Docking strategies for predicting protein-ligand interactions and their application to structure-based drug design

Abstract

Molecular docking stands as a pivotal element in the realm of computer-aided drug design (CADD), consistently contributing to advancements in pharmaceutical research. In essence, it employs computer algorithms to identify the “best” match between two molecules, akin to solving intricate three-dimensional jigsaw puzzles. At a more stringent level, the molecular docking challenge entails predicting the accurate bound association state based on the atomic coordinates of two molecules. This process assumes particular significance in unraveling the mechanistic intricacies of physicochemical interactions at the atomic scale. Notably, the application of docking, especially in the context of protein-small molecule interactions, holds wide-ranging implications for structure-based drug design, given the prevalent use of small compounds as drug candidates. This study provides an overview of docking methodologies, delves into recent key developments, elucidates the physicochemical underpinnings of molecular recognition in protein-ligand interactions, and concludes by addressing the applications of docking in virtual screening, alongside current challenges within existing docking methods.

1. Introduction

Proteins are incredibly diverse macromolecules made up of amino acids as the building blocks. In a cell, proteins interact with ions, small molecules (drugs), and even macromolecules (e.g., other proteins and nucleic acids). During such interactions, shape alteration (referred to as conformational change) almost occurs at the same time. It is known that a structure of protein dictates its function in living organisms [1]. Any dysfunction could lead to pathological changes. Therefore, studying protein interactions and solving the structures are crucial steps for understanding the mechanism of life, in addition to the investigation of biological systems and rational drug design. Direct methods for the determination of three-dimensional structures of the protein-ligand complexes rely on experiments. Traditional techniques include X-ray crystallography, neutron diffraction or electrons diffraction (which depends on the source type of beams used in interacting with the specimen) [2, 3]. The principle of this method is the analysis of diffraction patterns, which requires a crystalline form of the specimen. Proteins are very difficult to crystalize due to the numerous unstructured loops. Therefore, the limitation of X-ray crystallography is the quality of the crystallization. In this circumstance, cryo-electron microscopy (cryo-EM) has the advantage that a crystalline sample is not necessary. However, cryo-EM usually has lower resolution when compared with crystallography. Nevertheless, considerable improvements have been made very recently [4, 5, 6]. Also, magnetic resonance (NMR) spectroscopy has contributed to many structural details and dynamics of biological materials [7]. However, NMR technology has difficulty to solve the structures of large proteins. In summary, all experimental techniques utilized to obtain protein structures have limitations. In addition, they are time consuming and costly. Ultimately, these shortcomings spur the development of computational methods. Molecular docking is one of the examples that implement the integration of computational and experimental strategies [8, 9, 10, 11]. In the past decades, molecular docking has been a powerful approach for computer-aided drug design (CADD). Generally, molecular docking is performed between two molecules (see Figure 1). The target (receptor) is usually a macromolecule such as a protein, DNA or RNA. The other molecule can be a protein, peptide or small molecule. Molecular docking utilizes computational algorithms to predict the bound association of the two molecules. In this work, focus was essentially on docking of small molecules against a protein target, referred to as ligand-protein docking, because of its broad application to structure-based drug design (SBDD) (drug compounds are usually small molecules). With a protein of a known three-dimensional structure, the docking method can predict how ligands interact with their protein targets from a physical or chemical perspective. The determination of protein-ligand complex structures can be done routinely by docking algorithms. In other words, protein-ligand docking approaches allow an automatic way to manipulate the recognition of a drug by its protein target through capturing physical principles.

Figure 1:

Illustration of ligand-protein docking. Docking algorithms generate a variety of complex configurations. The right panel shows the surface representation of a protein binding pocket, with the ligands in different orientations.

Nowadays, protein-ligand docking is widely applied to structure-activity and mutagenesis studies, virtual screening (VS), and lead optimization [12]. With the rapid growth of the protein structures in the Protein Data Bank (as illustrated in Figure 2, data sourced from the Protein Data Bank [13, 14]), protein-ligand docking methods have become a valuable tool for mechanistic biological research and pharmaceutical drug discovery.

Figure 2:

Protein Data Bank growth statistics [13]. Number of structures deposited per year vs the number of total available structures.

