Virtual Screening of Pharmaceutical Compounds with hERG Inhibitory

Activity (Cardiotoxicity) using Ensemble Learning

Aditya Sarkar and Arnav Bhavsar

Indian Institute of Technology, Mandi, India

Keywords:

Ensemble Learning, Feature Selection, Virtual Screening, Cardiotoxicity, Pharmaceuticals.

Abstract:

In silico prediction of cardiotoxicity with high sensitivity and speciﬁcity for potential drug molecules can be

of immense value. Hence, building machine learning classiﬁcation models, based on some features extracted

from the molecular structure of drugs, which are capable of efﬁciently predicting cardiotoxicity is critical. In

this paper, we consider the application of various machine learning approaches, and then propose an ensem-

ble classiﬁer for the prediction of molecular activity on a Drug Discovery Hackathon (DDH) (1st reference)

dataset. We have used only 2-D descriptors of SMILE notations for our prediction. Our ensemble classiﬁca-

tion uses 5 classiﬁers (2 Random Forest Classiﬁers, 2 Support Vector Machines and a Dense Neural Network)

and uses Max-Voting technique and Weighted-Average technique for ﬁnal decision.

1 INTRODUCTION

It is well known that drug discovery is complex, long-

drawn, and requires interdisciplinary expertise to dis-

cover new molecules. Drug safety is an important

issue in the process of drug discovery. Failure in

Clinical trials in the 2000s was majorly due to efﬁ-

cacy and safety (approx 30%) (Kola, I. and Landis,

J., 2004). One important aspect of drug safety is

drug toxicity. Frequently observed toxicities are car-

diotoxicity, hepatotoxicity, genotoxicity, and photo-

toxicity (Keiji Ogura, 2019). Toxicological screening

is very important for the development of new drugs

and for the extension of the therapeutic potential of

existing molecules. The US Food and Drug Adminis-

tration (FDA) states that it is essential to screen new

molecules for pharmacological activity and toxicity

potential in animals (21CFR Part 314). The toxic ef-

fects of chemicals, food substances, pharmaceuticals,

etc., have attained great signiﬁcance in the 21st cen-

tury (Parasuraman, S, 2011).

The h-ERG (human Ether-

a-go-go-Related Gene)

is a gene that codes for a protein known as K

11.1,

the alpha subunit of a potassium ion channel (”hERG

safety”,2018). The h-ERG potassium channels (Sny-

ders, 1999) are essential for normal electrical activity

in the heart. When this channel’s conductivity of elec-

tric current is inhibited by some action of drugs, it can

lead to a fatal disorder called Long QT Syndrome.

It is found that many drugs have the h-ERG

inhibitory activity which can prolong the QT and

thereby resulting in irregularity of the heartbeat called

Torsades de Pointes (Keiji Ogura et al, 2019). As

a result, many drugs, that are inhibiting the h-ERG

channel’s conductivity, have been withdrawn from the

markets. Hence, it is regarded as a major anti-target

for drug discovery.

Over the years, many works have been done to

achieve the goal of classifying compounds having

h-ERG inhibitory activity. In early drug discovery

stages such as screening h-ERG inhibitory activity,

performing costly and time-consuming assays is difﬁ-

cult. Hence, developing an in-silico model to predict

hERG inhibition is very useful. (Keiji Ogura et al,

2019) Machine Learning techniques can be used for

classifying if any compound is having the inhibitory

activity or not. Machine Learning model can learn

features that can classify any compound on the basis

of h-ERG inhibition activity.

There have been some recent classiﬁcation models

reported for h-ERG inhibition which have used Neu-

ral Networks (B. Mehlig, 2019), Random Forest Clas-

siﬁers (RF) (Leo Breiman, 2001) and Support Vector

Machines (SVM) (Yongli Zhang et al, 2012). Below

we provide a brief review of these.

