[Link].

com/scientificreports

OPEN hERG toxicity prediction in early

drug discovery using extreme
gradient boosting and isometric
stratified ensemble mapping
Gabriela Falcón-Cano1, Aliuska Morales-Helguera1, Heather Lambert1,
Miguel-Ángel Cabrera-Pérez2 & Christophe Molina1
Blockade of the human Ether-à-go-go Related Gene (hERG) potassium channel by small molecules
can prolong the QT interval, leading to fatal cardiotoxicity. Numerous drugs have been withdrawn
from the market due to cardiac side effects, underscoring the need for early identification of hERG
toxicity. Despite several classification machine learning (ML) models having been developed to this
end, robustness, class imbalance, and interpretability are still challenges. Using the largest public
database of hERG inhibition, this work integrates eXtreme Gradient Boosting (XGBoost) with
Isometric Stratified Ensemble (ISE) mapping (XGB + ISE map) to enhance hERG prediction. An XGBoost
consensus model was developed using balanced training sets and diverse variable subsets, resulting
in robust models less affected by class imbalance. The model demonstrated competitive predictive
performance, achieving a balance between sensitivity (SE = 0.83) and specificity (SP = 0.90) through
exhaustive validation. ISE mapping estimated the model applicability domain and improved prediction
confidence evaluation and compound selection by stratifying data. Refined variable selection
procedures enhanced model interpretability. Variable importance analysis highlights key molecular
determinants (peoe_VSA8, ESOL, SdssC, MaxssO, nRNR2, MATS1i, nRNHR, nRNH2) associated with
hERG inhibition. The XGB + ISE map strategy provides an effective approach to identifying promising
molecules in drug discovery campaigns with reduced hERG inhibition risk.

Keywords Machine learning, Ensemble methods, Imbalanced dataset, Applicability domain, Variable
selection

Ether-à-go-go (EAG) proteins are potassium channels found in muscle tissues, the brain, endocrine cells, and
the heart. The EAG-related gene (ERG) subfamily includes three distinct isoforms: Kv11.1, Kv11.2, and Kv11.3.
The human isoform Kv11.1, or human Ether-à-go-go Related Gene (hERG), is crucial in drug development due
to its role in cardiac toxicity. Although different mechanisms have been associated with QT-related arrhythmias
like Torsades de Pointes (TdP), such as mitochondrial toxicity, inhibition of voltage-gated calcium channels, and
hERG trafficking inhibition, blockade of the hERG potassium channel is considered one of the major reasons
for drug-induced TdP1–3. Several marketed drugs, have been withdrawn due to toxicity associated with hERG
blockade4. The high sensitivity of the hERG channel to inhibition by a wide variety of structurally diverse
molecules reinforces the importance of early evaluation of hERG blockade during drug discovery5,6.
Traditionally, anti-target studies have relied on a combination of in vitro and in vivo methods. However,
ethical concerns and the high cost associated with in vivo methods have demanded the development of viable
alternatives7, aligning with the 4R principles: reduce, refine, replace and responsibility8,9. In vitro methods, like
radioligand binding and patch clamp techniques, assess hERG liability, but are costly and time-consuming. In
vivo hERG assays, while also expensive and time-consuming, further suffer from low throughput, making them
almost impractical for evaluating large datasets of compounds in the early stages of drug discovery10. In this
sense, the Food and Drug Administration has recently proposed a major initiative via the Comprehensive In
Vitro Pro-Arrhythmia Assay program that integrates in silico models with in vitro data from engineered human
cells and stem cell-derived cardiomyocytes to estimate cardiotoxicity in early drug development11.

1PIKAÏROS, S.A, 31650 Saint Orens de Gameville, France. 2Departamento de Ciencias Farmacéuticas,
Facultad de Ciencias, Universidad Católica del Norte, Angamos, 0610 Antofagasta, Chile. email:
[Link]@[Link]

Scientific Reports | (2025) 15:15585 | [Link] 1

[Link]/scientificreports/

Computational toxicity prediction has become a valuable tool for assessing cardiotoxicity and eliminating
molecules likely to fail in preclinical or clinical studies. Recent reviews, such as the one by Tran et al.7, highlight
that most hERG prediction models rely on public, diverse data from sources like ChEMBL, PubChem, and
BindingDB, often combining data from various cell lines (e.g., CHO, Hek293).
The largest public database for predicting hERG inhibition is derived from the study by Sato et al.2. In their
research, the authors meticulously compiled and integrated hERG-related data from ChEMBL, PubChem,
GOSTAR, and hERG Central, resulting in a structurally diverse database. For the pivotal endpoint, many
researchers have chosen to use half maximal inhibitory concentration (IC50) to binarize hERG activity. However,
the classification of compounds as inhibitors or non-inhibitors varies widely due to differing IC50 thresholds,
which can range from 1 to 40 μM7. In classification models, the selection of these thresholds is closely tied to the
issue of data imbalance, a recognized challenge in machine learning (ML). To address this, researchers in the
field of hERG prediction have employed various methods, such as simple data manipulation techniques (e.g.,
Synthetic Minority Over-sampling TEchnique (SMOTE), random oversampling and random undersampling).
More sophisticated approaches include weighted models and probability threshold optimization based on
classification metrics sensitive to class imbalance12–14.
In terms of molecular descriptors, most studies rely on classical physicochemical descriptors12,15–17 and/or
molecular fingerprints6,13,18–22. Additionally, some studies explore graph structures to uncover correlations with
hERG inhibition23,24. Modeling algorithms predominantly include deep learning (DL) neural networks (NN),
such as Recurrent NN, Graph Convolutional NN, and Generative Adversarial NN6,19–22,24, as well as ML methods
like Random Forest12,16,25, Support Vector Machine (SVM)13 and eXtreme Gradient Boosting (XGBoost)17,26.
Model comparison is often challenging due to the use of proprietary datasets and the lack of publicly available
code for reproducing results or evaluating new external sets. Regarding model performance, statistical results
from different approaches tend to show minimal variation in global performance indicators, such as accuracy
(ACC, ranging from 0.80 to 0.95) and Area Under the Curve (AUC, ranging from 0.80 to 0.95) that is calculated
from Receiver Operating Characteristic (ROC) curves. Although the published models demonstrate similar
ACC, they should be analyzed using statistical metrics that reflect class imbalance, as these metrics reveal
substantial differences in their ability to predict inhibitor and non-inhibitor samples. Nevertheless, a standardized
benchmarking test set remains essential for making meaningful model comparisons. Additionally, many current
models lack interpretability and explainability, which poses a challenge in modeling hERG, particularly given
the critical implications for drug safety7,16,27.
This work implements a refined ML algorithm integrating recursive methods for variable selection, XGBoost
and Isometric Stratified Ensemble (ISE) mapping to enhance hERG prediction. XGBoost, a gradient boosting
machine variant, has demonstrated superior predictive performance in hERG channel inhibition modeling,
outperforming other ML algorithms17. Its ability to handle imbalanced datasets and maintain robustness across
diverse chemical spaces28 makes it particularly suitable for cardiotoxicity screening in early drug discovery. By
integrating XGBoost with recursive feature selection and ISE mapping, this study aims to optimize predictive
performance, expand the applicability domain (AD), and enhance model interpretability. Inspired by our
previous work on antileishmanial activity and colloidal aggregation29,30, this strategy leverages the ISE Map
methodology, adapting it for hERG inhibition prediction by incorporating model-agnostic interpretability
techniques and enhancing screening power. This leads to an advanced strategy to integrate hERG inhibition risk
assessment into drug discovery campaigns to select promising compounds with reduced hERG liability.

