Ninety-two adult patients with a diagnosis of MDD (average age: 32.79 ± 12.21 years old; 34 males and 58 females) according to the DSM-IV criteria (22–25) and 38 healthy adults (average age: 37.12 ± 12.61 years old, 13 males and 25 females) with no history of neurological or psychiatric illness participated in this study. Details of inclusion criteria were provided in Supplementary material.

MDD patients were further divided into three subgroups according to their levels of suicide risk, which were estimated by the third item on the HAMD-17 (13, 26, 27). The low suicide risk group [not suicidal (NS) group, n = 30] consisted of patients with scores ≤1, indicating the absence of suicidal ideation or suicide attempts within last 2 weeks; the high suicide risk group (n = 62) consisted of patients with scores of 2, 3, or 4. The high suicide risk group was then divided into the suicide ideation (SI, n = 30) group (with scores of 2 or 3), indicating the presence of suicidal ideation but no suicide attempts within last 2 weeks and the suicide attempt (SA, n = 32) group (with scores of 4), indicating a history of suicide attempts within last 2 weeks.

The local Research Ethics Committee approved this study, and written informed consent was obtained from each participant before the study.

All of the structural and rs-fMRI images were acquired on a 3.0 Tesla MR scanner (GE Discovery 750) with an eight-channel head coil. Data were preprocessed using DPABI v5.1¹ (28) following previous studies (20, 29). Detailed imaging parameters and data preprocessing were provided in Supplementary materials.

Connectome gradient analysis had been developed to analyze the cortical hierarchy at the systems level, which anchors the unimodal sensory networks on the one end and the transmodal network on the other end (30, 31) and supports the transition from perception to more abstract cognitive functions (31, 32). Gradient mapping provides us a simplified representation in terms of main dimensions to characterize the macroscale cortical hierarchy (16, 31). The connectome gradient analysis was processed using the BrainSpace toolbox ² (33). More details can be found in the Supplementary materials and a previous study (20).

Stepwise functional connectivity (SFC) analysis is a graph theory-based method that counts the number of defined seeds to other brain regions in a given length of connectivity distance (19–21). More details for stepwise connectivity analysis can be found in the Supplementary materials.

No significant differences were found in education among healthy controls and MDD subgroups (one-way ANOVA, F value = 1.75, p = 0.160). However, sex and age differed significantly among them (sex, Chi-square test, χ² value =11.56, p < 0.05; age, one-way ANOVA, F value = 4.45, p < 0.05). Post hoc analysis revealed that the SA group had a higher proportion of women than the NS group and SI group (p < 0.05), while the HC group and NS group were older on average than the SI group and SA group (p < 0.05). Among MDD patients, the NS group had a longer duration of depressive episodes than the SI group and SA group (p < 0.05), while the SA group had a higher HAMD score than the NS group and SI group (p < 0.05). The demographic and clinical characteristics information is provided in Table 1.

To maximize interpretability, we only retained principal gradients explaining more than 10% of connectome variance in the dataset following previous studies (20, 21). The first and second principal gradients explained 11.33 and 10.18% of the connectome variance in our dataset, respectively (Figure 1A). The connectome gradient maps for four groups are shown in Figure 1B. The spatial patterns of connectome gradient maps were outstandingly similar among healthy controls and the three MDD subgroups, with the unimodal sensory networks (VIS, AUD, SMN) on the one end and the transmodal network [default mode network, (DMN)] on the other end.

Visual inspection of the global histogram of the principal gradients indicated that the extremes of the unimodal-to-transmodal gradient were contracted uniquely in MDD patients with diverse suicide risks compared with healthy controls. The SA group had a prominent compression from the sensorimotor system of the cortical hierarchy (gradient 1 and gradient 2), the SI group had a prominent compression from higher-level systems such as the frontoparietal network (FPN) and DMN (gradient 2), and the NS group had a compression from both the sensorimotor system and higher-level systems (gradient 2), although it was less prominent than that of patients with SI or SA (Figures 1C,D and Supplementary Figures S1A,B). Group comparison indicated that SI had a higher gradient score in the SMN (e.g., Cingulum_Mid_L) than in the HC (Gradient 1), while SI had a higher gradient score in the DMN (e.g., SupraMarginal_R) and a lower gradient score in the limbic network (e.g., ParaHippocampal_L) than in the HC (Gradient 2) (Figure 1F and Table 2). Compared to NS, SI had a higher gradient score in the VIS (e.g., cuneus), and SA had a lower gradient score in the DMN (e.g., medial frontal gyrus, precuneus) (Supplementary Figure S1D and Table 2). Of note, the whole-brain network allocations were based on the seven networks template by Yeo et al. (35) (Figure 1E).

Similar spatial transition patterns of stepwise connectivity were observed among healthy controls and the three MDD subgroups (Figure 2). Unimodal sensory-driven networks reached and stabilized at the transmodal networks (DMN) through seven steps (Figure 2).

Relative to HC, NS patients had an increased SFC degree of the DMN (e.g., precuneus, SupraMarginal_R) at steps 1–3 and increased SFC degree of VIS (e.g., lingual gyrus) at steps 4–7; SI patients had a decreased SFC degree of the limbic network (inferior frontal gyrus and SupraMarginal_R) as well as increased SFC degree of the DMN (precuneus); SA patients had an increased SFC degree of the VIS (e.g., cuneus, Calcarine_L) at steps 3–7, which also extended to the DMN (e.g., precuneus); a transient decrease in the SFC degree of the limbic network (e.g., inferior frontal gyrus) was also observed in step 2. More details are provided in Figure 3 and Table 3.

