Predicting Patient Response to the Antiarrhythmic Mexiletine Based on Genetic Variation: Personalized Medicine for Long QT Syndrome

The data that support the findings of this study are available from the corresponding author on reasonable request.

cRNA for human Nav1.5 α subunit was produced from the pMAX vector. All mutagenesis was achieved using overlap extension polymerase chain reaction, followed by Infusion cloning (Clonetech). All mutations were confirmed with sequencing (Genewiz). Each plasmid was then linearized with PacI restriction enzyme. Capped mRNA was synthesized using the mMESSAGE mMACHINE T7 Transcription Kit (Life Technologies) and purified via phenol-chloroform extraction.

Four previously developed constructs for VCF were used in recordings (DI: V215C, DII: S805C, DIII: M1296C, and DIV: S1618C). mRNA of the Na_V1.5 channel constructs was coinjected with the Na_V β1 subunit in Xenopus oocytes. Voltage clamp recordings were performed 4 to 5 days after injection. The recording setup, solutions, and recording protocols for VCF are the same as described previously.^13–15 Mexiletine hydrochloride powder (Sigma) was dissolved in extracellular recording solution to a stock concentration of 4 mmol/L. The pH for the solution is adjusted to 7.4. Mexiletine was further diluted from the stock solution to various concentrations (2–2000 μmol/L). During recordings, measurements were made from the same cell before and after addition of the indicated concentration of mexiletine. Mexiletine was manually perfused into the extracellular solution chamber in the cut-open voltage clamp setup.

Human Embryonic Kidney 293T (HEK 293T) cells, obtained from the American Tissue Culture Collection (Manassas, VA), were maintained in DMEM (Gibco) supplemented with 10% FBS and 100 U/mL Penicillin-Streptomycin, in 37°C, 5% CO₂ incubator. Cells from passage 30 to 32 were cotransfected with Na_V1.5 variant and β1 subunit with jetPRIME reagents (Polyplus). Patch clamp recordings were conducted 24 to 48 hours after transfection. Solutions and protocols used were the same as described previously.¹⁵

Data analyses were performed with Clampfit (v10; Molecular Devices), MATLAB (R2012a; MATLAB), and Excel (Microsoft). Conductance-voltage (G-V), fluorescence-voltage (F-V), and steady-state inactivation (SSI) curves were quantified by fitting a Boltzmann function: . DIII-VSD deactivation rate was quantified based on the time to 50% decay. Channel recovery from inactivation was fitted with a sum of exponents function:, which accounts for both fast and slower components of recovery.¹⁶ The V_1/2 of wild-type (WT) DIII F-V curves from 21 cells were tested with the 1-sample Kolmogorov-Smirnov test. The test implies that the V_1/2 values are normally distributed. Thus, comparison of V_1/2 values of F-V curves between conditions or constructs was performed using paired or independent Student t test, respectively (Microsoft Excel). Comparisons of % of mexiletine block were performed with the nonparametric Mann-Whitney U test. The error bars shown in the figures represent the SEM. The number of trials (n) reported in figure legends represent biological replicates for each experimental condition.

Channel parameters (predictor variables) and channel responses to mexiletine (response variables) were standardized using z score transformation. Partial least square (PLS) regression was performed using MATLAB function plsregress. Model stability was examined with leave-one-out cross validation. Each perturbation (channel variant) is individually removed from the dataset, and a PLS regression model is built on the rest of the variants. Using this model, mexiletine responses are predicted for the removed variant and compared with the measured responses. The model’s general ability to predict left-out data was measured by calculating the Q-squared (Q²) values, which is the sum of squares of the difference between predicted and real values, normalized by the total variability in data. Variable importance in the projection (VIP) scores for each gating parameter is ranked by its impact on model fitness (Q²). One gating parameter is removed at a time, and cross-validated model fitness Q² is calculated for the model based on the rest of parameters. The lower the Q² is, the higher VIP score the gating parameter has.

