Abstract

   Aging is the primary risk factor for many cancer types, including lung
   adenocarcinoma (LUAD). To understand how aging-related alterations in
   the regulation of key cellular processes might affect LUAD risk and
   survival, we built individual-specific gene regulatory networks
   integrating gene expression, transcription factor protein-protein
   interaction, and sequence motif data, using PANDA/LIONESS algorithms,
   for non-cancerous lung samples from GTEx project and LUAD samples from
   TCGA. In healthy lung, pathways involved in cell proliferation and
   immune response were increasingly targeted with age; these
   aging-associated alterations were accelerated by smoking and resembled
   oncogenic shifts observed in LUAD. Aging-associated genes showed
   greater aging-biased targeting patterns in individuals with LUAD
   compared to healthier counterparts, a pattern suggestive of age
   acceleration. Using drug repurposing tool CLUEreg, we found small
   molecule drugs that may potentially alter the accelerating aging
   profiles we found. We defined a network-informed aging signature that
   was associated with survival in LUAD.

   graphic file with name 41514_2025_247_Figa_HTML.jpg

   Subject terms: Cancer, Genetics

Introduction

   Lung cancer is second only to breast cancer worldwide in annual
   incidence and is the leading cause of cancer death. Lung cancer risk
   increases with age and as the average age of the population increases
   worldwide, the prevalence of lung cancer is expected to continue
   growing^[37]1. In 2021, 75% of lung cancer fatalities were reported in
   individuals aged 65 and older^[38]2. While lung adenocarcinoma (LUAD)
   in younger adults is often diagnosed at more advanced stages compared
   to those in older adults^[39]3, elderly individuals have more
   comorbidities and tend to be less tolerant of certain cancer
   therapeutics than younger individuals^[40]4. These differences are
   likely the result of aging-induced alterations in the regulation of key
   cellular processes^[41]5, but the mechanism by which age shifts the
   gene regulatory landscape to alter lung cancer risk and survival
   outcome is largely unknown. In this paper we address this critical gap
   in our understanding by building individual (person)-specific gene
   regulatory networks to gain insights into aging related changes in gene
   regulation that might influence the risk and prognosis of LUAD across
   all age groups. Additionally, we explore how aging-associated
   regulatory changes are further accelerated by tobacco smoking history,
   since lung diseases including LUAD are more prevalent among individuals
   with a history of smoking, compared to individuals who have never
   smoked in their lifetime^[42]6.

   Transcription factors (TF) have been established as known drivers of
   aging and cancer^[43]7. In recent research^[44]8 involving mouse
   single-cell RNA-sequencing atlas, it has been shown that within immune
   cells from different organs including lung, TF activity changes with
   age. This pattern was particularly dominant in tumor-associated
   macrophages. Aging-associated TF dysregulation was also observed in
   antigen processing and inflammation, both of which are key hallmarks in
   many age-related diseases including cancers. To understand
   aging-associated heterogeneity in the gene regulatory landscape of
   LUAD, we identified biological pathways that are differentially
   regulated by TFs in tumors from individuals of different ages and
   investigated if any of these age-associated changes in regulatory
   networks might influence survival and the response to chemotherapy,
   potentially favoring individuals exhibiting regulatory signatures akin
   to those found in younger individuals. We further validated our
   findings in two independent datasets of non-cancerous lung tissue and
   LUAD.

   Most studies investigating the role of aging in LUAD have focused on
   the mutational landscape of tumor among individuals across various age
   groups^[45]9,[46]10. Tumor mutations in several genes, including
   CDKN2A, KRAS, MDM2, MET, and PIK3CA, have been found to increase in
   frequency with the individual’s age, while the frequencies of mutation
   in ALK, ROS1, RET and ERBB2 show a decreasing trend with age^[47]11.
   ALK and EGFR mutations are high among younger individuals with LUAD,
   especially among females and nonsmokers^[48]10,[49]12,[50]13. Analysis
   of somatic interactions has indicated that EGFR-positive samples in
   younger individuals are more prone to concurrent mutations in PIK3CA,
   MET, TP53, and RB1 when compared to older individuals^[51]11. Age may
   influence both the number of mutations in a tumor and their
   evolutionary timing^[52]14. While germline mutations are more commonly
   identified in tumors from younger individuals, tumors in older
   individuals appear to be predominantly influenced by somatic
   mutations^[53]15. Such mutations clearly play a role in cancer risk and
   prognosis, acting in part, by altering the activity of biological
   pathways associated with cancer. However, changes in these pathways can
   only be partially explained by known mutations, indicating that other
   mechanisms of pathway activation might play a significant role^[54]16.

   Despite some studies in lung cancer that have reported altered
   expression of genes linked to survival^[55]17,[56]18, there has been
   limited research investigating aging-associated alterations in gene
   regulatory networks that influence the risk and prognosis of LUAD and
   lung cancer in general. We addressed this gap in our understanding by
   using the network-modeling approaches, PANDA^[57]19 and LIONESS^[58]20,
   to derive person-specific gene regulatory networks for non-cancerous
   lung tissue samples from the Genotype Tissue Expression project (GTEx)
   and LUAD tumor samples from The Cancer Genome Atlas (TCGA), with a
   focus on evaluating age-associated genes and their regulation by TFs.
   This approach was motivated by multiple earlier network-modeling
   analyses that identified disease relevant regulatory features in both
   healthy tissues as well as in tumor^[59]21–[60]24.

