Nanoscale organization in the cell membrane dynamically modulates the biophysics of voltage-gated sodium channels Authors: Mikhail Tarasov, Madison Ammon, Jan Otto Wirth, Christopher Hampton, Zoja Selimi, Rengasayee Veeraraghavan, Przemysław B Radwański Published: 2026 ABSTRACT AbstractPrecise regulation of ion channel biophysics is an essential life process that governs electrical signaling in excitable tissues. Many ion channels, including voltage-gated Na+ channels (NaVs), exist in the membrane as clusters, which show distinct biophysical behavior not predicted by single-channel measurements. In both heterologous and native systems, we report that single-channel-based predictions significantly overestimated Na+ current (INa) amplitudes from multi-channel clusters. Computational modeling suggested that these observations could reflect interactions between adjacent channels, such as those recently reported between NaVs, and identified specific biophysical consequences thereof. This updated model not only accurately predicted behaviors observed from NaV clusters and consequent cellular physiology, but also suggested the possibility that clustered NaVs may respond differently to use-dependent pharmacological agents. Experiments validated the latter prediction and further identified modulation of clustering as an approach to correcting macroscopic electrophysiological dysfunction resulting from NaV defects linked to life-threatening arrhythmias and seizures. Thus, our study not only motivates a fundamental revision of how ion channels behave when clustered but also highlights resulting biophysical effects as important considerations for pharmacology and a potential therapeutic target to address human disease.Subject terms: Cardiovascular biology, Ion channels, Ion channels in the nervous system, Permeation and transport, Single-channel recording Introduction Voltage-gated sodium channels (NaVs) play an essential role in the rapid depolarization of excitable tissues such as the heart and the brain1. While individual NaV isoforms display biophysical properties tuned to their sites and modes of operation, all are capable of rapid activation in response to depolarization and rapid inactivation, enabling robust excitability and refractoriness of such tissues. Disruption of NaV gating, whether through inherited or acquired defects, underlies serious and often life-threatening pathologies ranging from seizures to cardiac arrhythmia2,3. Thus, understanding NaV biophysics is of paramount importance in healthcare. As reported for other channel types4, NaVs form clusters in the membranes of multiple cell types. To wit, NaV1.5, the predominant cardiac isoform, shows location-specific clustering in particular niches within cardiac myocytes5–9, while neuronal isoforms, such as NaV1.6, form clusters at neuronal sites, such as nodes of Ranvier and axon initial segments10,11. Differential NaV clustering critically regulates physiology, modulating initiation and propagation of action potentials (APs)12–15, and cluster dispersal is linked with serious dysfunction, such as cardiac arrhythmia6,7,16. Likewise, in failing cardiac myocytes, changes in late INa (INa-L) have been localized to niches where NaV clustering is altered17. However, the biophysical basis of these phenomena remains unclear, given the canonical paradigm in electrophysiology, which envisions macroscopic phenomena as resulting from the linear summation of individual channels’ behavior. Available evidence suggests that clustering-dependence of NaV biophysics may derive from homomeric interaction(s) between adjacent NaVs18–26. One previous study suggested that NaVs dimerize, enabling pairs to gate together, a phenomenon termed coupled gating21. However, neither this study21, nor subsequent efforts26,27, could link coupled gating with modulation of macroscopic Na+ current (INa) at the cellular level. These results, along with the unexplained observations described above, suggest that homomeric interaction(s) between NaVs may (i) be more complex than dimerization alone, and (ii) modulate function by tuning channel gating. Support for these ideas comes from the reduction in whole-cell INa resulting from the interaction of a NaV fragment with endogenous NaVs21, and a recent structural study suggesting state-dependent interaction between NaVs in the membrane25. Thus, we hypothesized that interactions between clustered NaVs might regulate channel biophysics and underlie emergent INa behaviors observed at cellular and macroscopic levels. In this work, using a combined experimental and modeling approach, we demonstrate that homomeric interactions between NaVs modulate stability and speed of inactivation, thereby altering INa in a cluster size-dependent manner. We further demonstrate crucial implications these effects hold for macroscopic physiology, and as a proof-of-principle, demonstrate how modulating NaV clustering can rescue a congenital NaV defect linked with serious cardiac and neurological disorders. In summary, our results prompt a fundamental rethinking of how ion channel biophysics are regulated in native contexts, and integrated from single molecules through macroscopic levels. Importantly, we identify modulation of clustering as a heretofore unexplored therapeutic approach for cardiac, neurological, and other disorders. Results Sodium channel clustering modulates current Cardiac NaVs (NaV1.5) cluster in the membrane, as do other channels and NaV isoforms, with multiple studies suggesting homomeric interactions18–26. However, the functional ramifications of such interactions remain elusive. Therefore, we undertook structural and functional studies in Chinese hamster ovary (CHO) cells stably expressing human NaV1.5. NaV1.5 clusters in CHO cells were visualized by minimal photon flux (MINFLUX) nanoscopy (Supplementary Fig. 1a, b, 2.4 nm resolution achieved in our samples). In these experiments, NaV1.5 was labeled using a fluorescently-tagged nanobody directed to a short BC2 peptide tag28 inserted into the C-terminus of NaV1.5 to minimize linkage error (Fig. 1a, b and Supplementary Fig. 2 shows consistency of biophysical properties between wild-type (WT) and BC2-tagged NaV1.5). Next, we used cell-attached patch clamp to compare NaV activity between membrane locations with single vs multiple channels. Based on ≥100 sweeps, single-channel patches displayed significantly more openings, and greater normalized late Na+ current (INa-L; referred to as persistent INa in neuroscience) relative to multi-channel patches (Fig. 1c–f). In order to understand how the presence of multiple channels in close proximity might alter local NaV activity at a given membrane location, we experimentally altered NaV1.5 membrane clustering. To this end, we inhibited NaV1.5 forward trafficking by incubating cells with a microtubule stabilizer taxol (TXL, 100 µM, 2–3 h)29,30. MINFLUX revealed a significant reduction in NaV1.5 cluster density (Fig. 1a, b) following trafficking inhibition. Single-channel behavior remained unaltered, as measured by peak open probability (Po; Supplementary Fig. 3), and local INa-L (Fig. 1c–e), as well as the decay of ensemble-averaged peak INa (Supplementary Fig. 4). However, at multi-channel sites, smaller clusters evidenced increased local INa-L (Fig. 1c–f), while displaying the anticipated reduction in ensemble-averaged peak INa (Fig. 2a, b). Together, these data suggest that reducing the membrane density of NaV1.5 can paradoxically enhance INa-L by diminishing NaV1.5 cluster sizes. NaV channel clusters stabilize late Na+ current (INa-L). a Confocal microscopy and MINFLUX (minimal fluorescence photon fluxes) nanoscopy of CHO cells expressing BC2-tagged NaV1.5 channels obtained under control or after reduction in NaV1.5 surface expression (100 µM taxol (TXL) treatment for 2–3 h). The color of each pixel in the MINFLUX images indicates the number of NaV1.5 channel localizations recorded at that location. b Reduced NaV1.5 surface expression reduces channel cluster density. c–f Functional results with single- (red) and multi-channel (blue) NaV1.5 current (INa) traces. c Representative current traces, and (d) ensemble average currents normalized to peak INa recorded under control (bright colors) or after reduction in NaV1.5 surface expression (faded colors) reveal (e) higher normalized ensemble average late INa (INa-L) when NaV1.5 surface expression is reduced. The number of cells for measurements in (d, e) is indicated in (d). Dashed boxes in (d) show enlarged current traces near baselines. The mean normalized to peak INa (Inorm) over all cells are shown. f Correlation between ensemble average INa-L and peak INa in control (circles) and after reduction of NaV1.5 surface expression (triangles). Data were fitted to a two-phase decay function (solid lines). ####p < 0.0001, by two-tailed χ2-test for numbers of measurements in all cluster density intervals shown in (b) data from 9 and 8 cells in control and reduced surface expression conditions, respectively. *q < 0.0293, Kruskal–Wallis test with the original false discovery rate (FDR) method of Benjamini and Hochberg for post hoc multiple comparisons; #p = 0.0115, by two-sided Spearman correlation test for control (Spearman correlation coefficients 0.34, black trace) and p = 0.2906 by two-sided Spearman correlation test for reduced surface expression (Spearman correlation coefficients 0.23, gray trace). In (e) box-and-whiskers plots show: box—25th (lower bound), 50th (center), 75th (upper bound) percentiles, whiskers—minima and maxima. Divergence between observed and predicted behavior of NaV clusters. a Ensemble average INa (left; mean ± standard error of the mean, S.E.M) either generated by stochastic simulations of independent NaV1.5 activity (black), or obtained experimentally from NaV1.5 multi-channel recordings under control conditions (dark blue), or after reduction of NaV1.5 surface expression (faded blue). Corresponding levels of functional NaV1.5 interaction quantified by Kullback-Leibler divergence (DKL, right; mean ± S.E.M) suggest that (b) interacting channels (dark blue) do not conform with predicted function (black) and (c) exponentially diverge from predictions with increasing cluster size (solid lines—best fit exponential growth curves; shaded areas—95% confidence intervals). ***p = 0.0003, two-sided Mann–Whitney test, **q = 0.0048 for control vs reduced expression and *q = 0.0106 for control vs model, q = 0.2193 for reduced expression vs model, Kruskal–Wallis test with the original FDR method of Benjamini and Hochberg for post hoc multiple comparisons. &&&&p < 0.0001, two-sided F-test for the null hypothesis that one function fits all datasets well. In (a–c) 49 stochastic simulations of 2—190 independent channels, patch-clamp recordings from 52 and 17 cells in control and reduced surface expression conditions, respectively. In (b) box-and-whiskers plots show: box—25th (lower bound), 50th (center), 75th (upper bound) percentiles, whiskers—minima and maxima. Divergence of NaV1.5 cluster behavior from single-channel-based predictions Next, we tested whether the biophysical properties of NaV1.5 clusters vary in a cluster size-dependent fashion, indicating the influence of inter-channel interactions. To this end, we compared experimental observations from clusters against statistical predictions of ion channel gating based on the binomial distribution expected for non-interacting channels27,31–36. Specifically, we compared statistical deviation between the observed and expected distributions of the number of open channels (k) over the entire experimental duration using Kullback-Leibler divergence (DKL; Supplementary Fig. 6)27,36–38. A model parametrized by non-interacting channels (Supplementary Fig. 7) in clusters displayed overall low DKL (<0.01, Fig. 2a–c, black). In contrast, experimentally recorded multi-channel clusters displayed significantly greater DKL compared to these model predictions (Fig. 2a–c¸blue). Importantly, the magnitude of divergence increased with increasing peak INa (Fig. 2c). This suggests that the interaction between NaVs increases in proportion to the increase in channel density. Furthermore, this notion is supported by conformity of divergence (DKL) observed in smaller NaV1.5 clusters (after TXL exposure; Fig. 2a–c¸ light blue) with the predictions of a non-interacting multi-channel model (Fig. 2a–c, black). Taken together, these findings demonstrate how the biophysical behavior of NaVs observed experimentally deviates from the classical predictions based on models of non-interacting channels. A model of NaV1.5 interaction predicts NaV1.5 behavior Given the degree to which the experimentally-observed behavior of multi-channel NaV clusters diverged from predictions from models of non-interacting channels, we sought to develop a model for clustered and interacting NaV1.5 channels. To this end, we began by building composite Markov models of two to four non-interacting channels, as previously described (Fig. 3a) in ref. 36. Given previous experimental observations25, we modeled interaction between channels as occurring only in the closed state. Since protein-protein interactions reduce the free energy relative to non-interacting states39, we parameterized this interaction as a reduction in the energy of the composite closed state, and the energy of transition to that state from composite states contaning open and/or inactivated states (Fig. 3a). Using these models, we compared decay kinetics and divergence (DKL) between ensemble averaged INa predicted by models of interacting and non-interacting channels. Decay kinetics were similar between the two models (Fig. 3b), just as has been observed experimentally from single channels and multichannel clusters (Supplementary Fig. 4), while divergence was higher for multichannel clusters in both experiments and simulations (Figs. 2a–c and 3c). Furthermore, simulations with the updated model predicted a reduction of ensemble-averaged peak INa in interacting NaV1.5 channels compared to non-interacting channels (Fig. 3b). Revising the NaV model to capture cluster behavior. a Models of interacting (orange) and non-interacting (black) channels. A simplified Clancy–Rudy single-channel model was employed to model the kinetics of single channels104. Composite models of 2–4 channels were constructed by the Cartesian graph product of single-channel models. Channel interaction was simulated by conservatively scaling transition rates (α and β) in the composite model, which reflects inter-channel interaction(s) in the resting state25. b In comparison with non-interacting channels (black), interacting channels (orange) are predicted to exhibit lower ensemble average peak INa, with the reduction being proportional to the number of interacting channels. No significant alterations in the time constants of ensemble average INa decay (τ) are anticipated. c With an increasing number of interacting channels, the level of divergence (DKL) from a binomial distribution in the number of open channels increases relative to the non-interacting ones. d The fraction of interacting channels in the fast inactivated state is lower than that of non-interacting channels (inset, blue), resulting in reduced response to use-dependent blockade, while (e) the fraction of interacting channels in the slow inactivated state is higher relative to non-interacting channels (inset), resulting in reduced ensemble average INa-L. Precise experimental validation of this prediction requires a reliable estimation of the total number of channels and their open probabilities (Po), which contribute to multi-channel recordings. Previous results40,41 and our evaluation of errors inherent to such estimates (Supplementary Fig. 5) highlight the perilousness of this task. Therefore, we instead used the variance-to-mean ratio (VMR)42 of open channels contributing to peak ensemble average INa observed in multi-channel clusters (Supplementary Fig. 8a–c) as a surrogate measure. VMR has been previously demonstrated to inversely correlate with single-channel Po38. Thus, the reduction in VMR observed when cluster density was experimentally reduced (Supplementary Fig. 8c) implies increased average Po in these smaller NaV1.5 clusters after TXL exposure, an observation further supported by our updated model (Fig. 3b and Supplementary Fig. 8a), which is also paralleled by predictions of a previously published model of Naundorf et al.43, albeit in the opposite direction (Supplementary Fig. 8b). The main limitation of the approach that relies on VMR is potential inaccuracy arising from estimation of sample mean and variance. Thus, conclusions based on VMR require additional validation. Therefore, we sought to experimentally validate our updated model’s predictions on how NaV1.5 cluster size modulates Po. To this end, we performed experiments using the open state-dependent NaV activator, veratridine44,45, which should increase both whole-cell late INa, as well as tail INa. While late INa represents whole-cell averaged late channel openings of veratridine-modified and unmodified channels, tail INa represents only the population of veratridine-modified channels, with Po that slowly subsides after returning to holding potential from the depolarizing pulses44. Veratridine-dependent tail INa inversely correlated with the peak INa density (Supplementary Fig. 9), suggesting that larger NaV1.5 clusters (reflected by larger whole-cell peak INa density) exhibit lower Po (reflected in the smaller tail INa) compared to the smaller clusters. These experiments demonstrated remarkable agreement with numerical simulations estimating the impact of NaV1.5 cluster density on the relationship between ensemble-averaged peak INa and Po (Supplementary Fig. 8d, e). Together, our results demonstrate that cooperativity within channel clusters must be accounted for in order to accurately predict the peak Po and biophysical behavior of clustered NaVs. Clustering alters NaV response to pharmacology Our updated model, which accounts for inter-channel interactions, predicts fewer channels existing in the fast inactivated state compared to the model of non-interacting channels (Fig. 3d). When we tested this prediction experimentally, NaVs in multi-channel clusters displayed faster recovery from fast inactivation relative to single-channels (Fig. 4), consistent with shorter occupancy in the fast inactivated state predicted by our model of interacting channels (Fig. 3d). Given the altered kinetics of clustered NaVs in comparison with non-interacting channels, we reasoned that clustering may modulate NaVs’ response to drugs. In particular, shorter residence of clustered NaVs in the fast inactivated state could alter their response to use-dependent drugs, which bind during this state46. Indeed, in our simulations, use-dependent block, modeled as previously described in ref. 47–53, had reduced impact on ensemble-averaged INa from interacting channels relative to the non-interacting channels (Fig. 3d, inset). In experiments, the use-dependent NaV inhibitor lidocaine (50 µM) displayed blunted potency against multi-channel clusters vs single channels (Fig. 5), consistent with our model predictions. These results lend further support to state-specific homologous interactions modulating the inactivation of clustered NaVs, which impact channel response to state-dependent block. NaV channel clustering enhances recovery from fast inactivation. Recovery from inactivation assessed in (a) single- and (b) multi-channel clusters and (right) corresponding ensemble average INa reveal (c) faster recovery from fast inactivation in multi-channel clusters. d Computational prediction of recovery from fast inactivation based on two interacting (orange) or non-interacting (black) channels. Solid lines represent a one-phase exponential decay model fit to the data. The dashed line indicates a 5 ms inter-pulse interval. *p = 0.0173, by two-tailed unpaired t-test. The number of cells for measurements in (c) is indicated in the figure panel. In (c), box-and-whiskers plots show: box—25th (lower bound), 50th (center), 75th (upper bound) percentiles, whiskers—minima and maxima. Use-dependent NaV blockade is less potent against clustered channels. a Single- and (b) multi-channel clusters before (top left) and after (bottom left) exposure to use-dependent NaV blockade with lidocaine (lido, 50 µM). The number of active channels is confirmed by fitting a mixture of Gaussian distributions (solid curves, right) to the observed probability density functions of open channels (vertical scale-bars reflect square roots of probability densities). c Ensemble average INa measurements reveal (d) reduced sensitivity of multi-channel clusters to lidocaine. In (d), left—mean ± S.E.M. of ensemble average INa after lido normalized to peak ensemble average INa in control. n = 3 and 4 cells for single- and multi-channel patches, respectively. *p = 0.0138 with two-tailed unpaired t-test. Clustering mitigates pathological INa-L in a NaV1.5 gain-of-function mutant Impaired slow inactivation of NaV1.5 results in pathologically enhanced INa-L and, consequently, cardiac arrhythmias in both inherited and acquired pathologies. In this context, our model makes an interesting prediction that clustering should increase the fraction of channels in the slow inactivated (SI) state, reducing local INa-L (Fig. 3e). Given that this prediction aligns with experimental observations, where INa-L from multi-channel NaV1.5 clusters is smaller vs single channels (Fig. 1c–e), we hypothesized that modulating clustering may be a viable approach to mitigating the proarrhythmic consequences of NaV1.5 inactivation defects. First, we incorporated closed state inter-channel interaction into a composite model of two NaV1.5 channels harboring an inactivation defect, ΔKPQ (Fig. 