.
O P E N A C C E S S S O U R C E : bioXRiv
Abstract
Cancer is often called a disease of aging. There are numerous ways in which cancer epidemiology and behaviour change with the age of the patient. The molecular bases for these relationships remain largely underexplored. To characterize them we analyzed age-biases in the somatic mutational landscape of 12,774 tumours across 33 tumour-types. Age influences both the number of mutations in a tumour and their evolutionary timing. Specific mutational signatures are associated with age, reflecting differences in exogenous and endogenous oncogenic processes. A subset of known cancer driver genes were mutated in age-biased patterns, and these alter the transcriptome and predict for clinical outcomes. These effects were most striking in lower grade glioma where ATRX mutation is a strongly age-dependent prognostic biomarker. Though cancer genome sequencing data is not well-balanced in epidemiologic factors, these data suggest that age shapes the somatic mutational landscape of cancer, with clear clinical implications.
Introduction
Cancer health disparities across different population stratifiers are common through a wide range of measures. These include differences in incidence rates, mortality rates, response to treatment, and survival between individuals of different sexes1–6, races or ancestries7–11 and ages12–14, and these differences have been described across a range of tumour-types. Cancer disparities involving age are particularly well known. Aging is a leading risk factor for cancer, as it is associated with increased incidence of most tumour-types9,15. Older age is also associated with higher mortality and lower survival16,17. The links between older age and increased cancer burden such that cancer is often described as a disease of aging18,19.
However, there are many nuances in the relationship between aging and cancer. Pediatric cancers are an obvious exception, as cancers arising in children have different molecular and clinical characteristics9,20–22. Tumours arising in young adults (< 50 years of age) are often more aggressive: early onset prostate23, breast24, and colorectal25 cancers are diagnosed at higher stages and associated with lower survival. Molecular studies have described some striking differences in the mutational landscapes of early onset vs. later onset disease26–28, suggesting differences in the underlying oncogenic processes driving cancer at different ages.
The mechanisms of how age shapes the clinical behaviour of cancers has been subject to intense study. Many factors and behaviours closely tied to aging have been implicated in observed epidemiological and clinical cancer health disparities. For example, higher age is associated with a greater burden of comorbidities such as diabetes and cardiovascular disease29,30. Higher prevalence of chronic disease, frailty and increased likelihood of adverse drug reactions also influence the choices of clinical interventions given to older cancer patients31–33. Nevertheless differences remain even after accounting for these factors34. Previous work associating somatic molecular changes with age suggest differences in overall tumour mutation burden35, transcriptional profiles36, and some mutational differences26–28. These studies have focused on single tumour-types, relatively small cohorts, or have only evaluated fractions of the whole-genome, leaving the landscape of age-associated cancer mutations largely unknown.
To fill this gap, we perform a pan-cancer, genome-wide study of age-associated molecular differences in 10,218 tumours of 23 tumour-types from The Cancer Genome Atlas (TCGA) and 2,562 tumours of 30 tumour-types from the International Cancer Genome Consortium/The Cancer Genome Atlas Pan-cancer Analysis of Whole Genomes (PCAWG) projects. We quantified age-biases in measures of mutation density, subclonal architecture, mutation timing, mutational signatures and driver mutations in almost all tumour-types. We adjusted for potential confounding factors such as sex and ancestry. Many of these genomic age-biases were linked to clinical phenotypes. In particular, we identified genomic alterations that were prognostic in specific age contexts, suggesting the clinical utility of age-informed biomarkers.
Results
Age Biases in Mutation Density and Timing
We investigated TCGA and PCAWG datasets independently and performed pan-cancer analyses spanning all TCGA tumours (pan-TCGA), and all PCAWG tumours (pan-PCAWG); these were supplemented with tumour-type-specific analyses. We used the recorded age at diagnosis for both TCGA and PCAWG37 (Table 1). Our modeling accounted for a range of confounding variables for each cancer type including sex and genetic ancestry. We adapted a statistical approach previously applied to quantify sex-biases in cancer genomics38: we first used univariate methods to identify putative age-biases, then further modeled these putative hits with multivariate regression to evaluate age effects after adjusting for confounding factors. We modeled each genomic feature and tumour subtype based on available clinical data, a priori knowledge, variable collinearity and model convergence. Model and variable specifications, and results of association tests between model variables and age are presented in Supplementary Table 1.