2. Physical Basis of Molecular Docking

Protein-ligand interactions are central to the in-depth understanding of protein functions in biology because proteins accomplish molecular recognition through binding with various molecules [15, 16]. Drugs often act as inhibitors when interacting with proteins, because many diseases including cancer are attributed to the abnormally active protein-ligand interactions. In this circumstance, inhibitors can be used to prevent the abnormal interactions for specific therapy. For instance, in the treatment of chronic myelogenous leukemia (CML), STI571 (Gleevec^™, imatinib mesylate) is used to inhibit Bcr-Abl target protein activation, which also exemplifies the successful development of the rational drug design [17, 18]. Therefore, insights into protein-ligand interactions are vital for the development of drugs.

Molecular docking algorithms predict protein-ligand interactions in a complex that is formed non-covalently. In biological systems, hydrogen bonds, ionic bonds, Van der Waals interactions and hydrophobic bonds are the four main types of non-covalent interactions [19]. Non-covalent bonds, ranging from 1–5 kcal/mol [20], are weak in comparison with covalent bonds. Therefore, most of the time the term “non-covalent interactions” is used, rather than non-covalent bonds. However, the cumulative effect of these non-covalent interactions can be significant, which means multiple non-covalent bonds often act together at the binding interface of a complex to produce highly stable and specific associations [21].

2.1. Major types of non-covalent interactions

2.1.1. Hydrogen bonds

Hydrogen bonds are polar electrostatic interactions that can be described in the form of D—H … A. Here, D and A stand for an electron donor and acceptor atom, respectively. H is the hydrogen atom attached to a donor atom. Thus, the donor atom must be electronegative [22]. Hydrogen bonds have a strength of about 5 kcal/mol, which is weaker than the covalent bond between an oxygen atom and a hydrogen atom (approximately 110 kcal/mol) [20]. However, biomolecules exist in solvent surroundings, in which the extensive hydrogen bonds between adjacent molecules are broken and reformed constantly, resulting in the change of system enthalpy and entropy. Both the protein and the ligand interact with the solvent before binding, so the enthalpy and entropy influence their complex formation [8, 15]. Also, hydrogen bonds contribute a stabilizing force to macromolecules (e.g., proteins and nucleic acids), which is of great importance to the structure.

2.1.2. Ionic interaction

Ionic interaction is the electronic attraction between oppositely charged ionic pairs. Thus, ionic bonds are highly specific electrostatic interactions. In the aqueous solution, ions are surrounded by a shell of water molecules which is caused by dissolution in water [21, 23, 24].

2.1.3. Van der Waals interactions

When atoms in different molecule are close enough, the transient dipoles in the electron clouds lead to an intermolecular force and form nonspecific Van der Waals interactions. Roughly, the strength of Van der Waals interaction is 1 kcal/mol, which is even weaker than typical hydrogen bonds.

2.1.4. Hydrophobic interactions

Hydrophobic interactions result from the effect that nonpolar molecules tend to exclude from the solvent and aggregate in an adequate surrounding. Hydrophobic interactions are often considered as driven by the entropy gain [24]. However, the molecular mechanisms of the hydrophobic effect remain controversial. In the scaled-particle theory, the hydrophobic effect is multifaceted depending on the size of solute [24, 25].

2.1.5. Other types of interactions

Besides the aforementioned major types of non-covalent interactions, other types of non-covalent interactions also exist in protein-ligand complexes, such as CH-π in protein-sugar complexes due to the presence of aromatic rings [26].

2.2. Enthalpy-Entropy Compensation

The contributions of various non-covalent interactions are intimately linked to two fundamental thermodynamic quantities: entropy and enthalpy. The enthalpic and entropic changes occurring before and after the complex formation determine how tightly the protein and ligand bind together, which is quantified by the Gibbs binding free energy, as shown in Equation 1.

$$\mathrm{\Delta }{G}_{bind}=\mathrm{\Delta }H-T\mathrm{\Delta }S$$ (1)

