Artificial Intelligence Review (2022) 55:1947–1999

https://doi.org/10.1007/s10462-021-10058-4

Machine Learning in Drug Discovery: A Review

Suresh Dara1 · Swetha Dhamercherla1 · Surender Singh Jadav2 · CH Madhu Babu1 ·

Mohamed Jawed Ahsan3

Published online: 11 August 2021

© The Author(s), under exclusive licence to Springer Nature B.V. 2021

Abstract
This review provides the feasible literature on drug discovery through ML tools and tech-
niques that are enforced in every phase of drug development to accelerate the research pro-
cess and deduce the risk and expenditure in clinical trials. Machine learning techniques
improve the decision-making in pharmaceutical data across various applications like
QSAR analysis, hit discoveries, de novo drug architectures to retrieve accurate outcomes.
Target validation, prognostic biomarkers, digital pathology are considered under problem
statements in this review. ML challenges must be applicable for the main cause of inad-
equacy in interpretability outcomes that may restrict the applications in drug discovery.
In clinical trials, absolute and methodological data must be generated to tackle many puz-
zles in validating ML techniques, improving decision-making, promoting awareness in ML
approaches, and deducing risk failures in drug discovery.

Keywords Artificial intelligence · Drug discovery · Machine learning · Target validation ·

Prognostic biomarkers · Digital pathology

* Suresh Dara
[email protected]
* Surender Singh Jadav
[email protected]
Swetha Dhamercherla
[email protected]
CH Madhu Babu
[email protected]
Mohamed Jawed Ahsan
[email protected]
1
Department of Computer Science and Engineering, B V Raju Institute of Technology,
Narsapur, Medak 502313, Telangana, India
2
Centre for Molecular Cancer Research (CMCR) and Vishnu Institute of Pharmaceutical Education
and Research (VIPER), Narsapur, Medak 502313, Telangana, India
3
Department of Pharmaceutical Chemistry, Maharishi Arvind College of Pharmacy, Jaipur 302023,
Rajasthan, India

13
Vol.:(0123456789)
1948 S. Dara et al.

1 Introduction

In computer science, Artificial intelligence (AI) additionally attributed as machine intelli-

gence because machines are trained or customized to perform activities like a human brain
(Poole et al. 1998; Vinod and Anand 2021; Gopal 2018). Artificial Intelligence (AI) can
be categorized here as the field is dealing with a wide range of utilization and layouts of
numerous algorithms for interpreting and attaining knowledge from data. And the AI con-
cept is firmly related to many fields like pattern recognition, probability theory, statistics,
machine learning, and numerous procedures like fuzzy models, neural networks which are
collectively known as “Computational Intelligence” Vinod and Anand (2021), Engelbrecht
(2007), Konar (2006), Duda et al. (2012), Webb (2003), Friedman et al. (2001). Multiple
complicated usages engaged with AI strategies like classification, regression, predictions
and also optimization techniques. Machine learning needs to be modified well in the uti-
lization of any kind of information i.e., initially, a particular model must be characterized
along with parameters. So, machines can be gain proficiency in the model with accessible
parameters through the utilization of trained data. Furthermore, the model can predict the
data in the future for recovering information from data (Alpaydin 2020).
In this review, we are primarily focusing on qualities of AI approaches that are appro-
priate for drug development and discovery (Duch et al. 2007). Recently various factors
were developed due to greater enthusiasm for utilizing machine learning approaches in
the pharmaceutical industry. Figure 1 shows that the various fields of Drug Discovery and
advancements utilized through machine learning. Every phase was performed like a pipe-
line to represent therapeutic concepts. The respective phases represent unique iterations in
time and cost expenditure. Here each phase is carried out to prove the effectiveness of the
remedial treatment. The medical information was being mined and estimated accurately
by using some ‘omics’ and ‘smart automation tools’. Enlarging these techniques into the
biological field gives more opportunities as well as challenges in the pharmaceutical indus-
try. Since numerous pharmaceutical enterprises’ objective is to distinguish the persuasive
clinical hypothesis. With the obtained results, practitioners or clinicians can develop the

Fig. 1  Various fields in Drug

discovery by using Machine
Learning

13
Machine Learning in Drug Discovery: A Review 1949

medications. For establishing any type of drug in pharmaceutical industries, the usage of
machine learning approaches has checked out the performance. At this point, if included
with unlimited storage, improvement appeared in datasets like size, types can provide
premises to machine learning. In this way, it can access enormous data from pharmaceuti-
cal industries. Data types can have different configurations like textual data, images, assay
information, biometrics, and furthermore high dimensional omics data (Mamoshina et al.
2018).
Thus, the AI field has developed from theoretical knowledge to real-world data. Infor-
mation was widely improved for utilizing in PC hardware, for example, Graphical Process-
ing Units (GPU), which makes faster in processing (i.e., in computational techniques).
Recently, the deep learning model is one of the machine learning algorithms (LeCun et al.
2015), it develops the models for making more accomplishment in broad daylight chal-
lenges (Chen et al. 2018; Hinton 2018). For the past 2 years, the usage of ML algorithms
has a great extension within pharmaceutical enterprises.
In the clinical field, developing a new drug for persistent disease primarily relied on
new medications. As of late, various drugs are improvised for recognizing dynamic com-
ponents from traditional treatments such as penicillin. In chemical laboratories, it consists
of natural substances, small molecules that aid in therapeutic medicine to detect substances
such as cells or intact organisms. This procedure is called old-style pharmacology.
High throughput screening with multiple libraries has normally expanded because of
the human genome has permitted cloning strategies and furthermore improving refining
of proteins in huge quantities. Screening activity for large compounds through biological
targets can be used to achieve a change in a disease called reverse pharmacology. Multiple
hits can be generated from screening activity to provide cells and furthermore tests have
been conducted in creatures for adequacy. In modern days, drug discovery has engaged
with the performance of identifications on screening hits, optimization techniques can
build the drug effectiveness, affinity, stability of metabolic. If all requirements are satis-
fied by the compound, a particular drug will be developed in clinical trials if the drug is
successful. In process of drug development and discovery, it requires lead optimization,
target identification and validation, hit discovery, clinical trials (Vohora and Singh 2018).
In novel drug development, the cost expenditure can approximately 2.558 billion USD
(DiMasi et al. 2016) and it is a tedious procedure in light of the fact that about 10–15 years
have taken for selling in the market (Turner 2010). To accomplish a small number of mol-
ecules in drug development, many investors are putting a lot of cash in developing exact
progress in clinical trials. And still, 13% precision rate is lagging with disappointment.
So as to conquer this issue, clinicians have utilized the Computer-assisted Drug Design
CADD technique (Hassan Baig et al. 2016). By utilizing this strategy in drug discovery,
the artificial techniques not just provide the molecular properties (i.e., selectivity, distri-
bution, absorption, bioactivity, metabolism, side effects, and excretion in the theoretical
levels) but also provides the lead compounds such as ideal attributes in silico. Also, attri-
tion cost in the preclinical state can be decreased through the utilization of multi-objective
optimization techniques.
In drug discovery, computational intelligence provides various techniques for analyz-
ing, learning and furthermore clarifies how such pharmaceutical was identified with AI for
finding numerous medications in a programmed and integrated format (Duch et al. 2007).
Therefore, many pharmaceutical industries have shown greater enthusiasm for contributing
to technologies, resources for retrieving accurate results in drug discovery. At last, this sur-
vey proposes AI techniques in the drug discovery area for targeting multiple applications in
drug discovery and development by utilizing deep learning techniques. Along these lines,

13
1950 S. Dara et al.

the AI field provides expected outcomes in concern of computational intelligence in drug

development and discovery (Table 1).

1.1 Roadmap

The rest of the article is arranged in the following way: Sect. 2 describes the application of
AI in Drug design. Then, the various machine learning methods towards Drug discovery

Table 1  List of Major Abbreviations

ADMET Absorption, Distribution, Metabolism And ExcretionToxicology

AE AutoEncoder
AI Artificial Intellingence
ANN Artificial Neural Networks
AUC​ Area under the ROC Curve
CNN Convolution Neural Networks
CT Computed Tomography
DL Deep Learning
DNN Deep Neural Networks
DPI Drug Protein Interaction
GPCR G-Protein coupled receptors
GPU Graphical Processing Unit
HARF Heterogeneity Aware Random Forest
HTVS High-Throughput Virtual Screening
LDA Linear Discriminant Analysis
LR Logistic Regression
MAE Mean Absolute Error
ML Machine Learning
MLP Multi Layer Perceptron
MLR Multiple Linear Regression
MRI Magnetic Resonance Imaging
NBC Naive Bayesian Classification
NCE New Chemical Entities
PNN Probabilistic Neural Networks
PPI Protein to Protein Interaction
QSAR Quantitative Structure-Activity Relationship
RBN Radial Basis function Network
RF Random Forest
RMSE Root Mean Square Error
RNA Ribonucleic Acid
RNN Recurrent Neural Networks
ROC Receiver Operating Characteristic curve
SARS-CoV-2 Severe Acute Respiratory Syndrome CoronaVirus 2
SMILES Simplified Molecular Input Line Entry Specifications
SVM Support Vector Machines
VAE Variational AutoEncoders

13
Machine Learning in Drug Discovery: A Review 1951

are discussed in Sect. 3. Various Drug design applications are discussed in Sect. 4. In
Sect. 5, different Drug design problems have discussed. Finally, Sect. 6 presents the
research challenges with few possible suggestions in Drug discovery using Machine learn-
ing, and Sect. 7 concludes the article and provided some future directions.

2 Application of AI in drug design

This section discusses a few applications in AI which relate to drug study. The activity
of protein structure is considered as the application in drug design. Many impurities have
appeared in the human body due to protein dysfunctions. Structural drug design strategies
are used to differentiate small molecules in protein targets. Protein structure in 3D format
requires more money and time for predicting the 3D structure. And still, it faces the prob-
lem i.e., in making more exactness over de-novo prediction in 3D structure. By using deep
learning and feature extraction tools, it is mandatory to predict the secondary structure
(Spencer et al. 2014) and residing the protein contacts (Li et al. 2017). It precisely gains the
information on the connection among structure and sequence from feature extraction. The
further goal is to predict the 3D- protein structure by utilizing deep learning techniques for
improving the accuracy. To retrieve information from drug design of protein-protein com-
puter structure, then it is mandatory to conduct investigations on PPI interface (Xue et al.
2015).
Artificial Intelligence has been used in various applications like a prediction on
drug–protein interactions, the discovery of drug efficacy, ensuring the safety biomarkers.
The detailed discussion is given as follows

2.1 Prediction on drug–protein interactions

The crucial step of drug development in silico is consisting of multiple biological sources
for predicting drug–protein interactions. Here complications can be seen in large predic-
tions, which relied on the countless unknown interactions. Therefore, semi-supervised
training techniques should be used to address these unlabelled and labeled date complica-
tions. Usually, only labeled data will produce better results. In addition, the semi-super-
vised technology integrates chemical structure, drug–protein interaction network data, and
genome sequence data. Finally, in this article, drug–protein interactions of various data sets
such as ions, enzymes, and nuclear receptors provided well predictable results (Xia et al.
2010).
Drugs have an important priority in therapeutic activity, which is regulated by protein
interactions. The drug–protein interaction database (DPI) focuses primarily on therapeu-
tic protein targets, while knowledge of non-targets has been limited and resolved. Thus,
computational techniques can fill the knowledge gap for predicting protein targets for dis-
tributed drug molecules. In that study, the pool of 35 predictors had a major impact on
the similarity between protein and drug targets. Drug structure, target sequence, and drug
profile are three types of similarity developed from the results of 35 predictors. Finally, the
significant content, relationships, and implications between database sources are of great
importance for therapeutic activity (Wang and Kurgan 2020).
In drug repurposing, the unexpected detection of drug–protein interactions is essential.
Thus, the dominant drug may be useful for repurposing, while drug side effects are una-
voidable and about 1,000 human proteins can cause critical side effects. The proteomic

13
1952 S. Dara et al.

scale method was used to predict side effects and protein goals. FINDSITEcomb is used to
predict drug–protein interactions. The estimates showed greater disruption with a mean of
329 human targets for each drug (Zhou et al. 2015).

2.2 Discover of drug efficacy

Usually, a drug effect assessment looks at its biochemical activity. The effectiveness of
the therapeutic activity has posed a challenge to be properly coupled with the biochemical
activity. The collection of a large amount of data on the effects of cellular drugs was under-
taken to fill a gap that has been explored in the extensive content of cellular estimations
and while this estimation is classified as a psychotropic drug. Here, the microarray data can
be analyzed by applying random trees to the forest and classifying them, providing a profile
for the efficiency of biomarker gene expression. Accuracy of 88.9% of the classification
tree and 83.3% of the random forest model used this efficacy profile for a drug treatment
analysis. Therefore, at the cellular assessment level, general genomic data are acceptable to
reconcile the effects of new physiological drugs with clinical applications. Finally, in vitro
signatures of gene expression data can identify the effectiveness of therapeutic activities
that can help validate targeting and drug development (Gunther et al. 2003).
In drug development, increasing profitability by validating new drugs requires predict-
ing effectiveness and identifying targets. The proximity of medical illnesses helps to reduce
the effectiveness of the treatment and also releases drugs that are effective in therapeutic
activity. The study treated 78 diseases with 238 drugs to demonstrate the drug’s effective-
ness in therapeutic activity, as well as problems with gene efficacy and various disorders.
Here the network-based system is used to develop a drug-disease proximate measure that
assesses the interactions between the disease and the drug target. Therefore, the proximity
of network-based systems makes it possible to predict associations for novel drug diseases,
offering a wide range of possibilities for conflict detection and drug repurposing (Guney
et al. 2016).

2.3 Ensuring the safety biomarkers

In drug development, the use of biomarkers supports the provision of safety measures that
critically determine the biological and analytical indicators of a particular biomarker. In
this way, stakeholders can assess and manage whether claims are defended for a particular
purpose and whether the desired standards are being met. For shareholders in the imple-
mentation of evaluating the experiment agreement, a stakeholder evaluation process is
needed to adjust the unique characteristics of the biomarkers, as well as to determine how
these innovations are analyzed, integrated, and interpreted, and how improved biomarkers
and conventional comparators are measured (Sistare et al. 2010).
In the survey, we found that modern medicines are no safer than older drugs, even
though with longer medical trial programs. These trails are placed on the market and
impractical inspections are carried out which are not sufficient to be carried out system-
atically to ensure safety. Previous drug-related signals can help in improving drug safety
as well as identify underlying biomarkers, making them more toxic. However, the safety
markers can be different for different target systems. However, no other approach can pro-
vide assurance that medicines are very safe, but we can develop a common understanding
of benefit and risk assessment by communicating with the public (Rolan et al. 2007).

13
Machine Learning in Drug Discovery: A Review 1953

Various deep learning techniques are carried out here to predict the PPI interface and
show fantastic results when contrasted with the SVM technique (Du et al. 2016). Thus, the
PPI’s became more complex to utilize in biological techniques (Falchi et al. 2014; Scott
et al. 2016). Each PPI can be a mixture of various residues (Cukuroglu et al. 2014). New
PPI can act as a modern class for pharmaceutical targets where disparate for different tar-
gets i.e., ion channels, GPCRs (G-Protein coupled receptors), kinases (Higueruelo et al.
2013; Santos et al. 2017). iFitDock is a docking tool used for investigating a few hotspots
in PPIs. Further, AI techniques have been utilized for distinguishing structures and hotspots
in PPI interface (Fig. 2).

3 Machine learning methods to drug discovery

AI innovation has a high priority in drug design through the enhancement of ML

approaches and the collection of pharmacological data. AI does not rely upon any hypo-
thetical improvements, but it has more essence in transforming medical information into
studies like reusable methods. In general, there are different approaches such as Random
Forest, Naive Bayesian Classification (NBC), Multiple Linear Regression (MLR), Logis-
tic Regression (LR), Linear Discriminant Analysis (LDA), Probabilistic Neural Networks
(PNN), Multi-Layer Perceptron (MLP), Support Vector Machine (SVM), etc are consid-
ered in the context of ML (Lavecchia and Di Giovanni 2013). In order to gain capability
in feature extraction and feature generalization, AI advancements are specifically used as
a deep learning technique towards drug design. Also, Fig. 3 shows respective applications
which illustrate an outline of AI procedures utilized to respond to drug discovery queries
in the review. A scope of classifier and regression strategies i.e. supervised learning tech-
niques utilized to respond addresses desire expectations in continuous or categorical data
factors, also unsupervised methods utilized in creating a model which empowers the clus-
tering data.

Fig. 2  Applications of AI in Drug discovery depicts the Machine learning mechanisms

13
1954 S. Dara et al.

Fig. 3  Maximum-margin hyper-

plane and margins for an SVM
trained with samples from two
classes. Samples on the margin
are called the support vectors

Many designed features in traditional ML models are performed manually, but deep
learning approaches will accelerate various features through available initialized data
automatically because multi-layer feature extraction techniques are used to convert
straightforward features into complex features. One advantage of using deep learning
approaches was, presence of low quantity generalization blunders, so it recovers more
exact results. CNN, RNN, Auto Encoder, DNN, and RBN are considered as differ-
ent deep learning techniques. Summary of deep learning algorithms can be identified
(LeCun et al. 2015; Angermueller et al. 2016; Schmidhuber 2015) and provides detailed
information about deep learning techniques which are available in Deep Learning litera-
ture (Goodfellow et al. 2016).
In drug discovery and development, many AI calculations are associated to analyse
and predict the data. Here, few popular models like SVM, RF, and MLP discuss their
effective use in drug discovery.

3.1 Support vector machines

SVM model is a supervised learning algorithm basically utilized in predicting the class
labelled data i.e., binary data. In SVM, x is considered as feature vector i.e., input to
SVM model. At that point, x ∈ Rn where n is a dimension feature vector. Y acts as a
class i.e., output for svm. Y ∈ {−1, 1}. Here, Binary values are considered as classifica-
tion task. Parameters in SVM u and b have considered for learning data in training set.
In dataset, (x(i) , Y (i) ) are considered as ith sample. Y can be represented as follows:
{
if − 1 if UTX(i) + b ⩽ −1
Y (i) =
if 1 if UTX(i) + b ⩾ 1

A class Y can be written as Y (i) )(UTX(i) + b) ⩾ 1. Finally, SVM algorithm goal is to satisfy:

1. In SVM, seperation between any two boundaries ought to be augmented i.e., the distance
between two hyperplane uT x + b = −1 and uT x + b = 1 should be maximized. In this
way, Distance = ||U||2
maxU . Finally, it have to solve maxU ||U||
2 2
minU ||U||
2. Complete x(i) samples need to classify effectively in the SVM i.e.,
Y (i) )(UTX(i) + b) ⩾ 1∀ ∈ 1, 2, 3 … N

Then, it produces quadratic optimization problem i.e. min||U||

U,b2
. So that,
Y i (U T X i + b) ⩾ 1, ∀i ∈ 1, 2, 3, … , N .