We began by assessing whether measures of mutation density were associated with age. The accumulation of mutations with age is a well-known phenomenon in both cancer and non-cancer cells39–46. We examined both genome instability and SNV density to investigate trends across age and test the robustness of our statistical framework in detecting age-associated genomic events. Genome instability is a measure of copy number aberration (CNA) burden and approximated by percent of the genome altered by CNAs (PGA), a surrogate variable associated with poor outcome in several tumour-types47–49. We identified univariate age-biases in PGA using Spearman correlation. Putative age-associations identified at a false discovery rate (FDR) threshold of 10% were further analysed by multivariate linear regression (LNR) models to adjust for tumour-type-specific confounding effects (Supplementary Table 1).
We discovered significant associations between age and PGA in both pan-TCGA (ρ =0.14, adjusted LNR p = 1.1 × 10−7) and pan-PCAWG (ρ = 0.19, adjusted LNR p = 0.023) data. Positive correlations were also identified in three TCGA and three PCAWG tumour-types, with prostate cancer showing a statistically significant correlation in both datasets. (Figure 1A, 1B, Supplementary Table 2). Other tumour-type specific associations were statistically significant in only one dataset (Figure 1A, Supplementary Figure 1). For example, we detected similar correlations between age and PGA in TCGA lower grade glioma (LGG) and PCAWG (CNS-Oligo), but the association was significant in only TCGA data (Figure 1A, 1B). This is likely due in part to decreased statistical power in the PCAWG dataset because of smaller sample sizes. Surprisingly, in TCGA both adenocarcinomas and squamous cell carcinomas of the lung showed the inverse trend, with tumours arising in older patients harbouring fewer CNAs (LUAD: ρ = −0.18, adjusted LNR p = 6.0×10−4, LUSC: ρ = −0.10, adjusted LNR p = 0.039, Figure 1A). We observed similar negative correlations in the corresponding PCAWG lung data, though these associations were not statistically significant (Lung-AdenoCA: ρ = −0.13, Lung-SCC: ρ = −0.065, Figure 1A).
Figure 1. Mutation density and timing are biased to age.
Summary of associations between age and (A) percent genome altered and © SNV density in TCGA and PCAWG tumours. The dot size and colour show the Spearman correlation, and background shading indicate adjusted multivariate p-value. Only tumour-types with at least univariately significant associations are shown. Associations between (B) PGA and (D) coding SNV density with age in selected tumour-type specific analyses. Univariate Spearman correlation, adjusted correlation p-value and adjusted multivariate p-values shown. (E) Correlations between age and proportion of SNVs occurring in the truncal clone in four PCAWG tumour contexts.
Analogous to PGA, somatic single nucleotide variation (SNV) density measures the burden of somatic SNVs. SNV density frequently increased with age, as expected39,40. In addition to pan-cancer age-biases (pan-TCGA: ρ = 0.33, adjusted LNR p = 2.0 × 10−49, pan-PCAWG: ρ = 0.41, adjusted LNR p = 3.1 × 10−28), tumour-type-specific positive correlations occurred in 15/23 TCGA and 14/30 PCAWG tumour-types, including in prostate and gastric cancers (Figure 1C, 1D Supplementary Figure 1, Supplementary Table 2). Again, there was an inverse relationship in lung tumours, with more SNVs occurring in the squamous cell tumours of younger patients (LUSC: ρ = −0.15, adjusted LNR p = 0.064, Figure 1D). While not statistically significant, we observed similar negative associations in PCAWG lung tumours (Lung-SCC: ρ = −0.14). The negative association between age and both PGA and SNV density in lung cancers has been attributed to smoking exposure leading to hypermutation in younger lung cancer patients50, suggesting differences in disease aetiology between patients of different ages.
Another source of of hypermutation is microsatellite instability (MSI), which is frequently detected in colorectal and gastric cancers51,52. Since MSI-positive status is often associated with increased SNV density and age (Supplementary Figure 2), we investigated whether age-biases in MSI might explain the associations between age and SNV density in this data. We focused on four tumour-types with high frequency of MSI-positive tumours: stomach & esophageal, colorectal, pancreatic and endometrial cancers53,54. There was a significant association between age and MSI status in gastric cancers, where tumours arising in older individuals were more likely to have high levels of MSI (MSI-H; ANOVA q = 6.4×10−4; Supplementary Figure 2). While there were no statistically significant associations between age and MSI status in colorectal, pancreatic, or endometrial cancers, we nevertheless assessed the relationship between age and SNV density while accounting for MSI status in all four tumour-types. The associations between age and SNV density remained significant even after adjusting for MSI status in stomach & esophageal, colorectal, and pancreatic tumours (Supplementary Figure 2).
