Examining Functional Resting-State Connectivity in Psychosis and Its Subgroups in the Bipolar-Schizophrenia Network on Intermediate Phenotypes Cohort

Abstract

BACKGROUND:

We sought to examine resting-state functional magnetic resonance imaging connectivity measures in psychotic patients to both identify cumulative differences across psychosis and subsequently probe deficits across conventional DSM-IV diagnoses and a newly identified classification using cognitive/neurophysiological data (Biotypes).

METHODS:

We assessed 1125 subjects, including healthy control subjects, probands (schizophrenia, schizoaffective disorder, psychotic bipolar disorder), and relatives of probands. Probands and relatives were also segregated into Biotype groups (B1–B3, B1R–B3R using a method reported previously). Empirical resting-state functional magnetic resonance imaging networks were derived using independent component analysis. Global psychosis-related abnormalities were first identified. Subsequent post hoc t tests were performed across various diagnostic categories. Follow-up linear mixed model compared significance of within-proband differences across categories. Secondary analyses assessed correlations with biological profile scores.

RESULTS:

Voxelwise tests between proband and control subjects revealed nine abnormal networks. Post hoc analysis revealed lower connectivity in most networks for all proband subgroups (DSM and Biotypes). Within-proband effect sizes of discrimination were marginally better for Biotypes over DSM. Reduced connectivity was noted in relatives of patients with schizophrenia in two networks and relatives of patients with psychotic bipolar disorder in one network. Biotype relatives showed similar deficits in one network. Connectivity deficits across four networks were significantly associated with cognitive control profile scores.

CONCLUSIONS:

Overall, we found psychosis-related connectivity deficits in nine large-scale networks. Deficits in these networks tracked more closely with cognitive control factors, suggesting potential implications for disease profiling and therapeutic intervention. Biotypes performed marginally better in terms of separating out psychosis subgroups compared with conventional DSM or psychiatric diagnoses.

Traditional diagnostic classification of psychiatric disorders relies predominantly on clinical symptomatology. On the one hand, there may be considerable biological overlap among disorders traditionally regarded as distinct, for example, schizophrenia and bipolar disorder. On the other hand, it is likely that syndromes with similar clinical presentations capture etiopathologically distinct disorders (1,2). Acknowledging the latter, authors of the recently revised DSM-5 classification advocated for development of more dimensional assessments that cross-cut conventional diagnostic boundaries (3–5). There is a need for more objective and specific biologically based classifications of psychiatric brain disorders that could map to pathology and could indicate individualized treatments, informed by patient-specific etiopathology (4). Such biological signatures could be examined among biological relatives to inform underlying genetic/molecular risk mechanisms of psychotic disorders.

Accordingly, our group recently constructed a novel biologically driven disease classification of psychosis (Biotypes) (6). In brief, rather than relying on clinical symptoms, the Biotypes were instead constructed using a select panel of brain-based cognitive and electrophysiological biomarkers (not including resting-state functional magnetic resonance imaging [rs-fMRI]) spanning multiple psychosis-related neurocognitive/neurophysiological domains in a large cohort of N = 1872 subjects participating in a multisite, multiyear project, the Bipolar-Schizophrenia Network on Intermediate Phenotypes. Overall, Biotype separations were summarized by two functions, cognitive control and sensorimotor reactivity (6). These psychosis Biotypes did not respect traditional diagnostic boundaries, but rather each Biotype contained members of all DSM diagnoses. In the original study delineating the groupings, each Biotype demonstrated its own unique biofingerprint (6). Biotype group B1 manifested impaired cognitive control and low sensorimotor reactivity, group B2 demonstrated impaired cognitive control but exaggerated sensorimotor response, and group B3 exhibited near normal cognitive and sensorimotor characteristics. More details on Biotypes is provided in Supplemental Methods and elsewhere (6).

The purpose of the present study was to use rs-fMRI to examine connectivity profiles in psychosis as a whole and in both Biotype and DSM subclassification schemes. Resting fMRI has been identified as a robust biomarker in several diseases, including psychosis (7–17). Further, because core networks within rs-fMRI are familial and similarly abnormal in relatives of psychosis probands (14,18–20), we also determined whether significant proband abnormalities using the above classification systems were also observed in their relatives.

