Introduction
Schizophrenia (SCZ) is a serious mental disorder characterized by substantial inter-individual variability in long-term clinical outcomes (Molstrom et al., Reference Molstrom, Nordgaard, Urfer-Parnas, Handest, Berge and Henriksen2022). Its onset is thought to be determined by complex interactions between genetic and environmental factors (van Os, Rutten, & Poulton, Reference van Os, Rutten and Poulton2008), and extensive neuroimaging research has reported widespread structural and functional brain abnormalities in patients with SCZ (Dong et al., Reference Dong, Wang, Chang, Luo and Yao2018; Shepherd et al., Reference Shepherd, Laurens, Matheson, Carr and Green2012). Despite these advances, the relationship between neuroimaging findings and clinical symptoms remains unclear. In particular, although resting-state functional MRI studies have accumulated rapidly, associations between alterations in intra- or inter-network functional connectivity and clinical symptoms have been inconsistent, limiting their explanatory value for the diverse clinical manifestations of the disorder (Voineskos et al., Reference Voineskos, Hawco, Neufeld, Turner, Ameis, Anticevic, Buchanan, Cadenhead, Dazzan, Dickie, Gallucci, Lahti, Malhotra, Ongur, Lencz, Sarpal and Oliver2024).
Higher-order mental functions such as cognition and emotion depend not on isolated brain regions but on the coordinated transmission and integration of neural signals across large-scale brain networks (Park & Friston, Reference Park and Friston2013). Importantly, neural signal propagation does not occur randomly; instead, it is constrained by the large-scale organization of cortical networks and proceeds along preferred directions (Mesulam, Reference Mesulam2012). Accordingly, understanding SCZ requires moving beyond static measures of connectivity strength toward examining how neural signals propagate within a hierarchically organized network architecture and how this propagation relates to clinical symptomatology (Bassett et al., Reference Bassett, Bullmore, Verchinski, Mattay, Weinberger and Meyer-Lindenberg2008).
Neural signal transmission can be broadly conceptualized as comprising top–down and bottom–up processes (Friston, Reference Friston2005). Top–down processing refers to the influence of higher-order associative regions on lower-level sensory processing, supporting functions such as contextual integration, expectation, and cognitive control, and is primarily associated with frontoparietal (Corbetta & Shulman, Reference Corbetta and Shulman2002) and default mode networks (Greicius, Krasnow, Reiss, & Menon, Reference Greicius, Krasnow, Reiss and Menon2003; Raichle et al., Reference Raichle, MacLeod, Snyder, Powers, Gusnard and Shulman2001). In contrast, bottom–up processing involves the flow of sensory information from primary sensory cortices toward higher-order associative regions and is mainly mediated by visual, auditory, and somatosensory networks (Felleman & Van Essen, Reference Felleman and Van Essen1991; Raichle, Reference Raichle2010). In SCZ, converging evidence suggests abnormalities in both directions of information flow, including disturbances in early sensory processing (Javitt & Freedman, Reference Javitt and Freedman2015; Rissling et al., Reference Rissling, Braff, Swerdlow, Hellemann, Rassovsky, Sprock, Pela and Light2012), impaired contextual modulation (Dias, Butler, Hoptman, & Javitt, Reference Dias, Butler, Hoptman and Javitt2011; Seymour et al., Reference Seymour, Stein, Sanders, Guggenmos, Theophil and Sterzer2013), altered predictive processing (Fletcher & Frith, Reference Fletcher and Frith2009; Sterzer et al., Reference Sterzer, Adams, Fletcher, Frith, Lawrie, Muckli, Petrovic, Uhlhaas, Voss and Corlett2018), and executive dysfunction (Lesh, Niendam, Minzenberg, & Carter, Reference Lesh, Niendam, Minzenberg and Carter2011). These bidirectional abnormalities unfold within a hierarchical cortical organization that constrains the direction, timing, and efficiency of information transfer across networks.
Prior studies have demonstrated that the cerebral cortex is organized along a hierarchical axis extending from unimodal sensory and motor regions to transmodal association cortices (Margulies et al., Reference Margulies, Ghosh, Goulas, Falkiewicz, Huntenburg, Langs, Bezgin, Eickhoff, Castellanos, Petrides, Jefferies and Smallwood2016). This hierarchical architecture supports progressive integration of information, whereby incoming sensory signals are successively transformed and integrated with internally generated, abstract, and self-referential processes (Mesulam, Reference Mesulam2012). Recent methodological advances have enabled the empirical characterization of large-scale cortical hierarchy using functional connectivity gradient analysis. When applied to resting-state functional MRI data, gradient analysis summarizes whole-brain functional connectivity patterns into continuous, low-dimensional axes, providing a compact topographical representation of hierarchical network organization. In healthy individuals, the principal functional gradient reliably captures a unimodal–transmodal axis spanning primary sensory cortices to default mode regions, closely mirroring microstructural, genetic, and developmental features of the cortex (Dong, Margulies, Zuo, & Holmes, Reference Dong, Margulies, Zuo and Holmes2021; Haueis, Reference Haueis2021; Huntenburg, Bazin, & Margulies, Reference Huntenburg, Bazin and Margulies2018). Importantly, gradient representations may provide a multivariate framework to quantify large-scale brain organization while accommodating the substantial inter-individual heterogeneity characteristic of SCZ.
