Computational protein stability analysis of SCN1A missense variants reveals domain‐dependent stability patterns Authors: Youngkyu Shim, Eungu Kang, Naeun Kwak, Suhyun Kim Published: 2026 ABSTRACT AbstractObjectiveTo determine whether computational protein‐stability predictions discriminate pathogenic from benign SCN1A missense variants, and to characterize the structural distribution of predicted destabilization among pathogenic variants.MethodsOn an AlphaFold3‐predicted Nav1.1 structure, FoldX, and Rosetta Cartesian ΔΔG were computed for a single ClinVar snapshot of pathogenic/likely‐pathogenic (P/LP) and benign/likely‐benign (B/LB) missense variants and its extension to ClinVar variants of uncertain significance (VUS) and gnomAD v4.1 variants; membrane‐aware RosettaMP was applied to the patch‐clamp subgroup. Pathogenic variants were stratified by functional domain.ResultsPathogenic variants were more destabilizing than benign (FoldX 2.61 vs. 0.31 kcal/mol, p = 1.27 × 10−11; ROC‐AUC = 0.760), concordant with Rosetta (ROC‐AUC = 0.697; ρ = 0.660). Destabilization was domain‐dependent: pore (P‐loop/selectivity‐filter) pathogenic variants were depleted of stability‐neutral variants (0.40‐fold; Bonferroni‐adjusted p = 4.3 × 10−7), whereas S4 voltage‐sensor variants were enriched for them (2.32‐fold; p = 0.013). Across ~3300 non‐redundant variants, gnomAD‐common variants resembled benign controls and VUS were intermediate (mean ΔΔG 0.98 kcal/mol; 18.5% strongly destabilizing), with the domain pattern preserved. Among 64 patch‐clamp variants, stability did not separate gain‐ from loss‐of‐function, though gain‐of‐function variants clustered in voltage‐sensing domains and were absent from the pore.SignificanceComputational stability analysis thus adds a mechanistic layer complementary to the conventional gating‐dysfunction view, distinguishing a destabilized pore‐region subset—for which proteostasis impairment is a candidate, though unproven, mechanism—from a structurally tolerated S4 subset whose pathogenicity is stability‐independent. As a hypothesis‐generating rather than mechanism‐defining approach, this stratification prioritizes candidate variants—including the 18.5% of VUS that are strongly destabilizing—for direct functional and surface‐expression validation in SCN1A‐related epilepsies. Plain Language SummaryWe used computational modeling to predict how thousands of SCN1A genetic variants influence the stability of the Nav1.1 sodium channel protein. Disease‐causing variants tended to destabilize the protein more than benign variants, and variants in the pore region—where ions flow through the channel—were predominantly destabilizing. This is consistent with loss‐of‐function arising from misfolding and degradation of the channel protein in this subset of variants. By contrast, variants in the voltage‐sensing region were often structurally tolerated, indicating that their disease‐causing effects likely arise through a different mechanism that requires direct functional measurement to define. Accordingly, the analysis nominates a candidate pore‐region subset potentially affected by proteostasis impairment and a complementary stability‐neutral subset warranting functional evaluation. Key points Pathogenic SCN1A variants in pore‐associated regions (P‐loop/selectivity filter) tend to be more destabilizing—compatible with, though not proof of, a candidate proteostasis‐impaired mechanism complementary to gating dysfunction.Pathogenic variants in S4 voltage‐sensing domains are predominantly stability‐neutral, a pattern compatible with—but not direct evidence of—gating dysfunction.Domain‐stratified stability analysis is hypothesis‐generating, prioritizing candidate variants for direct functional and surface‐expression validation rather than assigning a mechanism. Key points Pathogenic SCN1A variants in pore‐associated regions (P‐loop/selectivity filter) tend to be more destabilizing—compatible with, though not proof of, a candidate proteostasis‐impaired mechanism complementary to gating dysfunction.Pathogenic variants in S4 voltage‐sensing domains are predominantly stability‐neutral, a pattern compatible with—but not direct evidence of—gating dysfunction.Domain‐stratified stability analysis is hypothesis‐generating, prioritizing candidate variants for direct functional and surface‐expression validation rather than assigning a mechanism. Pathogenic SCN1A variants in pore‐associated regions (P‐loop/selectivity filter) tend to be more destabilizing—compatible with, though not proof of, a candidate proteostasis‐impaired mechanism complementary to gating dysfunction. Pathogenic SCN1A variants in pore‐associated regions (P‐loop/selectivity filter) tend to be more destabilizing—compatible with, though not proof of, a candidate proteostasis‐impaired mechanism complementary to gating dysfunction. Pathogenic variants in S4 voltage‐sensing domains are predominantly stability‐neutral, a pattern compatible with—but not direct evidence of—gating dysfunction. Pathogenic variants in S4 voltage‐sensing domains are predominantly stability‐neutral, a pattern compatible with—but not direct evidence of—gating dysfunction. Domain‐stratified stability analysis is hypothesis‐generating, prioritizing candidate variants for direct functional and surface‐expression validation rather than assigning a mechanism. Domain‐stratified stability analysis is hypothesis‐generating, prioritizing candidate variants for direct functional and surface‐expression validation rather than assigning a mechanism. INTRODUCTION SCN1A encodes the voltage‐gated sodium channel Nav1.1, a 2009‐amino acid protein predominantly expressed in GABAergic inhibitory interneurons essential for maintaining excitation–inhibition balance. 