Introduction

Certain core cognitive impairments, such as working memory (WM) deficits [1], have a major influence on the long-term functional outcomes of individuals with schizophrenia (SZ) [2]. Visuospatial WM (vsWM) is predominantly mediated by coordinated neural activity across a distributed cortical neural network that includes nodes in primary (V1) and association (V2) visual regions of the occipital lobes which convey visual information to nodes in the posterior parietal cortex (PPC) and dorsolateral prefrontal cortex (DLPFC) [3, 4]. The activity of excitatory pyramidal neurons that provide the connections among these cortical nodes is shaped by the activity of local inhibitory neurons that are, in turn, influenced by excitatory inputs from neighboring pyramidal neurons [5,6,7].

In individuals with SZ, multiple postmortem studies have revealed structural and molecular alterations in cellular components that regulate the excitatory synaptic transmission required for WM performance [8, 9]. For example, a lower density of dendritic spines, the sites of most excitatory synaptic inputs to pyramidal neurons [10,11,12], and altered expression of transcripts for the regulators of actin cytoskeleton, which play a critical role in the formation and maintenance of dendritic spines [13,14,15,16], have been reported in the DLPFC of subjects with SZ.

In addition, ATP synthesis via mitochondrial oxidative phosphorylation (OXPHOS), which supports energetic demand of the neuronal activity and synaptic transmission required for WM performance [17,18,19], appears to be impaired in the DLPFC of subjects with SZ [20,21,22,23,24]. Consistent with these findings, lower levels of OXPHOS-related transcripts have been reported in DLPFC pyramidal neurons from subjects with SZ [25, 26].

Expression of actin- and OXPHOS-related transcripts would be expected to be correlated as dendritic spine density influences pyramidal neuron activity and hence the demand for energy production [17, 27], and at the same time, energy production is crucial for the maintenance of dendritic spines [28, 29]. Furthermore, since both the number of dendritic spines on layer 3 pyramidal neurons [30,31,32,33] and mitochondrial energy production [34, 35] have been reported to differ across cortical regions, levels of actin- and OXPHOS-related transcripts are expected to differ across regions of the vsWM network in UC subjects. Finally, in SZ subjects, lower spine density is more prominent in the DLPFC compared to V1 [10], suggesting that actin-related transcripts should exhibit regionally different alterations within the vsWM network in SZ subjects. Therefore, comparing the cross-regional alterations between actin- and OXPHOS-related transcripts in SZ subjects might provide an insight into relationship between lower dendritic spine density and less energy production across the regions of the vsWM network in SZ.

In the present study, using the RNA sequencing (RNAseq) data generated by the CommonMind Consortium in DLPFC total gray matter samples from 82 UC and 57 SZ subjects [36], we first examined transcript levels and cross correlations of a subset of key actin- and OXPHOS-related molecules. Next, we employed quantitative PCR (qPCR) to determine the levels of these same transcripts in total gray matter samples from the DLPFC and three other cortical regions (primary (V1) and secondary (V2) visual cortices, and PPC) of the vsWM network from 20 matched pairs of UC and SZ subjects. Then, we compared the effect of SZ on the expression of actin- and OXPHOS-related transcripts across these regions.

Methods

Human subjects

Brain specimens were obtained during routine autopsies conducted at the Allegheny County Medical Examiner’s Office (Pittsburgh, PA) following consent from the next-of-kin. An independent committee of experienced research clinicians made consensus DSM-IV diagnoses for each subject or confirmed the absence of any psychiatric or neurological diagnoses, based on the results of structured interviews conducted with family members and review of medical records, neuropathology exam and toxicology testing. All procedures were approved by the University of Pittsburgh’s Committee for the Oversight of Research and Clinical Training Involving Decedents and Institutional Review Board for Biomedical Research, as well as by the Ethics committees of Nara Medical University and Kanazawa University Graduate School of Medical Sciences.

