Modeling hERG Channel Liability: From Structural Insight to Highly Accurate Qualitative and Quantitative Models

Abstract

Drug-induced QT interval prolongation, most commonly resulting from the blockade of a voltage-dependent potassium ion channel encoded by the hERG (human ether-à-go-go–related gene), has been recognized as a critical side-effect of noncardiovascular therapeutic agents. This adverse effect has led to withdrawal of many drugs from the market. Early identification of potential hERG channel blockers is therefore essential to mitigate cardiotoxicity-related attrition during the later, more resource-intensive stages of drug development. In this paper, we aimed at understanding ligand-channel interactions, including a detailed analysis of the cryo-electron microscopy (cryo-EM) structures of hERG channels and pharmacophore models shared among known hERG blockers. The highly adaptive nature of the hERG ligand-binding site may poses challenges for structure-based approaches, such as molecular docking, yet also offers mechanistic insights into a longstanding question: why does hERG interact with such a wide variety of small-molecule drugs? To complement these structural observations, we summarized the benefits and limitations of both quantitative and qualitative models and their applications across various stages of drug discovery. We developed highly predictive classification and regression models built using customized atom-type descriptors. The regression model, trained on a large and curated data set (∼8,000 compounds), achieved an average absolute error (AAE) of 0.383 log units and root-mean-square error of prediction (RMSEP) of 0.548 log units on the test sets. Meanwhile, the classification model demonstrated strong performance as well, with a receiver operating characteristic (ROC) area under the curve (AUC) of 0.88. Validation on an external set of 1,133 compounds resulted in an AAE of 0.50 log units. Together, these complementary modeling strategies can significantly aid in the early assessment of cardiovascular liabilities associated with hERG channel blockade, thereby supporting safer and more efficient drug development.

Introduction

In drug discovery, the human ether-à-go-go-related gene (hERG, Kv11.1) potassium channel is a critical safety antitarget due to its central role in cardiac repolarization. The KCNH2 gene encodes the hERG1 protein, which conducts the rapid delayed rectifier potassium current (IKr) that contributes significantly to the repolarization phase of cardiac action potentials. Blockade of this channel by small molecules can lead to prolonged QT intervals and severe arrhythmias, most notably torsade de pointesposing a major cardiotoxicity risk.

The significance of hERG-related cardiotoxicity in drug development cannot be overstated. Several high-profile drugsincluding terfenadine, astemizole, cisapride, and haloperidolhave been withdrawn from the market due to their adverse cardiac effects. , This has prompted regulatory bodies like the U.S. Food and Drug Administration (FDA) and the International Council for Harmonisation (ICH) to recommend early evaluation of hERG activity for all drug candidates destined for human use.

Structurally, the first cryo-electron microscopy (cryo-EM) structure of the hERG channel (PDB: 5VA2) was resolved in 2017, using a truncated construct to minimize aggregation. A few years later, the first structure with a bound small moleculeastemizolerevealed a unique structural adaptivity of the channel, shedding light on why structurally diverse molecules across various therapeutic classes (e.g., antihistamines, antipsychotics, antianginals, antibiotics, and antidepressants) can all act as hERG blockers (Figure ). More recently, the cryo-EM structures of hERG in complex with three different hERG blockers, astemizole, E4031, and pimozide, became available, allowing comparison of intermolecular interactions between hERG channel and different inhibitors.

1.

Chemical structures of hERG blockers with varying activities and different therapeutic indications. The three ligands observed in cryo-EM hERG structures–astemizole, E-4031, and pimozide −together with terfenadine and fexofenadine, all share a piperidine core.

In vitro methods to assess hERG inhibitionparticularly manual patch-clamp electrophysiologyremain the gold standard. However, due to its low throughput, automated patch-clamp systems and high-throughput assays like the rubidium flux method have gained traction. ,, These assays have enabled large-scale screening of drug candidates, although differences in assay conditions, cell lines (e.g., CHO vs HEK293), and detection sensitivity introduce variability in reported IC₅₀ values.