Patients included in the study were referred to the Molecular Cardiology Division at ICS Maugeri Hospital in Pavia, Italy. Patients were genotyped and identified as carriers of a DNA variant in the coding sequence of the SCN5A gene and they were treated with mexiletine. Demographic data, personal and family clinical history, and follow-up data were stored in a customized database. The study was approved by the ethics committee of the Institution.²

All predictions were made without prior knowledge of clinical outcomes. Predictions were made based on patient QT_c baseline and their genetic variants. After measuring mutant channel gating properties, parameters were inputted into the mexiletine QT_c shortening PLS regression model, which then predicted postmexiletine QT_c intervals. The predicted values were then compared with clinical data.

We examined the effect of mexiletine on heterologously expressed Nav1.5 channels. The response to mexiletine was similar for some properties and different for others of the Class Ib antiarrhythmics. Like other Class Ib antiarrhythmics,¹⁷ mexiletine preferentially inhibited the late component of the Na⁺ current (I_Na) compared with the effect on the peak current¹⁷ (Figure 1A, top). Mexiletine also exhibited UDB (Figure 1A, bottom), which is a property of Class Ib antiarrhythmics that is reflected by an increase in the inhibitory effect as the channels are repetitively activated by depolarizing voltage pulses.

Binding of Class Ib drug lidocaine to the channel results in stabilization of the inactivated state, which is often reflected by a hyperpolarizing shift in the SSI curve.⁸ In contrast to lidocaine, mexiletine caused a minimal shift of the SSI curve (Figure 1B). However, similar to lidocaine, mexiletine delayed recovery from inactivation, especially the slow component of recovery (Figure 1C). Thus, mexiletine appeared to influence the inactivated state through a mechanism different from that of lidocaine.

Previously, multiple studies demonstrated that lidocaine block of Na_V channels enhances the stability of the DIII-VSD activated conformation.^11,12 Here, we tested mexiletine interaction with the Nav1.5 VSDs by VCF.¹⁵ In these experiments, a fluorophore is tethered to the charged S4 segment of 1 of the 4 VSDs. As the VSD changes conformation, the environment around the fluorophore is altered, which changes emission from the fluorophore, enabling measurement of the time and voltage dependence of the VSD conformation.¹⁴ The steady-state F-V curves represent the voltage dependence of the VSD activated conformation. Among 4 domains, only the DIII F-V curve displayed a large hyperpolarizing shift (ΔV_1/2 =−32.5±7.5 mV, P=0.04) after mexiletine block (Figure 1D), implying that the DIII-VSD remains in an activated conformation at more negative potentials on mexiletine binding. Comparing the fluorescence traces before and after mexiletine block showed that both DIII and DIV-VSDs have slower activation and deactivation kinetics in the presence of mexiletine (Figure 1D). Because the DIII and DIV-VSDs are tightly coupled,¹⁶ the mexiletine-induced alteration of the DIV-VSD kinetics may be a consequence of its effects on the DIII-VSD.

To account for mexiletine’s effects on the DIII-VSD, we proposed a model for its mechanism of action. Binding of mexiletine within the channel pore prevents the DIII-pore domain (S5-S6) from transitioning to a completely closed conformation during membrane repolarization (Figure 1E). The partially open conformation of the DIII-pore causes the DIII-VSD to remain in the activated conformation. This is like previously reported mechanisms of the interaction of lidocaine with Na_V channels.¹²

To understand the molecular mechanisms underlying differences in mexiletine sensitivity among LQT variants, we tested channels with single point mutations R1626P or M1652R (Figure 2A), which exhibit different responses to mexiletine.⁴ These 2 variants exhibited different responses to mexiletine under nonstimulated conditions, tonic block (TB), and under stimulated conditions, UDB. Consistent with the previous studies, when compared with the effect of mexiletine on WT channels, the drug exerted TB of R1626P at lower concentrations (EC₅₀=211 μmol/L) and required higher concentrations to exert TB of M1652R (EC₅₀=2035 μmol/L). Mexiletine had an EC₅₀=761 μmol/L for TB of WT channels (Figure 2B).

