Deep Learning-based Prediction of Drug-induced Cardiotoxicity

Abstract

Blockade of human ether-à-go-go-related gene (hERG) channel by small molecules induces the prolongation of the QT interval which leads to fatal cardiotoxicity, and account for the withdrawal or severe restrictions on the use of many approved drugs. In this study, we develop a deep learning approach, termed deephERG, for prediction of hERG blockers of small molecules in drug discovery and post-marketing surveillance. In total, we assemble 7,889 compounds with well-defined experimental data on the hERG and with diverse chemical structures. We find that deephERG models built by a multi-task deep neural network (DNN) algorithm are superior to those built by single-task DNN, naïve Bayes (NB), and support vector machine (SVM). Specifically, the area under the receiver operating characteristic curve (AUC) value for the best model of deephERG is 0.967 on the validation set. Furthermore, based on 1,824 U.S. Food and Drug Administration (FDA)-approved drugs, 29.6% drugs are computationally identified to have potential hERG inhibitory activities by deephERG, highlighting the importance of hERG risk assessment in the early drug discovery. Finally, we showcase several novel predicted hERG blockers on approved antineoplastic agents, which are validated by clinical case reports, experimental evidences, and literatures. In summary, this study presents a powerful deep learning-based tool for risk assessment of hERG-mediated cardiotoxicities in drug discovery and post-marketing surveillance.

Graphical Abstract

INTRODUCTION

The human ether-à-go-go-related gene (hERG) encodes the pore-forming α-subunit of rapid delayed rectifier current, playing crucial roles in the regulation of exchanges of the resting potential and action potential on cardiac myocyte.^{1, 2} Overwhelming experimental and clinical evidences have indicated that a blockade of hERG channel can induce long-QT syndrome (LQTS), which may lead to fatal cardiotoxicities, such as torsade depointes (TdP) arrhythmia.³ To date, several drugs, including astemizole, terfenadine, vardenafil, cisapride and ziprasidone, have been withdrawn or severely restricted on the use for the undesirable hERG-related cardiac side effects.^4–6 Since hERG channel is highly sensitive to be inhibited by a large amount of structurally diverse molecules, an early evaluation of hERG blockade has become a necessary step in drug discovery.^{7, 8}

According to the guideline (S7B) published by International Conference of Harmonization, all new drugs should be assessed pre-clinically for their hERG inhibitory activities before submitted to regulatory reviews.⁹ However, current in vivo and in vitro methods for screening hERG blockers, such as rubidium-flux assays, fluorescence-based assays, electrophysiology measurements and radioligand binding assays, are costly, laborious and time-consuming.¹⁰ Recent advances of in silico approaches and tools have offered possibilities for effective evaluation of drug ADMET (absorption, distribution, metabolism, excretion and toxicity) and pharmacokinetics and pharmacodynamics (PK/PD) properties at the early stages of drug discovery.^11–14 Over the past several years, a wide range of prediction models for hERG blockers have been published using various machine learning methods.^{4, 6, 15–24} For instance, in 2010, Doddareddy and co-workers developed classification models from 2,644 compounds using linear discriminant analysis and support vector machine (SVM) methods to estimate the hERG-related cardiotoxicity.²³ The area under the receiver operating characteristic curve (AUC) values of models ranged from 0.89 to 0.94 in 5-fold cross validation.²³ In 2016, Wang and co-workers utilized pharmacophore modeling combined with machine learning to build classification models for prediction of hERG active compounds. A accuracy for the hERG active and inactive compounds in the test set reached 83.6% and 78.2%, respectively.²⁴ Although some of these models showed acceptable performance on the training set and test set, a small space of chemical diversities has resulted in a limited application domain.²³ Meanwhile, most of the studies prepared decoy sets by randomly extracting compounds from the entire chemical database. The unknown experimental evidence of negative samples may cause potential false positive rate.

