Characteristics of Subjects and Brain Samples
We analyzed postmortem brain tissue from a carefully selected sample of 9 males with idiopathic autism and 9 age- and gender-matched typically developing controls (Table S1
, Table S2
). Eight pairs of cerebella from autistic and control subjects produced high quality gene expression microarray data; 6 controls and 4 cases from BA19 also yielded high quality microarray data (Table S2
). There were no significant differences between cases and controls in age, postmortem interval or cause of death (Table S1
Overview of Gene Expression Differences Between Autistic and Control Brains
On correspondence analysis, samples segregated on the first component by brain region, as expected (). The second principal component separated several autistic cases and one control from the others, with heterogeneity of autistic cases most apparent in cerebella. After controlling for brain region, there were 876 unique, annotated genes differentially expressed between autistic and control brains at a false discovery rate of 5%, 32 of which had log2
-fold changes ≥0.7 (Table S3
Autistic brain shows transcriptional heterogeneity and differential expression of genes of mitochondrial oxidative phosphorylation and protein translation.
Unsupervised clustering of samples by the top 50 differentially expressed probes (ranked by p-values) showed separation of most of the autism cases from controls (). Although there was significant transcriptional heterogeneity of the autistic brains, gene ontology enrichment analysis revealed two convergent themes: down-regulation of genes involved in mitochondrial oxidative phosphorylation and in protein translation ( and Tables S4
). This has not been reported in gene expression or protein interaction network analyses of autism 
, in gene ontology analyses of large cohorts of autistic individuals with monogenic disorders and rare genomic copy number variants 
, or in recent whole-exome sequencing studies 
Down-regulation of Genes of Mitochondrial Oxidative Phosphorylation and Protein Synthesis in Autistic Brain
Down-regulation of genes for mitochondrial respiratory chain complexes I, III and ATP synthase was particularly notable (Table S4
). Abnormalities of mitochondrial oxidative phosphorylation are among the more commonly identified biochemical findings in individuals with autism 
. In most instances it is unclear whether these findings are primary genetic defects in oxidative phosphorylation or secondary phenomena, although some individuals having autism and mutations in genes of oxidative phosphorylation have been reported 
. Recent work showed reduced levels of these respiratory chain complexes in autistic brain 
To our knowledge, down-regulation of genes of mitochondrial oxidative phosphorylation has not been reported in prior gene expression analyses of autistic brain. The differences in the results between our study and other whole genome transcriptomic analyses of autistic brain 
may be due to etiological heterogeneity together with the use of different criteria for sample selection in each study. Because of the known etiological heterogeneity of autism, we designed our study to reduce variability in the data and increase statistical power through stringent subject selection criteria (see Methods
). A more heterogeneous sample in other studies could limit the ability to identify differences in the expression of mitochondrial oxidative phosphorylation-related transcripts if the findings that we observed apply only to a subset of autistic persons (eg, males with idiopathic autism). The differences in the results of our study and other transcriptomic analyses of autistic brain might also be due to differences in the regions of brain that were analyzed. It is possible that the differences in metabolic demands or brain region specific pathophysiology could account for differences in expression of mitochondrial genes. The finding of down-regulation of genes of oxidative phosphorylation, if independently confirmed, will be important in future studies of pathophysiology of autistic brain.
We also noted significant down-regulation of genes of protein synthesis. This finding suggests a generalized and non-targeted process and may be a result of dysregulation of other pathways. Alterations of neuronal protein synthesis involving selected pathways are increasingly understood to be key aspects of the pathogenesis of some forms of syndromic autism 
and may be generalizable to many forms of cognitive and behavioral disease 
Our data do not inform whether the down-regulation of genes of mitochondrial oxidative phosphorylation or of protein synthesis mainly relates to glia and/or neurons. Of note, the differentially expressed genes of both mitochondrial oxidative phosphorylation and protein translation observed in this study are highly enriched for regulation by HNF4A
in transcription factor analysis (p
1.4E-8; Fisher’s exact test) (Figure S1
is involved in tissue-specific cell differentiation and energy metabolism and has been recently identified as a cerebellar-specific marker of the hypoxia response in mouse brain 
. To our knowledge, an involvement of HNF4A
in the pathophysiology of autism has not been previously noted.