Applying this framework to SCZ, prior studies have reported alterations in functional connectivity gradients, most notably compression of the unimodal–transmodal gradient (Dong et al., Reference Dong, Luo, Guell, Wang, He, Duan, Eickhoff and Yao2020, Reference Dong, Yao, Wang, Hong, Genon, Xin, Jung, He, Chang and Duan2023; Xiang, Ma, Chen, & Cheng, Reference Xiang, Ma, Chen and Cheng2025). In patients with SCZ, reduced differentiation between sensory and transmodal networks is reflected in relative positional shifts of visual, limbic, frontoparietal, and default mode networks (Yu et al., Reference Yu, Hawco, Bassman, Oliver, Argyelan, Gold, Tang, Foussias, Buchanan, Malhotra, Ameis, Voineskos and Dickie2025). These gradient alterations have been linked to genes involved in synaptic signaling and neurodevelopment (Xiang et al., Reference Xiang, Ma, Chen and Cheng2025; Yao et al., Reference Yao, Luo, Li, Feng, Liu and Xu2024), as well as to cognitive performance, symptom severity, and, in some cases, treatment response (Holmes et al., Reference Holmes, Levi, Chen, Chopra, Aquino, Pang and Fornito2023; Yao et al., Reference Yao, Luo, Li, Feng, Liu and Xu2024; Yu et al., Reference Yu, Hawco, Bassman, Oliver, Argyelan, Gold, Tang, Foussias, Buchanan, Malhotra, Ameis, Voineskos and Dickie2025). Studies encompassing different disease stages further suggest that early-stage SCZ is associated with more global gradient alterations, whereas chronic disease is characterized by more focal disruptions in secondary gradients involving attention, visual, frontoparietal, and limbic networks (Holmes et al., Reference Holmes, Levi, Chen, Chopra, Aquino, Pang and Fornito2023; Yao et al., Reference Yao, Luo, Li, Feng, Liu and Xu2024). Despite these insights, existing gradient studies have primarily focused on static shifts in hierarchical positioning or overall gradient compression. Consequently, a critical question remains unresolved: how alterations in cortical hierarchy translate into disrupted neural communication and information flow across large-scale brain networks.
Stepwise functional connectivity (SFC) analysis offers a complementary approach for modeling neural signal transmission by explicitly quantifying both direct and indirect pathways through which signals traverse the connectome (Sepulcre et al., Reference Sepulcre, Sabuncu, Yeo, Liu and Johnson2012). In healthy individuals, SFC reveals a consistent, orderly sequence of signal propagation from primary sensory regions through intermediate attentional systems to transmodal association cortices (Huntenburg et al., Reference Huntenburg, Bazin and Margulies2018). Rather than redundantly replicating gradient findings, this sequential propagation pattern provides a functional instantiation of hierarchical organization by illustrating how the macroscale architecture of the brain constrains both the routes and the efficiency of signal transmission. Together, these observations suggest that cortical gradients not only describe the geometric organization of brain networks but also impose constraints on hierarchical signal propagation. However, whether alterations in functional gradients are associated with disrupted signal propagation in SCZ has not been systematically examined.
A recent study integrating functional connectivity gradient analysis and SFC showed a compressed sensorimotor-to-transmodal cortical hierarchy in SCZ, accompanied by aberrant connectome-based propagation from sensory to higher-order association systems (Dong et al., Reference Dong, Yao, Wang, Hong, Genon, Xin, Jung, He, Chang and Duan2023). This work provides important evidence that hierarchical disorganization may disrupt the integration of bottom–up sensory information with attentional and executive networks. However, the study did not explicitly quantify how hierarchical alterations translate into network-level propagation delays or examine their associations with clinically meaningful phenotypes. It therefore remains unclear how large-scale cortical hierarchy and dynamic connectome-based propagation jointly contribute to SCZ and whether these mechanisms are linked to clinical symptoms.
In the present study, we addressed these gaps by integrating whole-brain functional connectivity gradient analysis with SFC in a large, multi-cohort sample. Specifically, we aimed to (i) characterize alterations in cortical hierarchy at the network and regional levels, (ii) determine whether these hierarchical alterations are associated with delayed connectome-based propagation across predefined hierarchical levels, and (iii) examine how these network-level mechanisms relate to clinical symptoms.
