Assessing the Relationship Between Leukocyte Telomere Length and Cancer Risk/Mortality in UK Biobank and TCGA Datasets With the Genetic Risk Score and Mendelian Randomization Approaches

Abstract

Background

Telomere length is an important indicator of tumor progression and survival for cancer patients. Previous work investigated the associations between genetically predicted telomere length and cancers; however, the types of cancers investigated in those studies were relatively limited or the telomere length-associated genetic variants employed often came from genome-wide association studies (GWASs) with small sample sizes.

Methods

We constructed the genetic risk score (GRS) for leukocyte telomere length based on 17 associated genetic variants available from the largest telomere length GWAS up to 78,592 individuals. Then, a comprehensive analysis was undertaken to evaluate the association between the constructed GRS and the risk or mortality of a wide range of cancers [i.e., 37 cancers in the UK Biobank and 33 cancers in The Cancer Genome Atlas (TCGA)]. We further applied the two-sample Mendelian randomization (MR) to estimate the causal effect of leukocyte telomere length on UK Biobank cancers via summary statistics.

Results

In the UK Biobank dataset, we found that the GRS of leukocyte telomere length was associated with a decreased risk of nine types of cancer (i.e., significant association with multiple myeloma, chronic lymphocytic leukemia, kidney/renal cell cancer, bladder cancer, malignant melanoma, basal cell carcinoma, and prostate cancer and suggestive association with sarcoma/fibrosarcoma and Hodgkin’s lymphoma/Hodgkin’s disease). In addition, we found that the GRS was suggestively associated with an increased risk of leukemia. In the TCGA dataset, we observed suggestive evidence that the GRS was associated with a high death hazard of rectum adenocarcinoma (READ), sarcoma (SARC), and skin cutaneous melanoma (SKCM), while the GRS was associated with a low death hazard of kidney renal papillary cell carcinoma (KIRP). The results of MR further supported the association for leukocyte telomere length on the risk of malignant melanoma, Hodgkin’s lymphoma/Hodgkin’s disease, chronic lymphocytic leukemia and multiple myeloma.

Conclusion

Our study reveals that telomere played diverse roles in different types of cancers. However, further validations in large-scale prospective studies and deeper investigations of the biologic mechanisms are warranted.

Introduction

Telomere is a special structure with a 6-bp TTAGGG repeat sequence and plays an important role in genomic stability by protecting DNA against damage and fusion 0 (de Lange, 2005). Due to the inability of DNA polymerase to fully extend the 3′ end of DNA strand, the telomere becomes progressively shorter during each round of cell division. The length of telomere is thus a biomarker of cellular and overall biological aging. Once a critically short telomere length is reached, the cell would be triggered to enter senescence, which would ultimately lead to cell growth arrest or apoptosis (Shay and Wright, 2019). In stem and progenitor cells, the length of telomere is maintained by enzyme telomerase (Hackett and Greider, 2002; Shawi and Autexier, 2008). It is shown that enzyme telomerase is activated in almost all human tumors; such an activation can result in the continuous division of cancer cells and is the key component of the tumorigenic phenotype of human cancer cells (Stewart and Weinberg, 2006; O’Sullivan and Karlseder, 2010).

Prior studies have demonstrated that telomere length is associated with a lot of age-related diseases and disorders (e.g., cancers and neurodegenerative disorders) (Zhu et al., 2011) and that a shorter telomere length in tumor tissues is an important indicator of tumor progression and survival for cancer patients (Ma et al., 2011; Xu et al., 2016). However, not all studies reported consistent findings (Supplementary Table S1), partly reflecting the complicated function of telomere on human cancers. The diversity in cancer types, ethnicities, study designs, measurement methods, and selected tissues for telomere length in previous work further complicates the observed association. Given the severe disease burden of cancers worldwide (Siegel et al., 2019), understanding the association between telomere length and cancers can provide valuable insights into the development of cancers and has the potential to improve the prevention and treatment strategies for cancers.

On the other hand, in the past few years, a number of single nucleotide polymorphisms (SNPs) have been identified to be associated with leukocyte telomere length through genome-wide association studies (GWASs) (Levy et al., 2010; Gu et al., 2011; Mangino et al., 2012; Codd et al., 2013; Pooley et al., 2013; Dorajoo et al., 2019). Relying on associated genetic variants, many studies have been undertaken to investigate the association between genetically predicted leukocyte telomere length and cancers. However, the types of cancers investigated in previous studies (Zhang et al., 2015; Li et al., 2020) were relatively limited. In addition, the telomere length-associated SNPs employed in previous studies (Zhang et al., 2015; Rode et al., 2016; Haycock et al., 2017) often came from GWASs with small sample sizes (Levy et al., 2010; Codd et al., 2013).

Recently, a large-scale GWAS of leukocyte telomere length was conducted with the largest sample size to date (up to ∼80,000) (Li et al., 2020), which allows us to choose more appropriate SNPs to study the multilocus genetic profile of leukocyte telomere length via the genetic risk score (GRS) approach (Ripatti et al., 2010; Dudbridge et al., 2013; Eusden et al., 2015; Guo et al., 2016; Goldman, 2017; Tosto et al., 2017; Bogdan et al., 2018; De La Vega and Bustamante, 2018; Zeng et al., 2019b). Briefly, GRS is an efficient and powerful genetic method to explore the association between an exposure and complex diseases by integrating multiple genetic variants with weak effects, and it dramatically enhances the predictability of complex diseases through genetic polymorphisms (Belsky et al., 2013; Khera et al., 2018; Duncan et al., 2019; Khera et al., 2019). Moreover, several cancer-relevant cohorts, such as The UK Biobank (Bycroft et al., 2018) and The Cancer Genome Atlas (TCGA) (Hoadley et al., 2018), have collected a variety of cancer-related omics and clinical information, which makes it feasible to systematically investigate a large number of types of cancers.

