Association of core brain networks with antipsychotic therapeutic effects in first-episode schizophrenia

Abstract

Elucidating neurobiological mechanisms underlying the heterogeneity of antipsychotic treatment will be of great value for precision medicine in schizophrenia, yet there has been limited progress. We combined static and dynamic functional connectivity (FC) analysis to examine the abnormal communications among core brain networks [default-mode network (DMN), central executive network (CEN), salience network (SN), primary network (PN), and subcortical network (SCN) in clinical subtypes of schizophrenia (responders and nonresponders to antipsychotic monotherapy). Resting-state functional magnetic resonance imaging data were collected from 79 first-episode schizophrenia and 90 healthy controls. All patients received antipsychotic monotherapy for up to 12 weeks and underwent a second scan. We found that significantly reduced static FC in CEN-DMN/SN and SN-SCN were observed in nonresponders after treatment, whereas almost no difference was observed in responders. The nonresponders showed significantly higher dynamic FC in PN-DMN/SN than responders at baseline. Further, the baseline FC in core brain networks were treated as moderators involved in symptom relief and distinguished response subtypes with high classification accuracy. Collectively, the current work highlights the potential of communications among five core brain networks in searching biomarkers of antipsychotic monotherapy response and neuroanatomical subtypes, advancing the understanding of antipsychotic treatment mechanisms in schizophrenia.

Introduction

Schizophrenia is considered a complex psychiatric disorder (Stephan et al. 2009). It has been characterized by aberrant communications within and between intrinsic brain networks, especially those that are involved in core cognitive and primary networks (PN) (Dong et al. 2018a). The cognitive network plays a fundamental role in psychopathological and higher cognitive processes, which includes the central executive network (CEN), default-mode network (DMN), and salience network (SN) (Liao et al. 2019; Supekar et al. 2019; Xi et al. 2021). The PN mainly includes the visual network, auditory network, sensorimotor network, and thalamus and is regarded as closely associated with symptomatology in schizophrenia, particularly auditory hallucinations (Javitt 2009; Dong et al. 2019). Increasing documents highlight the deficits of perceptual processing and multisensory integration, which further contribute to self-disorder in schizophrenia (Borda and Sass 2015). Specific patterns of integration and interactions between these bottom–up and top–down brain networks are known to contribute to complex symptom presentations in schizophrenia (McCutcheon et al. 2020). This substantial research utilizing advanced neuroimaging techniques over the past few decades has partially revealed the neurobiological substrates of schizophrenia (Bègue et al. 2020; McCutcheon et al. 2020; Liu et al. 2021; Sun et al. 2023). Yet, little is known about the brain network mechanisms underlying the clinical response of antipsychotics.

Longitudinal studies have shown the interactions normalization between networks after antipsychotic treatment in schizophrenia (Yang et al. 2021), particularly within triple-network (CEN, DMN, and SN) (Wang et al. 2021a), fronto-striatal-thalamic circuits (Duan et al. 2020), and corticolimbic and primary networks, which is sometimes correlated with symptom improvement (Anticevic et al. 2015; Sarpal et al. 2015). However, there is a considerable inter-individual variation in antipsychotic response, with more than 50% failing to achieve remission on initial treatment (Tarcijonas and Sarpal 2019) and even over a third of patients becoming treatment-resistant schizophrenia owing to a clinically negligible response to first-line antipsychotic medications (Lally et al. 2016; Howes et al. 2017). In addition to being associated with premature treatment discontinuation, symptom exacerbations, relapse, and increased risk of hospitalization, the more crucial reason for the heterogeneous symptom response is the complexity of the disease itself (Case et al. 2011). One interesting explanation is that schizophrenic patients with at least two biological subtypes have different responses to antipsychotic treatment (McCutcheon et al. 2019), showing variability not only in clinical presentation but also in functional outcome and biomarker expression (Davatzikos et al. 2020).

Indeed, to personalize treatment and prognostic outcomes, the recent advent of neuroimaging studies has highlighted the crucial role of neuroanatomical subtypes in schizophrenia (Sun et al. 2015; Dwyer et al. 2018; Chand et al. 2020; Jiang et al. 2023). A recent study has shown that the distinct differences in connectivity in the brain triple-network could distinguish the clinically relevant cognitive subtypes of schizophrenia (Liang et al. 2021). It should be noted that dysfunction and interaction in triple-network and PN may contribute to specific symptoms of schizophrenia, and subtyping dysfunction in this way may allow us to account for the heterogeneity of schizophrenia (Wada et al. 2022). Substantial research also found that baseline functional connectivity (FC) could serve as a predictive biomarker of antipsychotic treatment response in schizophrenia (Sarpal et al. 2016; Mehta et al. 2021), especially the static and dynamic features revealed by the complementary advantage approach of static and dynamic FC (dFC) (Oh et al. 2020; Zhang et al. 2021). However, few studies have focused on communications in core brain networks including the primary and subcortical cortex during antipsychotic treatment and why baseline FC can be considered an effective predictor of antipsychotic efficacy.