We used a group independent component analysis (ICA) approach to automatically extract, identify, and study intrinsic rs-fMRI networks (21). ICA inherently parses noise networks (22,23). Several recent studies have used network/pattern analysis in conjunction with conventional diagnostic criteria to uncover multiple overlapping networks across schizophrenia (SZ) and psychotic bipolar disorder (PBP) (13,14,19,24). Some of these networks are also seen in relatives of psychosis probands (11–14,20,25,26). Few studies have shown significant within-proband differences (8,11,27) when using conventional DSM-like diagnostic schemes. In the present project we expected to 1) find extensive commonalities across psychosis groups compared with healthy control (HC) subjects, 2) validate rs-fMRI connectivity in terms of neurobiological distinction among newer psychosis subgroups, 3) uncover both unique and shared network profiles across Biotype and DSM groups, and 4) explore additional neurobiological associations of connectivity deficits in psychosis.

METHODS AND MATERIALS

Subject recruitment and scanning were completed at five Bipolar-Schizophrenia Network on Intermediate Phenotypes sites. Demographic details are provided in Table 1. Clinically stable patients were diagnosed using the Structured Clinical Interview for DSM-IV. We examined 1125 subjects: 258 HC subjects, 206 SZ probands, 149 schizoaffective disorder (SAD) probands, 163 PBP probands, and their relatives (137 relatives of patients with SZ [SZR], 101 of SAD [SADR], and 111 of PBP [PBPR]). Probands were stably medicated outpatients. Relatives meeting criteria for Axis-I proband-like psychotic disorders (SZ, SAD, PBP) were grouped with the corresponding proband category. The remaining biological relatives comprised SZR, SADR, and PBPR groups. The above probands were also classified into three Biotype groups B1 (n = 143), B2 (n = 183), and B3 (n = 192) and their relatives B1R (n = 88), B2R (n = 116), and B3R (n = 145) detailed in the Supplement and elsewhere (6). The present study used a subset (60%) of the population analyzed in the original Biotype classification study (6). The study protocol was approved by the institutional review board at each local site. After a complete description of the study was given to subjects, their informed consent was obtained.

Table 1.

Demographic and Clinical Characteristics of the Overall Sample (N = 1125) Across DSM-IV and Biotypes

	Controls (n = 258)		Bipolar Probands (n = 163)		Schizoaffective Probands (n = 149)		Schizophrenia Probands (n = 206)		Bipolar Relatives (n = 111)		Schizoaffective Relatives (n = 101)		Schizophrenia Relatives (n = 137)		Statistics
Characteristic	Mean	SD	Mean	SD	Mean	SD	Mean	SD	Mean	SD	Mean	SD	Mean	SD	F	p Value
DSM
Age, years	37.99	12.4	36.34	12.65	35.78	12.24	34.8	12.52	38.3	16.19	42.17	16.05	44.27	15.79
Clinical profiles
Cognitive control
Sensorimotor reactivity	0.24	1.05	0.2	1.19	0.03	0.99	20.09	1.16	0.14	1.24	0.27	1.27	0.19	1.06	2.36	.03
	n	%	n	%	n	%	n	%	n	%	n	%	n	%	χ2	p Value
Sex
Male	114	44.19	56	34.36	68	45.64	139	67.48	41	36.94	33	32.67	42	30.66	69.85	<.001
Female	144	55.81	107	65.64	81	54.36	67	32.52	70	63.06	68	67.33	95	69.34
Ethnicity
Non-Hispanic	233	0.90	149	0.91	128	0.86	188	0.91	98	0.88	94	0.93	122	0.89	5.06	NS
Hispanic	25	0.10	14	0.09	21	0.14	18	0.09	13	0.12	7	0.07	15	0.11
Sites
HFD	68	0.26	36	0.22	54	0.36	57	0.28	46	0.22	43	0.43	57	0.42	114.53	<.001
BAL	54	0.21	32	0.20	29	0.19	74	0.36	16	0.08	21	0.21	34	0.25
CHI	55	0.21	58	0.36	25	0.17	35	0.17	35	0.17	14	0.14	18	0.13
DET	22	0.09	11	0.07	2	0.01	18	0.09	3	0.01	0	0.00	8	0.06
DAL	57	0.22	26	0.16	40	0.27	22	0.11	11	0.05	23	0.23	20	0.15
	Controls (n = 258)		Biotype1 (B1) (n = 143)		Biotype2 (B2) (n = 183)		Biotype3 (B3) (n = 192)		B1 Relatives (n = 88)		B2 Relatives (n = 116)		B3 Relatives (n = 192)		Statistics
	Mean	SD	Mean	SD	Mean	SD	Mean	SD	Mean	SD	Mean	SD	Mean	SD	F	p Value
Biotypes
Age, years	37.99	12.4	35.85	12.9	35.52	12.19	35.41	12.49	40.5	15.29	41.91	16.43	42.41	16.49	7.28	<.001
Clinical profiles
Cognitive control	1.49	1.38	−1.55	1.17	−0.79	1.13	1.21	1.1	0.36	1.44	1.16	1.33	1.72	1.37	155.43	<.001
Sensorimotor reactivity	0.24	1.05	−0.73	1.07	0.71	0.9	−0.04	0.98	0.01	1.03	0.51	1.24	0.06	1.17	29.31	<.001
	n	%	n	%	n	%	n	%	n	%	n	%	n	%	χ2	p Value
Sex
Male	114	0.44	77	0.54	87	0.48	99	0.52	33	0.38	32	0.28	51	0.27	29.81	<.001
Female	144	0.56	66	0.46	96	0.52	93	0.48	55	0.63	84	0.72	94	0.49
Ethnicity
Non-Hispanic	233	0.90	124	0.87	158	0.86	183	0.95	74	0.84	108	0.93	132	0.69	15.26	.02
Hispanic	25	0.10	19	0.13	25	0.14	9	0.05	14	0.16	8	0.07	13	0.07
Sites
HFD	68	0.26	31	0.22	63	0.34	52	0.27	27	0.31	57	0.49	62	0.32	83.47	<.001
BAL	54	0.21	53	0.37	33	0.18	49	0.26	24	0.27	12	0.10	35	0.18
CHI	55	0.21	28	0.20	34	0.19	56	0.29	17	0.19	22	0.19	28	0.15
DET	22	0.09	13	0.09	11	0.06	7	0.04	6	0.07	4	0.03	1	0.01
DAL	59	0.23	18	0.13	42	0.23	28	0.15	14	0.16	21	0.18	19	0.10