Differential Expression of Genes of Synapse and Other Brain Functions
Numerous genes of brain development and function have been associated with the pathophysiology of autism with varying levels of evidence; the specific genes that are implicated vary across studies 
. To explore specific genes that might be relevant in our dataset, we reviewed the list of probes that were differentially expressed at a log2
-fold change of ≥0.7. Of these 32 genes, there were two genes associated with synapse/neurotransmitters (LIN7B, SYN1
), three genes associated with vesicle transport (VPS29, HSPB1, TUBB2B
), one gene associated with brain patterning (GPR56
), four genes that are of significance for normal brain function based on monogenic neurologic disorders associated with mutations in those genes (ODZ3, ATP1A2, MOCS2, PTS
) and another gene with brain-specific expression (BEX5
) (Table S3
). Eight of these genes have some evidence in support of an association with autism (http://autismkb.cbi.pku.edu.cn/
, two of which (SYN1, PTS
) met thresholds for a high level of evidence 
, and concordant changes in the differential expression in autism vs control were found for three of the genes (GPR56, HSPB1, BEX5
) between this study and the work of others 
. These data provide additional support for the hypotheses that genes of synapse formation/function and of cortical development are differentially expressed in autism 
Brain Immune Gene Dysregulation in a Subset of Autistic Brains
We also separately analyzed the gene expression data of the three autistic cerebellar outlier samples. Differentially expressed genes between these outliers and matched controls were enriched for NF-kB signaling and cell cycle regulation pathways (Figure S2
). The top 300 differentially expressed probes were used to generate a non-redundant gene list and compared with two published differentially expressed autism gene lists in brain 
). There was a statistically significant overlap in the lists (P
1.4E-9; Fisher’s exact test), notably including genes associated with inflammation. While the role of neuroinflammation in autism pathophysiology remains unclear 
, our data emphasize the substantial interindividual heterogeneity of brain immune system gene dysregulation.
Brain Gene Expression Profiles and Associations with Specific Clinical Phenotypes
We next explored associations between phenotype and transcriptional profile with weighted genome co-expression network analysis (WGCNA) 
. This approach can potentially identify higher order, systems level correlations. We examined potential associations between expression of all gene modules (ie, clusters of co-expressed genes) in cerebella and ADI-R total or ADI-R domain scores, which are associated with phenotypic severity in the three cardinal domains of autism: impairment of social reciprocity, impairment of verbal or non-verbal communication, and stereotyped or repetitive behaviors. Two gene modules were significantly associated with the social interaction domain of the ADI-R and one gene module significantly associated with the stereotyped and repetitive behavior domain of the ADI-R (). These three modules are characterized by gene ontology enrichments relating to purinergic signaling/immune response, inflammation/immune response, and myelin/myelination, respectively (). There were no associations for other covariates tested (). Correlations between autism behavioral endophenotypes and CNS gene expression patterns have not been reported.
Specific domains of the Autism Diagnostic Interview-Revised (ADI-R) are associated with specific gene modules.
Several types of data support the observed associations. Altered immune responses and neuroinflammation are well documented in autism 
and there are reports of specific associations between immune parameters and deficits in social interactions and communication in autistic individuals 
. Purinergic signaling is involved with synaptic plasticity and neuron/glia interactions 
and recent genetic and clinical data implicates a role for purines/purinergic signaling in autism 
. In addition, abnormalities of brain white matter are commonly observed in diffusion tensor imaging-based neuroradiological studies of autism 
and abnormalities of cerebellar white matter have been associated with repetitive behaviors in autistic individuals 
. Determination of the pathophysiology and clinical meanings of these associations will be important next steps.
Validation of Gene Expression Microarray Data
Quantitative RT-PCR was used to validate the results of the microarray gene expression analyses (Figure S3
). We selected candidates based on strong effect size on the gene expression microarray analyses or biological plausibility based on published data. The results of all 5 statistically significant differentially expressed transcripts by RT-PCR were concordant with microarray data, confirming the validity of transcriptomic analysis by a gene microarray approach.
