Predicted leukocyte telomere length and risk of myeloid neoplasms

Abstract

Maintenance of telomere length has long been established to play a role in the biology of cancer and several studies suggest that it may be especially important in myeloid malignancies. To overcome potential bias in confounding and reverse causation of observational studies, we use both a polygenic risk score (PRS) and inverse-variance weighted (IVW) Mendelian randomization (MR) analyses to estimate the relationship between genetically predicted leukocyte telomere length (LTL) and acute myeloid leukemia (AML) risk in 498 cases and 2099 controls and myelodysplastic syndrome (MDS) risk in 610 cases and 1759 controls. Genetic instruments derived from four recent studies explaining 1.23–4.57% of telomere variability were considered. We used multivariable logistic regression to estimate odds ratios (OR, 95% confidence intervals [CI]) as the measure of association between individual single-nucleotide polymorphisms and myeloid malignancies. We observed a significant association between a PRS of longer predicted LTL and AML using three genetic instruments (OR = 4.03 per ~1200 base pair [bp] increase in LTL, 95% CI: 1.65, 9.85 using Codd et al. [Codd, V., Nelson, C.P., Albrecht, E., Mangino, M., Deelen, J., Buxton, J.L., Hottenga, J.J., Fischer, K., Esko, T., Surakka, I. et al. (2013) Identification of seven loci affecting mean telomere length and their association with disease. Nat. Genet., 45, 422–427 427e421–422.], OR = 3.48 per one-standard deviation increase in LTL, 95% CI: 1.74, 6.97 using Li et al. [Li, C., Stoma, S., Lotta, L.A., Warner, S., Albrecht, E., Allione, A., Arp, P.P., Broer, L., Buxton, J.L., Alves, A.D.S.C. et al. (2020) Genome-wide association analysis in humans links nucleotide metabolism to leukocyte telomere length. Am. J. Hum. Genet., 106, 389–404.] and OR = 2.59 per 1000 bp increase in LTL, 95% CI: 1.03, 6.52 using Taub et al. [Taub, M.A., Conomos, M.P., Keener, R., Iyer, K.R., Weinstock, J.S., Yanek, L.R., Lane, J., Miller-Fleming, T.W., Brody, J.A., Raffield, L.M. et al. (2022) Genetic determinants of telomere length from 109,122 ancestrally diverse whole-genome sequences in TOPMed. Cell Genom., 2.] genetic instruments). MR analyses further indicated an association between LTL and AML risk (P_IVW ≤ 0.049) but not MDS (all P_IVW ≥ 0.076). Findings suggest variation in genes relevant to telomere function and maintenance may be important in the etiology of AML but not MDS.

Introduction

Telomeres are repetitive DNA sequences that make up the terminal ends of the chromosomes. In normal cells, telomeres shorten progressively with each cell division due to the so-called ‘end-replication problem’ and therefore, telomere length is reflective of cellular aging (1). Once telomeres reach a critically shortened length at which they are unable to maintain genomic stability, cellular senescence or programmed cell death occurs (2). Telomere length is regulated by the enzyme telomerase, where the protein portion is encoded by TERT and the mRNA guide is encoded by TERC, which can allow cells to maintain telomere length and overcome cellular senescence when upregulated (3). Telomere length in somatic cells is heterogeneous but typically declines with age (4). Variation in telomere length also occurs between individuals and is influenced by age, disease status, genetics and lifestyle factors such as tobacco smoking (5–7).

Maintenance of telomere length has long been established to play a role in the biology of cancer and has been recognized as the mechanism facilitating the limitless replicative potential that serves as one of the original ‘Hallmarks of Cancer’ described by Hanahan and Weinberg (8,9). Cancer cells are able to maintain telomere lengths over successive generations, typically through the upregulation of telomerase (10). The role of telomere length has been evaluated in observational studies of cancer with mixed results (11–15); however, these studies are likely to be confounded by demographic and lifestyle factors that are known to influence telomere length (16). Retrospective studies are further limited by biases due to the measurement of telomere length after diagnosis and/or treatment (17). Notably, meta-analyses of published studies have reported significant associations between telomere length and cancer in retrospective studies while associations were weaker and non-significant in prospective studies (12,15).

