Introduction

In the absence of any neurodegenerative disease, the aging process of the human brain is inevitably and quintessentially characterized by memory and learning impairments. Unlike in the case of a neurodegenerative disease, normal aging has not been associated with neuronal loss1,2,3. Rather, it has been observed that the impairments induced by normal aging are associated with synaptic remodeling and that they are more likely to affect functions that are associated with the hippocampus, i.e. several areas of memory and learning1,2,4.

In order to study the process of human normal aging, this study focuses on the most vulnerable target of that process, namely, the hippocampus. Given the long, gradual course of the normal aging process, I arbitrarily defined the boundaries of the two groups as follows: 1) Old subjects (O) with age ≥ 60 years and 2) Young subjects (Y) with age ≤ 45 years. This 15-yr age gap, I theorized, would accentuate the contrast between the two groups in connection with this otherwise continuous and overlapping process. I analyzed the global gene expression data from post-mortem hippocampal tissue (harvested from the body of the hippocampus at the level of the lateral geniculate nucleus) of 25 old and 15 young cognitively intact human subjects, posted at the Gene Expression Omnibus (GSE11882)5. Demographical information pertaining to all 40 subjects is shown in Supplementary Table 1.

Results

Having employed three different and independent methods of statistical significance, namely, ROC curve analysis, fold change and P-value, I was able to identify 36 genes that were the most significant in terms of differential expression. Fig. 1b depicts the results of K-Means clustering analysis based on the expression of the top 36 most significant genes. All K-Means clustering analysis results (with respect to both the housekeeping genes and the 36 most significant genes) are shown in Supplementary Table 2. As can be seen in both Fig. 1b and Supplementary Table 2, there is a clear separation of the two groups. Fig. 2 depicts the heat map that resulted by plotting the expression of those 36 genes for all 40 subjects (15 young and 25 old). As can be seen by the relative intensities, all of the 36 most significant genes are over-expressed (red color) in the case of the old subjects as compared with the case of the young subjects (blue color). The direction of the differential expression of those 36 genes also appears in Table 1. Moreover, Fig. 3 provides a 3D representation of the differential expression of those 36 genes between the two groups in a surface-contour plot.

Table 1 Top 36 most significantly differentially expressed genes
Full size table
Figure 1
figure 1

Results of K-Means clustering analysis.

K-Means clustering of the hippocampal tissue gene expression of 25 old subjects (O) and 15 young subjects (Y) with respect to the house-keeping genes (a) and with respect to the 36 most significant genes (b). In (a), in connection with the house-keeping genes, the two groups are inseparable and indistinguishable; whereas in (b), in connection with the 36 most significant genes, the two groups are separated and are clearly distinguishable. D1 is subject distance from the centroid of cluster 1 and D2 is subject distance from the centroid of cluster 2.

Full size image
Figure 2
figure 2

Heat map of the hippocampal gene expression of all 40 subjects.

Heat map of the hippocampal tissue gene expression of 15 young subjects (columns # 1–15) and 25 old subjects (columns # 16–40) with respect to the 36 most significant genes (rows # 1–36). The intensity scale of the standardized expression values ranges from −2.5 (blue: low expression) to +2.5 (red: high expression), with 0 (white) representing the reference intensity value. As can be seen and based on the group mean expression values, all 36 most significant genes are over-expressed in the case of the old subjects as compared with the young subjects.

Full size image
Figure 3
figure 3

Surface-contour plot of the hippocampal gene expression of all 40 subjects.

Hippocampal tissue gene expression of 15 young subjects (columns # 1–15) (x-axis) and 25 old subjects (columns # 16–40) (x-axis) with respect to the 36 most significant genes (rows # 1–36) (y-axis). The intensity scale of the standardized expression values ranges from −4 (blue: low expression) to +4 (red: high expression) (z-axis). As can be seen and based on the group mean expression values, all 36 most significant genes are over-expressed in the case of the old subjects as compared with the young subjects.