13
Machine Learning in Drug Discovery: A Review 1955

The above equation was a hard-margin SVM, and we can avoid this problem through
applying linearly separable method. Using the slack variable 𝜖 (i) as constraints. In training
data, each sample has its own slack variable. Then,
N
min||U|| ∑
+C 𝜖 (i)
U, b 2 i=1
(1)
i.e., Y (i) (U T X (i) + b) ⩾ 1 − 𝜖 (i) , ∀i ∈ 1, 2, 3, … , N
𝜖 (i) ⩾ 0, ∀i ∈ 1, 2, 3, … , N
Now, it’s a soft margin SVM, where ‘C’ is considered as a penalty of the error term.
Involving function 𝜙 to allow more flexibility in mapping. So, it maps multiple features like
original space to high dimensional space (Noble 2006). Then, the quadratic optimization
problem updates Eq. 1 as the following:

Therefore, Y (i) (U T 𝜙X (i) + b) ⩾ 1 − 𝜖 (i) , ∀i ∈ 1, 2, 3, … , N

The SVM widely used in drug discovery using its various kernels (Smola and Schölkopf
2004). Various problems like Screen radiation protection and Gene Interaction using SVM-
RBF(Radial Basis Function) (Matsumoto et al. 2016; Guo et al. 2008), Assess target-ligand
interactions using Regression-SVM (Li et al. 2011), Identify drug target interaction by
Biased SVM (Wang et al. 2017), Predicting drug sensitivity prediction by Ensemble SVM,
and the Linear SVM used in Identify novel drug targets (Volkamer et al. 2012), Anti/non-
anticancer molecule classification (Kapoorb et al. 2020), Kinase mutaion activation (Patil
et al. 2021).
The SVM approach (Huang et al. 2018) was used to quantify anti-cancer drugs based
on cancer cell properties. To understand the relationship between cancer cell properties
and drug resistance, 24 drugs were tested on cancer cell lines (Gupta et al. 2016). In the
treatment of oral cancer, the SVM-RBF (Radial Basis Function) approach has been used
to find therapeutic compounds from a large collection of public databases (Bundela et al.
2015), the RBF is the popular kernel function used in various learning algorithms. The
RBF kernel takes two samples2 S1 and S2, represented as feature vectors in some input
space K(S1, S2) = exp( ||S1−S2||
2𝜎 2
) where ||S1 − S2||2 is used to recognized as the squared
Euclidean distance between two vectors and 𝜎 is a free parameter. Here the RBF is used
and hybridized as many variations with different parameter values.
In general, radiation therapy techniques help to protect against cancer. Therefore, the
SVM method is used in virtual screening (Matsumoto et al. 2016) to protect the radiation
function. Radiation therapy also has side effects on normal cells and tissues (Morita et al.
2014). In this study, we found that the SVM approach worked better than other techniques.
When the target protein is known, we can find a suitable compound for the target pro-
tein. However, the SVM technique is mainly used to predict the outcome of targeted drugs.
SVM has used sites to link global descriptors, taking into account various properties such
as compactness and size. These descriptors can determine drug scores for novel targets
(Volkamer et al. 2012; Li et al. 2011).
In therapeutic activities, the use of SVM helps to find the active ingredient at various
stages of the drug development process. In general, the active component of the connec-
tion is taken into account in the number of turns of the design process. The main goal is
to find different lead series in the active compound to improve them in parallel in thera-
peutic activity (Warmuth et al. 2003). In contrast to other artificial neural networks,

13
1956 S. Dara et al.

SVM demonstrated the ability to test drug similarity predictions of a wide variety of
compounds. Because of this set of descriptors, the SVM outperformed the task and also
reported that the SVM model predicted better enzyme inhibitor quality for conventional
QSAR (Zernov et al. 2003).
Right now, the SVM model is the best methodology for predicting organic and
compound properties. Recently, the SVM model has been utilized in the drug discov-
ery region and turned out to be more famous in drug discovery applications like a pre-
diction on properties, compound classification (Maltarollo et al. 2019). In designing
new structures, the SVM approach was utilized for retrieving higher predicted results
where depend on ligands (Hartenfeller and Schneider 2010). In the Activity process,
to improve scoring capacity execution, the SVM approach was utilized for clarifying
non-linear relationships of energy terms from eHiTS and binding data which shows a
lot of improvement in scoring power and screening power (Kinnings et al. 2011; Zsol-
dos et al. 2007). SVM model was frequently utilized in virtual screening (Leelananda
and Lindert 2016; Liew et al. 2009; Melville et al. 2009) and demonstrated best results
(in the predicted ratio called hits) and furthermore false-hit rates are decreased concur-
rently (counterfeit hit rates in the predicted hits) (Ma et al. 2009). Creating meta-clas-
sifiers with SVM-based methodology can coordinate different methods for exploiting
each complementarity and individual strengths (Maltarollo et al. 2019).

3.2 Random forest

The Random Forest algorithm was a supervised algorithm. The name itself says, ”This
is a way of creating a forest from various perspectives to make it random”. The signifi-
cant advantage of the Random Forest algorithm was, it can relevant for both regression
and classification issues. In the procedure of regression and classification tasks, overfit-
ting can happen normally, so the outcome will be in a worse state. We can defeat the
overfitting issue through the usage of random forests algorithm with the availability of
multiple trees in the forest. Random forests can apply trained algorithmic techniques
i.e., bagging. Training set comprises,
X = X1 , X2 , … , Xn , Y = Y1 , Y2 , … , Yn . Then, random samples can alternately selected
from training data for fitting random forest tree.

1. Alternate samples with n trained examples from X, Y then Xa , Ya.

2. Classification tree fb must be trained on Xa , Ya data. Here, a = 12, … , A.

After training the data, invisible samples x′ need to be predicted by averaging all indi-
vidual trees on x′:
A
1∑
f̂ = f (x� )
A a=1 a

In classification trees, majority voting can be considered. Finally, random forest model
produces better results due to the absence of increment in bias, it reduces variance in the
model. The equation for individual regression tree on x′ can be represented in standard
deviation form i.e.,

13
Machine Learning in Drug Discovery: A Review 1957

�
∑A
a=1 (fa (x
�) − f̂ )2
𝜎=
A−1
where ‘A’ is a free parameter. In view of the size, nature of the trained data, a large number
of trees can be used (Ho 1995). Also, the random forest can be appropriate in medication
for deciding the right segments of grouping in therapy, and; investigating patient records
can be supportive in recognizing the infections (Polamuri 2017). In ligand-protein bind-
ing affinity, using random forests can improve the scoring function performance (Kinnings
et al. 2011; Zsoldos et al. 2007). Representation of scientific models and chemical struc-
tures are the fundamental issues in QSAR model (Dudek et al. 2006). At that point when
descriptors are chosen, it is necessary to establish the best mathematical model for correct
fitting in structure-activity correlation. So as to improve fitting standards in mathematical
model (A Dobchev et al. 2014; Ning and Karypis 2011), a random forest algorithm was
utilized (Fig. 4).
The selection of molecular descriptors is seen as an important step in virtual screening
to identify bioactive molecules during the drug development process. Because this choice
of descriptors shows predictions with lower accuracy. Hence, the random forest technique
was used to improve prediction and then select naturally trained molecular descriptors for
kinase ligands, hormone receptors, enzymes, etc. (Cano et al. 2017).
In the pharmaceutical industry, when developing drugs, the question that arises natu-
rally is whether a prediction model trained with heterogeneous data is implemented as a
similar prediction model. Then the heterogeneity data were compiled for forecasting and
model training. In this study, heterogeneity was treated as a problem with the latent dis-
tribution, and the covariate-free allocation technique was distributed to be distributed
by means of an ensemble leaf node model. In general, an ensemble-based random forest
model has incorporated Heterogeneity Aware Random Forest (HARF) and assign specific
weights to tree-based categories. Of course, the technique proposed by HARF gives better
results than classical random forest, whereas drug feedback with the cancer disease types is
something peculiar (Rahman et al. 2017).
Immune network technology is to determine new compounds from drug molecules.
Using examples of sulfonamide properties, sulfonamides are divided into various prognos-
tic effects over a period of time. Using a random forest approach, we selected molecular

Fig. 4  The random forest visually generated a data point decision tree to extract estimations for each sample
to determine the best outcomes through voting

13
1958 S. Dara et al.

descriptors to achieve better accuracy than the simulation results for compounds designed
for the drug (Samigulina and Zarina 2017).

3.3 Multilayer perception

The Multilayer perception model is also known as a feed-forward neural network. MLP
provides an outcome based on a set of input sources. For training any sort of information,
the backpropagation approach is utilized. This model is similar to a directed graph because
of the essence of multiple layers as input nodes and output nodes are associated with some
weights (Pal et al. 1992). After processing the data, the perceptron can fluctuate each con-
nected weight in the network. In this way, the presence of error in actual output can be
compared with the expected output. Consider node }j� in output as degree of error in last
data point i.e., nth ej (n) = aj (n) − Yj (n) Where a → targetvalue, Y → the variable developed
from the perception. Based on some corrections, weights in each node can be adjusted
through decreasing error in the output i.e.,
1∑ 2
𝜀(n) = e (n)
2 j j

Also, every weight can be varied through the gradient descent approach i.e.,
𝜕𝜀(n))
𝛥Wij (n) = −𝜂 Y (n)
𝜕Vj(n)) i

where, 𝜂 → learning rate and weights can be converted into a response without any oscilla-
tions. Yi → previous neuron result.
Depending on }Vj� field, derivative can be calculated. Then, simplified derivative in out-
put node can be
𝜕𝜀(n))
− = ej (n)𝜙I (Vj (n))
𝜕Vj(n))

Here 𝜙 cannot be varied itself. Because changing each and every weight in hidden layer
becomes difficult; Also, it provides
𝜕𝜀(n)) ∑ 𝜕𝜀(n))
− = 𝜙I (Vj (n)) − W (n)
𝜕Vj(n)) k
𝜕Vj(n)) ij

where }k� is represented as the last node in the output layer. In case, changing any weights
in a hidden layer, the activation function can be varied the weights in the output layer.
Figure 5 performs specific computations to distinguish few features in input data. It learns

Fig. 5  Multilayer Perception

Architecture

13
Machine Learning in Drug Discovery: A Review 1959

optimal weights consequently and afterward input features will be increased with available
weights to decide specific neuron was terminated or not. In this way, Multilayer percep-
tron uses backpropagation strategy with the activation function (Rosenblatt 1961). In this
review, a multi-layer perceptron was utilized for predicting action between the drugs. This
model has one advantage i.e., it does not require any structural information on compounds
because of the fact that it uses experimental data for predicting the accuracy (Stokes et al.
2020). Additionally, MLP was utilized to generate a de-novo drug design. This model
having the capability to generate different compounds automatically with some advanced
properties (Gómez-Bombarelli et al. 2018).
In general, MLP can be used very easily and very quickly, but fulfilling its duties in
training is very difficult, and MLP also does not offer any guarantee of global minimum
performance (Gertrudes et al. 2012).
The secondary structure of proteins offers a greater advantage in determining protein
function, drug design. In that study, the MLP approach showed greater interest in classifi-
cation success. However, in the experimental area, determining the secondary structure is
more difficult and expensive. Finally, the results from the trained data were reported as a
positive success compared to the classification (Yavuz et al. 2018).

3.4 Deep learning

Deep learning is a part of machine learning, having the capability to extract a greater level
of features through utilization of multiple layers from input data (Deng and Dong 2014).
Deep learning is an immense field that is creating massive premiums nowadays. Recently,
deep learning techniques have been used in many research fields and have achieved higher
profitability in business strikes. But what exactly is deep learning? In general, deep learn-
ing is the same neural network architecture that consists of several layers, and data can be
transformed between these layers. It’s still a significant popular expression, but the innova-
tion behind it is genuine and very refined. So, models in deep learning can be developed
through a strategy called greedy layer-by-layer (Bengio et al. 2007). Figure 6 contrasts the
powerful deep learning approaches with pooling layers and figure outs the critical issues
and devise the most appropriate solution even problem was in a complex situation. In this
review, deep learning algorithms have presented numerous models like DNN, CNN, RNN,
Autoencoder in drug discovery areas. The pooling layer is another structure that hinders
the neural networks. The capacity of the pooling layer is to reduce the spatial size of the
representation to reduce boundary measurement and system computations and work inde-
pendently on each feature map (channel). The motivation behind why max-pooling layers
work so well in various networks is that it enables the system to recognize the features very
effectively after down-testing an input structure and it reduces the over-fitting.
DNN architecture was evolved from an extension of Artificial Neural Network (ANN),
contains multiple layers between input and output nodes (Bengio 2009). The DNN archi-
tecture traces the outcomes in a mathematical model either it can be a non-linear or linear
relationship. Here, each mathematical model expected as a layer, also multiple layers were
available in complex DNN, so that network is named as ‘deep’. Deep learning models are
introduced in QSAR modeling to retrieve feature extractions and capabilities in chemical
characters automatically. Dahl et.al had inspired by Kaggle’s results and improved investi-
gations on multi-task DNN. The results of multi-task DNN have demonstrated incredible
execution in learning general features of sharing parameters (Dahl et al. 2014).

13
1960 S. Dara et al.

Fig. 6  Deep Learning Architectures

Development of candidate drugs plays major desirable property in oral delivery. Mol-
ecules in intestinal permeability can be assessed by computational technology through
affording rapid and reasonable ways. Multiple studies focused on intestinal intake of
chemical composites for predicting the peptide sequence data. ML techniques like artifi-
cial neural networks have been adopted for predicting the intestinal permeabilities of pep-
tides. The intestinal permeable of peptides consists of positive controlled data obtained
through the peroral phage technique and random sequence data can be prepared through
negative controlled data. Multiple statistical indicators like specificity, sensitivity, ROC
score, enrichment curves, etc., are validated to produce appropriate predictions. And the
statistical results declared that models have good quality and can segregate in between ran-
dom sequences and permeable with great levels of confidence. Finally, the ANN models
demonstrated greater prediction than unpredictable one. So, this model can applicable for
intestinal permeable peptide selection to generate peptidomimetics (Jung et al. 2007).
Multi-task neural networks integrated into a platform called ‘DeepChem’, it helps the
multi-task neural network to perform in drug development process (Ramsundar et al.
2017). Along with this, networks have assessed performance in the multi-task deep net-
works was robust. Finally, the performance of deep learning algorithms in QSAR models
upgraded the prediction performance. Also, DNN played out a significant role in further
research of hit-to-hit lead optimization.
CNN is a subclass of DNN, ordinarily utilized for analyzing the visual images (Val-
ueva et al. 2020). CNN also called shift-invariant ANN because frequently rely on weights.
CNN is a regularized version of a multi-layer perceptron. The concept of multi-layer per-
ceptron characterizes fully connected networks, where each neuron in the first layer is asso-
ciated with the following layer. By using of a fully connected algorithm, a network can
conquer the overfitting problem. The CNN algorithm examines the clinical field so that,
every neuron in a human cell appears like the visual cortex (Venkatesan and Li 2017). In
ligand-protein interaction, many researchers utilized CNN model for predicting affinity in

13
Machine Learning in Drug Discovery: A Review 1961

protein-ligand (LeCun et al. 2015; Leelananda and Lindert 2016). The affinity prediction
indicated the best correlation in the dataset (Jiménez et al. 2018). In protein-ligand interac-
tion, the CNN algorithm predicted binding affinities which can further increase scoring
function but predictive capabilities must upgrade simultaneously.
RNN algorithm is an area of the artificial neural network, connections can occur
between the input node and the output node. In this way, a directed graph can be created
in the network along with a temporal sequence. Likewise, the RNN network utilizes the
internal memory to perform grouping in input variables (Dupond 2019). It also exhibits
dynamic performance Miljanovic (2012) because the RNN algorithm struggled for two
networks at a time with the general structure. Each network may contain various impulses
i.e., finite and infinite impulses.
Determining the functionality of protein structure will play a vital role in secondary
and tertiary structures. Previously, numerous algorithms relates to folding prediction have
improved to encode in the protein sequence experiment to develop protein structures. So,
Visibelli has found 𝛼 − helixes signals on a large dataset. To locate specific occurrences
in amino acids to characterize the specifications in secondary structure for deciding the
helical moieties boundaries. The 𝛼 − helixes occurrences are predicted through various ML
models for validating the hypothesis equipped with an attention mechanism. This mecha-
nism can interpret the weights of each input, model’s decision for prediction. At last, the
similar subsequences show the experimental outcomes, where input code-driven in second-
ary structure information (Visibelli et al. 2020).
Day by day, it has been turning out to be a challenge in improving affordable and effec-
tive treatments to humans without any prescience in drug target information. The deeDT-
net is one of the deep learning techniques that were embedded with 15 variations of phe-
notypic, chemicals, cellular profiles, genomics utilized to accelerate drug repurposing and
target identification. Due to the presence of high accuracy, deepDTnet has been approved
by U.S. Food and Drug Administration with the identification of novel targets to familiar
drugs. Through experimental results, topotecan was an approved inhibitor that can directly
be utilized for human retinoic-acid receptors to diminish transitional void in drug develop-
ment (Zeng et al. 2020).
In virtual screening, RNN utilized to cause new molecular libraries, so it got support-
ive in finding anticancer agents through molecular fingerprints (Kadurin et al. 2017). In
producing the de novo drug design, the prediction must be conducted on biological perfor-
mance. In this way, the RNN algorithm was utilized for generating molecules (Olivecrona
et al. 2017). In the ChEMBL dataset, molecules could be gathered. For sampling, gen-
erated molecules must be trained by the RNN algorithm through conditional probability.
Various classifiers performed data sampling however RNN with reinforcement learning has
given 95% accuracy in scoring function (Mnih et al. 2015).
’Deep Interact’ was an integrative domain-based approach is utilized to predict PPI’s
through Deep Neural Network. Assortment of multiple PPIs is extended out from (KUPS)
Kansas University Proteomics Service and (DIP) Database of Interacting Proteins. It’s
highly fundamental to discover and analyze the cellular components in the specificity of
interactions and explicit molecular protein complexes. The significant goal is to develop
enormous scope high-throughput experiments through silico approach to improve the
uncovering levels in PPI. From a dataset known as Saccharomyces cervisiae, 34,100 PPIs
have been validated to return promising results with a sensitivity of 86.85%, an accuracy
of 98.31%, a specificity of 98.51%, and an accuracy of 92.67%. At last, the Deep Interact
approach concluded to be better performed over existing ML approaches in PPI prediction
(Patel et al. 2017).

13
1962 S. Dara et al.

Autoencoder is a class of artificial neural network, it retrieves information through

unsupervised learning (Kramer 1991). Autoencoder objective is to represent the encod-
ing data format in dimensionality reduction for maintaining a strategic distance from the
‘noise’ signal in the network. Along with this, the autoencoder must explore input data and
then copied to the output layer. Autoencoder has two areas i.e., Encoder and Decoder; and
one hidden layer. Here, the hidden layer is considered as code. Encoder transfers input data
to the hidden layer. The decoder can retrieve information for reproducing the signal output.
Autoencoders was most appropriate in dimensionality reduction and learning the data from
generative models (Kingma and Welling 2013; Larsen et al. 2015).
Considering encoder as 𝜙 and decoder as 𝜓 , such that 𝜙 ∶ Y → E , 𝜓 ∶ E → Y

𝛷, 𝜓 ∶ arg𝛷,𝜓
min
||Y − (𝛷o𝜓)Y||2

In first hidden layer, encoder considered input as y ∈ Rd = Y and maps to h ∈ Rp = E .

h = 𝜎(Wy + b) Here, }h� considered as code, }W � as weight matrix, }b� as bias vector, 𝜎
acts as activation function. Basically, biases and weights are randomly utilized and updated
through backpropagation technique. Then, decoder maps }h� to }y� with same structure of
}y� ∶, Y � = 𝜎 � (W � h + b� )
Decoder consists 𝜎 ′ , W ′ , b′ coefficients may vary in encoder i.e., 𝜎, W, b coefficients.
Mainly autoencoders were trained to decrease reconstruction errors (loss).

