Figures
Abstract
Background
Identifying biological pathways that vary across the age spectrum can provide insight into fundamental mechanisms that impact disease and frailty in the elderly. Few methodological approaches offer the means to explore this question on as broad a scale as gene expression profiling. Here, we have evaluated mRNA expression profiles as a function of age in two populations; one consisting of 191 individuals with ages-at-death ranging from 65–100 years and with post-mortem brain mRNA measurements of 13,216 genes and a second with 1240 individuals ages 15–94 and lymphocyte mRNA estimates for 18,519 genes.
Principal Findings
Among negatively correlated transcripts, an enrichment of mitochondrial genes was evident in both populations, providing a replication of previous studies indicating this as a common signature of aging. Sample differences were prominent, the most significant being a decrease in expression of genes involved in translation in lymphocytes and an increase in genes involved in transcription in brain, suggesting that apart from energy metabolism other basic cell processes are affected by age but in a tissue-specific manner. In assessing genomic architecture, intron/exon sequence length ratios were larger among negatively regulated genes in both samples, suggesting that a decrease in the expression of non-compact genes may also be a general effect of aging. Variance in gene expression itself has been theorized to change with age due to accumulation of somatic mutations and/or increasingly heterogeneous environmental exposures, but we found no evidence for such a trend here.
Citation: Hong M-G, Myers AJ, Magnusson PKE, Prince JA (2008) Transcriptome-Wide Assessment of Human Brain and Lymphocyte Senescence. PLoS ONE 3(8): e3024. https://doi.org/10.1371/journal.pone.0003024
Editor: Winston Hide, University of the Western Cape, South Africa
Received: July 22, 2008; Accepted: August 1, 2008; Published: August 20, 2008
This is an open-access article distributed under the terms of the Creative Commons Public Domain declaration which stipulates that, once placed in the public domain, this work may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose.
Funding: We are grateful for financial support from The Swedish Medical Research Council, the Swedish Alzheimer Foundation (Alzheimerfonden), and from the Johnnie B Byrd Institute. For the brain sample, data and biomaterials were collected from several National Institute on Aging (NIA) and National Alzheimer's Coordinating Center (NACC, grant #U01 AG016976) funded sites. John A. Hardy, PhD, Marcelle Morrison-Bogorad, PhD, Tony Phelps, PhD and Walter Kukull PhD are thanked for helping to co-ordinate this collection. The directors, pathologist and technicians involved include: National Institute on Aging: Ruth Seemann, John Hopkins Alzheimer's Disease Research Center (NIA grant # AG05146): Juan C. Troncoso, MD, Dr. Olga Pletnikova, University of California, Los Angeles (NIA grant # P50 AG16570):Harry Vinters, MD, The Kathleen Price Bryan Brain Bank, Duke University Medical Center (NIA grant #AG05128, NINDS grant # NS39764, NIMH MH60451 also funded by Glaxo Smith Kline): Christine Hulette, MD, Director, John F. Ervin, Stanford University: Dikran Horoupian, MD, Ahmad Salehi, MD, PhD, New York Brain Bank, Taub Institute, Columbia University (NYBB): Jean Paul Vonsattel, MD, Katerina Mancevska, Massachusetts General Hospital: E. Tessa Hedley-Whyte, MD, Karlotta Fitch, University of Michigan (NIH grant P50-AG08671): Dr. Roger Albin, University of Kentucky (NIH #AG05144): William Markesbery, MD, Mayo Clinic, Jacksonville: Dennis W. Dickson, MD, University Southern California: Caroll A. Miller, MD, Washington University, St Louis Alzheimer's Disease Research Center (NIH #P50AG05681): Dan McKeel, MD, John C. Morris, MD, Eugene Johnson, Jr., PhD, Virginia Buckles, PhD, University of Washington, Seattle (NIH #P50 AG05136):Thomas Montine, MD, PhD, University of Pennsylvania School of Medicine, Alzheimer's Disease Research Center: John Q Trojanowski, MD, Virginia M Lee, MD, Vivianna Van Deerlin, MD, Boston University Alzheimer's Disease Research Center (NIH grant P30-AG13846): Ann C. McKee, Sun Health Research Institute, Arizona (NIA #P30 AG19610): Joseph Rogers, PhD, Thomas G. Beach, MD, PhD, Emory University: Bruce H. Wainer, MD, PhD, Marla Gearing, PhD, University of Texas, Southwestern Medical School: Charles L. White, III, M.D., Roger Rosenberg, University of California, Davis: William Ellis, MD, Rush University Medical Center, Rush Alzheimer's Disease Center (NIH #AG10161): David A. Bennett, M.D. Julie A. Schneider, MD, MS, University of Miami/NPF Brain Endowment Bank: Deborah C. Mash, MD, Margaret J Basile. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Introduction
A decline in cell function with advancing age is a ubiquitous characteristic of all organisms. In humans, the effects of aging become manifest on a variety of levels that extend from an accumulation of DNA mutations to lipid oxidation, protein modification, cell loss, and ultimately death that is primarily due to increased susceptibility to age-related diseases [1]. Apart from overt changes, such as declining muscle strength, extensive metabolic alterations also occur with aging, one of the most prominent being impaired glucose tolerance [2]. Two central evolutionary theories hypothesize that the detrimental effects of aging are due to an accumulation of mutations, or antagonistic pleiotropy, whereby genes with beneficial effects early in life become deleterious with age [3]. These are not necessarily exclusive, and there is at present relatively strong evidence for both in studies of model organisms and in natural populations [3], [4].