Selection of transcripts used in the current study

We selected 4 actin- (BAIAP2, CDC42, ARPC3, and ARPC4) and 4 OXPHOS-related (NDUFB3, COX4l1, COX7B and ATP5H) transcripts. Expression levels of these 8 transcripts were all previously reported to be altered in the DLPFC from SZ subjects in studies using different quantification methods at different anatomical resolutions [13,14,15,16, 25, 26, 37, 38]. Among the actin-related transcripts, the Rho GTPase CDC42, a central regulator of actin assembly that is essential for neuronal morphology, promotes spine growth and stability [39], whereas BAIAP2 inhibits actin assembly and thus suppresses spine growth [40, 41]. ARPC3 and APRC4 link CDC42-dependent signals to actin assembly and are required for dendritic spine formation and maintenance [42]. For OXPHOS-related transcripts, each transcript encodes a component of key complexes for mitochondrial electron transfer chain [43, 44]: NDUFB3 for NADH dehydrogenase (complex I), COX4l1 and COX7B for cytochrome oxidase (complex IV), and ATP5H for ATP synthase (complex V). The expression levels of these OXPHOS-related transcripts are normally coordinated in response to changes in the energetic demand in neurons [27, 45].

These actin- and OXPHOS-related transcripts have multiple isoforms derived from alternative splicing. Our qPCR primer sets for ARPC3, ARPC4, NDUFB3, COX7B, and ATP5H amplified a fragment that is common to all reported isoforms. For BAIAP2, CDC42 and COX4I1, our qPCR primer sets amplified a fragment that is selective to some isoforms. These fragments are adjacent or within the sequences recognized by the microarray probes that detected altered expression levels of these transcripts in our previous study [25].

RNA sequencing (RNAseq) subject cohort and analysis

RNAseq data used in these analyses included a subset of the data generated by the CommonMind Consortium (https://www.nimhgenetics.org/available_data/commonmind/). Principal component analysis of the DLPFC samples revealed institution as a major source of variance. Therefore, to eliminate this unwanted source of variance, only University of Pittsburgh subjects (n = 139) were used in the transcriptome analysis of the current study (Table 1, Supplementary Table 1a). Subject groups did not differ in mean age, pH, RNA integrity number (RIN) (Agilent Bioanalyzer, Agilent Technologies, CA), postmortem interval (PMI), or tissue storage time at −80 °C (Table 1). Total RNA was isolated from gray matter tissues of right DLPFC (area 9/46) of each subject and subjected to RNAseq as described previously [36] and in Supplementary Methods.

Table 1 Summary of cohort characteristics and methods for mRNA quantification.
Full size table

Quantitative PCR (qPCR) subject cohort and analysis

To control for experimental variance and to reduce biological variance between the groups of UC (n = 20) and SZ (n = 20) subjects, each SZ subject was individually matched to one UC subject for sex and as closely as possible for age (Table 1, Supplementary Table 1b). The two groups did not differ in mean age, pH, RIN, PMI, or tissue storage time (Table 1). Total RNA was isolated from gray matter samples of all four cortical regions (V1, V2, PPC, and DLPFC) of the right hemisphere of each subject (Supplementary Fig. 1) and used for qPCR amplification of transcripts of interest and two internal reference transcripts (β-actin and glyceraldehyde-3-phosphate dehydrogenase) with specific primer sets (Supplementary Table 3). Samples from members of a subject pair were always processed together throughout all stages of an experiment. Transcript levels were determined based on the difference between the cycle threshold (CT) of the target transcript and the mean CT of internal reference transcripts. See Supplementary Methods for details.

Statistical analyses

For RNAseq data analysis, genes with 0 counts in at least 50% of the samples were removed and the remaining 14,428 genes were subjected to normalization and confidence weight calculation using the LIMMA package [46] in R. The normalized read counts and confidence weights were used to determine differential gene expression using a basic linear regression model with age, sex, RIN, PMI, and five ancestry-related factors included as covariates. Statistical significance was determined using the Benjamini-Hochberg procedure on the resulting p values with a false discovery rate of 5%. Levels of actin- and OXPHOS-related molecules identified in the differential-expression analysis were subjected to Pearson correlation analysis.