L(y, y� ) = ||y − y� ||2 = ||y − 𝜎 � (W � (𝜎(Wy + b)) + b� ||2

Here, feature space }E� consists of less dimensionality than input space }Y �. Also 𝜙(y) is a
compressed format for input ‘y’. At whatever point, hidden layers are more prominent than
or equivalent to the input layer, it offers the adequate capability to learn identity function,
finally, it was useless. In Autoencoders, test results despite everything to learn numerous
valuable features from training set (Kingma and Welling 2019). In drug discovery, autoen-
coders utilized as unique architecture to deliver molecules through conducting experiments
right into vermin (Zhavoronkov et al. 2019). In designing of de-novo drug design, deep
learning model i.e., autoencoder have utilized for generating the molecules. So, the autoen-
coder approach was employed with various classifiers like multilayer perceptron for gener-
ating new compounds automatically with appropriate properties (Gómez-Bombarelli et al.
2018). In many situations, the drug produces invalid SMILES syntax, so as to defeat this
issue, grammar variational autoencoder was utilized for developing SMILES syntax with
more effectiveness (Pu et al. 2017) (Fig. 7).

4 Drug design applications

The review of drug discovery is further categorized on the basis of task performing of ML
and their applications like target identification, hit discovery, hit to lead, lead optimization
techniques are discussed out. The drug design techniques rely on the databases which are
inturn developed based on the different ML algorithms. The precise training, validation,
and application of ML algorithms in the drug discovery era provide an enthusiastic out-
come by easing the complicated error-prone protocols. The ML techniques are introduced
in most of the drug design processes to reduce the time as well as manual interference. The
best example is QSAR, in which the huge data collection and training of datasets are con-
sidered as rate-limiting steps in defining the ligand-based virtual screening protocols and

13
Machine Learning in Drug Discovery: A Review 1963

Fig. 7  Basic flowchart of an AutoEncoder with an example NCE

are now replaced by Denovo design techniques. The relationship between drug discovery
steps and algorithms is presented in Fig 8.

4.1 Homology modeling/prediction of protein folding

The folding of secondary structure like 𝛽 − sheets and 𝛼 − helices, which is formed by
the interaction of side-chain amino acid residues are very critical to regulating the smooth
functioning of three-dimensional proteins. An accurate protein folding along with its pre-
rogative active ligand site can be experimentally obtained by X-ray crystallography, NMR-
spectroscopy, and Cryogenic electron microscopic techniques (Cryo-EM).

Fig. 8  Primary, secondary, tertiary and quaternary structures of the protein highlighted with active site resi-
dues. The AmpC beta-lactamase (PDB:6DPZ) as case example is taken and depicted in the above figures

13
1964 S. Dara et al.

Information about the primary amino acid sequences of proteins/enzymes/receptors,

both dissolved / insoluble, is stored on the UNIPROT server along with their targets and
cellular functions. Based on medicinal chemistry or pharmacological or biochemical stud-
ies, the main role of proteins is identified, and this information is also the basic unit for
developing the protein folding prediction studies by software or experimental studies.
Whereas, the protein folding predictions in the provided aminoacid (UNIPROT) sequence
were compared with its experimentally derived PDB homologues which became a hopeful
technique to refine the new protein models computationally and is also termed as ”homol-
ogy modeling”. The homology modeling or comparative modeling is analyzed by the sev-
eral algorithms which need to be implemented in either software modules (PRIME) or web
servers (EXPASY, SWISS-MODEL) will definitely make a decision to predict the second-
ary structure folding with high accuracy within provided templates. However, the fine-tun-
ing for the obtained homology models or template-based models are again scrutinized by
Ramachandran analysis which can be sorted out by commercial modules (PRIME) or web
servers (QMEAN, PROCHEK). For further understanding, the homology of CHIKV nsP2
protease is described here (Fig. 9) which is obtained based on experimentally predicted
VEEV nsP2 protease template by using insilico techniques. The inisilico tool utilizes the
computational databases to dig the information about the homology templates and pro-
vides the best closest match as considering for more practical bioinformatics and medici-
nal chemistry applications. Figure 8 depicted the alignment of secondary structures such
as 𝛼-helices, 𝛽-pleated sheets, and loop representations present in tertiary complexes. The
surface view also useful for recognizing the hotspots present on the protein to bind with
incoming ligands/substrates. The sequence alignment mode also shows the mutations or
differences in their primary sequences, it can be employed in different chemo-informatics

Fig. 9  a Overlap of 3TRK with 2HWK; b surface view of 2HWK; c, d off-surface/ribbon diagram of finest
3TRK model; and e homology validation parameter obtained from SWISS-MODEL

13
Machine Learning in Drug Discovery: A Review 1965

approaches to identify the mutations similar kinds of viruses or any other pathogenic dis-
orders. The significance of chemo-informatics is playing a crucial role and prevailing as
an emerging tool in the current SARS-COV2 pandemic towards the identification of new
drug-like molecules (Fig. 10).
In addition, selecting the best homologous model obtained from the above process is
another major task that can be performed with SVQMA (Support-vector-machine Protein
single-model Quality Assessment ) servers or ProQ3 or ERRAT, which are operated by the
Deep-learning methods. After going through the above steps, the best 3D protein template
can be used for any basic drug chemistry study to identify hits that are part of a structure-
based virtual screening protocol.
To provide insight for homology modeling, the Q5XXP4 fasta sequence belongs to
CHIKV nsP2 protease domain has been employed as a template by overlapping its closest
VEEV nsP2 protease solved protein (PDB:2HWK) as reference model using the SWISS-
MODEL web server and the results are presented in Fig. 11. for understanding the above-
specified concepts. Further, the active site residue position analysis of the finest developed
model has been done and is found to have similar to VEEV nsP2 protease residues as
shown in Fig. 10. The SWISS-MODEL also provides the information about percentage
similarity along with structure alignment, the Fig. 10 shown the overlap of similar active
site residues consists of catalytic site (catalytic diad Cys and His). It also represents the
conformational changes present in the new template which also considered as an essential
parameter for drug interaction studies

Fig. 10  The overlap of active site residue of the CHIKV (homology model) (red sticks) and VEEV nsP2
protease (green sticks)

13
1966 S. Dara et al.

Fig. 11  Illustrating drug discovery design techniques and topics with AI models

4.2 Target identification

The target identification for NCE’s is an extreme task due to lack of knowledge on their
off-targets such as enzymes, ion channels, proteins, or receptors. The binding site recogni-
tion for the NCE’s is another key task for computational/bioinformatics experiments where
more than one active site has existed in the protein. In the above cases, the predefined most
popular web servers (FTMap), as well as specific modules such as ”Sitemap” developed
with the help of algorithms, can define the preferential binding site to speed up the drug
discovery process. A few other online programs like GHECOM, POCASA, Pocketome,
SURFNET, ConCavity, LIGSITE, Q-SiteFinder, Fpocket, and PASS predicts the feasible
binding sites located within the provided protein templates. Whereas, the metaPocket 2.0
program utilizes the above platforms to afford the most reliable ligand binding sites present
on templates. Further, AI models like FD/DCA can also predict the druggable sites in the
provided biological macromolecules. Recently, the DeepDTnet as a new target identifier
in drug repurposing has been tested. The DeepDTnet strategy is developed by amalgamat-
ing the multi-disease cellular targets, pathogenic genes (genomics), and drugs (chemical
spaces) being utilized for their treatment.

4.2.1 Prediction of protein folding

Patients who experienced illnesses can be recognized through protein dysfunctions. Here,
active molecules can recognize through a structure-based drug design approach. Time and
cost consumption should be required for 3D structural processing, and it is also important

13
Machine Learning in Drug Discovery: A Review 1967

to be aware of what algorithms are used to predict the 3D structure of proteins. Because
of the essence of the large amount of protein sequence data, it creates a problematic issue
in making 3D structure accuracy for de-novo prediction. For retrieving feature extraction
capabilities, deep learning approaches must apply prediction in backbone torsion angle (Li
et al. 2017), secondary structure (Spencer et al. 2014), and protein residue contacts (Wang
et al. 2017). At long last, the goal was to predict the 3D protein structure. Also, deep learn-
ing techniques have elaborated this field for improving 3D protein structure.

4.2.2 Prediction of protein–protein structure

PPI’s are essential for biological processes and infections (Falchi et al. 2014; Scott et al.
2016). PPI can be characterized as ‘it performs similar to networks for mathematical rep-
resentation of physical contacts between cell proteins. Composed contacts between bind-
ing regions in proteins have specific biological importance. Also, it obtains the experimen-
tal and bioinformatics strategies from PPI’s database (Li and Lai 2007; Szklarczyk et al.
2015). PPI interface is also referred to as a collection of multiple residues (Cukuroglu
et al. 2014). In this way, it turns into a new class for drug targets that are different from
mainstream pharmaceutical targets like ion channels, coupled receptors, G-protein, etc
(Higueruelo et al. 2013; Santos et al. 2017). At that point, a new class will extend the target
space for improving small molecule drugs (Shin et al. 2017). When contrasted with tradi-
tional drug targets, target PPI’s reduces harmful impacts because of increment in biological
selectivity of regulatory impacts (Valkov et al. 2011). It is mandatory to learn fundamental
ideas of the PPI interface on the protein-protein structure. Because of the less accessibility
of PPI’s data, it contributes many computational techniques for predicting PPI’s interface
(Xue et al. 2015). Those techniques are dependent on a template which makes it simple
for PPI interface protection (Zhang et al. 2010). For example, a website name “eFindSite”
(Maheshwari and Brylinski 2016) utilized for predicting PPI interfaces which consist of
templates, residues, and sequence-related features for improving SVM, NBC techniques. If
the chance of two interactive protein structures is vacant then it makes it easy for predicting
the PPI interface (Vakser 2014) where it mainly relies on complementarity rules of pro-
tein-protein docking (Chen et al. 2003) and SymmDock strategies (Schneidman-Duhovny
et al. 2005). When two unbound proteins are integrated and converged as one protein,
then a difficulty emerges for predicting the conformational change. When an equivalent
accent sequence needs to be derived, deep learning models are used to predict PPI and
better improvement is achieved compared to machine learning models such as SVM (Du
et al. 2016). Searching for druggable sites for interface in the buried zone (in the range of
1500-3000 A2) (Scott et al. 2016) was mandatory. Considering druggable sites as hotspots
because of providing an enormous amount of binding free energy to convince the medical
chemists (Cukuroglu et al. 2014).
Bai et al. (2016), utilized two techniques i.e., fragment docking and direct coupling
analysis for detecting druggable PPI sites. Fragment docking named “iFitDock”, utilized
for checking druggable hot spots(problem areas) in the PPI interface. Further improvement
for candidate binding locales needs to integrate similar small hot spots. At last, based on
the evolutionary conservative level, the scoring function must be located to provide the
finest protein-protein binding spots. The PPI interface objective was to improve computa-
tional methodologies for locating the best hot spots and significant structure of small mod-
ulator targets in the PPI interface.

13
1968 S. Dara et al.

4.3 Prophecy of protein–protein interactions

The Protein-Protein Interactions (PPI) is one of the major biological phenomena

through which the basic units of the body (cell) transports the signals, ions, substrates,
and energy production components that need to improve the pharmacological responses
needed by the body. In another circumstance, the PPI plays a critical role in the patho-
genesis of the disease such as various types of cancer, especially colorectal carcinoma.
The development of colorectal carcinoma in humans is purely dependent on lifestyle
as well as hereditary means. However, the pathogenesis of the colorectal carcinoma
is linked with the formation of malignant Adenomatous Polyposis Coli (APC) and its
migration in the entire colorectal portion in the body is majorly occurs due to the inter-
action of APC protein with Asef (guanine nucleotide exchange factor) and 𝛽 − catenin
with TCF4 component peptide are located in the pathogenic carcinoma cells. The exam-
ple APC-Asef, 𝛽 -catenin-TCF4 PPI has been illustrated in Fig. 12.
In recent years, the PPI-based drug discovery programs are experimentally produced
a hopeful pharmacological substance, in terms of cancer pathogenesis, APC-Asef PPI
inhibitors are the best example which are delivered the basic peptides as an initial point
to switch on the medicinal chemistry oriented drug design projects. The importance of
PPIs in understanding host pathogenic protein interactions is another extreme task that
excites most vaccination programs. Battling against SARS-CoV2 infection is a key para-
digm in the current scenario where the scientific community targets a protein spike from
SARS-CoV2 that preferentially binds to the human angiotensin converting enzyme-2
(hACE2) to enter into the alveoli mainstream of lungs and cause severe obstruction in
respiratory syndrome. However, the time and cost for experimental prediction of PPI
are considered as rate limiting barriers. In this regard, the different databases hosted the
web servers (few are publicly available) framed by targeting PPI which are prevailing as
preliminary PPI identification tools to accelerate the medicinal chemistry research.

Fig. 12  a, b Protein-protein interactions of APC-Asef (yellow surface/cartoon-APC & cyan surface/car-
toon-Asef); and c, d PPI of 𝛽-catenin/TCF4 in surface & cartoon forms (yellow surface cartoon- 𝛽-catenin
& cyan surface/cartoon-TCF4)

13
Machine Learning in Drug Discovery: A Review 1969

4.4 Hit discovery

The Hit discovery process is advanced in success which has been taken in drug discovery.
In this procedure, small molecules are considered as hits for target binding to identify the
best-altered functions. The detection of hit by diverse algorithms is currently prevailing as
a robust technique in the current drug discovery paradigm. An application of multivari-
ate parameters (K-nearest neighbors (K-NN) and support vector machine(SVM)) on high-
content screening (HCS) analysis in one such method produced a variety of hits against
neurological complications.

4.4.1 Drug repurposing

DeepDTnet’s training parameters outperform other existing target identification techniques

and rely on a minimum quantity of FDA-approved drugs (732 drugs) to produce beneficial
therapeutic effects (human retinoic acid receptor orphan receptor gamma t-ROR-𝛾 t) of the
existing topoisomerase inhibitor Topotecan (TPT). The deepDTnet strategy also transfers
several FDA drugs with different chemical scaffolds against GPCR with new targeted phar-
macological actions. (See in Figs. 13, 14, 15). The deepDTnet algorithm is considered to
be much more advantageous than NetLapRLS and KBMF2K methods as well as Naive
Bayes, SVM, KNN, and Random Forest algorithms.
“Repurpose” refers “reprocess/reused/recycle”. Drug Repurposing is characterized
as ‘locating new indications for drugs (Ashburn and Thor 2004; Lotfi Shahreza et al.
2018) which are as of now in the existence stage’. Because it reduces time and hazardous

Fig. 13  The FDA approved drugs under drug-target repurposing applications derived by deepDTnet

13
1970 S. Dara et al.

Fig. 14  The FDA approved drugs under drug-target repurposing applications derived by deepDTnet
(contd.)

Fig. 15  Basic overview of molecular docking sampling and scoring flowchart

13
Machine Learning in Drug Discovery: A Review 1971

circumstances in drug discovery (Ashburn and Thor 2004). A significant reason for uti-
lizing the drug repurposing concept in drug discovery, because it exceptionally support-
ive to have multiple targets (Susan et al. 2017) in each drug which corresponds to various
impacts. In this way, it provides high diversity in drug-disease relationships. Example: Few
drugs extend its life expectancy such as “Metformin” which is an approved medicine to
deal with diseases like “type 2 Diabetes”. In repurpose, essential elements are “drugs and
diseases” (Cabreiro et al. 2013, De Haes et al. 2014, Martin-Montalvo et al. 2013) utilized.
Drug targets and disease genes are other elements utilized in drug repurposing.
In order to show the interactions that have occurred in element (Lotfi Shahreza et al.
2018), this can be performed through the network investigations based on diversity inter-
actions. Nine sorts of networks arranged in drug design concept i.e., Gene regulatory
networks, target-disease networks, drug-adverse networks, metabolic networks, protein-
protein networks, drug-drug networks, drug-disease networks, disease-disease networks,
drug-target networks (Lotfi Shahreza et al. 2018). In general, the network’s model principle
was, indistinguishable drugs have similar targets/effects (Yamanishi et al. 2008). If data is
less or fragmented, in that situation drug repurposing is necessary. For repurposing, inte-
grating the entire multiple networks to create extraordinary (heterogeneous) networks. At
last, consolidate the drug repurposing with drug target prediction to generate drug target
(Wang et al. 2014). So, drug target assists with treating the sicknesses. To generate new tar-
gets and indications, then utilize the network diffusion algorithm and dimensionality reduc-
tion approach (Luo et al. 2017).

4.4.2 Virtual screening

It is an AI strategy utilized in the drug discovery process for locating small molecules to
distinguish bind structures for a drug target. In drug development, virtual screening also
utilized software as well as algorithms to recognize hits from private chemical collections
for retrieving unique hits inefficient way. After identification of new hits, a further step
needs to purify compounds with unfavorable scaffolds (framework) (Lavecchia and Di
Giovanni 2013). And furthermore incorporates hardly includes few strategies like docking-
based, similarity searching (Willett 2006), pharmacore-based (Willett 2006), and machine
learning methods (Leelananda and Lindert 2016). Based on the above techniques, clas-
sification has taken two strategies i.e., structure-based and ligand-based virtual screening.
When 3D-protein structure was accessible then molecular docking process can be
widely utilized (Chen 2015). Many applications related to docking-based virtual screen-
ing have built (Talele et al. 2010) effectively without any impacts. May some obstacles are
present in this strategy such as the scoring function. A scoring function cannot estimate
binding affinities (bond/relationship) with accuracy because insufficient arrangements and
entropy impacts (Huang and Zou 2010) have taken protein flexibility which makes it more
complicated (Chen 2015). Finally, many docking models considered binding affinities and
refuses remained like docking score, distance-time (Copeland 2010; Xing et al. 2017).
When compared to docking-based virtual screening, the ligand-based virtual screening
cannot confide to the 3D-protein structure. Its goal is to design bioactivity domains from
molecular features (Lavecchia and Di Giovanni 2013).
In this concept, the aim is to persistently improve yields and to decrease false hit rates
(Leelananda and Lindert 2016; Liew et al. 2009; Melville et al. 2009). To accomplish this
objective, the SVM technique was frequently utilized in virtual screening (Ma et al. 2009).
DL strategies have been applied to retrieve great classification capacity, low generalization

13
1972 S. Dara et al.

error (LeCun et al. 2015; Thomas et al. 2014) and powerful feature extraction ability.
Example: In virtual screening, sparse distribution method wastes a lot of time in searching
process (Ma et al. 2009; Segler et al. 2018). So as to conquer this issue, molecule librar-
ies must be provided along with unique training molecules (Thomas et al. 2014) among
the Simplified Molecular Input Line Entry Specifications (SMILES) and natural language
relies on long short-term memory network architecture. ML techniques like DNN and gra-
dient boosting trees provided the molecular libraries by RNN. Adversarial autoencoder
models the molecular fingerprints to locate potential anti-cancer agents (Kadurin et al.
2017).

4.5 High throughput virtual screening and scoring in molecular docking

techniques

Routine techniques used after target identification are high through virtual screening
(HTVS) and molecular docking techniques embedded in free energy perturbations, sam-
pling, and scoring algorithms. The knowledge of active site for the protein/receptor where
ligand would bind to mimic/antagonize the physiological role which is an essential task to
initiate the HTVS protocol. Similarly, the ligand-based virtual screening (LBVS) consid-
ered as another basic method relies on the Physico-chemical properties of chemical data-
bases (Fig. 15).

