Personalized transcriptional network analysis links age-related loss of gene coordination to individual biological aging

Abstract

Background

Aging is characterized by the decline in biological functions, accompanied by changes in gene-to-gene transcriptional coordination, which can be estimated by expression coordination in gene transcriptional network. Notably, gene networks and coordinated expression relationships (CERs) showed inter-individual variability, while personalized aging-related gene expression coordination dynamics in human cohorts have yet to be investigated.

Methods

In this study, we constructed 15,933 personalized transcriptional networks across 26 tissues from 967 donors aged 20 to 80 years old, using the sample-specific network (SSN) framework based on data from the Gene-Tissue Expression (GTEx) project.

Results

We identified gene–gene CERs and characterized their age-dependent dynamic trends across tissues, observing a universal trend of increased gene-to-gene coordination loss during aging across tissues. The count of lost CERs is also positively correlated with individual-level aging and senescence-related molecular phenotypes. Notably, we revealed that the lost CERs have potential as biomarkers for individual aging and health status. In addition, we identified gene coordination loss events exhibiting significant positive correlation with age, defined as aging-related lost relationships (ARLRs), which may be functionally associated with pathways related to proteolytic processes. Finally, we showed that ARLRs may contribute to deleterious effects and increased pathogenicity through gene dosage imbalances.

Conclusions

This study establishes, for the first time, a connection between the loss of gene-to-gene expression coordination and individual-level aging progress. It provides proof-of-principle evidence for using lost gene coordinated expression relationships as biomarkers of healthy aging and highlights the potential risks associated with coordination loss in specific biological pathways during aging. 

Supplementary Information

The online version contains supplementary material available at 10.1186/s13073-025-01533-6.

Background

Aging is characterized by widespread dysregulation across various biological levels [1, 2], including a decline in gene-to-gene transcriptional coordination [3, 4]. This decline leads to reconfigured interrelations within gene transcriptional networks [5], which warrants study to better understand the biological system disorders in the aging process. Previous studies have explored the global transcriptional coordination level [6] and gene-to-gene coordination dynamics during aging in specific tissues or across multiple tissues and cell types [7–11]. These studies highlighted network analysis as a powerful tool to investigate gene expression dynamics and suggested that aging is associated with decreased gene-to-gene transcriptional coordination. However, the gene pair coordinated expression relationships (CERs) in these analyses were estimated using correlation coefficients across a bulk of samples, capturing only population-level trends. Changes in CERs within individuals during aging remain unclear. Especially since such an analysis cannot determine whether two genes are coordinated in a single individual, it is difficult to connect the gene coordination to personalized biological functioning and health status, thus limiting its potential as an indicator or biomarker of an individual’s biological aging or even disease risk. Therefore, a study focusing on gene-to-gene transcriptional coordination at an individual-specific level is urgently needed to gain deeper insights into its biological significance in aging and even age-related outcomes. 

With advances in omics analysis algorithms, various methods have been proposed to interpret and extract personalized biological regulatory networks from single samples’ omics data [12]. These methods enable the investigation of gene-to-gene coordination dynamics related to aging at the individual level. Notably, a single-sample network (SSN) approach has been developed specifically to identify the significant differential gene relationships between a test sample and reference samples [13], such as baseline samples for longitudinal data. Thus, this approach is particularly well-suited for studying the individual-level gene-to-gene coordination change trajectory across different ages.

To systematically explore the individual-level gene–gene expression coordination dynamics during aging, we constructed 15,933 personalized transcriptional networks using the SSN method in 26 tissues from 967 donors (ranged from 20 to 80 years old), sourced from the Genotype-Tissue Expression (GTEx) project. We revealed that the loss of gene coordination is positively correlated to an individual’s senescence-related molecular traits (e.g., senescence-associated secretory phenotype (SASP) and immune cell infiltration) and biological functioning processes (e.g., reactive oxidative species (ROS) and oxidative phosphorylation). Notably, we provided evidence showing that age-related CER loss has the potential to serve as an indicator of an individual’s biological aging and health status. Moreover, we found that gene-to-gene relationship loss during aging leads to disrupted gene expression coordination in key pathways, such as proteolysis, which are closely related to longevity and healthy aging. Further analysis indicated that the aging-related CER loss may be pathogenic in a gene dosage-dependent manner.