Whatever the cause, there is value in charting the molecular sequelae of aging on as broad a scale as possible. Few other methodological approaches lend themselves as well to this as mRNA expression profiling. There have been a handful of studies that have attempted to catalogue how mRNA expression changes with age, the largest of which have been performed in kidney [5] and muscle samples [6]. An intriguing conclusion from the latter study is that there may be a common set of genes that change equivalently in different tissues. For example genes that make up the mitochondrial electron transport chain appear to decrease with age in different tissues, and this is supported in that decreases are also evident in mice and flies [6]. Importantly however, these studies remain relatively small in scale and few in number thus meriting larger studies in additional populations and tissues.
The effects of aging are particularly pronounced in the human brain where characteristic changes in morphology include a reduction in both neuronal size and synaptic density [7], [8]. On a behavioral level, decreases in motor and cognitive function are hallmarks of normal aging [9]. Dementia is the most prevalent disorder of the human brain, affecting 20–25 million people worldwide and representing a tremendous burden in terms of years of suffering with disability [10]. To date, there has been only one study, involving 30 individuals, on the effects of aging on gene expression at a transcriptome-level in the normal human brain [11]. There the authors described a number of pathways affected by aging and also concluded that genes for which down-regulation was evident had an over-representation of mutation in gene promoters. A recent study examined the effects of gene polymorphism on gene expression in a relatively large set of human brain samples [12]. An attractive feature of that study was that individuals, prior to death, were free of neurological disease. Our primary focus in the present study has been to use the above sample to investigate the question of whether general changes in gene expression occur in the aging human brain. This has been complemented with an analysis of lymphocyte mRNA expression in order to explore for common molecular themes in different tissues as well as to enable an assessment of a broader age range than that typically available in post-mortem samples.
Results
We began by performing several validation steps towards the initial goal of evaluating if age-related change in gene expression could be detected in the human brain. First, evidence was sought for systematic outlier effects among the 2,096,975 individual expression level estimates. We identified all expression values that were more than 3.4644 standard deviations from the mean expression value for that particular transcript (see methods for an explanation for choosing this threshold). With this, we expected to see approximately 1,086 outliers by chance in the entire set assuming normal distributions of the log10 transformed data. There were 7,079 outliers identified in this way, and these were eliminated from all further analyses. Second, the focus of the original study was primarily upon brain cortex, with 5 subsets being represented, 3 cortical, one group of 6 cerebellum samples and an additional group of samples for which no specific region could be assigned. We assessed whether these classifications differed systematically with regards to transcript detection rates and global expression levels using ANOVA (see methods). Across the various brain region categories there were no significant differences for any these covariates. Simple linear regression was also performed for age-at-death and pmi versus global expression level. For this, there was marginal evidence that age-at-death correlated with global expression (P = 0.080, r2 = 0.016). For pmi however, the trend was significant, but the direction of the curve suggested that expression was increasing with increasing pmi (P = 0.0039. r2 = 0.046). Of note, linear regression of the 2 global expression covariates (see methods) and detection rate metrics (one reflecting all 24,357 transcripts probed by the Illumina HumanRefseq-8 chip and the other reflecting the 14,078 transcripts detected in this brain sample) all showed strong correlation (r2 in excess of 0.5 for all 6 comparisons). Based upon all of the above tests, we resolved to use the global average expression level that includes all 14,078 transcripts as a primary metric of quality as a covariate for initial analyses.
For the brain sample, we regressed age-at-death on transcript levels one at a time for all 14,078 transcripts that had been detected in 5 or more individuals in the original study. The results of this are shown in Tables 1 and 2, where we document 54 transcripts that exhibited significant change after Bonferroni correction for 14,078 tests. For purposes of illustration, regression lines for the top 6 scoring genes (3 up-regulated and 3 down-regulated) are also displayed in Figure 1. For this analysis, the most significant finding was for the SVOP gene (encoding a synaptic vesicle protein), and the single strongest effect in terms of variance explained was for the TAC3 gene (encoding tachykinin 3). The smallest number of individuals for which a significant correlation with age was detected was 133 (also for the TAC3 gene). We noted that 3 of the top scoring 54 genes were also present in a list of 532 putative “housekeeping” genes (see [13]), these being PIN1, GUK1, and ARPC2. This proportion is not significantly different from what would be expected by chance (3 vs. 51 compared to 532 vs. 6968). At this stage, we also scrutinized this list of 54 genes more closely for the potential impact of covariates. For this, multiple regression models were fitted as above, but for each transcript in turn we also included terms for pmi, brain region, or gender. All 54 genes remained highly significant when any of these were individually tested (not shown). Among the top 54 transcripts, all were detectable in more than 100 individuals, reinforcing the importance of power in analyses such as this. There were however a few cases where a small number of observations gave rise to apparently large effects. The gene that ranked as the 139th most significant finding (falling below the multiple test correction threshold), ZIC3, had only 14 expression level observations and an r2 = 0.76 (P = 2.6×10−5). This highlighted the possibility that genes with lower maximum expression levels might be changing to a larger degree with age, but technologies with lower detection thresholds and/or larger sample sizes would be required to identify them.
