You are on page 1of 8

Drug Discovery Today  Volume 24, Number 3  March 2019 REVIEWS

Reviews  INFORMATICS
Artificial intelligence in drug
development: present status and
future prospects
Kit-Kay Mak1,2 and Mallikarjuna Rao Pichika2,3
1
School of Postgraduate Studies and Research, International Medical University, Kuala Lumpur, Malaysia
2
Department of Pharmaceutical Chemistry, School of Pharmacy, International Medical University, Kuala Lumpur, Malaysia
3
Head, Centre for Bioactive Molecules and Drug Delivery, Institute for Research, Development and Innovation (IRDI), International Medical University, Kuala
Lumpur, Malaysia

Artificial intelligence (AI) uses personified knowledge and learns from the solutions it produces to
address not only specific but also complex problems. Remarkable improvements in computational
power coupled with advancements in AI technology could be utilised to revolutionise the drug
development process. At present, the pharmaceutical industry is facing challenges in sustaining their
drug development programmes because of increased R&D costs and reduced efficiency. In this review, we
discuss the major causes of attrition rates in new drug approvals, the possible ways that AI can improve
the efficiency of the drug development process and collaboration of pharmaceutical industry giants with
AI-powered drug discovery firms.

Introduction subfield called machine learning (ML), which uses statistical meth-
Artificial intelligence (AI) is the simulation of the human intelli- ods with the ability to learn with or without being explicitly
gence process by computers. The process includes acquiring infor- programmed [1–3]. ML is categorised into supervised, unsuper-
mation, developing rules for using the information, drawing vised and reinforcement learning. Supervised learning comprises
approximate or definite conclusions and self-correction. The ad- classification and regression methods where the predictive model
vancement of AI can be seen as a double-edged sword: many fear is developed based upon the data from input and output sources.
that it will threaten their employment; by contrast, every advance Output from supervised ML entails disease diagnosis under the
in AI is celebrated because of the belief that it will vastly contribute subgroup classification; and drug efficacy and ADMET prediction
to the betterment of society. AI is used in various sectors from under the subgroup regression [4]. Unsupervised learning com-
innovating educational methods to automating business process- prises clustering and feature-finding methods by grouping and
es. The sprouting idea of adopting AI in the drug development interpreting data based solely on input data [5]. Through unsu-
process has shifted from hype to hope. In this review, the possible pervised ML, outputs such as disease subtype discovery from
applications of AI in the drug development pipeline in drug clustering and disease target discovery from feature-finding meth-
development strategies and processes, the pharmaceutical R&D ods can be attained [6]. Reinforcement learning is largely driven by
efficiency and attrition, and partnerships between AI and phar- decision making in a given environment and its execution to
maceutical companies are discussed. maximise its performance. The outputs from this type of ML
include de novo drug design under decision making and experi-
AI, machine learning and deep learning mental designs under execution – where both can be achieved via
The incorporation of AI into the healthcare industry is concisely modelling and quantum chemistry [7]. A further subfield of ML
shown in Fig. 1. AI is described as the use of techniques that enable called deep learning (DL) uses artificial neural networks that adapt
computers to mimic human behaviour (Fig. 1). AI also contains a and learn from the vast amount of experimental data [3,8]. The big
data and associated data mining and algorithm methods could
Corresponding author: Pichika, M.R. (mallikarjunarao@imu.edu.my) provide us with the capacity to discover new compounds that

1359-6446/ã 2018 Elsevier Ltd. All rights reserved.


https://doi.org/10.1016/j.drudis.2018.11.014 www.drugdiscoverytoday.com 773
REVIEWS Drug Discovery Today  Volume 24, Number 3  March 2019

Any technique which enables computers to mimic human brain.

It is simply a data analysis method that automates analytical model building using
Research
algorithms that iteratively learn from data. Computers can identify hidden insights
Wearables without being explicitly programmed.
Artificial inteligence

Drug discovery

Machine learning
Disease It is a deeper
Virtual Classification
Reviews  INFORMATICS

diagnosis
assistant Supervised subset of AI that
learning processes data

Deep learning
Drug
Healthcare

Patient data &


Regression efficacy/ADMET and creates
risk analysis

Machine learning
prediction patterns for
Hospital decision making
Disease subtype
management Clustering
Unsupervised
discovery purposes. It
Mental health learning comprises
Disease target
Feature finding networks which
discovery
Surgery are capable of
De novo drug learning from
Decision making
Lifestyle Reinforcement design
unstructured
learning
Digital health Experimental date.
Execution
monitoring & design
diagnostics

Drug Discovery Today

FIGURE 1
The applications of artificial intelligence (AI) and its subfields: machine learning and deep learning, in healthcare.