We assessed UDB by applying 400 ms depolarizing pulses at 2 Hz, mimicking conditions during VT.⁴ We found that WT and R1626P have comparable UDB (WT: EC₅₀=58 μmol/L and R1626P: EC₅₀=57 μmol/L), while M1652R has much lower UDB (EC₅₀=193 μmol/L; Figure 2C). We tested the UDB of mexiletine using the cut-open voltage clamp. The EC₅₀ values that we observed with this method are higher than those reported using patch clamp analysis of HEK 293 cells.⁴ We hypothesized that this difference is because of limited solution access to the cell membrane in the cut-open voltage clamp setup during perfusion. To test this hypothesis, we measured dose responses using a 2-electrode voltage clamp, which allows better access to the solution. Two-electrode voltage clamp recordings showed mexiletine EC₅₀ values for each variant that were similar to previously reported values (Online Table I; Online Figure I), suggesting that, in the cut-open setup, amount of mexiletine at the channel is ≈3-fold lower than the perfused concentration (Online Table I). With this information, we can account for the differences in EC₅₀ values that relate to methodology.

To probe the link between the DIII-VSD conformation and mexiletine block, we assessed the correlation between DIII-VSD conformation and sensitivity of the channel to mexiletine. We hypothesized that, if mexiletine block of the pore caused the DIII-VSD to remain in the activated position, then channels with an activated DIII-VSD conformation would facilitate mexiletine accessibility to the pore. VCF experiments showed that both mutations significantly affected DIII-VSD conformation (Figure 2D and 2F), despite their locations in DIV-VSD, which is distant from DIII-VSD (Figure 2A). Compared with WT channels, the mexiletine-sensitive⁴ R1626P mutant exhibited a hyperpolarized DIII F-V curve (ΔV_1/2=−38.9 mV, P=0.02), suggesting that more DIII-VSDs were in an activated conformation at the resting membrane potential. Conversely, the mexiletine-insensitive⁴ variant M1652R exhibited a depolarizing shift in the DIII F-V curve (ΔV_1/2=28.2, P=0.01), indicating that more DIII-VSDs were in deactivated conformation. The shifts in voltage dependence of the DIII-VSD activation mirrored the differences in block by mexiletine: The mutant with DIII-VSD in an activated conformation (R1626P) at the resting potential displayed higher TB. These results support our hypothesis of a reciprocal relationship between mexiletine block and DIII-VSD conformation of the mutants.

Conventionally, occupancy of the inactivated state has been considered the primary determinant of Class Ib drug action. Consequently, the modulated receptor model describes preferential drug binding to channels that are inactivated. To test this notion, we measured how variants affected the SSI curve. Although the most sensitive mutation R1626P shifted SSI prominently, the magnitude of the shifts by these 2 variants are not consistent with their differences in mexiletine block (Figure 2E).

To further test whether the conformation of DIII-VSD regulates mexiletine block independent of inactivation, we assessed TB of WT channels at various potentials ranging from −120 to −90 mV. At these holding potentials, the WT channels exhibited full conductance (none in the inactivated state; Figure 2E) and showed a range of DIII-VSD conformations (Figure 2D). At 4 different holding potentials, −120, −110, −100, and −90 mV, the channel showed altered TB by mexiletine (Figure 2H). Moreover, the amount of TB had a linear relationship with the fraction of DIII-VSDs in the activated conformation at those potentials (Figure 2H). This result showed that the proportion of channels in the inactivated state is not the only factor that determines effectiveness of mexiletine block.

Based on our results, we proposed a model that explains the difference in mexiletine sensitivity between the 2 LQT variants (Figure 2G). At resting membrane potential, DIII-VSD of the sensitive variant R1626P tends to occupy the activated conformation. Because the activated conformation of DIII-VSD is coupled to conformation of the DIII-pore, the pore adopts a conformation that facilitates mexiletine accessibility. In contrast, fewer of the DIII-VSDs of the insensitive mutant M1652R are in the activated conformation at these potentials, causing the DIII-pore to remain in a conformation that prevents mexiletine from binding.

The F1760K mutation eliminates UDB by lidocaine and prevents lidocaine from affecting gating currents. Based on this binding site, Hanck et al⁹ categorized lidocaine block into 2 components: a voltage-independent lipophilic block and a voltage-dependent block. Lipophilic block is independent of the putative binding site F1760. We tested whether the difference in the EC₅₀ for TB by mexiletine among the R1626P, M1652R, and WT channels is because of lipophilic or voltage-dependent block by monitoring their responses to mexiletine in the background of F1760K mutation.