For medication information on subjects, refer to Supplemental Tables S1 and S2.

BAL, Baltimore; CHI, Chicago; DAL, Dallas; DET, Detroit; HFD, Hartford; NS, nonsignificant.

Data Acquisition and Preprocessing

All subjects underwent a single 5-minute run of rs-fMRI on a 3T MRI scanner at each site. Participants were instructed to keep their eyes open, focus on a crosshair displayed on a monitor, and remain still during the scan. In addition, head motion was restricted with a custom-built head-coil cushion. Scanning protocols (accounted for during analysis) were mostly consistent as noted elsewhere (14).

We discarded the initial six images. Remaining images were reconstructed offline and realigned with INRIAlign (http://www-sop.inria.fr/epidaure/software/INRIAlign/index.php), implemented in Statistical Parametric Mapping (SPM8; http://www.fil.ion.ucl.ac.uk/spm/software/spm8/). Realigned images were then corrected for slice timing differences, spatially normalized to Montreal Neurological Institute space and smoothed by an 8-mm isotropic Gaussian kernel. At each stage the output was visually validated, and scans were discarded if they did not meet quality control standards (artifacts, noise, excessive motion [>1.5voxels translational and/or 3˚ rotational] missing data, etc). After quality control, a total of 1125 subjects were used for further analysis.

Primary Data Analysis

ICA.

fMRI data were analyzed using a group-ICA algorithm (GIFT v1.3f; http://mialab.mrn.org/software/gift) to identify spatially independent networks, by pooling data from all participants into a single ICA analysis. After data reduction through two principal component analysis stages (28), 20 mutually independent components were determined using minimum description length criteria (29). Time courses and spatial maps (SMs) were then back-reconstructed for each participant, maintaining interindividual variability. The SM loading coefficients from a group ICA represent the degree of connectivity of each voxel’s time course with the aggregate/overall time course of the network, thus inherently representing a functional connectivity map. Each network’s consistency metrics were derived from multiple (N = 20) ICA runs using a clustering approach called ICASSO (30).

Selection of rs-fMRI ICA Networks.

Credible rs-fMRI networks (as opposed to physiological/susceptibility artifacts) were identified using a method described previously (22). In brief, two metrics, the dynamic range and the low frequency-to-high frequency power ratio (31) were calculated for each network. An arbitrary cutoff was chosen based on the above metrics such that credible networks were separated from potential noisy networks (Supplemental Figure S1). Credible SMs were further screened visually (by two separate imaging raters) to ensure that there was low overlap with known vascular, ventricular, motion, and susceptibility artifacts before retaining for further analysis.

