Application of Machine Learning in Drug Discovery and Development Lifecycle

International Journal of Medical, Pharmacy and Drug Research (IJMPD)
[Vol-6, Issue-6, Nov-Dec 2022]
ISSN: 2456-8015
https://dx.doi.org/10.22161/ijmpd.6.6.4
Peer-Reviewed Journal
Int. J. Med. Phar. Drug Re. 2022 16
Vol-6, Issue-6; Online Available at: https://www.aipublications.com/ijmpd/
Application of Machine Learning in Drug Discovery and
Development Lifecycle
Geerisha Jain
Department of Computer Science Engineering, Vellore Institute of Technology, Vellore, India
Received: 11 Oct 2022; Received in revised form: 11 Oct 2022; Accepted: 15 Nov 2022; Available online: 20 Nov 2022
©2022 The Author(s). Published by AI Publications. This is an open access article under the CC BY license
(https://creativecommons.org/licenses/by/4.0/)
Abstract— Machine learning and Artificial Intelligence have significantly advanced in recent years owing
to their potential to considerably increase the quality of life while reducing human workload. The paper
demonstrates how AI and ML are used in the drug development process to shorten and enhance the overall
timeline. It contains pertinent information on a variety of Machine Learning approaches and algorithms
that are used across the whole drug development process to speed up research, save expenses, and reduce
risks related to clinical trials. A range of QSAR analysis, hit finding, and de novo drug design applications
are used in the pharmaceutical industry to enhance decision-making. As technologies like high-throughput
screening and computation analysis of databases used for lead and target identification and development
create and integrate vast volumes of data, machine learning and deep learning have grown in importance.
It has also been emphasized how these cognitive models and tools may be used in lead creation,
optimization, and thorough virtual screening. In this paper, problem statements and the corresponding
state-of-the-art models have been considered for target validation, prognostic biomarkers, and digital
pathology. Machine Learning models play a vital role in the various operations related to clinical trials
embracing protocol optimization, participant management, data analysis and storage, clinical trial data
verification, and surveillance. Post-development drug monitoring and unique industrially prevalent ML
applications of pharmacovigilance have also been discussed. As a result, the goal of this study is to
investigate the machine learning and deep learning algorithms utilised across the drug development
lifecycle as well as the supporting techniques that have the potential to be useful.
Keywords—Machine Learning, Artificial Intelligence, Drug Discovery, Drug Development,
Pharmacovigilance
I. INTRODUCTION
Over the last ten years, machine learning (ML) has been
more popular in the area of medicine. Since the middle of
the 20th century, machine learning has been explored, but
recent advances in computing power, data accessibility,
cutting-edge techniques, and a broad variety of technical
expertise have expedited its use in healthcare.
Machine learning methods have been used by drug
companies since 1962. These techniques make it easier to
gather pertinent characteristics, which advances our
knowledge of complex biological systems. The
pharmaceutical industry is increasingly using many
prediction models to enhance the drug development
process. We can finally acquire answers to topics that
present a higher challenge to chemists, all thanks to the
algorithms used by various computational methodologies.
They aid chemists in accurately modeling, analyzing, and
forecasting a variety of biological responses with regard to
drug design. With the help of the annotated data, machine
learning algorithms learn intricate patterns to predict the
annotations of new test data sets [1]. Genome association,
protein function prediction, and other tasks involve the
application of machine learning. It helps in comprehending
a diverse array of drug features such as solubility, binding,
and target-related assays. Despite the positive results, it is
never easy to apply machine learning to the complex

Jain / Application of Machine Learning in Drug Discovery and Development Lifecycle
www.aipublications.com Page | 17
problem of drug development. Drug development, in
contrast to other areas, has unique challenges in choosing
an appropriate representation for the targets in a
medication, such as the molecules and their complexes that
are important to the drug's intended effect. The lack of
bio-activity descriptions is one of the biggest problems. It
is crucial to consider how to use the data at hand to
accomplish the desired result.
Therefore, determining the correct representation is always
the most difficult task. The training data is particularly
important for machine learning techniques. This is made
much more difficult by the fact that the data used to make
the majority of the forecasts is often inconsistent, noisy,
and imprecise. It may become much more challenging due
to the scarce and uneven data produced by the chemical
experiments that were conducted. Recently, computational
approaches to deal with these challenges have been
created. Drug discovery and development may be sped up
in different ways by increasing the use of machine learning
to bioactivity data.
Due to the significant time and financial commitment, the
process of developing new drugs is exceedingly
challenging. Finding a drug to combat a target often takes
a huge number of years and billions of dollars. Even then,
regardless of a great deal of effort, the success rate is
extremely low. There is a risk that many long-term
research endeavors may fail, wasting tremendous effort.
Bestseller drugs are those that are frequently prescribed for
common conditions like the flu, diabetes, high blood
pressure, asthma, cold, etc. They are quite successful in the
pharmaceutical sector and generate great annual revenues
and daily profits. However, if the drug exhibits any side
effects, it might also pose problems for the company.
Drugs typically face competition from less-priced
substitutes when their patents expire. As a result, finding
new drugs is a difficult and risky process that is constantly
driven by the potential good it could do for millions of
individuals suffering from various ailments. The life cycle
of drug development is shown in Figure 1
Fig 1. Drug Discovery and Development Life Cycle
1. The first stage is target discovery. We now select the
illness target upon which to focus drug development. The
target helps us better understand how parasite infection
affects genes, proteins, RNA, and other cellular
components.
2. Phase two entails verifying the accuracy of the intended
target. During this stage, the discovered target is verified
to confirm that the drug being developed addresses the
right problem.
3. Discovery of HITs is the third phase. In this step, we
synthesize and purify the intended target-interacting
chemical compounds. Chemists and assay developers work
together to test the chosen substances at this step.
4. The fourth stage is the Hit to lead transformation. This
phase involves finding prospective lead compounds from
the molecules discovered as part of High Throughput
Screening (HTS) in the previous stage.
5. The fifth stage is lead optimization. This phase is
designed to provide a better and safer scaffold by
minimizing structural alteration while eliminating the
undesirable effects of the current active analogs.
6. The stage 5 is pre-clinical studies. This comprises
identifying medications and comprehending drug
mechanisms for reasonable patients, as well as applying
biomarkers to increase the effectiveness of clinical trials. It
clarifies us on the disease's activity and allows for more
precise functional imaging of its response to the drug
created to treat it.
7. The following step, clinical trials, involves testing the
drug on human subjects. If the medicine achieves its
intended results, then the process is complete.
8. Post-development Monitoring and pharmacovigilance
are the process's last steps. Medical professionals may
clinically prescribe the drug after it has been evaluated and
given FDA approval. After that, the drug is put on the
market for consumer purchase, and it needs to be
monitored continuously.
It takes many years to successfully complete each of these
phases. Continuous research is being done to increase the
efficiency and speed of this procedure. This paper is aimed
at discussing and reviewing the various applications of
cognitive sciences and machine learning in order to drive
productive benefits for the drug development lifecycle in

the pharma industry. The rest of the essay is divided into
the following sections: With sections describing the ML
models used in each phase and the constraints they impose,
Section II sheds insight on the types of ML algorithms
employed in different phases of drug development. This
section is then followed by a conclusion and a list of
references.
II. ML ALGORITHMS USED IN VARIOUS
STAGES OF DRUG DEVELOPMENT
Drug development has considerably advanced because of
Machine Learning algorithms. Consequently, the use of
multiple ML algorithms in drug discovery has significantly
benefited pharmaceutical companies. ML algorithms have
been used to construct many models for predicting the
chemical, biological, and physical characteristics of
compounds used in drug discovery. Over the duration of
drug discovery, these trained models will become
invaluable. Machine learning has been put to use in the
pharmaceutical sector for a variety of purposes, such as
drug efficacy identification, drug-protein interaction
prediction, safety biomarker confirmation, and molecule
bioactivity enhancement. Several ML methods have seen
extensive usage in the pharmaceutical industry recently.
These include the support vector machine (SVM), random
forest (RF), and naïve bayesian (NB).
Figure 2 depicts the four main categories of machine
learning algorithms: supervised, unsupervised, semi-
supervised, and reinforcement learning. [2][3]. Input data
must be provided for supervised learning, along with the
expected results. During the training phase, it also looks
after delivering accuracy rate predictions. Before using the
method on new test data, the features, instances, and
models to be utilized must be established. Learning can be
stopped once performance reaches an acceptable level.
The supervised learning framework can be categorized as
either classification or regression problems. Any situation
where the output is a category falls under the classification
problem, for example, YES or NO. A real-valued output
for instance height, weight, etc. falls under the category of
the regression problem. Unsupervised algorithms, on the
other hand, don't need to be trained for the intended result.
They model the underlying distribution via an iterative
process, giving them the opportunity to understand the
data better. These problems are classified as association or
clustering problems. We aim to define the rules for
understanding the vast data by defining the inherent
groupings in the data in clustering and by doing the same
in the association. Moreover, semi-supervised learning
employs input data with just a subset of labels for training.
Many of these issues really occur often in the real world.
To solve these issues, researchers use both supervised and
unsupervised study methods. In reinforcement learning,
the observations derived from environmental interaction
are utilised. The reinforcement learning system repeatedly
takes up new information from the environment until risk
is reduced. To learn the behavior of the environment, it
makes use of a feedback signal called a reinforcement
signal.
Fig 2. Types of Machine Learning Algorithm

There are various machine learning algorithms[4][5], some
of the popular approaches are:
• Decision Trees: This model uses data about various
decisions and their respective potential outcomes to form a
tree-like graph. By eliminating the low-value branches,
pruning can help a tree function better. This minimizes
both the over-fitting and the tree's complexity.
• Naive Bayes Classification: These Bayes theorem-based
classifiers are typically used when the inputs have a large
dimensionality. In comparison to other, more complex
models, this one has the greatest result.
• K-means Clustering: This technique aids in grouping the
data, and K stands for the group number. It uses given
features to iteratively assign each data point to a group.
The collected data is then clustered in terms of shared
characteristics. K-means clustering returns the data's labels
and cluster axes' centres.
• Logistic Regression: Methods of statistical analysis used
to identify the significance of one or more independent
variables in a data collection. It is a way to describe data
that is used for prediction, with the goal of learning more
about the association between a binary variable and other
independent factors.
• Support Vector Machines: In order to maximize the
separation between classes, this technique aims to
categorize and model the training data into a decision
boundary. For cases when linear data separation is not an
option, the kernel function is used.
• Neural Networks: These are parameterized non-linear
algorithms that classify input data at each layer using a
multi-layer perceptron. The accuracy of the model is
determined by the perceptron' and hidden layers' numerical
values.
2.1. Drug Discovery
A useful classification of the literature review is made
possible by the application of ML at every step of drug
development, from target identification and validation
through hit discovery and hit-to-lead optimization through
pre-clinical trials. The drug design methodologies rely on
datasets that were created using various ML algorithms.
When ML algorithms are properly trained, verified, and
used across the drug development phase, they may
expedite error-prone, previously difficult procedures and
provide insightful findings. The majority of drug design
processes now incorporate ML approaches to cut down on
time and manual interference, hence leading to optimal
results and timelines.
Fig 3. Machine Learning Models used in various stages of the Drug Discovery Process
2.1.1 Target discovery
The first step in the target identification and
characterization procedure is to determine the function and
significance of a gene or protein that may be used as a
therapeutic target. After a target has been identified, the
molecular pathways, it is expected to effect may be
described. Effectiveness, safety, and compliance with
clinical and business needs are some of the qualities of a
good target.
Producing drugs (small molecules, peptides, antibodies, or
more advanced techniques like short RNAs or cell
therapies) that will change the disease state by modifying
the activity of a biological target is the main objective of
drug discovery.[6]. The selection of a target with a valid
therapeutic hypothesis, that is, that modifying the target
would modify the disease state, is still important before
beginning a drug development program, despite the recent
revival of phenotypic screens. Target identification and
prioritization is the process of choosing this target based

