Agomelatine as augmentation therapy with second‐generation antipsychotics for negative symptoms of schizophrenia: An open‐label randomised clinical trial and electroencephalogram‐derived model for treatment response prediction Authors: Wanyan Zhou, Fei Liang, Nizhe Chen, Yao Zhang, Xiaoxiao Wang, Xin Li, Daofeng Lu, Fang Guo, Yuan Yao, Jiayu Zhu, Chengjia Shen, Chongze Wang, Peiyun Zhang, Lilan Zhao, Ying Wang, Jinyu Han, Wei Li, Haisu Wu, Peijuan Wang, Qi Yan, Ping Zhang, Qinyu Lv, Zhenghui Yi Published: 2026 ABSTRACT ABSTRACTBackgroundNegative symptoms are a core diagnostic domain of schizophrenia that affect functional outcomes and remain difficult to treat. Agomelatine, a novel antidepressant with melatonin receptor agonism and serotonin 2C receptor antagonism, may have therapeutic potential.AimsThis study aimed to evaluate the efficacy of agomelatine combined with second‐generation antipsychotics (SGAs) in reducing negative symptoms and to explore electroencephalogram (EEG)‐based biomarkers for predicting treatment response.MethodsAn open‐label randomised parallel‐group clinical trial was conducted at the Shanghai Mental Health Centre (December 2022 to September 2023), enrolling outpatient adults with schizophrenia and prominent negative symptoms, defined as a positive and negative syndrome scale factor score for negative symptoms (PANSS‐FSNS) higher than a PANSS factor score for positive symptoms (PANSS‐FSPS); PANSS‐FSNS ≥ 24; 7 < PANSS‐FSPS ≤ 28; at least two of three core PANSS negative symptoms ≥ 4 and a Calgary depression scale for schizophrenia score ≤ 18. Patients were randomised (1:1) using a random number table to a group receiving agomelatine (25 mg/day) plus SGAs or a group receiving SGAs alone for 12 weeks. Outcome assessors were blinded after assignment to interventions. The primary outcome was the change in score in the negative symptom factor score (NSFS) at 12 weeks. Longitudinal changes were assessed using linear mixed‐effects models.ResultsSeventy‐four patients were enrolled, with an agomelatine augmentation group (n = 37) and a control group (n = 37). The agomelatine augmentation group demonstrated significantly lower scores on the NSFS compared to the control group starting from Week 8 (estimated marginal mean [EMM] difference = −1.79, 95% confidence interval [CI] −3.00 to −0.59, p = 0.004) and sustaining through Week 12 (EMM difference = −2.65, 95% CI −3.86 to −1.44, p < 0.001). Exploratory support vector machine (SVM) models demonstrated high classification performance in distinguishing treatment responders from non‐responders (area under the curve = 0.91), highlighting notable EEG changes in the left frontotemporal region.ConclusionsAgomelatine augmentation significantly improves negative symptoms, encompassing affective deficits, expressive impairments and social‐withdrawal‐related symptoms. EEG‐based SVM models suggest potential utility for informing future personalised treatment strategies.Trial registration numberNCT05646264. INTRODUCTION Negative symptoms, according to The International Classification of Diseases‐11 (ICD‐11), are recognised as a core diagnostic domain of schizophrenia, alongside positive symptoms, disorganised cognition and behaviour, passivity experiences and psychomotor disturbances. 1 The focus on negative symptoms is driven by their persistent intractable presence throughout the illness course and a strong association with poorer prognosis. 2 Based on descriptive definitions, negative symptoms are divided into diminished emotional expression, avolition, alogia, anhedonia and asociality. 3 These dimensions differentially affect functional outcomes, reflecting multifaceted neurobiological substrates and inherent treatment resistance. Currently, there are no approved treatments that are specifically effective in reducing the negative symptoms of schizophrenia. However, certain second‐generation antipsychotics (SGAs), such as cariprazine, olanzapine, clozapine and amisulpride—used as monotherapy—demonstrate small to moderate effect‐size advantages for negative symptom control. 4 , 5 , 6 This has prompted further research into augmentation therapies, such as the addition of a second agent, cognitive remediation or psychosocial interventions. Antidepressants are currently the most commonly used augmentation therapy to treat negative symptoms, based on demonstrated phenotypic and pathophysiological overlaps between depressive and negative symptom domains. Evidence of dysfunction in the serotonergic (5‐HT) system in the pathophysiology of psychosis suggests that selective serotonin reuptake inhibitors (SSRIs) and serotonin‐norepinephrine reuptake inhibitors (SNRIs) may be beneficial. Meta‐analyses have shown that SSRIs (e.g., citalopram, fluvoxamine and sertraline) and SNRIs (e.g., duloxetine) produce marginal improvements in negative symptoms. 7 , 8 Mirtazapine exhibited limited efficacy, likely due to its complex interactions with histaminergic and adrenergic systems. 