4.5.1 Activity scoring

In virtual scoring, the scoring function is a fundamental component in molecular docking

for assessing binding affinities towards target (Huang and Zou 2010). In machine learn-
ing, mapping ability features can yield great accomplishment to extract physical, geomet-
ric, and chemical features (Khamis et al. (2015)) to retrieve scores. Based on scores, data-
driven black box models which are considered to predict interactions in binding affinities
and furthermore avoiding few concepts in docking like physical function are very hard to
study (Ain et al. 2015). Random Forest and SVM concepts identified with AI utilization
for better performance in the scoring function. For instance, an SVM model can be utilized
instead of a linear additive method related to the energy terms concept. Since an SVM can
characterize the relationship between experimental binding affinities and own energy terms
i.e., can be extracted from docking program eHiTS. Thus, data gives better execution in
scoring power and screening power (Kinnings et al. 2011; Zsoldos et al. 2007).
Numerous researchers initiated in utilizing the CNN model in image processing (LeCun
et al. 2015) field because CNN demonstrated better performance and protein-ligand inter-
actions providing numerous features to CNN for predicting protein-ligand affinities. In the
estimation of protein-ligand affinities, Jimenez et al. worked on the 3D visual representa-
tion of CNN model and binding affinities (Jiménez et al. 2018) which have indicated bet-
ter correlation behavior in data sets. And essentially, deep learning represents its genuine
intensity to increase abstract features from primitive features, since it’s necessary to rep-
resent fundamental features for a compound-protein structure like molecule types, parti-
cle separation (LeCun et al. 2015) etc. A structure Deep VS, reliant on CNN model, got
familiar with abstract features from fundamental features to provide docking programs like
GLIDE SP (Friesner et al. 2004) and ICM (Abagyan et al. 1994). Thus, the point in activity
scoring was, choosing few features among protein-ligand interaction for predicting binding

13
Machine Learning in Drug Discovery: A Review 1973

affinities with help of the CNN model, so it increases information scoring function but it
upgrades the predictive capabilities.

4.6 Hit to lead

It is also referred to as lead generation in the beginning phases of drug discovery. It locates
small molecules referred to as hits from the High Throughput Screen (HTS) through defi-
cient optimization to locate promising lead compounds. The practical interface of hit-to-
lead optimization approach integrated with chemical synthesis as well as mapping algo-
rithm ”design layer”/Random Forest regression applied to create new biologically active
chemical spaces through the utilization of existed kinase inhibitor library (Desai et al.
2013) (Fig. 16).

4.6.1 QSAR

QSAR analysis was used in the hit-to-lead optimization process to find potential lead
compounds from the hit analogs with the prediction of bioactivity analogs (Esposito et al.
2004). And primarily utilized in mathematical concepts to study quantitative mapping with
physicochemical or structural objects and biological activities. QSAR analysis taken apart
in foundation of mathematical models, selection and making the progression of molecular
descriptions, evaluation and interpretation methods, utilization techniques (Myint and Xie
2010). Here, mathematical models and chemical structure representations are considered
issues in QSAR demonstration. When descriptors are chosen, then locating mathematical
models is necessary to fit relationships in the structure-activity technique. In the year 1964,
Hansch equation was suggested by Hansch et al. For clarifying the 2D structure-activity
relationship, utilize the parameters like physicochemical descriptors and linear regression
models for presenting QSAR study as another section (Hansch and Fujita 1964).
In the same year, Free-Wilson model suggested by Free et al. He formulated the bioac-
tivity description and chemical structure relationships have hypothesis concept to contrib-
ute substituent in compound activities (Free and Wilson 1964). Contrasted with the Hansch
method, the Free-Wilson method can encode the chemical structures since it predicts legiti-
mately from the chemical structure without any physiochemical parameters. Random For-
est and SVM are machine learning procedures, used in mathematical models (A Dobchev
et al. 2014; Dudek et al. 2006; Ning and Karypis 2011).
Likewise, QSAR modeling utilized deep learning techniques to retrieve capabilities
in chemical strings and automatically extracts the features. Merck Molecular Activity

Fig. 16  Abl kinase inhibitor obtained from Hit-to-lead optimization protocol linked with ML algorithms

13
1974 S. Dara et al.

challenge was held in 2012 and a team called George Dahl’s won the challenge in ensem-
ble methods like gaussian progress regression, multi-task DNN, and gradient boosting
machine (Ma et al. 2015). Kaggle inspired the results in multi-task DNN. Along with this,
Dahl et al. proceeded to work on the multi-task DNN concept and shown excellent perfor-
mance in single-task neural systems.
Due to multi-task strategy, neural networks learn features from different parameters
however tasks can be similar (Dahl et al. 2014). Ramsundar et al. (2017) utilized multi-
task neural structures in drug development to assess the performance and finally, excel-
lent results appeared in the random forests algorithm. Since multi-task neural structures
consolidated towards platform called Deeepchem. Subramanian utilized canvas descriptors
for employing DNN. Prediction in binding affinities needs to reinforce the regression and
classification model to gain results in human 𝛽-secretase-1 inhibitors (Subramanian et al.
2016). Usage of DNN model gives great results in validation set i.e., classification capabil-
ity gives 0.82 accuracy, it exhibits regression ability R2 with 0.74, MAE (Mean Absolute
Error) is 0.52. DNN model utilizes the 2D descriptors and indicated better results when
compared with force-field-based strategies because of the utilization of partial capability
models in deep learning. At last, QSAR models rely upon deep learning techniques which
allots the better results in the future prediction role of hit-to-lead optimization research.

4.6.2 De novo drug architecture

De novo Drug Architecture progressed unique chemical structures by adjusting or balanc-

ing the target interest (Hartenfeller and Schneider 2010). To introduce a new molecule
from scratch using a popular De novo model called the fragment-based approach. If at
this point there are impracticalities and complexities in the molecular structure (Schneider
et al. 2017), the risk arises in the development of the structure and becomes difficult in the
assessment of bioactivity. Deep learning models utilized powerful knowledge and genera-
tive capabilities to introduce a new structure with appropriate properties (Mullard 2017).
In the De novo drug design process, the deep learning models acts as autoencoder to
generate an appropriate format for new chemical entities (NCE’s). Therefore, an embed-
ment of autoencoder with multilayer perceptron classifier is also a value-added technique
in the generation of NCE’s with predefined physicochemical properties. The syntax of the
drug/chemical structure is produced in SMILES format which might be difficult to under-
stand in many circumstances and grammar variational autoencoder (VAE) overcomes this
problem to accelerate the process (Fig. 17).
Deep reinforcement learning technique extended by Olivecrona et al. for predicting
biological activities to develop new molecules by adjusting RNN model (Olivecrona
et al. 2017). To obtain SMILES syntax, RNN model to be trained; where molecules can

Fig. 17  Smiles/SLN notation of

antiviral compound

13
Machine Learning in Drug Discovery: A Review 1975

collect from chemBL. In reinforcement learning, agents act through actions in activities
under certain conditions. At this point, if the agent gets a positive reward, the actions
made by the agent’s trend can be renewed (Mnih et al. 2015). To acquire a high reward
for activity scoring, then utilize the SVM technique to enhance few approaches rely-
ing upon ligands concept in the training set. Generate few molecules against dopamine
receptor 2-type for employing deep reinforcement learning model with RNN model.
Along with this, it observed predictions have taken over 95% for structures in the bioac-
tive region through the scoring capacity of SVM. By utilizing deep learning techniques,
unique molecules can be created through the auto-encoders technique. To generate new
molecules automatically with appropriate properties then, Gomez-Bombarelli et al.
(2018) integrated multilayer perceptron (MLP) and variational autoencoder (VAE).
In PPI prediction, numerous tackles have taken placed due to (i) spending low
expenditure in protein information, (ii) lack of known PPI to learn about the explicit
virus, (iii) inefficient strategies due to sequence dissimilarity in viral families. The de-
novo methodology motivation is to predict innovative PPI virus with its host. De-novo
was a sequence-based negative examining framework that learns the diverse viruses in
PPI to predict the innovative one, where the shared host proteins can exploit. For assess-
ing generalization, de-novo has endeavored to test the PPI’s with various domains. At
last, the De novo approach retrieved 81% accuracy in reducing the noisy negative asso-
ciations and 86% accuracy in the viral protein prediction that utilized in the training
period respectively. De-novo strategy accomplished more comparable in intra-species
and single virus-host prediction cases. In this way, it turns to be difficult to predict the
PPI for a contaminated person and optimal accuracy is obtained when carrying out tests
for the human-bacteria interactions (Eid et al. 2016).
To develop biological and chemical prospects, multi-objective optimization tech-
nique and AI has given promising outcomes through entrusting an automated De-novo
compound structure like a human-creative mechanism. In this study, innovative percep-
tion pair, which relies on multi-objective technology, is to apply the RNN algorithm to
automate unique molecules with a de-novo structure build on common properties found
among constant physicochem properties for leading trade-offs. In this view, multiple
chemical libraries related to de-novo structure targeting acetylcholinesterase and neu-
raminidase. For assessing chemical feasibility, validity, drug-likeness, and diversity
content were employed through numerous quality metrics. In the de-novo generative
molecules, molecular docking has taken place for the evaluation of posing and scoring
through X-ray cognate ligands with similar molecular counterparts. At last, multi-objec-
tive optimization and AI are provided to use easily for customizable design techniques
which especially effective for lead advancement and generation (Domenico et al. 2020).
For the most part, the network consists of 3 segments i.e., encoder, decoder, and pre-
dictor. Encoder plays a significant role in changing strings called discrete SMILES into
latent (inactive) space, where vectors are considered as constants. The decoder role was
considering vectors back to the past string stage i.e., discrete SMILES. In the predic-
tor stage, Multi-Layer Perceptron (MLP) approach is used for predicting the molecules.
For retrieving a high prediction ratio in constant vectors, then utilize the gradient-
based technique. To locate new molecules rapidly with appropriate properties, then uti-
lize 2 techniques i.e., Bayesian inference and gradient-based approach. By using both
approaches, a significant advantage was delivering a high predictive ratio consequently,
where humans can comprehend the chemical structure. It does not correlate to chemical
structure when SMILES syntax is invalid. To maintain a strategic distance from such

13
1976 S. Dara et al.

difficulties, make the result source more constrained; Pu et al. used variational autoen-
coder (VAE) for characterizing SMILES syntax (Pu et al. 2017).
For creating molecular fingerprints, Kadurin et al. have utilized the AAE model, were
later referred as druGAN. While using the AAE technique, it demonstrated excellent per-
formance in the VAE model in areas of generation ability, error in reconstruction area,
further extraction ability (Kadurin et al. 2017). Coley et al. (2018) suggested locating
whether the generated molecule was synthetically accessed or not. Depending upon the
reaction database, the neural network was trained because of the availability of excellent
approximation capabilities for retrieving synthetic complexity metrics. The fundamental
explanation behind synthetic reaction is to increase the reactant complexities i.e., the score
in product complexity must be greater than reactant (Andras 2017). Coley strived numer-
ous attempts to build scoring function through encoding chemicals response into product
pair and reactant pair for clarifying correlation inequalities between product and reactant
complexities. To become familiar with any scoring capacity at that point, neural networks
need to be trained where Coley utilized reactant and product pairs in a scope of 22 million.
Along with this, the outcome determined with huge complexities in the synthesis process.
At long last, generative models not just clarify drug activities in inverse synthetic plan-
ning yet additionally discloses synthetic complexities due to disposing of the non-realistic
molecules.

4.7 Lead optimization

The lead optimization is an essential step of the drug discovery process in which the best
medicinally active fragment hits are considered leads to extend the medicinal chemistry
projects. The main aim of the lead optimization is to eliminate the side effects/notorious
effects of the existing active analogues by a minimal structural modification to yield a bet-
ter and safer scaffold. One such example is the optimization of Autotaxin inhibitors such
as GLPG1690 clinical agent which is advanced in human clinical trials to combat pulmo-
nary fibrosis. Another example is to increase the potency by tailor-made approaches to pro-
vide better active analogue. Here, the various properties of ADME/T like Chemical and
physical properties, Absorption, distribution, metabolism and excretion, Toxicity, and the
ADME/T multi-task neural networks are discussed in the following sections.

4.7.1 Chemical and physical properties

In the drug discovery pipeline, physical and chemical properties have been utilized to
reduce significant failures. At that point, deep learning models are utilized lead optimi-
zation techniques to improve unique methodologies (Lusci et al. 2013). Duvenaud et al.
(2015) extracted data from molecular graph directly by adopting the CNN-ANN concept
to perform prediction i.e., (MAE is 0.53+0.07) due to relied upon interpretability concept.
Coley et al. inspired Duvenaud’s work and begun working for better results in molecu-
lar aqueous concepts. And furthermore used the tensor-based convolutional technique and
gave better outcomes as MAE (0.424+0.005).
It’s necessary to clarify molecular graph attribution since tensor-based techniques need
to integrate features like a bond, atom levels. For predicting molecular aqueous solution,
Coley’s employed an enormous number of atom level information compared to Duve-
naud’s model (Coley et al. 2017). Establishing a great correlation between Caco-2 per-
meability coefficients and oral drug absorption (P app) for predicting the candidate drug

13
Machine Learning in Drug Discovery: A Review 1977

(P app) (Artursson and Karlsson 1991; Hubatsch et al. 2007) in the estimation of phar-
macokinetic properties. To fabricate prediction templates with 30 descriptors (Wang et al.
2016) at that point, Wang et al. composed 1,272 components for permeability information
of Caco-2 including models like SVM regression, boosting. In the testing set, the boosting
model demonstrated the best outcomes with great expectation capability. It follows QSAR
principles from OECD (Organization for Economic Co-operation and Development). So as
to persuade reliability and rationality, then follow the sequence of OECD standards.

4.7.2 Absorption, distribution, metabolism and excretion

Entering medicines or drugs into veins of the human body under some activity site known
as drug absorption. For examining the degree of absorptions utilize the bioavailability
parameter. Numerous clinical departments clarified optimization of absorption properties
with a prediction of bioavailability molecules (Tian et al. 2011). In the usage of the MLR
model, Tian et al. employed 1,014 molecules for bioavailability prediction through molecu-
lar assets and structural fingerprints. By utilizing the genetic function technique, excellent
results appeared in predictive performance as RMSE = 0.2355 and correlation coefficient
is 0.71 respectively. Conveying drugs or medicine into the human body i.e., intracellular
and interstitial fluids along with few drug absorption (Sim 2015) properties called as drug
distribution. Drug distribution at steady state (VDss) is a proportion of dosage from vivo
stage into plasma reaction. The steady phase in drug distribution is the significant index
for evaluating the drug distribution process. Thus, VDss must be predicted; Lombardo
and Jing have created PLS and Random Forest techniques along with 1,096 molecules
(Lombardo and Jing 2016). Here, board members are not satisfied with prediction results
because 50% of molecules are accessible in twofold error. VDss may influence by the pres-
ence of obscure factors. To defeat this issue, intently taken as a challenge for VDss value
in molecular structural data. If a drug or drug enters the human body under the conditions
applied, the drug itself tries to produce the current toxic metabolite in order to successfully
structure the metabolism. To ensure the strength of the metabolic structure, use structural
optimization techniques to encourage the metabolism to make predictions with high accu-
racy. Many AI strategies adopted a huge amount of drug metabolism information to pre-
dict unique metabolic enzymes like UDP-glucuronosyltransferases (UGT’s), cytochrome
P450s, etc. Furthermore, neural networks trained in UGT metabolism at Xenosite (Matlock
et al. 2015; Zaretzki et al. 2013) platform for predicting the UGT metabolism (Dang et al.
2016). Eliminating dosage from drugs and also metabolites from the human body referred
to as drug excretion. Drug metabolites are wiped out from the human body either with the
usage of water (i.e., some drugs can be soluble in water) or it directly eliminated through
the absence of metabolism. For retrieving excellent results in unique mechanisms, Lom-
bardo et al. utilized the PCA technique with an expectation pace of 84% (Lombardo et al.
2014) accuracy.

4.7.3 Toxicity and the ADME/T multi‑task neural networks

In clinical and preclinical damage accomplishment was reduced the adequacy of about
33% of significant molecules in drug localization, optimizing the significant molecules
reducing risk hazards by predicting toxicity (Guengerich 2010). Prediction can perform
through techniques called structural alerts and rule-based expert knowledge for toxicity
profiles like kidney and liver. Here, deep learning models are required to produce better

13
1978 S. Dara et al.

results in toxicity prediction. Along these, Xu et al. created a prediction model named
acute-oral toxicity, for predicting results on molecular graph encoding CNN (MGE-CNN).
Predicted outcomes indicated as better when compared with SVM model (Youjun et al.
2017). Therefore, the MGE-CNN model succussed because of feature extraction, model
development, molecular encoding is similar in training for neural networks. The advantage
was, the issue can alter through molecular fingerprints because of accessibility of flexibil-
ity in the MGE-CNN model. For acquiring great fragments relates to structural alerts, Xu
et al. utilized toxic features for fingerprints which characterizes TOX Alerts (Sushko et al.
2012). If parameters were comparative, then it’s necessary to correlate with trained multi-
task neural networks and performance demonstrated better results contrasted with single
task neural networks (Mayr et al. 2016) because of sharing parameters and more supportive
towards multiple tasks for retrieving similar features. At last, some information is provided
to the human body when drug absorption, distribution, metabolism, and excretion has han-
dled and prediction improved through performing multi-tasking neural networks. Here, sin-
gle-task and multi was tasks contrasted by Kearnes et al. with ADME/T experimental data,
and outcome demonstrated better performance in multi-task model (Kearnes et al. 2016).

4.8 ML in e‑Resources for drug discovery

The AI and ML algorithms prevailed as the main computational scoring functions for eval-
uation when a predicted value was added as a parameter, which is involved in the basic
drug discovery paradigm (Stork et al. 2020), it illustrated in 18. The detailed applications
of the ML algorithms specified in the e-resource are described in the following sections
(Fig. 18).

4.8.1 ML in Pan‑assay interference screening (PAINS)

The precise information about hits can be obtained from primary or secondary biologi-
cal screening assays of purchasable/commercially available databases which were the most
important parameters before starting medicinal chemistry projects. Thus, elimination of the
compounds has been exhibited its presence in different cellular biological assays consid-
ered as pan-assay derived hits could reduce the cost and time of the medicinal chemists.
The pan assay information can be accessed from the PAINS database on request. There-
fore, the Hit Dexter 2.0 web server has been launched compiled from Pubchem library
and screening assays. The Hit Dexter 2.0 could be initially utilized to know the biological

Fig. 18  ML in e-Resources of

drug discovery platform

13
Machine Learning in Drug Discovery: A Review 1979

properties of the newly designed compound and thus anyone can easily eliminate the pan-
assay interfering compound at the initial stage itself (Stork et al. 2019).

4.8.2 ML in drug metabolite and metabolic site prediction

The identification of metabolic site for any kind of drug or new chemical entity is very
essential before its administration into the human body. The prediction of drug metabolism
can be done by animal models (preclinical studies) which was a rate-limiting step as well
as costly and it is mandatory to retrieve therapeutic approval of new chemical entities. The
site of metabolism can be predicted by several modules among ”ADMET Predictor” of
SimulationsPlus tools have gained attention and is pure works on the models compiled by
the artificial intelligence algorithms. The FAME3 is one of the online servers which pre-
dicts the region for the given drug/compound which undergoes metabolism validated data-
bases gathering phase-1/phase-2 metabolic parameters associated with several databases
validated by comparing with Matthews correlation coefficient (MCC) (Stork et al. 2020). It
is also important to have an overview of the chemical modification of drugs/NCE’s which
are undergone the metabolism and thus can be used in calculating dosage regimen, dosage
frequency, toxicity, and other beneficial side effects. The online services such as GLORY/
GLORYx provides the precise information about the possibilities of new metabolite and
their relevant formation data with respect to mitochondrial cytochromeP450 enzyme and
conjugations (de Bruyn Kops et al. 2019).

4.8.3 ML in skin sensitivity parameter prediction

The prediction of skin sensitivity is one of the essential criteria for assessing safety param-
eters of the new drugs/compounds and it is patient to patient specifications. In this regard,
the AI models such as Random Forest based MACCS (RF_MACCS) and support vec-
tor machine (SVM) based PaDEL (SVM_PaDEL) algorithms trained with approximately
1400 ligands linked with local lympho node assay (LLNA) information (Stork et al. 2020;
Vranic et al. 2019).