Methods

Data resource

Genotype-Tissue Expression (GTEx) v8 [14] transcriptome dataset and the corresponding metainfo files were downloaded from UCSC Xena data portal [15]. The gene-level expression level was normalized by transcripts per million (TPM). PEER (probabilistic estimation of expression residuals) factors were inferred from gene expression data using factor analysis methods, which account for hidden technical and other covariates such as batch effects, RNA quality, environmental influences, and other known or unknown factors [16]. Considering that many of the raw GTEx PEER factors were significantly correlated with sample age, which may hinder the correlation between interested phenotypes and age. So, we adopted the corrected GTEx PEER factors [17] as covariates (retrieved from https://github.com/sudmantlab/gene_expression_aging). Referring to previous studies like the recount2 project [18] (https://research.libd.org/recount-brain), we classified the sample health status according to the Hardy-scale death classification, with samples belonging to scales 3 and 4 categorized as unhealthy and the rest categorized as healthy.

RNA-seq data preprocess

We firstly adopted some filter steps. The retained genes must be listed in protein–protein interaction networks from the STRING v11 database[19], whose credential score > 700. We then filtered out genes whose expression level TPM < 1 in more than 30% of samples. After the preprocessing procedures, 10,702 genes were retained for downstream analysis.

Sample-specific network construction

The sample-specific network (SSN) was computed via Python scripts implemented by Liu et al. [13], which have been deposited on https://github.com/xp-liu/SSN. The samples from the youngest age group (i.e., 20–30 years old) were adopted as reference samples. We only retained those edges that meet the two criteria: (1) Storey-Taylor-Siegmund’s adaptive step-up FDR < 5% and (2) the raw P-value < 0.001.

We defined the coordinated expression relationship (CER) loss/gain by these procedures. For each tissue, we firstly obtained the significant edges by the above thresholds. Then for each edge, we examined if its sample-specific change direction (i.e., the sign of change of Pearson correlation coefficients, ΔPCC) is the same as its correlation in reference samples. For example, we assume an edge’s ΔPCC was < 0, while in reference samples, the PCC of the two genes was > 0; then we can define such CER as “lost” in the tested sample. If the sign of ΔPCC is the same as the sign of PCC in reference samples, which means the correlation was intensified in the test case, we defined the CER between the two member genes of the edge as “gaining CERs”

Construction of lost coordinated expression relationship-based clock

Because blood is the most readily obtainable tissue for molecular diagnosis, we used data from blood samples to demonstrate the potential of lost CER for biological clock construction. We identified that 303,216 gene–gene expression CERs were lost in at least one blood sample. We then retained relationships significantly positively correlated with age group (Kendall’s correlation test P < 0.05) and used their ΔPCC calculated by SSN as input features; the chronological age from Rath et al.[20] was set as the output variable. We randomly selected 80% of the blood samples as the training set, using the rest for testing. We used elastic net regression on the training dataset with eightfold cross-validation for parameters (i.e., alpha and lambda) tuning and informative feature selection. Pearson correlation coefficient and mean absolute error (MAE) were used to estimate the clock performance.

Gene annotations and molecular traits for senescence

The GenAge genes were retrieved from the Human Ageing Genomics Resources database [21] (HAGR, https://genomics.senescence.info/genes/index.html). The HALLMARK gene sets were from the MSigDB database [22] (https://www.gsea-msigdb.org/gsea/msigdb/human/genesets.jsp?collection=H). The GTEx sample-matched precomputed stemness scores and mitochondrial copy numbers were downloaded from the supplementary materials of Dos Santos et al. [23] and Rath et al. [20], respectively. Senescence-associated secretory phenotype (SASP) genes were downloaded from the REACTOME database [24] (https://reactome.org/content/detail/R-HSA-2559582).

Calculation of gene set-level gene expression levels

The aggregated average gene expression level from gene sets was computed by the GSVA v1.50.5 R package [25] with default parameters. The input transcriptome matrix was the TPM matrix.

Cellular deconvolution based on transcriptome

The immune cell proportions were computationally deconvoluted by CIBERSORT [26] (R script v1.03 from https://github.com/Moonerss/CIBERSORT), with the LM22 panel as the reference panel. The input for CIBERSORT was both the TPM matrix. The seven cell type deconvolution was conducted by xCell [27] (https://github.com/dviraran/xCell), which was retrieved from the GTEx portal [14] (https://www.gtexportal.org/home). Pearson correlations were calculated between the count of lost CERs and cellular proportions estimated by CIBERSORT and xCell.

