D virtual screening for thrombin ligands

Structure-based three-dimensional VS (3D-VS) is quite often identified solely with the application of docking methods once the structure of the target of interest has been resolved. In many cases, several years are required to obtain the target structure, which limits the applicability of VS from docking to quite late in a pharmaceutical project timeline. The recent development of similarity methods allowing for fast flexible ligand superposition [20,77,78] has moved the applicability of 3D-VS much earlier in the project stages. In most cases, 3D-VS from similarity could be applied now as soon as hits from high-throughput screening are confirmed or become publicly available from other sources. Therefore, taking docking as the method of reference, it is probably appropriate to compare the performance of similarity and docking methods in 3D-VS to extrapolate the validity of applying 3D similarity methods at those early stages. In this contribution, the performance of two particular implementations, MIMIC for similarity and DOCK for docking, will be compared in a virtual screening of 10 000 diverse compounds selected from the ACD database in search for thrombin-like compounds.

The thrombin case has been selected, on the one hand, because of its extensive use as a validation test by other implementations of similarity and docking methods, being thus relevant for comparative purposes. On the other hand, it has been shown recently that, despite having a well-defined active site, it can still be a challenge for docking methods aiming at obtaining small root-mean square deviations in binding mode assessment applications [20,30] and good active-molecule enrichments in molecular database screening applications [20,42]. As structural templates for the 3D-VS, the NAPAP ligand in MIMIC calculations and its complementary thrombin structure in DOCK calculations were extracted from the Protein Data Bank (lDWD, 3.0 A, resolution). Hydrogen atoms were automatically added to the crystal structure of the NAPAP ligand and Gasteiger-Marsili [90] atomic charges generated using Sybyl 6.5. Along with the structure of the NAPAP ligand, the active site residues included in a 10 A, sphere around NAPAP were also considered in DP-MIMIC calculations. In order to reduce bias towards the observed water-assisted ligand-protein interactions in 1 DWD, no crystallographic water molecules were considered in DOCK calculations.

A diverse selection of 10 000 molecules was obtained by clustering the BCI fingerprints [58] of the entire ACD database. For all molecules, three-dimensional structures were generated automatically by CORINA [9 1], and Gasteiger-Marsili [90] atomic charges calculated with Sybyl 6.5 [84]. This set of 10 000 molecules constitutes the virtual library that was screened with MIMIC and DP-MIMIC to identify molecules with similar steric and electrostatic characteristics to the NAPAP ligand and with DOCK to identify molecules with complementary shape and interaction sites to the thrombin active site. The computational cost of screening those 10 000 molecules with MIMIC and DP-MIMIC was about 50 h and with DOCK about 255 h on a single CPU of an SGI/R10000 machine. However, it is important to note that 3D-VS calculations can be fully performed in parallel by distributing different database chunks to the number of CPUs available.

An important issue in any analysis of virtual screening results is how to decide which fraction of the best-scoring molecules will be further considered for a more detailed inspection or directly submitted for a prioritized experimental testing. Commonly, the 5-10% highest-ranking molecules within the database are subjectively chosen. However, a more rational method of selecting the set of potentially most interesting molecules would be desirable. With this in mind, the use of the (normalized) similarity or docking scores is proposed as an objective metric for the (de)selection of molecules from 3D searches [20].

Figure 1. Normalized score obtained from MIMIC (in blue) and DOCK (in red) for the best-scoring 2000 molecules.

Figure 1 depicts the decline of the normalized MIMIC similarity and DOCK energy scores with increasing rank for the highest-ranking 2000 molecules obtained by the two 3D-VS runs. The distribution of scores obtained with DP-MIMIC is practically identical to that plotted in Figure 1 for the original MIMIC search. Thus, although the final rankings of individual molecules may change due to the active site steric-field correction, the overall discrimination between highly ranked molecules and the bulk of the database is maintained in DP-MIMIC with respect to MIMIC. This notwithstanding, it is interesting to observe from Figure 1 that both MIMIC and DOCK methods experience a sharp decrease in normalized scores within the first 200 molecules. After that, the scores experience an almost linear decrease [20]. In particular, the DOCK score provides a faster initial decay than the MIMIC score. This is most likely due to two factors. On one hand, the substantial emphasis put on charged contacts by the DOCK energy score (i.e. in the P1 pocket of thrombin), which is less pronounced in the MIMIC similarity score (based on a 'soft' Gaussian representation of the electrostatic potential). On the other hand, the fact that, by definition, molecules will always overlap to some extent in similarity methods. In any case, the shape of the curves in Figure 1 indicates that the first 100-200 molecules in the ranked list are clearly favoured by MIMIC and DOCK over those molecules constituting the linear plateau that is eventually reached. Therefore, it is important to check always whether the 3D-VS method being used provides a clear discrimination between a reduced number of high-scoring molecules and the rest of the database. If this is indeed the case, by simply setting a minimum for the gradient of the normalized score, an independent and rational metric can be obtained to select molecules from different virtual screening approaches for further analysis or experimental testing [20].

