Introduction

Antipsychotic (AP) medication is the gold standard in schizophrenia treatment. Unfortunately, overall estimations suggest that a fifth to half of patients do not respond to APs and are classified as having a treatment-refractory form of the disorder, also known as treatment-resistant schizophrenia (TRS) [1]. However, a potential issue in the field is the large heterogeneity in the criteria used to define treatment resistance differentially to treatment responsiveness [2, 3]. Since clozapine is the only medication with an indication from the US Food and Drug Administration (FDA) for TRS, clozapine treatment is the most widely applied criterion that is used as a proxy for TRS [1]. The consensus definition of TRS establishes the disease as showing a reduction of less than 20% in positive symptomatology after at least two trials of non-clozapine APs administered at adequate doses (i.e., equivalent to 400–600 mg of chlorpromazine per day) and duration (i.e., > 6 weeks) and with confirmed adherence [3].

Despite a relative consistency in the major aspects of the definition of TRS, it is not readily recognized as a distinct clinical entity in daily practice. The existence of confounding factors such as AP non-adherence, the heterogeneous nature of TRS (e.g., varying onset of treatment failure and the involvement of non-responsive positive, negative, and cognitive symptoms), and the lack of consensus on clinically relevant criteria for defining and treating TRS complicate the identification of TRS [1,2,3,4,5,6]. As a result, appropriate treatment is delayed or not offered at all. Current data show that a longer duration of untreated psychosis and the administration of multiple different AP treatments are indicators of poor prognosis [2, 7]. In this regard, early identification is a critical element in optimizing the treatment of TRS. Through early identification, the duration of inadequately controlled illness may be reduced, thereby improving long-term outcomes. Accordingly, although it is still poorly understood, clarifying the neurobiology of TRS and the consequent obtention of reliable biomarkers of the disease could allow an earlier identification of TRS and help reduce the duration of persistent psychosis and improve long-term treatment outcomes.

Several findings from a diverse range of research fields support the existence of distinct biological subtypes of schizophrenia, with TRS being one of them [8, 9]. Several lines of evidence highlight the potential existence of a distinct genetic susceptibility in TRS patients that could involve the activity of various neurotransmitters, such as dopamine, glutamate, and serotonin, with several pathways converging and possibly contributing to its neurobiology [8]. Therefore, genetic studies could yield valuable insights into the genetic differences underlying TRS. Several recent studies have attempted to identify the genetic and molecular differences underlying the TRS phenotype, using clozapine treatment as a proxy for TRS in pharmacogenetic or pharmacogenomic approaches [10,11,12,13]. Pardiñas et al. [14] recently demonstrated that TRS is a complex phenotype and has a polygenic heritability associated with common risk alleles [14].

In this context, gene expression profiling studies could be a valuable tool in identifying the specific genes and pathways involved in the mechanism of action of clozapine, leading to a better understanding of the molecular biology underlying TRS. To date, no study has investigated gene expression in the human brain specifically during clozapine treatment. The only gene expression analysis in the context of clozapine treatment was performed in whole blood [15]. There is a need for brain gene expression studies of clozapine treatment and, in dorsolateral prefrontal cortex (DLPFC) due to its role in cognitive functioning and emotional processing, both impaired in schizophrenia patients [16]. Further gene expression profiling in human postmortem brains from patients treated with clozapine could potentially help elucidate the neurobiology of TRS and, consequently, identify new biomarkers that could allow the development of personalized strategies and early interventions in patients with TRS.

Therefore, to explore the genomic architecture beyond clozapine mechanism of action, as a proxy for TRS, we analyzed gene co-expression modules in the DLPFC of schizophrenia subjects treated with APs that specifically included clozapine or not and compared the results with those obtained in DLPFC of control subjects. This experimental design aimed to identify the co-expressed modules (clusters of genes with highly correlated expression) that reflect the genetic differences between clozapine-treated and non-clozapine-treated patients with schizophrenia.