Methods
Study population
A total of 966 participants were recruited from multiple publicly available and institutional datasets, including the Center for Biomedical Research Excellence (COBRE), Neuromorphometry by Computer Algorithm Chicago (NMorphCH), and two Korean datasets from Asan Medical Center (AMC) and Jeonbuk National University Hospital (JBNU). Psychiatric diagnoses were established according to the DSM-IV or DSM-5 criteria. The included cohorts differed in sample composition, illness duration, scanner and acquisition protocol, and the availability of clinical and cognitive measures. AMC 1 included both healthy controls and patients with recent-onset schizophrenia, whereas AMC 4 included only patients with chronic schizophrenia. JBNU contributed the largest case–control sample, while COBRE and NMorphCH provided publicly available case–control datasets. Clinical and cognitive measures, including the Positive and Negative Syndrome Scale (PANSS) and Full-Scale Intelligence Quotient (FSIQ), as well as antipsychotic dose information, were available for subsets of participants. Cohort-specific details are provided in the Supplementary Material.
All procedures contributing to this work complied with the ethical standards of the relevant national and institutional committees on human experimentation and with the Declaration of Helsinki guidelines (1975) as revised in 2008. The study was approved by the Institutional Review Board of Asan Medical Center (IRB No. 2021-1128).
Image acquisition and processing
MRI acquisition parameters are summarized in Supplementary Table 1. T1-weighted images underwent quality control using the Euler number (Rosen et al., Reference Rosen, Roalf, Ruparel, Blake, Seelaus, Villa, Ciric, Cook, Davatzikos, Elliott, Garcia de la Garza, Gennatas, Quarmley, Schmitt, Shinohara, Tisdall, Craddock, Gur, Gur and Satterthwaite2018) derived from FreeSurfer, yielding 926 participants after the exclusion of 40 outliers (Supplementary Figure 1). Resting-state functional MRI data were preprocessed using SPM12 and the CONN toolbox (v22a) default pipeline, including realignment, slice-timing correction, normalization, and smoothing (8 mm FWHM). Participants exceeding motion thresholds (Heleven & Van Overwalle, Reference Heleven and Van Overwalle2016; Power et al., Reference Power, Barnes, Snyder, Schlaggar and Petersen2012) (mean framewise displacement ≥0.5 mm or invalid scan ratio ≥ 10%) were excluded (n = 106; Supplementary Figure 2), resulting in a final sample of 820 participants (310 healthy controls [HCs] and 510 patients with SCZ; Table 1). Denoising included white matter and ventricular signals, motion parameters, and outlier scans, followed by band-pass filtering (0.008–0.09 Hz). Further details of the preprocessing and denoising procedures are provided in prior publications (Nieto-Castanon, Reference Nieto-Castanon2020).
Demographic and clinical characteristics of the study population
Table 1. Long description
The table is structured with columns for five study sites: A M C 1, A M C 4, C O B R E, J B N U, and N Morph C H, followed by a Total column. Each site is subdivided into H C (healthy control) and S C Z (schizophrenia) groups.
Key data points include:
* Number of subjects: Total H C = 310, Total S C Z = 510.
* Age, mean (S D): Total H C = 36.3 (11.8), Total S C Z = 35.7 (11.8).
* Male, n (%): Total H C = 176 (56.8%), Total S C Z = 268 (52.6%).
* Illness duration, mean (S D): Total S C Z = 9.7 (9.8) years.
* P A N S S positive, mean (S D): Total S C Z = 14.0 (6.0).
* P A N S S negative, mean (S D): Total S C Z = 13.5 (6.6).
* F S I Q, mean (S D): Total H C = 113.3 (13.0), Total S C Z = 90.4 (19.2).
* Daily O L Z equivalent dose, mean (S D): Total S C Z = 17.5 (11.9).
Note that A M C 4 has 0 H C subjects. Some data points for J B N U and N Morph C H are listed as N A (not available).
Note: AMC, Asan Medical Center; COBRE, Center for Biomedical Research Excellence; JBNU, Jeonbuk National University Hospital; NMorphCH, Neuromorphometry by Computer Algorithm Chicago; HC, healthy control; SCZ, schizophrenia; PANSS, Positive and Negative Syndrome Scale; FSIQ, Full-Scale Intelligence Quotient; OLZ, olanzapine; SD, standard deviation; NA, not available.
a Information on illness duration, PANSS scores, FSIQ, and antipsychotic medications was available for 483, 500, 237, and 112 individuals, respectively.