Aging-related lost CERs and network

For every lost CER in each tissue, we performed the Kendall correlation analysis between its prevalence and the age group. The relationships whose P-value < 0.001 and Kendall’s τ > 0 were defined as aging-related lost relationships (ARLRs). All the retained ARLRs have Storey–Taylor–Siegmund’s adaptive step-up FDR < 5%. After getting the ARLRs, we collapsed these relationships to a network, whose edges were the ARLRs, and nodes were genes in ARLRs. The collapsed network was defined as the aging-related lost network (ARLN) for each tissue. The gene with the highest degree in the ARLN was defined as the tissue’s core gene, and its neighbors were defined as the core module.

Construction of comprehensive ARLN, gene set, and pathway enrichment analysis

The comprehensive ARLN was constructed based on the CERs whose loss was significantly associated with increasing age in more than five tissues (i.e., those were identified as ARLRs in more than five tissues). The hub gene was selected as the gene with the highest degree. Pathway enrichment analysis was performed by the g:profiler [28] website with 17,224 genes included in our study as background genes. We only included the pathways from KEGG and Wiki Pathways.

Two-sample Mendelian randomization analysis and stratified LD-score regression

For each tissue, we extracted the core module genes in its ARLN as a gene list for causality analysis. The two-sample Mendelian randomization (MR) analysis was performed by the R package TwoSampleMR [29]. Longevity summary statistics for the European (EUR) ancestry population were from the GWAS catalog [30] (https://www.ebi.ac.uk/gwas) under accession: GCST008598. Focal genes were extracted from each ARLN by retaining the neighbors of the core gene who have the highest degree. Tissue-matched eQTL data were retrieved from the GTEx database [14]. Stratified LD-score regression (s-LDSC) analysis was performed by LDSC v1.0.1 [31]. The ± 2 Mb regions of genes were used as the annotations, and the LD score window was 1 cM (default). The 1000 Genomes (1000G) Phase 3 data files were from https://zenodo.org/record/7768714[32].

Gene dosage risk analysis

The risk scores for gene dosage alteration were retrieved from the pHaplo and pTriplo scores trained by Collins et al.[33]. The rCNV association with 41 diseases was also retrieved from this paper. For each tissue and each disorder, we constructed a 2 × 2 contingency table containing the numbers of ARLR genes associated/not associated with the disorder and the numbers of non-ARLR genes associated/not associated with the disorder. The enrichment analyses were then performed by Fisher’s exact test. The sum for the 2 × 2 contingency table was 17,224 (number of the genes included in our study). The associations with P < 0.05 and FDR < 25% were considered significant.

Data visualization

The visualization of networks was performed by the igraph R package and tuned by Cytoscape v3.8.2. The heatmaps were plotted by the pheatmap v1.0.12 and ggplot2 v3.5.0 R packages.

Availability of data and materials

This study did not generate new datasets. The scripts for analysis in this study has been deposited on GitHub repository (https://github.com/ritianjiang/SSN_paper) [34] .

Results

Increased loss of gene-to-gene coordination with age

To gain insights into how aging influences gene-to-gene expression coordination, we retrieved transcriptomic data of 15,933 tissue samples across 26 organs from 967 donors in the GTEx (v8) database [14], with ages ranging from 20 to 80 years. We applied the SSN method to construct the personalized transcriptional network for each tissue sample, enabling the analysis of the gene coordination dynamics across various tissues during aging. Based on the proposed concept of a “ground state” in organismal aging [35], we defined the youngest age group (i.e., 20–30 years) in the cohort as the reference controls within the SSN framework, while the remaining samples were classified as cases. For each tissue, we then categorized gene-to-gene coordinated expression relationships (CERs) with decreased correlation coefficients in the cases as “lost CERs” and those with increased correlation coefficients as “gained CERs” (Fig. 1A, Table S1, and Methods), representing the weakening or strengthening of gene-to-gene transcriptional coordination, respectively. 

Fig. 1.