Full size image

Mathematical modeling of aging

Given the aforementioned biovariability in connection with hippocampal gene expression, I wanted to explore whether, via mathematical modeling, I could generate a function that could classify the 40 subjects with a high accuracy. Such a model would be valuable in future studies of global gene expression analysis of post-mortem hippocampal tissue investigating biological and chronological aging. To that end, I randomly selected approximately 70% of the subjects [11/15 young subjects and 18/25 old subjects] for the development of the function (henceforward referred to as super variable) and I used the remaining subjects (4 young and 7 old) solely for the purpose of validating the super variable. Employing a general methodology that I have previously introduced6,7, I was able to generate a super variable (function) that, based on the input of 7 genes from the 36 most significant genes, was able to identify/classify accurately all but one of the old subjects {subject # 33 [64 yrs (F)]} [sensitivity = (24/25) = 0.96] and all of the young subjects [specificity = (15/15) = 1.00]. Those overall results of the performance of the F1 super variable were obtained by combining the results from the development and the validation phases. According to the rank that appears in Table 1, the seven genes that provide the input to the F1 super variable are: C4A (C4B), ADORA3, MS4A7, BCL6, CD44, C3AR1 and HLA-DRB1. All of those seven genes are, in terms of biological function, either genes of inflammation or genes of immune system activation. Supplementary Fig. 1 shows the F1 super variable function in relation to its 7 input gene variables. Fig. 4 and Supplementary Table 3 show the overall results of the F1 super variable, i.e. the F1 scores of all 40 subjects used in this study, as well as their respective classification. Fig. 4 and Supplementary Table 3 were created by combining the results from the development phase (the F1 scores of all 29 subjects that were randomly selected and used exclusively for the development of the model) with the results from the validation phase (the F1 scores of all 11 subjects that were randomly selected and used exclusively for testing purposes). The results of the F1 super variable in the development phase are shown in Supplementary Fig. 2 and Supplementary Table 4, whereas the results in the validation phase are shown in Supplementary Fig. 3 and Supplementary Table 5.

Figure 4
figure 4

Overall results of the F1 super variable (function).

The F1 uses 7 of the 36 most significant genes as its input variables. Using the expression value of those 7 genes for a particular subject, the F1 yields the F1 score of that subject; and, based on the determined cut-off score of 53.450, the F1 classifies that subject as young if the F1 score is < 53.450 or as old if the F1 score is ≥ 53.450. As can be seen by the overall performance, the F1 classified correctly all subjects except one old one [sensitivity = (24/25) = 0.96 and specificity = (15/15) = 1.00]. The mean F1 score of the Y subjects was 45.028 (top of the blue bar) and their standard deviation (whiskers above or below the top of the blue bar) was 4.721. The mean F1 score of the O subjects was 64.212 (top of the red bar) and their standard deviation (whiskers above or below the top of the red bar) was 6.514. The significance level was set at α = 0.001 (two-tailed) and the probability of significance for the F1 was P = 4.18 × 10−12 (independent t-Test with T-value = 9.927). The F1 is parametrically distributed with respect to both groups. The F1 scores of all 40 subjects are shown in Supplementary Table 3.

Full size image

It is interesting to note here that, assessing and comparing the performance of the F1 super variable (Supplementary Table 3) with that of the supervised K-Means clustering (Supplementary Table 2), one can see that the former yielded one misclassification as opposed to four yielded by the latter.

Finally, it should also be noted here that, owing to the constraints of this study, namely, the paucity of healthy, normal human brain tissue samples and respective available data, the F1 super variable needs to be further validated with a larger, independent cohort.

Biovariability of aging

It has long been observed empirically that aging is not a steady-state, uniformly continuous process; that it is characterized by a relatively wide biovariability; and that biological age may not necessarily coincide with chronological age. The results of my study corroborate those observations. Looking at the expression of the 36 most significant hippocampal genes of all 40 subjects [15 young (columns 1–15) and 25 old (columns 16–40)] in Fig. 2, one notices that four old subjects {# 22, 27, 33 and 35 [80 yrs (M), 83 yrs (M), 64 yrs (F), & 86 yrs (M), respectively]} displayed gene expression patterns that were distinctly closer to those of the young subjects than the patterns of the rest of the old subjects. Conversely, the same observation, albeit in the opposite direction, can be made for one of the young subjects {# 2 [45 yrs (F)]}. The results of K-Means clustering analysis supported numerically those observations (Supplementary Table 2). Moreover, the aforementioned observations about the biovariability of the aging process were also supported by the results of hierarchical clustering analysis performed on the F1 scores of all 40 subjects (Supplementary Fig. 4).