Methods
General workflow
The Konstanz Information Miner (KNIME) open-source software version 4.7.8 (available free of charge at
[Link] and the KNIME Python Integration were used for the development and
automatization of the Quantitative Structure–Activity Relationship (QSAR) methodology. Installation and
usage instructions are available on the KNIME website. Figure 1 summarizes the general modeling pipeline
described in the following sections.

Dataset
To develop the classification model, the largest public dataset with experimental values of hERG inhibitory
activity was used2,13. This dataset made up of 291,219 molecules, comes from different in vitro assays and
consists of 9,890 hERG inhibitors (molecules with IC50 ≤ 10 μM; for molecules without IC50 information,
any molecule with % of inhibition ≥ 50% at 10 μM). The remaining 281,329 molecules are considered non-
inhibitors. Although the authors had previously outlined a detailed procedure for dataset curation, an additional
curation protocol was implemented as described in our previous paper31. In brief, the initial stage involved the
elimination of structures with erroneous representations, such as unusual valences, and structures containing
d-block elements. Major fragments were retained for hydrates and inorganic salts during the manipulation of
unconnected structures. The second stage focused on normalizing charges and bonds, achieved through the
standardization of specific chemotypes, tautomeric forms, and the neutralization of charges. To perform these
normalizations, the rules published by Kamel et al. were used32. These rules were manually drawn, saved as
RxN files, and applied in the RDKit Chemical Transformation node within the KNIME platform. The third
and final stage encompassed the removal of duplicates and the curation of experimental data. In this step, 3D
features were excluded before generating International Chemical Identifier (InChI) codes/InChIKeys to identify
duplicate molecules. Any molecule with an inconsistent class label was excluded from the dataset.

Scientific Reports | (2025) 15:15585 | [Link] 2

[Link]/scientificreports/

Fig. 1. General overview of the modeling pipeline.

Molecular descriptors
The molecular structure parameterization scheme was based solely on the 2D structure. Therefore, stereoisomers
of the same molecule were considered duplicates and removed in the previous steps. Various 2D molecular
representations were investigated to assess their significance in predicting hERG inhibition. Physicochemical
properties, Molecular Operating Environment (MOE) and Kappa type descriptors, as well as Morgan,
Feat Morgan, and MACCS fingerprints, were computed using the KNIME RDKit plugin33. Other groups of
physicochemical descriptors were calculated using alvaDesc provided by Alvascience SRL 34,35. These belong

Scientific Reports | (2025) 15:15585 | [Link] 3

[Link]/scientificreports/

to the following families: Constitutional Indices, Ring Descriptors, Topological Indices, Walk and Path Counts,
Connectivity Indices, Information Indices, 2D Matrix-Based Descriptors, 2D Autocorrelations, Burden
Eigenvalues, P_VSA-like Descriptors, Extended Topochemical Atom (ETA) indices, Functional Group Counts,
and Molecular Properties. The variable selection procedure was performed during the model building process
as explained in the Variable Selection I subsection.

Data partitioning
The test set employed in the study by Ogura et al. was subtracted from the original dataset to act as an external
test set (ET I), enabling a fair comparison with other models13. ET I represents 30% of the hERG dataset.
From the remaining 70%, referred to here as the Modeling set, an internal test set was created by randomly
selecting 10% of the samples. The remaining 90% formed the full training set. Given the size of our modeling
set (> 200,000 compounds), allocating 90% of the data to training allows the model to capture complex patterns
more effectively, while the 10% test set remains large enough to provide a reliable performance estimate. Our
previous studies support this decision29,30, which used this ratio for large datasets. The full training set, internal
test set, and ET I set are all highly imbalanced, with the same imbalance ratio (see Table 1 of the Results and
Discussion section). This guarantees an unbiased evaluation of model performance during training and testing
phases.

Prospective validation
To ensure comprehensive validation of the model and conduct a thorough analysis of its performance, two
additional external datasets were assembled. This validation was performed after constructing and optimizing
the model. The entire modeling set was utilized to make predictions on each new external dataset, ensuring
the removal of any overlaps between the new external data and the modeling set. Below, we provide a detailed
description of the two External Validation Sets:

• External Validation Set I (EV I). The PubChem AID 588,834 consists of a 1536-well cell-based assay for
high-throughput screening (HTS), specifically quantitative HTS (qHTS), designed to evaluate in vitro hERG
channel blockage as a measure of cardiotoxicity with small molecules36. This assay employs the U2OS cell
line derived from human osteosarcoma, using the FluxORTM thallium flux approach, and the same National
Center for Advancing Translational Science (NCATS) laboratory. This dataset consists of a total of 5,381 data
points taken from version 2.1 of PubChem AID 588,834. Among these data points, there are 664 compounds
identified as HERG inhibitors, with an IC50 value of less than or equal to 10 μM, indicating their potential in-
hibitory activity. It is worth noting that the Ogura dataset 2 used in this study contains 2,688 chemicals taken
from the same source, but from version 1.1. For this reason, EV-I is partially included in the original modeling
set, and its overlap was removed from the modeling set before the validation process. This dataset is a valuable
resource for model validation since it corresponds to an HTS assay.
• External Validation Set II (EV II). The second dataset is part of the US federal Tox21 program. The NCATS
applied a qHTS approach to screen the Tox21 library of 9667 chemicals at 15 concentrations in triplicate to
quantify IC50 values for hERG activity in the U2OS cell line using a thallium flux assay platform37. Missing
SMILES (Simplified Molecular Input Line Entry System) and inconclusive samples were removed. IC50 values
were used to define inhibitors (IC50 ≤ 10 μM).

Murcko
hERG Dataset Class Input Molecules Output Molecules Data Loss (%) Imbalance Ratio Scaffolds
Modeling Total 203,853 203,425 0.21 28.42 71,831
hERG 6,923 6,915
non-hERG 196,930 196,510
ET I Total 87,366 87,306 0.07 28.46 42,630
hERG 2,967 2,964
non-hERG 84,399 84,342
Prospective Validation
EV I Total 5,381 3,457 35.76 10.68 1,355
hERG 664 296
non-hERG 4,107 3,157
EV II Total 9,667 6,509 33.1 24.53 1,883
hERG 371 255
non-hERG 8,858 6,254

Table 1. Summary of the data curation pipeline for training, test and validation datasets. The modeling dataset
includes both training and internal test data. ET I is the External Test Set (Ogura’s Test Set)13, while EV I and
EV II are external validation sets extracted from PubChem assay AID 588,334 36 and Tox21 hERG data37,
respectively.

Scientific Reports | (2025) 15:15585 | [Link] 4

[Link]/scientificreports/

Balanced training sets

To prevent training bias toward non-inhibitory samples (the majority class) in the highly imbalanced full training
set, multiple balanced training sets were generated. Specifically, the majority class was randomly sampled and
divided into R subsets, where R represents the imbalance ratio of 29 (the number of non-inhibitor compounds
divided by the number of inhibitor compounds). This process resulted in 29 non-inhibitor subsets, each of which
was combined with the full set of inhibitor compounds to generate 29 balanced training sets. A more detailed
discussion of this strategy can be found in our previous paper29.

Quality analysis of the different datasets

A critical analysis of the different datasets was conducted following data curation to assess both experimental
and chemical quality. Regarding experimental quality, the number of structures with inconsistent classification
and missing data labels were determined. Subsequently, the overlap between the different datasets was analyzed.
To assess the inter-dataset consistency, an overlap analysis was performed to quantify the degree of agreement
in their hERG activity labels. In addition, an intra-dataset analysis was also conducted, focusing on molecules
with two or more experimental bioactivity measurements (inhibitor/non-inhibitor) to identify agreement and
potential inconsistencies within individual datasets. To measure the structural diversity of the compounds, the
number of Murcko frameworks was counted, as reported by Langdon et al.38.

