dimer that does nothing. Any future test of the association hypothesis needs a functional readout attached to it, not an association readout alone.

Which leaves the local-strain reconciliation, or the possibility that something in my own calculations is wrong. Both remain open.

Data: HYDROPHOBIC_EXPOSURE_RELAXED.csv, POS104_SIDECHAIN_RSA.csv, POS104_MIN_ENERGETICS.csv, POSITION_SPECIFICITY_CONTROL.csv, TRP_EXPOSURE_VS_PHENOTYPE.csv, ctrl_energetics.csv. Twelve frames for position 104, six for the control positions. Statistical error bars only; single relaxation protocol.

Correction, 7 August 2026 — that file list is struck through because five of its six names are not in the data archive and never were. Only HYDROPHOBIC_EXPOSURE_RELAXED.csv is deposited, and it reproduces the corrected-measurement table above exactly. POS104_SIDECHAIN_RSA.csv, POS104_MIN_ENERGETICS.csv, POSITION_SPECIFICITY_CONTROL.csv, TRP_EXPOSURE_VS_PHENOTYPE.csv and ctrl_energetics.csv are outputs of this Part's own protocol that were never written anywhere this project can still reach, so the seven-position control table, the two fa_rep tables and the side-chain accessibility figures in this Part exist only as printed here. The struck names are kept rather than deleted, because they are printed in the record deposited on 5 August 2026 and a reader who went looking for them should find out what happened rather than find a silence. Full account in the correction section below and in P8_REGENERATION_NOTE.md in the archive.

Part 3. No linear degradation or retention motif at Arg104

A negative with a built-in control.

Why this branch mattered

Gütter 2013 listed five candidate mechanisms for R104Q loss of function. this work has closed two of them. "Enhanced degradation" is one of the three that survive, and it is the one most directly testable from sequence: many degradation and endoplasmic-reticulum-retention signals are short linear motifs, so a variant that creates or destroys one has an obvious mechanism.

Result: R104 is not in any canonical motif

Scanning the canonical protein (NP_932173.1, 2016 aa, residue 104 = Arg, verified) for motifs spanning residue 104:

R104 is an isolated arginine. Its local sequence is NKGKTIF-R-FSATNAL, flanked by hydrophobics, not by basic residues. Since Arg→Gln can only destroy arginine-dependent motifs, and there are none to destroy, R104Q cannot act through this class of signal.

The scan was sanity-checked rather than trusted: the same regex finds 8 RxxL occurrences elsewhere in the protein, so the absence in this window is real rather than a broken pattern.

The control that makes this decisive

R121W is the project's other ΔΔG false negative, also an arginine, also in this domain, also pathogenic. If a linear retention motif explained pathogenicity here, the two should look alike.

They do not. R121 sits inside a dibasic pair (FHPI-RR-AAVKIL, R121-R122), which is a canonical retention-signal context. R104 sits alone.

So the two pathogenic arginines of this domain have opposite motif contexts, yet behave the same way: both pathogenic, both false negatives for folding ΔΔG, both with reduced current. A linear-motif mechanism cannot explain the pair, which argues against the class rather than just against R104.

What this does and does not close

Argues against: R104Q acting through a linear degron or a linear endoplasmic-reticulum retention signal.

Does not close the degradation branch. Degradation can be triggered by conformational quality control, misassembly recognised by chaperones, with no linear motif involved anywhere. That is in fact what the intracellular-dimer reconciliation would predict, since quality control there acts on an assembly state rather than a sequence feature. This result removes the simple version of the degradation hypothesis and leaves the conformational version untouched.

Running score on Gütter's five-way fork:

Four of five branches have been narrowed or closed, three of them, all from sequence and published data. The surviving branch is the one the coupled-gating literature already points at.

Sequence NP_932173.1. Motif consensus definitions from standard published forms; this is a sequence-level screen, not a proteomics measurement, and a real degron could exist without matching a canonical consensus.

Part 4. The N-terminal domain has no glycosylation sequon

A clean negative with a real consequence.

