Editor's evaluation: Common genetic variations in telomere length genes and lung cancer: a Mendelian randomisation study and its novel application in lung tumour transcriptome

Abstract

Article Figures and data Abstract Editor's evaluation Introduction Materials and methods Results Discussion Appendix 1 Data availability References Decision letter Author response Article and author information Metrics Abstract Background: Genome-wide association studies (GWASs) have identified genetic susceptibility variants for both leukocyte telomere length (LTL) and lung cancer susceptibility. Our study aims to explore the shared genetic basis between these traits and investigate their impact on somatic environment of lung tumours. Methods: We performed genetic correlation, Mendelian randomisation (MR), and colocalisation analyses using the largest available GWASs summary statistics of LTL (N=464,716) and lung cancer (N=29,239 cases and 56,450 controls). Principal components analysis based on RNA-sequencing data was used to summarise gene expression profile in lung adenocarcinoma cases from TCGA (N=343). Results: Although there was no genome-wide genetic correlation between LTL and lung cancer risk, longer LTL conferred an increased risk of lung cancer regardless of smoking status in the MR analyses, particularly for lung adenocarcinoma. Of the 144 LTL genetic instruments, 12 colocalised with lung adenocarcinoma risk and revealed novel susceptibility loci, including MPHOSPH6, PRPF6, and POLI. The polygenic risk score for LTL was associated with a specific gene expression profile (PC2) in lung adenocarcinoma tumours. The aspect of PC2 associated with longer LTL was also associated with being female, never smokers, and earlier tumour stages. PC2 was strongly associated with cell proliferation score and genomic features related to genome stability, including copy number changes and telomerase activity. Conclusions: This study identified an association between longer genetically predicted LTL and lung cancer and sheds light on the potential molecular mechanisms related to LTL in lung adenocarcinomas. Funding: Institut National du Cancer (GeniLuc2017-1-TABAC-03-CIRC-1-TABAC17‐022), INTEGRAL/NIH (5U19CA203654-03), CRUK (C18281/A29019), and Agence Nationale pour la Recherche (ANR-10-INBS-09). Editor's evaluation This study is of interest to epidemiologists and geneticists studying the association between telomere length and lung cancer risk. This work provides useful insight into risk factors for lung cancer. Overall, the results of this study are solid, as the genetic instrument used here is better powered and the battery of Mendelian randomization analysis makes this broad set of results convincing compared to previous work. https://doi.org/10.7554/eLife.83118.sa0 Decision letter Reviews on Sciety eLife's review process Introduction Telomeres are a complex of repetitive TTAGGG sequences and nucleoproteins located at the end of chromosomes and have an essential role in sustaining cell proliferation and preserving genome integrity (de Lange, 2009). Telomere length progressively shortens with age in proliferative somatic cells due to incomplete telomeric regions replication (Watson, 1972) and low activity of the telomerase TERT in adult cells. The shortening of the telomere length results in cell cycle arrest, cellular senescence, and apoptosis in somatic cells (Harley et al., 1990). The maintenance of telomere length, which allows cancer cells to escape the telomere-mediated cell death pathways, is one feature related to the hallmarks of cancer (Hanahan and Weinberg, 2011). Telomere length appears to vary between individuals and has been studied in relation to many diseases. In observational studies, telomere length is measured as the average length of telomeric sequences in a given tissue (Montpetit et al., 2014). Telomere length appears correlated across tissue types (Demanelis et al., 2020), and as such, leukocyte telomere length (LTL) is generally measured in epidemiologic studies as a proxy for telomere length in other tissues. Recently, LTL has been measured in 472,174 individuals from the UK Biobank (UKBB; Codd et al., 2021), and LTL was associated with multiple biomedical traits (i.e. pulmonary and