9 Modulators targeting the N‐methyl‐D‐aspartate receptor are also being explored, with ongoing clinical trials assessing their potential to address hypothesised deficits in glutamatergic neurotransmission. 4 These findings highlight the critical need for further research into antidepressants with novel mechanisms as augmentation therapy for negative symptoms in schizophrenia. Agomelatine, a relatively newer atypical antidepressant used mainly for major depressive disorder in adults, was approved for clinical use in China in 2011. Agomelatine acts as a melatonin MT1/MT2 receptor agonist and 5‐HT2C antagonist. Activation of melatonin receptors regulates circadian rhythm, whereas 5‐HT2C receptor antagonism increases dopamine and norepinephrine release in the prefrontal cortex. 10 Agomelatine may have therapeutic potential for negative symptoms in schizophrenia. Melatonin, as a myeloperoxidase inhibitor, 11 could reverse ketamine‐induced schizophrenia‐like symptoms, immune inflammation and oxidative stress in mouse models. 12 Magnetic resonance imaging (MRI) studies have revealed that disrupted functional connectivity between the cerebellum and dorsolateral prefrontal cortex in schizophrenia is correlated with negative symptom severity 13 ; significantly, these regions are rich in melatonin MT1/MT2 receptors. 14 The dual mechanism of agomelatine may enhance the prefrontal and hippocampal function associated with negative symptoms by promoting neurogenesis, upregulating brain‐derived neurotrophic factor (BDNF) and increasing glutamate release. 15 , 16 , 17 Adjunctive melatonin therapy has demonstrated beneficial effects on cognitive function, 18 sleep and metabolic profile in patients with schizophrenia, 19 though its efficacy for negative symptoms remains less established. A case series of seven patients with schizophrenia and comorbid depression showed significant reductions in both the Calgary depression scale for schizophrenia (CDSS) and the scale for the assessment of negative symptoms (SANS) scores following agomelatine augmentation treatment. 20 These results were supported by an open‐label prospective study (n = 27) that observed significant improvements in positive and negative syndrome scale (PANSS) negative subscale scores and SANS scores after 6 weeks of treatment with agomelatine added to stable doses of antipsychotic agents. 21 Furthermore, an open‐label uncontrolled trial of agomelatine augmentation to clozapine in partial‐responder schizophrenia (n = 20) found that agomelatine (50 mg/day) produced significant reductions in PANSS negative subscale scores after 16 weeks. 22 The reliability of these therapeutic outcomes is compromised by inherent study limitations, notably underpowered sample sizes and uncontrolled experimental conditions. Emerging evidence suggests that electroencephalogram (EEG) features may be associated with antipsychotic treatment response in schizophrenia 23 ; thus, EEG‐based measures were included as an exploratory component in the present study. EEG provides a direct non‐invasive measure of neural activity, enabling assessment of functional brain dynamics that potentially predict treatment response. Mechanistic studies have shown that alterations in EEG oscillatory patterns, such as frontal theta and beta activity, are associated with clinical improvement and cognitive changes during antipsychotic treatment, supporting their potential as biomarkers of therapeutic efficacy. 23 , 24 Specifically, we employed functional connectivity analysis of EEG data to assess synchronised neural activity across schizophrenia‐relevant brain networks, complemented by support vector machine (SVM) classification of differential entropy features. Differential entropy is used to measure the complexity of a continuous random variable, representing the entropy of a continuous signal. 25 It was selected over traditional EEG power spectral measures because it captures the dynamic information content of EEG signals. 26 Compared with conventional spectral features, differential entropy has shown promise in capturing subtle neural changes relevant to treatment response, 27 , 28 , 29 making it a suitable feature for predictive modelling in schizophrenia. However, evidence for the use of differential entropy to predict treatment response in schizophrenia remains limited. The present study aims to explore its potential in this context. This open‐label randomised controlled trial was set up to evaluate the efficacy and safety of agomelatine augmentation therapy with SGAs for negative symptoms in schizophrenia, while also exploring treatment response prediction using EEG‐based biomarkers. METHODS Participants This is an investigator‐initiated, randomised and controlled 1:1 parallel‐group trial. Participants were recruited from December 2022 to September 2023 at the outpatient department of the Shanghai Mental Health Centre. Patients were eligible if they were aged 18–65, with a confirmed diagnosis of schizophrenia according to ICD‐10 criteria by two experienced psychiatrists, with prominent negative symptoms defined as follows: (1) PANSS‐factor score for negative symptoms (PANSS‐FSNS; sum of items N1–N7) higher than PANSS‐Factor score for positive symptoms (PANSS‐FSPS; sum of items P1–P7); PANSS‐FSNS ≥ 24 and 7 < PANSS‐FSPS ≤ 28; (2) at least two of three core PANSS negative symptoms (items N1, N4 and N6) with a score ≥ 4 and (3) a CDSS score ≤ 18. Patients were eligible only if the psychiatrists deemed that treatment with either SGAs alone or in combination with agomelatine augmentation was clinically appropriate. Patients were excluded if they presented with any major medical or neurological comorbidities, concurrent severe psychiatric disorders, a documented history of substance abuse or alcohol dependence within the past year, intellectual disability, a current pregnancy or lactation status or a requirement for the use of other antidepressants during the study period. Procedures Site investigators were responsible for enrolling participants, obtaining informed consent and managing the randomisation process. Eligible participants were randomly allocated in a 1:1 ratio to either the agomelatine augmentation group (SGAs plus agomelatine 25 mg/day) or the control group (SGAs alone) using a random number table. Although the study employed an open‐label design and participants were aware of treatment allocation, outcome assessments were performed by independent raters who were not involved in treatment assignment or clinical care and were blinded to group allocation. All patients underwent standard assessments of demographic characteristics, medical history, laboratory values, PANSS, SANS and CDSS scores at enrolment, 4, 8 and 12 weeks. Outcomes The negative symptom factor score (NSFS), a refined PANSS subscale comprising items N1–N4 (affective blunting, emotional withdrawal, affective communication impairment, passive/indifferent social withdrawal), N6 (lack of spontaneity and fluency in conversation), G7 (psychomotor retardation) and G16 (active social avoidance), was applied to enhance the assessment of negative symptoms. The primary outcome was the change in the NSFS score from baseline to week 12 following the treatment. Secondary outcomes included changes in PANSS, PANSS‐FSNS, SANS and CDSS total scores from baseline to week 12. The safety outcomes included all‐cause adverse events at 12 weeks. Adverse events were defined as any unfavourable or unintended signs or symptoms occurring after treatment initiation and were collected non‐systematically at each scheduled follow‐up visit via routine clinical interviews and examinations. Treatment response was defined as a reduction of 20% or more in the NSFS score, in accordance with previously established criteria for clinical response in antipsychotic and adjunctive treatment studies. 30 Statistical analysis A linear mixed‐effects model (LMM) was used for the primary outcome analysis. However, because of limited information on within‐subject correlations, the sample size was approximated using a repeated‐measures ANOVA model. Based on prior findings suggesting a large effect size, 22 the effect size was set at f = 0.2. Assuming a correlation among repeated measures of 0.4 and a nonsphericity correction of ε = 0.5, a total sample size of 70 participants was estimated to achieve 80% power at a two‐sided significance level of 0.05 using G*Power (version 3.1.9.7; Heinrich‐Heine‐Universität Düsseldorf). Allowing an anticipated dropout rate of 10%, the target recruitment size was set at 77 participants. All data analyses were performed using SPSS (version 20.0; IBM Corp). Data normality was evaluated using the Kolmogorov–Smirnov test. Normally distributed continuous variables were expressed as mean (standard deviation), and non‐normally distributed data were presented as median (interquartile range). All primary efficacy analyses were conducted according to the intention‐to‐treat (ITT) principle. For longitudinal comparisons, LMMs were applied to include all available data, and results are presented as estimated marginal means (EMMs, standard error). Categorical variables were analysed using Pearson's Chi‐squared test or Fisher's exact test, as appropriate. All statistical tests were two‐tailed, and a significance level of 0.05 was used for all outcomes. The study is registered at ClinicalTrials.gov (NCT05646264). EEG acquisition and processing Resting‐state EEG data were acquired at baseline and 12‐week follow‐up using a Dry Sensor Interface‐24 wireless EEG system. Participants underwent a 5‐min protocol with open‐eye (fixating on a computer screen) and closed‐eye (maintaining alertness) conditions. EEG preprocessing and metric computation were performed in Python (version 3.9.2) using standard signal‐processing libraries. Signal artefacts were removed using a 1‐Hz high‐pass filter to eliminate low‐frequency drift and a 50‐Hz low‐pass filter to suppress high‐frequency noise. Artefact rejection was performed, including removal of electrooculographic, electrocardiographic and electromyographic artefacts. Independent component analysis was employed to isolate artefactual components. At least 1 min of artefact‐free continuous