We assessed the response of F1760K channels to mexiletine. Both TB and UDB by mexiletine are greatly reduced for the F1760K channel. Using VCF, we found that mexiletine did not alter the conformation of DIII-VSD of the F1760K mutant channel (Online Figure II). We measured mexiletine-induced TB at 500 µmol/L, a concentration at which the difference in TB among the channels was evident (Figure 2B), in WT, R1626P, and M1652R channels that also had the F1760K mutation. The TB achieved with 500 µmol/L mexiletine in channels with F1760K were similar (Figure 2I). Although not significant, with the F1760K background, R1626P even became slightly less sensitive than WT and M1652R. Thus, the differences in the EC₅₀ values for mexiletine among these LQT variants appeared because of voltage-dependent block rather than lipophilic block.

To further understand how the DIII-VSD affects mexiletine block, we used the A1326W mutation (Figure 3A), which decouples DIII-VSD from the pore.¹⁸ In channels with A1326W, mexiletine no longer affects the conformation of DIII-VSD (Online Figure IIIA), demonstrating that a connection between the DIII-VSD and the pore is required to observe the mexiletine effect on DIII-VSD conformation. Our hypothesis is that R1626P and M1652R have distinct mexiletine sensitivities because of differences in the voltage dependence of DIII-VSD activation, consequently altering pore accessibility by mexiletine. From this hypothesis, we predict that channels in which the DIII-pore is decoupled from the DIII-VSD by A1326W will exhibit similar mexiletine block. Indeed, we observed that, on the addition of A1326W, mexiletine caused similar TB and UDB for the R1626P, M1652R, and WT channels (Figure 3B and 3C).

In the presence of the A1326W mutation, the differences caused by R1626P or M1652R mutation in DIII-VSD activation and SSI are preserved (Figure 3D and 3E), suggesting that the A1326W mutation does not interfere with voltage-dependent DIII-VSD conformational changes. We proposed a model to explain the elimination of mexiletine sensitivity by the A1326W mutation in the 2 LQT3 variants (Figure 3F). Although R1626P stabilized and M1652R destabilized the activated conformation of DIII-VSD, the channel pore remains in the same conformation with the same mexiletine accessibility, because A1326W decoupled the DIII-pore from the DIII-VSD. These results indicated that the differences in mexiletine sensitivity of R1626P and M1652R are a consequence of the effects of the mutations on DIII-VSD activation, which are transmitted to the DIII-pore to increase accessibility to mexiletine. Thus, removing the coupling between the DIII-VSD and pore abolished the differences in mexiletine sensitivity.

By studying 2 LQT variants with extremely high or low mexiletine sensitivity, we showed that the voltage dependence of DIII-VSD activation strongly affects mexiletine block. We explored if this mechanism is generally applicable to common LQT3 variants. We first investigated how DIII-VSD activation modulates TB by mexiletine. TB is usually assessed at negative potentials (−100 mV), a voltage at which most LQT variants have similar level of inactivation. DIII-VSD activation occurs at much lower voltage range than closed-state inactivation. Consequently, at −100 mV, the variability in the proportion of channels with DIII-VSD in the activated conformation is high among LQT3 variants. We measured the gating properties of WT and 15 LQT3 variant channels and the mexiletine TB of these channels. Most variants that we analyzed are found in patients who were previously treated with mexiletine.² Although the LQT variants span the channel (Figure 4A), many of the variants exhibited altered DIII-VSD activation. We observed a strong correlation between the voltage dependence of DIII-VSD activation (V_1/2 of DIII F-V) and TB by (R²=0.90; Figure 4B). Higher TB occurred for channels that had DIII-VSD activation at more negative potentials (Figure 4B). Instead of a linear relationship, we fitted the data to a Hill function, because we expected that TB will saturate at the ends of the curve. Intriguingly, we found that the minimum TB saturated at 15% block, suggesting that 15% of mexiletine-mediated TB is lipophilic block¹² (low affinity, voltage-independent block).