cardiovascular diseases, haematological traits, lymphomas, kidney cancer, and other cancer types). Genetic analysis of LTL also revealed 138 genetic loci linked to LTL across a variety of different genes involved in telomere biology and DNA repair (Codd et al., 2021). In the context of lung cancer, genetic variants at several loci have been associated with both LTL and lung cancer risk, including variants near the TERT, TERC, OBFC1, and RTEL1 genes, fundamental to telomere length maintenance (McKay et al., 2008; Wang et al., 2008; Rafnar et al., 2009; Kachuri et al., 2016; McKay et al., 2017). The effects of the telomere-related variants appear more relevant to lung adenocarcinoma risk than other histologic subtypes (McKay et al., 2017; Landi et al., 2009). Accordingly, a causal relationship between LTL and susceptibility to lung cancer was observed using Mendelian randomisation (MR) approaches (Zhang et al., 2015; Haycock et al., 2017; Kachuri et al., 2019) as well as in observational studies that have associated directly measured telomere length with risk of lung cancer (Sanchez-Espiridion et al., 2014; Zhang et al., 2017). The aim of the current work was to investigate the relationship between genetically predicted LTL and lung cancer, including lung cancer histological subtypes and smoking status. To this end, we conducted genome-wide correlations, MR, and colocalisation analyses to explore the relationship between LTL and lung cancer. We additionally undertook polygenic risk score (PRS) analysis using the LTL genetic instrument to explore the influence of LTL on the demographic, clinical, and molecular features of lung adenocarcinoma tumours. Materials and methods Reporting guidelines Request a detailed protocol The current study has been reported according to the STROBE-MR guidelines (Reporting Standards Document). Data Genome-wide association studies (GWASs) summary statistics for lung cancer (29,239 cases and 56,450 controls) and stratified by histological subtype (squamous cell carcinoma, small-cell carcinoma, and adenocarcinoma) and smoking status (ever and never smokers) were obtained from the International Lung Cancer Consortium (ILCCO; McKay et al., 2017). All analyses of LTL requiring summary statistics used results from a GWAS of LTL in 464,716 individuals of European ancestry from the UKBB (Codd et al., 2021). Downstream analyses considered additional lung cancer risk factors, such as lung function and cigarette smoking. We obtained GWAS summary statistics for forced expiratory volume in 1 s (FEV1) and forced vital capacity (FVC) from a published UKBB analysis (Kachuri et al., 2020). For smoking behaviour traits, we used results from the GWAS and Sequencing Consortium of Alcohol and Nicotine use (GSCAN) consortium meta-analysis of cigarettes per day (continuous), smoking initiation (ever versus never), smoking cessation (successfully quit versus continuing), and age at smoking initiation (continuous; Liu et al., 2019) excluding the UKBB participants. For the obesity-related traits (continuous), we used the results from the UKBB and GIANT meta-analysis of BMI (Pulit et al., 2019) and waist-to-hip ratio (WHR; Pulit et al., 2019), or OpenGWAS data using UKBB participants (Elsworth et al., 2020) for high-density lipoprotein (HDL), triglycerides, and systolic and diastolic blood pressure. For the alcohol behaviour trait, we obtained the results from GSCAN phase 2 of drinks per week (continuous; Saunders et al., 2022). Colocalisation analyses of gene expression used lung tissue expression quantitative trait loci (eQTL) summary statistics from the Genotype-Tissue Expression (GTEx) data version 8. Analyses of molecular phenotypes were performed using 343 lung adenocarcinoma samples of European ancestry from The Cancer Genome Atlas (TCGA) cohort with germline profile, RNA-sequencing, and epidemiological data available. Genotyping and imputation of germline variants have been described elsewhere (Gabriel et al., 2022). The total somatic mutation burden of TCGA samples was obtained from Ellrott et al., 2018, and DNA mutational signatures were extracted and attributed, as previously described (Gabriel et al., 2022). RNA-sequencing data were obtained from TCGA data portal using TCGA biolinks package