Here, $\mathrm{\Delta }G$ represents the change of the free energy, and $\mathrm{T}$ is the absolute temperature in Kelvin. $\mathrm{\Delta }H$ represents the types and numbers of chemical bonds and noncovalent interactions that are broken and formed, and $\mathrm{\Delta }S$ is the change in the randomness of the system. The Gibbs binding free energy can be calculated using theoretical methods, or can be quantified experimentally by the reaction rate [27]:

$$\mathrm{\Delta }{G}_{bind}=-RTln{K}_{eq}=-RTln\frac{k_{on}}{k_{off}}$$ (2)

Equation 2 shows that complex stability can be measured through the equilibrium binding constant ${\mathrm{K}}_{eq}$, which is determined by the two kinetic rate constants, ${\mathrm{K}}_{on}$ (for the binding reaction) and ${\mathrm{K}}_{off}$ (for the dissociation reaction). $\mathrm{R}$ and $\mathrm{T}$ is the gas constant (8.314 J/mol·K) and the absolute temperature, respectively.

The net driving force for binding is balanced between the two factors, entropy (the tendency to achieve the highest degree of randomness) and enthalpy (the tendency to achieve the most stable bonding state). In fact, the protein-ligand binding process is driven by the decrease in the total Gibbs free energy of the system [15].

2.3. Molecular recognition models

Noncovalent interaction-driven molecular complexes show geometrical complementarity. Three conceptual models of ligand-protein binding have been proposed to explain the mechanism for molecular recognition (Figure 3).

Figure 3:

Illustration of the three conceptual protein-ligand interaction models accompanied with one extended interaction model: (a) The lock-and-key model; both protein and ligand are rigid. (b) The induced-fit model; the conformational change of the protein occurs. (c) The conformational selection model; the ligand binds to the most suitable conformation among an ensemble of protein conformations. (d) The extended conformational selection model; the ligand binds to one protein conformer first and then a subsequent conformational change occurs to the protein. To focus on protein flexibility, this figure does not show ligand flexibility.

Fisher’s lock-and-key model theorizes that the binding interface should be matched complementarily [28]. The protein and the ligand are rigid in this model, which means that their conformations are identical before and after binding. Du et al. suggested that the lock-and-key model is an entropy-dominated binding process [15].

In Koshland’s induced-fit model, conformational change occurs to the protein during binding in order to accommodate the ligand with the best amino acid configuration [29]. It should be noted that in most cases only minor conformational changes arise in proteins after ligand binding. Compared with the lock-and-key model, the induced-fit hypothesis can be linked to a “hand in glove” model, adding flexibility upon Fisher’s idea [30].

In the conformational selection model, ligands bind selectively to the most suitable conformational state among an ensemble substates [31, 32]. In the original conformational selection model, no further conformational rearrangement occurs. However, in an extended conformational selection recognition mechanism, ligands first bind to a protein in the initial favorable protein conformational state, which induces a protein conformational change subsequently. In other words, the ”induced-fit” and ”conformational selection” recognition mechanisms may co-exist in ligand binding [33, 34].

3. Key Components in Ligand-Protein Docking

The basic procedure of docking can be simplified to sampling and ranking, two critical ingredients in current docking programs. Sampling refers to searching putative ligand conformations in the binding site of a receptor. Proteins are usually held rigid in most docking algorithms. Generation of 3D conformers for the ligand is required for flexible (ligand) docking. A good sampling algorithm should be able to generate a diverse ensemble of ligand structures, including the bioactive, bound ligand structure. The second component, ranking, refers to assessing the fit between a ligand conformation and the target by scoring functions. These two docking components will be discussed in detail, respectively.

3.1. Conformational sampling

Biomolecules are inherently dynamic, existing either dissolved within the cytosol or interacting with other cellular components. In addition to the six degrees of freedom—three translational and three rotational—dihedral (torsional) angles significantly enhance their flexibility. This extensive range of motion enables biomolecules to adopt diverse conformations, complicating the computational prediction of their behaviors and properties. For instance, acetylcholine, a neurotransmitter essential for signal transmission across synapses in the nervous system, has a torsional angle between the acetyl group and the choline part that affects its structure (see Figure 4), which in turn influences how it fits into the binding site of acetylcholinesterase.

Figure 4:

Stick representation of acetylcholine with the elements in different colors: C(tan), O(red), N(blue). One of the torsion angles is annotated.