on the information available. The next stage is to
demonstrate the selected target's involvement in the illness
using ex vivo and in vivo models that are physiologically
relevant (target validation). Early target validation is
essential to concentrate on high-probability projects even
if clinical trials will eventually confirm the target.
The pharmaceutical sciences devote considerable attention
to the study of drug-target interactions (DTI) [7]. The
procedure of discovering new medicines is costly and
time-consuming. Therefore, the ability to anticipate drug-
target interactions is useful to biologists since it allows
them to focus their research. The first and most important
step in the drug development process is determining the
desired effects of the medicine. Medicable proteins that
play a role in illness make up the majority of these areas of
intervention. Drug-target interaction prediction is used to
find novel treatment approaches. Potential targets include
proteins with enzymes, ion channels, G protein-coupled
receptors (GPCRs), and nuclear receptors. Certain ligands
may alter the functioning of these groups. As a result,
studying the genomic space generated by these protein
classes enables us to precisely predict the likelihood of an
interaction. Drug discovery and drug repositioning for
novel targets are both possible using DTI. The three main
categories of DTI prediction tools are ligand-based,
docking-based, and chemogenomic strategies. Similarity
between ligands and target proteins is used as a predictor
of DTI in the Ligand-based Approach [8], [9]. The target
protein's three-dimensional structure may be used to
identify the probability of a pharmacological interaction.
The Docking-based approach is used for this, which
considers the relative stability and binding affinity of the
proteins [10, 11]. If the drug's chemical information,
protein genomic data, and known DTIs are all considered,
the chemogenomic technique is then used.[12][13]. In a
ligand-based technique, a target with a small number of
binding ligands frequently yields subpar DTI predictions.
This is a shortcoming of this approach. Similarly, the
docking-based approach is time-consuming and depends
on the target proteins' 3D structures. Due to these
drawbacks, the chemogenomic technique has recently
gained popularity for the identification of DTI. The DTI
problem is presented as a machine learning problem using
this method, and a classifier is often created and trained
using publicly accessible interaction data. In order to
forecast the unknown interactions, this classifier is
used[13]. The chemogenomic fully utilized a number of
techniques. Bipartite graphs[14], recommendation
systems[15], and supervised classification problems are
some of these[16]. However, when we look at the data, we
can see that there will only be a small number of favorable
interactions, and the other possible interactions are
unknown. For instance, there could only be 7000 favorable
drug interactions out of the 35 million potential drug
options[17]. The two types of computational
chemogenomic methodologies are feature-based and
similarity-based strategies. Features are the inputs for a set
of instances defined by a specific class label for feature-
based methods. In most cases, the targets are the features
and the instances are the drugs. The presence of a possible
associations is represented by the binary value of the class
label. Support vector machines, decision trees, and random
forests are a few examples of feature-based classification
techniques.[18]. Drug-target interactions are often
identified using Support Vector Machine or Random
Forest. [19].
Particularly, the target identification portion of the drug
development process significantly relies on the
categorization of biomedical data. The classification of
biomedical data, which is sometimes replete with
irrelevant information and data known as noise, has shown
excellent potential when using Naïve Bayes classifier (NB)
algorithms [20]. Lead discovery might be considerably
enhanced by applying NB approaches to predict ligand-
target interactions.[21]. In recent years, researchers have
been able to use NB strategies in many areas of the drug
development process. In a research aimed at finding new
breast cancer therapies, Pang et al. [22] employed NB
models and other methods to categorize compounds
according on their potential efficacy as estrogen receptor
antagonists. "The model produced impressive results when
used in conjunction with other techniques, such as the
extended-connectivity fingerprint-6. In a study by Wei et
al. [23], potential drugs that would be effective against the
targets of the hepatitis C virus and human
immunodeficiency virus type 1 were predicted using a mix
of NB and support vector machine methods. Their
approach included two distinct descriptor systems, one of
which was the extended-connectivity fingerprint-6, with
NB as a classifier technique. Utilizing NB in conjunction
with other approaches and technologies has proven
effective in implementing drug discovery processes.
2.1.2 Target validation
The concept of creating a medicine for a certain target is
also an important consideration for the pharmaceutical
companies. For instance, identifying targets with
properties that imply that these proteins can bind tiny
molecules is necessary for small-molecule drugs [24].
These druggable models can be created using various
target attributes. Using the physicochemical, structural,
and geometric characteristics of 1,187 drug-binding and 99
non-drug-binding cavities in a sample of 99 proteins,
Nayal and Honig [25] created a random forest classifier.
The most important characteristics were the size and shape

of the surface voids. On the basis of the protein sequences
of well-known drug and non-drug targets, some studies
have used SVMs [26][27] or biassed SVMs with stacked
autoencoders, a Deep Learning model [28] to forecast
druggable targets to assess different physicochemical
properties. Additionally, it has been discovered that
druggable proteins tend to be strongly linked and occupy
certain areas of protein-protein interaction
networks[29][30][31]. These ML methods reduced the
search area by generating lists of drug-binding targets, but
further research is needed to confirm these forecasts.
The holy grail of target identification or validation, namely
the ability to accurately anticipate the outcome of a drug's
clinical trials in advance, has not yet been attained.
Success indicators have been the subject of several non-
ML studies. Rouillard et al.[32] evaluated 332 targets that
were either successful or unsuccessful in phase III clinical
trials by analyzing their omics data using ML and selecting
multivariate characteristics. The gene expression data,
which was characterized by low mean RNA expression
and considerable heterogeneity across tissues, was shown
to be a strong predictor of effective targeting. This work
provided more evidence that optimal targets are expressed
selectively in diseased tissues [33]. In order to anticipate
de novo therapeutic targets, Ferrero et al. [34][35] trained a
variety of ML classifiers utilizing target-disease linkages
from the open target's platform. It was determined that
regardless of indication, the three most essential data
categories for therapeutic target prediction are gene
expression, genetic data, and the availability of an animal
model. . The sparseness of the data and the lack of
knowledge regarding the causes of failed programs,
however, pose a limitation to this technique.
Fundamentally, because it takes years to build a good drug
discovery plan and finally bring it to market, successful
programs are a reflection of earlier drug development
models. Considering the arrival of more recent therapeutic
modalities like biologics (including antibodies), it's
doubtful that the elements that contributed to the success
of small-molecule projects in the past will be relevant in
the present. Additional restrictions are imposed by
precision medicine's growing importance. For future
prediction tools to be effective, large amounts of
information on both successful and failed drug
development projects must be made accessible with
metadata in the public domain.
2.1.3 Hit Discovery
It is essential to execute comprehensive virtual and
experimental high-throughput screening of large chemical
pools to identify treatment candidates that inhibit or
activate the target protein of interest. The
pharmacodynamic, pharmacokinetic, and toxicological
characteristics of candidate structures are further
improved, as well as their target specificity and selectivity.
It is important to note, however, that there may be a lack of
enough high-quality data in this domain, which might limit
the use of ML to new chemistry. This is especially true for
macrocycles and proteolysis-targeting chimaeras
(PROTACs).
For ligand-based virtual screening, a lot of attention has
been paid to the use of Deep Learning models, such as
multi-task neural networks. Computational methods may
use a particular lead molecule to identify physically
comparable compounds with similar chemical
characteristics. The use of multi-task DNNs has shown to
be more effective than standard statistical approaches for
this job, which was previously performed. When it comes
to predicting the properties and functionalities of small
substances, DNNs may greatly increase predictive power
[37]. One-shot learning may significantly reduce the
amount of information needed to accurately forecast how a
molecule would read out in a new experimental
environment. The binding mechanism of opiates to the -
opioid receptor was previously unknown, however a
Markov State model and Machine Learning approaches
were able to pinpoint an allosteric region implicated in this
activation [38]. The advantages of multi-task models over
single-task models, however, vary depending on the data
set. The evaluation of ML algorithms has made use of
MoleculeNet [39], a large benchmarking data set produced
by Pande et al. to assist in the comparison of different ML
algorithms. Data on the characteristics of more than
700,000 molecules can be found in MoleculeNet. The
open-source DeepChem package now includes all of these
hand-picked data sets along with a number of additional
number of advantages.
Planning effective chemical synthesis routes can also be
done using DNNs and contemporary tree search
techniques. A target molecule is formally deconstructed
using reversed processes in order to plan its production
(retrosynthesis). In order to synthesise the target, this
method creates a sequence of processes that may be carried
out in a straightforward way in the laboratory. The
systematic application of synthetic chemistry skills to this
method is a tremendous task. The exponential growth of
chemical knowledge and the inadequate understanding of
the range and boundaries of many reactions have made the
manual insertion of transformation rules impracticable. A
database called Reaxys (with 11 million reactions and
300,000 rules) was utilised by Segler et al.[40] to
automatically extract the rules. He used a Monte Carlo tree
search (MCTS) to weight the tree's nodes and DNNs in
order to determine the most profitable paths for further
research. This strategy outperforms the industry norm for

best first search in quantitative analyses using two
different implementations (heuristic method and neural).
Furthermore, for around two-thirds of the examined
chemicals, MCTS is 30 times quicker than conventional
computer-aided search methods. In a double-blind
experiment, qualitative evaluations were also incorporated.
Organic chemists were required to pick between expected
and literature-based synthesis paths in a blind process. For
the first time, chemists agreed that the predicted routes'
quality was, on average, on par with routes selected from
the literature.
2.1.4 Hit to Lead
This procedure is sometimes referred to as ''lead
generation'' in the early phases of drug development
research. Insufficient optimization during the High
Throughput Screen (HTS) to find potential lead
compounds leads to the discovery of molecules, or ''hits.''
Using a preexisting kinase inhibitor library, the ''design
layer''/Random Forest regression mapping method is used
to construct new chemical spaces with biological activity.
This method of optimising hits into leads is a practical use
of chemical synthesis. [41]
By adjusting or rebalancing the target interest, de novo
drug design produced distinct chemical structures [42]. By
starting from scratch, de novo techniques introduce new
molecules using a fragment-based methodology. If the
molecular structure has impracticalities and complexity at
this stage [43], the risk emerges in the structure's
development and the evaluation of bioactivity becomes
challenging. In order to develop a novel structure with the
necessary properties, deep learning models could be used
in terms of their extensive knowledge and generative skills
[44]. The use of reinforcement learning in molecular de
novo design is another significant application of Machine
Learning. By modifying a sequence-based generative
model to produce molecules with almost ideal values for
solubility, pharmacokinetic characteristics, bioactivity, and
other factors, researchers at AstraZeneca were able to
expand the chemical space. Similar models were created
by Kadurin et al. utilizing deep GANs to extract chemical
features from very huge data sets [45]. It's important to
keep in mind, however, that reinforcement learning may
not be helpful when trying to find previously undiscovered
synthetic pathways.
Olivecrona et al’s[46] expansion of the use of deep
reinforcement learning to the prediction of biological
activities in the creation of new drugs included some RNN
model modifications. To understand the SMILES syntax,
an RNN model must be trained; chemBL compounds may
be gathered for training. Agents take engage in activities in
reinforcement learning under certain rules. If the agent is
rewarded enough, the trend of their actions will be revived.
[47]. Use the SVM methodology to improve a few
methods based on the ligands concept in the training set in
order to achieve a high benefit for activity scoring. Create
a few compounds that are antagonistic to the dopamine
receptor 2-type before using the RNN and deep
reinforcement learning model. Additionally, it was noted
that with SVM's scoring capability, predictions for
structures in the bioactive region have exceeded 95%. The
auto-encoders method can be used to produce unique
molecules by employing deep learning algorithms. Then,
Gomez-Bombarelli et al. [48] used the multilayer
perceptron (MLP) and variational autoencoder (VAE) to
automatically produce new molecules with the required
characteristics.
Kadurin et al. [49] used on the AAE model, now known as
druGAN, to create molecular fingerprints. The AAE
approach produced impressive results when applied to the
VAE model in terms of power production, reconstruction
inaccuracy, and subsequent extraction effectiveness. Coley
et al. [50] proposed analysing the synthetic molecule to
determine whether it was accessible synthetically. As a
result of the great approximation capabilities for producing
synthetic complexity measures, he postulated that the
neural network was trained in line with the response
database. The product complexity score must be higher
than the reactant complexity score for a synthetic reaction
to be successful. [51]. In order to demonstrate correlation
inequality between the complexity of the products and
reactants, Coley made several efforts to develop a scoring
function by encoding chemical responses into pairs of
products and reactants. In order for neural networks to feel
at ease with any kind of scoring capability at that time,
they must be trained using the reactant and product
pairings that Coley utilised across a scope of 22 million.
Additionally, the synthesis process's conclusion was
established with a great deal of complexity. Finally,
generative models reveal both the complexity of the
synthetic process caused by eliminating the non-realistic
molecules and the pharmacological actions in inverse
synthetic planning.
How to adequately explain the chemical structure is a
challenge in small molecule design. There are several
representations to select from, ranging from simple
circular fingerprints like the extended-connectivity
fingerprint (ECFP) to intricate symmetry functions [52].
Which structural representation is best for every small-
molecule design challenge is yet unclear. It would be
fascinating to see whether the extensive body of ML
research in cheminformatics provides any new information
on the most efficient approach for structural
representation.

