It is well known that the hERG channel resides in discrete functional states, minimally comprising closed, open, and inactivated states, which interconnect as a function of time and membrane voltage (Vandenberg et al., 2012). In the open state, the channel conducts K⁺ ions through a central pore. In contrast, the closed and inactivated states are non-conductive due to either a constricted pore at the intracellular gate (closed state) or a distorted selectivity filter (SF; inactivated state) (Vandenberg et al., 2012). So far, published experimental cryo-EM structures resolved the channel in an open state (Asai et al., 2021; Wang and MacKinnon, 2017). Starting with the experimental structure, computational studies explored hERG inactivation by simulating how different membrane voltages can change the SF and thus affect ion conduction through the channel (Li et al., 2021b; Miranda et al., 2020; Yang et al., 2020). These studies are essential but have some limitations, as the high voltages applied can force the channel into unnatural conformations, and the simulations are not long enough to allow observation of state transitions (Shi et al., 2020). To overcome these limitations, we adopted different modeling strategies to guide AlphaFold2 in producing diverse conformations relevant to specific functional states of the hERG channel.

The first modeling strategy involves using a structural fragment from an experimental structure of a homologous protein that exhibits the desired characteristics of the target state we aim to model in our channel. This fragment serves as a structural template, and AlphaFold2 is used to rebuild the rest of the channel while adhering to the constraints of the template. For example, to model the closed state of the hERG channel, it is known that channel closure requires the voltage sensor in the voltage-sensing domain to be in a deactivated conformation. To achieve this, we used the deactivated voltage-sensing domain from the closed-state rat EAG channel cryo-EM structure (PDB 8EP1, residues H208–H343) (Mandala and MacKinnon, 2022) and combined it with the SF from the open-state hERG cryo-EM structure (PDB 5VA2, residues I607–T634) (Wang and MacKinnon, 2017). This hybrid structure was used as the template for AlphaFold2 predictions, as illustrated in Figure 1a. Using these discrete structural fragments, AlphaFold2 was then applied to generate 100 models, specifically configured to encourage diverse prediction outcomes for further analysis.

For modeling the open hERG channel state, we utilized the existing cryo-EM structure of hERG (PDB 5VA2) (Wang and MacKinnon, 2017) and rebuilt the missing extracellular loops using Rosetta (Fleishman et al., 2011) with the results shown in Figure 1b. This reconstructed model served as a basis for MD and drug docking simulations.

The second modeling strategy addresses situations where structural information about state transitions is either limited or inconsistent. In this approach, we erase regions expected to undergo changes during state transition from an existing structure and use AlphaFold2 to sample potential conformations for these regions. This allows AlphaFold2 to identify possible substates, which are then grouped into clusters of structurally similar models for further analysis. For example, during hERG inactivation, the SF shifts from an open, conductive to a distorted, non-conductive conformation, as shown by numerous studies on hERG and other K⁺ channels (Butler et al., 2019; Cuello et al., 2010; Fan et al., 1999; Li et al., 2021b; Pettini et al., 2023; Schönherr and Heinemann, 1996; Tan et al., 2022; Wu et al., 2025). Moreover, there are a number of studies that do not uniformly suggest a single discrete structure of the inactivated-state SF but propose several alternative conformations (Lau et al., 2024; Li et al., 2021b).

To model the inactivated state of hERG, first we configured AlphaFold2 to introduce more uncertainty into the sampling process. As illustrated in Figure 1c, starting with the open-state cryo-EM structure (PDB 5VA2) (Wang and MacKinnon, 2017), we removed everything except for the cytosolic domain (S660–R863), then let AlphaFold2 reconstruct the transmembrane domain. In half of the resulting predictions, including the top-ranked model by prediction confidence, the SF showed a distinct lateral flip of the backbone carbonyl oxygens at residue V625 compared to the open-state structure. This flip created a potential barrier that could prevent K⁺ ions from crossing between the S3 and S2 ion-binding sites. To further investigate this conformation, we extracted the predicted SF region (residues Y607–T634) and merged them with the activated VSDs (W398–V549) and the cytosolic domain (S660–R863) from the open-state hERG structure (PDB 5VA2) (Wang and MacKinnon, 2017) to create a new structural template. Then, we generated 100 more models for further analysis.