1 Pathogenic SCN1A variants represent the most common monogenic cause of epilepsy, with a phenotypic spectrum from genetic epilepsy with febrile seizures plus (GEFS+) to Dravet syndrome. 2 , 3 Critically, cognitive decline frequently progresses despite adequate seizure control, 4 suggesting that mechanisms beyond altered neuronal excitability contribute to disease pathogenesis. The SCN1A field has predominantly characterized variant pathogenicity through electrophysiological measurements of gating properties. 2 However, missense variants can cause disease through at least three mechanisms: (1) protein destabilization leading to misfolding, endoplasmic reticulum retention, and proteostatic degradation, reducing functional channel density at the cell surface; (2) gating dysfunction, where structurally stable channels exhibit altered voltage‐dependent activation, inactivation, or persistent current; or (3) compound mechanisms combining both. 5 , 6 , 7 Proteostasis impairment may be more prevalent than currently appreciated: Rusconi et al. 6 and Bechi et al. 8 demonstrated rescuable folding defects in SCN1A variants, suggesting that a fraction of apparently loss‐of‐function variants operate through proteostatic mechanisms. Because gating dysfunction and proteostasis impairment may require different therapeutic approaches, distinguishing these mechanisms is clinically relevant. Protein stability changes can be computationally predicted using physics‐based methods such as FoldX 9 and Rosetta Cartesian ddG, 10 which estimate the free energy change upon mutation. The availability of AlphaFold3‐predicted structures now enables systematic stability analysis for membrane proteins lacking experimental structures. 11 Here, we test whether computational protein‐stability predictions discriminate pathogenic from benign SCN1A missense variants and characterize the structural distribution of predicted destabilization among pathogenic variants to explore whether protein destabilization—and hence possible proteostasis impairment—contributes to SCN1A pathogenicity alongside conventional gating dysfunction. MATERIALS AND METHODS Variant dataset SCN1A missense variants were curated from a single ClinVar extraction in May 2026 into pathogenic/likely‐pathogenic (P/LP), benign/likely‐benign (B/LB), and variants of uncertain significance (VUS) classes. All SCN1A missense variants observed in gnomAD v4.1 were additionally pulled via the public GraphQL endpoint and stratified by allele frequency into ultra‐rare (AF <1 × 10−5), rare (1 × 10−5 ≤ AF <1 × 10−4), and common (AF ≥ 1 × 10−4) tiers. gnomAD‐common variants were treated as a population‐level “likely‐benign” comparator independent of ClinVar curation. Variants observed in both ClinVar and gnomAD were assigned to ClinVar strata by precedence. Functional and phenotype data Published whole‐cell patch‐clamp data were collected from ClinVar‐linked and additional published studies and classified by functional effect following Knox et al. 12 : complete LOF (<5% wild‐type peak Na+ current), partial LOF (5–80%), GOF (≥ 80% with altered gating: hyperpolarized activation, depolarized fast inactivation, slowed inactivation, or persistent current), and mixed. Stability predictions for variants identified outside ClinVar used the same FoldX and Rosetta pipeline. Phenotype data were obtained from ClinVar submissions and original publications; pathogenic variants were classified as severe (Dravet syndrome or developmental and epileptic encephalopathy) or mild (GEFS+ or febrile seizures), and variants with ambiguous or absent annotations were excluded. Structural modeling The AlphaFold3‐predicted structure of human Nav1.1 (SCN1A; UniProt P35498, 2009 amino acids) was generated using AlphaFold Server with the full‐length sequence as input. Per‐residue structural confidence was assessed using pLDDT scores extracted from the model coordinates. Prior to mutagenesis calculations, the structure was prepared using FoldX RepairPDB to optimize local side‐chain geometry and reduce steric clashes. For variant‐level stability estimation, FoldX BuildModel and Rosetta cartesian_ddg were used because both methods perform mutation‐site‐centered side‐chain repacking and local conformational optimization as part of their standard ΔΔG calculation workflows. Conformational‐state characterization of the AlphaFold3 model Four structural descriptors of the FoldX‐input AlphaFold3 model were calculated in a pore‐aligned local coordinate frame (origin: Cα centroid of the DEKA selectivity‐filter residues D382, E951, K1432, A1724; positive z‐axis: extracellular): (i) the DEKA Cα cross‐section—six pairwise Cα–Cα distances among the DEKA residues; (ii) the activation‐gate geometry—minimum sterically corrected pore radius across the cytoplasmic S6 bundle‐crossing region (z = −15 to 0 Å), or the minimum heavy‐atom distance to the pore axis where steric correction could not be applied; (iii) the S4 outermost gating‐arginine z‐position per voltage‐sensing domain (DI–DIV)—pore‐axis projection of the most extracellular canonical S4 arginine; and (iv) the fast‐inactivation motif geometry—Cα distance between F1499 (IFM motif) and the DIV S6 receptor‐pocket centroid. Descriptors (i) and (ii) were additionally calculated on a panel of cryo‐EM Nav structures—7 K48, 13 6J8E, 14 7XVE, 15 6UZ3, 16 7DTC, 17 7XM9, 18 and 6AGF 19 for comparison with the AF3 values; the DEKA cross‐section was reported only for structures with