We also investigated the relationship between SSI and TB to test the classical theory that closed-state inactivation promotes Class Ib block. In contrast to DIII-VSD activation, closed-state inactivation (V_1/2 of SSI) did not correlate well with mexiletine TB (R²=0.48; Figure 4C). These results further support the hypothesis that voltage dependence of DIII-VSD activation rather than that of closed-state inactivation determines mexiletine TB.

UDB and late I_Na block are critical features of Class Ib drugs. UDB enables the drug to block channels during periods of heightened channel activity, as in tachycardia. Late I_Na block describes a unique property of Class Ib drug that preferentially blocks the late over the peak component of I_Na, which shortens the action potential duration, while minimally affects the AP upstroke or conduction. Unlike TB, which occurs at resting potentials at which channels undergo limited conformational changes, UDB and late I_Na block involve many complex gating transitions, including the activation of the other 3 VSDs, pore opening, pore closing, channel inactivation, and channel recovery from inactivation. Because of the complexity of the molecular movements that affect UDB and late I_Na block, using a single gating parameter, such as SSI or DIII-VSD activation, to predict them is insufficient (Online Figure IV). To address this challenge, we applied a data-driven modeling approach to identify the multivariate relationship between channel gating parameters and UDB or late I_Na block by mexiletine.

For each LQT variant, we quantified 14 gating parameters that describe gating processes, such as DIII-VSD, DIV-VSD activation, channel activation, and channel fast inactivation (Figure 5A). We also assessed UDB and late I_Na block by 250 μmol/L mexiletine for each variant. To understand how these gating phenomena related to drug block, we used the PLS regression approach. Gating parameters were used as predictive inputs, and the measured UDB or late I_Na block was used as an output for the PLS regression model. The PLS regression method has the ability to identify relationships between the measured gating parameters and drug block and also reduce redundancy among the input parameters

We first applied feature selection among the 14 gating parameters to identify the most important parameters in determining mexiletine UDB or late I_Na block. Feature selection was based on the VIP score of each parameter (Materials and Methods), which describes parameter impact on model fitness. Gating parameters with high VIP scores (Figure 5A bottom, red squares) were then extracted to build the final PLS regression model for prediction. The VIP scores suggest that 5 gating parameters are crucial for determining UDB, including voltage dependence of channel conductance (V_1/2 of G-V), DIII-VSD activation (V_1/2 of DIII F-V), DIV-VSD activation (V_1/2 of DIV F-V), time constant of slow recovery from inactivation (slow recovery τ), and late I_Na. Four gating parameters are important for predicting late I_Na block, including DIII-VSD activation (V_1/2 of DIII F-V), SSI (V_1/2 of the SSI), slow recovery from inactivation (slow recovery τ) and late I_Na. By reducing the number of gating parameters used in the model, feature selection not only helps prevent overfitting but also improves our understanding of the relationship between channel gating processes and drug response.

To reduce data dimensionality, we also decreased the number of principal components to 3 for UDB and 2 for late I_Na block, because the reduced components were sufficient to explain over 90% variants in the data. With the selected features and reduced components, the PLS regression model predicts the UDB with an R² of 0.9, suggesting the model fits the data well. We further validated the model with leave-one-out cross validation, as described in Methods. The cross-validated PLS regression model predicts the UDB with a Q² of 0.7 (Figure 5B). The Q² value calculated from cross validation measures the model’s ability to predict left-out data. A positive Q² indicates the model has predictive relevance.¹⁹ Compared with the best prediction using a single gating parameter (DIII-VSD activation), which has a Q² of 0.3, the PLS regression approach improved the prediction accuracy.

Likewise, the PLS regression model for late I_Na block has an R² of 0.6 and Q² of 0.5 (Figure 5C), suggesting the amount of late I_Na block by mexiletine can also be predicted from channel gating parameters.