could potentially be new drugs, uncover or repurpose drugs that specific enzyme or with an entire organism. The first compound
could be more potent when used individually or in combination that shows activity against a given biological target is called a ‘hit’.
and improve the area of personalised medicine based on genetic Hits are often found during the screening of chemical libraries,
markers. The emergence of DL was observed with the increasing computer simulation or screening of naturally isolated materials,
amount of data and the continuous growth of computer power. such as plants, bacteria and fungi [13]. The identification of a lead
The striking difference that makes DL a subfield of AI is the molecule is the second step in drug development. A lead is a
flexibility in the architecture of neural networks such as convolu- chemical compound that shows promising potential that can lead
tional neural network (CNNs), recurrent neural networks (RNNs) to the development of a new drug as a treatment for a disease.
and fully connected feed-forward networks [7]. It is believed, with Identified hits are screened in cell-based assays predictive of the
proper establishment of methods in AI, we will witness the transi- disease state and in animal models of disease to characterise the
tion into an era of minimised failures in clinical trials and a faster, efficacy of the compound and its probable safety profile. Once a
cheaper and effective drug development processes [9–11]. lead compound has been found, its chemical structure is used as a
starting point for chemical modifications with the objective of
Drug development process discovering compounds with maximal therapeutic benefit and
The feedback-driven drug development process starts from exist- minimal potential for harm [13,14].
ing results obtained from various sources such as high-throughput During the process of lead generation, hit molecules are sys-
compound and fragment screening, computational modelling and tematically modified to improve their activity and selectivity
information available in the literature. This process alternates towards specific biological targets, while reducing toxicity and
between induction and deduction. This inductive–deductive cycle unwanted effects. The chemically related compounds derived
eventually leads to optimised hit and lead compounds. Automa- from a hit are called analogues and the process is referred to as
tion of specific parts of the cycle reduces randomness and errors hit expansion [15]. Medicinal chemists undertake hit expansion
and improves the efficiency of drug development. De novo design using well-established organic chemistry techniques. To increase
methods require knowledge of organic chemistry for in silico the synthetic throughput, chemists focus on a specific reaction or
compound synthesis and virtual screening models that function set of reactions to assemble building blocks together to make a
as surrogates for biochemical and biological tests of efficacy and series of analogues quickly. A ‘building block’ is a compound that
toxicity [12]. Ultimately, active learning algorithms allow the possesses a reactive functional group and atoms that interact with
identification of new or novel compounds with promising activi- the active site of a biological target. This active site is a specific
ties against a given disease target. region in the biological target to which the compound (or sub-
The first step in drug development is the identification of novel strate) binds through interaction forces. The binding of a substrate
chemical compounds with biological activity. This biological ac- to an active site can be visualised as ‘lock and key’ or ‘induced-fit’
tivity can arise from the interaction of the compound with a models [14,16].

774 www.drugdiscoverytoday.com
Drug Discovery Today  Volume 24, Number 3  March 2019 REVIEWS

R&D efficiency and attrition rate in drug development 11% (i.e., one in nine compounds). Nearly 62% of new chemical
Well-established drug-likeness guidelines in drug development are entities (NCEs) in Phase IIb and Phase III clinical trials do not reach
strictly applied, yet pharmaceutical companies face considerable the clinic. The major underlying sources of attrition in late clinical
challenges in improving R&D efficiency. The R&D efficiency is stages are attributed to clinical safety and efficacy followed by
simply a term used to describe the number of new drugs approved formulation, pharmacokinetics and bioavailability, and toxicity.
by the FDA per billion US$ spent on R&D alone. The cost of The lack of efficacy that emerged in therapeutic areas with high
discovering and developing a drug has since escalated from US$ Phase IIb and Phase III failure rates includes drugs for central
800 million in the year 2001 to the current estimated figure of US$ nervous systems (CNS) and oncology applications, mainly owing

