WISDOM-II: Screening against multiple targets implicated in malaria using computational grid infrastructures

Due to very high costs associated to the drug discovery process as well as due to late stage attrition rates, novel and cost effective strategies are absolutely needed for combating the neglected diseases, like malaria. Virtual high throughput screening is a technique, which can screen millions of compounds rapidly, reliably and cost effectively on a computer [3, 4].

There are millions of chemical compounds available in the labs and also in 2D, 3D electronic databases due to advances in the combinatorial chemistry, but it is nearly impracticable to synthesize them [5]. Moreover it is labour-intensive and very expensive to screen such a high number of compounds in experimental labs by high throughput screening (HTS). Besides the heavy costs (required for developing efficient and reliable assays) the hit rate in HTS is quite low [5].

In addition to the availability of a huge number of chemical compounds, there is also a significant increase in the number of resolved X-ray crystal structures, most of which are freely available from the Brookhaven protein database [6].

The presence of sound open source electronic data of chemical compounds as well as the data for macromolecules like proteins, enzymes facilitated virtual high throughput screening (vHTS). Virtual high throughput screening (vHTS) by molecular docking serves as a complementary or alternative technique to experimental high throughput screening (HTS). In silico drug design, especially high throughput virtual screening is a widely-used and well-accepted technology in new lead identification and optimization [7, 8].

The downside to vHTS is that screening millions of chemical compounds is computationally intensive: it has a high storage demand and is, therefore, termed as computational data challenge. Screening each compound, depending on structural complexity, can take from one to few minutes on a standard PC, which means screening a database with millions of chemical compounds can take years of computation time. Computational grid infrastructures are the best attempt to solving this problem thus far. The computing resources from EGEE grid infrastructure were utilized extensively [9‐11], besides EGEE [12] several other grid infrastructures such as: AuverGrid [13], EELA [14], EUChinaGrid [15] and EUMedGrid [16] contributed significantly to the current project.

The combination of these two techniques (vHTS and Grid computing) will definitely decrease the financial and cost implications of rational drug design strategies. Several docking applications have already been run on grids and proved to be successful. Some of the success stories in in silico drug design on grid are the smallpox research grid [17], Anthrax research project and Cancer project [18‐20].

WISDOM-I [21], the first large scale deployment of molecular docking application on EGEE, which took place from August 2005 to September 2005 has seen 42 million dockings, which is equivalent to 80 years of CPU. Virtual screening of 500,000 chemical compounds was performed by using FlexX against different plasmepsins (aspartic protease implicated in haemoglobin degradation). On the biological front three scaffolds were identified, one of them is guanidino scaffold, which is likely to be novel as they have not been reported as plasmepsin inhibitors before [22]. Experimental results have proved that the some of the compounds selected from WISDOM-I function as sub-micromolar inhibitors against plasmepsin [23].

The main goals of WISDOM project were to identify inhibitors to be tested in the experimental laboratories and to develop fault tolerant WISDOM production environment [24], which is capable of deploying molecular docking application efficiently on grid infrastructure. Drug targets from malaria are chosen initially, but this could be expanded to determine ligand binding to any target protein.

The different steps of the virtual docking procedure will be described in the following section.

The initial coordinates for all the target structures are obtained from Brookhaven protein database. Depending upon the inclusion of the significant residues, active site is defined as 8.0 – 10Ǻ around the co-crystallized ligands.

The compound library used for WISDOM was obtained from the ZINC database [52, 53]. The ZINC database is a collection of 4.3 million chemical compounds from different vendors. ZINC library has been chosen because it is an open source database, and the data are available in different file formats (Sybyl mol2 format, sdf and smiles). So, basically, ZINC provides virtual compounds ready for virtual screening. A total of 4.3 million compounds were downloaded from the ZINC database and screened against targets mentioned in Table 1.

The docking software used in the current study is FlexX [54, 55]. It is extremely fast, robust and highly configurable computer program for predicting protein-ligand interactions. Standard parameter settings are used except for two cases: "Place particles" [56] and "Maximum overlap volume". These two parameters were subject to deliberate variation with FlexX. Since every target has a unique response to docking software parameters set, there is no generic solution to the parameter sets. Initial optimization experiments were done to arrive at a parameter set which gave best results in redocking experiment for a particular target.