Statistical Analysis

To determine voxelwise aberrant spatial profiles pertinent to psychosis, a two-sample t test across HC subjects and pooled probands was conducted across each of the credible intrinsic networks using SPM8. Bidirectional contrast maps were derived for the HC subject versus proband analysis for each network separately. Aberrant voxels were identified by using permutation-testing in conjunction with the threshold-free cluster enhancement option (32) as used in the threshold-free cluster enhancement toolbox (http://dbm.neuro.uni-jena.de/tfce/). Results were thresholded at a familywise error p <.05 level that stringently accounted for multiple comparisons across all brain voxels (Figure 1). The average loading coefficient across the set of aberrant voxels identified from the above was extracted using the REX toolbox (http://www.nitrc.org/projects/rex/) across individuals for each network separately. Average loading coefficients were then adjusted for age (based on HC subjects only), sex, and site. We further adjusted the data (at the group level) using mean framewise displacement measures to account for any remaining spurious connectivity as indicated in Power et al. (33). We also included site-by-group interactions in the full model. This term was later removed if it was not found to be significant.

Figure 1.

(Left) Spatial profiles of all valid independent component analysis (ICA)-driven intrinsic networks analyzed in the study. Thresholded and displayed at Z > 2. (Right) Voxelwise differences of within-network ICA connectivity between healthy control subjects and psychosis probands using the threshold-free cluster enhancement approach across all 13 valid ICA networks. Only 9 of 13 networks showed significant voxelwise differences across healthy control subjects and pooled probands, including cuneus-occipital (red), right frontoparietal network (light green), left frontoparietal network (dark blue) cerebellar-occipital (pink), anterior default mode network (yellow), inferior posterior DMN (orange), superior posterior DMN (purple), bilateral temporoparietal (dark green), and frontoparietal control network (white). Voxelwise results were thresholded at p < .05 familywise error rate for multiple comparison correction across the whole brain. RSN, resting state network.

One-Way Analysis of Variance.

An omnibus 1-way analysis of variance (ANOVA) was conducted on averaged loading coefficients across each network. Post hoc analyses were then conducted across both DSM-IV and Biotype classifications to assess HC subjects versus probands, HC subjects versus relatives, and between-proband differences. A false discovery rate (FDR) correction (34) was applied here to all relevant post hoc analyses to adjust for multiple comparisons. All post hoc statistical analysis was performed in SPSS version 19 (http://www-01.ibm.com/software/analytics/spss/).

Two-Way Linear Mixed Model Analysis Within Probands.

To document whether between-proband differences were significantly different between DSM and Biotypes, we followed up with a 2 × 3 linear mixed effects (LME) model analysis (in probands only) focusing on networks that showed significant within-proband differences in either DSM or Biotypes from the earlier one-way ANOVA analysis. The model consisted of DSM and Biotypes modeled as repeated measures, each with three between-subject levels (SZ, SAD, PBP, or B1, B2, B3) and most importantly interactions between the above two. We looked for significant interaction terms in the above model to assess if the differences between the three levels differed significantly between Biotypes and DSM.

Assessing Neurophysiological Relations

To further discern the neurophysiological correlates of connectivity deficits, we correlated connectivity values against the two primary composite factors, cognitive control and sensorimotor reactivity that characterized Biotype classes (6). Significance values were adjusted for multiple comparisons using an FDR correction as applied previously (34). In addition, regression analysis was also conducted on each DSM and Biotype subgroup separately. Pairwise within-proband difference of slopes were then assessed from these data.

Medication-Related Analyses

To reveal any medication effects in our study we converted all probands’ available antipsychotic data to their respective chlorpromazine dosage equivalents using the methods of Andreasen et al. (35). We then correlated ICA network signal to these measures after controlling for aforementioned variables. Because patients were also on other medication classes, we assessed rs-fMRI network relations with binary coded (on/off) medication status on four major drug classes, including primary antipsychotics, antidepressants, and mood stabilizers (with and without lithium). Separate t tests were conducted for each medication class to test their effects across probands’ rs-fMRI networks. Data were adjusted for multiple comparison using FDR. All medication data were based on current status and reported in Supplemental Tables S1 and S2. Distribution of Biotypes and DSM probands/relatives by diagnoses are reported in Supplemental Table S3.

RESULTS

Akin to the parent sample used to derive Biotypes, although asymmetric, each Biotype contained a significant representation of SZ, SAD, and PBP (Supplemental Table S3).