Reviews  INFORMATICS
3 billion. The cost for drug development includes the entirety of to the lack of available fit-for-purpose animal models [23] in
failure; thus the estimated cost is an average estimate for a new preclinical development. Additionally, another contributing fac-
drug to be introduced into the clinic [17]. The declining rate in the tor for attrition is the ‘narrow clinical research’ strategy adopted by
numbers of new drugs approved per billion US$ spent on R&D is pharmaceutical companies. This term refers to the extensive use of
alarming. Thus, Scannell et al. systematically analysed the primary in vitro and in silico models instead of animal models in the early
causes for such worrying statistics, and narrowed them down to stages of the drug development process. The opportunities for
four factors impacting the R&D inefficiency: (i) the ‘better than serendipitous discovery are also limited with the existing system
Beatles’ problem; (ii) the ‘cautious regulator’ problem; (iii) the because the current system considers a drug candidate that exerts
‘throw money at it’ tendency; and (iv) ‘basic-research–brute-force’ other pharmacological effects than the intended effect as a clinical
bias [18]. trial failure. Furthermore, multicentre clinical trials across the
The better than Beatles factor offers an analogy that describes globe are also preferred with the goal of obtaining diverse data
the improbable idea of developing a new drug for a particular [24]. However, this leads to thinly spread patients (smaller pools of
disease with pharmacological activity better than any of the samples) in which the doctors can miss out on the opportunity to
existing approved drugs. The hurdles explained here concern observe new pharmacological effects. By contrast, the clinical trial
the fate of every new drug that offers a better therapeutic effect. size for a new follow-up drug in any therapeutic area is relatively
With every new drug, the stakes are raised, and the R&D ineffi- high. This is because the follow-up drug should be more potent
ciency takes place here as it becomes more and more difficult to than the existing drugs for it to be eligible for registration with the
clear the ever-progressing hurdles. Cautious regulators describe FDA. As per the guidelines, the clinical trial size must be inversely
the increase in stringency and in the tightening of regulations proportional to the square of the difference in the dose. For
owing to drug misfortunes in the past. It is a progressive effort to instance, if the dose is halved the clinical trial size must be
lower the risk tolerance in new drugs and to create a guide for increased four times [25]. All these constraints contribute to an
drugs with safer profiles, which is appropriate for the consumers increase in cost of later-stage clinical trials.
but considerably increases R&D expenditure. This is followed by The return on R&D investment is reduced by the availability of
the throw money at it tendency, an inclination where R&D many treatment options for a given disease, resulting in the low
facilities increase their recruitment in human and other resources probability of making a new follow-up drug as a blockbuster drug.
in the hopes of getting good returns on investment by being the Additionally, the long cycle time for approval owing to the strin-
first to launch a new drug. This also contributes to their high R&D gent approval rules stipulated by the FDA also acts as a contribut-
expenditures. Lastly, basic-research–brute-force bias refers to ing factor for a decline in return on R&D investments. However,
overestimating the ability of advances in basic research – espe- these factors are not universal as supported by the scientific
cially target identification and validation and brute-force screen- literature that demonstrates that different therapeutic areas have
ing methods in preclinical research – to increase the success rate different rates of success. For example, biologics have a higher rate
in the clinical stages of drug development [18]. In addition, it is of success from first-in-human studies to the clinic, and licensing-
known that the therapeutic benefits of drugs arise from the in compounds [21] has relatively higher success rates, at 24%.
interactions with multiple proteins rather than a single protein. The rate of attrition of compounds with novel mechanisms of
Thus, it is challenging to design a drug that acts on multiple action is higher than that of those with precedent mechanisms of
targets [19]. Another theory proposed to explain the decline in action. Pharma companies can average out the different therapeu-
R&D efficiency is the ‘low-hanging fruit’ problem. This describes tic areas of varying success rates as a measure to balance the overall
the fate of different tractable drug targets. The drugs for easily success to reduce the attrition rates [26]. Thus, by understanding
tractable drug targets were already invented and that leaves us their respective underlying factors, the attrition rates could be
with the more difficult drug targets. Despite the exciting chal- reduced.
lenge for discovery, attempts to develop new drugs will require a
higher cost in R&D [20]. Applications of AI in drug development