Re-docking can be defined as the removal of the co-crystallized compound (inhibitor or substrate) and then using a specific parameter set to dock this compound back into the active site of its target protein to validate the programs ability to dock novel compounds into the active site. These experiments serve as positive controls before large scale docking since aids in defining the active site and other simulation conditions. The docking pose during these experiments is validated by comparing the pose based on the RMSD between the atoms of the co-crystallized pose and the docking pose, as well as visual inspection of the orientation of the ligand. The lower the RMSD value and the more similar the docking pose to the co-crystallized ligand the better the docking results. Ligand plot information obtained from Brookhaven database serves as a template to validate the docking pose. The ligand plots of all the targets used in the current project are displayed in Figure 1. Ligand plots displays the binding mode of the co-crystallized ligand within the active site of the receptor, besides this it also describe the atom-to-atom interaction between the co-crystallized ligand to its respective receptor. This information is later compared with the docking poses before large-scale screening.

The results of the re-docking experiments are displayed in Table 2. Results are analysed at three levels: the RMSD (root mean square deviation) between the docking pose and the co-crystallized ligand, the docking score and the interaction information between protein and ligand. FlexX has a unique ability to compute the atom-to-atom interaction between the protein and the ligand. This information is exploited and further used in analysing the results. The docking poses of the co-crystallized ligands generated by FlexX are manually visualized and compared to their respective ligand plots. Two aspects were considered; the binding mode of the docking pose should be similar to ligand plot and should make interactions to the key residues of the receptor as described in ligand plots. Ligand plots are displayed in Figure 1. Table 2 displays the docking score and RMSD of the best docking conformation. For target PfGST (1Q4J: PDB ID), parameter set 1 performed better compared to other parameter sets. However 3.68 Å is still a big deviation (ideal RMSD should be <2 Å), but the binding mode the co-crystallized ligand adopted was quite convincing, as the docking pose was making interactions to the key residue. Besides that, the docking pose made interactions to the key residues responsible for the activity of the protein. In case of P. vivax DHFR (2BLC and 2BL9), the docking of the co-crystallized ligand did not perform well. The RMSD deviations were high (>4 Å) and the binding modes were not convincing. This is due to clashes between the protein and ligand atom surfaces. For PfDHFR re-docking is performed against protein structures before and after minimization by Amber software. Docking software parameters were tuned accordingly. For PfDHFR, we increased the maximum allowed overlap between the protein and ligand atom to diminish the van der Waals clashes. Re-docking against minimized structures with the same parameter sets gave best results. All the parameter sets reproduced the actual binding mode of the ligand, further made interactions to key amino acids and RMS deviation were less than 2 Å. Re-docking results of PfDHFR minimized structures are displayed in Table 3 and 4. Besides docking the co-crystallized ligand, well-known inhibitors against PfDHFR are docked. Table 3 and 5 displays the results of cycloguanil and pyrimethamine, WR9 under different docking parameter sets. Parameter 8 (maximum allowed overlap volume between protein and ligand surface: 100 ų) gave the best results in terms of docking score, docking conformation and interactions to key amino acids. The results displayed in Table 4 correspond to the quadruple mutant results (1J3I: PDB ID) and Table 5 corresponds to the wild type results (1J3K: PDB ID). Figure 2 and 3 display the re-docking pose of WR9 against minimized structure of the Pf DHFR (1J3K) and Pf DHFR (1J3I), respectively, on the right hand side of the figure we can see the docking pose (CPK color) and reference co-ordinates in red color (IJ3K) and violet color (1J3I). On the left hand side protein ligand interactions are displayed. Highlighted are the interactions responsible for the activity of the protein (parameter sets 5, 6, 7, 8 correspond to maximum allowed overlap volume 10, 20, 30, 100 ųrespectively).