Thirteen ICA components were judged credible rs-fMRI networks using the above-described methods (Figure 1 and Supplemental Figure S1). Among the 13 rs-fMRI networks, voxelwise tests identified connectivity disturbances among probands in 9 networks (Figure 1B). These networks included 1) cuneus-occipital (red; visual), 2) right frontoparietal (rFPN; light green; attentional-working memory), 3) left frontoparietal (lFPN; white; attentional/working-memory), 4) cerebellar-occipital (pink; visuomotor integration), 5) anterior default mode network (aDMN; yellow), 6) inferior posterior DMN (ipDMN; orange), 7) superior posterior DMN (spDMN; purple), 8) bilateral temporoparietal (dark green), and 9) frontoparietal (white; cognitive control). Because we found no significant site-by-group interactions for any network they were not included in the final model as mentioned previously.

One-way ANOVA Analysis (Post Hoc Effects)

Post hoc effects on average connectivity measures derived from the above deviant voxels evaluated across both DSM-IV diagnoses and Biotypes are reported in the sections below.

Cuneus-Occipital Network (Red; Visual)

DSM:

This network only showed reduced connectivity in SZ probands compared with HC subjects. No other significant differences were detected in DSM probands or relatives.

Biotypes:

Biotypes B1 and B2 showed significantly diminished connectivity compared with HC subjects but not B3, consistent with their previously discussed neurobiological profiles. Within Biotype probands, results suggested significantly reduced connectivity in B2 compared with B3. No significant effects were found in relatives versus HC subjects.

rFPN (Light Green; Attentional-Working Memory)

DSM:

This network showed connectivity decreases across all three DSM groups. Similar connectivity differences were found in SZR but not PBPR or SADR. In addition, SZ and SAD differentiated significantly from each other.

Biotypes:

Similar to DSM, all three Biotype groups had reduced connectivity compared with HC subjects for this network. Within Biotype probands, both B1 and B2 showed significantly reduced connectivity compared with B3. In addition, both B1R and B2R showed lower connectivity with respect to HC subjects, consistent with patterns demonstrated by their respective probands.

lFPN (Dark Blue; Attentional-Working Memory)

DSM:

Effects were similar as those to the rFPN, with all three DSM groups having reduced connectivity compared with HC subjects. However, no relative group showed significant effects.

Biotypes:

Similar to DSM, all Biotype proband groups showed reduced connectivity with no effects seen in relatives.

Cerebellar-Occipital Network (Pink; Visuomotor Integration)

DSM:

Reduced cerebellar-occipital connectivity was noted across all proband DSM groups compared with HC subjects. Similar connectivity deficits were observed in relatives of PBP only.

Biotypes:

Similar to DSM, all Biotype probands showed reduced connectivity in the cerebellar-occipital network. No effects were seen in Biotype relatives.

aDMN (Yellow)

DSM:

Effects of decreased connectivity within this network were again similarly propagated through all proband groups relative to HC subjects in DSM. Similar connectivity deficits were observed in SZR only.

Biotypes:

Again, all three Biotype groups showed significantly reduced connectivity. However, no Biotype relative showed any significant effect.

ipDMN (Orange)

DSM:

All three proband groups were equally affected (decreased connectivity) across DSM groups. None of the relatives showed significant differences in connectivity profiles.

Biotypes:

All three Biotype proband groups showed decreased connectivity versus HC subjects. In addition, B1 and B2 showed significantly decreased connectivity compared with B3. No differences were observed in relatives.

spDMN (Purple)

DSM:

Diminished network connectivity was only observed in all proband groups relative to HC subjects. No differences in relatives were noted.

Biotypes:

Across Biotypes, all three groups showed decreased connectivity. Further, B2 showed significantly decreased connectivity compared with B3. No effects were seen in relatives.

Temporoparietal Network (Dark Green)

DSM:

All three DSM groups showed reduced connectivity compared with HC subjects. No relative group showed significant effects.

Biotypes:

Similar to DSM, all Biotype proband groups showed reduced connectivity with no such effects seen in the relative group.

Frontoparietal Control Network (White)

DSM:

All three DSM groups showed reduced connectivity compared with HC subjects. No relative group showed significant effects.

Biotypes:

Similar to DSM, all Biotype proband groups showed reduced connectivity with no such effects seen in the relative group.

Please see Figure 2 and Table 2 for a representation of significant post hoc effects across networks.

Figure 2.

Bar plots of post hoc effects of average connectivity measures (residuals adjusted for age, sex, and site) derived from the voxelwise tests. Data is presented across different networks separated by psychosis probands (left) and their relatives (right). Error bars represent SEM. a-DMN, anterior default mode network; B1, Biotype group 1; B1R, relative of Biotype group 1; B2, Biotype group 2; B2R, relative of Biotype group 2; B3, Biotype group 3; FPN, frontoparietal network; ip-DMN, anterior default mode network; PBP, psychotic bipolar disorder; PBPR, psychotic bipolar disorder relative; SADP, schizoaffective disorder proband; SADR, schizoaffective disorder relative; sp-DMN, superior posterior default mode network; SZP, schizophrenia proband; SZR, schizophrenia relative.

