Age-related tumor mutational burden and inflammaging: a pan-cancer multi-omics study of treatment-naïve patients

Abstract

Background

Aging is a major nonmodifiable risk factor for cancer, yet age-associated genomic instability, inflammatory activation, and tumor microenvironment interactions remain poorly characterized.

Patients and methods

The primary genomic cohort comprised 153 treatment-naïve patients with six cancer types, stratified at age 60 years; four additional patients without genomic profiling contributed to nongenomic analyses. Tumor mutational burden (TMB) and fraction genome altered were derived from whole-genome sequencing, and inflammatory pathway activity was determined from single-sample gene set enrichment analysis of RNA sequencing and proteomics in paired tumor and normal tissue. Because cancer type distribution differed between age groups, all principal associations were re-examined in multivariable models.

Results

TMB was higher in older patients (median 4.55 compared with 2.84 mutations per megabase; P < 0.001) and correlated with age (r_s = 0.334, P < 0.001) but was attenuated by adjustment for cancer type and lost significance once microsatellite instability (MSI) status, sex, and stage were added (P = 0.27). Normal-tissue inflammation correlated with paired tumor TMB (protein r_s =0.387, P < 0.001) but not after adjustment for age and cancer type (partial r_s = 0.07, P = 0.54). MSI-high tumors showed greater tumor-normal inflammation amplification than non-MSI-high tumors (+0.065, P = 0.015), strengthening after adjustment (+0.087, P < 0.001). An immune-hot cluster arose only in older patients but was not independently associated with age (odds ratio 1.03, P = 0.61).

Conclusions

Most age-associated molecular differences were substantially explained by cancer type. Only the age-related increase in normal-tissue inflammation at the protein level and the greater inflammation amplification of MSI-H tumors persisted. The correlation between normal-tissue inflammation and tumor TMB is not evidence that systemic inflammaging drives genomic instability.

Highlights

• •MSI-H tumors amplified tumor-normal inflammation more than non-MSI-H tumors; adjustment strengthened this.
• •Inflammation in normal tissue increased with age at the protein level after adjustment for cancer type.
• •The age-related increase in tumor mutational burden was not independent of cancer type and MSI status.
• •Correlation between normal-tissue inflammation and tumor mutational burden did not survive adjustment.
• •An immune-hot microenvironment arose only in older patients, but cancer type explained this distribution.

Introduction

Aging is a major nonmodifiable risk factor for cancer development.¹ Cancer incidence rises steadily with age, from fewer than 26 cases per 100 000 people aged <20 years to more than 1000 per 100 000 in those aged 60 and over.² ‘Inflammaging’—chronic, low-grade systemic inflammation associated with aging—has emerged as a key driver of age-related diseases including cancer3, 4, 5 and is characterized by elevated circulating cytokines, altered immune cell composition, and persistent inflammatory signaling.⁶

Large-scale genomic studies have documented that tumor mutational burden (TMB) increases with age, consistent with the accumulation of somatic mutations over time.⁷ This has therapeutic implications because TMB predicts immune checkpoint inhibitor response.8, 9, 10 The relationship between inflammation and genomic instability is bidirectional: chronic inflammation can induce DNA damage11, 12, 13 while genomically unstable tumors trigger inflammatory responses through cytosolic DNA sensing.¹⁴ Whether systemic inflammaging contributes to the age-related increase in TMB remains an open question.

The TITANIA (mulTi omIcs daTa cANcer dIagnostics therApies) framework provides an opportunity to investigate age-related molecular differences through matched multi-omics profiling. A distinctive feature is a sample handling protocol that minimizes cold ischemia time, which affects biomolecule stability in a modality-dependent manner: ischemia beyond 15 min has been shown to affect 12.5% of messenger RNA (mRNA), 25% of protein, and 50% of phosphosite differential expression signals.¹⁵ TITANIA employs rapid snap-freezing of fresh surgical specimens, preserving molecular signals otherwise lost in conventional biobanking.

Integrating whole-genome sequencing (WGS), transcriptomics, and proteomics from the same tissue specimens enables characterization of both tumor-intrinsic features and microenvironmental changes that may differ between younger and older patients. The treatment-naïve cohort analyzed here included patients with six cancer types: colorectal cancer (CRC), gastric cancer, renal cell carcinoma (RCC), non-small-cell lung cancer (NSCLC), ovarian cancer, and hepatocellular carcinoma (HCC).

We conducted a pan-cancer multi-omics analysis comparing patients aged ≤60 with those aged >60 years, characterizing genomic instability and inflammatory pathway activity at both transcriptomic and proteomic levels and exploring their relationships with tumor microenvironment (TME) characteristics. Our findings may inform future cancer type-stratified investigations of age-associated tumor biology.

Patients and methods

Study design and patient population

This study analyzed treatment-naïve cancer patients enrolled in the TITANIA project, a prospective multi-omics registry study conducted at the National Cancer Center Hospital East (Kashiwa, Japan) in collaboration with Indivumed GmbH (Hamburg, Germany), approved by the institutional review board (approval number: 2020-079) and conducted in accordance with the Declaration of Helsinki. All patients provided written informed consent for tissue and blood collection, multi-omics profiling, transfer of coded data to the collaborating institution, and publication of results in a form that does not identify individual participants.

Inclusion criteria were histologically confirmed primary malignancy, no prior systemic anticancer therapy, surgical resection with curative intent, and sufficient tissue for multi-omics analysis. Six cancer types were included: CRC, gastric cancer, RCC, NSCLC, ovarian cancer, and HCC. Patients were stratified at 60 years by age at surgery, a cut-off supported by prior literature.¹⁶ If a patient contributed more than one tumor specimen, the primary specimen was used. Transcriptomic and proteomic profiling was additionally available for four patients (three with CRC, one with ovarian cancer) in whom tumor genomic profiling was not evaluable; these are excluded from Table 1 and from every analysis using TMB, microsatellite instability (MSI) status, or the fraction of the genome altered (FGA), but were included in the pathway, microenvironment, and survival analyses, which therefore comprised 136 tumor transcriptomes, 141 tumor proteomes, 62 matched normal transcriptomes, and 82 matched normal proteomes, respectively.

Table 1.