We contrasted the individual transcript results of this analysis with the first study to address the impact of aging on expression at the transcriptome level [11]. In that study, 30 individuals were included with an age range of 26–106. Our set was scrutinized for the 148 genes previously highlighted to be associated with aging (see specifically table 1 from [11]). Only one gene from our set that survived strict multiple testing correction was also present in their set, this being MRPL28. With a relaxed uncorrected significance threshold in our set of 0.005, we found 26 overlapping genes. This proportion is significant given that we observed 1,141 transcripts in our set that were significant at the 0.005 threshold, suggesting that some replication exists, albeit limited (26 vs. 122 compared to 1,141 vs. 12,937). In our set, the majority of significant genes exhibited a decrease in expression with advancing age. An intriguing aspect of this comparison was that among the 26 overlapping genes, there were 9 for which expression increased with age in the original study, and for all 9 these were increased with age in the present study. Due to the consistency in the direction of the effects, this might be considered a stronger case for replication. Seven of these 9 genes were previously considered as a class of stress response genes.
We explored an expanded list of genes that exceeded a relaxed significance threshold (uncorrected P<0.05) for enrichment or deficit based upon Gene Ontology (GO) terms and the Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways. The entire set of genes was divided into two groups based upon whether expression was increasing or decreasing with age. We then searched for over- or under-represented terms and pathways in each list of genes using the DAVID web application [14]. For this we created a diminished non-redundant set of 13,216 base genes from the total set of 14,078 by eliminating multiple transcripts representing the same gene (see methods). The enriched terms (Bonferroni corrected P<0.05) are shown in Tables 3 and 4, excluding the ancestor terms that became significant mainly by an overrepresented descendant in the GO structure. A larger list of all overrepresented terms and pathways with a less stringent threshold (uncorrected P<0.01) is shown in Tables S1 and S2. In summarizing these results, the genes whose products are involved in DNA-binding were the most significantly overrepresented group among genes whose expression increases with age (Table 4). Genes with products involved in the regulation of DNA-dependent transcription and genes encoding proteins located in or a subcomponent of the nucleus were also significantly enriched. Among all genes negatively correlated with age, the most significantly enriched groups included genes involved in nervous system development, mitochondrial genes, and those whose products are constituents of the cytoplasm (Table 3). We also highlighted the degree of enrichment in terms of a fold change (the number of genes that give rise to this number are shown in Tables S1 and S2). Thus, in quantitative terms the most enriched genes occurred in the set of negatively correlated transcripts, where dendrite genes and genes associated with neuronal projections were enriched more than 3 fold (Table 3). In contrast, although highly significant in some cases, genes whose expression is increasing with age are enriched a maximum of 1.80 fold (Table 4). As a comparison, the probabilities for the enriched terms to be underrepresented by chance within the opposite group of genes are also shown in the last two columns of Tables 3 and 4. For example, from Table 4, we show that the term “nucleus” which is significantly enriched in the set of genes that are increasing with age is also in deficit in the set of genes that are decreasing with age.
We extended the above analytical scheme to a second sample consisting of 1240 individuals with lymphocyte mRNA measures for 19,648 transcripts representing 18,519 genes. We note that a rigorous standardization procedure had been applied to this set previously [15] and we thus elected not to perform any additional transformations, nor to include any additional covariates (see methods). Linear regression was performed as above on all 19,648 transcripts versus age for this sample. Given its size, and compared to the brain sample, there were many more transcripts that exceeded a strict Bonferroni significance threshold, and we chose to present only the top 50 (25 negatively and 25 positively correlated with age) which are shown in Tables 5 and 6, arbitrarily truncated from a total of 1080 (612 negative and 468 positive). The top 3 scoring negatively regulated genes and top 3 scoring positively regulated genes are shown in Figure 2. The most significant finding in this set was for the LRRN3 (encoding a membrane protein with unknown function), and this was also the gene upon which age exhibited the largest effect (28% variance explained). Interestingly, the second largest effect observed in this set was 14% variance explained, representing in our view an anomalous drop from the top ranking gene. As was done for the brain sample, we again divided this list into up- and down- regulated genes and performed term/pathway based analyses. An arbitrary significance level of 0.01 was chosen from the linear regression results to establish these 2 lists (this differed from the brain sample for which a threshold of 0.05 was set due to the number of genes exceeding significance in lymphocytes and a 2000 gene limit set by the DAVID web tool). The total set of 19,648 transcripts for which regression was performed, was reduced to 13,231 for which annotations could be obtained towards analysis using the DAVID web application. The results of this are shown in Tables 7 and 8 where Bonferroni corrected (P<0.05) enriched terms and pathways are documented. As was done for the brain sample, an expanded list of enriched terms and pathways according to relaxed uncorrected significance threshold of P<0.01 is shown in Tables S3 and S4. The main highlight in our view is the highly significant enrichment of mitochondrial genes in the negatively regulated group (Table 7), providing a replication of what was seen in the brain. In general, the magnitude of the fold-change in term/pathway enrichment was larger in the set of negatively regulated genes, also in agreement with what was seen in brain.