After setting up the docking platform, virtual screening was performed on 4.3 million compounds against the targets specified in Table 1. Screening 4.3 million compounds on multiple target structures demands huge computation and storage resource power. EGEE and its related grid infrastructures (AuverGrid, EELA, EUChinaGrid and EUMedGrid) provided the resources required for executing the data challenge. Deployment strategies and WISDOM production environment were developed for submission of jobs and retrieval of results on the EGEE grid. More information on deployment and wisdom production environment can be found in Jacq et al [21]. The significant achievement on the grid side is the improved WISDOM production environment, which enabled smooth and successful deployment of docking jobs on the grid. The major difference between our previous deployments (WISDOM-I and the data challenge on avian flu) is the automatic resubmission of the docking jobs on the grid. Besides the enormous gain in the computing resources available, grid infrastructures enable also to store the tera bytes of scientific data and to share this data between the research laboratories located not only in different countries but also on different continents.

The outputs of the docking results in FlexX are log files. All the results are stored and analysed by using MySQL databases. Three different forms of results are saved and analysed from each docking assay:

Moreover, the ranking process is the integral part of the docking software. FlexX have a post processing optimization of the docking solution and clustering. Clustering in FlexX is based on RMSD, angle, and distance deviation (default values are used for the clustering; the necessity of clustering docking poses is discussed in [22]).

During the first deployments (WISDOM-I and data challenges on Avian Flu) the results were stored on the grid storage elements using the grid data management, this format made the analysis of the results particularly difficult, which left some room for improvement.

Since docking and scoring results often need to be extracted and parsed by biologists, user-friendly data retrieval systems need to be put in place. Hence, it was decided to rank the information based on the docking scores and do some initial filtering of the compounds. The relevant information was parsed directly into a relational database. The database was designed around the docking table (Figure 4), where docking scores are stored, which represent the binding free energy. The individual energy contributions to the total free energy of binding are also stored. This was useful for the filtering of compounds and it also helped in the docking results analysis. As shown in Figure 4, the insertion of records is performed directly at the end of the jobs. A simple perl script, using perl DBI library, parses the result file and builds, from the useful information, a query to insert the data from the grid to a remote MySQL server. The raw result files are also stored and replicated on the grid storage elements.

The real interest of such a solution is that the useful data are immediately available for query and analysis during the process. The usage of relational database along with SQL eases the selection of the best compounds as they can be selected accordingly to any attribute of the database tables. As almost all programming languages offer the ability to access database management systems through APIs, it will also ease the interoperability with web servers, for instance, if one wants to be able to monitor and view the data on a web interface.

Result analysis is performed to identify a small number of promising compounds that can be tested further. During WISDOM-I, the compounds with best docking scores were visualized manually and interestingly observed that some of the top scoring compounds were making interactions to the key residues but the binding modes of the compounds were not optimal and not comparable with docking poses of co-crystallized ligands, one of the reason may be due to the rigid nature of the receptor. Consequently, the result analysis took place by first extracting the compounds based on docking scores and then by rescoring the best docked ligands with more sophisticated scoring functions. At this purpose, for this project, an automated refinement and rescoring procedure developed by Rastelli et al [57], which refines ligand-target complexes with molecular mechanics and molecular dynamics, and then calculates the binding free energies according to MM-PBSA and MM-GBSA methods are utilized. Such procedure, called BEAR (Binding Estimation After Refinement) [58] significantly improved the overall procedure and resulted in the identification of plasmepsin inhibitors (WISDOM-I). Hence, the same procedure was used to extract the best 5,000 (against GST), 15,000 (against DHFR) compounds based on docking scoring values. MM-PBSA and MM-GBSA (Molecular mechanics-Poisson Boltzmann surface area and Molecular Mechanics-Generalized Born surface area) procedures calculate the absolute free energies for non-covalent interactions of protein and ligand in solution by using force field based molecular mechanics method [59, 60].

The common filtering process employed in WISDOM project is displayed in Figure 5. After rescoring by molecular dynamics methods, the compounds are further manually visualized by using Chimera software [61] and other structural visualizing software. Interactions to key residues of the receptor and binding mode of the ligand are main criteria for further selecting the compounds to be tested in experimental laboratories.