Functional connectivity gradient analysis
Resting-state functional MRI data were parcellated into 1,000 cortical regions using the Schaefer atlas (Schaefer et al., Reference Schaefer, Kong, Gordon, Laumann, Zuo, Holmes, Eickhoff and Yeo2018). For each participant, functional connectivity matrices were constructed using Fisher z-transformed Pearson correlations. Scanner-related batch effects were harmonized using neuroCombat (Supplementary Figure 3) (Fortin et al., Reference Fortin, Cullen, Sheline, Taylor, Aselcioglu, Cook, Adams, Cooper, Fava, McGrath, McInnis, Phillips, Trivedi, Weissman and Shinohara2018). Matrices were subjected to a 10% proportional threshold per row (Margulies et al., Reference Margulies, Ghosh, Goulas, Falkiewicz, Huntenburg, Langs, Bezgin, Eickhoff, Castellanos, Petrides, Jefferies and Smallwood2016), and cosine similarity matrices were computed.
Diffusion map embedding was applied to derive functional gradients using BrainSpace (α = 0.5) (Coifman et al., Reference Coifman, Lafon, Lee, Maggioni, Nadler, Warner and Zucker2005; Vos de Wael et al., Reference Vos de Wael, Benkarim, Paquola, Lariviere, Royer, Tavakol, Xu, Hong, Langs, Valk, Misic, Milham, Margulies, Smallwood and Bernhardt2020). Individual gradients were aligned to a group template using Procrustes rotation (Margulies et al., Reference Margulies, Ghosh, Goulas, Falkiewicz, Huntenburg, Langs, Bezgin, Eickhoff, Castellanos, Petrides, Jefferies and Smallwood2016). The variance explained by each gradient is shown in Supplementary Figure 4. Subsequent analyses focused on the first two gradients (principal gradient [G1] and second gradient [G2]), which have been extensively examined in previous studies (Hong et al., Reference Hong, Xu, Nikolaidis, Smallwood, Margulies, Bernhardt, Vogelstein and Milham2020; Margulies et al., Reference Margulies, Ghosh, Goulas, Falkiewicz, Huntenburg, Langs, Bezgin, Eickhoff, Castellanos, Petrides, Jefferies and Smallwood2016). Distributional differences in gradient values were assessed using Kolmogorov–Smirnov tests, with pole-specific analyses defined relative to zero. Gradient compression was quantified by correlating HC gradient values with between-group differences (SCZ − HC). Network-level analyses were conducted using Yeo’s seven-network parcellation and regional analyses across all parcels, adjusting for age and sex. Multiple comparisons were corrected using the Benjamini–Hochberg false discovery rate (FDR), with statistical significance defined as FDR p < 0.05. Statistical analyses were performed using R software (version 4.2.1; R Development Core Team, Vienna, Austria).
Stepwise functional connectivity analysis
SFC analysis was performed following a previously established protocol (Hong et al., Reference Hong, de Wael, Bethlehem, Lariviere, Paquola, Valk, Milham, Di Martino, Margulies, Smallwood and Bernhardt2019). Seeds were defined in the primary sensory cortices, including the visual (V1; MNI ± 14, −78, 8), auditory (A1; −54/58, −14, 8), and somatosensory cortices (S1; −42/38, −29, 65). For each participant, SFC degree at each step was computed by iterative multiplication of a binarized, thresholded functional connectivity matrix, following the findwalks.m implementation in the Brain Connectivity Toolbox (Rubinov, Kötter, Hagmann, & Sporns, Reference Rubinov, Kötter, Hagmann and Sporns2009). SFC maps were Z-scored across the whole brain at each step to account for exponential increases in path counts. A two-step group-guided strategy was applied as previously described (Hong et al., Reference Hong, de Wael, Bethlehem, Lariviere, Paquola, Valk, Milham, Di Martino, Margulies, Smallwood and Bernhardt2019). Group-level SFC templates were generated by averaging the raw SFC trajectories within each diagnostic group. During reconstruction of individual SFC maps, negative values were set to zero before element-wise multiplication with the corresponding group-level template at each step. Connectome-based propagation was tracked over 100 steps. Cortical networks were assigned to four hierarchical levels based on a prior study (Margulies et al., Reference Margulies, Ghosh, Goulas, Falkiewicz, Huntenburg, Langs, Bezgin, Eickhoff, Castellanos, Petrides, Jefferies and Smallwood2016): visual and somatomotor (Level 1), dorsal attention and salience/ventral attention (Level 2), control and limbic (Level 3), and default mode network (Level 4). Propagation efficiency was quantified using a step count metric, defined as the iteration at which each level reached 50% of its maximum signal. Group differences in step counts were evaluated using multiple linear regression models with age and sex as covariates, with FDR correction applied across the four hierarchical levels.
Clinical associations and medication analysis
Clinical variables included the PANSS positive and negative subscale and FSIQ scores. Associations between step count at each hierarchical level and clinical variables were examined using multiple linear regression models, adjusting for age and sex, with FDR correction applied across PANSS positive, PANSS negative, and FSIQ. Analyses were performed in the full patient sample and repeated after stratification into early disease (<5 years) and chronic disease (≥5 years) subgroups.
