Expression differences between autopsy and resection samples

Gene expression profiles were determined in six resection samples from hippocampus and frontal cortex, and in four and six autopsy samples from hippocampus and frontal cortex, respectively, using Affymetrix^® HG U133plus2 arrays (see Materials and methods). Of the 54,613 probe sets on the microarray, 42,427 (77.69%) gave a detectable hybridization signal in at least one individual (see Materials and methods). Among these probe sets, we found 5,703 with a significant difference in expression (13.4%) using analysis of variance (ANOVA) with a nominal significance cutoff of 0.01 (false discovery rate (FDR) = 4.12%, permutation test) and 8,643 using significance analysis of microarrays (SAM) at the 5% FDR cutoff. Out of the 5,703 probe sets identified in ANOVA, 5,515 (96.7%) overlapped with the probe sets identified by SAM. Further, of these 5,703 probe sets, 4,508 differed significantly (p < 0.01) between autopsy and resection samples in both brain regions while 981 probe sets showed a significant difference between autopsy and resection samples as well as between brain regions (Figure 1). For none of these 5,489 probe sets did the differences between autopsy and resection samples depend significantly on the brain region. Finally, for 214 probe sets (0.5% of all detected ones), expression differences between autopsy and resection samples differed significantly (p < 0.01) depending on the brain region examined. This indicates that death-induced expression changes are highly consistent in both brain regions and influence only a small fraction of the total observed expression differences (214 out of 5,703).

Among 20 GO categories with significant under-representation of genes differently expressed between autopsies and resections, seven are involved in cell communication, three in response to stimulus, two in sensory perception, and four in development. In addition, 'cellular physiological process' and 'organismal physiological process' are among the GO categories that are significantly conserved in their expression between autopsy and resection samples (Table 2).

In contrast, no chromosome showed either an excess or lack of expression differences (two-sided binomial test, p < 0.341, corrected for multiple testing).

Expression differences between brain regions

To test whether in vivo expression differences between the brain regions are conserved in postmortem samples, we first considered the ANOVA results (Figure 1). Among 42,427 probe sets with hybridization signals detectable in at least one individual, 6,568 (15.5%) showed significant expression differences between the two brain regions analyzed (nominal significance p < 0.01, FDR = 3.6%, permutation test). Out of these probe sets, 6,431 (97.9%) overlapped with the ones identified by SAM (FDR = 5%). In 234 of these 6,431 probe sets, differences between brain regions depended significantly on the source of sample material (p < 0.01). Thus, although autopsy and resection samples differ substantially with regard to their gene expression profiles, the patterns of expression differences between the brain regions remain largely preserved.

We tested further whether in vivo expression differences between the brain regions are conserved in the postmortem samples by separately identifying, independent of the ANOVA results, probe sets differently expressed between the brain regions in the autopsy and in the resection samples. Using Student's t test with nominal significance p < 0.01, we found 788 and 3,943 probe sets with a significant difference in expression between the brain regions in the autopsy and in the resection samples, respectively (FDR = 22.8% and 4.3% respectively, permutation test). Similarly, using SAM with FDR = 5% we found 874 and 6,699 probe sets with a significant difference in expression between the brain regions in the autopsy and in the resection samples, respectively. This large discrepancy in the numbers of differences between the brain regions when the autopsy and resection samples are considered separately seems to contradict the ANOVA results. To address this, we examined whether probe sets that do not overlap between these two lists tend to show the same pattern of change between the brain regions or, alternatively, are completely uncorrelated in their expression behavior. For this purpose, we considered all probe sets present on either of the two lists and calculated the strength of correlation of the expression difference between the brain regions measured in the autopsy and in the resection samples. We found a strong and significant correlation between the expression differences for both t test (Pearson's correlation R = 0.763, N = 4,471, p < 10^-15) and SAM results (Pearson's correlation R = 0.726, N = 7,162, p < 10^-15) (Figure 2). Similarly, we found slightly reduced but still highly significant correlations using expression differences normalized to the average variation (effect size) (Pearson's correlation R = 0.566, p < 10^-15 and R = 0.584, p < 10^-15, respectively). Thus, expression differences betweenthe two brain regions are largely concordant in the autopsy and resection samples. Interestingly, the slopes of the regression lines (β) fitted through the distributions of the expression differences between the two brain regions in the autopsy and the resection samples equal 0.49 for both sets of genes (Figure 2). An even stronger effect was observed using the effect size measurements (β = 0.33 and β = 0.32 for t test and SAM results, respectively). Thus, despite an overall agreement of the measurements of expression differences in the two sources of sample material, the amplitude of expression differences measured in the autopsy samples is, on average, half of that observed in the resection samples. Limiting the regression to genes with a high expression difference amplitude in either autopsy or resection samples did not change this effect. Interestingly, it was even more pronounced for genes with lower expression in the frontal cortex compared to the hippocampus (β = 0.27 and β = 0.34 for t test and SAM results, respectively). Since the significance test depends on the effect size, smaller expression differences explain the reduced number of identified probe sets in the autopsy samples.