A major bottleneck in early drug discovery is the timely identification of hERG liability. Computational predictive models serve as valuable tools to flag potential cardiotoxic compounds prior to synthesis or biological testing. Early models were built on limited data sets and often relied on pharmacophore or binary classification approaches. − Recent advances in machine learning, including deep neural networks and ensemble methods like XGBoost, combined with the availability of larger curated data sets (e.g., from ChEMBL), have significantly improved predictive accuracy and expanded the models’ applicability domain. −

While classification models are commonly used to filter compounds in early stage screening (e.g., based on IC₅₀ cutoffs like 1 μM or 10 μM), quantitative models capable of predicting IC₅₀ values provide deeper insight, particularly in late-stage development where therapeutic index and exposure margins are critical. Regulatory guidelines often recommend a 30-fold safety margin between the predicted hERG IC₅₀ and the maximum unbound plasma concentration of the drug candidate, , where quantitative models are superior to qualitative models in providing more detailed activity estimation.

Thus, integrating reliable in silico models with experimental screening workflows has become an essential strategy to mitigate cardiotoxicity risk, streamline compound optimization, and reduce costly late-stage failures in modern drug discovery pipelines.

Experiments and Methods

Data Sets

The data sets used in this study were adapted from Creanza’s work published on JCIM. The original data sets were sourced from ChEMBL, focusing on compounds with reported hERG IC₅₀ values. Compounds with molecular weight (MW) outside the 200–600 Da range, as well as those metal–organic and silicon-containing compounds, were excluded. Patch-clamp data were obtained using HEK or CHO cell lines. After removing duplicates and unsuitable entries, a final data set of 8,333 structurally diverse, nonredundant molecules was obtained. Among them, 7,772 had quantitative IC₅₀ values, suitable for regression modeling. For validation, we collected 1,133 compounds from ChEMBL with reported hERG activity that were uploaded after 2021, were nonredundant, and had molecular weights below 600 Da.

Molecular Descriptors

An atom-type-based molecular descriptor system was employed for both regression and classification modeling. Atom types were assigned using a casting tree that considered local chemical environments, such as aromaticity, attached protons and bonding patterns. In conjugated systems, the environment extended up to four bonds. The tree structure was optimized recursively, in terms of where to further split a branch and where to merge existing branches, in order to achieve the best predictive performance in log P regression model. The final descriptor set comprised 221 atom types and 40 correction factors, capturing whole-molecule features like molecular flexibility, fraction of sp²-hybridized atoms in a molecule, etc.(Table S1).

Support Vector Machine (SVM) Modeling

SVM, a supervised learning algorithm, is outstanding for explicitly addressing the generalization problem. We employed the υ-SVC implementation from LIB-SVM, as proposed by Schölkopf et al. Model parameters υ and γ, the nonlinearity parameter in the kernel function of a Gaussian Radial Basis Function (RBF), were optimized via a grid-based search to minimize the mean standard error (MSE) in 5-fold cross-validation (CV). To evaluate model stability, the data set was randomly split into training and test sets at various ratios, with each split scheme repeated 10 times.

Support vector regression (SVR), implemented via the ε-SVR module in LIB-SVM, is similar to SVC in the sense that SVC segregates data points belonging to different classes into different sides of hyper-planes and maximizes the margin between the planes, whereas SVR fits the data points in between the minimized ε-tube was applied for regression.

Results and Discussion

Structural Insights into hERG and Ligand Interactions

The overall structures of the four newly solved hERG channels–one apo and three ligand bound–are highly similar, with the pairwise RMSD (root-mean-square deviation) of the Cα atoms less than 1 Å (Figure a). Electrostatic surface analysis revealed a highly electron-rich binding environment (Figure b), which helps explain the prevalence of positively charged moieties in potent blockers (Figure ). In contrast, anions can significantly reduce binding affinity, as illustrated by the rescue of the withdrawn drug terfenadine through the introduction of an acidic group (Figure ). Notably, no negatively charged residues are found in the binding pocket; instead, electron density arises primarily from the backbone carbonyls (S624, V625) and the side chain hydroxyl groups (S624, Y652) (Figure S1).

2.

(a) Superposition of the apo hERG channel (PDB ID: 8ZYN, light blue) and hERG structures in complex with astemizole (PDB ID: 8ZYO, green), E-4031 (PDB ID: 8ZYP, yellow), and pimozide (PDB ID: 8ZYN, teal). (b) Electrostatic surfaces of the ligand-binding pockets of apo-hERG (PDB ID: 8ZYN), illustrating the C4 symmetry.