in R (version 2.22.3; Colaprico et al., 2016). Telomere length measurement by whole-genome sequencing (WGS-measured TL, 655 samples across cancer sites) was retrieved from Barthel et al., 2017. Tumour genomic characteristics were defined by the analyses of the TCGA data, including gene expression-based scores of telomerase activity (Barthel et al., 2017) and cellular proliferation (Thorsson et al., 2018), as well as the observed frequency of somatic homologous recombination-related events (represented as a homologous recombination repair deficiency score), and the average number of somatic copy number alteration within the tumours (Knijnenburg et al., 2018). Linkage disequilibrium score regression Request a detailed protocol Genetic correlations across traits were calculated using linkage disequilibrium score regression (LDSC) by the LDSC package (v1.0.0; Bulik-Sullivan et al., 2015). Linkage disequilibrium (LD) scores were generated on the 1000 Genomes Project Phase 3 reference panel with the HLA region excluded as provided by the package due to long range LD patterns. The genome-wide correlations that passed Bonferroni correction (adjusted p-values<0.05) were considered statically significant. Mendelian randomisation Request a detailed protocol MR is a method for interrogating relationships between putative risk factors and health outcomes by using genetic variants associated with the exposure of interest, typically obtained from GWAS, as instrumental variables. Assuming that fundamental MR assumptions are satisfied, this approach can be said to identify unbiased causal estimates. The genetic instrument for LTL was defined as the set of 144 genetic variants that were genome-wide significant (p<5x10–08) but not in linkage disequilibrium with each other (r2<0.01) and restricted to common genetic variation (minor allele frequency >1%) in European populations. Proxy variants in LD (r2>0.8, whenever possible) were chosen when a genetic variant was not available in the lung cancer GWAS. Primary MR analyses were conducted using the inverse-variance method with multiplicative random-effects (Yavorska and Burgess, 2017). Sensitivity analyses to horizontal pleiotropy and other violations of MR assumptions were performed using other MR estimation methods, such as weighted median, MR-Egger, contamination mixture model, MR-PRESSO, and MR-RAPS (Yavorska and Burgess, 2017; Sanderson et al., 2022). Multivariable MR (MVMR) methods included the inverse-variance weighted, MR-Egger, and least absolute shrinkage and selection operator (LASSO)-based methods (Yavorska and Burgess, 2017; Sanderson et al., 2019). Colocalisation methods Request a detailed protocol Unlike MR, where the goal is to assess the evidence for a causal effect of an exposure on an outcome, colocalisation is agnostic with respect to direction of effect and only assesses the probability that the two traits are affected by the same genetic variants at a given locus. Colocalisation can be viewed as a complementary approach for evaluating MR assumptions within specific genes or regions since strong evidence of colocalisation indicates overlap in genetic mechanisms affecting LTL and lung cancer. We used COLOC (v5.1.0; Wallace, 2020) to estimate the posterior probability for two traits sharing the same causal variant (PP4) in a 150 kb LD window, with PP4 >0.70 corresponding to strong evidence of colocalisation, as previously suggested (Wallace, 2020; Lopes et al., 2022). Priors chosen for the colocalisation analyses were p1=10–3, p2=10–4, and p12=10–5, or approximately, a 75% prior belief that a signal will only be observed in the LTL GWAS and less than 0.01% prior belief in favour of colocalisation between the two traits at a given locus (Giambartolomei et al., 2014). Conditioning and masking colocalisation methods were also used as they may identify putative shared causal variants in the presence of multiple causal variants present in a defined LD window (Wallace, 2021). We present the average PP4 from all methods as our posterior belief in favour of colocalisation between LTL and lung cancer risk. Multi-trait colocalisation based on a clustering algorithm was also performed using HyPrColoc (v1.0) to identify shared genetic signals with other