1.1 Related Work

Czodrowski (Czodrowski,,2013) constructed RF

models using descriptors calculated by Rdkit , based

152

Sarkar, A. and Bhavsar, A.

Virtual Screening of Pharmaceutical Compounds with hERG Inhibitory Activity (Cardiotoxicity) using Ensemble Learning.

DOI: 10.5220/0010267701520159

In Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2021) - Volume 2: BIOIMAGING, pages 152-159

ISBN: 978-989-758-490-9

Table 1: Details of the various datasets.

Integrated Dataset

Database hERG inhibitors Inactive

compounds

ChEMBL (version 22) 4793 5275

GOSTAR 3260 3509

NCGC 232 1234

hERGCentral 4,321 274,536

hERG integrated dataset 9,890 281,329

on 3,721 compounds measured in a binding assay

and 765 compounds measured in a functional assay

collected from ChEMBL. The prediction models

were constructed separately from each data set, and

showed prediction accuracies of 79.7%–80.1% and

69.2%–90.7%, respectively.

Wang et al. (Wang, S. et al, 2012) developed

hERG classiﬁcation models using naive Bayesian

classiﬁcation and recursive partitioning based on

molecular properties and the ECFP 8 fngerprints, and

recorded 85% accuracy (Wang, S. et al, 2016).

Schyman et al. (Schyman. P, 2016) combined 3D

(David C. Kombo et al,2012) similarity conformation

and 2D similarity ensemble approach, and achieved

69% sensitivity and 95% speciﬁcity on an indepen-

dent external data set.

Recently, Keiji Ogura, Tomohiro Sato, Hitomi

Yuki and Teruki Honma (Keiji Ogura et al, 2019) used

Support Vector Machines (SVMs) on an integrated h-

ERG database having more than 291,000 structurally

diverse compounds. They achieved kappa statistics of

0.733 and accuracy of 98.4%. Table 1 provides a sum-

mary of various datasets used across existing works.

They have made the dataset publicly available for re-

search purposes.

1.2 Contributions of this Work

Most of the works (Wang, S. et al, 2012),(Wang,

S. et al, 2016),(Schyman. P, 2016),(Keiji Ogura et

al, 2019) which we have reviewed have either used

descriptors (2-D, 3-D and 4-D) or ﬁngerprints. On

the other hand, unlike the above, we have used only

2-D descriptors for our classiﬁcation model. 2D-

descriptors deal with the molecular topology of the

compounds i.e. topological indices and fragment

counts. 2D-Descriptors incorporate precious chemi-

cal information like size, degree of branching, ﬂexi-

bility etc. Generating 2D descriptors of the SMILES

compounds usually takes less time than 3D descrip-

tors. Even with just 2D descriptors, we demonstrate

that the proposed ensemble model achieves a very

good performance.

In this study, we have developed an ensemble

model having two Random Forest Classiﬁers, two

Support Vector Machines and one Dense Neural Net-

work which achieved a AUC score (Area Under the

ROC Curve) of 0.96 and Cohen’s Kappa of 0.9195.

Most of the existing models have used Sup-

port Vector Machines, Random Forest Classiﬁers and

Naive Bayesian Classiﬁers for prediction. However,

in addition to these models, we have also used Deep

Neural Networks and two different Ensemble classi-

ﬁers for the task. We have found that the Deep Neural

Networks and the Ensemble classiﬁer yield the high-

est performance.

We have also worked with data augmentation for

our class-imbalanced dataset. We have used SMOTE

(Synthetic Minority Oversampling Technique) (N. V.

Chawla, et al, 2011) for augmenting data. Data aug-

mentation is a very useful procedure because the data

that it generates is almost similar to the training data.

For some diseases, the drugs available can be quiet

less, and doing prediction with less data points can

lead to over-ﬁtting. Data Augmentation can prove

useful by not only creating new data but may also

help in understanding the underlying distribution of

each property (descriptors or ﬁngerprints) of drugs.

