Machine Learning and Artificial Intelligence in Therapeutics and Drug Development Life Cycle

Abstract

In recent years, the pharmaceutical business has seen a considerable increase in data digitization. With digitization, however, comes the challenge of obtaining, analyzing, and applying knowledge to solve complex clinical problems. Artificial intelligence (AI), which entails a variety of advanced tools and networks that can mimic human intellect, can overcome such challenges with traditional pharmaceutical development. Artificial intelligence and machine learning have a vast role in therapeutic development, including the prediction of drug target and properties of small molecules. By predicting the 3D protein structure, AI techniques, such as Alpha Fold, can help with structure-based drug development. Machine learning algorithms have been utilized to anticipate the properties of small molecules based on their chemical structure. Many researches have shown the importance of using in silico predictive ADMET (absorption, distribution, metabolism, excretion, and toxicity) models to speed up the discovery of small compounds with enhanced efficacy, safety, and dosage. This chapter discusses various roles of these methods in the development of effective therapeutics.

1. Introduction

The goal of the drug development process is to find bioactive molecules that can help in disease therapy [1, 2]. The drug discovery life cycle has multiple steps: drug target identification, target validation, hit identification, lead optimization, preclinical development, clinical trial, approval, and postmarketing monitoring (Figure 1A). Because going through all of these stages of developing a new drug can cost between $1 and $2 billion and take 10–17 years, drug discovery is a big issue in the pharmaceutical business [3]. To speed up the drug discovery process, a considerable number of developments were made in the 1990s using combinatorial and high-throughput screening (HTS) approaches. These approaches were widely used since they allowed for the quick synthesis and screening of vast libraries, but no meaningful success was achieved, and little progress was made toward the discovery of new compounds. To aid the discovery process, a combination of modern computer approaches, biological research, and chemical synthesis was developed, and this combined approach increased the scope of discovery. The term “computer-aided drug design” (CADD) was eventually used to describe the use of computers in drug discovery. Computer-aided drug design (CADD) is one of the most widely utilized approach for reducing drug development costs and time. CADD is a specialized field, in which various computational approaches are employed to mimic receptor–drug interactions in order to identify binding affinities (Figure 1B). The approach, however, is not just for studying chemical interactions and predicting binding affinity; it can be used for everything from designing compounds with desired physiochemical features to managing digital databases of chemicals. CADD is a wide term that encompasses both structure- and ligand-based drug developments. Virtual screening (VS) is a computational method for screening large databases of compounds that has successfully supplemented HTS in drug discovery. The fundamental purpose of VS is to make it feasible to quickly and cheaply screen enormous virtual chemical databases for potential leads for synthesis and future study [4]. Computer-assisted virtual screenings have been a widely used method for estimating various types of ligands to bind with target over time [5, 6, 7, 8]. Additionally, in order to investigate atomistic level of protein/nucleic acid-small molecule interactions, one of the widely utilized computational biophysics tools is molecular dynamics (MD) simulation [9, 10]. MD simulation finds its relevance in shedding lights on the conformational ensembles either of the small molecule or of the target. The technique is seldom utilized to capture the dynamics of proteins and/or to check the stability of modeled protein structures enabling CADD for designing efficient inhibitors [11, 12]. It is also leveraged to investigate a comparative binding of small molecule to different proteins along with complementing experimental observations [13].

Figure 1.

Flowchart of events taken place in (A) drug discovery lifecycle and (B) computer aided drug discovery.

The recent expansion of make-on-demand libraries to billions of synthesizable molecules has piqued the interest of the drug-discovery community, as such massive databases allow access to previously unexplored chemical realms. The introduction of ultralarge libraries, on the other hand, has revealed substantial limitations of traditional docking techniques, which typically work on the scale of millions of molecules at a high cost of computation. This aspect depicts CADD as a very useful process, in which only a small portion of the highest-scoring compounds often leaving low scoring but potential compounds are considered for experimental examination. Artificial intelligence (AI) and machine learning (ML) aided approaches provide a low-cost, high-reliability solution to a variety of problems (Figure 2), from protein three-dimensional (3D) structure prediction to physiochemical property calculation and bioactivity prediction to ultralarge docking [14, 15].

Figure 2.

Biomolecular analysis core scheme for drug discovery via AI/ML.