sequence‐unambiguous DEKA‐equivalent residues. The four AF3 descriptors were jointly interpreted against geometric patterns characteristic of established Nav conformational states (resting, activated/open, fast‐inactivated). FoldX stability calculations FoldX v5.1 9 was used to calculate ΔΔG for all variants. BuildModel was run with 5 independent iterations per mutation; values were averaged and treated as a continuous variable. Positive ΔΔG indicates destabilization; estimated prediction error is ±0.46 kcal/mol. 20 , 21 The optimal threshold for pathogenicity discrimination was determined by Youden's J statistic. Rosetta Cartesian ddG calculations Rosetta Cartesian ΔΔG 10 was used as an independent validation method with the ref2015_cart scoring function, 3 iterations per mutation, and Cartesian space minimization. ΔΔG values in Rosetta Energy Units (REU) were treated as a continuous variable. RosettaMP membrane‐aware stability calculations To assess whether membrane environment modeling improves predictions, RosettaMP with the franklin2019 scoring function 22 was applied to the electrophysiologically characterized pathogenic variants and all benign variants. The AlphaFold3 structure was positioned in an implicit lipid bilayer based on the Nav1.1 transmembrane topology (24 TM helices across DI–DIV), with a span file generated from UniProt P35498 annotations. For each variant, local FastRelax with franklin2019 was applied to the 8 Å neighborhood in torsion space (3 rounds with different random seeds), adapting the local refinement strategy of the cartesian_ddg protocol 10 to the membrane context. Comparison with sequence‐based pathogenicity predictors FoldX and Rosetta were compared with two deep‐learning pathogenicity predictors: AlphaMissense, 23 an AlphaFold‐derived model that integrates protein structural context with evolutionary conservation and outputs a per‐variant pathogenicity score from 0 (predicted benign) to 1 (predicted pathogenic), and EVE (Evolutionary model of Variant Effect 24 ;), a deep generative model that predicts variant pathogenicity from multiple sequence alignments of evolutionarily related protein families. Precomputed scores for all variants were obtained from respective databases; AlphaMissense pathogenicity scores were extracted in May 2026 from the official release for UniProt P35498, covering all possible SCN1A missense substitutions. Statistical significance of AUC differences was assessed by bootstrap resampling (10 000 iterations). Statistical analysis Analyses were performed using Python 3 (SciPy, scikit‐learn). Between‐group comparisons used Mann–Whitney U tests; correlations were assessed using Pearson and Spearman coefficients. ROC‐AUC 95% confidence intervals were estimated by bootstrap resampling (2000 iterations). Optimal thresholds were determined by Youden's J statistic; effect sizes by Cohen's d. For domain‐stratified enrichment analyses across five functional regions, Bonferroni correction was applied (adjusted α = 0.01). Statistical significance was set at α = 0.05 for all other analyses. Ethical considerations This study utilized publicly available data from ClinVar and published literature. Variant classifications followed ACMG standards 25 RESULTS Dataset and structural model The single‐snapshot calibration set comprised 1045 SCN1A missense variants (985 P/LP, 60 B/LB). gnomAD v4.1 contributed 1803 unique protein‐level missense variants (1543 PASS‐filter; 13 of these gnomAD‐common, AF ≥1 × 10−4), and ClinVar contributed 1275 VUS; the total non‐redundant extended set comprised approximately 3300 unique missense substitutions. The AlphaFold3 model achieved pTM = 0.73 and ranking score = 0.86, with approximately 25% of the structure classified as disordered, primarily in intracellular loops. Transmembrane segments, pore helices, and P‐loop regions showed high confidence (mean pLDDT 84–92), while intracellular linkers (DI–DII mean pLDDT 37.0; DII–DIII 46.7) and the C‐terminus (65.4) showed very low to low confidence. Prediction reliability rose monotonically with AlphaFold3 structural confidence across the three pLDDT bins shown in Figure 1A. FoldX predictions in disordered regions were compressed toward zero rather than randomly noisy (Spearman ρ ≈ 0.15 for pLDDT vs. |ΔΔG|). The core findings below derive predominantly from variants in very‐high‐ and confident‐pLDDT regions. Protein stability analysis of the SCN1A missense calibration set (n = 985 P/LP + 60 B/LB). (A) ROC‐AUC stratified by AlphaFold3 structural confidence (pLDDT) across three bins: Low (< 70), confident (70–90), and very high (≥ 90). Bar groups show FoldX, Rosetta, and combined predictor AUC for each pLDDT bin; n = number of variants per bin. (B) Violin plots of FoldX ΔΔG (left y‐axis, kcal/mol) and Rosetta Cartesian ΔΔG (right y‐axis, REU) for P/LP versus B/LB. P/LP variants demonstrated significantly greater destabilization by both methods (Mann–Whitney ***p < 0.001). (C) Scatter plot of FoldX versus Rosetta ΔΔG for all calibration variants, colored by ClinVar classification (Spearman ρ = 0.660, p < 10−100). (D) ROC curves comparing FoldX (AUC = 0.760), Rosetta (AUC = 0.697), and the combined Z‐score predictor. Optimal Youden thresholds are marked by filled circles. B/LB, benign/likely benign; P/LP, pathogenic/likely pathogenic. Protein stability and pathogenicity Pathogenic variants showed significantly greater predicted destabilization than benign variants by FoldX (2.61 ± 3.98 vs. 0.31 ± 0.97 kcal/mol, p = 1.27 × 10−11, Cohen's d = 0.79). Rosetta Cartesian ΔΔG was concordant (FoldX–Rosetta Spearman ρ = 0.660, p < 10−100). Among pathogenic variants, 46.5% had FoldX ΔΔG > 2.0 kcal/mol versus 1.7% of benign variants; 43.3% of benign variants had ΔΔG within ±0.5 kcal/mol versus 17.8% of pathogenic variants. FoldX achieved ROC‐AUC = 0.760 for P/LP discrimination (Figure 1B–D). The optimal FoldX threshold of ΔΔG > 1.60 kcal/mol (Youden's J) achieved high specificity (95.1%) at moderate sensitivity (53%), supporting use as high‐confidence supporting evidence for pathogenicity rather than as a standalone classifier. Domain‐stratified stability analysis Stability predictions stratified by functional domain revealed pore‐associated regions with substantially greater destabilization of pathogenic variants than VSDs (Figure 2; P‐loop vs. S4 mean FoldX ΔΔG, Mann–Whitney p < 0.001). Domain‐stratified stability effects in P/LP SCN1A variants. Box‐and‐strip plots of predicted stability changes for P/LP variants stratified by Nav1.1 functional domain (5 categories), with FoldX ΔΔG (left y‐axis, solid boxes, kcal/mol) and Rosetta Cartesian ΔΔG on the cross‐method subset (right y‐axis, dashed boxes, REU) displayed side‐by‐side. S4 voltage‐sensor variants showed significantly lower destabilization than P‐loop/selectivity‐filter variants by both methods (Mann–Whitney ***p < 0.001; bracket marked). To quantify domain‐specific patterns independently of benign variant counts, we compared the within‐pathogenic proportion of near‐neutral (|ΔΔG| ≤ 0.5 kcal/mol) versus strongly destabilizing (ΔΔG > 2.0 kcal/mol) variants across functional regions (Fisher's exact test, Bonferroni correction; Table 1). P‐loops/selectivity filter showed a 0.40‐fold depletion of stability‐neutral pathogenic variants (Bonferroni‐adjusted p = 4.3 × 10−7), whereas S4 voltage sensors showed a 2.32‐fold enrichment (adjusted p = 0.013); among the S4 variants, several occurred at canonical gating‐charge positions (e.g., R1639C, R1645Q, K1313T). VSD helices (S1–S3) showed weaker enrichment (1.73‐fold, adjusted p = 0.029). Figure S1 maps the variant‐level stability predictions across the full Nav1.1 linear topology. TABLE 1Domain‐stratified enrichment analysis of stability‐neutral versus strongly destabilizing P/LP variants.Functional domain n (P/LP)Near‐neutral (%) a Strongly destabilizing (%) b Enrichment ratio c Bonferroni‐adj p S4 voltage sensors7032.937.12.320.013*VSD helices (S1–S3)16023.135.01.730.029*Pore helices (S5–S6)25516.945.90.961.000P‐loops/selectivity filter2748.858.00.404.3 × 10−7 ***Loops/termini22621.244.21.260.713 Note: Ratio > 1.0 indicates over‐representation of stability‐neutral pathogenic variants; ratio < 1.0 indicates over‐representation of destabilizing variants. Fisher's exact test with Bonferroni correction for five comparisons (adjusted α = 0.01). *p < 0.05, **p < 0.01, ***p < 0.001. a Near‐neutral: FoldX ΔΔG within ±0.5 kcal/mol. b Strongly destabilizing: FoldX ΔΔG > 2.0 kcal/mol. c Enrichment ratio = (% of near‐neutral variants in domain/% of strongly destabilizing variants in domain). Domain‐stratified enrichment analysis of stability‐neutral versus strongly destabilizing P/LP variants. Note: Ratio > 1.0 indicates over‐representation of stability‐neutral pathogenic variants; ratio < 1.0 indicates over‐representation of destabilizing variants. Fisher's exact test with Bonferroni correction for five comparisons (adjusted α = 0.01). *p < 0.05, **p < 0.01, ***p < 0.001. Near‐neutral: FoldX ΔΔG within ±0.5 kcal/mol. Strongly destabilizing: FoldX ΔΔG > 2.0 kcal/mol. Enrichment ratio = (% of near‐neutral variants in domain/% of strongly destabilizing variants in domain). Exploratory domain‐stratified ROC‐AUC analysis, constrained by small benign samples per domain (range 2–40), showed a consistent pattern: the highest point estimates occurred in P‐loops (AUC = 0.823; 95% CI 0.726–0.910; 266 pathogenic, 6 benign) and pore helices (0.805; 0.737–0.873; 213 pathogenic, 2 benign), and the lowest in S4 voltage sensors (0.397; 0.293–0.500; 97 pathogenic, 2 benign). These estimates should be interpreted with caution given the limited precision at such small benign sample sizes. Generalization to ClinVar VUS and gnomAD population variants ClinVar VUS distributed continuously across the FoldX ΔΔG range (Figure 3) with mean 0.98 kcal/mol, significantly intermediate between B/LB and P/LP (Kruskal–Wallis p = 1.5 × 10−47). A fraction of 18.5% crossed the strongly destabilizing threshold (ΔΔG > 2.0 kcal/mol), highlighting these as candidate VUS for prioritized functional follow‐up—a directly actionable output of the analysis for clinical variant interpretation. Domain‐stratified enrichment of stability‐neutral pathogenic variants in S4 (2.32‐fold) and depletion in P‐loops (0.40‐fold) was qualitatively unchanged when computed on the expanded dataset. FoldX ΔΔG distribution across ClinVar P/LP, VUS, and B/LB. (A) Violin plots of per‐variant FoldX ΔΔG for the three ClinVar categories (P/LP n = 985; VUS n = 1275; B/LB n = 60); the 2.0 kcal/mol strongly‐destabilizing threshold is marked. VUS distribute continuously between B/LB and P/LP (Kruskal–Wallis p = 1.5 × 10−47; mean 0.98 kcal/mol). (B) Within‐category composition: Proportion of variants in stabilizing/near‐neutral (ΔΔG ≤ 0.5), mildly destabilizing (0.5 < ΔΔG ≤ 2.0), and strongly destabilizing (ΔΔG > 2.0) bins. Variants observed only in gnomAD with allele frequency ≥ 1 × 10−4 (“gnomAD‐common”, n = 13) provide a population‐level “likely‐benign” comparator that is independent of the ClinVar B/LB curation used for predictor training and exceeds the maximum credible allele frequency for a highly penetrant SCN1A pathogenic variant. The