Preliminary studies have shown that multi-task deep neural network (DNN) has better learning and adaptive ability compared to conventional machine learning approaches for drug discovery.^25–28 For instance, recently, Li and co-workers developed DNN models using multi-task deep autoencoder neural network for concurrent inhibition prediction of five major CYP450 isoforms. The predictive power of multi-task deep neural network outperformed other machine learning methods including logistic regression, support vector machine, C4.5 DT and kNN.²⁸

In this study, we proposed a multi-task deep neural network framework for comprehensive assessment of hERG blockers, termed deephERG (Figure 1). Firstly, we collected comprehensive data on the hERG blockade activities of 7,889 diverse chemicals with well-defined experimental endpoints. All the chemicals were depicted through integration of Molecular Operating Environment (MOE) descriptors²⁹ and Mol2vec descriptors.³⁰ For more rigorous concern, we split the data set according to different decoy threshold values (10 μM, 20 μM, 40 μM, 60 μM, 80 μM, and 100 μM) for building multi-task DNN models. For each task, training set, test set, and validation set were used for model training, optimization and evaluation of generalization abilities, respectively. After systematic comparison, the multi-task DNN models offer the best performance comparing to both single-task DNN models and traditional machine learning models. Finally, we applied the best deephERG model to 1,824 FDA approved drugs for risk assessment of hERG-related cardiotoxicities.³¹ In summary, deephERG offers a powerful tool for cardiotoxic risk assessment in drug discovery and post-marketing surveillance.

Figure 1. A diagram illustrating deephERG framework.

A comprehensive collection on the hERG blockers (IC₅₀ ≤ 10 μM) and non-blockers with different decoy thresholds values (IC₅₀ > 10 μM, 20 μM, 40 μM, 60 μM, 80 μM, and 100 μM, respectively) are divided into training sets, test sets, and validation sets. The deep learning models are built based on Mol2vec and MOE descriptors using a multi-task deep neural network containing three hidden layers.

MATERIALS AND METHODS

Data preparation

The original compounds with experimental hERG blocking bioactivities were assembled from various well-defined experimental assays: (i) patch-clamp measurements from ChEMBL bioactivity database; (ii) radioligand binding measurements on mammalian and non-mammalian cell lines; (iii) hERG K⁺ channel binding affinity, and (iv) literature-derived data (Supporting Information, Table S1).^{20, 23, 24, 32} We then implemented three criteria: (i) compounds without well-defined experimental hERG blocking bioactivities were eliminated; (ii) incompatible measuring units were converted to unified IC₅₀ value (μM); (iii) only compounds with IC₅₀ value ≤ 10 μM were considered as hERG blockers, while the rest were regarded as “decoy pool” for further screening according to different threshold settings.

All compounds were converted into SMILES format and duplicate ones were removed via comparing their InChI keys. In this study, for any two duplicated compounds with inconsistent inhibitory activity values derived from different assays, the one with higher IC₅₀ value was preserved for a more rigorous criterion for hERG blockers. In addition, compounds were further processed by molecular washing and energy minimizing using MOE 2010 software²⁹ for protonating strong bases, deprotonating strong acids, removing inorganic counterions, adding hydrogen atoms, generating stereoisomers, and validating single 3D conformers. Finally, a comprehensive collection consisted of 7,889 compounds with well-defined experimental hERG blocking bioactivities was obtained, and 4,355 of the compounds whose experimental values were less than or equal to 10 μM were regarded as hERG blockers (Supporting Information, Table S2 and Table S3).

It is undetermined as to how to define a compound as hERG non-blocker, since what decoy threshold should be used to distinguish blockers and non-blockers has not yet reached a consensus. For example, in Wang’s paper, the threshold was 40 μM, while Didziapetris set it to 10 μM (Supporting Information, Table S1). Herein, for more rigorous concern, we evaluated several thresholds of decoys, including 10 μM, 20 μM, 40 μM, 60 μM, 80 μM and 100 μM for building the negative data sets. Subsequently, all compounds were split into three sets: training set, test set, and validation set with ratios of 8:1:1 via chemical diversity analysis. Chemical diversity analyses were performed by Tanimoto Coefficient based on MACCS fingerprint in MOE 2010. Such a split would assign the chemical structures uniformly and avoid potential data bias. Following standard practice, the training sets were used to build deep learning models, and the validation sets for final evaluation of trained models. It is well known that the performance of deep neural networks is highly sensitive to the selected hyperparameters,³³ thus test sets were applied for tuning model hyperparameters. In this study, hERG blockers were set as “positive” samples, whereas decoys were defined as “negative” samples. The chemical information of the data sets is provided in Supporting Information, Table S4, and the detailed statistics of all data sets used in this study is shown in Table 1.