The increase in R&D costs as well as high attrition rates in the The tasks of finding successful new drugs is daunting and predom-
development process of a new drug pose challenges to the phar- inantly the most difficult part of drug development. This is caused
maceutical industry. The underlying causes and possible measures by the vast size of what is known as chemical space, which is
to reduce attrition rates have previously been reviewed [21,22]. estimated to be in the order of 1060 molecules [27]. The technolo-
The rate of attrition was described using first-in-human trials to gies that incorporate AI have become versatile tools that can be
register ten established pharma companies in the period 1991– applied ubiquitously in various stages of drug development, such
2000, where the regular success rate of products that make it as identification and validation of drug targets, designing of new
through development and approval from the regulators is nearly drugs, drug repurposing, improving the R&D efficiency, aggregat-

www.drugdiscoverytoday.com 775
REVIEWS Drug Discovery Today  Volume 24, Number 3  March 2019

ing and analysing biomedicine information and refining the that animal models exhibiting a disease-relevant phenotype with a
decision-making process to recruit patients for clinical trials neural network classifier of >71% accuracy provided the most
[28–30]. These potential uses of AI provide the opportunity to predictive power [45]. The application of AI in the process of drug
counter the inefficiencies and uncertainties that arise in the clas- development is proposed in Fig. 2. IBM Watson for Drug Discov-
sical drug development methods while minimising bias and hu- ery, an AI platform, has identified five new RNA-binding proteins
man intervention in the process [31]. (RBPs) linked to pathogenesis of a neurodegenerative disease
The other uses of AI in drug development include the prediction known as amyotrophic lateral sclerosis (ALS) [46].
of feasible synthetic routes for drug-like molecules [32], pharma-
Reviews  INFORMATICS

cological properties [33], protein characteristics as well as efficacy AI in finding the hit or lead
[34], drug combination and drug–target association [35] and drug The implementation of AI in the discovery of small drug-like
repurposing [36]. Also, the identification of new pathways and molecules concerns the utilisation of chemical space. Chemical
targets using omics analysis becomes possible via the generation of space provides the stage for identifying novel and high-quality
novel biomarkers and therapeutic targets, personalised medicine molecules because it is possible to computationally enumerate the
based on omics markers and discovering the connections between probable organic molecules [47]. Additionally, ML techniques and
drugs and diseases [37,38]. predictive model software also contribute to the identification of
DL has demonstrated outstanding success in proposing potent target-specific virtual molecules and association of the molecules
drug candidates and accurately predicting their properties and the with their respective target while optimising the safety and effica-
possible toxicity risks [39]. Circumventing past problems in drug cy attributes.
development – such as analysis of large datasets, laborious screen- AI systems can reduce the attrition rates and the R&D expendi-
ing of compounds while minimising standard error, requiring ture by decreasing the number of synthesised compounds that are
large amounts of R&D cost and time of over US$2.5 billion and subsequently tested in either in vitro or in vivo systems [48]. A
a more than a decade [40] – are now possible using AI methods variety of in silico techniques for profile selection such as virtual
[41]. With AI technology, new studies can be carried out in ligand or structure-based design approaches can be used with the
assisting the identification of new drug targets, rational drug available data on small-molecule modulator probes or their struc-
designing and drug repurposing [42,43]. tural biology. DL becomes useful in instances where structural data
are insufficient. Thus, phenotypic data or disease, biology or
AI in understanding the pathway or finding molecular molecule network-based algorithms can be used. The validated
targets AI techniques can be used to increase the success rates in drug
In drug development, AI has transformed the methods of pathway development, whereas the AI techniques that are in development
or target identification to treat diseases. This was possible owing to must be validated before applying to the drug development pro-
the incorporation of genomics information, biochemical attri- cess. The most crucial part in the drug development process is the
butes and target tractability [44]. One study determined the plau- synthesis of chosen molecules. Thus, AI is valuable owing to its
sibility of predicting therapeutic targets using a computational ability to prioritise molecules based on the ease of synthesis or
prediction application known as ‘Open Targets’ – a platform develop tools that are effective for the optimal synthetic route
consisting of gene–disease association data – and it was reported [48,49].

