525
15 (How to) Profit from Molecular Dynamics-based Ensemble Docking
poses [41]. Averaging the scoring values, these rarely sampled conformations and
the respective scoring values contribute less to the final score [41].
In a virtual screening study on influenza neuraminidase, Cheng et al. used 27
cluster representatives of the holo ensemble and the score of binding to each representative was weighted by cluster size [26]. Known inhibitors, zanamivir and
oseltamivir, were ranked within the top compounds, and superior to the less active inhibitor DANA and the natural substrate sialic acid. However, peramivir was
scored more unfavorable compared to experimental expectations independent from
the weighting method. In this publication also other ranking metrics were evaluated.
Taking the arithmetical mean assumes a normal distribution of the energies of the
conformations underlying the docking experiment. Alternatively, an averaging assuming a harmonic distribution for the K i values [26] or a Boltzmann distribution
for the predicted binding free energy [86, 115] can be applied. Xu et al. discuss
the effects of averaging regarding enrichment of active versus inactive compounds
[115]. They used two ways of merging the hit list: either selection of the best score
from multiple hit list or averaging over multiple scores clustered by docking poses.
This averaging strategy performed slightly better for two cases (thrombin, Pneumocystis carinii dihydrofolate reductase) and both methods performed equally on
the other three out of five cases (acetylcholinesterase, cyclin-dependent kinase 2,
estrogen receptor α) [115].
Average scores for our test scenarios and most favorable scores are shown in
Table 15.1. Regarding the bell-shaped symmetric distribution of scoring values for
periodically extracted structures (Fig. 15.4c, d) we calculated the arithmetic mean
of these values. Zanamivir preserves a higher average score than compound 2 in
docking to conformations from the two closed simulations, which is in agreement
with experimental data. The average scores for compound 2 are in the range of the
score calculated for the near-native pose (Fig. 15.4f), as docking scores between
50 and 60 dominate the distribution of scoring values (Fig. 15.4d). To average the
results for cluster representatives, the scores for the representatives were weighted
by the population of the respective cluster. The ranking trends between the different simulations are preserved independently from the strategy of representation or
score extraction.
Choosing the best scoring pose to determine the docking result corresponds to
the dominant state approximation [119]. In the picture of conformational selection
the optimal conformation for formation of the complex imposes a shift in state
populations. Moreover, the paradigm of conformational selection states that this
optimal conformation will be part of the ligand-free as well as the ligand-bound
ensemble. All other states with less favorable score for the docking pose sampled
during the MD are ignored. This assumes that a ligand selects one single ideal conformation from the MD ensemble. Classical scoring functions are trained to reflect
the free energy of the bound state based on a single complex structure. The selection
procedure would allow integrating force field energies for the protein conformation
in analogy to the ligand conformational strain term in scoring functions.
The free energy of a conformation is reflected in its state probability within
the MD ensemble. Going with the recently formulated implicit ligand theory, an
15 (How to) Profit from Molecular Dynamics-based Ensemble Docking
poses [41]. Averaging the scoring values, these rarely sampled conformations and
the respective scoring values contribute less to the final score [41].
In a virtual screening study on influenza neuraminidase, Cheng et al. used 27
cluster representatives of the holo ensemble and the score of binding to each representative was weighted by cluster size [26]. Known inhibitors, zanamivir and
oseltamivir, were ranked within the top compounds, and superior to the less active inhibitor DANA and the natural substrate sialic acid. However, peramivir was
scored more unfavorable compared to experimental expectations independent from
the weighting method. In this publication also other ranking metrics were evaluated.
Taking the arithmetical mean assumes a normal distribution of the energies of the
conformations underlying the docking experiment. Alternatively, an averaging assuming a harmonic distribution for the K i values [26] or a Boltzmann distribution
for the predicted binding free energy [86, 115] can be applied. Xu et al. discuss
the effects of averaging regarding enrichment of active versus inactive compounds
[115]. They used two ways of merging the hit list: either selection of the best score
from multiple hit list or averaging over multiple scores clustered by docking poses.
This averaging strategy performed slightly better for two cases (thrombin, Pneumocystis carinii dihydrofolate reductase) and both methods performed equally on
the other three out of five cases (acetylcholinesterase, cyclin-dependent kinase 2,
estrogen receptor α) [115].
Average scores for our test scenarios and most favorable scores are shown in
Table 15.1. Regarding the bell-shaped symmetric distribution of scoring values for
periodically extracted structures (Fig. 15.4c, d) we calculated the arithmetic mean
of these values. Zanamivir preserves a higher average score than compound 2 in
docking to conformations from the two closed simulations, which is in agreement
with experimental data. The average scores for compound 2 are in the range of the
score calculated for the near-native pose (Fig. 15.4f), as docking scores between
50 and 60 dominate the distribution of scoring values (Fig. 15.4d). To average the
results for cluster representatives, the scores for the representatives were weighted
by the population of the respective cluster. The ranking trends between the different simulations are preserved independently from the strategy of representation or
score extraction.
Choosing the best scoring pose to determine the docking result corresponds to
the dominant state approximation [119]. In the picture of conformational selection
the optimal conformation for formation of the complex imposes a shift in state
populations. Moreover, the paradigm of conformational selection states that this
optimal conformation will be part of the ligand-free as well as the ligand-bound
ensemble. All other states with less favorable score for the docking pose sampled
during the MD are ignored. This assumes that a ligand selects one single ideal conformation from the MD ensemble. Classical scoring functions are trained to reflect
the free energy of the bound state based on a single complex structure. The selection
procedure would allow integrating force field energies for the protein conformation
in analogy to the ligand conformational strain term in scoring functions.
The free energy of a conformation is reflected in its state probability within
the MD ensemble. Going with the recently formulated implicit ligand theory, an