Similarly, proteins are polymers of amino acids that link together through peptide bonds. Therefore, even small variations in the conformations of protein subunits could lead to extensive flexibility in a protein. Figure 5 shows the dihedral angles of an amino acid in a protein. The third backbone dihedral angle $\chi$ (see Figure 5 for its definition) is typically close to 180° or 0° in proteins. The flexible side chain dihedral angle for each residue has various distributions [35]. Furthermore, the task is more complicated because proteins in solution exist as an ensemble of conformations and fluctuate over time, which make it even harder to simulate on computer. While more extensive samplings are needed to account for the induced fit and conformational selection models as introduced in Section 2.3, Zou group has developed a docking package MDock [36, 37, 38] to consider many different conformations in an ensemble without time consuming strain. This will be further delved into in Section 3.3 and Section 4.

Figure 5:

Illustration of dihedral angles in one amino acid residue of a protein: two backbone dihedral angles $\phi ,\psi$ and one side chain dihedral angle $\chi$. The elements are in different colors: C(grey), O(red), and N(blue). Hydrogen atoms are not shown. The backbone of this amino acid is colored green.

3.2. Sampling algorithms

Computational conformational sampling has been an active realm for many years because of evolving small molecule modeling and design in pharmaceutical work [39, 40]. Sampling algorithms explore the molecule’s conformational space from two perspectives. One of the categories exhaustively enumerates all possible torsions, which is necessary due to our limited knowledge of pharmaceutically relevant conformational space. However, this method may cause exponential growth of the search space. To overcome high degrees of freedom, systematic search methods store the physicochemical properties in an applicate grid to prevent combinatorial explosion [41]. The typical procedures divide the ligand into fragments first and then dock into the active site sequentially. An application of automated calculating grids is incremental construction (IC) methods (anchor and grow), which are implemented in the application DOCK [42]. DOCK is the first and most widely used docking program, and the newest version is DOCK 6 [43, 44]. Other well known flexible ligand searching programs such as FlexX and Glide also use incremental construction methods [45, 46].

The other perspective uses a random element to change the system degrees of freedom, which is also referred to as stochastic search. Most common stochastic methods include genetic algorithms (GAs) and Monte Carlo (MC) search [47, 48]. Also, simulation methods have been developed and widely used in molecular docking sampling procedures. Molecular dynamics (MD) is one example of simulation methods that the generated state has to be lower in energy than the initial state to move to the next step. Therefore, it may often get trapped in the local minima that cannot cross high energy barriers. An alternative strategy is to identify the conformation preliminarily, and followed by further MD simulations to implement local optimization. Other methods such as increasing the simulation temperature and smoothing potential energy surfaces are also introduced in previous studies [8, 40].

3.3. Scoring functions

Another key aspect of molecular docking is to rank the ligand conformations that have been generated. A good scoring function should be capable of distinguishing between the true binding modes and the decoy modes. Due to the large degrees of freedom in the protein-ligand system, balancing the speed and accuracy is necessary for the scoring step. The scoring functions are mainly divided into four categories: force field-based, empirical, knowledge-based, and machine learning-based. Different scoring functions are listed in Table 1.

Table 1:

Example of scoring functions

Scoring function	Category	Prominent features	Ref
GOLDScore	force field-based	Default scoring function for GOLD	[48]
MedusaScore	force field-based	Including a hydrogen-bonding model and implicit solvent model	[49]
DOCK6	force field-based and footprint similarity	Updated internal energy function	[44]
Glide	force field-based and empirical	Modifying and expanding the ChemScore function	[46]
AutoDocka	empirical	Allowing side chains to be flexible	[50]
AutoDockFRa	empirical	Considering partial receptor flexibility	[51]
AutoDock VinaXB	halogen bond scoring function	Adding halogen bonding parameters	[52]
ICM	empirical	Allowing flexible side chains which are sampled simultaneously with the ligand	[53]
LigScore1 /LigScore2	empirical	Consisting of three terms that describe the van der Waals interaction	[54]
Vina	empirical	Summation of the energetic contributions from both inter-and intramolecular interactions	[55]
Vinardo	empirical	Modifying the interaction terms and the atomic radii of the Vina scoring function	[56]
Vina-Carb	empirical	Adding a CHI-energy term to the scoring function of AutoDock Vina to improve protein-carbohydrate docking	[57]
SLICK	empirical	A scoring function for protein-carbohydrate docking by implementing a CH-n stacking interaction energy term in BALLDock	[58]
DrugScore	knowledge-based	Extracting both short-range pair and solvent accessible surface potentials	[59]
PFM	knowledge-based	Potentials of mean force for fast calculation of atom pair interaction potentials	[60]
ITScore	knowledge-based	An iterative method to derive atomic pair potential energies	[37]
ROTA	knowledge-based	pair potentials in GalaxyDocka	[61, 62]
AffiScore	knowledge-based	A scoring function in SLIDE for the secondary scoring step	[63]
X-CSCORE	consensus	Combine three empirical scoring functions	[64]
RF-Score-VS	machine-learning-based	Terms from RF-Score-v3	[65]
ΔvinaRF20	machine-learning-based	10 terms from Vina and 10 terms related to buried solvent-accessible surface area (bSASA)	[66]
ΔvinaXGB	machine-learning-based	58 terms from Vina, 30 terms related to bSASA, 3 features related to water effect, 2 features related to ligand stability, 1 feature related to ions	[67]
Kdeep	machine-learning-based	Voxelized 24 A representation of the binding site considering 8 pharmacophoric-like properties	[68]
Pafnucy	machine-learning-based	Atomic coordinates and 19 features associated with atom type, hybridization bonds, pharmacophoric-like properties or partial charges	[69]
DeepDock	machine-learning-based	Ligand-based information and structure-based information	[70]
XGB-Score	machine-learning-based	Terms from RF-Score and Vina	[71]

^aAllows certain protein flexibility in docking.

3.3.1. Force field-based scoring functions

Classical force field-based scoring functions calculate Coulombic (for electrostatics) and Lennard-Jones (for Van der Waals) terms for non-covalent interactions. More sophisticated forms as an extension of force field-based scoring functions also consider the terms of hydrogen bonds, solvations and entropic contributions.

3.3.2. Empirical scoring functions

Empirical scoring functions are often related to force field-based scoring functions because they consider similar molecular contributions such as hydrogen bonds, ionic interactions, hydrophobic effects and entropic change. Summation of all these components results in the binding free energy, which is ultimately calculated by empirical scoring functions. The empirical coefficients are trained by fitting the scoring function to the experimentally measured binding affinities. Therefore, these coefficients strongly rely on the experimental data set for fitting, which is the main drawback of empirical scoring functions.

3.3.3. Knowledge-based scoring functions

Knowledge-based scoring functions derive atomic pair potentials by statistical principles to extract preferred interaction geometries [72]. High efficiency is one of the major characteristics of knowledge-based scoring functions, in addition to the implicit consideration of the solvent effect, resulting in more common application of knowledge-based scoring functions than other scoring functions [36, 59]. Example knowledge-based scoring functions include DrugScore [59], PMFScore [60], and ITScore [37, 38, 73, 74, 75].