Based on these valuable data resources, in the present work, we evaluated the association between leukocyte telomere length and 37 cancers from the UK Biobank cohort as well as 33 cancers from the TCGA dataset using the genetic risk score method. We further applied the two-sample Mendelian randomization (MR) (Burgess et al., 2017; Hartwig et al., 2017) to assess the association between leukocyte telomere length and multiple cancers, for which the summary statistics can be available from the UK Biobank cohort. Our study revealed that telomere played cancer-specific roles and that a shorter leukocyte telomere length can either increase or decrease the risk/mortality of cancers. However, further validations in large-scale prospective studies and deeper investigations of the biological mechanism of leukocyte telomere length on various types of cancers are warranted.

Materials and Methods

Selection of Instrumental Variables for Leukocyte Telomere Length

We obtained the summary statistics (e.g., effect size and effect allele) of leukocyte telomere length from the ENGAGE consortium as well as the EPIC-CVD and EPIC-InterAct cohorts (Supplementary Table S2; Li et al., 2020), which was the largest GWAS of telomere length (N = 78,592) undertaken in the European population to date. In this study, leukocyte telomere length was measured as a continuous variable and the linear additive regression was implemented to investigate the association for each genetic variant (Li et al., 2020). Particularly, in the association analysis, the age of participants was considered as a covariate to remove the influence of biological age. We selected 17 independent index SNPs that were strongly associated with leukocyte telomere length (p < 5.00E-8; see Table 1) to construct GRS. Note that, given the fact that the length of telomere would shorten progressively with age, to facilitate the explanation of our results, we made a sign transformation for the effect sizes of these used SNPs so that the relationship under investigation corresponded to a shorter leukocyte telomere length.

TABLE 1.

Independent index single nucleotide polymorphisms (SNPs) associated with leukocyte telomere length in the European population.

SNP	Chr	Position	Gene	A1/A2	EAF	Beta	SE	p	PVE	F
rs3219104	1	226,562,621	PARP1	C/A	0.83	–0.042	0.006	9.60E-11	6.23E-04	49.0
rs55749605	3	101,232,093	SENP7	A/C	0.58	0.037	0.007	2.45E-08	3.55E-04	27.9
rs10936600	3	169,514,585	TERC	T/A	0.24	0.086	0.006	7.18E-51	2.61E-03	205.4
rs13137667	4	71,774,347	MOB1B	C/T	0.96	–0.077	0.014	2.43E-08	3.85E-04	30.2
rs4691895	4	164,048,199	NAF1	C/G	0.78	–0.058	0.006	1.58E-21	1.19E-03	93.4
rs7705526	5	1,285,974	TERT	A/C	0.33	–0.082	0.006	5.34E-45	2.37E-03	186.8
rs34991172	6	25,480,328	CARMIL1	G/T	0.07	0.061	0.011	6.19E-09	3.91E-04	30.8
rs2736176	6	31,587,561	PRRC2A	C/G	0.31	–0.035	0.006	3.53E-10	4.33E-04	34.0
rs59294613	7	124,554,267	POT1	A/C	0.29	0.041	0.006	1.17E-13	5.94E-04	46.7
rs9419958	10	105,675,946	OBFC1	C/T	0.86	0.064	0.007	5.05E-19	1.06E-03	83.6
rs228595	11	108,105,593	ATM	A/G	0.42	0.029	0.005	1.43E-08	4.28E-04	33.6
rs2302588	14	73,404,752	DCAF4	C/G	0.10	–0.048	0.008	1.68E-08	4.58E-04	36.0
rs3785074	16	69,406,986	TERF2	G/A	0.26	–0.035	0.006	4.64E-10	4.33E-04	34.0
rs62053580	16	74,680,074	RFWD3	G/A	0.17	0.039	0.007	4.08E-08	3.95E-04	31.0
rs7194734	16	82,199,980	MPHOSPH6	T/C	0.78	0.037	0.006	6.94E-10	4.84E-04	38.0
rs8105767	19	22,215,441	ZNF208	G/A	0.30	–0.039	0.005	5.42E-13	7.74E-04	60.8
rs75691080	20	62,269,750	STMN3	T/C	0.09	0.067	0.009	5.99E-14	7.05E-04	55.4

Chr, chromosome; A1, effect allele; A2, alternative allele; EAF, frequency of the effect allele; PVE, proportion of variance explained by the SNP [i.e., ${\text{PVE}}_j\mathrm{\_}={({\widehat{\mathrm{\beta }}}_j^{\text{X}})}^2/({({\widehat{\mathrm{\beta }}}_j^{\text{X}})}^2+\mathrm{var}({\widehat{\mathrm{\beta }}}_j^{\text{X}})\times N_j)$, where ${\widehat{\mathrm{\beta }}}_j^{\text{X}}$ and $\text{var}({\widehat{\mathrm{\beta }}}_j^{\text{X}})$ are the estimated effect size and variance, respectively, for instrument j (Shim et al., 2015)]; F, F statistic [i.e., F_j=PVE_j(N_j−1−k)/(k−k×PVE_j), where Nj is the sample size for instrument j (i.e., Nj = 78,592) and k is the number of instruments (Burgess et al., 2011; Burgess and Thompson, 2012). Both PVE and F statistic are calculated to validate the issue of weak instruments].

Construction of Genetic Risk Score

The genetic risk score for leukocyte telomere length is calculated in a weighted way (Ripatti et al., 2010; Guo et al., 2016; Zeng et al., 2019b).

$$\text{GRS}=\sum _{j=\mathrm{\hspace{0.33em}1}}^{17}{G}_j{\widehat{\mathrm{\beta }}}_j$$ (1)

where ${\widehat{\mathrm{\beta }}}_j$ is the estimated marginal SNP effect on the shorter leukocyte telomere length for the jth selected index SNP (e.g., Table 1) (Li et al., 2020). G_j is the individual-level genotype of the same SNP in the UK Biobank (Bycroft et al., 2018) or TCGA dataset (Hoadley et al., 2018) and is coded to be 0, 1, and 2, representing the number of effect allele. Following prior work (Zeng et al., 2019b), we do not directly rescale the GRS as its p-value would not be altered regardless of whether the GRS is scaled or not. We instead standardize the GRS so that its mean is zero and the variance is equal to 1.