A key distinction of our approach is that, rather than relying solely on single-model predictions, we generated a diverse population of models to better explore the conformational landscape. By clustering these models, we identified predominant substates, represented as clusters of structurally similar models. To determine which of these substates are likely to be physiologically relevant, we quantitatively assessed the structural reliability within each cluster using the predicted Local Distance Difference Test (pLDDT). Higher pLDDT scores indicate more reliable and accurate structural predictions (Jumper et al., 2021a). This clustering approach helps to capture a range of conformations that might represent stable states of the protein.

For each predicted conformational state of hERG, we clustered 100 predicted structural models based by their degree of similarity, quantified by the root-mean-square deviation (RMSD), as shown in Figure 2.

After further structural refinement in Rosetta (Fleishman et al., 2011; Leman et al., 2020) to resolve steric clashes, the resulting models are compared in Figure 3. In Figure 3a, b, the closed-state model displayed the most constricted channel pore, followed by the inactivated state and then the open-state model. In the pore-lining S6 helix, the canonical drug-binding residue Y652 (Vandenberg et al., 2012) retains a relatively consistent position with minor variation across all channel state models. The rotation and shift of the S6 helix in the inactivated and closed states affect the position of another canonical drug-binding residue, F656 (Vandenberg et al., 2012). The adjustment caused the F656 side chain to extend more into the hERG inner cavity in both the closed and inactivated states, compared to the open state.

We performed all-atom MD simulations on two hERG channel models described above, one in the open state and the other in the predicted inactivated state, to evaluate their ion conduction capabilities. Unlike the closed-state model, both open- and inactivated-state models should allow ions and water to enter and traverse the channel pore reaching the SF region. However, only the open-state model is expected to facilitate ion conduction through its SF.

To investigate ion conduction in the SF, we considered two conditions, as shown in Figure 5—figure supplement 1a one in which the SF initially contained only K⁺ ions and another in which both ions and water were present to test previously proposed direct (or Coulombic) and water-mediated K⁺ conduction knock-on mechanisms (Lam and de Groot, 2023; Roux, 2017) as in a previous study (Miranda et al., 2020). In the direct knock-on (ions-only) scenario, we manually positioned K⁺ atoms in the putative K⁺-binding sites of S0, S2, S3, and S4 within the SF. For the water-mediated knock-on (the alternating ions and water molecules) scenario, K⁺ ions were placed in the S1, S3, and S_cav positions, while water molecules were inserted into the S0, S2, and S4 positions. These models were incorporated into phospholipid bilayers consisting of 1-palmitoyl-2-oleoylphosphatidylcholine (POPC) molecules and hydrated by 0.30 M KCl, as depicted in Figure 5—figure supplement 1b. Subsequently, we conducted MD simulations for each case under three membrane voltage conditions: 0, 500, and 750 mV, each lasting 1 μs. This resulted in a total of six MD simulations for each model.

In all instances where a non-zero membrane voltage was applied after equilibration, we observed K⁺ conduction for the open-state model (Figure 5a, c and Figure 5—figure supplement 2a, c), whereas such conduction was not observed for the inactivated-state model (Figure 5b, d and Figure 5—figure supplement 2b, d). For ions-only initial SF arrangement, we observed that all K⁺ ions initially located in the SF went across during 1 μs MD runs under applied 750 and 500 mV membrane voltages (Figure 5a and Figure 5—figure supplement 2), whereas for the alternating water-ion initial SF configuration, we observed conduction of SF ions as well as additional K⁺ ions moving all the way across the channel pore (Figure 5c and Figure 5—figure supplement 2c). In both cases, we saw a combination of direct and water-mediated knock-on mechanisms, as in our previous hERG channel MD simulations (Miranda et al., 2020; Yang et al., 2020). Control MD simulations conducted under zero voltage conditions revealed a single K⁺ SF conduction event for the open-state model (Figure 5—figure supplement 2g) when the SF was initially filled with water molecules and ions, while no conduction events were observed in the remaining cases (Figure 5—figure supplement 2e, f, h).

