Inference of cell-type composition from human brain transcriptomic datasets illuminates the effects of age, manner of death, dissection, and psychiatric diagnosis

Most neuroscientists would agree that psychiatric illness is unlikely to arise from pathological changes that occur uniformly across all cells in a given brain region. Despite this fact, the majority of transcriptomic analyses of the human brain to date are conducted using macro-dissected tissue due to the difficulty of conducting single-cell level analyses on donated post-mortem brains. To address this issue statistically, we compiled a database of several thousand transcripts that were specifically-enriched in one of 10 primary brain cell types identified in published single cell type transcriptomic experiments. Using this database, we predicted the relative cell type composition for 157 human dorsolateral prefrontal cortex samples using Affymetrix microarray data collected by the Pritzker Neuropsychiatric Consortium, as well as for 841 samples spanning 160 brain regions included in an Agilent microarray dataset collected by the Allen Brain Atlas. These predictions were generated by averaging normalized expression levels across the transcripts specific to each primary cell type to create a “cell type index”. Using this method, we determined that the expression of cell type specific transcripts identified by different experiments, methodologies, and species clustered into three main cell type groups: neurons, oligodendrocytes, and astrocytes/support cells. Overall, the principal components of variation in the data were largely explained by the neuron to glia ratio of the samples. When comparing across brain regions, we were able to easily capture canonical cell type signatures – increased endothelial cells and vasculature in the choroid plexus, oligodendrocytes in the corpus callosum, astrocytes in the central glial substance, neurons and immature cells in the dentate gyrus, and oligodendrocytes and interneurons in the globus pallidus. The relative balance of these cell types was influenced by a variety of demographic, pre- and post-mortem variables. Age and prolonged anaerobic conditions around the time of death were associated with decreased neuronal content and increased astrocytic and endothelial content in the tissue, replicating the known higher vulnerability of neurons to aging and adverse conditions, and illustrating the proliferation of vasculature in a hypoxic environment. We also found that the red blood cell content was reduced in individuals who died in a manner that involved systemic blood loss. Finally, statistically accounting for cell type improved both the sensitivity and interpretability of diagnosis effects within the data. We were able to observe a decrease in astrocytic content in subjects with Major Depressive Disorder, mirroring what had been previously observed morphometrically. By including a set of “cell type indices” in a larger model examining the relationship between gene expression and neuropsychiatric illness, we were able to successfully detect almost twice as many genes with previously identified relationships to bipolar disorder and schizophrenia than using traditional analysis methods. Sample collection, including human subject recruitment and characterization, tissue dissection, and RNA extraction, was described previously (see Evans 2003 Neurobiol Dis 14:240-250, Vawter 2003 Neuropsychopharm 29: 373-384, Li 2004 Biol Psychiatry 55:346-352, Li 2013 PNAS 110: 9950-9955). The original sample size included dorsolateral prefrontal cortex (DLPFC) tissue from 172 subjects. We ran each sample on at least two microarrays (chip count=367) using the Affymetrix U133-A or U133Plus-v2 GeneChips, however, we only used the U133A part of the U133Plus-v2 GeneChip in our analysis and also 4 samples have been excluded by quality filtering. We applied RMA (Robust Multi-array Analysis) (see Irizarry 2003 Biostatistics 4:249-264, Irizarry 2003 Nucleic Acids Res 31:e15) to summarize probe set expression levels, using a custom Chip Definition File, resulting in expression data for 11,911 ENTREZ transcripts. All downstream analyses were performed in R. Details of the data processing, including data cleaning and normalization, are located in the Supplementary Materials for Li 2013 PNAS 110: 9950-9955 and Hagenauer et al. (submitted). Please note that [1] the data from the control subjects was previously uploaded to GEO as part of GSE45642 (GSM1124358-GSM1124518) [2] the original raw data for the 'DLPFC_Male_Control_r_p_UCDavis_10622' (GSM2431810) sample is unavailable and thus, the annotated version CEL file is provided.