FlexX scoring function versus a potential of mean force

As has been demonstrated above, the performance of FlexX in library prioritization is often still moderate, especially when activity differences between highly potent and less potent inhibitors have to be estimated. Alternative scoring schemes are therefore of great interest. Recently, a new knowledge based scoring function has been proposed by Muegge and Martin [47,78].

We were interested to compare its performance to that of the FlexX scoring function, because it relies on a completely different approach and has a different functional form. We have re-implemented this scoring function as a stand-alone C program. FlexX rank 1 solutions for the thrombin, DHFR diaminopyrimidine and COX2 libraries were used as input structures for scoring in order to be able to compare scores of identical complex structures. Structures for which no docking solutions could be found are omitted in the enrichment calculations. We are aware of the fact that this procedure contains a hidden bias, because the generation of the ligand poses was driven by one of the two scoring functions that are compared. In our hands, however, the Muegge scoring function has proven to be less sensitive to small coordinate changes than the FlexX scoring function. We have repeated the re-scoring experiment by using a Powell optimizer [79] in conjunction with the Muegge scoring function and found very minor changes in enrichment factors.

Enrichment plots for the complete thrombin database are shown in Figure 4a. Both scoring functions give very similar curves, reaching almost 50% of the maximum achievable enrichment factor of 6.1 at 15% of the database. Depending on the value of AGrol used, enrichment calculated with the FlexX scoring function can vary significantly, while there is no such flexibility penalty term in the Muegge function. Figure 4b shows enrichment curves for the COX2 database (activity threshold at pK = 7, maximum ef= 6). In this case, FlexX fails completely for the following reasons: The COX2 binding site is a very confined and inaccessible cavity. Inhibitor activity is mainly due to lipophilic interactions within the tight cavity. Small changes in cavity size can significantly alter the size of the contact surface and make it difficult for FlexX to generate accurate ligand placements. In addition, some inhibitors form hydrogen bonds with Arg and Glu residues at one end of the pocket. These residues can move substantially to accommodate different types of inhibitors. Since hydrogen bond scores depend strongly on angle and distance deviations from an ideal value, in this docking experiment the total FlexX score is a meaningless number. In spite of the large uncertainties in ligand placement and protein conformation, moderate enrichment is obtained with the Muegge score. This result supports our experience that the Muegge function is a rather robust scoring function. In this experiment, the neglect of directionality in potentials of mean force is certainly an advantage.

In the third example, there is less doubt about the correct orientation of the ligand than in the two cases discussed above. All compounds of the DHFR library contain a diaminopyrimidine moiety, whose orientation could be expected to remain constant and was taken from an X-ray structure. The remainder of most ligands has little conformational freedom and extends along a deep active site cleft that stretches from the position NADPH cofactor

% of ranked library
Figure 4. Enrichment plots calculated with the FlexX and Muegge scoring functions for the thrombin library (a) and the COX2 library (b).

to Arg57. All in-house X-ray structures of DHFR complexes could be well reproduced as rank 1 solutions of FlexX employing standard settings.

As can be seen from Figure 5d, high enrichment is obtained. In this case, the activity threshold was chosen to be pK = 8.5 (maximum ef= 5.8), since high affinity can be quite easily achieved with diaminopyrimidine ligands. There is a decent correlation (r2 = 0.61) between calculated and experimental binding affinities for both Muegge and FlexX score (Figures 5a and b). There is also good agreement between FlexX and Muegge score, supporting the fundamental validity ofboth approaches (Figure 5c). The correlation between FlexX and Muegge score (r2 = 0.67) is better than that between experimental

Figure 5. Plots of calculated FlexX score (a) and Muegge score (b) versus experimentally determined log(IC50) values for diaminopyrimidines binding to S. aureus DHFR. (c) Correlation between Muegge score and FlexX score. (d) Enrichment plots obtained with FlexX and Muegge scores.

data and each score. Two reasons can be made responsible for this fact: Firstly, systematic errors in the determination of IC50 values can affect the correlation. Secondly, both scoring functions assess only static interactions between protein and ligand and omit many solvation and conformational aspects of ligand binding. It is certainly encouraging to see that protein-ligand interaction terms calculated by two very different methods correlate well. On the other hand, it becomes clear that accurate affinity predictions cannot be based on such interaction terms alone.

0 0

Post a comment