Setting up a large set of protein-ligand PDB complexes for the development and validation of knowledge-based docking algorithms
© Diago et al; licensee BioMed Central Ltd. 2007
Received: 23 October 2006
Accepted: 25 August 2007
Published: 25 August 2007
The number of algorithms available to predict ligand-protein interactions is large and ever-increasing. The number of test cases used to validate these methods is usually small and problem dependent. Recently, several databases have been released for further understanding of protein-ligand interactions, having the Protein Data Bank as backend support. Nevertheless, it appears to be difficult to test docking methods on a large variety of complexes. In this paper we report the development of a new database of protein-ligand complexes tailored for testing of docking algorithms.
Using a new definition of molecular contact, small ligands contained in the 2005 PDB edition were identified and processed. The database was enriched in molecular properties. In particular, an automated typing of ligand atoms was performed. A filtering procedure was applied to select a non-redundant dataset of complexes. Data mining was performed to obtain information on the frequencies of different types of atomic contacts. Docking simulations were run with the program DOCK.
We compiled a large database of small ligand-protein complexes, enriched with different calculated properties, that currently contains more than 6000 non-redundant structures. As an example to demonstrate the value of the new database, we derived a new set of chemical matching rules to be used in the context of the program DOCK, based on contact frequencies between ligand atoms and points representing the protein surface, and proved their enhanced efficiency with respect to the default set of rules included in that program.
The new database constitutes a valuable resource for the development of knowledge-based docking algorithms and for testing docking programs on large sets of protein-ligand complexes. The new chemical matching rules proposed in this work significantly increase the success rate in DOCKing simulations. The database developed in this work is available at http://cimlcsext.cim.sld.cu:8080/screeningbrowser/.
Improving our understanding of protein-ligand interactions at the molecular level plays an important role in the discovery process of new drug candidates. The Protein Data Bank (PDB)  is the main source of structural information on protein-ligand complexes. It is constantly being improved through the addition of on-line tools and links to complementary datasets.
Current virtual screening methodologies for in-silico discovery process of new leads rely on databases of chemical complexes with structural information and also chemical, physical and biological properties, when available. Examples of these datasets are the Cambridge Structural Database , NCI , ZINC  and ChemStar . These databases, however, do not provide clues on the ways these molecules may interact with proteins or other macromolecules, which is important for the task of developing knowledge-based docking algorithms.
Several databases containing information on ligand molecules found in the PDB, and on protein-ligand interactions (e.g. PDB-ligand , PLD , PDBsum , Hic-Up , Relibase [10, 11] and MOAD ) have been released. Furthermore, a few databases collecting data on protein binding sites found in the PDB have been recently reported [13–16]. Creative data mining may provide useful knowledge on protein-ligand interactions from these databases, with potential use in the design of docking algorithms or scoring functions. Testing of docking algorithms, on the other hand, requires a particular processing of both the protein and ligand structures, so that the given docking program has all the data it needs (and in the required format) to perform the simulation. The above mentioned databases were not intended for this purpose.
The number of algorithms available to assess and rationalize protein-ligand interactions has increased during the last years [17, 18], but the number of test cases used to validate methods keeps being small (from tens to a few hundreds of complexes) and problem dependent . A few datasets have been developed for testing the performance of some of the most popular docking programs. One of them is the GOLD validation test set , comprising 100 different complexes extracted from the Protein Data Bank. FlexX-200  is also a dataset consisting of 200 protein-ligand complexes selected and modified by hand from original PDB files. The developers of the DOCK program  also compiled a small dataset of 49 small molecules in complex with a variety of macromolecular targets . In these datasets, the structural and chemical information is stored in the right format used by the corresponding programs.
In this work, we report the compilation of a large set of protein-ligand complexes extracted from the Protein Data Bank, which is tailored both for extracting knowledge on these interactions via data mining and for large scale testing of docking algorithms. The term "large scale" means the use of thousands of protein-ligand complexes that were identified in the PDB and processed in an automatic manner. A key element in the extraction procedure was the use of our own definition of molecular contact to identify protein-ligand interactions at the atomic level. The database of protein-ligand complexes was further enriched in molecular properties such as atom types, atomic charges, etc, needed for the scoring functions of many of the docking programs.
In the final part of this paper, we demonstrate how our new database can be a valuable resource in the process of developing knowledge-based docking algorithms. We carried out a research exercise that started with data mining of protein-ligand interatomic interactions, leading then to a modification of the matching rules of the chemical filter used by the program DOCK to speed up the generation of ligand orientations in the binding site, and finished with testing the new rules and comparing them with the old ones by running docking simulations for several hundreds of protein-ligand complexes.
