Introduction

Schizophrenia (SZ) is a severe, chronic mental illness with a lifetime prevalence of about 1% [1]. The economic and social burdens of SZ are enormous, as SZ usually strikes in early adulthood and leads to long-term social and occupational dysfunction as well as a reduction in overall life expectancy [2]. Antipsychotic medications are the mainstay treatments for SZ, but patient responses to these medications are variable and unpredictable, with 20 to 30% of patients showing clinically significant treatment resistance [3]. Therefore, predicting individual patient response prior to treatment based on easily measured clinical variables could be of immense help for developing individualized treatment plans for improved long-term quality of life. Further, it is likely that such markers will provide clues to the pathogenesis of SZ and the processes underlying treatment resistance.

Antipsychotics are dopamine receptor antagonists that mitigate SZ symptoms at least in part by promoting structural and functional plasticity within brain circuits. Thus, it is theorized that neurobiological metrics of brain circuit function may be effective predictors of antipsychotic drug response. Among potential circuit markers, indices of connectivity derived from functional magnetic resonance imaging (fMRI) may be particularly advantageous as these can be measured in awake humans multiple times during treatment under a variety of conditions, including while performing specific cognitive tasks. Thus, serial fMRI can track changes in whole-brain signals across cognitive and behavioral states, or depict differences associated with specific traits or clinical states (e.g., drug-free, medicated, stable, remission, or relapse) [4]. Some fMRI paradigms have revealed correlations between baseline brain activity or connectivity and antipsychotic medication response. For instance, resting-state (rs)-fMRI studies have reported correlations between treatment response and a striatum-to-cortex functional connectivity (FC) node [5], low-frequency fluctuating fractional amplitude in the left crustal nucleus [6], and FC of the frontoparietal region [7], suggesting that these connectivity metrics may have clinical utility as prognostic biomarkers. In addition to resting-state measures, task-based fMRI has also proven useful for studying potential predictors of treatment response. For example, one study using imaging data obtained during a cognitive task found that task-related activity in frontal-parietal regions at baseline predicted symptom reduction at one year [8], while another using multiple cognitive tasks found that strong connectivity in the default mode network and weak connectivity in cerebellar-cortical circuits at baseline predicted better patient treatment response [9]. Collectively, these findings suggest that task-based imaging metrics at baseline may predict subsequent antipsychotic treatment response. However, the current research findings are inconsistent, potentially due to insufficient sample size, differences in methodology, or sample heterogeneity.

We have identified some imaging biomarkers related to the symptoms and therapeutic effects of SZ in recent work, such as the right superior frontal gyrus [10] and right cingulate gyrus [11], but we have not found and validated imaging features in machine learning models. Therefore, in the current study, we recruited two independent clinical samples in the acute phase of SZ with relatively low heterogeneity. Then, by combining rs-fMRI and machine learning, we explored whether neural features of the individual functional connectome can predict patient outcome from antipsychotic treatment using a connectome-based predictive model (CPM) [12].

Materials and methods

Participants

Two independent SZ patient cohorts were recruited using similar sampling methods, a 105-patient discovery group and a 52-patient validation group. Both cohorts were drawn from northern Henan Han patients treated at the Second Affiliated Hospital of Xinxiang Medical University. The enrollment criteria for both samples were (1) SZ diagnosed by two psychiatrists or attending physicians according to the Diagnostic and Statistical Manual of Mental Disorders, 4th edition, (2) no somatic diseases such as hypertension and diabetes, (3) Han Chinese ethnicity, (4) right-handed, unlimited gender, (5) between 18 and 55 years of age, and (6) Positive and Negative Symptom Scale (PANSS) total score > 60. Exclusion criteria for both cohorts were as follows: (1) a history of severe unstable physical illness, comorbidities, organic brain disease, or other psychiatric disorders; (2) a well-documented history of epilepsy; (3)meet the diagnostic criteria for alcohol and drug (methamphetamine, ketamine, and cocaine) dependence in the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition, Text Revision (4) electroconvulsive therapy within the past 6 months; (5) pregnancy or breastfeeding; (6) prior suicide attempt; (7) severe symptoms of euphoria or agitation within a week prior to the MRI scan. All participants provided written informed consent after receiving a full explanation of study goals and methods. The study protocol was approved by the Ethics Committee of the Second Affiliated Hospital of Xinxiang Medical University (Registration number: ChiCTR2200057153).

