Conformational search methods

One of the most challenging aspect of pose prediction is the conformational search algorithm. In general, there are two main classes of conformational search methods, systematic and stochastic. The users do not always have a choice to select the search algorithm because the developers usually select based on their expertise and code development. Nevertheless, there are programs, such as AutoDock [7], that allow users to select the search algorithm. Here, we succinctly discuss commonly used conformational search methods. However, this list is far from being comprehensive, interested readers are directed to consult the following papers for detailed discussion [8, 9].

Molecular dynamics simulations

At this point, it is appropriate to gently introduce another powerful conformational search method, namely molecular dynamics (MD) simulations. It simulates the atomic motions as a function of time by solving the Newton’s equation of motion [17]. Molecular docking and MD simulations are complementary methods commonly used in structure-based drug discovery to understand protein-ligand interactions and their dynamics [18, 19]. For computational efficiency, molecular docking algorithms treat receptors atoms rigid and the ligand as flexible. This leads to incorrect pose prediction when induced fit binding is observed. Therefore, MD simulation is an elegant method to incorporate such induced fit effects. This can be achieved in two ways, as shown below:

• Pre-docking step to sample various conformations of the receptor without the influence of the ligand. This can serve as multiple receptor conformations for docking.
• Post-docking step to refine the docked receptor-ligand conformation by allowing the complex to evolve to their realistic conformation, which would be more probable physiologically.

Implementation of artificial intelligence in molecular docking

With a near exponential increase in the tertiary structure information in the last 3 years attributed to several artificial intelligence based methods, particularly Alphafold [20] and RoseTTAFold [21], many efforts have been made to incorporate machine learning methods in drug discovery to improve the quality of therapeutic agents and reduce the drug discovery timelines significantly [22]. Developers are also constantly developing methods that utilize machine learning techniques to enhance the accuracy of molecular docking predictions. Incorporating the state-of-the-art deep learning neural networks have shown to improve conformational search algorithms and develop better and more generalised scoring functions. These improvements have helped ameliorate challenges posed by limited structural and small molecule data [23, 24]. AI techniques enhance traditional molecular docking methods by introducing innovative strategies such as network-based sampling [25] and unsupervised pre-training [26, 27]. AI-Bind [28], for instance, combines network science with unsupervised learning to mitigate issues like over-fitting and annotation imbalance. It identifies negative protein-ligand pairs using shortest path distances and learns node feature representations without relying heavily on limited binding data. By utilizing node embeddings from extensive chemical and protein structure collections, AI-Bind captures a broader range of structural patterns. Additionally, models like IGModel [29] leverage geometric graph neural networks to incorporate spatial features of interacting atoms, improving binding pocket descriptions. AI-Bind can also predict active binding sites from amino acid sequences alone, without needing 3D structures. Overall, these AI-driven approaches significantly improve the accuracy and generalization of predicting protein-ligand interactions, representing a major advancement over traditional methods.

Scoring functions

Scoring functions are designed to reproduce the binding thermodynamics [30, 31], i.e., binding enthalpy and entropy (see Eq 1).

(1)

Where is the enthalpy component of binding and the entropy component of binding at temperature “T”. The enthalpy of binding can be estimated by the number and type of interactions between the receptor and ligand at the atomistic level. Hence, scoring functions often estimate the enthalpy component by summing all interactions of different types. However, this is a non-trivial task, which has also been criticised for being treated as purely additive [32, 33]. Nevertheless, different types of scoring functions employ different methods to fit experimental binding data to develop scoring functions. This leads to different types of scoring functions that can be classified into four types. We briefly discuss the scoring functions and their types, for more detailed discussions the readers are directed to consults these papers [34, 35].

