The characteristics of the 90 libraries used for the analysis are shown in Table 1. The logarithms of the ratios of observed over expected transcript abundances (log obs/exp) are given in Additional file 1. These "log obs/exp" data were subjected to a Principal Component Analysis (PCA) to determine whether there were any major factors (in particular the two quantitative variables of the database, donor age and proteasome expression level) related to variability. Since the three most important factors accounted for only 31,2%, 18,2% and 15,7% of the variability observed, one cannot say that there is one major principal component involved.

However, a significant correlation between proteasome expression level (i.e. the total of transcript abundances found for all the proteasome genes in a given library) and a few of the factor loadings of the PCA: F2 (P = 2,65 × 10^-5), F3 (P = 0,0016), F4 (P = 3,94 × 10^-5) and F9 (P = 0,0039) was observed and together these factors are responsible for 46% of the variability.

Age also correlates with the loadings of some of the remaining factors: F1 (P = 0,016), F5 (P = 0,044) and F13 (P = 0,012), which cumulatively account for 40% of the variability. The higher P-values found for age could be due to the lower number of libraries, since age was known for only 68 of the 90 libraries.

Therefore both age and proteasome expression levels affect the proteasome transcriptional profile. The lack of correlation between proteasome expression level and age (P = 0,288) suggests that these factors exert independent effects within the database on the expression profile. The remainder of the analysis aims to visualize these effects.

As can be inferred from Table 1, there was an approximately 4-fold variation (49–203) in proteasome expression level. Since considerable variation will be due to regulation by transcription factors, variability in log obs/exp is only in part caused by epigenetic structures. Possibly the variation seen in proteasome expression level, seen in individual tissues, e.g. adrenals (75–174) and salivary glands (63–124), reflects largely variation due to gene regulation but the present data are not informative at this particular point.

To establish whether the variation in proteasome expression level affects the transcriptional profile of the proteasome, regression lines between the proteasome expression level and the "log obs/exp"of each individual gene-probe were calculated (Table 2).

The slopes of the regression lines differ widely. Three probes even have negative slopes, which means that tissues with a high proteasome expression level have less transcripts for these probes than tissues with a lower proteasome expression level, in other words: the transcriptional profile is different for tissues with different proteasome expression levels.

An example of a regression is shown for PSMA4 (Figure 1) and it can be seen that one library (kidney 2) falls well outside the general range and should therefore be considered an outlier. Some other gene probes have similar outliers; all representing strong under-expressions. In total, 8 libraries that deviated from the overall expression pattern were identified in this way. Seven of these deviated for 1 gene-probe with no preferential gene-probe being involved (liver 4, kidney 2, liver 2, heart 5, kidney 1, uterus 4 and ovary 4) and one library deviated for 5 gene-probes simultaneously (pancreas 2). These 8 outlying libraries were omitted from the subsequent analysis. Recalculation of the regression lines then revealed 4 probes with a significant negative slope (Table 2).

To determine the differences between genes in terms of absolute abundance of transcripts, regression lines were also calculated for proteasome expression level and absolute abundances (data not shown). These regression lines have been used to demonstrate the effect of proteasome expression level on the transcriptional profile of the proteasomal genes. Since Table 1 shows that the proteasome expression level lies within the range 49 – 203, the absolute abundances of transcripts at specific proteasome expression levels of 50 and 200 were calculated for each probe using its regression line (Table 3). This allows comparing the degree of transcription at the highest induction level (200) with the transcription at the lowest level (50). The figures differ for each probe but the response to high induction can be arbitrarily classified into 4 classes: some probes (B7c and A3a) are not induced and thus behave as constitutive genes (class A). Two probes (A4 and A1), already highly abundant at level 50, are enhanced about 2-fold at level 200. Other probes (A2a, A2b, A3b, A5, A6, B2, B6, B7a), which are moderately abundant at level 50 are induced to the expected 4-fold level (Class C), whilst others (A7a, A7b, B3, B5, B7b, B7d) which have extremely low levels of transcript abundance at level 50 are induced much more than 4-fold at level 200 (Class D).

One could therefore speculate that class D expression acts as a rate-limiting step in upgrading the number of proteasomes in the cell.

In order to determine the contribution of variation in proteasome expression level to the variability in "log obs/exp", the regression lines from the 82 libraries (Table 2) were used to calculate expected values for "log obs/exp" which were then subtracted from the observed ones. The resulting values (provided in Additional file 2 and designated "2nd-log obs/exp") are free from the effect of proteasome expression level. Since the variance of "log obs/exp" is 0,3558 and that of "2nd-log obs/exp is 0,2629, it is apparent that about 26,1% (1-(0,2629/0,3558) = 0,261) of the total variability in "log obs/exp" is due to the level of expression of the proteasomal genes.