2.1.5 Lead Optimization
The optimization of potential drug leads is an important
step in the drug development process. If a fragment has the
potential to be used in medicinal chemistry, it will be
evaluated as a potential future step in the research process.
Lead optimization aims to offer a better and safer scaffold
by reducing structural modification and removing negative
effects of existing active analogues. An illustration of this
is the advancement of Autotoxin inhibitors, such as the
investigational drug GLPG1690, in human clinical trials to
treat pulmonary fibrosis. Figure 4 below provides an
overview of the factors that can make active analogs more
potent by using customized methods. Here, we evaluate a
substance's ADME/T features, including its toxicity,
chemical makeup, physical attributes, and rates of
absorption, distribution, metabolism, and excretion.
Fig 4. Factors Affecting Lead Optimization
• Chemical and physical properties
Physical and chemical properties have been used in the
drug development process to lessen the number of
significant failure. To this end, researchers have turned to
lead optimization strategies powered by deep learning
models [53]. Due to their dependence on the
interpretability principle, Duvenaud et al. [54] directly
gathered data from the molecular graph using the CNN-
ANN idea to generate a prediction, i.e. (MAE =
0.53+0.07). Duvenaud's study was motivated by Coley et
al efforts’ to improve molecular aqueous approaches.
Additionally, the tensor-based convolutional approach was
used, and the improved results were MAE (0.424+0.005).
Clearly defining molecular graph attribution is crucial
since tensor-based approaches must incorporate properties
like bonds and atom levels. To predict molecular aqueous
solutions, Coley's model utilised a lot more atom-level
information than Duvenaud's [55]. It was shown that Caco-
2 permeability coefficients had a good correlation with
oral drug absorption (P app) for predicting the candidate
drug while pharmacokinetic parameters were being
evaluated.[56, 57]. Using the Caco-2 permeability data,
Wang et al’s [58] attempt to generate 30 descriptors'
worth of prediction templates necessitated the building of
1,272 components, including models like SVM regression
and boosting. In the test set, where it also had the
maximum expectation capacity, the boosting model fared
the best. It conforms with the OECD's (Organization for
Economic Co-operation and Development) standards for
promoting reliability and logical arguments since it
adheres to the QSAR principles set out by the OECD.
• Absorption, distribution, metabolism, and excretion
Injecting pharmaceuticals or treatments into a person's
veins is a method of absorption. Bioavailability parameter
is used to examine the level of absorptions. Several clinical
departments explained how to increase absorption
properties using molecular predictions for bioavailability
[59]. Tian et al. used 1,014 compounds to predict
bioavailability using molecular resources and structural
fingerprints using the MLR model. The predicted
performance of applying the genetic function approach
was excellent, with RMSE = 0.2355 and a correlation
value of 0.71. The distribution of medications or
treatments within the human body is influenced by
intracellular and interstitial fluids as well as specific drug
absorption characteristics.[60]. The steady state
distribution of a drug is the amount of drug that makes it
from the in vivo phase to the plasma reaction (VDss). The
steady phase is a crucial indicator for evaluating the drug
distribution mechanism. Lombardo and Jing used 1,096
molecules and the PLS and Random Forest methodologies
to make predictions about VDss. [61]. The board members
in this case are dissatisfied with the prediction findings
since 50% of the compounds are accessible in a twofold
mistake. Because of the presence of such obscure
components, VDss may be affected. The use of VDss in
molecular structure data is intentionally put to the test by
this issue. Any drug or treatment taken by a person under
these circumstances will try to produce the already-
existing toxic metabolite as a result of the metabolic
system's inbuilt redundancies. It is important to maintain
the integrity of the metabolic structure, hence structural
optimization methodologies are utilised to motivate the
metabolism to make very accurate prediction. Numerous
machine learning (ML) methods were used to predict
specific metabolic enzymes, such as UDP-
glucuronosyltransferases (UGTs), cytochrome P450s, etc.,
using a vast quantity of drug metabolism data. In addition,
Xenosite's platform has UGT-trained neural networks for
predicting UGT metabolism [62][63][64]. When a drug is
digested, it is eliminated from the body in a process known
as excretion. Because certain medications are soluble in
water, water may be used to flush them out of the body, or
in the absence of metabolism, the metabolites can be
eliminated directly. The PCA method was utilised by

Lombardo et al. to get excellent results in innovative
approaches, with a predicted rate of accuracy of 84%.[65].
• Toxicity and the ADME/T multi-task neural networks
In clinical and preclinical damage completion, about one-
third of the most important compounds utilised in drug
localization were shown to be inadequate. Risks were
reduced by improved toxicity prediction and molecular
optimization [66]. Kidney and liver toxicity profiles are
among those that may be predicted using tools like
structural warnings and rule-based expert knowledge. To
improve the accuracy of toxicity predictions, deep learning
models are required. Similarly, Xu et al., anticipating
results from CNN molecular graph encoding, created the
acute-oral toxicity prediction model (MGE-CNN). When
compared to the SVM model, predicted results were
shown to be better [67]. The similarity in training neural
networks feature extraction, model construction, and
molecule encoding resulted in the success of the MGE-
CNN model. Due to the adaptability of the MGE-CNN
model, the issue was reformulated in terms of molecular
fingerprints. To categorise TOX Alerts and collect high-
quality data on structural alerts, Xu et al. [68] employed
hazardous characteristics for fingerprints. When
comparing parameters, multi-task neural networks that
have been trained to retrieve comparable characteristics
outperform single-task neural networks. [69]. This is
because the neural networks are more supportive of
multiple tasks and share parameters. The human body
receives data after the drug's absorption, distribution,
metabolism, and excretion have all been taken care of and
prediction has been enhanced using multi-tasking neural
networks. In this study, Kearnes et al. examined single-
task and multi-task performance using ADME/T
experimental data. The results demonstrated that the multi-
task approach was superior. [70].
2.1.6 Pre-Clinical Studies
Through the use of ML models, biomarker discovery
increases the effectiveness of clinical trials by identifying
drugs and understanding how they work for reasonable
people [71][72][73]. The completion of a clinical trial
requires a lot of money and time. Throughout the first
stages of clinical trials, expected models must be used,
developed, and validated in order to solve this problem. In
preclinical data collection, ML systems enable the
prediction of translational biomarkers. Following data
validation, corresponding biomarkers and models may
examine the patient's symptoms and provide a treatment
strategy. Although many scholarly articles proposed
predictive models and biomarkers, only a few of those
articles were actually implemented in clinical trials. For a
clinical situation, it is required to consider model
development, design, data access, data quality, software,
and model selection. The main problem was how ML
methods evaluated the effectiveness of community-driven
efforts to create regression and classification models. The
US Food and Drug Administration-led (MAQC II)
MicroArray Quality Control [74] analyzed ML algorithms
for predicting gene expression data in the last step of
clinical trials a number of years ago. 36 independent
organizations that examined 6 microarray data sets created
predictive models for categorizing clinical locations
nearing completion of development. Information simulates
the best methods for clinical trials by including high-
quality data, trained scientists, and control systems.
Patients with multiple myeloma had poor prognosis and
their treatment was stopped after 24 months owing to an
incomplete application. Multiple myeloma and gene
expression are continuous variables, hence their future
behavior may be predicted using a regression-based
method. A gene expression profile may be utilized in
combination with Cox regression models to identify a
patient's illness risk factors, as has been shown [75]. Here,
the advantage of using regression models is emphasized
due to the lack of specified classes that might perform
prediction in clinical trials. [76][77][78][79]. The National
Cancer Institute (NCI) finds it challenging to develop
medicine prediction models in order to assess regression
models. [80]. The best model with key parameters must be
used to acquire training data (for instance, treating 35
breast cancer cells with 31 medicines), and models must be
validated using identical blind testing data (i.e., treating 18
breast tumour cells with similar 31 drugs). Using data from
six different data profiles—RNA sequencing, RNA
microarray, reverse protein phase array, SNP (Single
Nucleotide Polymorphism) array, DNA methylation status,
and exome sequencing—better prediction algorithms are
created". These profiles are used to conduct multivariate
statistical analyses on 44 sets of data using a variety of
regression models, including sparse linear regression,
kernel methods, regression trees, and principal component
analysis. The MAQC II findings showed that certain
groups performed very well, while other groups utilized
similar models. While some teams concentrated on
technical issues like feature selection, quality control, data
reduction, modifying ML parameters, and splitting
strategy, others utilized biological information like gene
expression data to set themselves apart from the
competition. A huge number of medications are feasible
for creating a prediction model when compared to other
approaches.
2.2. Clinical Trials
Proper Clinical drug development follows the completion
of preclinical research and includes investigations with

human volunteers and clinical trials to further perfect the
drug for human consumption. The intricacy of clinical trial
design, the cost of conducting such a study, and the
difficulties inherent in putting it into practise are all factors
that may affect trials performed at this level. Trials must
be safe and effective, done within the budget given for
drug development, and adhere to a certain methodology to
guarantee the medicine is useful and practical for its
intended application. For this rigorous process to be
successful, it needs to be properly set up and involve a
significant volunteer base. In order to successfully carry
out these various tasks involved in clinical trials as shown
in Figure 5; ML algorithms have been extensively used in
each sphere thereby aiding the process as a whole.
Fig. 5 Various Tasks associated with Clinical Trials
2.2.1 Clinical study protocol optimization
The success and effectiveness of human clinical trials may
be improved by using ML to ease the formulation of trial
protocols in advance using simulation techniques on a
significant quantity of data from prior studies. As shown in
reinforcement learning approaches for Alzheimer's disease
and non-small cell lung cancer [81, 82], study simulation,
for instance, may optimize the selection of treatment
regimens for trials. Researchers may submit protocols
using AI, which employs natural language processing, in
order to detect potential roadblocks and obstacles to
successfully completing trials (such as inclusion/exclusion
criteria or outcome indicators). Although the use of ML in
research planning may theoretically ensure that a particular
trial design is best suited to the needs of the stakeholders,
this is just a promise since the effectiveness of these
sample models has not been assessed in a peer-reviewed
manner. In conclusion, machine learning clearly has the
potential to improve the effectiveness and productivity of
preclinical research and the planning of clinical trials.
However, rather than focusing on the planning of clinical
trials, the great majority of peer-reviewed studies on the
use of ML in this context are preclinical research and
development-focused. This could be because there are
more large, high-dimensional datasets available in
translational contexts or because using ML in clinical trial
settings comes with higher costs, hazards, and regulatory
requirements. We require scholarly research on the
effectiveness of ML in clinical trial design to solve these
challenges.
2.2.2 Clinical trial participant management
Clinical trial participant management involves selecting
research populations, enlisting patients, and maintaining
their participation. Despite significant investment in
participant management, studies often run over budget,
take longer than expected, or fail to provide useful data
due to patient drop-out and non-adherence. There is a total
failure rate of 13.8 percent for medications evaluated in
phase I, with estimates indicating that between 33.6
percent and 52.4 percent of clinical studies underpinning
drug development that take place during stages 1-3 are
unsuccessful.. [83]. ML techniques can help with
participant identification recruiting and retention
and choosing the study demographics of patients. If
individuals were more carefully selected for trials, the
sample size required to detect an impact may be less.
Alternatively stated, improved techniques of selecting the
patient population may lead to fewer individuals being
offered therapies for which they are not likely to improve
outcomes. Previous studies have indicated that for every
expected response, there are anywhere from three to
twenty-four non-responders for the most commonly
prescribed drugs, making progress in this field a continual
challenge. Many people who use these drugs end up
having unintended consequences [84]. Unsupervised
machine learning of patient populations may quickly
analyze large databases of existing research and in turn aid
with patient population selection as well as reveal patterns
in patient features that may be utilized to choose patient
phenotypes that are best suited for treatment [85]. A cross-
modal inference learning model technique may more
successfully match patients to trials using EHR data by
concurrently encoding enrollment criteria (text) and patient
records (tabular data) into a shared latent space [86]. The
utility of these procedures is questioned by the lack of
peer-reviewed documentation of their development and
performance measures [87]. Mendel.AI and Deep6AI are
two businesses that provide comparable services. This
method may have the benefit of not requiring participants
to be identified precisely by structured data fields, which
has been demonstrated to dramatically skew trial
populations. [88, 89]. There are two basic strategies to
boost retention and policy adherence using ML models, as