Influence of death on expression variation

All microarray studies involving postmortem human samples report substantial biological variation among individuals. We asked whether death-induced expression changes contribute to this variation by affecting different individuals to different degrees. To do this, we examined published gene expression data from 40 brain autopsy samples [13]. First, we asked whether probe sets that differ in expression between autopsy and resection samples vary more among individuals in this dataset than other probe sets. From the 16,376 probe sets with a detectable hybridization signal in at least one of the 40 individuals, 1,752 overlap with the probe sets showing significant differences in expression between autopsy and resection samples. Using logarithm transformed variation measures, we found no significant difference between the expression variation among these probe sets and among the remaining probe sets (Student's t test, p = 0.916). Thus, genes that differ in expression between autopsy and resection samples do not vary more among postmortem samples compared to the other genes.

Next, we asked whether the amplitude of death-induced expression changes correlates with the duration of postmortem interval. To test this, we computed correlations between gene expression levels and postmortem delay in the 40 brain autopsy samples for 1,752 probe sets that differ in expression between autopsy and resection samples and for 1,000 subsets of the same size randomly sampled from the other 14,624 probe sets. In 837 out of 1,000 random subsets, the correlation was greater or equal to the one observed for probe sets with significant difference in expression between autopsy and resection samples. Thus, genes that differ in expression between autopsy and resection samples do not correlate more with duration of postmortem interval than the rest of the detected genes.

Tissue samples and microarray data collection

Human postmortem samples were obtained from the National Disease Research Interchange. Informed consent for use of the tissues for research was obtained in writing from all donors or the next of kin. None of the subjects had a history of neurological disease or had indications of brain abnormalities at the tissue level as determined at autopsy. All individuals suffered sudden death for reasons other than their participation in this study and without any relation to the tissues used. Human resection samples were obtained from patients with brain tumors and/or chronic pharmaco-resistant epilepsy who underwent surgical treatment in the Surgery/Epilepsy Surgery Programs at the University of Bonn Medical Center. In all patients, surgical removal of the tumor/lesion tissue was necessary. Informed consent for additional studies was obtained in writing from all patients. The diagnosis of the individual patients is presented in Table 1. All procedures were conducted in accordance with the Declaration of Helsinki and approved by the ethics committees of the respective institutions. Representative tissue sections were snap frozen at -80°C. Based on neuropathological analyses by means of hematoxilin and eosin stainings, normal tissue adjacent to the tumor or lesions was used for subsequent experiments. Intense care was taken to avoid tumor infiltrated tissue. None of the surgically obtained tissue samples used in this study, with the exception of four frontal cortex samples with focal cortical dysplasia, showed any histological abnormalities. Age, sex, and degree of relatedness of all individuals are listed in Table 1.

All samples were processed in parallel starting from the frozen tissue by the same person (HF) in random order with respect to brain region and the source of sample material. Total RNA was isolated from approximately 50 mg of frozen tissue using TRIZol^® (GIBCO, San Diego, CA, USA) reagent according to the manufacturer's instructions and purified with QIAGEN^® RNeasy^® kit (Valencis, CA, USA) following the 'RNA cleanup' protocol. All RNA samples were of high and comparable quality as determined by the ratio of 28S to 18S ribosomal RNAs estimated using the Agilent^® (Palo Alto, CA, USA) 2100 Bioanalyser^® system and by the signal ratios between the probes for the 5' and 3' ends of the mRNAs of GAPDH used as quality controls on Affymetrix^® (Santa Clara, CA< USA) microarrays (Table 1). Labeling of 1.2 μg of total RNA, hybridization to Affymetrix^® HG U133plus2 arrays, staining, washing and array scanning were carried out following Affymetrix^® protocols. All primary expression data are publicly available at the ArrayExpress database (accession number E-TABM-20) [22].