Methods

Subjects

The study included brain samples of 26 subjects with an antemortem diagnosis of schizophrenia based on DSM and ICD criteria. Samples were obtained at autopsies performed at the Basque Institute of Legal Medicine (IVML, Bilbao, Spain). The study was carried out in accordance with the Spanish ethical standards for postmortem brain studies (give informed consent and their anonymity was preserved) and with the Declaration of Helsinki. The study was approved by the research ethics committee of the Hospital Clinic (Barcelona, Spain) (HCB/2020/1415).

Samples of the DLPFC were collected, dissected, and immediately frozen and stored at −70 °C. The causes of death were suicide (n = 12), natural (n = 10), accidental (n = 3), or homicide (n = 1). Information on possible diagnosis and treatment was obtained from the medical examiner’s report and antemortem medical records. A blood toxicology screen at the time of death (detection of antidepressants, antipsychotics, psychotropic drugs, and ethanol) was performed at the National Institute of Toxicology, Madrid, Spain. All the reported results were confirmed by ultrahigh-performance liquid chromatography coupled to tandem mass spectrometry (UPLC-MS/MS). Clozapine was detected in the blood samples of 13 subjects with schizophrenia (clozapine-treated group). Nine patients were treated with clozapine in monotherapy, two were treated in combination with quetiapine, one in combination with sulpiride, and one in combination with paliperidone. Each clozapine-treated subject with schizophrenia was carefully matched with a clozapine-free subject with schizophrenia (non-clozapine-treated group) for sex, age, the postmortem interval (PMI; time interval between death and autopsy), and, whenever possible, storage time (Table 1). Non-clozapine-treated patients were treated with quetiapine (n = 5), olanzapine (n = 3), clotiapine (n = 3), levomepromazine (n = 1), or a combination of quetiapine and olanzapine (n = 1). This cohort had not been previously included in other studies.

Table 1 Characteristics of the study participants.
Full size table

Sample collection, RNA isolation, and microarray hybridization

Total RNA was extracted from all human brain samples using a commercial RiboPure™ kit (Thermo Fisher Scientific, Waltham, MA, USA) according to the manufacturer’s instructions. RNA concentration and quality (260/280 and 260/230 absorbance ratios, respectively) were measured in a NanoDrop 1000 Spectrophotometer (Thermo Fisher Scientific). The RNA integrity number (RIN) was also assessed for RNA quality in an Agilent 2100 Bioanalyzer (Agilent, Santa Clara, CA, USA), using an Agilent RNA 6000 Nano Kit and RNA Nano chips according to the manufacturer’s instructions. The mean RIN was 6.95 ± 1.11. One sample was discarded due to a low RIN number (RIN < 5). Samples from matched clozapine-treated and non-clozapine-treated subjects with schizophrenia were always processed in parallel.

A total of 1 μg of purified RNA from each of the samples was submitted to the Kompetenzzentrum für Fluoreszente Bioanalytik Microarray Technology (KFB, BioPark Regensburg GmbH, Regensburg, Germany) for labeling and hybridization to the Clariom S Human Array (Affymetrix, Santa Clara, CA, USA), following the manufacturer’s protocols. The Clariom S Human Array comprises more than 211,300 probes covering over 337,100 transcripts and variants, which represent 20,800 genes. Microarray data was deposited in the GEO database (GSE244782).

Genome-wide expression analysis and the WGCNA procedure

Microarray data preprocessing was performed using the Oligo package and the package:clariomshumantranscriptcluster.db in R [17]. The data were standardized using a robust multichip analysis. Microarray probes were mapped to an Ensembl gene ID using the biomaRt package in R [18]. Multiple probes mapping to the same gene were merged using the average as the summary of the hybridization values. Probes that do not match known genes were discarded from the analysis. The obtained data matrix included 18,478 genes. Unwanted sources of variability were removed with the sva R package [19].