Why this was asked

Two 2026 papers converge on glycosylation as central to Naᵥ1.5 α-α association:

• Iamshanova et al. 2026 (PMID 42082654) report that heterologous Naᵥ1.5 interactions "depend on nascent N-linked glycosylation" and are supported by non-native intermolecular disulfide bonds.
• Li & Schmalzing 2026 (PPR1279120) report an intracellular antiparallel homodimer whose surface form is monomeric, with mature Golgi-processed glycosylation confirmed by Endo H / PNGase F.

A third 2026 paper, added here at revision, asks a different question of the same dimer and is the reason the section below ends where it does:

• Tano et al. 2026 (PMID 41582807, PMC12911491) report co-immunoprecipitation and proximity ligation signals for every pairwise combination of wild-type and two variant Naᵥ1.5 channels, in their words irrespective of the electrophysiological characteristics, alongside a biotinylation assay finding no significant difference in cell-surface channel between wild type and either variant.

If the dimer that quality control acts on depends on N-linked glycosylation, and if R104Q sits in or near a glycosylation determinant, that would be a direct mechanistic link from the variant to the dimer. It is a sequence question, answerable in minutes.

Result: there is no site to affect

The canonical N-glycosylation sequon is N-X-S/T where X ≠ P. Scanning residues 1-130 of the canonical sequence (NP_932173.1, 2016 aa, matching transcript NM_000335.5):

Four asparagines, zero sequons. The Naᵥ1.5 N-terminal domain carries no N-glycosylation site at all. Substituting position 104 with Gln, Trp or Lys neither creates nor destroys one.

What this rules out, and what it does not

Ruled out: any mechanism in which R104Q alters dimer behaviour by changing N-linked glycosylation of the N-terminal domain itself. There is nothing there to change.

Not ruled out, and this is the more interesting reading. Both papers report strong binding between full-length channel and the isolated N-terminal domain peptide. If that association is glycosylation-dependent yet the N-terminal domain is itself unglycosylated, then the glycosylation determinant must lie on the partner, the transmembrane regions, which carry the known Naᵥ1.5 glycosylation sites. That makes the interaction asymmetric: an unglycosylated N-terminal domain binding a glycosylated partner surface.

That asymmetry is a testable prediction rather than a conclusion, and it is not something this analysis establishes.

What Tano 2026 adds to the reconciliation, including one thing it takes away

The dimer is not the discriminating observable, and that is the useful finding. The three 2026 papers ask three different questions of the same object. Iamshanova asks what the association depends on and answers nascent glycosylation. Li and Schmalzing ask where it lives and answer intracellular, antiparallel, monomeric at the surface. Tano asks what it does, and answers that it can do nothing: p.G833R dimerises with wild type by both co-immunoprecipitation and proximity ligation while producing no dominant-negative effect and no coupled-gating effect, and the second variant in the same study, p.T1396P, dimerises and reduces peak sodium current by 37 percent. Physical association is therefore necessary for the dimer model and not sufficient for the phenotype. Any argument in this project that moves from "these subunits associate" to "therefore the dominant-negative effect follows" is missing a step, and Tano is the citation that shows the step is real rather than pedantic.

The functional coupling is pharmacologically separable from the dimer, and it runs through 14-3-3. Difopein, a high-affinity inhibitor of the 14-3-3 mediated Naᵥ1.5 interaction, reversed p.T1396P's prolonged current decay and its 5.6 mV rightward shift in steady-state inactivation and raised peak current about 1.5-fold, while the proximity ligation signals were unchanged by difopein. The dimer stays; the interference goes. This bears on Part 2 of this paper from an independent direction. The hypothesis refuted there was that an exposed hydrophobic patch at position 104 drives direct α-α association. If the functional consequence of association is carried by an adaptor protein that a defined inhibitor can remove, then a direct hydrophobic contact between protomers has less work to do in the mechanism than the hypothesis required. The refutation in Part 2 did not need this and does not rest on it. It is corroboration, not evidence, and it is recorded as such.

On trafficking, it constrains the class and not the variant. Tano's biotinylation shows normal cell-surface expression in the presence of a dominant-negative effect, which is the pattern "defective gating in correctly targeted channels" predicts and the pattern "disturbed trafficking" does not. But the variants are p.T1396P and p.G833R, neither in the N-terminal domain and neither at residue 104, so this does not add a fourth surface measurement for R104Q. It says the surviving branch of Gütter's fork is a branch other variants have actually been shown to take.