Variable selection I
A molecular descriptor selection procedure, named Recursive Decorrelated Variable Selection (RVS), was
applied to each of the 29 balanced training sets. The procedure determined the most important variables
using a DT algorithm, followed by a recursive selection to remove redundant variables (see Variable selection
II subsection below). Specifically, before variable selection, each descriptor column was duplicated, and their
descriptor values were shuffled. This way, artificial random columns were appended to the original descriptor
columns. To evaluate variable importance, the DT model was used to compute the number of occurrences of
each original and artificial descriptor. A variable was selected if the ratio between the number of occurrences of
the original variable and the number of occurrences of the counterpart artificial variable exceeded the threshold
of 1.5. This means that the original variable was at least 1.5 times more important than the artificial counterpart.
Afterwards, the Pearson correlation coefficient was calculated between the selected original variables. Through
a recursive loop, the number of variables was reduced until no more correlated pairs existed above the defined
Pearson correlation threshold (r = 0.6)29–31.

Ensemble model
The strategy employed in this work relied on the construction of an ad hoc ensemble consisting of homogeneous
base models trained on balanced training sets and validated on highly imbalanced data29,30.
The ensemble model was composed of 29 XGBoost39 base models, which were trained using 29 different
balanced training sets (see Fig. 1 and Balanced Training Sets subsection). The parameters of the algorithm
were configured as follows: booster = ‘gbtree’, learning_rate = 0.3, max_depth = 10, subsample = 1.0, colsample_
bytree = 1.0, scale_pos_weight = 1.0, objective = ‘binary’, num_boost_round = 101. Finally, the XGBoost ensemble
was first tested on the internal test set and then evaluated with ET I.

Applicability domain level

The concept of Applicability Domain Level (ADL) was introduced in our recent paper to refer to stratified
intervals of molecular distance30. ADL is not a direct quantification of the Applicability Domain (AD) itself,
but rather a practical framework based on “Molecule Distance Levels” that enables the evaluation of model
performance relative to molecular novelty. This operational interpretation facilitates a more nuanced reliability
analysis across a gradient of similarity to the training compounds.
To establish ADLs, Euclidean distances were calculated from each test set molecule to its nearest neighbor
in the full training set. Prior to this calculation, molecular descriptors were normalized. Based on these nearest
neighbor distances, eight ADLs were created, numbered from 0 to 7.

Consensus level
A consensus level (CL) was used to set the confidence level or probability of class labeling. Each one of the 29
XGBoost base models was used to predict the new data. Once the predictions of base models were completed, the
probabilities for the inhibitor class were averaged, and from this, the confidence score or maximum probability
of belonging to the winning class was calculated. Based on this probability, samples were categorized into subsets
called ‘strata’, ranked from the highest to the lowest possible probability of belonging to the winning class.
Winner class probability values ranging from 0.5 to 1 were linearly distributed into six different CLs numbered
from zero to five, where zero represents the stratum with the lowest confidence and five represents the stratum
with the highest confidence.

ISE map for estimating AD and prediction confidence

The ISE method integrates the CL and ADL into a 2D discrete map, the ISE map. In this map, compounds
are positioned according to their CL (horizontal axis) and ADL (vertical axis) values. This grid organization
enhances compound selection and ranking by considering both probability and AD criteria. With known
statistics and compound cardinality in each cell, new compounds can be projected onto the ISE map in order to
predict their statistics, and make informed decisions about compound selection and triage30.
The ISE map can be analyzed from either a partial or an incremental perspective. In a partial ISE map,
each sample is assigned to only one stratum, allowing for the calculation of individual statistics per stratum. In

Scientific Reports | (2025) 15:15585 | [Link] 5

[Link]/scientificreports/

contrast, in an incremental ISE map, a sample can belong to multiple cumulative categories. Both approaches,
partial and incremental, enhance the screening power of the method.

Variable selection II
Starting with the initial subset of key descriptors identified through the previous RVS, a second variable selection
method named Recursive Variable Elimination (RVE) was applied to determine the minimum number of
descriptors needed to preserve model performance. The process began with an initial model encompassing
the previously selected descriptors, which were sorted in descending order based on their importance derived
from the RVS method. Following this, an iterative process of descriptor elimination was implemented. In each
iteration, the descriptor with the lowest importance was identified and removed. After the removal of each
variable, its impact on the internal test set prediction was evaluated to assess whether its removal negatively
affected or maintained the model performance. At each step, the importance of the remaining variables was
recalculated, and the process was repeated, continuously removing the next least important variable.
With the iterative results of this method, a statistical comparison across iterations was conducted to determine
the minimum number of variables needed for a simpler and more interpretable model, while still maintaining its
performance. The Friedman non-parametric test was applied first, and when significant differences were found,
a Dunn-Bonferroni test was performed for further verification.

Performance measures and evaluation

Several statistical measures were computed, encompassing Recall or Sensitivity (SE), Specificity (SP), Positive
Predictive Value or Precision (PPV), Negative Predictive Value (NPV), F-measure, ACC, Balanced Accuracy
(BACC) and Cohen’s Kappa (CK). Additionally, the G-Mean and Matthews correlation coefficient (MCC) were
determined. The AUC was derived from the ROC curve. Mathematical definitions for all statistics can be found
in Table S1 in Supplementary Information (SI) 1.
To address the problem of model evaluation based on imbalanced data, F-measure, G-Mean, and MCC
statistics were employed. The entire modeling pipeline was independently executed three times with different
random seeds while consistently maintaining the same internal and ET I sets. The mean and standard deviation
of the performance measures for the models were calculated and recorded for all training, internal test, and
external test sets. Throughout the study, the ET I remained untouched during model training. The internal
test set served multiple purposes, including the primary evaluation of model performance, the definition of
isometric strata levels based on CL and ADL, and the selection of the final list of descriptors through the RVE
procedure.

Statistical analysis and model retraining

The modeling and validation process were repeated independently three times using different random seeds.
This approach facilitated the provision of mean and standard deviation values for all statistical parameters and
frequency metrics related to the importance of the selected descriptors. Only those descriptors that appeared
in all three runs and in at least 9 of the 29 base models (1/3 of the base models), were retained. At the end of
this procedure, the first list of the most important variables was obtained (Selected Variables I). Subsequently,
the RVE method was applied to the Selected Variables I set, resulting in a second list, the Interpretable Set of
Variables (refer to Fig. 1). This led to having two additional models, one trained with Selected Variables I and the
second trained using the Interpretable Set of Variables.