Brain Whole Genome Methylation Analysis
To explore a possible epigenetic basis of the gene expression findings, we obtained high quality methylation data for 9 pairs of age- and gender-matched cerebella and 8 pairs of BA19 cortex and analyzed >27,000 probes for differential methylation at individual CpG dinucleotides enriched in promoter regions of genes. Correspondence analysis indicated that region and age were the major sources of variability in the data () and, after controlling for these covariates, there were no significantly differentially methylated genes in either autistic cerebellar or BA19 cortex (data not shown). We also examined the genomic proportion of methylated loci in cerebellar and BA19 cortex. There was no significant difference between autism cases and controls within each region, but there was a statistically significant difference of the proportion of global DNA methylation between cerebellar and BA19 cortex, consistent with previous analyses of normal brain 
(). We then used WGCNA to analyze the full DNA methylation dataset; there were no modules associated with the autistic phenotype in either cerebellar or BA19 cortex (data not shown).
No differential methylation of genomic DNA was identified between control and autism cerebellar cortex or BA19 cortex.
Methylation Analyses of Candidate Autism Genes
Several investigators previously reported epigenetic alterations in RORA
in lymphoblasts 
, and OXTR
in brain tissue 
from small numbers of individuals with autism. We evaluated each of these genes using the DNA methylation microarray data and did not find differential methylation for any of the multiple probes associated with each of the genes (Table S9
). Although CpG dinucleotides within a particular promoter CpG island are generally correlated with each other and with gene regulation, we could not exclude the possibility that some unmeasured CpG dinucleotides within an island we investigated were differentially methylated. We therefore performed bisulfite pyrosequencing of the most relevant regions of the CpG islands; no differentially methylated CpG dinucleotides were noted in MECP2
(, Figure S4
, Table S7
Strengths and Limitations of the Current Work
We recognize several limitations to our study. First, the sample size is necessarily relatively small as it is difficult to procure brain specimens from numerous persons with autism, particularly after stringent selection criteria are used. Second, as is the case for any study, we cannot exclude that our results are artifacts of an unmeasured covariate. However, we tried to exclude potential confounders by exploring postmortem interval, cause of death, seizure comorbidity and age, as well as tissue variables such as RNA quality and tissue sampling error. Third, although we had excellent genomic coverage on both our methylation and gene expression arrays, we did not have complete coverage and we may have missed small epigenomic or transcriptional changes with pathologic relevance. We did not evaluate histone modifications, which can account for expression differences in the absence of differential DNA methylation. Fourth, we evaluated two carefully chosen brain regions but in so doing may have missed important gene expression or DNA methylation changes in other brain regions. Fifth, we evaluated whole tissue extracts and might have missed epigenetic or transcriptional dysregulation specific to a subpopulation of clinically relevant cells.
This study has several strengths. More than 100 uncommon or rare causes of autism are currently described yet autistic individuals with defined syndromic, metabolic or chromosomal etiologies comprise a minority of all individuals with autism. The pathophysiologic processes operative in some, or perhaps most, persons with these uncommon causes of autism may not necessarily be the same as the disease processes present in persons with idiopathic autism. Our subject selection process was, therefore, rigorous and included evaluation of clinical and molecular biologic data to produce a set of stringently defined idiopathic autistic male brains. We used high resolution microarrays to provide a detailed map of the epigenetic and transcriptional profiles of the analyzed brain regions. We also used complementary confirmatory techniques to evaluate candidate genes.
Taken together, our analyses demonstrate significant transcriptional heterogeneity in autistic brains with decreased expression of mitochondrial oxidative phosphorylation and protein synthesis-related genes. The latter suggests that convergence of pathobiology in autism may occur at the level of fundamental cell processes such as mitochondrial energy metabolism and protein translation. Nonetheless, the transcriptional heterogeneity noted here also raises the possibility there may be diverse sets of biological pathways that are also dysregulated in different autistic individuals and this, in turn, may be relevant in subtyping individuals for clinical trials. Additionally, our data suggest that there are particular biological pathways associated with different symptom domains in autism. Perhaps clinical heterogeneity in autism is the result of differences in these specific pathways which act to control domain severities semi-independently. We found that DNA methylation of gene promoters is not widely disrupted in the autistic cerebellum or BA19 brain regions, suggesting that transcriptional changes are largely the result of other regulatory mechanisms.