We also built a PLS regression model to predict mexiletine-induced corrected QT_c shortening (ΔQT_c) in patients with the LQT3 variants for which we measured channel gating parameters. We obtained QT_c interval data before and after mexiletine for 32 patients with 13 different genetic variants from a previously published study.⁶ The data were used as training and validation dataset for the model. Although some patients also received β blocker metoprolol along with mexiletine, a review of their records showed that QT_c is not significantly different between patients treated with β block and mexiletine and those treated with only mexiletine.² The VIP scores for ΔQT_c showed that only 2 gating parameters are important for determining the ΔQT_c: DIII-VSD activation and τ of slow recovery from inactivation (Figure 5A, bottom). With these 2 parameters as inputs, the PLS regression model has an R² of 0.7, suggesting the model fit the training dataset well. To test the model prediction performance, we further cross-validated with leave-one-out validation. The model predicted QT_c shortening in patients with a Q² of 0.6 (Figure 5D), demonstrating that the model has strong predictive value. To further test whether electrophysiology data collected from mammalian cell lines can improve the model prediction accuracy, we also recorded I_Na from 13 variants and WT channels expressed in HEK 293T cells (Online Table III). Notably, the important gating parameters for each variant correlate well between the 2 expression systems (Online Figure V). Using the gating parameters quantified from HEK cell recordings, and DIII-VSD parameters from VCF recordings in Xenopus oocytes, we reconstructed the PLS regression model for ΔQTc (Online Figure VA). The model has an R² of 0.7 and a Q² of 0.6 (Online Figure VB), which are comparable to the model based on gating parameters quantified from Xenopus oocytes, suggesting that the data collected with the oocyte expression system were sufficient for building the PLS regression model for QT_c prediction.

Our results indicated that the clinical efficacy of mexiletine can be predicted from measurements of gating parameters of the Na_V1.5 variants, supporting the notion that in vitro testing may help predict a patient’s specific response to mexiletine.⁴ We extended this observation by building a precise model for predicting patient-specific QT_c shortening by mexiletine based on detailed biophysical parameters.

Notably, LQT3 is usually inherited in an autosomal dominant manner, suggesting that the patients have heterogenous population of WT and LQT3-linked Na_V1.5 channels. As mexiletine preferentially inhibits late I_Na and channels with more activated DIII-VSD, its effects on the WT channels are relatively weak compared with LQT3-linked channels. Although the model only focuses on the LQT3-linked channels, it should be sufficient to explain and predict the differences in QT_c shortening among patients.

To validate if the gating parameters selected based on the VIP scores improved the prediction accuracy (model fitness) of the PLS regression models, we evaluated 1000 models with randomly selected parameters (blue bars, Online Figure VIA through VIC). Among these models, those that contain most of the preselected parameters (>3 out of 5 for UDB, 3 out of 4 for late I_Na block, and 1 out 2 for ΔQT_c) had overall improved prediction accuracy, implying that VIP score is an effective method to rank gating parameter importance.

To test the performance of the PLS regression model in predicting a set of new variants in a blinded setting, we conducted a blind retrospective clinical trial that involved 8 LQT3 patients carrying 5 distinct SCN5A variants (Table; Figure 6), which were not variants included in the training/cross-validation dataset. The 8 patients were previously treated with mexiletine, and their electrocardiograms were recorded before and after treatment. Evaluators were blinded from these clinical data during prediction. We expressed the Na_V1.5 channel variants from those patients and tested them in vitro to obtain the 2 essential electrophysiological parameters for the prediction: DIII-VSD activation and slow recovery τ. We generated the predicted postmexiletine QT_c is with the mexiletine QT_c shortening PLS regression model, with an upper and lower bound based on the 95% CI calculated from the cross-validated model.

Strikingly, 7 out 8 patients had postmexiletine QT_c that aligned with our predictions (Table; Figure 6). We noted that the one outlier (N1774D) patient that we failed to predict had a baseline QTc of 814 ms, which is higher than most of the training data. Assembling patient data from both training and trial datasets, we observed a trend that patients with high QTc baselines (>650 ms) tend to have higher percentage of QT_c shortening by mexiletine, independent of their genetic variants (Online Figure VII). As a result, our current model is not suitable for predicting patients with high baseline QT_c (>650 ms). From this blind clinical trial, we validated that our PLS model accurately predicted patients’ response to mexiletine therapy for those patients with a baseline lower than 650 ms.