Patients in both cohorts received a comprehensive clinical assessment at baseline, including administration of the PANSS scale and rs-fMRI acquisition. After 8 weeks of antipsychotic medication, the PANSS assessment was repeated to determine the treatment response. The scales were collected by 2 attending psychiatrists who had at least 2 weeks of training in PANSS scale administration, and participating physicians passed a consistency reliability test.

Treatment response assessment

Since SZ symptoms are the primary therapeutic target of current antipsychotic drugs, we focused on prediction of symptom mitigation. The degree of symptom improvement was expressed as a percentage reduction in the total score of the three psychiatric programs (positive symptoms, negative symptoms, and general psychopathology) of the PANSS as follows:

$$\Delta =\frac{{{\rm{PANSS}}}_{{\rm{baseline}}}-{{\rm{PANSS}}}_{{\rm{followup}}}}{{{\rm{PANSS}}}_{{\rm{baseline}}}-30}\times 100 \%$$

Subtraction of 30 from the denominator is necessary because 30 is the lowest score for patients (i.e., a score = 30 indicates “asymptomatic” status).

MRI data acquisition and processing

All MRI data were acquired using a Siemens 3.0-T MRI scanner (Siemens, Germany). Functional MRI data were acquired using an echo planar imaging sequence sensitive to blood oxygen level-dependent signals. Sequence settings were as follows: repetition time, 2000 ms; echo time, 30 ms; flip angle, 90°; matrix size, 64 × 64; resolution of axial slice, 3.4 × 3.4 mm2; slice thickness, 4 mm; slice gap, 0.6 mm. Resting-state function data were acquired over a period of 8 minutes and contained a total of 240 time points.

Resting-state images were preprocessed using the BRANT toolkit. Briefly, preprocessing steps included discarding the first ten time points, slice timing correction, realignment, coregistration, normalization to Montreal Neurological Institute (MNI) space, resampling, regressing out linear trends, global signals, and head motion parameters, and temporal band-pass filtering from 0.01–0.08 Hz.

For each patient, we used the BRANT toolkit to perform whole-brain FC computations based on the Brainnetome atlas [13]. Whole-brain images were segmented according to the 273 nodes (210 cortical, 36 subcortical, 27 cerebellar) of the Brainnetome atlas. Pearson’s correlation coefficients were then calculated between the time series extracted from each node, yielding a 273 × 273 connectivity matrix for each patient.

Connectome-based predictive modeling

To investigate whether individual resting-state FC matrices can predict the post-treatment reduction in symptom score, a CPM was constructed and trained using connectome matrix data from the discovery cohort. Given the moderate size of this cohort, we first validated model performance using the leave-one-out cross-validation strategy, which has been used for similar tasks with small datasets [14]. Specifically, we used one sample as the test set and the remaining samples as the training set until all samples had been tested once. Pearson correlation coefficients were then calculated between the values of each connection in the subjects’ FC matrix (N = 273 × 272/2 = 37,128) and the % reduction in total PANSS score (%Δ). Connections that were significantly correlated (p < 0.001) were subsequently selected as features. These selected features were categorized as positive if connectivity was positively associated with PANSS total score reduction, or as negative if connectivity was negatively associated with PANSS total score reduction. Subsequently, we will set these two types of features as independent variables and take the PANSS total score reduction as the dependent variable, and then incorporate them into a linear regression model for analysis. During this process, it is necessary to process these two types of features. Specifically, for each sample, we sum up the two types of connection features respectively to generate two separate subject scores, namely the positive and the negative subject score, and use them as independent variables. The fitted model parameters were then applied to the discovery dataset and used to predict the score reduction for each discovery set patient. After cross-validation, Pearson correlation coefficients were calculated between the predicted and true score reduction for all patients in the discovery cohort as a measure of predictive performance. Finally, the independence of the model was determined by calculating r values for 5000 permutations in which the data were randomly shuffled, and the proportion of correlation coefficients within the distribution of permutations greater than the true value was calculated as the P-value (threshold = 0.01).