EEG segments per participant was retained for analysis. The primary analytical metric for EEG data was differential entropy, which quantifies the signal energy of fixed‐length EEG segments within specific frequency bands and assesses the complexity of neural oscillations. Differential entropy values were computed for each channel using a 4‐s window and a 2‐s step size, resulting in an 18‐dimensional feature vector. Machine learning methods SVM was employed to differentiate between treatment responders and non‐responders across three EEG recording conditions (open‐eye, closed‐eye or combined) in both the agomelatine augmentation and control groups. To minimise information leakage, training and validation sets were strictly separated at the subject level. In each validation iteration, we randomly selected two responders and two non‐responders to constitute the validation set, utilising the remaining participants for model training. Given the pronounced class imbalance, responder samples (n = 28) were randomly distributed into nine clusters (three per cluster), whereas control samples (n = 29) were organised into six clusters (five per cluster). Within the training set, these clusters were iteratively combined with non‐responder samples to ensure that all responder data contributed to model training. SVM hyperparameter C was tuned within a predefined range (0.01–1.00) using repeated training on the training data only, and the highest mean validation accuracy across three repeated runs was used to select the final model. RESULTS Seventy‐four participants, randomised to the agomelatine augmentation group (n = 37) and the control group (n = 37), were included in the ITT analysis. At 12 weeks, five participants exhibited medication non‐adherence according to the study protocol; 69 participants (33 in the agomelatine augmentation group and 36 in the control group) were included in per‐protocol (PP) analysis (figure 1). There were no statistically significant differences in demographics or baseline characteristics between the two groups (table 1). The median time from enrolment to receiving treatment did not differ between the agomelatine augmentation and control groups. Flowchart of the study enrolment and analysis. TABLE 1Demographics and baseline characteristicsCategoryAgomelatine augmentation group (n = 37)Control group (n = 37) t/χ 2 p‐valueGenderMale25260.063 a 0.802Female1211Age (years)Range19–3718–391.075 b 0.286Mean (SD)25.41 (4.49)24.16 (5.45)Height (cm)Range150.00–191.00160.00–185.00−0.369 b 0.713Mean (SD)170.27 (7.39)170.86 (6.45)Weight (kg)Range37.30–119.0052.50–115.00−0.096 b 0.924Mean (SD)69.70 (14.59)70.01 (13.70)BMIRange17.82–41.1817.93–38.420.068 b 0.946Mean (SD)23.99 (4.62)23.91 (4.15)Years of educationRange8.00–16.009.00–16.000.451 b 0.654Mean (SD)12.27 (1.97)12.08 (1.64)Marital statusUnmarried32350.631 c 0.427Married52Smoking historyNegative37352.028 a 0.154Positive02Alcohol use historyNegative3536< 0.001 c 1.000Positive21Family history of mental disordersNegative21< 0.001 c 1.000Positive3536Chlorpromazine equivalent dose for second‐generation antipsychotics (mg)251.17 (88.77)254.77 (85.42)−0.178 b 0.859 Note: Chlorpromazine equivalents were calculated uniformly according to Leucht et al. 31 Abbreviations: BMI, body mass index; SD, standard deviation. a Uses Pearson's Chi‐squared test. b Uses the independent samples t‐test. c Uses Yates' continuity correction. Demographics and baseline characteristics Note: Chlorpromazine equivalents were calculated uniformly according to Leucht et al. 31 Abbreviations: BMI, body mass index; SD, standard deviation. Uses Pearson's Chi‐squared test. Uses the independent samples t‐test. Uses Yates' continuity correction. LMM analysis revealed a significant group‐by‐time interaction on NSFS scores (F = 11.728, p < 0.001). Although no significant differences were observed at baseline (p = 0.822) or Week 4 (p = 0.217), the agomelatine augmentation group demonstrated significantly lower scores compared to the control group starting from Week 8 (EMM difference = −1.79, 95% confidence interval [CI] −3.00 to −0.59, p = 0.004) and persisting through Week 12 (EMM difference = −2.65, 95% CI −3.86 to −1.44, p < 0.001). A significant main effect of time was observed (F = 269.168, p < 0.001), indicating a significant overall change over time across both groups (table 2). TABLE 2Efficacy outcome measures (intention‐to‐treat)Follow‐upAgomelatine augmentation group (n = 37)Control group (n = 37)Difference (95% CI) p‐value (between‐group) F‐value (p‐value)TimeGroupTime × groupNSFSBaseline25.81 (0.42)25.68 (0.42)0.14 (−1.05 to 1.32)0.822269.168 (< 0.001**)5.837 (0.018*)11.728 (< 0.001**)Week 422.85 (0.43)23.59 (0.42)−0.75 (−1.94 to 0.44)0.217Week 820.39 (0.43)22.19 (0.43)−1.79 (−3.00 to −0.59)0.004*Week 1217.56 (0.44)20.21 (0.43)−2.65 (−3.86 to −1.44)< 0.001**CDSSBaseline6.27 (0.25)6.11 (0.25)0.16 (−0.55 to 0.87)0.651398.386 (< 0.001**)0.048 (0.827)2.467 (0.063)Week 45.05 (0.25)5.27 (0.25)−0.22 (−0.94 to 0.49)0.533Week 84.21 (0.25)4.33 (0.25)−0.11 (−0.82 to 0.60)0.756Week 