Cuello et al. in their study of KcsA channel identified a similar constriction at G77 within the SF and a corresponding reorientation of the V76 carbonyl, resulting in a dilation in the SF at this location and corresponding loss of the S2 and S3 ion-binding sites (Cuello et al., 2010). They suggested this backbone rearrangement as a fundamental molecular mechanism underlying C-type inactivation in K⁺ channels (Cuello et al., 2010). In other studies on Shaker and K_v1.3 channels, dilation in the upper SF that disrupts the S1 and S2 K⁺-binding sites has been proposed to be a potential C-type inactivation mechanism (Chandy et al., 2023; Selvakumar et al., 2022; Tan et al., 2022; Tyagi et al., 2022). Similar dilations in the SF are also predicted by AlphaFold2, particularly within Cluster 3 of the predicted inactivated-state hERG channel clusters shown in Figure 2a. Although these models were not simulated under our study, such dilated conformations of the SF also emerged during our MD simulations of the open-state model under applied voltage.

We further compared SF conformations by plotting the cross-subunit distances between the carbonyl oxygen atoms of SF residues of open-state hERG channel MD simulations at 750 and 500 mV applied voltages, as shown in Figure 5—figure supplement 6. The dilation observed in the hERG channel, which also occurs in the upper SF, differs from that in the aforementioned K⁺ channels. In Shaker-family channels, the most considerable widening occurs at the SF tyrosine residue (Y445 in Shaker/Y377 in K_V1.2) immediately below the topmost SF residue (G446 in Shaker/G378 in K_V1.2) (Tan et al., 2022; Wu et al., 2025). Conversely, in the hERG channel, the topmost SF residue (G628) exhibits the most significant widening, followed by the residue immediately below it (F627). Our MD simulations of the hERG channel reveal that its dilation process involves two sequential steps: SF near residues F627 dilates first, followed by SF near topmost G628 residues. The latter step occurs faster at higher voltages (750 mV) compared to lower voltages (500 mV). We present these steps in Figure 5—figure supplement 7. Notably, despite the dilation of the hERG SF, ion conduction is still observed across all replicas, in contrast to the Shaker channel (Chandy et al., 2023; Selvakumar et al., 2022; Tan et al., 2022; Tyagi et al., 2022).

We utilized Rosetta GALigandDock software (Park et al., 2021) to dock 19 drugs from different classes, considering their multiple protonation states, into our hERG state-specific channel models. This process aimed to evaluate and corroborate state-dependent binding interactions with experimental studies, specifically in terms of relative binding affinities. Figure 6—figure supplement 1 presents these findings in the form of Rosetta GALigandDock (Park et al., 2021) binding energies (lower, more negative values mean more favorable binding). Consistent with published studies, most drugs showed stronger binding to the inactivated-state hERG channel model, including astemizole, terfenadine, cisapride, d/l-sotalol, dofetilide (Ficker et al., 1998; Kamiya et al., 2008; Perrin et al., 2008), haloperidol (Suessbrich et al., 1997), and E-4031 (Numaguchi et al., 2000; Wang et al., 1997). Drugs like moxifloxacin (Alexandrou et al., 2006), quinidine (Perrin et al., 2008), verapamil (Duan et al., 2007), and perhexiline (Perrin et al., 2008) did not show strong preference for the inactivated-state model, aligning with findings from hERG experimental studies using inactivation-deficient mutants (Perrin et al., 2008) or ‘step-ramp’ voltage protocol (Alexandrou et al., 2006).

As a control, we also included docking results for the presumed open-state model from inactivated-state-sampling Cluster 3 (referred to as Open, AF ic3). Although its SF differs from both the experimental open and predicted inactivated-state models, our previous simulations confirmed that it supports ion conduction. Structurally, its pore most closely resembles the open state (Figure 5—figure supplement 5b) with only minor differences, and accordingly, its drug docking profile aligns well with that of the open-state model. These results further support the interpretation that the Cluster 3 model represents an alternative open-state conformation. Table 3 provides an overview of all models examined, along with qualitative insights into their observed behaviors thus far.

In our GALigandDock docking results, most drugs exhibited increased binding affinity to the closed-state hERG channel model compared to the open-state hERG channel model. Drugs are unable to bind to the closed state from the intracellular space because the pore is closed. However, they can become trapped if they are already bound when the channel transitions from an open to a closed state, as shown in experiments for dofetilide (Windley et al., 2017), cisapride (Windley et al., 2017), terfenadine (Windisch et al., 2011; Windley et al., 2017), E-4031 (Windisch et al., 2011), and nifekalant (Kamiya et al., 2006).