shown by the monitoring of participants and their
adherence to protocols. The first stage is to use ML to
identify, investigate, and penalise participants who are
likely to infringe upon the terms of the research. The
second approach is to use ML to make the research easier
for participants and to improve their overall experiences.
AiCure is a company that employs face recognition
technology to track whether or not patients really take their
prescription. AiCure was shown to be more efficient than
directly observed modified therapy in detecting and
improving patient adherence in studies on schizophrenia
patients and recent stroke survivors using anticoagulation
[90, 91]. AiCure's performance may differ across patient
subgroups since its model building and validation
technique is not publicly known, as has been shown in
previous computer vision applications. [92]. Additionally,
data obtained during routine clinical care might be
analyzed using ML techniques to provide data that can be
utilized for study. For instance, rather of exposing all
participants to the additional strain and cost of more in-
depth and multiplexed imaging, generative adversarial
network modeling of typically clinically stained slides
with hematoxylin and eosin may identify the ones who
need it. [93]. Natural Language Processing may also make
it simpler to repurpose clinical data for research purposes
by automatically filling out study case report forms when
used often with the Unified Medical Language System
[94]. There are two examples of how patients produce
useful content outside of the clinical trial context that ML
can process into study data to lessen the burden of data
collection for trial participants: natural language
processing of social media posts to identify serious drug
reactions with high fidelity [95]. The International
Parkinson and Movement Disorders Society's Unified
Parkinson's Disease Rating Scale has been found to
correlate participant activity with wearable device data,
which can also be used to distinguish between
neuropsychiatric symptom ontology patterns, identify
patient falls, and identify participant activity [96]. In
summary, ML and NLP have shown promise for a number
of tasks related to improved participant management in
clinical trials; nevertheless, additional research comparing
various approaches to participant management is required
to further improve clinical trial quality and participant
experience.
2.2.3 Data collection and management
Applying ML to clinical trials has the potential to enhance
the methods used to gather, handle, and analyse trial data.
ML techniques can also aid in addressing some of the
challenges related to collecting real-world data and dealing
with corresponding missing data. Wearable and other
mobile/electronic device data on patients' health may
supplement or even replace data collected via more
conventional means, such as in-person visits for a research.
The usage and validation of new, patient-centered
biomarkers may be made possible by wearables and other
devices. When creating new ''digital biomarkers'' from the
data acquired by the device's numerous sensors, ML
processing is often necessary since the data provided by
mobile devices might be sparse and inconsistent in quality,
accessibility, and synchronisation (such as cameras, audio
recorders, accelerometers, and
photoplethysmography).Therefore, in order to analyse the
massive and complicated data created by wearables and
other devices, appropriate data collecting, storage,
validation, and analysis procedures are required [97].
Patients with atopic dermatitis had their accelerometer data
processed using a recurrent neural network [98], a mobile
single-lead electrocardiogram platform's input was
processed using a deep neural network, and an audio signal
from a Parkinson's disease patient was processed using a
random forest model. [99]. These cutting-edge digital
biomarkers might make clinical studies run more smoothly
and with a focus on patients, but there are risks associated
with this strategy. Although this risk exists for all data,
regardless of processing technique, using machine learning
to evaluate wearable sensor output to define research goals
involves the possibility of producing false results, as was
shown to happen with an electrocardiogram classification
model[100]. Lack of awareness of participant privacy
attitudes in relation to the sharing and use of device data,
as well as a lack of a precise description of the overlap
between authorised clinical aims and patient-centric digital
biomarkers, are obstacles to ML processing of device data
implementation.
2.2.4 Study data collection, verification, and
surveillance
An intriguing use of ML is in automating data collecting
into case report forms, which may save time, money, and
human error in either prospective trials or retrospective
evaluations. Specifically, Natural Language Processing is
very important for this kind of data administration.
Depression [101], epilepsy [102], and cancer [103] are just
a few examples of diseases where this application has
showed early promise despite having to overcome varied
data formats and provenances. Regardless of the method
used for data collection, ML might support risk-based
monitoring techniques for clinical trial surveillance. This
allows for the avoidance or early detection of issues like
site failure, fraud, and inconsistent or nonsensical data that
might otherwise delay database lock and subsequent
analysis. For instance, when people fill out case report
forms, the accuracy of the information acquired for result
determination may be evaluated by combining optical

character recognition with natural language processing
(usually supplied in PDF form). Clinical trials and
observational studies may benefit from auto-encoders
since they can be used to identify potentially fraudulent
data patterns by classifying them as plausible or
improbable [104]. Endpoint detection, adjudication, and
safety signal detection are all examples of how machine
learning may be used in data processing. Currently, events
are manually adjudicated by a committee of doctors.
However, there may be time, money, and complexity
savings with semi-automated endpoint identification and
adjudication. While categorising events into useful
categories has typically been the domain of semi-
automated ML systems, adjudicating endpoints has
historically required a significant amount of human labor.
Although this technique has not been peer-reviewed,
IQVIA Inc. has described the capacity to automatically
treat certain adverse events connected to pharmacological
therapy utilizing a mix of optical character recognition and
natural language processing [105]. A classification model
would theoretically need to be retrained for each new
experiment due to the fact that endpoint criteria and the
data needed to support them often alter across research.
This might be a roadblock in the way of fully automated
event adjudication (which is not a viable approach).
Although not all studies adhere to these objectives, there
have been recent attempts to standardize outcomes in the
area of cardiovascular research. The majority of areas have
not combined trial data to enable model training for
cardiovascular endpoints [106]. For this area to go further,
stakeholders must establish consensus definitions, really
accept the definitions of events, and be prepared to provide
the right data from several trials for model training.
The issue of missing data may be solved using different
ML applications. This may be accomplished by thinking
about the data's context, the assumptions and objectives
made about the data, the methods used to acquire the data,
and the analyses that will be conducted. Goals could
include computing other important quantities by averaging
over a large number of potential values from a learning
distribution or directly calculating precise estimates of the
missing covariate values. Though more modern
approaches are still in their infancy and thorough
comparisons are needed, preliminary studies show that
complex ML methods may not always be superior than
simple imputation strategies like the population mean
estimate. [107]. One use of missing value algorithms is the
analysis of sparse datasets like those found in registries,
electronic health records, ergonomic studies, and data
collected from wearable devices. [108][109]. Data
augmentation solutions may mitigate the effects of missing
data or values, but they should be used with caution lest
they lead to models that are only partly generalizable to
newly collected data that has inherent flaws. Therefore,
using ML to enhance data gathering while conducting
research itself might be a more fruitful approach.
2.2.5 Data analysis
Rich sources of information for study design, risk
modeling, and counterfactual simulation include data
collected in clinical trials, registries, and clinical practices.
These projects are ideally suited for machine learning.
Unsupervised learning, for instance, might find phenotypic
clusters in real-world data that can be explored further in
clinical research [110]. Additionally, ML has the potential
to advance the established practice of secondary trial
analysis by more accurately identifying treatment
heterogeneity while still providing some (albeit
insufficient) protection against false-positive findings,
thereby revealing more intriguing areas for further research
[111]. Additionally, machine learning may provide risk
predictions that may be evaluated in the future with the
proper utilization of previous data. For instance, a random
forest model in the COMPANION trial data performed
better at identifying individuals who might benefit from
cardiac resynchronization treatment than a multiple
logistic regression [112]. The results demonstrated that
random forests may explain feature interactions that are
often missed by simpler models.
ML shows considerable promise in this area by increasing
the precision with which it can distinguish real-world
evidence from real-world data, even if it is still a highly
desired (and extremely difficult) objective (i.e., draw
causal inferences). A vital and important endeavor is the
creation of predictive models that can predict future
occurrences. A few of the methods suggested in the
literature include optimal discriminant analysis, targeted
maximum likelihood estimate, and propensity score
weighting made possible by ML [113][114].
The use of ML to provide counterfactual policy estimates,
where existing data is used to anticipate outcomes under
circumstances that do not now exist or may not, is
particularly fascinating. For instance, reinforcement
learning suggests better treatment plans based on prior
unsuccessful treatments and outcomes, and trees of
predictors may provide survival predictions for heart
failure patients under the conditions of obtaining or not
receiving a heart transplant. [115]. Risky data sharing
agreements that restrict the amount of data accessible for
model training and a lack of compliance with EHR data
systems are the key obstacles to adoption. [116]. In
conclusion, there are many efficient ML algorithms for
managing, processing, and analyzing data from clinical
trials, but there are much less methods for enhancing data

quality from the start. High-quality trials must be
conducted in order to enable more advanced ML
processing since the availability and quality of data are the
foundations of ML techniques.
2.3 Post Drug Development Sector and
Pharmacovigilance
Once the results of clinical studies have been compiled and
the treatment has been developed to achieve maximal
effectiveness and safety, the FDA will move it forward for
comprehensive assessment. Currently, the FDA examines
the drug application that the pharmaceutical company has
submitted and decides whether to approve it or not. Once
the pharmaceutical company has received permission, it
can start selling drugs and continue to manage its products.
There is a completely different sector or area of
technology, processes, and advanced improvements that
open up once the drug hits the market and is ready for use.
Figure 6 below shows some of the practical illustrations of
how companies can and have applied AI and ML
technologies to the post-drug development and
pharmacovigilance arena.
Fig 6. Machine Learning Applications in
Pharmacovigilance
• Predictive Analytics
ML aids in making predictions based on that analysis
while AI aids in managing enormous amounts of data. The
time it takes for new drugs to reach the market has been
cut in half with the help of AI and ML. Typically, the
lifecycle of a drug design lasts 10–15 years[117]. Artificial
intelligence and machine learning allow specialists to use
statistical models to learn from the past, present, and
future, speeding up the process of discovering and testing
new treatments. SciBite makes the most of the predictive
analytics that AI and ML offer [118]. The company
reduced the amount of time it took for new
pharmaceuticals to hit the market by integrating AI into its
R&D methodology. According to New York University,
80 percent of clinical data is unstructured [119]. To speed
up operations in the post-drug development sector, AI and
ML are the tools that can operate with such a vast
information segment.
• Social Listening for Accurate Health and Drug-
Related Information
Social media sites may provide a wealth of vital
information if the correct tools are available. The article
that was published following the Pacific Symposium on
Biocomputing demonstrates how AI can provide important
insights into the effectiveness of antidepressants by
analyzing five million posts[120]. The study also
emphasizes the value of social listening in identifying drug
safety combinations and adverse drug reactions (ADRs).
In fact, researchers were able to identify some new side
effects of prescription medications. They used artificial
intelligence (AI) to examine public posts from users and
learned various patterns.
• Smarter Individual Case Safety Report (ICSR)
Collection
ICSR report collection constitutes a significant problem.
An even bigger challenge is their analysis. Over 20 million
ICSRs are said to be stored in the WHO database, which
might prove to be an invaluable resource for studying drug
side effects and other potential dangers. According to a
study published in the journal Clinical Pharmacology &
Therapeutics[121], the ICSR collection process is made
smarter overall with the use of AI and ML. The experts
forecast that ICSR reporting would be significantly more
advanced than it is presently by 2030. Massive amounts of
unstructured text in ICSRs can be analyzed using AI-based
technologies like Natural Language Processing (NLP),
resulting in ICSR management that is enhanced by AI.
• Cloud-Based Reporting
AI pharmacovigilance and cloud computing go hand in
hand. The experts believe that cloud technology will be
used to gather and analyze data. It is anticipated that the
cost-efficiency, scalability, and simplicity of
Pharmacovigilance will increase with the integration of
cloud-based computing with AI and ML.
• Personalized Medicine
In order to create individually customized treatments,
personalized drugs can be done by identifying a person's
biological, physical, physiological, and genetic markers.
Healthcare practitioners may evaluate thousands of
markers using artificial intelligence in pharmacovigilance
to produce considerably more precise predictions about
how particular drugs will affect particular people[122].

