A review of the current trends in computational approaches in drug design and metabolism

For a long time, plants have been known to have therapeutic properties, and most plant-derived medicines have been used to manage various pathological conditions [1, 2]. While these medicines have been effective when used as concentrated plant extracts or concoctions, modern medicine requires the isolation and purification of active compounds [3]. Despite the identification of drugs for treating hypertension, malaria, HIV/AIDS, cancer, and diabetes, these diseases continue to pose global health concerns with modern medicine struggling to find sustainable cures [4]. Research activities on medicinal chemistry show that antimalarial drugs such as Artemisinin (Artemisia annua), quinine (Cinchona spp.), and anticancer drugs such as Vinblastine (Catharanthus roseus) and Taxol (Taxus brevifolia) were discovered from natural products and were effective in managing and treating serious health problems and diseases [5, 6]. Similar reports show that natural products are relevant in drug discovery for communicable and non-communicable diseases [1]. By leveraging the recent advances in technology, the success rate for new therapeutic moieties will be increased, hence solving global health challenges. Modern drug discovery methods in the realm of natural products stand as a major discovery in achieving the United Nation sustainable development goals (SDGs) on health [7]. In particular, aspects of drug discovery and development are important translational science activities that contribute to SDG number 3 (good health and well-being) and number 9 (sustainable innovation and industrialization) [8]. The urgency of the provision of affordable essential vaccines and medicine in accordance with the Doha Declaration on the Trade-Related Aspects of Intellectual Property Rights (TRIPS) Agreement and Public Health cannot be overstated within the context of sustainable healthcare [9].

Conventional drug design and discovery methods are well known to be cost-intensive and time-consuming, taking an average of 10 to 15 years and cobbling more than US$2 billion before they reach the market [10]. The long-standing impediment in drug design and discovery is the reproducibility of the experimental designs [11,12,13]. With the rapid development of computer algorithms, hardware, and software, drug development has significantly benefited from vast computational methods that have drastically reduced drug discovery time. Studying biological systems requires techniques that provide substantial information within a limited period of time. More prominently, computational tools have come into the limelight to represent essential and valuable tools in the discovery of chemical compounds with potential therapeutic effects [14,15,16]. Several computational tools have tremendously helped researchers identify and optimize new drug candidates and predict their potential interactions. They have allowed for the investigations of thermodynamic and structural features of target proteins on different levels, which helps study drug action mechanisms and drug binding sites [17].

The high failure rate of the traditional pathways for drug discovery and design has prompted the scientific community to use computer-aided drug discovery (CADD), which involves system-based, structure-based and ligand-based drug strategies [18, 19]. The scientific progress involving CADD begins with wet-lab experiments, which comprise the identification of the hit or target compound and subsequently through high-throughput screening (HTS). Typically, the side effects associated with the use of drugs necessitate the screening of adsorption, distribution, metabolism, excretion, and toxicity (ADMET) properties at early stages [20,21,22]. Accordingly, the CADD screening of library compounds narrows drug candidates into a smaller cluster, thereby increasing the success rates and reducing the time used in screening potential drugs with therapeutic impacts. However, CADD screening is associated with high-resource requirements and extensive costs [23], prompting the use of virtual screening methods, including molecular docking, whereby a known target is screened against a virtual library of compounds [24, 25]. The first step in drug design usually identifies the target molecule (a protein of a biochemical pathway associated with the disease). The lead compounds are then designed or identified to promote or inhibit the biochemical pathway [26]. Another crucial step in drug design and discovery is lead optimization so as to achieve maximum interaction between the target molecules. Consequently, CADD plays a central role in lead optimization [27].

As noted in Fig. 1, CADD is broadly classified into ligand-based and system-based approaches. The ligand-based approach exploits the experimental nature of inactive/active molecules. This strategy needs only bioactivity data and the structural information of small molecules rather than prior knowledge of mechanisms of action. In a ligand-based approach, similar molecules likely have similar properties [28]. Numerous studies have reported that in the absence of the experimental three-dimensional (3D) structure, a ligand-based approach is used to understand the physicochemical and structural properties of the ligand associated with the desired pharmacological activity [29, 30]. Besides the known ligand molecules, ligand-based methods are also applied to substrate analogues or natural products that interact with target molecules to yield the required pharmacological effects [30]. Similarly, structure-based methods such as in silico chemical alteration or molecular docking are used in the presence of 3D structures for lead optimization [29]. Structure-based methods help identify the structural requirements of the ligands and the nature of target-ligand interactions to optimize the interaction [30]. Figure 1 presents a summary of drug discovery processes and the corresponding computation tools used.

Drug discovery process and identification using various computational methods [11, 28]

Full size image

Multiscale simulations of biomolecular systems used in drug design depend on the biological problem being investigated. For instance, a combination of molecular mechanics (MM) and quantum mechanics (QM), (QM/MM), have been more prominently used to simulate chemical reactions such as enzyme catalysis, calculate spectra, and study electronic properties in a single simulation [31, 32], which can be used to elucidate drug action mechanisms [33, 34]. Another popular biomolecular simulation technique is molecular dynamics (MD), which applies molecular mechanics force fields and is based on classical Newtonian physics. Typically, MD simulations have been used to calculate the binding free energy between the drug molecules and target proteins to identify the potential drug binding sites and drug molecule action mechanisms [35, 36].

Further to this, depending on different accuracy requirements, coarse-grained (CG), united-atom (UA), and all-atom (AA) MD simulations, as well as implicit/explicit solvent models, have been used in simulations of spatial and temporal scales in drug design [38, 39]. Various biomolecular simulation methods are required to work on a specific problem. Accordingly, each of these methods is associated with practical limitations, disadvantages, and advantages in terms of the type of phenomena that can be modelled, the size of the biomolecular system that can be modelled, and the length of simulation required [38, 40]. In drug design and discovery, the scientific understanding of the action mechanism of a drug, the dynamics of drug targets, and the elucidation of binding sites on the target protein require multiple simulation methods for the analysis of cross-scale connections [38]. Figure 2 shows computational strategies, including MADE prediction, QM/MM, structure and ligand-based structure designs, protein structure design, and MD.

The computation resources used in predicting molecules with potential therapeutic effects [37]

Full size image

A significant thrust in most preclinical drug discovery research activities is the identification of orally efficacious and safe drug candidates amenable to low daily dosing. In this pursuit, drug metabolism has been critical in identifying chemical compounds with optimal ADMET characteristics [41, 42]. Research efforts to reduce drug metabolism are not limited to introducing conformational constraints, altering stereochemistry, modulating electronic and steric factors, and reducing lipophilicity [43]. Computational resources are doubtlessly pertinent to drug discovery, and researchers are always excited to study new kinds of interactions [44]. In view of this, docking using software such as AutoDock, Glide, and GOLD–is first conducted, and then MD simulations are performed to gain insights into the effects of interactions at a temporal scale [45,46,47]. Despite the wide range of computational methods, a free energy perturbation (FEP), in combination with Monte Carlo (MC) with MD, is an attractive option because it offers a rigorous theoretical means to calculate changes in binding free energy [48, 49].