123.67 (0.25)3.80 (0.25)−0.13 (−0.84 to 0.58)0.718SANSBaseline56.97 (0.88)58.51 (0.88)−1.54 (−3.99 to 0.91)0.217376.058 (< 0.001**)24.152 (< 0.001**)13.443 (< 0.001**)Week 449.57 (0.89)53.65 (0.88)−4.08 (−6.55 to −1.62)0.001*Week 841.73 (0.90)48.77 (0.88)−7.04 (−9.53 to −4.55)< 0.001**Week 1235.94 (0.91)43.88 (0.88)−7.94 (−10.45 to −5.44)< 0.001**PANSSBaseline71.84 (1.10)70.54 (1.10)1.30 (−1.80 to 4.39)0.408252.303 (< 0.001**)2.612 (0.110)11.576 (< 0.001**)Week 463.64 (1.11)66.76 (1.10)−3.12 (−6.23 to −0.02)0.049*Week 860.21 (1.12)63.31 (1.11)−3.10 (−6.22 to 0.03)0.052Week 1255.14 (1.12)59.50 (1.11)−4.36 (−7.49 to −1.23)0.007*PANSS‐FSNSBaseline26.68 (0.37)26.89 (0.37)−0.22 (−1.24 to 0.81)0.677342.155 (< 0.001**)23.378 (< 0.001**)18.829 (< 0.001**)Week 423.13 (0.37)24.81 (0.37)−1.68 (−2.71 to −0.65)0.002*Week 820.53 (0.38)23.18 (0.37)−2.65 (−3.69 to −1.61)< 0.001**Week 1217.33 (0.38)21.10 (0.37)−3.76 (−4.81 to −2.72)< 0.001** Note: Data are presented as EMM (SE) or EMM difference (95% CI).Abbreviations: CDSS, Calgary depression scale for schizophrenia; CI, confidence interval; EMM, estimated marginal mean; NSFS, negative symptom factor score; PANSS, positive and negative syndrome scale; PANSS‐FSNS, PANSS factor score for negative symptoms; SANS, scale for the assessment of negative symptoms; SE, standard error.*p < 0.05, **p < 0.01. Efficacy outcome measures (intention‐to‐treat) Note: Data are presented as EMM (SE) or EMM difference (95% CI). Abbreviations: CDSS, Calgary depression scale for schizophrenia; CI, confidence interval; EMM, estimated marginal mean; NSFS, negative symptom factor score; PANSS, positive and negative syndrome scale; PANSS‐FSNS, PANSS factor score for negative symptoms; SANS, scale for the assessment of negative symptoms; SE, standard error. *p < 0.05, **p < 0.01. For the PANSS total score, LMM analysis revealed a significant group‐by‐time interaction (F = 11.576, p < 0.001). The agomelatine augmentation group demonstrated significantly lower PANSS scores compared to the control group at Week 12 (EMM difference = −4.36, 95% CI −7.49 to −1.23, p = 0.007). For the PANSS‐FSNS score and SANS total score, the group‐by‐time interactions were significant (F = 18.829 and F = 13.443, respectively; both p < 0.001), and between‐group differences at Week 12 showed a similar pattern to that observed for the PANSS total score. For the CDSS total score, only the main effect of time was significant (F = 398.386, p < 0.001). The scores of PANSS items related to NSFS are shown in Supporting Information S1: table S1. The safety analysis showed that the overall incidence of adverse reactions in the agomelatine augmentation group was lower than that in the control group, but the difference was not statistically significant (Supporting Information S1: table S2). At Week 12, the response rate was 84.85% in the agomelatine augmentation group (28 responders and 5 non‐responders) and 80.56% in the control group (29 responders and 7 non‐responders). The Pearson Chi‐squared test indicated no significant difference between the groups (χ 2 = 0.22, p = 0.638), consistent with Fisher's exact test (p = 0.755). Results for all endpoints in the PP dataset are presented in Supporting Information S1: table S3. The SVM model demonstrated strong classification performance in distinguishing responders from non‐responders with respect to negative symptom improvement under both open‐eye and closed‐eye conditions, as evidenced by receiver operating characteristic analysis (figure 2). The predictive efficacy of the support vector machine model. The receiver operating characteristic curve illustrates the model's performance metrics across different conditions—open‐eye, closed‐eye, or combined (open‐eye and closed‐eye electroencephalogram data combined for model training)—in both the agomelatine augmentation and control groups. ROC, receiver operating characteristic. In the agomelatine augmentation group, model performance was strongest under the open‐eye condition (accuracy = 0.76, 95% CI 0.73–0.80, recall = 0.71 and F1‐score = 0.80), followed by the closed‐eye condition (accuracy = 0.78, 95% CI 0.75–0.81, recall = 0.70 and F1‐score = 0.81), and the combined condition showed reduced predictive power (accuracy = 0.62, 95% CI 0.59–0.65, recall = 0.61 and F1‐score = 0.65). The corresponding area under the curve (AUC) values were 0.91, 0.80 and 0.66, respectively. In the control group, the closed‐eye condition yielded optimal performance (accuracy = 0.84, 95% CI 0.81–0.86, recall = 0.96, F1‐score = 0.82 and AUC = 0.90), with poorer results in the open‐eye condition (accuracy = 0.62, 95% CI 0.58–0.66, recall = 0.69, F1‐score = 0.54 and AUC = 0.65) and the combined condition (accuracy = 0.57, 95% CI 0.55–0.60, recall = 0.67, F1‐score = 0.39 and AUC = 0.60). The Kappa values, specificity values and precision values are shown in Supporting Information S1: table S4. Using the machine learning results, we generated a brain topographic map to visualise the temporally aggregated mean differential entropy. The 18‐dimensional