Results and discussion

Data analysis
The reliability and generalizability of results in QSAR modeling heavily depend on the quality and
comprehensiveness of the dataset used. Table 1 offers a comprehensive overview of the datasets used for model
development and prospective validation after the completion of data curation.
The Modeling set, used as the primary data source for training and internal testing, covers a diverse spectrum
of structural classes and chemical functionalities. This diversity ensured that the Modeling set accurately
represented the broad chemical space relevant to the biological endpoint under study2,13. The final ratio of
non-hERG/hERG inhibitor compounds (IR = 29:1) highlights a considerable class imbalance, with 96.6% of the
curated dataset biased toward non-inhibitor molecules. In silico modeling with such an imbalance typically
exhibits weak predictive performance for the minority class (hERG inhibitors). To address this challenge, a
strategy was implemented based on multiple balanced data sampling (MBDS), as described in prior published
studies29,30. To implement MBDS, 29 base models were created, each using a different balanced training set with
all inhibitors and an equal number of randomly sampled non-inhibitors. Internal and external sets (ET I, EV I
and EV II) exhibited a bias toward non-inhibitor molecules, with imbalance ratios ranging from 11 to 25 (Table
1). The curated datasets, along with their set label (training, internal test, and ET I) are available in SI 2, while the
curated EV I and EV II datasets are available in SI 3.
An overlap analysis between hERG datasets is provided in Fig. 2. Note that the dataset sourced from
PubChem AID 588,334 (EV I) shows the highest number of duplicate molecules and intra-set classification
agreement. The near‑perfect agreements (> 90% in all cases) obtained for all datasets confirmed that, despite
originating from different sources and experimental conditions, these datasets yielded closely aligned results for
overlapping molecules, indicating generally favorable agreement in biological activity.
Figure S1 in SI 1 illustrates cumulative frequency plots of Murcko scaffolds across each dataset. Acyclic
molecules were excluded from the analysis of Murcko scaffolds, which apply exclusively to ring-containing
structures38,41. Nonetheless, the percentage of acyclic molecules for each dataset was relatively consistent and
low, ranging from 0.17—0.20%. The steep initial gradient observed across all curves suggests the dominance

Scientific Reports | (2025) 15:15585 | [Link] 6

[Link]/scientificreports/

Fig. 2. Inter-set and intra-set hERG classification agreement for molecules with two or more experimental
measurements. The diagonal of the matrix represents the classification agreement within each dataset (intra-
set) after analyzing molecules with two or more experimental measurements (duplicates). The upper half of the
matrix shows the number of overlapping molecules between sets (inter-set) and the percentage of duplicates
with the same label for each set combination.

of specific structural motifs in the chemical space under study. The bottom diagram provides a matrix-like
visualization of the number of shared Murcko scaffolds between datasets, where the size of each square is
proportional to the number of common scaffolds. This visualization highlights scaffold diversity across datasets
and their structural similarities, which are crucial for assessing the applicability and generalizability of predictive
models.

Model performance and predictive power

The application of the methods described in the previous Variable Selection I subsection resulted in a refined
list of 22 important variables (Selected Variables I). This reduced list was derived from an initial pool of 4,298
variables and became the initial pool of variables for the training of a new model in this lower dimensional space.
Utilizing an ensemble of 29 base classification models with these 22 variables demonstrated robust
performance across all validation sets. Table 2 outlines the classification statistics for both internal set and ET
I using the retrained model with the 22 variables. A detailed description of the model performance on Internal
and ET I sets is provided in SI 4 and SI 5, respectively. The ACC remains consistently high, ranging from 0.87 to
0.90, and the model stands out for its balanced SP and SE values. To mitigate biased evaluation, statistics were
also generated for 29 balanced internal and external test subsets. In each iteration, a balanced internal set was
created by maintaining the active subset while changing the non-inhibitor samples. The results are presented
in Table 2. In this experiment, the CK and MCC values exhibit a substantial increase, reaching up to 0.74. It is
noteworthy that both metrics are influenced by the imbalance phenomenon.

Scientific Reports | (2025) 15:15585 | [Link] 7

[Link]/scientificreports/

Test Set N N variables SE SP ACC MCC CK

Internal 20,343 22 0.83 ± 0.01 0.91 ± 0.01 0.90 ± 0.01 0.40 ± 0.01 0.32 ± 0.01
(Balanced) 1,345 22 0.83 ± 0.01 0.91 ± 0.01 0.87 ± 0.01 0.74 ± 0.01 0.73 ± 0.01
ET I 87,306 22 0.83 ± 0.01 0.90 ± 0.01 0.90 ± 0.01 0.41 ± 0.01 0.33 ± 0.01
(Balanced) 5,976 22 0.83 ± 0.01 0.90 ± 0.01 0.87 ± 0.01 0.74 ± 0.01 0.74 ± 0.01

Table 2. Mean and standard deviation statistics for several metrics on internal, external and their respective
balanced test sets, considering three independent runs of the ensemble with the majority vote as the output
model. The abbreviations are detailed as follows: ET I: External Test Set (Ogura’s Test Set)13, N: number
of samples, N variables: Number of input variables, SE: Sensitivity, SP: Specificity, ACC: Accuracy, MCC:
Matthews correlation coefficient and CK: Cohen’s kappa.

A comparative analysis of the results obtained in this work with other models available in the literature is
presented in Table S3 in SI 1. This analysis includes models that: i) used the full hERG dataset13,27, and ii) are
commonly used and freely accessible such as DeepHIT22 and the model of Delre et al.14.
The ET I of the present study was used as the reference dataset for the comparison. Notable variations in
performance metrics of models are found across different studies (Table S3 in SI 1). Ogura et al.13 and Feng et
al.27 reported BACC values of 0.8 and 0.75 respectively, accompanied by a high SP of 0.99. However, their SE
values were relatively low, standing at 0.67 and 0.51, respectively, indicating a suboptimal trade-off between SP
and SE. Ogura et al. proposed an SVM model using 72 descriptors and ECFP_4 structural fingerprints. Feng
et al. implemented two natural language processing methods, an autoencoder and a transformer, to embed
molecular sequences. Both studies used the full hERG dataset from this study, originally compiled by Sato et al.2.
The predictive power of both DeepHIT22 and the model of Delre et al.14 is moderate when applied to ET I.
DeepHIT integrates three DL models with three different variables: molecular descriptors, fingerprints, and
graph-based descriptors to produce high NPV. Delre et al.14 used SMOTE, balanced random forest (BRF),
gradient boosted trees (GBT) and SVM techniques. However, both studies rely on small, balanced datasets
limiting the generalizability to highly imbalanced datasets like ET I. These results highlight the importance of
the nature of training data in developing ML models for drug discovery. In the previously cited studies (DeepHit
and Delre et al.), hERG data were collected from ChEMBL. Recent studies have shown that biological data in
ChEMBL are biased toward the representation of active (inhibitor) compounds, whereas the opposite is observed
in HTS data, which are generally highly imbalanced, with a small ratio of active to inactive compounds, as is the
case in the present study42. The data distribution in this study is representative of a fully unbiased HTS approach.
The results of our study reveal a different performance profile. The model demonstrates a trade-off between
SE (0.83) and SP (0.91), yet the challenge of class imbalance remains evident, as the MCC is around 0.4. This
indicates that while the model is effective at identifying hERG inhibitors, the number of false positives (FP)
remains high. This situation serves as a reason for the further implementation of the ISE map methodology (see
below), as a method to elucidate at which point the model preserves both SE and PPV for the inhibitor class. The
balance between SE and SP in such scenarios is strongly dependent on the balancing method. For instance, in the
paper by Ogura et al., a model-based approach was employed for handling class imbalance, specifically utilizing
the class weight option during the training of the SVM. This method imposed a greater penalty for errors in the
minority class, resulting in improved balance between SP and PPV. However, the SE remained low around 0.6713.
The findings of the present study closely align with those reported by Ylipää et al.43 in which, a graph neural
network was applied as their top-performing model. By optimizing the discretization probability threshold
through Youden’s J-statistic, their model achieved more balanced performance metrics (ACC: 0.92, SE: 0.88, SP:
0.92). However, a direct comparison of the two models is not feasible due to differences in dataset composition.
Specifically, they utilized only 93% of the Ogura dataset, excluding 5,638 chemicals during molecule preprocessing.
Without access to detailed information regarding the excluded compounds, a meaningful comparison of model
performance across the two studies remains unviable.
An essential consideration in anti-target prediction, such as hERG inhibition, is the cost associated with
false predictions. For safety-focused in silico models, minimizing false negatives (FN) while maximizing true
negatives is crucial to reduce cardiotoxicity risks during early drug development. The proposed model effectively
reduces the likelihood of misclassifying a potentially hERG-inhibiting molecule as safe. As a result, the consensus
methodology employed here enhances the reliability and accuracy of predictions for safety–critical targets,
aligning with the stringent requirements of early-stage drug safety assessments. In summary, the comparison of
our approach with other methods demonstrates that the model performs competitively. Our approach proves to
be an effective and well-rounded model for the prediction of hERG inhibition, offering a solid balance between
predictive performance and computational efficiency.
The OECD40 emphasizes the importance of transparency and mechanistic relevance in computational
models to facilitate interpretability. Our approach embodies these principles by operating in a low-dimensional,
interpretable space that seamlessly balances mechanistic insights with robust predictive power. Specifically, the
"Refinement: Prospective Validation and Model Interpretation" subsection presents a comprehensive variable
importance analysis that highlights the key descriptors significantly contributing to the prediction of hERG
activity. This analysis, conducted after applying two recursive variable selection methodologies, substantially
enhances the interpretability of our model.