6a). Consistent with predictions for wild-type NaV1.5 (Fig. 3b, e), our model predicted that inter-channel interactions would lower peak INa and INa-L levels from ΔKPQ channels (Fig. 6b). Consistent with this prediction, our experiments in CHO cells expressing NaV1.5 ΔKPQ revealed smaller INa-L in multi-channel patches relative to single-channel patches (Fig. 6c). Taken together with predictions and observations with wild-type NaV1.5 channels (Fig. 1), our results suggest that clustering-dependent modulation of NaV biophysical properties occurs independently of mutations, provided they do not alter inter-channel interactions. Clustering mitigates late current in NaV1.5 gain-of-function mutant. a A simplified Clancy–Rudy model of a single ΔKPQ-NaV1.5 channel104 was used to construct composite models of two interacting (orange) or non-interacting (black) channels as in Fig. 3, which (b) revealed that interacting channels evidence lower ensemble average peak INa and INa-L. c Cell-attached patch clamp recordings in CHO cells transiently expressing ΔKPQ-NaV1.5 evidenced reduced ensemble average INa-L in multi-channel relative to single-channel recordings (left and center), and an inverse correlation between ensemble average peak INa and INa-L (right). The latter data, which combined single- and mult-channel recordings (black circles), were fitted to a two-phase exponential decay function (solid line). &&p = 0.0019, by two-tailed Mann–Whitney test. ****p < 0.0001 by two-sided Spearman correlation test (correlation coefficient 0.76 in c). n = 11 cells each for single-channel and multi-channel patch recordings in (c). In (c), box-and-whiskers plots show: box—25th (lower bound), 50th (center), 75th (upper bound) percentiles, whiskers—minima and maxima. Next, we sought to determine how the probability of inter-channel interactions varies with the membrane density of ΔKPQ-NaV1.5. To this end, we undertook Monte Carlo simulations of channel localizations over a 1 µm2 membrane, assuming previously reported diameters (20 nm in closed and 10 nm in open/inactivated states, respectively25) and membrane densities in the range from 25 to 2000 channels/µm2 for NaV1.5 (Fig. 7a, top)14. In line with previous estimates25, our model predicted interactions occurring between 40% of the open/inactivated channels, and 80% of channels in the closed state for a membrane density of 1200 channels/µm2, which is comparable to that reported in nodes of Ranvier14. These simulations suggest that, as the membrane density of ΔKPQ-NaV1.5 increases, peak INa should increase and INa-L should decline, respectively, in direct and inverse proportion to cluster size (a surrogate for the probability of inter-channel interactions; Fig. 7a, bottom). Clustering mitigates late current in NaV gain-of-function mutants. a Numerical simulations predict that increasing NaV channel density (blue and red are predicted channel diameters25) increases the probability of channel-channel interaction(s) (solid squares and circles). Increasing the level of interaction proportionally dampens the increase in the integral of ensemble average peak INa (open squares and circles) relative to non-interacting channels (open triangles; top). Increased NaV channel interaction and corresponding changes in ensemble average peak INa are negatively correlated to INa-L (blue and red circles; linear regressions with 95% confidence envelopes). Non-interacting channels (black circles) are expected to exhibit no relationship between ensemble average peak INa and INa-L (bottom). b Experimental evidence from ΔKPQ-NaV1.5 reveals an inverse relationship between whole-cell peak INa and INa-L. c This inverse relationship between whole-cell peak INa and INa-L was recapitulated in measurements from a neuronal NaV isoform harboring a pathogenic mutation, N1768D-NaV1.6. Solid lines in (a–c) represent the best linear regression, and areas between dashed lines represent 95% confidence intervals. ****p < 0.0001 and **p < 0.01, by two-sided Spearman correlation test (correlation coefficients are 0.996 for both 20 nm and 10 nm in (a); while correlation coefficient and p value are 0.71 and 0.003, respectively in (b), and 0.80 and 0.0014, respectively in (c). n = 6 deterministic simulations in each condition (a), n = 16 cells (b), and n = 13 cells (c). Increased clustering rescues pathogenic phenotypes resulting from dysfunctional NaV gating and translatability of cooperative gating to other NaV isoforms The results described above suggest that modulating clustering may represent a generalizable approach to mitigating functional consequences resulting from NaV gating defects. To experimentally evaluate this, we further investigated the influence of clustering on ΔKPQ-NaV1.5 behavior in CHO cells. In these experiments, we observed significant inverse correlation between INa-L and peak INa at both the cell-attached (i.e., local) and whole-cell levels (Fig. 7b), paralleling the cooperative behavior observed of WT NaV1.5 on cell-attached (Fig. 1e, f) and whole-cell levels (Supplementary Fig. 10a). We then tested the generalizability of this approach to other NaV isoforms. This approach revealed negative cooperativity in the form of an inverse relationship between INa-L and peak INa in WT NaV1.6 (Supplementary Fig. 10b, c) and NaV1.6 variant harboring the N1768D gain-of-function defect (Fig. 7c)54. These results, taken together, argue that modulation of clustering is a generalizable means to mitigating the pathological consequences of gain-of-function defects in NaVs independent of the isoform. Based on the aforementioned principle, we reasoned that pharmacological enhancement of NaV1.5 channel clustering in cardiac myocytes from transgenic mice harboring ΔKPQ-NaV1.5 (ΔKPQ) may help mitigate pathological INa-L and the proarrhythmic consequences thereof. To this end, we treated ΔKPQ cardiac myocytes with SB216763 (SB2, 5 µM, 2–3 h), a drug reported to enhance membrane densities of NaV1.555. MINFLUX revealed greater numbers of NaV1.5 cluster localizations (cluster mass) and numbers of NaV1.5 channels per unit area in clusters (cluster density) in SB2-treated ΔKPQ cardiac myocytes relative to vehicle-treated ΔKPQ myocytes (controls; Supplementary Fig. 1b, c and Fig. 8a, b). Functionally, SB2 significantly reduced whole-cell INa-L relative to untreated ΔKPQ myocytes (Fig. 8c) without impacting peak INa (Supplementary Fig. 11). This reduction in INa-L translated into shorter APs (Fig. 8d, e) and abrogated early and delayed after-depolarizations, consistent with reduced arrhythmia potential (Fig. 8f, g). Given the history of clinical challenges posed by direct pharmacological modulation of NaV gating, these results highlight NaV clustering as a potentially attractive target for pharmacological modulation to prevent organ-level dysfunction, such as seizures and arrhythmia. Increased clustering rescues cellular arrhythmia in a murine model with dysfunctional NaV gating. a MINFLUX nanoscopy of immunolabeled NaV1.5 channels in ΔKPQ myocytes under control conditions (left) and with enhanced NaV1.5 surface expression (middle). Right: enlarged views of regions highlighted by dashed boxes. The color of each pixel in the MINFLUX images indicates the number of NaV1.5 channel localizations recorded at that location. Cardiac myocytes are contoured with dashed gray lines. b Distribution of NaV1.5 channel cluster density. Enhancing NaV1.5 surface expression (blue) increased cluster density relative to untreated ΔKPQ myocytes (red). c Representative whole-cell INa from ΔKPQ cardiomyocytes demonstrates that increasing NaV1.5 clustering mitigates INa-L. d Action potential (AP) recordings with (bottom) and without (top) early (EADs) and delayed after depolarizations (DADs). Black vertical lines above the AP recordings mark the electrical stimuli. Pacing cycle length was 3 s for ΔKPQ with enhanced NaV1.5 surface expression. Due to a high frequency of EADs in untreated ΔKPQ cardiomyocytes, which prevented measurement of the AP duration (APD), the pacing cycle length for these myocytes was 20 s. Brackets and red asterisks indicate EADs, red arrow - DAD. e–g Increasing NaV1.5 clusters mitigates APD at 90% of repolarization (APD90) prolongation (e), EADs (f), and DADs (g) in ΔKPQ cardiomyocytes. #p = 0.0238, by two-tailed χ2-test for numbers of measurements in all cluster density intervals shown in (b); &&&p < 0.009, by two-tailed unpaired t-test; *p = 0.0192, **p = 0.007, ***p = 0.0003, by the two-tailed Mann–Whitney test. Data from 4 and 5 cells, respectively, isolated from a single mouse for ΔKPQ and ΔKPQ with enhanced NaV1.5 surface expression (SB2) in (b). Numbers of cells, which were obtained from 2 mice for ΔKPQ and 3 mice for ΔKPQ with enhanced NaV1.5 surface expression (SB2), are shown below the corresponding box and whiskers plots in (c, e–g). In (c, e–g), box-and-whiskers plots show: box—25th (lower bound), 50th (center), 75th (upper bound) percentiles, whiskers—minima and maxima. Discussion Over the last half century, single-channel biophysical studies have dramatically deepened our understanding of cellular and organ -scale electrophysiology. Despite these advances, sodium channels continue to exhibit unexplained behaviors at the cellular scale21,43,56–58, hinting at emergent behavior driven by mesoscale phenomena. Importantly, a lack of clear understanding of such mechanisms poses a critical barrier to effective treatment of disorders impacting excitable tissues such as the heart and the brain9,17. A plausible explanation supported by mounting evidence implicates interaction(s) between sodium channels in the plasma membrane in modulating sodium currents at the cellular scale18,19,21–26. However, early attempts to reveal the functional effects of inter-channel interactions proved inconclusive21,27,36, motivating continued search for an updated, unifying theoretical framework for ion channel physiology. We undertook multiscale experimental and modeling studies of how clustering alters the biophysics of NaVs across single molecules through whole-cell levels, and the consequences for health and disease at the whole-organ level. Specifically, we demonstrate that homologous interactions modulate sodium channel kinetics in previously unanticipated ways, thus enabling nanostructure (clustering) to influence excitability and refractoriness from the single molecule to the whole-cell levels in ways not predictable from single-channel properties. Beginning with simple heterologous systems expressing a single sodium channel type, we systematically progress to more complex, terminally differentiated native cells (cardiac myocytes) to demonstrate that clustering tunes sodium channel biophysics, particularly inactivation (the key determinant of potentially pathogenic late activity), determining the difference between health and disease even in the presence of genetic defects. We further demonstrate how clustering