feature vector was spatially mapped according to anatomical location and displayed as a brain topography plot. Figure 3A–D display the brain topographic map, where the dataset was divided into 8 groups based on experiment/control, open‐eye/closed‐eye and responder/non‐responder conditions. The mean differential entropy feature for each group was calculated, with the rightmost position in each group representing the difference between the responder and non‐responder groups. Colour intensity reflects the relative significance of brain activity patterns, with shades approaching purple indicating higher levels of topographic significance, whereas green represents lower levels of significance. Brain topographic map and EEG connectivity maps. (A) Control group under the closed‐eye condition. (B) Control group under the open‐eye condition. (C) Agomelatine augmentation group under the closed‐eye condition. (D) Agomelatine augmentation group under the open‐eye condition. (E) EEG connectivity maps of responders. (F) EEG connectivity maps of non‐responders. EEG, electroencephalogram. The connectivity maps illustrate the degree of correlation between any two EEG electrodes. Similar to the mean differential entropy feature map, the dataset was divided into eight groups. Connectivity was calculated for each participant within each group, using a frequency range of 1–50 Hz. The resulting correlation matrices were then averaged across frequencies, yielding the 8 connectivity maps shown in figure 3E, F. In these maps, line colours denote the magnitude of correlation coefficients, with lighter colours (closer to white) representing stronger functional connectivity between electrode pairs. The spatial arrangement corresponds to standard EEG electrode positions. Across the five resting‐state frequency bands, the spatial distribution of differential entropy features was generally consistent. In conditions with higher predictive performance, specifically the open‐eye and closed‐eye conditions in the agomelatine augmentation group and the closed‐eye condition in the control group, differences in differential entropy between responders and non‐responders were more pronounced. These differences were primarily observed as lower entropy in the occipital and right temporal regions and higher entropy in the left prefrontal and left temporal regions, suggesting that baseline entropy patterns in these regions may contribute to discrimination of treatment response. In conditions with lower predictive performance, differences in entropy between responders and non‐responders were less distinct, indicating that these spatial patterns may be relevant when the SVM model predicts treatment outcomes. Functional connectivity analyses revealed that, across all participants, connectivity in the occipital and right temporal regions tended to be higher, whereas connectivity in the left prefrontal and left temporal regions tended to be lower. Although these patterns broadly corresponded to the spatial distribution of differential entropy features, the differences in connectivity between responders and non‐responders were relatively small. DISCUSSION This randomised controlled trial provides robust evidence supporting the efficacy of agomelatine augmentation therapy for negative symptoms in schizophrenia, with parallel advancements in EEG‐based predictive modelling. Our findings demonstrate the following three key insights: (1) Agomelatine combined with SGAs significantly outperformed SGA monotherapy in reducing NSFS scores from week 8 onward, indicating improvements across negative symptom domains including blunted affect, emotional and social withdrawal, reduced spontaneity of speech, psychomotor retardation and social avoidance; (2) exploratory analyses using resting‐state EEG differential entropy features combined with SVM algorithms demonstrated high classification performance in distinguishing responders from non‐responders and (3) preliminary neurophysiological differences were observed between responders and non‐responders, particularly in left frontal‐temporal regions under closed‐eye conditions. These results align with emerging neurobiological frameworks for negative symptoms and address critical gaps in treatment personalisation. At the pathophysiological level, the negative symptoms of schizophrenia and depressive symptoms demonstrate distinct yet interconnected pathophysiological mechanisms. Negative symptoms in schizophrenia and depressive symptoms may share common neurobiological pathways, particularly involving dopamine and serotonin systems. Moreover, biological studies have reported increased levels of pro‐inflammatory cytokines and oxidative stress markers, as well as their association with disorder‐specific genetic risk, in both sets of symptoms. 22 , 32 , 33 Neuroimaging findings further elucidate these relationships. Functional connectivity studies revealed symptom‐specific neural circuits: Negative symptoms are associated primarily with alterations in the ventral lateral prefrontal cortex‐striatum‐thalamus network, whereas depressive symptoms involve mainly the left motor cortex‐parietal lobe circuitry. 