For the cross-regional analysis by qPCR, we used mixed models treating observations from the four regions of each subject as repeated measurements to account for the within-subject correlation among four regions and to characterize the within- and between-subject variability (see Supplementary Methods for details). To assess differences in the target transcript levels across the four regions in UC subjects, the model included transcript as the dependent variable, region as a fixed effect, and age, sex, brain pH, RIN, PMI and RNA ratio as covariates. F-tests were used to assess the overall regional effect, followed by post hoc pairwise comparisons between regions. To determine if the expression of these transcripts was altered in SZ, the mixed model included transcript level as the dependent variable, diagnosis, region, and diagnosis-by-region interaction as the fixed effects, and the same set of covariates. If the diagnosis-by-region interaction was significant, we also assessed the diagnosis effect in each region. Benjamini-Hochberg adjusted p values were computed to control for multiple comparison across eight transcripts [47], and reported p values for each transcript have been adjusted to correspond to the false discovery rate of 5%. In each of V1, V2, PPC and DLPFC, cross correlations among actin- and OXPHOS-related transcripts were assessed by Pearson correlation analyses. All analyses of individual transcripts were conducted on log-transformed data. The potential confounding effect of antipsychotic medications, nicotine, or other substances of abuse, suicide, and other factors frequently comorbid with SZ was also examined by using Student t test (Supplementary Table 4).

Composite measures for actin assembly and OXPHOS were computed by the sum of the normalized (Z-score) expression levels for each of the actin- and OXPHOS-related transcripts, respectively (see Supplementary Methods). Since BAIAP2 has an inhibitory regulation, as opposed to the promoting effects of CDC42, ARPC3 and ARPC4, on actin assembly, we added minus signs to its Z-scores before summing Z-scores for all transcripts. For each composite measure, the mixed model was performed as described above. In order to address the risk of false positive findings, we conducted a Bayesian repeated measures analysis of variance (ANOVA) on the composite measures using JASP (https://jasp-stats.org) [48]. This analysis included the composite measures from the four regions of each subject as repeated measures, diagnosis as the between-subject factor and the same set of covariates as in the mixed model.

Results

Transcript levels for actin- and OXPHOS-related pathways in DLPFC of UC and SZ subjects by RNAseq

Our analysis of RNAseq data found altered expression levels of the 4 actin- and 4 OXPHOS-related transcripts (all q < 0.05) (Fig. 1A). Among actin-related transcripts, BAIAP2 mRNA levels were higher by 26%, whereas levels of other three transcripts were lower by –12% to -17% in SZ subjects. All OXPHOS transcripts exhibited lower levels by -15 to -22% in SZ subjects. These results are consistent with prior reports [13,14,15,16, 25, 26, 37, 38], including our use of a different analytical strategy to assess OXPHOS-related transcripts in the same UC and SZ subjects [26]. Importantly, levels of all actin- and OXPHOS-related transcripts were significantly correlated (all r > |0.21 | ) with each other in the UC, SZ and combined groups of subjects (Fig. 1B), consistent with the idea that dendritic spine density and energy production are related in the DLPFC. All transcripts were positively correlated with each other except for BAIAP2 which, as expected based on its inhibitory role in actin assembly, was negatively correlated with the other actin- and all OXPHOS-related transcripts.

Fig. 1: Differential expression and cross-correlation analysis of actin- and OXPHOS-related transcripts assessed by RNAseq in DLPFC gray matter of unaffected comparison and schizophrenia subjects.
figure 1

A RNAseq data of mean differences between SZ and UC subjects in actin- and OXPHOS-related transcripts in the DLPFC. B Matrix of cross-correlations among the transcripts in total (left), unaffected comparison (middle) and schizophrenia (right) subject groups. Correlation coefficients for each combination of two transcripts are shown in the lower left part of the matrix according to the color bar at the bottom, and corresponding p values are shown in the upper right part.