The apo-structure of the tetrameric hERG channel exhibits C4 symmetry, apart from loop regions between α-helices (Figure S2). The ligand-binding pocket in the apoprotein is also C4 symmetric (Figure b), comprising seven core residues: S621, T623, S624, V625 (forming the pocket floor), and S649, Y652, F656 (forming the sidewalls; Figure ).

3.

Seven core residues form the box-shaped ligand-binding pocket of the hERG channel. (a) Side view and (b) top view of the bottom of the binding pocket formed by residues S621, T623, S624, and V625, colored in light blue. (c) Side view and (d) top view of the pocket walls formed by residues S649, Y652, and F656, colored in green. The pocket coordinates are taken from the apo hERG channel (PDB ID: 8ZYN), whereas the ligand astemizole, shown in stick representation, is taken from PDB ID 8ZYO.

To understand the role of ligand-associated cations in molecular interactions, the three hERG-bound ligands were overlaid in their binding modes (Figure a). All three ligands contain a piperidine core bearing a cation (Figure ), and these cationic centers superimpose well across the binding poses. However, only in the case of astemizole does the cation form a weak hydrogen bond with Y652 (Figure b); otherwise, no energetically favorable interactions  such as charge–charge or π-cation interactions  are observed (Figure S3). The side chains of Y652 exhibit substantial flexibility as reflected in the dihedral angles, χ₁ and χ₂ that connect the phenyl ring to the backbone (Table ). In the apo hERG channel, the ligand-binding site is open to solvent, but it becomes partially or nearly fully enclosed upon ligand binding (Figure c), largely due to displacement of the Y652 side chains in all four monomers. The piperidine core shared by the three hERG blockers likely contributes to stabilizing this closed conformation, or to initiating rotation of the Y652 side chains. This is because the hydroxyl groups of Y652 attract delocalized electrons, and the four electron-rich moieties would otherwise repel one another in the absence of a cation participating in stabilizing interactions. In contrast, the other aromatic residue in the pocket, F656, remains more conformationally stable upon ligand binding (Figure d). In other words, the four F656 residues form a rigid wall of the pocket, whereas the Y652 side chains act as movable lids that open or close the pocket.

4.

(a) Superposition of the three hERG blockers in the binding mode: astemizole (green), E-4031 (yellow) and pimozide (teal). (b) Astemizole interacting with the hERG ligand binding pocket, with key hydrogen bonds indicated by yellow dashed lines. (c) Ligand binding sites of apo, astemizole-bound, E-4031-bound, and pimozide-bound hERG channels, shown with semitransparent surfaces. The side chains of the key residue Y652 were displayed as lines, and the ligands as sticks. For comparison, astemizole was positioned in the ligand binding site of the apo hERG structure. (d) Superposition of ligand binding pockets of apo (light blue) and astemizole-bound (green) hERG channels, with the side chains of F656 shown in a space-filling representation.

1. Dihedral Angles, χ₁ and χ₂, of Y652, Which Connect the Phenyl Ring and the Backbone, in the Four Monomers of Apo- and Ligand-Bound-hERG Structures .

^aThe Angles that vary more than 25° upon ligand binding are highlighted.

Piperidine is one of the most favorite rings frequently used by medicinal chemists, actually it is the third most frequently used ring systems, after phenyl and pyridine, among the marketed drugs. A significant fraction of molecules in the drug-discovery pipeline contain piperidine or piperazine ringstwo of the four most common ring systemsas well as aromatic moieties. This prevalence helps explain why hERG blockers span a wide range of therapeutic areas.

The adaptive orientation of Y652 enables favorable π–π stacking and potential π-cation interactions with ligand aromatic and charged groups. However, such intermolecular interactions are not observed in the three hERG-ligand complex structures (Figure S3), suggesting that the primary driving force for ligand recognition is the sequestration of hydrophobic and aromatic moieties from the aqueous environment into a less hydrophilic pocket. The resulting metastable complex is subsequently stabilized by the cation–Y652 interaction. The nonspecific nature of these interactions makes conventional pharmacophore modeling and molecular docking inadequate for broadly characterizing hERG blockers. In contrast, QSAR modelingwhere molecular descriptors capture the presence of chemical features rather than their precise spatial arrangementmay be more suitable for predicting hERG activity.