Identifying small ligands in complex with proteins in the PDB
Determining whether a PDB entry contains a non-covalent complex of a protein with one or several small molecules (ligands) may be a complicated task for several reasons. First, there is not explicit information in the PDB entries that would make clear which records contain atomic data belonging to each small molecule, if present in the structure. Usually, small ligands are listed within HETATM records, but this doesn't apply to amino acids or nucleic acids, which may also constitute small ligands. Furthermore, some small molecules are a mixture of amino acids and/or nucleic acids with other chemical groups, being listed both under HETATM and ATOM records. On the other hand, some of the small molecules present in the PDB, especially many sugars, are covalently bound to proteins, being therefore not of interest for the purpose of this work. The solution given here to these problems was to analyse the atomic connectivity within a PDB entry.
We wrote our own program (called 'complex_info') to process the whole PDB. For peptides and nucleotide chains, we set a limit of eight and four residues, respectively, to consider them as ligands. First, all the amino acid and/or nucleotide chains present in an entry were identified. Entries containing nucleotide chains having more than four residues were not further analysed. Water molecules and metal ions were stored apart. Hydrogen atoms, if present, were deleted (and constructed later for ligands, see below).
The remaining heteroatoms, as well as the small peptides and nucleotide chains were all processed together and grouped by their connectivity into fragments (molecules). Two atoms were considered to be covalently bonded if the distance between them was less than 2.1 Å. This number was selected to cover the larger covalent distances we found in the PDB (between sulphur atoms: 2 – 2.05 Å). Fragments having less than 10 atoms were removed, as well as those fragments that were covalently bonded to a protein chain.
Ligands were classified in seven categories according to their chemical nature: peptide, nucleotide, hetero, peptide-nucleotide, peptide-hetero, nucleotide-hetero and peptide-nucleotide-hetero. Ligand residues listed under the ATOM keyword in the PDB files were identified as peptide or nucleotide by checking their residue names. Every residue listed under the HETATM keyword was included in the broad category of 'hetero'.
Computing protein-ligand contacts
Atomic contacts between two molecules are typically determined based on a distance criterion. Thus, in several reports a contact between two atoms is registered if the distance between them is below a fixed cut-off value. For example, in Relibase [10, 11], a value of 7 Å was used. In a previous work, we used a cut-off of 4.3 Å to collect protein-small ligand interactions . In a stricter approach, a contact is postulated if the distance between the two atoms is not larger than the sum of their van der Waals radii, plus certain tolerance value (generally in the range of 0.5 – 1.0 Å) [7, 10, 11, 24].
Our definition of atomic contact takes also into account a distance criterion, but with an additional refinement. In our approach, the Connolly's molecular surfaces  of the two interacting molecules are employed to determine whether two atoms are in direct contact. The surfaces were calculated using a dot density of 4 Å-2 and a radius of 1.4 Å for the rolling probe. The van der Waals atomic radii were taken (with just a slight rounding) from the AMBER united-atom force field .
So, two atoms belonging to different molecules are said to be in contact if: a) the distance between them is not greater than the sum of their van der Waals radii plus a tolerance distance (0.5 Å), and b) if both atoms have associated surface points located around a line joining their centres, and covering at least an area of 1.5 Å2 for each of the two atoms (see Figure 1B).
Following this definition, the program complex_info computes, for each protein-ligand complex, a list of atomic contacts, as well as a more concise list of the protein residues that make contact with the ligand.
Interface water molecules and ions. Buried surfaces
The program complex_info also determines which of the water molecules, if present in a PDB entry, are located at a protein-ligand interface. In addition, it gives the total surface area of the ligand and the surface areas buried on the ligand and on the protein as a result of the interaction. This analysis was based on the calculated Connolly's surfaces, since his program MS DOT  specifies in the output file which of the surface dots (and their associated areas) are buried by other atoms surrounding the molecule of interest.
For each water oxygen that is close to the ligand (distance < 3.5 Å), its Connolly surface was computed taking into account all the surrounding atoms within a cut-off distance of 8 Å. Water molecules having > 90% of their surface buried by ligand and protein atoms were considered as interface water. The same type of analysis was performed for any metal ion present in the structure.
The calculation of the ligand surface buried by the protein was performed including the interface water molecules and ions as if they were part of the protein. Conversely, the interface water and ions were considered as extensions of the ligand surface to compute the area of the protein surface buried by the ligand.
Typing protein and ligand atoms
When studying molecular interfaces it is important to consider the nature of the atomic interactions, for example, the presence of hydrogen bonds, salt bridges or hydrophobic contacts. For this purpose, it is necessary to know what chemical properties the atoms at the molecular interface have. This information, however, is not included in a PDB file. In particular, the absence of hydrogen atoms in most PDB structures makes it difficult in many cases to determine whether a polar atom can be a donor or an acceptor of a hydrogen bond. In order to overcome these problems, we implemented a procedure for automatic atom type assignment.