Pharmacological therapy will inevitably become more
personalized, reducing ADRs and boosting drug
effectiveness.
• Nanomedicine and Drug Delivery
Nanomedicine is currently a reality, not just a concept
from science fiction. The research, which was published in
the academic journal Drug Discovery Today[117],
demonstrates how pioneers are using nanotechnology and
medicine in tandem to identify, treat, and keep track of a
variety of complex illnesses. Specialists deal with asthma,
cancer, malaria, and HIV. Although nanomedicine is still
in its infancy, advances have been made in the field of
medication delivery by nanoparticle modification.
According to a recently released study from a scholarly
magazine[123], engineers and scientists are striving to
construct implantable nanorobots that will improve drug
delivery. Fuzzy logic, integrations, and neural networks
are examples of AI techniques that can simplify the overall
process.
III. CONCLUSION
The pharmaceutical industry is experiencing difficulties
with drug development projects due to rising drug
development costs and fewer chances of discovering new
drug molecules. This finding has led to an increase in the
number of pharmaceutical corporations and research
institutions investigating the use of ML and robotics
techniques to hasten the development of novel therapies
and make the exchange of observational data and clinical
trial outcomes easier. Multiple points in a drug's life cycle
are amenable to ML algorithms. This has been
demonstrated in detail in the preceding sections, where we
discussed ML applications beginning with the drug
discovery phases, such as target prediction and validation,
discovering therapeutic and toxicity effect profiles of
drugs, for prediction of, structure, bioactivity, and mode of
action. More data on high-risk populations, long-term
effects, food and drug interactions, and the escalation of
known and unknown adverse effects of the drug over time
are revealed by post-market drug monitoring. The use of
ML in drug post-market monitoring increases compliance
adherence and reduces expenditure significantly for each
and every ICSR. Pharmacovigilance that uses AI can
classify the harmful nature of reported events in addition
to evaluating their quality.
Despite the fact that ML, augmented intelligence, and a
variety of medical data from around the world are paving
the road for unified global healthcare, some challenges in
the utilization of Machine Learning algorithms for drug
development lifecycle still persist today. For the
construction and training of ML models, high-quality,
precise, and painstakingly vetted data is necessary. The
intricacy of the data type and the problem to be solved
dictate the requirements for the necessary data quantity
and accuracy. As a result, producing large data sets might
be costly. It's important to keep in mind that when training,
numerous neural network parameters are adjusted, some
theoretical and practical frameworks for enhancing these
models are not yet available. Another area where ML
models fall short is in the prediction of novel paradigms.
Because ML relies on training data to produce usable
models, these models can only make predictions within the
training data's predefined framework.
Drug research might be sped up and saved money by using
AI technology. Although ML might not be a solution for
all issues in drug discovery, it is unquestionably a useful
tool when used appropriately with the right data. The
power of artificial intelligence (AI) technology will
undoubtedly be used to complement human intelligence
and enhance our capabilities, thereby transforming the way
we approach drug development.
REFERENCES
[1] Burbidge, R., Trotter, M.W.B., Buxton, B.F. & Holden, S.
(2001) Drug design by machine learning: Support vector
machines for pharmaceutical data analysis. Computers and
Chemistry, 26, 5–14 [DOI: 10.1016/s0097-8485(01)00094-
8] [PubMed: 11765851].
[2] Olivier, Blondel, Mathieu, Prettenhofer, Peter, Weiss, Ron,
Dubourg, Vincent, Vanderplas, Jake, Passos, Alexandre,
Cournapeau, David, Brucher, Matthieu, Perrot, Matthieu,
Duchesnay, Edouard, Louppe & Gilles (2012). Scikit-
learn: Machine Learning in Python Pedregosa, Fabian &
Varoquaux, Gael & Gramfort. Journal of Machine
Learning Research. Alexandre & Michel: Vincent &
Thirion: Bertrand, USA, 12.
[3] Domingos, P. (2012) A few useful things to know about
machine learning. Communications of the ACM, 55, 78–87
[DOI: 10.1145/2347736.2347755].
[4] Vamathevan, J., Clark, D., Czodrowski, P., Dunham, I.,
Ferran, E., Lee, G., Li, B., Madabhushi, A., Shah, P.,
Spitzer, M. & Zhao, S. (2019) Applications of machine
learning in drug discovery and development. Nature
Reviews. Drug Discovery, 18, 463–477 [DOI:
10.1038/s41573-019-0024-5] [PubMed: 30976107]
[PubMed Central: PMC6552674].
[5] Fakhraei, S., Huang, B., Raschid, L. & Getoor, L. (2014)
Network-based drug-target interaction prediction with
probabilistic soft logic. IEEE/ACM Transactions on
Computational Biology and Bioinformatics, 11, 775–787
[DOI: 10.1109/TCBB.2014.2325031] [PubMed:
26356852].
[6] Bolton, E.E., Wang, Y., Thiessen, P.A. & Bryant, S.H. (2008)
PubChem: Integrated platform of small molecules and
biological activities. Annual Reports in Computational

Chemistry, 4, 217–241 [DOI: 10.1016/S1574-
1400(08)00012-1].
[7] Xia, Z., Wu, L.Y., Zhou, X. & Wong, S.T. (2010) Semi-
supervised drug-protein interaction prediction from
heterogeneous biological spaces. BMC Systems Biology.
BioMed Central, 4 (Supplement 2), S6 [DOI:
10.1186/1752-0509-4-S2-S6] [PubMed: 20840733].
[8] Keiser, M.J., Roth, B.L., Armbruster, B.N., Ernsberger, P.,
Irwin, J.J. & Shoichet, B.K. (2007) Relating protein
pharmacology by ligand chemistry. Nature Biotechnology,
25, 197–206 [DOI: 10.1038/nbt1284] [PubMed:
17287757].
[9] Keiser, M.J., Setola, V., Irwin, J.J., Laggner, C., Abbas, A.I.,
Hufeisen, S.J., Jensen, N.H., Kuijer, M.B., Matos, R.C.,
Tran, T.B., Whaley, R., Glennon, R.A., Hert, J., Thomas,
K.L., Edwards, D.D., Shoichet, B.K. & Roth, B.L. (2009)
Predicting new molecular targets for known drugs. Nature,
462, 175–181 [DOI: 10.1038/nature08506]. Epub 1
November 2009 [PubMed: 19881490] [PubMed Central:
PMC2784146].
[10] Cheng, A.C., Coleman, R.G., Smyth, K.T., Cao, Q., Soulard,
P., Caffrey, D.R., Salzberg, A.C. & Huang, E.S. (2007)
Structure-based maximal affinity model predicts small-
molecule druggability. Nature Biotechnology, 25, 71–75
[DOI: 10.1038/nbt1273] [PubMed: 17211405].
[11] Morris, G.M., Huey, R., Lindstrom, W., Sanner, M.F.,
Belew, R.K., Goodsell, D.S. & Olson, A.J. (2009)
Autodock4 and autodocktools4: Automated docking with
selective receptor flexibility. Journal of Computational
Chemistry, 30, 2785–2791 [DOI: 10.1002/jcc.21256]
[PubMed: 19399780].
[12] Mousavian, Z. & Masoudi-Nejad, A. (2014) Drug–target
interaction prediction via chemogenomic space: Learning-
based methods. Expert Opinion on Drug Metabolism and
Toxicology, 10, 1273–1287 [DOI:
10.1517/17425255.2014.950222] [PubMed: 25112457].
[13] Ding, H., Takigawa, I., Mamitsuka, H. & Zhu, S. (2014)
Similarity-based machine learning methods for predicting
drug–target interactions: A brief review. Briefings in
Bioinformatics, 15, 734–747 [DOI: 10.1093/bib/bbt056]
[PubMed: 23933754].
[14] Bleakley, K. & Yamanishi, Y. (2009) Supervised prediction
of drug–target interactions using bipartite local models.
Bioinformatics, 25, 2397–2403 [DOI:
10.1093/bioinformatics/btp433] [PubMed: 19605421].
[15] Alaimo, S., Giugno, R. & Pulvirenti, A. (2016).
Recommendation Techniques for Drug–Target Interaction
Prediction and Drug Repositioning. Data Mining
Techniques for the Life Sciences, 441–462.
[16] Wen, M., Zhang, Z., Niu, S., Sha, H., Yang, R., Yun, Y. &
Lu, H. (2017) Deep-learning-based drug-target interaction
prediction. Journal of Proteome Research, 16, 1401–1409
[DOI: 10.1021/acs.jproteome.6b00618] [PubMed:
28264154].
[17] Bolton, E.E., Wang, Y., Thiessen, P.A. & Bryant, S.H.
(2008) PubChem: Integrated platform of small molecules
and biological activities. Annual Reports in Computational
Chemistry, 4, 217–241 [DOI: 10.1016/S1574-
1400(08)00012-1].
[18] Cristianini, N. & Shawe-Taylor, J. (2000). An Introduction
to Support Vector Machines and Other Kernel-Based
Learning Methods. Cambridge University Press:
Cambridge.
[19] Yu, H., Chen, J., Xu, X., Li, Y., Zhao, H., Fang, Y., Li, X.,
Zhou, W., Wang, W. & Wang, Y. (2012) A systematic
prediction of multiple drug-target interactions from
chemical, genomic, and pharmacological data. PLOS ONE,
7, e37608 [DOI: 10.1371/journal.pone.0037608] [PubMed:
22666371].
[20] Anagaw, A. & Chang, Y.-L. (2019) A new complement
naïve Bayesian approach for biomedical data classification.
Journal of Ambient Intelligence and Humanized
Computing, 10, 3889–3897 [DOI: 10.1007/s12652-018-
1160-1].
[21] Nigsch, F., Bender, A., Jenkins, J.L. & Mitchell, J.B.O.
(2008) Ligand-target prediction using winnow and naïve
Bayesian algorithms and the implications of overall
performance statistics. Journal of Chemical Information
and Modeling, 48, 2313–2325 [DOI: 10.1021/ci800079x]
[PubMed: 19055411].
[22] Pang, X., Fu, W., Wang, J., Kang, D., Xu, L., Zhao, Y., Liu,
A.L. & Du, G.H. (2018) Identification of estrogen receptor
α antagonists from natural products via in vitro and in
silico approaches. Oxidative Medicine and Cellular
Longevity, 2018, 6040149 [DOI: 10.1155/2018/6040149]
[PubMed: 29861831].
[23] Wei, Y., Li, W., Du, T., Hong, Z. & Lin, J. (2019) Targeting
HIV/HCV coinfection using a machine learning-based
multiple quantitative structure–activity relationships
(multiple QSAR) method. International Journal of
Molecular Sciences, 20, 3572 [DOI:
10.3390/ijms20143572] [PubMed: 31336592].
[24] Lengauer, T. (2007) Bioinformatics—From genomes to
therapies. Bioinformatics‐ from Genomes to Therapies, 1–
24.
[25] Nayal, M. & Honig, B. (2006) On the nature of cavities on
protein surfaces: Application to the identification of drug-
binding sites. Proteins, 63, 892–906 [DOI:
10.1002/prot.20897] [PubMed: 16477622]
[26] Li, Q. & Lai, L. (2007) Prediction of potential drug targets
based on simple sequence properties. BMC Bioinformatics,
8, 353 [DOI: 10.1186/1471-2105-8-353] [PubMed:
17883836].
[27] Bakheet, T.M. & Doig, A.J. (2009) Properties and
identification of human protein drug targets.
Bioinformatics, 25, 451–457 [DOI:
10.1093/bioinformatics/btp002] [PubMed: 19164304].
[28] Wang, Q., Feng, Y., Huang, J., Wang, T. & Cheng, G.
(2017) A novel framework for the identification of drug
target proteins: Combining stacked auto-encoders with a
biased support vector machine. PLOS ONE, 12, e0176486
[DOI: 10.1371/journal.pone.0176486] [PubMed:
28453576].
[29] Jeon, J., Nim, S., Teyra, J., Datti, A., Wrana, J.L., Sidhu,
S.S., Moffat, J. & Kim, P.M. (2014) A systematic approach