can alter NaVs’ response to pharmacology and, most importantly, how clustering can be manipulated to correct potentially life-threatening defects. Our study was motivated by our initial observation that single NaVs are more likely to open repeatedly compared to NaV clusters, with the result that single NaVs contribute more to late currents compared to NaV clusters. To address the mechanism underlying this phenomenon, we assessed the divergence of experimentally observed NaV1.5 gating from a standard model assuming independently operating channels. Our modeling and experimental data, taken together, suggest that homomeric closed-state-specific interaction between NaVs in channel clusters underlies their observed noncanonical behavior. Mainly, interactions between clustered channels decrease occupancy of the open and fast inactivated states, while increasing occupancy of the SI state. This results in multiple biophysical effects, including reduced late sodium current, simultaneously resulting in previously unexplained physiology and pathophysiology, while also providing avenues for therapy (Fig. 9). Homomeric interaction(s) modulate voltage-gated sodium channel open probability within clusters. Negative cooperativity of interacting NaVs channels. Closed state channel interaction increases transition to the closed state (green arrows). In turn, this stabilizes the non-conductive state, thereby reducing open channel probability and reducing INa-L. Elements used in this figure were downloaded from https://smart.servier.com/ and licensed under CC-BY 3.0 Unported https://creativecommons.org/licenses/by/3.0/ and CC-BY 4.0 Unported https://creativecommons.org/licenses/by/4.0/. Non-canonical behavior of clustered NaVs: state-dependent interaction Canonically, regulation of ion channel biophysics has been assumed to occur within each channel, independent of interactions with other channels59,60. However, mounting evidence challenges this view and motivates a rethinking of mechanisms. In pyramidal neurons, for example, the axon initial segment has a lower sodium channel density relative to the soma and yet produces a larger INa-L61. Likewise, in cardiomyocytes, an acute reduction in NaV1.5 membrane density can paradoxically increase INa-L29. When comparing different regions of the heart, atrial myocytes display higher NaV1.5 membrane density62 but lower INa-L amplitude63 relative to ventricular myocytes64. Consistent with this non-monotonic relationship between NaV1.5 membrane density and INa-L amplitude, an inactivation defect in NaV1.5 (ΔKPQ) prolongs APs less in atrial myocytes65 than in ventricular counterparts66,67. Taken together, these results challenge the canonical paradigm, which envisions macroscopic currents as reflecting the linear summation of single-channel conductances. Instead, they reveal an inverse relationship between NaV surface density and INa-L amplitude. The potential for functional interactions between NaVs was suggested as early as the 1980’s. Kiss and Nagy reported that the binomial distribution expected for non-interacting channels failed to predict experimentally observed numbers of open sodium channels33. Although inconclusive about the mechanisms underlying this effect, their study suggested the possibility that adjacent channels may interact and therefore, deviate from single-channel-like behavior. Consistent with this observation, but in diametric opposition to canonical expectations, Neumcke and Stampfli subsequently reported an increase in Na+ conductance when the number of available channels was reduced, suggestive of negative cooperativity between the channels32. Our study expands on these observations and provides a conceptual framework for negative cooperativity between NaVs, where increased interaction between clustered channels promotes stability of the non-conductive state and thereby reduces Po. Notably, this prediction from our updated model of interacting NaVs was validated by experimental assessments of peak and tail INa in the presence of the open state-dependent NaV activator, veratridine. Likewise, Naundorf et al. demonstrated variability in the timing of neuronal excitation, which could not be predicted by the Hodgkin and Huxley model, and suggested the possibility of cooperative NaV gating43. In line with these observations, Clatot et al.21 reported that in membrane patches containing two channels, they more often observed both channels open than they did only one. Further, they reported abrogation of this effect secondary to reduced interaction of NaV1.5 with 14-3-3 regulatory proteins. While the authors interpreted these results as suggestive of dimerization and consequent coupled gating through open-state dependent interactions21, subsequent numerical analysis revealed that observed numbers of open channels deviated from single-channel-based predictions, but not to the full extent expected with coupled gating36. Indeed, Clatot and colleagues reported no effect of 14-3-3 binding on whole-cell peak INa nor any difference in ensemble average current amplitudes between patches with and without coupled gating21. In the light of subsequent studies, where 14-3-3-dependent coupled gating was not detected26,27, observed non-canonical NaV behaviors cannot be explained well by coupled gating and open-state-dependent interactions. On the other hand, our findings point to clustering-dependent negative cooperativity, which reduces Po. Thus, previous reports and our observations here motivate questions about the role of closed-state-dependent interactions between clustered NaVs in determining macroscopic sodium current biophysics. Lastly, previous in silico studies36 suggested that interaction between NaVs in the inactivated state results in negligible changes in transient gating properties. That said, stabilization of the inactivated state should reduce late INa, prolong recovery from inactivation, and enhance the potency of inactivated state-dependent NaVs blockers. In our study, we observed a NaV cluster-dependent reduction in late INa, which is consistent with potential inactivation state interaction. However, recovery from inactivation in clustered channels was hastened, and the inactivated state-dependent sodium channel blocker, lidocaine, demonstrated reduced potency relative to single non-interacting channels, which diverges from predictions based on interactions in the inactivated state. Additionally, inactivation-deficient NaV variants retained the inverse relationship between peak INa and late INa as observed in WT channels. One would expect this to be altered if the variants disrupted state-dependent interaction(s). Thus, these results suggest that interaction(s) during the inactivated state may have a negligible contribution to the cooperativity between NaVs we report here. Function follows form: what gating reveals about interactions between clustered NaVs As noted above, clustered NaVs display reduced occupancy of the open and fast inactivated states and increased occupancy of the SI state relative to single NaVs. These suggest interactions between NaVs within a cluster may occur in a state-dependent fashion, an idea also supported by recent atomic force microscopy studies25. Specifically, structural evidence indicated interactions between channels in the closed state occurring via voltage sensor domains (via sites blocked by the pore domain in other states). These findings offer a structural explanation for our results, specifically implicating more stable slow inactivation of clustered NaVs. Given that such interactions between channels in non-conductive states alter the physical energy barrier for the transition to conductive states, it follows that such state-specific interactions can exert pervasive biophysical effects. We further note that we observed significant interaction between NaV1.5 predominantly in membrane patches containing ten or more channels (Fig. 2c). This suggests that any potential influence from coupled gating eludes detection, given the low probability of inter-channel interaction within a membrane patch containing just two channels. This argument is further underscored by our MINFLUX results (Fig. 1a, b), where single isolated localizations and larger clusters are more prevalent. Additionally, we observed sodium channel biophysics deviating from single-channel predictions as a continuous function of cluster size (Fig. 2c), indicating more complex and sweeping roles for inter-channel interactions. Even while questions remain about the precise nature of interactions between clustered NaVs, our data suggest that macroscopic electrophysiology is sensitive to dynamic changes in NaV clustering independent of NaV isoform (Supplementary Fig. 10). Indeed, when we enhanced NaV1.5 inter-channel interactions, we observed substantial shifts in cellular physiology. Specifically, this intervention mitigated late INa phenotypes in cardiac myocytes from mice harboring a proarrhythmic defect, NaV1.5-ΔKPQ (Fig. 8), as predicted by our model incorporating closed-state interactions (Figs. 3 and 6). These results highlight the translational promise of targeting NaV clustering to correct aberrant biophysics. Sodium channels are part of a multiprotein macromolecular complex whose components range from cell adhesion/junction proteins (such as N-cadherin, NaVβx, and Connexin 43) to intracellular interacting/regulatory proteins (such as calmodulin and 14-3-3). These can modulate the probability of interactions between NaVs, which scale with cluster size as evidenced, for instance, when NaVβx are co-expressed with NaV pore-forming subunits68. Connexin 43 gap junctions can also provide scaffolding for NaV cluster assembly by forming perinexal electrical synapses between cells, where sodium channel alpha and beta subunits are enriched6,69. Notably, the organization and function of such nanodomains have been previously modulated using modulatory peptides targeting NaVβ1-mediated adhesion6,68, highlighting these auxiliary subunits as an attractive potential therapeutic target. On the other hand, important mechanistic questions remain to be answered about the local regulation by passive biophysics and active processes (most notably regulated intramembrane proteolysis70,71) of these NaV cluster-rich nanodomains, the pursuit of which poses significant technical challenges. Looking beyond NaVβ subunits, cytosolic proteins, which may potentially facilitate interactions between NaVs, represent more tractable avenues for modulating NaV cooperativity. The aforementioned 14-3-3, thought to modulate biophysical coupling between NaVs21, is one such target21,26,27. Another is calmodulin, which regulates many ion channels, facilitates coupling of CaV1.2, and may influence NaV clustering72. Impairment of NaV-calmodulin interaction altered NaV surface expression, clustering, and channel function73,74. Intriguingly, calmodulin has been reported to