The Aging Transcriptome: Read Between the Lines

Abstract

The increasing sophistication of gene expression technologies has given rise to the idea that aging could be understood by analyzing transcriptomes. Mapping trajectories of gene expression changes in aging organisms, across different tissues and brain regions has provided insights on how biological functions change with age. However, recent publications suggest that transcriptional regulation itself deteriorates with age. Loss of transcriptional regulation will lead to non-regulated gene expression changes, but current analysis strategies were not designed to disentangle mixtures of regulated and non-regulated changes. Disentangling transcriptional data to distinguish adaptive, regulatory changes, from those that are the consequence of the age-associated deterioration is likely to create an analytical challenge but promises to unlock yet poorly understood aspects of many age-associated transcriptomes.

Introduction

Biological aging is not simply a process of organismal decay. Organisms have established mechanisms of self-repair that attempt to prevent physical deterioration and to restore homeostasis. However, for reasons that are poorly understood, these mechanisms of self-repair slowly expire leaving old organisms vulnerable to deterioration, damage and death [1,2]. Maybe surprisingly, aging is not only driven by external forces but also by endogenous pathways such as mTOR or insulin signaling [3]. The theory of “antagonistic pleiotropy” suggests that pathways promoting reproductive fitness that also promote aging will remain in populations because their detrimental effects occur in the “selection shadow” after reproduction, and thus are immune to selective pressure [4]. Many genetic or pharmacological interventions that extend lifespan show detrimental effects on reproductive fitness, consistent with the notion of “antagonistic pleiotropy.” Thus, age- associated transcriptomes are comprised of expression changes that actively promote aging, as well as expression changes that occur because the organism’s attempts to re-establish homeostasis through self-repair and stress response mechanisms that antagonize aging.

The rise of -omics technologies, that enable studies of gene expression, chromatin modifications, metabolomes, and proteomes, present a seemingly ideal opportunity to comprehensively describe molecular patterns of aging. To better understand the opposing forces that promote or antagonize aging, many studies recorded gene expression changes across different ages, brain regions and species. Recording gene expression changes as a function of age has been one of the most frequently conducted -omics experiments with thousands of age-associated gene expression profiles publicly available in data repositories. In spite of such a plethora of data, understanding the observed changes and how they lead to aging or age-associated dementia has been surprisingly difficult. More recent studies, including some using single cell RNAseq, suggest that aging directly impairs chromatin and transcriptional regulation itself [5–10]. These findings suggest, that the aging transcriptome contains a third class of gene expression changes that neither antagonizes nor promotes aging, but that is the result of transcriptional deterioration itself. To determine the meaning of this third type of altered gene expression, we suggest that new analytical strategies will be necessary to distinguish gene expression changes caused by regulatory programs from those that are caused by the deterioration of transcriptional regulation. We believe that separating regulated from deteriorative gene expression changes is likely to generate many new insights into already existing age-associated transcriptome data and unlock currently hidden aspects of aging transcriptomes.

What do aging transcriptomes reveal?

Analysis of age- associated gene expression changes in yeast, flies, worms, rats mice and humans have generally shown an age-associated decrease in genes related to the electron transport chain, protein translation, growth signaling and an increase in the expression of genes related to innate immunity, inflammation and DNA damage [11–14]. Similar gene expression changes are also found in progeroid mice, but less so in long lived mouse models, suggesting that these gene expression signatures reflect the consequences of ongoing pathology that is less prominent in mice with extended longevity [13,15]. However, it is important to note that changes in these pathways explain but a small fraction of the generally hundreds or thousands of genes that change expression with age, and most of the observed changes remain unexplained.

Despite our inability to fully understand many or most of the observed gene expression changes or to formulate meaningful hypothesis explaining them, transcriptomes do contain latent information about biological age. In most studies, principle component analysis (PCA) sorts gene expression datasets according to age, along one of the first 1 to 3 principle components. For this to occur, the information defining biological age must be contained within the transcriptional data [7,12,16–18]. This feature of aging transcriptomes has been exploited in studies attempting to identify gene expression signatures that can be used to quantify chronological or biological age. For example, recently Tyshkovskiy et al. reported a gene expression signature that can be used to predict the potential of pharmacological or genetic interventions to induce longevity in mice [16]. In humans transcriptional signatures of peripheral blood samples from large cohorts have been leveraged to develop expression signatures that are highly correlated with age across different ethnicities and can serve as a transcriptional clocks reporting age [12,19]. It is important to note that these signatures were constructed through iterative correlations with chronological age, with the goal to generate molecular clocks, rather than to understand how gene-expression drives aging. Signatures generated by different groups show relatively little overlap, either with each other, or with epigenetic clocks, also arrived at through correlation [20–22]. Still, the strong correlation with chronological age shows that transcriptomes must contain information about age that we are currently unable to unlock.