Unlike most existing works, we also sug-

gest an automatic approach based on information

gain/entropy to shortlist (or select) features from a

larger set. To our knowledge, the only exception

among the existing methods, which are considered

such a selection is work by (Keiji Ogura et al, 2019)

which involves the NSGA-II (Non-dominated Sorting

Genetic Algorithm-II) for descriptor selection.

Finally, towards the end of the paper, we pro-

vide a consolidated summary of the various works in

this domain, the datasets used, the feature descriptors

and methods employed, and their performance across

several metrics. We note that although this work is

not analysing imaging data, it involves core machine

learning on an important problem in biology. Con-

sidering that this is a relatively recent application do-

main, such an overview provides a good perspective

of the trade-offs of the approaches and paves the way

for more standardized benchmarking and extensions

in this area.

2 DATASET AND DESCRIPTORS

In this work, we have used a dataset provided

in one of the competitions of the Drug Discovery

Hackathon, organized by the Govt. of India; The

dataset has been made by Dr Kunal Roy, profes-

sor from Department of Pharmaceutical Technology,

Jadavpur University. http://sites.google.com/

Virtual Screening of Pharmaceutical Compounds with hERG Inhibitory Activity (Cardiotoxicity) using Ensemble Learning

153

(a) Pandas view of Dataset (b) Pandas view of cleaned Dataset

Figure 1: Pandas Dataframe images.

site/kunalroyindia/home. It contains the SMILE

notations of 8227 pharmaceutical compounds, along

with their h-ERG inhibitory activity label (i.e. blocker

or non-blocker). Out of these, 6878 compounds were

blockers and 1349 were non-blockers.

We have used mordred (Moriwaki H, et al, 2018)

Python module to decrypt the SMILE (Anderson, E.

et al, 1987) notations to 2-D descriptors. As a result,

we got 1613 features for each compound. Figure 1(a)

shows a snippet of the Pandas view of the dataset.

It shows the IDs of the pharmaceutical compounds,

the class to which it belongs i.e. Blocker or Non-

Blocker (Blocker means that the compound possesses

the h-ERG inhibitory activity and Non-Blocker

means that it does not possess the h-ERG inhibitory

activity) and few 2D descriptors. The list of de-

scriptors along with their descriptions can be found

at mordred documentation link: https://mordred-

descriptor.github.io/documentation/master/ descrip-

tors.html. Initially, we have extracted all the 2-D

descriptors that mordred has offered. In section III-C,

we have described a method for further selecting

features out of these.

3 METHOD

We have ﬁrst worked on data cleaning, then aug-

mented our data because it was having an imbal-

anced class problem, selected important features us-

ing an Entropy/Information Gain based method, and

ﬁnally performed the classiﬁcation using two ensem-

bles - Max-Voting Ensemble and Weighted-Average

Ensemble using two variants of Random Forest Clas-

siﬁers, two variants of Support Vector Machines and

one Deep Neural Network. We have divided this

section into 5 subsections - Data Cleaning, SMOTE

Application, Feature Selection and Base Models and

Ensemble-Learning.

3.1 Data Cleaning

Data Cleaning is important part of analysing this data.

This is done so as to eliminate outliers present in the

data. These outliers are due to miscalculations made

by the mordred python module. Since the range each

column is different, so to normalize them, Z-score/

Standard Score is used.

Z-score or standard score of a particular column

is deﬁned as the number of standard deviations by

which the value of a datapoint value is above or be-

low the mean value of datapoints present in the col-

umn. Raw scores above the mean have positive Z-

scores, while those below the mean have negative

Z-scores (Spiegel, Murray R.; Stephens, Larry J ,

2008),. Mathematically, it is deﬁned as -

Z-Score =

X−µ

where x is the sample datapoint, µ is mean of all the

datapoints in the sample column and σ is standard de-

viation in the sample column.