These computational tools can be synergistically employed in pre-clinical drug design and discovery. For instance, Amsterdam density functional (ADF) and DFT are used to study reactivity, molecular properties, and electronic structures of molecules [50, 51]. Conversely, MM simulations are pivotal in studying the dynamics involved in protein–ligand interactions [52]. MD simulations help to predict the binding modes of ligands to target proteins [53]. On the other hand, hybrid methods like QM/MM combine efficiency and accuracy in studying enzyme-catalyzed reactions [54]. The objective of this review is to evaluate the current trends in using robust computational tools in pre-clinical drug discovery. This review also aims to explore the current use of ADF and DFT, MD, MM (Monte Carlo Simulations and MD), and hybrid methods (QM/MM and Quantum MC) which are remarkable computational tools in drug design and discovery. Also, the review provides an overview on the use of artificial intelligence in drug design and discovery. We believe this contribution will provide valuable insights into the use of computational resources in drug design and development.

The methodology used in this review paper entails a thorough examination of research studies on relevant literature on drug design and development using computational resources. By embracing diverse viewpoints, this work provides an in-depth summary of computational resources that have been used in the identification of potential compounds with therapeutic properties. In addition, this paper offers a concise synthesis of numerous research findings in this field, showcasing the progress that has been made in leveraging emerging technologies such as AI, machine learning, and deep learning in medicinal chemistry. A systematic literature was conducted on research reports and empirical and theoretical studies that focused on drug discovery and development. Extensive searches using Scopus, International Pharmaceutical Abstracts (OvidSp, WHO Global Health Library, Cochrane, Google Scholar, Web of Science, ScienceDirect, ProQuest dissertation & thesis, Worldwide Political Science Abstracts (CSA), and PubMed was carried out. The identified studies were cross-checked to ensure that they were peer-reviewed, written in English, and published in scientific journals. Due to the long history of drug design and development, search strategies combining keywords “drug discovery and development” with concepts of “drug metabolism”, “artificial intelligence”, “machine learning”, “computational”, “in silico methods” or “theoretical” were used; nevertheless, elaborate details of the search strategies were available on request. The search keywords and terms were also not restricted to “docking”, “therapeutic agents”, and “virtual screening”.

Over 800 scientific papers in the medical field, published between 1990 and 2024 were identified for this review paper but only 263 were relevant. Papers were selected in two phases using pre-defined inclusion and exclusion criteria. Oftentimes, the titles of literature studies are vague, thus, in the first phase we selected studies by reading abstracts and titles. Abstracts and titles that mentioned drug design and development, in silico methods, metabolism, and pharmacokinetics, were included. Studies that combined studies on drug discovery, computational drug design, drug metabolism, AI-assisted drug design, and drug design were included. If the abstract was not available electronically, the articles were passed along to the second phase of full-text review. A standardized template was used to ensure that the selected papers met the inclusion criteria. Owing to the long-history of medicinal chemistry and computational tools, year of publication was not strictly adhered to. The template collected data on drug discovery studied using theoretical/computational methods, years of publication, and drug classes. Studies that only focused on protein structure prediction or traditional medicines without drug development projects or those that examined wet laboratory approaches describing drug discovery projects also met the inclusion criteria. The search results were then imported into EndNote Version X8, Clarivate Analytics, USA.

2.1 Systematic approach to data synthesis

Two authors extracted data from the publications that met the inclusion criteria. The data included first author, year of publication, computational method used in drug discovery, molecular docking tool, and metabolism in drug design, potential therapeutic candidates, the visualization tool used, in silico approach employed in pre-clinical drug design trials, drug pharmacokinetics, and AI in drug discovery. In case of any discrepancies or disagreements, the two authors discussed and sought consensus.

There are two drug discovery strategies–drug re-positioning and de novo drug discovery strategies. Drug repurposing refers to an approach of identifying new uses of the already existing products [55]. Drug repurposing (also called drug repositioning) has various inherent advantages including cost efficient and faster drug development time due to prior knowledge toxicity profiles, dosage, and safety of existing medications [56]. The interest of drug repositioning has risen in recent times. For instance, thalidomide which was formally known to cause birth defects has been repurposed to treat multiple myeloma and leprosy [57, 58]. Similarly, bufulfone and chlorambucil which were initially developed as alkylating agents based on toxic chemical warfare agent mustard gas, have been repurposed for treating leukemia [55]. Thus, drug repositioning may be a viable and compelling strategy for enhancing treatment options for various diseases. conversely, de novo drug design involves creating a pool of drug candidates for sequential analysis in order to create “anew”, “from the beginning” or “afresh” molecules with desirable therapeutic properties from scratch [59].

Similarly, drug research and development follow three experimentation pathways: in silico, in vivo and in vitro [60]. In silico refers to the studies performed via computer simulation while in vitro refers to the studies performed with biological molecules, cells or microorganisms in flasks and test tubes [61]. These experimentations are performed outside living organisms. On the other hand in vivo experimentations are mainly involved clinical trials and animal studies [62]. Accordingly, in silico study is easy occasioned with the rapid development of numerical algorithms, ability of pharmaceutical companies to reproduce human environments, the advancement of computational power and medical imaging.

Novel computational drug design methods, better machine algorithms, and faster architectures of Graphic Processing Unit (GPU)-based clusters that have been developed until recently have significantly helped researchers perform high-level computations, aiding in the discovery of promising drug candidates by improving arears such as scoring functions [63, 64]. Undoubtedly, computational strategies in drug discovery have uncovered numerous therapeutic undertakings. In particular, advanced in nuclear magnetic resonance spectroscopy, advances in protein purification and structural analysis, and crystallography stand out as successes that have scaled the drug-discovery process. Despite the tremendous research undertakings in the academic environment for concept validation, there are a variety of obstacles that limit their application. Accordingly, this work compiles succinct discussion on developments and challenges on in silico approaches in drug discovery and discovery, hybrid methods based on QM/MM, drug metabolism, machine learning, and AI in pre-clinical drug research.

3.1 In Silico approaches in pre-clinical drug design

In Silico approaches have been successfully applied to the in vivo prediction of ADMET and have been developed to replace in vivo models in the study of pharmacokinetics [65]. Two widely used in Silico methods for ADMET prediction are molecular modelling and data modelling [65]. Molecular modelling is based on 3D protein structures and includes techniques such as QM simulations, MD simulations, pharmacophore modelling, and molecular docking calculations [66,67,68]. There is some evidence that the computational ability of molecular modelling can complement or even surpass the quantitative structure–activity relationship (QSAR) as more 3D proteins become available [69]. It is still a scientific challenge to predict ADMET properties using molecular modelling because of the large binding cavities and the flexible nature of proteins. Nonetheless, other studies have successfully used molecular modelling to predict metabolism [70].

3.2 Applications of molecular dynamics in drug design

Molecular dynamics (MD) is a computational resource that employs Newtonian second law of motion to evaluate motions of small molecules, ions, water, complex systems or macromolecules to represent the biological environment, including lipid membranes and water molecules [35, 71]. Therefore, MD simulations have been widely used to study drugs against specific proteins, considering the movement of proteins in a solvated environment [35]. These simulations generate data on conformational, energetics, and binding free energy changes, proving useful in drug design [71].

The MD simulations have been partnered with experiments to understand the underlying behaviour of biological and chemical systems in their ever-growing role in drug discovery [49]. Notably, MD simulations appear destined to impact drug discovery significantly [35, 52, 72]. MD simulations follow virtual screening simulations. It is considered an advanced technique that complements docking, and most research scholars have used it to complement virtual screening. MD employs Newtonian mechanics in improving the efficacy and the binding properties of lead compounds [44]. For this reason, researchers have popularly used MD to confirm the validity of the docked results. Figure 3 presents the steps involved in MD simulations.