Furthermore, previous research has indicated that baseline FC plays a critical role in shaping subsequent changes following antipsychotic treatment (Sarpal et al. 2015). The magnitude of connectivity alterations itself may be a key determinant of symptom improvement, with some research suggesting that a higher degree of normalization in connectivity patterns is associated with better treatment outcomes (Sarpal et al. 2016; Kraguljac et al. 2021). Previous studies have also demonstrated that factors such as the duration of illness (Buoli et al. 2012) and medication dosage (Leucht et al. 2020) are significantly associated with the variability in treatment response among schizophrenia patients. Despite these findings, the underlying mechanisms through which these factors modulate treatment response, particularly their impact on functional brain network reorganization and subsequent symptom alleviation, remain to be elucidated. Moderation analysis is recognized as a robust framework for investigating whether and how specific factors influence the strength or direction of the association between FC changes and clinical symptoms (Rolle et al. 2020). This approach not only advances our understanding of treatment heterogeneity but also facilitates the development of precision medicine strategies in schizophrenia.

To address those questions, we employed the static and dynamic FC approach to investigate the interactions between brain core networks associated with antipsychotic treatment outcomes in first-episode schizophrenia (FES). We first defined two subgroups (responders and nonresponders) according to the clinical symptom relief after antipsychotic monotherapy. We hypothesized that there would be significant differences in coupling among core brain networks before and post-treatment in responders and nonresponders. Second, we used Spearman correlation analysis to test the hypothesis that the pattern of association between altered FC and symptom relief differed in two subgroups. Further, to explore the potential influencing factors of symptom improvement in the two treatment response subgroups, such as baseline FC, illness duration, or medication dosage, we applied moderation analysis to capture potential interactions among these variables. Finally, we performed a support vector machine (SVM) analysis to assess the predictive power of baseline FC in classifying treatment response subtypes to antipsychotic treatment.

Materials and methods

Study design and participants

The current study included 79 patients with medication-naïve FES (gender: 43 males and 36 females; average age: 24.57 ± 6.82 years; range:15–42 years) and 90 healthy controls (HCs) (gender: 50 males and 40 females; average age: 24.04 ± 6.35 years; range:14–39 years) matched for age, gender, education, and handedness. All participants satisfied the selection criteria with no brain injury and history of any central nervous system disease, no history of neurological illness and comorbid alcohol/substance dependency as defined by the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-IV) criteria, and no epileptic seizure and any contraindication to magnetic resonance imaging (MRI) scanning. The data collection was performed at the Shanghai Mental Health Center. The Institutional Review Board of Shanghai Mental Health Center approved the current second analysis of prospectively acquired data (ethics number: 2017-36R), and written informed consents were obtained from all participants.

All patients were diagnosed according to the DSM-IV. Following the baseline MRI scanning, only patients received monotherapy with second-generation antipsychotics for up to 12 weeks. Subsequently, a second follow-up MRI scan was performed in the patient group. The Positive and Negative Syndrome Scale (PANSS) (Kay et al. 1987) was administered to assess the symptomatic severity of patients by the same psychiatrist at baseline and 12-week follow-up. Symptom relief was measured using the reduction of PANSS total scores (⁠|$\Delta{PANSS}_T$|⁠) defined by the following formula (Obermeier et al. 2010):

where |${PANSS}_{T1}$| is the PANSS total score at baseline and |${PANSS}_{T2}$| is the PANSS total score at post-treatment. Patients were divided into responders (SR subgroup, n = 46) and nonresponders (NR subgroup, n = 33) according to the criterion of more or less than 50% |$\Delta{PANSS}_T$| (Boter et al. 2009). More details are shown in Table 1 and Supplementary Table S1. The data from this cohort were also included in other recent studies (Wang et al. 2021a, 2021b; Jiang et al. 2022); however, the mechanisms of interaction among five core brain networks associated with the efficacy of antipsychotic monotherapy were the first to be investigated on this dataset. The experimental design and method pipeline is shown in Fig. 1.