We have used Z-scores for ﬁnding the out-

liers and ﬁnally replaced the outliers with the

mean of column of the Dataframe, to which it

belongs. The Dataframe is two-dimensional,

size-mutable, potentially heterogeneous tab-

ular data (https://pandas.pydata.org/pandas-

docs/stable/reference/api/pandas.DataFrame.html).

This can be interpreted as removing the outlier

samples, and augmenting the rest of the samples,

with a mean estimate.

We have considered a datapoint, an outlier when

BIOIMAGING 2021 - 8th International Conference on Bioimaging

154

Z − Score(datapoint)

(PeruriVenkataAnusha et. al.). A section of the

dataset view obtained after cleaning is shown in Fig-

ure 1(b). It shows the class to which the pharmaceuti-

cal compound belongs to i.e. Blocker or Non-Blocker

and few 2D descriptors. However, unlike Figure 1(a),

this contains Z-score. Also, we have removed those

features which were containing NaN values. Hence,

our number of features decreased from 1613 to 1375.

3.2 SMOTE for Data Augmentation

As mentioned earlier, our dataset has 6878 Blocker

compounds and 1349 non-Blocker compounds.

Hence, it is imbalanced and can lead to high bias.

To tackle this problem, we made use of the pop-

ular data augmentation method SMOTE (Synthetic

Minority Oversampling Technique). SMOTE (N. V.

Chawla, 2011) can be used to create synthetic exam-

ples for the minority class. It works by ﬁrst choosing

a random example from the minority class Then k of

the nearest neighbors for that example are found. A

randomly selected neighbor is interpolated between

the two examples in feature space. We chose SMOTE

considering its popularity, with which we were also

getting an improvement in our results. However, one

can also consider other data augmentation methods.

We have used Imblearn Python Module (Guillaume

Lema

ıtre et al, 2017) for applying SMOTE. After

SMOTE, we achieved a total size of 13756 datapoints,

including 6878 Blocker compounds and 6878 Non-

Blocker compounds.

3.3 Feature Selection

Feature selection is an important part of the model.

Considering we have numerous 2-D features, it is im-

portant to consider the ones which can contribute for

the task of the classiﬁcation.

Gini-index and Entropy are used as criterion for

calculating information gain. Decision tree algo-

rithms use information gain to split a node. Entropy

and gini are used for measuring impurity of a node.

Node having multiple classes is considered impure

and node having single class is considered pure. In

this project, we have used Entropy as our impurity

measure. While training a tree, we can compute how

much each feature decreases the impurity. The more

a feature decreases the impurity, the more important

that feature is. In random forests, the impurity de-

crease from each feature can be averaged across trees

to determine the ﬁnal importance of the variable. For

this we have used the features selected by Random

Forests with measure of impurity as Entropy. Initially,

we started with 1375 features for each compound. Af-

ter applying feature selection, we had 592 features,

which is a signiﬁcant reduction.

3.4 Base Models and Ensemble

Learning

We have used 5 base models in our ensemble - 2 Ran-

dom Forest Classiﬁers, 2 Support Vector Machines

and 1 Dense Neural Network. These well known

methods are discussed brieﬂy in the following subsec-

tions, and the corresponding parameters are provided

in Table 2. We have used Scikit-Learn Python mod-

ule for Random Forest Classiﬁers and Support Vec-

tor Machines. For implementing Dense Neural Net-

works, we have used Tensorﬂow 2 and Keras.

3.4.1 Random Forest Classiﬁers

The random forest classiﬁer is essentially a ensemble

of decision tree-based classiﬁers. It operates by con-

structing a number of decision trees during its train-

ing and outputs a class decision that is the mode of the

classes estimates of each decision tree. It is based on

the principle of bagging to mitigate the bias-variance

trade-off in decision tree-based classiﬁers.

3.4.2 Support Vector Machines

SVM is a popular maximum-margin classiﬁcation

framework boasting advantages of good generaliza-

tion and non-linear classiﬁcation via the use of kernel

functions. Support Vector machines ﬁnd a maximum-

margin-hyperplane that divides the data points of the

classes such that the distance between the hyperplane

and the nearest point of either class to this hyperplane

is maximised.