Schematic representation of molecular dynamics simulations [44]

Full size image

The computational chemistry methods based on MD are used to study various types of macromolecules – carbohydrates, nucleic acids, proteins – of medicinal or biological interest [73]. They are well-suited to the study of membrane proteins and their behaviour. In MD simulations, dihedral angles are depicted using a sinusoidal function, while bond angles and chemical bonds are handled using simple virtual springs. MD applications, including free energy perturbation methods (FEP), linear interaction energy (LIE), and molecular mechanics Poisson-Boltzmann surface area (MM/PBSA), are used to calculate free energy to correlate calculated and experimental binding affinities of small molecules of proteins [74, 75].

MD simulations employ force fields like CHARMM, GROMOS, Amber, OPLS, and coarse-grained (CG) [28]. These force field models include Assisted Model Building with Energy Refinement (AMBER) which was developed in the 1970s by a group of scholars led by Peter Kollman at the University of California, San Francisco [76]. This toolkit has been commonly used to explore the dynamic behaviour of carbohydrates, organic molecules, lipids, RNA, DNA, and proteins, among other macromolecules [77]. AMBER has the computational ability to compute the potential energy of a molecular system. It uses water models such as OPC, PIP4PEW, SPC/E, and TIP3P to represent the solvent environment when performing biomolecular simulations [28]. Another pivotal program in MD simulations is CHARMM (Chemistry at Harvard Macromolecular Mechanics), which Martin Karplus and colleagues at Harvard University developed. CHARMM accurately describes molecular interactions of complex biomolecular systems [78]. Another force field model is optimized potentials for liquid simulations (OPLS), which offers accurate prediction of thermodynamic and structural properties obtained by MC computations compared to experimental results [79]. OPLS-AA was developed to reproduce QM conformational energy profiles of small molecules, whereas Groningen Molecular Simulation (GROMOS) is a multipurpose computational resource used to study biomolecular systems, including sugars, nucleotides, and proteins [73]. GROMOS has been commonly used by researchers to model polymers, liquid crystals, solutions of biomolecules, and glasses. Coarse-grained (CG) is a popular computation approach that simplifies the cost of computational calculations by reducing the number of degrees of freedom in the model. Similarly, CG Martini has the computational ability of simulating large-scale systems [28, 73]. These force fields have been summarized in Table 1.

3.3 Application of molecular docking in drug design

Molecular docking encompasses three key objectives: binding affinity estimation, pose prediction, and virtual prediction [103]. Reliable molecular docking techniques must differentiate non-binding sites from binding sites and their molecular interactions [47]. Besides, these techniques must also classify molecules as non-binding or binding and rate binding molecules among the best compounds when dealing with large compound libraries. The success of virtual screening relies on the accuracy and the amount of structural information known about the ligand and the target protein being docked. The first step in molecular docking is the analysis of protein–ligand to identify novel binding pockets using in silico methods such as PATCH-SURFER [104], AFT, POCKET SURFER [105], catalytic site Atlas [106], and SURFACE. Molecular docking based on virtual screening is the most preferred, and a helpful technique if the binding site of the target biomolecule is known [107, 108]. Therefore, ligands with inhibitory potential against the target protein could be docked against it in search of novel therapeutic candidates. Further to this, blind docking techniques could also be employed when the binding site in a protein is not known [109]. However, blind docking is associated with downsides such as low success rates and high time requirements compared to docking into a known binding pocket [47].

Accuracy and speed are key features required for a successful molecular docking simulation. Typically, docking algorithms are developed to speed up the discovery of novel lead compounds in virtual screening or to validate experimental data at higher accuracies [107, 110, 111]. There are various docking programs: Zdock [112], GOLD [113], MSDOCK [114], FLExX [115], AUTODOCK [116], MOE-DOCK [117], FRED [118] Surflex [119] among others. These computational packages have specific search algorithms such as MC, genetic algorithm, incremental construction, and cavity detection algorithm [47]. These algorithms have their specific search method and specific parameters. Docking has computational capabilities of searching for the best fit between molecules, considering various parameters obtained from ligand and receptor input coordinates as: ligand or receptor structure flexibility, interatomic interactions such as hydrophobic contacts and hydrogen bonds, and geometric complementary [47, 120, 121]. Docking applications usually return the pose (predicted predictions) of a ligand in the binding site of the target. Figure 4 represents the scheme of molecular docking techniques.

Active site determination of 3D-receptors, protein–ligand docking simulation (blind and site-specific docking), and analysis of docking complex [47]

Full size image

Docking applications can be classified according to defined parameters and rules applied to predict the conformations. For instance, the flexibility of receptor and/or ligand docking algorithm can be classified as flexible or rigid-body docking [122, 123]. Rigid-body docking does not account for the flexibility of either the receptor or ligand, limiting the accuracy and specificity of results, considering geometrical complementarities between two molecules [123]. However, rigid-body docking is still capable of predicting the correct position of ligands when compared with crystallographic structures. It has also been used to perform initial screening of a small molecule database in a rapid manner. Conversely, flexible docking considers several possible conformations of the receptor or ligand for both molecules simultaneously, although with higher computation time [120].

Rigid-body, also called the "lock-and-key model", emphasizes the importance of geometric complementary, and in flexible docking, the process has to be flexible that ligands and receptors change their conformations to fit each other well, and thus, an "induced-fit model" (c.f. Fig. 5) [124].

Two models of molecular docking—(a) rigid-body and (b) flexible (induced-fit model) [124]

Full size image

As ascribed in Fig. 5, rigid-body docking is a traditional docking technique that assumes that the ligand and the receptor are rigid and, thus, relies on the scoring function to obtain the optimal binding orientation of the ligand to the receptor pockets [125]. Based on their geometric complementarity, the scoring function considers the electrostatic and steric interactions between the ligand and the receptor [124, 125]. This technique is simpler and more efficient when generating multiple docking poses. The drawbacks of rigid-body docking is its inability to account for the conformational changes that may arise when the ligand and the receptor bind [124]. It is known that several protein receptors undergo conformational changes that promote the binding of ligands, thus lacking the computational ability to predict the binding affinity of ligands [124]. These limitations has prompted the research community to use flexible-rigid docking, which allows for the flexibility of the receptor structure [126]. This docking technique considers specific predefined conformational changes in the receptors. The specific rotatable bonds in the receptor structure may be redefined, and the algorithm defines various conformational changes of these bonds to enhance docking prediction. It offers realistic representations of the receptor-ligand interactions. Also, it provides an accurate prediction of the binding affinity [122].

Flexible (soft) is also an advanced docking technique that explores the full flexibility of the ligand and the receptor structures. Its dynamic nature accounts for the receptor and docking flexibility to investigate possible conformations, making it more accurate than flexible-rigid and rigid-docking methods [124]. Docking programs are known to use one or a combination of the following ways: MC, distance geometry, incremental construction, simulated annealing, geometric algorithms, etc. [103]. Figure 6 demonstrates the three types of docking techniques; rigid-body, flexible-rigid, and flexible docking techniques.