Two-Stage Regression Model in the UK Biobank and TCGA Using GRS

To link GRS with the risk of cancers from the UK Biobank (Table 2; Bycroft et al., 2018), we apply an additive logistic regression while adjusting for a set of available covariates (i.e., age, gender, smoke, drink, and BMI).

TABLE 2.

Association between the genetic risk score (GRS) of leukocyte telomere length and the risk of 37 UK Biobank cancers.

Types of cancer	OR (95%CI)	p	FDR	Case	M/F	Age (years)
Leukemia	1.20 (1.02–1.41)	0.025	0.058	147	79/68	67.99 ± 8.17
Rectal cancer	1.10 (0.96–1.25)	0.165	0.193	231	134/97	70.64 ± 6.17
Tongue cancer	1.06 (0.88–1.29)	0.526	0.407	102	65/37	68.89 ± 7.35
Squamous cell carcinoma	1.04 (0.93–1.15)	0.514	0.401	332	168/164	70.89 ± 6.21
Testicular cancer	1.02 (0.95–1.11)	0.549	0.417	595	595/0	64.78 ± 8.02
Primary bone cancer	1.02 (0.81–1.29)	0.845	0.524	72	44/28	67.56 ± 8.02
Non-melanoma skin cancer	1.02 (0.93–1.12)	0.648	0.458	472	280/192	70.01 ± 6.97
Large bowel cancer/Colorectal cancer	1.02 (0.93–1.12)	0.739	0.490	440	260/180	71.54 ± 5.80
Rodent ulcer	1.01 (0.92–1.11)	0.893	0.538	437	203/234	70.50 ± 5.77
Esophageal cancer	1.01 (0.85–1.19)	0.946	0.552	137	110/27	71.98 ± 6.30
Cervical cancer	1.01 (0.95–1.06)	0.857	0.527	1273	0/1,273	66.16 ± 7.65
Non-Hodgkin’s lymphoma	0.99 (0.92–1.08)	0.945	0.552	593	355/238	69.40 ± 7.34
Pre-cancer cells cervix	0.99 (0.94–1.06)	0.922	0.546	1117	1/1,116	63.99 ± 7.98
Breast cancer	0.98 (0.96–1.01)	0.164	0.193	7330	37/7,293	70.08 ± 6.48
Colon cancer/sigmoid cancer	0.97 (0.92–1.04)	0.399	0.342	1055	631/424	72.30 ± 5.68
Uterine/endometrial cancer	0.95 (0.89–1.02)	0.176	0.200	752	0/752	71.27 ± 5.89
Ovarian cancer	0.95 (0.87–1.03)	0.222	0.224	512	0/512	69.11 ± 7.33
Brain cancer/primary malignant brain tumor	0.95 (0.80–1.13)	0.539	0.412	128	62/66	64.77 ± 8.81
Prostate cancer	0.94 (0.91–0.98)	0.005	0.020	2410	2,410/0	73.96 ± 4.08
Skin cancer	0.94 (0.88–1.00)	0.065	0.106	943	478/465	71.21 ± 6.38
Basal cell carcinoma	0.94 (0.90–0.97)	0.001	0.010	2916	1,206/1,710	70.02 ± 6.84
Stomach cancer	0.94 (0.77–1.14)	0.516	0.402	96	56/40	71.33 ± 6.22
Malignant melanoma	0.91 (0.88–0.95)	4.57E-06	9.56E-05	2526	1,031/1,495	68.95 ± 7.41
Larynx/throat cancer	0.91 (0.80–1.04)	0.161	0.191	228	190/38	71.12 ± 6.43
Bladder cancer	0.91 (0.84–0.98)	0.010	0.030	725	548/177	72.30 ± 5.87
Eye and/or adnexal cancer	0.90 (0.74–1.11)	0.325	0.297	95	44/51	68.82 ± 7.48
Thyroid cancer	0.90 (0.80–1.01)	0.067	0.108	293	52/241	67.27 ± 7.63
Small intestine/small bowel cancer	0.90 (0.76–1.06)	0.206	0.216	133	77/56	72.28 ± 5.55
Hodgkin’s lymphoma/Hodgkin’s disease	0.89 (0.79–0.99)	0.033	0.069	321	184/137	64.96 ± 8.13
Chronic myeloid leukemia	0.88 (0.71–1.10)	0.273	0.262	81	44/37	68.35 ± 7.98
Lung cancer	0.88 (0.74–1.06)	0.172	0.197	123	82/41	72.60 ± 5.70
Kidney/renal cell cancer	0.86 (0.78–0.95)	0.003	0.017	401	261/140	70.22 ± 6.36
Cancer of lip/mouth/pharynx/oral/cavity	0.86 (0.68–1.09)	0.213	0.220	69	43/26	70.42 ± 5.98
Sarcoma/fibrosarcoma	0.84 (0.72–0.98)	0.028	0.063	164	76/88	66.73 ± 7.58
Chronic lymphocytic leukemia	0.82 (0.71–0.94)	0.005	0.020	206	131/75	71.28 ± 6.28
Lymphoma	0.80 (0.64–1.01)	0.057	0.098	78	51/27	68.69 ± 8.27
Multiple myeloma	0.77 (0.63–0.93)	0.006	0.021	108	62/46	70.37 ± 7.08

The cancers were sorted by the estimated odds ratios (ORs). CI, confidence internal; p, the original p-value; FDR, false discovery rate; M, male; F, female. In bold are significant (i.e., FDR < 0.05) or suggestive associations (i.e., p < 0.05).

$$\mathrm{logit}({\mathrm{\mu }}_i)={\text{GRS}}_i\times \mathrm{\theta }+X_i^T\mathrm{\alpha }$$ (2)

where μ_i is the expectation of y_i, with y_i = 1 or 0 representing the status of individual i with or without cancer; θ is the effect size of GRS; and X_i is the vector of standardized covariates with effect sizes α. Of note, we assume that all of the entries in the first column of X are 1, representing the intercept term.