Table 1.

Detailed description of all data sets used in this study.

Threshold value of decoy (μM)	Training set			Test set			Validation set
	Positive	Negative	Total	Positive	Negative	Total	Positive	Negative	Total
10	3,485	2,826	6,311	435	354	789	435	354	789
20	3,485	1,755	5,240	435	219	654	435	219	654
40	3,485	863	4,348	435	108	543	435	108	543
60	3,485	644	4,129	435	81	516	435	81	516
80	3,485	469	3,954	435	58	493	435	58	493
100	3,485	380	3,865	435	48	483	435	48	483

Chemical representation

We calculated two different types of descriptors for each compound, including MOE 2010 and Mol2vec descriptors.

Mol2vec descriptors

Mol2vec³⁰ is an unsupervised machine learning approach to learn vector representations of molecular substructures. It is inspired by the natural language processing technique Word2vec.³⁴ Mol2vec learns vector representations of molecular substructures that are pointing to similar directions for chemically related substructures. Compounds can be encoded as vectors via summing up the vectors of the individual substructures and then fed into modeling approaches for compound properties prediction. In this study, the Mol2vec model was pre-trained based on a corpus containing 19.9 million compounds and then utilized to feature new samples. Skip-gram method was used with a window size of 10. Finally, 100-dimensional embeddings were generated for all compounds.

MOE descriptors

Although Mol2vec can well represent the chemical structures of molecules, it still has some limitations. Molecular properties, which are also of great importance for hERG blockers prediction, cannot be comprehensively reflected by Mol2vec. Thus, as a common approach to depict molecular properties, 185 two dimensional (2D) molecular descriptors were generated by MOE 2010 software for each compound and concatenated to current Mol2vec descriptors to further enhance the representation. Specifically, 185 2D MOE descriptors cover physical property descriptors, subdivided surface area descriptors, atom count and bond count descriptors, adjacency and distance matrix descriptors, Kier and Hall connectivity and Kappa shape indices descriptors, pharmacophore feature descriptors, and partial charge descriptors. The detailed description of these descriptors can be found in a recent study.²⁹

Methods for model building

In this study, we evaluated two deep learning network algorithms: multi-task deep neural network (DNN) and single-task DNN.³⁵ In addition, we also evaluated two traditional machine learning methods: naïve Bayes (NB) and support vector machine (SVM) implemented by Orange canvas (v3.13.0).³⁶

Multi-task deep neural network

A typical multi-task DNN is composed of interconnected neurons which are arranged hierarchically as layers. The number of neurons in each layer is referred to as “size”. The input layer comprises the neurons for the input vectors based on molecular structure and properties. In the intermediate layers (also known as hidden layers), each hidden neuron applies a weighted sum of the output of the neurons from the previous layer. The output of a hidden layer can be seen as an abstraction of the features of its previous layer. The last layer outputs the prediction results of the model.³⁴ The processed outputs of each hidden layer are shared across all learning tasks and then input to separate models for each different task.³⁵ Since multi-task networks are trained on the joint data, they try to generalize from the data for multiple tasks. The parameters of the shared layers are encouraged to produce joint representations which share information between the tasks.³⁵

One important advantage of multi-task deep network is that it can handle multiple tasks, and simultaneously improves the generalization by leveraging the domain-specific information contained in the training samples of related tasks. Thus, we used different thresholds for decoys, including 10 μM, 20 μM, 40 μM, 60 μM, 80 μM, and 100 μM. Each threshold led to a set of decoys (as negative samples), which was considered as one task when combined with positive samples. The six tasks generated from different thresholds were trained simultaneously by the multi-task DNN.