facilitate interaction between two isolated NaV C-terminal domains in vitro75; however, calmodulin-mediated interaction between full-length NaV proteins has yet to be demonstrated. In this context, proteins such as fibroblast growth factor homologous factors (FHFs), which modulate NaV function in a calmodulin-dependent manner60,76 also merit further attention. Overall, our results here motivate a fresh look at the local proteome surrounding NaVs to elucidate biophysical and biological control and modulation of clustering. Translational perspective: manipulating clustering to improve human health In excitable tissues, such as the heart and the brain, NaVs are known to cluster within functionally specialized niches, where they perform diverse and vital physiological functions. Disruption of such clusters, whether in animal models7,16 or in patients77, has been linked with potentially life-threatening dysfunction at cellular through organ scales. Our results, in this context, raise the possibility that differential and dynamic cluster organization at such sites may enable localized fine-tuning of electrophysiology. Perhaps, even more important is the possibility that aberrant electrophysiology at macroscopic scales could be corrected by manipulating NaV clustering. Thus, we explored the idea of therapeutically enhancing NaV clustering and found that this approach effectively suppressed cellular precursors of cellular arrhythmia in mice harboring a NaV1.5 defect associated with deadly arrhythmia in humans. This observation parallels findings obtained with other NaV isoforms, where glycogen synthase kinase 3β (GSK3β) inhibition increased expression of NaV1.6 and NaV1.2 through post-translational modification of these channels78,79. On the other hand, the detrimental effects of acute GSK3β inhibition in human hearts have also been reported80. Thus, GSK3β requires in-depth analysis of safety and efficacy as a potential therapeutic target. However, our proof-of-concept results highlight clustering enhancement as a potentially valuable therapeutic avenue for correcting neurological, cardiac, and other issues arising from NaV dysfunction, particularly where conventional therapies fall short on efficacy and/or safety. Limitations Current methods preclude direct orthogonal validation of the exact number of ion channels within a given membrane patch, concurrent with electrical interrogation via patch clamp. However, we sought to mitigate this using unpaired MINFLUX nanoscopy measurements to compare NaV cluster sizes observed optically with electrophysiologic readouts. Furthermore, limited precision in counting ion channels within a membrane patch does not preclude comparison of current properties between single-channel and multi-channel patches, as done in this study. Although MINFLUX achieves high precision in localizing NaVs, localization precision is limited by linkage error. While the use of the BC2 tag limits linkage error to a few nanometers, it precludes the angstrom precision necessary to relate functional observations to exact intermolecular distances. Thus, the development of methods that rely on direct labeling of ion channels with fluorophores could overcome this limitation. Another major technical limitation arises from the inability of current technologies to directly quantify inter-channel interactions in situ. Thus, the choice of the single-channel model and the magnitudes of energy changes in composite models of interacting channels may influence the precision of model predictions. However, qualitative predictions from our models demonstrate agreement with multiple lines of experimental evidence, bolstering confidence in our mechanistic findings. Our studies here, performed in the simplified conditions afforded by a heterologous model, motivate further studies in more complex and diverse settings, such as those present within terminally differentiated cells like neurons and myocytes. Nevertheless, our work identifies negative cooperativity resulting from homomeric interactions as a factor modulating NaV biophysics regardless of cellular context. Methods Cell culture and transfection A CHO cell line stably expressing human NaV1.5 or NaV1.6 (Mayflower Bioscience, vendor for B’SYS GmbH cat. # BSYS-NAV1.5 DUO-CHO-C and BSYS-NAV1.6-CHO-C, respectively) was cultured in F-12 medium with glutamine (ThermoFisher Scientific) supplemented with 10% (v/v) FBS (Millipore Sigma). G-418 sulfate (250 μg/mL) and 1 μg/mL puromycin (both from ThermoFisher Scientific) were added to the medium to select for NaV1.5-expressing cells, and hygromycin b (500 μg/mL) for NaV1.6. Cells were cultured at 37 °C in 5% CO2 in a humidified atmosphere. A blank CHO cell line (CHO-K1, ATCC cat. # CCL-61) was cultured under the same conditions but without antibiotics. Mouse cDNA for WT SCN5A (NCBI Nucleotide database accession number: XM_006511997.1) and BC2-tagged (PDRKAAVSHWQQ; inserted after p.P1966)28 Scn5a were synthesized de novo and separately cloned into pcDNA3.1(+) P2A-eGFP plasmids (GenScript). These plasmids were then separately transfected into CHO-K1 cells (ATCC) in 24-well plates with 500 ng of DNA per well using Lipofectamine 3000 (ThermoFisher Scientific) according to the manufacturer’s protocol. GFP-positive cells were used for assays within 48−96 h post-transfection. Human SCN5A cDNA (NCBI Nucleotide database accession number: NM_198056.2) was cloned into pcDNA3.1(+) P2A-eGFP plasmid, and directed mutagenesis was used to make a deletion of codons 1505–1507 encoding the amino acids KPQ (GenScript)66,81. This ΔKPQ-NaV1.5 expressing plasmid was then transiently transfected into CHO-K1 cells in 24-well plates with 500 ng of DNA per well using Lipofectamine 3000 (ThermoFisher Scientific) according to the manufacturer’s protocol. As above, GFP-positive cells were used for assays within 48−96 h post-transfection. Human SCN8A cDNA (NCBI Nucleotide database accession number: NM_014191.4) incorporating the c.5302 A > G (p.N1768D) mutation54 was synthesized de novo and cloned into pcDNA3.1(+) P2A-mCherry plasmid (GenScript). This N1768D-NaV1.6 expressing plasmid was transiently co-transfected with NaVβ1 expressing plasmid (SCN1B, NCBI Nucleotide database accession number: NM_001037.5, cloned into pcDNA3.1(+) P2A-eGFP (GenScript)) into CHO-K1 cells in 24-well plates with 500 ng of each plasmid per well using Lipofectamine 3000 (ThermoFisher Scientific) according to the manufacturer’s protocol. Cells exhibiting both GFP and mCherry fluorescence were used for assays within 48–96 h post-transfection. Cells were detached by trypsin-EDTA solution (ThermoFisher Scientific, 0.25% for 5 min), resuspended in the culture medium, and plated on glass coverslips for experiments. Taxol (TXL, 100 µM, Sigma-Aldrich) or DMSO (vehicle, 0.5%, Sigma-Aldrich) was added to cells in culture 2 h before detachment and the start of the experiments. Mouse models Transgenic mice harboring a knock-in heterozygous deletion of amino acid residues 1505–1507 in the NaV1.5 channel (ΔKPQ, obtained from Vlaams Instituut voor Biotechnologie [VIB])66 were maintained on a 129S6/SvEvTac (Taconic, model#:129SVE) background. Mice were housed in a vivarium with a standard 12/12 light-dark cycle, a temperature range of 68–79 °F, and humidity between 30% and 70%. They received standard chow (Teklad Irradiated LM-485 Mouse/Rat Diet #7912, Envigo) and water ad libitum. Mice of both sexes were used for our studies at 12–36 weeks of age. Cardiac myocytes isolation Ventricular cardiomyocytes were isolated as previously described in refs. 63,74,82. Briefly, mice were anesthetized with 5% isoflurane mixed with 100% oxygen (1 L/min), and once a deep level of anesthesia was confirmed, hearts were rapidly excised and submerged in ice-cold Ca2+-free Tyrode’s solution consisting of (mM):133.5 NaCl, 4 KCl, 1 MgCl2, 10 glucose, 10 HEPES, and the pH was adjusted to 7.4 with NaOH. Subsequently, the aorta was cannulated using a blunted 24-gauge needle, transferred to a Langendorff apparatus, and perfused with Ca2+-free Tyrode’s solution at 37 °C to wash out the remaining blood. Next, the heart was perfused with Ca2+-free Tyrode’s solution containing Liberase TH (Roche). The heart was then removed from the perfusion system, the ventricles isolated, minced in Tyrode’s solution containing 2% BSA (Millipore Sigma), dispersed by gentle agitation, and filtered through a nylon mesh. Cardiomyocytes were then resuspended in low-Ca2+ Tyrode’s solution consisting of (mM): 133.5 NaCl, 4 KCl, 1 MgCl2, 0.1 CaCl2, 10 glucose, and 10 HEPES, and the pH was adjusted to 7.4 with NaOH. Isolated cardiomyocytes were incubated at room temperature in low-Ca2+ Tyrode’s solution with either 5 µM of SB216763 (SB2, Sigma-Aldrich), or 0.05% DMSO (Sigma-Aldrich) as a control, within 2–3 h before the commencement of experiments55. MINFLUX data acquisition and analysis CHO-K1 cells (ATCC, cat # CCL-61) expressing BC2-tagged NaV1.5 were grown on acid-washed #1.5 glass coverslips and fixed in 2% paraformaldehyde (PFA) in PEM buffer83 (5 min at room temperature (RT)) followed by a PBS wash (3 × 10 min at RT). They were then immunolabeled using an Alexa 647-labeled nanobody directed against the BC2 tag84 (1:800; Chromotek, #EBAF6467-50). Murine ventricular cardiomyocytes, isolated as described above, were immunolabeled for NaV1.5 as previously described in refs. 62,85 using our high fidelity custom antibody against NaV1.5 (1:100; custom rabbit polyclonal anti-NaV1.5)6 along with connexin43 (Cx43; mouse monoclonal antibody, dilution 1:250, EMD Millipore Corp., cat. #: MAB3067), or N-cadherin (N-cad; mouse monoclonal antibody, dilution 1:100; BD Biosciences, cat. #: 610920). Secondary immunolabeling of cardiomyocytes was performed using goat anti-rabbit secondary antibody conjugated to Flux 640 (1:2000; Abberior, cat. #: FX640) and goat anti-mouse secondary antibody conjugated to Alexa 568 (1:4000; ThermoFisher Scientific Cat #: A-11004). All MINFLUX samples were finally incubated for 10 min with 100 µL of saline buffer containing gold nanobeads (BBI Solutions), which provided fiducials for drift correction. MINFLUX data were recorded on a commercial Abberior Instruments MINFLUX setup (Abberior Instruments GmbH), similar to the one reported by Schmidt et al.86. The microscope was equipped with a 60×/1.4 NA magnification oil immersion lens, a 640 nm continuous-wave (cw) laser for both confocal and MINFLUX imaging, both a 488 nm cw and a 560 nm pulsed laser for confocal imaging, as well as a 405 nm cw laser for controlling the blinking speed, and a 980 nm infrared laser for actively stabilizing the sample. In contrast to the microscope described by Schmidt et al., the microscope used in this study employs two avalanche photodiode detectors with spectrally tunable detection. MINFLUX imaging was performed as follows: Measurements in CHO cells were performed using the standard 3D imaging sequence (version Nov2022, details