Pharmacological and genetic studies in different organisms clearly point to a strong conservation of aging [3,23]. While the genetic and pharmacological data are very clear, comparisons of age-associated transcriptomes between species often suggest otherwise. For example, substantial efforts were made to characterize gene expression changes in aging brains across different species. These studies identified changes in genes related to metabolism, DNA damage, stress response, and inflammation, similar to those seen in other tissues. However, these studies also indicate changes in synaptic transmission, providing a more direct link to age-associated cognitive decline [11,24–28]. In addition, the transcriptomes also suggested species-specific changes.

For example, Loerch et al. compared cortex expression patterns among mice, humans and rhesus macaques showing only few conserved changes in gene expression between aging mice and humans [27]. Furthermore, many changes specific to synaptic transmission showed increased down-regulation in humans compared to that in mice, suggesting a recently evolved primate-specific aging signature. In a follow up study the same group conducted an in depth analysis of human data and identified the chromatin remodeling factor REST to be a key protective player in the aging brain. REST was the first factor to distinguish between normal aging and neuro-degeneration [29]. This finding added an interesting twist to the previous evolutionary study [27]. The gero-protective activity of REST was functionally conserved in mice and even in C. elegans, despite a clear divergence in age-associated gene expression observed in their previous paper. Thus, the direct comparison of cortical transcription data between mouse and humans showed little overlap and pointed to a recent evolutionary divergence of brain aging. In contrast, later findings that identified the chromatin remodeling factor REST, showed that its function is conserved in mice and humans and even in far more distant species such as C. elegans [29]. Were it not for the persistent effort by the authors that identified REST, the interpretation of the gene expression data would have suggested that aging in mouse and human brains is very different with minor conservation.

Reading between the lines of aging transcriptomes

The studies above are examples that illustrate why we believe that current analysis strategies are insufficient to interpret much of the information within aging transcriptomes. This is unlikely to be the result of technical issues such as tissue extraction because the same data identify signatures that closely correlate with chronological or biological age and can detect interventions that extend lifespan [7,12,16]. Perhaps, the information related to biological age is contained within the transcriptional data, but that confounding factors are blurring the picture. Indeed, a few recent studies may have uncovered hints on what may be missing.

A hidden assumption behind pathway or gene set analysis, a standard approach in most analysis pipelines, is that the observed changes in gene expression actively contribute to a meaningful biological function. Consider the opposite, which would be to analyze randomly generated gene expression data by pathway analysis. Even if statistically significant correlations were found, they would have no sensible interpretation. Thus, in order for gene set enrichment or pathway analysis to be meaningful, the observed variation must be associated with functional changes reflecting a regulated biological activity that changes in response to age.

An early study on transcriptional noise suggested that transcription itself is subject to age associated deterioration. Bahar et al. quantified the transcripts of specific genes in single cells of young and aged hearts and showed that transcriptional control itself degenerates as a function of age, a phenomenon called transcriptional noise [6]. The increase in availability of single cell RNAseq has recently enabled measurements of thousands of transcripts across different cell types and thousands of cells. Single cell RNAseq studies have confirmed that transcriptional noise increases in many cell types with age [8,9]. However, some studies have shown that the phenomenon is not universal and instead identified cell types in which transcriptional noise seem to remain stable or even decreases [10,30,31]. Despite the findings that transcriptional noise does not increase in some cell types, the evidence suggests that transcriptional regulation deteriorates with age and therefore, age-associated transcriptomes not only consist of gene expression changes caused by regulated adaptive responses but also of changes that are caused by the deterioration of transcriptional regulation [7–10,30,32,33].

From the standpoint of biological processes, the increase in stochastic noise is likely to have an overall corrosive effect (Fig. 1). However, with regards to analyzing data, transcriptional noise can be relatively easy to account for as it is averaged out when combining data from many cells into a mean value within an experiment. Much more difficult are changes that are non-random but still non-regulated, a phenomenon we named transcriptional drift (Fig. 1)[7]. Gene expression changes due to transcriptional drift are reproducible across cells, samples and experiments but are nevertheless the result of regulatory breakdown. For example, in yeast, aging causes a reproducible and widespread loss of histones along large sections of DNA [32,34]. Such a loss will increase the transcription of genes that are usually actively repressed. Similarly, age-associated mitochondrial dysfunction will impair the synthesis of chromatin remodeling co-factors and therefore impact chromatin remodeling [5,35]. The change in the availability of chromatin remodeling co-factors would then result in consistent non-random changes in gene expression (drift) that are not the result of a transcriptional response and only affect parts of pathways. As a result, expression changes due to transcriptional drift will look similar to regulated changes even though the mitochondrial dysfunction was the result of stochastic non-regulated processes such as oxidative stress caused by metabolic activities. Gene expression changes due to deterioration of biological functions that systemically influence transcription are likely to appear as species specific aging patterns, as they are caused by stochastic processes that are unlikely to replicate between species.

Fig. 1. Random and non-random forms of transcriptional degeneration.