To model drug trapping, we placed the drug in a pocket beneath the SF in the closed pore configuration before docking. However, this method does not consider how the conformational shift from the open to the closed state might influence drug binding. Under physiological conditions, the pore gating motion from open to closed might expel drugs from the pore instead of pushing them deeper. This limitation might account for some inconsistencies noted in our docking study, particularly regarding the apparent trapping of drugs such as amiodarone and haloperidol, which is at odds with experimental results (Stork et al., 2007). However, these preliminary results could pave the way for more thorough investigations, employing advanced computational techniques to delve deeper into the dynamics of drug trapping (Branduardi and Faraldo-Gómez, 2013; Miao et al., 2020).

Figure 6 highlights the binding profiles of terfenadine, dofetilide, and moxifloxacin. Terfenadine and dofetilide are modeled in their cationic forms, while moxifloxacin is in its zwitterionic form. Experimental evidence (Ficker et al., 1998; Kamiya et al., 2008; Perrin et al., 2008) indicates that terfenadine and dofetilide preferentially bind to the inactivated state of hERG, whereas moxifloxacin does not show this state-specific preference. Notably, both terfenadine and dofetilide have been associated with TdP arrhythmia and have been withdrawn or restricted in clinical use, while moxifloxacin is generally considered safer (Alexandrou et al., 2006; Jaiswal and Goldbarg, 2014; Monahan et al., 1990; Orvos et al., 2019; Yang et al., 2020). Here, we investigate whether molecular differences in state-dependent binding modes, particularly to the inactivated state, and corresponding differences in binding affinities may help explain their varying proarrhythmic risks.

There are several complications that make it difficult to directly compare experimental binding affinities with predicted affinities from simulations. During electrophysiological recordings of hERG inhibition by various drugs, the hERG channel has been shown to adopt various functional states, presumably corresponding to protein conformation states. These states can be bound by drugs with varying affinities, with drug ionization state also being a contributing factor. Additionally, the variability in experimental protocols affects the measured affinities (Gomis-Tena et al., 2020). In general, electrophysiological measurements report the IC₅₀, the drug concentration required for 50% inhibition of current. However, the IC₅₀ value is not directly comparable to computed affinities from drug docking.

To address these challenges, we developed a novel computational approach that combines modeling and simulation to predict hERG channel conformational state probabilities (open, closed, and inactivated) over time. First, we collected a comprehensive set of experimental data and employed a hERG functional model with five functional states, which was extensively validated in our earlier study (Romero et al., 2015). For each drug, we ran in silico electrophysiological experiments under the same conditions as the experimental studies, allowing us to calculate the relative probabilities of the various hERG channel states specific to the drug and protocol. These state probabilities were then used to refine the computed binding affinities from docking simulations. We adjusted the affinities for both neutral and charged forms of each drug according to their prevalence in each conformational state. This method allowed us to scale the binding predictions based on the likelihood of each channel state occurring during the experimental protocols. Finally, we compared the simulated binding affinities with experimental hERG drug potencies (Table 4), offering a new validation technique that enhances the accuracy of our predictions and helps reconcile the differences between experimental IC₅₀ measurements and computed affinities.

We compared the experimental drug potencies with the simulated binding affinities, starting with the traditional approach of using only open-state docking simulations (Figure 7a, c), commonly employed in ion channel pharmacology due to a scarcity of multi-state models, and then extended our analysis to include drug binding to different states (Figure 7b, d). Using only the open-state model (PDB 5VA2) yielded a moderate correlation with experimental data (R² = 0.43, r = 0.66, Figure 7a). Incorporating multi-state binding (weighted by their experimental distributions) improved the correlation substantially (R² = 0.63, r = 0.79, Figure 7b), boosting predictive power by 47% and underscoring the value of multi-state modeling. Importantly, this improvement was achieved without considering potential drug-induced allosteric effects on hERG channel conformation and gating, which will be addressed in future work.