Baseline characteristics of the 153 patients with genomic profiling, stratified by age group

Characteristic	Total (N = 153)	Aged ≤60 years (n = 40)	Aged >60 years (n = 113)	P value
Age, years	72 (60-77)	52.5 (44-55)	74 (71-79)	<0.001
Sex				0.051
Male	87 (56.9)	17 (42.5)	70 (61.9)
Female	66 (43.1)	23 (57.5)	43 (38.1)
Body mass index, kg/m2	23.1 (20.8-25.6)	22.9 (20.6-27.0)	23.2 (20.8-25.4)	0.751
Cancer type				0.002
CRC	52 (34.0)	13 (32.5)	39 (34.5)
Gastric	19 (12.4)	2 (5.0)	17 (15.0)
RCC	32 (20.9)	11 (27.5)	21 (18.6)
NSCLC	27 (17.6)	2 (5.0)	25 (22.1)
Ovarian	19 (12.4)	11 (27.5)	8 (7.1)
HCC	4 (2.6)	1 (2.5)	3 (2.7)
MSI status				0.033
MSI-H	11 (7.2)	0 (0)	11 (9.7)
Non-MSI-H	121 (79.1)	38 (95.0)	83 (73.5)
Not evaluable	21 (13.7)	2 (5.0)	19 (16.8)
Tumor stage				0.464
I	51 (33.3)	16 (40.0)	35 (31.0)
II	41 (26.8)	7 (17.5)	34 (30.1)
III	50 (32.7)	14 (35.0)	36 (31.9)
IV	11 (7.2)	3 (7.5)	8 (7.1)
Follow-up, days	756 (731-1103)	753 (610-1109)	758 (732-1102)	0.929
PFS events	33 (21.6)	11 (27.5)	22 (19.5)	0.402
Overall survival events	7 (4.6)	3 (7.5)	4 (3.5)	0.555
Cold ischemia time, minutes	12 (10-14)	12 (10-14)	13 (10-14)	0.198
Omics data availability
RNA-seq (tumor)	132	37	95
RNA-seq (normal)	61	16	45
Proteomics (tumor)	137	37	100
Proteomics (normal)	81	20	61

Values are n, n (%), or median (interquartile range). Tumor mutational burden and mutation data were available for all patients (N = 153). Tumor RNA-seq data were available for CRC (n = 43), RCC (n = 30), NSCLC (n = 23), Ovarian (n = 19), Gastric (n = 13), and HCC (n = 4). Tumor proteomics data were available for CRC (n = 45), RCC (n = 30), NSCLC (n = 23), Ovarian (n = 19), Gastric (n = 16), and HCC (n = 4). Matched normal tissue samples were available for a subset of patients. All counts in this table refer to the 153 patients with genomic profiling; the analysis sets used for the differential expression, pathway, microenvironment, and survival analyses are given in the Methods. P values were calculated using the Wilcoxon rank-sum test for continuous variables and the two-sided Fisher's exact test or chi-squared test for categorical variables. The MSI comparison used the 132 patients with an evaluable call.

CRC, colorectal cancer; HCC, hepatocellular carcinoma; MSI, microsatellite instability; MSI-H, microsatellite instability-high; NSCLC, non-small cell lung cancer; PFS, progression-free survival; RCC, renal cell carcinoma.

Tissue collection and processing

Tumor and matched normal tissues were collected intraoperatively using a standardized protocol designed to minimize cold ischemia time,¹⁵ with a median of 12 min [interquartile range (IQR) 10-14]. Pathological quality control required tumor cell content ≥30% and necrosis ≤30%. Normal tissue was sampled from a site distant from the tumor within the same resection specimen rather than from tissue immediately adjacent to the tumor, which may partially mitigate field cancerization effects. Full details of tissue handling, extraction, sequencing and mass spectrometry are given in Supplementary Methods, available at https://doi.org/10.1016/j.iotech.2026.101616.

DNA and RNA extraction and sequencing

DNA and RNA were coextracted from the same tissue aliquots, enabling direct comparison of genomic and transcriptomic data from identical cellular populations.

WGS achieved ≥60× mean coverage for tumor and ≥30× for matched normal samples. RNA sequencing (RNA-seq) was carried out after ribosomal RNA depletion to ≥100 million reads per sample.

Mass spectrometry-based proteomics

Proteomes were measured by data-independent acquisition liquid chromatography–tandem mass spectrometry and processed with DIA-NN version 1.8.1 using tissue-specific spectral libraries. Protein group-level quantification was used for downstream analysis, with normalized log₂ intensities representing relative protein abundances.

Bioinformatics processing

WGS data were aligned to the GRCh38 reference genome. Somatic mutations were called by a consensus of at least two of four callers, and copy number alterations were analyzed with TitanCNA (Supplementary Methods, available at https://doi.org/10.1016/j.iotech.2026.101616). TMB, MSI status, and FGA were computed by the data provider and supplied with the dataset. TMB was the number of nonsynonymous mutations in protein-coding genes divided by the exome size in megabases, expressed as mutations per megabase (mut/Mb). FGA, used throughout as the chromosomal instability measure, was derived from the whole-genome copy-number segmentation with the R package CINmetrics¹⁷ at its default settings: the summed length of segments whose absolute segment mean was at least 0.2, divided by the package default genome size of 2 873 203 431 bp (hg38 excluding the sex chromosomes). MSI status was determined with MSIseq¹⁸ from the somatic mutation calls, with samples classified as MSI-high (MSI-H) or non-MSI-H (including MSI-low and microsatellite stable). MSI status could not be called for 21 of the 153 patients; these were treated as missing in all MSI-related analyses rather than grouped with non-MSI-H tumors.

RNA-seq count data underwent batch effect correction using ComBat-seq, with normalization carried out within each cancer type based on the global Indivumed cohort. Transcripts per million (TPM) values were calculated from batch-corrected counts for downstream pathway analysis.

Missing data handling in proteomics

Proteomics data contained missing values arising from the stochastic nature of data-independent acquisition detection. We applied k-nearest neighbors imputation with k = 10 to proteins with <50% missing data across samples; proteins with more missing values were excluded.

Pathway enrichment analysis

Gene set enrichment analysis (GSEA) was carried out with the fgsea R package using Hallmark gene sets from the Molecular Signatures Database,¹⁹ with genes preranked by the DESeq2 Wald statistic²⁰ and 10 000 permutations. Pathway-level significance used a single primary threshold of Benjamini–Hochberg false discovery rate (FDR) <0.05. The correction family was the 50 Hallmark pathways tested within one analysis and one omics-tissue layer, as this is the set a reader interprets when examining a single panel; for analyses restricted to the six inflammatory pathways comprising the composite inflammation score, correction was applied across those six within each layer. Results significant only at FDR <0.25 are reported as exploratory. A master table of every pathway-level test is provided as Supplementary Table S1, available at https://doi.org/10.1016/j.iotech.2026.101616. One analysis departs from this threshold: the comparison of cell type abundance between MSI-H and non-MSI-H tumors was carried out at FDR <0.1 as an exploratory screen.

Single-sample GSEA (ssGSEA) was carried out using the Gene Set Variation Analysis (GSVA) R package with Hallmark gene sets.²¹ For RNA-seq data, log₂(TPM + 1) transformed expression values were used as input. For proteomics, log₂ normalized intensities were used directly.