Rigid-body, flexible-rigid, and flexible docking methods [124]

Full size image

3.3.1 Docking algorithms

Various docking algorithms are used depending on the features of the optimization problem and the problem domain. For instance, the genetic algorithm (GA) uses a population-based approach inspired by biological evolution, where chromosomes evolve over generations [127]. This method is influenced by the Darwin's Theory of Evolution [128]. The solutions are chosen and combined with others to develop new solutions and are mainly implemented in AUTODOCK 4.0 and GOLD 3.1 [28]. Stochastic, or MC, is another essential algorithm that employs random sampling to obtain solutions to numerical problems [129]. Its working principle is based on probability distribution to solve numerical integrals. MC algorithm uses probabilistic selection to obtain global optima [130]. Incremental construction (IC) is also an important algorithm that obtains new solutions by adding small modifications iteratively to a solution and evaluating it [131].

The conformation selection algorithm also finds valuable applications in biomolecular simulations. It works by exploring the conformations of biomolecules with the lowest energy defined by the degrees of freedom of a system [28]. It applies a combination of energy minimization and conformational sampling to obtain the global minimum energy conformation of biomolecules. Besides, the non-stochastic methods do not apply probability or randomness but employ mathematical tools such as gradients and derivatives in the search process [132]. Cavity detection is another algorithm that employs a random initial solution and iteratively uses local search techniques to obtain new solutions. The various docking algorithms used by docking software are reported in Table 2.

The over 40 years evolution of molecular docking algorithms, together with the advancement of computing power and architecture, has led to more sophisticated applications in docking nucleic acids (NA) as potential therapeutic targets [157]. The druggability of particular ribonucleic acid (RNA) has piqued research activities on RNA-ligand binding with a specific focus on docking-and-scoring methods. With the ever-increasing findings on powerful methods of RNA structure determination, RNA-based therapeutics has become a prospective approach to treating human diseases such as cancer and human immunodeficiency virus (HIV) [158, 159]. Typically, there are two types of RNA-based therapeutics: (1) RNA molecule serves as the target for the drug binding, which is analogous to protein-targeted drug discovery, and (2) the therapeutic RNAs (i.e. guide RNAs, small interfering RNAs, RNA aptamers, antisense oligonucleotides) bind to target (protein targets, DNA targets, and RNA transcripts) to induce or inhibit targeted biochemical reactions [157]. Studies show that genes that are difficult to drug or undruggable by targeting their associated proteins may be inhibited by drugs targeting the corresponding messenger (coding) RNA sequence [160, 161]. Unlike proteins, RNAs show broader druggability, garnering tremendous interest in gene therapy [157].

Computational approaches based on small-molecule docking are aimed at predicting the binding poses of biomacromolecules and small molecules such as NA or protein targets such as RNA or DNA [162, 163]. Various docking algorithms help in predicting and understanding binding modes (molecular recognition) and predicting scoring (estimation of binding energy). While the molecular methods for proteins are well-developed, docking methods for NA are comparatively underdeveloped [164]. Researchers in medicinal chemistry have used docking programs such as GRAMM, ZDOCK, HDOCK, and FTDOCK, initially developed for protein–ligand docking to RNA-ligand docking and developed new scoring functions and methods for NA-docking [144]. Emerging transcriptomics studies, particularly studies on non-coding RNA such as small interfering RNA (siRNA) and microRNAs (miRNA) in the development of anti-cancer drugs [158]. Machine learning (ML) and algorithms such as HNADOCK for RNA-ligand docking, though still in their infancy, have shown promising performance, and the improved in vivo efficacy would accelerate RNA-targeted drug discovery [165]. A noteworthy interdisciplinary collaboration and cutting-edge scientific knowledge in medicinal chemistry is the development of these powerful docking software currently in use [166]. However no sole docking algorithms fits in every system and thus the use of more than one software increases the quality of the output [166].

While molecular docking was used in the study of interactions between targets and the ligands, its application scope currently is wider including drug repurposing, polypharmacology, virtual screening drug side effect prediction, target profiling and discovery (c.f Fig. 7) [167, 168]. Remarkably, there exists vast opportunities provided by molecular docking in drug discovery process. It has been typically integrated with workflow that includes a variety of experimental and in silico techniques. Wet lab experiments and computational methods in conjunction with docking may potentially be utilized in elucidating drug metabolism to gain meaningful insights from cytochrome P450 system, development of new antibacterial agents such as DNA gyrase [169]. Typically, the protein–ligand docking algorithm is twofold; conformation generation and scoring [170]. During conformation generation, different ligand orientations are generated at different positions inside the protein binding pocket [171]. Besides the various applications of docking in drug discovery, they have also been used in target fishing and target (identifying a series of targets for which the ligands exhibit desirable complementarity, drug repositioning (identifying uses of novel pharmaceutical products with already optimized profiles), and in polypharmacology (identification of ligands that simultaneously bind to a pool of selected targets of interest [168]. Consequently, these conformations are evaluated by a scoring function, as illustrated in Fig. 7.

Applications of molecular docking in drug discovery and development [168]

Full size image

Docking involves a combination of a search algorithm and scoring function. Molecular docking is a method that employs various docking algorithms based on molecular dynamics, Monte Carlo, fragment-based and genetic algorithm [171]. The methodology employed in molecular docking is as presented in Fig. 8.

Molecular docking steps: conformation generation and scoring

Full size image

Nevertheless, the shortcomings of docking are well-documented in the literature [172]. Docking is associated with challenges such as receptor flexibility because typically a protein/biomolecule adopts a variety of conformations depending on the ligand with to which it binds [168, 173]. Conformational states of proteins accounts for better affinity to be attained between the target and the drug [174]. Another scientific challenge directly linked to docking is imperfect scoring function because electrostatic interactions and entropy among other physical phenomenon are disregard by scoring schemes [172]. Other studies have also reported that the success of molecular docking is impeded by incomplete understanding of the underlying molecular protein–ligand binding mechanism involving various classes of ligands and modeling receptor flexibility [168]. Most docking software employs force field calculations that estimate binding energy guided by experimental data and quantum mechanics. Molecular dynamics (MD) and MM simulate motions of molecules and atoms in a complex [45]. The binding affinity is estimated by computing the energy of the systems using force fields like GROMOS, CHARMM, or AMBER [175, 176]. However, it is known that accurate binding energies can only be obtained from ab initio methods such as MM simulations and DFT [177]. Most docking programs remove protons (hydrogens) of inhibitors or enzymes under study, potentially excluding important information and leading to inaccuracies [178]. MD simulations have been coupled with docking to obtain important information on solvated and protonation on docking preliminary results [179]. Remarkably, the cutting-edge advancement in force fields and MD simulations has improved the accuracy of these simulations.

The significance of bioinformatics within the context of drug discovery cannot be overstated in the context of drug discovery and development. Bioinformatics entails the study of proteomic and genomic data that in analyzing vast amount of biological information generated through various experimental and sequencing techniques [180], whereas docking predicts how small molecules interact with biological macromolecules. The integration of docking and bioinformatics has proven to be a powerful approach in understanding structural biology and computational drug design [181]. These techniques have been successfully integrated in network pharmacology where multi-target drug candidates have been identified, analysis of protein–ligand interactions, and combined with AI and machine learning to improve accuracy hits identification and enhance the accuracy of docking results [182].