Next, we substituted the PDB 5VA2-based open-state model with the AF ic3 open-state model. Docking to this alternative model alone produced similar performance (R² = 0.44, r = 0.66, Figure 7c), and incorporating it into the multi-state ensemble further improved the correlation with experiments (R² = 0.64, r = 0.80, Figure 7d), representing a 45% gain in R² and matching the performance of multi-state docking results based on the PDB 5VA2-derived model.

These findings suggest that the predictive power of computational drug docking is enhanced not merely by the accuracy of individual models, but by the structural diversity and complementarity provided by an ensemble of conformations. Rather than relying solely on a single experimentally determined structure, the ensemble benefits from incorporating AlphaFold-predicted models that capture alternative conformations identified through our state-specific sampling approach. These diverse models reflect different structural features, which together offer a more comprehensive representation of the channel’s binding landscape and enhance the predictive performance of computational drug docking. Overall, these results reinforce that multi-state modeling offers a more realistic and predictive framework for understanding drug-channel interactions than traditional single-state approaches, emphasizing the value of both individual model evaluation and their collective integration.

In this study, we introduced a methodology to extend AlphaFold2’s predictive capabilities by guiding it to sample multiple protein conformations reminiscent of different ion channel states, using the hERG potassium channel as a proof of concept. While other studies have focused on generating diverse protein conformations using AlphaFold (Del Alamo et al., 2022; Kalakoti and Wallner, 2024; Lidbrink et al., 2024; Sala et al., 2023), we aimed to take this further by validating the physiological relevance of these predicted states for ion channels through computational simulations and experimental data. By incorporating multiple structural templates and refined input parameters, we directed AlphaFold2 to predict distinct functional states, including the closed and inactivated forms of hERG. This approach is significant for pharmacology, particularly for the hERG channel, an anti-target notorious for state-specific drug interactions linked to cardiotoxicity (Li and Ramos, 2017; Perrin et al., 2008; Sanguinetti and Tristani-Firouzi, 2006; Vandenberg et al., 2012; Yang et al., 2020).

We employed two complementary strategies to guide AlphaFold2 in predicting physiologically relevant hERG channel conformations. In the first, we used structural fragments representative of specific functional states as templates, prompting AlphaFold2 to reconstruct the full hERG channel accordingly, successfully yielding models with features characteristic of the closed state, such as a constricted pore and deactivated voltage-sensing domains. In the second, we allowed AlphaFold2 to explore a broader conformational space, particularly useful in scenarios where structural knowledge is limited. Remarkably, AlphaFold2 managed to generate conformations that had been previously documented in the literature but not included in the training dataset, showcasing its broad predictive capacity. Among the most compelling findings is the strong correspondence between our AlphaFold2-predicted inactivated-state conformations and those previously proposed through experimental and simulation studies (Lau et al., 2024; Li et al., 2021b). These structures were identifiable in discrete protein clusters (Figure 2a) with high prediction confidence metrics and likely together constitute components of the hERG inactivation mechanism. In this way, AlphaFold2 provided a strong method to reconcile apparently disparate previous experimental and computational data.

Together, these strategies produced models that captured key features of closed, open, and inactivated channel states. Our study goes beyond modeling distinct hERG channel conformations; it also provides extensive validation of their functional and pharmacological relevance through MD simulations and integrated docking analyses. The closed-state model exhibits a constricted pore and deactivated voltage-sensing domains, providing a structural basis for hERG channel deactivation. The open-state model displays a widened, conductive pore consistent with physiological ion flow during cardiac repolarization. The inactivated-state model features a non-conductive SF primarily characterized by SF residue V625 carbonyl flipping away from the conduction pathway. It also reveals notable rearrangements in the pore cavity that enhance drug binding, which is consistent with experimentally measured increases in drug affinity and arrhythmia risk (Ficker et al., 1998; Perrin et al., 2008; Perrin et al., 2008; Yang et al., 2020), yet previously lacking a defined structural explanation. Nonetheless, these alterations may not occur concurrently under physiological conditions and could represent only a subset of the broader conformational changes that accompany channel inactivation.