mean FoldX ΔΔG of gnomAD‐common variants was 0.14 kcal/mol—close to the ClinVar B/LB mean (0.31 kcal/mol) and significantly less destabilizing than ClinVar P/LP variants (Mann–Whitney p = 5.4 × 10−4, Cohen's d = 0.85; Figure 4). Augmenting the ClinVar‐curated B/LB set with gnomAD‐common as a ClinVar‐independent orthogonal benign reference (combined n = 73) yielded FoldX ROC‐AUC = 0.763 for P/LP discrimination, essentially identical to the B/LB‐only baseline (AUC = 0.760)—confirming that the pathogenic‐vs‐benign separation is not an artifact of ClinVar B/LB ascertainment. Population‐level orthogonal benign reference and predictor comparison. (A) FoldX ΔΔG distributions for ClinVar P/LP, ClinVar B/LB, and gnomAD allele‐frequency‐stratified subsets (common AF ≥ 1 × 10−4; rare 1 × 10−5 ≤ AF <1 × 10−4; ultra‐rare AF < 1 × 10−5). gnomAD‐common variants cluster near zero ΔΔG, behaving like ClinVar B/LB (mean 0.14 vs. 0.31 kcal/mol). (B) AlphaMissense pathogenicity score distributions for the same strata, illustrating the complementary predictor's pattern: High AM scores across P/LP, lower scores across gnomAD‐common, supporting the use of gnomAD‐common as a ClinVar‐independent likely‐benign comparator. Sensitivity to AlphaFold3 structural confidence To assess whether the observed associations were driven by structurally uncertain regions of the AlphaFold3 model, we performed sensitivity analyses restricted to variants at residues with high local model confidence (mutated‐residue pLDDT ≥ 70; n = 949: 918 pathogenic, 31 benign). The main pathogenicity and domain‐stratified trends were materially unchanged. Pathogenic‐benign separation remained significant for both FoldX (p = 1.19 × 10−5) and Rosetta (p = 1.02 × 10−3), with modest AUC decreases (FoldX 0.760 → 0.731; Rosetta 0.697 → 0.673) attributable to the halving of the benign sample (60 → 31) rather than loss of predictive signal. Domain enrichment patterns were not only maintained but strengthened: S4 near‐neutral enrichment increased from 2.32‐ to 2.60‐fold (Bonferroni‐adjusted p = 0.004), P‐loop depletion remained robust (0.44‐fold, adjusted p < 0.001), and VSD helix (S1–S3) enrichment remained significant (1.93‐fold, adjusted p = 0.009 vs. 0.029 in the full dataset). These results confirm that the core findings derive from structurally reliable regions and are not artifacts of low‐confidence AlphaFold3 modeling. Conformational‐state characterization of the AF3 model Geometric comparison with cryo‐EM Nav structures The AF3 activation‐gate minimum radius (2.0 Å) was narrower than a hydrated Na+ ion, consistent with a non‐conducting S6 bundle‐crossing. The cryo‐EM reference values were systematically smaller (7 K48 0.78, 6J8E 0.51, 7XVE 0.64, 6UZ3 0.61, 7DTC 0.25, 7XM9 0.75, 6AGF 1.35 Å), but several panel structures are pore‐blocker complexes (KIIIA in 6J8E; XEN907 in 7XM9)—in which the minimum lumen distance reflects the bound ligand rather than the protein gate—and 7 K48 carries a bacterial NavAb pore module; the metric therefore supports non‐conduction of the AF3 model without providing a quantitative ranking against the panel. The DEKA selectivity‐filter Cα geometry (DI–DIII 10.7/DII–DIV 11.8/DI–DII 5.6/DIII–DIV 6.8 Å) matched the Nav1.2–KIIIA reference 6J8E (10.8/11.9/5.9/7.1 Å) within ~0.3 Å on all six pairwise distances, indicating a canonical mammalian‐Nav selectivity‐filter backbone configuration. State assessment of the AF3 model S4 outermost‐arginine z‐positions were heterogeneous across the four VSDs (DI/DII−15.2/−13.1 Å, cytoplasm‐shifted; DIII/DIV+11.3/+13.9 Å, extracellular‐shifted), and the IFM–receptor‐pocket distance was 28.7 Å. Because matched S4 z‐projections were not available across the panel and no validated cross‐paralog receptor‐pocket definition was applied, these descriptors are reported on the AF3 model only. Taken together, the AF3 model presents as a non‐conducting structure with internally heterogeneous S4 positions across the four VSDs and does not uniquely match any single experimentally captured Nav state. Comparison with patch‐clamp electrophysiology data Stability predictions were compared with published patch‐clamp data for 64 epilepsy‐associated variants (45 from ClinVar, 19 additional; Table 2). All four functional classes showed elevated mean FoldX destabilization relative to benign controls (0.31 kcal/mol): complete LOF (n = 21, mean 3.07 kcal/mol), partial LOF (n = 20, 1.28), GOF (n = 12, 1.93), and mixed (n = 11, 0.82). FoldX ΔΔG did not, however, discriminate among these functional classes (Kruskal–Wallis H = 4.79, p = 0.188; all pairwise Mann–Whitney p > 0.09): the proportion that were strongly destabilizing (ΔΔG > 2.0 kcal/mol) was comparable across classes (27.3–57.1%, chi‐square p = 0.31), within‐class standard deviations (1.28–4.41 kcal/mol) exceeded between‐class mean differences, and both GOF and complete LOF variants spanned the full stability spectrum (GOF, R1648H 5.93 to T782I−2.44 kcal/mol; complete LOF, R101W 14.32 to G1674R−3.46). TABLE 2Stability prediction method comparison across functional classes for 64 SCN1A epilepsy‐associated variants.MethodComplete LOF (n = 21)Partial LOF (n = 20)GOF (n = 12)Mixed (n = 11)KW p a cLOF vs. GOF p b Pathogenicity AUC c FoldX (kcal/mol)3.07 ± 4.411.28 ± 1.281.93 ± 2.410.82 ± 1.900.1880.6940.736Rosetta Cartesian (REU)3.95 ± 4.912.96 ± 3.561.83 ± 5.48−0.30 ± 4.230.1290.2860.621RosettaMP (REU)9.31 ± 13.403.24 ± 11.933.50 ± 9.102.35 ± 4.660.1340.4210.640 Note: Values are mean ± SD. RosettaMP KW p and partial LOF values reported after exclusion of L479P (pLDDT 30.5; RosettaMP 100.97 