Co-expression modules were identified using the R package for WGCNA [20]. Firstly, to remove outlier samples, the distance-based adjacency matrices of the samples were estimated and the sample network connectivity according to the distances was standardized. Samples with a connectivity less than –5 would have been considered outliers and excluded. In our study, no samples were excluded. The co-expression analysis involved constructing a matrix of pairwise correlations between all the pairs of genes across all the selected samples. Next, the matrix was raised to a soft-thresholding power (β = 10 in this study) to obtain an adjacency matrix (Supplementary Figure S1). To identify modules of co-expressed genes, we constructed a topological overlap-based dissimilarity, which was then used as input to average linkage hierarchical clustering. This step resulted in a clustering tree (dendrogram), whose branches were identified for cutting based on their shape using the dynamic tree-cutting algorithm (Supplementary Figure S2). The above steps were performed using the automatic network construction and module detection function (blockwiseModules in WGCNA), with the following parameters: minModuleSize of 30, reassignThreshold of 0, and mergeCutHeight of 0.25.

Validation of the gene co-expression network

To assess whether the resulting co-expression modules were robustly defined in our cohort, we performed a subsampling analysis (Supplementary Figure S3). This analysis consisted of a network construction and module identification using the previous parameters, with 50 iterations including randomly drawn individuals, as implemented in the sampleBlockwiseModules function in the WGCNA R package. For each gene, consistency was calculated as the percentage of iterations in which it was assigned to the original module. Finally, the stability of each module was defined as the average gene consistency of all the genes in the given module.

An external validation was performed, assessing the replication of the identified co-expression modules in the DLPFC of large cohorts of subjects with schizophrenia, including those from studies by Fromer et al. [21] (n = 159 patients with schizophrenia) [21] and Gandal et al. [22] (n = 258 patients with schizophrenia) [22]. To this end, we tested the degree of overlap, using the userListEnrichment function of the WGCNA package. This function measures list enrichment between inputted lists of genes and files containing user-defined lists of genes. Significant enrichment is measured using a hypergeometric test. P-values were corrected for multiple comparisons using Bonferroni method.

Functional validation

Cell type-specific expression analysis of each module was performed using the userListEnrichment function of the WGCNA package and three pre-made sets of brain-derived enrichment lists: Cahoy (definite (10+ fold) and probable (1.5+ fold) enrichment from Cahoy et al. [23]), CTX (modules from the cortex network from Oldham et al. [24]), and HumanMeta (modules from the human network from Miller et al. [25]).

We also tested the enrichment of the modules with genes significantly associated with TRS, schizophrenia, and educational attainment in large public genome wide association studies (GWAS) [14, 26, 27]. To this end, the single nucleotide polymorphisms (SNPs) of the GWAS summary statistics with p < 0.01 were annotated to genes according to their chromosome position using Magma v1.06 [28]. The window of annotation was set at 5 and 1.5 kilobases upstream and downstream of the target gene, respectively.

Finally, to characterize the modules that were significantly associated with clozapine treatment, the genes from significant modules were imported to ClueGO v2.1 for a gene set enrichment analysis [29]. The biological processes of the Gene Ontology (GO) database were selected for the enrichment analysis. The genes involved in each network were mapped to their enriched biological processes based on the hypergeometric test (two-sided).

Enrichment analysis p-values were corrected using the Benjamini-Hochberg method (adjusted p values <0.05 were considered significant).

Statistical analysis

Means and standard deviations were computed for continuous variables. The normality of continuous variables was tested using the Kolmogorov–Smirnov and Shapiro–Wilk tests, and the equality of the variance between groups was assessed using Levene’s test. To generate association statistics reflecting the differences in each module between clozapine-treated and non-clozapine-treated schizophrenia subjects, we used generalized linear mixed models with the module eigengenes (MEs, defined as the first principal component of each module) as the independent variables, and sex, age, RIN number, cause of death and PMI as the covariates.

Results

We identified 13 modules of co-expressed genes (Supplementary Figure S2). The inferred modules showed different sizes ranging from 34 (tan module) to 4445 genes (turquoise module). A further 129 genes were assigned to the gray module, which represents the genes that were not co-expressed based on gene dissimilarity. Importantly, the organization of our co-expression modules was robustly defined in our cohort (Supplementary Figure S3) and significantly overlapped with the modules originally reported in the DLPFC of patients with schizophrenia (Supplementary Table S1). The 12 modules identified in our analysis (excluding the gray module) showed significant overlap with the 14 modules identified by Gandal et al. [22] and with 22 of the 35 modules identified by Fromer et al. [21].