For a docking tool to be effective and yield reliable result, the search algorithm and the scoring function are crucial and carefully developed. The single most important question that is often asked about a docking algorithm is “How efficiently can the algorithm differentiate between binders and non-binders?â€#157; For the docking algorithm to efficiently differentiate between them, both the conformation search algorithm and the scoring functions must work in tandem at their best capacity. A huge data set of chemical decoys is employed to validate the accuracy, which is spiked with known binders. Chemical decoys are chemical structures that physically resemble known binders but have a different topological arrangements of atoms. The accuracy of the docking algorithm is assessed by its ability to recover known binders from the pool of decoys. However, it is overambitious to expect any docking algorithm to recover all the known binders and none of the decoys. To quantify the accuracy of the docking algorithms and the parameters used, an enrichment factor (EF) [36] is used, which is computed using Eq 2.

(2)

Here, ligands_(total) and denote the total number of known binders and binders recovered, respectively, and N_(total) and denote the total number of decoys and recovered decoys, respectively. From the formula, we can infer that a higher value suggests an effective docking algorithm and parameters thereof. A comprehensive discussion on search algorithms and scoring functions is beyond the scope of this article, and the readers are advised to read the listed papers in the literature [37, 38]. However, we enlist the scoring functions with their succinct descriptions.

Force field-based scoring function.

Force fields [39] are mathematical frameworks that describe the potential energy of a molecular system as a function of the positions of its atoms. In other words, they describe chemical information in the numerical information that is more suitable for computations. To do so, various components, each representing a specific type of interaction between atoms. Each component is characterized by its own parameters, such as the strength of the interaction and the range of distances (r) over which it is effective. A force field typically consists of:

1. Bonded Interactions
   1. (a) Bond stretching
   2. (b) Angle bending
   3. (c) Torsional interactions
   4. (d) Improper torsions
2. Non-Bonded Interactions
   1. (a) Van der Waals forces
   2. (b) Electrostatics

The bond stretching and angle bending are generally considered hard terms because the energy required to bring about a small change is generally very large. Moreover, to improve the computational speed, these are treated as “fixed” or “constant”. The torsional and improper terms are changed to bring about a conformational change by the conformational search algorithm. Owing to such changes, the non-bonded interactions vary as different conformations lead to different atomic interactions, which is show in Eq 3 as a docking score.

(3)

Where r_ij is the atomic distance between protein atom i and ligand atom j, A_ij and B_ij are vdW repulsive and attractive parameters, respectively. q_i and q_j are atomic partial charges on atoms i and j. The first two terms estimate the van der Waals interactions using Lennard-Jones 12â€"6 potentials, which are commonly used in most force fields. The last term estimates the electrostatic component of the interaction, with solvent effects implicitly defined using the term , which is the distance-dependent dielectric constant. It is known that incorporating solvation effects is far from perfect, and without it, the electrostatics will be overestimated for the charged ligands owing to the absence or incorrect dielectric effect seen in the solvated systems. There are several approximations introduced such as Generalised-Born (GB) [40] and Poisson-Boltzmann (PB) models [41, 42], which are popular and computationally efficient methods that estimate solvent effects.

Empirical scoring function.

(4)

Where , , and , depict the free energy contribution of van der Waals’, electrostatic, solvation/desolvation, etc. , , and are the corresponding coefficients that are adjusted to fit the experimental binding affinity data. The final score is obtained as the sum of all the energy components. Different scoring functions in this class include or exclude certain components, while others use a different fitting method that best fits the binding affinity data.

Knowledge-based scoring function.

(5)

Where w(r) is the final score which is dependent on the atomic distances r. k_B is the Boltzmann constant, T is the absolute temperature in K. is the frequency of protein-ligand atom pairs at distance r, and is the frequency of protein-ligand atom pairs in the reference state at distance r. The reference state is generally the frequency that a particular atom pair is observed in the experimental structures. The type of interaction (intermolecular atom pairs) that appears more frequently in the reference state will be scored higher. Therefore, the ratio can be considered as a pair distribution function.

Consensus scoring function.

