Strong correlation of downregulated genes related to synaptic transmission and mitochondria in post-mortem autism cerebral cortex

We analyzed gene expression in autism and control cerebral cortex using genes from Parikshak et al., an RNA-seq study [19], and microarray data from three other published studies [16‐18]. The primary microarray dataset was from Voineagu et al. [17] and was downloaded from the Gene Expression Omnibus (GEO, GSE28521) [25]. We limited the Voineagu et al. samples to those with information on RNA integrity number (RIN) and post-mortem interval (PMI). This dataset consisted of prefrontal and/or superior temporal gyrus samples from 15 autism and 16 control subjects (n = 29 control samples, n = 27 autism samples). The Voineagu et al. dataset also contained samples from the cerebellum, which were not used in our study. The two other microarray datasets were from Chow et al. (dorsolateral prefrontal cortex, n = 18 control samples, n = 15 autism samples) [18] and Garbett et al. (superior temporal gyrus, n = 6 control samples, n = 6 autism samples) [16]. We downloaded the Chow et al. dataset from GEO (GSE28475), and the Garbett et al. authors sent us their dataset directly. A Venn diagram depicting overlap of subjects among these datasets is shown in Additional file 1: Figure S1. See Additional file 2: Table S1 for more details about each study used for these analyses.

The Parikshak et al. dataset samples came from the National Institute of Child Health and Human Development-funded University of Maryland Brain and Tissue Bank and the Autism Tissue Program [19]. The Voineagu et al. dataset samples came from the Autism Tissue Program and the Harvard Brain Bank [17]. The Garbett et al. dataset samples also came from the Autism Tissue Program [16], and the Chow et al. dataset samples came from the National Institute of Child Health and Human Development-funded University of Maryland Brain and Tissue Bank and the Autism Tissue Program [18]. In all studies, RNA was extracted from frozen samples. For the Parikshak et al. study, RNA sequencing was performed using an Illumina HiSeq 2000 or 2500 machine, with reads mapped to hg19 using Gencode v18 annotations. For the Voineagu et al. and Chow et al. studies, RNA was hybridized to the Illumina HumanRef8v3 microarray, which contains 24,526 probes. For the Garbett et al. study, RNA was hybridized to Affymetrix Human Genome 133 plus 2 microarrays, which has 54,675 probe sets.

Each microarray dataset was downloaded or received in its normalized form, except that we renormalized the Voineagu et al. dataset to include more probes. Our renormalized version of the Voineagu et al. dataset was the same except that we eliminated probes that did not have significant expression (detection p < 0.05) in at least half of the autism or control samples, rather than half of the samples overall. All three studies for the microarray datasets used log2 transformation and quantile normalization. Voineagu et al. showed that all samples met quality control parameters, specifically, if the interarray Pearson correlation was not greater than 0.85 and if the array was an outlier in hierarchical clustering [17]. Chow et al. similarly eliminated samples based on interarray correlation but also used ComBat [26] to correct for batch effects. The Chow et al. data preprocessing pipeline is described in greater detail separately [27]. After data processing, the Voineagu et al. dataset had 12,632 probes and 10,901 unique Entrez gene identifiers. Chow et al. and Garbett et al. did not eliminate probes, resulting in 18,491 and 20,750 Entrez genes, respectively. All other conversion between gene or probe identifiers were performed using the R package biomaRt [28]. Other than in the Voineagu et al. dataset, we used all available genes in the arrays of these datasets.

Using Ensembl gene identifiers from Parikshak et al., functional annotation clustering of gene sets was performed in DAVID (Database for Annotation, Visualization, and Integrated Discovery) [29] using all available gene pathways, including all Gene Ontology (GO) [30] gene sets, and default parameters of DAVID, including medium classification stringency [29]. The background set of genes in DAVID analysis was the list of all protein-coding genes in the Parikshak et al. RNA-seq dataset, and for the Voineagu et al. study, the background was the list of unique Entrez identifiers in the dataset. All other statistical analyses were performed using R 3.3.2. We performed differential gene expression analysis using the Bioconductor package limma with an empirical Bayes adjustment [31], and we adjusted for RIN, PMI, age, sex, and cortical location (temporal vs. frontal). p values were corrected for multiple testing using the Benjamini-Hochberg method [32]. For DAVID analysis of the differentially expressed genes from the Voineagu et al. data, if multiple Entrez identifiers mapped to the same Illumina probe, which was true for nine of the downregulated probes and six of the upregulated probes, then a single Entrez identifier was chosen at random to avoid over-representing a single genomic feature. Additionally, all duplicate Entrez identifiers, which was true for 15 of the downregulated genes and five of the upregulated genes, were removed prior to DAVID analysis. In validation analyses, we used all available genes from a pre-specified pathway. We then calculated mean expression of these genes and determined a signature’s differential expression using a t test. Heatmaps were generated using the made4 package [33], with Euclidian distance as the distance function.

We set out to discover other biological processes that were not previously reported to be differentially regulated in the Parikshak et al. [19] or Voineagu et al. [17] studies. Parikshak et al. provided a list of genes that are differentially expressed between autism and control cerebral cortex, adjusted for RNA quality, age, sex, brain region, and batch. These genes were up- or downregulated in autism cerebral cortex compared to control (see Additional file 3: Table S2 from Parikshak et al. [19]).