The remaining variation in "2nd-log obs/exp" could be due either to factors associated with nuclear programming, such as gender, age and tissue specificity or to other factors such as post-mortem changes (if the biopsy was obtained by autopsy), 'noise' in transcription or experimental error. Also individual specificity in the transcription profile is, in theory, possible and belongs with the first factor type. To determine the possible influence of such factors, the libraries were divided into 2 classes by k-means clustering. The mean "2^nd log obs/exp", that indicate the deviations from expected transcription profiles, are depicted in Figure 2, with class 1 and class 2 showing major over-expressions or under-expressions respectively for 4 of the probes. These 4 probes belong to 2 genes and also belong to the group of 5 probes that were the most variable in "2nd-log obs/exp" (Figure 3B). The factors mentioned above were assessed for an eventual contribution to the variation in "2nd-log obs/exp" by comparing the numbers of factor variants in both classes and determining their probability by either a chi-square or Wilcoxon test. Since a comparison of the many variants in tissue specificity is not possible using this method, the mesodermic or endodermic origin of the tissues was used instead. The data in Table 4 demonstrate that with the exception of age (P = 3,64 × 10^-6) none of these factors was significant, thereby suggesting that there is an increase in under-expression with age. The mean ages in classes 1 and 2 were 53,5 and 61,9 years respectively.

To further analyze the influence of aging, the libraries were assigned to 3 separate groups as follows: A (22 – 50 years, 19 libraries), B (51 – 60 years, 19 libraries) and C (61 – 87 years, 24 libraries). The mean and standard deviation of the "2nd-log obs/exp" for each age group are shown in Figures 3A and 3B respectively. A number of observations can be made from the data presented:

1) it appears that the standard deviations (Fig. 3B) may decrease with age, since when subjected to the Wilcoxon test a significant decrease was observed for group B versus C (P = 0,001) whereas the P-value of the decrease for A versus C was 0,094; 2) the standard deviations are much larger for some of the probes (A7b, B3, B5, B7a, B7b, B7c); 3) the same probes seem to show differences between the age groups with respect to mean "2^nd-log obs/exp" and 4) the patterns of the mean "2^nd-log obs/exp" over the three age groups appear different for different probes. This latter observation was investigated further by subjecting the age-group data from Figure 3A to agglomerative hierarchical clustering (AHC). This analysis (Figure 4) identified three probe-types, i.e. type 1: A1, A2a, A3a, A4; type II: A2b, A3b, A6, B2, B3, B7d, and type III: A5, A7a, A7b, B5, B6, B7a, B7b, B7c. Consequently the data for mean "2^nd-log obs/exp" was pooled for each probe-type. The results, for the different age groups, shown in Figure 5 reveal that the three probe-types differ in relation to age. Therefore, for each probe-type, regressions were determined between age and the mean of the "2^nd-log obs/exp" for the contributing probes (Figure 6). Although the regressions for type 1 probes for mean "2^nd-log obs/exp" and mean abundance respectively, are not significant (P = 0,528 and P = 0,995), they are highly significant for type II probes (P = 0,0020 and P = 0,0006) and moderately significant for type III probes (P = 0,249 and P = 0,031.

It therefore appears that the type II and type III probes differ in their relation to age with type II showing an increase and type III a decrease in transcript abundance with aging. Two of the probes stand out in terms of a positive and negative correlation between age and transcript abundance, i.e. type II probe B2 (P = 0,020) and type III probe B7b (P = 0,009) respectively.

To analyse the observed decrease in standard deviation (Fig. 3B) the relationship between age and the 'deviation index' was assessed. The 'deviation index' represents the standard deviation of all "2^nd-log obs/exp" in a given library and indicates the extent to which the transcription profile of that library deviates from the mean standard transcription profile. It has previously been shown that this index is significantly enhanced in cancer tissues 2. If the aging process results from the increasing accumulation of mutational damage one would expect to observe a parallel increase in the deviation index during aging. The opposite, however, appears to be the case, since the regression of age with the deviation index of "2^nd-log obs/exp" (Figure. 7) shows that the deviation index decreases significantly with increasing age (P = 0,0117). This decrease might suggest that during aging, transcription becomes more accurate and/or less variable.

In order to determine whether all probes are similarly involved in this aging process, a new deviation index was first calculated in which the data from one probe was omitted and then the regression with age was recalculated. The rationale for this was that if a probe contributes to the aging effect, the P-value of the resulting regression will increase. This procedure was applied for each probe. The differences in the P-values (z-values) obtained in this way, are shown in table 5A. The data show that the probes differ widely in their contribution to the age effect. More than half are not involved and the deviation index calculated from these 10 probes does not correlate with age (P = 0,798). However the deviation index calculated from the 8 probes which do contribute, correlates strongly with age (P = 0,00027) with the 6 most influential probes being possibly even more strongly correlated (P = 0,00024). The reason for this probe specificity in aging remains a mystery, but the z-values (Table 5A) of the 8 probes providing the greatest contribution correlate significantly (P = 0,0086) with the slopes of the regression lines relating proteasome expression level and "log obs/exp" (Table 2). Therefore it appears that this aging phenomenon might be connected to the degree of upregulation and downregulation in the expression of these genes.