We next evaluate the effect of GRS on the mortality of cancers from TCGA (Table 3; Hoadley et al., 2018) with the Cox proportional hazards model (Cox, 1972) while controlling for available clinical covariates (i.e., age at diagnosis, gender, and stage).

TABLE 3.

Association between the genetic risk score (GRS) of leukocyte telomere length and the mortality of 33 TCGA cancers.

Cancer	HR (95%CI)	p	FDR	N	Median survival time			M/F	Age at diagnosis (years)	Stage or grade (1/2/3/4/5)
					All	Event	Censor
DLBC	2.24 (0.88–5.67)	0.090	0.317	42	31.85	19.83	32.4	19/23	55.33 ± 14.39	8/17/5/12
PCPG	2.16 (0.95–4.92)	0.068	0.283	178	25.28	15.08	25.6	78/100	47.30 ± 15.12	NA
READ	1.72 (1.09–2.73)	0.020	0.138	157	21.2	24.33	21.02	85/72	64.34 ± 11.67	30/51/51/25
UVM	1.47 (0.94–2.30)	0.092	0.320	79	25.77	19.68	27.37	44/35	61.68 ± 13.94	0/39/36/4
PRAD	1.44 (0.72–2.87)	0.306	0.610	501	30.8	29.17	30.87	501/0	60.93 ± 6.81	NA
SARC	1.29 (1.06–1.58)	0.011	0.138	260	31.77	21.6	36.4	119/141	60.80 ± 14.61	NA
ESCA	1.28 (0.99–1.66)	0.063	0.274	162	13.57	13.38	13.57	137/25	62.40 ± 11.74	18/79/56/9
TGCT	1.23 (0.10–15.59)	0.870	0.816	81	37.53	116.48	37.53	81/0	32.85 ± 10.18	55/12/14/0
SKCM	1.19 (1.03–1.37)	0.018	0.138	411	33.2	31.93	34.5	256/155	58.82 ± 15.51	77/140/171/23
KICH	1.17 (0.46–2.99)	0.743	0.792	65	74.93	28.5	90.43	38/27	51.15 ± 13.99	20/25/14/6
CESC	1.14 (0.90–1.45)	0.274	0.583	295	21.27	20.23	23.12	0/295	47.88 ± 13.47	160/69/46/20
THCA	1.12 (0.66–1.89)	0.676	0.776	503	31.67	34.03	31.47	136/367	47.28 ± 15.78	284/52/113/54
BRCA	1.11 (0.93–1.32)	0.258	0.569	924	26.38	44.13	24.23	0/924	58.84 ± 13.14	156/523/219/14/12
LUSC	1.06 (0.93–1.21)	0.378	0.659	487	21.77	18.13	24.58	359/128	67.31 ± 8.58	239/157/84/7
MESO	1.05 (0.82–1.35)	0.705	0.783	86	17.1	15.23	38.93	70/16	63.08 ± 9.72	10/16/44/16
LUAD	1.04 (0.90–1.21)	0.584	0.749	503	21.87	20.47	22.33	232/271	65.16 ± 10.07	277/121/80/25
UCEC	1.04 (0.84–1.28)	0.742	0.791	546	30.47	23.63	32.2	0/546	63.99 ± 11.13	338/52/127/29
KIRC	1.04 (0.90–1.20)	0.639	0.766	532	39.2	27.35	48.1	342/190	60.57 ± 12.07	267/57/125/83
HNSC	1.03 (0.90–1.19)	0.655	0.770	450	21.37	14.08	27.4	324/126	60.90 ± 12.13	27/73/82/268
LIHC	1.00 (0.83–1.21)	0.963	0.831	350	19.43	13.67	21.47	239/111	59.03 ± 13.30	174/86/85/5
GBM	1.00 (0.91–1.10)	0.949	0.829	595	12.27	12.7	8.67	364/231	57.87 ± 14.41	NA
ACC	0.99 (0.68–1.44)	0.967	0.832	88	37.93	18.38	48.45	29/59	47.07 ± 16.43	9/43/18/18
OV	0.98 (0.88–1.08)	0.629	0.763	569	33.57	35.77	28.57	0/569	59.71 ± 11.46	16/30/437/86
PAAD	0.97 (0.80–1.19)	0.792	0.802	182	15.55	13.13	16.92	100/82	64.92 ± 11.06	21/152/4/5
STAD	0.96 (0.83–1.11)	0.576	0.746	407	14.53	11.6	18.87	260/147	65.37 ± 10.70	55/128/181/43
BLCA	0.96 (0.82–1.12)	0.576	0.746	411	17.87	13.68	21.27	303/108	68.10 ± 10.58	3/131/141/136
COAD	0.93 (0.76–1.13)	0.448	0.696	458	22.32	13.47	24.33	239/219	67.03 ± 13.06	79/183/131/65
LGG	0.89 (0.73–1.09)	0.270	0.580	512	22.47	27.13	20.97	284/228	42.99 ± 13.34	0/247/265/0
LAML	0.89 (0.73–1.09)	0.252	0.563	186	12.17	9.1	23.3	102/84	55.53 ± 16.06	NA
UCS	0.85 (0.59–1.23)	0.395	0.669	56	20.25	16.72	27.6	0/56	69.38 ± 8.89	21/5/20/10
THYM	0.85 (0.41–1.76)	0.654	0.770	121	41.77	28.43	42.33	62/59	58.37 ± 12.94	37/61/15/8
KIRP	0.66 (0.47–0.93)	0.019	0.138	257	24.67	20.8	25.37	190/67	61.50 ± 12.03	171/20/51/15
CHOL	0.64 (0.38–1.08)	0.097	0.331	36	21.5	16.67	31.42	16/20	63.03 ± 12.67	19/9/1/7