To demonstrate the broader feasibility of this approach, we applied it to another ion channel system, Na_V1.5, as illustrated in Figure 1—figure supplement 1. In this example, a deactivated VSD II from the cryo-EM structure of Na_V1.7 (PDB 6N4R) (Xu et al., 2019), which was trapped in a deactivated state by a bound toxin, was used as a structural template. This guided AlphaFold to generate an Na_V1.5 model in which all four voltage sensor domains (VSD I–IV) exhibit S4 helices in varying degrees of deactivation. Compared to the cryo-EM open-state Na_V1.5 structure (PDB 6LQA) (Li et al., 2021a), the predicted model displays a visibly narrower pore, representing a plausible closed state. This example underscores the versatility of our strategy in modeling alternative conformational states across diverse ion channels.

While previous studies have proposed potential molecular mechanisms for hERG C-type inactivation (Lau et al., 2024; Li et al., 2021b), they have primarily concentrated on conformational shifts leading to a non-conductive SF. Our study takes a step further and sheds light on how these alterations extend to the pore region and subsequently impact drug binding as seen in experiments, which is an outstanding issue in safety pharmacology and drug development as hERG channel is a notorious drug anti-target (Perrin et al., 2008; Sanguinetti and Tristani-Firouzi, 2006; Vandenberg et al., 2012; Yang et al., 2020).

Our results suggest that specific structural rearrangements occurring during the open-to-inactivated transition of the hERG channel may fine-tune the geometry of the central cavity in ways that enhance ligand binding. In particular, the movement of the canonical drug-binding residue Y652, together with the reorientation of the S624 sidechain hydroxyl groups, slightly expands the hydrophobic pockets beneath the SF (Figure 3a, b, at Z = 21 Å), allowing certain drugs to bind more deeply. Additionally, a slight inward rotation of the S6 helix repositions F656 toward the pore axis, which likely enhances drug interactions through π-stacking and hydrophobic contacts and helps to reduce the likelihood of drug dissociation from the cavity. These changes may not occur simultaneously or uniformly and may vary in magnitude, as suggested by the range of conformations observed in the AlphaFold models. Importantly, this local remodeling likely represents just one aspect of the broader inactivation process predicted by AlphaFold and validated in our study, contributing to drug stabilization without necessarily capturing the full inactivated state.

Together, these structural changes may account for the increased binding affinity observed for some compounds in the hERG channel inactivated state (Figure 6—figure supplement 1), offering a structural mechanism that links this enhanced binding to elevated arrhythmogenic potential (Ficker et al., 1998; Jaiswal and Goldbarg, 2014; Perrin et al., 2008; Yang et al., 2020). It is important to note that drug interaction with other cardiac ion channels may modulate or offset this risk in vivo, and the net proarrhythmic liability is determined by a more complex interplay of multi-channel effects (Colatsky et al., 2016). Nevertheless, given the exceptional sensitivity of the hERG channel to a wide range of compounds, understanding its state-dependent binding mechanisms remains critical for predicting and mitigating cardiac safety risks during drug development.

An additional novelty of our study was the development of a new computational modeling and simulation approach that allowed us to use the predicted affinities of drug binding from docking simulations and compare them directly to measured hERG channel inhibition potencies from electrophysiology experiments. We employed a simulated hERG functional model comprising five functional states that we have extensively validated in our earlier studies (Romero et al., 2015) and performed in silico electrophysiological experiments under the same conditions as reported in the experimental papers. In doing so, we could compute the relative hERG channel state probabilities during the experimental protocol. The channel state probabilities were then used to scale the computed affinities in each state from the docking simulations, allowing comparison of the overall computed relative affinity with experimentally reported relative potencies. Such an analysis created an opportunity for an ‘apples-to-apples’ comparison between structurally derived affinity predictions and abundant functional measurements that have been conducted for half a century. This novel linkage can be readily expanded to other protein targets and any variety of drugs. The intersection of structural modeling, molecular docking, functional simulations, and supporting experimental data offers a comprehensive approach to understanding protein structures and their links to biological functions.