Inflammation and immune system activation in aging

Remarkably, of the 30 known genes out of the 36 most significant genes, 27 were – in terms of function – either genes of inflammation or genes of immune system activation (Table 1). This suggests that – to a large extent and insofar as it pertains to the hippocampal area of the brain – the dual process of a chronic inflammation and the elicited chronic immune-system response and activity can differentiate between old and young brains with a high accuracy. This is further supported by the fact that the aforementioned seven genes employed by the F1 super variable, all of which are genes of inflammation or genes of immune system activation, can discriminate between old and young brains with almost a perfect accuracy [sensitivity = (24/25) = 0.96 and specificity = (15/15) = 1.00].

Discussion

The seven genes [C4A (C4B), ADORA3, MS4A7, BCL6, CD44, C3AR1 and HLA-DRB1], which are the constituent input variables of the model (F1 super variable) and all of which are – in terms of function – inflammation or immune system activation genes (Table 1), were all found to be over-expressed in the old subjects compared with the young subjects (Table 1). C4A (C4B) has been observed to be over-expressed in patients with Huntington disease8 and Alzheimer disease9, in mice with rheumatoid arthritis10, etc. ADORA3 has been found to be over-expressed in the hippocampus of patients with Parkinson disease11, in patients with astrocytomas12, etc. MS4A7 has been observed to be over-expressed in mice with rheumatoid arthritis10. BCL6 has been observed to be over-expressed in patients with Huntington disease13, with ischemic stroke14, with rheumatoid arthritis15, with B-cell lymphoma16, etc. CD44 has been found to be over-expressed in patients with systemic lupus erythematosus17, with immune thrombocytopenia18, with schwannomas19, with Huntington disease8 and numerous other diseases and conditions. Over-expression of C3AR1 has been observed in patients with severe acute respiratory syndrome20, with asthma21, etc., while over-expression of HLA-DRB1 has been observed in patients with multiple sclerosis22, with rheumatoid arthritis23, with Duchenne muscular dystrophy24, etc.

Previous studies using animal models have observed associations between aging and inflammation in connection with the hippocampus, the neocortex and the cerebellum25,26. Using animal models or human subjects with early-stage neurodegenerative diseases, such as Alzheimer, other studies have observed a link between neuroinflammation and deficits in synaptic plasticity, especially long-term potentiation (LTP) in the hippocampus, which is associated with long-term memory consolidation27,28,29,30,31,32,2. The fact that definitive causality cannot be established here notwithstanding – in other words, whether it is the normal aging process that induces inflammation/immune-system-overactivity, or whether the vice versa occurs, or whether another, hitherto unspecified, process engenders the normal aging process, which in turn induces inflammation/immune-system-overactivity, or whether that unspecified process engenders inflammation/immune-system-overactivity, which in turn induces the normal aging process – the results of my study support a direct causal link between the normal aging process and the process of inflammation/immune-system-overactivity. When considered collectively, therefore, the results of my study and all of the above observations from the other aforementioned studies point to a plausible theory on the normal aging process. At some point in time, chronic, low-level inflammation establishes itself and elicits a corresponding chronic immune response and activity. These two conjugate processes ultimately are responsible for a gradual loss of synaptic plasticity, particularly LTP in the hippocampus, accompanied with a minimal neuronal loss33,34,1,2,3. It is this loss of synaptic plasticity – at least in the hippocampus part of the brain – that is associated with the phenotypical changes of normal aging.

The results of my study, in addition to providing evidence for this dual process of chronic, low-level neuroinflammation/immune-system-activation in connection with normal aging, suggest a means of a potential treatment. Regardless of the exact causal sequence of the events, administration of anti-inflammatory drugs/chemicals that can normalize the expression of the aforementioned 27 genes of inflammation/immune-system-activation may decelerate the onset of the aging process, as well as the aging process itself and mitigate its symptoms by restoring synaptic plasticity throughout the hippocampus and possibly throughout the rest of the brain. Supplementary Table 6 lists all those 27 most significant genes as possible targets for the development of such an anti-inflammatory treatment, along with potential candidate drugs/chemicals that are known (via Ingenuity Pathway Analysis) to interact with those genes.