The cancers were sorted by the estimated hazard ratios (HRs). CI, confidence internal; p, the original p-value; FDR, false discovery rate; M, male; F, female. Cancer types: DLBC, lymphoid neoplasm diffuse large B-cell lymphoma; PCPG, pheochromocytoma and paraganglioma; READ, rectum adenocarcinoma; UVM, uveal melanoma; PRAD, prostate adenocarcinoma; SARC, sarcoma; ESCA, esophageal carcinoma; TGCT, testicular germ cell tumor; SKCM, skin cutaneous melanoma; KICH, kidney chromophobe; CESC, cervical squamous cell carcinoma and endocervical adenocarcinoma; THCA, thyroid carcinoma; BRCA, breast invasive carcinoma; LUSC, lung squamous cell carcinoma; MESO, mesothelioma; LUAD, lung adenocarcinoma; UCEC, uterine corpus endometrial carcinoma; KIRC, kidney renal clear cell carcinoma; HNSC, head and neck squamous cell carcinoma; LIHC, liver hepatocellular carcinoma; GBM, glioblastoma multiforme; ACC, adrenocortical carcinoma; OV, ovarian serous cystadenocarcinoma; PAAD, pancreatic adenocarcinoma; STAD, stomach adenocarcinoma; BLCA, bladder urothelial carcinoma; COAD, colon adenocarcinoma; LGG, brain lower grade glioma; LAML, acute myeloid leukemia; UCS, uterine carcinosarcoma; THYM, thymoma; KIRP, kidney renal papillary cell carcinoma; CHOL, cholangiocarcinoma. In bold are suggestive associations (i.e., p < 0.05).

$$h(t_i|{\text{GRS}}_i,X_i)=h_0(t_i)e^{{\text{GRS}}_i\times \mathrm{\theta }+X_i^T\mathrm{\alpha }}$$ (3)

where t_i is the observed survival time and h₀(t) is an arbitrary baseline hazard function. Cancer-specific covariates are considered for some cancers in TCGA [e.g., the status of estrogen and progesterone receptors for breast invasive carcinoma (BRCA)]. In the logistic or Cox model, we are mainly interested in estimating θ and testing for the null hypothesis H₀: θ = 0. We further examine the interaction effect between GRS and each of the clinical covariates (e.g., GRS × gender) if GRS is detected to be associated with some cancer.

Two-Sample MR Analysis

Besides the GRS method, we also perform the two-sample MR analysis to estimate the causal effect of leukocyte telomere length on cancers in the UK Biobank using summary statistics (Sudlow et al., 2015). In observational studies, MR is a flexible approach for causal inference to avert confounding and reverse causality (Zeng et al., 2019a; Yu et al., 2020). In brief, we estimate the causal effect of leukocyte telomere length (again, denoted as θ) relying on all the available instrumental variables (Table 1) through the commonly employed inverse-variance weighted (IVW) method (Burgess et al., 2017; Hartwig et al., 2017).

$$\begin{array}{c}\widehat{\mathrm{\theta }}=\frac{1}{{\sum }_{j=1}^{17}\text{var}{({\widehat{\mathrm{\beta }}}_j^{\text{Y}})}^{-1}{({\widehat{\mathrm{\beta }}}_j^{\text{X}})}^2}\sum _{j=1}^{17}\text{var}{({\widehat{\mathrm{\beta }}}_j^{\text{Y}})}^{-1}{\widehat{\mathrm{\beta }}}_j^{\text{Y}}{\widehat{\mathrm{\beta }}}_j^{\text{X}}\hfill \\ \text{var}(\widehat{\mathrm{\theta }})=\frac{1}{{\sum }_{j=1}^{17}\text{var}{({\widehat{\mathrm{\beta }}}_j^{\text{Y}})}^{-1}{({\widehat{\mathrm{\beta }}}_j^{\text{X}})}^2}\hfill \end{array}$$ (4)

where ${\widehat{\mathrm{\beta }}}_j^{\text{X}}$ and $\text{var}({\widehat{\mathrm{\beta }}}_j^{\text{X}})$ are the effect size and the variance, respectively, of the instrumental variable j for the exposure X (i.e., leukocyte telomere length; Li et al., 2020), and ${\widehat{\mathrm{\beta }}}_j^{\text{Y}}$ and $\text{var}({\widehat{\mathrm{\beta }}}_j^{\text{Y}})$ are the effect size and the variance, respectively, for the same instrumental variable j on the outcome Y (i.e., cancer in the UK Biobank; Sudlow et al., 2015).

To guarantee the validity of our MR analysis, before the formal analysis, we examine the pleiotropic effects of instruments by removing index SNPs that may be potentially related to individual cancers if the Bonferroni-adjusted p-values are less than 0.05. We also conduct a series of sensitivity analyses: (i) weighted median-based (Bowden et al., 2016b) and maximum likelihood methods (Burgess et al., 2013), which are robust when some instrumental variables might be invalid; (ii) MR-Egger regression (Bowden et al., 2016a; Burgess and Thompson, 2017), which guards against horizontal pleiotropic effects; and (iii) leave-one-out (LOO) analysis (Noyce et al., 2017) and Mendelian randomization pleiotropy residual sum and outlier (MR-PRESSO) test (Verbanck et al., 2018) to examine potential instrumental outliers.