Based upon the identified lists of age-related genes for both brain and lymphocyte samples, we also explored for differences in basic genomic architecture. Our guide for this was a detailed analysis recently presented on characteristics of housekeeping genes [13]. The results of this analysis are presented in Tables 9 and 10, where we document the differences between the up and down-regulated sets in comparison with the characteristics of the non-regulated genes. In summarizing these results, while we found some striking differences, for example with coding sequence length in brain, we note that this was not replicated in lymphocytes. Instead we highlight a single category that does appear to be replicating, namely the ratio of average intron to average exon sequence length. We used second-order factorial ANOVA models to perform combined analyses for all variables. For this, the intron/exon sequence ratio is highly significant (F2,24158 = 41.57, P = 9.5×10−19) with no evidence of interaction between the two sets (P = 0.69). This can be taken in context with the same analysis for coding sequence length, where the tissue by group (group defined by up, down, or non-regulated genes) interaction term was highly significant (P<0.0001). For all other combined analyses, the main group effect was either non-significant or the interaction term was highly significant (P<0.0001).
Finally, we used these two samples to pursue the question of whether variance in gene expression itself changes with age. Our hypothesis was that variance might increase with age as a consequence of accumulating somatic mutation [11] and/or increasingly heterogeneous environmental exposures. For the brain sample, an F-test for equality of variances was conducted on all 14,078 transcripts by dividing the sample in two groups according to the median age. We observed only 10 genes that exceeded a strict multiple test correction threshold and in each case there was evidence that the distributions of these genes deviated from normality. There was no overlap in this set for the highest scoring candidates with genes found to change with age in the linear regression analyses. There were 6 cases where an age-related gene also had a borderline significant F-test (P<0.05). This proportion was not significantly different than what could be found in the entire set (not shown). For the lymphocyte set, we took a slightly different approach due to its size and rather than dividing the sample at the median chose to examine decile bins and applied a Levene's robust variance test to explore for differences across age groups. There was again no strong evidence that variance differed across these groupings, the most significant finding being for the NLP gene (P = 2.3×10−6), and with only 2 genes in total attaining significance after multiple test correction. As a final closing note, we enumerated the number of statistical tests used in this study in its entirety, arriving at an approximate number of 110,000. This may be used as a reference for any of the un-corrected P-values that are presented.
Discussion
We consider the most important finding in this study to be that expression levels of genes involved in mitochondrial processes are decreasing with age, as originally proposed by Zahn et al. and also supported by Miller et al. [6], [16]. In the former study, a key observation was that this is apparent in multiple tissues as well as in species other than humans. The present analysis provides a confirmation of this as a common characteristic of aging, in that this evident in both brain and lymphocytes. For these two samples there are also tissue specific themes that have emerged that we consider a validation of the quality of the expression phenotypes originally obtained. In the brain sample, consistent with previous results [11], a decrease in genes involved in synaptic function was observed, which follows from the documented changes in synaptic function that occur with age [8]. In lymphocytes, there was evidence that genes involved in the immune response increase expression with age, which might be regarded as a reflection of chronic persisting viruses such as cytomegalovirus (CMV) [17]. While we acknowledge central differences in these samples that include ethnicity, the age ranges, sample size, and the cell types represented, in this discussion we highlight these and other pathways with a comparison of the two different tissues as a guide.
A key distinction in contrasting the study by Zahn et al. was the specific focus on genes of the mitochondrial respiratory chain [6]. While many of these are included in our emergent lists of genes of the mitochondrion (see Tables S1 and S3), results indicate that the effects of age on mitochondrial function may be broader. For example, in the brain the most significant finding across all pathways was for cytoplasm genes (Table 3). Although this categorization includes all genes of the mitochondrion, and indeed our results suggest it is the latter grouping that contributes to the strength of the statistic, there remain additional genes that may be enriched in this set (see Table S1). Oxidative phosphorylation itself did emerge as a highly significant term associated with aging in the brain sample, but not in lymphocytes, indicating that there are likely to be tissue specific differences in the component genes related to mitochondrial function (Tables 3 and 7). Nonetheless, these data together with previous studies provide an intriguing foundation to investigate if this is a primary event in the aging process or derivative of a general decrease in energy metabolism and activity that accompanies old age (e.g. [18]).
That transcript levels are both decreasing and increasing is an important indication that the aging process does not lead to a unidirectional decline in expression. While this was evident in both brain and lymphocytes, in each case the relative fold-enrichment of gene categories was higher among negatively correlated transcripts. Thus, the up-regulated categories consisted of more unique transcripts in general than those that were down-regulated. In brain we noted several pathways that appear to have component genes that increase in expression with advancing age, the foremost among these encompassing genes that encode nuclear proteins. The most significant categorizations that emerged included a large contingent of genes related to DNA-binding and transcription, the most abundant being zinc-finger proteins. Although genes related to transcription were highlighted previously, results suggested a mix of both positive and negative regulation [11]. In contrast however, Miller et al. did observe an overlap of genes related to transcription that were increasing in relation to both ageing and to Alzheimer's disease (AD) [16].