Inflammation scores were calculated as the arithmetic mean of ssGSEA enrichment scores for six inflammatory Hallmark pathways: HALLMARK_INFLAMMATORY_RESPONSE, HALLMARK_INTERFERON_ALPHA_RESPONSE, HALLMARK_INTERFERON_GAMMA_RESPONSE, HALLMARK_IL6_JAK_STAT3_SIGNALING, HALLMARK_IL2_STAT5_SIGNALING, and HALLMARK_TNFA_SIGNALING_VIA_NFKB. The inflammation score is a composite ssGSEA summary constructed for this study and has not been independently validated as a biomarker. Its construction follows the general approach used in previous pan-cancer immune characterizations in which Hallmark or interferon (IFN)-related gene set scores are averaged to summarize inflammatory tone,^22,23 and all analyses based on it are therefore exploratory.

Differential expression analysis

Age-related differential expression was analyzed separately within CRC, RCC, and ovarian cancer. RNA-seq counts were analyzed with DESeq2,²⁰ retaining genes with a count of ≥10 in at least 20% of samples, and proteomic intensities with limma,²⁴ retaining proteins quantified in at least half of the samples. Significance was Benjamini–Hochberg adjusted P < 0.05 with an absolute log₂ fold change >0.5, except in Figure 1B, where a stricter effect-size threshold ≥1 was used for visualization. To assess whether the number of differentially expressed genes (DEGs) differed between cancer types because of sample size, each cancer type was down-sampled to the same number of patients per age group and the analysis repeated over 208 random draws, with the number of draws kept equal across cancer types.

Figure 1.

Cancer type-specific mutational and transcriptomic analysis. (A) Lollipop visualization of the top three mutated genes per cancer type. Mutation frequency (%) is shown for all patients (circle), younger patients aged ≤60 years (triangle), and older patients aged >60 years (square). The top three genes are ranked separately within each group, so a gene is shown only for the groups in which it ranks among the top three, and markers coincide where the frequency is identical between groups. Cancer types are indicated by color: CRC (green), gastric (red), RCC (orange), NSCLC (blue), and ovarian (purple). Hepatocellular carcinoma was excluded due to limited sample size (n = 4). (B) Venn diagrams showing overlap of DEGs between CRC, RCC, and ovarian cancer. Upper panel, DEGs with increased expression in younger patients (≤60 years); lower panel, DEGs with increased expression in older patients (>60 years). Numbers indicate gene counts with percentages in parentheses. This panel applies a stricter threshold than the Results, a Benjamini–Hochberg adjusted P < 0.05 together with an absolute log₂ fold change of at least one after shrinkage, to show the genes with the largest effect sizes; counts therefore differ from those reported in the text. (C) Heatmap of age-associated Hallmark pathway enrichment comparing patients aged ≤60 compared with >60 years in the RCC, CRC, and ovarian cancer groups. Colors represent NESs from gene set enrichment analysis: red indicates enrichment in younger patients; blue indicates enrichment in older patients. Asterisks denote Benjamini–Hochberg adjusted significance across the 50 Hallmark pathways within each cancer type (∗P_adj < 0.05, ∗∗P_adj < 0.01, ∗∗∗P_adj < 0.001). (D) Dot plot of shared Hallmark pathways across cancer types. NES values are shown for CRC (green), RCC (orange), and ovarian (purple). Point size represents −log₁₀(adjusted P value). Pathways are ordered by mean NES across cancer types.

CRC, colorectal cancer; DEG, differentially expressed gene; NES, normalized enrichment score; NSCLC, non-small-cell lung cancer; RCC, renal cell carcinoma.

TME deconvolution and clustering

Cell type abundances were estimated from RNA-seq TPM data using xCell,²⁵ which provides scores for 64 immune and stromal cell types. For clustering analysis, we selected cell types with detectable expression (nonzero scores) in >50% of samples and carried out z-score normalization across samples.

To explore TME subtypes we applied k-means clustering to the normalized xCell scores, evaluating k = 2 to k = 10 by silhouette width; k = 3 gave the highest average width (0.21). This width is low in absolute terms and indicates limited separation, so the three-cluster solution was treated throughout as an exploratory partition rather than a definitive classification. Clusters were labeled by their predominant cell type enrichment: immune-hot, immune-cold, and vascular-rich. Robustness was assessed in three ways: cluster-wise bootstrap Jaccard coefficients over 1000 resamples (clusterboot, fpc package), with ≥0.75 taken as stable and <0.60 as unstable²⁶; consensus clustering with ConsensusClusterPlus compared with the k-means solution by the adjusted Rand index²⁷; and reclustering on the residuals after regressing each xCell feature on cancer type, to quantify how much structure was attributable to tumor entity.

Statistical analysis

Continuous variables were summarized as median with IQR and compared using the Wilcoxon rank-sum test (two groups) or Kruskal–Wallis test (multiple groups). Categorical variables were compared using the two-sided Fisher’s exact test or chi-square test as appropriate. Correlations were assessed using Spearman’s rank correlation coefficient (r_s).

Because cancer type distribution differed significantly between age groups (Table 1), all principal age-related associations were re-examined in multivariable models. TMB was modeled on the log₁₀ scale by linear regression with age (continuous and dichotomized at 60 years), cancer type, MSI status, sex, and stage; CRC and stage I were the reference categories. Inflammation scores were modeled with the same covariates in each of the four omics-tissue layers, and a linear mixed model with a patient-level random intercept was fitted so that tumor and normal samples could be used jointly. The association between normal-tissue inflammation and tumor TMB was assessed by partial Spearman correlation and by linear regression in both directions. Cluster membership was modeled by multinomial logistic regression and, for the immune-hot membership, by binary logistic regression, with age, cancer type, MSI status, and log₁₀(TMB) as covariates; coefficients, confidence intervals (CIs), and P values are Wald-based throughout. Variance inflation factors were computed for all linear models. All patients with MSI-H tumors were older than 60 years, so models containing both terms are reported with the caveat that the two variables overlap only in the older group. As a sensitivity analysis, all principal analyses were repeated after excluding the four patients with HCC.

Survival analysis used the Kaplan–Meier method for progression-free survival (PFS), defined as time from surgery to progression or death from any cause, with the log-rank test for univariate comparisons. Optimal cut points for dichotomizing continuous pathway scores were determined using maximally selected log-rank statistics (R package maxstat) with a minimum group proportion of 0.15. The concordance index (C-index) was calculated from Cox models. Because the cut point is selected from the data, the log-rank P value obtained at that cut point is anticonservative²⁸; we therefore also report a permutation-corrected P value obtained by permuting the survival outcome 100 000 times and repeating the entire cut-point search on each replicate, together with the bootstrap distribution of the cut point over 1000 resamples. The continuous, cut-point-free Cox model is the primary survival analysis. Events per variable are reported for every multivariable model because values <10 indicate a substantial risk of overfitting.²⁹ Cancer type-specific survival analyses were carried out only when at least five events were observed.

All analyses were carried out in R version 4.4.1, using DESeq2 and limma for differential expression, fgsea for GSEA, GSVA for ssGSEA, xCell for cell type deconvolution, survival and survminer with maxstat for survival analysis, lme4 and lmerTest for mixed models, nnet for multinomial regression, fpc and ConsensusClusterPlus for cluster stability, and ComplexHeatmap and ggplot2 for visualization.