However, bioinformatics and docking are not without challenges which require considerations during drug development process. These drawbacks are not limited to conformation changes (docking typically assumes rigid protein structures), achieving high scoring function accuracy is challenging due to dynamic protein ligand interactions, and these simulations require high computational demands which hinder large-scale virtual screening undertakings [182]. The frequently cited mentionable shortcomings of docking techniques include assumptions and simplifications in scoring functions that lead to inaccuracies; the programs fail to account for entropic, solvent, directional interactions, and hydrogen bonding interactions [168]. Other limitations include flexibility issues related to conformation and structural flexibility, and inadequate resolution of crystallography targets [168].

QM method is a computational technique that models covalent and non-covalent intermolecular interactions [183]. Computer experiments based on QM methods, such as ab initio molecular orbital or DFT, have been used to explore interactions of molecular systems of up to hundreds of atoms [184]. They have also been used to relate the structure of proteins and enzymes to their biological function. On the other hand, MM methods have been used in modelling substrate-protein binding [185]. MM simulations have been reported to lack accuracy and precision in describing bond formation and breaking, which is essential in drug discovery [186]. Modelling large systems using QM methods has been an enduring scientific challenge, prompting the need to combine MM and QM methods in the QM/MM model [187,188,189,190]. Accordingly, MM and QM calculations are invaluable in studying long-range electrostatic interactions affecting biological macromolecules' electronic structures [190].

Several programs are available for QM/MM calculations. For example, Gaussian is used for QM calculations, while CHARMM and AMBER are widely used for MM calculations [85]. However, an effective way to perform QM/MM calculations is to jointly use the existing MM and QM packages with an interface program such as PUPIL, QoMMM, and ChemShell [191]. The hybrid QM/MM method allows a combination of computational abilities that each package has (e.g., algorithms to obtain stationary points and optimal reaction paths on potential energy surfaces implemented in QM packages and algorithms for expanding conformational spaces sampled in MD simulations executed in MM packages. The hybrid QM/MM approach is suitable for biophysical and biochemical mechanisms by combining the low computation cost of MM (with additive classical force fields) and the high accuracy obtained from the QM part (with first principles) [190].

Hybrid methods based on QM/MM have allowed researchers to study enzyme catalysis in drug discovery [54, 192]. They have also been used to gain meaningful insights into the binding between drug candidates and target proteins. Hybrid QM/MM calculations have helped obtain detailed information on binding pathways, identifying specific interactions, and studying ligand binding [193]. Furthermore, they have also been used to account for the impact of protein and solvent dynamics on drug behaviour. Previous authors, for instance, Alonso-Cotchico et al. [194], studied the metalloenzymes involved in drug research using the hybrid QM/MM method. QM/MM calculations can be used to explore the influence of metal ions in enzymes, which offers crucial information on metal-related toxicity [191]. This method has also been used to describe phenomena such as tautomerization, electron transfer, and proton design during drug design [191].

With the demand to enhance the speed of the drug discovery process, QM/MM has been used to describe the binding ability of the ligand with the receptor and the ligand-receptor interactions (c.f. Fig. 9).

The partitioning of the protein-ligand complex into the MM applied region, QM applied region, and QQ/MM applied regions [195]

Full size image

The Hamiltonian operator is essential for quantum calculations and corresponds to the total energy of the system [195, 196]. Equation 1 is the Hamiltonian of the system.

Here, \({\text{H}}_{\text{QM}}\), \({\text{H}}_{\text{MM}}\), and \({\text{H}}_{\text{QM}/\text{MM}}\) are the Hamiltonians accounting for all QM particles of the ligand, MM particles of the protein, and the interactions between QM and MM particles within the system, respectively. Simple functions like Leonard-Jones potential describe the van der Waal's forces at the MM level. Conversely, the electrostatic term enters Fock Matrices as a self-consistent field method. The hybrid QM/MM calculations are divided into additive QM/MM coupling and substantive QM/MM coupling [195].

4.1 Subtractive and additive QM/MM coupling

Subtractive QM/MM coupling includes three steps in calculating the energy of systems: calculation of the total energy at the MM level, the addition of QM energy of the isolated system, and calculating the energy value of the MM system and subtracting this value [197, 198]. It is a simple method which mainly uses the ONION method in calculations. However, it is limited by the fact that it requires a flexible force field to describe the effect of chemical changes during reactions. Moreover, modelling biological charge transfer is challenging because of the absence of polarization by the MM environment on QM electron density [195].

The QM system is embedded within the MM system, and therefore, the sum of MM, QM, and QM/MM energy terms gives the energy of the system. Equation 2 gives the mathematical expression of additive QM/MM coupling.

Various approaches have been used to describe these interactions – mechanical embedding, polarization embedding, and electrostatic embedding [199]. Mechanical embedding extends the computational ability of QM/MM by accounting for mechanical degrees of freedom, treating electronic interactions quantum mechanically, and describing molecular deformations and vibrations [199]. However, mechanical embedding has limitations: it requires accurate MM parameters for MM and QM systems and ignores the perturbation of the electronic structure of the MM system [195]. Polarization embedding accounts for the influence of the surrounding polarizable environments on the quantum regions. Electrostatic embedding is used within the ONIOM framework [200]. It describes the electrostatic interactions between MM and QM regions within a system. It does not require MM or QM electrostatic parameters as in mechanical embedding and offers a more sophisticated treatment of electrostatics [200]. The hybrid QM/MM methods have been employed in drug design. For instance, Tuttle [201] used this approach to explore the binding modes of Latrunculin A, Latrunculin B (naturally occurring analogue), and synthetic l32 to G actin. The research group concluded that L32 possess biological activity like naturally occurring latrunculins.

Drug metabolism scientists have been able to control the pharmacokinetic profiles of drugs such as half-life through metabolism-guided drug design, and subsequently in reducing the attrition rates [202,203,204]. The well-established in vivo and in vitro methods, along with the in silico computational predictions, have been fundamental in reducing toxicological factors at early stages. Studies have shown that drug toxicity is mainly due to their metabolites, either as a result of free radicals or reactive electrophiles [205]. Therefore, the identification of potential liabilities in new chemical series is arguably one of the critical roles of preclinical drug metabolism prediction. To date, many chemical compounds with therapeutic effects have been identified from quantum chemical, docking and MD chemical calculations. Pharmaceutical companies have leveraged in silico tools such as Autodock combined with other methods to introduce drugs in the market [206]. Lead optimization for optimal pharmacodynamics (PD) and pharmacokinetics (PK) and comparison of the preclinical metabolism in in animals with humans for supporting human dose prediction are the main breakthroughs in medicinal chemistry to date [207].

The two phases of drug metabolism–phase I and phase II metabolism are well documented in the literature [209,210,211,212,213]. Typically, phase I metabolism involves hydrolysis, reduction, and oxidation of the drug, which produces metabolites that can be toxic, active or inactive [212]. On the other hand, phase II metabolism involves the conjugation of the modified drug with another molecule, such as amino acid, sulfate, and glucuronic acid [213]. Phase II metabolism increases the water solubility of the drug, easing its excretion. Various enzymes carry out these phases of metabolism. Cytochrome P450 (CPY) family of enzymes carry out phase I metabolism while a variety of enzymes such as glutathione S-transferases (GSTs), sulfotransferases, and UDP-glucuronosyltransferases (UGTs) carry out phase II metabolism (c.f Fig. 10) [208]. Various factors, including disease status and individual characteristics such as age, genetics, and sex, influence drug metabolism [214].