Full size image

Expression of actin- and OXPHOS-related transcripts in the vsWM network in UC subjects

Among actin-related transcripts, BAIAP2 mRNA levels were highest in V1 and V2 and progressively declined from posterior to anterior regions, with intermediate levels in PPC and lowest levels in the DLPFC (F3,57 = 49.6, p < 0.001) (Fig. 2A). In contrast, CDC42 mRNA levels (F3,57 = 17.0, p < 0.001) were highest in DLPFC, intermediate in V1 and lowest in V2 and PPC (Fig. 2B). Levels of both ARPC3 (F3,57 = 3.33, p = 0.026) and ARPC4 (F3,57 = 11.2, p < 0.001) mRNAs were highest in DLPFC and did not differ across the other three regions (Fig. 2C, D).

Fig. 2: Expression levels of actin- and OXPHOS-related transcripts across the four cortical regions of the visuospatial WM network in unaffected comparison subjects.
figure 2

For each (A–H), transcript name and the mixed model results are shown at the top left. Transcript levels of individual subjects are shown by the same symbol in all graphs. Horizontal bars represent group means. For each transcript, regions within each graph that do not share the same letter are significantly different by post hoc comparisons (p < 0.05). V1 primary visual cortex, V2 association visual cortex, PPC posterior parietal cortex, DLPFC dorsolateral prefrontal cortex.

Full size image

All four OXPHOS-related transcripts showed declined from posterior to anterior regions. Levels of NDUFB3 (F3,57 = 15.3, p < 0.001), COX4I1 (F3,57 = 34.3, p < 0.001) and ATP5H (F3,57 = 24.3, p < 0.001) mRNAs were highest in V1, intermediate in V2, and lowest in PPC and DLPFC (Fig. 2E, F, H). Similarly, COX7B mRNA levels (F3,57 = 24.9, p < 0.001) were highest in V1, intermediate in V2 and PPC, and lowest in DLPFC (Fig. 2G).

Effect of SZ on actin- and OXPHOS-related transcript levels in the vsWM network

For BAIAP2 mRNA, our mixed model detected significant effects of region (F3,114 = 83.2, p < 0.001) and a diagnosis-by-region interaction (F3,114 = 4.93, p = 0.006) without a significant diagnosis effect (F1,32 = 1.46, p = 0.27) (Fig. 3A). Post hoc analysis revealed that BAIAP2 mRNA levels were unaltered in V2, PPC and DLPFC, but were significantly upregulated in V1 (p = 0.007) of SZ subjects. Levels of CDC42 and ARPC4 mRNAs significantly differed by diagnosis (F1,32 = 10.27, p = 0.005 and F1,32 = 5.78, p = 0.029, respectively) and by region (F3,114 = 16.1, p < 0.001 and F3,114 = 23.1, p < 0.001, respectively), but the interaction term was not significant (Fig. 3B, D). CDC42 and ARPC4 mRNA levels were lower in all four cortical regions of the vsWM network in SZ subjects. Finally, ARPC3 mRNA expression showed neither an effect of diagnosis nor a diagnosis-by-region interaction, but the effect of region on mRNA level was significant (F3,114 = 5.41, p = 0.002) (Fig. 3C).

Fig. 3: Effect of schizophrenia on actin- and OXPHOS-related transcripts across the four cortical regions of the visuospatial WM network.
figure 3

For each (A–H), transcript name is at the top left and the mixed model result at the top center. Across regions, % difference of transcript levels between unaffected comparison (UC) and schizophrenia (SZ) subjects in each subject pair are shown by the symbols from Fig. 2 that corresponds to the unaffected comparison subject in that pair. Horizontal bars represent mean percent difference across subject pairs. Values in parentheses represent Cohen’s D effect size in each region for each transcript between SZ and UC subjects. An asterisk (*) indicates significant difference between the UC and schizophrenia subjects (p < 0.05) in each region on post hoc comparisons that were performed when the mixed model detected a significant diagnosis-by-region interaction. V1 primary visual cortex, V2 association visual cortex, PPC posterior parietal cortex, DLPFC dorsolateral prefrontal cortex.