UK Biobank and TCGA Cancer Datasets

The UK Biobank dataset consists of approximately 500,000 individuals (Bycroft et al., 2018). We selected age, gender, smoke, drink, and BMI as covariates and originally chose 79 self-reported cancers up to 337,198 independent individuals (28,820 cases and 308,378 controls) of European ancestry, but only included cancers with at least 60 cases (to some extent, this cutoff value was used arbitrarily) and treated cancer-free individuals to be controls. Finally, a total of 37 cancers were left up to 335,036 individuals (27,641 cases for various cancers and 307,395 shared cancer-free controls after removing individuals with missing values). The genotypes were provided by the UK Biobank after the research application was approved. However, we can only obtain 15 SNPs because two were missing (i.e., rs3219104 on PARP1 and rs55749605 on SENP7) in the UK Biobank. In addition, because summary-level statistics are necessary for the two-sample MR analysis, herein we can only consider 28 cancers from the UK Biobank (n = 420,473) (Sudlow et al., 2015; Supplementary Table S6). The summary statistics of these cancers were obtained from https://pan.ukbb.broadinstitute.org/.

Then, we obtained the survival and clinical information of 33 cancers from TCGA (Hoadley et al., 2018). We selected the overall survival time and status as the outcome and primarily included age at diagnosis, gender, and pathologic tumor stage as covariates because many other important clinical covariates were missing for most of the patients. When the pathologic tumor stage cannot be available, we instead employed the clinical stage (i.e., for CESC, DLBC, OV, THYM, UCEC, and UCS) or histological grade (i.e., for LGG). It needs to be stated that all three stage variables were missing in five cancers (i.e., GBM, LAML, PCPG, PRAD, and SARC). For each cancer, we only kept samples from the primary cancer tissue and excluded those with missing values in clinical covariates. More details about these TCGA cancers are demonstrated in Table 3 and Supplementary Table S3. For each cancer, we filtered out SNPs that had a missingness rate >0.95 across individuals, genotype calling rate <0.95, minor allele frequency (MAF) > 0.01, or Hardy–Weinberg equilibrium (HWE) p-value < 10^–4. We next performed an imputation procedure by first phasing the genotypes with SHAPEIT (Delaneau et al., 2013), then imputed the SNPs based on the Haplotype Reference Consortium panel (McCarthy et al., 2016) on the Michigan Imputation Server using minimac3 (Das et al., 2016). The filtering procedure for the imputed genotypes included an HWE p-value < 10^–4, a genotype call rate <95%, a MAF < 0.01, and an imputation score <0.30. After the imputation of genotypes, all of the 17 SNPs were yielded in TCGA.

Power Evaluation

Finally, we performed power calculation to detect a non-zero causal effect for GRS with regards to cancers based on the UK Biobank and TCGA datasets. Firstly, we simulated genotypes for 17 independent SNPs with varying MAFs (Table 1) and then calculated the GRS. Two independent covariates (i.e., one was binary and the other was continuous) were also included, with each having an effect size of 0.5. We generated a case–control variable y with the probability of exp(η)/(1 + exp(η)) and η = GRS × θ + 0.5X₁ + 0.5X₂. We created 2,000,000 individuals to be the population and then randomly sampled 50 (or 100 and 150) cases and 300,000 controls (as well as their GRS and covariates) to be a subset for the final simulation analysis.

Secondly, to simulate survival datasets, we first generated genotypes and calculated the GRS in the same way as described above. Again, two independent covariates were included, with each having an effect size of 0.5. Then, we employed the inverse probability method (Bender et al., 2005) to create survival time which followed a Weibull distribution, with the shape parameter being 1 and the scale parameter being 0.01. The location parameter of this Weibull distribution was determined by the GRS and the two covariates [i.e., μ = exp(η), with η = GRS × θ + 0.5X₁ + 0.5X₂]. The censored rate was fixed to be 50% in a random manner (the high censored rate corresponded to a similar situation observed in the TCGA cancer dataset). The sample size varied from 100, 300, to 500.

In both simulations, the effect size of GRS θ was set to 0.05, 0.10, or 0.20, approximately corresponding to odds ratios (ORs) [or hazard ratio (HR)] of 1.05, 1.10, and 1.20. The simulation was repeated 1,000 times, and the power calculated by the proportion of the p-value of GRS was less than 1.67E-3, approximately equal to the significance level after the Bonferroni correction of 30 types of cancers.

Throughout our study, we utilized the R software (version 3.6.1) to implement all the analyses. The association was declared to be statistically significant if the false discovery rate (FDR) is <0.05 (Benjamini and Hochberg, 1995), while the association was deemed to be suggestive if the unadjusted p-value is <0.05.

Results

Association Between GRS and UK Biobank Cancers

The 17 selected index SNPs collectively explain about 1.37% phenotypic variance of leukocyte telomere length, and all the F statistics are above 10 (ranging from 27.9 to 205.4, with an average of 63.3) (Table 1), largely ruling out the possibility of weak instrument bias (Cragg and Donald, 1993; Burgess et al., 2017; Zeng and Zhou, 2019a). Based on the constructed GRS, we first investigate the association between leukocyte telomere length and the risk of UK Biobank cancers (Table 2). We detect that the GRS of leukocyte telomere length is significantly associated with a decreased risk of seven types of cancers (Table 2), including multiple myeloma [OR = 0.77, 95% confidence interval (CI) = 0.63–0.93, FDR = 0.021], chronic lymphocytic leukemia (OR = 0.82, 95%CI = 0.71–0.94, FDR = 0.020), kidney/renal cell cancer (OR = 0.86, 95%CI = 0.78–0.95, FDR = 0.017), bladder cancer (OR = 0.91, 95%CI = 0.84–0.98, FDR = 0.030), malignant melanoma (OR = 0.91, 95%CI = 0.88–0.95, FDR = 9.56E-05), basal cell carcinoma (OR = 0.94, 95%CI = 0.90–0.97, FDR = 0.010), and prostate cancer (OR = 0.94, 95%CI = 0.91–0.98, FDR = 0.020). Suggestive associations are observed for two types of cancers including sarcoma/fibrosarcoma (OR = 0.84, 95%CI = 0.72–0.98, FDR = 0.063) and Hodgkin’s lymphoma/Hodgkin’s disease (OR = 0.89, 95%CI = 0.79–0.99, FDR = 0.069). In addition, we discover that the GRS of leukocyte telomere length is also marginally related to an increased risk of leukemia (OR = 1.20, 95%CI = 1.02–1.41, FDR = 0.058).