In this study, all multi-task DNN models comprised three hidden layers with varying input layer sizes according to the number of input features. The first hidden layer had a size of twice as much as that of the input layer. The second and third layers had halved size as the previous ones. The size of the hidden layers were MOE: [400, 200,100], Mol2vec: [200, 100, 50], and MOE+Mol2vec: [600, 300, 150]. Adam³⁸ algorithm was used for optimization and L² normalization term was used for regularization to avoid overfitting. We selected Rectified Linear Unit (ReLU) as the activation function, which was defined as the positive part of its argument:

$$f(x)=x^{+}=max(0,x)$$ (1)

where x is the weighted sum of a neuron.³⁹

Cross entropy was applied as the loss function for the classification task:

$$E=-{\sum }_{j=1}^T{y}_{i}l{n}_{P_j}$$ (2)

where y₁ is the ground truth for total error of the batch as described in eq 3:

$$E=-\frac{1}{N}\sum _{i=1}^N\sum _{j=1}^T{y}_{ij}l{n}_{P_{ij}}$$ (3)

Number of epochs is an important hyperparameter for the training. A larger number of epochs may bring the overfitting problem, while not enough number of epochs may lead to underfitting. According to the average loss of multi-task DNN models based on experimenting with different batch sizes as shown in Supporting Information, Figure S1, the number of epochs was set to 20, since the average loss converged with relatively smaller average loss and lower oscillatory. Hyperparameters, including learning rate, weight decay for the L₂ normalization, dropout rate, and weight initialization were tuned on the test sets with a combination of random hyperparameter search and manual hyperparameter tuning by the hyperparam_search method.^{35, 40}

Single-task deep neural network (Single-task DNN)

An independent single-task neural network for each learning task was trained.⁴¹ The only difference between the multi-task DNN versus single-task DNN is the number of outputs. In the case that a data set contains only a single task, multi-task networks are just single-task network.³⁵ In this study, all parameter settings and architecture of single-task DNN were consistent with those using in multi-task DNN.

In addition, support vector machine (SVM) and naïve Bayes (NB) were also utilized to construct models using the same data sets for comparison. SVM defines a decision boundary that is expressed as a separating hyperplane on the basis of a linear combination of functions parametrized by support vectors.⁴² NB algorithm is a robust classification approach derived from the Bayes theorem with the strong independence assumption that each attribute contributes equally and independently.⁴³ Default parameter settings of these two algorithms were used in this study.

Model evaluation

Model performance was assessed in terms of true positive (TP), true negative (TN), false positive (FP), and false negative (FN). In addition, five metrics, including sensitivity (SE), specificity (SP), prediction accuracy of blockers (Q+), prediction accuracy of non-blockers (Q−), and overall predictive accuracy (Q) were calculated using the following equations.

$$SE=\frac{TP}{TP+FN}$$ (4)

$$SP=\frac{TN}{TN+FP}$$ (5)

$$Q+=\frac{TP}{TP+FP}$$ (6)

$$Q-=\frac{TN}{TN+FN}$$ (7)

$$Q=\frac{TP+TN}{TP+FN+FP+TN}$$ (8)

Moreover, the area under the receiver operating characteristic (ROC) curve (AUC), which indicates the ability of a classifier to distinguish between two classes, was also computed.⁴⁴ A perfect model has an AUC value of 1, whereas random classifier has a value of 0.5.

RESULTS

Chemical diversity analysis

Prior to comparing the performances of different approaches and models, it is pertinent to verify the diversity of the chemical space of the data sets. A large chemical space, to a certain extent, reflects the effectiveness of the application domain of a model.⁴⁵ In this study, the chemical space was analyzed using principal component analysis (PCA)⁴⁶ with 185 MOE descriptors as input. As demonstrated by the chemical space defined by the first two principal components in Figure 2A and Supporting Information, Figure S2, high chemical diversity of datasets as well as expected homogenous distribution (overlap) among the compounds within training sets, test sets, and validation sets are observed. Linear fitting also confirmed the similarities among these data sets. In addition, PCA approach was also employed to observe the difference of chemical space between blocker compounds and decoys of different threshold values (Figure 2B). Obviously, as the value of decoy threshold increases, the distinction between blockers and decoys become more significant. Compare to the linear fitting line of blockers, the fitting line of decoys with threshold value of 10 μM shows highest parallelism, while 100 μM exhibits the lowest.

