Biology Reference
In-Depth Information
PSN connectivities and (b) selecting the shortest path(s) that contains at least one
residue correlated (e.g., with a correlation coefficient
0.8 in the case study shown
herein) with either one of the two extremities. With the PSN-MD approach, all the
shortest paths that pass the filter of correlation of motions are subjected to a further
filter based upon path frequency, that is, number of frames containing the selected
path divided by the total number of frames in the trajectory.
Collectively, the main differences between PSN-MD and PSN-ENM in terms of
path search include the way in which the cross correlations of atomic motions for
path filtering are computed and the application of a frequency-based extra
filter. In detail, whereas with PSN-MD the cross correlations of atomic motions
are computed on the trajectory frames by means of the LMI method, with PSN-
ENM, they are extracted from the covariance matrix of the deformation modes
computed by ENM-NMA. As for the path filtering issue, PSN-MD refilters those
paths that pass the motion correlation filter by finally keeping only those that exceed
a recurrence cutoff (i.e., presence in a given number of trajectory frames); in
contrast, PSN-ENM applies only the motion correlation filter. Thus, whereas
with PSN-MD, recurrence of network parameters (i.e., links and paths) in the trajec-
tory frames dictates the composition of both PSG and path pool, recurrence-based
filtering does not apply to PSN-ENM. In spite of these significant differences,
the two different approaches tend to produce overlapping outcomes concerning those
nodes and links that recur the most in the predicted pathways (
Raimondi, Felline,
Seeber, et al., 2013
). The PSN-ENM approach tends to predict a more extended com-
munication, likely due to the less heavy filtering applied to the communication
pathways.
Thus, the paths that pass the filtering stage(s) constitute the pool of paths of a
system at given
I
min
and correlation coefficient cutoffs. The statistical analysis of
such pool of paths can lead to the building of global meta paths constituted by the
most recurrent nodes and links in the pool. In this case study, the communication
pathways characterizing the dark and MII states of rhodopsin are, indeed, repre-
sented as meta paths made of nodes that recur in
20% of the retrieved paths
(i.e., “frequent nodes”) and of links satisfying conditions both of being present in
20% of the paths and of connecting “frequent nodes.”
Cluster analysis may provide finer information on the predicted pathways. We
implemented two path clusterization methods differing both in the clusterization al-
gorithms and in the score employed to evaluate the similarity between path pairs (i.e.,
similarity score) (
Raimondi, Felline, Portella, et al., 2013; Raimondi, Felline, Seeber,
et al., 2013
). Irrespective of the clusterization method, for each cluster, following a
pairwise comparison of all cluster members, the center is computed, which is the path
with the highest average similarity among all the paths in the cluster. The center can
be employed as a representative of a given cluster (
Raimondi, Felline, Seeber, et al.,
2013
). Computational indices describing path features can be used as well to choose
representative paths. These indices include the mean square distance fluctuation
(MSDF) either computed between the extreme nodes (this case study) or averaged
over all node pairs in a path (
Raimondi, Felline, Seeber, et al., 2013
).