(6)

Where W_1,2,3,.. are weights assigned to each of the scoring function 1, 2, 3, etc. These weights may not be directly linked to the atomic physicochemical properties. As we have seen so far, each type of scoring function comes with its own set of advantages and disadvantages. There have been many attempts to use more than one type of scoring function simultaneously to improve their overall scoring accuracy. An ideal consensus scoring function will use one or more scoring functions that complement the pros and cons. This is expected to minimise the cons and amplify the pros of each scoring function. However, this is far from reality, and often times, new limitations are observed.

Few questions that frequently arise, especially among the beginners, is what are generally the units of the docking score? The units are usually units of energy, expressed as “kcal/mol” or “kJ/mol”. However, it is not unusual to observe scoring functions that yield unitless docking scores. Irrespective, of the units of the docking score, a generally accepted convention is such that a more negative score indicates a better binding score. We would also point out that not always a more negative score indicates “correct” binding conformation (also referred to as “pose”). Are docking scores obtained from different scoring functions directly comparable? This is rather difficult to answer holistically because scoring functions could be comparable as they share common philosophy of development or datasets. In our experience, we have always avoided any head-to-head comparison of scores, nevertheless, we compare the rank-ordering given by different scoring functions [43].

Tip 1: Know your biological target and its functional state

The quality and accuracy of the results obtained from molecular docking, largely depends on the quality of the tertiary structure of the biological target. The structures are resolved using X-ray crystallography, Nuclear Magnetic Resonance (NMR) spectroscopy, and cryogenic electron microscopy (CryoEM). These methods provide with high quality structural data, however, introduce errors such as missing atom information or multiple amino acid rotamers [44, 45]. Furthermore, X-ray crystallographic and CryoEM resolved structures are single snapshot of the receptor structure, while NMR derived structures provide insights in to the dynamics. Hence, it is crucial to evaluate and correct these structural defects before using them for molecular docking. Emphasis should be on the binding site amino acids, their protonation states, and appropriate rotameric states. Furthermore, the receptor’s conformational state or functional states are important. Since most of the experimental structures do not contain this information, use of multiple structures resolved experimentally under different conditions or use MD simulations is highly recommended to include target flexibility. The worldwide Protein Data Bank (wwPDB) is the primary source for receptor structures [46]. In the absence of experimental data, computational methods, like template-based modelling and ab initio predictions, can be used to predict the tertiary structures from primary sequences. These theoretical models require thorough validation to ensure accuracy of the predicted tertiary folds. With recent developments in AI-based methods in protein structure prediction, such as AlphaFold [20], high quality structure are readily available in the wwPDB annotated as Computed Structure Models (CSMs).

Tip 2: Choose the receptor entry in the wwPDB that is closer to the biological problem at hand

It is possible on many occasions that several entries for the same receptor are available on the wwPDB. One must choose based on the conformational state of the receptor that is desired. Then selecting the entry with a good resolution, in our opinion, a resolution below 2.0 Šis a good choice whenever possible. There are other structure quality parameters, apart from the resolution, that provide useful information about the target. The b-factors or beta-factors provide key information on the flexibility of the atoms. It also holds key information on the confidence with which the crystallographers are able to narrow down the atomic positions while fitting the electron density maps, for examples atoms with b-factors Ų provide good confidence in their assignments. While using the CSMs, predicted local distance difference test (pLDDT), ranges from 0 to 100, depicts a per-residue measure of local confidence that must be carefully examined. pLDDT values suggest good confidence in the prediction, while values <50 indicate very low confidence. The users are directed to the Guide to Understanding PDB Data with a section dedicated to CSMs. The choice of receptor should also bear in mind if there exists an entry with a co-crystallised ligand in the binding site of interest. While preparing the receptor, the atomic resolution limitation makes it difficult to locate the position of hydrogens, which must be added in the preparation step. Partial atomic charges are calculated and assigned for individual atoms. The decision to keep important metals, co-factors, and water molecules in the docking site is made at this stage. If they are crucial for ligand binding, they must be retained; otherwise, they can be removed. Such decisions are far more complex than they appear at first sight. To come to a meaningful decision, the computational chemist must be well aware of the biology and functioning of the receptor in question ( see Tip 1). Recently, there has been a particular interest in developing tools that systematic evaluate the energetics of displacing water molecules from the binding site by the ligands. In doing so, they are well placed to predict crucial water molecules that may lead to significant entropic gains. The tools such as waterMap [47, 48], WaterDock [49] and SSTmap [50] are available that predicts the energetics of displacing water molecules from the binding site. Deleting co-factors, waters, or metal ions deemed important can lead to meaningless results, even though very high or favourable docking scores are observed.