lung cancer-related traits (Foley et al., 2021). Principal component analysis based on RNA-sequencing data Request a detailed protocol Read counts of RNA-sequencing data were normalised within (GC-content and gene length) and between (sequencing depth) lane procedures by EDASeq R package (version 2.28.0; Risso et al., 2011) and excluding low read counts. Principal component analysis was applied using singular value decomposition method, after excluding extreme outliers. Pathway analyses were conducted using Gene Set Enrichment Analysis software (GSEA, version 4.2.3; Subramanian et al., 2005) on gene annotations from Gene Ontology database. Pathway analyses were restricted to the top 500 genes positively and negatively correlated with each principal component that passed multiple-testing correction (Bonferroni-adjusted p-value<0.05 for 74,465 tests), which is the maximum number of genes supported by the online software. The PRS for LTL was composed of the same 144 variants used in the MR analysis and was computed as the sum of the individual’s beta-weighted genotypes using PRSice-2 software (Choi and O’Reilly, 2019). Associations were estimated per SD increase in the PRS, which was normalised to have a mean of zero across lung adenocarcinoma samples of European ancestry within the TCGA cohort. The associations between the eigenvalues of the gene expression principal components (outcome) and demographic, clinical, and genomic features related to genome stability (predictors derived from TCGA published papers and TCGA data portal, except for the DNA mutational signatures [Gabriel et al., 2022]) were calculated using a multivariate linear regression model. Inferring PC2 gene expression signature based on RNA-sequencing data Request a detailed protocol The TCGA-lung adenocarcinoma (LUAD) tumour samples were split into training (70%, N=255) and validation (30%, N=108) datasets. Principal components analysis based on RNA-sequencing data was performed to summarise the gene expression profiles of lung adenocarcinoma tumours into five principal components in the training and validation datasets separately, as previously described (see ‘Principal component analysis based on RNA-sequencing data’ section in methods). Subsequently, we applied the partial least squares-based method called rigid transformation (Hubert and Branden, 2003) to align the first five principal components in both datasets. This method compares embeddings, low-dimensional representations in both datasets, performing a slightly rotation of principal components in order to translate and match them in both training and validation datasets. To select the most informative genes of PC2 in the training dataset, RNA levels of the genes correlated with PC2 (N=3914 out of 14,893 genes, FDR <0.05 for 14,893 tests) were selected as variables for the LASSO regression models. The LASSO tune parameters were chosen by resampling the training dataset (1000 bootstraps: root mean of SE=0.12 ± 0.0004, Lambda = 10 x 10–10, r2=0.99 ± 0.00007) using the tidymodels metapackage in R (v1.0.0; wrapper of glmnet). The 10 genes selected by the LASSO model were used to infer the gene expression signature of PC2 by adding up the scaled values of the log-normalised read counts of each gene multiplied by the respective LASSO regression coefficients. For validation purpose, the inferred PC2 signature was calculated in the validation dataset and compared with the observed principal components. After validation in the subset of the TCGA-LUAD cohort, the inferred PC2 signature was calculated in TCGA-LUSC dataset to compare differences between lung cancer histological subtypes. Results Genome-wide genetic correlations The design of the study is represented in Figure 1—figure supplement 1. We first assessed the shared genetic basis of telomere length, lung cancer risk, and other putative lung cancer risk factors, such as smoking behaviours (age start smoking, smoking cessation, smoking initiation, and cigarettes per day) and lung function (FEV1 and FVC) using genome-wide correlations (Figure 1A). There was little evidence for genetic correlations by LDSC between LTL variants and lung cancer (rg = −0.01, p=0.88) or when stratified by histologic subtypes (Figure 1A). Increasing