Scientific Reports | (2025) 15:15585 | [Link] 8

[Link]/scientificreports/

Enhancing predictive reliability of the XGBoost ensemble with the ISE map approach
The proposed model trains an ensemble of 29 base XGBoost classifiers and the ISE map provides a CL based
on majority probability values and an ADL based on Euclidean distance for each sample. Figure 3 illustrates the
distribution of key metrics, ACC, G-Mean, and MCC, across different CLs for both internal and external sets,
based on partial and incremental statistical performance. Results reveal a clear trend: at the lowest consensus
level (CL 0), where agreement among base models is minimal, statistical values for ACC, G-Mean, and MCC
are at their lowest. In contrast, the highest CL (CL = 5), corresponds to the highest values for these metrics,
indicating the most reliable performance. For instance, when Incremental Consensus at level 0 was evaluated on
the external set, the ACC, G-Mean, and MCC were 90%, 87%, and 41%, respectively. These metrics improved
when the same Consensus strategy was applied at level 5 (ACC = 98%, G-Mean = 94%, and MCC = 72%). This
improvement was due to the ensemble making predictions in the chemical space where it is most accurate,
resulting in a reduction of coverage from 100 to 61% and an increase in prediction reliability. The same tendency
was observed with partial consensus (Fig. 3), but with worse performance results due to its restrictive nature. The
partial and incremental CL results are presented in SI files 4 − 7.
The incremental CL approach offers a refined strategy for selecting compounds based on the statistical
performance of the model. Specifically, when the objective is to identify compounds where the G-Mean of the
model surpasses 90%, focusing on CL 4 and CL 5 proves effective. Compounds within these levels benefit from
heightened predictive confidence, as they represent cases where nearly all base models consistently agree on
class labels. At CL 4 and CL 5, approximately 78% of compounds across both internal and external sets meet
this stringent reliability criterion, making these levels optimal for virtual screening aimed at prioritizing safe
(hERG non-inhibitors) compounds, such as those in combinatorial or virtual chemical collections44. This high-
confidence selection process ensures that the compounds retained for further testing exhibit strong predictive
ACC, thus aligning well with early-stage safety assessments that demand high precision in anti-target screening.
The ISE-map organizes compounds on a discrete 2D grid based on their CL and ADL. The combination
of the XGBoost ensemble with the ISE map improves previous approaches, because it incorporates, besides
stratification by CL, AD information based on molecular similarity. According to the OECD requirements,
computational models should have a well-defined AD to ensure their reliability.
Figure 4 provides a 2D representation of the ISE map, where the F-measure values for both the internal
and ET I sets are plotted according to two key axes: CL on the x-axis and ADL on the y-axis. F-measure was
selected because it provides a single score that balances the estimations of both PPV and SE in one number,
making it particularly useful for imbalanced datasets. As CL increases, especially around levels 4 and 5 on the
x-axis, and when ADL is 0 or 1, the color moves to darker shades of gray, indicating F-measure values greater
than 0.55 and approaching 0.85. These values correspond to the best-classified compounds of the internal Set,
whereas the worst-classified compounds are located at the opposite corner of the contour ISE map, with strongly
decreasing coverage (see coverage values in Fig. 4). By analyzing the internal set ISE map, the strata with the
best classification performance and highest coverage are (ADL, CL): (0, 4), (0, 5) and (1, 5). These strata can be
selected and extrapolated to the ET I, from which the best classified molecules can then be sorted and triaged.
Figure 4 shows the resulting ET I ISE map with calculated F-measure cell values. Notably, there is a clear
correspondence between the ISE map for the internal and ET I sets. Using the best-selected strata from the
internal set, a coverage of 64% of the ET I was obtained (55,885 of 87,306 compounds), resulting in a global
F-measure of 0.71, with ACC, CK, G-Mean, and MCC values of 0.98, 0.70, 0.93 and 0.72, respectively. These
metrics represent a substantial improvement over our overall model (F-measure: 0.36, ACC: 0.90, CK: 0.33,
G-Mean: 0.87, MCC: 0.41), increasing prediction reliability, despite the reduction in coverage from 100 to 64%.

Fig. 3. Distribution of key metrics, Accuracy (ACC), Geometric Mean (G-Mean), Matthews Correlation
Coefficient (MCC) and coverage, across different levels of consensus for both internal and external sets based
on partial and incremental statistical performance. The error bars represent the standard deviation for each
metric, calculated from three independent runs of each consensus model. All numerical statistics with means
& standard deviations are available in SI files 4 and 5.

Scientific Reports | (2025) 15:15585 | [Link] 9

[Link]/scientificreports/

Fig. 4. ISE map of F-measure values per stratum represented as a 2D contour plot for the (a) Internal
Set, and (b) External Test Set (ET I). In the contour plot, the color gradient ranges from lighter shades of
gray (low F-measure values) to darker shades of gray (high F-measure values). The contour lines delineate
these transitions between the F-measure values throughout the plot. In the table, F-measures > 0.55 and
coverage > 10% are marked with a gray background. The abbreviations are detailed as follows: ADL:
Applicability Domain Level and ISE: Isometric Stratified Ensemble.

By organizing data into strata based on CLs and incorporating Nearest Neighbor distance values, the ISE
map not only establishes the model AD but also facilitates a nuanced understanding of prediction confidence,
which is another standout feature of this study. This approach contributes to the broader field of predictive
toxicology modeling by providing a graphical tool for assessing the reliability of model predictions across
different consensus and AD levels.
Overall, our methodology effectively addresses data imbalance and enhances model reliability through a
refined validation framework incorporating the ISE map. While similar performance levels have been reported
in other studies, our approach reduces FN, critical in early drug discovery, and provides a robust method for
hERG screening. The refined selection method based on the ISE map helps identify promising molecules while
minimizing the risk of selecting those with undesired hERG inhibition.

Refinement: Prospective validation and model interpretation

Reducing the number of input variables lowers computational costs and improves interpretability. In this study,
22 descriptors were selected from an initial set of 4,298 using the RVS technique, followed by RVE to identify the
minimum number of variables needed for a compact, interpretable model of hERG inhibition.
The statistical analysis described in Variable Selection II in Method section indicated that only eight variables
are strictly significant for predicting the target variable using the current methodology. A graphical representation
is provided in Figure S2 in SI 1. The sorted list of the most important variables for hERG prediction is shown in
Table S4 in SI 1. Additionally, Figure S3 in SI 1 presents the Pearson and Spearman rank correlation coefficients
for each pair of variables, illustrating low correlations between them. These results suggest that the two variable
selection algorithms effectively reduce model complexity by selecting the most important and non-redundant
variables, thereby enhancing interpretability.
Table 3 shows performance comparisons when reducing descriptors from 4,298 ((RVS + XGB)_a) to 22
((RVS + XGB)_b), and ultimately to 8 ((RVS + XGB)_c), highlighting the interpretability gains with fewer
variables. The model developed from the complete pool of descriptors exhibited only a marginal improvement
over the model with 22 descriptors across the three external sets, demonstrating their equivalence in performance.
However, the second model was adopted and further developed, adhering to the principle of parsimony and
favoring a low-dimensional descriptor space. Table 3 also summarizes the results obtained from external sets
using these model predictions against the output of the ISE map methodology, based on the previously selected
strata. Our XGBoost ISE approach outperforms the overall models, enhancing screening power despite a
decrease in coverage.