One thing I could not resolve. Li and Schmalzing report the surface form as monomeric while the intracellular form is dimeric. Tano's co-immunoprecipitation and proximity ligation results, as reported in the full text I read, are not resolved by compartment, so I cannot tell whether their positive signals come from the intracellular pool, the surface pool, or both. That matters, because a 14-3-3 mediated coupled-gating effect at the membrane and an intracellular antiparallel dimer seen by quality control are different objects with the same assay signature. I state the gap rather than pick a reading.

Why the negative is worth recording

It closes a plausible-sounding hypothesis in under an hour and prevents a longer investigation into "does R104Q disrupt a glycosylation signal." It also constrains the fourth reconciliation: whatever couples R104Q to the intracellular dimer, it is not a glycosylation-site effect in the domain where the variant lives.

Methods note, including one error caught

The first attempt used NP_000326.2, which returned 2015 residues rather than the expected 2016. That is isoform b, which lacks Q1077. Re-run against NP_932173.1 (isoform a, 2016 aa, matching the transcript whose CDS was verified as 210..6257 = 2016 codons). The two isoforms are identical across residues 1-130, so the result is unaffected, but the mismatch was checked rather than assumed, because a silent isoform substitution is exactly the kind of error that propagates.

Sequences from NCBI (NP_932173.1, NP_000326.2). Sequon definition N-X-S/T, X ≠ P. Companion sources: PMID 42082654; Europe PMC PPR1279120 (abstract only there; the full text is open on bioRxiv under CC BY-NC-ND).

Part 5. The empagliflozin site and Arg104 are structurally independent

A structural test of the empagliflozin lead, run on 8VYJ. Cheap, decisive, and it makes the therapeutic argument sharper in both directions.

The question

Empagliflozin raises peak sodium current by binding Y1767, the local-anaesthetic site, the same site mexiletine uses, and mutating it abolishes both drugs' effect. R104Q sits in the N-terminal domain. Whether the drug could help depends on whether the two sites are coupled.

The site is intact, and it is far away

Checked against the 8VYJ cryo-EM structure (1,395 modelled residues, 12-1882):

The local-anaesthetic site is present and correctly composed. R104 to Y1767 is 50.7 Å Cα-Cα.

They share no structural neighbourhood at all

Residues within 12 Å of each site:

• R104 shell: 36 residues, segments 58-61, 78-84, 88-96, 101-107, 125, 128-129, 178-179, 185-188
• Y1767 shell: 32 residues, segments 398-410, 1466-1470, 1653-1660, 1760-1774
• Overlap: zero.

Why this is good news for the lead

The drug's mechanism cannot be blocked by the variant. Empagliflozin acts through a site that shares no contact surface with R104, so the N-terminal defect cannot interfere with binding or with the conformational change that follows. Whatever is wrong at position 104, the local-anaesthetic site is intact and available.

That is the best possible structural answer for transferability: the drug does not need the N-terminal domain to be functional in order to work.

Why it is also the problem

By the same argument, the drug cannot repair the defect. It has no contact with the damaged region. It would act by increasing delivery of whatever channel protein exists, mutant and wild-type alike.

In a heterozygote, that means more mutant subunit at the membrane. If the mechanism is dominant-negative interference occurring at the membrane, more mutant protein is the wrong direction. The structural independence that makes the drug robust to the variant is exactly what prevents it from being selective.

An incidental structural result

R104's 8 Å shell contains only its own domain (residues 60, 92-95, 102-106). Its nearest out-of-domain contacts sit at 8-12 Å: A178, R179, A185, F186, T187, F188, the start of the channel body.

So R104 is not buried in an isolated appendage, but nor is it in direct contact with the pore-forming regions. It sits near the domain boundary with second-shell reach into the channel body. That is consistent with the earlier finding that this residue is not N-terminal-domain-internal, and it is the structural basis for how a defect there could influence the whole channel without touching it directly.

What would falsify these conclusions

Each refutation rests on a control, and each control is the thing to attack.

The local-strain refutation would be overturned by a strain metric on which the benign controls R34C and V125L score below the dominant-negative variants. Mine did not; a better-constructed metric might.