Predictive Models for hERG Blockers

A number of hERG models have been developed over the past 30 years, often limited by small data sets and narrow applicability domains. While larger data sets offer broader chemical space coverage, they can also introduce greater variability, potentially reducing predictive accuracy. This study leveraged one of the largest curated hERG data sets to build both classification and regression models.

In classification modeling, the data set was randomly split into training and test sets at ratios of 90/10, 80/20, 70/30, 60/40, and 50/50, with each split scheme repeated ten times to ensure reproducibility. Under the 90/10 split, the models achieved an average AUC-ROC of 0.88 (Figure S4) and an accuracy of 0.80, both with a standard deviation of 0.01 (Table ). These results are comparable to those reported from structure-based models but offer superior computational efficiency.

2. Optimized Parameters and Performance Metrics for the Ten SVR Models .

data set #	υ	γ	AUC	accuracy	sensitivity	specificity	cutoff
1	0.50	0.25	0.88	0.81	0.76	0.85	0.61
2	0.50	0.25	0.87	0.80	0.82	0.77	0.56
3	0.25	0.50	0.89	0.82	0.82	0.82	0.49
4	0.13	0.25	0.88	0.79	0.73	0.86	0.65
5	0.50	0.50	0.87	0.79	0.74	0.84	0.65
6	0.13	0.50	0.88	0.81	0.76	0.85	0.61
7	0.50	0.50	0.89	0.82	0.78	0.86	0.59
8	0.50	0.50	0.88	0.80	0.83	0.78	0.51
9	0.50	0.50	0.88	0.80	0.77	0.83	0.62
10	0.13	0.50	0.89	0.81	0.79	0.83	0.56
average			0.88	0.80	0.78	0.83	0.59

^aThe dataset was randomly split into training and test sets (90/10 ratio) and repeated ten times.

As the training set size decreased, AUC-ROC dropped slightly to 0.85 for a 50/50 split, and accuracy declined modestly to 0.78 (Figure a). Interestingly, the 80/20 split yielded slightly better accuracy than 90/10, suggesting the data set is sufficiently large and internally consistent. Among the misclassified quantitative entries, 73.7% deviated by ± 0.5 log units, and 95.1% fell within ± 1.0 log units of the experimental values (Figure b). Considering the known sources of measurement variability due to instrument calibration, cell health and density, and temperature, an error range of ± 0.5 log units is considered acceptable for high-throughput screening (HTS) using the same electrophysiology platform. − Additional noise could be introduced into the data set if hERG measurements from different laboratories using different HTS platforms and different cell lines. Large variations were also observed between manual patch clamp and HTS measurements.

5.

(a) Area under the ROC curve (ROC-AUC) and accuracy for five classification models using different split ratios (90/10, 80/20, 70/30, 60/40, and 50/50). Error bars represent the standard deviation from ten repeated experiments for each split scheme. (b) Distribution of pIC₅₀ values for the misclassified compounds in the test sets.

In short, compilation of data sets from different sources often introduces variability due to differences in assay methods (e.g., manual patch clamp vs HTS platforms), cell line selection (CHO or HEK), and other experimental factors. As a result, a variability of ± 0.5 ∼ ± 1.0 log units is a reasonable estimate for the uncertainty inherent in the data set used in this study. Therefore, misclassifications within the classification models can largely be attributed to noise in the experimental measurements.

For regression analysis, the Support Vector Regression (SVR) model was trained using 90% of the data set, with predictions made on the remaining 10%, repeated ten times. Figure a shows the correlation between predicted versus experimental pIC₅₀ values for 7,619 compounds. The R ² value was 0.63 (Figure a), but this metric is limited due to the non-normal distribution of pIC₅₀ values (Figure b). As such, more informative metrics including the Average Absolute Error (AAE = 0.383 log units) and Root Mean Square Error of Prediction (RMSEP = 0.548 log units) were used. Among the predictions, 74.3% fell within ± 0.5 log units, and 93.2% were within ± 1.0 log units (Figure a), demonstrating that the models’ predictive performance aligns with the expected experimental variability.