   PANDA (Passing Attributes between Networks for Data Assimilation) is a
   computational method to synthesize three sources of information about
   the regulatory process and infer condition-specific regulatory networks
   connecting regulatory TFs and the genes they control. PANDA starts with
   a “motif prior” based on mapping TFs to likely promotor regions of the
   genome to establish candidate TF-gene regulatory connections, uses
   pairwise correlations between gene expression levels as a measure of
   possible co-regulation by the same TF, considers TF-TF protein-protein
   interaction (PPI) data as a sign of regulatory cooperation, and then
   uses a machine learning “message passing” framework to iteratively
   update the weights of regulatory edges, finally producing a bipartite
   network of directed edges from TFs to target genes. LIONESS (Linear
   Interpolation to Obtain Network Estimates for Single Samples) builds
   upon PANDA by enabling estimation of sample-specific gene regulatory
   networks. The conceptual framework underlying LIONESS is that a gene
   regulatory network estimated on a population is an average over the
   networks for individuals in that population such that removing an
   individual perturbs that average estimate in a way that is, to first
   order, linear in the network edge weights. LIONESS begins with a PANDA
   network for the entire population. LIONESS then systematically removes
   each sample, one at a time, reconstructing the PANDA network for the
   remaining data, and uses linear interpolation between the gene
   regulatory networks with and without that sample to estimate the
   network for the left-out individual; the process repeats iteratively
   until a PANDA network has been generated for each member of the
   starting population.

   By analyzing individual-specific regulatory networks, we found
   increased TF targeting of pathways related to intracellular adhesion,
   cell proliferation, and immune response with age in healthy lung
   tissue. These aging-associated alterations are further increased by
   tobacco smoking and resemble oncogenic shifts in the regulatory
   landscape observed in LUAD, thereby suggesting a potential association
   between aging-associated dysregulation of biological pathways and an
   elevated risk of developing LUAD.

   Using a web-based drug repurposing tool CLUEreg, we also found
   potential geroprotective small molecule drug candidates that may be
   useful in reducing the risk of LUAD by reversing the aging-associated
   regulatory signatures. By constructing a network-informed aging
   signature for tumor samples based on the TF-targeting patterns of key
   biological pathways significantly changing with age in LUAD, we found
   that a lower aging signature is associated with better survival
   probability. In contrast, chronological age was not predictive of
   survival, thus demonstrating that the aging signature captures aspects
   of tumor biology not captured by chronological age alone. Using
   CLUEreg, we also found distinct small molecule drug candidates tailored
   to LUAD samples with varying aging signatures. In conclusion, our
   findings not only highlight the mechanisms underlying increased risk
   and poorer prognosis of LUAD associated with aging-induced gene
   regulatory alterations but also establish a potential avenue for
   leveraging individual-specific gene regulatory networks in designing
   personalized therapeutic interventions.

Results

Identifying aging-associated gene regulatory alterations in healthy human
lung and geropropective drug candidates

   We used the PANDA + LIONESS algorithm to infer individual-specific gene
   regulatory networks using gene expression data from non-cancerous lung
   tissue samples from GTEx (Fig. [61]1), For every individual, the
   algorithm infers a bipartite gene regulatory network linking TFs to
   their target genes within that individual. By analyzing these networks
   across samples, we identified several genes (Fig. [62]2a) that are
   differentially targeted by TFs as a function of age, in the lung. Among
   these genes, there are 1018 that exhibit significantly increased
   targeting by TFs with age (p < 0.05). Most significant among them are
   NNAT^[63]25, FBLN7^[64]26, SH3BP1^[65]27, CNTN1^[66]28, THEM5^[67]29,
   and FOXP4^[68]30; upregulation of these genes has been previously
   reported to be associated with cell proliferation and poorer prognosis
   in multiple cancers. We also find 404 genes that exhibit significantly
   decreased targeting by TFs with age (p < 0.05) including DUSP15^[69]31,
   ALDH1L2^[70]32, HPD^[71]33, GSTT2^[72]34, FOXI3^[73]35, and
   ZIC2^[74]36, all of which have been shown to be influential in
   predicting tumor progression and therapeutic efficacy in various cancer
   types, including non-small cell lung cancer.

Fig. 1. Schematic overview of the study.

   [75]Fig. 1
   [76]Open in a new tab

   Top box, overview of our approach to constructing individual specific
   networks using PANDA and LIONESS which integrate information on
   protein-protein interactions (PPIs) between transcription factors
   (TFs), prior information on TF-Gene motif binding, and gene expression
   data – in this case, from GTEx lung tissues and TCGA LUAD primary tumor
   samples - downloaded from Recount3. Bottom box, overview of the
   differential targeting analysis.

Fig. 2. Change in TF-targeting of genes with age.

   [77]Fig. 2
   [78]Open in a new tab

   a Left: Volcano plot of genes that are differentially (increasingly or
   decreasingly) targeted by TFs over varying age in lung tissue samples
   from GTEx. The x-axis represents log fold change (logFC), which is
   defined as the change in gene indegree in response to a unit change in
   age. The y-axis represents negative of logarithm of p values
   (−log10(P.Value)). b Right: Boxplot of the t-statistics of the age
   coefficient associated with the oncogenes and tumor suppressor genes
   (listed in the COSMIC database) from the limma analysis in GTEx.
   Positive value means these genes are targeted more with age on average
   and negative value means these genes are targeted less with age on
   average. For comparison, we also show the same t-statistics for
   non-cancer genes (genes that are not annotated as oncogenes and/or
   tumor suppressor genes in the COSMIC database). Each boxplot ranges
   from the upper and lower quartiles with the median as the horizontal
   line. Outliers are marked by points. In our analysis we include genes
   that are explicitly marked as either “Oncogene” or “TSG” respectively
   in the COSMIC database, thus excluding all the genes that can work
   either as an oncogene or as a tumor suppressor gene, depending on the
   mutation. The reported p values correspond to the hypothesis testing
   with respect to alternative hypotheses reported in parentheses. The
   alternative hypotheses “<” or “>” denote the hypotheses “mean > 0” and
   “mean < 0” respectively.