For other K⁺ channels, dilation in the SF has been proposed to be potential C-type inactivation mechanisms (Chandy et al., 2023; Cuello et al., 2010; Selvakumar et al., 2022; Tan et al., 2022; Tyagi et al., 2022). While inactivation processes across various K⁺ channels may share some similarities, the associated conformational changes can adopt distinct differences due to small variations in the SF sequences, which could explain the observed variability in inactivation rates among hERG and other ion channels (Vandenberg et al., 2012). As an example, a study on Shaker-family channels suggested that a two-step widening process in the upper SF could be a mechanism for C-type inactivation (Wu et al., 2025). The SF of the Shaker-family channel has the sequence ‘TVGYGD’. The first step involves a partial twist of the P-loop backbone, particularly involving the upmost SF residue D30’ (D379 in K_V1.2), which originally stabilizes itself by interacting with W17’ (W336 in K_V1.2). The second step is the reorientation of the upper-middle SF residue Y28’ (Y377 in K_V1.2) upward, which normally participates in hydrogen bonds with nearby pore-helix residues, to fill some of the original volume occupied by D30’.

In contrast, the hERG SF is characterized by the sequence ‘SVGFGN’. In the upper SF of the hERG channel, phenylalanine (F) replaces the tyrosine (Y) found in many K⁺ channels including Shaker and K_V1.2. Phenylalanine has a non-polar benzyl side chain, whereas tyrosine has a polar hydroxyl group attached to its benzene ring. The hydroxyl group in tyrosine can form additional hydrogen bonds, which may stabilize different SF conformations in other K⁺ channels compared to hERG. Similarly, at the outermost end of the SF of hERG, asparagine (N) replaces the aspartate (D) found in many other K⁺ channels. Asparagine is uncharged, while aspartate introduces a negative charge through its carboxylate group, which could explain the differences in ion coordination and gating dynamics. These structural differences may explain why the hERG channel adopts a similar but distinct SF rearrangement compared to other K⁺ channels (such as Shaker and K_V1.2) and can have a slightly different structural mechanism of the C-type inactivation.

Despite the promising results, our study is not without limitations. While AlphaFold has demonstrated remarkable accuracy in numerous instances (AlQuraishi, 2019; Jumper et al., 2021b; Varadi and Velankar, 2023), it is important to note that the predicted models may not always be reliably accurate to assess drug binding (Karelina et al., 2023). Moreover, hERG channel inactivation and closure might encompass a range of states, as was shown for other ion channels (Catterall et al., 2020; Goldschen-Ohm et al., 2013; Hite and MacKinnon, 2017; Li et al., 2024; Yao et al., 2024), and the conformations we have identified could potentially represent just a few possibilities within this broad spectrum. Our models excluded the N-terminal PAS domain due to GPU memory limitations, despite its inclusion in initial templates. This omission may overlook its potential roles in gating kinetics and allosteric effects on drug binding (Abi-Gerges et al., 2011; Goversen et al., 2019; Gustina and Trudeau, 2013; Harchi et al., 2018; Perissinotti et al., 2018). Future research will explore the full-length hERG channel with enhanced computational resources to assess these regions’ contributions to conformational state transitions and pharmacology.

As noted in recent studies, pLDDT scores are not reliable indicators for selecting alternative conformations (Bryant and Noé, 2024; Chakravarty et al., 2024). To address this, we performed a protein backbone dihedral angle analysis in the regions of interest to ensure that our evaluation captured a representative range of sampled conformations. GALigandDock docking results, while insightful, are provisional (Maly et al., 2022) and limited by a rigid protein backbone assumption, thus preventing observation of drug-binding-induced allosteric modifications (Harris et al., 2024). As such, the results presented here should be interpreted as qualitative indicators of state-dependent binding trends rather than definitive quantitative predictions. To achieve more accurate binding affinity estimates, future studies could leverage MD simulations, incorporating methods like Molecular Mechanics/Poisson–Boltzmann Surface Area to assess relative ligand-binding energies during MD trajectories (Miller et al., 2012; Ngo et al., 2024; Wang et al., 2016). These insights could be further extended by integrating MD results with multiscale functional modeling approaches, as demonstrated in our earlier work (DeMarco et al., 2021; Yang et al., 2020).

Our approach currently relies on well-characterized systems with ample static structures, MD simulation data, and mutagenesis insights, as demonstrated with the hERG channel, which may limit its applicability to less-studied proteins. Recently, AlphaFold3 was released, incorporating a diffusion model that enables the prediction of proteins in complex with other proteins, small molecules, nucleic acids, and ions (Abramson et al., 2024). We plan to explore the applicability of our template-guided methodology in a follow-up study, leveraging AlphaFold3’s advanced diffusion-based architecture to enhance protein conformational state predictions and state-specific drug docking, particularly given its improved capabilities for modeling small molecule–protein interactions.