Age-related gene–gene coordination decline associated with age and individuals’ senescent levels. A Schematic diagram of the identification of CER loss and its downstream analysis. Tissue icons were retrieved from https://reactome.org/icon-lib (CC BY 4.0). B The pattern of changed CERs in younger (30–40 years) and older (> 60 years) groups, using the 20–30 years group as baseline. (Left) The distribution of proportions of lost CERs in total changed CERs. (Right) The number of changed CERs in the 30–40 years and > 60 years groups. BH-adjusted *P<0.05, **P<0.01, ***P<0.001; Wilcoxon’s test. C Results of Pearson’s correlation analysis illustrating the relationship between chronological age and counts of gaining/lost CERs (Signif., BH-adjusted P < 0.05). D Correlation between SASP gene expression levels and counts of lost CERs among various tissues (BH-adjusted *P < 0.05, **P < 0.01, ***P < 0.001; Pearson’s correlation test). E Scatter plots showing positive correlations between SASP expression and lost CER counts in breast, salivary gland, and spleen. F Heatmap showing correlation between expression of HALLMARK pathways (gene sets) and lost CER counts. Spearman’s correlation coefficients were shown. G Scatter plots showing correlations between the expression level of certain representative pathways and lost CER counts in ovary, heart, lung, and blood. H Heatmap illustrating correlation between cellular proportions and lost CER counts (BH-adjusted *P < 0.05, **P < 0.01, ***P < 0.001; Pearson’s correlation test). The cell proportions were computed using xCell deconvoluted methods

To determine whether alterations in coordination patterns are correlated with chronological age, we compared the counts of CERs between older (≥ 60 years) and younger (30–40 years) samples, using the 20–30 years group as the baseline. We found that lost CERs account for the majority of changes (87.5 on average; Fig. 1B). Notably, older samples showed significantly more changed CERs than younger samples in 18 of 26 (69.2%) tissues (Fig. 1B, S1). Subsequently, we performed correlation analysis for all 15,933 samples across various age groups and observed a significant positive correlation between ordinal age and the counts of lost CERs in 15 of the 26 (57.7%) tissues (Fig. 1C). This finding was further validated by adjusting for PEER factors, which estimate unobserved transcriptome confounders (e.g., batch effect, RNA quality) [16] (Fig. S2). Additionally, we also investigated the association between CER gain and age and found that only one tissue (i.e., esophagus) exhibited a significant association between age and the count of gained CERs (Fig. 1C). These results jointly suggest that the loss of gene coordination is a common and dominant characteristic shared by most tissues during aging.

Gene-to-gene coordination loss is associated with senescence-related biological processes

The personalized network method allows us to link the gene coordination status to phenotypes in each sample. Therefore, we investigated whether individual-level gene CER loss is associated with the individual’s aging level across different tissues. Since the overall expression level of cell senescence genes has been used to assess tissue degeneration and aging[36], we calculated the Gene Set Variation Analysis (GSVA) normalized expression level of the senescence gene set from CellAge [37] of each sample. Of the 26 tissues, 11 displayed a significant positive correlation between the overall expression level of senescence genes and the count of lost CERs (Fig. S3), suggesting an association between senescence/aging and gene coordination loss.

In addition, we also analyzed the association between gene-to-gene coordination loss and the expression of senescence-associated secretory phenotype (SASP) factors, which act as key mediators of the senescent tissue microenvironment and important indicators of tissue aging [38]. Our analysis revealed that in 23 of 26 tissues, the count of lost CERs significantly increased alongside the overall expression levels of SASP genes, with exceptions observed in the spleen, muscle, and testis (Fig. 1D and E). Since stem cell exhaustion and mitochondrion copy number (mtCN) decline contribute to physiological dysfunction during tissue aging [23, 38], we further analyzed the sample-matched stemness scores [23] and mtCN [20], finding that lost CER counts were widely negatively correlated with these two traits across different tissues. Specifically, the negative correlation with stemness scores was found to be statistically significant in 17 out of 26 tissues (Fig. S4), and with mtCN in 7 out of 20 tested tissues (Fig. S5). Together, these findings highlight that the loss of gene-to-gene coordination is positively correlated with an individual’s aging degree across multiple tissues.