3.3.4. Machine learning-based scoring functions

In recent years, methods based on machine learning (ML), including deep learning (DL), have been applied in scoring functions and demonstrated tremendous success [76, 77, 78, 79]. For example, RFScore [80] and NNScore [81] are two pioneering ML-based scoring functions. These methods have been demonstrated to perform better and more effective than classical methods [82, 83]. However, despite their significant advancements, machine learning scoring functions (MLSFs) have faced skepticism due to issues with generalization [84, 85, 86, 87, 88, 89]. For example, Gabel et al. [84] found that their MLSFs trained on random forest (RF) and support vector machine (SVM) could reproduce the claimed excellent scoring power (the capability for binding affinity prediction) of RFScore [80], but their docking power (the capability for discrimination of near-native poses from decoy poses) and screening power (the capability for discrimination of active ligands from decoy compounds) are rather bad. Nonetheless, efforts persist in enhancing MLSFs to broaden their applicability [66, 67, 70, 90, 91, 92, 93]. Zhang’s group successfully developed ${\mathrm{\Delta }}_{\text{Vina}}{\mathrm{R}\mathrm{F}}_{20}$ [66] and ${\mathrm{\Delta }}_{\text{Vina}}\mathrm{X}\mathrm{G}\mathrm{B}$ [67], which employ ML algorithms to fit correction terms to AutoDock Vina scores rather than conventionally fit the final binding scores. These methods have excelled in ranking the top three in all four tasks (scoring, docking, ranking, and screening) in the Comparative Assessment of Scoring Functions (CASF) benchmark. They further proposed ${\mathrm{\Delta }}_{\text{Lin\_F9}}\mathrm{X}\mathrm{G}\mathrm{B}$ [90], a novel scoring function that employed a similar parametrization strategy based on another clasical scoring function Lin_F9. This new scoring function achieve excellent scoring, ranking, screening powers on the CASF benchmark. OnionNet-SFCT proposed by Zheng et al. [91] also incorporated a correction term to AutoDock Vina, but this term was calculated based on a binding pose prediction model rather than binding affinity prediction, implying that the training of the model still relies on the involvement of decoy poses. The linear combination of Vina scores and root-mean square error (RMSD) values produced by the ML model enabled the method to yield impressive results in docking and screening tasks. More recently, Méndez-Lucio et al. [70] introduced DeepDock, a geometric DL-based method that introduced a mixture density network (MDN) to learn the probability density distribution of the distance between each ligand atom and each point in the surface of the binding site. This distance likelihood potential performed competitively in docking and screening tasks, hinting at its potential for capturing protein-ligand interactions effectively. Shen et al. [94] reported a DL approach named RTMScore based on residue-atom distance likelihood potential and graph transformer for the prediction of protein-ligand interactions. Their method achieved remarkable performance on CASF-2016 benchmark in terms of docking and screening powers. These advancements underscore the continuous evolution of ML and DL techniques in refining scoring functions and their applications in drug discovery and protein-ligand interaction studies.

The significant growth that machine learning, including deep learning has experienced in the field of drug discovery over recent years is propelled by advancements in data-driven techniques [95, 96]. While molecular docking has demonstrated success in various applications such as protein-ligand docking and drug discovery, physics-based models have been employed to comprehend the interactions between macromolecules and small molecules [96, 97, 98]. However, these models are not well-suited for handling the rapidly expanding experimental datasets. In contrast, machine learning screenings frequently surpass molecular docking predictions [96]. They adeptly forecast drug-related properties and predict drug-target interactions, thereby identifying potential repurposable drugs [96]. The chEMBL database [99] comprises inhibitor datasets for various targets, featuring experimentally established binding labels that serve as the foundation for machine learning predictions [96]. DrugBank, an openly accessible database, houses a diverse collection of compounds, encompassing approved drugs, investigational or off-market drugs, along with their interactions with targets [100, 101]. In a recent investigation, machine learning was employed to repurpose DrugBank compounds for addressing opioid use disorder [96]. These models were applied to screen 8865 compounds within the DrugBank database against four opioid receptors for their potency [96].

The popularity of deep learning has surged, primarily driven by its ability for universal approximation and the increasing availability of powerful, high-performing computational tools [102]. More recently this method in drug discovery has immensely benefitted from the incorporation of Artificial Intelligence (AI) Deep Learning techniques [103, 104, 105], culminating in the phenomenal success of Alpha-Fold2 (AF2), a DL algorithm developed by Google DeepMind [106]. Additionally, DeepMind has collaborated with the European Bioinformatics Institute (EBI) to establish AlphaFoldDB [107], providing open access to a repository containing over 200 million predicted protein structures generated by AlphaFold, providing large coverage of UniProt [108]. DeepMind’s AlphaFold, a prominent deep learning algorithm, gained widespread attention in 2020 for its exceptional performance in the 14th Critical Assessment of Structure Prediction (CASP14) competition, demonstrating remarkable accuracy in predicting protein structures. CASP comprises various categories, including template-guided modeling and template-free modeling [75, 109] and plays a pivotal role in assessing the effectiveness of computational methods in predicting three-dimensional protein structures. In a recent investigation, a new modeling category known as protein-ligand complex modeling was introduced in Round 15 of the Critical Assessment of Structure Prediction (CASP), which was traditionally centered on protein modeling [75]. Zou group participated in predicting the protein-ligand targets in the CASP15, and using a novel template-guided approach that combines a physiochemical, molecular docking method and a bioinformatics-based ligand similarity method, they achieved commendable performance in the CASP15 protein-ligand complex structure prediction [75].