Benchmark compounds set design


Predict target’s role in a disease
Al in drug Design of in silico compound libraries Prediction of structure-activity Drug repurposing
development Identify novel targets relationship (SAR) Selection of patient population in Pharmacovigilance
Predict druggability of targets Prediction of ADMET properties clinical trials to increase success rates
Predict signalling systems

FDA approval,
Clinical candidate,
Validated Assay Lead molecule shown Drug, safe Drug,shows Drug,shows effect manufacture, Observation of
Outcomes shown effect in
target developed effect on target in humans effect in humans in large poplation post-marketing adverse effects
an animal model
survey

Phase ll Phase lll Phase lV


Drug development Target Biological Compound Hit/lead Lead Preclinical Phase I Approved drug
(proof of (multicentre (postmarketing
process validation assay screening identification optimisation development (1st in human) for clinical use
concept) trials) & surveillance)

Drug development Target-based drug discovery Opportunity for serendipitous identification of new therapeutic uses
strategies Phenotype-based discovery
Identification of new therapeutic uses

Drug development Approved drug


Phase l Phase ll Phase lll
process for clinical use

Drug development Drug repurposing


strategies
Al in drug
Interrogation of transcriptomic data
development

Drug Discovery Today

FIGURE 2
Utilisation of artificial intelligence (AI) in the drug development process. The outcomes and the strategies of the various components of the drug development
process are described. The applications of AI at each stage of drug development are also shown.

776 www.drugdiscoverytoday.com
Drug Discovery Today  Volume 24, Number 3  March 2019 REVIEWS

AI in synthesis of drug-like compounds reported to accurately predict the toxicity of unknown com-
Drug-like molecules are compounds that obey Lipinski’s rule of pounds.
five: (i) molecular weight <500 Da; (ii) H-bond donors <5; (iii) H-
bond acceptor <10; and (iv) calculated Log P (cLogP) <5. For the AI in selection of a population for clinical trials
synthesis of drug-like molecules, retrosynthesis is used extensively An ideal AI tool to assist in clinical trials should recognise the
by chemists. The first step in the retrosynthetic approach is to disease in patients, identify the gene targets and predict the effect
analyse the target compounds recursively and to sequentially of the molecule designed as well as the on- and off-target effects. A
convert them into smaller fragments or building blocks that can novel AI platform called AiCure was also developed as a mobile

Reviews  INFORMATICS
be easily purchased or prepared. The second step is to identify the application to measure medication adherence in a Phase II trial of
reactions that will convert these fragments into target compounds. subjects suffering from schizophrenia, where it was reported that
The second step is the most challenging because it is difficult for AiCure increased adherence 25% compared with the traditional
the human brain to interrogate the vast number of relevant ‘modified directly observed therapy’ [55]. Patient selection for a
organic reactions available in the literature to pick the most clinical trial is a crucial process. Interrogating the relationship
plausible reaction. AI would aid in predicting the best sought-after between human-relevant biomarkers and in vitro phenotypes
reactions by filling the voids that cause high failure in expected affords a more predictable, quantifiable assessment of the uncer-
organic synthesis (commonly known as ‘out of scope’ com- tainty of therapeutic responses in a specific patient. The develop-
pounds). The voids in organic synthesis are mainly the result of ment of AI approaches to identify and predict human-relevant
unpredictable steric and electronic effects and incomplete under- biomarkers of disease allows the recruitment of a specific patient
standing of the reaction mechanism. Currently, several computer- population in Phase II and III clinical trials. The AI predictive
aided organic compound synthesis (CAOCS) systems are available modelling in selection of a patient population would increase the
to assist chemists in selecting the synthesis route; however, it is not success rate in clinical trials [56,57].
a component of the computer-aided drug discovery (CADD) work-
flow. AI in drug repurposing
Seglar et al. have developed a new AI platform named 3N-MCTS, With AI, the drug repurposing process becomes more attractive
which combines three different deep neural networks with Monte and pragmatic. The concept of applying an existing therapeutic
Carlo Tree Search (MCTS) for CAOCS. This platform can filter out to a new disease is advantageous because the new drug is
the most promising building blocks and select only well-known qualified go directly to Phase II trials for a different indication
reactions for the synthesis of target compounds. The platform, 3N- without having to pass through Phase I clinical trials and
MCTS, was proven to be much faster and better than that of toxicology testing again [58,59]. In silico methods predicting
traditional computer-assisted retrosynthesis systems. The plat- pharmacological properties of drug and drug repurposing using
form was able to propose feasible synthesis routes without unrea- transcriptomic data comprising various biological systems and
sonable steps in a relatively short time. However, quantitative conditions through DL applications were reported. The meth-
prediction of enantiomeric or diastereomeric ratios and devising ods described are based on high-level representations of data
natural product syntheses plans are an unmet need. utilising deep neural networks (DNNs), which is essentially a
highly adaptive multilayer system comprising connected and
Predicting the mode-of-action of compounds using AI interacting artificial neurons that perform various data trans-
The prospect of having an AI platform that can predict the on- and formation [60]. In a study performed by Aliper et al., it was
off-target effects and in vivo safety profile of compounds before demonstrated that DNNs could classify complex drug action
they are synthesised excites those involved in the drug develop- mechanisms on the pathway level, thus classifying drugs into
ment process – particularly medicinal chemists. The availability of therapeutic categories according to their functional class, effi-
such platforms reduces the drug development time, R&D costs and cacy, therapeutic use and toxicity [61]. Additionally, the
attrition rates. A few examples of such platforms are DeepTox advances in precision medicine have resulted in the creation
(predicts toxicity of new compounds) and PrOCTOR (predicts the of next-generation AI that offers the ability to design drug
probability of toxicity in clinical trials) [51,52]. The predictive molecules from the generative adversarial networks (GANs)
accuracy of these platforms could be improved if a bigger and [62]. GANs are an astounding technology that uses DL to
refined dataset on toxicity and therapeutic profile of a varied set of produce photo-realistic pictures from text descriptions [63].
compounds is made available. However, this can only be achieved Thus, this platform can perform remarkable tasks beyond ana-
if there is a willingness to share data among the industry. lysing data, such as imagining or creating new data modelled
Recently, an innovative AI tool, SPiDER, was developed [53] as on real data. Essentially, the GAN technique is an adversarial
an alternative to chemoproteomics to advance natural products game between two DNNs where, principally, one DNN evalu-
for drug discovery. For proof-of-concept, SPiDER was used to ates the output of the other iteratively and, in that adversarial
predict the molecular target of b-lapachone, a clinical-stage natu- game, the two networks learn how to generate more perfect
ral naphthoquinone with antitumour activity. The platform pre- molecules [62].
dicted b-lapachone as an allosteric and reversible modulator of 5- Another next-generation AI method used in in silico medicine is
lipoxygenase (5-LO). The prediction is validated using a 5-LO reinforcement learning. The advantage of this AI technique is that
functional assay. Another AI tool: read-across structure–activity it is less dependent on learning from datasets, thus the networks
relationships (RASAR) [54], which links molecular structures and can recognise certain strategies in drug molecule design.
toxic properties by mining a large database of chemicals, was Zhavoronknov’s team designed algorithms that can reconstruct