Results

Patient characteristics and data availability

The primary genomic cohort comprised 153 treatment-naïve patients with a median age of 72 years (IQR 60-77), of whom 40 (26.1%) were aged ≤60 years (Table 1). Cancer type distribution differed significantly between age groups (P = 0.002): the younger patient group was enriched for ovarian cancer (27.5% compared with 7.1%) and the older patient group for NSCLC (22.1% compared with 5.0%) and gastric cancer (15.0% compared with 5.0%). Sex, stage, body mass index, and cold ischemia time were balanced (all P > 0.05). MSI-H tumors were identified in 11 of the 132 patients with an evaluable MSI call (8.3%); all were older than 60 years, and all arose in patients with CRC (n = 7) or gastric cancer (n = 4) (two-sided Fisher’s exact P = 0.033; Supplementary Table S2, available at https://doi.org/10.1016/j.iotech.2026.101616). Because no patient aged ≤60 years had an MSI-H tumor, age and MSI status overlap only in the older group, and their independent associations cannot be reliably separated.

Multi-omics data availability varied by sample type: WGS-based mutation data were available for all 153 patients, tumor RNA-seq for 132 (86.3%), tumor proteomics for 137 (89.5%), matched normal RNA-seq for 61 (39.9%), and matched normal proteomics for 81 (52.9%). No patient with ovarian cancer had matched normal tissue, so every tumor-versus-normal comparison is structurally restricted to the remaining five cancer types (Supplementary Figure S1 and Table S2, available at https://doi.org/10.1016/j.iotech.2026.101616). Four patients with CRC or ovarian cancer had transcriptomic and proteomic but not genomic profiling and are included in the analyses that do not require genomic data.

Cancer type-specific molecular analysis

We first examined the mutational landscape, focusing on the genes ranking among the top three within each age-group stratum (Figure 1A); HCC was excluded because of its sample size (n = 4). The top mutated genes were largely consistent between age groups within each cancer type. Fisher’s exact tests were carried out for every gene mutated in at least three patients within a cancer type, and no gene showed an age-group difference surviving Benjamini–Hochberg correction (Supplementary Table S3, available at https://doi.org/10.1016/j.iotech.2026.101616). ARID1A in ovarian cancer was the most striking nominal result (82% compared with 25%, P = 0.024) but did not survive correction.

Differential expression and pathway analyses were limited to CRC, RCC, and ovarian cancer because of insufficient younger patient representation in gastric cancer and NSCLC. At the individual gene level (Figure 1B), each cancer type showed distinct age-associated DEGs with no genes common to all three. Volcano plots are provided for the transcriptome (Supplementary Figure S2, available at https://doi.org/10.1016/j.iotech.2026.101616) and proteome (Supplementary Figure S3, available at https://doi.org/10.1016/j.iotech.2026.101616), and the age-associated DEGs are listed in Supplementary Table S4, available at https://doi.org/10.1016/j.iotech.2026.101616. Ovarian cancer yielded substantially more DEGs (548) than CRC (56) or RCC (55). It was the only one of the three with a majority of younger patients and so provided the most balanced age comparison, but group size is not the whole explanation: down-sampled to identical group sizes over 208 draws, ovarian cancer still returned the most DEGs (median 413 against 56 and 46), although the distributions were wide and overlapped (Supplementary Figure S4 and Table S5, available at https://doi.org/10.1016/j.iotech.2026.101616).

In contrast, GSEA using Hallmark pathways revealed partially consistent age-related patterns across the three cancer types (Figure 1C and D), with asterisks based on Benjamini–Hochberg adjusted P values within each cancer type. Eleven of the 150 pathway-by-cancer combinations had a Benjamini–Hochberg adjusted P value <0.05. In CRC, IFN-γ response, allograft rejection, and complement were all enriched in older patients, together with pancreas β cell enrichment in younger patients. In ovarian cancer, TNFα signaling via NF-κB, oxidative phosphorylation, MYC targets V1, hypoxia, IL6-JAK-STAT3 signaling, and allograft rejection were all enriched in older patients, together with WNT/β-catenin signaling enriched in younger patients. No pathway had an adjusted P value <0.05 in RCC. Inflammatory and IFN-related pathways therefore showed higher enrichment in older patients in CRC and ovarian cancer, but this pattern was not shared by all three entities (Supplementary Table S1, available at https://doi.org/10.1016/j.iotech.2026.101616).

Integrated multi-omics ssGSEA analysis

To examine pathway-level changes at both omics levels, we carried out ssGSEA for all 50 Hallmark pathways (Figure 2A). Hierarchical clustering showed that samples clustered primarily by cancer type rather than by age group.

Figure 2.

Integrated multi-omics pathway analysis. (A) Heatmap of integrated RNA sequencing and proteomics Hallmark ssGSEA scores. Samples are arranged by hierarchical clustering. Annotation tracks (top to bottom): age group (≤60 in orange, >60 in cyan), age in years, progression-free survival (PFS) event status, tumor stage (I-IV), sex, MSI status, and cancer type. (B) PC analysis of Hallmark ssGSEA scores at RNA (upper panel) and protein (lower panel) levels. Points are colored by cancer type and shaped by age group (circle: >60 years; triangle: ≤60 years). The percentage of variance explained by each PC is indicated on the axes. (C) Sankey diagram illustrating patient flow across clinical and molecular features. From left to right: age group, cancer type, MSI status, TMB level (High/Medium/Low), RNA-based pathway cluster, protein-based pathway cluster, and disease progression status. MSI-H patients are highlighted in red.

CRC, colorectal cancer; EMT, epithelial-mesenchymal transition; HCC, hepatocellular carcinoma; MSI, microsatellite instability; MSI-H, microsatellite instability-high; NSCLC, non-small-cell lung cancer; PC, principal component; RCC, renal cell carcinoma; ssGSEA, single-sample gene set enrichment analysis; TMB, tumor mutational burden.

Principal component analysis of the ssGSEA scores separated samples primarily by cancer type at both RNA and protein levels, with no distinct clustering by age group (Figure 2B).

A Sankey diagram of patient flow across age group, cancer type, MSI status, TMB level, and pathway cluster (Figure 2C) showed that MSI-H tumors predominantly flowed through high TMB.

Age-associated increase in TMB

TMB correlated positively with age across the pan-cancer cohort (Spearman r_s = 0.334, P < 0.001; Figure 3A), whereas FGA did not (r_s = 0.05, P = 0.54), suggesting that the age-TMB relationship is more attributable to point mutations than to large-scale chromosomal alterations. The absence of an FGA association persisted after adjustment for cancer type (partial r_s = 0.04, P = 0.63), although ovarian cancer showed a positive association when analyzed alone (r_s = 0.58, adjusted P = 0.046; Supplementary Figure S5 and Table S6, available at https://doi.org/10.1016/j.iotech.2026.101616).