4.8.4 ML in natural product identification

The ML trained with 265000 natural product isolates and synthetic libraries validated by
MCC is being used as a basic predictive model NP Scout online server will reveal the prob-
able identity of the newly discovered analogs. The application of NP Scout in the predic-
tion of sources for the query molecule might provide information about their natural prod-
uct sources and could become a part of natural product-based drug discovery (An et al.
2019).

5 Drug discovery problems

In drug development and discovery, numerous clinicians and specialists confronted chal-
lenges towards target validation, computational pathology data, identification of prognostic
biomarkers in clinical preliminaries.

13
1980 S. Dara et al.

5.1 Target validation

By regulating the molecular target activity, drugs can be developed through the utili-
zation of ultimate methodologies in drug discovery for altering the infection state. By
inaugurating a program in drug development, target identification requires a therapeutic
hypothesis for modulating target regulation in the outcome of the infection state. When
available evidence is identified for that target, it can be considered as target identifica-
tion. Based on fundamental decisions, in vivo and ex vivo models are utilized to vali-
date the target disease. In target validation, outcomes can be retrieved through clinical
preliminaries, yet it’s necessary to concentrate on target validation efforts for successful
projects. The diseases incorporate metabolomic, transcriptomic, proteomic profiles that
are available in-patient clinical material. With the clinical database, the capability of re-
utilizing data through public databases provides the primitive target identification and
target validation. For predicting target identification, it requires appropriate strategies
for yielding legitimate statistical models.
ML approaches are used in target identification because of the increment of data-
driven target identification experiments. In target identification, recognizing causal con-
federation among disease and target is the initial step. Target disease modulates either
naturally or artificially (experimental). By using ML approaches, prediction can be
taken placed on known properties of targets, causalities, driven targets. ML techniques
can apply from various perspectives in the target identification field. For predicting
genes with dysphoria, a decision-tree classifier need to be trained on a protein-protein
localization network (Costa et al. 2010). So, distinguished few key parameters in deci-
sion-tree inspection i.e., extracellular path, transcription factors, metabolic paths. John
et al. improved a classifier model called SVM with genomic details for classifying pro-
teins towards non-drug and drug spots in ovarian and breast cancer (Jeon et al. 2014).
mRNA expression, network topology, protein-protein interaction, DNA copy numbers
are the key segments in classification and recognized 122 cancer targets globally. Tar-
gets identified as 462, 266, and 355 related to pancreatic, breast, and ovarian tumors.
Peptide inhibitors were validated through the prediction of two targets. Outcomes in
the cell culture approach were identified as more prominent anti-proliferative effects.
Although, in pancreatic tumors, usage of inhibitors shown twice greater inhibition on
cells.
To distinguish transcriptional changes in Huntington’s disease, Ament et al. devel-
oped a model called mouse transcription factor site with transcriptome information
(Ament et al. 2018). By utilizing LASSO and regression models in mouse striatum, a
genome-scale has been created for 718 transcription factors. Transcriptional factor mod-
ules are recognized to provide treatment in the early phases of Huntington’s disease. In
tissue-related anti-aging treatments, Mamoshina et al. (2018) identified molecular tar-
gets for comparing gene-expression signature with old and new muscles. When con-
trasted with supervised machine learning models, SVM exposed feature selection and
linear kernels are generally appropriate for identifying biomarkers. Predicted targets
can be developed through ML i.e., blind drugs can furtherly be utilized for therapeutic
assumptions. For identifying affiliations like gene-disease, drug-disease, target-drugs,
then apply NLP kernel strategies in Medline concept (Bravo et al. 2015). Many super-
vised learning techniques rely upon EU-ADR [European Union Adverse Drug Reaction]
database for disease genes identification in the Medline concept. NLP technique is used
in the extraction of biological entity events (Kim et al. 2017).

13
Machine Learning in Drug Discovery: A Review 1981

For identifying therapeutic treatment through novel targets, ML is the best extension for
understanding biological aspects. The splicing signal model is an example had in curing Alz-
heimer’s disease. DL splicing signal model is utilized to predict alternate signal (Leung et al.
2014). Binding the integrative splicing signals (Jha et al. 2017) like RNA sequencing data
and CLIP-seq splicing data indicated knock-down results. To identify variations in Alzheimer
disease (Vaquero-Garcia et al. 2016), then code models like complex variants and de-novo
designs must integrate for prediction. ML can predict cancer-related drug impacts (Iorio et al.
2016). So that, ML investigated how DNA-methylation, somatic mutation data, genome-wide
data impacts the drug feedback. To identify molecular features, then utilize logical models,
ANOVA, and machine learning models like random forests for predicting the drug response.
Gene expression, DNA methylation are recognized as the best predictive data types in can-
cer regions. Data utilized from RNAi screens to locate molecular features from 501 cancer
lines, so it predicts 769 genes from cancer cells (Tsherniak et al. 2017). 171 chemicals are nec-
essary to locate in genetic affiliations because targetable vulnerabilities revealed as oncotypes
don’t influence cancer therapy (McMillan et al. 2018). The models used in predictive data
types how therapy in cancer-intrinsic medicine. Many queries emerge for developers i.e., how
specific drugs are developed for the given target. For identifying targets in small molecular
design, proteins suggested integrating with small molecules for delivering drugs. In this way,
a random forest algorithm must train on genomic attributes like physicochemical and cavi-
ties of 1,187 compounds in non-drug adhesive sites against 99 protein collection (Nayal and
Honig 2006). Additionally, length and configuration are considered significant features in sur-
face cavities. For predicting drug targets, distinctive physicochemical properties from protein
sequences applied SVM’s (Li and Lai 2007; Bakheet and Doig 2009) DL model (Bakheet and
Doig 2009). Proteins occupy explicit locations in PPI network to associate exceptionally (Jeon
et al. 2014; Costa et al. 2010; Kandoi et al. 2015). ML algorithms utilized newly developed
targets to predict blind drugs for reducing search space, but drug target requires more endorse-
ments. Predicting the clinical trial success in drug targets is a complicated goal for target vali-
dation and identification. Along ML approaches, omics information utilized 332 drug targets,
so it can come up short or accomplishment in the third phase of clinical trials through multi-
variate compound selection (Rouillard et al. 2018).
Gene-expression data is identified as successful prediction across tissue layers with high
variance and less RNA mean expression in clinical trials. In this way, the drug target was
confirmed that specific disease expression can influence tissue region (Kumar et al. 2016).
For predicting de-novo therapeutic drug targets, (Koscielny et al. 2017) ML classifiers should
train from open platform (Ferrero et al. 2017). Significant indications are key data types such
as genetic data, gene expression for predicting therapeutic drug targets. In such cases, ML
approaches constrained because of data absence and sparse data are fundamental purposes
behind failure in drug development programs. Practically, to initiate any drug in the market, it
considers the length of time period due to more advancement in technology, new models like
biologics (antibodies were included) can accessible and small molecular drug design may not
same as today. Additional constraints are developed to predict medicine because it can fail or
succeed with accessible metadata in public space.

5.2 Prognostic biomarkers

Using the ML approach, biomarker discovery is used to improve clinical trial performance
by differentiating drugs and understanding drug mechanisms for reasonable patients (Li
et al. 2015; van Gool et al. 2017; Kraus 2018). It consumes a lot of time and cost in the

13
1982 S. Dara et al.

final stages of clinical trials. To defeat this issue, necessary to apply, build and validate
predicted models in the early stages of clinical trials. Usage of ML algorithms allows pre-
dicting translational biomarkers in preclinical data assortment. After data validation, cor-
responding biomarkers and models must investigate the patient indications and lastly pro-
pose the medication. In literature, several papers provided information relates to predictive
models and biomarkers, and last, few were utilized in clinical trials. Various factors like
model rebuilding, designing, data accessing, data quality and software, model selection
are necessary for a clinical setting. The principal issue was, ML approaches assess com-
munity endeavors for developing regression and classification models. Many years ago,
in US FDA (Food and Drug Administration) led (MAQC II) MicroArray Quality Control
evaluated ML algorithms for predicting gene expression data (Shi et al. 2010) in the final
stage of clinical trials. In this project, 6 microarray data collections were analyzed by 36
independent groups to develop predictive models for classifying in the end stage of clini-
cal sites. For modelling appropriate approaches in a clinical trial, information incorporates
data quality, skilled scientists, control processes. Multiple myeloma is a poor prediction
in patients and cut-off within 24 months due to partially applied. Here, the regression-
based approach is appropriate for prediction because multiple myeloma and gene expres-
sion are continuous variables. By utilizing Cox regression models, it confirmed to predict
(Zhan et al. 2006) patient risk factors through gene expression signature. In this review,
the advantage was, utilizing regression models (Shaughnessy et al. 2007; Zhan et al. 2008;
Decaux et al. 2008; Mulligan et al. 2007) can be highlighted due to the absence of prede-
fined classes that can perform prediction in clinical trials. To evaluate regression models,
NCI (National Cancer Institute) challenge is to build drug predictive models (Costello et al.
2014). Each group must utilize the best model with key parameters in training data col-
lection (i.e., treating 35 breast tumor cells with 31 drugs) and models ought to be veri-
fied through similar blind testing data collection (i.e., treating 18 breast tumor cells with
similar 31 drugs). For generating more predictive techniques, six sorts of data profiles are
considered i.e., RNA sequencing, RNA microarray, reverse protein phase array, SNP (Sin-
gle Nucleotide Polymorphism) array, DNA methylation status, exome sequencing for 44
groups are utilized for applying multiple regression models like sparse linear regression,
kernel methods, regression trees, principal component methods. In MAQC II results, indi-
vidual groups performed well and other groups utilized similar models. In differentiating,
few groups maintained technical details like feature selection, quality control, data reduc-
tion, tuning ML parameters, splitting strategy, and biological data like gene expression data
to improve the predictive model. Numerous drugs are convenient in the development of the
predictive model when compared to other strategies.
Challenge of NCI-DREAM needs to maintain a data collection and outcomes for evalu-
ating, improving group factor analyses in validation (Bunte et al. 2016), Random forest
framework (Rahman et al. 2017) and other approaches (Huang et al. (2017); Hejase and
Chan (2015)). Predictive ML models were published in several papers where biomarkers
play a significant role in drug development and discovery. A conference was conducted
in utilizing the tumor cell screen data to create drug sensitivity models (i.e., sorafenib and
erlotinib) (Li et al. 2015). In BATTLE clinical trials (Kim et al. 2011), improved models
ought to apply to patients for finalizing whether these approaches are drug-specific and
predictive. In this case, study, utilizing ML models helps in recognizing key parameters in
drug sensitivity sites across tumors in tissue cells. PD1 (Programmed cell Death 1) inhibi-
tor endorsed by FDA in 2017, at that situation, genetic biomarkers utilized s pembroli-
zumab as inhibitors for tumors. It was the first endorsement made by FDA that relates to
genetic biomarkers other than tumor type (Boyiadzis et al. 2018), which can highlight the

13
Machine Learning in Drug Discovery: A Review 1983

biomarker disclosure. Recently, predictive biomarkers indicated improvement in ML other

than different oncology data types. For improving drug responses in patients, ML algo-
rithms ought to apply multi-omics data (Tasaki et al. 2018). And gradient regression tree is
utilized for improving polygenic risk scores in predicting clinical trials (Paré et al. 2017).
Tested outcomes in UK Biobank explained the presentation of SNP model is indicated
polygenic variance as 46.9% for height, 32.7% for BMI. For distinguishing high complexes
in individuals such as cardiac arrests, breast cancers, inflammatory bowel cancers, at that
point, genome-wide scored data must develop (Khera et al. 2018).
RNA sequencing for single-cell innovation is widely utilized in advanced biomarker
discoveries and gene clustering. This technique is utilized to locate lineages of trace devel-
opment, determining cell states, novel cell varieties. Here, reducing estimations in gene
expression from thousand cells into the low-dimensional regions was the unresolved issue.
For reducing high-dimensional into low dimensional form, Ding et al introduced probabil-
istic generative structure in gene expression of single-cell data accompanied by unpredict-
able estimations (Ding et al. 2018). Here, a probabilistic model is widely utilized to exam-
ine RNA sequencing for four single cells data. Along with, it develops 2D structure in the
multi-dimensional regions for distinguishing cell patterns in RNA sequencing single-cell
data. Transformation of RNA sequencing single-cell data into the encoded feature of latent
space, VAE’s (Variational autoencoders) utilized for determining subpopulations in hid-
den tumour (Sabrina et al. 2019). Encoded features assessed few relationships in gene cell
subpopulations. This strategy contains a data pre-processing technique since it relies upon
unsupervised learning. RNA sequencing of single-cell data utilized the VASC model for
data visualization (Wang and Jin 2018).
When testing was conducted on 20 informational sets, results indicated more superior
to VASC model other than SIMLR (Wang et al. 2017) and ZIFA (Pierson and Yau 2015)
reduction models. By utilizing ML approaches, feature selection received huge advance-
ments in biomarker discovery. For extracting appropriate structures in clusters (Tan et al.
2016), many specialists have claimed unsupervised deep learning methods. To locate
explicit structures in VAE encoded features, then the VAE technique must compete with
TCGA (The Cancer Genome Atlas) data in RNA sequencing (Way and Greene 2017). To
upgrade identifications in carcinoma disease, Beck et al. (2011) explained data integra-
tion techniques, image analysis with gene expression data to identify the squamous cells
in lungs. And CNN model showed better execution in predicting the cardiac failures i.e.,
(AUC=0.97) from endomyocardial biopsy data other than (AUC=0.73 and 0.75) trained
samples (Nirschl et al. 2018). From the above examples, the usage of ML approaches has
shown success in biomarker discovery and still, numerous issues need to be rectified. A
few issues considered as; one classifier must understandable by end-users for clinical adop-
tions. Another key issue was, every approach needs to validate the multi-institutional,
multi-site data sets for determining the generalizability approach. Many community par-
ties tended to key issues and providing a quick advancement like model extraction and
interpretations in biological sites (Finnegan and Song 2017), key optimization and training
algorithms (Angermueller et al. 2016), model reproducib0ility (Hutson 2018).

5.3 Digital pathology

The word pathology refers to a realistic field, each pathologist clarified what can see from a
glass slide through visual assessment. A lot of information is produced through glass slides
for example, which cell type is arranged in tissue layer and spatial context. In this way, it

13
1984 S. Dara et al.

is generally imperative to examine relationships between immune cells and immune-oncol-

ogy cancers. In clinical trials, before choosing a patient to test with thousands of com-
pounds, pharmaceutical industries must realize how the particular drug can treat patient
cells and tissues in the body. Because of rapid advancements in clinical trials, locating
biomarkers became more significant for victims i.e., who can ready to react to the therapy.
Fast improvement in digital pathology can discover new biomarkers with more reasonable,
precise, and high-throughput behavior for reducing time in drug development, and also vic-
tims can access therapy very fast. Prior to applying deep learning models, many algorithms
related to image analysis propelled me to collaborate with pathologists. For classifying tis-
sue layers, numerous computer scientists are required to handcraft graphical features in
computers.
The objective of digital pathology study is to recognize etymological descriptors largely
utilized in hematoxylin and easin (H&E) structures. Here, Nuclear morphometry is an
implementation in the digital study for explaining relationships between prognosis (Vel-
tri et al. 2000) and features created by PCs. From the spatial context, Beck et al. (2011)
identified tissues in stroma cancer and stroma survival features in breast cancer. Recently,
the Nuclear orientation structure was explained by LU et al. (2017) for clarifying survival
features in oral cancers and breast cancers (Cheng et al. 2018). In many conditions, anti-
bodies utilized immunohistochemical stains for targeting image proteins. With the absence
of deep learning tools, morphology can detect tissues in sophisticated data. Investigation
of immuno-oncology permits ML approaches for generating high throughput features to
explain thousands of cells associated with a spatial context, and impossible tasks given
for pathologists. Usage of DL methods shows improvement more precisely for tissue and
cell detection in cancer environments. Many different features are explained spatial con-
text associations for cells and tissues through scale estimations. Understanding hetero-
geneity concept in breast-cancer population to utilize lymphocytes in biomarkers (Mani
et al. 2016). The cell-cell relationship was examined and delivered outcomes through cell
locations like CD8+, PD1+ and cell densities for distinguishing carcinoma Merkel cell to
respond in pembrolizumab (Giraldo et al. 2017). For leading a trial, utilized the number
of tissues for each stain. If thousands of features are examined, then cell-cell interaction
increases in each stain. In this circumstance, ML models and feature selections must be
incorporated to predict the therapeutic response.
The CNN model is well applicable for digital pathology works since a single biopsy was
utilized to train feasible pixels. So, DL models automatically learn structured features from
various classification tasks (Janowczyk and Madabhushi 2016). Here model was, M-CNN
(Multi-scale CNN) considered as a supervised learning technique for phenotyping images
with high-content cells (Godinez et al. 2017), where it restricts a few models with their
customized steps. Converting image pixel values to phenotype images, then the M-CNN
approach demonstrated more accuracy at classification levels. For creating objectives in
image analysis, numerous DL methods utilized in tubules (Romo-Bucheli et al. 2016), lym-
phocytes (Saltz et al. 2018; Corredor et al. 2019), mitotic activity (Romo-Bucheli et al.
2016), cancer tumours (Sharma et al. 2017; Korbar et al. 2017; Bychkov et al. 2018; Cruz-
Roa et al. 2017) situated in lung and breast cancers. In digital pathology, DL models pro-
vide information related to other methodologies. Utilization of deep learning models can
stimulate data acquisition (Cohen et al. 2018) of MRI (Magnetic Resonance Imaging) or it
diminishes dosage for radiation in CT (Computed Tomography) image process (Chen et al.
2017). The quality of images improved a lot in noise signal ratio, spatial resolution; so,
applications like victim stratification, disease prediction, image qualification have corre-
spondingly improved. The deep learning framework is another study (Coudray et al. 2018)

13
Machine Learning in Drug Discovery: A Review 1985

which determines to predict the usage of mutated genes called lung cancers from hema-
toxylin & eosin (H & E)-stained images.
In image analysis, numerous deep learning procedures are required to perform explicit
tasks; So, integration of image analysis and deep learning algorithms can be accommo-
dated for problem-solving. In numerous issues, usage of DL techniques can outperform the
results, however, it was not an image analysis tool because of lack of flexibility. Likewise,
many scientific experts are accessible for any classification tasks. However, it consumes a
lot of money to generate. To defeat this challenge (Turkki et al. 2016) immunohistochem-
istry staining would utilize to mitigate this problem. Due to community tasks, it provides
more data for pathologists to build annotations for many use-cases. The transparency
issue is another challenge to digital pathology. Black-box is a known methodology in deep
learning strategies. In classification tasks, decision-making is unclear. For understanding
numerous mechanisms in drug development, interpretable outcomes can be accommodat-
ing in locating potential biomarkers and drug targets for predictive response in therapy.
Additionally, trust should be improved in generating assembled features with interpretabil-
ity. In clinical trials, the large sample size required to apply DL techniques legitimately for
predictive response in therapy is a further challenge. The DL requires countless sample
examples in clinical trials. Sometimes, integrating data in clinical trials can be possible
however the existence of bias can make the outcomes difficult for interpretation. Corredor
et al. (2019) and Saltz et al. (2018) explained numerous models related to image analysis
and DL models for predictive response in therapy, at that point CNN model used to iden-
tify features in sub-sequent graph and lymphocytes situated in H&E-stained cells. In the
future, DL consists of more capabilities to replace nuclear detection and traditional seg-
mentation algorithms for providing spatial context features (Table 2).

6 Challenges

Many challenges are there in Drug discovery, most of the challenges can be solved by
using Machine Learning Techniques. Here, some of the challenges are being given with
possible suggestions.