We further examine the interaction effect of GRS and one of the covariates (e.g., age, gender, smoke, drink, or BMI) for each of the 10 cancers. We observe that the interaction term is statistically significant between smoke and GRS for sarcoma/fibrosarcoma (OR = 0.83, 95%CI = 0.71–0.97) as well as between drink and GRS for leukemia (OR = 0.82, 95%CI = 0.69–0.97) (Supplementary Table S4).

Association Between GRS and TCGA Cancers

We now examine the effect size of GRS on 33 TCGA cancers through the Cox proportional hazards model. We observe suggestive evidence that the GRS of leukocyte telomere length is related to a higher death hazard of READ (HR = 1.72, 95%CI = 1.09–2.73, p = 0.020), SARC (HR = 1.29, 95%CI = 1.06–1.58, p = 0.011), and SKCM (HR = 1.19, 95%CI = 1.03–1.37, p = 0.018) and is associated with a lower death hazard of KIRP (HR = 0.66, 95%CI = 0.47–0.93, p = 0.019), suggesting that a genetically decreased leukocyte telomere length can lead to a worse overall survival of READ, SARC, and SKCM while can result in a better overall survival of KIRP. However, all these associations become non-significant after accounting for multiple comparisons (FDR > 0.05). Neither suggestive nor significant associations are identified between GRS and the remaining cancers (Table 3). We further examine the interaction effect of GRS and each of the covariates (e.g., age at diagnosis, gender, or stage) for each of the four cancers. We do not identify any statistically significant interactions (Supplementary Table S5).

Association Between Leukocyte Telomere Length and UK Biobank Cancers Using the Two-Sample MR

With the selected 17 instrumental variables, we further perform MR analysis to investigate the causal effect of leukocyte telomere length on each of the 28 cancers from the UK Biobank. As no evidence of effect heterogeneity is presented across instruments (all the p-values for the Cochran’s Q test are greater than 0.05), thus, only the results estimated via the fixed-effects IVW method are displayed below. Among the 28 cancers, we identify that leukocyte telomere length is associated with a decreased risk of nine cancers (Supplementary Table S6), including basal cell carcinoma, malignant melanoma, skin cancer, bladder cancer, kidney/renal cell cancer, Hodgkin’s lymphoma/Hodgkin’s disease, thyroid cancer, chronic lymphocytic leukemia, and multiple myeloma. We also observe that leukocyte telomere length is associated with an increased risk of leukemia (Supplementary Table S6).

We now validate the observed causal associations shown above through various sensitivity analyses (Supplementary Table S6). Here, we focus on the associations that are significant in all sensitivity analyses (i.e., P_{Weighted  median} and P_Likelihood < 0.05) and have no horizontal pleiotropic effects (i.e., P_{Egger–intercept} > 0.05). Then, four types of cancers are left, including malignant melanoma (OR = 0.58, 95%CI = 0.44–0.79, FDR = 0.004), Hodgkin’s lymphoma/Hodgkin’s disease (OR = 0.30, 95%CI = 0.13–0.69, FDR = 0.008), chronic lymphocytic leukemia (OR = 0.20, 95%CI = 0.08–0.54, FDR = 0.004), and multiple myeloma (OR = 0.18, 95%CI = 0.05–0.66, FDR = 0.018). Of note is that both the weighted median method and the maximum likelihood method generate consistent causal effect estimates compared with the IVW method (Supplementary Table S6). In addition, we create scatter plots for the SNP effect sizes of leukocyte telomere length and these four cancers (Figure 1); we find that no instruments may be potential outliers. The finding is also supported by MR-PRESSO, which displays the absence of instrument outliers at the significance level of 0.05.

FIGURE 1.

Relationship between the single nucleotide polymorphism (SNP) effect sizes of leukocyte telomere length (LTL) (x-axis) and the corresponding effect sizes of cancer (y-axis). (A) Malignant melanoma. (B) Hodgkin’s lymphoma/Hodgkin’s disease. (C) Chronic lymphocytic leukemia. (D) Multiple myeloma. In the plot, horizontal/vertical lines represent the 95% confidence interval.

To further examine whether a single instrumental variable may strongly influence the causal effects of leukocyte telomere length on these four cancers, we performed the LOO analysis. Again, the LOO analysis results demonstrate that none of the 17 instruments can substantially impact the estimated casual effect. Therefore, we can conclude that it is likely that a shorter leukocyte telomere length can decrease the risk of malignant melanoma, Hodgkin’s lymphoma/Hodgkin’s disease, chronic lymphocytic leukemia, and multiple myeloma. This finding here is also consistent with the results derived by the GRS regression above.

Power Calculation for the Association Between GRS and Cancers in the UK Biobank/TCGA Datasets

In terms of our simulations, we have sufficient power to detect the association in the UK Biobank as the total sample size is large, although only a few of the cancer cases are included. Specifically, we observe that the estimated power approaches 100% even when the number of cases is only 50 and the OR is only 1.05. In contrast, due to the relatively weak effect size and small sample size in the simulated TCGA cancer dataset, under our simulation settings, we have only low to moderate power to detect the association between GRS and the survival risk of cancer (Figure 2). For example, when the sample size is 300, the statistical power is only 3.0 or 10.7% when the HR was set to be 1.05 or 1.10. As can be expected, the power improves with the increase in the sample sizes and effect sizes.

FIGURE 2.

Estimated power in the simulation to evaluate the association between genetic risk score (GRS) and cancers in The Cancer Genome Atlas (TCGA). In the simulation, the effect sizes of GRS were set to 0.05, 0.10, and 0.20 and the sample sizes of cancer were set to 100, 300, and 500.