Figure 3.

Age-associated TMB and inflammatory pathway activation. (A) Scatter plot showing the correlation between age at surgery and TMB across all cancer types. TMB is displayed on a log scale. Points are colored by cancer type, shaped by MSI status (square: MSI-H; circle: non-MSI-H), and sized by chromosomal instability (CIN-FGA). The correlation was calculated for all 153 patients; the 21 without an evaluable MSI call carry no symbol and are therefore not displayed, so 132 points appear. The dashed red line represents the linear regression fit with 95% confidence interval (gray shading). Spearman correlation coefficient (ρ) and P value are shown. (B) Violin plot comparing TMB between younger (≤60 years) and older (>60 years) patients. Box plots within violins display median and interquartile range. Statistical comparison by Wilcoxon rank-sum test. ∗∗∗P < 0.001. (C) Scatter plots showing the correlation between age and inflammation score in tumor (red) and normal (blue) tissues at RNA (left panel) and protein (right panel) levels. Inflammation score represents the mean of six Hallmark inflammatory pathway ssGSEA scores. Points are shaped by MSI status (triangle: MSI-H; circle: non-MSI-H); correlation coefficients are calculated on all available samples. Regression lines with 95% confidence intervals are shown. Spearman correlation coefficients and P values are displayed for tumor and normal tissues separately. (D) Box plots comparing inflammation scores between age groups in normal and tumor tissues at RNA (left panel) and protein (right panel) levels. ∗P < 0.05, ∗∗P < 0.01.

CIN-FGA, chromosomal instability–fraction genome altered; CRC, colorectal cancer; HCC, hepatocellular carcinoma; MSI, microsatellite instability; MSI-H, microsatellite instability-high; NSCLC, non-small-cell lung cancer; RCC, renal cell carcinoma; ssGSEA, single-sample gene set enrichment analysis; TMB, tumor mutational burden.

Older patients showed significantly higher TMB than younger patients (P < 0.001; Figure 3B), and median TMB increased progressively across age decades (Kruskal–Wallis P = 0.00046; Supplementary Figure S6, available at https://doi.org/10.1016/j.iotech.2026.101616). Cancer type-specific analysis (Supplementary Figure S7, available at https://doi.org/10.1016/j.iotech.2026.101616) showed significant age-TMB associations in RCC (P < 0.001) and CRC (P < 0.05). A nominally significant difference in gastric cancer was attributed to only two patients aged ≤60 years and was not carried forward. With age as a continuous variable, the association varied markedly between entities (Supplementary Figure S8, available at https://doi.org/10.1016/j.iotech.2026.101616): strong in RCC (r_s = 0.79) and moderate in CRC (r_s = 0.39), but absent in NSCLC and gastric cancer and inverse in ovarian cancer (r_s = −0.31).

In multivariable models (Figure 4A and Supplementary Table S7, available at https://doi.org/10.1016/j.iotech.2026.101616) the age coefficient for log₁₀(TMB) decreased from 0.077 per 10 years unadjusted (P = 0.007), to 0.059 after adjustment for cancer type (P = 0.034), and to 0.027 (P = 0.27) after further adjustment for MSI status, sex, and stage. MSI-H status was by far the strongest predictor (β = 1.08, P < 0.001, an ∼12-fold higher TMB), and because all MSI-H cases were older than 60 years the age and MSI terms compete for the same variance; excluding MSI-H cases did not restore the age effect (β = 0.025, P = 0.30). Restricting the cancer type-adjusted model to the 132 patients with an evaluable MSI call left the age coefficient essentially unchanged (0.066 per 10 years, P = 0.032), so it was the addition of MSI status rather than the change in sample that removed it. The standard error of the age coefficient did not increase across these models (0.028 to 0.023 per 10 years), and variance inflation factors were <1.2 throughout: the loss of significance reflects attenuation of the estimate, not loss of precision. The age-related increase in TMB is therefore not independent of tumor entity and MSI status.

Figure 4.

Multivariable adjustment of age-associated findings. (A) Linear regression coefficients for age (per 10 years) predicting log₁₀(TMB) across four models of increasing adjustment. Points are coefficients and horizontal lines 95% CIs; red indicates P < 0.05. (B) Linear regression coefficients for age (per 10 years) predicting the composite inflammation score in each of the four omics-tissue layers. Colors denote the omics layer (blue, RNA; orange, protein); filled markers with a solid line are unadjusted estimates and open markers with a thin line are adjusted for cancer type. ∗P < 0.05. Estimates with further adjustment for MSI status, sex and stage are given in Supplementary Table S8, available at https://doi.org/10.1016/j.iotech.2026.101616. (C) Spearman correlation between normal tissue inflammation score and paired tumor TMB, unadjusted and as partial correlations adjusting for age, cancer type, or both, at RNA (blue) and protein (orange) levels. Error bars are 95% CIs; ∗P < 0.05. (D) Difference in ΔInflammation (tumor minus normal) between MSI-H and non-MSI-H tumors at the protein level: unadjusted, adjusted for cancer type, and within colorectal cancer only. Points are differences in the composite inflammation score with 95% CIs. (E) Left, odds ratio for immune-hot cluster membership per year of age, from binary logistic regression before and after adjustment for cancer type, MSI status and log₁₀(TMB); points are odds ratios and horizontal lines Wald 95% CIs. Right, cancer type of the 14 tumors assigned to the immune-hot cluster.

CI, confidence interval; CRC, colorectal cancer; MSI, microsatellite instability; MSI-H, microsatellite instability-high; NSCLC, non-small-cell lung cancer; RCC, renal cell carcinoma; TMB, tumor mutational burden.

Age-related inflammatory pathway activation

Inflammation scores were analyzed in relation to age in both tumor and normal tissues (Figure 3C). Tumor inflammation showed a weak positive correlation with age of near-identical magnitude at the two omics levels (RNA r_s = 0.223, P = 0.0091; protein r_s = 0.233, P = 0.0054). In normal tissue, the correlation was of similar size and reached significance at the protein level (r_s = 0.279, P = 0.011) but not at the RNA level (r_s = 0.241, P = 0.059).

Using the 60-year cut-off (Figure 3D), older patients had higher inflammation scores in tumor tissue at both RNA (P < 0.05) and protein (P < 0.01) levels and in normal tissue at the protein level (P < 0.01).

TMB did not correlate with tumor inflammation at the RNA level (r_s = 0.114, P = 0.194) but showed a weak positive correlation at the protein level (r_s = 0.249, P = 0.004). Normal-tissue inflammation also correlated with paired tumor TMB at the protein (r_s = 0.387, P < 0.001) and RNA (r_s = 0.283, P = 0.027) levels (Supplementary Figures S9-S11, available at https://doi.org/10.1016/j.iotech.2026.101616).