REU); without L479P: partial LOF 3.24 ± 11.93, KW H = 5.58, p = 0.134. With L479P included: partial LOF 8.13 ± 24.74, KW H = 4.28, p = 0.232. All pairwise p‐values are uncorrected, with complete LOF vs. GOF designated as the pre‐specified primary comparison based on therapeutic relevance. FoldX: no pairwise comparisons between functional classes reached significance (all p > 0.09). Rosetta Cartesian: strongest pairwise separation was complete LOF vs. mixed (p = 0.033). RosettaMP: no pairwise comparisons reached significance (cLOF vs. GOF p = 0.421). a Kruskal–Wallis test across four functional classes. b Mann–Whitney U test. c ROC‐AUC for pathogenicity discrimination (64 pathogenic vs. 60 benign variants). Stability prediction method comparison across functional classes for 64 SCN1A epilepsy‐associated variants. Note: Values are mean ± SD. RosettaMP KW p and partial LOF values reported after exclusion of L479P (pLDDT 30.5; RosettaMP 100.97 REU); without L479P: partial LOF 3.24 ± 11.93, KW H = 5.58, p = 0.134. With L479P included: partial LOF 8.13 ± 24.74, KW H = 4.28, p = 0.232. All pairwise p‐values are uncorrected, with complete LOF vs. GOF designated as the pre‐specified primary comparison based on therapeutic relevance. FoldX: no pairwise comparisons between functional classes reached significance (all p > 0.09). Rosetta Cartesian: strongest pairwise separation was complete LOF vs. mixed (p = 0.033). RosettaMP: no pairwise comparisons reached significance (cLOF vs. GOF p = 0.421). Kruskal–Wallis test across four functional classes. Mann–Whitney U test. ROC‐AUC for pathogenicity discrimination (64 pathogenic vs. 60 benign variants). Rosetta Cartesian ΔΔG showed a monotonic but non‐significant destabilization gradient (complete LOF 3.95 ± 4.91 > partial LOF 2.96 ± 3.56 > GOF 1.83 ± 5.48 > mixed −0.30 ± 4.23 REU; Kruskal–Wallis H = 5.66, p = 0.129), with the strongest pairwise separation between complete LOF and mixed (p = 0.033). Supplementary membrane‐aware RosettaMP calculations likewise did not improve pathogenicity or functional‐class discrimination (Table 2). Within this limited dataset, domain position showed a suggestive association with functional outcome: no GOF variant was located in pore regions (0 of 17 pore variants), whereas all pore variants exhibited LOF (76.5%) or mixed (23.5%) profiles, and GOF variants were enriched in VSDs (OR = 4.80, p = 0.028). Per‐variant electrophysiology parameters, functional classifications, reported phenotypes, current density, domain assignments, and stability predictions for all 64 variants are provided in Supplementary Table S1. Phenotype severity Among 777 pathogenic variants with severity classification (588 severe, 189 mild), phenotype severity was not predicted by stability (Mann–Whitney p = 0.19), domain position (χ 2 p = 0.21), or their interaction (logistic regression LOO‐AUC = 0.439). The proportion of severe phenotypes was uniformly distributed across the ΔΔG spectrum (68.6–80.4% by quartile) and all domains (65.8–79.4%). Neither stability alone (LOO‐AUC = 0.222), domain position alone (LOO‐AUC = 0.404), nor their combination improved severity prediction beyond chance. Comparison with sequence‐based predictors and mechanistic complementarity For P/LP discrimination against the ClinVar B/LB + gnomAD‐common benign reference (P/LP n = 985 vs. benign n = 73), AlphaMissense achieved ROC‐AUC = 0.972, FoldX 0.763, and EVE 0.808. EVE was evaluated on the ClinVar B/LB subset only because precomputed EVE scores were not available for the gnomAD‐derived variants. Specificities for correctly classifying benign‐reference variants as non‐pathogenic at each predictor's Youden‐optimal threshold were FoldX 0.959, AlphaMissense 0.918, and EVE 0.828; at the FoldX ΔΔG > 2.0 kcal/mol threshold, FoldX specificity rose to 0.986. Among 1903 SCN1A variants called pathogenic by AlphaMissense (pathogenicity score ≥ 0.564, the default cut‐off for the “likely pathogenic” call), the AlphaMissense scores were near‐saturating in both P‐loop and S4 voltage‐sensor variants (medians 0.989 vs. 0.993), providing no resolution between these domain subsets. DISCUSSION Destabilization is significantly associated with pathogenicity Pathogenic SCN1A variants exhibited 8.4‐fold greater mean predicted destabilization than benign variants (FoldX: 2.61 vs. 0.31 kcal/mol, p = 1.27 × 10−11). Cross‐method concordance with Rosetta Cartesian ΔΔG (Spearman ρ = 0.660) supports prediction reliability. At ΔΔG > 2.0 kcal/mol, specificity reached 98.3%, indicating that strong destabilization provides high‐confidence supporting evidence for pathogenicity. The moderate sensitivity reflects a substantial fraction of pathogenic variants for whom the stability signal is not the dominant contributor—a finding that motivated the domain‐stratified analysis. Domain‐associated differences in predicted stability Within‐pathogenic enrichment analysis—which does not depend on benign variant counts — showed domain‐associated differences in the predicted stability profiles of pathogenic variants. Two patterns are central to the conclusions, and we describe them with explicit attention to what the data do and do not support. Stability‐neutral pathogenic variants were strongly depleted in the P‐loop/selectivity‐filter region (0.40‐fold, Bonferroni‐adjusted p = 4.3 × 10−7), indicating that pathogenic variants in the pore are predominantly destabilizing. Precise side‐chain packing in the pore is essential for ion permeation and structural integrity, and we interpret this pronounced destabilization as compatible with—though not proof of—misfolding, ER retention, and proteostatic degradation. 