We performed separate DAVID functional annotation clustering analyses for the up- and downregulated genes from Parikshak et al. For the upregulated genes, few gene sets were significantly enriched after Benjamini-Hochberg adjustment [32] (Additional file 3: Table S2). However, each of the top 2 clusters among the downregulated genes included multiple gene sets that were significantly enriched after Benjamini-Hochberg adjustment (Additional file 4: Table S3). For the downregulated genes, the gene set cluster with the highest score was largely related to synapse function and the gene set cluster with the second highest score was related to mitochondrial function. In both the Voineagu et al. [17] and Parikshak et al. [19] studies, the authors described differential expression of genes related to synaptic function. Given that a mitochondrial pathway had not previously been reported by Voineagu et al. or Parikshak et al., we decided to focus on this next. We defined the “synapse pathway” as the downregulated genes that overlapped with the UniProt keyword Synapse and the “mitochondria pathway” as those that overlapped with the GO term “Mitochondrion” (Additional file 5: Table S4). To ensure that the mitochondria pathway did not describe synaptic function, we excluded from the mitochondria pathway any genes that were in the synapse pathway or in a related gene set, module M12 from Voineagu et al. [17]

To exclude the possibility that the mitochondria pathway’s downregulation was unique to the Parikshak et al. study, we next validated this downregulation in other genomic datasets. We used three microarray studies for validation (Additional file 1: Figure S1 and Additional file 2: Table S1). Because the Voineagu et al. study’s subjects were nearly all in the Parikshak et al., it served largely as technical validation (see Fig. 1 for a heatmap of the genes in the Voineagu et al. dataset and Additional file 6: Figure S2 for similar heatmaps regarding the Chow et al. and Garbett et al. datasets). The mean expression of these mitochondria pathway genes was downregulated in the Voineagu et al. dataset (t test p = 0.001), Chow et al. dataset (p = 0.039), and Garbett et al. dataset (p = 0.076) (Fig. 2). To ensure that the downregulation of the mitochondria pathway in the Voineagu et al. dataset was not due to confounders, we also did a separate logistic regression adjusting for RIN, PMI, age, sex, and cortical location (temporal vs. frontal) and found that the mitochondria pathway was still downregulated in autism (p = 0.0054). It was also downregulated in a similar multivariate analysis after limiting the dataset only to frontal cortex (p = 0.041) or temporal cortex (p = 0.057). While the mitochondria pathway was associated with autism, it was not associated with seizures, speech delay, motor delay, or global functioning in the Voineagu et al. dataset (p > 0.3 for each comparison).

Because these studies’ subjects overlapped, we did a separate validation analysis of the Chow et al. and Garbett et al. datasets after removing all but the subjects unique to these studies. The mitochondria pathway was still downregulated in these analyses, although the p values were not significant (p = 0.12 for the Chow et al. dataset and p = 0.33 for the Garbett et al. dataset), likely because of reduced sample sizes (n = 20 for the Chow et al. dataset and n = 7 for the Garbett et al. dataset).

In our reanalysis of the Parikshak et al. genes, the synapse-related gene sets were the most strongly enriched in those downregulated in autism, so we next determined the relationship between those synapse-related gene sets and the mitochondria pathway. Across all three microarray datasets, these two pathways had strong Pearson correlation (Fig. 3). To exclude the possibility that such correlation was common, we also used the Voineagu et al. dataset to randomly sample without replacement 10,000 gene sets of similar size to the synapse pathway, and we found that the mitochondria pathway had greater correlation with the synapse pathway than all but 0.37% of random gene sets. For a specific example of correlated genes, in the Voineagu et al. dataset, GABRA1, which codes for a gamma-aminobutyric acid (GABA) receptor subunit, and ATP5A1, which codes for an ATP synthase subunit, were strongly correlated (correlation = 0.876).

Given that the mitochondria pathway was among the most enriched in the Parikshak et al. downregulated genes, we did a separate analysis to confirm the importance of this pathway in autism cerebral cortex. Using the Voineagu et al. dataset, we performed a differential gene expression analysis using limma [31] between autism and control cerebral cortex, adjusting for RIN, PMI, age, sex, and cortical location (temporal vs. frontal). This produced 185 upregulated and 247 downregulated unique genes (Additional file 7: Table S5). Given that several individuals were represented twice in this analysis (both for frontal and temporal cortexes), we also did separate analyses using only temporal or frontal cortex samples (Additional file 7: Table S5). We did not adjust for multiple testing in these cortex-specific analyses because no genes were differentially expressed after Benjamini-Hochberg adjustment in these limited samples. At least 84% of the original up- and downregulated genes were in the respective cortex-specific up- or downregulated genes, suggesting broad similarity in these three differential expression analyses.

We next performed DAVID functional annotation clustering of the original up- and downregulated genes. The upregulated genes were enriched in only one gene set (Additional file 8: Table S6), but the downregulated genes were enriched in several gene sets, and all of the top 5 highest scoring gene set clusters were related to mitochondria (Additional file 9: Table S7).

The Voineagu et al. study limited their differential gene expression analysis to genes that had a fold change > 1.3. The mitochondria pathway may have previously gone unreported in that study because in the Voineagu et al. dataset, the mitochondria pathway genes had on average 1.13-fold change in gene expression while the synapse pathway genes had 1.20-fold change. Similarly, in the Parikshak et al. study, the synapse pathway genes showed on average 1.25-fold change while the mitochondria pathway genes showed 1.16-fold change. Thus, the enrichment may not have been detected because of the greater fold change for the synapse genes.

Finally, we also observed that the GABA-related genes in particular were differentially expressed. The synapse pathway included genes coding for two different GABA receptor subunits, and the gene coding for parvalbumin, which is a marker of inhibitory interneurons [34], was the most strongly downregulated gene. The gene coding for parvalbumin was also the most strongly downregulated gene in the Parikshak et al. study [19].