Scientific Reports | (2025) 15:15585 | [Link] 10

[Link]/scientificreports/

Set Method Input Variables Cov (%) IR SE SP ACC MCC F-1 CK

ET (RVS + XGB)_a 4,298 100 28 0.85 0.92 0.92 0.46 0.42 0.39
(RVS + XGB)_b 22 100 28 0.83 0.90 0.90 0.41 0.36 0.33
(RVS + XGB)_c 8 100 28 0.80 0.85 0.85 0.32 0.27 0.22
(RVS + XGB)_b + ISE 22 64 36 0.88 0.98 0.98 0.72 0.71 0.70
EV I (RVS + XGB)_a 4,298 100 11 0.75 0.96 0.94 0.65 0.68 0.64
(RVS + XGB)_b 22 100 11 0.74 0.94 0.93 0.60 0.63 0.59
(RVS + XGB)_c 8 100 11 0.71 0.93 0.91 0.53 0.56 0.51
(RVS + XGB)_b + ISE 22 69 12 0.82 0.98 0.97 0.80 0.81 0.80
EV II (RVS + XGB)_a 4,298 100 24 0.84 0.95 0.94 0.55 0.54 0.51
(RVS + XGB)_b 22 100 24 0.77 0.94 0.94 0.50 0.49 0.46
(RVS + XGB)_c 8 100 24 0.75 0.93 0.92 0.44 0.42 0.39
(RVS + XGB)_b + ISE 22 71 28 0.97 0.97 0.97 0.67 0.67 0.65

Table 3. XGBoost (XGB) ensemble performance on external sets following the application of variable
selection procedures (RVS). (RVS + XGB)_a is the model obtained using all initial molecular descriptors (4,298
variables) with RVS, (RVS + XGB)_b is the model obtained using a reduced initial set of molecular descriptors
(22 variables), (RVS + XGB)_c is the model with the lowest dimensionality (8 variables) and (RVS + XGB)
_b + ISE is the model derived from a reduced initial set of molecular descriptors (22 variables) with RVS,
considering only the predictions made in the previously selected strata (ADL, CL): (0, 4), (0, 5), and (1, 5).
Input variables are the initial pool of molecular descriptors from which the model is built. The abbreviations
are detailed as follows: Cov: Coverage, IR: Imbalance Ratio, SE: Sensitivity, SP: Specificity, ACC: Accuracy,
MCC: Matthews correlation coefficient, F-1: F-measure and CK: Cohen’s kappa. ET I is the External Test Set
(Ogura’s Test Set)13, while EV I and EV II are external validation sets extracted from PubChem assay AID
58833436 and Tox21 hERG data36, respectively.

Median Value
Rank Variables Inhibitor Non-inhibitor Type Description
1 peoe_VSA8 30.714 17.92 VSA Descriptors P_VSA- like on PEOE (partial charges), bin 8
2 ESOL -4.672 -3.829 Molecular Properties Estimated Solubility (logS) for aqueous solubility using LOGPcons
3 SdssC 0 -0.217 Atom-type E-state indices Sum of E-states of atoms of type = C < within a molecule
4 MaxssO 4.880 4.780 Atom-type E-state indices Maximum atom-type E-State: -O-
5 nRNR2 1 (0–4) 0 (0–4) Functional group counts Number of tertiary amines (aliphatic)
6 MATS1i -0.139 -0.108 2D autocorrelations Moran autocorrelation of lag 1 weighted by ionization potential
7 nRNHR 0 (0–2) 0 (0–4) Functional group counts Number of secondary amines (aliphatic)
8 nRNH2 0 (0–2) 0 (0–5) Functional group counts Number of primary amines (aliphatic)

Table 4. List of selected variables/descriptors from the Recursive Variable Elimination algorithm, including
median descriptor values for hERG inhibitors and non-inhibitors in the training set. A Mann–Whitney U
test showed significant differences (p < 0.05) in continuous descriptors between hERG inhibitors and non-
inhibitors. A Chi-squared analysis revealed a significant association between the variables nRNH2, nRNHR
and nRNR2 and the hERG activity (p < 0.001). The rank refers to the order of importance of the 8 descriptors
included in the lowest dimensionality model. All descriptors shown here were calculated with alvaDesc
software, except for peoe_VSA8, which was calculated using RDKit software. In parentheses, the maximum
and minimum values of the molecular descriptors nRNH2, nRNHR, and nRNR2 are noted.

Table 3 also presents a model with significantly lower dimensionality (8 variables) that can predict outcomes
for EV I and EV II with strong statistical performance. These results support the ability of the model to predict
hERG inhibition potential, as EV I and EV II were derived from standardized assays. The predictive power of
the model remained robust despite the reduced set of descriptors, emphasizing the efficiency of the employed
methodology. However, the model with 22 variables remains the one that achieves the best balance between the
number of variables used and model performance across all external sets.
To gain insights into the structure-hERG inhibition relationship, the 8 most important variables were
examined. Table 4 lists these descriptors, belonging to five logical blocks: molecular properties, autocorrelation
descriptors, atom-type electrotopological state (E-state) indices, VSA descriptors and functional group counts.
Although no single descriptor explains hERG activity alone, their combination provides valuable insights.
The variable MATS1i is a Moran (M) autocorrelation descriptor with lag 1 that reflects spatial autocorrelation
in a molecular graph where atoms are weighted by ionization potential (i)45. When lag = 1, MATS provides
a measure of Nearest Neighbor correlations; when lag = 2, next Nearest Neighbor, and so on. In general,