In lymphocytes, immune response was the most significant category associated with increasing expression. However, there are several issues that are impossible to resolve for this latter result, especially regarding which cell types are represented in this sample (e.g. [19]). For example, this might relate to an accumulation of highly differentiated T cells due to acceleration by persistent CMV infection but this would require extensive further study [17]. There were also some interesting highlights among the many individual genes that were up-regulated with age in lymphocytes. In particular the most significant up-regulated gene IGFBP3 is intriguing in the context that another member of the insulin-like growth factor binding protein family, IGFBP7 has recently been implicated as an inducer of apoptosis in human melanoma cell lines [20], [21]. IGFBP3 itself has also been shown to be increased in senescent human fibroblasts [22]. Those findings together with results of the present study provide support for the role of insulin-like growth factor signaling in cellular senescence.
For the brain sample, we were intrigued by the emergent pattern that expression of cytoplasm genes may be decreasing with age, while expression of nuclear genes may be increasing with age. While this might reflect the large number of mitochondrial and transcription genes in these particular sets, we cannot ignore the possibility that a basic change in cell morphology might be at play. One explanation may be that neuronal number remains stable with age, but synaptic vesicle density decreases [7], [8]. This relative change might give rise to our observations here. In support of this, genes related to neuronal projections (which includes dendrite genes) and synapse genes are major categories that appear to decrease with age (see Table 3). Also in support of this, the highest ranking individual gene, SVOP, encoding the SV2 related protein, is localized to the synaptic vesicle and appears to be an ion transporter. We had anticipated that more neuronal specific genes would exhibit a decreasing pattern, but noted that even some common neuronal reference genes, such as ENO2 (ranking 7597th), were not associated with age. Thus, in contrast to the enrichment of neuronal projection genes mentioned above (and see Table 3), this might support the concept that a decrease in neuronal number is not a major feature of aging [23]. A potentially important source of confounding in analyses of post-mortem brain samples is the mode of death, which we have not examined in more detail in this study. In particular, agonal state and pH have been highlighted as contributors to brain mRNA expression [24]–[26]. The present study has strong similarities with the study by Li et al., (2004)[25]. However, the relationship between pH and age was previously explored, but there was no evidence of a significant correlation [24], [27]. Interestingly, although it was a small sample Vawter et al. also noted significant correlations of age with mitochondrial genes regardless of agonal state (see specifically supplementary table 4 from [24]). Nonetheless, one explanation might be that transcription factor activation and mitochondrial deactivation is a natural response to hypoxic stress which may be more common in elderly individuals (e.g. [25], [28]). Another possibility might be that both aging and extended hypoxia dependent on the mode of death share similarities in terms of gene regulation.
For the lymphocyte sample, the most significantly enriched terms were found in the class of negatively regulated genes, where the ribosomal and translation machinery appears to be strongly affected by age. This could potentially contribute to “immunosenescence”, whereby the immune system decays with time, rendering elderly individuals more susceptible to infectious diseases [29]. While changes at the expression level for this class of genes have to our knowledge not been previously reported on this scale, the concept that protein synthesis is involved in immunosenescence is not new [30]. This result is in our view the most striking difference between lymphocytes and brain. Although this suggests a more general decreases in metabolism in aging lymphocytes, we consider this an important reminder that tissue specificity can play a major role in gene expression profiling [31].
Another important feature of this study is the relative degree of stability in gene expression across the age spectrum. To put this in context, the strongest effects of age on any genes were 18% variance explained for TAC3 in brain and 29% for LRRN3 in lymphocytes, but in examining Tables 1, 2, 5 and 6, it is apparent that effect sizes drop off rapidly. For each of these top-ranked genes there was an approximately 2-fold change in average expression levels over the age ranges. This is comparable to well documented changes in hormone levels with age, such as for IGF1 [2], [32]. Also of note, genes claimed to be changing with age previously exhibited differences between young and older groups in the range of 2–3 fold [11]. This again might suggest that cell-loss and/or changes in organ function/morphology are more important than broad changes in gene expression in the decrease in health that typically accompanies increasing age. The possibility however cannot be excluded that relatively small changes in gene expression have a large impact on cell function. As an example, disorders such as Parkinson's disease (PD) can be caused by simple gene dosage effects, as has now been shown for the SNCA gene which encodes alpha-synuclein [33]. Although rare, we noted several genes that may be changing to a larger degree, but which fell below a strict significance threshold, for example ZIC3 in brain. However, there were no similar cases in the much more powerful lymphocyte sample.
Finally, we also explored basic gene architecture characteristics in an attempt to provide some insight into why expression might vary with age. The only parameter that emerged as significant and equivalent in both brain and lymphocytes was the intron/exon length ratio, this representing a metric of gene “compactness” [13]. This might be interpreted as suggesting that non-compact genes are more susceptible to mutations that disrupt regulation, and thus lead to decreasing expression with advancing age. However, we cannot exclude the possibility that what we are seeing is simply a result of the gene pathways that are over-represented, these being enriched for compact or non-compact genes. In other words, if specific pathways are affected by age, the genes that represent those pathways may have similar features. We consider this nonetheless an intriguing finding in that it is replicating in different tissues.