It is worth noting here that various anti-inflammatory drugs have been used in an effort to slow down the progression of neurodegenerative diseases, such as Alzheimer, with various degrees of success35,36. The magnitude of the neuroinflammatory processes in the case of Alzheimer disease or other neurodegenerative diseases, however, cannot be compared to that of the neuroinflammation in the normal aging process; and by virtue of the same argument, the task of halting neuronal cell loss during the course of a neurodegenerative disease cannot be compared to that of restoring synaptic plasticity during the course of the normal aging process. It would stand to reason, therefore, that anti-inflammatory treatment strategies may be more successful and efficacious than those employed against the progression of neurodegenerative diseases. Finally, I should point out that, based on increased evidence over the last twenty years or so, neuroinflammation seems to be the common denominator of normal aging and neurodegenerative diseases, such as Alzheimer. Understanding the causal circumstances under which a chronic, low-level neuroinflammatory process can transition to a major neuroinflammation conducive to neuronal degeneration and death would be of paramount importance. According to the latest evidence37, even a single substitution mutation on a single inflammation gene might suffice to trigger that transition in a small percentage of the population.

Methods

Data acquisition

I downloaded the raw intensity microarray data (CEL files) of 25 old subjects (age ≥ 60 years) and 15 young subjects (age ≤ 45 years) posted at the Gene Expression Omnibus (accession number: GSE11882)5.

Data processing

I processed the original raw intensity data (CEL files) using the Expression Console software by Affymetrix with the library for the HG-U133 Plus 2.0 microarray chip and choosing the RMA algorithm (510 (k) FDA approved) with the standard settings.

Statistical methods

I first assessed the quality of the data by examining the expression of all housekeeping genes by all 40 subjects (25 old and 15 young). With regard to the housekeeping genes, there was no statistically significant differential expression between the two groups (Supplementary Table 7). As can be seen by the results of K-Means clustering analysis in Fig. 1a, the two groups cannot be discriminated based on the expression of the housekeeping genes. Having, thus, established the quality of the data, I investigated for any differential expression among all gene variables using three different and independent methods. 1) Using a methodology that I have developed and introduced previously38,39,40,41,42, I performed ROC curve analysis on all gene variables in order to assess their discriminating power with respect to the two groups (old vs. young) and with respect to this method, I set statistical significance at ROC AUC ≥ 0.920. 2) For a given gene variable, I defined fold change (FC) as the mean expression value of the old subjects (O) over the mean expression value of the young subjects (Y) and I set statistical significance at FC ≥ 1.10 (if O > Y) or FC ≤ 0.91 (if O < Y). 3) I performed independent t-Tests, Mann-Whitney U tests, or Aspin-Welch unequal-variance tests (depending on how a particular gene variable was distributed) and having accounted for all possible comparisons (there are 54,675 probe sets in the Affymetrix HG-U133 Plus 2.0 chip), I set the significance level at α = 9.15 × 10−7. The results according to those three different and independent methods and prior to the final selection, are shown in Supplementary Table 8. In order to minimize the number of false negatives in the case of the third method43,44, for the final selection of significant variables, I imposed the condition that if a given gene variable met the significance criteria of all three methods, or those of the first method and those of only one of the other two methods, it would be deemed significant. Excluding multiplicities (different transcripts that corresponded to the same genes), thirty six genes made up the final list of the most significantly differentially expressed genes between the two groups, as assessed by the aforementioned three different and independent methods of statistical significance (Table 1).

In greater detail, to assess statistical significance, I used to assess statistical significance, I used the following three different and independent methods. 1) ROC curve analysis. I performed ROC curve analysis on all gene variables in order to assess their discriminating power with respect to the two groups (old vs. young) and with respect to this method, I set statistical significance at ROC AUC ≥ 0.920. 2) Fold Change. For all gene variables, I defined fold change (FC) as the mean expression value of the old subjects over the mean expression value of the young subjects and I set statistical significance at FC ≥ 1.10 (if O > Y) or FC ≤ 0.91 (if O < Y). 3) P-value. I used the independent t-Test for parametric gene variables (both normality and homogeneity of variance conditions were met); the Aspin-Welch unequal-variance test (AW) for gene variables that met the normality condition but not the homogeneity of variance condition; and the Mann-Whitney U test (MW) for the non-parametric gene variables, i.e., for those variables that i) the normality condition was not met or ii) the normality and the homogeneity of variance conditions were not met. Taking into account that there are 54,675 probe sets (including those of the housekeeping genes) in the Affymetrix HG-U133 Plus 2.0 chip and using the Bonferroni correction, I set the significance level for the entire study at α = 9.15 × 10−7. Therefore, in order for any variable to be deemed significant according to the P-value method, the following condition must be met: P < α. Regarding the Mann-Whitney U test (MW), since none of the non-parametric variables had any sets of ties (a subject from one group having the same expression value as a subject from the other group), I used the exact probability for all MW tests.