Validation of prediction performance

We then validated the predictive model in a separate validation cohort. The connectivity features used for predictions were those selected in at least 80% of the 105 CPM repetitions. These connections were used to calculate two subject scores and fit the above model using the entire discovery cohort. Thereafter, the derived model parameters were applied unmodified to the validation cohort to predict individual patient score reduction. Predictive performance was again evaluated by calculating the Pearson correlation coefficient between predicted and true score reduction and significance tested by performing 5000 random permutations.

Results

Demographics, clinical characteristics, and flow charts

As shown in Fig. 1, the overall flowchart of this study was including eight steps. This study included two sample groups, namely the discovery sample containing 105 SZ patients and the validation sample containing 52 SZ patients. Table 1 summarizes the demographic and clinical characteristics of these two groups. Both sample groups were from independent clinical cohorts, and it was found that 16 and 19 patients in the discovery sample and the validation sample, respectively, did not take medication at baseline, while the remaining patients were taking second-generation antipsychotic drugs before enrollment. Specific antipsychotic use is shown in Supplementary Table 1. There are differences in age and the dose of antipsychotic drugs between the two groups of samples. However, the features we discovered and the trained model can still excellently predict the treatment efficacy of patients in the validation sample, demonstrating that this group of features and the model have strong generalization ability in independent samples.

Fig. 1
figure 1

Flow chart of the study.

Full size image
Table 1 Demographic and clinical characteristics of the studied samples.
Full size table

Prediction of treatment response

In the discovery cohort, predictive performance of the CPM after cross-validation (as evaluated by the correlation between predicted and real post-treatment PANSS total score reduction) was r = 0.59, P = 3.582 × 10−11 (MSE = 0.021) (Fig. 2a), which was statistically significant according to 5000-permutations testing (P = 0.0002) (Fig. 2b). We also tested the correlation between the individual subject scores obtained by summing the positive and negative features separately, and PANSS total score reduction. In the discovery cohort, the correlation for the positive features was r = 0.53, P = 6.343 × 10−9 (Fig. 2c) whereas for the negative features it was r = −0.54, P = 2.239 × 10−9 (Fig. 2d).

Fig. 2: The connectome-based features and their prediction performance in the discovery cohort.
figure 2

a Scatter plot showing the correlation between predicted and true PANSS total score reduction in the discovery cohort. b The correlation is significant according to 5000-permuations testing. The green dashed lines in the permutation histograms indicate the position of the observed correlation. c Scatter plots showing the correlations of positive features with the true PANSS total score reduction in the discovery cohort. d Scatter plots showing the correlations of negative features with the true PANSS total score reduction in the discovery cohort. The positive and negative features are marked in red and blue, respectively.

Full size image

Feature information

Twelve neural connections were consistently selected by the CPM on at least 80% of the replicated experiments, forming the final identifying features of the predictive model. Four of these connections exhibited a positive correlation between baseline strength and symptom improvement (PANSS total score reduction), while eight exhibited a negative correlation between connectivity strength and symptom improvement (termed positive and negative features, respectively). In other words, stronger connectivity at baseline between nodes of positive features predicted a greater score reduction and therefore a superior treatment response, whereas weaker connectivity at baseline between nodes of negative features predicted a greater score reduction and therefore a superior treatment outcome.

Positive features mainly involved connections between the postcentral gyrus (PoG) and superior parietal gyrus, amygdala and basal ganglia, and between the mid-occipital ventral cortex and cerebellum. In contrast, negative features consisted mainly of connections between the frontal lobe and parietal lobe, cerebellum and temporal lobe, parietal lobe and thalamus, and between the insula and cerebellum (Fig. 3). Detailed information on the brain regions included in these features is provided in Table 2.

Fig. 3: Locations of connector-based features for predicting antipsychotic drug response according to the Brainnetome atlas.
figure 3

Final features (four positive and eight negative) selected from the 105 repetitions of cross-validated connectome-based predictive modeling in the discovery cohort. Red indicates positive traits and blue indicates negative traits.