Several lines of evidence suggest that telomere maintenance may be especially important in myeloid malignancies. First, families with telomere biology disorders are at increased risk of myelodysplastic syndromes (MDS) and acute myeloid leukemia (AML) (18–24). Further, telomere erosion has been documented in sporadic MDS and AML and appears to be associated with increasing genomic complexity and risk of progression (25–27). Genetic variation in telomere biology genes has not been evaluated extensively in sporadic MDS/AML cases; however, several previous studies have reported on associations between TERT variants and MDS and AML (19,28,29). Direct analysis of leukocyte telomere length (LTL) in cases is problematic because it is likely to be influenced by the disease process.

Telomere length is a trait with high heritability (30,31), and recent genome-wide association studies (GWASs) of more than 37 000 individuals have led to the development of robust genetic predictors of LTL that explain 1.23–4.57% of the variation in telomere length (32–35). These genetic variants have been used as a proxy for telomere length in a number of studies of cancer risk, using either a genetic risk score or Mendelian randomization (MR) approach (36–46). In these studies, telomere length genetic risk scores have been established as risk factors for multiple types of cancer (37,38,44). To date, few studies have evaluated the association between genetically predicted telomere length and myeloid malignancy (42,43).

In this analysis, we use both a polygenic risk score (PRS) of telomere length–associated variants and MR methods to evaluate the association between genetically predicted LTL using four validated genetic instruments of LTL (32–35) and the risk of myeloid malignancy in case–control studies of AML and MDS.

Results

AML results

In this genetically inferred European ancestry-only analysis, we included 498 AML cases (238 Discovery and 260 Replication) and 2099 controls (1059 Discovery and 1040 Replication). Approximately 60% of the Discovery AML sample population were males, while the Replication sample was more evenly distributed by sex (Table 1). The majority of AML cases were older than 60 years of age at diagnosis.

Table 1

Selected characteristics of the study population.

Characteristic	Discovery			Replication
	Controls	AML	MDS	AML controls	AML cases	MDS controls	MDS cases
N	1059	238	435	1040	260	700	175
Sex
Male	606 (57)	151 (63)	287 (66)	548 (53)	137 (53)	444 (63)	111 (63)
Female	452 (43)	87 (37)	148 (34)	492 (47)	123 (47)	256 (37)	64 (37)
Age (years)
<50	151 (14)	72 (30)	12 (3)	17 (2)	4 (2)	7 (1)	1 (1)
50–59	194 (18)	46 (19)	31 (7)	188 (18)	50 (19)	41 (6)	9 (5)
60–69	309 (29)	75 (32)	119 (27)	381 (37)	86 (33)	266 (38)	60 (34)
70–79	332 (31)	45 (19)	174 (40)	454 (44)	120 (46)	386 (55)	105 (60)
≥ 80	72 (7)	0 (0)	99 (23)	-	-	-	-

Abbreviations: AML: acute myeloid leukemia, MDS: myelodysplastic syndrome

In the single-nucleotide polymorphism (SNP) meta-analysis of each of the LTL-associated SNPs, we observed a statistically significant association between AML and SMC4/rs201009932 (P = 0.0001) from the Codd et al. (35) genetic instrument after Bonferroni correction (Supplementary Material, Table S1). No other SNPs reached significance after Bonferroni correction. A nominal association (P ≤ 0.05) was observed for 19 SNPs including three (out of four) TERC and seven (out of 20) TERT SNPs that were included in the analysis.

Compared to controls, AML cases had a mean difference in genetically predicted LTL of +0.02 per ~1200 base pairs (bp) using Codd et al. (32), +0.02 per one-standard deviation (SD) using Li et al. (33), +0.07 per one-SD using Codd et al. (35) and +0.02 per 1000 bp using Taub et al. (34) genetic instruments.