Mediation analyses were conducted using the pingouin library in Python, with network-level gradient values as the independent variable, step count as the mediator, and clinical variables as the dependent variables, adjusting for age and sex. Indirect effects were assessed using 5,000 bootstrap iterations, with significance defined by a 95% bias-corrected confidence interval (CI) excluding zero.
Robustness analyses
The robustness of the main findings was assessed using several validation analyses: (1) alternative connectivity thresholds (5%, 15%, and 20%); (2) mixed-effects linear models with cohort specified as a random effect and age and sex as fixed covariates; (3) group comparisons additionally adjusted for mean framewise displacement; (4) leave-one-cohort-out analyses excluding each cohort in turn; and (5) analyses restricted to the largest single cohort (JBNU). For both leave-one-cohort-out validations and the single-cohort, all primary analyses were repeated.
Results
Functional gradient alterations in schizophrenia
Gradient decomposition recapitulated the canonical cortical hierarchy in both groups (Figure 1a). The principal gradient (G1) explained 29.4% of the variance and spanned a unimodal–transmodal axis, whereas the second gradient (G2; 19.2%) separated visual from nonvisual systems.
Functional gradient alterations in schizophrenia. (a) The first two functional connectivity gradients (G1 and G2) illustrating the canonical cortical hierarchy, with G1 representing a unimodal–transmodal axis and G2 differentiating visual from nonvisual networks. (b) No significant group differences were observed in the global range of G1 or G2. (c) Patients with schizophrenia showed a density shift toward lower G1 values. (d) Network-level analysis revealed significantly reduced G1 values in the dorsal attention network in patients with schizophrenia. (e) Regional G1 alterations were predominantly localized to dorsal attention and somatomotor regions.
Figure 1. Long description
Panel A shows brain surface maps for H C and S C Z groups across Gradient 1 and Gradient 2 for left and right hemispheres. A color scale from negative 15.0 blue to positive 15.0 red indicates gradient values. Panel B contains two scatter plots of Gradient 1 on the x-axis versus Gradient 2 on the y-axis, with data points colored by seven networks: Vis, Som Mot, Dors Attn, Sal Vent Attn, Limbic, Cont, and Default. Panel C is a histogram of mean Gradient 1 values showing density for H C in blue and S C Z in red, with S C Z showing a higher density at lower G 1 values. Panel D is a bar chart of network-based analysis for G 1. It shows significantly reduced values in the Dors Attn network for S C Z compared to H C with an F D R p-value of 0.006. Panel E displays region-based analysis with brain maps highlighting specific areas in red and blue, accompanied by a table of R O I coordinates, t-values, and F D R p-values. Significant regions include Som Mot and Dors Attn areas in both hemispheres.
Global extrema of G1 and G2 did not differ between groups (Supplementary Table 2; Figure 1b). Distribution analyses revealed selective alterations in G1. Although the overall G1 distribution did not differ (D = 0.048, p = 0.200), density shifts toward lower G1 values were observed in SCZ (Figure 1c). Pole-specific analyses showed significant deviations at the sensorimotor pole (D = 0.087, p = 0.030) but not at the transmodal pole. G1 compression was further indicated by a negative correlation between HC G1 values and between-group differences (r = −0.293, p < 0.001), driven by the sensorimotor pole (r = −0.412, p < 0.001; Supplementary Figure 5). Since no group differences were observed for G2 (D = 0.032, p = 0.685; Supplementary Figure 6), subsequent analyses focused on G1.
The network-level analysis demonstrated topologically selective alterations. Mean G1 values were significantly reduced in the DAN in SCZ compared with HC (t = −3.366, FDR p = 0.006; Figure 1d), with no other networks surviving correction (Supplementary Table 3). Region-wise analyses showed spatial clustering of G1 alterations predominantly within dorsal attention and somatomotor regions (Figure 1e), indicating a positive shift in G1 values within the somatomotor network and a negative shift within the DAN in patients with SCZ relative to controls.
Delayed connectome-based propagation in schizophrenia
Step count analysis demonstrated significantly increased steps required to reach higher hierarchical levels in SCZ, with robust group differences at Level 3 (t = 3.050, FDR p = 0.005) and Level 4 (t = 11.796, FDR p < 0.001), while no group differences were observed at lower levels (Figure 2a, 2b; Supplementary Table 6). These findings indicate delayed transitions from lower to higher hierarchical levels in SCZ despite preserved propagation within Level 1 and Level 2. Figure 2c illustrates a relatively orderly pattern of connectome-based propagation in HCs, whereas propagation toward higher-order association networks is delayed in SCZ.