Atom types defined in this work and their correspondence with MOL2 types
Aliphatic carbon – sp3 hybridized
C.1, C.2, C.3
Aromatic carbon – sp2 hybridized, included in aromatic rings and in guanidine or carboxyl groups
C.ar, C.cat, C.2
Positive nitrogen – sp 2 or sp 3 hybridized, protonated, forming a group with a positive charge (e.g. guanidine group)
Nitrogen donor of a hydrogen bond – sp 2 hybridized, protonated (e.g. amine group)
N.2, N.ar, N.am
Nitrogen acceptor of hydrogen bond – not all valences covered by heavy atoms, not protonated
N.2, N.ar, N.3, N.1
Aromatic nitrogen – sp 2 hybridized, with all its valences covered by heavy atoms
"Hydroxyl" nitrogen in histidines – sp 2 hybridized, undefined protonation
Oxygen acceptor of hydrogen bond – not protonated
Negative oxygen – negatively charged (e.g. nitro group, carboxyl group)
Hydroxyl oxygen – protonated (OH group)
Protein atoms were classified according to seven categories: donor, positive, acceptor, negative, hydroxyl, hydrophobic and aromatic, without indicating explicitly the chemical element. This classification matches the definition of chemical types for the "attached point" representation of protein binding sites, as described in , designed to be used with the program DOCK. The assignment of atom types for the protein receptors was performed simply from the standard PDB atom names for amino acids. Aspartic and glutamic acids were considered as not protonated, and therefore their carboxylic oxygens were classified as negative. The delta and epsilon nitrogens of histidine were classified as "hydroxyl", as a way to leave undefined their protonation state.
For ligand molecules, atom typing was carried out using the well-known program Babel-1.6 , with a few modifications. One of the changes made in the program had the purpose of translating its internal atom types into the newly defined types. Other changes were introduced to correct some errors that we obtained when processing a group of test molecules, as explained below.
Babel-1.6 uses bond distances, bond angles and torsion angles to assign atom types, using methods similar to those described in [29, 30]. After determining the chemical element and hybridization corresponding to each atom, hydrogens are placed on atoms to fill up the empty valences. The algorithms implemented in Babel work fine when the geometry of the molecule does not deviate much from the ideal values. The small molecules found in the PDB, however, often show large deviations from optimal bond distances and angles.
Statistics from the Babel 1.6 runs on the DOCK 4.0 dataset to test the protonation
Number of atoms
The assignment of 'donor' and 'acceptor' atom types performed by Babel 1.6 for some nitrogen atoms, which depends on their protonation, was not always correct. This kind of problem appeared mainly with heterocyclic structures. To improve the capabilities of Babel to protonate correctly these structures, we included some empirical rules generally accepted in organic chemistry .
Once a molecule is protonated Babel 1.6 can also assign partial charges to atoms, using the Gasteiger algorithm . The information on atom types, hybridization and partial charges was stored in a separate file for each ligand.
Testing the atom typing with reference datasets
Two reference datasets were used to test the typing of ligand atoms with the modified Babel-1.6 program: the DOCK 4.0 dataset , consisting of 49 small molecules randomly selected from the Current Medicinal Chemicals (CMC) molecular database (MDL Information Systems, San Leandro, CA), which was used for testing the implementation of the empirical rules to protonate nitrogen atoms in heterocyclic structures, and the FlexX-200 validation dataset , which consists of 200 protein-ligand complexes selected from the PDB. For each complex in the FlexX-200 dataset, an active site file for the proteins as well as two MOL2 files for the ligand (one file with the crystal geometry and another file with a protonated, energy minimized structure) are provided. The FlexX-200 dataset was used to check the performance of the empirical rules introduced in Babel to protonate nitrogen atoms and the final refinement of the protonation against the structures of the protein-ligand complexes. To perform these tests, hydrogen atoms were removed from the original files, which were converted from the MOL2 to the PDB format. The obtained PDB files were then used as input to the Babel program.
Filtering to create a "non-redundant" dataset
A filtering procedure was applied to the collected protein-ligand complexes aiming to create a non-redundant dataset for data mining. The term "non-redundant", for this particular work, means that the dataset should not contain entries for which both the structures of the ligand and the protein-binding site are similar. Following this principle, the comparison of the complexes was focused mainly on the protein amino acids and the ligand atoms that are in contact with each other. Thus, we compared the ligand structures and the protein-ligand contact tables. The comparison took into account the atom types previously assigned.
First, the filtering procedure was performed within each PDB entry. Afterwards, all pairs of protein-ligand complexes obtained after the first round were compared against each other. If both the ligand structures and the contact tables for two complexes were similar, the complex with lower resolution was discarded.