Nine of the 12 modules identified in our analysis (excluding the gray module) were enriched in cell type-specific markers (Table 2).

Table 2 Cell type-specific enrichment of the identified co-expression modules using pre-made sets of brain-derived enrichment lists: Cahoy (Cahoy et al. [23]), CTX (Oldham et al. [24]), and HumanMeta (Miller et al. [25]).
Full size table

Three modules (brown, green, and pink) were enriched in genes associated with TRS. In addition, the brown module was also enriched in genes associated with schizophrenia and educational attainment, while the pink module was enriched in genes associated with schizophrenia (Table 3).

Table 3 Gene enrichment analysis of the co-expression modules with the genes significantly associated with TRS (Pardiñas et al., [14]), schizophrenia (Trubetskoy et al. [26]), and educational attainment (Lee et al., 2019) in large public GWAS.
Full size table

Our analysis identified the green module to be significantly associated with clozapine treatment (Table 4). The genes in the green module were enriched in 76 biological processes from the GO database (Supplementary Table 2), which were grouped in 15 clusters (Fig. 1). These clusters included processes related to cellular detoxification, neuronal differentiation and proliferation processes, organ development, fatty acid metabolism, and the activity of kinases including phosphatidylinositol 3-kinase and MAP kinase.

Fig. 1: Biological processes obtained from the genes in the green module according to ClueGO.
Fig. 1: Biological processes obtained from the genes in the green module according to ClueGO.
Full size image

Each node represents a GO biological process. The node size represents the enriched p-value corrected with the Benjamini-Hochberg method. The edge between the nodes is based on their kappa score level. Colors represent clusters of biological processes.

Table 4 Results of the test generalized linear mixed models comparing clozapine vs non-clozapine patients with the module eigengenes adjusted for sex, age, RIN number, cause of death and PMI as the covariates.
Full size table

Discussion

As a result of our analysis of the gene co-expression architecture in the DLPFC, we identified one module (green) of co-expressed genes that was significantly associated with clozapine treatment. This module was significantly enriched in astrocyte markers and genes involved in the polygenic architecture of TRS.

Our results provide evidence of cell type-specific associations with clozapine treatment that could enable the study of the cellular basis of TRS. Our study pointed to a pivotal role of astrocytes in the mechanism of action of clozapine and potentially in the neurobiology of TRS. Astrocytes are glial cells with highly diverse functional roles in the brain, actively participating in synaptic transmission through neurotransmitter synthesis, buffering, and recycling, as well as in synaptogenesis and synapse elimination [30]. Several lines of evidence, including RNA sequencing data and the findings of genome-wide association studies, demonstrate astrocyte dysregulation in schizophrenia [31]. The functional role of astrocytes in regulating glutamatergic neurotransmission has been proposed to contribute to the glutamate dysfunction associated with schizophrenia [32]. Several findings suggest that glutamatergic dysfunction in the frontal cortex may occur temporally and spatially upstream of dopaminergic dysfunction in the substantia nigra and striatum through the regulation of hippocampal and amygdala activities via thalamic circuits [33]. There is also evidence that abnormalities in glutamate regulation may specifically play a role in TRS. Neuroimaging studies have shown that glutamate levels in several brain regions are higher in patients with TRS compared to healthy controls or patients with schizophrenia who are treatment responsive [33,34,35,36,37].

The biological processes enriched in the green module correspond to the functional roles of astrocytes. Some processes are involved in the critical role of astrocytes in synapse elimination such as fatty acid metabolism, kinase activity (including phosphatidylinositol 3-kinase and MAP kinase), and Ca2+ mobilization [30]. Additionally, the pivotal role of astrocytes as partners of neurons in homeostatic and metabolic processes is reflected in cellular detoxification processes and also in processes related to neuronal proliferation and differentiation [30, 31].