Another point worth analyzing when comparing 3D-VS results from different methods or implementations of the same method is their degree of consensus or redundancy within the highest-scoring molecules. As recently pointed out [1], one way to get around the problem of discriminating a set of potentially interesting molecules is to perform consensus scoring by selecting only compounds that obtain a certain threshold score in a number of different scoring functions. In the case presented here, however, consensus will be demanded between different 3D-VS methodologies rather than between protein-ligand scoring functions [20]. By following this strategy it is expected that the likelihood of selecting potentially interesting compounds will be higher.

Figure 2 provides a chemical diversity analysis of the molecules favoured by both MIMIC and DOCK. For each of the 10 000 molecules of the database, 45 descriptors were computed, which were previously selected by cluster significance analysis of a larger number of descriptors. Subsequently, a principal component analysis was done. The axes in Figure 2 represent the first two principal components (explaining 45.8% and 12.6% of the total variance), the values of which have been scaled in order to produce a uniform density of molecules along the two axes. A more detailed description of the diversity analysis performed here can be found elsewhere [55].

There are two aspects worth remarking from Figure 2. The first one is that the molecules that have been ranked by both MIMIC and DOCK within the top 1000 scores (in red) cover almost the same diversity space as that described by the entire set of 10 000 diverse compounds from the ACD (in gray). Although one may have expected molecules selected on the basis of 3D similarity or docking to be limited to a confined region of chemical space, a diverse range of structures is retrieved by both methods. The second one is that some of the consensus molecules overlap well with the property-space regions populated by known thrombin inhibitors (in blue). In fact, the consensus compounds cover more diversity space than the selection of known thrombin inhibitors. However, it is not possible to state a priori that this is due to a true diversification in active scaffolds for thrombin or to a lack of specificity of the scoring functions used by MIMIC and DOCK. Selection of these compounds for further assay testing should give the answer a posteriori.

At this point, it is interesting to address the question of whether combining 3D similarity and protein structure-based approaches could improve the retrieval of potential hits from chemical databases with respect to the original similarity and docking implementations. Our recent experiences with includ-

Figure 2. First two principal components of the diversity analysis of the 470 consensus molecules between MIMIC and DOCK within the best-scoring 1000 molecules (in red) with respect to the entire set of 10 000 ACD compounds (in gray) and 32 known thrombin inhibitors (in blue). The descriptor space has been scaled to produce a uniform density of compounds along the two axes.

Figure 2. First two principal components of the diversity analysis of the 470 consensus molecules between MIMIC and DOCK within the best-scoring 1000 molecules (in red) with respect to the entire set of 10 000 ACD compounds (in gray) and 32 known thrombin inhibitors (in blue). The descriptor space has been scaled to produce a uniform density of compounds along the two axes.

ing a ligand-ligand similarity term in the context of flexible protein-ligand docking have shown promising results [20]. Here, we report the novel use of introducing information on the active site environment in 3D similarity searches, as implemented in the program MIMIC (vide supra). The percentages of consensus molecules between MIMICLIP-MIMIC, MIMIC/DOCK, DP-MIMIC/DOCK and MIMIC/DP-MIMIC/DOCK within the 1000 best-scoring molecules (10% of the database) are 66.9, 47.0, 43.5 and 36.0%, respectively. Interestingly, MIMIC and DP-MIMIC show less than 67% of consensus molecules. As a result of the inclusion of an active site bump penalty in the similarity scoring, a different ranking of molecules is clearly obtained from MIMIC and DP-MIMIC. Unfortunately, this is not translated in a higher number of consensus molecules between DP-MIMIC and DOCK

than between MIMIC and DOCK, probably due to the fact that the active site penalty is incorporated only after, and not during, the similarity optimization. It is, though, gratifying to find out that two conceptually different methods, such as MIMIC and DOCK, still reach consensus on almost half (47%) of the 1000 top-scoring molecules and that more than one third (36%) of the 1000 top-scoring molecules are independently identified by MIMIC, DP-MIMIC, and DOCK.