see Supplementary Table 2) starting at 400 μW excitation power. Fluorescent signal was detected between 650 nm and 750 nm and at a pinhole setting of 0.60 AU. Measurements in murine ventricular cardiomyocytes were mostly performed by a shortened 2D imaging sequence (iteration 0–3 of the standard 2D tracking sequence, version Nov2022, details see Supplementary Table 1) starting at 160 μW excitation power. Fluorescence signal was detected between 650 nm and 750 nm and at a pinhole setting of 0.43 AU. Three measurements were performed using the 3D imaging sequence described for CHO cells at a pinhole setting of 0.43 AU. All samples were imaged either in a GLOX buffer with methyl ethyl ammonium (MEA) at a final concentration of 10–30 mM or using the Everspark 2.0 (Idylle Labs, Paris, France) buffer diluted in PBS to a final MEA concentration of 15–20 mM. The sample position was actively stabilized on the back-scattered light of gold beads illuminated with 980 nm in widefield mode. The Abberior Instruments Imspector software (v16.3.15645-m2205) with MINFLUX drivers was used to operate the system. Individual MINFLUX localizations (molecule position estimated by the final MINFLUX iteration) of every recorded single molecule blink were converted into images with Gaussian rendering, using a pixel size of 1 nm and a Gaussian blurring with a full width at half maximum of 4 nm. No filtering (by density or otherwise) was applied. Images thus generated show 4 pixel intensities corresponding to 0, 1, 2, and 3 localizations (respectively colored black, red, yellow, and white). Object-based segmentation by K-nearest neighbors (KNN) clustering was performed, enabling cluster-wise assessment of density (# of localizations/area). These data were then grouped, enabling comparison between study groups by χ2-test. Details of the used MINFLUX sequences are shown in Supplementary Tables1 and 2. Cell-attached patch clamp Cells were plated on glass coverslips in the electrophysiology chamber 10–20 min prior to the start of experiments to ensure attachment. During attachment, cells were maintained in the standard Tyrode’s solution (mM): 140 NaCl, 5 KCl, 2 CaCl2, 1 MgCl2, 10 HEPES, 10 Glucose, pH 7.4 with NaOH. Immediately before the start of patch clamp experiments, the chamber was thoroughly perfused by 10–15 mL of depolarizing bath solution (mM): 140 KCl, 2 CaCl2, 1 MgCl2, 10 HEPES, pH 7.4 with KOH. Patch clamp pipettes were pulled from borosilicate glass capillaries (Sutter Instrument Company) using a P-2000 puller (Sutter Instrument Company), fire-polished with a CPM-2 micro forge (ALA Scientific Instruments), and Sylgard 184 elastomer (Sigma-Aldrich) was used to coat pipette tips. Pipette solution contained (mM): 280 NaCl, 4 CsCl, 2 CaCl2, 1 MgCl2, 0.05 CdCl2, 10 HEPES, pH 7.4 with CsOH. The pipette resistance was 5–10 MΩ. Ionic current signals were recorded using a Multiclamp 700B amplifier and digitized using a Digidata 1440 A using pClamp 10 acquisition software (all from Molecular Devices). Instrumental subtraction of pipette capacitance was applied after obtaining a gigaohm (GΩ) seal. All recordings included in the analysis had a seal resistance of more than 5 GΩ. The signal was sampled at 100 kHz (at 0.01 ms time step) and low-pass filtered at 4 kHz using an 8-pole Bessel filter. No additional filtering was applied. Cell-attached currents were elicited by the following voltage protocol: From a holding potential of 80 mV, a 200 ms pre-pulse to 120 mV was applied, which was followed by a step to a test potential of 40 mV for 500 ms. This stimulus protocol was repeated with a frequency of 0.3 Hz. At least 100 current sweeps recorded with this protocol were obtained from each membrane patch included in the analysis. For estimation of the effect of lidocaine, cells were stimulated using the same protocol but at 10 Hz frequency to achieve the maximal effect of lidocaine (Sigma-Aldrich)87. The effect of lidocaine was estimated after 5 min of wash-in of 50 µM of the drug (the concentration was chosen to achieve the maximum effect based on previous lidocaine dose-response measurements87,88). To measure recovery from fast inactivation in the cell-attached configuration, a two-pulse protocol was applied. Cells were held at 80 mV, and a 200 ms pre-pulse to 120 mV was applied before the first pulse to 40 mV for 30 ms was applied. After the first pulse, the potential was returned to 120 mV for 5 ms (corresponds to the time constant of recovery from inactivation estimated on whole-cell NaV1.5 currents74). This was followed by a second 30 ms pulse to a test potential of 40 mV. The frequency of repetition of this protocol was 0.3 Hz. For presentation, the polarity of currents and potentials for cell-attached experiments was reversed following the convention of inward currents being negative. Whole-cell INa measurements Current recordings in the voltage-clamp mode were obtained with a MultiClamp 700B amplifier and Digidata 1440 A digitizer (Molecular Devices). The pipette solution contained (mM): 10 NaCl, 20 TEA-Cl, 123 CsCl, 1 MgCl2, 0.1 Tris-GTP, 5 MgATP, 10 HEPES, 10 EGTA, pH 7.2 (adjusted with CsOH). Cells were equilibrated for 5 min after establishing the whole-cell configuration. For INa recordings in CHO cells and INa-L in cardiac myocytes, the bath solution contained (mM): 140 NaCl, 4 CsCl, 1 CaCl2, 2 MgCl2, 0.05 CdCl2, 10 HEPES, 10 glucose, 0.03 niflumic acid, 0.004 strophanthidin (both Sigma Aldrich), pH 7.4 with CsOH. For peak INa recordings in mouse cardiomyocytes, the extracellular solution was altered by reducing NaCl to 10 mM, increasing CsCl to 123 mM, and adding 20 mM TEA-Cl. Patch pipettes had a resistance of 1.2–1.6 MΩ after fire-polishing. Compensation for whole-cell capacitance and series resistance (≥60%) was applied along with leak subtraction. Signals were sampled at 20 kHz and filtered with a 10 kHz 8-pole Bessel filter. To measure the voltage dependence of peak INa, cells were held at −90 mV and a 200 ms pre-pulse to −120 mV was applied followed by a step potential from −100 mV to 80 mV (for CHO cells) or −100 mV to 40 mV (for cardiac myocytes) in 5 mV increment. This protocol was repeated with a frequency of 0.3 Hz. To measure INa-L, the cells were repeatedly stepped to −30 mV (for cardiac myocytes and NaV1.5-expressing CHO cells) or −10 mV (for NaV1.6-expressing CHO cells) using the same holding and pre-pulse potentials. The voltage-dependence of steady-state inactivation of peak INa was measured at −40 mV (for cardiac myocytes and NaV1.5-expressing CHO cells) or −10 mV (for NaV1.6-expressing CHO cells) after 250 ms pre-pulses to potentials ranging from −140 to −20 mV in 5 mV increments. The holding potential for these experiments was −90 mV, and the pulse protocol was repeated at a frequency of 0.3 Hz. To assess the effect of veratridine (VTD, 25 µM, MCE) on tail INa, CHO cells stably expressing human NaV1.5 (B’SYS GmbH) were held at −90 mV and a 200 ms pre-pulse to −120 mV was applied followed by depolarizing step to −40 mV for 100 ms. The potential was then returned to −90 mV and the protocol repeated every 5 s. Baseline INa was measured during 20 consecutive pulses, then VTD was applied. Tail INa was measured 3 ms after stepping from −40 mV to −90 mV and reaching a steady-state level (about 30–40 pulses after VTD application) as previously described in ref. 44. Action potential measurements APs were recorded from freshly isolated mouse cardiac myocytes at room temperature by whole-cell patch clamp in the current-clamp mode using a MultiClamp 700B amplifier and Digidata 1440 A digitizer (Molecular Devices). The bath solution contained (mM): 133.5 NaCl, 4 KCl, 1 MgCl2, 2 CaCl2, 10 glucose, and 10 HEPES, and the pH was adjusted to 7.4 with NaOH. Pipette solution contained (mM): 50 KCl, 90 K-aspartate, 5 NaCl, 5 ATP-Mg, 1 MgCl2, 0.1 Tris-GTP, 10 HEPES, pH 7.3 with KOH. The junctional potential was corrected by 10 mV before establishment of a giga-ohm seal. APs were elicited by 1 ms current stimuli (0.6–1.5 nA). In ΔKPQ-NaV1.5 cardiomyocytes, current stimuli typically elicited APs, which did not repolarize for more than 3 s due to a high frequency of early afterdepolarizations (EADs), as was previously reported in ref. 89. Thus, in these cells, we used a prolonged pacing cycle (20 s) to allow the membrane potential to return to baseline between measurements. In ΔKPQ-NaV1.5 cardiomyocytes treated with SB216763 (5 µM for 3 h), APs typically did not exhibit such failure to repolarize; therefore, we used the standard pacing cycle length (3 s)74. Action potential duration at 90% repolarization (APD90) was measured only in APs with no EADs. An EAD was defined as a transient slowing or reversal of the membrane potential during repolarization. A DAD was defined as a fluctuation from the resting membrane potential of more than 10 mV. Patch clamp data analysis and modeling Idealization of cell-attached patch clamp recordings For cell-attached patch clamp recordings, we performed post hoc subtraction of capacitive transients from each current sweep using the Adam method90 for optimization of the previously suggested convex function with L1 regularization27,91. Next, to isolate the NaV signal from instrument noise, we idealized all current sweeps using Bayesian hidden Markov models (HMM)92 as described below. First, single-channel current amplitude and standard deviations of noise within recordings were estimated on a training set of current sweeps exhibiting only single-channel openings. The presence of only single channel openings in training sweeps was confirmed by fitting Gaussian mixture models to all current amplitudes in each of the sweeps and subsequent model selection using a Bayesian information criterion93. Estimation of the single-channel noise and amplitude parameters in training sweeps was performed by Bayesian inference. A regularizing Laplace prior distribution was put on the standard deviations of noise and the mean current amplitudes for a closed and open channel. An HMM likelihood calculated with the forward algorithm was used as the Bayesian model likelihood function92. The likelihood HMM had a constant uniform initial distribution, and transition probabilities to a different state were set to 0.05 for all latent states. In the likelihood HMM, a robust Student’s t distribution was used as an observation distribution. A vague exponential prior was put on the ν (degree of freedom) parameter of the observation distribution. Posterior sampling was performed using Parno and Marzouk’s modified Hamiltonian Monte Carlo (HMC) method using the softplus transformation for all priors in the Bayesian model94. Additionally, a dual averaging step size adaptation was applied to the model parameters to improve convergence of HMC sampling95. Finally, the posterior mean values were used for idealization of channel openings. Idealization of single- and multi-channel patch clamp recordings