www.drugdiscoverytoday.com 777
REVIEWS Drug Discovery Today  Volume 24, Number 3  March 2019

missing features from half-full datasets and interpret the differ- extraordinary speed, accuracy and diversity in structural chemis-
ences in normal and diseased profiles within complex data [64]. try using DL, Atomwise has assisted in the invention of new
One aspect of interest is the possibility of AI in designing drugs potential medicines for 27 disease targets and is working with top
with fewer side effects. The AI algorithm is also being trained to institutions such as Harvard University and Stanford University,
differentiate between cardiotoxic and non-cardiotoxic drugs using as well as pharmaceutical companies. Benevolent AI is one of the
gene expression data from cells incubated with different drugs. AI tools currently used in drug development. It utilises the
The team is currently competing their AI-designed molecules strategy of text mining to analyse the available patents and other
against those designed by chemists. An attempt was made to genetics and biological information to extrapolate the relation-
Reviews  INFORMATICS

capture innate knowledge that assists experienced medicinal che- ships between those entities to create highly informative graphs
mists in identifying promising drug candidates by looking using consisting of dynamic maps with over one billion relationships.
mobile electroencephalography to measure responses to the struc- The resulting graphs consisting of highly complex relationships
ture and numerical properties of a molecule. Apart from that, it are believed to have the ability to provide information or identify
also searches for signs of bias in the types of drug molecules new links or knowledge gaps that can generate new hypotheses.
preferred by chemists, which would not be apparent in AI- In contrast to human intervention where the element of bias is
designed molecule selection. present, AI adopts the disease-agonistic approach which permits
the unbiased views of scientific data that produce unbiased
AI in polypharmacology hypotheses.
Currently, the ‘one-disease–multiple-targets’ paradigm domi- Exscientia is an AI company that specialises in phenotypic drug
nates over the ‘one-disease–one-target’ paradigm because of dee- discovery. Human analysis of extremely complicated datasets for
per understanding of pathological processes in diseases at the high-content phenotypic drug discovery is outperformed by AI by
molecular level. One-disease–multiple-targets is termed poly- a large margin [69]. The ease in rapidly evolving drug designs is
pharmacology [19]. Many databases, such as ZINC, PubChem, achievable through the testing of each newly designed compound
Ligand Expo, KEGG, ChEMBL, DrugBank, STITCH, BindingDB, and then comparing it with its anticipated performance and other
Supertarget, PDB, among others, are available to integrate diverse molecules. Another notable AI start-up is Numerate. Numerate
information of molecular pathways, crystal structures, binding focuses on ligand chemistry, ADMET and combinatorial ML with
affinities, drug targets, disease relevance, chemical properties and classical approaches, giving emphasis on the transformation of
biological activities. AI could be used to probe these databases to new medicine discovery by filling significant therapeutic gaps
design polypharmacological agents. A success story of an AI through analysing large datasets pertaining to drug development
application in designing polypharmacological agents was recent- by applying algorithms.
ly published in the literature – the authors developed a compu- A healthy spread of such partnerships covering various research
tational platform, DeepDDI, for better understanding of drug– areas, such as the identification of novel small molecules [32], the
drug interactions and associated mechanisms and prediction of discovery of new treatment methods, monitoring the health data
alternative drugs for intended clinical use without negative via wearable technologies [70], among others, is concisely shown
health effects [65]. in Fig. 3. These advances are forecasted to serve as contributing
factors to the betterment of healthcare services, improvement in
Partnerships between the pharma industry and AI terms of efficiency in clinical trials, enhancement in stratified
companies medicine and more [71]. Presently, up to US$3 billion over the
With the rapid introduction of AI in healthcare, especially in the course of 15 years will be spent to bring a new drug to market. This
years 2016 and 2017, numerous pharmaceutical companies have trend is unsustainable and a change is inevitable because consu-
made investments in and have joint ventures with AI companies mers are not willing to pay more for medicines and for the cost of
in the hopes of developing better healthcare tools. These include the failures [72,73]. Thus, a change in the business model is
the improvements in diagnostics or biomarkers and the identifi- necessary, and AI offers such opportunity. These collaborations
cation of drug targets and designing new drugs [66]. The transi- demonstrate the importance of AI technology in allowing us to
tion from general medicine to modern AI healthcare focuses on explore a much bigger design space and discover rare molecules
the basis of the data [67]. The analyses of these underlying data that have properties, which is beyond the conventional finds if we
coupled with ML or DL are subsequently formulated into algo- relied solely on traditional HTS.
rithms – thereby strongly contributing to progressive modern
healthcare that incorporates AI. Thus, numerous partnerships Concluding remarks
between pharmaceutical industries and AI companies were re- Currently, there are no developed drugs that have utilised AI
cently developed on a global scale. For instance, DeepMind approaches but, based on the advances described in this review,
Technologies, a subsidiary of Google, collaborated with Royal it is likely that it will take a further 2–3 years for a drug to be
Free London NHS Foundation Trust to assist in the management developed. Interestingly, experts strongly believe that AI will
of acute kidney injury [68]. A global project that utilises data and permanently change the pharmaceutical industry and the way
AI from NHS patients with rare diseases is UK’s 100 000 Genomes drugs are discovered. However, for an individual to be efficient in
Project; this project is in collaboration with Roche, Berg, Merck drug development using AI, the individual should know how to
and Biogen. train algorithms, requiring domain expertise. This creates the
Atomwise is a pioneer in healthcare AI and is the first DL suitable workspace whereby AI and medicinal chemists can work
technology for novel small molecule discovery. Known for its closely together, because the former will be able to help in analys-

778 www.drugdiscoverytoday.com
Drug Discovery Today  Volume 24, Number 3  March 2019 REVIEWS

Al Pharma Al

Confidential

Drug repurposing To announce the mechanisms


of Al-based patient monitoring
platform can improve adherence

To track real-time data via smartphones To predict how cancer types


and other wearable technologies will respond to treatment

Reviews  INFORMATICS
To target prsonalised medicine
using machine learning and
large-scale genome sequencing

To identify new treatments


for psychiatric diseases

To discover novel and


selective small molecules To discover new treatment
To create bispecific small molecules options and to improve
for immuno-oncolog therapies chronic disease management

To identify novel biological To accelerate drug discovery


targets and pathway in immuno-oncology

To license the rights to develop, To identify new drug targets,


manufacture and commercialise combination therapies and
clinical stage drug candidates patient selection strategies
in immuno-oncology
To find new drug candidates for glaucoma
To improve health outcomes
To accelerate clinical development for breast cancer patients
of new therapies for breast cancer
To design small molecule modulators
of ryanodine receptor 2
To find and validate potential
cancer drug targets To identify candidates for oncology,
gastroenterology and CNS disorder
To evaluate novel targets for for clinical trials
neurodegenerative disorders such
as parkinson’s disease To generate novel small molecule
drug leads from an unnamed
To assess potential biomarkers for
cardiovascular disease target
seasonal flu vaccine performance
To announce the mechanisms of
Al-based patient monitoring Confidential
platform can imrove adherence
To optimise clinical trial
planning with AI

Drug Discovery Today

FIGURE 3
Partnerships between artificial intelligence (AI) and pharmaceutical companies and areas of collaboration in drug development.

ing huge datasets and the latter can train machines, set algorithms Conflicts of interest
or optimise the analysed data for a speedier and accurate drug There are no funding sources to declare for the writing of this
development process. Despite the benefit of AI in speeding up drug manuscript. Both authors contributed to the drafting, analysis,
development, real experiments still need to be conducted. Addi- editing and submission of the manuscript.
tionally, AI can be used in assisting gene therapy or other therapies
that are currently not available to us as tools in healthcare. With Acknowledgements
AI, the possibility of combining regenerative medicine with phar- The authors thank the International Medical University for the
macology and gene therapy emerges. library facilities.