Full size image
Table 2 Connector-based features in the Brainnetome atlas.
Full size table

Importance of identifying features

To assess the relative importance of each identified feature, predictive performance was also recalculated following removing of features in 6 specific brain regions: orbital gyrus, inferior temporal gyrus, inferior parietal lobe, postcentral gyrus, subcortical nuclei (including the amygdala, basal ganglia, and thalamus), and the cerebellum. Deletion of the features including the postcentral gyrus (Fig. 4) produced the greatest reduction in model predictive performance (Supplementary Table 2), while training the model using only features including the postcentral gyrus resulted in a reasonable cross-validated prediction performance of r = 0.35, P = 2.484 × 10−4. Therefore, features including the postcentral gyrus are essential to the predictive performance of the model.

Fig. 4: The set of features that contribute most to model predictive performance.
figure 4

The connections of the right postcentral gyrus (blue) to the bilateral superior parietal lobule (red) belong to the positive feature group. L, Left; R, Right; SPL, the Superior Parietal Lobule; POG, the Postcentral Gyrus.

Full size image

Validation using an independent cohort

When the same model was applied to an independent cohort, these same 12 connectivity features predicted individual patient score reductions with reasonable accuracy (r = 0.41, P = 0.002 and MSE = 0.036) (Fig. 5a) and high statistical significance according to 5000-permutations testing (P = 0.003) (Fig. 5b). Among them, the correlation of the positive features was r = 0.37, P = 0.006 (Fig. 5c) while the negative features were r = −0.11, P = 0.441 (Fig. 5d). This result supporting the generalizability of our prediction model to similarly designed clinical studies.

Fig. 5: The connectome-based features and their prediction performance in the validation cohort.
figure 5

a Scatter plot showing the correlation between predicted and true PANSS total score reduction in the validation cohort. b The green dashed lines in the permutation histograms indicate the position of the observed correlation. c Scatter plots showing the correlations of positive features with the true PANSS total score reduction in the validation cohort. d Scatter plots showing the correlations of negative features with the true PANSS total score reduction in the validation cohort. The positive and negative features are marked in red and blue, respectively.

Full size image

Discussion

Our connectome-based predictive model identified a set of functional circuit features strongly associated with antipsychotic drug treatment response in SZ. These included parietal intracortical and cerebellar-occipital cortical circuits as positive predictors and parietal-frontal, parietal-thalamic, and cerebellar-cortical circuits as negative predictors. We further demonstrated the general utility of these predictors using an independent validation cohort of patients. Taken together, these results identify a number of promising functional biomarkers that could serve as predictors of antipsychotic drug efficacy for individualized psychiatric treatment.

Resting-state and task-dependent fMRI hold unique advantages for characterizing brain function during various clinical states, and a growing number of fMRI studies have identified predictors of SZ treatment effects [6, 7, 15, 16]. To date, however, rs-fMRI has identified most disease and prognostic biomarker candidates [17] owing to superior temporal stability during scanning [18], higher test retest reliability on repeated scans [19], and the practical benefits of acquiring scans in different clinical settings with limited resources [20]. In addition, rs-fMRI yields a variety of useful metrics, among which FC is widely selected as a clinical predictor because it is strongly correlated with the symptoms of SZ [21]. Moreover, resting-state network metrics are especially useful for the development of predictive models [22]. Thus, our findings based on individual rs-fMRI-derived connectomes may reveal fundamental neural mechanisms regulating the clinical responses to antipsychotic medications in SZ. Another significant advantage of rs-fMRI is the potential clinical translational value. In contrast to task-state fMRI, rs-fMRI can generate metrics that are not influenced by clinicodemographic factors, leading to greater generalizability.

Our study also used an individual-level approach to treatment prediction rather than categorizing patients into a responder group (RG) and non-responder group (NRG) as in most previous studies. Therefore, antipsychotic efficacy could be treated as a continuous variable rather than a binary variable, which is not only subjective but also statistically inefficient [23, 24]. In addition, we chose individual-patient PANSS total score reduction as the predictive label because it encompasses positive, negative, and general psychopathologic symptoms of SZ, thereby providing broader symptomatologic coverage and higher clinical value than positive [9] or negative [25] symptoms alone as used in many previous studies.