Hierarchical signal propagation revealed by stepwise functional connectivity analysis. (a, b) Step count analysis showing the number of steps required to reach each hierarchical level. Patients with schizophrenia exhibited significantly increased step counts to higher hierarchical levels (Level 3 and Level 4), while no group differences were observed at lower levels (Level 1 and Level 2). (c) Stepwise functional connectivity maps illustrating hierarchical signal propagation. Orderly progression from unimodal to transmodal networks was observed in healthy controls, but delayed propagation toward higher-order association networks was noticed in schizophrenia.
Figure 2. Long description
Panel A is a line graph of Normalized S F C versus Number of steps (0 to 40). Four colored curves represent hierarchical levels: L 1 (red), L 2 (orange), L 3 (green), and L 4 (blue). Circles represent H C and triangles represent S C Z. Vertical dashed lines indicate that S C Z peaks occur at higher step counts than H C for L 3 and L 4, marked with double and triple asterisks respectively.
Panel B is a violin plot showing the Number of steps across Functional hierarchy levels L 1 through L 4. H C (teal) and S C Z (salmon) distributions are side-by-side. At L 1 and L 2, distributions are similar. At L 3 and L 4, the S C Z group shows a significantly higher mean and wider distribution of steps compared to H C.
Panel C displays brain surface maps for H C and S C Z across five rows representing Step 0, Step 5, Step 10, Step 15, and Step 20. Each row shows lateral and medial views of the left and right hemispheres. An orange color scale indicates S F C intensity from 0.00 to 3.00. In H C, the signal propagates rapidly from sensory areas to association cortex by Step 15. In S C Z, the propagation is visibly delayed, with lower intensity in higher-order regions at the same step intervals.
Relationships among functional gradients, connectome-based propagation, and clinical variables
Within the DAN, lower G1 values were associated with increased step counts to Level 4 (β = −0.674, p < 0.001) but not to Level 3 (β = 0.028, p = 0.848). Associations between step count and clinical measures were dependent on illness duration (Supplementary Table 9). In patients with an illness duration <5 years, higher step counts to Level 3 were associated with more severe negative symptoms (Figure 3a; β = 0.150, FDR p = 0.040), while delays to Level 4 were associated with both negative (Figure 3b; β = 0.172, FDR p = 0.024) and positive symptoms (Figure 3c; β = 0.182, FDR p = 0.024). No significant associations were observed in patients with chronic illness.
Relationships among functional gradients, signal propagation, and clinical variables. (a–c) Associations between step counts to higher hierarchical levels and clinical symptom severity in patients with schizophrenia with illness duration <5 years. Longer step counts to Level 3 were associated with higher PANSS negative scores (a), whereas longer step counts to Level 4 were associated with higher PANSS negative (b) and positive (c) scores. (d) Mediation analysis illustrating that step count to Level 4 significantly mediates the association between dorsal attention network G1 values and PANSS positive and negative symptom severity in patients with schizophrenia with illness duration <5 years. PANSS, Positive and Negative Syndrome Scale.
Figure 3. Long description
The figure is divided into four panels labeled A through D.
Panel A is a scatter plot with orange data points. The x-axis is Steps to L 3 (0 to 40) and the y-axis is P A N S S Negative (5 to 35). A red linear regression line shows a positive correlation. A text box indicates Coef equals 0.150 and F D R p equals 0.040.
Panel B is a scatter plot with orange data points. The x-axis is Steps to L 4 (10 to 50) and the y-axis is P A N S S Negative (5 to 35). A red linear regression line shows a positive correlation. A text box indicates Coef equals 0.172 and F D R p equals 0.024.
Panel C is a scatter plot with teal data points. The x-axis is Steps to L 4 (10 to 50) and the y-axis is P A N S S Positive (5 to 40). A blue linear regression line shows a positive correlation. A text box indicates Coef equals 0.182 and F D R p equals 0.024.
Panel D contains two mediation diagrams.
Diagram (A) P A N S S positive: A triangle of arrows connects three boxes. The bottom-left box G 1 value (Dors Attn) points up to Steps to L 4 with a value of minus 0.862 triple asterisk. Steps to L 4 points down to the bottom-right box P A N S S positive with a value of 0.182 asterisk. A dashed arrow connects the bottom boxes directly with a value of minus 0.011. The Indirect Effect is minus 0.156 asterisk.
Diagram (B) P A N S S negative: Follows the same structure. The arrow from G 1 value to Steps to L 4 is minus 0.862 triple asterisk. The arrow from Steps to L 4 to P A N S S negative is 0.172 asterisk. The direct dashed arrow is 0.034. The Indirect Effect is minus 0.151 asterisk.
Mediation analyses in patients with an illness duration <5 years showed significant indirect effects of DAN G1 values on PANSS positive and negative subscale scores via step count (PANSS positive: −0.156, 95% CI [−0.398, −0.016]; PANSS negative: −0.151, 95% CI [−0.352, −0.021]), with no significant direct effects, indicating full mediation (Figure 3d).