1. Numerous ML strategies produced precise results, despite the fact that a couple of
parameters and structures lead to trouble during the training period. Especially when
data is insufficient during the training period, the particular algorithm cannot fulfill the
accuracy and local optimum.
  To defeat this issue, a deep belief architecture, which is an unsupervised pre-trained
model needs to be implemented for improving parameters, so the results can be created
with more effectiveness Ghasemi et al. (2018)).
2. The transparency issue is another challenge in drug discovery. Because decision-making
is unclear in different classification models. In drug development, numerous mecha-
nisms need to comprehend for interpreting the outcomes. So, it makes more supportive
in locating new drug targets and multiple assembled features need to improve trust in
interpretability Vamathevan et al. (2019)).
  In drug development, numerous mechanisms like SVM, MLR, RF, and Deep learn-
ing techniques can be implemented to comprehend for interpreting the outcomes. So,
it makes more supportive in locating new drug targets and multiple assembled features
for developing trust in interpretability.

13
 1986

Table 2  Different ML methods related to various tasks in Drug discovery

13
Method Element/features Task Refs.

MACCS Molecular fingerprints Locating anti-cancer molecules Kadurin et al. (2017)

CNN Molecular graph Identifying graph convolutional fingerprints Duvenaud et al. (2015), Coley et al. (2017)
CNN Subsequent graph- based features HE-stained To predict the disease response in lympho- Saltz et al. (2018), Corredor et al. (2019)
tissues cytes
CNN 2D chemical structure image Biological Activity/toxicity Pu et al. (2017)
RNN Molecular Graph Generating molecules with predicted biologi- Olivecrona et al. (2017)
cal activity
RNN SMILES To predict molecular properties Pu et al. (2017)
RNN Molecular fingerprints and protein sequence Compound protein interaction Wang et al. (2016)
RNN SMILES Generating novel molecules Olivecrona et al. (2017)
RNAi 501 cancer cell lines To identify molecular markers for predicting Tsherniak et al. (2017)
cancer dependencies
DNN Atom pair descriptor and donor-acceptor pair To predict molecular bioactivity Ma et al. (2015)
descriptor
DNN Molecular descriptors and protein features Drug target interactions Wang et al. (2014)
M-DNN Molecular descriptor To predict the chemical descriptors Mayr et al. (2016)
Univariate Cox regression Gene expression signatures It identifies the predicted high-risk subgroup Zhan et al. (2006)
of victims
Gradient boosted regression trees Genome-wide polygenic scores To identify high-risks of breast cancer, Khera et al. (2018)
coronary artery type 2 diabetes diseases in
patients.
VAEs RNA sequencing dataset Differentiating hidden cancer subpopulations Sabrina et al. (2019)
with effectively
Auto Encoder Fingerprints Virtual screening Kadurin et al. (2017)
S. Dara et al.
Machine Learning in Drug Discovery: A Review 1987

2. Integrated data can accessible from many references, especially from the ‘omics’ region.
It’s turning out to be more challenging in day-by-day, because not only expanding the
data as well as this data type contains profoundly heterogeneity in pharmaceutical com-
panies (Searls 2005).
  Public databases are available like ZINC, BindingDB, PUBCHEM, Drugbank, and
REAL chemical databases, developers need to create a pipeline architecture to integrate
these heterogeneous data sources. However, the Data warehousing tools which work
based on ETL (Extract Transform and Load) are Integrated Genomic Database, Adapt-
able Clinical Trail Database, DataFoundry, SWISS-PROT, SCoP, and dbEST. Genome
Information Management System, BIOMOLQUEST, PDB, SWISS-PORT, ENZIME
and CATH data (Cornell et al. 2001; Bukhman and Skolnick 2001).
4. Additionally, Homogeneous data can generate integration challenges, commencing with
testing and logical issues, cross-platform normalization, and statistical issues can expand
enormous heterogeneity information (Searls 2005).
  So, ML with Big data analytic can be utilized for integrating homogenous data
sources. Some Ontology-based integration tools are available like Ontology Web Lan-
guage, Extensive Markup Language (XML), RDF Schema or Resource Description
Language (RDF), Unified Medical Language System, etc (A Seoane et al. 2013). Some
weblink based integration tools available like Sequence Retrieval System (Etzold et al.
1996, ChEMBL (Gaulton et al. 2012, NCBI Entrez, PubChem, Integr8, DisaseCard and
EMBL-EBI search and Sequence analysis (A Seoane et al. 2013; Madeira et al. 2019).
Some visualization tools are also available like Microsoft Power BI, IBM Cognos, Tab-
leau, Zoho Analytics, Sisense, SAS Business Intelligence, etc. Because integration and
visualization tools help in identifying bottlenecks and potential problems before which
affects important processes (Soukup and Davidson 2002).
5. In pharmaceutical companies, research was stretched out from huge molecules to indi-
viduals, and generally relied upon integration of heterogenous data which sustain its
own challenges in varying contexts and scales (Searls 2005).
  A high level of artificial intelligence needs to be obtained for managing various
sources and must be improved with a better understanding of the gathered data. So that,
modern data connectors are suggested to centralize the dissimilar data and at last, these
data connectors help in allotting original data.

7 Conclusion and future directions

The AI technology is utilized in pharmaceutical industries including ML algorithms

and deep learning techniques in daily life. ML techniques in drug development regions
and health service centers have encountered numerous conflicts, especially in image
analysis and omics data. In medical science, ML models predict the trained data in a
known framework i.e., the compound structure can perform alternative tools like PPT
inhibitors, macrocycles with traditional algorithms. Additionally, deep learning models
can be considered the chemical structures and QSAR models from pharmaceutical data
which was pertinent for molecules with appropriate properties, because to the forward
success rate in clinical trials. AI technology has taken a forward step in entering into

13
1988 S. Dara et al.

computer-aided drug development to retrieve the powerful capabilities in data mining.

Some issues still existed i.e.,

1. The performance of deep learning methods can directly influence the innovation of data
mining because multiple deep neural networks are effectively trained on a large volume
of data. The main aim is to tackle the transfer learning automatic problem.
2. “Black-Box” model became confused in deep learning concepts. The Local Interpretable
Model-Explanations (LIME) is an example of a counterfactual probe. LIME was utilized
to unlock the black-box model (Voosen 2017). Here, restricted data was mandatory to
explain through deep learning models (Tishby and Zaslavsky 2015). However, revealing
data by deep learning techniques perform only in the initial stages.
3. Many parameters are adjusted during the training period of neural networks but some
theoretical and practical frameworks are out of reach to optimize these models.

7.1 Future directions

Web innovation was integrated with medical science to improve predictive power in deci-
sion-making and deep learning algorithms about biomarkers, side effects in therapies,
therapeutic benefits. In clinical trials, success is achieved through the utilization of par-
ticular applications. So, motivation is performed for future investment in pharmaceutical
companies. In the future, drug discovery and development, looking forward to covering
all aspects by AI technology. Automated AI needs to coordinate theoretical results such as
chemistry information, omics data, and medical data for emerging. Also, we are anticipat-
ing that more confirmations should be rebuilt for the medication revelation campaign.

References
Abagyan R, Totrov M, Kuznetsov D (1994) Icm–a new method for protein modeling and design: applica-
tions to docking and structure prediction from the distorted native conformation. J Comput Chem
15(5):488–506
Ain QU, Aleksandrova A, Roessler FD, Ballester PJ (2015) Machine-learning scoring functions to improve
structure-based binding affinity prediction and virtual screening. Wiley Interdisciplinary Reviews.
Comput Molec Sci 5(6):405–424
Alpaydin E (2020) Introduction to machine learning. MIT press, Cambridge
Ament SA, Pearl JR, Cantle JP, Bragg RM, Skene PJ, Coffey SR, Bergey DE, Wheeler VC, MacDonald ME,
Baliga NS et al (2018) Transcriptional regulatory networks underlying gene expression changes in
huntington’s disease. Mol Syst Biol 14(3):e7435
An H, Li M, Gao J, Zhang Z, Ma S, Chen Y (2019) Incorporation of biomolecules in metal-organic frame-
works for advanced applications. Coord Chem Rev 384:90–106
Andras P (2017) High-dimensional function approximation with neural networks for large volumes of data.
IEEE Trans Neural Netw Learn Syst 29(2):500–508
Angermueller C, Pärnamaa T, Parts L, Stegle O (2016) Deep learning for computational biology. Mol Syst
Biol 12(7):878
Artursson P, Karlsson J (1991) Correlation between oral drug absorption in humans and apparent drug per-
meability coefficients in human intestinal epithelial (caco-2) cells. Biochem Biophys Res Commun
175(3):880–885
Ashburn TT, Thor KB (2004) Drug repositioning: identifying and developing new uses for existing drugs.
Nat Rev Drug Discovery 3(8):673–683
Asher M (2017) The drug-maker’s guide to the galaxy. Nature News 549(7673):445
Bai F, Morcos F, Cheng RR, Jiang H, Onuchic JN (2016) Elucidating the druggable interface of pro-
tein- protein interactions using fragment docking and coevolutionary analysis. Proc Natl Acad Sci
113(50):E8051–E8058

13
Machine Learning in Drug Discovery: A Review 1989

Bakheet TM, Doig AJ (2009) Properties and identification of human protein drug targets. Bioinformatics
25(4):451–457
Beck AH, Sangoi AR, Leung S, Marinelli RJ, Nielsen TO, Van De Vijver MJ, West RB, Van De Rijn M,
Koller D (2011) Systematic analysis of breast cancer morphology uncovers stromal features asso-
ciated with survival. Sci Trans Med 3(108):108ra113-108ra113
Bengio Y (2009) Learning deep architectures for AI. Now Publishers Inc, Norwell
Bengio Y, Lamblin P, Popovici D, Larochelle H (2007) Greedy layer-wise training of deep networks
advances in neural information processing systems
Boyiadzis MM, Kirkwood JM, Marshall JL, Pritchard CC, Azad NS, Gulley JL (2018) Significance and
implications of fda approval of pembrolizumab for biomarker-defined disease. J Immunother Can-
cer 6(1):1–7
Bravo À, Piñero J, Queralt-Rosinach N, Rautschka M, Furlong LI (2015) Extraction of relations between
genes and diseases from text and large-scale data analysis: implications for translational research.
BMC Bioinform 16(1):55
Bukhman YV, Skolnick J (2001) Biomolquest: integrated database-based retrieval of protein structural
and functional information. Bioinformatics 17(5):468–478
Bundela S, Sharma A, Bisen PS (2015) Potential compounds for oral cancer treatment: resveratrol, nim-
bolide, lovastatin, bortezomib, vorinostat, berberine, pterostilbene, deguelin, andrographolide, and
colchicine. PLoS ONE 10(11):e0141719
Bunte K, Leppäaho E, Saarinen I, Kaski S (2016) Sparse group factor analysis for biclustering of multi-
ple data sources. Bioinformatics 32(16):2457–2463
Bychkov D, Linder N, Turkki R, Nordling S, Kovanen PE, Verrill C, Walliander M, Lundin M, Haglund
C, Lundin J (2018) Deep learning based tissue analysis predicts outcome in colorectal cancer. Sci
Rep 8(1):1–11
Cabreiro F, Au C, Leung K-Y, Vergara-Irigaray N, Cochemé HM, Noori T, Weinkove D, Schuster E,
Greene NDE, Gems D (2013) Metformin retards aging in c. elegans by altering microbial folate
and methionine metabolism. Cell 153(1):228–239
Cano G, Garcia-Rodriguez J, Garcia-Garcia A, Perez-Sanchez H, Benediktsson JA, Thapa A, Barr A
(2017) Automatic selection of molecular descriptors using random forest: Application to drug dis-
covery. Expert Syst Appl 72:151–159
Chen Y-C (2015) Beware of docking! Trends Pharmacol Sci 36(2):78–95
Chen H, Engkvist O, Wang Y, Olivecrona M, Blaschke T (2018) The rise of deep learning in drug dis-
covery. Drug Discovery Today 23(6):1241–1250
Chen R, Li L, Weng Z (2003) Zdock: an initial-stage protein-docking algorithm. Proteins Struct Funct
Bioinf 52(1):80–87
Chen H, Zhang Y, Kalra MK, Lin F, Chen Y, Liao P, Zhou J, Wang G (2017) Low-dose ct with a residual
encoder-decoder convolutional neural network. IEEE Trans Med Imaging 36(12):2524–2535
Cheng L, Lewis JS, Dupont WD, Plummer WD, Janowczyk A, Madabhushi A (2017) An oral cavity
squamous cell carcinoma quantitative histomorphometric-based image classifier of nuclear mor-
phology can risk stratify patients for disease-specific survival. Mod Pathol 30(12):1655–1665
Cheng L, Romo-Bucheli D, Wang X, Janowczyk A, Ganesan S, Gilmore H, Rimm D, Madabhushi A
(2018) Nuclear shape and orientation features from h&e images predict survival in early-stage
estrogen receptor-positive breast cancers. Lab Invest 98(11):1438–1448
Cohen O, Zhu B, Rosen MS (2018) Mr fingerprinting deep reconstruction network (drone). Magn Reson
Med 80(3):885–894
Coley CW, Barzilay R, Green WH, Jaakkola TS, Jensen KF (2017) Convolutional embedding of attrib-
uted molecular graphs for physical property prediction. J Chem Inf Model 57(8):1757–1772
Coley CW, Rogers L, Green WH, Jensen KF (2018) Scscore: synthetic complexity learned from a reac-
tion corpus. J Chem Inf Model 58(2):252–261
Copeland RA (2010) The dynamics of drug-target interactions: drug-target residence time and its impact
on efficacy and safety. Expert Opin Drug Discov 5(4):305–310
Cornell M, Paton NW, Wu S, Goble CA, Miller CJ, Kirby P, Eilbeck K, Brass A, Hayes A, Oliver SG
(2001) Gims-a data warehouse for storage and analysis of genome sequence and functional data.
In: Proceedings 2nd Annual IEEE International Symposium on Bioinformatics and Bioengineer-
ing (BIBE 2001). IEEE, pp 15–22
Corredor G, Xiangxue Wang Yu, Zhou CL, Pingfu F, Syrigos K, Rimm DL, Yang M, Romero E,
Schalper KA et al (2019) Spatial architecture and arrangement of tumor-infiltrating lymphocytes
for predicting likelihood of recurrence in early-stage non-small cell lung cancer. Clin Cancer Res
25(5):1526–1534

13
1990 S. Dara et al.

Costa PR, Acencio ML, Lemke N (2010) A machine learning approach for genome-wide prediction
of morbid and druggable human genes based on systems-level data. In: BMC genomics, vol 11.
Springer, Berlin, p S9
Costello JC, Heiser LM, Georgii E, Gönen M, Menden MP, Wang NJ, Bansal M, Hintsanen P, Khan
SA, Mpindi J-P et al (2014) A community effort to assess and improve drug sensitivity prediction
algorithms. Nat Biotechnol 32(12):1202–1212
Coudray N, Ocampo PS, Sakellaropoulos T, Narula N, Snuderl M, Fenyö D, Moreira AL, Razavian N,
Tsirigos A (2018) Classification and mutation prediction from non-small cell lung cancer histopa-
thology images using deep learning. Nat Med 24(10):1559–1567
Cruz-Roa A, Gilmore H, Basavanhally A, Feldman M, Ganesan S, Shih NNC, Tomaszewski J, González
FA, Madabhushi A (2017) Accurate and reproducible invasive breast cancer detection in whole-
slide images: A deep learning approach for quantifying tumor extent. Sci Rep 7:46450
Cukuroglu E, Engin HB, Gursoy A, Keskin O (2014) Hot spots in protein-protein interfaces: towards
drug discovery. Prog Biophys Mol Biol 116(2–3):165–173
Dahl GE, Jaitly N, Salakhutdinov R (2014) Multi-task neural networks for qsar predictions. arXiv pre-
print arXiv:​1406.​1231
Dang NL, Hughes TB, Krishnamurthy V, Swamidass SJ (2016) A simple model predicts ugt-mediated
metabolism. Bioinformatics 32(20):3183–3189
de Bruyn KC, Stork C, Šícho M, Kochev N, Svozil D, Jeliazkova N, Kirchmair J (2019) Glory: generator
of the structures of likely cytochrome p450 metabolites based on predicted sites of metabolism.
Front Chem 7:402
De Haes W, Frooninckx L, Van Assche R, Smolders A, Depuydt G, Billen J, Braeckman BP, Schoofs L,
Temmerman L (2014) Metformin promotes lifespan through mitohormesis via the peroxiredoxin
prdx-2. Proc Natl Acad Sci 111(24):E2501–E2509
Decaux O, Lodé L, Magrangeas F, Charbonnel C, Gouraud W, Jézéquel P, Attal M, Harousseau J-L,
Moreau P, Bataille R et al (2008) Prediction of survival in multiple myeloma based on gene
expression profiles reveals cell cycle and chromosomal instability signatures in high-risk patients
and hyperdiploid signatures in low-risk patients: a study of the intergroupe francophone du mye-
lome. J Clin Oncol 26(29):4798–4805
Deng L, Dong Y (2014) Deep learning: methods and applications. Found Trends Sign Process
7(3–4):197–387
Desai B, Dixon K, Farrant E, Feng Q, Gibson KR, van Hoorn WP, Mills J, Morgan T, Parry DM, Ram-
jee MK et al (2013) Rapid discovery of a novel series of abl kinase inhibitors by application of an
integrated microfluidic synthesis and screening platform. J Med Chem 56(7):3033–3047
DiMasi JA, Grabowski HG, Hansen RW (2016) Innovation in the pharmaceutical industry: new esti-
mates of r&d costs. J Health Econ 47:20–33
Ding J, Condon A, Shah SP (2018) Interpretable dimensionality reduction of single cell transcriptome
data with deep generative models. Nat Commun 9(1):1–13
Dobchev DA, Pillai G, Karelson M (2014) In silico machine learning methods in drug development.
Curr Top Med Chem 14(16):1913–1922
Domenico A, Nicola G, Daniela T, Fulvio C, Nicola A, Orazio N (2020) De novo drug design of targeted
chemical libraries based on artificial intelligence and pair based multi-objective optimization. J
Chem Inform Model
Du T, Liao L, Wu CH, Sun B (2016) Prediction of residue-residue contact matrix for protein–protein
interaction with fisher score features and deep learning. Methods 110:97–105
Duch W, Swaminathan K, Meller J (2007) Artificial intelligence approaches for rational drug design and
discovery. Curr Pharm Des 13(14):1497–1508
Duda RO, Hart PE, Stork DG (2012) Pattern classification. John Wiley & Sons, New Jersy
Dudek AZ, Arodz T, Gálvez J (2006) Computational methods in developing quantitative structure-activ-
ity relationships (qsar): a review. Comb Chem High Throughput Screen 9(3):213–228
Dupond S (2019) A thorough review on the current advance of neural network structures. Annu Rev
Control 14:200–230
Duvenaud DK, Maclaurin D, Iparraguirre J, Bombarell R, Hirzel T, Aspuru-Guzik A, Adams RP (2015)
Convolutional networks on graphs for learning molecular fingerprints. In: Advances in neural
information processing systems, pp 2224–2232
Eid F-E, ElHefnawi M, Heath LS (2016) Denovo: virus-host sequence-based protein–protein interaction
prediction. Bioinformatics 32(8):1144–1150
Engelbrecht AP (2007) Computational intelligence: an introduction. John Wiley & Sons, New Jersy
Esposito EX, Hopfinger AJ, Madura JD (2004) Methods for applying the quantitative structure-activity
relationship paradigm. In: Chemoinformatics. Springer, pp 131–213