3.4.3 Dense Neural Networks

Dense Neural Networks have gained popularity as

contemporary classiﬁers due to their ability to learn

highly non-linear classiﬁcation models, given enough

data. A neural network is a network of neurons that

can well approximates a highly nonlinear boundary

between the classes, given enough data.

3.4.4 Ensemble Learning

For our Ensemble, we have used Max-Voting and

Weighted-Average with our 5 base models described

above. In max-voting, each base model makes a pre-

diction and votes for each sample. Only the sam-

ple class with the highest votes is included in the ﬁ-

nal predictive class. In weighted-average, we have

Virtual Screening of Pharmaceutical Compounds with hERG Inhibitory Activity (Cardiotoxicity) using Ensemble Learning

155

Figure 2: ROC curves for all models.

Table 2: Accuracy of various classiﬁcation models.

Base Model Accuracies and ROC-AUC

Models Features

Dense Neural Networks 4 dense layers : 3 with L1 param 10

−5

L2 param 10

−4

, dropout of 0.5,0.3, 0.3

resp. 1 with L1 param 10

−4

, L2 param

−4

, Optimiser: ADAM, Activation

func: Sigmoid.

Random Forest Classiﬁer-1 89 trees, min no. of samples at leaf node

: 2 and to split an internal node : 4, feat.

selection criteria : Entropy-Based.

Random Forest Classiﬁer-2 70 trees, min no. of samples at leaf node

: 10 and to split an internal node : 5, feat.

selection criteria : Gini-Based.

Support Vector Machines-1 regularization parameter : 1

Support Vector Machines-2 regularization parameter : 0.8

placed weights on predictions of each of the base

model for the ﬁnal prediction. The weights we have

placed is 0.75 for DNN, 0.1 for RF-1, 0.07 for RF-

2, 0.05 for SV-1 and 0.03 for SV-2. We have given

more weights to that classiﬁers among the base mod-

els, which yields a higher accuracy. Since, our Neu-

ral network is showing maximum accuracy, we have

placed maximum weightage to it.

4 EXPERIMENTS AND

OUTCOMES

For our experiments, we have divided our dataset in

the training-testing ratio of 70% and 30%.As a result,

our training set contains 9629 datapoints and testing

set has 4127 datapoints. The training data is split into

training and validation sets, automatically by the in-

built models in the packages that we employ. We use

testing data only for prediction.

We have tested the performance of the base classi-

ﬁers as well as the ensemble classiﬁers using several

metrics, in addition to the overall accuracy. These are

deﬁned below:

4.1 Description of Metrics

The metrics we have used for classifying our ﬁnal

classiﬁcation models are descrbed below:

4.1.1 AUC-ROC Score

AUC-ROC score computes Area Under the Receiver

Operating Characteristic Curve (ROC) from predic-

tion scores. The ROC curve is shown in Figure 3. We

note from all ROC curves that a high True Positive is

achieved at fairly low value of False positives.

BIOIMAGING 2021 - 8th International Conference on Bioimaging

156

4.1.2 Sensitivity, Precision and Speciﬁcity

Abbreviating TN for True Negative, TP for True Pos-

itive, FN for False Negative and FP for False Positive,

the other metrics are deﬁned as

Sensitivity =

T P

T P+F N

, Speciﬁcity =

T N

T N+FP

Bal. Acc. =

Sensitivity+Speci ficity

, Precision =

T P

T P+F P

4.1.3 Cohens-Kappa (κ)

We have also used Cohens-Kappa (κ) (J. Cohen,

1960), which determines the level of agreement be-

tween two annotators in a classiﬁcation problem.

κ =

−p

1−p

where p

is deﬁned as empirical probability of agree-

ment on the label assigned to any sample and p

deﬁned as expected agreement when both annotators

assign labels randomly.