Full size image

All OXPHOS-related transcripts (Fig. 3E–H) exhibited significant effects of diagnosis (all F1,32 > 10.2, all p < 0.006) and region (all F3,114 > 3.2, all p < 0.03), as well as a diagnosis-by-region interaction (all F3,114 > 4.7, all p < 0.006). Post hoc analyses showed that NDUFB3 and ATP5H mRNA levels (Fig. 3E, H) were significantly lower in V1 and V2 (both p < 0.001), but were unaltered in PPC and DLPFC (all p > 0.07), of SZ subjects. COX4I1 mRNA levels (Fig. 3F) were significantly lower in V1, V2 and DLPFC (all p < 0.05), respectively, but were unaltered in PPC (p = 0.08), and COX7B mRNA levels (Fig. 3G) were significantly lower in V1, V2 and PPC (all p < 0.03), but were unaltered in DLPFC of SZ subjects.

We tested cross correlations between actin- and OXPHOS-related transcripts in V1, V2, PPC and DLPFC using qPCR data. In contrast to the robust correlations detected in DLPFC RNAseq data of 82 comparison and 57 schizophrenia subjects, correlations among transcript levels were much less prominent in each region, presumably due to the limited sample size (20 pairs of UC and SZ subjects). However, we detected that cross-correlations between actin- and OXPHOS-related transcripts were weaker in A17 and A18 than in A46 and A7 (Supplementary Fig. 2).

Composite measures of actin- and OXPHOS-related transcripts across the vsWM network

To obtain aggregate indices of actin assembly and mitochondrial energy production across regions of the visuospatial WM network, we computed composite actin and OXPHOS measures from the normalized expression levels of the four transcripts in each category (Fig. 4). In UC subjects, the composite actin (F3,114 = 25.8, p < 0.001) and OXPHOS (F3,114 = 20.1, p < 0.001) measures significantly differed across regions (Fig. 4A, B). The composite actin measure was greatest in DLPFC and did not differ across the other three regions, whereas the composite OXPHOS measure was greatest in V1 and progressively decreased from posterior to anterior regions of the vsWM network (Fig. 4A, B).

Fig. 4: Effect of schizophrenia on composite measures of actin- and OXPHOS-related transcripts in the visuospatial WM network.
figure 4

In unaffected comparison (UC) subjects, the composite (A) actin (F3,114 = 25.8, p < 0.001) and (B) OXPHOS (F3,114 = 20.1, p < 0.001) significantly differed across regions. In comparisons between UC and SZ subjects, the actin composite measure significantly differed by diagnosis (F1,32 = 5.95, p = 0.02) and region (F3,114 = 40.9, p < 0.001), but the interaction term was not significant (F3,114 = 0.8, p = 0.50). For the OXPHOS composite measure, the caudal-to-rostral gradient was absent in the SZ subjects, as indicated by significant effects of diagnosis (F1,32 = 17.4, p < 0.001), region (F1,32 = 12.1, p < 0.001) and diagnosis-by-region interaction (F3,114 = 8.25, p < 0.001). Error bars indicate the standard error of the mean (SEM). V1 primary visual cortex, V2 association visual cortex, PPC posterior parietal cortex, DLPFC dorsolateral prefrontal cortex.

Full size image

Relative to UC subjects, the composite actin measure (Fig. 4A) was lower in all regions of SZ subjects, exhibiting significant differences by diagnosis (F1,32 = 5.95, p = 0.02) and region (F3,114 = 40.9, p < 0.001), without a significant interaction between diagnosis and region (F3,114 = 0.8, p = 0.50). For the composite OXPHOS measure (Fig. 4B), the posterior-to-anterior gradient was absent in the SZ subjects, as indicated by significant effects of diagnosis (F1,32 = 17.4, p < 0.001), region (F3,114 = 12.1, p < 0.001) and diagnosis-by-region interaction (F3,114 = 8.25, p < 0.001).

Bayesian repeated measures ANOVA [48] on the composite measures supported strongly the likelihood of the models including the factors with the significant effects in the mixed models, rejecting the alternative models without these factors. For the composite actin measures, the Bayesian Factors computed for diagnosis and regions were 4.6 and 5.4 × 1015, respectively, providing substantial and decisive evidence of the effects of these factors [49]. For the composite OXPHOS measures, the Bayesian Factors for diagnosis, region, and diagnosis-by-region interaction were 702, 4124 and 443, respectively, decisively supporting the effects of these factors.