The modulated receptor theory proposed by Hille has been applied for 40 years to describe Class Ib drug interaction with Na_V channels.²¹ This theory includes 3 basic and modulated channel states that illustrate drug interactions with the channel pore: closed, open, and inactivated. The modulated receptor theory emphasizes the primary role of the inactivation gate in promoting and stabilizing drug blockade. As more information about channel structure has become available, it is apparent that many conformational changes that are spread throughout the channel work together to cause channel gating. The 4 VSDs exhibit varied behavior during gating and each is coupled to the channel pore. Thus, subtle changes in VSD dynamics can affect pore conformation and vice versa.

Previous studies showed that lidocaine binding to the pore affected DIII-VSD dynamics.^11,12 We demonstrated that LQT3 variants alter the voltage dependence of the DIII-VSD activation and channels that populated an activated DIII-VSD conformation exhibit increased mexiletine block. To form a more complete understanding of channel and Class Ib drug interactions, we considered various components of the channel. Based on these results, we propose an updated VSD-modulated receptor model that describes how conformations of the DIII-VSD, the pore, and the inactivation gate alter Class Ib drug blockade. In this updated model, we included 5 states: CR (pore closed, DIII-VSD resting), CA (pore closed, DIII-VSD activated), OA (pore open, DIII-VSD activated), IA (pore inactivated, DIII-VSD activated), and IR (pore inactivated, DIII-VSD resting; Figure 7). Drugs can block the channel from each state but with different binding and unbinding affinities. When channels are in the CR state, drugs have a low binding rate, because the hydrophilic pathway is unavailable. When the DIII-VSD activates, channels enter the CA state, in which drugs have much higher accessibility to the pore. Finally, in the channel open (OA) and inactivated (IA) states drugs exhibit high binding rates. After membrane repolarization, channel recovery from inactivation has slower kinetics compare to the DIII-VSD deactivation,²² causing the channel to enter the IR state. Drugs have both low binding and unbinding rates when the channel occupies this state. When channels are in CR and IR states, only the hydrophobic pathway is accessible.

Because of the structural similarity between lidocaine and mexiletine, patients with VT that respond well to intravenous lidocaine are often prescribed mexiletine for long-term treatment. However, mexiletine fails to prevent arrhythmia in a large fraction (50%) of these patients,²³ and even induces severe arrhythmia in some cases.²⁴ This clinical outcome suggests that mexiletine and lidocaine have distinct interactions with the channel that were not previously defined. Notably, mexiletine (pK_a 9.52) is mostly charged and hydrophilic, whereas lidocaine (pK_a 7.6) is partly uncharged and hydrophobic at physiological pH. Here, our results showed that the activated conformation of the DIII-VSD is required for hydrophilic, not hydrophobic, drug access to the channel pore. Many LQT3 variants stabilize the DIII-VSD in its activated position, promoting mexiletine block through the hydrophilic voltage-dependent pathway, which explains why mexiletine is effective in managing LQT3 syndrome. In contrast, for treating VT patients with normal Na_V1.5 channels that do not have activated DIII-VSDs, mexiletine may be less effective than lidocaine. These results suggest that response to lidocaine may not be a good predictor of the clinical response to mexiletine therapy.

The most severe side effect of Class I drugs is proarrhythmia, which is potentially because of conduction slowing, as peak I_Na is inhibited.²⁵ The prediction of peak I_Na block from the PLS regression models can help to identify the sensitivity for each variant. If a certain variant is susceptible to peak I_Na block, the dose of mexiletine can be adjusted accordingly to avoid the proarrhythmic effects.

