Knowledge-based Optimization of Protein-Ligand-Complex Geometries
The aim of this work was to develop a tool to optimize insilico generated protein-ligand complexes according to DrugScore (DS) potentials. DS is typically used to rescore ligand geometries that were generated by docking. Thus, these poses are optimized according to the scoring function used by the s...
|PDF Full Text
No Tags, Be the first to tag this record!
|The aim of this work was to develop a tool to optimize insilico generated protein-ligand complexes according to DrugScore (DS) potentials. DS is typically used to rescore ligand geometries that were generated by docking. Thus, these poses are optimized according to the scoring function used by the selected docking algorithm. Applying DS to such a geometry does not necessarily guarantee reliable and relevant scoring. Considering the steepness of the DS potentials, even small variations in the atomic positions can lead to large differences in the resulting scores. Thus, a local optimization with respect to DS is strongly recommended in this case.
In 2009, O’Boyle et al. stated, that a local optimization is always constrained to the energy well on the potential surface in which the original pose resides. So there may be even deeper wells nearby which are not considered in the local optimization but would be equally valid. The new tool MiniMuDS, developed in this thesis, should account for this problem. On the other hand, MiniMuDS is not intended to perform a global optimization since this would, at the end, result in a new docking algorithm. Instead, the new algorithm is supposed to stay close to the pose generated by the original docking engine and simply adapts it to the DS function, a task typically addressed by local search methods.
MiniMuDS was to combine these two tasks by avoiding a strictly local optimization without extending to a fully global search at the same time. Therefore, a strategy was implemented, that contains elements of a global optimization, but is still restricted to a local part of the search space. Simply speaking, the applied algorithm can overcome small hills on the potential surface, but only if the following valley is deeper than the current one. Thus, major energetic barriers between basically different conformations will not be passed. In the validation of MiniMuDS several important properties were shown:
1. The optima of the applied energy model correspond impressively well to the experimentally determined native states of the evaluated complexes. This was shown by the optimization of the original crystal structures, which resulted in an average rmsd of about 0.5Å, a value much smaller than the one observed in case of in-silico generated geometries. This deviation has to be seen in light of the positional accuracy estimated for experimental structure determination. The observed deviations virtually fall into the same range.
2. The aim of conserving the given binding modes was achieved. The presented method allows for modifications up to 2Å rmsd compared to the input geometry. Remarkably, not even 5% of the optimized docking poses fully exploited this available space. On average a modified geometry shows an rmsd of about 1Å to the input structure.
3. MiniMuDS improves a given docking solution by about 0.1Å on average. The best performance was observed for well docked poses between 1 and 2Å rmsd which could be improved by up to 0.3Å on average.
4. It was shown that an optimization exceeding the restrictions of a strictly local search can improve the resulting ranking. Up to 4.7% better success rates at a 2Å cutoff and an improvement of up to 9.3% at the 1Å level were received when comparing MiniMuDS to a local optimization.
5. Taking into account not only the top ranked solution but the whole ranking, it was shown that MiniMuDS strongly improves the discrimination between near-native and misplaced poses. Geometries with lower rmsd values to the crystal structure are more likely to be placed within the first positions of the ranking.
6. The inclusion of additional flexible components into the optimization is easy to manage while results can strongly benefit. This was shown using the example of protein side chain flexibility and binding relevant water molecules.
7. Considering computational efforts, it was shown that it is sufficient to only subject the 10 top-ranked docking solutions to an optimization. This consistently yielded slightly better ranking results for all applied protocols compared to an optimization of all generated docking solutions. At 80% less computational effort, up to 4.7% higher success rates at 2Å and 2.1% higher once at a 1Å cutoff were recorded.
Especially the last aspect confirms that it is advisable to focus only on those docking poses that were already ranked high by another scoring function. This way, only poses that score well with respect to two different scoring functions are considered, taking thereby advantage of some kind of consensus effect.
In light of these findings, the usage of at least a local optimization has to be strongly recommend before applying DS for rescoring purposes. Beyond that, the application of a more sophisticated search strategy like the one implemented in MiniMuDS is suggested. In particular, when dealing with small, lead-like structures, the presented method showed to substantially improve the results.