Docking results of PfGST are represented in Table 5. Six amino acids were considered to be responsible for the activity of the target: Tyr9, Gln58, Val59, Ser72, Gln71 and Lys15. Chemical compounds interacting with these amino acids were of significance and hence computed. All ten top scoring compounds displayed in Table 5 made interactions to these key amino acids. A binary scoring mode was adopted for the residue reactions in Table 5, column 3: "0" represents false (no interaction with the specified amino acid) and "1" represents true (either a hydrogen bond or a hydrophobic interaction, was made). From Table 5, it is clear that all the top scoring compounds are making interactions with at least one of the key amino acids. These observations are later compared to the standard protein ligand interaction information obtained from ligand plots displayed in Figure 1. This particular method not only allowed us to select compounds based on scoring but also based on interaction information (hydrophobic and hydrophilic interactions), which is very significant from the structural point of view for the identification of hits.

To give wide overview on the results obtained by docking, diversity analysis against the PfGST best 5,000 compounds and PfDHFR best 15,000 compounds by docking score was performed by using MOE software [62]. Finger prints of all the compounds were created by using FP: BIT MACCS and then used Tanimoto coefficient (TC) for calculating the diversity among the compounds [63]. At similarity cut off of Tanimoto coefficient 0.7, out of 5,000 compounds of PfGST, 3,394 different clusters were identified by this method, which indicates the best 5,000 compounds diverse and dissimilar. The Tanimoto coefficient (similarity = the number of bits set in both molecules divided by the number of bits set in either molecule) is a validated and most commonly used similarity coefficient in chemical informatics while calculating diversity of the chemical compound database. It ranges from values 0 to 1, while value "1" corresponds to completely similar compound and "0" completely disssimilar). Diversity analysis is performed to demonstrate the best compounds by docking score are diverse enough for further analysis and for the identification of novel scaffolds. Pair wise frequency (Y-axis) and Tanimoto coefficient value (X-axis) are plotted and displayed in Figure 6. The values of mean, median, 1^st quartile, 3^rd quartile of the histogram are 0.44, 0.43, 0.37, 0.50 respectively. The 1^st quartile and 3^rd quartile values signify that 25% of the compounds possess TC values of 0.37 and 75% of the compounds possess TC values of 0.5. For PfDHFR diversity analysis is performed against 15,000 top scoring compounds. Pair wise frequency (Y-axis) and Tanimoto coefficient value (TC) (X-axis) are plotted and displayed in Figure 7. The values of mean, median, 1^st quartile, 3^rd quartile of the histogram are 0.42, 0.40, 0.34, 0.48 respectively. The 1^st quartile and 3^rd quartile values signify that 25% of the compounds possess TC values of 0.34 and 75% of the compounds possess TC values of 0.48. These observations and figures indicate that the top scoring compounds are diverse and have potential to find novel compounds. The frequency on the Y-axis represents pair wise similarity of each compound against all the compounds in the database (5,000 × 5,000 times for PfGST).

To understand the interactions between PfGST and final hits, the ligand plots for each complex (PfGST and the compound) were generated and further visualized manually. Protein ligand interactions are studied in three dimensions and for clarity in displaying they are depicted as 2D interaction diagrams. These interactions presented here are generated using the ligand plot module of MOE software. It is evident from Figure 8 that inhibitors are located in the center of the active site, and are stabilized by hydrogen bonding interactions. The hydrogen bonding information along with their distances is listed in Table 6. Figure 8 displays the binding modes of the five best compounds in the active site of the PfGST_a chain. To allow the comparison of binding mode of the compounds and co-crystallized ligand, ligand plot and interactions information is generated for GTX (Cocrystallized ligand of PfGST). It is obvious from Table 6 and Figure 8 that the compounds listed here possess comparable binding poses and patterns. Especially compounds ZINC03533756, ZINC03830430, ZINC03580546, ZINC02305869 generated interaction patterns very similar to the one observed with GTX; making hydrogen bonding to Val59 and Ser72 with backbone as well as with side chains of the amino acids.