The hydrophobic-surface refutation would be overturned by a cross-position test in which apolar gain does track phenotype, or by evidence that position 34 is not in fact benign. Both are checkable. The benign status of R34C rests on its appearance in unrelated healthy individuals, including in the report that first described R104Q (PMID 11960580).

The degron refutation would be overturned by a non-canonical or structurally-defined retention determinant at Arg104 that a linear motif scan cannot see, which is a real possibility and the reason I state the claim as being about linear motifs specifically.

The glycosylation conclusion would be overturned by a glycosylation site outside the canonical N-X-S/T sequon, or by evidence that the relevant sequon lies on an interaction partner rather than in this domain.

The drug-independence conclusion would be overturned by an allosteric path between the two sites, which a distance and shell analysis cannot exclude. Fifty Angstroms of separation with no shared shell residue argues against direct competition, not against long-range coupling.

A challenge to the framing rather than to any result above, stated at its own strength. Every mechanism tested and eliminated in this paper is a candidate explanation for a dominant-negative effect, and the value of eliminating them assumes the dominant-negative effect is what sets how ill a carrier becomes. Tano et al. 2026 (PMID 41582807) argue directly against that assumption. In their family the carrier of the variant that has the dominant-negative effect and the coupled-gating signature, p.T1396P, was the milder of the two: sinus bradycardia from her forties, no permanent pacing device at seventy. The proband, carrying that variant plus p.G833R, which has neither effect, presented at thirty-two with severe sinus bradycardia and recurrent ventricular tachycardia. Their reading is that losing coupled gating is itself harmful, which inverts the framing rather than softening it. The bounds on that claim are real and I will not use them to bury it: it is sick sinus syndrome and bradycardia rather than Brugada syndrome, one pedigree, a severity comparison between two people of different ages and different genotypes, a compound heterozygote against a single heterozygote so that allelic burden is not separated from the coupled-gating interpretation, both variants classified as variants of uncertain significance, every measurement in HEK293 cells by the authors' own first limitation, and the one other affected family member never genotyped. None of that touches the structural refutations in this paper, which stand or fall on their benign controls. What it touches is how much a resolved mechanism would be worth, and the honest answer is that this has not been established for any SCN5A variant, including this one.

Correction, 7 August 2026: the data availability statement was partly false, and so was the file list closing Part 2

Version 1 of this paper, deposited at 10.5281/zenodo.21799867 on 5 August 2026, told its readers that four kinds of derived table were in the shared data archive. One of the four was not, and separately the italic data line closing Part 2 named six CSV files of which five were not. This was found on 7 August 2026, in the last of an eleven-paper audit of data availability statements begun on 6 August, which found the same class of defect in six other papers.

Lead with what was wrong rather than with what has been fixed.

And a fifth thing was never claimed and was also missing: Part 1 has no deposited data at all. The 1,008-model Rosetta and OpenMM analysis behind sections 2 through 6a has no file in the archive. That was tested rather than assumed — HYDROPHOBIC_PATCH_ANALYSIS.csv was checked against Part 1's printed tables under every plausible reading of its columns and reproduces none of them, because Part 1 carries a three-seed axis and a self-mutation control the deposited file has no column for. Every effect size, confidence interval and Cliff's δ in Part 1 therefore exists only as printed here. The statement below says so.

What was regenerated, and how far it verifies. Everything in this paper that derives from a public structure or a public sequence was regenerated by p8_regen_structure_and_sequence.py, written from the method as this paper states it, from PDB entry 8VYJ chain A and the 2,016-residue canonical sequence. Forty-three of the forty-five quantities checked reproduce exactly, including the shell residue lists themselves, the 50.7 Å separation of Arg104 from Tyr1767, the zero overlap between the two 12 Å shells, all seven motif classes, the sanity-check count of eight RxxL occurrences elsewhere in the protein, the four asparagines and the zero sequons in residues 1–130, and Arg104's 13.4 % Shrake–Rupley accessibility. The two that do not reproduce are in one sentence of Part 2's validation paragraph — Asp82's accessibility, printed 8.0 % and recomputing to 9.2 %, and the median accessibility across the 171-residue carve-out, printed 45.5 % and recomputing to 41.0 %. Both are consistent with a different choice of reference maxima for those two figures than for the Arg104 figure beside them, which reproduces exactly. The printed values are left as printed and the deviations are recorded rather than quietly adjusted. Neither bears on any conclusion; the sentence's purpose is to show that Shrake–Rupley agrees with an independently measured accessibility, and on the figure it is actually comparing it does.