   We compared the list of aging-related genes with a known list of
   cancer-associated proto-oncogenes and tumor suppressor genes downloaded
   from the Catalog of Somatic Mutations in Cancer (COSMIC)^[79]37. Among
   genes increasingly targeted with age, there was only 1 proto-oncogene
   XPO1 and among genes decreasingly targeted with age, there were 4 tumor
   suppressor genes (TSGs) DICER1, PIK3R1, POT1 and PTEN.

   We further examined the network topology surrounding all cancer-related
   genes annotated in the COSMIC database by analyzing gene targeting
   score in relation to age. Specifically, we compared the t-statistics of
   the age coefficients (normalized age gradients) for TF targeting of
   oncogenes and TSGs (Fig. [80]2b) and found that among the healthy
   samples in GTEx, TF-targeting increases with age on average for both
   oncogenes and TSGs (Wilcoxon rank sum test gives a p value of <4e-09
   for oncogenes and 0.001 for TSGs).

   For comparison, it should be noted that, non-cancer genes (that is,
   genes not annotated as oncogene or TSG in the COSMIC database) are also
   increasingly targeted by TFs with age (p value of Wilcoxon rank sum
   test is 2.2e-16), meaning that the gene regulatory networks inferred
   for individuals in GTEx, increase in TF regulatory density as the age
   of the individual increases. Nevertheless, the mean TF-targeting with
   aging is the greatest for oncogenes (aging slope of oncogenes is
   significantly larger than those of non-cancer genes and tumor
   suppressor genes with p values being equal to 0.001 and <2.2e-16,
   respectively). This indicates that although changes in regulation are a
   natural consequence of aging, the greatest changes occur in the
   regulatory neighborhoods of oncogenes, including genes that are common
   drug targets^[81]38 in LUAD, such as MYCN, ERBB3, and AKT1 (Fig.
   S.[82]1).

   We performed gene set enrichment analysis (GSEA) with genes, ranked by
   how much their targeting patterns change with age, and found (Fig.
   [83]3 leftmost column) that biological pathways associated with
   intracellular adhesion and cell proliferation, cell growth, and death
   have increasing TF targeting with age, including pathways annotated to
   adherens junction, apoptosis, hematopoietic cell lineage, cell adhesion
   molecules, and pathways in cancer. Pathways associated with immune
   response, including B-cell receptor signaling pathway,
   cytokine-cytokine receptor interaction, chemokine signaling pathway and
   intestinal immune network for IgA production, also show increased TF
   targeting with age. We confirmed these findings on an independent
   dataset LGRC (Fig. [84]3 middle column).

Fig. 3. Aging-related changes in pathway targeting by transcription factors
is similar to those as in lung adenocarcinoma.

   [85]Fig. 3
   [86]Open in a new tab

   Heatmap of normalized enrichment scores (NES) for pathways that are
   significantly (at FDR cutoff 0.05) differentially targeted by
   transcription factors with age among non-cancerous lung samples (GTEx).
   The first two columns exhibit NES from GSEA based on the age
   coefficients from the limma analysis of GTEx and LGRC. The third column
   shows NES from GSEA based on difference between tumor samples from TCGA
   and healthy samples from GTEx.

   Using genes that are differentially targeted by age as input (1018
   increasingly targeted and 404 decreasingly targeted by TFs with age) to
   a web-based drug repurposing tool CLUEreg^[87]39, we identified small
   molecule drug candidates (Supplementary Data [88]S1) with potential to
   reverse the aging-associated regulatory alterations in the gene
   regulatory networks from non-cancerous lung samples. CLUEreg compares
   these lists to the library of 19,791 drug-specific gene regulatory
   networks to identify drug candidates that are likely to reverse the
   targeting score of the gene in the input list. For each drug, CLUEreg
   computes a p value by resampling 10,000 random inputs of varying
   lengths and using them to construct a null distribution. For our given
   input gene lists, CLUEreg gave a list of 150 drugs as output. Among
   these drugs, we selected 112 drugs with a p < 0.05 as our candidate
   drugs with the potential to reverse gene regulatory signatures of
   aging. Some of these drugs, henceforth referred to as “geroprotective
   drugs”, including Carnosol^[89]40, Curcumin^[90]41, Cucurbitacin
   B^[91]42, Isonicotinamide^[92]43, Meclofenoxate^[93]44,
   Scriptaid^[94]45, and Withaferin A^[95]46 have already been shown to
   have potential geroprotective effects in various animal models,
   including humans. Among these geroprotective drug candidates, we found
   several FDA-approved anti-cancer drugs, including Trametinib,
   Doxorubicin, Alisertib, Actinomycin-d, Toremifene, and Plumbagin, as
   well as several investigational drugs with potential anti-tumor effects
   including Avrainvillamide analogs^[96]47, aurora kinase inhibitors
   (MK-5108, AT-9283)^[97]48, Avicins^[98]49, HMN-214^[99]50,
   Chaetocin^[100]51, ron kinase inhibitors^[101]52, and
   Linifanib^[102]53, among others.

Tobacco smoking is associated with accelerated aging

   To explore whether tobacco smoking is associated with an acceleration
   of the aging process, we compared the gene regulatory networks from
   non-cancerous samples, between individuals with a history of smoking
   and individuals who have never smoked in their lifetime. We split the
   1422 aging-associated genes we had previously identified into two sets:
   genes that exhibit increasing TF targeting with age (1018 genes) and
   those that show decreasing TF targeting (404 genes). For every gene in
   these two sets, we used limma^[103]54, to compute the age coefficient
   (Fig. [104]4) in a linear model for individuals with and without a
   history of smoking.

Fig. 4. Aging-related transcription factor-targeting of genes by smoking
history.