Comparison of the ligand structures was centred on the atoms contacting the protein. From the list of protein-ligand contacts computed by our contact definition, a graph representation of the interaction is built to find the molecular similarity between two structures. The maximum common sub-graph (MCS) between two structures is used to express molecular similarity. From the MCS, an analysis of the protein-binding site is carried out to eliminate redundancies. Since finding the MCS is a NP-complete graph-matching problem, it is computationally expensive to locate an optimal mapping in a realistic time frame without the application of heuristics. Several approaches have been reported to tackle this kind of problem. Here, our heuristic to find MCS starts with a clique detection algorithm, as described in , which constrains the match using our previous atom type definition and successively adds small subgraphs of both ligands. The similarity of the ligands in terms of three-dimensional structure is also taken into account by checking the root mean square deviation (RMSD) of the MCS for every newly included subgraph.
Data mining study on interactions between atom types
The assignment of atom types carried out in this work, together with all the structural information stored in the database, allows performing different types of studies regarding the chemical nature of the protein-ligand interactions. The extracted knowledge may have important applications, for example, in the design of docking algorithms. In this paper, we explored the data mining possibilities given by our new database of protein-ligand complexes by compiling statistics about the frequencies of interaction between different protein and ligand atom types, and using the obtained data to tune the chemical matching filter implemented in the well known program DOCK .
In this study we introduced the term "relative contact frequency" (FCrij) for different pairs of ligand-protein atom types, e.g. N_d-acceptor, which was defined as follows:
FCrij = k Cij/(pi pj), where Cij is the number of contacts counted for ligand atom type "i" and protein atom type "j", divided by the total number of computed ligand-protein contacts; pi and pj are the proportions in which atom types i and j, respectively, are found in the database; and k is a normalization constant, computed so that the sum of the frequencies by all protein-ligand pairs yields one.
Simulations using the program DOCK with a few modifications
To test the performance of different versions of the chemical filter implemented in the program DOCK, we ran a series of simulations with this program for a selected set of hundreds of protein-ligand complexes.
The algorithm implemented in the program DOCK uses a representation of the binding site based on points ("spheres"), which are intended to occupy positions favourable for ligand atoms to be placed there. Then, sets of ligand atoms are matched onto sets of spheres to generate different orientations of the ligand . A chemical filter may be used in this process to speed up the calculations, by discarding matches that do not fulfil a defined set of chemical complementarity, or matching rules . In order to use this chemical filter, chemical labels (colours) must be assigned both to site points and ligand atoms.
Modifying the chemical filter that comes by default with the program DOCK was easily done by just replacing the original input files containing atom type definitions and matching rules.
To create a negative image of the active site, we used our own method for generating binding site points, called "attached points" (ATPTS) . In this approach, the site points are generated in an automated way from templates constructed for each amino acid type, and chemical labels are assigned to these points also in an automatic fashion. Each ATPTS point is linked to a protein atom (its "parent atom") and its label or colour carries chemical information on this atom in its particular chemical context. Furthermore, each ATPTS point is located at a distance from its parent atom which is optimal for an intermolecular atomic interaction.
ATPTS binding site representations were generated in a totally automated way, following the algorithm described in . The selection of the protein residues that form the binding site was made on the basis of their proximity to the ligand in the crystal complex: every protein residue within a distance of 5 Å from at least one of the ligand atoms was included. Bump distances of 2.2 and 2.8 Å were used for polar and nonpolar protein atoms, respectively. A merging distance of 1 Å was used to fuse points of the same colour that were too close to each other.
Since we were interested only in the functioning of the orientation part of the program DOCK with different versions of the chemical filter, we did not perform any scoring of the generated ligand geometries. Instead, we ran "truncated" docking simulations – the code of the program was slightly modified to output to a file the RMSD value, with respect to the crystal structure geometry, of every generated ligand orientation, and then skip the scoring routines. The docking simulations were carried out also in automated way, using simple scripts.
Results and discussion
We processed the whole PDB release of August 2005 (32069 files) using our program 'complex_info' in order to detect ligand molecules bound to proteins, extract a variety of information on these protein-ligand complexes and calculate several atomic and molecular properties. As result, 22513 complexes from 10694 files were extracted and analysed.
Typing of ligand atoms
One important part of the calculations performed on the extracted protein-ligand complexes was the assignment of atom types to both the protein and the ligand. We defined a simple set of atom types, as described in the Methods section (Table 1), aiming mainly at classifying the molecular interactions in a few elementary types such as hydrophobic contacts and hydrogen bonding. We decided to employ the popular program Babel-1.6 to perform the typing of ligand atoms. One important reason behind this decision was the possibility of modifying the code to implement our new atom type definitions, and making other changes as necessary.
Typing of protein atoms was easily done from the PDB atom names. Typing of ligand atoms, however, presented some difficulties due mainly to the imperfect stereo chemistry of ligand molecules extracted from the PDB and imprecisions in the protonation of nitrogen atoms.
Errors in atom typing/protonation process with FlexX-200 dataset, for the original and modified versions of the Babel 1.6 program
Number of atoms
The final checking of the protonation performed against the crystal structures of the protein-ligand complexes allowed to correct some of the remaining errors in the protonation of ligand nitrogen atoms, but only for those atoms in contact with the protein receptors. Therefore, some errors still remained for nitrogen atoms that are outside the protein-ligand contact tables. Similar problems have been reported for other atom-typing programs ).