Table 2.

Overall Significant Post Hoc Effects Observed in the Study for Both Probands/Relatives Versus Control and Between Probands Across DSM-IV and Biotypes for Each of the Nine Intrinsic Networks

Network	DSM-IV		Biotypes
	Controls	Between Probands	Controls	Between Probands
Red (Cuneus-occipital)	HC > SZ	—	HC > B1/B2	B3 > B2
Light Green (rFPN)	HC > SZ/SAD/PBP/SZR	SZ > SAD	HC > B1/B2/B3/B1R/B2R	B3 > B1/B2
Dark Blue (lFPN)	HC > SZ/SAD/PBP	—	HC > B1/B2/B3	—
Pink (Cerebellar-occipital)	HC > SZ/SAD/PBP/PBPR	—	HC > B1/B2/B3	—
Yellow (aDMN)	HC > SZ/SAD/PBP/SZR	—	HC > B1/B2/B3	—
Orange (ipDMN)	HC > SZ/SAD/PBP	—	HC > B1/B2/B3	B3 > B1/B2
Purple (spDMN)	HC > SZ/SAD/PBP	—	HC > B1/B2/B3	B3 > B2
Dark Green (Frontotemporal Motor)	HC > SZ/SAD/PBP	—	HC > B1/B2/B3	—
White (Frontoparietal Control)	HC > SZ/SAD/PBP	—	HC > B1/B2/B3	—

Post hoc effects were adjusted using a false discovery rate correction p < .05 to account for multiple comparisons. Note the lack of significant within-proband difference across DSM. However, Biotypes show four of nine networks with at least one significant within-proband post hoc difference.

aDMN, anterior default mode network; B1, Biotype group 1; B1R, relative of Biotype group 1; B2, Biotype group 2; B2R, relative of Biotype group 2; B3, Biotype group 3; HC, healthy control; ipDMN, inferior posterior default mode network; lFPN, left frontoparietal network; PBP, psychotic bipolar disorder; PBPR, relative of psychotic bipolar disorder; rFPN, right frontoparietal network; SAD, schizoaffective disorder; spDMN, superior posterior default mode network; SZ, schizophrenia; SZR, schizophrenia relative.

2 × 3 LME Analysis

Among the four networks (cuneus-occipital, lFPN, ipDMN, and spDMN) further tested, LME revealed significant interactions among the three proband levels across DSM and Biotypes (p = .024) in the cuneus-occipital network. Further, we saw a strong trend for interaction in both the ipDMN (p = .07) and spDMN (p = .06). We saw no significant interaction for the lFPN, suggesting that the within-proband differences were not sufficiently different between DSM and Biotypes here.

Glass effect size calculations revealed generally larger deficits in B1 and B2 across networks than in HC subjects. Similarly for DSM, we found SAD probands and SZ capturing larger deficits (Supplemental Figure S2). Similar trends were also noticed for relatives. Within-proband effect sizes (Cohen’s d) for the two categories are shown in Supplemental Figure S3.

Association With Neurobiological Profiles

Secondary analyses revealed that deficits found in all networks (except for aDMN and frontotemporal motor) showed significant positive correlations with cognitive control scores; however, no such associations were found for the sensorimotor reactivity profile scores (Table 3). Additional analyses revealed no differences in slopes related to cognitive control or sensorimotor scores across both DSM and Biotype proband subgroups when FDR corrected for multiple comparisons.

Table 3.

Correlation of ICA Network Coefficients With Cognitive and Sensorimotor Profile Scores as Derived in Clementz et al. (6) (Across all Probands)

Network	Cognitive Control		Sensorimotor Reactivity
	Pearson Correlation (r)	FDR Adjusted p Value (q Value)	Pearson Correlation (r)	FDR Adjusted p Value (q Value)
Cuneus-occipital (Visual)	.174	<.001	−.062	NS
Right Frontoparietal (Attentional Working Memory)	.159	<.001	−.025	NS
Left Frontoparietal (Attentional Working Memory)	.088	<.05	.018	NS
Cerebellar-occipital (Visuomotor Integration)	.082	<.05	.062	NS
aDMN	.068	NS	−.022	NS
ipDMN	.166	<.001	−.009	NS
spDMN	.168	<.001	−.064	NS
Frontotemporal Motor	.066	NS	.030	NS
Frontoparietal Control Network	.097	<.05	−.003	NS

Note the disparity in ICA association between the two neurobiological domains.

aDMN, anterior default mode network; FDR, false discovery rate; ICA, independent component analysis; ipDMN, inferior posterior default mode network; NS, nonsignificant; spDMN, superior posterior default mode network.