   [105]Fig. 4
   [106]Open in a new tab

   Boxplot of the rates of change in TF-targeting with age in GTEx
   (designated by the t-statistics from the limma analysis with
   interaction between age and smoking status) a Left: for 1018 genes that
   are increasingly targeted with age in healthy human lung (based on
   evidence from GTEx) and b Right: the same boxplot for 404 genes that
   are decreasingly targeted with age in healthy human lung (based on
   evidence from GTEx). “Ever” refers to the group of individuals with a
   prior history (both past and current) of tobacco smoking and “Never”
   refers to the group of individuals without a prior history of smoking.

   We found that for genes with increasing age-associated TF targeting,
   the t-statistics of the age coefficients among individuals with a
   history of smoking have significantly greater positive values than
   among never-smokers (p < 2.2e-16). Similarly, for genes with decreasing
   TF-targeting with age, the t-statistic of the age coefficients among
   individuals with a history of smoking have significantly (p < 2.2e-16)
   larger negative values than those among individuals who have never
   smoked. In other words, for both kinds of aging-associated genes, the
   age gradients are significantly steeper for the individuals with a
   history of smoking, than those for the individuals who have never
   smoked in their lifetime. The steeper age gradients mean that the rate
   of change in gene regulation with age is faster among individuals with
   a history of smoking, comparted to individuals who have never smoked.

   To visually represent the continuous changes in gene regulation with
   age, we plotted the aging trajectories (as described in “Methods”) for
   the non-cancerous samples from GTEx (Fig. S.[107]2 in the
   [108]Supplementary Material). For the genes which are increasingly
   targeted by TFs with age (left plot on Fig. [109]S2), the slope of the
   regression line is steeper among individuals who have a history of
   smoking, compared to individuals who have never smoked. Although for
   genes that are decreasingly targeted by TFs with age (right plot of
   Fig. [110]S2), we do not observe any significant difference between the
   slopes of the aging trajectories among individuals with or without a
   history of smoking. Nevertheless, these findings indicate that tobacco
   smoking is associated with an acceleration of the aging-induced
   alterations in gene regulation. We also validated these findings in the
   non-cancerous lung samples from the independent LGRC (a.k.a.
   [111]GSE47460) validation dataset (Fig. [112]S3).

   Taken together, we find that even in individuals without evidence of
   lung cancer, there are aging-associated changes in the TF-targeting of
   genes that are further accelerated by tobacco smoking, which may be
   linked to an increased risk of developing LUAD at a younger age among
   individuals with a history of smoking.

Aging-associated gene regulatory alterations in non-cancerous lung resemble
oncogenic gene regulatory shifts observed in LUAD

   To explore the association of aging with LUAD risk, we compared the
   targeting patterns of aging-associated biological pathways between
   non-cancerous samples from GTEx and LUAD samples from TCGA. We observe
   that aging-associated pathways involved in cell adhesion, cell
   proliferation and immune response (except for type I diabetes mellitus
   and allograft rejection pathways), are also highly targeted in LUAD
   tumor, compared to non-cancerous lung (Fig. [113]3 rightmost column).
   This indicates that increased TF-targeting of these pathways with age
   might be a contributing factor to an elevated risk of developing LUAD
   among older adults and that those with the greatest regulatory
   targeting of these pathways might be at the greatest risk.

   LUAD among younger individuals, although less frequent, is often
   detected at more advanced stages compared to their older
   counterparts^[114]3. Given that we found age-acceleration of gene
   targeting to correlate with LUAD, we tested whether LUAD in younger
   individuals was also associated with patterns of accelerated aging,
   compared to healthier individuals of similar age. To confirm this
   hypothesis, we compared the TF-targeting pattern of 1422
   aging-associated genes in normal adjacent lung samples from individuals
   with LUAD (from TCGA) versus non-cancerous lung samples from GTEx (Fig.
   [115]5). For the 1018 genes that exhibited increased targeting with
   age, we found that their mean TF-targeting was significantly higher
   (p < 2.2e-16) in the normal adjacent lung samples from younger
   individuals (age less than median age 66) with LUAD, compared to
   individuals of similar age without LUAD (Fig. [116]5 left). In
   contrast, for older individuals with LUAD (age greater than median age
   66) we did not find a significantly higher mean targeting of aging
   genes compared to healthy individuals of similar age.

Fig. 5. Individuals with lung adenocarcinoma show transcription factor
targeting patterns indicative of accelerated aging, even in normal lung
tissue, compared to those without the disease.

   [117]Fig. 5
   [118]Open in a new tab

   Boxplot of the difference in transcription factor-targeting of genes in
   TCGA normal adjacent samples compared to GTEx non-cancerous lung for a
   Left: 1018 genes that are increasingly targeted with age in healthy
   human lung (based on evidence from GTEx) and b Right: for 404 genes
   that are decreasingly targeted with age in healthy human lung (based on
   evidence from GTEx). The reported p values correspond to the hypothesis
   testing with respect to alternative hypotheses reported in parentheses.
   The alternative hypotheses “<” or “>” denote the hypotheses “mean > 0”
   and “mean < 0” respectively.

   In other words, the gene regulatory patterns observed in the
   normal-adjacent lung tissues of younger individuals with LUAD are more
   like those found in older individuals, than they are to their healthy
   counterparts of the same age. This suggests that LUAD in younger
   individuals may be driven, in part, by age-accelerated changes in gene
   regulation, and that this acceleration may also be associated with more
   aggressive tumor biology at diagnosis we see in younger individuals.

Biological pathways differentially regulated in LUAD tumor across varying age

   We analyzed gene regulatory networks of LUAD samples from TCGA and
   performed GSEA to identify biological pathways that are differentially
   regulated by TFs with age (Fig. [119]6 left column).