In summary, we validate previous findings that a decrease in mitochondrial gene expression appears to be a common theme in the aging process. Whether this is a primary event that causes a decline in health with advancing age or a result of a general decrease in metabolism in the elderly remains a topic for further investigation. We also highlight additional novel pathways that may be age dependent but with dramatic differences between tissues, in particular with genes related to transcription and to translation. These results may provide a valuable foundation for understanding the molecular consequences of aging and emphasize the development of catalogues of senescence-related genes in additional tissues.
Materials and Methods
Human Samples
The primary sample used in this study consisted of 191 individuals with ages-at-death data ranging from 65–100 years for which brain autopsy tissue was obtained. Expression level data were obtained using the Illumina HumanRefSeq-8 expression BeadChip platform for a total of 14,078 transcripts in which expression was detected in 5 or more individuals. Detailed descriptions of the human samples as well as the expression profiling protocol are provided in the original publication for which this sample was presented [12]. To create a working dataset, expression phenotypes and covariate data were merged that included age-at-death, post-mortem interval (pmi), gender, brain region, and transcript detection rates. From this, we generated 2 additional covariates for 1) total average expression level from all transcripts and 2) total average expression level for all transcripts that were detected in all individuals (5269 transcripts fulfilled the latter criteria). All individual transcript levels, the global average transcript levels, and pmi were log10 transformed prior to inclusion in analyses.
The second sample consisted of 1240 individuals ranging in age from 15-94 in which fresh blood lymphocytes had been obtained and mRNA extracted. Expression phenotyping was conducted using the Illumina Sentrix Human Whole Genome (WG-6) Series I BeadChip platform. In total there were 19,648 individual transcript measures, representing approximately 18,500 genes. Details of the sample and expression protocols have been provided previously in the original publication describing this sample [15]. For our analysis, we used the normalized expression phenotypes without further transformation.
Correlation of Age and Transcript Level
We classified the extremes as outliers that are expected to be observed once or more in 1930 individual log transformed transcript estimates with assumption of independence and normal distribution of 193 measures for each transcript.where α′ is the alpha level for each measure, α is 0.05 and k 193. Based on normal distribution assumption,
where Φ is standard normal cumulative distribution fuction, z is z-score of each measure.
The normality of log transformed observed values for each transcript was tested by means of a Shapiro-Wilk W test. To assess differences of transcript detection rates or global expression levels across brain regions, as well as to test differences in characteristics of genomic architecture between age-related gene sets, ANOVA was used. Contingency tables (for example comparing the proportions of housekeeping genes) were evaluated by means of a standard chi-square test. All statistical analyses not related to pathway based tests were performed using STATA se 9.0. For the brain sample, the dependency of expression level on age-at-death was tested by fitting a linear regression model for each transcript:where Y is the log10 transformed expression level of the transcript, x1 age-at-death, x2 log10 of global expression, β0 intercept, β1 and β2 slope of each independent variable, and ε random error. For the lymphocyte sample, the same test was applied but without the global expression covariate according to the following:
where Y is the normalized expression level of each transcript, x1 age, β0 intercept, β1 slope of age, and ε random error. We also allowed for non-independence of expression measures given the relatedness of family members in the lymphocyte sample [34].
Gene Ontology and KEGG Pathway Analyses
The Entrez Gene IDs, symbols, and descriptions of genes for all tested transcripts were attained by Entrez Programming Utilities (eUtils) using GI number with the aid of a Perl script. This facilitated a search for replaced sequence identifiers and extracted information of interest by scanning the output text files from Entrez eUtils on the transcripts. Using the attained Entrez Gene IDs as identifiers of genes, we obtained the total lists of both Gene Ontology terms and KEGG pathway descriptors, with which our selected set of genes was annotated and analyzed using the DAVID functional annotation tool [14] with the most recent data update (January, 2008). The EASE scores [35] from the DAVID tool were used in trimming to over-represented term lists, with Bonferroni corrections of the scores from the tool as an inclusion criteria of Tables 3, 4, 7 and 8. We scrutinized ancestor-descendant relationships in the Gene Ontology structure among the terms in the enrichment lists on the basis of is_a, part_of, and regulates relationship by scanning the master ontology file which was updated in April, 2008. Our goal with this was to determine if ancestor terms had emerged as significant primarily because of enrichment of one of their descendant terms. We detail our strategy for this in Figure S1. Among the GO terms in the list, those which had no descendants were labeled as ‘end-terms’. For every term excluding end-terms, we created the artificial descendants. These were intended to represent the complement of the set of genes annotated with a descendant term in the list with respect to all genes in the ancestor. The artificial term for every descendant of every ancestor in the list was then tested to determine if it was over-represented by applying a hypergeometric test [36]. The ancestors for which all artificial descendant terms were over-represented at P<0.005 were labeled ‘significant ancestor terms’. The terms whose artificial descendants for end-terms or significant ancestors were over-represented were added to the set of ‘significant ancestor terms’. After modifying the set, the terms in the set were checked if they fulfilled inclusion criteria (enriched complement at P<0.005) and the set was updated iteratively until there was no change. The end-terms and significant ancestor terms are listed in Tables S1, S2, S3, S4. The remaining terms in the enrichment list follow in the same tables that are significantly over-represented mainly by a highly enriched descendant. P-values of the enrichment test for the artificial terms are also shown in the Tables S1, S2, S3, S4.