Association With Medication Data

In terms of medication use across Biotypes, assessed using chi-squared tests, we found no significant association in antidepressants or mood stabilizers (with and without lithium) use across the different Biotype groups. Although we found a significant difference in primary antipsychotic usage among Biotype groups (p <.001), no associations between chlorpromazine equivalents and ICA network connectivity were found in the present study. Further, independent t tests computed across probands using binary on/off medication status in four major classes of drugs (including primary antipsychotics) revealed no significant differences in ICA network connectivity across networks when corrected for multiple comparisons using FDR.

DISCUSSION

As expected, we found abnormal connectivity deficits related to psychosis as a whole in 9 of 13 large-scale ICA networks. Given that psychotic illnesses likely stem from core brain circuit disruptions, it is intuitive to agnostically derive patient groups or classes based on brain biology rather than symptom presentations. Thus, the newly developed Biotype classification used a battery of neurobehavioral biomarkers to separate psychosis probands into three discrete classes transcending traditional symptom-based boundaries. As previously noted, Biotype group B1 manifested impaired cognitive control and low sensorimotor reactivity, B2 demonstrated impaired cognitive control but exaggerated sensorimotor response, and B3 exhibited near normal cognitive and sensorimotor characteristics (6). On testing the above characteristics in the present sample, results were consistent with previous reports with the exception of B3 showing decreased cognitive and sensorimotor characteristics, albeit much less marked compared with that observed for B1 in the present sample.

Network Abnormalities in Psychosis

Consistent with previous reports of ours and others (8,11,13,14,16,27), we found extensive connectivity deficits in psychosis probands across nine large-scale ICA networks (Figure 1). Effect sizes across networks were medium to large, ranging from glass delta of 0.22 to 0.64. Functional domains of these networks encompassed a variety of areas known to be compromised in psychiatric disorders, including cognitive control, working memory, attention, and introspective thought maintenance (36).

Effects Grouped by DSM-IV

In relation to HC subjects, all networks showed reduced connectivity across all three proband groups. Interestingly, connectivity deficits found were significantly larger for SAD probands in the right frontoparietal (attentional) networks than for SZ probands. In other networks SAD probands showed similar effect sizes as SZ probands and/or PBP. Given the varying effect sizes of the SAD proband group, it is recommended that, similar to this study, the SZ and SAD proband groups should be analyzed separately in future studies. Among relatives, two networks (rFPN, aDMN) showed connectivity abnormalities in SZR only; similarly, the cerebellar-occipital network was affected in PBPR only, suggesting these networks might be good endophenotypic candidates for a study of SZ and/or PBP, whereas other networks seem to be suitable biomarker choices.

Overall, these results are highly consistent with a recently published report showing the intrinsic rFPN to be compromised in both SZ and their relatives (25). Our results also agree with prior reports showing cerebellar functional connectivity abnormalities in SZ and PBP using both ICA and similar connectivity methods (19,25,37–39). Our finding is also in line with a recent report showing decreased cerebellar connectivity in subjects who were at ultra-high risk of psychosis (40). Further, anomalous neural connectivity within this region has also been attributed to abnormal emotional and persistent auditory hallucinations in SZ (41). Although we saw no within-proband separation in both Biotypes and DSM for the temporoparietal network, all proband groups showed diminished network connectivity, consistent with prior studies showing similar dysfunction in behavior and biological function of this network in psychosis (11,13,36,41).

Effects Grouped by Biotypes

Compared with HC subjects, all three Biotype groups had reduced connectivity across all networks except cuneus-occipital, which was affected in only B1 and B2. Biotypes showed a significant within-proband difference in four networks. Specifically, two networks (rFPN, ipDMN) showed significant reductions in both B1 and B2 compared with B3. Two networks (spDMN and cuneus-occipital) showed significantly diminished connectivity in B2 compared with B3. An important observation in regard to the widely studied DMNs was that the three data-driven networks reported here seemed to demonstrate varying connectivity abnormalities across DSM/Biotypes, emphasizing that studying these DMNs separately is not only justified but also recommended in future studies to make proper interpretations regarding abnormalities within this network. Interestingly, we also demonstrated a hemispheric sensitivity within the frontoparietal attentional networks in terms of capturing within-proband differences across Biotypes.