Next, we analyzed the correlation between the count of lost CERs in each sample and the individual’s expression levels of 50 REACTOME Hallmark pathways, which represent well-defined biological states or processes. The results revealed that the association between coordination loss and pathway expression levels varied considerably among tissues. Notably, key energy production pathways, such as oxidative phosphorylation (MSigDB: M5936), and proteolytic pathways including peroxisomes (MSigDB: M5949), showed a broadly negative association with CER loss in various tissues (Fig. 1F and G). In contrast, there is a widespread positive correlation between CER loss and the expression of certain canonical cancer-related pathways, such as the epithelial-mesenchymal transition (EMT, MSigDB: M5930) and the P53 pathway (MSigDB: M5939), and pathways associated with cellular damage, such as the reactive oxygen species (ROS) pathway (MSigDB: M2932) (Fig. 1F and G).

Furthermore, among the Hallmark pathways, we noticed that the individual’s expression levels of several immunity and inflammation-related pathways were generally associated with the corresponding sample’s count of lost CERs, including TNFα signaling via NFκB (MSigDB: M5890), IL2 STAT5 signaling (MSigDB: M5947), and inflammatory response (MSigDB: M2932) (Fig. 1F and G). These findings align with the notion that tissue aging was accompanied by a pro-inflammatory microenvironment [39]. Subsequently, we estimated the compositions of 22 types of immune cells for each sample by computational deconvolution and explored the association between individual’s CER loss and immune infiltration. Our analysis revealed a significant positive correlation between neutrophil infiltration, a hallmark of immune aging [40, 41], and lost CER counts in 16 of the 26 tissues (Fig S6). In addition, we found that the M2-like macrophage tended to decrease alongside CER loss (Fig S6), consistent with recent findings suggesting that the enhancement of M2-like macrophages may have an anti-aging effect [42].

Since aging-related organ functional decline can also be mediated by dysregulated proportions of non-immunes cells [43], we expanded deconvolution analysis to include non-immune cells in the tissues using the xCell algorithm. The proportions of seven main cell types were obtained. Among them, the positive correlations between neutrophil proportion and lost CER count were further validated in 14 of the 26 tissues (Fig. 1H). For the remaining non-immune cells, the associations between lost CER count and cell type compositions varied across tissues, while we observed that in many tissues, counts of lost CERs were associated with decreased proportions of major cell types (Fig. 1H). For example, in the brain, the proportion of neurons exhibited a significant negative correlation with lost CER counts, while in the lung and small intestine, the proportions of epithelial cells were negatively correlated with CER loss (Fig. 1H). These results suggest the age-related loss of gene-to-gene transcriptional coordination is associated with progressive biological function decline during tissue aging.

Gene coordination loss relates to an individual’s age and health

The association between the loss of gene coordination and individual aging suggests its potential as an aging indicator. To evaluate this, we conducted a proof-of-principle validation leveraging lost CERs for individuals age prediction. Focusing on blood, the most readily obtainable tissue for molecular diagnosis, we built a lost CER-based clock (Lost Relationship Age, LRAge) (Fig. 2A and Methods). LRAge, comprising 25 lost CERs, accurately predicted chronological ages in both the training dataset (Fig. 2B, mean absolute error MAE = 7.44 years old) and the test dataset (Fig. 2C, MAE = 7.95 years old).

Fig. 2.

Lost coordinated expression relationship-based clock and CER loss among stratified healthy/unhealthy individuals. A Concept and construction pipeline of LRAge. B, C Accuracy of LRAge in the training dataset (B) and test dataset (C); P-values were calculated by the Pearson correlation test. D The average count of lost CERs of unhealthy individuals minus that of unhealthy individuals among 14 tissues and five age groups (BH-adjusted *P < 0.05, **P < 0.01, ***P < 0.001; Wilcoxon rank-sum test). Dot size and color represent the difference in lost relationship counts E, F The exceeding counts of lost CERs of unhealthy individuals in blood (E) and brain (F) (BH-adjusted *P < 0.05, **P < 0.01, ***P < 0.001; Wilcoxon rank-sum test). The horizontal line represents the median value in each violin plot

Since individuals of the same chronological age often differ in health status [44], we examined whether CER loss correlates with poorer health in donors of similar chronological ages. Using the cause of death of GTEx donors, we divided the samples into healthy and unhealthy groups for each tissue, focusing on 14 tissues with at least 100 samples in both groups for downstream analysis. We observed a widespread increase in lost CER counts in unhealthy individuals (Fig. 2D). For example, unhealthy individuals exhibited significantly more relationship loss in blood across three age groups (Fig. 2E), with an even more pronounced trend in brain across all five age groups (Fig. 2F). Together, these findings emphasize the potential of personalized transcriptional network-derived gene coordination loss as a biomarker for individual healthy aging.