Recently, Stark et al. [110] developed EQUIBIND, a deep neural model which relies on SE(3)-equivariant graph neural networks to predict bound protein-ligand conformations in a single shot, treating docking as a regression problem. EQUIBIND tried to tackle the blind docking task by directly predicting pocket keypoints on both ligand and protein and aligning them. Lu et al. [111] proposed Trigonometry-Aware Neural Networks, TANKBind, which incorporates trigonometry constraints as a rigorous inductive bias into the model and explicitly attends to all possible binding sites for each protein by segmenting the whole protein into functional blocks. In terms of blind docking performance, TANKBind achieved state-of-the-art performance, outperforming EQUIBIND. Stark et al. [112] presented DIFFDOCK, a diffusion generative model over the non-Euclidean manifold of ligand poses. DIFFDOCK significantly outperforms all previous methods on PDBBind [113], and provides confidence estimates with high selective accuracy. More recently, Lu et al. [114] presented DynamicBind, a deep learning method that employs equivariant geometric diffusion networks to construct a smooth energy landscape, thereby promoting efficient transitions between different equilibrium states. DynamicBind unifies two conventionally separated steps: protein conformation generation and ligand pose prediction. This method outperforms previous methods [110, 111, 112] in predicting ligand poses for both the PDBBind dataset and major drug targets (MDT) dataset.

3.3.5. Recent development in scoring functions

It is worth noting that there is no universal scoring function that can fit all biological systems, and it is not appropriate to do so because the existing scoring functions remain to have limitations and most of them are sensitive to the training set or have a limited scope of application. Therefore, when docking against a specific target family, training a scoring function with the particular protein family can usually improve binding mode and binding affinity prediction, which is referred to as target-specific scoring functions [115]. The consensus scoring approach is another way to overcome the deficiencies of individual scoring functions to some extent. Consensus scoring functions can mitigate the bias of a single scoring function and improve the hit rate [116, 117]. The consensus appoach should not include multiple scoring functions that yield highly correlated results; otherwise the bias of these scoring functions would be amplified rather than diminished [116]. Other ongoing methodology development includes adding additional terms to the existing scoring functions to improve the scoring performance [118, 56]. Numerous comparison studies have been reported to evaluate the performance of different scoring functions [36, 119, 120, 121, 122].

4. Docking Programs

In the past two decades, more than 60 docking programs have become available [123], such as SLIDE [63], X-CSCORE [64] and ConsDock [124]. Most of the existing docking algorithms treat ligands as flexible molecules and proteins as rigid molecules, because a protein usually consists of thousands of atoms. In some programs such as GalaxyDock [61], protein side-chains can be optimized.

Moreover, there are indirect ways to consider the protein induced fit effect in ligand binding. A common approach is to utilize ensemble protein conformations. These ensemble protein structures can be obtained either from computational analysis such as molecular dynamics (MD) simulations or from experimental measurements such as NMR or X-ray diffraction. The MDock software incorporates a method referred to as “ensemble docking” to implicitly account for protein flexibility. In this method, the protein conformational number is introduced as an additional variable during docking optimization. This ensemble docking algorithm is computationally efficient; its computational speed is comparable to that of single-protein docking [37, 38, 125]. Mizutani et al. used an enlarged binding pocket in their docking program ADAM to allow for certain protein flexibility [126]. In this approach, the van der Waals energy curve for each atom pair is shiftied to widen the protein cavity uniformly, followed by structure optimization to remove atomic clashes and to improve interaction energy. Similarly, Bottegoni et al. adopted the SCARE (SCan Alanines and REfine) protocol, which enlarges the binding pocket by alanine mutation within the binding pocket [127].