Phase I and Phase II Drug Metabolism [208]

Full size image

Figure 10 depicts CYP-mediated metabolism, which accounts for 75% of the overall metabolism. The human CYP family has 57 isoenzymes, and the CYP-mediated metabolism plays a significant role in drug discovery and development because it significantly affects the safety profile, desired activity, and bioavailability [208, 215]. These enzymes are highly concentrated in the liver, and this is where the majority of the drug metabolism takes place [215, 216]. Evidence in the literature suggests that the drug-drug interaction due to the uptake of two drugs, whereby one is an inducer and the other is an inducer of drug metabolism, potentially has adverse pharmacological effects on the body [217, 218]. For this reason, enzymatic metabolic studies are critical to identifying and quantifying main metabolites, assessing potential drug-drug interactions and resolving metabolic stability [219]. The prediction of sites of metabolism and metabolite structures in CYP-mediated reactions is mainly employed in silico, and AI approaches are the starting point of the metabolic research pathway towards lead optimization [208]. There exist vast software tools to predict sites of metabolism, including BioTransformer, FAME 3, GLORYx, CypReact, PreMetabo, virtual Rat, FP-ADMET, and Cyproduct [208]. These software employ various methods include various methods including rule-based, machine learning (ML), knowledge-based, and Random Forest, among others. Table 3 summarises the public metabolism tools and their corresponding methods [208].

5.1 The role of machine learning algorithms in drug metabolism prediction

The metabolic reactions that are mediated by the enzymes in the human body may transform the drug administered into metabolites that show different biological activity [216]. The effect of metabolic reactions in deactivating the administered drug cannot be overlooked [226]. At the same time, metabolism is essential for the formation of active substances in prodrugs (biologically inactive compounds). To this end, the metabolic fate of drug candidates requires thorough investigation. The growing body of literature on drug metabolism has made data available, and machine learning (ML) models such as Random Forests and Support Vector machines to become the main choices that offer faster inference [226]. The rule-based methods have been used alongside ML models to predict drug metabolism to reduce false positives and filter unlikely predictions. The recent ML models are trained on data to cover the entire human metabolism, including endogenous compounds. For instance, researchers have used Graph Convolutional Neural Networks to provide insights on molecule interactions with the enzyme [227]. These models also postulate reaction outcomes as well as reveal valuable information on reaction mechanisms [228].

However, Litsa and others [226] reported that the current ML approaches used in metabolism prediction of drug candidates are based on shallow ML models and mainly classification models for distinguishing sites of metabolism and non-sites of metabolism and enzyme binders from non-binders. Besides, the existing ML models also fail to provide valuable insights on why and how a prediction is made like in computational approaches. ML models also suffer limitations related to inconsistent labelling of the sites of metabolism due to the different limitations. Also, these models fail to classify metabolites as primary or secondary [226]. These issues often avert medicinal chemists from not only taking an active role in developing novel models for predicting drug metabolism but also the comparative assessment and evaluation of methods.

Cancer is the most challenging and devastating disease, threatening the health and life of millions of people [229]. It is the prominent cause of high mortality rates in the world. Cancer develops when abnormal cells proliferate, invading the surrounding tissues, which then spread to body organs and other parts of the body through lymphatic and circulatory systems [230]. Several strategies, including hormonal therapy, radiation therapy, chemotherapy, immunotherapy, and surgery, have been used in fighting against cancer [230]. Despite these strategies being effective in fighting cancer cells, they are known to have potential toxicity, accompanied by adverse drug reactions, pain, and cost, limiting their application in clinical settings and public health in general [231]. Further to this, the risk of tumours acquiring multidrug resistance prompts medicinal chemists to develop efficient, safer, and novel antitumor agents for fighting cancer [232]. Pharmacokinetic (PK) studies are often conducted to determine the schedule and dose of treatment of the drug relations between drug concentrations and functional or biochemical effects, route of administration, and drug interaction when more than one drug is administered [3].

In the current drug development paradigm, animal studies provide a framework for human clinical trials [233]; however, a drug that works in animals may be ineffective in humans due to inappropriate translation of a drug dose from animals to humans [3]. Scientific experts in the pharmaceutical industry employ pharmacodynamics (PD, effect vs. time) and pharmacokinetics (PK, concentration vs. time) for optimal design and the discovery of new drugs to combat cancer-cell growth [3]. While PD has been defined as “how the drug affects the body”, PK has been defined as “how the body handles the drug”. Accordingly, PK/PD data is becoming more available in literature and consequently has piqued significant scholarly attention in industry, academia, and regulatory authorities as an advanced method for exposure–response analysis [234]. Often time, a combination of therapies is widely used in treating cardiovascular diseases, cancer, and infectious diseases.

Research activities in developing anticancer treatments have yielded desirable outcomes, but many of these drugs have been reported to have significant systemic toxicity [235]. This challenge is attributed to the drug’s lack of selectivity, having healthy tissues and cells that undergo fast turnover, resulting in toxicity. For instance, approximately 30% of colorectal cancer patients develop tumour metastasis, therefore posing a great challenge in diagnosis and treatment [3]. Whereas chemotherapy is known to be a conventional treatment for patients with this type of cancer, the therapeutic benefits are limited by toxicity to normal drugs and drug-drug resistance [236,237,238]. Capecitabine (CAP) is a tumour-selective pro-drug which has been approved by the Food and Drug Administration (FDA) for the management of pancreatic cancer, breast cancer, gastric cancer, and colorectal cancer, tumours known to be resistant to 5-fluorouracil (5-FU) among other malignancies [3, 239]. CAP is an oral chemotherapy tumor-selective pro-drug which is preferentially converted to 5-FU (most active compound) in targeting tumour tissues via three metabolic steps [240]. After oral administration, CAP is absorbed in the intestine and metabolized to 5’-deoxy-5-fluorocytidine (5’-5’-DFCR) by carboxylesterase (CES) enzyme located in the liver [3]. The ubiquitous enzyme cytidine deaminase (CyD) then converts 5’-DFCR to 5’deoxy-5-fluorouridine (5’-DFUR). Finally, 5’-DFUR is then converted to toxic and active metabolite 5-FU by thymidine phosphorylase (TP) (Fig. 11), which is more concentrated in solid tumours than normal adjacent tissues, decreasing the effect of 5-FU on normal cells.

The structure of CAP and metabolic conversion of CAP to 5-FU [3]

Full size image

Previous scholars performed a pharmacology study on CAP to assess clinical pharmacokinetics and disposition of CAP and its metabolites 5’-DFUR, 5’-5’-DFCR and 5-FU. Their study aimed to determine the initial and steady-state PK parameters, including area under the curve, maximum plasma concentrations (Cmax), volume of distribution (Vd), and clearance (Cl) of CAP and its metabolites [241]. The authors reported that CAP and its metabolites PK were linear and consistent over time when studied in a dosage range of 500–3500 mg/m²/day among cancer patients. The study also found that lower CAP depicted an acceptable response rate which may also suggest that higher CAP doses are not always required to achieve the required therapeutic effects [241].