Scientific Reports | (2025) 15:15585 | [Link] 11

[Link]/scientificreports/

-1 < MATS < 1. Positive autocorrelations correspond to positive values of the Moran coefficient, indicating that
atoms with high ionization potential are likely to follow each other along the atom sequence at topological
distance of 1. Conversely, negative autocorrelation produces negative values, meaning that a high value is
followed by a low value, or vice versa, at lag = 1. For instance, transformations that incorporate atoms with high
ionization potential, such as nitrogen and/or oxygen, may reduce hERG inhibitory activity when these atoms are
adjacent in the molecular structure. These changes are reflected in the MATS1i descriptor, which shows higher
median values in non-inhibitors than in inhibitors.
Other significant descriptors include MaxssO, the maximum E-state value among all -O- groups in the
molecule, and SdssC, the sum of E-state values for carbon atoms with sp2 hybridization. E-state values capture
both electronic and topological information, reflecting the intrinsic electronic state of the carbon and oxygen
atom and the influence of surrounding atoms. The E-state of an atom Si is formulated as the sum of its intrinsic
state Ii, and a perturbation term ΔIij which represents the influence of the j-th atom on the i-th intrinsic state.
This perturbation term acts as an electronegative gradient, with its sign indicating the direction of influence from
surrounding atom intrinsic states. Thus, large positive E-state values are associated with highly electronegative
and/or terminal atoms that lie on the periphery of the molecule. In contrast, small or negative E-state values
correspond to atoms involved in sigma bonding, atoms that are located in the core of the molecule or atoms close
to highly electronegative atoms45. Table 4 shows that hERG non-inhibitors exhibit lower median values for both
descriptors compared to the inhibitors. For instance, introducing = C < groups near highly electronegative atoms
and/or in the interior of the molecule reduces these descriptor values, thereby contributing to non-inhibition.
This reduces the electronic accessibility of these atoms and limits their interaction with other molecules, such as
the hERG channel. These findings align with previous studies that emphasize the role of electrostatic interactions
with amino acid residues around the hERG cavity, such as Y652, as the key mechanism in hERG inhibition46.
The peoe_VSA8 descriptor exhibits a similar trend to that of E-state indices. This descriptor has higher
median values for inhibitor compounds compared to non-inhibitors. This descriptor is calculated as the sum
of atomic contributions to the van der Waals surface area (VSA) from atoms within specific ranges, in this case,
bin 847. This suggests that the contact surfaces related to charge localization are significant factors in hERG
inhibition. PEOE charges depend on the connectivity of the input structures, including elements, formal charges,
and bond orders. The observation that uncharged molecules exhibit suboptimal hERG inhibition may be related
to their electrostatic characteristics. Such molecules may lack the essential charge distribution patterns required
for effective hERG inhibition.
The ESOL descriptor is primarily designed to predict the aqueous solubility of compounds based on their
molecular structure, utilizing properties such as the octanol/water logP, molecular weight, etc.48. The trend
observed in the ESOL descriptor shows that many hERG inhibitors tend to exhibit lower aqueous solubility. This
finding is also related to the behavior of the nRNR2 descriptor, which is directly associated with the presence
of tertiary amines in the structural moieties. The presence of basic nitrogen atoms with additional hydrophobic
groups is a classical characteristic of many hERG blockers. This is justified by the hydrogen bond between the
protonated nitrogen of the channel blocker and the carbonyl oxygen of hERG residue T62346.
Despite the non-linear relationships between the selected variables and hERG inhibition, certain tendencies
in the distribution of descriptors can still be observed, as described above. However, their practical implications
necessitate further analysis.

Conclusions
This study presents a robust and comprehensive framework for the QSAR modeling of hERG inhibition. By
generating multiple balanced training subsets, we ensured that our models were not skewed toward the majority
class, thus enhancing their predictive performance, particularly reducing the false negative rate and minimizing
potential cardiotoxicity risks in early drug discovery. The use of recursive methods for variable selection enhances
model interpretability and is important for translating computational insights into actionable knowledge in
drug design. Prospective validation on two external datasets offers a comprehensive assessment of the model
generalization capabilities, showing robust predictive power even with fewer descriptors. A standout feature of
this research is the application of the ISE map for estimating the applicability domain and prediction confidence.
By organizing data into strata based on consensus level and incorporating Nearest Neighbor distances, the
ISE map serves as a model-agnostic method that enhances model interpretability and facilitates a nuanced
understanding of prediction confidence. This methodological innovation contributes significantly to the field of
predictive modeling by providing a graphical tool for assessing the reliability of model prediction across various
consensus levels. The XGB + ISE map strategy provides an effective approach to identify and prioritize promising
molecules in drug discovery campaigns with reduced hERG inhibition risk.

Data availability
All data generated or analyzed during this study are included in the main text of this article and its Supplemen-
tary Information files.

Code availability
The computational method used to generate the results in this study is accessible through a web server de-
ployed at [Link] This server allows users to submit data and receive predictions based
on the hERG model described in this manuscript. Detailed instructions for using the server are available on
the web page. The code is available under license for academic research and for commercial use subject to
negotiation. For further information, please contact the corresponding author CM.