6.

(a) Experimental pIC₅₀ of 7,619 molecules in the test sets of ten SVR models plotted against predicted values with the regression line (red), and deviation lines (±0.5 log units in green). (b) Distribution of pIC₅₀ values of the 7,772 compounds used in regression.

Regression models provide quantitative predictions of hERG activity, which are crucial for assessing the safety profile of drug candidates in late-stage development where therapeutic index and exposure margins are critical. In contrast, qualitative classification models with preset cutoffs are inadequate for assessing dose-dependent risk and cannot effectively characterize the therapeutic window (Table ). However, a key limitation of regression modeling is that it requires quantitative data, which means inactive compounds or those with IC₅₀ values outside of the titration range that are reported qualitatively (e.g., > 10.0 μM or <1.0 nM) are not usable. These compounds are often excluded from regression models but are typically included in classification models depending on the cutoff selected. As a result, inactive compounds are generally underrepresented in regression models compared to classification models, given the large number of inactive test compounds. To overcome this limitation, a hybrid approach can be applied, where classification models are first applied to predict whether a compound is active, and if it is, a regression model is then employed to estimate its IC₅₀ value.

3. Comparison of Strengths and Weaknesses of Regression and Classification Modeling.

	regression models	classification models
prediction	quantitative	qualitative
data and availability	numerical, usually less	include qualitative, more
cutoff issue	no	yes
precision	good	Poor
assess of performance	R 2, RMS, AAE	AUC, accuracy
application scenario	late stage	early stage

Validation of Regression Models

To validate the hERG predictive models, the SVR model was applied to a data set of 1,133 nonredundant compounds uploaded to ChEMBL after 2021. The predictive performance was acceptable, with an AAE of 0.59 log units (Figure a). The ten largest outliers belonged to two chemotypes: two naltrexamines and eight piperazine–thiazoles. Examination of the original research articles revealed that the hERG IC₅₀ values reported in ChEMBL were 1,000-fold more potent than those described in the publications. , Correcting these errors improved the AAE to 0.50 log units (Figure b). Additional annotation errors may exist in the validation set; however, full manual curation of the data set is beyond our current capability.

7.

Experimental pIC₅₀ values of the 1,133 molecules in the validation set plotted against the predicted values, with the regression line shown in red, for (a) the original data set and (b) the corrected data set.

Summary

The structural features shared by hERG channel blockers have been analyzed in detail using cryo-electron microscopy (cryo-EM) structures of the hERG channel with and without ligand binding. The highly adaptive nature of the hERG ligand-binding site provides an explanation for the wide spectrum of structurally diverse small-molecule drugs that interact with hERG. Both quantitative and qualitative models were developed and evaluated using one of the largest curated data sets. The classification models achieved a strong predictive performance, with an average area under the receiver operating characteristic (ROC) curve (AUC) of 0.88 and an accuracy of 0.80. Most misclassified compounds (95.1%) were close to the cutoff pIC₅₀ of 5.0 (±1.0 log units). The regression models, trained on 90% of a large data set of 7,772 compounds, achieved an AAE of 0.383 log units and an RMSEP of 0.548 log units on the test sets. Validation on an external set of 1,133 compounds resulted in an AAE of 0.50 log units for the regression model. The accuracy of the predictive models is comparable to the level of uncertainty introduced by experimental variability. Accurate prediction of hERG channel blockade will significantly enhance the efficiency of discovering safer drug candidates.

All the data sets used to build and validate the regression and classification models are available at Supporting Information-.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.jcim.6c00163.

• Herg-JCIM (XLSX)

• Data sets used for training, testing, and validation of the classification and regression models; a detailed description of the atom types and correction factors employed as molecular descriptors; the ROC curves of the ten classification models; and additional information on the symmetric and electrostatic features of the hERG channel, along with 2D representations of ligand interaction maps (PDF)

†.

School of Life Science and Health Engineering, Jiangnan University, Wuxi 214122, P.R. China

H.S. and Y.W. developed and implemented the hERG predictive models; H.S. and M.S. wrote and edited the manuscript.

The authors declare no competing financial interest.