Tip 3: Druggability is the key when defining a potential binding site for docking

Docking on a receptor structure is performed in a fixed space, called the binding site. Generally, one must choose the receptor structure from the wwPDB that contains a co-crystallised ligand within the binding site of interest. The search space for docking is defined by selecting amino acid residues 3-6 Å around the ligand. However, we may not always have information on the location of the binding site; in such cases, we can perform a blind docking approach, wherein, the whole receptor surface is defined as the search space. Alternatively, several algorithms are available that can locate potential binding sites, which are ranked according to their likelihood of being a druggable site. Here are a few examples of programs used to locate potential binding sites, SiteMap [51], LIGSITE [52] and Fpocket [53] If the docking program allows target flexibility to accommodate conformational changes induced by the ligand, the number and identity of flexible amino acids and the degree of flexibility are defined.

Tip 4: Learn the fundamentals of the docking algorithm

There are a plethora of docking programs available (Table 1), some with commercial licenses and others with open-source licenses. Knowing every program to make a choice that best suits you is almost always impossible. However, this is also not necessary as there exists ample literature reporting the use of many of the tools. Irrespective of the choice of the docking program, the general steps will be same; for example, preparation of the receptor ( see Tip 2), defining the binding site as search space( see Tip 3), preparation of ligands ( see Tip 5), specifying the docking parameters, validating the docking procedure, performing the docking calculations, and analysis and dissemination of the results. In a nutshell, the general principles remain unchanged; however, there are differences in ligand format, docking methods, and scoring algorithms that are well-documented in the user manual. We strongly recommend that before using any docking algorithm, the user manual must be consulted to understand each parameter’s importance. The developers present their report on several validation studies as a part of the software development which is a great resource to understand the effect of each parameter on the outcome. A literature survey of the program’s usage will also be covered in the user manual.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. List of common molecular docking software.

https://doi.org/10.1371/journal.pcbi.1013030.t001

Tip 5: Filter and prepare ligands to enhance accuracy and speed

In this case, the small molecules that are docked into a receptor pocket are referred to as ligands. It is not uncommon for the terms “ligands” and “drugs” to be used interchangeably. However, it is important to note that while all drugs are ligands, not all ligands are drugs. Ligands are typically represented as two-dimensional connectivity of atoms, and hence, called a 2D graph. These are converted to three-dimensional representations and energy-minimized for docking. When dealing with a large library of ligands [54] (ranging from millions to billions), an initial filtering process is applied to select drug-like molecules. Various criteria are considered, such as molecular weight, log P, the number of rotatable bonds, and the number of hydrogen bond donors and acceptors. One such filtering rule or criterion named “Lipinski’s Rule of Five” was put forth by Christopher Lipinski [55]. Since the first publication there have been significant amendments to this rule that have improved the ligand filtering criteria [56]. These filtering criteria are applicable for orally administered drugs that do not involve any active or receptor-mediated transport across the bilayer membranes. Nevertheless, these filters may be used to discard chemical structures that are unlikely to be drug or drug-like molecules. Pan Assay INterference compound S (PAINS) [57, 58] is another filter that helps to weed out molecules that tend to show up as actives in all assays and hence must be treated as “False Positives.”