34 Cross‐diagnostic structural analyses identified both shared grey matter reductions in the orbitofrontal cortex across schizophrenia and major depressive disorder, and disorder‐specific volumetric changes. 35 These findings demonstrate that negative symptoms and depressive symptoms are interconnected through shared pathophysiological mechanisms involving neurotransmitter dysregulation, inflammatory pathways, genetic risk factors and functionally integrated neural circuitry. However, negative symptoms are increasingly recognised as a partially independent psychopathological construct, characterised by distinct neural circuits and treatment responsiveness. Despite these distinctions, the phenotypic overlap between depressive and negative symptoms presents a potential therapeutic rationale for the use of antidepressants in schizophrenia, particularly given the limited efficacy of current antipsychotic medications in managing negative symptoms. Notably, SSRI and SNRI augmentation has shown greater efficacy for negative symptoms when combined with first‐generation antipsychotics compared to SGAs. This differential response can be explained by the hypothesis that serotonergic mechanisms may partially mediate negative symptoms, as SGAs already combine both dopaminergic and serotonergic activity through 5‐HT2A/2C receptor antagonism. Consequently, the reduced efficacy of adjunctive antidepressants with SGAs may stem from pharmacological redundancy in serotonin‐targeted pathways, prompting researchers to explore antidepressants with non‐serotonergic mechanisms of action. The efficacy of agomelatine likely stems from its dual mechanism of action as a melatonin MT1/MT2 receptor agonist and 5‐HT2C antagonist. Preclinical evidence suggests MT1/MT2 agonism enhances prefrontal–hippocampal neuroplasticity by upregulating BDNF and glutamate release, potentially counteracting the mesocortical dopaminergic hypofunction implicated in negative symptoms. This aligns with our EEG findings showing left frontal–temporal differential entropy differences in responders, regions that have been implicated in both negative‐symptom circuitry and melatonin‐related neurophysiological regulation. However, causality cannot be inferred from EEG topography alone, and whether these neurophysiological changes reflect MT1/MT2 receptor activity requires confirmation in future mechanistic studies. The spatial patterns of differential entropy and functional connectivity may reflect underlying neurophysiological mechanisms of treatment response. Lower occipital‐right temporal and higher left prefrontal‐left temporal entropy in the agomelatine group were associated with improved negative symptom outcomes, suggesting that baseline cortical dynamics may contribute to treatment responsiveness. Functional connectivity broadly aligned with these patterns, though differences between responders and non‐responders were modest. Incorporating these EEG features into SVM models allowed accurate prediction of individual treatment response, highlighting their potential as objective biomarkers to guide personalised interventions. These findings provide preliminary evidence supporting a potential approach for identifying patients who might benefit from agomelatine augmentation, which is particularly relevant given the 20% non‐response rate observed in our cohort. Notably, previous research has highlighted sex differences in schizophrenia across epidemiological, clinical and neurobiological domains. 36 Although the present study was insufficiently powered to examine sex‐specific effects, future research with more balanced samples should incorporate sex‐based or interaction analyses to clarify the impact of sex on treatment response and EEG‐based biomarkers. The results of this study align with similar machine learning studies in the field. One study recruited 37 patients with schizophrenia who were treated with clozapine and used kernel partial least squares regression to predict clozapine efficacy from baseline resting‐state EEG features, achieving an accuracy rate of 85%. 37 In another study using SVM to construct MRI efficacy prediction models, the functional feature model achieved 80.38% accuracy, whereas the structural feature model achieved 69.68% accuracy. 