After adjustment for cancer type, the association between age and inflammation score persisted only in normal tissue at the protein level (β = 0.0007 per year, P = 0.006), and it remained after further adjustment for MSI status, sex, and stage (β = 0.0008, 95% CI 0.0003-0.0013, P = 0.003, n = 74). It was not significant in tumor tissue at either the RNA (P = 0.56) or protein level (P = 0.81) or in normal tissue at the RNA level (P = 0.32) (Figure 4B and Supplementary Table S8, available at https://doi.org/10.1016/j.iotech.2026.101616). A linear mixed model using tumor and normal samples jointly gave the same picture at the protein level. MSI-H status was itself associated with higher inflammation in tumor tissue at the protein level (β = 0.081, P < 0.001) but not in normal tissue (β = −0.011, P = 0.29), consistent with the tumor-restricted amplification described below. Of the six individual Hallmark inflammatory pathways, only IL6-JAK-STAT3 signaling in normal tissue at the protein level survived correction within its layer (β = 0.0014 per year, adjusted P = 0.002).

Normal-tissue inflammation and TMB after adjustment

Because the correlation between normal-tissue inflammation and paired tumor TMB is the observation from which a systemic inflammaging effect might be inferred, we examined it in detail. Adjustment for age alone reduced the protein correlation only modestly (partial r_s = 0.276, P = 0.013), adjustment for cancer type alone removed it (partial r_s = 0.165, P = 0.15), and adjustment for both left no association (partial r_s = 0.071, P = 0.54); all four estimates are derived from the same 81 patients, so the decline reflects adjustment rather than any change in precision. The RNA correlation behaved similarly (Figure 4C and Supplementary Table S9, available at https://doi.org/10.1016/j.iotech.2026.101616), and linear regression in both directions gave the same result. In the cancer type-stratified analysis, only CRC retained a significant correlation at the protein level (r_s = 0.51, adjusted P = 0.012, n = 35), and ovarian cancer contributed no paired normal tissue (Supplementary Figure S12, available at https://doi.org/10.1016/j.iotech.2026.101616). The apparent pan-cancer correlation therefore reflects differences between tumor entities in both typical TMB and the inflammatory tone of the organ of origin, rather than a consistent relationship within them; among the five entities with matched normal tissue, only CRC retained a significant correlation.

MSI-H status and inflammation amplification

Normal-tissue inflammation scores did not differ between MSI-H and non-MSI-H tumors at either the RNA (P = 0.79) or protein (P = 0.98) level (Supplementary Figure S13, available at https://doi.org/10.1016/j.iotech.2026.101616). However, the tumor-normal inflammation difference (ΔInflammation = Tumor − Normal) was significantly greater in MSI-H tumors at the protein level (Wilcoxon P = 0.016), whereas the difference at the RNA level was not significant (Wilcoxon P = 0.54; Figure 4D). Pathway-specific analysis (Supplementary Figure S14, available at https://doi.org/10.1016/j.iotech.2026.101616) showed that this amplification was more pronounced for IFN-α response and TNFα signaling via NF-κB at the protein level.

Patients whose MSI status could not be called were excluded from these comparisons rather than grouped with non-MSI-H tumors. In a linear model restricted to the 74 patients with paired tumor and normal proteomes and an evaluable MSI call, the ΔInflammation difference was 0.065 before adjustment (P = 0.015) and increased to 0.087 after adjustment for cancer type (P < 0.001); restricting to CRC gave 0.076 (P = 0.005) (Figure 4D; Supplementary Figure S15 and Table S10, available at https://doi.org/10.1016/j.iotech.2026.101616). At the pathway level, the amplification remained significant after adjustment and correction for IFN-α response, IFN-γ response, TNFα signaling via NF-κB, and inflammatory response. Only four patients had both an MSI-H call and paired normal RNA, so no adjusted model was fitted at the RNA level. Age was not included because no MSI-H tumor occurred in the younger age group.

Multi-omics concordance of age-related pathway changes

To identify pathways with age-related changes at both omics levels, we plotted RNA against protein age effects for all Hallmark pathways (Supplementary Figure S16, available at https://doi.org/10.1016/j.iotech.2026.101616). Pathways in the upper-right quadrant showed concordant higher scores in older patients at both levels. This is presented as concordance validation rather than as an independent result because the two omics layers were measured from the same specimens and were not statistically independent.

Several inflammatory pathways, including TNFα signaling via NF-κB, IL6-JAK-STAT3 signaling, IFN-γ response, inflammatory response, and allograft rejection, clustered in the upper-right quadrant, showing positive age effects at both levels.

The correlation matrix (Figure 5A) shows weak to modest positive correlations of age with inflammatory pathway scores and positive interpathway correlations.

Figure 5.

Inflammatory correlation structure and TME clustering. (A) Correlation matrix displaying Spearman correlation coefficients among age, TMB, and inflammatory pathway ssGSEA scores at RNA and protein levels. Circle size and color intensity indicate correlation strength. Red indicates positive correlation; blue indicates negative correlation. (B) Cluster-wise bootstrap stability of the three TME clusters, expressed as the mean Jaccard similarity over 1000 bootstrap resamples. The dashed line marks the conventional threshold of 0.75, above which a cluster is regarded as stable, and the dotted line represents the threshold of 0.60, below which it is regarded as unstable. Values are printed above each bar. (C) PC analysis of xCell-derived cell type abundances. Points are colored by k-means cluster assignment (k = 3): Cluster 1 (immune-hot, red), Cluster 2 (immune-cold, blue), and Cluster 3 (vascular-rich, green). Point size represents TMB. (D) Dot plot showing tumor microenvironment cluster distribution across cancer types and age groups. Dot size represents the percentage of patients in each cluster. Columns are grouped by cluster, with age groups (>60 and ≤60 years) shown within each cluster. (E) Violin plot comparing TMB across the three TME clusters. Statistical significance was assessed by Kruskal–Wallis test (P < 0.0001) followed by Dunn’s post hoc test with Bonferroni correction. ∗P < 0.05, ∗∗∗∗P < 0.0001. (F) Bar plot of Spearman correlation coefficients between TMB and xCell-derived cell type scores, in the 132 patients with both xCell data and TMB. Positive correlations (red) indicate cell types associated with higher TMB; negative correlations (blue) indicate cell types associated with lower TMB. Asterisks denote Benjamini–Hochberg adjusted P < 0.05 across the 14 cell types shown.

CRC, colorectal cancer; HCC, hepatocellular carcinoma; JAK, Janus kinase; IFN, interferon; IL, interleukin; NK, natural killer; NSCLC, non-small-cell lung cancer; PC, principal component; RCC, renal cell cancer; ssGSEA, single-sample gene set enrichment analysis; STAT, signal transducer and activator of transcription; TMB, tumor mutational burden; TME, tumor microenvironment; TNFα, tumor necrosis factor alpha.