to identify novel cancer drug targets using machine
learning, inhibitor design and high-throughput screening.
Genome Medicine, 6, 57 [DOI: 10.1186/s13073-014-0057-
7] [PubMed: 25165489].
[30] Costa, P.R., Acencio, M.L. & Lemke, N. (2010) A machine
learning approach for genome-wide prediction of morbid
and druggable human genes based on systems-level data.
BMC Genomics, 11 (Supplement 5), S9–S9 [DOI:
10.1186/1471-2164-11-S5-S9] [PubMed: 21210975].
[31] Kandoi, G., Acencio, M.L. & Lemke, N. (2015) Prediction
of druggable proteins using machine learning and systems
biology: A mini-review. Frontiers in Physiology, 6, 366–
366 [DOI: 10.3389/fphys.2015.00366] [PubMed:
26696900].
[32] Rouillard, A.D., Hurle, M.R. & Agarwal, P. (2018)
Systematic interrogation of diverse Omic data reveals
interpretable, robust, and generalizable transcriptomic
features of clinically successful therapeutic targets. PLOS
Computational Biology, 14, e1006142 [DOI:
10.1371/journal.pcbi.1006142] [PubMed: 29782487].
[33] Kumar, V., Sanseau, P., Simola, D.F., Hurle, M.R. &
Agarwal, P. (2016) Systematic analysis of drug targets
confirms expression in disease-relevant tissues. Scientific
Reports, 6, 36205 [DOI: 10.1038/srep36205] [PubMed:
27824084].
[34] Ferrero, E., Dunham, I. & Sanseau, P. (2017) In silico
prediction of novel therapeutic targets using gene-disease
association data. Journal of Translational Medicine, 15,
182 [DOI: 10.1186/s12967-017-1285-6] [PubMed:
28851378].
[35] Koscielny, G., An, P., Carvalho-Silva, D., Cham, J.A.,
Fumis, L., Gasparyan, R., Hasan, S., Karamanis, N.,
Maguire, M., Papa, E., Pierleoni, A., Pignatelli, M., Platt,
T., Rowland, F., Wankar, P., Bento, A.P., Burdett, T.,
Fabregat, A., Forbes, S., Gaulton, A., Gonzalez, C.Y.,
Hermjakob, H., Hersey, A., Jupe, S., Kafkas, Ş, Keays, M.,
Leroy, C., Lopez, F.J., Magarinos, M.P., Malone, J.,
McEntyre, J., Munoz-Pomer Fuentes, A., O’Donovan, C.,
Papatheodorou, I., Parkinson, H., Palka, B., Paschall, J.,
Petryszak, R., Pratanwanich, N., Sarntivijal, S., Saunders,
G., Sidiropoulos, K., Smith, T., Sondka, Z., Stegle, O.,
Tang, Y.A., Turner, E., Vaughan, B., Vrousgou, O.,
Watkins, X., Martin, M.J., Sanseau, P., Vamathevan, J.,
Birney, E., Barrett, J. & Dunham, I. (2017) Open targets: A
platform for therapeutic target identification and validation.
Nucleic Acids Research, 45, D985–D994 [DOI:
10.1093/nar/gkw1055] [PubMed: 27899665].
[36] Ramsundar, B., Liu, B., Wu, Z., Verras, A., Tudor, M.,
Sheridan, R.P. & Pande, V. (2017) Is multitask deep
learning practical for pharma? Journal of Chemical
Information and Modeling, 57, 2068–2076 [DOI:
10.1021/acs.jcim.7b00146] [PubMed: 28692267].
[37] Ma, J., Sheridan, R.P., Liaw, A., Dahl, G.E. & Svetnik, V.
(2015) Deep neural nets as a method for quantitative
structure–activity relationships. Journal of Chemical
10.1021/ci500747n] [PubMed: 25635324].
[38] Barati Farimani, A., Feinberg, E. & Pande, V. (2018)
Binding pathway of opiates to μ-opioid receptors revealed
by machine learning. Biophysical Journal, 114, 62a–63a
[DOI: 10.1016/j.bpj.2017.11.390].
[39] Wu, Z., Ramsundar, B., Feinberg, E.N., Gomes, J.,
Geniesse, C., Pappu, A.S., Leswing, K. & Pande, V. (2018)
MoleculeNet: A benchmark for molecular machine
learning. Chemical Science, 9, 513–530 [DOI:
10.1039/c7sc02664a] [PubMed: 29629118].
[40] Segler, M.H.S., Preuss, M. & Waller, M.P. (2018) Planning
chemical syntheses with deep neural networks and
symbolic AI. Nature, 555, 604–610 [DOI:
10.1038/nature25978] [PubMed: 29595767].
[41] Desai, B., Dixon, K., Farrant, E., Feng, Q., Gibson, K.R.,
van Hoorn, W.P., Mills, J., Morgan, T., Parry, D.M.,
Ramjee, M.K., Selway, C.N., Tarver, G.J., Whitlock, G. &
Wright, A.G. (2013) Rapid discovery of a novel series of
abl kinase inhibitors by application of an integrated
microfluidic synthesis and screening platform. Journal of
Medicinal Chemistry, 56, 3033–3047 [DOI:
10.1021/jm400099d] [PubMed: 23441572].
[42] Hartenfeller, M. & Schneider, G. (2011) De novo drug
design. In: Chemoinformatics and computational chemical
biology. Methods in Molecular Biology. Springer: Berlin,
672, 299–323 [DOI: 10.1007/978-1-60761-839-3_12]
[PubMed: 20838974].
[43] Schneider, G., Funatsu, K., Okuno, Y. & Winkler, D. (2017)
De novo drug design-ye olde scoring problem revisited.
Molecular Informatics, 36, 1681031 [DOI:
10.1002/minf.201681031] [PubMed: 28124833].
[44] Mullard, A. (2017) The drug-maker’s guide to the galaxy.
Nature, 549, 445–447 [DOI: 10.1038/549445a].
[45] 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, 10883–10890 [DOI:
10.18632/oncotarget.14073] [PubMed: 28029644].
[46] Olivecrona, M., Blaschke, T., Engkvist, O. & Chen, H.
(2017) Molecular de-novo design through deep
reinforcement learning. Journal of Cheminformatics, 9, 48
[DOI: 10.1186/s13321-017-0235-x] [PubMed: 29086083].
[47] Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A.A., Veness,
J., Bellemare, M.G., Graves, A., Riedmiller, M., Fidjeland,
A.K., Ostrovski, G., Petersen, S., Beattie, C., Sadik, A.,
Antonoglou, I., King, H., Kumaran, D., Wierstra, D., Legg,
S. & Hassabis, D. (2015) Human-level control through
deep reinforcement learning. Nature, 518, 529–533 [DOI:
10.1038/nature14236] [PubMed: 25719670].
[48] Gómez-Bombarelli, R., Wei, J.N., Duvenaud, D.,
Hernández-Lobato, J.M., Sánchez-Lengeling, B., Sheberla,
D., Aguilera-Iparraguirre, J., Hirzel, T.D., Adams, R.P. &
Aspuru-Guzik, A. (2018) Automatic chemical design using
a data-driven continuous representation of molecules. ACS
Central Science, 4, 268–276 [DOI:
10.1021/acscentsci.7b00572] [PubMed: 29532027].
[49] 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, 10883–10890 [DOI:
10.18632/oncotarget.14073] [PubMed: 28029644].
[50] Coley, C.W., Rogers, L., Green, W.H. & Jensen, K.F. (2018)
Scscore: Synthetic complexity learned from a reaction
corpus. Journal of Chemical Information and Modeling,
58, 252–261 [DOI: 10.1021/acs.jcim.7b00622] [PubMed:
29309147].
[51] Andras, P. (2018) High-dimensional function approximation
with neural networks for large volumes of data. IEEE
Transactions on Neural Networks and Learning Systems,
29, 500–508 [DOI: 10.1109/TNNLS.2017.2651985]
[PubMed: 28129193].
[52] Vamathevan, J., Clark, D., Czodrowski, P., Dunham, I.,
Ferran, E., Lee, G., Li, B., Madabhushi, A., Shah, P.,
Spitzer, M. & Zhao, S. (2019) Applications of machine
learning in drug discovery and development. Nature
Reviews. Drug Discovery, 18, 463–477. Gale Academic
OneFile accessed 2 October, 2022 [DOI: 10.1038/s41573-
019-0024-5] [PubMed: 30976107].
[53] Lusci, A., Pollastri, G. & Baldi, P. (2013) Deep architectures
and deep learning in chemoinformatics: The prediction of
aqueous solubility for drug-like molecules. Journal of
Chemical Information and Modeling, 53, 1563–1575 [DOI:
10.1021/ci400187y] [PubMed: 23795551].
[54] Duvenaud, D.K., Maclaurin, D., Iparraguirre, J., Bombarell,
R., Hirzel, T., Aspuru-Guzik, A. & Adams, R.P. (2015)
Convolutional networks on graphs for learning molecular
fingerprints. In: Advances in Neural Information
Processing Systems 28 (edited by C. Cortes, N. D.
Lawrence, D. D. Lee, M. Sugiyama & R. Garnett). Curran
Associates, Inc, (2224–2232).
[55] Coley, C.W., Barzilay, R., Green, W.H., Jaakkola, T.S. &
Jensen, K.F. (2017) Convolutional embedding of attributed
molecular graphs for physical property prediction. Journal
of Chemical Information and Modeling, 57, 1757–1772
[DOI: 10.1021/acs.jcim.6b00601] [PubMed: 28696688].
[56] Artursson, P. & Karlsson, J. (1991) Correlation between oral
drug absorption in humans and apparent drug permeability
coefficients in human intestinal epithelial (caco-2) cells.
Biochemical and Biophysical Research Communications,
175, 880–885 [DOI: 10.1016/0006-291x(91)91647-u]
[PubMed: 1673839].
[57] Hubatsch, I., Ragnarsson, E.G.E. & Artursson, P. (2007)
Determination of drug permeability and prediction of drug
absorption in caco-2 monolayers. Nature Protocols, 2,
2111–2119 [DOI: 10.1038/nprot.2007.303] [PubMed:
17853866].
[58] Wang, N.N., Dong, J., Deng, Y.H., Zhu, M.F., Wen, M.,
Yao, Z.J., Lu, A.P., Wang, J.B. & Cao, D.S. (2016) Adme
properties evaluation in drug discovery: Prediction of caco-
2 cell permeability using a combination of nsga-ii and
boosting. Journal of Chemical Information and Modeling,
56, 763–773 [DOI: 10.1021/acs.jcim.5b00642] [PubMed:
27018227].
[59] Tian, S., Li, Y., Wang, J., Zhang, J. & Hou, T. (2011) Adme
evaluation in drug discovery. 9. prediction of oral
bioavailability in humans based on molecular properties
and structural fingerprints. Molecular Pharmaceutics, 8,
841–851 [DOI: 10.1021/mp100444g] [PubMed:
21548635].
[60] Sim, D.S.M. (2015) Drug distribution. In: Pharmacological
Basis of Acute Care. Springer: Berlin, pp. 27–36.
[61] 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. Journal of
Chemical Information and Modeling, 56, 2042–2052 [DOI:
10.1021/acs.jcim.6b00044] [PubMed: 27602694].
[62] Matlock, M.K., Hughes, T.B. & Swamidass, S.J. (2015)
Xenosite server: A web-available site of metabolism
prediction tool. Bioinformatics, 31, 1136–1137 [DOI:
10.1093/bioinformatics/btu761] [PubMed: 25411327].
[63] Zaretzki, J., Matlock, M. & Swamidass, S.J. (2013)
Xenosite: Accurately predicting cyp-mediated sites of
metabolism with neural networks. Journal of Chemical
10.1021/ci400518g] [PubMed: 24224933].
[64] Dang, N.L., Hughes, T.B., Krishnamurthy, V. & Swamidass,
S.J. (2016) A simple model predicts ugt-mediated
metabolism. Bioinformatics, 32, 3183–3189 [DOI:
10.1093/bioinformatics/btw350] [PubMed: 27324196].
[65] Lombardo, F., Obach, R.S., Varma, M.V., Stringer, R. &
Berellini, G. (2014) Clearance mechanism assignment and
total clearance prediction in human based upon in silico
models. Journal of Medicinal Chemistry, 57, 4397–4405
[DOI: 10.1021/jm500436v] [PubMed: 24773013].
[66] Guengerich, F.P. (2011) Mechanisms of drug toxicity and
relevance to pharmaceutical development. Drug
Metabolism and Pharmacokinetics, p 1010210090, 26, 3–
14 [DOI: 10.2133/dmpk.dmpk-10-rv-062] [PubMed:
20978361].
[67] Xu, Y., Pei, J. & Lai, L. (2017) Deep learning based
regression and multiclass models for acute oral toxicity
prediction with automatic chemical feature extraction.
Journal of Chemical Information and Modeling, 57, 2672–
2685 [DOI: 10.1021/acs.jcim.7b00244] [PubMed:
29019671].
[68] Sushko, I., Salmina, E., Potemkin, V.A., Poda, G. & Tetko,
I.V. (2012) Toxalerts: A web server of structural alerts for
toxic chemicals and compounds with potential adverse
reactions. Journal of Chemical Information and Modeling,
52, 2310–2316 [DOI: 10.1021/ci300245q] [PubMed:
22876798].
[69] Mayr, A., Klambauer, G., Unterthiner, T. & Hochreiter, S.
(2016) Deeptox: Toxicity prediction using deep learning.
Frontiers in Environmental Science, 3, 80 [DOI:
10.3389/fenvs.2015.00080].
[70] Kearnes, S., Goldman, B. & Pande, V. (2016). Modeling
Industrial ADMET Data with Multitask Networks.
arXiv:1606.08793 [DOI: 10.48550/arXiv.1606.08793]
[71] Li, B., Shin, H., Gulbekyan, G., Pustovalova, O., Nikolsky,
Y., Hope, A., Bessarabova, M., Schu, M., Kolpakova-Hart,
E., Merberg, D., Dorner, A. & Trepicchio, W.L. (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, e0130700 [DOI: 10.1371/journal.pone.0130700]
[PubMed: 26107615].
[72] van Gool, A.J., Bietrix, F., Caldenhoven, E., Zatloukal, K.,
Scherer, A., Litton, J.E., Meijer, G., Blomberg, N., Smith,
A., Mons, B., Heringa, J., Koot, W.J., Smit, M.J., Hajduch,
M., Rijnders, T. & Ussi, A. (2017) Bridging the
translational innovation gap through good biomarker
practice. Nature Reviews. Drug Discovery, 16, 587–588
[DOI: 10.1038/nrd.2017.72] [PubMed: 28450744].
[73] Kraus, V.B. (2018) Biomarkers as drug development tools:
Discovery, validation, qualification and use. Nature
Reviews. Rheumatology, 14, 354–362 [DOI:
10.1038/s41584-018-0005-9] [PubMed: 29760435].
[74] 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.
[75] Zhan, F., Huang, Y., Colla, S., Stewart, J.P., Hanamura, I.,
Gupta, S., Epstein, J., Yaccoby, S., Sawyer, J., Burington,
B., Anaissie, E., Hollmig, K., Pineda-Roman, M., Tricot,
G., van Rhee, F., Walker, R., Zangari, M., Crowley, J.,
Barlogie, B. & Shaughnessy, J.D. (2006) The molecular
classification of multiple myeloma. Blood, 108, 2020–2028
[DOI: 10.1182/blood-2005-11-013458] [PubMed:
16728703].
[76] Shaughnessy, J.D., Jr, Zhan, F., Burington, B.E., Huang, Y.,
Colla, S., Hanamura, I., Stewart, J.P., Kordsmeier, B.,
Randolph, C., Williams, D.R., Xiao, Y., Xu, H., Epstein, J.,
Anaissie, E., Krishna, S.G., Cottler-Fox, M., Hollmig, K.,
Mohiuddin, A., Pineda-Roman, M., Tricot, G., van Rhee,
F., Sawyer, J., Alsayed, Y., Walker, R., Zangari, M.,
Crowley, J. & Barlogie, B. (2007) A validated gene
expression model of high-risk multiple myeloma is defined
by deregulated expression of genes mapping to
chromosome 1. Blood, 109, 2276–2284 [DOI:
10.1182/blood-2006-07-038430] [PubMed: 17105813].
[77] Zhan, F., Barlogie, B., Mulligan, G., Shaughnessy, J.D., 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 dexamethasone.
Blood J Am. Blood, 111, 968–969 [DOI: 10.1182/blood-
2007-10-119321] [PubMed: 18182586].
[78] Decaux, O., Lodé, L., Magrangeas, F., Charbonnel, C.,
Gouraud, W., Jézéquel, P., Attal, M., Harousseau, J.L.,
Moreau, P., Bataille, R., Campion, L., Avet-Loiseau, H.,
Minvielle, S. & Intergroupe Francophone du Myélome
(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 myelome. Journal of
Clinical Oncology, 26, 4798–4805 [DOI:
10.1200/JCO.2007.13.8545] [PubMed: 18591550].
[79] Mulligan, G., Mitsiades, C., Bryant, B., Zhan, F., Chng,
W.J., Roels, S., Koenig, E., Fergus, A., Huang, Y.,
Richardson, P., Trepicchio, W.L., Broyl, A., Sonneveld, P.,
Shaughnessy, J.D., Bergsagel, P.L., Schenkein, D.,
Esseltine, D.L., Boral, A. & Anderson, K.C. (2007) Gene
expression profiling and correlation with outcome in
clinical trials of the proteasome inhibitor bortezomib.
Blood, 109, 3177–3188 [DOI: 10.1182/blood-2006-09-
044974] [PubMed: 17185464].
[80] Costello, J.C., Heiser, L.M., Georgii, E., Gönen, M.,
Menden, M.P., Wang, N.J., Bansal, M., Ammad-ud-din,
M., Hintsanen, P., Khan, S.A., Mpindi, J.P., Kallioniemi,
O., Honkela, A., Aittokallio, T., Wennerberg, K., NCI
DREAM Community, Collins, J.J., Gallahan, D., Singer,
D., Saez-Rodriguez, J., Kaski, S., Gray, J.W. &
Stolovitzky, G. (2014) A community effort to assess and
improve drug sensitivity prediction algorithms. Nature
Biotechnology, 32, 1202–1212 [DOI: 10.1038/nbt.2877]
[PubMed: 24880487].
[81] Romero, K., Ito, K., Rogers, J.A., Polhamus, D., Qiu, R.,
Stephenson, D., Mohs, R., Lalonde, R., Sinha, V., Wang,
Y., Brown, D., Isaac, M., Vamvakas, S., Hemmings, R.,
Pani, L., Bain, L.J., Corrigan, B., Alzheimer's Disease
Neuroimaging Initiative & Coalition Against Major
Diseases (2015) The future is now: Model-based clinical
trial design for Alzheimer’s disease. Clinical
Pharmacology and Therapeutics, 97, 210–214 [DOI:
10.1002/cpt.16] [PubMed: 25669145].
[82] Zhao, Y., Zeng, D., Socinski, M.A. & Kosorok, M.R. (2011)
Reinforcement learning strategies for clinical trials in
nonsmall cell lung cancer. Biometrics, 67, 1422–1433
[DOI: 10.1111/j.1541-0420.2011.01572.x] [PubMed:
21385164].
[83] Wong, C.H., Siah, K.W. & Lo, A.W. (2019) Estimation of
clinical trial success rates and related parameters.
Biostatistics, 20, 273–286 [DOI:
10.1093/biostatistics/kxx069] [PubMed: 29394327].
[84] Schork, N.J. (2015) Personalized medicine: Time for one-
person trials. Nature, 520, 609–611 [DOI:
10.1038/520609a] [PubMed: 25925459].
[85] Vasudev, Naveen & Selby, Peter & Banks, Rosamonde.
(2012). Renal cancer biomarkers: The promise of
personalized care. BMC medicine. 10. 112. 10.1186/1741-
7015-10-112.
[86] Zhang, X., Xiao, C., Glass, L. M., & Sun, J. (2020).
DeepEnroll: Patient-Trial Matching with Deep Embedding
and Entailment Prediction. In The Web Conference 2020 -
Proceedings of the World Wide Web Conference, WWW
2020 (pp. 1029-1037). (The Web Conference 2020 -
Proceedings of the World Wide Web Conference, WWW
2020). Association for Computing Machinery, Inc.
https://doi.org/10.1145/3366423.3380181
[87] Calaprice-Whitty, D., Galil, K., Salloum, W., Zariv, A. &
Jimenez, B. (2020) Improving clinical trial participant
prescreening with artificial intelligence (AI): A comparison
of the results of AI-assisted vs standard methods in 3
oncology trials. Therapeutic Innovation and Regulatory