Correlating simulated drug-binding affinities with experimental results remains inherently challenging. As demonstrated in multiple studies, drug-binding potency is highly dependent on the measurement technique used, resulting in different IC₅₀ values being reported for the same channel–drug pairing (Alexandrou et al., 2006; Asai et al., 2021; Cameron et al., 2021; Chiu et al., 2004; Ficker et al., 1998; Hansen et al., 2006; Johnson and Trudeau, 2023; Kushida et al., 2002; Orvos et al., 2019; Paul et al., 2002; Perrin et al., 2008; Rampe et al., 1997; Suessbrich et al., 1997; Tanaka et al., 2014; Walker et al., 1999; Zhang et al., 1999; Zhou et al., 1998) as was explored in detail in our recent study (Gomis-Tena et al., 2020). Additionally, knowing the binding free energies of a drug is not the complete story; binding rates such as k_on (association rate) and k_off (dissociation rate) are also crucial for a quantitative evaluation of drug binding to the channel.

In conclusion, this study advances our understanding of hERG channel structural dynamics and state-dependent drug binding, while also demonstrating the broader potential of AlphaFold2-based modeling workflows. Our findings provide a foundation for integrating deep learning-based structure prediction with simulation and functional modeling to study other ion channels and membrane proteins. As computational methods continue to evolve, including alternatives like RoseTTAFold (Baek et al., 2021) and ESMFold (Lin et al., 2023), such integrated approaches will be increasingly valuable for addressing complex questions in ion channel physiology and pharmacology, with important implications for cardiac drug safety and therapeutic development.

AlphaFold2 employs a deep learning architecture that integrates several innovative components, including residue pair representation comprising an architecture module that represents each possible pair of amino acid residues in the sequence. This pairwise representation captures the interactions between residues that determine protein folding. AlphaFold2 also applies an attention mechanism, which constitutes a transformer-based model (similar to the architecture used in natural language processing) to weigh the influence of different parts of the input data differently (Jumper et al., 2021a). In AlphaFold2, the effect is to emphasize interactions between certain amino acid residues more than others based on how they might impact folding. There is also a so-called Evoformer block within the learning model that specifically processes the evolutionary data from MSAs, enabling the model to effectively incorporate evolutionary information (Jumper et al., 2021a). After processing through the Evoformer, intermediate representations are used to predict the distances and angles between residues as part of an iterative feedback process. A critical feature of AlphaFold2 is its iterative refinement, where pairwise residue representations, MSAs, and initial structural predictions are recycled through the model multiple times, improving accuracy with each iteration.

Due to the time-intensive process of creating MSAs for AlphaFold2, the ColabFold (Mirdita et al., 2022) webserver was made to streamline protein structure prediction by combining MMseqs2 (Mirdita et al., 2019) sequence search toolkit with AlphaFold2, enhancing runtime efficiency while preserving high prediction accuracy. ColabFold is available at https://github.com/sokrypton/ColabFold (Ovchinnikov and Steinegger, 2025).

We modeled the hERG potassium channel in three functional states (closed, open, and inactivated) using ColabFold with tailored structural templates and configurations. The structural templates were assembled by first using ChimeraX (Pettersen et al., 2021) to superimpose all relevant PDB entries with the ‘matchmaker’ command. Unnecessary regions were then removed from the aligned models, leaving only the fragments intended for use as templates. These remaining segments were merged into a single model using the ‘combine’ command. This approach preserves the spatial arrangement of template regions as they are expected to appear in the AlphaFold prediction, which can be important for guiding the model toward a specific conformation. The following sections describe the template selection process for each conformational state.

Molecular graphics visualization was performed using ChimeraX (Pettersen et al., 2021). MD trajectory and simulation images were visualized using VMD (Humphrey et al., 1996). Interaction network analysis was performed using the Protein-Ligand Interaction Profiler (PLIP) (Salentin et al., 2015) with criteria outlined in Table 1.