Figure 2. Distribution of chemical diversity.

(A) Principal component analysis (PCA) of chemical diversity analysis across training sets, test sets, and validation sets. (B) PCA of chemical space between hERG blockers and hERG non-blockers. Linear fittings were utilized to compare the differentiation among different data sets. The closer the fitting lines they present, the more similarity of the chemical space they have.

Evaluation of multi-task DNN models

The hyperparameter batch size defines the number of events that a model reads from the input at a time for stochasticity. In general, smaller batch size leads to a less accurate estimated gradient, but models trained with an oversize batch are more prone be stuck in local optimums.⁴⁷ As discussed above, decoys from a higher threshold can be more distinct from the blockers, but also lead to less samples. Insufficient sample size may bring unsatisfactory performance. Put together, we constructed various multi-task DNN models with different batch size (32, 64, 128, 256, 512, and 1024) as well as various threshold values of decoy (10 μM, 20 μM, 40 μM, 60 μM, 80 μM and 100 μM) for building the best model. We found that the best multi-task DNN model was built with the batch size of 256, which produced the best AUC value in most tasks (Figure 3). Detailed performance of multi-DNN models is listed in Table 2. As expected, the increasing threshold value of decoys does enhance the performance of model, but stops at 80 μM, which may be a result of the lack of negative samples. To sum up, the model with batch size of 256 and decoy threshold value of 80 μM demonstrates greater predictive power than others by achieving the highest AUC values (0.967) and Q values (0.925) on the validation set (Table 3).

Figure 3. Comparison of the area under the receiver operating characteristic curve (AUC) of multi-task deep neural network (DNN) models.

Multi-task DNN were built by different batch sizes on training sets and validation sets. The size of circles denotes the value of AUC. The best results (surrounded by the red dotted box) were achieved with the batch size of 256 and decoy threshold value of 80 μM.

Table 2.

The area under the receiver operating characteristic curve (AUC) values of different thresholds of multi-task DNN models constructed with different batch sizes.

Batch size	Training set						Validation set
	10μM	20μM	40μM	60μM	80μM	100μM	10μM	20μM	40μM	60μM	80μM	100μM
32	0.783	0.840	0.902	0.925	0.942	0.941	0.861	0.889	0.939	0.959	0.964	0.960
64	0.789	0.847	0.912	0.933	0.946	0.946	0.873	0.895	0.948	0.956	0.962	0.955
128	0.783	0.841	0.906	0.929	0.943	0.941	0.867	0.894	0.942	0.957	0.960	0.952
256	0.792	0.847	0.911	0.931	0.944	0.943	0.883	0.899	0.950	0.962	0.967	0.958
512	0.785	0.844	0.910	0.931	0.945	0.945	0.871	0.896	0.948	0.960	0.965	0.959
1,024	0.761	0.820	0.888	0.910	0.930	0.927	0.842	0.881	0.931	0.953	0.960	0.951

Table 3.

Performance of the best multi-task DNN models (batch size = 256) for the training set and validation set.