Microarray data analyses

Affymetrix^® microarray image data were collected with Affymetrix^® GeneChip^® Operating Software version 1.1 using default parameters. We used the robust multichip average (rma) procedure [23] for array normalization and calculation of expression base two logarithm transformed intensity values. Since logarithm-transformed intensity values are approximately normally distributed, we used them for all analyses. We calculated detection p values using the Bioconductor 'affy' software package [24]. We defined probe sets having a detectable hybridization signal using Affymetrix default detection cutoff of 0.065.

We used ANOVA to identify probe sets that showed a statistically significant change in expression depending on the brain region or on the source of sample material among human samples using the following model: Y _ij = μ _j + source _i + region _i + (source*region) _i + ε _ij . In this equation, Y _ij is the base two logarithm of the expression level for probe set j in sample i, μ is the mean expression level of a probe set j, source _i is the term for the effect of the source of sample material, region _i is the term for the effect of the source of the brain region, (source*region) _i is the term for the interaction effect of the two factors, and ε _ij is the error term. For each term we used a nominal significance level of 0.01. In order to estimate an average number of probe sets expected by chance at this significance cutoff, we applied the same ANOVA approach to 1,000 datasets constructed by random permutation of the sample labels in the original data.

Alternatively, differently expressed probe sets were determined using SAM software version 2.01 with 5% FDR cutoff [25]. In all cases except the analysis of epilepsy effects, we performed t statistics on the logarithm transformed expression values. FDR estimates were based on 500 permutations of the samples within the set. We used block permutation design for the two-factor analysis and time course for the analysis of epilepsy effects. Effect of epilepsy was scored based on the diagnosis and seizure type: 0, no diagnosed epilepsy; 1, simple partial seizures; 2, simple and complex partial seizures; 3, complex partial seizures; 4, simple and complex partial seizures, grand mal; 5, complex partial seizures. Effect size was calculated as a difference between means divided by the pooled standard deviation. The pooled standard deviation was defined as the square root of the average of the squared standard deviations.

Functional analysis and distribution on chromosomes

To functionally annotate the probe sets on the Affymetrix^® HG U133plus2 arrays, we integrated information from four public databases: Affymetrix^® NetAffx™ (12/2004 release) [26], LocusLink (12/2004 release) [27], and Gene Ontology (12/2004 release) [28]. Affymetrix^® probe sets were linked to the corresponding genes using LocusLink annotation provided by NetAffx™. When a single gene was represented by multiple probe sets, the gene was classified as detected if at least one probe set was detected and classified as differentially expressed if at least one probe set was both detected and differentially expressed. Genes were assigned to their GO annotations from each of the three GO taxonomies ('molecular function', 'biological process', and 'cellular component') using GenMapper [29, 30]. Note that a gene belongs to its assigned GO group as well as all higher groups in the taxonomy.

To assess if the overall distribution of genes differentially expressed between autopsy and resection samples across the groups in a GO taxonomy differs significantly from the distribution of all detected genes, we compared it with 10,000 random sets in which the same number of differentially expressed genes was randomly drawn from the annotated detected genes as described elsewhere [18]. GO groups with significant excess and with significant lack of expression differences between autopsy and resection samples were determined independently using the hypergeometric distribution [18]. The percentage of false positive GO groups was estimated from the ratio of the number of significant groups in the observed data to the average number of the significant groups in 10,000 random sets. In the GO taxonomy 'biological process', we expect 20% false positives for the groups with significant excess and 5.8% false positives for the groups with significant lack of expression differences between autopsy and resection samples. Significant over-representation of up- or down-regulated genes in GO groups with significant excess of expression differences was determined by binomial test. Probability of up- and down-regulation within a group was based on distribution of all differently expressed genes. To assign chromosomal location to genes we used annotation provided by NetAffx™. Genes differently expressed between autopsy and resection samples were defined the same way as for the functional analysis.