Fig. 6. Transcription-factor targeting patterns of genes in lung
adenocarcinoma tumor samples vary by age.

   [120]Fig. 6
   [121]Open in a new tab

   Normalized enrichment scores (NES) of the biological pathways that are
   significantly (at an FDR cutoff 0.05) differentially targeted with age
   by transcription factors in tumor samples from TCGA. Left column shows
   NES from GSEA on TCGA samples, and the right column shows NES for the
   same pathways from GSEA on [122]GSE68465 samples.

   We found several pathways involved in cell signaling and cell
   proliferation that were increasingly targeted by TFs with age. When we
   compared these differentially targeted pathways to those differentially
   targeted with age in non-cancerous lung samples from GTEx, we found
   that both had identified the pathway associated with cell adhesion
   molecules.

   However, there were many pathways with aging-associated regulatory
   changes found exclusively in tumor samples and not in healthy lung
   samples, including the NOD-like receptor signaling pathway, FC-epsilon
   RI signaling pathway, toll-like receptor signaling pathway and JAK-STAT
   signaling pathway—all of which have been associated with LUAD
   development, progression and outcome. In contrast, we found that
   metabolic pathways including oxidative phosphorylation, nitrogen
   metabolism, arginine and proline metabolism, ascorbate and alderate
   metabolism were decreasingly targeted by TFs with age. These results
   were validated in an independent dataset [123]GSE68465 (Fig. [124]6
   right column).

   Biological pathways associated with immune response that had been
   previously identified as increasingly targeted (at an FDR cutoff 0.05)
   with age in non-cancerous lung samples, also showed increased targeting
   by TFs with age in tumors. We also found several additional
   immune-related pathways to have age-dependent regulatory patterns in
   tumor, that were not evident in non-cancerous samples. Such
   immune-related pathways include those involved in antigen processing
   and presentation, graft versus host disease, JAK-STAT signaling
   pathway, natural killer cell mediated cytotoxicity, primary
   immunodeficiency and T-cell receptor signaling pathway, all of which
   were increasingly targeting by TFs with age. Aging-related gene
   regulatory dysregulation in antigen processing was previously observed
   in murine models^[125]8.

   This difference between non-cancerous lung versus LUAD tumor in the
   age-associated targeting patterns of immune pathways can be partially
   attributed to differences in immune cell infiltration by age in
   non-cancerous lung tissue as compared to tumor. Immune infiltration
   analysis (Fig. [126]S4) showed that the proportion of CD8+ central
   memory cells increased with age in both non-cancerous lungs, and in
   tumor. However, aging-related changes in immune cell composition in
   tumor were distinct from those in healthy lungs for most immune cell
   types. For example, while the proportions of activated myeloid
   dendritic cells and B cells increased with age in LUAD tumors, in
   non-cancerous lung the proportion of these cells did not change
   significantly with age. The proportion of macrophages also increased
   with age in tumor; in non-cancerous lung, macrophages were more
   abundant in samples from younger individuals. This higher infiltration
   of immune cells in tumor with age might be associated with a higher
   targeting of immune pathways among older individuals with LUAD, as
   evidenced by the positive correlation between immune score and
   TF-targeting score of immune pathways (Table S.[127]1). In contrast,
   the proportion of CD8+ naïve T-cells and common myeloid progenitors
   showed a decreasing trend with age in tumor, while exhibiting no
   significant difference in composition across healthy lung samples of
   varying age (Fig. [128]S4)

Gene regulatory network-informed aging signature of tumor predicts survival
in LUAD

   We conducted survival analysis using the Cox proportional hazard model
   to understand whether the aging-associated regulatory patterns of
   biological pathways have any influence on the prognosis of LUAD for
   individuals of varying age. First, using randomly chosen 50% of the
   TCGA samples as training data, we constructed an “aging signature” for
   tumor samples as follows: we fitted a penalized Cox proportional hazard
   model with LASSO penalty, using the targeting score (as defined in
   “Methods”) of 28 biological pathways (Fig. [129]6) as input. These
   pathways were earlier discovered to be significantly differentially
   targeted by TFs as a function of age in tumors. The penalized model
   selected 4 out of 28 of these pathways as significantly predictive of
   survival outcome in the training data: pathways associated to asthma,
   nitrogen metabolism, oxidative phosphorylation and ribosome. We defined
   biological aging signature as the predicted value from this cox
   regression model. This network-informed “aging signature” of tumors is
   a linear combination of the TF-targeting scores of these 4 biological
   pathways and is uncorrelated with both chronological age
   (correlation = 0.021 with p = 0.7397) and clinical tumor stage (p value
   from ANOVA = 0.842).

   We implemented the fitted penalized Cox regression model on the TCGA
   test data and the [130]GSE68465 independent validation dataset to
   compute the biological aging signature for each sample in these
   datasets. After dividing all test samples into two parts depending on
   if their aging signature value is higher or lower than the median
   across all samples, we found that samples with a lower aging signature
   had significantly better survival probability than samples with a
   higher aging signature (Fig. [131]7 left; p = 0.015 from Kaplan-Meier
   survival model). For comparison, we split these same samples into two
   chronological subgroups based on whether samples were above or below
   the median chronological age (Fig. [132]7 right) and not find any
   significant difference (p = 0.147) in survival probability. This
   indicates that the aging signature, defined using gene regulatory
   networks is more informative than chronological age in predicting LUAD
   survival in TCGA. The results remained consistent even after adjusting
   for self-reported gender, race, smoking status, clinical tumor stage
   and therapy status (p values for the network-informed aging signature
   and chronological age were 0.049 and 0.095, respectively). In the
   validation dataset [133]GSE68465 also (Fig. [134]S5), we observed that
   individuals with lower biological aging signature had better survival
   outcome than individuals with higher biological aging signature,
   however the difference was not statistically significant (p = 0.124).