In the AML PRS analysis, we observed a positive association between longer genetically predicted LTL and risk of AML for three of the genetic instruments [OR = 4.03 per ~1200 bp increase in LTL, 95% CI: 1.65, 9.85 using Codd et al. (35), OR = 3.48 per one-SD increase in LTL, 95% CI: 1.74, 6.97 using Li et al. (33) and OR = 2.59 per 1000 bp increase in LTL, 95% CI: 1.03, 6.52 using Taub et al. (34) genetic instruments; Table 2]. We did not observe a statistically significant association using the Codd et al. (35) genetic instrument, although the effect estimate was consistent in direction. The effect estimates were similar in magnitude in both the Discovery and Replication datasets. To better characterize the association of telomere length with AML risk, we investigated the magnitude of the association of the telomere length–associated PRS with AML risk by sex. For all four genetic instruments, females had a strong positive association between longer genetically predicted LTL and increase risk of AML [OR = 8.97 per ~1200 bp increase in LTL, 95% CI: 2.35, 34.29 using Codd et al. (32), OR = 6.51 per one-SD increase in LTL, 95% CI: 2.26, 18.76 using Li et al. (33), OR = 2.38 per 1 SD increase in LTL, 95% CI: 1.24, 4.56 using Codd et al. (35) and OR = 7.22 per 1000 bp increase in LTL, 95% CI: 1.66, 31.30 using Taub et al. (34) genetic instruments; Supplementary Material, Table S2].

Table 2

Association between myeloid malignancies (acute myeloid leukemia [AML] and myelodysplastic syndrome [MDS]) and predicted leukocyte telomere length (LTL).

	Cases	Controls	N SNPs (% of signal captured)a	Polygenic risk score (PRS)
				Odds ratiob (95% CI)	P-value
Codd et al. (32)—7 sentinel SNP instrument
AML
Meta-analysis	498	2099	7 (100)	4.03 (1.65, 9.85)	0.002
Discovery	238	1059	7 (100)	4.52 (1.24, 16.48)	0.022
Replication	260	1040	7 (100)	3.64 (1.06, 12.48)	0.040
MDS
Meta-analysis	610	1759	7 (100)	1.33 (0.55, 3.20)	0.524
Discovery	435	1059	7 (100)	0.94 (0.32, 2.76)	0.908
Replication	175	700	7 (100)	2.61 (0.58, 11.71)	0.211
Li et al. (33)—20 sentinel SNP instrument
AML
Meta-analysis	498	2099	20 (100)	3.48 (1.74, 6.97)	0.0004
Discovery	238	1059	20 (100)	3.65 (1.33, 10.00)	0.012
Replication	260	1040	20 (100)	3.33 (1.27, 8.70)	0.014
MDS
Meta-analysis	610	1759	20 (100)	1.19 (0.60, 2.35)	0.619
Discovery	435	1059	20 (100)	0.66 (0.29, 1.48)	0.309
Replication	175	700	20 (100)	4.82 (1.38, 16.77)	0.013
Codd et al. (35)—197 sentinel SNP instrument
AML
Meta-analysis	498	2099	197 (100)	1.49 (0.98, 2.27)	0.065
Discovery	238	1059	162 (87.1)	1.34 (0.71, 2.51)	0.368
Replication	260	1040	197 (100)	1.62 (0.92, 2.86)	0.094
MDS
Meta-analysis	610	1759	197 (100)	0.67 (0.44, 1.01)	0.057
Discovery	435	1059	162 (87.1)	0.50 (0.30, 0.84)	0.009
Replication	175	700	197 (100)	1.13 (0.56, 2.28)	0.730
Taub et al. (34)—59 sentinel SNP instrument
AML
Meta-analysis	498	2099	56 (99.8)	2.59 (1.03, 6.52)	0.043
Discovery	238	1059	55 (98.3)	2.42 (0.66, 8.97)	0.185
Replication	260	1040	53 (93.3)	2.77 (0.75, 10.16)	0.126
MDS
Meta-analysis	610	1759	57 (99.9)	0.84 (0.34, 2.03)	0.694
Discovery	435	1059	55 (98.3)	0.48 (0.17, 1.37)	0.168
Replication	175	700	52 (93.0)	3.32 (0.64, 17.28)	0.153

Abbreviations: AML: acute myeloid leukemia, MDS: myelodysplastic syndrome, SNP: single-nucleotide polymorphism, CI: confidence intervals.

The Discovery sample was adjusted for sex, age (only MDS) and the first two ancestry-specific principal components.

Bolded values signify statistical significance at P ≤ 0.05.

^aSome SNPs were not present on the imputation reference panel or were not well imputed in the Discovery and Replication cohorts and therefore could not be analyzed. The SNPs used in each PRS are reported in Supplementary Material, Table S1 and the proportion of the signal captured (summed proportion of variance explained [PVE] of the variants used divided by the summed PVE of all variants in the original score) was computed.