Age-related loss of gene expression coordination is negatively correlated with longevity and healthy aging

To further explore the health-related biological significance of the age-related gene coordination loss, we identified the aging-related lost relationships (ARLRs) by testing the correlation between CER loss and chronological age for each tissue (Fig. 3A and Methods). The number of ARLRs varies more than 100-fold across different tissues (Fig. 3B, Table S2). These CERs can be conceptualized as lost networks in each tissue, with ARLRs representing edges and genes serving as nodes (Fig. 3A; Fig S7). Given that the core module genes in gene expression networks play crucial roles, we hypothesized that the overall expression of core module genes in lost networks may be associated with healthy aging. We then conducted a two-sample Mendelian randomization (MR) analysis, combining longevity GWAS summary statistics with tissue-specific cis-expression quantitative loci (cis-eQTLs) for 26 tissues’ lost networks core module genes. Our MR analysis revealed that, in two tissues (brain and adrenal gland), the expression levels of core module genes in lost networks were causally associated with the likelihood of longevity, despite significant differences in their network scales (Fig. 3C–F). We further validated the associations of brain and adrenal gland with longevity using stratified LD score regression (s-LDSC) analysis, which showed that longevity heritability is preferentially enriched in core module genes of both tissues (Table S3). Moreover, the core module gene expression showed a positive association with longevity in the brain but a negative association in the adrenal gland (Fig. 3D and F), which is in accordance with the expression dynamics of core module genes in the two tissues (Fig. S8).

Fig. 3.

Aging-related lost relationships (ARLRs) among tissues. A Schematic diagram illustrating the ARLR identification and the construction of the lost network. B Numbers of ARLRs in different tissues. C,D The connected ARLRs in the brain were contacted as the aging-related lost network (ARLN) (C) and the Mendelian randomization (MR) results of the ARLN hub genes’ expression and surviving beyond age corresponding to the 90th survival percentile (D). E,F The ARLN in the adrenal gland (E) and the corresponding MR result for its hub genes (F). G The constructed common ARLN. The hub module zoomed in. H Pathway enrichment results of the hub module of the comprehensive lost network. P-values were adjusted by the BH method

Since our results indicate that gene coordination loss is a common characteristic of tissue aging, we next explored whether the biological functions of lost CERs shared any similarities across multiple tissues. To address this, we integrated the ARLRs present in more than five tissues to construct a common aging-related lost network, referred to as the comprehensive lost network (CLN) (Fig. 3G). This network compromised 1746 relationships and 1674 genes, with significant enrichment in pathways such as mRNA processing, ribosome, and pathways of neurodegeneration (Table S4). The hub gene of CLN was UBC, which encodes a canonical polyubiquitin precursor ubiquitin C [45], indicating that this CLN may function in protein degradation (Fig. 3G). Further pathway analysis of the neighbor genes of UBC within the CLN revealed that the two most significantly enriched terms were proteasome (P = 1.01 × 10⁻⁹) and ubiquitin-mediated proteolysis (P = 4.78 × 10⁻⁵) (Fig. 3H). Collectively, these findings suggest that gene-to-gene coordination loss during aging leads to disrupted function in key pathways that may be essential for healthy aging and longevity.

Gene coordination loss could be pathogenic in a gene dosage-dependent manner

We further explored the underlying biological basis for age-related loss of gene coordination by analyzing the expression changes of member genes in lost CERs. In ARLRs, particularly those with mutual positive co-expression in reference samples (i.e., the young controls in 20–30 years), we observed that one gene often exhibited reduced expression compared to its paired gene during aging, likely leading to the coordination loss through gene-to-gene expression dosage imbalances [46] (Fig. S9). To estimate the effects of ARLR genes’ dosage alteration during aging, we retrieved the risk scores of gene dosage alteration from a recently published large-scale CNV-disorder association analysis [33]. We found that the dosage risk scores for ARLRs were significantly higher than those of other genes in all 26 tissues, both for haploinsufficiency and triplosensitivity (Fig. 4A–F). These findings indicate that gene coordination loss during aging may contribute to adverse health effects via a gene dosage-dependent manner.