is considered as an estimation of the true number of open channels at each time point in a recording. To this end, we employed the maximum a posterior estimate of a sequence of hidden states resulting in an observed sequence of current amplitudes using the Viterbi algorithm in conjunction with an HMM96. Here, we considered the true number of open channels at each time point as a latent state of an HMM. For the Viterbi algorithm, HMMs were parametrized as described above for Bayesian likelihood HMMs, except a robust Cauchy distribution97 was used as an observation distribution to eliminate the ν parameter. Before idealization, a maximal number of open channels in a recording was set to a value apparently exceeding the actual number of observed open channels. Thus, this method does not require a precise prior estimation of the maximal number of open channels. Next, amplitudes of several simultaneously open channels were set to a sum of the single channel amplitudes (e.g., the amplitude of 5 open channels equals the single channel amplitude multiplied by 5). Thus, the initial probability vector and transition probability matrix in HMM were automatically modified according to the chosen maximal number of channels. For simplicity, open channel noise was considered to be independent of the number of open channels. We validated this idealization method on simulated single- and multi-channel NaV1.5 currents. To simulate channel activity, we used the NaV1.5 Markov model reported by Maltsev and Undrovinas98. After simulation of current sweeps, we added to them 8-pole Bessel-filtered noise to mimic a real patch clamp signal. In these tests, our algorithm demonstrated higher precision in comparison to the half-amplitude threshold-passing method (Supplementary Fig. 12). Additionally, we validated our method on NaV1.5 recordings experimentally obtained from CHO cells and demonstrated a statistically significant separation of current amplitudes assigned to different numbers of open channels both in single- and multi-channel membrane patches (Supplementary Fig. 13). For idealization of single- and multi-channel INa-L (50–500 ms after the test potential application), we employed the standard half-amplitude threshold-passing algorithm due to instability of the Viterbi algorithm for long time series. For the half-amplitude threshold-passing algorithm, we used a single-channel current amplitude estimated with Bayesian modeling, as described above. Analysis of ion channel kinetics Kinetics of INa in cell-attached recordings were analyzed after ensemble-averaging idealized current sweeps. At least 100 current sweeps recorded from one membrane patch were used for analysis. The presence of only one active channel in each of the membrane patches was confirmed by fitting Gaussian mixture models to all current amplitudes in a set of 100 or more current sweeps obtained from that specific patch. If the two-component model was selected as the best model based on the Bayesian information criterion93, then a membrane patch was defined as a single-channel patch. Single-channel open probability (Po) was calculated only in single-channel membrane patches by ensemble averaging of the number of open channels at each time point during a test potential. Analysis of experimental and simulated ion channel activity was performed identically. In cell-attached and whole-cell recordings from CHO cells, INa-L was calculated as the average INa over 50–500 ms after the test potential application, and then normalized to peak INa. To estimate the time constant of peak INa decay, we fitted a mono-exponential decay function to 10 ms time intervals of the INa time course during peak INa decay. To eliminate the potential contribution of differences in peak INa amplitudes in different membrane patches, we normalized INa to peak INa in each of the analyzed patches. VMR (also known as index of dispersion42) of a number of open channels at peak INa was calculated as a mean value of VMR in the time interval of ±0.5 ms from the time point of peak INa. In whole cell recordings from cardiac myocytes, INa-L was quantified as an integral of INa over time from 50 to 500 ms after test potential application. This integral INa was then normalized to the cell capacitance74. Voltage-dependence of steady-state inactivation was fitted to the Boltzmann sigmoidal function to calculate Vhalf of inactivation. Parametrization of the single-channel Markov model To model gating of the single NaV1.5 channel, we applied Bayesian inference to estimate transition rates of the single NaV1.5 channel Markov model99 from a set of 971 idealized current sweeps recorded from 8 single-channel membrane patches in CHO cells stably expressing human NaV1.5. In the Bayesian model, we used independent normal priors on each of the voltage-independent auxiliary variables, which determine voltage-dependent transition rates according to equations detailed in Moreno et al.99. The mean and standard deviation parameters of these priors were set to the corresponding best-fit values provided in Moreno et al.99. An HMM likelihood was used as the likelihood function. The HMM was parametrized in the following way: transition rates were organized in the transition rate matrix (frequently named Q-matrix) to represent the Moreno et al. model graph99. This matrix was used to calculate the transition probability matrix (frequently named A-matrix) by applying a matrix exponential operation with the time step of 0.01 ms as previously described in ref. 36. The initial probability of the state C3 was forced to 1, while the remainder of the states were kept at 0. The deterministic distribution with a mean of 1 for the open state, and 0 for all other states was used as the HMM observation distribution. Inference was performed on all 971 current sweeps simultaneously by HMC as described above. The posterior mean values of the inferred parameters were used for deterministic and stochastic simulations of ion channel gating. We found that the inferred transition rates better predicted the experimentally observed single-channel open probability time course and distributions of open and closed dwell times in comparison to the initial transition rates (Supplementary Fig. 7). Deterministic and stochastic simulations of ion channel gating In deterministic and stochastic simulations of channel activity after application of a test potential, an initial probability vector was calculated as the distribution of a Markov model at the holding potential preceding the test potential using vector and matrix multiplication36. In deterministic simulations, the distribution of a Markov model at each time point was calculated using the transition rate matrix exponential and a time step of 0.01 ms36. In stochastic simulations, the time series of ion channel activity was simulated using stochastic transition and emission mechanisms of HMM92. For HMM parametrization, the transition probability matrix was calculated by the transition rate matrix exponential using a time step of 0.01 ms100. As an observation distribution, we used a deterministic distribution with the mean parameter of 1 for the open state, and 0 for the remaining ones. For stochastic simulations of interacting channels using the Naundorf et al. model43, we simulated the activity of each of two interacting channels simultaneously by running two parallel Markov chains of the channel conformational states and corresponding observations. These Markov chains were driven by the HMM based on the Moreno et al. model99 parametrized by Bayesian inference on our NaV1.5 single channel recordings as described above. The time step in simulations was 0.01 ms. At each time step, a voltage shift parameter was calculated, given a number of open channels in the pair of interacting channels at that time step and the constant J, as described by Naundorf et al.43. Then, transition to the next step was performed under a transition probability matrix recalculated given the voltage shift parameter value at the current time step. J was set at 70 mV as previously described in ref.38. To estimate the magnitude of the peak current, the simulation was run for 5 ms after the test potential (−40 mV) application. To simulate two non-interacting channels, the same algorithm was used, but with a constant transition probability matrix. Monte Carlo simulations of channel localization and integral INa in the membrane To correlate the probability of interaction between channels in the membrane and INa, paired nanoscopy and INa measurements are needed. Since such type of experiments remain beyond the limitations of these techniques, we simulated the localization of channels in the membrane and the probability of contact between them to validate the predictions of our model of channel-channel interaction. To this end, we represented 1 µm2 of a membrane area as a 1 µm × 1 µm square grid. Square cells of this grid had a side of 10 nm or 20 nm for simulations of channels with corresponding diameters. Then, grid cells were independently and randomly occupied by no more than 1 channel, given specific total numbers of channels covering the physiologically relevant interval: 100, 200, 500, 1000, 1500, and 2000 per 1 µm2 grids14. Channels localized in directly adjacent grid cells were considered capable of interaction. Next, we estimated the total INa carried by interacting channels in pairs in each grid. To this end, we first counted the number of channels occupying two adjacent grid cells. Second, we multiplied the fraction of open channels simulated by composite models of two interacting channels by the number of channels in interacting pairs and by the single channel current amplitude (−1.8 pA at −40 mV). Next, we estimated a total INa carried by non-interacting channels. To achieve this, we counted the number of channels occupying non-adjacent grid cells. Then, we multiplied this by the fraction of open channels predicted by the model of non-interacting channels and by the single channel current amplitude (−1.8 pA at −40 mV). Finally, we summed INa from interacting and non-interacting channels to obtain INa through all channels in the grid. The simulated integral INa was further analyzed as those obtained from CHO cells as described above. Quantification of functional interaction between channels based on the binomial distribution expected for non-interacting channels To quantify nonstationary interaction between ion channels operating in the plasma membrane, we employed a previously suggested approach based on the statistical analysis of a number of open channels over time27,36. The analysis is based on the idea that in the absence of functional interaction between channels, the distribution of numbers of open channels must be binomial40,41,101. On the other hand, any interaction between channels affecting transitions between their conformational states (i.e., functional interaction) must produce