Discussion

Summary of the Results of the Present Study

The main objective of our study was to investigate whether there existed associations between genetically predicted leukocyte telomere length and various types of cancers. To achieve this, we first constructed the GRS of leukocyte telomere length based on associated SNPs from a large-scale GWAS and evaluated the effect of GRS on the risk and mortality of cancers. We found statistical evidence supporting the existence of associations between GRS and cancers in the UK Biobank and TCGA. Briefly, based on the GRS, a shorter leukocyte telomere length was identified to be associated with the decreased risk of some cancers (i.e., multiple myeloma, chronic lymphocytic leukemia, kidney/renal cell cancer, bladder cancer, malignant melanoma, basal cell carcinoma, prostate cancer, sarcoma/fibrosarcoma, and Hodgkin’s lymphoma/Hodgkin’s disease) as well as related to the decreased mortality of KIRP. In addition, inverse associations were observed for shorter leukocyte telomere length on the risk of leukemia as well as on the mortality of READ, SARC, and SKCM. The results of the MR analysis also supported the existence of an association between leukocyte telomere length and various cancers, including malignant melanoma, Hodgkin’s lymphoma/Hodgkin’s disease, chronic lymphocytic leukemia, and multiple myeloma. The diverse associations between leukocyte telomere length and cancers may in part reflect the different carcinogenic mechanisms acted by telomere in specific cancer types, further suggesting that telomere length is a valuable indicator of cancer risk and prognosis.

Discoveries Combined With the Previous Study

We found that the observed associations between leukocyte telomere length and cancers in the present study (i.e., multiple myeloma, chronic lymphocytic leukemia, kidney/renal cell cancer, bladder cancer, malignant melanoma, basal cell carcinoma, and prostate cancer) are greatly consistent with prior findings obtained in terms of MR (Supplementary Table S1; Zhang et al., 2015; Ojha et al., 2016; Haycock et al., 2017; Machiela et al., 2017; Li et al., 2020; Went et al., 2020). Particularly, several previous studies demonstrated that a shorter telomere length was associated with a decreased lung cancer risk or mortality and that the association was present in adenocarcinoma while absent in squamous cell carcinoma (Supplementary Table S1; Zhang et al., 2015; Haycock et al., 2017; Kachuri et al., 2018; Yuan et al., 2018), which may be attributed to the discrepancy in the biological characteristics of various subtypes of lung cancer. In the present study, inconsistent correlations were also identified within different subtypes of cancer. For example, we discovered that leukocyte telomere length had an opposite effect on the risk of leukemia and chronic lymphocytic leukemia. However, we observed that leukocyte telomere length displayed similar effects on the risk of malignant melanoma and basal cell carcinoma. These findings suggest that leukocyte telomere may influence the risk or mortality of cancer in a histologic way and also emphasize the unique roles of leukocyte telomere in the development of cancers.

Although the molecular mechanism remains unclear, some prior studies implied that both short and long telomere length played an important role in the etiology of cancers (Cui et al., 2012; Cheng et al., 2017; Nelson and Codd, 2020). Cells with longer telomere lengths have greater proliferative potential and more probability of accruing mutations (Hanahan and Weinberg, 2011); therefore, telomere shortening is generally considered to be a protective mechanism against tumorigenesis (Rode et al., 2016; Zhang et al., 2017; Kuo et al., 2019). However, it has been proposed that telomere shortening can generally give rise to end-to-end chromosome fusions and attenuates DNA damage response, thus increasing genomic instability and finally initiating carcinogenesis (Wu et al., 2003). These findings indicate that telomere plays a dual role in cancer development, and such role seems to depend on the types of cancers and the balance of the proliferation and senescence of cells in cancers.

Strengths and Limitations of Our Study

One advantage of our study is that more than 50 diverse types of cancers were investigated; it is thus feasible to undertake a systematic evaluation in the present analysis. In addition, methodologically, the GRS analysis can be viewed to be a two-stage regression model within the framework of instrumental variable-based causal inference (Baum et al., 2003; Hernán and Robins, 2006; Zeng et al., 2019a). Specifically, leukocyte telomere length is the exposure of interest and the associated SNPs are the carefully selected instrumental variables which are supposed to satisfy the necessary assumptions of instruments (Lawlor et al., 2008; Sheehan et al., 2008; Zeng et al., 2019a; Zeng and Zhou, 2019a, b). In the first stage, the effect size of each instrumental variable is estimated with an external large-scale GWAS dataset; in the second stage, the influence of leukocyte telomere length on various cancers is assessed based on the genetically determined leukocyte telomere length which is predicted with the chosen instrumental variables. Therefore, in terms of the principle of instrumental variable inference, the estimated effect of GRS can be interpreted as causal. In this sense, besides the MR method, we are actually investigating the causal association between leukocyte telomere length and cancers by constructing a GRS.

Finally, some shortcomings of this study should also be mentioned. Firstly, the majority of the individuals of the UK Biobank and TCGA were of European ancestry, so our results may not be applicable to other populations. Secondly, in our study, telomere length measured in blood leukocytes was employed and not in all cell types in vivo; however, leukocyte telomere length was demonstrated to be highly correlated with that in cells from other tissues (Friedrich et al., 2000; Wilson et al., 2008; Butt et al., 2010). Thirdly, as described before, the effect sizes of leukocyte telomere length on the mortality of TCGA cancers were only suggestive and the sample size of these cancers was not sufficiently large to maintain high power to detect weak associations. Therefore, further investigations with a larger sample size are required to validate our results.

Conclusion

Our study reveals that telomere played diverse roles in different types of cancers; however, further validations in large-scale prospective studies and deeper investigations of the biologic mechanisms are warranted.

Data Availability Statement

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found in the article/Supplementary Material.

Author Contributions

PZ conceived the idea for the study. PZ, YW, XZ, SH, and HZ obtained the data. PZ and YG cleared up the datasets, performed the data analyses, and drafted the manuscript. PZ, YG, and YW interpreted the results of the data analyses. All authors approved the manuscript and provided relevant suggestions.

Conflict of Interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Supplementary Material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fgene.2020.583106/full#supplementary-material