Fig. 7. The network-informed aging signature is more informative about
survival probability in lung adenocarcinoma (LUAD), than chronological age.

   [135]Fig. 7
   [136]Open in a new tab

   Kaplan-Meier plot for survival outcome in TCGA, a Left: split by median
   network-informed aging signature and b Right: by median chronological
   age into younger and older individuals with LUAD.

Drug repurposing with CLUEreg identifies distinct small molecule drugs for
tumor samples with lower versus higher aging signature

   To find potential targeted cancer therapeutics that might differ in
   efficacy depending on aging signatures, we split the TCGA tumor samples
   into two groups – one above and one below the median value of the
   network-informed aging signature. For each group, we separately used
   CLUEreg^[137]39 to identify small molecule drug candidates depending on
   the differential regulatory patterns between tumor and healthy samples
   and obtained a list of 150 small molecule drugs each for the two aging
   signature groups (Fig. [138]S6).

   While 59 small-molecule cancer therapeutics including FDA-approved
   Cisplatin and Amifostine and investigational drugs Timosaponin and
   Cardamonin appeared as potential drug candidates for both
   aging-signature groups, several other drugs appeared exclusively in
   only one of the two groups. Drugs including Homoharringtonine, Ingenol,
   Vatalanib (investigational), Midostaurin (investigational) and Ubenimex
   (investigational) appeared only for individuals with lower aging
   signature. Other drugs including Leucovorin, Actinomycin-d and
   Plumbagin appeared for the higher aging signature group alone. Several
   potential geroprotective drug candidates, including Meclofenoxate and
   Isonicotinamide, also appear in the lists of anti-tumor drugs. It is
   worth noting that we found 28 geroprotective drugs in the list for
   higher aging signature group while only 5 geroprotective drugs for the
   lower aging signature group. This suggests that the information
   captured by the aging signature encompasses disease-relevant processes
   driving LUAD, which are intertwined with other aging-related processes
   that also contribute to the development and progression of the disease.

Discussion

   LUAD, like most other solid tumors, is an age-biased disease in which
   individuals generally have a greater risk, poorer prognosis and poorer
   response to most therapies compared to their younger
   counterparts^[139]4. However, tumors diagnosed in younger individuals
   are often detected at more advanced clinical stages implying more
   severe disease biology. Earlier studies have found that across diverse
   cancer types, the mutational landscape of tumors in younger individuals
   is very different from that in older individuals^[140]11. However,
   mutational burden alone does not fully explain the mechanisms of
   disease, driven by the activity of biological pathways that are
   activated leading up to and during disease. To bridge this gap in our
   understanding of LUAD, we investigated how regulation of various genes
   and biological pathways change with age using individual specific gene
   regulatory networks. Analyzing gene regulatory networks that integrate
   gene expression, TF-binding motif and TF PPI data from both
   non-cancerous human lung and LUAD samples, we found aging-associated
   alterations in regulatory mechanisms involving key biological
   processes.

   Analyzing gene regulatory networks in GTEx lung samples, we found
   several genes with known relevance to cancer incidence and prognosis
   that were differentially targeted by TFs with age, including
   proto-oncogenes AKT1, ERBB3 and MYCN. Using pathway enrichment
   analysis, we saw a clearer picture of how age affects cancer-related
   processes in non-cancerous lung tissue, including altered regulation of
   intracellular adhesion, cell proliferation, and immune response. By
   conducting a differential targeting analysis on non-cancerous lung
   samples and LUAD samples, we confirmed that targeting of these same
   biological processes also changes in tumor and in the same direction as
   they do in “normal” aging. This suggests that aging-associated
   alterations in gene regulatory patterns of these pathways might be a
   contributing factor to a higher risk of LUAD development among older
   individuals. Further, we found that tobacco smoking was associated with
   an acceleration in the aging-associated gene regulatory changes,
   helping to explain the increased risk of LUAD incidence at a younger
   age among individuals with a history of smoking.

   Gene regulatory network analysis of LUAD identified an age-associated
   higher targeting of several biological pathways associated with immune
   response, these associations were not detected in non-cancerous lung
   samples from GTEx. Greater targeting of immune pathways with age was
   correlated with a higher infiltration of immune cells including active
   myeloid progenitors, B cells, macrophages and CD8+ central memory
   cells. It is worth noting that an increased infiltration of CD8+
   central memory cells with age has been previously observed in
   peripheral blood, indicating that this pattern is not unique to lung
   tissues, but rather a systematic effect prevalent in other tissues as
   well^[141]55. We suspect that a higher targeting of immune pathways in
   conjunction with a higher proportion of immune cells among older
   individuals might contribute to an age-biased response to
   immunotherapy. This is concordant with evidence from earlier studies
   which demonstrated that while chemotherapy is more beneficial for
   younger individuals^[142]5, some immune checkpoint inhibitors provide
   greater benefit to adults with age 65 or older, compared to younger
   adults^[143]56,[144]57.

   We constructed a network-informed “aging signature” for tumor samples,
   based on the TF-targeting patterns of 4 biological pathways (identified
   in TCGA and validated in [145]GSE68465) that exhibited significant
   differential targeting by TFs with age and impacted survival outcome in
   the training data. We found that individuals with a lower aging
   signature had better survival outcome compared to those with a higher
   aging signature in both the test data and validation data. In fact,
   within TCGA, this biological aging signature had better efficacy in
   predicting survival, compared to the chronological age at the time of
   tumor diagnosis.