References
1 Bishop, C.M. (2013) Model-based machine learning. Philos. Trans. A Math. Phys. Eng. 7 Chen, H. et al. (2018) The rise of deep learning in drug discovery. Drug Discov. Today
Sci. 371 http://dx.doi.org/10.1098/rsta.2012.0222 23, 1241–1250
2 VoPham, T. et al. (2018) Emerging trends in geospatial artificial intelligence (geoAI): 8 Grys, B.T. et al. (2017) Machine learning and computer vision approaches for
potential applications for environmental epidemiology. Environ. Health 17, 40 phenotypic profiling. J. Cell Biol. 216, 65–71
3 Lee, J.-G. et al. (2017) Deep learning in medical imaging: general overview. Korean J. 9 Labovitz, D.L. et al. (2017) Using artificial intelligence to reduce the risk of
Radiol. 18, 570–584 nonadherence in patients on anticoagulation therapy. Stroke 48, 1416–1419
4 Gun9car, G. et al. (2018) An application of machine learning to haematological 10 Jiang, F. et al. (2017) Artificial intelligence in healthcare: past, present and future.
diagnosis. Sci. Rep. 8, 411 Stroke Vasc. Neurol. 2, 230–243
5 Koohy, H. (2017) The rise and fall of machine learning methods in biomedical 11 Fleming, N. (2018) How artificial intelligence is changing drug discovery. Nature
research. F1000 Res. 6 http://dx.doi.org/10.12688/f1000research.13016.2 557, S55–S57
6 Young, J.D. et al. (2017) Unsupervised deep learning reveals prognostically relevant 12 Yuan, Y. et al. (2011) LigBuilder 2: a practical de novo drug design approach. J. Chem.
subtypes of glioblastoma. BMC Bioinf. 18, 381 Inf. Model. 51, 1083–1091

www.drugdiscoverytoday.com 779
REVIEWS Drug Discovery Today  Volume 24, Number 3  March 2019

13 Zhu, T. et al. (2013) Hit identification and optimization in virtual screening: 45 Ferrero, E. et al. (2017) In silico prediction of novel therapeutic targets using gene-
practical recommendations based on a critical literature analysis. J. Med. Chem. 56, disease association data. J. Transl. Med. 15, 182
6560–6572 46 Bakkar, N. et al. (2018) Artificial intelligence in neurodegenerative disease research:
14 Anderson, A.C. (2012) Structure-based functional design of drugs: from target to use of IBM Watson to identify additional RNA-binding proteins altered in
lead compound. Methods Mol. Biol. 823, 359–366 amyotrophic lateral sclerosis. Acta Neuropathol. 135, 227–247
15 Hall, D.R. et al. (2012) Hot spot analysis for driving the development of hits into 47 Reymond, J.-L. et al. (2010) Chemical space as a source for new drugs. Med. Chem.
leads in fragment-based drug discovery. J. Chem. Inf. Model. 52, 199–209 Commun. 1, 30–38
16 Alanine, A. et al. (2003) Lead generation-enhancing the success of drug discovery by 48 Okafo, G. et al. (2018) Adapting drug discovery to artificial intelligence. Drug Target
investing in the hit to lead process. Comb. Chem. High Throughput Screen. 6, 51–66 Rev. 2018, 50–52
17 DiMasi, J.A. et al. (2015) The cost of drug development. N. Engl. J. Med. 372 http://dx. 49 Segler, M.H.S. et al. (2018) Planning chemical syntheses with deep neural networks
doi.org/10.1056/NEJMc1504317 and symbolic AI. Nature 555, 604–610
Reviews  INFORMATICS