Robustness analyses
The results were robust across multiple validation analyses. Connectome-based propagation patterns varied according to the connectivity threshold (Supplementary Figure 7): a stringent threshold (5%) constrained propagation to lower hierarchical levels (Level 1 and Level 2), whereas less stringent thresholds (15–20%) selectively attenuated propagation to higher-order levels (Level 3 and Level 4), indicating threshold-dependent modulation of hierarchical transitions.
Key group differences remained significant when cohort effects were taken into account using mixed-effects models. Specifically, reduced G1 values in the DAN (t = −3.366, FDR p = 0.006) and increased step counts to Level 3 (t = 3.575, FDR p = 0.001) and Level 4 (t = 11.663, FDR p < 0.001) in SCZ persisted after modeling cohort as a random effect.
Group differences remained robust after additionally adjusting for mean framewise displacement, with persistent reductions in DAN G1 values (t = −3.672, FDR p = 0.002; Supplementary Table 4) and increased step counts to Level 3 (t = 3.468, FDR p = 0.001) and Level 4 (t = 11.617, FDR p < 0.001; Supplementary Table 7) in SCZ.
Leave-one-cohort-out validation analyses indicated that the main findings were not driven by any single cohort. Although the statistical strength varied across models, DAN G1 alterations were generally preserved and group differences in step counts to higher hierarchical levels, particularly Level 4, were largely maintained (Supplementary Tables 5 and 8). Clinical and cognitive associations were more variable, likely reflecting differences in sample composition and data availability across cohorts, but remained broadly consistent with the primary analyses (Supplementary Table 10).
Analyses restricted to the largest single cohort (JBNU) replicated the main findings. Patients with SCZ showed significantly reduced G1 values in the DAN and somatomotor networks (Supplementary Figure 8 and Supplementary Table 11), along with delayed propagation to Level 3 (t = 4.046, p < 0.001) and Level 4 (t = 9.659, p < 0.001) (Supplementary Figure 9 and Supplementary Table 12). In contrast, associations between step counts and clinical variables were attenuated in these analyses, and none of the associations reached statistical significance (Supplementary Figure 10).
Discussion
In this study, we examined alterations in large-scale cortical hierarchy and connectome-based propagation in SCZ by integrating functional connectivity gradient analysis with SFC. Patients with SCZ exhibited selective compression of the principal gradient, primarily localized to the DAN, rather than a global disruption of the unimodal–transmodal organization. In parallel, hierarchical propagation from sensory systems to higher-order association networks was delayed at intermediate-to-transmodal stages, as reflected by increased step counts to Level 3 and Level 4. Among patients with shorter illness duration (<5 years), delayed access to higher hierarchical levels was associated with greater symptom severity, and mediation analyses suggested that connectome-based propagation accounted for the association between DAN hierarchical positioning and the severity of both positive and negative symptoms. Collectively, these findings support the hypothesis that SCZ involves not only alterations in cortical hierarchy but also inefficiencies in translating hierarchical organization into effective large-scale neural communication.
Consistent with previous reports, alterations of the principal unimodal–transmodal gradient were observed in SCZ (Dong et al., Reference Dong, Luo, Guell, Wang, He, Duan, Eickhoff and Yao2020, Reference Dong, Yao, Wang, Hong, Genon, Xin, Jung, He, Chang and Duan2023; Xiang et al., Reference Xiang, Ma, Chen and Cheng2025). These alterations were reflected by an increased representation of lower gradient values in distribution analyses, along with network- and region-specific shifts involving the somatomotor and dorsal attention networks. This pattern suggests that alterations in cortical hierarchy preferentially involve attentional control systems positioned at intermediate levels of the hierarchy rather than reflecting a uniform global collapse. The DAN occupies an intermediate position between early sensory cortices and higher-order control systems and comprises distinct frontoparietal pathways that support the flexible routing of attentional priorities across sensory-driven and internally guided reference frames (Szczepanski et al., Reference Szczepanski, Pinsk, Douglas, Kastner and Saalmann2013). Reduced G1 values within this network may therefore reflect diminished hierarchical differentiation of frontoparietal attentional pathways, potentially resulting in less efficient gating and routing of information from sensory systems toward higher-order association networks. In this context, a previous gradient study reported reduced within-network dispersion in the DAN, as well as in visual, frontoparietal, and limbic networks; however, these effects were observed along the secondary visual-to-sensorimotor gradient rather than along the principal gradient (Holmes et al., Reference Holmes, Levi, Chen, Chopra, Aquino, Pang and Fornito2023).