Science, 54, 69–74 [DOI: 10.1007/s43441-019-00030-4]
[PubMed: 32008227].
[88] Vassy, J.L., Ho, Y.L., Honerlaw, J., Cho, K., Gaziano, J.M.,
Wilson, P.W.F. & Gagnon, D.R. (2018) Yield and bias in
defining a cohort study baseline from electronic health
record data. Journal of Biomedical Informatics, 78, 54–59
[DOI: 10.1016/j.jbi.2017.12.017] [PubMed: 29305952].
[89] Weber, G.M., Adams, W.G., Bernstam, E.V., Bickel, J.P.,
Fox, K.P., Marsolo, K., Raghavan, V.A., Turchin, A.,
Zhou, X., Murphy, S.N. & Mandl, K.D. (2017) Biases
introduced by filtering electronic health records for patients
with ‘complete data’. Journal of the American Medical
Informatics Association, 24, 1134–1141 [DOI:
10.1093/jamia/ocx071] [PubMed: 29016972].
[90] Bain, E.E., Shafner, L., Walling, D.P., Othman, A.A.,
Chuang-Stein, C., Hinkle, J. & Hanina, A. (2017) Use of a
novel artificial intelligence platform on mobile devices to
assess dosing compliance in a phase 2 clinical trial in
subjects with schizophrenia. JMIR mHealth and uHealth,
5, e18 [DOI: 10.2196/mhealth.7030] [PubMed: 28223265].
[91] Labovitz, D.L., Shafner, L., Reyes Gil, M., Virmani, D. &
Hanina, A. (2017) Using artificial intelligence to reduce
the risk of nonadherence in patients on anticoagulation
therapy. Stroke, 48, 1416–1419 [DOI:
10.1161/STROKEAHA.116.016281] [PubMed:
28386037].
[92] Adamson, A.S. & Smith, A. (2018) Machine learning and
health care disparities in dermatology. JAMA
Dermatology, 154, 1247–1248 [DOI:
10.1001/jamadermatol.2018.2348] [PubMed: 30073260].
[93] Burlingame, E.A., Margolin, A.A., Gray, J.W. & Chang,
Y.H. (2018) SHIFT: Speedy histopathological-to-
immunofluorescent translation of whole slide images using
conditional generative adversarial networks. Proceedings
of SPIE–The International Society for Optical Engineering,
10581 [DOI: 10.1117/12.2293249] [PubMed: 30283195].
[94] Han, J., Chen, K., Fang, L., Zhang, S., Wang, F., Ma, H.,
Zhao, L. & Liu, S. (2019) Improving the efficacy of the
data entry process for clinical research with a natural
language processing-driven medical information extraction
system: Quantitative field research. JMIR Medical
Informatics, 7, e13331 [DOI: 10.2196/13331] [PubMed:
31313661].
[95] Gavrielov-Yusim, N., Kürzinger, M.L., Nishikawa, C., Pan,
C., Pouget, J., Epstein, L.B., Golant, Y., Tcherny-Lessenot,
S., Lin, S., Hamelin, B. & Juhaeri, J. (2019) Comparison of
text processing methods in social media-based signal
detection. Pharmacoepidemiology and Drug Safety, 28,
1309–1317 [DOI: 10.1002/pds.4857] [PubMed:
31392844].
[96] Barnett, I., Torous, J., Staples, P., Sandoval, L., Keshavan,
M. & Onnela, J.P. (2018) Relapse prediction in
schizophrenia through digital phenotyping: A pilot study.
Neuropsychopharmacology, 43, 1660–1666 [DOI:
10.1038/s41386-018-0030-z] [PubMed: 29511333].
[97] Yurtman, A., Barshan, B. & Fidan, B. (2018) Activity
recognition invariant to wearable sensor unit orientation
using differential rotational transformations represented by
quaternions. Sensors, 18, 2725 [DOI: 10.3390/s18082725]
[PubMed: 30126235].
[98] Hannun, A.Y., Rajpurkar, P., Haghpanahi, M., Tison, G.H.,
Bourn, C., Turakhia, M.P. & Ng, A.Y. (2019) Cardiologist-
level arrhythmia detection and classification in ambulatory
electrocardiograms using a deep neural network. Nature
Medicine, 25, 65–69 [DOI: 10.1038/s41591-018-0268-3]
[PubMed: 30617320].
[99] Ozkanca, Y., Öztürk, M.G., Ekmekci, M.N., Atkins, D.C.,
Demiroglu, C. & Ghomi, R.H. (2019) Depression
screening from voice samples of patients affected by
Parkinson’s disease. Digital Biomarkers, 3, 72–82 [DOI:
10.1159/000500354] [PubMed: 31872172].
[100] Han, X., Hu, Y., Foschini, L., Chinitz, L., Jankelson, L. &
Ranganath, R. (2020) Deep learning models for
electrocardiograms are susceptible to adversarial attack.
Nature Medicine, 26, 360–363 [DOI: 10.1038/s41591-020-
0791-x] [PubMed: 32152582].
[101] Vaci, N., Liu, Q., Kormilitzin, A., De Crescenzo, F.,
Kurtulmus, A., Harvey, J., O’Dell, B., Innocent, S.,
Tomlinson, A., Cipriani, A. & Nevado-Holgado, A. (2020)
Natural language processing for structuring clinical text
data on depression using UK-CRIS. Evidence-Based
Mental Health, 23, 21–26 [DOI: 10.1136/ebmental-2019-
300134] [PubMed: 32046989].
[102] Fonferko-Shadrach, B., Lacey, A.S., Roberts, A., Akbari,
A., Thompson, S., Ford, D.V., Lyons, R.A., Rees, M.I. &
Pickrell, W.O. (2019) Using natural language processing to
extract structured epilepsy data from unstructured clinic
letters: Development and validation of the ExECT
(extraction of epilepsy clinical text) system. BMJ Open, 9,
e023232 [DOI: 10.1136/bmjopen-2018-023232] [PubMed:
30940752].
[103] Savova, G.K., Danciu, I., Alamudun, F., Miller, T., Lin, C.,
Bitterman, D.S., Tourassi, G. & Warner, J.L. (2019) Use of
natural language processing to extract clinical cancer
phenotypes from electronic medical records. Cancer
Research, 79, 5463–5470 [DOI: 10.1158/0008-5472.CAN-
19-0579] [PubMed: 31395609].
[104] Estiri, H. & Murphy, S.N. (2019) Semi-supervised
encoding for outlier detection in clinical observation data.
Computer Methods and Programs in Biomedicine, 181,
104830 [DOI: 10.1016/j.cmpb.2019.01.002] [PubMed:
30658851].
[105] Glass & LMS (2019) G; Patil, R. Ai in Clinical
Development: Improving Safety and Accelerating Results
[White paper].
[106] Hicks, K.A., Mahaffey, K.W., Mehran, R., Nissen, S.E.,
Wiviott, S.D., Dunn, B., Solomon, S.D., Marler, J.R.,
Teerlink, J.R., Farb, A., Morrow, D.A., Targum, S.L., Sila,
C.A., Hai, M.T.T., Jaff, M.R., Joffe, H.V., Cutlip, D.E.,
Desai, A.S., Lewis, E.F., Gibson, C.M., Landray, M.J.,
Lincoff, A.M., White, C.J., Brooks, S.S., Rosenfield, K.,
Domanski, M.J., Lansky, A.J., McMurray, J.J.V., Tcheng,
J.E., Steinhubl, S.R., Burton, P., Mauri, L., O’Connor,
C.M., Pfeffer, M.A., Hung, H.M.J., Stockbridge, N.L.,
Chaitman, B.R., Temple, R.J. & Standardized Data
Collection for Cardiovascular Trials Initiative (SCTI)

