Supplementary Materials
Brain Network Informed Subject Community Detection In Early-Onset Schizophrenia
Zhi Yang, Yong Xu, Ting Xu, Colin W. Hoy, Daniel A. Handwerker,
Gang Chen, Georg Northoff, Xi-Nian Zuo, Peter A. Bandettini
Contents
Supplementary Methods .................................................................................................. 1
Supplementary Figure 1 ................................................................................................... 3
Supplementary Figure 2 ................................................................................................... 5
Supplementary Methods
Evaluating the Influence of Parameters on Community Profiles
To examine the “mean effect” of different choices of k parameters, we summed up the
binary community profiles obtained with k from 2 to 6, under the threshold of p=0.005.
Larger numbers in the summed profile for a given subject indicate that the subject is
more consistently present in the community, regardless of the model parameters. As
shown in Supplementary Figure 2, the distribution of the subjects agreed with the
community profiles obtained with the optimal thresholds (k=5). Furthermore, a cut-off at
2.5 (center of full range in the summed profile) resulted in exactly the same community
profiles as reflected by the optimal thresholds (Figures 1C and 2C). These observations
indicate that the subject communities reflected in the two ICNs are robust across
different choices of parameters.
Linking ICN-derived subject communities to clinical characteristics
The subject communities derived based on ICNs suggested potential subgroups among
subjects. In order to interpret the ICN-derived subject communities, we aimed to link
these subject communities to their clinical characteristics. We hypothesized clinical
characteristics (i.e., the clinical diagnosis separated subjects in two groups; relative
strength of positive and negative symptoms further separate patients into two
subgroups), according to which the subjects could be classified into subgroups that
agree with the subject communities derived solely based on brain networks.
1
A multivariate distance matrix regression (MDMR) analysis (Zapala and Schork, 2012)
was employed to examine these hypotheses. Specifically, a regression model was
constructed to examine the agreement between the subject communities derived based
on ICNs and those indicated by clinical characteristics. The similarity matrix of a given
ICN was converted into a vector and treated as the dependent variable in the regression.
The independent variable was constructed by converting a hypothetical clinical similarity
matrix into a vector. This hypothetical clinical similarity matrix was constructed by
labeling the similarity between the subjects in a given community as one and all other
inter-subject connections as zero, where the subject community was determined by the
hypothetical clinical characteristics (e.g., diagnosis, symptom pattern). With this setting,
correlation coefficient between the ICN-derived similarity matrix and the hypothesized
clinical similarity matrix was computed and then examined for significance. Since ICNderived subject communities informed the hypotheses on clinical characteristics, we
present the specific clinical characteristics we investigated in the Results section.
2
Supplementary Figure 1
Supplementary Figure 1. Summary of other intrinsic functional networks found in gRAICAR.
For each network, a surface map showing constituent brain regions and a similarity matrix
showing the inter-subject consistency of the connectivity pattern are presented. (A) Visual
network; (B) Posterior parietal-prefrontal network; (C) Default mode network; (D) Left-lateralized
parietal-frontal network; (E) Primary motor network; (F) Right-lateralized parietal-frontal network;
(G) Precuneus; (H) Dorsal prefrontal-superior temporal network; (I) The anterior cingulate gyrusangular gyri-dorsal prefrontal cortex network; (J) Motor network; (K) Cingulate gyrus-superior
temporal gyrus network; (L) Primary visual network; (M) Visual network.
3
4
Supplementary Figure 2
Supplementary Figure 2. Overall subject community profiles across different max-clique
merging thresholds (k=2-6) for the PCU-AG network (A) and for the STG-IFG network (B).
These profiles were obtained by summing up the binary subject community profiles with
different thresholds for k with p=0.005. A value of 5 indicates that a given subject is included in
the community for all thresholds, and a value of 0 indicates that the subject is excluded from the
community regardless of the threshold. A majority of the subjects exhibit independence from the
threshold. Using 2.5 (middle point of the value range) as a boundary to binarize these subject
profiles, the subject community profile is exactly the same as the most reproducible profiles
reported in Figure 1C and Figure 2C.
5