13
Machine Learning in Drug Discovery: A Review 1991

Etzold T, Ulyanov A, Argos P (1996) [8] srs: information retrieval system for molecular biology data
banks. Methods Enzymol 266:114–128
Falchi F, Caporuscio F, Recanatini M (2014) Structure-based design of small-molecule protein–protein
interaction modulators: the story so far. Future Med Chem 6(3):343–357
Ferrero E, Dunham I, Sanseau P (2017) In silico prediction of novel therapeutic targets using gene-disease
association data. J Transl Med 15(1):182
Finnegan A, Song JS (2017) Maximum entropy methods for extracting the learned features of deep neural
networks. PLoS Comput Biol 13(10):e1005836
Free SM, Wilson JW (1964) A mathematical contribution to structure-activity studies. J Med Chem
7(4):395–399
Friedman J, Hastie T, Tibshirani R (2001) The elements of statistical learning, vol 1. Springer series in sta-
tistics, New York
Friesner RA, Banks JL, Murphy RB, Halgren TA, Klicic JJ, Mainz DT, Repasky MP, Knoll EH, Shelley M,
Perry JK et al (2004) Glide: a new approach for rapid, accurate docking and scoring. 1. method and
assessment of docking accuracy. J Med Chem 47(7):1739–1749
Gaulton A, Bellis LJ, Bento AP, Chambers J, Davies M, Hersey A, Light Y, McGlinchey S, Michalovich
D, Al-Lazikani B et al (2012) Chembl: a large-scale bioactivity database for drug discovery. Nucleic
Acids Res 40(D1):D1100–D1107
Gertrudes JC, Maltarollo VG, Silva RA, Oliveira PR, Honorio KM, Da Silva ABF (2012) Machine learning
techniques and drug design. Curr Med Chem 19(25):4289–4297
Ghasemi F, Mehridehnavi A, Fassihi A, Pérez-Sánchez H (2018) Deep neural network in qsar studies using
deep belief network. Appl Soft Comput 62:251–258
Giraldo NA, Kaunitz GJ, Cottrell TR, Berry S, Sunshine JC, Nguyen P, Xu H, Orgutsova A, Church CD,
Miller NJ et al. (2017) The differential association of pd-1, pd-l1, and cd8+ cells with response to
pembrolizumab and presence of merkel cell polyomavirus (mcpyv) in patients with merkel cell carci-
noma (mcc)
Godinez WJ, Hossain I, Lazic SE, Davies JW, Zhang X (2017) A multi-scale convolutional neural network
for phenotyping high-content cellular images. Bioinformatics 33(13):2010–2019
Goodfellow I, Bengio Y, Courville A, Bengio Y (2016) Deep learning, vol 1. MIT press, Cambridge
Gopal M (2018) Applied machine learning. McGraw-Hill Education, Chennai
Guengerich FP (2010) Mechanisms of drug toxicity and relevance to pharmaceutical development. Drug
metabolism and pharmacokinetics, p 1010210090
Guney E, Menche J, Vidal M, Barábasi A-L (2016) Network-based in silico drug efficacy screening. Nat
Commun 7(1):1–13
Gunther EC, Stone DJ, Gerwien RW, Bento P, Heyes MP (2003) Prediction of clinical drug effi-
cacy by classification of drug-induced genomic expression profiles in vitro. Proc Natl Acad Sci
100(16):9608–9613
Guo Y, Lezheng Yu, Wen Z, Li M (2008) Using support vector machine combined with auto covariance to
predict protein-protein interactions from protein sequences. Nucleic Acids Res 36(9):3025–3030
Gupta S, Chaudhary K, Kumar R, Gautam A, Nanda JS, Dhanda SK, Brahmachari SK, Raghava GPS
(2016) Prioritization of anticancer drugs against a cancer using genomic features of cancer cells: A
step towards personalized medicine. Sci Rep 6(1):1–11
Gómez-Bombarelli R, Wei JN, Duvenaud D, Hernández-Lobato JM, Sánchez-Lengeling B, Sheberla D,
Aguilera-Iparraguirre J, Hirzel TD, Adams RP, Aspuru-Guzik A (2018) Automatic chemical design
using a data-driven continuous representation of molecules. ACS Cent Sci 4(2):268–276
Hansch C, Fujita T (1964) Additions and corrections-analysis. a method for the correlation of biological
activity and chemical structure. J Am Chem Soc 86(24):5710
Hartenfeller M, Schneider G (2010) De novo drug design. In: Chemoinformatics and computational chemi-
cal biology. Springer, Berlin, pp 299–323
Hassan BM, Ahmad K, Roy S, Mohammad Ashraf J, Adil M, Haris Siddiqui M, Khan S, Amjad Kamal M,
Provazník I, Choi I (2016) Computer aided drug design: success and limitations. Curr Pharm Des
22(5):572–581
Hejase HA, Chan C (2015) Improving drug sensitivity prediction using different types of data. CPT: Phar-
macometrics Syst Pharmacol 4(2):98–105
Higueruelo AP, Jubb H, Blundell TL (2013) Protein-protein interactions as druggable targets: recent techno-
logical advances. Curr Opin Pharmacol 13(5):791–796
Hinton G (2018) Deep learning–a technology with the potential to transform health care. JAMA
320(11):1101–1102
Ho Tin K (1995) Random decision forests. In: Proceedings of 3rd international conference on document
analysis and recognition, vol 1. IEEE, pp 278–282

13
1992 S. Dara et al.

Huang S, Cai N, Pacheco PP, Narrandes S, Wang Y, Wayne X (2018) Applications of support vector
machine (svm) learning in cancer genomics. Cancer Genom Proteom 15(1):41–51
Huang C, Mezencev R, McDonald JF, Vannberg F (2017) Open source machine-learning algorithms for
the prediction of optimal cancer drug therapies. PLoS ONE 12(10):e0186906e0186906
Huang S-Y, Zou X (2010) Inclusion of solvation and entropy in the knowledge-based scoring function
for protein–ligand interactions. J Chem Inf Model 50(2):262–273
Hubatsch I, Ragnarsson EGE, Artursson P (2007) Determination of drug permeability and prediction of
drug absorption in caco-2 monolayers. Nat Protoc 2(9):2111
Hutson M (2018) Artificial intelligence faces reproducibility crisis
Iorio F, Knijnenburg TA, Vis DJ, Bignell GR, Menden MP, Schubert M, Aben N, Gonçalves E, Bar-
thorpe S, Lightfoot H et al (2016) A landscape of pharmacogenomic interactions in cancer. Cell
166(3):740–754
Janowczyk A, Madabhushi A (2016) Deep learning for digital pathology image analysis: a comprehen-
sive tutorial with selected use cases. J Pathol Inform 7
Jeon J, Nim S, Teyra J, Datti A, Wrana JL, Sidhu SS, Moffat J, Kim PM (2014) A systematic approach
to identify novel cancer drug targets using machine learning, inhibitor design and high-throughput
screening. Genome Med 6(7):1–18
Jha A, Gazzara MR, Barash Y (2017) Integrative deep models for alternative splicing. Bioinformatics
33(14):i274–i282
Jiménez J, Skalic M, Martinez-Rosell G, De Fabritiis G (2018) K deep: Protein-ligand absolute binding
affinity prediction via 3d-convolutional neural networks. J Chem Inf Model 58(2):287–296
Jung E, Kim J, Kim M, Jung DH, Rhee H, Shin J-M, Choi K, Kang S-K, Kim M-K, Yun C-H et al
(2007) Artificial neural network models for prediction of intestinal permeability of oligopeptides.
BMC Bioinform 8(1):245
Kadurin A, Aliper A, Kazennov A, Mamoshina P, Vanhaelen Q, Khrabrov K, Zhavoronkov A (2017)
The cornucopia of meaningful leads: Applying deep adversarial autoencoders for new molecule
development in oncology. Oncotarget 8(7):10883
Kandoi G, Acencio ML, Lemke N (2015) Prediction of druggable proteins using machine learning and
systems biology: a mini-review. Front Physiol 6:366
Kapoorb R, Haganb M, Paltab J, Ghosha P (2020) Artificial intelligence methods in computer-aided
diagnostic tools and decision support analytics for clinical informatics. Artif Intell Prec Health
From Conc Appl, p 31
Kearnes S, Goldman B, Pande V (2016) Modeling industrial admet data with multitask networks. arXiv
preprint arXiv:​1606.​08793
Khamis MA, Gomaa W, Ahmed WF (2015) Machine learning in computational docking. Artif Intell
Med 63(3):135–152
Khera AV, Chaffin M, Aragam KG, Haas ME, Roselli C, Choi SH, Natarajan P, Lander ES, Lubitz SA,
Ellinor PT et al (2018) Genome-wide polygenic scores for common diseases identify individuals
with risk equivalent to monogenic mutations. Nat Genet 50(9):1219–1224
Kim ES, Herbst RS, Wistuba II, Lee JJ, Blumenschein GR, Tsao A, Stewart DJ, Hicks ME, Erasmus
J, Gupta S et al (2011) The battle trial: personalizing therapy for lung cancer. Cancer Discov
1(1):44–53
Kim J, Kim J, Lee H (2017) An analysis of disease-gene relationship from medline abstracts by digsee.
Sci Rep 7(1):1–13
Kingma DP, Welling M (2013) Auto-encoding variational bayes. arXiv preprint arXiv:​1312.​6114
Kingma DP, Welling M (2019) An introduction to variational autoencoders. arXiv preprint arXiv:​1906.​
02691
Kinnings SL, Liu N, Tonge PJ, Jackson RM, Xie L, Bourne PE (2011) A machine learning-based method
to improve docking scoring functions and its application to drug repurposing. J Chem Inf Model
51(2):408–419
Konar A (2006) Computational intelligence: principles, techniques and applications. Springer Science &
Business Media, Berlin
Korbar B, Olofson AM, Miraflor AP, Nicka CM, Suriawinata MA, Torresani L, Suriawinata AA, Has-
sanpour S (2017) Deep learning for classification of colorectal polyps on whole-slide images. J
Pathol Inform 8
Koscielny G, An P, Carvalho-Silva D, Cham JA, Fumis L, Gasparyan R, Hasan S, Karamanis N, Magu-
ire M, Papa E et al (2017) Open targets: a platform for therapeutic target identification and valida-
tion. Nucleic Acids Res 45(D1):D985–D994
Kramer MA (1991) Nonlinear principal component analysis using autoassociative neural networks. AIChE
J 37(2):233–243

13
Machine Learning in Drug Discovery: A Review 1993

Kraus VB (2018) Biomarkers as drug development tools: discovery, validation, qualification and use. Nat
Rev Rheumatol 14(6):354–362
Kumar V, Sanseau P, Simola DF, Hurle MR, Agarwal P (2016) Systematic analysis of drug targets confirms
expression in disease-relevant tissues. Sci Rep 6:36205
Larsen ABL, Sønderby SK (2015) Generating faces with torch. URL http://​torch.​ch/​blog/​2015/​11/​13/​gan.​
html
Lavecchia A, Di Giovanni C (2013) Virtual screening strategies in drug discovery: a critical review. Curr
Med Chem 20(23):2839–2860
LeCun Y, Bengio Y, Hinton G (2015) Deep learning. Nature 521(7553):436–444
Leelananda SP, Lindert S (2016) Computational methods in drug discovery. Beilstein J Org Chem
12(1):2694–2718
Leung MKK, Xiong HY, Lee LJ, Frey BJ (2014) Deep learning of the tissue-regulated splicing code. Bioin-
formatics 30(12):i121–i129
Li H, Hou J, Adhikari B, Lyu Q, Cheng J (2017) Deep learning methods for protein torsion angle prediction.
BMC Bioinform 18(1):417
Li Q, Lai L (2007) Prediction of potential drug targets based on simple sequence properties. BMC Bioin-
form 8(1):353
Li B, Shin H, Gulbekyan G, Pustovalova O, Nikolsky Y, Hope A, Bessarabova M, Schu M, Kolpakova-Hart
E, Merberg D et al (2015) Development of a drug-response modeling framework to identify cell line
derived translational biomarkers that can predict treatment outcome to erlotinib or sorafenib. PLoS
ONE 10(6):e0130700e0130700
Li L, Wang B, Meroueh SO (2011) Support vector regression scoring of receptor-ligand complexes for rank-
ordering and virtual screening of chemical libraries. J Chem Inf Model 51(9):2132–2138
Liew CY, Ma XH, Liu X, Yap CW (2009) Svm model for virtual screening of lck inhibitors. J Chem Inf
Model 49(4):877–885
Lombardo F, Jing Y (2016) In silico prediction of vol of distribution in humans. extensive data set and the
exploration of linear and nonlinear methods coupled with molecular interaction fields descriptors. J
Chem Inf Model 56(10):2042–2052
Lombardo F, Obach RS, Varma MV, Stringer R, Berellini G (2014) Clearance mechanism assignment and
total clearance prediction in human based upon in silico models. J Med Chem 57(10):4397–4405
Lotfi SM, Ghadiri N, Mousavi SR, Varshosaz J, Green JR (2018) A review of network-based approaches to
drug repositioning. Brief Bioinform 19(5):878–892
Luo Y, Zhao X, Zhou J, Yang J, Zhang Y, Kuang W, Peng J, Chen L, Zeng J (2017) A network integration
approach for drug-target interaction prediction and computational drug repositioning from heteroge-
neous information. Nat Commun 8(1):1–13
Lusci A, Pollastri G, Baldi P (2013) Deep architectures and deep learning in chemoinformatics: the predic-
tion of aqueous solubility for drug-like molecules. J Chem Inf Model 53(7):1563–1575
Ma XH, Jia J, Zhu F, Xue Y, Li ZR, Chen YZ (2009) Comparative analysis of machine learning methods
in ligand-based virtual screening of large compound libraries. Comb Chem High Throughput Screen
12(4):344–357
Ma J, Sheridan RP, Liaw A, Dahl GE, Svetnik V (2015) Deep neural nets as a method for quantitative
structure-activity relationships. J Chem Inf Model 55(2):263–274
Madeira F, Park YM, Lee J, Buso N, Gur T, Madhusoodanan N, Basutkar P, Tivey ARN, Potter SC, Finn
RD et al (2019) The embl-ebi search and sequence analysis tools apis in 2019. Nucleic Acids Res
47(W1):W636–W641
Maheshwari S, Brylinski M (2016) Template-based identification of protein–protein interfaces using efind-
siteppi. Methods 93:64–71
Maltarollo VG, Kronenberger T, Espinoza GZ, Oliveira PR, Honorio KM (2019) Advances with support
vector machines for novel drug discovery. Expert Opin Drug Discov 14(1):23–33
Mamoshina P, Volosnikova M, Ozerov IV, Putin E, Skibina E, Cortese F, Zhavoronkov A (2018) Machine
learning on human muscle transcriptomic data for biomarker discovery and tissue-specific drug target
identification. Front Genet 9:242
Mani NL, Schalper KA, Hatzis C, Saglam O, Tavassoli F, Butler M, Chagpar AB, Pusztai L, Rimm DL
(2016) Quantitative assessment of the spatial heterogeneity of tumor-infiltrating lymphocytes in
breast cancer. Breast Cancer Res 18(1):78
Martin-Montalvo A, Mercken EM, Mitchell SJ, Palacios HH, Mote PL, Scheibye-Knudsen M, Gomes AP,
Ward TM, Minor RK, Blouin M-J et al (2013) Metformin improves healthspan and lifespan in mice.
Nat Commun 4(1):1–9
Matlock MK, Hughes TB, Swamidass SJ (2015) Xenosite server: a web-available site of metabolism predic-
tion tool. Bioinformatics 31(7):1136–1137

13
1994 S. Dara et al.

Matsumoto A, Aoki S, Ohwada H (2016) Comparison of random forest and svm for raw data in drug dis-
covery: prediction of radiation protection and toxicity case study. Int J Mach Learn Comput 6(2):145
Mayr A, Klambauer G, Unterthiner T, Hochreiter S (2016) Deeptox: toxicity prediction using deep learning.
Front Environ Sci 3:80
McMillan EA, Ryu M-J, Diep CH, Mendiratta S, Clemenceau JR, Vaden RM, Kim J-H, Motoyaji T, Cov-
ington KR, Peyton M et al (2018) Chemistry-first approach for nomination of personalized treatment
in lung cancer. Cell 173(4):864–878
Melville JL, Burke EK, Hirst JD (2009) Machine learning in virtual screening. Comb Chem High Through-
put Screen 12(4):332–343
Miljanovic M (2012) Comparative analysis of recurrent and finite impulse response neural networks in time
series prediction. Indian J Comput Sci Eng 3(1):180–191
Mnih V, Kavukcuoglu K, Silver D, Rusu AA, Veness J, Bellemare MG, Graves A, Riedmiller M, Fidje-
land AK, Ostrovski G et al (2015) Human-level control through deep reinforcement learning. Nature
518(7540):529–533
Morita A, Ariyasu S, Wang B, Asanuma T, Onoda T, Sawa A, Tanaka K, Takahashi I, Togami S, Nenoi
M et al (2014) As-2, a novel inhibitor of p53-dependent apoptosis, prevents apoptotic mitochondrial
dysfunction in a transcription-independent manner and protects mice from a lethal dose of ionizing
radiation. Biochem Biophys Res Commun 450(4):1498–1504
Mulligan G, Mitsiades C, Bryant B, Zhan F, Chng WJ, Roels S, Koenig E, Fergus A, Huang Y, Richardson
P et al (2007) Gene expression profiling and correlation with outcome in clinical trials of the protea-
some inhibitor bortezomib. Blood 109(8):3177–3188
Myint KZ, Xie X-Q (2010) Recent advances in fragment-based qsar and multi-dimensional qsar methods.
Int J Mol Sci 11(10):3846–3866
Nayal M, Honig B (2006) On the nature of cavities on protein surfaces: application to the identification of
drug-binding sites. Proteins Struct Funct Bioinf 63(4):892–906
Ning X, Karypis G (2011) In silico structure-activity-relationship (sar) models from machine learning: a
review. Drug Dev Res 72(2):138–146
Nirschl JJ, Janowczyk A, Peyster EG, Frank R, Margulies KB, Feldman MD, Madabhushi A (2018) A deep-
learning classifier identifies patients with clinical heart failure using whole-slide images of h&e tis-
sue. PLoS ONE 13(4):e0192726
Noble WS (2006) What is a support vector machine? Nat Biotechnol 24(12):1565–1567
Olivecrona M, Blaschke T, Engkvist O, Chen H (2017) Molecular de-novo design through deep reinforce-
ment learning. J Cheminform 9(1):48
Pal SK, Mitra S (1992) Multilayer perceptron, fuzzy sets, classifiaction
Paré G, Mao S, Deng WQ (2017) A machine-learning heuristic to improve gene score prediction of poly-
genic traits. Sci Rep 7(1):1–11
Patel S, Tripathi R, Kumari V, Varadwaj P (2017) Deepinteract: deep neural network based protein-protein
interaction prediction tool. Curr Bioinform 12(6):551–557
Patil K, Jordan EJ, Park JH, Suresh K, Smith CM, Lemmon AA, Mossé Yaël P, Lemmon MA, Rad-
hakrishnan R (2021) Computational studies of anaplastic lymphoma kinase mutations reveal common
mechanisms of oncogenic activation. Proc Natl Acad Sci 118(10)
Pierson E, Yau C (2015) Zifa: Dimensionality reduction for zero-inflated single-cell gene expression analy-
sis. Genome Biol 16(1):1–10
Polamuri S (2017) How the random forest algorithm works in machine learning. Retrieved December, 21
Poole D, Mackworth A, Goebel R (1998) Computational intelligence
Pu Y, Wang W, Henao R, Chen L, Gan Z, Li C, Carin L (2017) Adversarial symmetric variational autoen-
coder. In: Advances in neural information processing systems, pp 4330–4339
Rahman R, Matlock K, Ghosh S, Pal R (2017) Heterogeneity aware random forest for drug sensitivity pre-
diction. Sci Rep 7(1):1–11
Rahman R, Otridge J, Pal R (2017) Integratedmrf: random forest-based framework for integrating prediction
from different data types. Bioinformatics 33(9):1407–1410
Ramsundar B, Liu B, Zhenqin W, Verras A, Tudor M, Sheridan RP, Pande V (2017) Is multitask deep learn-
ing practical for pharma? J Chem Inf Model 57(8):2068–2076
Rolan P, Danhof M, Stanski D, Peck C (2007) Current issues relating to drug safety especially with regard
to the use of biomarkers: A meeting report and progress update. Eur J Pharm Sci 30(2):107–112
Romo-Bucheli D, Janowczyk A, Gilmore H, Romero E, Madabhushi A (2016) Automated tubule nuclei
quantification and correlation with oncotype dx risk categories in er+ breast cancer whole slide
images. Sci Rep 6:32706
Rosenblatt F (1961) Principles of neurodynamics. perceptrons and the theory of brain mechanisms. Techni-
cal report, Cornell Aeronautical Lab Inc Buffalo NY