One notational looseness fell out of the regeneration and is corrected here rather than in the text above, because the text above is not wrong. Section 5 prints the Arg104 12 Å shell as contiguous runs and the Tyr1767 12 Å shell as spans, and does not say so. 398-410 covers 398, 401–407, 409 and 410 — ten residues, not thirteen; 1466-1470 is three residues and 1653-1660 is four. The residue count of 32 is exact, every residue listed really is in the shell, and the zero-overlap conclusion is unaffected — but a reader adding up the printed ranges gets 41 rather than 32. P8_SHELL_RESIDUES_8VYJ.csv lists all three shells residue by residue, which removes the ambiguity.

Two conventions had to be recovered from the numbers because this paper does not state them, and they are stated here so that the next reader does not have to. The shells are Cα-to-Cα and include the target residue itself — that is what gives 36, 32 and the printed range 101-107, which contains 104; a closest-heavy-atom shell at the same cutoffs gives 67 and 72. And the RxxL sanity count is a non-overlapping scan, which gives 8; an overlapping scan gives 9, because Arg1913 and Arg1914 each start one.

One further defect, found by the same audit, stated because it is checkable and not resolved because it cannot be. This paper prints two different sets of numbers for the same quantity. Section 6a's cross-position table gives the apolar gain on tryptophan substitution as +68.4 at position 34, +32.7 at 124, +30.0 at 104, +18.9 at 87, +14.0 at 121, +9.1 at 84 and −3.2 at 125. Part 2's replication table, reporting the same run in its Independent run column, gives +84.2, +51.0, +27.8, +29.6, +9.1, +14.1 and −1.0 — different magnitudes, and 84 and 121 the other way round, so position 104 ranks third in one table and fourth in the other. Section 6a's own text quotes +30.0 as this run's value at position 104 and Part 2 attributes +27.8 to it.

The deposited file settles which of the two is reproducible and not which of the two is right. Part 2's column reproduces from HYDROPHOBIC_PATCH_ANALYSIS.csv to the printed digit, at every one of the seven positions, in both structural contexts. Section 6a's table reproduces from no reading of that file that was tried — not either apolar measure, not four normalisations, not either context, not either relaxation prep separately or pooled. That is consistent with section 6a having been computed on a different run of the same protocol whose output is not deposited, and it is equally consistent with an error in one of the two tables. Part 1's own data does not exist, so this project cannot tell which, and it will not guess in order to look tidy.

What does not change either way, and it is the whole of section 6a's argument. In both tables the benign control at position 34 ranks first by a wide margin, position 125 is at the bottom, and the dominant-negative-versus-benign split runs the wrong way. The refutation of hypothesis B rests on the ordering, and the two tables agree on the ordering wherever the argument uses it. The discrepancy is recorded here because a reader who adds up this paper's two tables will find it, and finding it undocumented is worse than finding it named.

Why the checker did not catch this, which is the part worth carrying forward. presubmit_check.py check 5 compares files cited by a paper against the deposit, and it extracts citations with a regular expression that matches backticked filenames only. Part 2's six names are in plain italic prose. The check therefore had nothing to match and reported nothing wrong, for a paper that named five missing files explicitly. Every filename in the statement below is in backticks so that the check has teeth on this paper from version 2 onward, and widening that regex to catch bare .csv tokens is the obvious follow-up and has not been done.

None of this changes a result. No mechanism moves into or out of the eliminated set. The corrections are to what this paper says about where its data are, and to one piece of notation.

This section is not in the record deposited at 10.5281/zenodo.21799867 on 5 August 2026, which carries the false statement.

Data availability

This statement was rewritten on 7 August 2026 because the version published on 5 August 2026 named a table that was not in the archive, and Part 2 named five more; see the correction section above.