We demonstrated several LQT3 variants are insensitive to mexiletine because of the less activated DIII-VSD conformation of the channel. To rescue their insensitivity, a new therapeutic strategy that uses a combination of drugs can potentially be employed. A drug that promotes DIII-VSD activation can be used in combination with mexiletine to improve antiarrhythmic efficacy in the insensitive variants. Several combination therapies of Na⁺ channel blockers have been tested in clinical settings.^26,27 For instance, an early study suggested that a combination of oral mexiletine and flecainide prevents recurrence of VT in patients that are nonresponsive to monotherapy.²⁷ From the present study, one may conjecture that flecainide may improve mexiletine efficacy by promoting DIII-VSD activation. However, the mechanism of why certain combinations improve efficacy is not known. More work needs to be done to systematically evaluate how different combination therapies alter channel conformations and gating to improve drug outcomes. As induced pluripotent stem cell–derived cardiomyocytes have emerged as useful model for studying LQT syndrome,²⁸ in the future studies, it may be possible to validate the PLS regression model based therapy predictions in the LQT3 patient induced pluripotent stem cell–derived cardiomyocytes.

For more general ventricular arrhythmias other than LQT3 syndrome, mexiletine has been prescribed to patients with recurrent VT postmyocardial infarction and ischemic heart diseases that are resistant to other conventional antiarrhythmic agents in early clinical studies.²⁹ Although mexiletine effectively suppressed episodes of premature ventricular contraction, it induced adverse side effects in some patients resulting in them withdrawing from therapy.³⁰ Side effects included severe nausea and tremor. The incidence of side effects is dosage dependent. To reduce the side effects of mexiletine, the dose must be lowered while preserving its blocking efficiency. A combined antiarrhythmic therapy that promotes DIII-VSD activation could resolve this challenge. Moreover, the patient-specific responses can be attributed to disparity in expression of Na_V1.5 isoforms (polymorphisms) and accessory Na_V β subunits.^31,32 The impacts of common polymorphisms³³ and interacting proteins on Na_V channels’ responses to mexiletine need to be explored in the future studies.

Data-driven modeling is commonly applied in the field of systems biology, due to the large-scale nature of nonintuitive experimental data from biological assays, such as microarray and gene sequencing.³⁴ Although the scale of data from channel electrophysiology recordings is much smaller, data-driven modeling can still be useful, because ion channels themselves are complex systems, in which many parts of the channel work in concert to generate time- and voltage-dependent gating. Different voltage protocols can isolate different channel gating processes. Although effective, these voltage protocols are not ideal, because the properties they measure overlap. For example, an inactivation change can affect protocols that measure activation by shutting down channels before the channels open maximally.³⁵ Statistical tools that can reduce data dimensionality, such as principal component analysis and PLS regression can reduce redundant information in data recorded with different voltage protocols.

In addition to dimensionality reduction, the methods that we applied in this study have the advantage of recognizing the multivariate relationships between input (independent) and output (dependent) variables. It rotates the input data to new optimal dimensions that maximize the covariance between input and output data. Thus, PLS regression models can be trained with existing data and then used to predict output of new input data. In this study, we built PLS regression models that predict mexiletine response using channel gating parameters. Because other Class I drugs, such as lidocaine and flecainide, have distinct chemical properties that potentially result in distinct interactions with the channel, the PLS regression model we developed for mexiletine cannot be directly implemented to predict patients’ responses to other drugs. However, the same methodology can be applied to predict other Class I drug responses. In the future, if a series of PLS regression models for common Class I drugs is established, this series could be used to predict a patient’s response to available drugs. When a patient is identified with a new genetic variant, a patient’s response to various drugs could be predicted with the established PLS regression models. Moreover, a systems biology approach has the advantage of testing the outcomes of a combination of scenarios.³⁶ Different drug prediction models can be combined to discover an optimal therapy for a certain patient.

Mexiletine is widely used as an antiarrhythmic drug for patients with LQT3 syndrome and VT. Patient-specific responses to mexiletine have been observed in clinical studies. However, the underlying mechanism is not well defined because of the lack of understanding of mexiletine’s molecular interactions with the Na_V1.5 channel. We demonstrate that the conformation of the DIII-VSD of the Na_V1.5 channel is crucial for determining mexiletine blockade. Using biophysical data and a systems biology approach, we built a model that can predict LQT3 patient QT_c shortening with mexiletine therapy based on channel gating parameters. Our findings also suggest a new antiarrhythmic strategy of combination therapies that target molecular conformations of the channel to increase mexiletine efficacy.