Image acquisition and preprocessing

Resting-state functional magnetic resonance imaging (fMRI) images were collected on a 3-T MRI scanner (Siemens MR B17). Details about image acquisition are described in Supplementary Methods. Data were preprocessed using Statistical Parametric Mapping (SPM12) (http://www.fil.ion.ucl.ac.uk/spm/) and the NIT toolbox (Dong et al. 2018b). Preprocessing of images included the following steps: (i) removing first five time points; (ii) slice-time correction; (iii) realignment; (iv) performing spatial normalization (3 × 3 × 3 mm) to the Montreal Neurological Institute space; (v) smoothing with an 8-mm full-width at half-maximum 3D isotropic Gaussian kernel; (vi) removal of linear trends; (vii) regressing nuisance signal (including 24-parameter motion correction, white matter and cerebrospinal fluid signals); and (viii) band-pass filtering (0.01–0.08 Hz). In addition, we excluded participants with unacceptable data quality (artifacts such as signal dropout or distortion) and those with maximum motion >2.5 mm or 2.5° and mean frame-wise displacement (mFD) > 0.5 mm (Power et al. 2012). Details about the formula of mFD are shown in Supplementary Methods.

Regions of interest definition of five-core brain networks

The nodes were selected from our previous work (Luo et al. 2012) and literature examining schizophrenia (Liao et al. 2019). A total of 35 nodes were selected, of which CEN consisted of 4 nodes in the bilateral dorsolateral prefrontal cortex and inferior parietal lobule (IPL), DMN included posterior cingulate cortex (PCC) and other 9 nodes, SN contained the right dorsal anterior insular (R.dAI) and other 6 nodes, the subcortical network (SCN) consisted of 6 including caudate, putamen, and thalamus, the PN consisted of 10 nodes including visual, auditory, and somatosensory cortex. The mean time series of each region of interest (ROI) were extracted from the 8-mm-radius spheres centered on the coordinates for subsequent FC evaluation. The detailed coordinates of all ROIs are listed in Supplemental Table S2.

Static and dynamic FC estimation

First, to measure the static interaction between nodes, Pearson’s correlation coefficients between each pair of ROIs were calculated and further transformed to obtain static functional connectivity (sFC) values through Fisher’s r-to-z transformation. This resulted in a 35 × 35 symmetric sFC matrix for each participant, with each element representing the connectivity strength between a pair of nodes.

Second, we conducted a sliding window dFC analysis to assess the dynamic interaction between nodes. The steps included the following: (i) the time course of each node was divided into windows of 50 TR (100 s) size according to the optimal window length (no less than 1/|${f}_{min}$|⁠, |${f}_{min}$| = 0.01 Hz) (Leonardi and Van De Ville 2015) with sliding the onset by 1TR (2 s); (ii) within each window, the Pearson correlation analysis was performed between each pair of nodes; and (iii) the dFC was estimated by calculating the variability across all sliding window, the indicator variability came from our previous work (Luo et al. 2019):

where |${f}_t$| was the FC between |${ROI}_i$| and |${ROI}_j$| at the time window t, t = 1, 2, … n, |${f}_{static}$| was the sFC between |${ROI}_i$| and |${ROI}_j$|⁠. Finally, a 35 × 35 symmetric dFC matrix was obtained for each participant.

Statistical analysis

First, to examine the interaction effect on FC changes after monotherapy, the two-way repeated-measures analysis of variance (ANOVA) were carried out in the SR and NR subgroups. To correct for multiple comparisons, we used permutation tests to assess the effect of factors on repeat-measures ANOVA while controlling the family-wise error rate (Groppe et al. 2011; Frossard and Renaud 2021) (Supplementary Methods). A statistical significance level was set at P < 0.005. Two-sample t tests and paired t tests were performed for the post hoc analysis. Subsequently, we focused on the edges with interactive effects and performed baseline statistical analysis using two-sample t tests. All patient groups, SR subgroup, and NR subgroup were compared with HCs at baseline. Age, gender, education years, and the mFD were included as covariates.

