top of page

VoteDock: Consensus docking method for prediction of protein–ligand interactions

Molecular recognition plays a fundamental role in all biological processes, and that is why great efforts have been made to understand and predict protein-ligand interactions. Finding a molecule that can potentially bind to a target protein is particularly essential in drug discovery and still remains an expensive and time-consuming task. In silico, tools are frequently used to screen molecular libraries to identify new lead compounds, and if protein structure is known, various protein–ligand docking programs can be used. The aim of docking procedure is to predict the correct poses of ligand in the binding site of the protein as well as to score them according to the strength of interaction in a reasonable time frame. The purpose of our studies was to present the novel consensus approach to predict both protein–ligand complex structure and its corresponding binding affinity. Our method used as the input the results from seven docking programs (Surflex, LigandFit, Glide, GOLD, FlexX, eHiTS, and AutoDock) that are widely used for docking of ligands. They evaluated it on the extensive benchmark dataset of 1300 protein–ligands pairs from refined PDBbind database for which the structural and affinity data was available. They compared independently its ability of proper scoring and posing to the previously proposed methods. In most cases, our method is able to dock properly approximately 20% of pairs more than docking methods on average, and over 10% of pairs more than the best single program. The RMSD value of the predicted complex conformation versus its native one is reduced by a factor of 0.5 Å. Finally, they were able to increase the Pearson correlation of the predicted binding affinity in comparison with the experimental value up to 0.5.

The VoteDock protein–ligand docking algorithm. The main goal of the VoteDock is to provide fast and accurate prediction method for 3D structure of a protein–ligand complex. It facilitates data exchange between various prediction docking methods, publicly available software, evaluation programs and visualization modules. The general model of the information flow and components of the algorithm are presented in the following diagram.

A typical drug design campaign requires substantial costs and is time consuming due to the fact that for thousands of chemical compounds biochemical screening has to be performed before proceeding to a more refined analysis. The in silico methods promise to shorten the time and decrease the amount of work needed when searching for a new inhibitor. One of the most important methods used here is the molecular docking that predicts a preferable conformation of a ligand when bound to a receptor molecule. Docking is used frequently in a high-throughput virtual screening where large libraries of commercially available compounds are searched to find the most active compound for a selected protein target. The aim of docking procedure is to predict the correct pose of a ligand in the binding site of the protein as well as to score it according to the strength of the interaction in a reasonable time frame. As all programs exploit empirically based scoring functions and algorithms, docking results are sometimes far from reality.

Among the most frequently reported is the docking accuracy of small organic compounds to a given protein, yet the nucleic acids can also be considered as a target for ligand molecules. In the protein-protein docking, the interactions between two identical or different proteins are studied. In the case of protein-ligand docking, various algorithms address different representations of a ligand and a receptor, their intrinsic chemical properties, and detailed characteristics of intramolecular interactions between their atoms. As in recent years, the crystallography and multidimensional NMR provided a wealth of structural information about various biological targets, several protein–ligand docking programs have been proposed. Usually, the receptor is treated as a rigid molecule because of high computational costs, whereas conformational flexibility of ligands is taken into account leading to different placement algorithms.The scoring procedure of such docked conformers is still regarded as one of the most difficult tasks in molecular docking because of their empirical nature. In our work, they used only software that considers flexibility of ligands, not proteins, and thus structure of protein before docking was not changed in comparison with original pdb file, assuring that protein is already in bounded state.

There are three major goals of docking simulations: (1) the native conformation of ligand in the active site should be predicted; (2) the binding energy should be estimated allowing for arrangement of the tested set of ligands according to their affinity toward the protein target; (3) in addition, it should be fast enough to screen large collections of small chemical molecules. The typical docking procedure is performed in two steps. The first step is focused on placing a small molecule into the binding site of the protein using mostly geometrical features and searching for its best three-dimensional (3D) conformation inside the cavity. The second step is performed using different scoring functions and it leads to the estimation of the binding affinity between the protein and the ligand.

During the last two decades, a set of different docking programs has become available both for commercial and academic use, such as DOCK, AutoDock, FlexX, Surflex, GOLD,ICM, Glide, CDocker,LigandFit,MCDock,and many others. They are based on different algorithms and can be grouped into four general categories: stochastic Monte Carlo, fragment-based, genetic algorithms, and, finally, shape complementary methods. None of those programs uses systematic search to fully explore all degrees of freedom in both receptor and ligand molecules because of enormous computational cost of such a procedure.That is why docking programs avoid systematic search and perform only guided search in conformational space. Our consensus algorithm attempts to combine those independent docking approaches into a single and powerful prediction method. They select a set of representative conformations from each docking algorithm to efficiently inspect different guided search algorithms for correct conformation of a protein–ligand complex.