LTL was genetically correlated with older age at smoking initiation (rg = 0.13, p=3.0 × 10–3), and negatively correlated with smoking cessation: (rg = −0.21, p=6.9 × 10–09), smoking initiation (rg = −0.16, p=1.3 × 10–10), and cigarettes per day (rg = −0.19, p=2.1 × 10–08). Longer LTL was genetically correlated with improved lung function, as indicated by increasing values of FEV1 (rg = 0.09, p=5.1 × 10–07) and FVC (rg = 0.09, p=1.1 × 10–05). To better understand the absence of genome-wide correlations between LTL and lung cancer, we visualised the Z-scores for each trait for approximately 1.2 million variants included in the LDSC analyses (Figure 1B). A subgroup of variants associated with longer LTL was correlated with increased lung adenocarcinoma risk, while the subgroup of smoking-behaviour associated variants, which also conferred an increased risk of lung adenocarcinoma, tended to have lower LTL. Figure 1 with 1 supplement see all Download asset Open asset Genetic correlations between leukocyte telomere length (LTL) and lung cancer (LC) related traits. (A) Heatmap representing the genetic correlation analyses (rg) for LTL across LC, histological subtypes (lung adenocarcinoma [ADE], squamous cell carcinoma [SQC], and small-cell carcinoma [SCC]), smoking propensity (cigarettes per day [CPD], smoking cessation [SmkCes], Smoking initiation [SmkInit], and age of smoking initiation [AgeSmk]), and lung function related (forced vital capacity [FVC] and forced expiratory volume [FEV1]) traits. The black star indicates correlations that passed Bonferroni correction (p<4x10–04). Heritability (h2) as the proportion of the phenotypic variance caused by SNPs. (B) Plot of Z-scores (ADE versus LTL), restricting to the Hapmap SNPs (~1.2 million) but excluding HLA region. Genome-wide significant SNPs (p<5x10–08) for each trait were coloured (CPD in red, SmkInit in dark red, LTL in dark blue, AgeSmk in blue, SmkCes in lightblue, and not genome-wide hits for LTL or any other selected trait in white). Linear regression line was coloured in red. MR analyses From the 490 genetic instruments associated with LTL at genome-wide significance (p<5x10–08), 144 LTL genetic instruments, that explained ~3.5% of the variance in LTL, and were in low-linkage disequilibrium (r2<0.01) were used in MR analysis (Supplementary file 1a). As a sensitivity analysis, a PRS composed of these genetic instruments was associated with TL estimated from WGS in blood samples across TCGA cohorts (Beta = 0.03, 95%CI = 0.01–0.05, p=0.001) but was not associated with TL in tumour material from the same patients (Figure 2—figure supplement 1). MR analyses demonstrated that longer genetically predicted LTL was associated with increased lung cancer risk (OR = 1.62, 95% CI = 1.44–1.84, p=9.91 × 10–15) (Figure 2; Supplementary file 1). Longer LTL conferred the largest increase in risk for lung adenocarcinoma tumours (OR = 2.43, 95% CI = 2.02–2.92, p=3.76 × 10–21), but there was limited evidence of a causal relationship for other histologic subtypes, such as squamous cell carcinoma (OR = 1.00, 95% CI = 0.84–1.19, p=0.98) and small-cell carcinoma (OR = 1.13, 95% CI = 0.87–1.45, p=0.34; Figure 2, Supplementary file 1). However, our study was underpowered to detect an association between lung small-cell carcinoma and LTL at OR of 1.13 and considering alpha type-1 error rate of 5% (Figure 2—figure supplement 1). When stratifying the analyses by smoking status, LTL was associated with lung cancer risk in both never (OR = 2.02, 95% CI = 1.45–2.83, p=3.78 × 10–05) and ever smokers (OR = 1.54, 95% CI = 1.34–1.76, p=7.75 × 10–10; Figure 2, Supplementary file 1). Evidence for negative pleiotropy (supplementary file 1c) and heterogeneity (supplementary file 1d) was observed for all lung cancer outcomes except for squamous cell carcinoma. However, a significant association for LTL and lung cancer risk was found for methods robust to the significant directional pleiotropy (MR-Egger: lung cancer [OR = 2.35, p=3.37 × 10–13]; lung adenocarcinoma [OR = 4.48, p=7.30 × 10–17]; never smokers [OR = 6.84, p=2.07 × 10–10]; supplementary file 1b). Leave-one-out analyses detected only one outlier, rs7705526 in TERT, resulting in >10% change in MR effect size for associated