Genomic Architecture of Age-related Gene Sets
The genomic positions of start and end points of transcripts, their coding sequences, and their exons and introns in both brain and lymphocyte gene sets were collected from the UCSC genome browser using Genbank accession numbers (e.g. NM_018711) as identifiers. Information on some of the transcripts that were detected in both samples was not available from the UCSC browser due to the records being suppressed. Therefore, the number of transcripts in the genomic architecture analysis was reduced to 12,520 and 12,242 for brain and lymphocyte samples, respectively. Based on the positional data, lengths of pre-mRNA sequences, coding sequence, total intron, 5′UTR, 3′UTR, average intron and average exon of each transcript, as well as the number of exons and the ratio of intron per coding sequence length were calculated and transformed on a log10 scale. For comparative analyses of gene characteristics, the transcripts were divided into 3 groups, up-regulated genes (positively correlated with age at a P<0.05 threshold), down-regulated (negatively correlated with age at a P<0.05 threshold), and a set of genes that were not significantly altered over the age spectrum. Differences between groups were assessed using ANOVA. Combined analyses were performed on both sets using second order factorial ANOVA with sample source as a covariate.
URLs
Entrez Gene: http://www.ncbi.nlm.nih.gov/sites/entrezdbgene; Entrez eUtils: http://eutils.ncbi.nlm.nih.gov/entrez/eutils; DAVID 2008: http://david.abcc.ncifcrf.gov/; Gene Ontology (GO): http://www.geneontology.org/; UCSC genome browser: http://genome.ucsc.edu/
Supporting Information
Figure S1.
Overrepresented GO term analysis
https://doi.org/10.1371/journal.pone.0003024.s001
(0.06 MB PDF)
Table S1.
Terms in the Gene Ontology and KEGG pathway databases enriched among genes that decreased expression with advancing age in brain (1450 genes in total 13,216)
https://doi.org/10.1371/journal.pone.0003024.s002
(0.15 MB PDF)
Table S2.
Terms in the Gene Ontology and KEGG pathway databases enriched among genes that increased expression with advancing age in brain (1943 genes in total 13,216)
https://doi.org/10.1371/journal.pone.0003024.s003
(0.09 MB PDF)
Table S3.
Terms in the Gene Ontology and KEGG pathway databases enriched among genes that decreased expression with advancing age in the human lymphocytes (1878 genes in total 13,232)
https://doi.org/10.1371/journal.pone.0003024.s004
(0.13 MB PDF)
Table S4.
Terms in the Gene Ontology and KEGG pathway databases enriched among genes that increased expression with advancing age in the human lymphocytes (1430 genes in total 13,232)
https://doi.org/10.1371/journal.pone.0003024.s005
(0.12 MB PDF)
References
- 1. Partridge L, Gems D (2002) Mechanisms of ageing: public or private? Nat Rev Genet 3: 165–175.
- 2. Lamberts SW, van den Beld AW, van der Lely AJ (1997) The endocrinology of aging. Science 278: 419–424.
- 3. Hughes KA, Alipaz JA, Drnevich JM, Reynolds RM (2002) A test of evolutionary theories of aging. Proc Natl Acad Sci U S A 99: 14286–14291.
- 4. Charmantier A, Perrins C, McCleery RH, Sheldon BC (2006) Quantitative genetics of age at reproduction in wild swans: support for antagonistic pleiotropy models of senescence. Proc Natl Acad Sci U S A 103: 6587–6592.
- 5. Rodwell GE, Sonu R, Zahn JM, Lund J, Wilhelmy J, et al. (2004) A transcriptional profile of aging in the human kidney. PLoS Biol 2: e427.
- 6. Zahn JM, Sonu R, Vogel H, Crane E, Mazan-Mamczarz K, et al. (2006) Transcriptional profiling of aging in human muscle reveals a common aging signature. PLoS Genet 2: e115.
- 7. Terry RD, DeTeresa R, Hansen LA (1987) Neocortical cell counts in normal human adult aging. Ann Neurol 21: 530–539.
- 8. Masliah E, Mallory M, Hansen L, DeTeresa R, Terry RD (1993) Quantitative synaptic alterations in the human neocortex during normal aging. Neurology 43: 192–197.
- 9. Kluger A, Gianutsos JG, Golomb J, Ferris SH, George AE, et al. (1997) Patterns of motor impairement in normal aging, mild cognitive decline, and early Alzheimer's disease. J Gerontol B Psychol Sci Soc Sci 52: P28–39.
- 10. Ferri CP, Prince M, Brayne C, Brodaty H, Fratiglioni L, et al. (2005) Global prevalence of dementia: a Delphi consensus study. Lancet 366: 2112–2117.