All primary data are public. The structure is Protein Data Bank entry 8VYJ, chain A, with cryo-EM map EMD-43662 for density-guided relaxation. The reference protein sequence is RefSeq NP_932173.1 and the transcript NM_000335.5. Paralogue sequences are the nine human Nav alpha subunits retrieved from RefSeq, accessions as listed in the relevant section. Variant classifications came from ClinVar via NCBI E-utilities. Published functional data are cited by PMID throughout. Solvent accessibility was computed with Bio.PDB ShrakeRupley; relaxation used Rosetta and OpenMM. Tool version numbers are not recorded in my working notes for every step and I report that gap rather than reconstruct it.

The derived tables are deposited in the data archive whose identifier is recorded in DATA_DOI.txt alongside this manuscript and should be cited as the data source. From version 2 of that archive onward they are these eleven files, named individually so that a reader can check this statement against the archive rather than take it on trust:

• HYDROPHOBIC_PATCH_ANALYSIS.csv — the per-model accessibility and energetics table, 2,592 rows and 56 columns, covering both the position-104 substitution ladder and the seven-position tryptophan scan, in both the assembled-channel and free-N-terminal-domain contexts, at all three protocol tiers. This is the run Part 2 calls the second run; its column of that table — position 34 +84.2, 124 +51.0, 87 +29.6, 104 +27.8, 84 +14.1, 121 +9.1, 125 −1.0 Ų — is this file filtered to run = position_specificity_control, context = assembled, tier = tier2_cart, column sc_apolar.
• HYDROPHOBIC_EXPOSURE_RELAXED.csv — Part 2's own position-104 measurement, 12 frames × 7 substitutions, from which the corrected-measurement table in Part 2 reproduces exactly.
• NTD_PARALOG_CONSERVATION.csv — the per-position conservation table across the nine human Nav paralogues, which supports the abstract's claim that Arg104 and Asp84 are both invariant across them.
• P8_SHELL_RESIDUES_8VYJ.csv — the shell residue lists. All three shells of Part 5 residue by residue: Arg104 at 8 Å, Arg104 at 12 Å and Tyr1767 at 12 Å, each row carrying the Cα–Cα distance, the closest-heavy-atom distance and whether the residue is inside the N-terminal domain. This table is a 7 August 2026 regeneration, not the 5 August original, which does not exist.
• P8_SITE_GEOMETRY_8VYJ.csv — every structural quantity Part 5 prints, each beside the value printed here: the modelled residue count and range, the Arg104-to-Tyr1767 distance, the two shell sizes, the overlap, the 8 Å shell, the six out-of-domain contacts and the composition of the local-anaesthetic site.
• P8_LINEAR_MOTIF_SCAN.csv — Part 3's scan, one row per motif class, with occurrence counts under both the overlapping and non-overlapping conventions, whether the motif spans residue 104, and every hit in the ±25-residue window.
• P8_GLYCOSYLATION_SEQUON_SCAN.csv — Part 4's scan, one row per asparagine in residues 1–130.
• P8_RESIDUE_SASA_8VYJ.csv — Shrake–Rupley solvent-accessible area and relative accessibility for all 1,395 modelled residues of chain A and for the 171-residue carve-out, which is the reference behind Part 2's validation paragraph.
• p8_regen_structure_and_sequence.py — the script that writes the five P8_* tables from the public 8VYJ coordinates and the public canonical sequence, and checks them against every quantity printed in this paper.
• P8_VERIFICATION_OUTPUT.txt — that script's output, setting each reproduced quantity beside the value printed here. 43 of 45 reproduce; the two that do not are named in the correction section above.
• P8_REGENERATION_NOTE.md — the provenance note, which states that the P8_* tables are a 7 August 2026 reconstruction and why the originals are unrecoverable.

Four things this paper reports are not deposited, and this statement says so rather than implying otherwise.