Associations with treatment outcomes

To comprehensively investigate the effect of treatment, we calculated the Spearman’s correlation coefficient between the changed FC (⁠|$\Delta FC$|⁠) of edges with interaction effects and the reduced PANSS scores (⁠|$\Delta PANSS$|⁠) in SR and NR subgroups, respectively. The |$\Delta FC$| was defined as

Sub-symptom relief was defined as similar to the |$\Delta{PANSS}_T$| (detailed formulas were described in Supplementary Methods). Age, gender, education years, and mFD were controlled as covariates. Further, we investigated the differences in brain–behavior correlation patterns between the SR and NR subgroups. To be specific, we calculated the difference in correlation values between the SR and NR subgroup. Then, we performed permutation tests to construct the null distribution of the correlation value differences and obtain their significance. In each permutation, we randomly shuffled the labels of the SR and NR subgroups, reconstructed the SR and NR subgroups, and recalculated the correlation value differences. This process was repeated 10,000 times, and the position of the observed correlation value difference in the null distribution was used to determine statistical significance.

Moderation analysis

To investigate whether the significant correlations between the |$\Delta FC$| and |$\Delta PANSS$| were modulated by other factors, such as baseline FC, |$\Delta FC$|⁠, illness duration, and chlorpromazine (CPZ) equivalents. The moderation analysis was performed using the linear multiple regression analysis (Aiken et al. 1991; Fairchild and MacKinnon 2009). The moderation model was as follows (Landis 2003):

Here, the reduced PANSS scores (⁠|$\Delta PANSS$|⁠) were entered as a predictor (X), the changed FC (⁠|$\Delta FC$|⁠) as an outcome variable (Y), other factors (the baseline FC and |$\Delta FC$| of other edges, illness duration, and CPZ) as a moderator variable (M), and |${\beta}_3$| as the modulation contribution. Age, gender, education years, and the average mFD (baseline and post-treatment) were included as covariates for controlling. Evidence for moderation was obtained if the addition of the interaction term (X|$\times$|M) significantly accounted for an increase in R² (Hayes and Matthes 2009; Pereira et al. 2021). Permutation tests (10,000 times) were used for significance testing of moderation effects.

Classification and prediction analysis of baseline FC

To test the power of baseline FC to classify and predict clinical response subtypes of schizophrenia, we conducted correlation analysis and classification analysis using baseline FC. The classification analysis adopted a nonlinear SVM approach, whose feature space was characterized by baseline FC between core brain networks in SR and NR subgroups. SVM aims to find the maximum marginally separated hyperplane in a high-dimensional space for data classification by mapping the original features via a kernel function (Ben-Hur et al. 2008). We adopted the Gaussian radial basis functional kernel as kernel type (Wang et al. 2020):

where |${x}_i$| and |${x}_j$| are two samples and |$\mathrm{\sigma}$| controls the width of the Gaussian kernel. Another important hyperparameter is the regularization parameter C, which controls the trade-off between maximal margin and misclassification rate. In the current study, the Python toolkit sklearn’s GridSearchCV was conducted to complete the optimal parameter exploration, with the parameter C and |$\mathrm{\sigma}$| set to {10⁻⁴, 10⁻³, 10⁻²,10⁻¹, 10⁰, 10¹, 10², 10³}. First, we randomly selected the baseline FC four-fifths of subjects as feature selection, which included sFC and dFC. Second, the features that were dimensionally reduced by principal component analysis were screened to the SVM model to establish the hyperplane separately or together. Finally, the “leave-one-out” cross-validation approach was employed to access the SVM model. Accuracy, sensitivity, specificity, and receiver operating characteristic curves were utilized to quantify the performance of the classifier (Shen et al. 2010). Furthermore, based on the classification scores output by the SVM model, we performed statistical analyses using two-sample t tests and paired t tests between two subgroups and each feature separately (sFC, dFC, and the combination of dFC and sFC (dsFC)).

Results

Significant interaction effects between SR and NR subgroup

Significant interactions between clinical response subgroups (SR and NR) and status (baseline and post-treatment) in terms of core brain networks coupling were observed: the interactions of sFC centered on the CEN and dFC centered on the PN (seven edges, P < 0.005, corrected; Figs 2a and3a and Table 2). Specifically, for the post hoc analysis of sFC, we found that the NR subgroup showed significant decreases in connectivity between CEN-DMN, CEN-SN, and SN-SCN after treatment compared to the baseline state, while the SR subgroup showed no significant differences. Moreover, in the post-treatment state, the NR subgroup displayed consistently lower connectivity compared to the SR subgroup (Fig. 2d; Table 2).