Discussion

Normal vsWM processes require information transfer across multiple nodes in the neocortex, including visual, posterior parietal, and dorsolateral prefrontal regions. Previous studies reported altered levels of transcripts involved in the regulation of actin assembly and ATP synthesis via OHPHOS in the DLPFC of subjects with SZ [13,14,15,16, 20, 25, 37], but the relationships among these transcripts and whether these transcript levels differ across regions in the vsWM network of UC and SZ subjects were unknown.

Actin- and OXPHOS-related transcripts in the DLPFC of UC and SZ subjects

A subset of representative molecules in the actin signaling (BAIAP2, CDC42, ARPC3 and ARPC4) and OXPHOS pathways (NDUFB3, COX4l1, COX7B and ATP5H) were reported to be differently expressed in the DLPFC from subjects with SZ using different levels of anatomical resolution and different quantification methods [13,14,15,16, 25, 37]. Thus, we investigated the expression levels of these 8 transcripts in the RNAseq data from DLPFC total gray matter in a larger cohort of 57 subjects with schizophrenia and 82 unaffected comparison subjects. In our RNAseq data, we verified altered expression levels of all actin-related transcripts in the DLPFC of SZ subjects (Fig. 1), providing both biological and technical replication of the prior findings. Given that BAIAP2 is a negative regulator of actin assembly, the reciprocal alterations between BAIAP2 and the other actin-related transcripts (CDC42, ARPC3 and ARPC4) in the DLPFC of SZ subjects might represent a molecular mechanism for deficient actin assembly that could contribute to the morphological alterations in DLPFC pyramidal cells, such as smaller cell bodies, shorter dendritic length and lower density of dendritic spines reported in SZ [10,11,12, 50]. We recently reported that many transcripts in the OXPHOS pathway were down-regulated in a coordinated manner in the DLPFC of SZ subjects [26], and the present reanalysis of these data using a different statistical model confirmed a conserved effect of SZ on the four OXPHOS-related transcripts examined here. Furthermore, the correlated expression among actin- and OXPHOS-related transcripts in both subject groups is consistent with the idea that disturbed actin assembly in dendritic spines might result in reduced pyramidal cell activity and a lower requirement for mitochondrial energy production [25, 26, 29, 51].

As genome wide-association studies have implicated genes involved in the regulation of dendritic spines and mitochondrial functions in the pathogenesis of SZ [52, 53], we evaluated the association between polygenic risk scores (PRS) that were determined based on the data of PGC and CLOZUK cohort [54] and our DLPFC RNAseq data from 139 subjects. We did not find any significant correlations between PRS determined with three different statistical thresholds and DLPFC levels of any of the transcripts of interest in either control (all r < |0.19 | ) or schizophrenia (all r < |0.22 | ) subject groups.

Expression of transcripts across regions of the cortical WM network in UC subjects

Both actin- and OXPHOS-related transcripts exhibited distinctive cross-regional expression patterns in vsWM network of UC subjects. Transcript levels of CDC42, ARPC3 and ARPC4, positive regulators of actin assembly [42], were highest in DLPFC and lowest in V1, whereas levels of BAIAP2, which negatively regulates actin assembly and spine formation [40, 41], were highest in occipital cortex (V1 and V2) and declined from posterior to anterior regions. Together, these findings suggest that across the vsWM cortical network the molecule substrate for actin regulation is greatest in DLPFC and lowest in V1. Interestingly, in nonhuman primates the dendritic field areas of pyramidal neurons were 1.3 times, 2.5–3.5 times, and >5 times larger in V2, PPC, and DLPFC, respectively, than in V1 [32, 33]. The total number of dendritic branches per pyramidal neuron, as well as the number and density of spines, showed a similar posterior-to-anterior increase [30,31,32,33]. Therefore, relative to the other cortical regions, the lower expression of BAIAP2 and the higher expression of CDC42, ARPC3 and ARPC4 (Fig. 2A-D) in the DLPFC could contribute to the greater spine number and dendritic arborization of DLPFC pyramidal neurons relative to those in more posterior regions of the vsWM network.