Figure 1 shows the transcripts levels of three genes that encode proteins A, B, C forming a hypothetical complex. It illustrates how age-associated transcriptional noise and drift affect subunit expression and therefore complex stoichiometry and complex formation. The complex consists of 4 subunits, 2 x A’s, 1x B and 1x C that require stoichiometric expression. As the animal ages, loss of transcriptional control leads to transcriptional noise and drift. These gene expression changes are not due to a cell-regulated response, but are caused by the deterioration of transcriptional control. Transcriptional noise describes random/stochastic changes that increase heterogeneity across cells, with each cell expressing different transcripts levels. While the mean transcript level may not change with age, the variance between cells will. However, even if the mean transcript level remains constant for each subunit, complex formation will vary from cell to cell dependent on the expression level of each subunit. Transcriptional drift causes non-regulated gene expression changes due to systemic problems, such as mitochondrial decline, increased Ca ²⁺, loss of histones and other effects of deterioration, that are common during aging and interfere with aspects of transcriptional regulation. Because transcriptional drift is the result of damage to components important for some, but not all aspects of transcriptional regulation, it might differentially affect different complex subunits, thereby disrupting stoichiometric expression and dramatically reducing the ability of the cell to form complexes. Interventions that induce longevity prevent transcriptional drift in aged animals and consequently preserve stoichiometric expression patterns and the ability to form complexes into later age.

A major difficulty is that changes in which the expression of a gene drifts away from its original level due to a lack of regulation are not easily distinguishable from adaptive changes that are initiated by regulated processes. One way to distinguish these two possibilities is to realize that co-expression patterns of functional modules should remain stable over time. An early pioneering study revealed an age-associated decrease in the correlation of gene expression patterns within genetic modules [36]. The study found that the expression of subsets of genes within a regulatory module showed a strong correlation in expression in young animals, but with age gene expression continuously drifted, leading to a decline in the correlated expression patterns. Importantly, the genes that showed decreases in correlated expression patterns over age were generally clustered in specific areas of the chromosome, hinting at the possibility of an non-regulated and non-random decline in chromatin structure as the animal got older.

If so, how could an age-associated decline in correlated expression patterns negatively affect the animal? Most genes act in concert with others to accomplish a biological function. For example, protein complexes like that of the enzyme ATP synthase consists of multiple subunits that must combine with a specific stoichiometry to form a functional complex. Either too little or too much expression of one subunit relative to other subunits can be detrimental to complex function either because subunits are missing, or because too much of a subunit will act as a competitive decoy, titrating away other subunits (Fig. 1). Loss of stoichiometry of ATP synthase genes can be expected to result in functional decline of the complex.

A more recent, study showed that transcriptional drift affected thousands of genes in aging transcriptomes across several species, and that it disrupted the stoichiometric co-expression of many pathways including the co-expression of synaptic vesicle release factors [7]. Importantly, treatment of worms, flies or mice with lifespan extending interventions, such as small molecules or intermittent fasting, stabilizes neuronal transcriptomes against transcriptional drift and preserves youthful gene expression in aging animals [7,18,37]. A proteomic study of aging C. elegans further confirmed that age-associated drift leads to a loss of normal protein complex stoichiometry, further supporting the notion that drift is the consequence of deterioration with age and unlikely to be a regulated process [38]. The functional consequences of transcriptional drift have also been studied and shown to alter metabolic pathways leading to imbalanced metabolite synthesis. Metabolomics studies in flies, fish and in brain tissues of mice and humans showed that aging is associated with dramatic relative changes in metabolite levels and with the appearance of metabolites in old animals that are undetectable in the young or detrimental to the cell [5,37,39,40]. Together, these findings provide considerable evidence for transcriptional drift and therefore for the existence of non-random and non-regulated gene expression changes associated with aging.

Conclusions

The co-existence of adaptive gene expression changes and those caused by transcriptional drift within the same transcriptome creates a considerable challenge for the interpretation of gene expression data. Gene expression changes due to drift are most likely non-functional because they do not lead to the functional expression of the entire pathway and, hence, should be excluded from pathway or gene set enrichment analysis. They should further be excluded from evolutionary or tissue comparisons, as they are caused by deterioration and not evolutionary pressure. It may well be that many of the observed species or tissue specific changes are caused by transcriptional drift. If the fraction of genes that drift is substantial across a transcriptome, multiple hypothesis testing correction will have to be adapted accordingly because the total number of genes included will substantially decrease. Instead of being analyzed for their biological function, genes whose expression drifts, should be analyzed with respect to chromatin modification or chromosomal location for insights into the mechanisms responsible for the disruption of their regulation. Thus, future efforts should develop new analysis strategies for transcriptomes that account for the unique biology of aging and consider co-expression values, chromosomal location and histone modification data to separate regulated from dysregulated gene expression changes. To unlock aging transcriptomes, we will have to start to read between the lines.

Highlights.

• Aging transcriptomes consist of regulated and dysregulated gene expression changes

• Dysregulated transcriptional changes can be random (noise) or non-random (drift)

• Novel strategies will be required to disentangle regulated from non-regulated changes