After identifying age-biases in mutation density, we next asked whether there were differences in the timing of when these mutations occurred during tumour evolution. We leveraged data describing the evolutionary history of PCAWG tumours55 and first investigated polyclonality, or the number of cancer cell populations detected in each tumour as assessed by multiple methods in this dataset. Monoclonal tumours, or those where all tumour cells are derived from one ancestral cell, are associated with better survival in several tumours types56–58. We also investigated mutation timing in polyclonal tumours by comparing how frequently SVNs, indels and structural variants (SVs) occurred as clonal mutations in the trunk or as subclonal ones in branches. While there were intriguing univariate associations between age and polyclonality in non-Hodgkin lymphoma and prostate cancer, they were not significant after multivariate adjustment (Supplementary Figure 2).
Focusing on polyclonal tumours, we compared how frequently mutations occurred in the trunk subclone vs. in branch subclones. Differences in the proportion of truncal mutations suggest difference in mutation timing over the evolution of a tumour. We identified several significant associations between age and mutation timing. In pan-PCAWG analysis, we found positive associations between age and proportion of clonal SNVs (ρ = 0.20, adjusted LNR p = 1.4×10−3, Figure 1E) and proportion of clonal indels (ρ = 0.14, LNR p = 0.013, Supplementary Table 2). Age was also associated with increasing clonal SNV proportion in two tumour-types: stomach cancer (Stomach-AdenoCA: ρ =0.44, adjusted LNR p = 0.028), and medulloblastoma (CNS-Medullo: ρ = 0.34, adjusted LNR p = 2.5 × 10−3, Figure 1E). A positive correlation results suggest that in these tumour-types, tumours arising in older individuals accumulate a greater fraction of SNVs earlier in tumour evolution. In contrast, an inverted trend occurred in melanoma, where tumours of younger patients tended to accumulate more subclonal than clonal SNVs (ρ = −0.47, adjusted LNR p = 7.8×10−3).
Age Biases in Mutational Processes
Differences in mutation density and timing suggest that different oncogenic processes might be preferentially active depending on the age of a patient. These processes can result in distinctive mutational patterns, which can be deconvolved and quantified59. We analysed age-biases in three types of mutational signatures generated by the PCAWG project: 49 single base substitution (SBS), 11 doublet base substitution (DBS) and 17 small insertion and deletion (ID) signatures60. We also investigated SBS signatures for TCGA tumours. For each signature, we examined both the proportion of signature-positive tumours as well as relative mutation activity, or the proportion of mutations attributed to each signature.
Across all 2,562 PCAWG tumours, we identified twelve mutational signatures with age-biased detection frequency (Figure 2A, left) and ten with age-biased mutation activity (Figure 2B, left). For example, tumours arising in older patients were more likely to be SBS3-positive (marginal log odds change = 0.0085, 95%CI = 0.0024-0.015, adjusted LGR p = 0.075), but in these SBS-positive tumours, the proportion of SBS3-attributed mutations decreased with age (ρ = −0.20, adjusted LNR p = 3.2×10−3). SBS3 mutations are thought to be caused by defective homologous recombination-based DNA damage repair. These results imply that while tumours derived from older individuals are more likely to harbour defective DNA damage repair, its relative impact on the burden of SNVs is lower compared with tumours derived of younger individuals. A similar relationship was seen for ID8, which is linked to defective non-homologous DNA end-joining (marginal log odds change = 0.024, 95%CI = 0.020-0.028, adjusted LGR p = 3.4 × 10−3; ρ = −0.099, adjusted LNR p = 3.7 × 10−5) and ID1, associated with slippage during DNA replication (marginal log odds change = 0.013, 95%CI = 0.0059-0.020, adjusted LGR p = 0.018; ρ = −0.059, adjusted LNR p = 0.048). We also identified positive associations between higher age and the tobacco-related signatures SBS4, DBS2 and ID3. Conversely, tumours arising in older individuals were less likely to exhibit defective base excision repair (SBS36). All mutations signatures findings are in Supplementary Table 2.
Figure 2. Biases in mutational signatures suggest differences in underlying mutational processes.
(A) Summary of associations between age and the proportion of signature-positive tumours, where dot size shows the marginal log odds from logistic regression and background shading show adjusted multivariate p-values. PCAWG data is on left and TCGA on right. (B) Similarly, the summary of associations between age and relative signature activity, with dot size showing Spearman correlations and background indicating adjusted linear regression p-values. © Comparison of PCAWG and TCGA signature detection frequency. Filled in and open circles indicate comparisons where the differences are statistically significant (proportion test q < 0.05) and not, respectively.