Accumulating evidence from oncology research shows that PK-PD studies have been pivotal in tailoring drug dosage schedules to minimize toxicity and achieve desirable pharmacodynamic effects [242,243,244]. They also guide in selection of drug candidates for further pre-clinical studies. Medicinal chemists have used PK-PD data to identify and prioritize compounds with favourable profiles, and streamlining the drug discovery process. However, there are scientific challenges associated with PK-PD studies including complexity of PK-PD data, variability in drug metabolism, response, and distribution [245, 246]. These impede the extrapolation of laboratory findings to real-life applications. Further to this, analysis of anticancer drugs using PK-PD data is sophisticated and cost-intensive, limiting the scalability of PK-PD studies in preclinical drug research.

The Hartree–Fock (HF) methods exclude electron correlations and thus may not describe some properties of drug candidates precisely [247]. Similarly, the post-HF methods, configuration interaction (CI) and Møller-Plesset perturbation (MPn) theory consider electron correlations flawed by high computational costs. DFT is cost-effective, with high computational accuracy and less computational time than MPn, HF, and CI methods [247]. In view of this, DFT has inspired intense scholarly interest in studying molecular properties [248, 249]. DFT has demonstrated its efficacy in studying the geometries of smaller drug molecules, demonstrating its suitability in drug design [247].

A crucial step in drug design using DFT is the study of energetic properties such as metal–ligand bond strengths, relative energies, ionization energies, and electron affinities [247, 250]. The properties and the interactions of drug molecules with their receptors are governed by their molecular structures. Therefore, the DFT technique has been conveniently used in computational drug design to predict relative conformational energies [251]. Numerous studies have reported that possible interactions between a drug and a receptor include charge transfer, dipole–dipole interactions, ionic interactions, ion–dipole interactions, covalent bonds, hydrophobic interactions, and hydrogen bonding [252,253,254]. DFT method (c.f Fig. 12) provides reasonable predictions for covalent, ionic, and hydrogen bonding; however, interactions for weaker bonds have been difficult to predict [255].

The DFT method (B3LYP and CAM-B3YLP) in Drug Design [256]

Full size image

Quantum chemistry software packages such as ADF employ DFT in electronic structure calculations and elucidate spectroscopic properties such as nuclear magnetic resonance (NMR), infrared (IR), and vibrational frequency spectra [50, 51]. Further, ADF can also be used for molecular geometry optimization, to explore the role of catalysis in reactions, and to study reaction channels and activation energies of systems being studied [257]. There are a significant number of research reports that have shown that DFT has been used for structure elucidation and modelling the interactions between the drug and the receptor [44, 258,259,260]. This computational tool has also shown proficiency in modelling metal-containing biological systems in the study of inorganic therapeutics.

Figure 12 depicts a graphical representation of ligands’ structural accuracy and binding affinity. The ligands are predicted using two DFT-based energy functionals, namely B3LYP and CAM-B3LYP [256]. The predictions were then compared to experimental data. The x-axis represents the experimental binding affinity value, while the y-axis represents the calculated binding affinity value using the two DFT methods [256]. Data points closer to the diagonal line indicate better structural accuracy and binding affinity prediction as they are more closely aligned with the experimental data. The graph demonstrates that both methods exhibit strong performance in predicting structural accuracy and binding affinity [256]. However, B3LYP yields slightly superior results compared to the CAM-B3LYP method. The effectiveness and usefulness of DFT methods in drug design are demonstrated by their ability to predict the binding affinity of ligands accurately [186, 247]. This prediction is crucial for the development of new and effective drugs. The image also demonstrates the potential of DFT in reducing the time and cost associated with traditional drug design methods. This is achieved through the provision of accurate computational predictions [256].

7.1 Deficiencies of computational methods in pre-clinical drug design

It is well-established in the literature that each in silico method has its application scope and features [261, 262]. Therefore, in drug design and development, appropriate methods must be selected for accurate prediction; however, some methods have deficiencies that may affect the prediction of results. Theoretical limitations associated with different approaches such as ADMET, QSAR, cavity prediction, pharmacophore building, modeling, and structure optimization have been reported [263]. Molecular modeling has been widely used in identifying possible interactions between metabolic enzymes and compounds, as well as predicting metabolism [264]. However, it is known that the scoring function in molecular docking affects the accuracy of ADMET prediction [265]. On the other hand, the accuracy of data modeling methods such as QSAR depends on the quantity and quality of data [263]. These models cannot predict chemicals outside the chemical space where the models were developed. Furthermore, the simplistic nature of data models limits their ability to predict drug interaction and behavior in a whole system. QSAR prediction methods assume that similar molecules have similar properties, but in other instances such as CYP metabolism, similar molecules exhibit different activities [266]. Although ADMET software has been widely used in predicting multiple properties and qualitative analysis of compounds [267], it cannot give an accurate prediction of the quantitative values [268].

DFT, on the other hand, has been important in drug design and development but has its limitations. The DFT computation approach introduces errors for systems with dispersion interactions and electron correlation [269]. DFT, in particular, diffusion quantum Monte Carlo (DQMC) is also associated with high computational requirements in terms of time and cost, thus medicinal chemists often prefer lower-level computational methods [270, 271]. Incorrect choices of binding sites for ligands give false DFT results, affecting the accuracy of the approach. Another notable challenge is that for a drug candidate to be licensed as an effecting drug or lead, it has to satisfy various therapeutic necessities. CADD relies on predefined codes, algorithms, and theoretical parameters/principles [272]. Therefore, all CADD tools that are used for pharmacophore modeling, QSAR, virtual screening, molecular modeling, and MD have limitations and merits [272]. The accuracy of these models relies on experimental data and knowledge to make them accurate [273]. Hybrid QM/MM methods are limited lack of new methodologies for virtual screening and high cost of computational resources [274]. These methods have shown to be more efficient for medium-sized systems unlike complex drug-protein interactions and large biomolecules. These hybrid methods are also known to have deficiencies in capturing conformational changes and binding modes [275]. Whereas is known to be accurate, QM computations involving extensive conformational sampling or large systems require high expensive computational resources [189].

Applying Artificial Intelligence (AI) and computational methods in drug design has fundamentally changed the scientific landscape in drug design. With the help of new technology, drug discovery is now more efficient, accurate, and cost-effective than before. By using AI, scientists can design and verify new medicines without first making synthesizing them in the laboratory. Conventional drug discovery approaches are labor-intensive and expensive [12, 36, 276]. However, AI is a game changer since it can quickly screen a range of molecules and determine their characteristics in interacting with specific targets by analyzing their chemical and structural features [12]. This helps scientists to reduce the list of possible medications and focus on the ones most crucial to work. The use of AI in analyzing chemical compounds is done mainly through employing algorithms and utilizing computer simulation capabilities. By employing these techniques, researchers can screen many potential drugs more accurately than in traditional approaches [276]. This helps determine the effectiveness of drug designs and make better predictions regarding the outcome, saving resources and avoiding potential risks. The use of AI and computational methods in drug design allows for the examination of considerable chemical space with different compounds and elements. In the past, drug discovery programs could only examine a limited number of substances because of the high cost and time it took to create and evaluate each drug. But, with the incorporation of AI and computational technologies, it is now possible to scrutinize countless compounds, thus enhancing the chances of detecting a powerful and harmless drug applicant [277].