Age-related pathway analysis in normal tissue (Supplementary Figure S17, available at https://doi.org/10.1016/j.iotech.2026.101616) showed that core inflammatory pathways were also elevated in older patients, most reaching significance at the protein level. Similar age-related patterns were therefore present in both tumor and matched normal tissue. Because the normal tissue analyzed here is organ tissue rather than blood or circulating mediators, these findings describe the tissue of origin and do not establish a systemic inflammatory process.

TME clustering

Using xCell-derived cell type abundances from tumor RNA-seq, k-means clustering identified three TME subtypes whose cluster-wise bootstrap stability is shown in Figure 5B. Principal component analysis (Figure 5C) showed partial separation of the clusters in the first two principal components, with point size indicating TMB.

The three-cluster solution was internally reproducible: repeating k-means with the same settings yielded clusters of 92, 14, and 30 samples with an average silhouette width of 0.208 (Supplementary Figure S18, available at https://doi.org/10.1016/j.iotech.2026.101616), and consensus clustering over 1000 resamples gave an almost identical partition (adjusted Rand index 0.948; Supplementary Figure S19, available at https://doi.org/10.1016/j.iotech.2026.101616). Cluster-wise bootstrap stability was high for the immune-cold cluster (Jaccard 0.89) and borderline for the immune-hot (0.75) and vascular-rich (0.74) clusters (Figure 5B and Supplementary Table S11, available at https://doi.org/10.1016/j.iotech.2026.101616). However, when cancer type was regressed out of every xCell feature and the samples reclustered, agreement with the original assignment fell to an adjusted Rand index of 0.36. A substantial part of this classification therefore reflects tumor entity rather than an entity-independent immune phenotype.

The three clusters were characterized as follows based on cell type distributions (Supplementary Figures S20 and S21, available at https://doi.org/10.1016/j.iotech.2026.101616):

• •Cluster 1 (immune-hot): High infiltration of immune cells including B cells, CD4+ T cells, CD8+ T cells, and M1 macrophages. This cluster showed the highest median TMB (Figure 5E) and was found exclusively in older patients.
• •Cluster 2 (immune-cold): Depleted immune cell infiltration across most cell types with lowest median TMB.
• •Cluster 3 (vascular-rich): Elevated endothelial cells and stromal components with intermediate TMB.

The immune-hot cluster was observed only in older patients across multiple cancer types, whereas the immune-cold and vascular-rich clusters showed more balanced age distribution (Figure 5D). The distribution of cancer types across age groups was markedly uneven in this cohort, so this panel is descriptive, and the formal analysis below should be used for inference.

Because cancer type distribution differs strongly between age groups, we tested whether age was independently associated with immune-hot membership. All 14 immune-hot tumors occurred in patients older than 60 years, and none of the 38 patients aged ≤60 years with xCell data fell into this cluster (Fisher’s exact P = 0.011); however, 9 of the 14 were NSCLC, an entity in which 25 of 27 patients in this cohort were older than 60 years. In binary logistic regression, the odds ratio per year of age was 1.063 (95% CI 0.997-1.134, P = 0.060) unadjusted and 1.026 (95% CI 0.930-1.132, P = 0.61) after adjustment for cancer type, MSI status, and log₁₀(TMB). In the multinomial model only cancer type reached significance (likelihood ratio P < 0.001), whereas age did not (P = 0.48) (Figure 4E and Supplementary Table S11, available at https://doi.org/10.1016/j.iotech.2026.101616). Several cancer type cells were empty, producing quasi-complete separation, so the individual odds ratios are not interpretable, and the likelihood ratio tests are the meaningful summary.

TMB differed significantly among clusters (Kruskal–Wallis P < 0.0001; Figure 5E), and correlated positively with B cells, neutrophils, M1 macrophages, CD4+ T cells, and overall immune score and negatively with endothelial cells, M2 macrophages, and stromal score (Figure 5F), indicating that high-TMB tumors are immune-infiltrated rather than stromal-dominant.

MSI-H tumors showed increased infiltration of CD4+ T cells, memory B cells, and CD8+ effector memory T cells alongside decreased stromal and endothelial signatures (Supplementary Figures S22 and S23, available at https://doi.org/10.1016/j.iotech.2026.101616); this screen was carried out at an exploratory threshold of FDR <0.1.

Prognostic pathway analysis

As an exploratory analysis, we evaluated the prognostic value of all 50 Hallmark pathways by C-index at both omics levels (Supplementary Figure S24, available at https://doi.org/10.1016/j.iotech.2026.101616). The epithelial-mesenchymal transition (EMT) pathway showed the highest C-index at both RNA (0.63) and protein (0.61) levels, with concordant hazard ratio (HR) direction between the two. Across all 100 pathway-by-layer models, only one bootstrap 95% CI for the C-index lay entirely above 0.5 (Supplementary Table S12, available at https://doi.org/10.1016/j.iotech.2026.101616).

Kaplan–Meier analysis showed that high EMT scores were associated with worse PFS at both RNA (HR 3.19, 95% CI 1.51-6.71, log-rank P = 0.0013; Supplementary Figure S25, available at https://doi.org/10.1016/j.iotech.2026.101616) and protein (HR 2.92, 95% CI 1.39-6.12, P = 0.0029; Supplementary Figure S26, available at https://doi.org/10.1016/j.iotech.2026.101616) levels, with 3-year PFS rates of 82.5% compared with 59.8% at the RNA level and 81.7% compared with 57.3% at the protein level.

These estimates require substantial qualification. Treating the EMT score as a continuous variable, which requires no cut point, the HR per standard deviation was 1.47 at the RNA level (P = 0.053) and 1.36 at the protein level (P = 0.11); multivariable models gave 1.50 (P = 0.059) and 1.40 (P = 0.10), with only 2.8 and 2.9 events per variable, respectively. The log-rank P values above were obtained at a data-derived optimal cut point; after permutation correction, they become P = 0.026 at the RNA level and P = 0.052 at the protein level (100 000 permutations), and the bootstrap distribution of the cut point was wide (Supplementary Figure S27 and Table S12, available at https://doi.org/10.1016/j.iotech.2026.101616). Cancer type-specific analysis was possible only for CRC (10 events), where the dichotomization was associated with PFS at both omics levels (P = 0.036 and P = 0.029), and NSCLC (6 events), where neither reached significance (Supplementary Figure S28 and Table S12, available at https://doi.org/10.1016/j.iotech.2026.101616). The analysis is hypothesis-generating, with no claim of independent prognostic value.

Sensitivity analyses

Because only four patients had HCC, all principal regression, survival, and clustering analyses were repeated after excluding them (n = 149). Of the 20 models compared, the significance conclusion changed in only one (the univariable Cox model for continuous EMT score at the RNA level, P = 0.053 to P = 0.037). The tumor-normal inflammation amplification of MSI-H tumors was unchanged (cancer type-adjusted difference 0.087 with and without HCC, P < 0.001), and rederiving the tumor microenvironment clusters without HCC reproduced the original partition exactly (adjusted Rand index 1.000). Full comparisons are given in Supplementary Figure S29 and Supplementary Table S13, available at https://doi.org/10.1016/j.iotech.2026.101616.