4.1.4 Matthews Correlation Coefﬁcient (MCC)

Matthews correlation coefﬁcient (MCC) (Matthews,

B. W., 1975) is generally regarded as a balanced mea-

sure which can be used in case of data imbalance be-

tween classes.

4.2 Results

The accuracy and the ROC-AUC values of the the

base models as well as the ensemble models are pro-

vided in Table 3. We note that among the base clas-

siﬁers, the neural network model performs the best.

The accuracy of the ensemble model with average

weighting is similar to the DNN model. It is likely

that the ensemble learning method with max voting is

performing relatively lower, because of the some low

performing base classiﬁers.

In Table 4, we provide the results of the top two

performing classiﬁers from Table 3., for the other

metrics that were deﬁned above. We note that for

both the approaches yield high quality classiﬁcation

across all the metrics, and their performance is close

to each other. While for this dataset, the DNN model

performs somewhat better than the ensemble learn-

ing approach, it is important to acknowledge the high

performance of the ensemble learning, as for larger

datasets, it is known that the ensemble strategies can

typically better mitigate the bias-variance tradeoff.

4.3 A Review of the Results for Existing

Methods

In Table 5, we provide a overall summary of the var-

ious existing methods for the task of classiﬁcation of

Table 3: Accuracy of various classiﬁcation models.

Base Model Accuracies and ROC-AUC

Models Accuracies AUC

Dense Neural Networks 95.98% 0.994

Random Forest Classiﬁer-1 94.3% 0.944

Random Forest Classiﬁer-2 93.51% 0.932

Support Vector Machines-1 90.87% 0.914

Ensemble learning (Max voting) 93.96% 0.94

Ensemble learning (Avg. weighting) 95.97% 0.96

Table 4: Metrics for Ensembles.

Metrics for different Ensembles

Metrics Weighted-Average DNN

AUC 0.960 0.994

Sensitivity 0.9436 0.975

Speciﬁcity 0.9755 0.983

Balanced Accuracy 0.9596 0.979

MCC 0.9199 0.956

Cohen’s kappa 0.9195 0.956

F1-score 0.96 0.98

Precision 0.96 0.975

Recall 0.96 0.975

pharmaceutical compounds based on their hERG in-

hibition activity. We note that while there have been

a few (but not many) methods to address this task,

these involve different datasets, and different features.

Thus, while such a summary is not a comparison, it

does provide, under one roof, a perspective on data,

methods, and can help in identifying scope of im-

provements in this area.

In the table, some cells are blank as not all met-

rics are provided for all methods. We note that the

methods do not yield a high performance across all

the metrics. Some methods employ relatively small

data. The work by Ogura et al, involves the largest

data, but yields a low sensitivity and kappa coefﬁ-

cient. Such discrepancy suggests an issue with the

data imbalance. Importantly, most of the approaches

use various different features in their methodology.

In contrast our approach uses only 2D features, and

yields a high performance across all metrics. A lim-

itation of this work is that it also involves relatively

less data, which we plan to address in future.

5 CONCLUSIONS AND FUTURE

WORK

In this work we have compared various standard ma-

chine learning methods for the task of pharmaceutical

compound classiﬁcation based on their hERG inhibi-

Virtual Screening of Pharmaceutical Compounds with hERG Inhibitory Activity (Cardiotoxicity) using Ensemble Learning

157

Table 5: A summary of the various approaches and their performance.

Reference Database Data size Classiﬁers Features AUC Sensiti-

vity

Speciﬁ-

city

Cohen’s

Kappa

Acc.