A final aspect of the comparison between similarity and docking approaches to 3D-VS concerns the binding modes proposed by each method. Despite being computationally demanding, one of the attractive characteristics of structure-based methods is their ability to provide a 3D alignment (in similarity methods) or orientation into the active site (in docking methods), from which a ligand optimization program can be more intuitively guided. It would be, thus, interesting to analyze the degree of resemblance between the binding modes proposed from similarity and docking for some of the consensus molecules identified during the 3D-VS.

A detailed comparison of the performance of MIMIC and DOCK for assessing the binding mode of a set of 32 known thrombin inhibitors was recently performed [20]. For the sake of completeness, the effect of introducing a ligand-ligand similarity penalty term into DOCK (SP-DOCK) was also investigated [20]. Here, emphasis will be put in analyzing the effect of introducing a protein-ligand docking penalty term into MIMIC (DP-MIMIC) on the final binding mode. For this purpose, the best binding mode solutions for two consensus compounds, MFCD00038940 and MFCD00063402, independently identified within the top-scoring 100 molecules (1 % of the database) by MIMIC, DP-MIMIC, and DOCK, are depicted in Figure 3.

In general, three situations have been encountered when introducing a docking penalty term into the similarity score in DP-MIMIC as compared to the original similarity score in MIMIC. First, inclusion of a bump penalty retains the best alignment identified from MIMIC using similarity only. This is especially likely if the ligands are of similar or smaller size and shape than the reference inhibitor and additional clashes with the protein are not expected. Second, inclusion of the bump penalty promotes a lower-scoring solution from MIMIC to the best alignment solution from DP-MIMIC, but the new best alignment solution is just a refinement of the previous one, thus resulting in a binding mode which is qualitatively equivalent to the one identified by MIMIC. This is the case of MFCD00063402. As can be observed in Figure 3, because some steric clash with the active site is identified in the best solution from MIMIC, the second-best solution from MIMIC becomes the best solution from DP-MIMIC, approaching in this case the best solution from DOCK. Third, inclusion of the bump penalty promotes a lower-scoring solution from

. \ r ■ v

,vv r-ii-'i

((K\r'

k

'-i-i'' w

V f

« l>

f

0

MFCD00038940 from MIMIC

MFCD00063402 from MIMIC

(V

/

)f '

f

MFCD00038940 from DP-MIMIC

MFCD00063402 from DP-MIMIC

0

t (^ rj

(U'%

* }

y

i

MFCD00038940 from DOCK MFCD00063402 from DOCK

Figure 3. Binding mode solutions for two consensus compounds, MFCD00038940 and MFCD00063402 (with carbon atoms in white), identified within the best-scoring 100 molecules from MIMIC, DP-MIMIC, and DOCK. The structure of NAPAP (with carbon atoms in green) is provided as a visual reference in all cases. The active site of thrombin is represented by a blue translucent surface.

MIMIC to the best alignment solution from DP-MIMIC, resulting in an essentially different binding mode. This is the case of MFCD00038940. As can be observed in Figure 3, the placement of the naphthalene ring system in MFCD00038940 clashes into the protein active site in the best solution found by MIMIC. Due to this dramatic steric clash, the second-best solution from MIMIC becomes the best solution from DP-MIMIC, where the spatial positions of the naphthalene and benzene rings have been interchanged and a completely different binding mode is proposed. Analysis of the 10 best orientations identified by DOCK reveals that they define two major binding modes. Interestingly, in this case DOCK favors a binding mode closer to that identified by MIMIC, albeit with an orientation with much less 3D overlap with NAPAP. However, solutions resembling the binding mode proposed by DP-MIMIC are also found with lower chemical score. Therefore, hybrid similarity-docking methods as DP-MIMIC or SP-DOCK [20] do represent real alternatives to the pure MIMIC and DOCK methods. With the number of ligand-bound protein structures available in the Protein Data Bank constantly increasing, the possibility of using structure-based methods that maximally exploit all the information present in ligand-bound structures will become more attractive.

0 0

Post a comment