We measured the expression levels of up to 18,462 transcripts, representing approximately 15,106 genes, using Affymetrix GeneChip^® human U133A microarrays. These transcripts corresponded to 20,261 probe sets, excluding 2,023 Affymetrix bacterial and housekeeping control probes and probes that do not map to any chromosomes. We performed principal components analysis (PCA) to explore the gene expression profiles from four regions (cerebrum, cerebellum, heart, and cerebrum-derived astrocyte cell lines) in human fetal samples diagnosed with TS21 and matched euploid controls (see Additional data file 1). PCA allows the visualization of highly dimensional data along principal component (PC) axes. These axes reflect the degree of variance in the data, allowing the identification of groups of data points having possible biological relevance. For example, two points corresponding to tissue samples that are close together in PCA space are likely to have highly similar overall gene expression profiles. Figure 1 shows the 25 tissue samples mapped from high-dimensional space to three dimensions for exploratory visualization. The first three PCs are displayed on the x-, y-, and z-axes, respectively. The percentage of total variance explained by each PC is displayed on the corresponding axis. This analysis was performed on 253 probe sets (chromosome 21) and 20,008 probe sets (non-chromosome 21) separately. Figure 1 shows that for chromosome 21 and non-chromosome 21 genes, the samples clustered primarily by tissue or cell type. Thus, the largest differences in overall gene expression between the samples exhibited by PCA are attributable to the different tissues or cells. For genes on chromosome 21, TS21 is distinguishable from euploid controls on the third PC, which accounts for 17.2% of the total variation in 253-dimensional data (Figure 1b). In contrast, PCA mapping of non-chromosome 21 genes (Figure 1c,d) showed no distinction between TS21 and euploid controls. Although only the first three PCs are displayed in Figure 1, a difference between TS21 and euploid controls was not significant on any of the PCs (based on a t test performed on each PC; data not shown).

To further explore the relationships between samples based upon gene expression profiles, we performed hierarchical clustering using average linkage with Euclidean distance (Figure 2). Hierarchical clustering and PCA are 'unsupervised' methods, which do not consider the known sample attributes such as tissue type or disease state when organizing the data. We superimposed the sample information using color coding. Consistent with PCA, cluster analysis indicated that the samples clustered primarily by tissue source in both chromosome 21 genes and non-chromosome 21 genes. The clustering for the chromosome 21 genes showed a tendency to cluster by disease type within the tissue clusters (Figure 2a), whereas no obvious clustering by disease type was evident in the primary clusters or sub-clusters of genes not on chromosome 21 (Figure 2b). Cluster analysis and PCA results are consistent with the hypothesis that TS21 samples are distinguishable from matched euploid samples based upon differences in the expression of genes assigned to chromosome 21. Additionally, these exploratory analyses revealed no substantial outliers or other anomalies in the data.

We used a mixed-model ANOVA to test the first three hypotheses stated in the introduction. The hypotheses tested included multiple tests on chromosomes or individual genes. Therefore, to protect against false discoveries due to multiple testing, we used the step-up 'false discovery rate' (FDR) 30. We set the FDR at 0.05, meaning that the list of significant genes after applying FDR is expected to contain 5% false positives.

For the first hypothesis, we assessed whether genes assigned to each chromosome displayed overall differential gene expression. Only chromosome 21 showed significant mean overall differential expression between TS21 and euploid controls (Figure 3). Genes on chromosome 21 were expressed at 1.37 ± 0.02 fold (mean ± standard error), while the ratio of TS21/control across the other chromosomes was 1.00 ± 0.02 (ranging from 0.96 ± 0.03 to 1.02 ± 0.03). For this first hypothesis, 23 chromosomes were tested (chromosomes X and Y were combined), so the FDR is based on n = 23 tests.