38 It is also noteworthy that relatively high response rates were observed in both the agomelatine augmentation group and the control group in the present study. This finding may be partly attributable to the predefined responder criterion, which is defined as a ≥ 20% reduction in the NSFS score. Compared with more stringent response thresholds adopted in some previous trials, this criterion may be more sensitive to early or partial improvements in negative symptoms, thereby yielding higher overall response rates. In addition, the inclusion of clinically stable patients receiving optimised SGA treatment may have further contributed to the favourable response observed across both groups. These methodological considerations should be considered when comparing response rates across studies. The temporal progression of symptom improvement warrants emphasis: significant reductions in NSFS were observed at week 8. This delayed response aligns with agomelatine's neurotrophic effects, requiring weeks to manifest, contrasting with immediate dopamine‐mediated antipsychotic actions. Such temporal dissociation supports the NSFS as a specific metric for negative symptoms. Limitations The study has several limitations that should be considered. Although randomisation minimised selection bias at group assignment, the open‐label design and lack of a placebo control may have introduced expectancy effects. Furthermore, although outcome raters were blinded to treatment allocation, residual observer bias cannot be entirely excluded. These factors should be considered when interpreting the clinical outcomes. In addition, participants received different SGAs, and residual confounding related to heterogeneous antipsychotic mechanisms cannot be entirely ruled out. Future studies with tightly controlled antipsychotic regimens will be necessary to further clarify the specific contributions of individual SGAs from adjunctive agomelatine effects. Moreover, only resting‐state EEG was collected; consequently, task‐activated networks implicated in reward processing remain unexplored. Finally, the EEG biomarker component was exploratory and not pre‐specified in the trial registration. From a clinical translation perspective, although the EEG‐based model demonstrated acceptable discriminative performance, its utility for individualised decision‐making remains limited, and validation in larger, prospectively collected cohorts is needed to enable a more comprehensive evaluation of model reliability. The relatively small and imbalanced sex distribution in our cohort also precluded adequately powered analyses of potential sex differences. Expanding the study population to include a broader range of individuals, including children and older adults, could further enhance the generalisability of the findings. 39 , 40 Future studies should adopt triple‐blind designs with active comparators (e.g., cognitive behavioral therapy augmentation) to isolate agomelatine‐specific effects. Multimodal imaging combining EEG with functional MRI could elucidate how melatonin receptor modulation translates to macroscale network changes. Longitudinal tracking of differential entropy features may further optimise SVM models for dynamic treatment adjustment. Conclusion Agomelatine augmentation demonstrates statistically and clinically significant improvements in schizophrenia negative symptoms, encompassing affective deficits, expressive impairments and social‐withdrawal‐related symptoms, compared to SGA monotherapy, with effects emerging at week 8 and peaking by week 12. SVM classification of resting‐state EEG differential entropy features achieves high predictive accuracy, offering a preliminary framework for future development of EEG‐informed personalised treatment strategies. Neurophysiological divergence in left frontotemporal entropy between responders and non‐responders suggests agomelatine may modulate distinct neural circuits, warranting further mechanistic investigations. AUTHOR CONTRIBUTIONS Wanyan Zhou: clinical investigation; writing of original draft. Fei Liang and Nizhe Chen: writing of original draft. Yao Zhang, Xiaoxiao Wang, Xin Li, Daofeng Lu, Fang Guo and Yuan Yao: clinical investigation and EEG data acquisition. Jiayu Zhu, Chengjia Shen and Chongze Wang: EEG data preprocessing and statistical analysis. Peiyun Zhang, Lilan Zhao, Ying Wang, Jinyu Han and Wei Li: data collection and curation. Haisu Wu, Peijuan Wang, Qi Yan and Ping Zhang: writing—review and editing. Qinyu Lv and Zhenghui Yi: conceptualization and study design, writing, review and editing. FUNDING This work was supported by multiple grants: STI 2030‐Major Projects (2022ZD0208500); the Fundamental Research Funds for the Central Universities (2024GDND02); Health Commission of Nantong City of China (MSZ2024074); Shanghai Science and Technology Innovation Action Plan Natural Science Fund Project (21ZR1455400); Shanghai Jiao Tong University ‘Jiao Tong University Star’ Programme Medical‐Industrial Crossover Research Fund Project (YG2019QNB07); Elderly Health Project of the Health Commission of Jiangsu Province, China (LKM2022062); Health Commission of Nantong City, China (MSZ2023061) and Science Foundation of Kangda College of Nanjing Medical University (KD2024KYJJ306). The funders had no role in the design, data collection, analysis, interpretation or writing of this study, despite the authors' affiliations with the funders. ETHICS STATEMENT The study was approved by the Ethics Committee of Shanghai Mental Health Centre. INFORMED CONSENT Written informed consent was obtained from all participants or their legally authorised representatives at enrolment. TRIAL REGISTRATION The trial was registered at ClinicalTrials.gov (NCT05646264). Supporting information Supporting Information S1