To enhance the success of docking procedures, certain docking programs focus on specific protein or ligand categories. Rather than creating a “one-size-fits-all scoring function”, improving prediction accuracy can be achieved by training a system-specific scoring function or incorporating additional factors into energy calculations. For instance, the ITScore in MDock has been effectively retrained using 2897 protein-ligand complexes sourced from the Protein Data Bank, proving valuable in screening for inhibitors and assessing their selectivities across multiple target proteins [38, 128]. Additionally, Vina-carb [57] integrated the Carbohydrate Intrinsic (CHI) energy terms into the AutoDock Vina scoring function (ADV), improving the docking accuracy to 74%, compared with the 55% when relying soley on the ADV score.

5. Structure-based virtual screening and inverse virtual screening

The applications of molecular docking in therapy interventions can be bidirectional depending on the purpose of docking. Conventional docking searches for a compound from a DrugBank [100, 101] or from a compound library, a process referred to as virtual screening (VS) [116, 129]. On the other hand, for a specific ligand of interest, inverse docking is useful for identification of the targets from thousands of protein candidates, a process called inverse virtual screening (IVS) [117, 130, 131, 132]. Both prediction tasks are challenging and require immense calculations.

Virtual screening methods are classified into structure-based virtual screening and ligand-based virtual screening [133]. Structure-based virtual screening is explored by docking methods on a large scale. In contrast, ligand-based virtual screening methods utilize all the structural and chemical information (e.g., by searching the common features or similarity) rather than using a docking strategy. These two methods are not competitive, and can be used individually or in combination. For instance, a ligand-based prescreening step can be done for preselection in order to reduce the size of the compound database in many structure-based virtual screening applications [116]. Tan et al. integrated the two methods into parallel selection and postulated that structure-based and ligand-based screening could be more effective in a complementarity manner than rank fusion [134].

IVS reverse docking has been widely applied to drug development in recent years [130, 135]. Screening from a drug target database can help rapidly identify potential target candidates for a specific molecule. The molecule can be a non-drug ligand or a known drug. The major safety issues of drug candidates are possible side effects (referred to as off-target effects) and toxicity. IVS can help investigate the interactions of the drug candidate with multiple targets in early-stage drug development. The purpose of IVS studies on known drugs is to search for alternative uses of FDA approved drugs, which reduce costs and risks in comparison with new chemical compounds. This approach is referred to as drug-repositioning (also known as drug repurposing or reprofiling) [136, 137, 138]. Drug development has been trending from “one gene, one drug, one disease” to polypharmacology, which means that drugs are designed to interact with multiple targets. The main challenge of IVS is the costly computational time. Recently, Zou group integrated their fast docking application, MDock, into a webserver platform in order to undertake the IVS tasks. The server is built-in with 3268 protein-ligand complexes (containing 3349 distinct binding sites) associated with 537 protein targets. A typical IVS run for a given ligand takes 1 to 4 hours. For the example, the IVS study on the progesterone receptor takes approxiamately one hour on a workstation with 24 Intel Xeon cores [Intel(R) Xeon(R) CPU E5–2650 v3 @ 2.30GHz]. In addition, the screening results from our web server provide searchable target diseases information [139].

6. Challenges

Molecular docking methodology is rapidly evolving and improving with the exponential increase of computing power. Recently, machine learning algorithms (including neural network and deep learning) [140, 141] and hybrid approaches from high-content screening and virtual screening [142] have been proposed for docking predictions and drug discovery. A critical question in docking is how reliable the docking predictions are [143]. Consequently, benchmarking excises for blind docking prediction have been held, such as Community Structure-Activity Resource (CSAR) [144], Drug Design Data Resource (D3R) [145, 146, 147] and Continuous Evaluation of Ligand Pose Prediction (CELPP) [148]. Zou group participated in the weekly CELPP competition using a fully automated systematic strategy [149]. These large-scale protein-ligand complex structure predictions are statistically significant. Through our analysis, the current sampling (80.4%) is much better than the scoring (46.3%) [149]. The inaccuracy of current scoring functions likely arises from various sources, such as solvent effects, entropic effects, protonation or tautomer states, and so on. Development of new scoring schemes to address these issues is at large. The biggest hurdle in molecular docking stems from protein flexibility. The computational cost is immense for sampling all possible protein conformations. In conclusion, despite decades of docking algorithm development, prediction accuracy remains a major issue.