For the second hypothesis, we tested whether individual genes assigned to chromosome 21 were differentially expressed in TS21 versus euploid samples. A mixed-model ANOVA (see Materials and methods) identified 26 out of 253 chromosome 21 probe sets (10.2%) with statistically significant differential expression at a FDR of 0.05. These most consistently dysregulated genes are listed in Table 1. For 104 gene expression comparisons listed in Table 1, 103 were increased in TS21 relative to controls. For this hypothesis, the FDR was based on n = 253 tests (for the number of probe sets assigned to chromosome 21).

For the third hypothesis, we tested whether individual genes not assigned to chromosome 21 were differentially expressed in TS21 relative to euploid samples. The presence of such genes would indicate whether the condition of TS21 causes changes in the transcriptome on chromosomes other than 21, possibly as a secondary consequence of the trisomy. Out of 20,008 non-chromosome 21 probe sets, 14 exhibited statistically significant differential expression at a FDR of 0.05 (Table 2). Using an alternative approach, we performed FDR on each chromosome separately with similar results (Additional data file 2). The same 14 genes passed FDR at the 0.05 level, as well as three additional genes (2,4-dienoyl CoA reductase 1 (NM_001359) and cholinergic receptor, nicotinic, alpha polypeptide 2 (NM_000742), both assigned to chromosome 8, and small inducible cytokine subfamily A (Cys-Cys), member 21 (NM_002989), assigned to chromosome 9). For chromosome 21 genes, 10.3% passed FDR at 0.05; for all other chromosomes, the greatest number of genes passing was 0.3% (chromosome 18) (Additional data file 2).

Based on the mixed-model ANOVA, a large proportion of chromosome 21 genes (n = 26 probe sets/253) showed significant altered expression at a FDR of 0.05, while a very small proportion of non-chromosome 21 genes (n = 14 probe sets/20,008) were significantly regulated. We further visualized this phenomenon by plotting a histogram of all the p values obtained for chromosome 21 genes (n = 253; Figure 4a) and for non-chromosome 21 genes (n = 20,008; Figure 4b). The histogram in Figure 4a contains 20 bins, at intervals of 0.05. If there were no truly differentially regulated genes, each bin would contain 253 × 0.05 = 12.65 transcripts (horizontal line on the figure). The figure indicates that there are many more small p values than expected by chance; there are 62 transcripts with p < 0.05, while only about 13 would be expected to be less than 0.05 by chance. For non-chromosome 21 genes (Figure 4b), the expected number of genes having a p value less than 0.05 by chance was 1000.4 (20,008 × 0.05), whereas the observed number of genes having p < 0.05 was 1,419. Although there was some tendency for the p values to be smaller than expected by chance, these two histograms provide a visual display of the extent to which the expression of many chromosome 21 genes are significantly different between TS21 and controls, whereas few genes assigned to other chromosomes were significantly regulated.

We asked whether there were regional differences among the significantly regulated genes. For those genes assigned to chromosome 21 (Table 1), the mean ratio of TS21/euploid mRNA level was 1.58 ± 0.05 (mean ± standard error) in the fetal brain tissues and astrocyte cell lines derived from the frontal cortex. Similarly, the TS21/euploid expression ratio in fetal heart was 1.60 ± 0.09 (with the exception of TMEM1, for which the TS21/euploid ratio was 9.58). These results are consistent for a gene expression dosage effect caused by trisomy. However, for significantly regulated genes that were not assigned to chromosome 21 (Table 2), a large percent were abundantly expressed and significantly different between TS21 and euploid samples only in the heart, but not in the brain. These genes included myomesin 1, myoglobin, calsequestrin 2, cardiac troponin I and T2, and alpha 1 actin.

To more completely assess differential gene expression, we investigated the ability to classify tissue samples as TS21 or euploid controls using genes on chromosome 21 and genes on chromosomes other than 21. The accuracy estimate for classification using chromosome 21 genes was 99.91% correct, whereas the estimate for classification using non-chromosome 21 genes was only 48.63% correct. Tables 3 and 4 show the classification results for the nested cross-validation using chromosome 21 genes and those using non-chromosome 21 genes (see Materials and methods and Additional data file 3). As expected, we were able to classify the tissue samples with very high accuracy using chromosome 21 genes (Table 3). The classification accuracy when using non-chromosome 21 genes was, however, approximately equal to the accuracy expected by chance (Table 4).