1. Part 1's per-model table does not exist. The 1,008-model Rosetta and OpenMM analysis behind sections 2 through 6a was never written to disk in a form this project can still reach, and it cannot be regenerated: it needs 12 Rosetta density-guided relaxations against map EMD-43662 and 12 restrained-OpenMM relaxations, and those ensembles are not retained. A fresh ensemble would be a different ensemble whose agreement with the printed effect sizes is unknown, so producing one and depositing it as Part 1's data would replace one false statement with another. Every excess local |ΔE|, every Asp84 donor count, every backbone RMSD, every DSSP agreement figure, every packstat and buried-unsatisfied count, and both apolar-exposure tables in sections 6 and 6a therefore exist only as printed in this paper.
2. Five files named in Part 2's own data line do not exist: POS104_SIDECHAIN_RSA.csv, POS104_MIN_ENERGETICS.csv, POSITION_SPECIFICITY_CONTROL.csv, TRP_EXPOSURE_VS_PHENOTYPE.csv and ctrl_energetics.csv. Those five names are deliberately not in backticks, and every filename in this paper that names a deposited file is. presubmit_check.py check 5 reads a backticked CSV name as a claim that the file is in the deposit and fails the kit when it is not; these five are the exact opposite of that claim, so backticking them would either break the check or teach a reader that the two notations mean the same thing. In this paper, backticks mean deposited. They are outputs of Part 2's own protocol, which is not the protocol behind HYDROPHOBIC_PATCH_ANALYSIS.csv and does not reproduce from it. Part 2's seven-position control table, its two fa_rep tables including the +197.82 clash figure, and its side-chain relative accessibilities exist only as printed.
3. The relaxed structural ensembles are not deposited. Neither the 12 Rosetta nor the 12 OpenMM frames, in either context, are retained anywhere in this project.
4. Tool versions are not recorded for Rosetta, OpenMM or PyRosetta, as the paragraph above already states. The regeneration script records the version of every library it uses itself, which covers the P8_* tables and nothing else.

The two inputs to the regeneration are public and are identified by checksum inside the script: 8VYJ chain A as 8vyj_full.pdb, md5 63b182c23e25e65148445e17b9c36310, and the 2,016-residue canonical sequence as nav15_canonical.fasta, md5 a6d7a83147964ee337df4e05514354ae. Neither file is redeposited in the archive, because both are copies of public reference records — PDB entry 8VYJ and the canonical SCN5A protein sequence — and this archive does not republish third-party reference data. The checksums are what make the regeneration reproducible against a fresh download. Residue identity is checked at every position this paper names before anything is computed, because a silent isoform substitution is the error Part 4's methods note records having made once.

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

Competing interests

I am a heterozygous carrier of SCN5A p.Arg104Gln, the variant at the centre of this analysis, and I have a clinical diagnosis of Brugada syndrome. No funding was received.

Nothing here is clinical guidance for any person, including me. In particular, the empagliflozin section reports a structural relationship and is not a statement about whether any person should or should not take that drug. That decision belongs to a treating clinician.

Use of AI tools

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

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

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

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

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

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

References

1. Clatot J, et al. Dominant-negative effect of SCN5A N-terminal mutations through the interaction of Nav1.5 alpha-subunits. Cardiovasc Res 2012;96(1):53-63. PMID 22739120.
2. Wang Z, et al. Calmodulin binds to the N-terminal domain of the cardiac sodium channel Nav1.5. Channels (Austin) 2020;14(1):268-286. PMID 32815768.
3. O'Neill MJ, et al. Dominant negative effects of SCN5A missense variants. Genet Med 2022;24(6):1238-1248. PMID 35305865.
4. Moreau A, et al. Mexiletine differentially restores the trafficking defects caused by two Brugada syndrome mutations. Front Pharmacol 2012;3:62. PMID 22529811.
5. Levy-Nissenbaum E, et al. Genetic analysis of Brugada syndrome in Israel: two novel mutations and possible genetic heterogeneity. Genet Test 2001;5(4):331-334. PMID 11960580.
6. Gutter 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. PMID 23805106.
7. Protein Data Bank entry 8VYJ, human Nav1.5 cryo-EM structure, with map EMD-43662.
8. RefSeq NP_932173.1 (SCN5A protein) and NM_000335.5 (transcript).
9. Tano A, Kato K, Yamauchi K, Jinzai H, Iguchi T, Toyoda F, Baba Y, Kubo T, Ohno S, Makiyama T, Nakagawa Y, Horie M. Phenotypic severity of SCN5A-related bradycardia is independent of dominant-negative and coupled gating effects. Circ Arrhythm Electrophysiol 2026. doi:10.1161/CIRCEP.125.014270. PMID 41582807, PMC12911491.

Additional PMIDs cited at the point of use in the text are not repeated here.

Back to all ten papers