The post hoc analysis of dFC revealed significant differences between the SR and NR subgroup at baseline in the PN-DMN/SCN and DMN-SN connections (Fig. 3d). Specifically, the NR subgroup exhibited significantly higher dFC than the SR subgroup in the PN-DMN/SCN, while the NR subgroup showed lower dFC than the SR subgroup in the DMN-SN. Interestingly, we found that after antipsychotic treatment, the SR and NR subgroups exhibited imbalanced changes in dFC in the PN-DMN/SCN and DMN-SN connections. Specifically, compared to the pre-treatment state, the SR subgroup showed an increase in dFC in the PN-DMN/SCN, while the NR subgroup showed a decreasing trend. In contrast, for the DMN-SN dFC changes, the SR subgroup exhibited a decrease after treatment compared to baseline, while the NR subgroup showed an increase (Fig. 3d; Table 2).

Additionally, we also found significant main effects in the ANOVA analysis (P < 0.005, corrected; Figs. 2b and c and 3b and c). The post hoc analyses found that the FC between PN and cognitive networks in the NR subgroup showed significant reduction relative to SR subgroup. More detailed results are provided in Supplementary Table S3. Besides, given the comparison with HCs at baseline, the SR, NR as well as all patient groups consistently showed decreased sFC between the R.dAI and left putamen (L.PUT) (SR: P = 0.040, t = −2.072; NR: P = 0.066, t = −1.854) (Figs. 1d and 3a; Table 2).

Changed FC associated with symptom relief and moderation effects

We observed significant relationships between changed FC (⁠|$\Delta FC$|⁠) and symptom improvement in both the SR and NR subgroups (Figs 2e and3e). In the SR subgroup, the changed sFC (⁠|$\Delta \mathrm{s} FC$|⁠) between the left inferior parietal lobule (L.IPL) and the PCC, as well as the pregenual anterior cingulate cortex (pgACC), were significantly negatively correlated with |$\Delta{PANSS}_T$| (L.IPL-PCC: P = 0.023, r = −0.358; L.IPL-pgACC: P = 0.005, r = −0.461). In contrast, no such correlations were found in the NR subgroup and the differences in correlations between the SR and NR subgroups were significant (P < 0.00001 for L.IPL-PCC and P < 0.001 for L.IPL-pgACC; see Supplementary Fig. S4). Furthermore, moderation analyses revealed that the communication between networks at baseline, particularly those centered on L.IPL and PCC, moderated the relationship between |$\Delta \mathrm{s} FC$| and |$\Delta{PANSS}_T$| in the SR subgroup. When the moderation effect of the |$\Delta \mathrm{s} FC$| was taken into account, the |$\Delta \mathrm{s} FC$| of L.IPL-R.vAI (right ventral anterior insular) acted as a positive moderator, weakening the negative correlation between the |$\Delta sFC$| of L.IPL-PCC and |$\Delta{PANSS}_T$|⁠. In contrast, the |$\Delta sFC$| of pgACC-L.MI (left primary motor) served as a negative moderator, strengthening the negative association between |$\Delta sFC$| of L.IPL-pgACC, and |$\Delta{PANSS}_T$| (see Fig. 2e). Detailed moderation parameters are provided Supplementary Table S4. Other correlations and moderations with the mean CPZ at two time points and the relief of general symptom are provided in Supplementary Fig. S2.

In the NR subgroup, the changed dFC (⁠|$\Delta \mathrm{d} FC$|⁠) between the left angular gyrus (L.AG) and right posterior insular (R.PI) was negatively correlated with positive symptom relief (⁠|$\Delta{PANSS}_P$|⁠) (P = 0.033, r = −0.415). Additionally, baseline dFC between R.PI and L.AG acted as a significant negative moderator, enhancing this negative correlation (see Fig. 4e and Supplementary Table S4). The correlation pattern between |$\Delta \mathrm{d} FC$| of L.AG-R.PI and positive symptom relief showed significant differences between the SR and NR subgroups (P = 0.020; Supplementary Fig. S4).

Additionally, considering the continuity of symptom improvement, for the entire patient group, we also observed significant correlations between the |$\Delta FC$| and the |$\Delta PANSS$|⁠. The changed static functional connectivity (∆sFC) between L.IPL and PCC/pgACC was significantly negatively associated with the mean CPZ at two time points. Moreover, the ∆sFC between R.dAI and L.PUT/L.THA was positively correlated with both total (⁠|$\Delta{PANSS}_T$|⁠) and negative (⁠|$\Delta{PANSS}_N$|⁠) symptom relief (Supplementary Fig. S3).