We should bear in mind that the transcriptome may either reflect the cause of TRS or be a consequence of clozapine treatment. In this regard, indirect evidence for the role of glutamate in clozapine mechanism of action or TRS is provided by studies examining the effect of clozapine on the glutamatergic system [38, 39]. However, the enrichment of the genes involved in the GWAS of TRS could indicate that the genes belonging to the green module reflect the cause of TRS rather than the consequence of clozapine treatment.

Interestingly, the green module was enriched in genes significantly associated with TRS, but not in genes associated with schizophrenia. One possible interpretation of this is that patients with TRS may constitute a subset of individuals with distinct genetic characteristics that fundamentally differ from those that contribute to the risk of developing schizophrenia. These findings indicate a partial overlap between the genetic basis of TRS, defined by clozapine treatment as a proxy for TRS, and the genetic susceptibility to schizophrenia and cognitive performance. This agrees with the conclusions drawn by Pardiñas et al. [14] that support the existence of a polygenic contribution to TRS that is largely distinct from susceptibility to schizophrenia (captured by the green module in our study), but partially correlates with schizophrenia and cognitive performance (measured as the educational attainment polygenic risk score (PRS), and captured by the brown and pink modules of our study) [14]. In previous studies, in a first episode schizophrenia cohort, we had shown that educational attainment PRS is a better predictor than schizophrenia PRS of the one-year longitudinal assessment of antipsychotic response [40]. Considering the association between the genetics of cognitive performance and short term response to antipsychotics we hypothesized that this genetic architecture could partially overlap with the genetic basis of clozapine treatment as proxy of TRS, as could be observed in the brown module.

Several limitations should be considered in the interpretation of our results that should be considered the result of an exploratory and preliminary analysis. For this reason, no correction for multiple testing was applied. The main limitations that could affect our results were: firstly, the small sample size, which significantly reduces the statistical power of the study and limits the ability to detect small to moderate effects, potentially increasing the risk of type II errors (false negatives). This constraint is particularly relevant given the inherent variability in gene expression data from brain samples and the complex nature of the biological processes involved in schizophrenia and clozapine response. Secondly, the higher rate of suicide in our sample that could affect the genetics and the gene expression of our sample. Despite these limitations, the functional validation of our results reinforces its validity. Finally, the use of clozapine treatment as a proxy for TRS could lead to some misclassifications in the non-clozapine-treated dataset because individuals who are not treated with clozapine may still have TRS but are not administered the drug due to contraindications or treatment preferences. Such misclassification could potentially dilute true associations or introduce bias into our findings. Future studies would include clinical assessment of treatment response or the use of standardized criteria such as the Treatment Response and Resistance in Psychosis (TRRIP) guidelines. Additionally, obtaining comprehensive clinical data to document contraindications or treatment preferences could help refine the classification process and minimize potential bias. Additionally, is well known that the postmortem tissues used for the study of gene expression could have been affected by autolysis and RNA degradation. The main strength of the study was that AP treatment was defined according to the measured levels of drugs instead of the information from clinical records. Moreover, our data analysis pipeline involved internal, external, and functional validation of the results.

In conclusion, our analysis provides evidence of an association between clozapine treatment and astrocyte function in the DLPFC. This finding provides cell type-specific associations that could help in the interpretation of the neurobiological basis of TRS. A better understanding of the specific DLPFC cell types involved in the mechanism of action of clozapine will contribute to the study of potential pathways and ultimately help improve psychiatric classification tools in personalized medicine. While our findings provide valuable preliminary insights, the limited sample size restricts the generalizability of the results and underscores the need for validation in larger, independent cohorts to confirm these observations. Moreover, replication and validation of these results in other tissues or models could provide complementary insights. For instance, the use of peripheral tissues such as blood could facilitate larger-scale analyses while enabling exploration of systemic molecular mechanisms related to treatment response. Additionally, integrating findings with data derived from induced pluripotent stem cells (iPSCs), animal models, or computational approaches may help address current limitations and improve the biological understanding of clozapine’s effects. Future research efforts should focus on assembling larger and more diverse datasets, as well as leveraging publicly available databases and collaborative networks to strengthen the robustness and reproducibility of these findings.