Predicting a compound's physical and chemical properties using AI has been critical in drug design and research. Computational methods suggest the various properties of a chemical substance like whether it will dissolve in water or not, if it is safe to use, and how it might interact with living cells [277]. Scientists can leverage this data to adjust the compound's formulation to perform better and increase the potential of achieving the desired outcomes. With the help of AI, researchers can design and optimize drug delivery systems based on factors such as particle size, surface charge, and drug release kinetics [276]. Another benefit of using AI and computational techniques is the ability to individualize drug treatment by tailoring them to a particular disease condition, and the causing agent. Pharmacogenomics is a technique that uses an individual's genetic information to predict the efficacy and side effects of a particular drug for a known group of patients [277]. This is important in the medicinal treatment of complex diseases, where a general solution may need to be revised. Applying AI and computational approaches can help detect disease-specific biomarkers, leading to the development of targeted treatment strategies [276,277,278].

Whereas AI and computational methods have many advantages, they also have some drawbacks that cannot be overlooked. These techniques are dependent on the accuracy of the data used in programming the AI. Data bias or incomplete data can lead to incorrect model predictions [278, 279]. Hence, researchers must guarantee that the data used is genuine and uninfluenced [278]. Moreover, AI and computational methods cannot account for unique chemical reactions and biological pathways. While these methods can predict the behavior of a compound under certain conditions, they may need to reflect the complex interactions that occur in living systems accurately. To overcome this limitation, experimental validation of the predictions is crucial, and a combination of both computational and experimental methods is often used for the most accurate results [279].

There have been numerous successful applications of AI and computational methods in drug design. For example, in developing the HIV protease inhibitor, darunavir, computational methods were used to design and screen a library of compounds for potential inhibitors [280]. Chuntakaruk and colleagues [281] investigated the ligand–protein interaction of the darunavir/HIV-1 protease using computational methods and noted that while darunavir and its three analogues 19–0–14–3, 19–8–10–0 and 19–8–14–3 are potential protease inhibitors. This led to the identification of darunavir, which is now widely used in the treatment of HIV. Similarly, the development of the influenza drug, Relenza was aided by computational methods that predicted the binding affinity of the drug with the viral protein (HIV-1 protease) [280].

Figure 13 illustrates the integration of AI into drug discovery and development processes and the various novel applications of AI in the pharmaceutical industry. The drug discovery process can be divided into four main components: drug design, polypharmacology, drug repurposing, and drug screening [282]. The main application of AI is in predicting drug properties. This can potentially decrease the necessity for clinical trials and the involvement of live study participants. Such a development would bring benefits from both financial and ethical perspectives. This section discusses the studies identified in this review that support the integration of AI into the drug discovery procedure. These studies highlight how AI can improve efficiency, accuracy, and productivity [282].

Application of AI in various aspect of pharmaceutical industry in the context of drug design [282]

Full size image

8.1 Limitations of AI approach in pre-clinical drug research

Accumulated evidence suggest that the use of AI in pre-clinical drug research is slowly living up to the expectations of the scientific community, with noteworthy advances in de novo molecular design, QSAR modeling, and synthesis planning, among other crucial pre-clinical research activities [283, 284]. Nevertheless, the efficacy of AI in accelerating drug research is yet to be demonstrated. There exist challenges that impede the integration and implementation of these technologies in search of compounds with therapeutic properties. One key challenge is inefficient data integration as a result of diversity between datasets, which include raw data, metadata, candidate data, race data or processed data [282]. For efficient analysis, these datasets must be collated or collected, and there exists no method of doing so as it currently standards. Without appropriately formatted data, the current AI systems often yield erroneous results [277]. The second impediment is the immobility of experts and occupational experts as many people in the realm of medicinal chemistry lack the required skills and qualifications required to operate AI systems [282]. Few experts are competent in a combination of molecular biology and chemistry, as well as data science to generate AI algorithms in a pharmaceutical context. Another notable problem is the lack of adequate financing for the development of novel AI systems in drug discovery and development [285]. The hesitancy to invest in AI is attributed to the skepticism surrounding the results of machine learning and AI in drug research, slowing AI-relevant advances [282]. Towards this end, lack of understanding of the methodology in AI systems causes distrust and researchers will be skeptical about using this technology, which may hold the industry back for the next few years [282, 286].

Over the past decade, the rates of identifying disease-associated targets have been higher than identifying novel compounds with promising therapeutic effects. There is a drastic increase in computational methods, such as docking or virtual screening, and MD that have accelerated drug design and discovery methods in the pharmaceutical industry. Well-developed computational tools have been used in the pharmaceutic industry and academia and have shown success and remarkably huge prospects for providing even a faster and cheaper approach in the drug discovery landscape. The computational capability of predicting molecules to be optimized and filtering a large molecular database has reduced the amount of time and challenges involved. Nowadays, structure-based methods have been popularly used in studying protein targets. Molecular docking and MD simulations are prominent examples of these methods, and they have been applied in the characterization of binding sites, determining the thermodynamics and kinetics involved in ligand-target recognition. Conversely, ligand-based methods such as QSAR have been popularly used to determine and improve the activity of active molecules. They have also been used to determine bioactive molecules in the drug discovery journey. CADD has significant applications, including the identification of hit compounds, lead optimization, and assessment of the ADMET profile of bioactive compounds. Considering drug discovery as a succession of decisions made in identifying the right target and dosing regimen, in silico and ab initio methods can support these decisions. Computational chemistry will undoubtedly yield an unprecedented drug design and development revolution, making the complex and time-consuming drug discovery more effective and affordable. There is an exponential growth in the number of pharmaceutical companies adopting these approaches, and numerous opportunities exist for further growth. Notably, drug discovery and design rely on human expertise and the synergy between CADD and human intelligence for success. CADD approaches are known to have vast computational capabilities in conducting unsupervised analysis of databases; human judgement and input are necessary to consider ethical concerns related to drug research. Human expertise is also required to validate real-world clinical studies and wet-lab experiments. All the classical and new methods jointly used with the existing computational disciplines are given an advantage at nearly every drug discovery and development stage. With the required expertise in biophysics, biochemistry and biology, significant investment in time and money, and the availability of the necessary computational software and hardware, drug discovery has become an affordable task in the pharmaceutic industry and research institutions. Accordingly, we foresee the impact of computational chemistry in the form of accelerated drug discovery and the reduction of attrition rates due to negative ADMET results.

With the entry of AI in the field of drug design and development, the drug industry will be inspired to develop drugs not only in a faster manner but also drugs that can target complex diseases with unrivaled accuracy. The synergistic application of computation tools and AI technologies has contributed to public health improvement and aligning the United Nation’s SDG number 3 on good health and well-being. Innovative computational resources have accelerated the drug research pipeline, and subsequently, cancer patients have benefited from innovative cancer therapies. The AI-driven predictions have enabled medicinal chemists to tailor cancer management regimens to patient characteristics such as drug response patterns and tumour heterogeneity, reducing any potential adverse reactions and promoting healthy access and equity globally. The emerging cutting-edge approaches in oncology research have also contributed to the growth of the biotechnology and pharmaceutical sectors, which aligns with UN SDG innovation and sustainable industrialization. By promoting oncology research, medicinal chemists can address the gaps in medicinal research, improve accessibility and affordability of healthcare services and contribute towards the achievement of SDGs related to public health and industrialization.