Based upon Gene Ontology (GO) annotations 313233, each of the probe sets represented on the Affymetrix GeneChip^® human U133A microarray, having a signal intensity above a background cutoff level, was either assigned to a GO functional group, or else defined as a member of a set excluding that functional group ('non-group members') (see Materials and methods). We asked whether our microarray data might indicate any particular functional groups of genes that were dysregulated in the TS21 samples compared to euploid controls. To address this question, we first performed permutation tests to establish the presence of a signal in the data. Due to the acyclic tree structure of the GO database, with multilevel interconnecting nodes, it is unclear which further permutation test might be performed to optimally define regulated groups. We therefore next applied a t test (or Wilcoxon's rank test for groups with only one or two members) to the gene expression data for two groups of probe sets: each given functional group, and the non-group members. This process was then repeated for all the functional groups. We found 1,141 functional groups for the cerebrum, 1,179 functional groups for the cerebellum, 1,126 functional groups for the astrocyte cell lines, and 1,180 functional groups for the heart.

The first 15 functional groups with the smallest p values for each tissue/cell type are listed in Tables 5, 6, 7, 8. In particular, the mitochondrion group (n = 417 probe sets) in the fetal cerebrum and heart tissues had the smallest p values from our functional group statistical analyses (Tables 5 and 8). Several other groups related to metabolic pathways, such as oxidoreductase activity (n = 299, in the cerebrum), NADH dehydrogenase activity (n = 31, in the cerebrum and heart), and mitochondrial inner membrane (n = 74, in the heart) were also among the most statistically significantly regulated functional groups (Tables 5 and 8).

To establish that there is signal in the data, we also performed permutation tests. For each functional group, a two sample t test was carried out, testing for a difference in expression for genes associated with this functional group compared to all other observed gene expression levels. If there were no signal in the data, a random assignment of the expression levels (obtained for example by randomly shuffling the observed expression levels) would yield comparable results. However, the distribution of p values obtained from 100 permutation tests (indicated by 100 black lines in the plots) are vastly different from those observed in the original data, indicating that the assumption of no signal in the data was wrong (Additional data files 4 and 5).

For GO functional groups having only one or two genes we applied a Wilcoxon rank test. In each tissue the lowest p value ranged from 0.0006 to 0.0726 for the top 20 GO functional groups having only one member, and 0.0001 to 0.1394 for groups having only two members. After correction for multiple comparisons, none of these values is significant (Additional data file 6), suggesting that none of the GO groups comprising one or two members was significantly regulated in TS21 samples from any tissue.

To confirm the altered expression levels of genes detected by microarrays, we performed over 5,600 quantitative real-time PCRs of cDNA derived from total RNA of the fetal samples. We selected a total of 28 genes from those that had shown the most consistent regulation by ANOVA (Tables 1 and 2), including 18 chromosome 21 genes and 10 non-chromosome 21 genes, based upon their abundance, fold regulation, and p values. We measured their mRNA levels by quantitative real-time PCR in four tissue/cell types, and compared these levels between TS21 and euploid samples. The hypoxanthine phosphoribosyltransferase (HPRT) housekeeping gene was used as a control gene for normalization between samples. Melting curves and gel electrophoresis of PCR products confirmed the identity of the amplification products (data not shown). The directions of dysregulation and fold changes from real-time PCR results were generally consistent with our microarray findings (Tables 9 and 10). Most genes showed increased transcript levels by both microarray and real-time PCR. Two non-chromosome 21 genes, RRAD and ADAMTS8, were down-regulated in the fetal TS21 heart consistently in microarray and PCR experiments. An example of the results from one real-time PCR experiment for the ZNF 294 gene is shown in Additional data file 7.