(2018) 2017 Cardiovascular and stroke endpoint
definitions for clinical trials. Circulation, 137, 961–972
[DOI: 10.1161/CIRCULATIONAHA.117.033502]
[PubMed: 29483172].
[107] Liu, Y. & Gopalakrishnan, V. (2017) An overview and
evaluation of recent machine learning imputation methods
using cardiac imaging data. Data, 2, 8 [DOI:
10.3390/data2010008] [PubMed: 28243594].
[108] Feng, T. & Narayanan, S. (2019) Imputing missing data in
large-scale multivariate biomedical wearable recordings
using bidirectional recurrent neural networks with temporal
activation regularization. In:. Annual International
Conference of the IEEE Engineering in Medicine and
Biology Society. IEEE Engineering in Medicine and
Biology Society. Annual International Conference 41st
Annual International Conference of the IEEE Engineering
in Medicine and Biology Society (EMBC), 2019, 2529–
2534 [DOI: 10.1109/EMBC.2019.8856966] [PubMed:
31946412].
[109] Qiu, Y.L., Zheng, H. & Gevaert, O.J. (2018). A Deep
Learning Framework for Imputing Missing Values in
Genomic Data.
[110] Tomic, A., Tomic, I., Rosenberg-Hasson, Y., Dekker, C.L.,
Maecker, H.T. & Davis, M.M. (2019) SIMON, an
automated machine learning system, reveals immune
signatures of influenza vaccine responses. Journal of
Immunology, 203, 749–759 [DOI:
10.4049/jimmunol.1900033] [PubMed: 31201239].
[111] Rigdon, J., Baiocchi, M. & Basu, S. (2018) Preventing
false discovery of heterogeneous treatment effect
subgroups in randomized trials. Trials, 19, 382 [DOI:
10.1186/s13063-018-2774-5] [PubMed: 30012181].
[112] Kalscheur, M.M., Kipp, R.T., Tattersall, M.C., Mei, C.,
Buhr, K.A., DeMets, D.L., Field, M.E., Eckhardt, L.L. &
Page, C.D. (2018) Machine learning algorithm predicts
cardiac resynchronization therapy outcomes: Lessons from
the companion trial. Circulation. Arrhythmia and
Electrophysiology, 11, e005499 [DOI:
10.1161/CIRCEP.117.005499] [PubMed: 29326129].
[113] Linden, A. & Yarnold, P.R. (2016) Combining machine
learning and propensity score weighting to estimate causal
effects in multivalued treatments. Journal of Evaluation in
Clinical Practice, 22, 871–881 [DOI: 10.1111/jep.12610]
[PubMed: 27421786].
[114] Schuler, M.S. & Rose, S. (2017) Targeted maximum
likelihood estimation for causal inference in observational
studies. American Journal of Epidemiology, 185, 65–73
[DOI: 10.1093/aje/kww165] [PubMed: 27941068].
[115] Komorowski, M., Celi, L.A., Badawi, O., Gordon, A.C. &
Faisal, A.A. (2018) The artificial intelligence clinician
learns optimal treatment strategies for sepsis in intensive
care. Nature Medicine, 24, 1716–1720 [DOI:
10.1038/s41591-018-0213-5] [PubMed: 30349085].
[116] Ghassemi, M., Naumann, T., Schulam, P., Beam, A.L.,
Chen, I.Y. & Ranganath, R. (2019) Practical guidance on
artificial intelligence for health-care data. Lancet. Digital
Health, 1, e157–e159 [DOI: 10.1016/S2589-
7500(19)30084-6] [PubMed: 33323184].
[117] Paul, D., Sanap, G., Shenoy, S., Kalyane, D., Kalia, K. &
Tekade, R.K. (2021) Artificial intelligence in drug
discovery and development. Drug Discovery Today, 26,
80–93 [DOI: 10.1016/j.drudis.2020.10.010] [PubMed:
33099022].
[118] SciBite Limited (2022). Available at:
https://www.scibite.com/
[119] Hauben, M. (2020) The potential of artificial intelligence in
pharmacovigilance. Journal of the Faculty of
Pharmaceutical Medicine. Available at:
https://www.fpm.org.uk/journals/the-potential-of-artificial-
intelligence-in-pharmacovigilance/. 10 November 2020.
[120] Correia, R.B., Li, L. & Rocha, L.M. (2016) Monitoring
potential drug interactions and reactions via network
analysis of Instagram user timelines. Pacific Symposium
on Biocomputing. Pacific Symposium on Biocomputing,
21, 492–503 [PubMed: 26776212] [PubMed Central:
PMC4720984].
[121] Arlett, P., Straus, S. & Rasi, G. (2020) Pharmacovigilance
2030: Invited commentary for the January 2020 ‘futures’
edition of Clinical Pharmacology and therapeutics. Clinical
Pharmacology and Therapeutics, 107, 89–91 [DOI:
10.1002/cpt.1689]. Epub 22 November 2019 [PubMed:
31758540] [PubMed Central: PMC6977396].
[122] Mishra, V., Chanda, P., Tambuwala, M.M. & Suttee, A.
(2019)Personalized medicine: An overview. International
Journal of Pharmaceutical Quality Assurance, 10, 290–294
[DOI: 10.25258/ijpqa.10.2.13].
[123] Fu, J. & Yan, H. (2012) Controlled drug release by a
nanorobot. Nature Biotechnology, 30, 407–408 [DOI:
10.1038/nbt.2206] [PubMed: 22565965].

Application of Machine Learning in Drug Discovery and Development Lifecycle

More Related Content

What's hot (20)

Similar to Application of Machine Learning in Drug Discovery and Development Lifecycle (20)

More from AI Publications (20)

Recently uploaded (20)

Application of Machine Learning in Drug Discovery and Development Lifecycle