All four OXPHOS-related mRNAs exhibited decreasing posterior-to-anterior expression gradients across the cortical WM network (Fig. 2E–H). This observation is consistent with a resting state positron emission tomography study [55] showing that the rate of ATP production, calculated from cerebral metabolic rates of oxygen and glucose, was high in V1, intermediate in V2, and low in PPC and DLPFC in healthy human subjects. In general, neurons adjust the expression of OXPHOS-related molecules to meet the changes in ATP demand, supporting their firing and synaptic neurotransmission [17, 18]. For example, in monkey visual cortex, a higher activity of cytochrome oxidase, an enzyme of OXPHOS, was detected in areas with relatively abundant glutamatergic inputs and high neuronal activity [56]. Therefore, the decreasing posterior-to-anterior gradients of OXPHOS-related transcripts could reflect a similar gradient of neuronal activity across the regions. Interestingly, the density of parvalbumin (PV)-positive subsets of GABA neurons is higher in occipital visual regions relative to DLPFC in monkeys [57, 58], which is consistent with a decreasing posterior-to-anterior gradient of PV mRNA levels across cortical regions of the vsWM network in humans [59]. PV neurons are considered to be very energy-demanding due to their high-frequency firing [60] and contain greater numbers of cytochrome oxidase-enriched mitochondria compared with other types of neurons [61]. Therefore, the gradient of OXPHOS-related transcript levels across cortical regions of the WM network could reflect, at least in part, regional differences in PV neuron density. Finally, higher OXPHOS-related transcripts in V1 might reflect the higher ratio of GABA to pyramidal neurons in V1 (1:3) than in other regions (1:4) [62], as GABA neurons have higher cytochrome oxidase activity than pyramidal neurons in monkeys [63, 64].

Alterations of transcripts across cortical regions of the WM network in SZ

The RNAseq and qPCR data showed the same pattern of differences between SZ and UC subjects in the DLPFC. The magnitudes of the disease effect were smaller for all transcripts in the qPCR data (Figs. 1 and 3), but were consistent with the size of the group differences found in prior qPCR studies of these transcripts [14, 16]. The different effect sizes in the RNAseq and qPCR data might reflect differences between studies in assay sensitivity and/or the composition of the subject groups. However, all transcript levels between RNAseq and qPCR measures were positively correlated in the subjects (see Supplementary Table 1) included in both studies (all r > 0.30, all p < 0.04 with one-tailed t test, n = 37).

In SZ, the deficit in the composite actin measure, which is intended to index the activity of actin assembly that promotes the formation and maintenance of spines, was much greater in DLPFC than the other regions (Fig. 4A). This regional difference could be associated with the previously reported greater deficit in spine density of layer 3 pyramidal neurons in DLPFC relative to V1 [10]. This interpretation would appear to not be supported by the significant up-regulation of BAIAP2 in V1 versus DLPFC of SZ subjects. However, in addition to its inhibitory role in spine formation [40, 41], BAIAP2 regulates glutamate transmission as a component of the postsynaptic density [65,66,67]. In a previous study of the vsWM network of the same subject cohort, transcripts involved in glutamate neurotransmission exhibited the greatest up-regulation in V1 of SZ subjects [47], suggesting that the higher BAIAP2 transcript levels might contribute to altered glutamatergic signaling in V1 of SZ subjects.

In contrast to actin-related transcripts, OXPHOS transcripts were very similar across regions in SZ subjects, suggesting that the magnitude of the disease effect progressively decreased from posterior to anterior regions (Fig. 3E–H). Interestingly, inhibition of OXPHOS enzymes enhanced depolarization-induced glutamate release from isolated synaptosomes [68, 69]. Thus, the prominent down-regulation of OXPHOS transcripts in posterior cortical regions might be associated with the greater alterations in markers of glutamate neurotransmission in these regions in SZ.