   The consistent theme of age-associated alterations in regulation being
   linked to LUAD suggested that network-based aging signatures might
   identify aging-related tailored therapeutics. Separately analyzing LUAD
   samples partitioned into low-and-high-aging signature groups, we found
   59 small-molecule cancer therapeutics, including FDA-approved Cisplatin
   and Amifostine, common to both aging-signature groups. But we also
   found several drugs were exclusive to one aging signature group alone,
   meaning that considering age-related regulatory changes might be useful
   in determining personalized therapeutic protocols. Certain potential
   geroprotective drugs including Meclofenoxate and Isonicotinamide
   appeared in the lists of anti-tumor drugs, mostly for the higher aging
   signature group, as did a number of candidate drugs such as
   Curcumin^[146]41, that have been shown have geroprotective effects.
   Unfortunately, older adults are severely underrepresented in clinical
   trials for most cancers including LUAD, thereby impacting the validity
   of clinical guidelines in diseases with an age effect^[147]58. Our
   analysis underscores the importance of including individuals across the
   spectrum of disease-associated demographics in clinical trials.

   It is important to note that our analysis is based on observational
   data alone and hence experimental validation is required to establish a
   causal relationship between the aging-associated regulatory changes
   identified by our analysis and the manifestation of tumor. Another
   limitation of our study is that the datasets used for discovery (GTEx
   and TCGA) and validation primarily consist of individuals of white and
   African American descent. Despite adjusting for the impact of race in
   our analysis, the applicability of our findings to other ethnicities
   may still be constrained due to underrepresentation in our data and the
   confounding effects of various social determinants of health on lung
   cancer. Further studies involving more diverse populations are
   necessary to confirm the validity of our results across a broader range
   of racial and ethnic backgrounds. Additionally, a more complete
   inclusion of social determinants such as individual socio-economic
   background^[148]59, is essential for the generalizability of our
   findings.

   Despite these limitations, our analysis provides interesting insights
   into the aging-associated alterations in gene regulation and their
   relevance in the clinical manifestation of LUAD, including some
   immediate implications in the context of personalized cancer
   therapy^[149]60. Based on our analysis we infer that aging related
   changes in regulation of key biological processes involved in
   intra-cellular adhesion, cell proliferation and immune responses are
   associated with altered risk, prognosis and response to therapy in LUAD
   among individuals of varying age. Notably, we observed that even among
   individuals of similar age, individuals with lower network-informed
   aging signature had better prognostic outcome, than individuals with
   higher aging signature. This observation implies that chronological age
   alone does not provide substantial information on prescribing
   personalized therapy for LUAD and gene regulatory networks can prove to
   be effective tools in facilitating more efficient personalized therapy
   design and improving prognosis in LUAD for individuals across varying
   age.

   What emerges from our analysis is an interesting picture of how aging
   influences LUAD. “Normal” aging in the lung is associated with
   alterations in the regulation of particular biological processes, and
   indeed, by inferring and analyzing gene regulatory network structure,
   we identified genes and biological processes that exhibit altered
   patterns of regulation with age. But as has been known, not all
   individuals age at the same rate. When we examine LUAD, we find that
   greater changes in age-associated patterns of gene regulation are more
   strongly associated with disease than is chronological age. We also
   find that smoking results in an apparent acceleration of the
   aging-associated patterns of gene regulation in both the lungs of
   “healthy individuals with a history of smoking” and in “normal
   adjacent” tissue from individuals suffering from LUAD, consistent with
   the fact that smoking dramatically increases risk, progression, and
   severity, and affects response to therapy. This suggests that the
   regulatory changes that are captured in the aging signature we derived
   are, at the least, correlated if not causally linked to LUAD disease
   processes. Looking at younger people with LUAD, we find that they also
   exhibit age -acceleration in their “normal” lung tissue relative to
   their peers without LUAD. Differential regulation in tumors of younger
   people with LUAD represented a subset of the changes we saw in the
   tumors of older patients, which suggests that the pathways associated
   with these changes might be particularly important in understanding the
   severity of disease in younger individuals. Finally, we found that even
   among individuals of similar age, those with a lower network-informed
   aging signature had better prognostic outcome, than individuals with
   higher aging signature. What all these means is that while
   chronological aging might have some effect on the risk of developing
   LUAD and its properties, the changes in aging-related regulation are
   far more important in estimating disease risk, in understanding disease
   processes, in identifying candidate therapies, and in designing
   aging-aware precision treatment protocols. This observation implies
   that chronological age alone does not provide substantial information
   on prescribing personalized therapy for LUAD and gene regulatory
   networks can prove to be effective tools in facilitating more efficient
   personalized therapy design and improving prognosis in LUAD for
   individuals across varying age.

Methods

Discovery dataset

   Uniformly processed RNA-Seq data were downloaded from the Recount3
   database^[150]61 for two discovery datasets using R package “recount3”
   (version 1.4.0): (i) lung tissue samples from the GTEx Project^[151]62
   (version 8) and (ii) LUAD samples from TCGA^[152]63 on May 26, 2022. We
   accessed clinical data for GTEx samples from the dbGap website
   ([153]https://dbgap.ncbi.nlm.nih.gov/) under study accession
   phs000424.v8.p2. Clinical data for TCGA samples were downloaded from
   Recount3. We refer to the GTEx samples as “non-cancerous lung samples”
   throughout our analysis.

   From 655 lung samples, 71 samples were removed because they were
   designated as “biological outliers” in the GTEx portal
   ([154]https://gtexportal.org/) for various reasons (as described in
   [155]https://gtexportal.org/home/faq). We analyzed the remaining 584
   lung samples (187 female and 397 male) from GTEx. We removed two
   recurrent tumor samples from the TCGA data and included the remaining
   541 primary tumor samples (293 female and 248 male) and 59 normal
   adjacent (34 female and 25 male) samples.

   Table [156]1 summarizes the clinical characteristics of the datasets.

Table 1.