In general, for Biotypes, the degree of connectivity abnormalities was much higher for B1/B2 than for B3. In addition, none of the networks showed significant differences between B1 and B2. Although it is true that rs-fMRI connectivity differed only between B1 and B3 and B2 and B3, it is important to note that this was not a global phenomenon and was restricted to only select networks, which might have implications for future studies examining data using neurobiological classifications such as Biotypes. In other words, not all networks sufficiently discriminated B3 from B1/B2. Further, although B3 had lesser cognitive and sensorimotor abnormalities than B1 and B2, it is important to note that this group still contained individuals with severe psychosis (e.g., although there are mild decrements in severity from B1 to B3, symptomatic level scores were quite abnormal in all three groups on the Positive and Negative Syndrome Scale and the Birchwood Social Functioning Scale (6). Statistically, within-proband effects were different between Biotypes and DSM in the cuneus-occipital network and demonstrated strong trends in two of three DMNs. These differential patterns of rs-fMRI manifestations across Biotypes might suggest improved sensitivity to subpsychosis-related deficits in these specific networks when clustered based on neurobiology rather than pure symptomology.

Our results also indicate that abnormalities seen in relatives across different networks were selective and varying. This might have important consequences in terms of future studies using recruitment strategies (i.e., choosing which groups to investigate) as well as what brain networks to study from an etiological and genetic risk point of view. For example, our results indicate that recruitment strategies concentrated on targeting probands/relatives similar to B1 and B2 and evaluating the biological connectivity of the rFPN among those subjects might result in a more fruitful endeavor from an endophenotypic perspective versus say recruiting subjects similar to B3 and studying the connectivity of inferior DMN.

The present study also illuminates some challenges associated with validating a new diagnostic nosology. Multiple factors, including epigenetics, pleiotropy, and variable expressivity, might influence rs-fMRI connectivity, making it difficult to parse complex psychosis heterogeneity using a single phenotype. Nevertheless, how biomarkers sort differentially with traditional versus biologically determined illness entities is important to determine. Our present results may provide specific and novel clues concerning the neurobiology of psychoses, which may be a promising starting point for disentangling their neurobiological and etiological heterogeneity going forward.

Associations of ICA Deficits in Psychosis Probands With Neurophysiological Profiles

It was notable that our secondary analysis showed that all intrinsic network differences except for those in the aDMN and the temporoparietal network correlated with the cognitive control factor, and none were associated with sensorimotor reactivity. Further, we were able to validate that the difference in slopes for cognitive control/sensorimotor versus network connectivity between the various probands subgroups in DSM and Biotypes was not different, suggesting that both categorizations equally conformed to the dimensional approach espoused in the Research Domain Criteria framework. In the original analysis the sensorimotor reactivity component mainly comprised of electrophysiological and eye movement measures. Our results therefore suggest that deficits captured by resting fMRI track more closely with the cognitive domain and might be complementary to those captured using electrophysiological recordings. However, it is also important to point out that the largest correlation (r = .17) noted here for cognitive control was weak but significant, accounting for about 3% of the variance in network connectivity.

Our study had the following limitations. 1) Significant interaction of rs-fMRI connectivity with medications have been shown in some studies (42) but not all (27); therefore, medication status might have affected signal response within groups. Although we tried to partially address this issue by using unmedicated relatives and computing associations with available medication data, this could further be disentangled by performing a similar analysis in antipsychotic-naive or -off patients in future studies. 2) We only examined within-network connectivity, but future studies could also examine between-network ICA connectivity (13,39). 3) Acquiring a complex set of multivariate phenotypes for Biotype creation in clinical practice could be challenging. 4) Finally, similar to our recent work (14), investigations examining genetic links in Biotypes across all intrinsic networks could be helpful in resolving at least a part of the etiological/pathophysiological puzzle that plagues clinical psychosis classification.

Conclusions

In summary, we were able to validate psychosis-related deficits in nine large-scale data-driven networks. Further, select networks benefitted from using a novel biologically driven grouping system in regard to connectivity deficits across the psychosis spectrum. Psychosis-related rs-fMRI connectivity deficits seemed to track more closely with cognitive profile metrics over sensorimotor scores, and this might have important implications for developing targeted treatment plans for different psychosis classes.

Supplementary Material

Supplementary material cited in this article is available online at http://dx.doi.org/10.1016/j.bpsc.2016.07.001.