Selecting a non-redundant set of protein-ligand complexes
Once the atom typing was performed for all the extracted complexes, a filtering procedure was carried out with the purpose of creating a non-redundant dataset, following the principles delineated in the Methods section. First, obvious redundancies such as repeated complexes at different resolutions and repeated copies of the same ligand within an entry were eliminated. In most cases, these copies are found for proteins having several identical binding sites, for instance, the cholera toxin B pentamer (entry 3chb), or for entries containing copies of the same complex in the asymmetric unit of the crystal. In either case, the protein-ligand contact patterns displayed by the different ligand copies are almost identical.
On the other hand, there are several small groups of entries containing the same protein, but in complex with different ligands, or mutants of the same protein in complex with the same or different ligands. We did not eliminate any of these complexes a priori because each of them might contribute different protein-ligand contacts. For instance, different versions of the same protein very often contain mutations in the binding site region, which most likely results in distinct new contacts with the ligand.
After a first filtering round where copies of the same ligand within the same PDB entry were removed, 12360 complexes out of 22513 remained. Afterward, a second filtering round was carried out where all complexes were compared against each other. As result, 6586 complexes were obtained.
Some statistics extracted from the collected protein-ligand complexes
As shown in Figure 5B, in most of the complexes of the database the ligand has more than 60% of its surface in contact (buried) with the protein. Fifteen protein residues as average are contacting ligand atoms (Figure 5C), with a contact area smaller than 1000 Å2 for all complexes. The buried ligand surface is always slightly smaller than the corresponding buried protein surface because the ligand is embedded within the protein site and because of the presence, in many cases, of water molecules at the interface. Figure 5D shows the important role that water molecules play in many cases, serving as bridges in protein-ligand interactions. About 5500 complexes out from the selected 6586 contain interface waters, with several hundreds of these complexes containing five or more. There are a few complexes with more than 20 water molecules at the protein-ligand interface.
Some of these properties are correlated, for example, Figure 6A shows that the higher the ligand size the higher the number of ligand atoms making contact with the protein. However, there is no correlation between the fraction of ligand surface buried by the protein and the ligand size, as shown in Figure 6B. Most of the smallest ligands are almost or completely embedded in the protein, but there are also many of them that make contact with only a fraction of their surfaces. For larger ligands (> 50 heavy atoms), in most cases a big part of the molecule is outside the binding site. Figure 6C shows that there is no correlation between the number of water molecules at the protein-ligand interface and the buried ligand area.
Number of ligand atoms and number of contacts with protein amino acids by ligand atom type, extracted from the non-redundant set selected from the database
Number of atoms (Na)
Data mining in the database: Application in docking
We further performed a series of data mining and docking simulations, in order to show the potentialities of our database in terms of extracting useful data that could be incorporated into knowledge-based docking algorithms, and testing these algorithms using a large number of protein-ligand complexes. For this purpose we used two different non-redundant data sets compiled from the database. One of them, called PDB2003 and containing 4243 complexes, was obtained from the complexes deposited in the PDB before March 2003, using the filtering procedure explained in the Methods section, while the second dataset, called PDB2005 and having 2931 complexes, was obtained from the rest of the structures, up to August 2005. PDB2003 was used to derive statistics on protein-ligand interactions by atom type, whereas PDB2005 was employed in testing newly designed chemical matching rules, as described below.
Protein-ligand interactions by atom type
Number of contacts in the database between the defined ligand atom types and protein atom types, as extracted from the PDB2003 dataset
Ligand atom type
Number of contacts
Number of contacts with protein atom types *
Probabilities of interaction between the ligand atom types and protein atom types
Ligand atom types
Protein atom types
After the performed normalization, it turned out that atomic interaction pairs formed by a positively charged nitrogen in the ligand and a negatively charged atom in the protein (an oxygen atom from the carboxyl group of asparagine or glutamine) have the highest relative frequency in protein-ligand complexes. In general, positively charged and hydrogen bond donor nitrogen atoms in ligands have high relative frequencies of interaction with hydrogen bond acceptors and negatively charged atoms in the protein. On the other hand, negatively charged oxygens in ligands interact with high relative frequencies with positively charged nitrogens and hydroxyl oxygens. Ligand hydroxyl oxygens are more prompt to donate a hydrogen bond that accepting one from the protein.
In summary, interactions between charged and polar atoms have much higher relative frequencies than interactions with or between hydrophobic or aromatic atoms. This means that when present both in the ligand molecule and in the protein binding site, nitrogens and oxygens have a much higher probability of being involved in interactions with the protein than individual carbon atoms.