Pairwise comparison of patients with diverse suicide risks revealed that NS patients had increased SFC degree of the VIS (e.g., Lingual_L, Lingual_R) and decreased SFC degree of the SMN (e.g., Postcentral_L, Postcentral_R) at step 1 when compared to SI patients. NS had decreased SFC degree of the SMN (e.g., Postcentral_L) at step 1 when compared to SA patients. SI patients had increased SFC degree of the DMN (e.g., Frontal_Sup_Medial_L) at step 5 when compared to SA patients. More details are provided in Supplementary Figure S2 and Table 3.

In order to investigate the potential influences of global signal regression (GSR) on gradient analysis and SFC analysis. We further conducted a control analysis without GSR in the functional MRI pre-processing. Furthermore, in order to minimize the risk of Type-I errors or false positives, we re-analyzed the data using a multiple comparison correction threshold of voxel-level p < 0.01 and cluster-level p < 0.008. Similar results were found between the validation analysis and primary analysis. More details are provided in Supplementary Figures S3–S9 and Supplementary Tables S1, S2.

Compared to HC, MDD patients with SA exhibited gradient compression mainly in the primary network (visual network) in both the first and second gradient. Furthermore, the SFC analysis revealed information transition dysfunction across all of the link-step distances (steps 2–7), which also mainly involved the visual networks, as well as the DMN. The visual network plays important roles in information feedforward and feedback (43, 44). The contact visual network is helpful for resisting disturbance and making a correct response during tasks (43, 44), while a compromised visual network can lead to misconceptions and incorrect decisions in MDD patients with SA (45, 46). Consistent with present results, Holmes et al. indicated that suicide-related imagery, which resulted from the disruption of the visual network, was an essential factor causing suicidal behavior (47, 48). Wagner et al. found SA exhibited attenuated regional brain function as well as decreased functional connectivity involved in the occipital regions (visual related network), although they also revealed other brain regions abnormalities in SA group when compared to health control and MDD patients without SA (13, 49). Stumps et al. found the connection between the visual network and amygdala to be significantly altered in posttraumatic stress disorder patients with SA but not in the psychiatric control group, and they concluded that the disruption of visual network was an important factor causing SA (9). A prior meta-analysis also demonstrated that the hypoactivation of the visual network (e.g., left superior occipital gyrus) was closely related to suicidal thoughts and behaviors (50). Our findings complement previous literature with that MDD with SA possessed abnormal cortical hierarchy organization within the visual network, this disorganization would lead to SA be more concentrated in negative imagery and be more sensitive with negative imagery than HC (48), which would directly influence the emotional system (51) and finally lead to suicidal behavior (52).

Compared to HC, MDD with SI exhibited gradient compression mainly in the positive pole, where the transmodal network (DMN) is located. The SFC analysis revealed similar findings, which indicated that the information transition dysfunction was also mainly involved in the DMN. The DMN plays a critical role in ruminative processes (53) and eliminating negative information (54); the abnormal cortical hierarchy organization of the DMN observed in SI patients may be responsible for the failure to eliminate negative information (54). Once depressed patients are deeply embedded in negative emotions, they are more likely to have suicidal thoughts (55). The DMN had also been widely recognized as an important node correlating with the severity of SI (46, 56, 57). Schreiner et al. found that an increased connection between the precuneus and primary sensorimotor regions as well as a decreased connection between the posterior cingulate cortex (PCC) and visual attention regions was associated with greater severity of SI in MDD adolescents (58). In MDD adults with SI, the severity of SI was closely correlated with an ongoing low-frequency BOLD signal in the dorsal PCC, ventral PCC, and dorsal anterior cingulate cortex, as well as the decreased connections among these three regions (59). Thus, it is plausible that abnormal cortical hierarchy organization of the DMN may be an important feature for MDD patients with SI.

Compared to HC, MDD with NS exhibited gradient compression in both the unimodal and transmodal networks, although it was less severe than that of MDD with high suicide risk levels (Supplementary Figures S1A,B). SFC analysis revealed that MDD patients with NS had unique disruption during information transition from unimodal to transmodal networks, and the dysfunction of transmodal networks (DMN) occurred at the early link-step distances, while the dysfunction of the visual network occurred at medium–late link-step distances. Although the significance of dysfunction from unimodal to transmodal networks is not clear, our findings were in line with those of recent studies. Using a large cohort of 2,414 participants (1,276 patients with MDD and 1,138 healthy controls), Xia et al. also demonstrated that patients with MDD exhibited abnormal global topography of the principal primary-to-transmodal gradient, as indicated by a reduced explanation ratio, gradient range, and gradient variation (16). Moreover, they found that focal alterations of gradient scores were mostly in the primary systems involved in sensory processing and in the transmodal systems implicated in high-order cognition (16). Our results also agreed with a recent effective connectivity study by Ray et al. using dynamic causal modeling, also demonstrating that depressed patients had hierarchy organization disruption, including compromises in both the backward connections (high-level network to primary network) and forward connections (primary network to high-level network), which were correlated with depressive symptomatology (17). This widespread dysfunction from the unimodal to transmodal networks would lead to failure in the processing of abstract representations (60) and information integration (61), which are characteristics of MDD (17).