   Clinical characteristics of the discovery and validation datasets
   GTEx TCGA TCGA LGRC [157]GSE68465
   Type Healthy Tumor Normal adjacent Healthy Tumor
   Sample size 584 539 59 108 443
   Age
   Mean ± std (range) 54 ± 11.81 (21–70) 65 ± 9.91 (33–88) 66 ± 10.83
   (42–86) 64 ± 11.35 (32–87) 64 ± 10.1 (33–87)
   Gender
   Female (%) 187 (32.02%) 291 (54.00%) 34 (57.63%) 59 (54.63%) 220
   (49.66%)
   Male (%) 397 (67.98%) 248 (46.00%) 25 (42.37%) 49 (45.37%) 223 (50.34%)
   Race
   White (%) 499 (85.44%) 411 (76.25%) 55 (93.22%) - 295 (66.60%)
   Black or African American (%) 70 (11.99%) 53 (9.83%) 4 (6.78%) - 12
   (2.71%)
   Others (%) 15 (2.57%) 75 (13.92%) 0 - 7 (1.58%)
   Unknown (%) - - - 129 (29.11%)
   Smoking status
   History of smoking (%) 386 (66.10%) 448 (83.11%) 46 (77.97%) 65
   (60.19%) 300 (67.72%)
   No history of smoking (%) 182 (31.16%) 77 (14.29%) 7 (11.86%) 32
   (29.63%) 49 (11.06%)
   NA (%) 16 (2.74%) 14 (2.60%) 6 (10.17%) 12 (10.18%) 94 (21.22%)
   Tumor stage
   I (%) - 295 (54.73%) 30 (50.85%) - 150 (33.86%)
   II (%) - 126 (23.38%) 13 (22.03%) - 251 (56.66%)
   III (%) - 84 (15.59%) 13 (22.03%) - 28 (6.32%)
   IV (%) - 26 (4.82%) 2 (3.39%) - 12 (2.71%)
   NA (%) - 8 (1.48%) 1 (1.69%) - 2 (0.45%)
   Ischemic time (hours)
   Mean ± std (range) 8.04 ± 6.98 (0.0–24.4) - - - -
   [158]Open in a new tab

   Both GTEx and TCGA gene expression data were normalized by transcript
   per million (TPM), using the “getTPM” function in the Bioconductor
   package “recount” (version 1.20.0)^[159]64 on R version 4.1.2. Lowly
   expressed genes were filtered out by removing genes with counts <1 TPM
   in at least 10% of the samples (126 samples) in GTEx and TCGA combined,
   removing 36386 genes, and keeping 27,470 genes. To construct gene
   regulatory networks, we further removed those genes that were not
   present in the TF/target gene regulatory prior used for creating the
   gene regulatory networks (see section “Differential targeting analysis
   using sample-specific gene regulatory networks”). This filtering left
   with 27,162 genes, including those on allosomes, which were used in
   subsequent analysis. For female samples in both GTEx and TCGA, gene
   expression values of all genes on the Y chromosome (36 genes in total)
   were replaced by “NA”. Principal component analysis (PCA) did not show
   any visible batch effect.

Validation dataset

   We chose two independent studies for validation from the Gene
   Expression Omnibus (GEO) repository: [160]GSE47460^[161]65 (downloaded
   on Feb 12, 2023) and [162]GSE68465^[163]66 (downloaded on Jan 24,
   2023). For validating our results on the lung samples from GTEx, we
   used [164]GSE47460, which consisted of microarray gene expression for
   582 samples in total from the Lung Genomics Research Consortium (LGRC).
   This study used Agilent-014850 Whole Human Genome Microarray 4x44K
   G4112F and Agilent-028004 SurePrint G3 Human GE 8x60K Microarray for
   sequencing. Among these 582 samples, we used only 108 samples (59
   female and 49 male), who have no chronic lung disease by CT or
   pathology and hence, were designated as “controls” within the study.
   [165]GSE68465 consists of microarray gene expression for 443 LUAD
   samples (220 female and 222 male). This study used Affymetrix Human
   Genome U133A Array for expression profiling.

   Table [166]1 summarizes the clinical characteristics of the datasets
   analyzed.

   Normalized expression data and clinical data were downloaded from GEO
   using R package “GEOquery” version 2.62.2. Within every dataset, for
   genes with multiple probe sets, we kept the probe set with the highest
   standard deviation in expression levels across samples. We discarded
   any genes that were not in the TF/target gene regulatory network prior
   that we used for creating the gene regulatory networks. This process
   left 13,575 genes in [167]GSE47460 (LGRC) dataset and 11,725 genes in
   [168]GSE68465 dataset, that we used to build gene regulatory network
   models. The LGRC data were not batch corrected because no visible batch
   effect was detected from PCA. The [169]GSE68465 dataset includes LUAD
   specimens from the following sources: University of Michigan Cancer
   Center (100 samples), University of Minnesota VA/CALGB (77 samples),
   Moffitt Cancer Center (79 samples), Memorial Sloan-Kettering Cancer
   Center (104 samples), and Toronto/Dana-Farber Cancer Institute (82
   samples). The [170]GSE68465 data were batch corrected for these sources
   using “ComBat” function implemented in the R package “sva” (version
   3.42.0).

Constructing sample-specific gene regulatory networks

   Gene regulatory networks for each sample were reconstructed by the
   PANDA^[171]19 and LIONESS^[172]20 algorithms using Python package
   netZooPy^[173]67 version 0.9.10, in both the discovery and the
   validation datasets. A schematic diagram of our network construction
   pipeline is given in Fig. [174]1. Three types of data were integrated
   to derive the regulatory networks: TF/target gene regulatory prior
   (obtained by mapping TF motifs from the Catalog of Inferred Sequence
   Binding Preferences (CIS-BP)^[175]68 to the promoter of their putative