Prediction and classification performance of baseline FC

We found that baseline sFC between the R.dAI and L.PUT predicted the duration of untreated psychosis (DUP) in the NR subgroup (P = 0.017, r = −0.491), but no such effect was observed in the SR subgroup (Fig. 4c). Further, using SVM classification, we found that baseline FC features could effectively classify antipsychotic treatment subtypes in FES. The classification performance was as follows: for sFC: Accuracy = 68.354%, Specificity = 84.783%, Sensitivity = 45.455%, AUC = 0.653; for dFC: Accuracy = 88.608%, Specificity = 93.478%, Sensitivity = 81.818%, AUC = 0.924; for combined dsFC: Accuracy = 91.139%, Specificity = 97.826%, Sensitivity = 81.818%, AUC = 0.921 (Supplementary Table S5). The dsFC features provided the best overall classification performance, as indicated by the highest AUC (Fig. 4d). In terms of classification scores, the NR subgroup had significantly higher scores compared to the SR subgroup across all features (sFC: t = 4.301, P < 0.0001, dFC: t = 10.576, P < 0.0001, dsFC: t = 14.978, P < 0.0001). Within the SR subgroup, there was a significant increase in classification scores between sFC and dFC (t = 4.401, P < 0.0001) and between sFC and dsFC (t = 5.620, P < 0.0001). However, within the NR subgroup, there were significant reductions in classification scores between sFC and dFC (t = −5.559, P < 0.0001) and between sFC and dsFC (t = −8.665, P < 0.0001) (Fig. 4e).

Discussion

This study first elucidates the differences in static and dynamic communications among the five core brain networks in treatment response subtypes of antipsychotic therapy for schizophrenia. It confirms the close relationship between neuroimaging subtypes and treatment outcomes, uncovering specific response patterns in brain connectivity induced by antipsychotics: static interactions centered on CEN and dynamic interactions centered on PN (Fig. 5). Specifically, after antipsychotic monotherapy, we found that the NR subgroup presented significantly reduced sFC in CEN-DMN/SN and SN-SCN compared with the baseline state, while there was no significant difference in the SR subgroup. From a dynamic perspective, we mainly found that the NR subgroup at baseline showed significantly enhanced dFC in PN-DMN/SCN and reduced dFC in DMN-SN relative to the SR subgroup. Further, association analysis revealed negative correlations between changed sFC of CEN-DMN/SN and symptomatic improvement in the SR subgroup, as well as a negative correlation between changed dFC of DMN-SN and positive symptom remission in the NR subgroup. These correlations were modulated by baseline FC and |$\Delta FC$| between related nodes and other core networks. Finally, the classification analysis further revealed a tight link between neuroimaging subtypes and a clinical subgroup of response in schizophrenia (the static and dynamic binding baseline FC in five core brain networks could effectively classify subtypes of response to antipsychotic).

Prior studies have demonstrated that cognitive networks like the CEN tend to exhibit stable FC, while sensory networks often display greater temporal variability (Shine et al. 2016; Keller et al. 2023). Our findings of static interactions centered on CEN and dynamic interactions centered on PN are consistent with this framework, highlighting the concept that schizophrenia involves a specific core network connectivity disorder (Van Den Heuvel and Fornito 2014), and our study extends how these networks static and dynamic interactions are related to treatment response. Specifically, the static interactions mainly reflect the effect before and after antipsychotics treatment, while the dynamic interactions primarily reflect the difference between the SR and NR subgroups at baseline. The predominance of static connectivity in the CEN may be attributed to its role in cognitive control and higher-order executive functions, which tend to remain stable over time (Menon 2011; Cole et al. 2013). Previous studies have identified alterations in the CEN as being associated with cognitive deficits in schizophrenia, suggesting that the static nature of CEN connectivity may reflect its core function in stabilizing cognitive processes (Pietrzykowski et al. 2022). In contrast, the PN exhibits more dynamic interactions, likely due to its essential role in processing sensory stimuli (Sadaghiani and Kleinschmidt 2013). Sensory and motor networks must rapidly adapt to continuously changing environmental inputs, necessitating a high degree of functional flexibility. These dynamic interactions may be particularly relevant to schizophrenia, as altered sensory processing and adaptability to external stimuli may be either impaired or exaggerated in different patient subgroups (Uhlhaas and Singer 2010). These findings have important implications for understanding treatment response. Individuals with more stable CEN connectivity may benefit from cognitive interventions aimed at enhancing executive function, whereas those with greater PN flexibility may be better suited for treatments targeting sensory processing abnormalities.