Our Model DDH 8227 SVM, RF,

DNN

2D descr- iptors 0.994 0.9436 0.9755 0.9195 0.959

Czodrowski P,

2013

ChemBL 11958 RF RDKIT descrip-

tors

0.564 0.029-

0.243

- - 0.907

Wang S et al, 2016 - 587 Na

ıve

Bayes,

SVM

Pharmaco- phore

hypothesis

0.899 0.943 0.596 - 0.782-

0.836

Ogura et al, 2019 hERG-Integrated Dataset 291219 SVM 2D 3D descrip-

tors, ECFP-4

structural

ﬁngerprints,

Pipeline Pilot

descriptors

0.966 0.715 0.933 0.733 0.98

Schyman P., 2016 National Cancer Insti.

Database

25000 - Accelrys

extended

connectivity

ﬁngerprints,

conformations

- 0.69 0.95 - 0.79

Doddareddy et al Dubus203, Literature368,

Thai313 datasets

7360 LDA, SVM Extended

connectivity

ﬁngerprints,

functional class

ﬁngerprints

0.94 - - - 0.91

Kwang-Eun et al Pipeline Pilot

(PP),FCFP 2, FCFP 4 and

FCFP 6, R package

5252 DNN, NB,

SVM, RF,

Bagging

integer and bi-

nary type ﬁnger-

prints

0.95 0.626 0.986 - -

Chuipu Cai et al ChEMBL, hERG K+ chan-

nel binding afﬁnity, radi-

oligand binding measure-

ments on mammalian and

non- mammalian cell lines,

literature-derived data

7889 DNN,

GCNN

Molecular

Operating

Environment

descriptors,

Mol2vec

descriptors

0.97 0.912 0.817 - 0.93

tion activity. Some of the important aspects that we

have considered is the used of only 2D features, data

augmentation, feature selection, and ensemble learn-

ing. The accuracies we have achieved with our model

is quite high for a small dataset.

The work encourages us to further explore more

ensemble strategies considering DNN features, stack-

ing, bagging etc.

REFERENCES

Anderson, E., G.D. Veith, and D. Weininger. 1987.

SMILES: A line notation and computerized in-

terpreter for chemical structures. Report No.

EPA/600/M-87/021. U.S. Environmental Protection

Agency, Environmental Research Laboratory-Duluth,

Duluth, MN 55804

B. Mehlig, 2019, Artiﬁcial Neural Networks. arxiv -

1901.05639

Chuipu Cai, Pengfei Guo, Yadi Zhou, Jingwei Zhou, Qi

Wang, Fengxue Zhang, Jiansong Fang, and Feixiong

Cheng, 2015. Deep Learning-based Prediction of

Drug-induced Cardiotoxicity. J Chem Inf Model. 2019

Mar 25; 59(3): 1073–1084.

Czodrowski, P., 2013 HERG Me Out. J. Chem. Inf.

Model.53, 2240–2251 (2013).

David C. Kombo, Kartik Tallapragada, Rachit Jain, Joseph

Chewning, Anatoly A. Mazurov, Jason D. Speake,

Terry A. Hauser, and Steve Toler,2012. 3D Molecular

Descriptors Important for Clinical Success. J. Chem.

Inf. Model. 2013, 53, 2, 327–342

Drug Discovery Hackathon, https://innovateindia.mygov.

in/drug-ps/track-2-general-drug-discovery-including-

covid/ddt2-13/

Doddareddy MR; Klaasse EC; Ijzerman AP; 2010. Ben-

der A Prospective Validation of a Comprehensive in

Silico hERG Model and its Applications to Commer-

cial Compound and Drug Databases. Chemmedchem

2010, 5, 716–729 [PubMed: 20349498].

Guillaume Lema

ıtre and Fernando Nogueira and Christos

K. Aridas,2017. Imbalanced-learn: A Python Tool-

box to Tackle the Curse of Imbalanced Datasets in

Machine Learning Journal of Machine Learning Re-

search.

”hERG safety”. Cyprotex, 9 October 2018. https://www.cy

protex.com/toxicology/cardiotoxicity/hergsafety

J. Cohen (1960). “A coefﬁcient of agreement for nominal

scales”. Educational and Psychological Measurement

20(1):37-46.