lung cancer subtypes (supplementary file 1e). MVMR analyses considering instruments related to LTL and WHR, HDL, total triglycerides, systolic blood pressure, smoking, and alcohol intake, as well as multiple traits combined, suggested that the association between LTL and lung adenocarcinoma risk is independent of smoking propensity, obesity-related, and alcohol intake-related traits (supplementary file 1f, 1g). Figure 2 with 1 supplement see all Download asset Open asset Genetically predicted leukocyte telomere length (LTL) association with lung cancer. Lung cancer (by histology or by smoking status) risk associations with the LTL instrument from the inverse-variance-weighted MR analyses are expressed as OR per SD increase in genetically predicted LTL. Statistically significant associations with p-values<0.05 (red square). Heterogeneity is estimated by the statistic I2, tau variance of subgroups (τ2), and p-values for Cochran’s Q heterogeneity measure. Colocalisation analyses We investigated whether there was evidence of shared genetic signals between LTL and lung adenocarcinoma at loci centred on the 144 genetic instruments used in MR analyses using colocalisation methods (Figure 3A, supplementary file 1h). Loci with evidence of colocalisation between LTL and lung adenocarcinoma tended to be near genes that encode telomerase subunits and its associated complex, including genetic variants at TERT (5p15.33; rs116539972, rs7705526, rs61748181, rs71593392, and rs140648021), TERC (3q26.2; rs12638862 and rs146546514), and OBFC1 (10q24.33; rs9419958 and rs139122544). Several colocalised loci mapped to genes that have not been previously linked to lung cancer risk at genome-wide significant level: MPHOSPH6 (16q23.3; rs2303262), PRPF6 (20q13.33; rs80150989), and POLI (18q21.2; rs2276182). Other telomere maintenance genes showed limited evidence of colocalisation with lung adenocarcinoma (i.e. TERF1 and PIF1). For instance, while the RTEL1 locus (20q13.33: rs117238689, rs115610405, rs35640778, and rs35902944) harboured variants associated with both LTL and lung adenocarcinoma (Figure 3—figure supplement 1), these signals appeared to be distinct and independent of each other (avg_PP3=0.999, avg_PP4=0.001; Figure 3A, supplementary file 1h). Figure 3 with 1 supplement see all Download asset Open asset Colocalisation analyses for the genetic loci defined by the 144 leukocyte telomere length (LTL) variants. (A) Distribution of the average posterior probability for shared genetic loci between LTL and lung adenocarcinoma, highlighting in orange the telomere maintenance loci that colocalised (avg_PP4≥0.70) and in blue the ones where there was limited evidence for colocalisation (avg_PP4<0.70). Dashed red line represents the arbitrary avg_PP4 cutoff of 0.70. Representative stack plots for the multi-trait colocalisation results within (B) MPHOSPH6 and (C) OBFC1 loci, centred on a 150 kb LD window of rs2303262 and rs9419958 variants, respectively. Left Y-axis represents the –log10(p-values) of the association in the respective genome-wide association study for a given trait. The right Y-axis represents the recombination rate for the genetic loci. The X-axis represents the chromosome position. SNPs are coloured by the linkage disequilibrium correlation threshold (r2) with the query labelled SNP in European population. Sentinel SNPs within the defined LD window were labelled in each trait. We further evaluated whether the loci colocalised between LTL and lung adenocarcinoma also shared genetic signals with other traits related to lung cancer susceptibility (supplementary file 1i). Multi-trait analyses at the 16q23.3 locus colocalised rs2303262 with MPHOSPH6 expression in lung tissue, FVC and FEV1, but not with any of the traits related to smoking behaviour (p=0.72; Figure 3B, supplementary file 1i). We additionally identified evidence of colocalisation (p=0.74) between lung adenocarcinoma, LTL, and gene expression in lung epithelial cells for two variants at the OBFC1 locus: rs139122544 and rs9419958 (Figure 3C, supplementary file 1i). Genetically predicted LTL association with tumour features We investigated the impact of genetically predicted LTL on lung adenocarcinoma tumour features by estimating molecular