Classification of the interactions based on interaction probabilities
Evaluating chemical complementarity in protein-ligand complexes
With the aim of making a practical use of the obtained probabilities of interaction by atom types, we designed a set of chemical complementarity rules that could be used as a filter to discard wrong ligand orientations by a docking program, such as DOCK .
As explained in Methods, DOCK uses a representation of the binding site based on points. Then sets of ligand atoms are matched onto sets of points to generate different orientations of the ligand. A chemical filter implemented in the program helps discarding matches that do not fulfil a defined set of chemical (or matching) rules, thus speeding up the calculations.
Here we used our own method to generate binding site points, called "attached points" (ATPTS), as briefly explained in Methods and in more details in ref. . Each ATPTS point carries a label that reflects the chemical nature of its parent protein atom. Not as a casual coincidence, these labels correspond to the atom types that were defined in this work for protein atoms. Since a spatial matching of an ATPTS point with a ligand atom corresponds to a contacting interaction between its parent atom and this ligand atom, we can extrapolate the statistics obtained for ligand-protein contacts to ligand-ATPTS matches.
Thus, ATPTS binding site representations were generated for the 4243 protein-ligand complexes included in the PDB2003 dataset, in a totally automated way. The calculation time per structure was, on average, less than 0.1 s.
The quality of the ATPTS representation of protein binding sites, in terms of its success in matching point and ligand atom positions for the correct (crystallographic) orientation of ligand molecules, was thoroughly demonstrated in a previous paper . For this purpose, we defined and calculated a magnitude called "number of matching atom-point pairs" (Nmp). A matching atom-point pair is formed by a site point and a ligand atom being within certain "matching cut-off distance" (1.5 Å) from each other. Then we showed that for the vast majority of the tested complexes, more than 70% of the ligand atoms that are in contact with the protein are matched by ATPTS points. Here we repeated the same type of calculation for the PDB2003 dataset and obtained roughly the same result (not shown).
Classification of the different types of protein-ligand atom contacts in three categories: less probable (0), probable (1) and very probable (2)
Ligand atom types
Protein atom types
Defining and testing new matching rules for DOCKing simulations
If the number of matching pairs with probable and very probable interactions is high enough, the stringency of the matching rules to be used in the docking runs may be increased, for example, by enabling only matching of very probable pairs in the ligand orientation process. This would significantly accelerate the calculations by reducing the number of ligand orientations to generate and evaluate. On the other hand, the use of these rules must allow the generation of a sufficient number of correct orientations, otherwise the program may fail in predicting a right solution.
Here we defined two different sets of matching rules: one that permits very probable and probable interactions (R1), and a second set that permits only very probable matches (R2). Then we tested their performance as compared with the chemical complementarity rules coming with the DOCK program (RDOCK), which are much more permissive.
The test of the three sets of matching rules was carried out for a selected set of complexes from the PDB2005 dataset. To rule out other variables that may affect the results of a docking simulation, such as big ligand size and a small contact area with the protein, which are not the subject of this particular study, we restricted the test set to complexes in which ligands have between twenty and forty heavy atoms and more than 70% of their molecular surface embedded in the protein binding site. As we were interested in evaluating the accuracy of the postulated matching rules, the chosen complexes also should have atom types included in the rules sets R1 and R2. Finally, 542 complexes were selected from the PDB2005 dataset, and ATPTS binding site representations were generated for all of them.
For all the selected complexes we ran "truncated" docking simulations, as explained in Methods. No scoring was carried out since we were not interested in the performance of the scoring functions, but only in evaluating the number of correct ligand orientations that are produced by the program. The maximum number of orientations to be generated was limited to 105. This number was high enough to explore all the combinations generated by the matching algorithm for most of the complexes, and was much higher than the value used by default (103). Ligands were kept rigid in these simulations.
In order to check the efficiency of the chemical rules, we computed the number of "good" ligand-receptor configurations generated by the DOCK program in the ligand orientation process. As "good orientations" were considered those whose RMSD value with respect to the crystal geometry was less than 2 Å.
When using its default set of chemical matching rules, the program DOCK fails in producing a good solution for about 30% of the complexes (Figure 8A), even though the program reached, for almost all the complexes, the maximum number of ligand orientations that was allowed (Figure 8D). For the 70% of complexes showing a successful result, an average number of 2985 good solutions was produced by the program.
Using the R1 set of matching rules significantly improves the performance of the simulations, with only a very few complexes having their best orientation with an RMSD > 2 Å (Figure 8B). For this set of rules the number of possible ligand atom-site point matches was still high enough as to make the program produce the allowed 105 orientations for most of the complexes (Figure 8E). The average number of produced good solutions was 4062.
Finally, when the R2 rules were employed, a good solution was always found (Figure 8C). For this more restrictive set of rules, the number of explored orientations was reduced notably, as shown in Figure 8F, and nevertheless, a high number of good orientations was obtained (2138 as average) for each protein-ligand complex. These results demonstrate that the matching rules used to limit the exploration of the conformational space between the protein and the ligand can be more stringent and still produce very good results if the ligand molecules have atom types with statistically high probabilities of interaction with proteins.