These pan-cancer differences persisted across individual tumour-types. We identified 23 age-associated signatures across eleven tumour-types, including six significant signatures in melanoma. In this tumour-type, tumours arising in older patients were preferentially SBS2-positive (marginal log odds change = 0.051, 95%CI = 0.013-0.095, adjusted LGR p = 0.029, Figure 2A), attributed to APOBEC cytidine deaminase activity61. Melanomas arising in younger patients were more likely to be positive for signatures related to UV damage (SBS 7a, 7b, 7d, Figure 2A, Supplementary Table 2). The proportion of mutations attributed to UV damage was also higher in younger patients (DBS1, ρ = −0.29, adjusted LNR p = 0.019, Figure 2B), while the proportion of mutations attributed to slippage during DNA replication was higher in older patients (ID1, ρ = 0.27, adjusted LNR p = 0.019, Figure 2B). These results suggest that melanomas in younger patients more frequently involve UV exposure and damage, while melanomas in older patients are more influenced by endogenous sources of mutation.
Leveraging data describing SBS signatures in TCGA data, we repeated this analysis to identify age-associations in signatures derived from whole exome sequencing (WXS) data. Across pan-TCGA tumours, we detected five signatures that occurred more frequently in older individuals, and three that occurred more frequently in younger individuals (Figure 2A). We also identified five signatures with higher relative activity in younger patients (Figure 2B). In cancer-specific analysis, we identified age-biased SBS signatures across eleven tumour-types, including negative associations between the tobacco-associated signature SBS4 and age in lung adenocarcinoma. SBS4 was more frequently detected in younger patients (LUAD: marginal log odds change = −0.041, 95%CI = −0.062 – −0.021, adjusted LGR p = 4.2 × 10−3, Figure 2A) and also had higher relative activity in younger lung squamous cell cancer patients (ρ = −0.17, adjusted LNR p = 0.015, Figure 2B). SBS4 activity was similarly negatively associated with age in PCAWG lung squamous cell cancers (Lung-SCC: ρ = −0.35, adjusted LNR p = 0.099, Figure 2B). Indeed, SBS4 and age were consistently negatively associated across both subtypes of lung cancer and both datasets though not all associations were statistically significant after multiple testing adjustment. This supports previous findings that tobacco has a larger tumorigenesis role in younger patients, with tobacco-associated mutations contributing to a greater portion of the mutational landscape of tumours derived from younger individuals50.
There was moderate agreement between TCGA and PCAWG findings: some signature like SBS2 and SBS4 were age-biased in the same or closely related tumour subtypes. Other signatures, such as SBS1 and SBS5 were age-biased in detection and relative activity across a range of tumour-types. Still others were age-biased exclusively in either TCGA or PCAWG data. We hypothesized that this was due to differences in signature detection rates between WXS and whole genome sequencing (WGS) data and compared how frequently each signature was detected across all samples (Figure 2C). Signatures with high agreement between datasets had similar detection rates, as observed for SBS2 (detection difference = 1.5%) and SBS4 (detection difference = 1.1%). Signatures where findings did not replicate had vastly different detection rates, as was seen for SBS1 (detection difference = 7.2%) and SBS5 (detection difference = 10%). We further examined this by comparing signatures data from non-PCAWG WGS and non-TCGA WXS data. Differences in signature detection rates between PCAWG and TCGA data were reflected in non-PCAWG WGS and non-TCGA WXS data (Supplementary Figure 3). We also looked specifically at identified age-biases and found high agreement in data generated by the same sequencing strategy (Supplementary Figure 3). The differences in signature detection, sequencing strategy, multivariate models, sample size, and geographic variation distinguishing PCAWG and TCGA datasets motivated our continued analysis of each dataset separately.
CNA Differences Associated with Transcriptomic Changes
Global mutation characteristics such as genome instability are features of later stages in a tumour’s evolutionary history. In contrast, the early stages are often driven by chromosome- or gene-specific events such as loss of specific chromosomes or mutation of driver genes55. We therefore narrowed our focus to chromosome segment and gene-level events. We applied our statistical framework to identify putative age-biased copy number gains and losses using univariate logistic regression (ULR). Putative events identified with a false discovery rate threshold of 10% were further analysed by multivariable logistic regression to account for confounding factors.
.../...
.
Edited by Engadin, 14 July 2020 - 07:58 PM.