The core five brain networks (CEN, DMN, SN, SCN, and PN) have been consistently implicated in the pathophysiology of schizophrenia (Chen et al. 2015; Dong et al. 2018a; Supekar et al. 2019) and were the critical target networks for the effects of antipsychotics (Sarpal et al. 2015; Chopra et al. 2021). Within these affected efficacy networks, we found that the NR subgroup showed significantly lower sFC among triple-network at post-treatment than the SR subgroup and baseline. However, we did not observe correlations between these significantly reduced sFC and symptom improvement in the NR subgroup; instead, a significant negative association was found in the SR subgroup. A growing body of evidence highlights the heterogeneity of antipsychotic response in schizophrenia (Cao et al. 2018; McCutcheon et al. 2019; Mizuno et al. 2019), and a hypothesis has been proposed that distinct subtypes of schizophrenia are defined by treatment response (Farooq et al. 2013; Lee et al. 2016). A recent study defined two subtypes of FES using the resting-state connectivity profiles of the triple-network model and found that the subtype of schizophrenia marked by hypoconnectivity within the triple-network tended to have more severe and persistent negative symptoms after antipsychotic treatment (Liang et al. 2021). Research of antipsychotic treatment-resistance has also demonstrated that the widespread hypoconnectivity in regions including DMN, frontal lobe, insula, and thalamus was especially associated with nonresponse to antipsychotics in schizophrenia (Chan et al. 2019; Tuovinen and Hofer 2023). Consistent with these findings, our results also support the presence of subtypes of treatment-response in schizophrenia, and the widespread reduction in FC of core brain networks may be a biomarker for an inefficient subtype of antipsychotic medication response.

Furthermore, our findings on dFC and the subgroup’s main effects also support this conclusion of the efficacy subtype. To be specific, we found that the significantly decreased FC in the NR subgroup relative to the SR subgroup mainly appeared in the communication of PN and other high-order cognitive networks. At baseline state, the NR subgroup showed significantly enhanced dFC in PN-DMN/SCN and reduced dFC in DMN-SN relative to the SR subgroup. A significant negative association between changed dFC in DMN-SN and reduced scores in positive symptoms was only observed in the NR subgroup. The concept of stratified remission suggests that distinct subtypes of psychosis patients have different therapeutic responses and alleviative pathways (Martinuzzi et al. 2019). Our earlier work found that schizophrenia showed elevated instability of information communication in the primary sensory and perceptual system, which is closely related to the self-disorder of patients (Dong et al. 2019). In line with these studies, the dysfunctional communication between primary sensory and cognitive processes may be a critical feature between clinical subtypes and may represent different aspects of self-disorder in schizophrenia.

Further taking into account the effects of baseline FC, we found that both subgroups, as well as the overall patient group, showed significantly reduced FC in R.dAI-L.PUT relative to HCs. And a negative correlation between the DUP and baseline sFC of R.dAI-L.PUT was observed in the NR subgroup. Research on risperidone monotherapy found that the FC between the striatum and SN could predict treatment response and negative symptom improvement in schizophrenia (Han et al. 2020). In addition, previous studies have confirmed that the decreased FC related to the striatum and SN was associated with longer DUP and worsened response to treatment (Sarpal et al. 2017; Maximo et al. 2020). Consistent with these findings, our results also highlight the association between reduced FC of SN-striatum and worse treatment outcomes and longer DUP, although there was no statistically significant DUP between our efficacy subgroups.

A plethora of neuroimaging studies have shown that resting-state FC at baseline can serve as a potential predictor of treatment response to antipsychotic drugs in schizophrenia (Sarpal et al. 2016; Blessing et al. 2020; Mehta et al. 2021). Prior research has demonstrated that the intra-integration of DMN and SN and their cross-network interactions with the PN and SCN at baseline emerged as key predictors of symptomatic improvement in response to antipsychotic treatment in schizophrenia (Doucet et al. 2018). In the present study, the difference between the SR and NR subgroup at baseline FC, and the moderation/classification effects of baseline FC may explain why baseline FC can predict efficacy. These findings complement previous observations and provide evidence on how intrinsic connectivity relates to outcomes from subtypes with different responses to antipsychotics in schizophrenia and facilitate individualized treatment of patients with schizophrenia.