On the other hand, the fact that using the matching rules given by default for the DOCK runs made the program fail in about 30% of the tested cases causes concern. Since the ensemble of ligand atom-site point matches produced with the R1 or R2 rules must be a subset of the ensemble of matches produced using more permissive rules, or not using any filter at all, the reason for the failure must be that the maximum allowed number of ligand orientations was not enough to give the program the chance of exploring the right sector of the orientational space. Even for small and medium size ligands, the total number of atom-point combinations that can be produced may be so large, that the probability of obtaining correct matches within the first 105, or even more unfavourably, within the first 103 explored conformations, can be small.
We compiled a large database of small ligand-protein complexes derived from the PDB that currently contains more than 6000 non-redundant structures. In processing the PDB structures, we characterized in details the protein-ligand interactions, using a new definition based on the proximity of the molecular surfaces. Furthermore, the structural data was enriched with chemical properties that we calculated using our own as well as existing algorithms. As an example to demonstrate the value of the new database for data mining and testing of docking algorithms, we derived a new set of chemical matching rules to be used in the context of the program DOCK, based on contact frequencies between ligand atoms and points representing the protein surface, and proved their enhanced efficiency with respect to the default set of rules included in the program. One important finding in this study is the high rate of failure of the program DOCK in finding orientations close to the crystal structure when using the default chemical filter, even in the case when the maximum number of allowed ligand orientations is set 100 times larger than the default value.
- Berman HM, Westbrook J, Feng Z, Gilliland G, Bhat TN, Weissig H, Shindyalov IN, Bourne PE: The Protein Data Bank. Nucleic Acids Res 2000, 28: 235–242. 10.1093/nar/28.1.235PubMed CentralView ArticlePubMedGoogle Scholar
- Cambridge Structural Database (CSD)[http://www.ccdc.cam.ac.uk/products/csd]
- Enhanced NCI Database Browser[http://cactus.nci.nih.gov/ncidb2]
- ZINC – A free database for virtual screening[http://zinc.docking.org]
- ChemStar, Ltd[http://www.chemstar.ru]
- Shin JM, Cho DH: PDB-Ligand: a ligand database based on PDB for the automated and customized classification of ligand-binding structures. Nucleic Acids Res 2005, 33: D238-D241. 10.1093/nar/gki059PubMed CentralView ArticlePubMedGoogle Scholar
- Puvanendrampillai D, Mitchell JB: L/D Protein Ligand Database (PLD): additional understanding of the nature and specificity of protein-ligand complexes. Bioinformatics 2003, 19: 1856–1857. 10.1093/bioinformatics/btg243View ArticlePubMedGoogle Scholar
- Laskowski RA: PDBsum: summaries and analyses of PDB structures. Nucleic Acids Resh 2001, 29: 221–222. 10.1093/nar/29.1.221View ArticleGoogle Scholar
- Kleywegt G, Jones T: Databases in protein crystallography. Acta Crystallogr D Biol Crystallogr 1998, 54: 1119–1131. 10.1107/S0907444998007100View ArticlePubMedGoogle Scholar
- Hendlich M, Bergner A, Günther J, Klebe G: Relibase: design and development of a database for comprehensive analysis of protein-ligand interactions. J Mol Biol 2003, 326: 607–620. 10.1016/S0022-2836(02)01408-0View ArticlePubMedGoogle Scholar
- Günther J, Bergner A, Hendlich M, Klebe G: Utilising structural knowledge in drug design strategies: applications using Relibase. J Mol Biol 2003, 326: 621–636. 10.1016/S0022-2836(02)01409-2View ArticlePubMedGoogle Scholar
- Smith RD, Hu L, Falkner JA, Benson ML, Nerothin JP, Carlson HA: Exploring protein-ligand recognition with Binding MOAD. J Mol Graph Model 2006, 24: 414–425. 10.1016/j.jmgm.2005.08.002View ArticlePubMedGoogle Scholar
- Chen YZ, Zhi DG: Ligand-protein inverse docking and its potential use in the computer search of protein targets of a small molecule. Proteins 2001, 43: 217–226. 10.1002/1097-0134(20010501)43:2<217::AID-PROT1032>3.0.CO;2-GView ArticlePubMedGoogle Scholar
- Gold ND, Jackson RM: SitesBase: a database for structure-based protein-ligand binding site comparisons. Nucl Acids Res 2006, 34: D231–234. 10.1093/nar/gkj062PubMed CentralView ArticlePubMedGoogle Scholar