Threshold value of decoy (μM)	Training set						Validation set
	SE	SP	Q+	Q−	Q	AUC	SE	SP	Q+	Q−	Q	AUC
10	0.767	0.636	0.722	0.689	0.709	0.792	0.816	0.808	0.839	0.781	0.812	0.883
20	0.813	0.697	0.842	0.653	0.775	0.847	0.848	0.781	0.885	0.722	0.826	0.899
40	0.875	0.753	0.935	0.598	0.851	0.911	0.894	0.880	0.968	0.674	0.891	0.950
60	0.884	0.793	0.959	0.559	0.870	0.931	0.906	0.889	0.978	0.637	0.903	0.962
80	0.912	0.817	0.974	0.556	0.901	0.944	0.926	0.914	0.988	0.624	0.925	0.967
100	0.895	0.826	0.979	0.461	0.888	0.943	0.915	0.896	0.988	0.538	0.913	0.958

In this study, molecular properties and structural information were represented using MOE together with Mol2vec descriptors. To further clarify, multi-task DNN models using the MOE descriptors only, Mol2vec descriptors only, and the combination of MOE and Mol2vec descriptors (MOE+Mol2vec) were developed independently with consistent hyperparameters and settings. We found that models building on MOE+Mol2vec demonstrated a higher accuracy across all tasks on both training sets and validation sets (Figure 4). We thus selected the multi-task DNN model building on MOE+Mol2vec descriptors with decoy threshold value of 80 μM as deephERG for further evaluation.

Figure 4. Comparison of the area under the receiver operating characteristic curve (AUC) of multi-task deep neural network (DNN) models on training set (A) and validation set (B).

Multi-task DNN models were built by different threshold values and three different combinations of descriptors: MOE+Mol2vec, Mol2vec alone or MOE alone.

Comparison of different approaches

We further compared the performance of multi-task DNN and several different approaches, including single-task DNN, SVM and NB (Supporting Information, Table S5). We found that multi-task DNN models offer better accuracies consistently over single-task DNN models across all tasks on both training set and validation set (Figure 5 and Supporting Information, Figure S3).

Figure 5. Receiver operating characteristic (ROC) curves of multi-task deep neural network (DNN) model, single-task DNN model and the two traditional machine learning methods: support vector machine (SVM) and naïve Bayes.

All models were built and evaluated based on the biological endpoint of hERG at 80 μM on training set (A) and validation set (B). AUC: the area under the receiver operating characteristic curve.

Furthermore, comparison was drawn between deephERG and two traditional machine learning methods (SVM and NB). The performance of deephERG significantly outperforms traditional machine learning methods as well (Supporting Information, Table S5). For example, as shown in Figure 5, the AUC value of 0.967 for deephERG on validation set with decoy threshold values of 80 μM is higher than those for NB (AUC = 0.930) and SVM (AUC = 0.876).

Identification of potential hERG blockers from approved drugs

Via deephERG, we next turned to perform risk assessment of hERG blockers for 1,824 FDA approved small molecule drugs available from DrugBank database.³¹ All these drugs were processed by molecular washing and energy minimizing as described previously after eliminating antibodies or ions drugs. Among the 1,824 drugs, 539 drugs were computationally predicted to have potential hERG inhibitory activities by deephERG (Figure 6). Among 15 drug categories defined by Anatomical Therapeutic Chemical Classification System (ATC code), five categories with top number of predicted hERG blockers included: nervous system (N), cardiovascular system (C), alimentary tract and metabolism (A), antineoplastic and immunomodulating agents (L), and respiratory system (R). Literature evidences confirmed some of our predictions. For example, prochlorperazine, a typical antipsychotic agent, induced a concentration-dependent decrease in current amplitudes at the end of the voltage steps and tail currents of hERG.⁴⁸ Ivermectin, as a broad-spectrum antiparasitic agent, in vitro assay showed that it has an effect on the rapid delayed rectifier current mediated by the K(+) ion channel encoded by hERG.⁴⁹ In addition, a hERG inhibitory effect was observed on somatostatin and its derivatives as well.⁵⁰ The detailed prediction results for 1,824 approved drugs by deephERG are provided in Supporting Information, Table S6. We next turned to focus on the deephERG-predicted cardiotoxicity risk on approved anticancer agents (colored in red in Figure 6).

Figure 6. hERG-mediated cardiotoxic risk assessment for 1,824 FDA-approved drugs by deephERG.