Fig. 4.

ARLR genes have a dosage-dependent effect on health and diseases. A,B Risk scores of ARLR genes and background genes for haploinsufficiency (A) and triplosensitivity (B). C–F Details of dosage risk scores in the adrenal gland (C), brain (D), small intestine (E), and pancreas (F). P-values were all BH-adjusted. G Heatmap representing the over-representation of tissues’ ARLR genes in diseases’ potential dosage pathogenic genes. Over-representations with P < 0.05 and FDR < 25% were labeled red. H The over-representation results of tissues’ ARLR genes for four diseases. I The overlap of small intestine ARLR genes and seven diseases’ copy-number associated genes. J ARLRs that were linked to CDIPT and PI4KA. The panel below represents the lost ratio of two CERs among different age groups. K Expression profiles of PI4KA among age groups. L A diagram representing the roles of CDIPT and PI4KA in phospholipid metabolism

Furthermore, we investigated whether the dosage of ARLR genes is associated with diseases by retrieving disease-specific dosage-sensitive genes for 41 diseases [33]. For each tissue, we conducted one-by-one association analyses to test for significant overlap between the ARLR genes and disease-specific dosage-sensitive genes. ARLR genes from the small intestine exhibited the highest numbers of significant associations to diseases (Fig. 4G) and had higher dosage-dependent risk scores compared to background genes (Fig. 4E). Moreover, we found that ARLR genes in the small intestine were not only associated with digestive system disorders, such as abnormality of digestive system morphology (HPO: 0025033), but also with certain motor and behavioral disorders, including abnormal central motor function, seizure, and schizophrenia (Fig. 4H). By examining the overlap between small intestine ARLR genes and dosage-sensitive genes in diseases, we found that PI4KA and CDIPT, both key players in phosphoinositide signaling pathways, overlapped with the highest number of diseases (7 and 4 diseases, respectively; Fig. 4I). Additionally, both PI4KA and CDIPT exhibited a loss of CER with PI4K2B (Fig. 4J) during small intestine aging, as CDIPT and PI4K2B remained relatively stable in expression (Fig. S10), while PI4KA was significantly downregulated (Fig. 4K). These suggested that the loss of transcriptional coordination between PI4KA, CDIPT, and PI4K2B during small intestine aging may adversely affect human health in the digestive system and even other systems, both directly and indirectly. Our results altogether indicate that the age-related loss of gene-to-gene coordination is pathogenic, driven at least in part by an imbalance in the expression dosage of gene pairs.

Discussion

This study constructed personalized networks for ~16,000 tissue samples across 26 organs and systematically characterized the individual-level dynamic changes in gene regulatory coordination during human aging. To our knowledge, this is the first analysis of aging-related gene-to-gene transcriptional coordination changes at the individual level across multiple tissues. Unlike previous studies [6–11] that provide only population-level insights into gene coordination, our study revealed the positive correlation between individual-level gene coordination loss and donor-specific aging and/or aging-related traits. Specifically, we found that an increased count of CER losses in an individual was associated with elevated levels of SASP, decreased tissue stemness, and mtCNs, with this pattern remaining relatively consistent among tissues. Taken together, these findings support that gene coordination loss represents not only a macro-scale aging-related pattern but also relates to biological functions at the individual level. Recent studies have suggested that age-related increases in cell-to-cell gene expression variation and regulatory changes (e.g., DNA methylation) in certain tissues [47–49]. The loss of transcriptional coordination observed at the bulk level may partly reflect the progressive accumulation of regulatory randomness in the cell population during aging.

Given the unique advantage of a personalized network, we proposed using individual-level gene coordination loss as a network-based biomarker for healthy aging. In a proof-of-principle study, we constructed the LRAge clock based on blood loss CERs, showing its high performance in predicting an individual’s age with a MAE of 7.95 years. Importantly, we showed that the count of CER losses is increased in unhealthy individuals, even though they were of similar chronological ages to healthy individuals. Therefore, our study advances the concept of gene-to-gene transcriptional loss from a somewhat vague notion to a potential operational biomarker for aging or healthy aging, leveraging the advancements in personalized networks.