18 Scannell, J.W. et al. (2012) Diagnosing the decline in pharmaceutical R&D 51 Mayr, A. et al. (2016) DeepTox: toxicity prediction using deep learning. Front.
efficiency. Nat. Rev. Drug Discov. 11, 191–200 Environ. Sci. 3, 80
19 Reddy, A.S. and Zhang, S. (2013) Polypharmacology: drug discovery for the future. 52 Gayvert, K.M. et al. (2016) A data-driven approach to predicting successes and
Expert Rev. Clin. Pharmacol. 6, 41–47 failures of clinical trials. Cell. Chem. Biol. 23, 1294–1301
20 Tormay, P. (2015) Big data in pharmaceutical R&D: creating a sustainable R&D 53 Rodrigues, T. et al. (2018) Machine intelligence decrypts b-lapachone as an
engine. Pharm. Med. 29, 87–92 allosteric 5-lipoxygenase inhibitor. Chem. Sci. 9, 6899–6903
21 Kola, I. and Landis, J. (2004) Can the pharmaceutical industry reduce attrition rates? 54 Luechtefeld, T. et al. (2018) Machine learning of toxicological big data enables read-
Nat. Rev. Drug Discov. 3, 711–716 across structure activity relationships (RASAR) outperforming animal test
22 Esch, E.W. et al. (2015) Organs-on-chips at the frontiers of drug discovery. Nat. Rev. reproducibility. Toxicol. Sci. 165, 198–212
Drug Discov. 14, 248–260 55 Bain, E.E. et al. (2017) Use of a novel artificial intelligence platform on mobile
23 Acharya, K.R. et al. (2003) Ace revisited: a new target for structure-based drug design. devices to assess dosing compliance in a Phase 2 clinical trial in subjects with
Nat. Rev. Drug Discov. 2, 891–902 schizophrenia. JMIR Mhealth Uhealth 5, e18
24 Black, N. et al. (2004) Cross sectional survey of multicentre clinical databases in the 56 Perez-Gracia, J.L. et al. (2017) Strategies to design clinical studies to identify
United Kingdom. BMJ 328, 1478 predictive biomarkers in cancer research. Cancer Treat. Rev. 53, 79–97
25 Noordzij, M. et al. (2010) Sample size calculations: basic principles and common 57 Deliberato, R.O. et al. (2017) Clinical note creation, binning, and artificial
pitfalls. Nephrol. Dial. Transplant. 25, 1388–1393 intelligence. JMIR Med. Inf. 5, e24
26 Mullard, A. (2016) Parsing clinical success rates. Nat. Rev. Drug Discov. 15, 447 58 Corsello, S.M. et al. (2017) The Drug Repurposing Hub: a next-generation drug
27 Segler, M.H.S. et al. (2018) Generating focused molecule libraries for drug discovery library and information resource. Nat. Med. 23, 405–408
with recurrent neural networks. ACS Cent. Sci. 4, 120–131 59 Hernandez, J.J. et al. (2017) Giving drugs a second chance: overcoming regulatory
28 Huang, Z. et al. (2017) Data mining for biomedicine and healthcare. J. Healthc. Eng. and financial hurdles in repurposing approved drugs as cancer therapeutics. Front.
2017 http://dx.doi.org/10.1155/2017/7107629 Article ID 7107629, 2 pages Oncol. 7, 273
29 Zhang, Y. et al. (2017) Computer-aided clinical trial recruitment based on domain- 60 Lozano-Diez, A. et al. (2017) An analysis of the influence of deep neural network
specific language translation: a case study of retinopathy of prematurity. J. Healthc. (DNN) topology in bottleneck feature based language recognition. PLoS One 12,
Eng. 2017, 7862672 e0182580
30 Mamoshina, P. et al. (2016) Applications of deep learning in biomedicine. Mol. 61 Aliper, A. et al. (2016) Deep learning applications for predicting pharmacological
Pharm. 13, 1445–1454 properties of drugs and drug repurposing using transcriptomic data. Mol. Pharm. 13,
31 Seddon, G. et al. (2012) Drug design for ever, from hype to hope. J. Comput. Aided 2524–2530
Mol. Des. 26, 137–150 62 Kadurin, A. et al. (2017) druGAN: an advanced generative adversarial autoencoder
32 Merk, D. et al. (2018) De novo design of bioactive small molecules by artificial model for de novo generation of new molecules with desired molecular properties
intelligence. Mol. Inform. 37, 1700153 in silico. Mol. Pharm. 14, 3098–3104
33 Klopman, G. et al. (2004) ESP: a method to predict toxicity and pharmacological 63 Galbusera, F. et al. (2018) Exploring the potential of generative adversarial networks
properties of chemicals using multiple MCASE databases. J. Chem. Inf. Comput. Sci. for synthesizing radiological images of the spine to be used in in silico trials. Front.
44, 704–715 Bioeng. Biotechnol. 6, 53
34 Menden, M.P. et al. (2013) Machine learning prediction of cancer cell sensitivity to 64 Ozerov, I.V. et al. (2016) In silico pathway activation network decomposition
drugs based on genomic and chemical properties. PLoS One 8, e61318 analysis (iPANDA) as a method for biomarker development. Nat. Commun. 7, 13427
35 Nascimento, A.C.A. et al. (2016) A multiple kernel learning algorithm for drug- 65 Ryu, J.Y. et al. (2018) Deep learning improves prediction of drug–drug and drug–
target interaction prediction. BMC Bioinf. 17, 46 food interactions. Proc. Natl. Acad. Sci. U. S. A. 115, 201803294
36 Schneider, G. (2017) Automating drug discovery. Nat. Rev. Drug Discov. 17, 97–113 66 Swan, A.L. et al. (2015) A machine learning heuristic to identify biologically relevant
37 Matthews, H. et al. (2016) Omics-informed drug and biomarker discovery: and minimal biomarker panels from omics data. BMC Genomics 16 (Suppl. 1), 2
opportunities, challenges and future perspectives. Proteomes 4 http://dx.doi.org/ 67 Krittanawong, C. (2018) The rise of artificial intelligence and the uncertain future
10.3390/proteomes4030028 for physicians. Eur. J. Intern. Med. 48, e13–e14
38 Hamet, P. and Tremblay, J. (2017) Artificial intelligence in medicine. Metabolism 69, 68 Powles, J. and Hodson, H. (2017) Google DeepMind and healthcare in an age of
S36–S40 algorithms. Health Technol. 7, 351–367
39 Hughes, J.P. et al. (2011) Principles of early drug discovery. Br. J. Pharmacol. 162, 69 Lee, E.-J. et al. (2017) Deep into the brain: artificial intelligence in stroke imaging. J.
1239–1249 Stroke 19, 277–285
40 Mohs, R.C. and Greig, N.H. (2017) Drug discovery and development: role of basic 70 Haghi, M. et al. (2017) Wearable devices in medical internet of things: scientific
biological research. Alzheimers Dement. 3, 651–657 research and commercially available devices. Healthc. Inform. Res. 23, 4–15
41 Katsila, T. et al. (2016) Computational approaches in target identification and drug 71 Curioni-Fontecedro, A. (2017) A new era of oncology through artificial intelligence.
discovery. Comput. Struct. Biotechnol. J. 14, 177–184 ESMO Open 2, e000198
42 Emig, D. et al. (2013) Drug target prediction and repositioning using an integrated 72 Lexchin, J. (2012) Models for financing the regulation of pharmaceutical
network-based approach. PLoS One 8, e60618 promotion. Global Health 8, 24
43 Duch, W. et al. (2007) Artificial intelligence approaches for rational drug design and 73 Martin, L. et al. (2017) How much do clinical trials cost? Nat. Rev. Drug Discov. 16,
discovery. Curr. Pharm. Des. 13, 1497–1508 381–382
44 Wang, Q. et al. (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

780 www.drugdiscoverytoday.com

You might also like