2. Artificial intelligence (AI) and machine learning (ML)

The ultimate goal of artificial intelligence (AI) is to train computer programs with human-like intellect. For this, AI uses computers to learn human behaviors, such as learning, judgment, and decision-making by simulating human intelligent behavior with computers. The term artificial intelligence was first proposed in 1956 at a conference at Dartmouth University; however, the major AI-related research started since the end of the twentieth century. AI has provided enormous economic benefits to humanity and has helped all parts of life, while also considerably promoted social growth and ushered in a new era of social development [16, 17]. Both the volume and the multidimensionality of data have increased dramatically as a result of the advent of numerous high-throughput technologies. Big data is both a requirement and a key component for AI to improve its recognition rate and accuracy [16].

Machine learning (ML), a branch of AI, is the use of an algorithm that improves its performance by learning from data. Machine learning, according to Arthur Samuel, is described as a computer's ability to analyze without being explicitly programmed [16, 18]. Supervised learning, unsupervised learning, semisupervised learning, and reinforcement learning are the four types of machine learning algorithms [16]. In supervised learning, the test data are trained with a labeled dataset to predict the type or value of new data, whereas unsupervised learning uses unlabeled data based on the input pattern. Support vector machine (SVM), linear discrimination, and decision tree are some of the types of supervised learning algorithms, whereas k-clustering and principal component analysis are the examples of unsupervised learning algorithms. Semisupervised learning combines the benefits of both supervised and unsupervised learning. It can be useful if there exist unlabeled data and collecting the labeled data is a time-consuming procedure. Reinforcement learning seeks to solve a problem through a hit-and-trial strategy, including feedback and decisions, with the ultimate goal of increasing total reward [16, 18, 19]. Deep learning (DL), a subset of machine learning, is one of the most cutting-edge areas of research and development in practically every scientific and technical discipline. Many problems that normal ML algorithms could not solve, such as image recognition and speech recognition, can be solved with the help of DL methods. DL methods also have immense role in the drug discovery pipelines, including drug activity prediction, target identification, and lead molecule discovery. The foundations of DL are frequently implicated in neural network systems, where they are employed to develop systems capable of complicated data recognition, interpretation, and production [20].

3. AI and ML in protein structure modeling

Drug design is based on the idea of creating compounds with a regulated interaction profile against a variety of target and off-target proteins in an organism. To understand the mode of action of a candidate drug, three-dimensional (3D) details are of paramount importance. Despite the availability of a variety of experimental methods to decipher the 3D structure of proteins, such as X-ray crystallography, nuclear magnetic resonance (NMR), and cryo-electron microscopy, the sequence–structure gap, in which protein sequences vastly outnumber the number of corresponding 3D structures, frequently causes problems. In such cases, protein structure prediction methods come as a remedy [21, 22].

Over the past 30 years, AI and ML methods have been used to predict protein structure. AI programs have helped to assess and identify most accurate models. To compare the predicted models to known crystal structures, these programs are trained utilizing numerous numerically represented atomic parameters from the models, such as bond lengths, inter-residue interactions, physiochemical properties, and so on. The Critical Assessment of Protein Structure Prediction (CASP) contests have been held biannually since 1994 for the blind evaluation of cutting-edge methods for predicting three-dimensional (3D) protein structures from protein sequences [23, 24]. For the cases where a template is not available for modeling, two approaches are considered: fragment-based assembly and ab initio folding. Fragment-based assembly is advantageous than ab initio folding due to its higher accuracy and higher capability [25].

With near-experimental precision, Alphabet's DeepMind won the 13th edition of CASP in 2018 with its latest artificial intelligence (AI) system, AlphaFold [25]. The 3D structure prediction by assembling the most probable fragments by AlphaFold is done by using co-evolution analysis of a multiple sequence alignment and using deep neural networks (DNNs) to discover coevolutionary patterns in protein sequences as contact distributions and transform them into protein-specific statistical energy potentials [23]. DeepFragLib, a fragment library constructed utilizing deep contextual learning techniques to give high-quality, native-like fragments for every segment of a protein for the efficient assembly of near-native conformations, is another example of AI breakthrough in the field of structure prediction. Table 1 represent applications of some AI-ML approaches for protein structure prediction/quality assessment.