From a clinical perspective, we also found a significantly higher negative symptom score in the NR subgroup relative to the SR subgroup at baseline. The earlier research confirmed that the benefits of antipsychotics in terms of symptom reduction were greatest in those with the most severe symptoms, as they have more room for improvement (Furukawa et al. 2015). However, our result did not seem to support this view, as the total clinical symptoms at baseline showed no significant differences between the two subgroups, except for negative symptoms. In addition, prior research has found that the greater antipsychotic effect in schizophrenia is associated with total and positive symptoms, while the less efficacy is related to negative symptoms, depressive symptoms, social functioning, and quality of life (Haddad and Correll 2018). Several recent studies of stratified remission in first-episode psychosis found that patients with less severe negative symptoms at baseline were more likely to be remitters than others (Jäger et al. 2009; Martinuzzi et al. 2019). Neuroimaging research has also found that negative symptoms were closely related to communications among the CEN, DMN, SN, striatum, and cortical networks and were associated with poor long-term prognosis (Forlim et al. 2020; Brakowski et al. 2022; Wang et al. 2023). Consistent with these findings, our results also support the conclusion that patients with an overall milder negative symptom at baseline (SR subgroup) were more likely to benefit from antipsychotic therapy and to achieve symptom remission and that decreased FC centered on five core networks may serve as potential functional biomarkers for lower efficacy of antipsychotic therapy.

Several limitations of this study should be considered. First, our sample size for longitudinal subtypes of patients was relatively small, and it was necessary to conduct longitudinal research with a large sample in the future. Second, previous evidence has shown that most FES patients respond between weeks 8 and 16 of treatment with a single antipsychotic medication (Gallego et al. 2011; Demjaha et al. 2017). However, the current study was limited to two longitudinal time points, and it was possible that some patients would have slower medication onset, which required long-term continuous follow-up to track the efficacy of antipsychotics. Third, the underlying neurobiology of schizophrenia may differ from treatment-responsive, the findings of this study can be further confirmed by combining neurotransmitters and genes. Fourth, all patients were not taking the same antipsychotic medication, which might be a factor in variation, and future studies could be carried out specifically for a specific single antipsychotic. Fifth, the current study used a 50% reduction rate to for the symptom remission criterion, which was based on expert consensus from previous studies (Andreasen et al. 2005) and extensive research (van Os et al. 2006; Emsley et al. 2011; Jiang et al. 2022) but may still be at risk of being arbitrary. Future research may consider treating symptoms as a continuous variable and perform a more granular, hierarchical analysis.

Conclusion

In summary, our study characterizes five core brain network differences associated with subtypes of clinical antipsychotic efficacy using both static and dynamic FC methods based on resting-state fMRI data. Results suggest that abnormal communications between primary sensory and cognitive core networks may be a specific manifestation between the two clinical subtypes (responders and nonresponders) and that the decreased FC at the overall network level may be a critical feature of poor response to antipsychotics. Moreover, the presentation of an unbalanced correlation pattern between the changed FC and symptomatic improvement in two clinical subtypes, as well as more severe negative symptoms at baseline in the NR subgroup support the idea that the subtypes of patients with milder negative symptoms at baseline were more likely to be remitters. The baseline FC not only plays a critical role as a moderator in symptom relief but also accurately classifies therapeutic subtypes, underscoring the importance of baseline FC in predicting therapeutic response. These findings highlight the significant potential of interaction patterns between primary and higher-order cognitive core brain networks in exploring biomarkers and neuroanatomical subtypes related to the efficacy of antipsychotics. They provide new evidence for stratification and personalized treatment of patients with schizophrenia, laying the foundation for early intervention in this debilitating disorder through precision medicine approaches.

Acknowledgments

We are grateful to all the participants in this study.

Author contributions

Yuling Luo and Tianyuan Zhu (Conceptualization, Methodology, Formal analysis, Visualization, Writing-original draft); Yu Zhang, Jiamin Fan, Xiaojun Zuo, and Xiaorong Feng (Formal analysis, Data curation, Software); Jinnan Gong (Software and Methodology); Dezhong Yao (Project administration, Funding acquisition, Supervision); Jijun Wang (Conceptualization, Resources, Supervision); Cheng Luo (Conceptualization, Funding acquisition, Project administration, Supervision).

Funding

This work was supported by the National Key R&D Plan of China (2024YFE0215100), the National Nature Science Foundation of China (82371560, 62401124, and 62201133), the Project of Science and Technology Department of Sichuan Province (23NSFSC0016), the grant from Chengdu Science and Technology Bureau (2024-YF05-02056-SN), and the CAMS Innovation Fund for Medical Sciences (CIFMS) (2019-I2M-5-039).

Conflict of interest statement. None declared.

Data availability

The code and data of this study are available from the corresponding author (Jijun Wang and Cheng Luo) upon reasonable request.

References

Author notes