While functional gradients describe the spatial embedding of networks within the cortical hierarchy, they do not directly capture how signals propagate across this architecture. When gradient analysis was combined with SFC, hierarchical alterations in SCZ were accompanied by delayed propagation toward higher-order association networks. These delays were most evident in transitions from attentional and control systems to limbic and default mode networks. This pattern suggests that early sensory processing may be preserved, whereas the integration of information into higher-order, transmodal systems may be less efficient. This is consistent with evidence indicating that SCZ is characterized by disrupted integration of information across distributed brain systems (Braff, Reference Braff1993; Tseng et al., Reference Tseng, Bossong, Modinos, Chen, McGuire and Allen2015). In addition, the association between reduced DAN G1 values and delayed propagation to Level 4 suggests that variations in the hierarchical positioning of attentional networks may influence the efficiency by which information reaches default mode regions. Collectively, these observations support the notion that cortical gradients provide a structural context that may shape large-scale patterns of signal propagation.
Connectome-based propagation was associated with clinical symptom severity, with delayed propagation linked to greater positive and negative symptoms. This finding accounts for the relationship between DAN hierarchical positioning and clinical measures. These observations are consistent with predictive coding frameworks of psychosis (Harding et al., Reference Harding, Wolpe, Brugger, Navarro, Teufel and Fletcher2024; Sterzer et al., Reference Sterzer, Adams, Fletcher, Frith, Lawrie, Muckli, Petrovic, Uhlhaas, Voss and Corlett2018), which conceptualize SCZ as involving disturbances in hierarchical information integration resulting from altered coordination between sensory evidence and higher-order predictions. Within this framework, delayed propagation to transmodal networks may constrain the timely integration of contextual, self-referential, and value-related information, potentially biasing inference processes toward aberrant prediction errors. These alterations are hypothesized to contribute to aberrant salience attribution, hallucinations, delusional belief formation, and motivational and affective disturbances relevant to negative symptoms. Notably, these associations were evident in patients with a shorter duration of illness but not in those with a longer duration. The absence of significant associations in chronic patients may reflect multiple factors that do not necessarily represent disease-related processes, including chronicity-related compensatory changes and cumulative medication effects. Additionally, a range of covariate-related factors may affect neuroimaging measures, representing an inherent limitation of cross-sectional neuroimaging studies (Lawrie, Reference Lawrie2025).
Methodologically, the present findings illustrate the potential utility of integrating complementary connectomic approaches. Gradient analysis characterizes continuous, multivariate features of cortical organization, whereas SFC provides a framework for examining indirect pathways and sequential patterns of signal propagation. Considered together, these approaches offer complementary perspectives on large-scale brain organization that may facilitate interpretation beyond the capacity of either method alone. However, several limitations should also be acknowledged. First, although data harmonization procedures were applied to mitigate the site-related variability inherent to a multisite study, residual site effects may persist and could have influenced the neuroimaging measures. Second, the potential effects of antipsychotic medication cannot be fully excluded. Medication dosage information was available for only a subset of the sample and was therefore not incorporated into the analyses; moreover, the magnitude and direction of the effects of antipsychotics on neuroimaging markers remain uncertain, warranting cautious interpretation of the results. Third, SFC provides an abstract representation of signal propagation across the connectome and does not directly capture neurophysiological timing, directionality, or causal interactions between brain regions. Fourth, the use of group-specific SFC templates may have influenced the magnitude of between-group step-count differences. Although a common-template approach provides an identical reference across groups, it may also average out diagnostic group-level SFC characteristics. Future studies using an independent normative template may help clarify patient-specific deviations in connectome-based propagation. Finally, the cross-sectional design limits inferences regarding illness progression, and longitudinal studies are necessary to determine whether alterations in hierarchical signal propagation precede symptom emergence or change with treatment and clinical course.
In summary, the present study suggests that SCZ involves selective alterations in cortical hierarchy centered on the DAN, accompanied by delayed connectome-based propagation to higher-order association systems. These propagation delays are associated with clinical symptom severity and appear to account for the relationship between hierarchical organization and symptoms. By linking large-scale cortical hierarchy to patterns of connectome-based propagation and clinical features, the present findings support a view of SCZ characterized by impaired hierarchical integration within the functional connectome.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S0033291726106011.
Funding statement
SWJ was supported by the National Research Foundation (NRF) of Korea grant (grant numbers: 2021R1F1A11057227 and RS-2026-25482642). JL was supported by the NRF of Korea grant (grant numbers: 2012R1A1006514 and 2017R1A1B03032707) and the Korea Health Technology Research and Development Project through the Korea Health Industry Development Institute, funded by the Ministry of Health and Welfare, Republic of Korea (grant number: HI22C0108).
Competing interests
The authors declare no conflicts of interest.
Ethical standards
The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008. The present study was approved by the institutional review board (IRB) of the Asan Medical Center (IRB No. 2021-1128).
Open access