Discussion

Matched tumor and normal tissue from the same treatment-naïve patients, collected under stringent cold ischemia control and profiled at both the transcriptomic and proteomic levels, allowed us to ask which age-associated differences belong to the tumor and which belong to the tumor entity. In unadjusted analyses, older patients exhibited higher TMB and higher inflammatory pathway activity at both omics levels. However, cancer type distribution differed markedly between age groups, and once cancer type was taken into account, most of these differences were substantially attenuated, with standard errors that did not increase. What persisted was the age-related increase in inflammatory signaling in normal tissue at the protein level and the greater tumor-normal inflammation amplification of MSI-H tumors.

Older patients showed significantly higher TMB than younger patients (median 4.55 compared with 2.84 mut/Mb; P < 0.001), consistent with previous large-scale genomic studies⁷ and likely reflecting the cumulative accumulation of somatic mutations over a lifetime. We did not observe a correlation between age and chromosomal instability as measured by FGA. In multivariable models, the age coefficient for TMB was more than halved by adjustment for cancer type and lost significance once MSI status, sex, and stage were added. This does not contradict the large-scale genomic literature, in which the relationship is estimated in far larger and better balanced cohorts,⁷ but it does mean that the present cohort cannot demonstrate an entity-independent age effect. The absence of an age-FGA association should likewise be interpreted with caution: the burden of large-scale alterations is strongly determined by tumor entity and by whole-genome doubling status,^30,31 and a single summary metric may be insufficiently sensitive to detect a modest age effect.

A notable observation was the correlation between normal-tissue inflammation and paired tumor TMB, statistically significant at the protein level before adjustment (r_s = 0.387, P < 0.001). It did not survive adjustment: it was removed by adjustment for cancer type alone and was absent after adjustment for both age and cancer type (partial r_s = 0.071, P = 0.54), and in the stratified analysis it was retained only in CRC. The most parsimonious explanation is compositional: tumor entities differ systematically both in typical TMB and in the inflammatory tone of their organ of origin, so pooling them generates a pan-cancer correlation that is not consistently observed within individual tumor entities. These data therefore do not support the interpretation that systemic inflammaging contributes to genomic instability. MSI-H tumors showed similar normal-tissue inflammation to non-MSI-H tumors, yet greater tumor-normal inflammation amplification at the protein level. This persisted after adjustment for cancer type, where it strengthened rather than weakened (+0.087, P < 0.001), and within CRC alone (+0.076, P = 0.005), although both estimates rest on small numbers. A stronger intratumoral inflammatory response in MSI-H tumors is biologically expected given their high neoantigen load, and the value of the present observation lies less in its novelty than in the demonstration that the difference is confined to the tumor compartment: baseline inflammation in matched normal tissue did not differ by MSI status. This separation is what matched tumor-normal profiling makes possible. Reverse causation cannot be excluded in this cross-sectional design.

Cell type deconvolution analysis revealed that TMB was positively associated with infiltration of B cells, neutrophils, M1 macrophages, and CD4+ T cells, while negatively associated with stromal and endothelial cell signatures. These findings suggest that high-TMB tumors tend to exhibit immune-infiltrated rather than stromal-dominant microenvironments.

High immune infiltration and elevated TMB have been associated with immunotherapy response in some contexts,8, 9, 10 but whether older patients in our cohort would benefit from checkpoint inhibitors remains speculative. The adjusted analyses indicate that this distribution is largely a function of cancer type: nine of the 14 immune-hot tumors were NSCLC, an entity represented almost entirely by older patients here, and age was not independently associated with immune-hot membership after adjustment (odds ratio 1.03, P = 0.61). The clustering is internally reproducible, but reclustering after removing the cancer type effect recovered the original partition only weakly (adjusted Rand index 0.36).

The six cancer types differ substantially in the extent to which checkpoint inhibition has entered routine practice. In CRC, it is standard only for the mismatch repair-deficient or MSI-H subgroup, in which pembrolizumab is superior to chemotherapy in the first line³²; the seven patients with MSI-H CRC here represent precisely that subgroup. In gastric cancer, first-line nivolumab with chemotherapy is standard for tumors with sufficient PD-L1 expression.³³ In RCC, checkpoint inhibitor combinations are first-line standard of care irrespective of TMB, which is characteristically low,³⁴ and RCC showed the strongest age-TMB correlation here while remaining a low-TMB entity. In NSCLC, chemoimmunotherapy is standard first-line treatment³⁵ and NSCLC contributed most of the immune-hot cluster. By contrast, checkpoint inhibition has not shown first-line benefit in ovarian cancer,³⁶ and in HCC, atezolizumab–bevacizumab is standard but not TMB-selected.³⁷ TMB is a licensed selection biomarker only in the tissue-agnostic high-TMB indication, while PD-L1 expression, mismatch repair status, and histology remain the operative criteria. That TMB rises with age largely as a function of tumor entity therefore does not support age as a stratification variable for immunotherapy, consistent with meta-analyses showing comparable efficacy above and below age 65 years.³⁸ No patient here received a checkpoint inhibitor, so any inference about response is indirect.

Our study has several limitations. The sample size, particularly for younger patients (n = 40), limited statistical power, and the cross-sectional design precludes causal inference. Most importantly, cancer type is a strong confounder of every age comparison here: the six entities differ in age distribution, in baseline TMB, and in the inflammatory tone of the tissue of origin, and several unadjusted findings did not survive adjustment for it. All 11 patients with MSI-H tumors were older than 60 years, so the independent associations of age and MSI status cannot be reliably separated. No normal tissue was available for ovarian cancer, and only four patients had HCC, so tumor-versus-normal analyses cover five entities. The composite inflammation score has not been independently validated. With 33 progression events, the survival analyses are underpowered and have fewer than three events per variable in multivariable models.

The strengths of this study are the matched tumor-normal design, the rigorous tissue handling with a median cold ischemia time of 12 min, and a treatment-naïve population that eliminates therapy-induced molecular changes. Together, these are what allowed the MSI-H inflammation difference to be localized to the tumor compartment rather than inferred from tumor tissue alone.

In conclusion, this matched tumor-normal multi-omics analysis shows that most apparent age-related differences in a pan-cancer surgical cohort are substantially explained by the differing composition of cancer types across age groups. Two findings persisted after adjustment: inflammatory signaling in normal tissue increases with age at the protein level, and MSI-H tumors amplify inflammation relative to their own normal tissue more than non-MSI-H tumors. The correlation between normal-tissue inflammation and TMB did not persist and should not be interpreted as evidence that systemic inflammaging drives genomic instability. Testing the inflammaging hypothesis will require cohorts large enough to estimate age effects within individual tumor entities.