6 , 8 We emphasize that the present data demonstrate destabilization only; the downstream misfolding and degradation that proteostasis impairment would entail are inferred from this prior experimental evidence and remain to be confirmed by direct measurement. We therefore present this proteostasis‐impaired interpretation as the reading most consistent with the stability data—offered as complementary to, rather than a replacement for, the gating‐dysfunction mechanism through which SCN1A pathogenicity is conventionally understood—and as a hypothesis‐generating basis on which the analysis can flag a candidate proteostasis‐impaired subset for clinical prioritization and functional validation. S4 voltage sensors showed a 2.32‐fold enrichment of stability‐neutral pathogenic variants (adjusted p = 0.013), indicating that pathogenic variants in this region are frequently structurally tolerated. Among these, several occurred at canonical gating‐charge positions known to be electromechanical transducers, 26 and substitutions at such positions could in principle disrupt voltage sensing without perturbing protein stability 27 —a pattern compatible with, but not direct evidence of, gating dysfunction. The data presented here demonstrate only that these variants are not strongly destabilizing; they do not directly demonstrate gating dysfunction or distinguish it from other stability‐independent mechanisms (e.g., altered subunit interactions, post‐translational regulation). Confirmation of a gating‐dysfunction interpretation for any specific S4 variant would require direct functional measurement. The domain‐stratified patterns are therefore correlative. They flag a candidate proteostasis‐impaired subset—supported by a statistically robust depletion of stability‐neutral variants in the pore—alongside a complementary stability‐neutral subset whose mechanism cannot be resolved by stability prediction alone. The domain‐stratified AUC point estimates reported in earlier sections (P‐loop 0.823, S4 0.397) are constrained by small benign samples per domain (range 2–40) and should be interpreted as supporting trend information rather than precision estimates. Domain position provides probabilistic functional guidance Published patch‐clamp data for 64 variants indicated that stability alone cannot predict functional direction: the proportion exceeding the pathogenicity threshold was similar across all four functional classes (36.4–66.7%, chi‐square p = 0.34), and each class spanned the full stability spectrum. Within this limited dataset, domain position showed a suggestive association with functional outcome. No GOF variant was located in pore regions (0 of 17 pore variants), whereas all pore variants exhibited LOF (76.5%) or mixed (23.5%) profiles. GOF variants were enriched in VSDs (OR = 4.80, p = 0.028). These observations, if replicated in larger datasets, would suggest that the combination of domain position and stability may offer probabilistic guidance for functional inference. However, the small number of experimentally characterized variants per domain limits the generalizability of these patterns. Complementarity with sequence‐based predictors Sequence‐based predictors (AlphaMissense and EVE) returned higher nominal AUC than FoldX and Rosetta on the SCN1A categorical‐pathogenicity task; we caution, however, that AlphaMissense was trained on ClinVar P/LP and B/LB labels, so the apparent margin on this gene's labeled variants substantially reflects training‐set inclusion rather than independent generalization. The question with practical bearing on variant interpretation is therefore not which predictor returns the higher categorical AUC, but whether stability prediction provides a complementary mechanistic layer that AlphaMissense does not. Two lines of evidence in our expanded dataset support this complementarity. First, AlphaMissense and FoldX rank pathogenic variants by domain in opposite directions: AlphaMissense scores S4 voltage‐sensor pathogenic variants higher than P‐loop pathogenic variants (reflecting evolutionary constraint on gating arginines), whereas FoldX scores them lower (reflecting that gating‐charge substitutions can preserve fold while abolishing function). Second, within the pool of AlphaMissense‐pathogenic variants, the FoldX × domain‐position decomposition partitions variants into mechanistically distinct subgroups (proteostasis‐impaired pore‐region variants vs. gating‐dysfunctional S4 variants) that AlphaMissense alone does not resolve. The two predictors therefore answer different questions: AlphaMissense answers “is this variant pathogenic?”; stability analysis combined with domain context answers “if this variant is pathogenic, by what mechanism?” The latter question is the one that matters for mechanism‐based therapeutic stratification (pharmacological chaperone and/or ion channel modulator) and is not addressable by sequence‐based predictors alone. Therapeutic implications If the proteostasis‐impaired interpretation for pore‐region variants were experimentally confirmed, pharmacological chaperones that enhance folding efficiency 6 , 8 might increase functional channel expression in carriers of those variants. For the stability‐neutral S4 subset, no therapeutic claim can be made without independent functional characterization, since the underlying mechanism is not resolved by the present analysis. This distinction may be clinically important given that seizure control alone does not prevent cognitive decline in SCN1A‐related disorders 4 : if proteostatic degradation reduces