- Powers R, Copeland JC, Germer K, Mercier KA, Ramanathan V, Revesz P: Comparison of protein active site structures for functional annotation of proteins and drug design. Proteins 2006, 65: 124–135. 10.1002/prot.21092View ArticlePubMedGoogle Scholar
- Kellenberger E, Muller P, Schalon C, Bret G, Foata N, Rognan D: sc-PDB: an annotated database of druggable binding sites from the Protein Data Bank. J Chem Inf Model 2006, 46: 717–727. 10.1021/ci050372xView ArticlePubMedGoogle Scholar
- Sousa SF, Fernandes PA, Ramos MJ: Protein-ligand docking: current status and future challenges. Proteins 2006, 65: 15–26. 10.1002/prot.21082View ArticlePubMedGoogle Scholar
- Leach AR, Shoichet BK, Peishoff CE: Prediction of protein-ligand interactions. Docking and scoring: successes and gaps. J Med Chem 2006, 49: 5851–5855. 10.1021/jm060999mView ArticlePubMedGoogle Scholar
- Cole JC, Murray CW, Nissink JW, Taylor RD, Taylor R: Comparing protein-ligand docking programs is difficult. Proteins 2005, 60: 325–332. 10.1002/prot.20497View ArticlePubMedGoogle Scholar
- Jones G, Willett P, Glen RC, Leach AR, Taylor R: Development and validation of a genetic algorithm for flexible docking. J Mol Biol 1997, 267: 727–748. 10.1006/jmbi.1996.0897View ArticlePubMedGoogle Scholar
- Kramer B, Rarey M, Lengauer T: Evaluation of the FLEXX incremental construction algorithm for protein-ligand docking. Proteins 1999, 37: 228–241. 10.1002/(SICI)1097-0134(19991101)37:2<228::AID-PROT8>3.0.CO;2-8View ArticlePubMedGoogle Scholar
- Kuntz ID, Blaney JM, Oatley SJ, Langridge R, Ferrin TE: A geometric approach to macromolecule-ligand interactions. J Mol Biol 1982, 161: 269–288. 10.1016/0022-2836(82)90153-XView ArticlePubMedGoogle Scholar
- Ewing TJ, Makino S, Skillman AG, Kuntz ID: DOCK 4.0: search strategies for automated molecular docking of flexible molecule databases. J Comput Aided Mol Des 2001, 15: 411–428. 10.1023/A:1011115820450View ArticlePubMedGoogle Scholar
- Moreno E, Leon K: Geometric and chemical patterns of interaction in protein-ligand complexes and their application in docking. Proteins 2002, 47: 1–13. 10.1002/prot.10026View ArticlePubMedGoogle Scholar
- Connolly ML: Solvent-accessible surfaces of proteins and nucleic acids. Science 1983, 221: 709–713. 10.1126/science.6879170View ArticlePubMedGoogle Scholar
- Weiner SJ, Kollman PA, Case DA, Singh UC, Ghio C, Alagona G, Profeta S Jr, Weiner PA: A new force field for molecular mechanical simulation of nucleic acids and proteins. J Amer Chem Soc 1984, 106: 765–784. 10.1021/ja00315a051View ArticleGoogle Scholar
- Sobolev V, Sorokine A, Prilusky J, Abola EE, Edelman M: Automated analysis of interatomic contacts in proteins. Bioinformatics 1999, 15: 327–332. 10.1093/bioinformatics/15.4.327View ArticlePubMedGoogle Scholar
- Openeye Scientific Software[http://www.eyesopen.com/products/applications/babel.html]
- Meng EC, Lewis RA: Determination of molecular topology and atomic hybridization states from heavy atom coordinates. J Comput Chem 1991, 12: 891–898. 10.1002/jcc.540120716View ArticleGoogle Scholar
- Baber JC, Hodgkin EE: Automatic assignment of chemical connectivity to organic molecules in the Cambridge Structural Database. J Chem Inf Comput Sci 1992, 32: 401–406. 10.1021/ci00009a001View ArticleGoogle Scholar
- Joule JA, Smith GF: Heterocyclic Chemistry. Blackwell Science; 2000.Google Scholar
- Gasteiger J, Marsili M: A new model for calculating atomic charges in molecules. Tetrahedron Lett 1978, 34: 3181–3184. 10.1016/S0040-4039(01)94977-9View ArticleGoogle Scholar
- Gardiner EJ, Artymiuk PJ, Willett P: Clique-detection algorithms for matching three-dimensional molecular structures. J Mol Graph Model 1997, 15: 245–253. 10.1016/S1093-3263(97)00089-2View ArticlePubMedGoogle Scholar
- Shoichet BK, Kuntz ID: Matching chemistry and shape in molecular docking. Protein Eng 1993, 6: 723–732. 10.1093/protein/6.7.723View ArticlePubMedGoogle Scholar
- van Aalten DM, Bywater R, Findlay JB, Hendlich M, Hooft RW, Vriend G: PRODRG, a program for generating molecular topologies and unique molecular descriptors from coordinates of small molecules. J Comput Aided Mol Des 1996, 10: 255–62. 10.1007/BF00355047View ArticlePubMedGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.