Note: A : alimentary tract and metabolism; B: blood and blood forming organs; C: cardiovascular system; D: dermatologicals; E: genito-urinary system and sex hormones; G: systemic hormonal preparations; H: excluding sex hormones and insulins; J: antiinfectives for systemic use; L: antineoplastic and immunomodulating agents; M: musculo-skeletal system; N: nervous system ;P: antiparasitic products, insecticides and repellents; R: respiratory system; S:sensory organs; V: various; Z: unknown. The deephERG-predicted drugs with predicted positive probabilities more than 0.5 were illustrated. Drugs with the top 10% highest predicted positive probability in each category are displayed enlarged on the outer circle, while all the predicted positive drugs are presented on the inner circle. Drugs existing in the positives of training sets are underlined. The high-resolution version is provided in Supporting Information.

Evaluation of cardiotoxicity risk of anticancer agents by deephERG

The growing awareness of cardiovascular complications associated with cancer or cancer treatment has led to the emerging field of cardio-oncology (also known onco-cardiology), which centers on screening, monitoring, and treating cancer patients with cardiac dysfunction. Furthermore, it is also an exciting field because there is no guidelines and non-available FDA-approved therapeutics in terms of how to prevent new cardiotoxicity in cancer survivors. In addition to myocardial cell, hERG channel is also expressed highly in tumor cells and is related to the regulation of cell apoptosis and proliferation.^{51, 52} Thus, we further investigated whether we could identify new hERG channel blockers on approved anticancer agents by deephERG (Supporting Information, Table S7). Figure 7 shows the more intuitive prediction results for 49 anticancer drugs by deephERG. In total, 15 out of 49 drugs whose hERG inhibitory activities have been validated by clinical case report or reported pre-clinical data are highlighted in bold (Supporting Information, Table S8).

Figure 7. hERG-mediated cardiotoxic risk assessment for 49 approved antineoplastic drugs (including immunomodulating agents) by deephERG.

Drugs which hERG inhibitory activity had been validated by reported experimental data are highlighted in bold font (Supporting Information, Table S8). Drugs existing in the positives of training sets are underlined, and tyrosine kinase inhibitors are marked with ‘*’. The deephERG-predicted drugs with predicted positive probabilities more than 0.5 were illustrated.

Among protein kinase inhibitors, tyrosine kinase inhibitors (TKIs) occupy a large proportion. TKIs have played important roles in the molecularly targeted treatment in multiple cancer types. We found that 24 TKIs were predicted as hERG blockers by deephERG and 9 of them had been validated by clinical case reports, in vitro assays, and literatures (Figure 7 and Supporting Information, Table S8). Among them, erlotinib was detected to have potential hERG inhibition at the concentration of 10 μM by an automated patch-clamp assay, consistent with deephERG-prediction.⁵³ In addition, sunitinib, crizotinib, and nilotinib were also confirmed to block potently the hERG channel with IC₅₀ values of 0.5 μM, 1.7 μM, and 0.7 μM, respectively.⁵³ Ponatinib, a multi-targeted TKI and potent pan-ABL inhibitor, was temporarily withdrawn from the U.S. market due to severe vascular adverse events.⁵⁴ Herein ponatinib was predicted as a hERG blocker by deephERG. An in vitro assay conducted in human embryonic kidney cells stably expressing the hERG potassium channel revealed that ponatinib inhibits hERG at a concentration above 1 μM.⁵⁵ Vandetanib, a multi-kinase inhibitor approved for treatment of multiple cancer types, has been reported to inhibit the hERG at 3 μM detected by a whole-cell patch-clamp assay in transiently transfected HEK293 cells.⁵⁶ Finally, several new deephERG-predicted TKIs without literature data also need to be highly vigilant, such as bosutinib, gefitinib, afatinib and dasatinib, which are warranted by experimental or clinical validation in the future.