Ligands undergo checks for proper geometry, ensuring that bond distances and angles are reasonable. If necessary, the conformation of the ligands can be minimized to achieve the correct geometry. However, depending on the conformational search algorithm of the docking program, it would warrant the use multiple fixed conformations or allowing the conformations to be generated on the fly. The formers case is relevant for docking programs that inherently treat ligand and receptor rigid bodies, and their user-defined multiple conformations serves as a source of ligand flexibility. These days most of the docking programs are bundled with their own conformational search engine, while user can still instruct which of the torsional degrees are flexible or rigid (default settings treat all torsions are flexible).

Ligands containing stereocenters are treated as independent enantiomers. A point to remember, in molecular mechanics the tautomers are treated as different molecules as opposed to quantum mechanics. When one wants to evaluate the binding affinity of tautomers, they must be prepared and docked separately. This is particularly important in evaluating which of the tautomeric states are more suitable for binding. A case study on why tautomeric states should be treated with great care is exemplified by chemical class of biguanides and can be found in ref [59]. Furthermore, the protonation states are selected based on the pH that closely mimics the physiological or experimental conditions. Protonation states can be predicted using software tool like Epik [60, 61], PlayMolecule pKAce [62] or Dimorphite-DL [63]. Table 2 enlists commonly used tools for ligand preparation and format conversion.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. List of Tools for Ligand preparation and format conversion.

https://doi.org/10.1371/journal.pcbi.1013030.t002

Tip 6: Compare the docking outcome of known molecules with experimental data to validate the docking protocol

One of the most important, albeit neglected, steps in molecular docking is validation. If the receptor was co-crystallised with a ligand in the receptor site of interest, the protocol must be validated by re-docking it and checking if the experimental binding conformation can be reproduced. However, when there are no bound ligands reported, we advise gathering from the literature a few molecules targeting the receptor in question with experimental binding affinity values reported. After docking these molecules, a rank ordering is carried out to understand if the docking parameters can reproduce the experimental binding affinities in the correct order. We strongly recommend that trying to reproduce the binding affinity values as determined by experimental methods is extremely difficult due to inherent limitations in the scoring functions. These limitations are outlined briefly in the previous section and we also direct the readers towards a comprehensive article for further explanation [65] A large collection of data can be found in the BindingDB or ChemBL. However, it must be noted that these data may contain results from different experimental protocols making a direct comparison difficult.

Another crucial yet often overlooked question is the number of poses that must be generated to ensure more accurate predictions. The number of poses generated in molecular docking varies based on the docking software and the objective of the study. Generally, protocols aim for 10â€"20 poses per ligand (in our group the standard practice is to generate 10 poses minimum) [54] to explore binding conformations effectively. While working with completely new scaffolds, it becomes difficult to establish which docking pose would be more physiologically relevant. In such cases, we recommend evaluating at least two to three diverse poses and then using MD simulations followed by MM-GB/PB-SA methods to estimate the binding affinity (see Tip 7)

Tip 7: Rescore the ligands with independent scoring methods to ensure reliability

It’s important to note that docking scores are mere estimations of the true binding constant (vide supra). They rely on the non-covalent interactions between the ligand and receptor and often disregard the entropic component of binding. Therefore, rescoring must be performed with an independent scoring function to ensure reliability. Three post-docking corrections are outlined here:

• Estimating the impact of solvent
• Using consensus scoring, and
• Calculating the absolute or relative free energy of binding.

Solvent plays a crucial role in ligand binding by forming specific interactions with ligands. The strength of electrostatic interactions depends on the dielectric screening effect of the solvent. Some docking algorithms consider solvent effects, while others do not. If solvent effects are not considered, including a solvation correction to the score can enhance the accuracy. MM-GB/SA and MM-PB/SA are well-known methods that can be used as rescoring methods that use implicit solvation correction to the gas phase energies [66]. Consensus scoring, where the top hits from the docking exercise are rescored using different scoring algorithms, has been effective in improving accuracy. Hits that appear at the top of multiple lists are selected for further investigation. Free energy perturbation (FEP) calculations offer a rigorous way to measure the changes in free energy between unbound and bound complexes in the solvent.