Nav1.1 density in interneurons throughout development, this could impair neurodevelopment independently of seizures. In such cases, therapeutic approaches aimed at restoring protein stability might complement conventional seizure management by addressing the underlying molecular deficit. However, compound mechanism variants present challenges. R1648H illustrates this complexity: its extreme destabilization (ΔΔG 5.93 kcal/mol) combined with well‐documented persistent current 28 suggests that a single variant may simultaneously reduce channel density through proteostatic degradation and alter biophysical properties of channels that escape degradation. Rescuing folding of such variants could worsen outcomes if the rescued channels exhibit GOF properties. Additionally, phenotype severity was not predicted by stability, domain position, functional class, or their combination (all models LOO‐AUC < 0.44), likely reflecting both ClinVar ascertainment bias (76% severe) and genuine biological complexity where severity is incompletely explained by genotype alone—even identical SCN1A missense variants can show phenotypic discrepancy across carriers, and modifier‐gene/genetic‐background effects are a plausible contributor. 29 Mechanism‐based therapeutic stratification will therefore require variant‐specific functional characterization on a case‐by‐case basis. Limitations Several limitations should be considered. FoldX and Rosetta Cartesian ddG have benchmark prediction errors of ±0.46 kcal/mol 10 , 20 , 21 and were calibrated on soluble globular proteins, so per‐variant interpretation in Nav1.1 should be made with that caveat in mind; the optimal threshold (ΔΔG > 1.60 kcal/mol) was derived without cross‐validation. The AlphaFold3‐predicted template presents as a non‐conducting structure with heterogeneous S4 positions and is not uniquely assignable to a single Nav functional state, so ΔΔG on this template captures local stability perturbations on this conformer rather than mutation‐dependent shifts in state preference. Multi‐state ΔΔG analysis on Nav1.1 will require cryo‐EM coverage of additional functional states and/or molecular‐dynamics simulation with enhanced sampling. The ClinVar set is class‐imbalanced with a larger P/LP portion, which inflates positive predictive value estimates. We therefore report specificity at ΔΔG > 2.0 and the prevalence‐independent positive likelihood ratio as the primary discrimination metrics. The gnomAD‐common population variants provide a ClinVar‐independent benign comparator that partially offsets the imbalance. Patch‐clamp data were available for only 64 variants, and phenotype severity analysis was constrained by class imbalance (76% severe) and cross‐sectional ClinVar annotations that may not capture the full neurodevelopmental trajectory. CONCLUSIONS Computational stability analysis stratifies SCN1A pathogenic missense variants into a destabilized pore‐region subset—for which proteostasis impairment is a candidate, though unproven, mechanism complementary to gating dysfunction—and a structurally tolerated S4 subset whose pathogenicity is stability‐independent. The stratification is hypothesis‐generating rather than mechanism‐defining, nominating priority candidates for functional follow‐up, including the 18.5% of ClinVar VUS that cross the strongly destabilizing threshold. Direct surface‐expression and patch‐clamp validation, combined with multi‐state molecular dynamics simulation, will be needed to confirm these interpretations. AUTHOR CONTRIBUTIONS Y.S. conceived the study, performed computational analyses, and wrote the manuscript. E.K, N.K, and S.K. contributed to data interpretation and critically revised the manuscript. All authors approved the final version. FUNDING INFORMATION This work was supported by Korea University Ansan Hospital under grant number K2618071. The funding organization had no role in study design, data collection, data analysis, decision to publish, or preparation of the manuscript. CONFLICT OF INTEREST STATEMENT The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. ETHICS STATEMENT We confirm that we have read the Journal's position on issues involved in ethical publication and affirm that this report is consistent with those guidelines. Supporting information Figure S1. Nav1.1 variant stability landscape. FoldX ΔΔG for each calibration variant (P/LP n = 985; B/LB n = 60) mapped to amino‐acid position. Top: lollipop plot colored by structural domain (S4 voltage sensors, orange; VSD helices S1–S3, blue; pore helices S5–S6, green; P‐loops/selectivity filter, purple; loops/termini, gray); green diamonds, B/LB variants; domain brackets above denote the four homologous repeats (DI–DIV). To preserve legibility across > 1000 variants, only six variants central to the main‐text arguments are labeled: the stability extremes R101W (complete LOF, most destabilizing; ΔΔG ≈ 14.3) and T782I (GOF, stabilizing), the compound‐mechanism exemplar R1648H, and the stability‐neutral S4 gating‐charge variants R1639C, R1645Q, and K1313T. Variants with ΔΔG > 10 kcal/mol are clipped to the axis maximum (red triangles); the solid red line marks the strongly‐destabilizing threshold (ΔΔG > 2.0 kcal/mol). Bottom: per‐segment mean ΔΔG of P/LP variants (warmer = greater destabilization). C, C‐terminus; N, N‐terminus. Table S1. Integrated patch‐clamp electrophysiology parameters, functional classifications, reported phenotypes, domain assignments, and predicted stability effects (FoldX, Rosetta Cartesian) for 64 SCN1A epilepsy‐associated variants (45 from ClinVar, 19 additional) from published studies.