In addition to TKIs, we also computationally predicted multiple cytotoxic chemotherapeutic agents as potential hERG blockers as well by deephERG, such as dactinomycin, tamoxifen and amsacrine. For example, the hERG inhibitory activity of dactinomycin from cytotoxic antibiotics and tamoxifen from endocrine therapy had been validated by in vitro assays.^{57, 58} Tamoxifen inhibited hERG at −50mV in a concentration-dependent manner with IC₅₀ value of 1.2 μM.⁵⁷ Dactinomycin markedly reduced the hERG mRNA levels as well.⁵⁸ Moreover, amsacrine, an antineoplastic agent used in acute lymphoblastic leukemia, had been confirmed to block hERG in HEK 293 cells with IC₅₀ values of 209.4 nM.⁵⁹ Finally, we computationally identified multiple novel hERG blockers by deephERG, such as picamycin and bleommycin. In summary, 23% (10/44, excluding 5 duplicated drugs existing in the positives of training sets) of our predictions of antineoplastic and immunomodulating agents had been validated successfully by clinical case reports, experiment evidences, and literatures, indicating that deephERG offer a useful tool for risk assessment of potential hERG-mediated cardiotoxicities in drug discovery and post-marketing surveillance.

DISCUSSION

In this study, we developed a deep learning approach, termed deephERG, for cardiotoxicity risk assessment of small molecules mediated by hERG blockers. We highlight several improvements compared to previous studies.^{20, 23, 24, 32} First, we assembled chemical diverse hERG channel blockers with well-defined experimental endpoints by our sizeable efforts, which increase the application domains of deephERG models. To our knowledge, this is the largest publicly available data set with well-defined experimental hERG blocking bioactivity values. We believed that this dataset is expected to represent comprehensive coverage of known hERG blockers. Second, we demonstrated that deephERG models built by multi-task DNN algorithm outperformed single-task DNN and traditional machine learning models, such as SVM and NB. Finally, all codes (multi-task DNN and single-task DNN) and the Mol2vec approach used in this study can be conveniently implemented using a single python script, which is freely available at: https://github.com/ChengF-Lab/deephERG.

Several potential shortcomings should be acknowledged. First, although we assembled a large-scale compounds with well-defined experimental data on hERG channel from published literatures based on our sizeable efforts, the incompleteness of samples still exists, especially for the negative compounds without known hERG inhibitory activities. As we discussed above, models trained based on higher threshold value of decoys had shown stronger predictive ability, but the insufficiency of learning samples hinder the advance. Therefore, a more comprehensive collection deriving from multiple sources should be further integrated in the future. Second, the Mol2vec approach applied for chemical representation in this study was based on a pre-trained compound corpus of 19.9 million compounds, which may contain inefficient and redundant features. Replacement of the currently used corpus by a trimmed-down data set, such as marketed drugs, may characterize chemical structures more pointedly and efficiently. Third, although the influence of model performance by different decoy threshold settings had been discussed, their underlying cause should be further delved. Fréchet ChemNet distance⁶⁰ is an innovative approach to compare the distributions of sets chemical compounds, which may facilitate to explain the reason from the chemical space angle. Finally, although some of our predictions on approved drugs have already been confirmed by in vitro assay, clinical reports, and literatures, some novel deephERG-predicted hERG inhibitory drugs should be further validated by experimental assays or pharmacoepidemiologic analyses from the real-world patient data in the future.^{11, 61}

CONCLUSIONS

In this study, we proposed a deephERG framework to build predictive models for evaluation of hERG channel blockers based on the largest dataset containing 7,889 compounds with well-defined experimental endpoints. DeephERG model built by a multi-task DNN algorithm shows a satisfactory predictive ability with AUC values of 0.944 and 0.967 on training set and validation set respectively, which is superior than single-task DNN and traditional machine learning approaches, including support vector machine and naïve Bayes. Finally, we utilized the best deephERG model for potential hERG-mediated cardiotoxicity risk assessment on over 1,800 FDA-approved drugs. In case studies, we showed that the predicted cardiotoxicities for approved antineoplastic by deephERG were validated by reported experimental data from in vitro assays and clinical case reports. If broadly applied, deephERG presented here offer useful in silico tools for hERG-mediated cardiotoxicity risk assessment of small molecules in drug discovery and post-marketing surveillance.