The effect of SZ differed between the actin and OXPHOS composite measures across regions of vsWM network, with the disease effect on actin-related transcripts largest in DLPFC and on OXPHOS transcripts largest in V1 (Fig. 4). These opposed regional patterns of disease effect do not seem to support a global cortical disease process of altered actin regulation leading to lower dendritic spine density, reduced pyramidal cell activity and less drive for mitochondrial ATP energy production [26], although the present data are consistent with that idea in the DLPFC (Fig. 4). Furthermore, a comparison of our RNAseq data in the DLPFC to a comparable data set from the anterior cingulate cortex (area 24) of the same subjects [70] revealed that the magnitude of the disease effect on both action- and OXPHOS-related transcripts differs between these regions (Supplementary Table 2). Thus, the dissociation of actin- and OXPHOS-related composite measures in V1 of SZ subjects suggests that other neuron types, such as PV neurons that exhibited a similar cross-regional pattern in PV mRNA deficits between V1 and DLPFC [59], might contribute to the transcript levels detected in total gray matter in the present study. As PV mRNA levels are regulated by neuronal activity [71], the lower levels of both PV and OXPHOS-related mRNAs in V1 are consistent with lower activity of PV neurons in that region. Although the major excitatory drive to PV neurons is provided by neighboring pyramidal neurons [72], the deficits in actin-related transcripts, which could be associated with lower spine density, were relatively small in V1 compared with DLPFC. Therefore, the prominent deficits in OXPHOS-related transcripts are not likely to be explained solely by lower excitatory drive from pyramidal neurons to PV neurons. Greater deficits in OXPHOS-related transcripts in V1 than DLPFC might not simply reflect a uniform elevation or reduction of the excitatory/inhibitory balance across different cell types and cortical regions in SZ. Indeed, these findings point to the need for cell type-specific assessments of actin- and OXPHOS-related transcripts across regions of the vsWM network in both UC and SZ subjects.

Several lines of evidence suggest that altered transcript levels across cortical regions in SZ are unlikely to be due to other factors commonly associated with the illness. First, in our previous studies, none of the comorbid factors assessed, including substance abuse/dependence at the time of death; use of antidepressants, benzodiazepine/anticonvulsant or antipsychotics at the time of death; or death by suicide had significant effects on actin- or OXPHOS-related transcript levels in the DLPFC of SZ subjects [13, 15, 16, 25, 26, 37, 73]. Second, none of these transcripts exhibited altered expression in the DLPFC of monkeys chronically exposed to antipsychotics at clinically relevant doses [13, 16, 25, 26, 37, 73]. Third, using the current data, none of comorbid factors, such as substance abuse/dependence at the time of death; use of antidepressants, benzodiazepine/anticonvulsants or antipsychotics at the time of death; or death by suicide, had a significant effect on the RNAseq data in DLPFC or on the qPCR data in any cortical region (all p > 0.1) (Supplementary Table 4). Fourth, in SZ subjects, none of transcripts showed a significant correlation between the expression levels and the duration of illness in any of the four regions of the WM network (all | r | < 0.33, all p > 0.16), suggesting that the findings are not a consequence of illness chronicity. Fifth, all individuals with SZ had been living in their communities outside of psychiatric facilities at the time of death. Finally, we did not detect a significant effect of socio-economic status measured by Hollingshead categorial rankings [74] (Supplementary Tables 1a, b and 4), which might reflect the level of environmental enrichment for each subject, on actin- nor OXPHOS-related transcripts as measured in the DLPFC by RNAseq (all p > 0.9) or across the four regions by qPCR (all p > 0.5), suggesting that the findings in the individuals with SZ are not due to an impoverished environment.

In conclusion, our analyses of transcript levels across cortical regions in the vsWM network revealed that the transcripts associated with dendritic spine number or energy production had distinct regional expression patterns in UC subjects, suggesting that the strength of local recurrent excitation and energetic demand required for information processing normally differs across cortical regions of the WM network. In SZ, the differential alterations in markers of actin assembly and energy production across regions of the cortical vsWM network suggest that each region may make specific contributions to vsWM impairments in the illness.