- 11. Lu T, Pan Y, Kao S-Y, Li C, Kohane I, et al. (2004) Gene regulation and DNA damage in the ageing human brain. Nature 429: 883–891.
- 12. Myers AJ, Gibbs JR, Webster JA, Rohrer K, Zhao A, et al. (2007) A survey of genetic human cortical gene expression. Nat Genet 39: 1494–1499.
- 13. Eisenberg E, Levanon EY (2003) Human housekeeping genes are compact. Trends Genet 19: 362–365.
- 14. Dennis G Jr, Sherman BT, Hosack DA, Yang J, Gao W, et al. (2003) DAVID: Database for Annotation, Visualization, and Integrated Discovery. Genome Biol 4: P3.
- 15. Göring HH, Curran JE, Johnson MP, Dyer TD, Charlesworth J, et al. (2007) Discovery of expression QTLs using large-scale transcriptional profiling in human lymphocytes. Nat Genet 39: 1208–1216.
- 16. Miller JA, Oldham MC, Geschwind DH (2008) A systems level analysis of transcriptional changes in Alzheimer's disease and normal aging. J Neurosci 28: 1410–1420.
- 17. Akbar AN, Fletcher JM (2005) Memory T cell homeostasis and senescence during aging. Curr Opin Immunol 17: 480–485.
- 18. Tumeh PC, Alavi A, Houseni M, Greenfield A, Chryssikos T, et al. (2007) Structural and functional imaging correlates for age-related changes in the brain. Semin Nucl Med 37: 69–87.
- 19. Czesnikiewicz-Guzik M, Lee WW, Cui D, Hiruma Y, Lamar DL, et al. (2008) T cell subset-specific susceptibility to aging. Clin Immunol 127: 107–118.
- 20. Wajapeyee N, Serra RW, Zhu X, Mahalingam M, Green MR (2008) Oncogenic BRAF induces senescence and apoptosis through pathways mediated by the secreted protein IGFBP7. Cell 132: 363–374.
- 21. Cichowski K, Hahn WC (2008) Unexpected pieces to the senescence puzzle. Cell 133: 958–961.
- 22. Goldstein S, Moerman EJ, Jones RA, Baxter RC (1991) Insulin-like growth factor binding protein 3 accumulates to high levels in culture medium of senescent and quiescent human fibroblasts. Proc Natl Acad Sci U S A 88: 9680–9684.
- 23. Terry RD (2006) Alzheimer's disease and the aging brain. J Geriatr Psychiatry Neurol 19: 125–128.
- 24. Vawter MP, Tomita H, Meng F, Bolstad B, Li J, et al. (2006) Mitochondrial-related gene expression changes are sensitive to agonal-pH state: implications for brain disorders. Mol Psychiatry 11: 615, 663–679.
- 25. Li JZ, Vawter MP, Walsh DM, Tomita H, Evans SJ, et al. (2004) Systematic changes in gene expression in postmortem human brains associated with tissue pH and terminal medical conditions. Hum Mol Genet 13: 609–616.
- 26. Obara M, Szeliga M, Albrecht J (2008) Regulation of pH in the mammalian central nervous system under normal and pathological conditions: facts and hypotheses. Neurochem Int 52: 905–919.
- 27. Preece P, Cairns NJ (2003) Quantifying mRNA in postmortem human brain: influence of gender, age at death, postmortem interval, brain pH, agonal state and inter-lobe mRNA variance. Brain Res Mol Brain Res 118: 60–71.
- 28. Rocha S (2007) Gene regulation under low oxygen: holding your breath for transcription. Trends Biochem Sci 32: 389–397.
- 29. Larbi A, Franceschi C, Mazzatti D, Solana R, Wikby A, et al. (2008) Aging of the immune system as a prognostic factor for human longevity. Physiology (Bethesda) 23: 64–74.
- 30. Tollefsbol TO, Cohen HJ (1986) Expression of intracellular biochemical defects of lymphocytes in aging: proposal of a general aging mechanism which is not cell-specific. Exp Gerontol 21: 129–148.
- 31. Yanai I, Korbel JO, Boue S, McWeeney SK, Bork P, et al. (2006) Similar gene expression profiles do not imply similar tissue functions. Trends Genet 22: 132–138.
- 32. Corpas E, Harman SM, Blackman MR (1993) Human growth hormone and human aging. Endocr Rev 14: 20–39.
- 33. Eriksen JL, Przedborski S, Petrucelli L (2005) Gene dosage and pathogenesis of Parkinson's disease. Trends Mol Med 11: 91–96.
- 34. Curran JE, Johnson MP, Dyer TD, Göring HH, Kent JW, et al. (2007) Genetic determinants of mitochondrial content. Hum Mol Genet 16: 1504–1514.
- 35. Hosack DA, Dennis G Jr, Sherman BT, Lane HC, Lempicki RA (2003) Identifying biological themes within lists of genes with EASE. Genome Biol 4: R70.
- 36. Draghici S, Khatri P, Martins RP, Ostermeier GC, Krawetz SA (2003) Global functional profiling of gene expression. Genomics 81: 98–104.