expression patterns within 343 lung adenocarcinomas tumours using principal component analysis in RNA-sequencing data. The first five components explained ~54% of the observed variance in the RNA-sequencing data (Figure 4, Figure 4—figure supplement 1). To explore the biological meaning of the five components, we performed pathway analyses for the top 500 genes with the highest loadings in each component (supplementary file 1j, supplementary file 1k). Overall, the genes correlated with each component tended to be enriched for specific cell signaling pathways (PC1: RNA processing; PC2: cell-cycle; PC3: metabolic processes; PC4: immune response; PC5: cellular response to stress and DNA damage; false discovery rate <5%; supplementary file 1l). Figure 4 with 1 supplement see all Download asset Open asset Associations between molecular expression patterns of lung adenocarcinoma tumours, LTL PRS, and The Cancer Genome Atlas (TCGA) features. (A) LTL PRS association with the first five principal components based on RNA-sequencing data of lung adenocarcinomas tumours (n=343). Results are expressed as beta estimate per SD increase in genetically predicted LTL. Linear regression model adjusted by sex, age, smoking status, and PC1-5 (genetic ancestry) covariates. Statistically significant associations with p-values<0.05 (red square). (B) Heatmap representing the correlations among PC2 and selected molecular features related to telomere length canonical roles. LTL = leukocyte telomere length; PRS = polygenic risk score; PC = principal component; TMB = tumour total mutation burden; HRD = homologous recombination deficiency, SBS (single base substitution DNA mutational signatures). SBS1 and SBS5 are DNA mutational signatures associated with age-related processes, and SBS4 is associated with tobacco smoking exposure. X-shaped marker to cross correlations with p-value>0.05. We then tested the association between the PRS composed of the 144 genetic instruments selected for MR analysis and the five components of gene expression within lung adenocarcinoma tumours (Figure 4A). The LTL PRS was positively associated with the second component (PC2) of tumour expression (Beta = 0.17, 95% CI = 0.12–0.19, p=1.0 × 10–3; Figure 4A). In multivariate analysis, higher values of PC2 tended to be associated with patients older at diagnosis (p=0.001), female (p=0.005), being never smokers (p=0.04), and diagnosed with early-stage tumours (p=0.002; Table 1). PC2 was also highly correlated with gene expression-based measure of cell proliferation and several genomic features related to genomic stability (Figure 4B). In multivariate analysis, higher values of PC2 were associated with reduced tumour proliferation (p=3.7 × 10–30), lower somatic copy number alternations (p=1.6 × 10–05), and higher tumour telomerase activity scores (p=1.6 × 10–5). Multivariate analysis also indicated that LTL PRS remained an independent predictor of PC2 when considering these genomic features (p=0.009; Table 1). It is noteworthy only nominal associations between LTL PRS and above-mentioned features and none remained statistically significant after correction for multiple testing (supplementary file 1m). We next inferred the gene expression signature of PC2, based on 10 genes informative of this component selected by the LASSO regression models, in both lung adenocarcinoma (TCGA-LUAD) and squamous cell carcinoma (TCGA-LUSC) cohorts (Figure 5, Figure 5—figure supplement 1). The association between LTL PRS and inferred PC2 was observed in TCGA-LUAD (p=0.001) but not in TCGA-LUSC (p=0.729) cases (Figure 5A). The inferred PC2 signature levels were higher in TCGA-LUAD than in TCGA-LUSC (Figure 5B), while higher proliferation rate (Figure 5C, p=1.45 × 10–141) and TERT activity (Figure 5D, p=1.36 × 10–20) were observed in TCGA-LUSC than in TCGA-LUAD cases. Of note, the low RNA levels of the telomere-related genes (less than five read counts), such as TERT and TERC, in both TCGA-LUAD and TCGA-LUSC tumour samples limited the direct comparison of these genes between these cohorts. Figure 5 with 1 supplement see all Download asset Open asset Comparing inferred PC2 gene expression signature by lung cancer histological subtypes. (A) Leukocyte telomere length (LTL) po