S. No.	Tools	Description	Web-link
1	AlphaFold [26]	Protein structure prediction using AI	https://github.com/deepmind/alphafold
2	DeepFragLib [27]	Fragment library construction software by DNN	https://github.com/ElwynWang/DeepFragLib
3	ProQ3/ProQ3D [28]	Protein quality assessment using deep learning	https://proq3.bioinfo.se/
4	QACon [29]	Protein model quality assessment using ML techniques	https://swmath.org/software/34249
5	DeepQA [30]	Protein model quality assessment using deep belief networks	https://swmath.org/software/15927
6	DEFMap [31]	DL-based method for extracting the dynamics associated with atomic fluctuations concealed in cryo-EM density maps	https://github.com/clinfo/DEFMap

Table 1.

Few examples of applications of AI–ML methods in protein structure methods.

4. ML and AI in physiochemical property calculation

In the biopharmaceutical sectors, the effective and precise forecasting of molecular characteristics of drug compounds is indeed a fundamental component of rationalized compound synthesis. Current techniques span from basic atom summation through bond energy additions, paired interatomic configurations, and more sophisticated machine learning systems capable of representing aggregate reactions among several particles or bonds. In addition, simple correlation force fields show predictive performance comparable to reference energy sources determined utilizing density functional theory with hybrid exchange-correlation functional for steady-state geometric models; even so, properly accounting for the collaborative many-body connections is required for advancing the “magic formula” of compound accuracy of 1 kcal/mol for both the steady-state and out-of-equilibrium topologies [32]. In the years 2010–2012, the initial machine learning (ML) methods for molecular modeling relied on tiny datasets with quantum mechanical (QM) features for 102–103 molecule systems. It is believed that the chemical compound space has 1060–10,100 molecular systems. Chemical spaces have grown in size and complexity during the previous decade. Data are being generated at an astonishing rate owing to large-scale QM and MD methodologies, as well as developments in high-throughput studies [33].

Machine learning models predict small molecule’s properties based on their chemical structure. Because of their ease of interpretation and effectiveness on small datasets, linear models were initially used. However, over time, nonlinear models were developed to capture more complex relationships between structure and activity. The nonlinear approaches include support vector machines, recursive partitioning methods, and deep learning methods. With the availability of standardized large-scale data, deep-learning-based techniques for ADMET (absorption, distribution, metabolism, excretion, and toxicity) prediction are showing growing promise and utility. The ability to identify small compounds with increased efficacy, safety, and dosage is considerably aided by understanding ADMET characteristics. In terms of consistency and predictive performance, the graph convolutional DNN (GCNN) technique is proposed to be superior to existing approaches such as random forest (RF), Cubist, and support vector machine (SVM) for calculating ADMET characteristics [34]. The AI-based ADMET predictors utilize cellular permeability data from a diverse class of molecules generated by different cell lines. To predict acid dissociation constant of compounds, artificial neural network (ANN)-based models, graph kernels, and kernel-ridge-based models have been used. To predict the solubility of the compounds, undirected graph recursive neural networks and GCNN have been used. GCNN methods are also used to predict cytotoxicity, which is one of the important properties used in drug discovery to avoid toxic effects [35]. The latest ANN studies support immunoinformatics and chemoinformatics analysis for novel vaccine and drug discovery [11]. Examples of the AI-based tools for molecular property calculation are DeepTox (www.bioinf.jku.at/research/DeepTox), Chemputer (https://zenodo.org/record/1481731), and ORGANIC (https://github.com/aspuru-guzik-group/ORGANIC) [35].

5. ML and AI in bioactivity prediction

In order to prioritize compounds for synthesis and/or biological evaluation, quantitative structure-activity relationship (QSAR) modeling has been used [36]. The goal of QSAR models is to find a mathematical relationship between the physicochemical qualities of substances, which are represented by molecular descriptors, and their biological activity. These models are important in drug optimization because they provide a preliminary in silico assessment of key qualities such as activity, selectivity, and toxicity of candidate compounds. In QSAR modeling, AI/ML techniques (such as RF, SVM, Naïve Bayesian, and ANN) have been widely used. Among these techniques, the RF algorithm has been regarded as a gold standard in QSAR studies [23]. In the case of bioactivity prediction, DL approaches have shown improved performance compared with ML [35]. Few examples of AI-based tools to determine bioactivity are WDL-RF (integration of DL and RF) (https://zhanglab.ccmb.med.umich.edu/WDL-RF/), pairwiseMKL (multiple-kernel-learning-based method) (https://github.com/aalto-ics-kepaco), and DeepMalaria [37] (DL based).