a non-binomial distribution of numbers of open channels27,36. Thus, to quantify interaction between channels, the distribution of numbers of open channels is compared to a binomial distribution. However, previous work identifies challenges in precisely estimating the binomial parameter that describes the total number of channels in the membrane (n) for more than one channel41. The problem becomes compounded by the difficulty of assessing the number of ion channels with a low steady-state open probability40, such as NaV1.5. In line with the previous studies40,41, our estimation of n in stochastic simulations of NaV1.5 channel gating demonstrated that the nonstationary noise analysis (also known as mean-variance analysis), as well as the “k-max method”41 cannot reliably estimate n of all n > 1 (Supplementary Fig. 5). Notably, the “k-max method” previously used for quantification of inter-channel interaction27,41, generated a progressively increasing error (Supplementary Fig. 5b, right), while the error of the nonstationary noise analysis randomly fluctuated independently of the true n (Supplementary Fig. 5b, left). Despite an imprecise estimation of n associated with nonstationary noise analysis, fitting the mean-variance relationship to a binomial mean-variance relationship distribution was virtually ideal in the case of non-interacting channels, as was previously shown in ref.40 and confirmed here by us (Supplementary Fig. 5a). In order to minimize the influence of errors intrinsic to such estimates of n40,41, we examined the goodness-of-fit of an actual mean-variance relationship in the number of open channels (k) to the binomial mean-variance relationship, which allows us to bypass the problem of estimating n for the detection of interaction between channels. For quantification of goodness-of-fit, we used the Kullback-Leibler divergence (DKL, also known as “entropy difference”27,36,37) of the observed data relative to the best-fit (“expected”) binomial distribution of k (Supplementary Fig. 6). Identification of the number of open channels (k) is performed by idealization of raw patch clamp current traces (Supplementary Fig. 6-1). Mean (µ, dark blue) and variance (σ2, light blue) in k are calculated from at least 100 idealized current sweeps per membrane patch (Supplementary Fig. 6-2). All experimentally measured µ to σ2 relationships are fitted to the theoretically expected function (shown in the panel), assuming the absence of any channel-channel interaction (Supplementary Fig. 6-2, inset). The best-fit value of the total number of channels (n) is used to calculate expected probabilities for observations of k (Supplementary Fig. 6-3ii). The ratios of expected and observed probabilities for each value of k are then taken into consideration in the DKL calculation (Supplementary Fig. 6-3iii). Finally, DKL is calculated for all time points within 5 ms post-test potential application. DKL at the peak of ensemble-average INa was calculated as the mean value of DKL in the time interval of ±0.5 ms from the time point of peak (maximum) INa. We defined the expected DKL range for non-interacting channels (Fig. 2a–c, black) based on the stochastic activity of 2–190 non-interacting NaV1.5 channels simulated using a single-channel Markov model parametrized using our experimental observations from single channels (Supplementary Fig. 7)99. Composite Markov models Composite Markov models of two or more channels were obtained by applying the Cartesian graph product and Kronecker matrix sum to a single-channel Markov model graph and transition rates matrix, respectively, as previously described in ref.36. To model single channels, we used a simplified Clancy–Rudy model102, in which the first two closed states (C1, C2) were omitted. This simplification was made to reduce the number of composite states and make simulations of composite models of 3 and 4 channels computationally tractable. Additionally, in order to more closely reflect our observations and previous findings91, the transition rate from the closed to open state in the model was scaled by a factor of two, which slightly increased persistent activity of channels in simulations. Recently, interactions between NaV channels were described to occur only in the closed state25, thus making the closed state more energetically favorable. This, in turn, would translate into an increased rate of transition into, and reduced transition out of, the interacting closed state. To incorporate this interaction between channels, we scaled the transition rates between composite states so as to preserve microscopic reversibility in the composite models, as previously described in ref.36. Specifically, we lowered the energy of the composite closed state (CC, CCC, CCCC for 2, 3, and 4 channels, respectively) by scaling all rates of transition out of the composite closed state by e2E/kT, where E is energy, and k and T are the Boltzmann constant and temperature. Then we lowered the energy barriers between the composite closed state and all other states by scaling all transition rates between them by e−E/kT. For all simulations, we used E = −0.5kT. To model the effect of lidocaine on interacting and non-interacting channels, we built upon previous findings, which demonstrated that lidocaine binds to NaV1.5 in the inactivated state, thereby trapping the channel in that state49,53. Thus, we modeled the single NaV1.5 channel bound to lidocaine by scaling the transition rates out of fast (FI) and SI states by 0.001 in the Clancy–Rudy model. Then, composite Markov models of non-interacting and interacting channels were constructed as described above. Statistics Statistical analysis was performed with GraphPad Prism 10 (GraphPad Software Inc.) and Scipy103. The normality of the data was tested (Shapiro-Wilk test), and appropriate methods were chosen for comparative statistics. For comparison of 2 independent data sets, an unpaired, 2-tailed Student’s t or Mann–Whitney U-test was used for normally and non-normally distributed data, respectively. For comparison of 2 paired data sets, the Wilcoxon matched-pairs or signed-rank test was used depending on normality. The two-sample Kolmogorov-Smirnov test was used to compare distributions. A p value of less than 0.05 was considered significant. For comparison between >2 datasets, ordinary 1-way ANOVA or Kruskal–Wallis test was used for normally and non-normally distributed data, respectively. Post hoc multiple comparisons were performed with the original FDR method of Benjamini and Hochberg. A q value of less than 0.05 was considered significant. The χ2-test was used to compare categorical data. All data are expressed as mean ± SEM or as box and whiskers plots, where the box represents the first and third quartiles, the line within the box reflects the sample median, and the whiskers reflect the minimum and the maximum values, unless otherwise indicated. Computational approaches Modeling and analysis of ion channels in cell-attached recordings was performed in Python 3.10 on the Google Colaboratory (Alphabet Inc.) platform using the following libraries: HMC sampling for Bayesian models and generation of random time series by HMM was performed in TensorFlow Probability104; fitting of functions to data and operations with matrices were performed by Scipy103, NumPy105, and TensorFlow104. Whole cell currents were analyzed in Clampfit 11.0 (Molecular Devices). Plotting and statistical hypothesis testing were performed in Matplotlib, Scipy, and GraphPad Prism 10 (GraphPad Software Inc.). Ethical approval All animal procedures were approved by the Institutional Animal Care and Use Committee at The Ohio State University and performed in accordance with the Guide for the Care and Use of Laboratory Animals published by the U.S. National Institutes of Health (NIH Publication No. 85-23, revised 2011). Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Reporting Summary Transparent Peer Review file Source data Source Data Acknowledgements The authors thank Joseph Waterman for maintaining the mouse colony used in this study and Jessica Matthias for valuable advice on sample preparation for MINFLUX imaging experiments. Funding: R01HL155378 (PBR), R01NS121234 (PBR), and R01HL165751 (RV). Author contributions M.T., R.V., and P.B.R. conceived and designed the study. M.T., M.A., C.H., Z.S., and J.O.W. performed experiments. J.O.W. and R.V. contributed tools. M.T., M.A., R.V., and Z.S. analyzed data. M.T., R.V., and P.B.R. drafted the manuscript. All authors read and approved the final manuscript. Peer review Peer review information Nature Communications thanks Igor Efimov and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available. Data availability Source data associated with the study are provided as a Source Data file. The raw data used in this study are available in the DRYAD database under accession code: [10.5061/dryad.0cfxpnwgh]. cDNA accession codes: mouse SCN5A cDNA (NCBI Nucleotide database accession number: XM_006511997.1), human SCN5A cDNA (NCBI Nucleotide database accession number: NM_198056.2), human SCN8A cDNA (NCBI Nucleotide database accession number: NM_014191.4), human SCN1B (NCBI Nucleotide database accession number: NM_001037.5). Source data are provided with this paper. Code availability Custom code available at GitHub: Mikhail Tarasov (2026) “tarasov4/Analysis-and-modeling-of-voltage-gated-sodium-channels-clusters: analysis and modeling of voltage-gated ion channels”. A permanent reference to the version of the code used in this study is available on Zenodo. 10.5281/zenodo.18499384. Competing interests Dr. Otto Wirth is an employee of Abberior Instruments GmbH, the manufacturer of the MINFLUX microscope used for imaging in this study. The remaining authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary information The online version contains supplementary material available at 10.1038/s41467-026-72387-8. References Associated Data Supplementary Materials Supplementary Information Reporting Summary Transparent Peer Review file Source Data Data Availability Statement Source data associated with the study are provided as a Source Data file. The raw data used in this study are available in the DRYAD database under accession code: [10.5061/dryad.0cfxpnwgh]. cDNA accession codes: mouse SCN5A cDNA (NCBI Nucleotide database accession number: XM_006511997.1), human SCN5A cDNA (NCBI Nucleotide database accession number: NM_198056.2), human SCN8A cDNA (NCBI Nucleotide database accession number: NM_014191.4), human SCN1B (NCBI Nucleotide database accession number: NM_001037.5). Source data are provided with this paper. Custom code available at GitHub: Mikhail Tarasov (2026) “tarasov4/Analysis-and-modeling-of-voltage-gated-sodium-channels-clusters: analysis and modeling of voltage-gated ion channels”. A permanent reference to the version of the code used in this study is available on Zenodo. 10.5281/zenodo.18499384.