Tip 8: Be consistent with the atomic representations and figures

In addition to advanced rescoring methods, it is also important to visually inspect the docking solution to assess if the protein-ligands show prominent interactions such as hydrogen bonds, salt bridges, water-mediated interactions, cation- and -. Since the results need to be communicated to a wider audience, the colour coding for receptor and ligand atoms must be distinct. The atom colouring and type of interactions must be explained clearly to avoid ambiguity. Additionally, the docking figures should be consistent, and the orientation of the binding poses must clearly illustrate the molecular interactions. This approach will also enable straightforward comparisons of docking images within a congeneric series. Whenever feasible, the docking results for theoretical compounds should be discussed alongside the known molecules, as this will enhance confidence in the findings.

Following are our recommendations while preparing the figures to illustrate the results in either 3-dimensional or 2-dimensional interactions.

1. How to Visualize 3D Interactions?:
   • Use molecular visualization tools like PyMOL [67], Chimera [68], or NGL Viewer [69] to display protein-ligand complexes in 3D.
   • Highlight key interactions such as hydrogen bonds, hydrophobic contacts, and -stacking.
   • Label interacting residues and measure distances between ligand atoms and receptor residues to provide quantitative insights.
2. How to generate 2D Interaction Diagrams?:
   • Tools like LigPlot+ [70], PoseView [71], and Protein-Ligand Interaction Profiler (PLIP) [72] can create clear 2D diagrams.
   • Display interactions such as hydrogen bonds and hydrophobic contacts with same orientation for intuitive comparison.

Tip 9: Communicate the results to the experimentalists with a biological and/or chemical viewpoint

The results from molecular docking and other computational studies aim to inspire new experimental research or provide more insights to biological observations. While experimentalists may not be experts in computational methods, they are well-versed in the biological aspects being investigated. It is essential to communicate findings without using technical jargon that could hinder understanding. It is also important to acknowledge that experimentalists are not entirely unfamiliar with computational techniques. When presenting the results, each score should be explained as clearly as possible with its physiological significance. The experimentalists are more accustomed to reading new results with respect to standard data, in the context of molecular docking, the scores for new molecular designs can be explained in comparison to known drugs or reference molecules. This presents the results with the correct context and self-explanatory manner.

Tip 10: Share input files and docking parameters to comply with FAIR guidelines

There have been significant questions raised on the reproducibility of results reported in the area of computational science [73, 74]. Often times, the use of a proprietary software or data not available in public domain makes it difficult to assess the reliability of the methods, and hence the results. A recent criticism on publishing the AlphaFold3 paper in Nature [75] has renewed the need to adopt policies that make data and methods accessible to all, to validate the findings and thus enhance its scope beyond what was envisioned by the authors. The FAIR (Findability, Accessibility, Interoperability, and Reusability) principles are guidelines established in 2016 to improve data management and sharing in scientific research [76]. We briefly explain each term as follows:

• Findable: Data should be easily locatable by humans and machines through unique identifiers, rich metadata, and registration in searchable resources.
• Accessible: Data must be retrievable using open, standardized protocols, with metadata remaining accessible even if the data is not.
• Interoperable: Data should integrate smoothly with other datasets using shared languages and controlled vocabularies, including references to related data.
• Reusable: Data should be well-documented, include clear licensing, and provide provenance information to support transparency and reproducibility.

All the input files, preparation steps, and docking parameters should be made available to the scientific community. This ensures fair dissemination of the results, and those who wish can reproduce the results and build upon their own set of experiments to continue and expand the current work. We also advise that even minor modifications must be reported in the methods. This may appear seemingly small but can be crucial for reproducing the results obtained. This can be indicated in the printed pages of the manuscript but would be more helpful in a repository listed in (Table 3) wherein users can download and test it themselves. Moreover, many of the repositories listed below allow to raise issues whenever users encounter errors or inconsistencies not foreseen by the authors or developers.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. List of repositories for share files in the public domain.

https://doi.org/10.1371/journal.pcbi.1013030.t003