Incorporating the three aforementioned independent methods of statistical significance assessment and in order to minimize the number of false negatives in the case of the third method43,44, I set the overall significance criterion as follows: in order for any variable to be included in the final list of the most significant variables, it would have to meet the significance criteria of the first method (ROC AUC ≥ 0.920) and those of at least one of the other two methods [FC ≥ 1.10 (or FC ≤ 0.91) and/or P < 9.15 × 10−7].

Mathematical modeling

Utilizing the final 36 most significant genes, I wanted to explore the possibility of developing – via mathematical modeling – a function that could identify as correctly as possible the age status (O or Y) of an unknown subject based on the expression of any combination of those 36 most significant genes. To that end, I randomly selected approximately 70% of the subjects [11/15 young subjects and 18/25 old subjects] that could be used only for the development phase of such function. In other words, a function could be developed only by the exclusive use of those 29 subjects. The remaining 11 subjects (4 young and 7 old ones) were designated unknown (test) subjects and were used solely for the purpose of validating any promising function generated in the development phase. This split into two fixed sets, whereby one is used only for training and the other only for validation, represents the simplest implementation of K-fold cross validation45,46. A function was deemed promising in the development phase only if it exhibited a sensitivity ≥ 0.90 and a specificity ≥ 0.90 in connection with the 29 subjects of the development phase. Pertaining to the validation phase and in connection with the 11 unknown subjects, a promising function would have to exhibit the same minimum classification accuracy (a sensitivity ≥ 0.90 and a specificity ≥ 0.90) in order to be accepted. I was able to generate one such function (F1 – henceforward also referred to as super variable) that fulfilled all of the aforementioned criteria. Supplementary Fig. 1 shows the equation of F1 as a function of 7 genes.

The cut-off score of the F1 was determined by taking into account the results of the following two analyses: 1) calculation of the optimal point on the ROC curve based on the 29 F1 scores of the 29 subjects used in the development phase [optimal point is defined as the point with the highest sensitivity and the lowest false positive rate (1-specificity)] and 2) calculation of the 99.99% confidence intervals for the mean F1 scores of the two groups (O and Y) of those 29 subjects and their respective standard deviations. The 99.99% confidence intervals were calculated based on a bootstrap sample size of 100,000. Taking into account the aforementioned ROC optimal point, as well as the relative overlap of MO and MY [MO = LLO – SDO and MY = ULY + SDY] (LLO: the 99.99% confidence lower limit for the mean of the O group; SDO: standard deviation of the O group; ULY: the 99.99% confidence upper limit for the mean of the Y group; SDY: standard deviation of the Y group), the cut-off score of the F1 super variable was determined to be 53.450. If a subject's F1 score is < 53.450, then that subject is classified as Y (young); otherwise, if the F1 score is ≥ 53.450, then that subject is classified as O (old). It should be pointed out here that, based on the equation of the F1 (Supplementary Fig. 1), a given F1 score is just a numerical value and does not signify age or number of years.

In addition to the main validation method explained above and in order to further assess the performance of the F1 super variable, I employed two other and different cross validation methods: 1) a 10-fold cross validation and 2) a leave-one-out cross validation47. Both of those methods yielded a misclassification rate of 0.05 and a mean-squared error of 0.05 in connection with the F1 super variable. The results of those methods, along with the confusion matrices generated by them, are shown in Supplementary Table 9. As can be seen in Supplementary Table 9, each one of those two and different validation methods resulted in a correct classification of all of the young subjects and in a correct classification of all but two of the old subjects.

Computer software

All analyses in this study were carried out with custom software written in MATLAB R2012b. All computer programs in connection with the model were also created using MATLAB R2012b.