Scientific Reports | (2025) 15:15585 | [Link] 12

[Link]/scientificreports/

Received: 20 December 2024; Accepted: 21 April 2025

References
1. Creanza, T. M. et al. Structure-based prediction of hERG-related cardiotoxicity: A benchmark study. J. Chem. Inf. Model. 61,
4758–4770 (2021).
2. Sato, T., Yuki, H., Ogura, K. & Honma, T. Construction of an integrated database for hERG blocking small molecules. PLoS ONE
13, e0199348 (2018).
3. Mamoshina, P., Rodriguez, B. & Bueno-Orovio, A. Toward a broader view of mechanisms of drug cardiotoxicity. Cell Rep. Med. 2,
100216 (2021).
4. Kalyaanamoorthy, S. & Barakat, K. H. Development of safe drugs: The hERG challenge. Med. Res. Rev. 38, 525–555 (2018).
5. Kalyaanamoorthy, S. et al. A structure-based computational workflow to predict liability and binding modes of small molecules to
hERG. Sci. Rep. 10, 16262 (2020).
6. Cai, C. et al. Deep learning-based prediction of drug-induced cardiotoxicity. J. Chem. Inf. Model. 59, 1073–1084 (2019).
7. Tran, T. T. V., Surya Wibowo, A., Tayara, H. & Chong, K. T. Artificial intelligence in drug toxicity prediction: Recent advances,
challenges, and future perspectives. J. Chem. Inf. Model. 63(2628), 2643 (2023).
8. Kang, M. et al. A Review of the ethical use of animals in functional experimental research in china based on the ‘four r’ principles
of reduction, replacement, refinement, and responsibility. Med. Sci. Monit. 28, e938807 (2022).
9. Lee, K. H., Lee, D. W. & Kang, B. C. The ‘R’ principles in laboratory animal experiments. Lab. Anim. Res. 36, 45 (2020).
10. Villoutreix, B. O. & Taboureau, O. Computational investigations of hERG channel blockers: New insights and current predictive
models. Adv. Drug. Deliv. Rev. 86, 72–82 (2015).
11. Colatsky, T. et al. The Comprehensive in Vitro Proarrhythmia Assay (CiPA) initiative - Update on progress.J. Pharmacol. Toxicol.
Methods81, 15–20 (2016).
12. Arab, I. & Barakat, K. ToxTree: descriptor-based machine learning models for both hERG and Nav1.5 cardiotoxicity liability
predictions. arXiv preprint arXiv:2112.13467 (2021)
13. Ogura, K., Sato, T., Yuki, H. & Honma, T. Support vector machine model for hERG inhibitory activities based on the integrated
hERG database using descriptor selection by NSGA-II. Sci. Rep. 9, 12220 (2019).
14. Delre, P. et al. Ligand-based prediction of hERG-mediated cardiotoxicity based on the integration of different machine learning
techniques. Front. Pharmacol. 13, 951083 (2022).
15. Hsiao, Y., Su, B. H. & Tseng, Y. J. Current development of integrated web servers for preclinical safety and pharmacokinetics
assessments in drug development. Brief. Bioinform. 22(3), bbaa160 (2021).
16. Konda, L. S. K., Keerthi Praba, S. & Kristam, R. hERG liability classification models using machine learning techniques. Comput.
Toxicol. 12, 100089 (2019).
17. Siramshetty, V. B. et al. Critical assessment of artificial intelligence methods for prediction of hERG channel inhibition in the “Big
Data” era. J. Chem. Inf. Model. 60, 6007–6019 (2020).
18. Tian, S., Li, Y., Wang, J., Zhang, J. & Hou, T. ADME evaluation in drug discovery. 9. Prediction of oral bioavailability in humans
based on molecular properties and structural fingerprints. Mol. Pharm. 8, 841–851s (2011).
19. Shan, M. et al. Predicting hERG channel blockers with directed message passing neural networks. RSC Adv. 12, 3423–3430 (2022).
20. Karim, A., Lee, M., Balle, T. & Sattar, A. CardioTox net: a robust predictor for hERG channel blockade based on deep learning
meta-feature ensembles. J. Cheminform. 13, 60 (2021).
21. Zhang, X., Mao, J., Wei, M., Qi, Y. & Zhang, J. Z. H. HergSPred: Accurate classification of hERG blockers/nonblockers with
machine-learning models. J. Chem. Inf. Model. 62, 1830–1839 (2022).
22. Ryu, J. Y., Lee, M. Y., Lee, J. H., Lee, B. H. & Oh, K.-S. DeepHIT: A deep learning framework for prediction of hERG-induced
cardiotoxicity. Bioinformatics 36, 3049–3055 (2020).
23. Sun, H. et al. Highly predictive and interpretable models for PAMPA permeability. Bioorg. Med. Chem. 25, 1266–1276 (2017).
24. Xiong, G. et al. ADMETlab 2.0: An integrated online platform for accurate and comprehensive predictions of ADMET properties.
Nucleic Acids Res. 49, W5–W14 (2021).
25. Venkatraman, V. FP-ADMET: A compendium of fingerprint-based ADMET prediction models. J. Cheminform. 13, 75 (2021).
26. Ishihara, K., Sone, H. & Hayamizu, K. Gradient boosting machine-based model for predicting hERG K+ channel inhibitory
activities. J. Comp. Sci. Appl. Inform. Technol. 5, 1–9 (2022).
27. Feng, H. & Wei, G.-W. Virtual screening of drugbank database for hERG blockers using topological laplacian-assisted AI models.
Comput. Biol. Med. 153, 106491 (2023).
28. Babajide Mustapha, I. & Saeed, F. Bioactive molecule prediction using extreme gradient boosting. Molecules 21, 983 (2016).
29. Casanova-Alvarez, O., Morales-Helguera, A., Cabrera-Pérez, M. Á., Molina-Ruiz, R. & Molina, C. A novel automated framework
for QSAR modeling of highly imbalanced leishmania high-throughput screening data. J. Chem. Inf. Model. 61, 3213–3231 (2021).
30. Molina, C., Ait-Ouarab, L. & Minoux, H. Isometric stratified ensembles: a partial and incremental adaptive applicability domain
and consensus-based classification strategy for highly imbalanced data sets with application to colloidal aggregation. J. Chem. Inf.
Model. 62, 1849–1862 (2022).
31. Falcón-Cano, G., Molina, C. & Cabrera-Pérez, M. Á. Reliable prediction of caco-2 permeability by supervised recursive machine
learning approaches. Pharmaceutics 14, 1998 (2022).
32. Kamel, M. et al. Automated workflows for data curation and standardization of chemical structures for QSAR modeling. In
(American Chemical Society meeting, Orleans, LA, 2018). [Link] 2018
33. RDKit Nodes for KNIME (trusted extension) KNIME. [Link]
34. Mauri, A. & Bertola, M. Alvascience: A New software suite for the QSAR workflow applied to the blood-brain barrier permeability.
Int. J. Mol. Sci. 23, 12882 (2022).
35. alvaDesc - KNIME - Alvascience. [Link]
36. National Center for Biotechnology Information. PubChem Bioassay Record for AID 588834, qHTS Assay for Small Molecule
Inhibitors of the Human hERG Channel Activity. [Link] (2017).
37. NCATS, EPA, NIEHS, NTP, & FDA,. Tox21 Data Challenge. Human Ether-a-go-go-Related Gene (hERG) channel small molecule
antagonists: fluorescence cell-based qHTS assay U.S. Tox21 Program. [Link]
38. Langdon, S. R., Brown, N. & Blagg, J. Scaffold diversity of exemplified medicinal chemistry space. J. Chem. Inf. Model. 51, 2174–
2185 (2011).
39. XGBoost Documentation — xgboost 2.1.1 documentation. [Link]
40. OECD. Guidance Document on the Validation of (Quantitative) Structure-Activity Relationship [(Q)SAR] Models. OECD Series
on Testing and Assessment. No 69. OECD Publishing. Paris (2014).
41. Bemis, G. W. & Murcko, M. A. The properties of known drugs. 1. Mol. Framew. J. Med. Chem. 39, 2887–2893 (1996).
42. Zakharov, A. V., Peach, M. L., Sitzmann, M. & Nicklaus, M. C. QSAR modeling of imbalanced high-throughput screening data in
PubChem. J. Chem. Inf. Model 54, 705–712 (2014).
43. Ylipää, E. et al. hERG-toxicity prediction using traditional machine learning and advanced deep learning techniques. Curr. Res.
Toxicol. 5, 100121 (2023).
44. Compound Collections - Enamine. [Link]

Scientific Reports | (2025) 15:15585 | [Link] 13

[Link]/scientificreports/

45. Todeschini, R. & Consonni, V. Handbook of molecular descriptors (John Wiley & Sons Ltd, 2000). ​h​t​t​p​s​:​//​ ​d​o​i​.​o​r​g​/​10​ ​.​1​0​0​2​/​9​7​83​ ​5​2​7​
6​1​3​1​0​6.​ ​c​h​1​a​​​​.​​​
46. Garrido, A., Lepailleur, A., Mignani, S. M., Dallemagne, P. & Rochais, C. hERG toxicity assessment: Useful guidelines for drug
design. Eur. J. Med. Chem. 195, 112290 (2020).
47. Labute, P. A widely applicable set of descriptors. J. Mol. Graph. Model. 18, 464–477 (2000).
48. Delaney, J. S. ESOL: Estimating aqueous solubility directly from molecular structure. J. Chem. Inf. Comput. Sci. 44, 1000–1005
(2004).

Acknowledgements
All figures presented in the main text and Supplementary Information were originally created by the authors,
G.F.-C., A.M.-H., and C.M. We thank Silvia Romanelli for her professional contribution to the graphical design
of this work. The authors also acknowledge KNIME and its contributors for making the KNIME data mining
environment available free of charge, as well as Alvascience for providing an academic license for alvaDesc.

Author contributions
Research design: C.M., G.F-C. & A.M-H.; Methodology: C.M., G.F-C., A.M-H. & H.L.; Data Curation: H.L.
& G.F-C.; Data Analysis: C.M., G.F-C. & A.M-H.; Manuscript Drafting: G.F-C., A.M-H, C.M. & M-A.C-P. All
authors reviewed the manuscript.

Declarations

Competing interests
The authors declare the following conflict of interest: Pikaïros is a commercial entity, and the corresponding
author, CM, is the sole shareholder of the company. The other authors declare no competing financial interest.

Additional information
Supplementary Information The online version contains supplementary material available at ​h​t​t​p​s​:​/​/​do
​ ​i​.​o​r​g​/​1​
0​.1​ ​0​3​8​/​s​4​1​5​98​ ​-​0​2​5​-​9​9​7​66​ ​-​3​​​​.​​
Correspondence and requests for materials should be addressed to C.M.
Reprints and permissions information is available at [Link]/reprints.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives
4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in
any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide
a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have
permission under this licence to share adapted material derived from this article or parts of it. The images or
other third party material in this article are included in the article’s Creative Commons licence, unless indicated
otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence
and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to
obtain permission directly from the copyright holder. To view a copy of this licence, visit ​h​t​t​p:​ ​/​/​c​r​e​a​t​i​ve​ ​c​o​m​m​o​
n​s​.​or​ ​g​/​l​i​c​e​n​s​es​ ​/​b​y​-​n​c​-​n​d​/​4​.​0​/​​​​.​​

© The Author(s) 2025

Scientific Reports | (2025) 15:15585 | [Link] 14