Testing the feature weights identified intraparietal connections, particularly connections involving the PoG, as the strongest predictors of the model. This is the new report suggesting that intraparietal connectivity with PoG influences antipsychotic medication efficacy in SZ, although previous studies have identified associations with PoG activity. For example, Yao and colleagues found a significant difference in PoG regional homogeneity between RG and NRG groups post-medication [26], while Cui and coworkers reported significantly higher baseline amplitude of low-frequency fluctuations in the PoG among RG compared to NRG [27]. In the current study, connections involving the PoG were all positive features, implying that stronger connectivity at baseline was associated with better drug response. The PoG, also known as the somatosensory cortex, is the primary receptive area for tactile and kinesthetic sensations and plays an important role in the sensorimotor network (SMN) involved in motion selection and execution, shape perception, color processing, stereotactic orientation, and depth perception [28]. Defects in the PoG may thus result in persistent disturbances within the SMN, leading to clinical symptoms of SZ [29]. We speculate that maintaining stronger connectivity between the PoG and interior parietal lobe may be a compensatory response that diminishes the effect of impaired somatosensory input on the SMN and facilitates the recovery of sensory deficits after treatment. Regardless of underlying cause, the current results suggest that FC with the PoG predicts the outcome of drug therapy in SZ patients.

With the exception of parietal internal connections, the other predictors identified in this study have been reported in previous neuromarker studies, including fronto-parietal and cerebello-thalamo-cortical (CTC) connections [6, 16, 30]. Several studies have reported reduced dopamine release in certain areas of the frontal lobe [31] and reduced expression of frontal lobe dopamine D1, D3, and D4 receptors linked to negative symptom expression [32, 33]. Reduced dopamine activity in the frontal lobe may lead to overactivity of midbrain dopaminergic neurons projecting to the limbic system, resulting in positive symptoms [34]. The parietal cortex is critical for generating mental imagery, ‌thus two core symptoms of SZ, cognitive deficits and hallucinations, may be associated with parietal dysfunction [35,36,37]. In light of these characteristics, connectivity of the frontoparietal control network (FPN) may be a good predictor of SZ onset or symptom exacerbation. Indeed, a study has found that inter-network connections of the FPN make the largest contribution to predicting positive symptom changes after medication, while connections of the FPN with other networks also contribute significantly to predicting negative symptom changes after medication [38]. In our study, frontoparietal connections were predominantly components of negative features (those for which weaker connectivity at baseline was associated with better antipsychotic drug efficacy). This may be because dopamine gates information input and output for balancing goal-directed signals and background noise [39, 40], this stronger connectivity may reflect dopamine-associated dysregulation of the cognitive system within frontoparietal cortex. Impaired dopamine signaling could disrupt the capacity of large-scale brain networks to adaptively switch in a context-sensitive manner, leading to the attentional and working memory deficits observed in SZ patients [41, 42]. All of these cognitive deficits fall under the category of reduced cognitive plasticity, and cognitive dysfunction is the primary reason for prolonged therapy [43].

Stronger connectivity within CTC loops was also negatively associated with clinical response (lower baseline connectivity predicted better efficacy), consistent with previous reports that increased FC within CTC circuits may be an inherent, state-independent neural feature of SZ [44, 45] and can be used to categorize and predict long-term treatment outcomes [30]. According to the N-methyl-D-aspartate receptor (NMDAR) hypofunction hypothesis, increased CTC connectivity is a downstream effect of decreased NMDAR function [46, 47] that leads to overactivity of pyramidal glutamatergic neurons, which in turn disrupts error-processing signals transmitted between the cerebral cortex and cerebellum [48] and results in a range of abnormal thoughts and behaviors (cognitive deficits) [49]. Since the function of dopaminergic neurons is regulated downstream of glutamate signaling [50], this property may also influence the response to antipsychotic medication.

This study has several limitations. First, the types and dosages of antipsychotics used were not strictly limited, so possible confounding effects caused by different drugs cannot be ruled out. Second, duration of illness was highly variable, which may also confound the results. Future studies exclusively recruiting first-episode unmedicated SZ patients are warranted to validate our results.

In summary, we identified and validated a set of connectome features as individualized predictors of antipsychotic drug response in SZ using rs-fMRI data from two clinically independent patient cohorts. These predictors emphasize the potential contributions of FPN, intraparietal PoG, and CTC circuits to SZ pathology. Additional studies are needed to further validate the translational potential of these neural features for clinical practice.