13
Machine Learning in Drug Discovery: A Review 1995

Rouillard AD, Hurle MR, Agarwal P (2018) Systematic interrogation of diverse omic data reveals inter-
pretable, robust, and generalizable transcriptomic features of clinically successful therapeutic targets.
PLoS Comput Biol 14(5):e1006142
Sabrina R, Sohrab S, Ziv BJ, Ravi P (2019) Dhaka: variational autoencoder for unmasking tumor heteroge-
neity from single cell genomic data. Bioinformatics
Saltz J, Gupta R, Hou L, Kurc T, Singh P, Nguyen V, Samaras D, Shroyer KR, Zhao T, Batiste R et al
(2018) Spatial organization and molecular correlation of tumor-infiltrating lymphocytes using deep
learning on pathology images. Cell Rep 23(1):181–193
Samigulina G, Zarina S (2017) Immune network technology on the basis of random forest algorithm for
computer-aided drug design. In: International Conference on Bioinformatics and Biomedical Engi-
neering. Springer, pp 50–61
Santos R, Ursu O, Gaulton A, Bento AP, Donadi RS, Bologa CG, Karlsson A, Al-Lazikani B, Hersey A,
Oprea TI et al (2017) A comprehensive map of molecular drug targets. Nat Rev Drug Discovery
16(1):19–34
Schmidhuber J (2015) Deep learning in neural networks: an overview. Neural Netw 61:85–117
Schneider G, Funatsu K, Okuno Y, Winkler D (2017) De novo drug design-ye olde scoring problem revis-
ited. Mol Inf 36(1–2):1681031
Schneidman-Duhovny D, Inbar Y, Nussinov R, Wolfson HJ (2005) Patchdock and symmdock: servers for
rigid and symmetric docking. Nucleic Acids Res 33(suppl-2):W363–W367
Scott DE, Bayly AR, Abell C, Skidmore J (2016) Small molecules, big targets: drug discovery faces the
protein-protein interaction challenge. Nat Rev Drug Discovery 15(8):533
Searls DB (2005) Data integration: challenges for drug discovery. Nat Rev Drug Discovery 4(1):45–58
Segler MHS, Kogej T, Tyrchan C, Waller MP (2018) Generating focused molecule libraries for drug discov-
ery with recurrent neural networks. ACS Cent Sci 4(1):120–131
Seoane JA, Aguiar-Pulido V, Munteanu C, Rivero D, Rabunal J, Dorado J, Pazos A (2013) Biomedical
data integration in computational drug design and bioinformatics. Curr Comput Aided Drug Des
9(1):108–117
Sharma H, Zerbe N, Klempert I, Hellwich O, Hufnagl P (2017) Deep convolutional neural networks for
automatic classification of gastric carcinoma using whole slide images in digital histopathology.
Comput Med Imag Graph 61:2–13
Shaughnessy JD Jr, Zhan F, Burington BE, Huang Y, Colla S, Hanamura I, Stewart JP, Kordsmeier B, Ran-
dolph C, Williams DR et al (2007) A validated gene expression model of high-risk multiple myeloma
is defined by deregulated expression of genes mapping to chromosome 1. Blood 109(6):2276–2284
Shi L, Campbell G, Jones W, Campagne F, Wen Z, Walker S, Su Z, Chu T, Goodsaid F, Pusztai L, et al.
(2010) The maqc-ii project: a comprehensive study of common practices for the development and
validation of microarray-based predictive models
Shin W-H, Christoffer CW, Kihara D (2017) In silico structure-based approaches to discover protein-protein
interaction-targeting drugs. Methods 131:22–32
Sim, DSM (2015) Drug distribution. In: Pharmacological Basis of Acute Care, Springer, Berlin, pp 27–36
Sistare FD, Dieterle F, Troth S, Holder DJ, Gerhold D, Andrews-Cleavenger D, Baer W, Betton G, Bounous
D, Carl K et al (2010) Towards consensus practices to qualify safety biomarkers for use in early drug
development. Nat Biotechnol 28(5):446–454
Smola AJ, Schölkopf B (2004) A tutorial on support vector regression. Stat Comput 14(3):199–222
Soukup T, Davidson I (2002) Visual data mining: techniques and tools for data visualization and mining.
John Wiley & Sons, New Jersey
Spencer M, Eickholt J, Cheng J (2014) A deep learning network approach to ab initio protein secondary
structure prediction. IEEE/ACM Trans Comput Biol Bioinf 12(1):103–112
Stokes A, Hum W, Zaslavsky J (2020) A minimal-input multilayer perceptron for predicting drug-drug
interactions without knowledge of drug structure. arXiv preprint arXiv:​2005.​10644
Stork C, Chen Y, Sicho M, Kirchmair J (2019) Hit dexter 2.0: machine-learning models for the prediction of
frequent hitters. J Chem Inf Model 59(3):1030–1043
Stork C, Embruch G, Šícho M, de Bruyn Kops C, Chen Y, Svozil D, Kirchmair J (2020) Nerdd: A web por-
tal providing access to in silico tools for drug discovery. Bioinformatics 36(4):1291–1292
Subramanian G, Ramsundar B, Pande V, Denny RA (2016) Computational modeling of 𝛽-secretase 1 (bace-
1) inhibitors using ligand based approaches. J Chem Inf Model 56(10):1936–1949
Susan K, Stephanie H, Mathias W, Harald P, Binje V, Paul-Albert K, Maria R, Benjamin R, Svenja P, Chen
M et al (2017) The target landscape of clinical kinase drugs. Science 358(6367)
Sushko I, Salmina E, Potemkin VA, Poda G, Tetko IV (2012) Toxalerts: a web server of structural alerts for
toxic chemicals and compounds with potential adverse reactions

13
1996 S. Dara et al.

Szklarczyk D, Franceschini A, Wyder S, Forslund K, Heller D, Huerta-Cepas J, Simonovic M, Roth A, Santos

A, Tsafou KP et al (2015) String v10: protein-protein interaction networks, integrated over the tree of life.
Nucleic Acids Res 43(D1):D447–D452
Talele TT, Khedkar SA, Rigby AC (2010) Successful applications of computer aided drug discovery: moving
drugs from concept to the clinic. Curr Top Med Chem 10(1):127–141
Tan J, Hammond JH, Hogan DA, Greene Casey S (2016) Adage-based integration of publicly available pseu-
domonas aeruginosa gene expression data with denoising autoencoders illuminates microbe-host interac-
tions. MSystems 1(1)
Tasaki S, Suzuki K, Kassai Y, Takeshita M, Murota A, Kondo Y, Ando T, Nakayama Y, Okuzono Y, Takiguchi
M et al (2018) Multi-omics monitoring of drug response in rheumatoid arthritis in pursuit of molecular
remission. Nat Commun 9(1):1–12
Thomas U, Andreas M, Günter K, Marvin S, Wegner Jörg K, Hugo C, Sepp H (2014) Deep learning as an
opportunity in virtual screening. Proc Deep Learn Workshop NIPS 27:1–9
Tian S, Li Y, Wang J, Zhang J, Hou T (2011) Adme evaluation in drug discovery. 9. prediction of oral bioavail-
ability in humans based on molecular properties and structural fingerprints. Mol Pharm 8(3):841–851
Tishby N, Zaslavsky N (2015) Deep learning and the information bottleneck principle. In: 2015 IEEE Informa-
tion Theory Workshop (ITW). IEEE, pp 1–5
Tsherniak A, Vazquez F, Montgomery PG, Weir BA, Kryukov G, Cowley GS, Gill S, Harrington WF, Pantel S,
Krill-Burger JM et al (2017) Defining a cancer dependency map. Cell 170(3):564–576
Turkki R, Linder N, Kovanen PE, Pellinen T, Lundin J (2016) Antibody-supervised deep learning for quan-
tification of tumor-infiltrating immune cells in hematoxylin and eosin stained breast cancer samples. J
Pathol Inform 7
Turner JR (2010) New drug development: an introduction to clinical trials. Springer Science & Business Media,
Berlin
Vakser IA (2014) Protein-protein docking: From interaction to interactome. Biophys J 107(8):1785–1793
Valkov E, Sharpe T, Marsh M, Greive S, Hyvönen M (2011) Targeting protein–protein interactions and frag-
ment-based drug discovery. In: Fragment-Based Drug Discovery and X-Ray Crystallography. Springer,
pp 145–179
Valueva MV, Nagornov NN, Lyakhov PA, Valuev GV, Chervyakov NI (2020) Application of the residue num-
ber system to reduce hardware costs of the convolutional neural network implementation. Math Comput
Simul
Vamathevan J, Clark D, Czodrowski P, Dunham I, Ferran E, Lee G, Li B, Madabhushi A, Shah P, Spitzer M
et al (2019) Applications of machine learning in drug discovery and development. Nat Rev Drug Discov-
ery 18(6):463–477
van Gool AJ, Bietrix F, Caldenhoven E, Zatloukal K, Scherer A, Litton J-E, Meijer G, Blomberg N, Smith
A, Mons B et al (2017) Bridging the translational innovation gap through good biomarker practice.
Nat Rev Drug Discovery 16(9):587–588
Vaquero-Garcia J, Barrera A, Gazzara MR, Gonzalez-Vallinas J, Lahens NF, Hogenesch JB, Lynch KW, Barash
Y (2016) A new view of transcriptome complexity and regulation through the lens of local splicing vari-
ations. Elife 5:e11752
Veltri RW, Partin AW, Miller MC (2000) Quantitative nuclear grade (qng): A new image analysis-based bio-
marker of clinically relevant nuclear structure alterations. J Cell Biochem 79(S35):151–157
Venkatesan R, Li B (2017) Convolutional neural networks in visual computing: a concise guide. CRC Press,
London
Vinod CSS, Anad Hareendran S (2021) Artificial intelligence: a practitioner’s approach. PHI Learning Pvt Ltd,
Delhi
Vinod CSS, Anand Hareendran S (2021) Machine learning: a practitioner’s approach. PHI Learning Pvt Ltd,
Delhi
Visibelli A, Bongini P, Rossi A, Niccolai N, Bianchini M (2020) A deep attention network for predicting amino
acid signals in the formation of [formula: see text]-helices. J Bioinform Comput Biol:2050028
Vohora D, Singh G (2018) Pharmaceutical medicine and translational clinical research. Academic Press,
London
Volkamer A, Kuhn D, Grombacher T, Rippmann F, Rarey M (2012) Combining global and local measures for
structure-based druggability predictions. J Chem Inf Model 52(2):360–372
Voosen P (2017) The ai detectives
Vranic S, Shimada Y, Ichihara S, Kimata M, Wenting W, Tanaka T, Boland S, Tran L, Ichihara G (2019) Toxi-
cological evaluation of sio2 nanoparticles by zebrafish embryo toxicity test. Int J Mol Sci 20(4):882
Wang N-N, Dong J, Deng Y-H, Zhu M-F, Wen M, Yao Z-J, Ai-Ping L, Wang J-B, Cao D-S (2016) Adme prop-
erties evaluation in drug discovery: prediction of caco-2 cell permeability using a combination of nsga-ii
and boosting. J Chem Inf Model 56(4):763–773

13
Machine Learning in Drug Discovery: A Review 1997

Wang Q, Feng YH, Huang JC, Wang TJ, Cheng GQ (2017) A novel framework for the identification of drug
target proteins: Combining stacked auto-encoders with a biased support vector machine. PLoS ONE
12(4):e0176486
Wang D, Jin G (2018) Vasc: dimension reduction and visualization of single-cell rna-seq data by deep vari-
ational autoencoder. Genom Proteom Bioinform 16(5):320–331
Wang C, Kurgan L (2020) Survey of similarity-based prediction of drug-protein interactions. Curr Med Chem
27(35):5856–5886
Wang S, Sun S, Li Z, Zhang R, Jinbo X (2017) Accurate de novo prediction of protein contact map by ultra-
deep learning model. PLoS Comput Biol 13(1):e1005324
Wang W, Yang S, Zhang X, Li J (2014) Drug repositioning by integrating target information through a hetero-
geneous network model. Bioinformatics 30(20):2923–2930
Wang B, Zhu J, Pierson E, Ramazzotti D, Batzoglou S (2017) Visualization and analysis of single-cell rna-seq
data by kernel-based similarity learning. Nat Methods 14(4):414–416
Warmuth MK, Liao J, Rätsch G, Mathieson M, Putta S, Lemmen C (2003) Active learning with support vector
machines in the drug discovery process. J Chem Inf Comput Sci 43(2):667–673
Way GP, Greene CS (2017) Extracting a biologically relevant latent space from cancer transcriptomes with vari-
ational autoencoders. BioRxiv, p 174474
Webb AR (2003) Statistical pattern recognition. John Wiley & Sons, New Jersy
Willett P (2006) Similarity-based virtual screening using 2d fingerprints. Drug Discovery Today
11(23–24):1046–1053
Xia Z, Wu L-Y, Zhou X, Wong STC (2010) Semi-supervised drug-protein interaction prediction from heteroge-
neous biological spaces. In: BMC systems biology, vol 4. BioMed Central, pp 1–16
Xing J, Wenchao L, Liu R, Wang Y, Xie Y, Zhang H, Shi Z, Jiang H, Liu Y-C, Chen K et al (2017) Machine-
learning-assisted approach for discovering novel inhibitors targeting bromodomain-containing protein 4.
J Chem Inf Model 57(7):1677–1690
Xue LC, Dobbs D, Bonvin AMJJ, Honavar V (2015) Computational prediction of protein interfaces: A review
of data driven methods. FEBS Lett 589(23):3516–3526
Yamanishi Y, Araki M, Gutteridge A, Honda W, Kanehisa M (2008) Prediction of drug–target interaction net-
works from the integration of chemical and genomic spaces. Bioinformatics 24(13):i232–i240
Yavuz BÇ, Yurtay N, Ozkan O (2018) Prediction of protein secondary structure with clonal selection algo-
rithm and multilayer perceptron. IEEE Access 6:45256–45261
Youjun X, Pei J, Lai L (2017) Deep learning based regression and multiclass models for acute oral toxicity pre-
diction with automatic chemical feature extraction. J Chem Inf Model 57(11):2672–2685
Zaretzki J, Matlock M, Swamidass SJ (2013) Xenosite: accurately predicting cyp-mediated sites of metabolism
with neural networks. J Chem Inf Model 53(12):3373–3383
Zeng X, Zhu S, Weiqiang L, Liu Z, Huang J, Zhou Y, Fang J, Huang Y, Guo H, Li L et al (2020) Target identi-
fication among known drugs by deep learning from heterogeneous networks. Chem Sci 11(7):1775–1797
Zernov VV, Balakin KV, Ivaschenko AA, Savchuk NP, Pletnev IV (2003) Drug discovery using support vector
machines. the case studies of drug-likeness, agrochemical-likeness, and enzyme inhibition predictions. J
Chem Inf Comput Sci 43(6):2048–2056
Zhan F, Barlogie B, Mulligan G, Shaughnessy JD Jr, Bryant B (2008) High-risk myeloma: a gene expression-
based risk-stratification model for newly diagnosed multiple myeloma treated with high-dose therapy is
predictive of outcome in relapsed disease treated with single-agent bortezomib or high-dose dexametha-
sone. Blood J Am Soc Hematol 111(2):968–969
Zhan F, Huang Y, Colla S, Stewart JP, Hanamura I, Gupta S, Epstein J, Yaccoby S, Sawyer J, Burington B et al
(2006) The molecular classification of multiple myeloma. Blood 108(6):2020–2028
Zhang QC, Petrey D, Norel R, Honig BH (2010) Protein interface conservation across structure space. Proc Natl
Acad Sci 107(24):10896–10901
Zhavoronkov A, Ivanenkov YA, Aliper A, Veselov MS, Aladinskiy VA, Aladinskaya AV, Terentiev VA, Pol-
ykovskiy DA, Kuznetsov MD, Asadulaev A et al (2019) Deep learning enables rapid identification of
potent ddr1 kinase inhibitors. Nat Biotechnol 37(9):1038–1040
Zhou H, Gao M, Skolnick J (2015) Comprehensive prediction of drug-protein interactions and side effects for
the human proteome. Sci Rep 5(1):1–13
Zsoldos Z, Reid D, Simon A, Sadjad SB, Johnson AP (2007) ehits: a new fast, exhaustive flexible ligand dock-
ing system. J Mol Graph Model 26(1):198–212

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

13
1998 S. Dara et al.

Suresh Dara, currently working as Professor in B V Raju Institute of

Technology, Narsapur, Medak -502313, India. He has completed his
Ph.D. from Indian Institute of Technology (ISM), Dhanbad, India in
2015, completed his Masters in CST from Andhra University, India in
2008. He has published many research articles in reputed journals,
International conferences, Books and Book chapters. He served as
reviewer for many reputed journals. His current research interests are
Soft computing, Machine Learning, Deep Learning, Evolutionary
Computation, and Bioinformatics.

Swetha Dhamercherla, currently Post Graduate student and working

towards her academic Project. She did her Under Graduate in Com-
puter Science and Engineering from JNTU Hyderabad, India in 2019.
Her research interests include Machine Learning, Deep Learning, Evo-
lutionary Computation, and Bioinformatics.

Surender Singh Jadav, currently working as Scientist in Centre for

Molecular Cancer Research (CMCR), Hyderabad, India, and Visiting
Professor at Vishnu Institute of Pharmaceutical Education and
Research (VIPER), Narsapur, Medak-502313, Telangana. He has done
his Ph.D. from Birla Institute of Technology, Mesra, India in 2015. He
has published many research articles in reputed journals, International
conferences, Books and Book chapters. His current research interests
are Drug Discovery, Machine Learning, Deep Learning and
Bioinformatics.

13
Machine Learning in Drug Discovery: A Review 1999

CH Madhu Babu, currently working as Professor and HOD/CSE in B

V Raju Institute of Technology, Narsapur, Medak-502313, India. He
has completed his Ph.D. from Acharya Nagarjuna University, India in
2017. He has published many research articles in reputed journals,
International conferences, Books and Book chapters. His current
research interests are Software Engineering, Machine Learning, Deep
Learning and Bioinformatics.

Mohamed Jawed Ahsan has earned his B. Pharmacy and M. Phar-

macy (Pharmaceutical Chemistry) from Jamia Hamdard University,
New Delhi; and Ph.D. (Pharmaceutical Sciences) from NIMS Univer-
sity, Jaipur, India. Presently he is working as Professor and Head,
Department of Pharmaceutical Chemistry, Maharishi Arvind College
of Pharmacy, Jaipur. He has a remarkable research profile in the area
of Pharmaceutical and Medicinal Chemistry. His area of interest is in
design and discovery of novel molecules against cancer and
tuberculosis.

13