6. ML and AI in drug–target interactions

To fully comprehend a drug's efficacy and usefulness, it is important to determine how it interacts with a receptor or target. Drug-protein interactions have recently been a hot topic in drug repurposing research [38]. ML algorithms have become the advanced approach for estimation of drug–target interactions due to the huge amount of obtainable drugs and target information in huge datasets, advancing as well as innovative computer networking, and inherent characteristics of different types of deep learning. A vast number of proteins have indeed been sequenced, and numerous molecules have now been synthesized since the advent of sequencing technology, high-throughput technologies, and computer-aided drug design methods. Actual information has been organized, and multiple databases have been developed based on existing related efforts and acquired expertise. The majority of data in these sources is open to the public and free to download, thus providing a strong data basis for using deep learning to solve drug-target contact predictions issues. PubChem presently comprises 109 million chemicals and is the world’s biggest database with open access to chemical characterization. PubChem has grown in importance as a source of chemical knowledge for researchers, learners, and the general public. Artificial intelligence can be used to train deep learning models for drug discovery using known drug data [39]. Several ML techniques have been used to predict drug–target interactions including SVM, DL, DNN, convolutional neural network (CNN), etc. The de novo drug design approach has been frequently employed to create therapeutic compounds in recent years. The old approach of de novo drug design is being phased out in favor of emerging DL methodologies, which have the advantages of less complicated synthesis routes and easier prediction of novel molecule bioactivity [35]. However, classical algorithms cannot be completely ignored, as studies point out that the classical algorithms show higher and more stable performance than the machine-learning-based methods at different similarity levels of training sets. Hence, many tools have been developed combining both classical and deep learning models [40]. Few of the tools and servers available for finding drug-protein interactions using AI and ML methods are mentioned in Table 2.

S. No.	Tools	Description	Web-link
1	Ligdream [41]	For de novo drug design through generative shape-based neural network decoding	https://playmolecule.com/LigDream/
2	WADDAICA [40]	Uses both deep learning model and classical algorithms for drug design	https://heisenberg.ucam.edu:5000/
3	MolAICal [42]	Uses both deep learning model and classical algorithms for drug design	https://molaical.github.io/
4	OpenChem [39]	A deep learning toolkit for computational chemistry	https://github.com/Mariewelt/OpenChem
5	DeepAffinity [43]	A combination of RNN and CNN methods for ligand–protein affinity	https://github.com/Shen-Lab/DeepAffinity
6	DeepFrag [44]	Uses deep CNN for fragment-based lead optimization	https://durrantlab.pitt.edu/deepfragmodel/

Table 2.

Few examples of applications of AI–ML methods in drug–protein interactions.

7. Conclusion

Although AI is frequently portrayed as a magic wand that can provide flawless output regardless of the quality of the input, it is not the solution to every problem. The ultimate goal of using AI and machine learning approaches to drug development challenges is to bring the best drugs to market. Throughout the drug discovery process, the combined effort of different AI methods allows for a better understanding and design of novel inputs [45]. The AI-based applications are getting more intelligent, cost-effective, and time-efficient while increasing efficacy, because of more precise algorithms, more powerful supercomputers, and significant private and public investment in the sector [20]. To properly leverage AI in drug development, one must increase the quality of decisions we make regarding compounds that are progressed to clinical trials. However, in many circumstances, the data available to make those decisions are not totally sufficient for this purpose [46]. Since the entire success of AI depends on the availability of a substantial amount of data, we need to conduct trials more efficiently, which can be supported by computational methods [35, 46]. Major challenges faced by AI methods include data accuracy and availability, reproducibility, model appropriateness, etc. Despite the challenges, AI is projected to advance the field of personalized/precision medicine to the point where it becomes regular practice even in the treatment of minor illnesses in the future [47]. By 2028, AI is expected to save the pharmaceutical industry more than US$70 billion in drug discovery costs [48]. With more clinical data and improved AI calculations, AI is projected to improve many elements of drug discovery and development and will eventually become the standard computer-assisted technique for drug discovery.

Acknowledgments

ANJ would like to acknowledge Department of Science and Technology (DST-SERB File no. CRG/2020/001829), Government of India, for providing computational facilities.