The binding affinity of generated output protein–ligand conformations is calculated here by using different scoring functions. More than 30 different scoring functions were published until 2009 and they can be classified into three major categories. The first group applies force fields functions to calculate the energy of a complex as the sum of the ligand and the receptor internal interaction energies and also the energy of intermolecular interactions between those two molecules. Typically, the force fields such as Assisted Model Building With Energy Refinement (AMBER) or Tripos are employed, considering two energy terms, i.e., van der Waals and electrostatic interactions between molecules. Additionally, to improve the accuracy of those functions, sometimes the solvation energy term is also included, usually using a distance-dependent dielectric function.Most of the docking programs do not support ligand binding to protein via covalent bond. However, when applied to protein-ligand complexes, the force fields are often found to overestimate the binding affinity, even when using very precise and time-consuming procedures. Therefore, the scaling coefficients multiplying both terms are used to resolve this problem.

The second group, i.e., the empirical scoring functions, describes interactions between a protein and a ligand as scalable parameters. Almost all of the proposed parameters exploit hydrogen bonds, hydrophobic interaction, metal bonds energy, typical force fields energies, and finally, the solvation energy term. The scaling parameters together with the empirical functions are trained on the selected dataset of complexes with known binding affinity for which scaling factors for each energy term can be optimized. Empirical scoring functions are often able to recreate binding affinities of original training dataset with very high accuracy, yet the results on previously unconsidered protein–ligand complexes are not always successful. The third group, namely knowledge-based scoring functions, is developed from the statistical analysis of X-ray and NMR structures of protein–ligand complexes. The distribution of different pairs of atom types is gathered using a set of pairs of atoms, one from a protein and the other from a ligand, and then converted into pairwise atom–atom statistical potentials. The final interaction energy is calculated as the sum of all pairwise interactions between atoms from a ligand and a protein lying within the sphere of the given cutoff (usually from 6 Å up to 12 Å).

The consensus is a novel technique recently used in various applications, mostly in bioinformatics. The main rationale behind is that although individual approaches can generate some misleading results, yet the distribution of those errors is random.That is why even a simple majority voting of a set of programs providing different results can be in principle much closer to the correct answer, than even the best single program. In the context of docking problem, several attempts to transfer that approach have been made. However, as the general opinion is that posing is not the main drawback of docking programs, typically consensus approach is applied in prediction of ligands activities. Nevertheless, some cases where authors applied this technique to poses selection were also reported. For example, Wolf et al.merged two docking algorithms, namely genetic- and fragment-based method into a single AutoxX protocol. The software used FlexX and AutoDock algorithms for choosing optimal ligand conformation, and it was able to decrease the mean root mean square distance (RMSD) of top score conformations by 0.3 Å in comparison with best individual program from those two. This approach allowed to predict correct conformation of ligand for 126 pairs of the 206 tested (RMSD below 2 Å), more than six for AutoDock alone. However, no consensus scoring was proposed there, thus, scoring functions were omitted and reported separately from those two programs.

Up to now, the research community has focused mostly on improving scoring predictions, because in common opinion, calculating a ligand in vitro activity is very difficult task. Therefore, typical strategy is to gather data from diverse set of scoring functions representing different approaches to create new function using simple linear regression technique. Typically, this procedure allows for development of the function working for specific protein families; therefore, it cannot be transferred from one family to another. Similar approach was used by Teramoto et al. where authors used the support vector regression performed on three protein families, acetylcholine esterase, thrombin, phosphodiesterrase 5, and proliferator-activated receptor gamma. New functions were used as an input scoring results obtained from F-score, D-score, Potential of Mean Force (PMF), G-score, and ChemScore. Those authors in 2007 used “rank-by-vote” approach, where instead of the absolute scores values, each ligand was given the rank based on its position in ligand list ordered by particular scoring function. Ligand with lowest average rank from the set of scoring functions was then chosen in this method as the most active one. Similar approach is also used in Sybyl's Consensus Score (CS) model. The successful modification of “rank-by-vote” approach was implemented in SeleX-CS algorithm developed by Bar-Haim. Here, the Monte Carlo simulated annealing is used to choose functions that can vote for a particular ligand. Two types of votes are allowed: “primary” rank-by-vote value, and “secondary” rank-by-number value. Authors reported three times increase in enrichment factor value obtained for studied small set of proteins. Summarizing, according to our knowledge, no single workflow that combines consensus both in pose prediction and score prediction has been introduced up to now.

Here they propose the consensus docking protocol that allows for massive docking of ligands into their corresponding protein targets using several independent docking algorithms and scoring functions running in parallel. Our approach combines the results from various programs into a single consensus prediction of the 3D protein–ligand complex structure. The clustering of results from those several docking algorithms is performed to select the poses that are close to the corresponding native conformation, and then the consensus scoring is performed using the multivariate linear regression to select the strongly binding conformations. The consensus docking method is evaluated here in terms of both posing and scoring abilities on the large dataset of protein–ligand complexes with known 3D structures and binding affinities.

Plewczynski D, Łażniewski M, Grotthuss MV, Rychlewski L, Ginalski K. VoteDock: Consensus docking method for prediction of protein-ligand interactions. Journal of Computational Chemistry 2011;32:568–81. doi:10.1002/jcc.21642.


bottom of page