Kola, I. & Landis, J.,2004. Can the Pharmaceutical Industry

Reduce Attrition Rates? Nat. Rev. Drug Discov. 3,

711–715 (2004)

Laura E. Raileanu, Kilian Stoffel, 2004. Theoretical Com-

parison between the Gini Index and Information Gain

Criteria. Annals of Mathematics and Artiﬁcial Intelli-

gence 41, 77–93 (2004)

Leo Breiman, 2001. Random Forests, Statistics Depart-

ment, University of California, Berkeley, CA 94720.

Springer

BIOIMAGING 2021 - 8th International Conference on Bioimaging

158

Marchese Robinson, R. L.; Glen, R. C.; Mitchell, J. B.,

2011. Development and comparison of hERG blocker

classiﬁers: Assessment on different datasets yields

markedly different results. Mol. Inform. 2011, 30,

443–458.

Matthews, B. W. (1975). Comparison of the predicted and

observed secondary structure of T4 phage lysozyme.

Biochimica et Biophysica Acta (BBA) - Protein Struc-

ture. 405 (2): 442–451.

Moriwaki H, Tian Y-S, Kawashita N, Takagi T (2018). Mor-

dred: a molecular descriptor calculator. Journal of

Cheminformatics 10:4.

N. V. Chawla, K. W. Bowyer, L. O. Hall, W. P. Kegelmeyer.,

2011. SMOTE: Synthetic Minority Over-sampling

Technique. arxiv - 1106.1813

Ogura, K., Sato, T., Yuki, H. et al.,2019. Support Vector

Machine model for hERG inhibitory activities based

on the integrated hERG database using descriptor se-

lection by NSGA-II. Sci Rep 9, 12220 (2019).

Parasuraman, S.,2011. Toxicological screening.” Journal

of pharmacology & pharmacotherapeutics vol. 2,2

(2011): 74-9.

Peruri Venkata Anusha et. al., 2019. Detecting Outliers in

High Dimensional Data Sets Using Z-Score Method-

ology. Int. Journal of Innovative Tech. and Exploring

Engg.

RDKit. Open-source Chemiformatics http://www.rdkit.org.

Schyman, P., Liu, R. & Wallqvist, A.,2016. General

Purpose 2D and 3D Similarity Approach to Identify

hERG Blockers. J. Chem. Inf. Model. 56, 213–222

(2016)

Spiegel, Murray R.; Stephens, Larry J. (2008). Schaum’s

Outlines Statistics (Fourth ed.), McGraw Hill.

Dirk J. Snyders, 1999. Structure and Function of Cardiac

Potassium Channels. Cardiovasc. Res 42, 377–390

(1999)

Wang, S. et al., 2012, ADMET Evaluation in Drug Discov-

ery.12. Development of Binary Classifcation Models

forPrediction of hERG Potassium Channel Blockage.

Mol.Pharmaceutics 9, 996–1010 (2012).

Wang, S. et al.,2016. ADMET Evaluation in Drug Dis-

covery. 16. Predicting hERG Blockers by Combining

Multiple Pharmacophores and Machine Learning Ap-

proaches. Mol. Pharmaceutics 13, 2855–2866 (2016).

Yoon, Hye & Balupuri, Anand & Choi, Kwang-Eun &

Kang, Nam (2020). Small Molecule Inhibitors of

DYRK1A Identiﬁed by Computational and Experi-

mental Approaches. International journal of molec-

ular sciences. 21. 10.3390/ijms21186826.

Yongli Zhang, 2012. Qinggong College, Hebei United

University, Tangshan, China. Support Vector Ma-

chine Classiﬁcation Algorithm and Its Applica-

tion.Information Computing and Applications. ICICA

2012. Communications in Computer and Information

Science, vol 308. Springer, Berlin, Heidelberg.

Virtual Screening of Pharmaceutical Compounds with hERG Inhibitory Activity (Cardiotoxicity) using Ensemble Learning

159