Furthermore, we identified a series of age-related gene–gene CER loss events, termed ARLR, for each tissue. Those ARLRs shared by multiple tissues appeared to be functionally associated with proteolytic pathways, such as proteasome and ubiquitin-mediated proteolysis. This loss of gene coordination among genes involved in these pathways aligns with recent studies that indicate a distinct loss of proteolytic complex stoichiometric balance with age across various species and organs [50, 51]. In addition, the biogenesis aspect of protein homeostasis has been broadly linked to aging [52]. Factors that increase translational fidelity and decrease translation levels have been shown to causally extend lifespan [53, 54]. Therefore, our study underscores that proteolytic pathways represent a cross-tissue characteristic of gene expression reconfiguration during aging, providing additional insights into the decline in proteolysis maintenance associated with aging.

Additionally, our findings indicate that the attrition of gene–gene transcriptional coordination during aging results from a progressive imbalance in gene expression dosage, which may have adverse health effects and increase susceptibility to certain diseases. This relationship is particularly evident in the small intestine, the primary organ responsible for nutrient absorption. We found that the loss of gene coordination between CDIPT, PI4KA, and PI4K2B is a distinguishing feature of aging in the small intestine, associated with a broad spectrum of diseases. This loss is primarily due to the age-related downregulation of PI4KA. Within phospholipid metabolism pathways, CDIPT converts myo-inositol to phosphatidylinositol, which is further transformed into phosphatidylinositol 4-phosphate (PI(4)P) by enzymes including PI4KA (Fig. 4L). Recent study indicates that functional attrition of PI4KA can lead to epithelial defects and inflammation in the small intestine[55], suggesting that the gene expression coordination loss in the aged small intestine may propagate the pathogenic effect via ways like inflammation and nutrient absorption dysregulation [56].

Conclusions

In this study, we construct individual-level gene expression networks throughout the aging process in a large human cohort. Our findings reveal a cross-tissue pattern of declining gene-to-gene expression coordination and underscore the potential risks associated with the deterioration of such coordination.

Limitation

We acknowledge that our study has certain limitations. Many of the findings, such as the association between lost coordinated relationships and donor health, and the MR results linking core module genes and longevity, remain largely correlational. Cause relationships still need to be further disentangled. The study primarily focuses on protein-coding genes, thus limiting our examination to non-coding RNAs that may also play important roles in expression regulation. Additionally, our study mainly focused on the loss of coordinated expression relationships during aging. Whether gains in coordinated expression relationships impact aging and health remains unclear and warrants further exploration. 

Abbreviations

ARLR

Aging-related lost relationship

CER

Coordinated expression relationship

GTEx

Genotype-Tissue Expression

LRAge

Lost relationship age

mtCN

Mitochondrion copy number

MR

Mendelian randomization

PEER

Probabilistic estimation of expression residuals

SASP

Senescence-associated secretory phenotype

s-LDSC

Stratified LD score regression

SSN

Sample-specific network

TPM

Transcripts per million

Authors’ contributions

Q.-P.K., F.-H.X. and H.-T.W. conceived and designed this study. H.-T.W. and F.-H.X. analyzed the data. L.Z., Q.S., T.-R.X., L.-Q.Y. and S.-Y.M. assisted the data analysis. Q.-P.K., F.-H.X. and H.-T.W. had access to the data. F.-H.X. and H.-T.W. verified the data. H.-T.W., F.-H.X. and Q.-P.K. wrote the draft. Q.-P.K. and F.-H.X. supervised this study. All authors read and approved the final manuscript.

Funding

This work was supported by grants from the National Key R&D Program of China (2023YFC3603400), the National Natural Science Foundation of China (82430049, 82401833, 82371580, 82071595), the CAS Project for Young Scientists in Basic Research (YSBR-076) and the West Light Foundation (to F.-H.X.) of the Chinese Academy of Sciences, the Yunnan Fundamental Research Projects (202201AS070080, 202401AW070011, 202301AT070281), High level Talent Promotion and Training Project of Kunming (Spring City Plan; 2020SCP001), the Yunnan Revitalization Talent Support Program Yunling Scholar Project (to Q.-P.K.), and the Reserve Talent Project of Young and Middle-aged Academic and Technical Leaders in Yunnan Province (202305AC160029). The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the manuscript.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.