^bOdds ratio (OR) and 95% confidence interval (CI) represents the increase in risk per extra ~1200 base pair of telomere length for Codd et al. (32), per one-standard deviation for Li et al. (33) and Codd et al. (35), and per 1000 base pair of telomere length for Taub et al. (34) genetic instruments.

The MR analyses produced similar effect estimates as the PRS analysis and further indicated an association between longer genetically predicted LTL and AML using two genetic instruments [OR = 4.15 per ~1200 bp increase in LTL, 95% CI: 1.68, 10.25 using Codd et al. (32) and OR = 3.50 per one-SD increase in LTL, 95% CI: 1.59, 7.70 using Li et al. (33) genetic instruments; Supplementary Material, Table S3]. Sensitivity and pleiotropy analyses demonstrated these results were relatively stable and that there was no significant heterogeneity or pleiotropy across the SNPs in either instrument. The results from the MR analysis using Codd et al. (35) and Taub et al. (34) genetic instruments showed a consistent but non-significant trend in the IVW analysis. For the Codd et al. (35) genetic instrument, the MR-Egger and weighted median effect estimates were significant and sensitivity analyses did not indicate heterogeneity or pleiotropy within the genetic instrument. There was, however, evidence of heterogeneity (Cochran’s Q P-value = 0.025) and notable horizontal pleiotropy (MR-PRESSO global P-value = 0.023) across SNPs using the Taub et al. (34) genetic instrument. No significant outliers were detected using MR-PRESSO. The effect estimates were similar in magnitude in both the Discovery and Replication datasets (Supplementary Material, Table S3; Figs S3–S6).

MDS results

In the analysis of MDS, we included 610 MDS cases (435 Discovery and 175 Replication) and 1759 controls (1059 Discovery and 700 Replication). Males comprised more than half of the MDS study population in all groups (Table 1). The majority of MDS cases were older than 60 years of age at diagnosis.

Similar to the findings from the AML study, in the SNP meta-analysis of each of the LTL-associated SNPs, we observed a statistically significant association between MDS and SMC4/rs201009932 (P = 0.0003) from the Codd et al. (35) genetic instrument after Bonferroni correction (Supplementary Material, Table S1). Additionally, we observed a statistically significant association between MDS and ACYP2/rs11125529 (P = 0.006) from the Codd et al. (32) genetic instrument. No other SNPs reached significance after Bonferroni correction. A nominal association (P ≤ 0.05) was observed for 15 SNPs including three (out of the four) TERC SNPs that were included in the analysis.

MDS cases as compared to controls had a mean difference between genetically predicted LTL of +0.002 per ~1200 bp using Codd et al. (32), +0.001 per one-SD using Li et al. (33), −0.22 per one-SD using Codd et al. (35) and −0.05 per 1000 bp using the Taub et al. (34).

In the MDS PRS analysis, we did not observe an association between longer genetically predicted LTL and MDS for any of the four genetic instruments (Table 2). The effect estimates were not similar in direction across studies, with the Discovery sample demonstrating a non-significant decreased risk between MDS and longer LTL and the Replication sample demonstrating a non-significant increased risk with imprecise confidence intervals. To better characterize the association of telomere length with MDS risk, we investigated the magnitude of the association of the telomere length–associated PRS with risk by sex. As with the overall PRS, we did not observe an association between longer genetically predicted LTL and increased risk for MDS by sex (Supplementary Material, Table S2).

The MR analyses produced similar effect estimates as the PRS analysis and further indicated no association between longer genetically predicted LTL and MDS for any of the four genetic instruments genetic instruments (Supplementary Material, Table S3). Sensitivity and pleiotropy analyses demonstrated heterogeneity within the Codd et al. (32) genetic instrument for both the IVW (Cochran’s Q P-value = 0.026) and MR Egger (Cochran’s Q P-value = 0.048) methods. However, no significant outliers were detected using MR-PRESSO (global P-value = 0.179 and 0.103, respectively). Similar to the PRS analysis, the effect estimates were not similar in direction across studies with the Discovery sample demonstrating a non-significant decreased risk between MDS and longer LTL and the Replication sample demonstrating a non-significant increase in MDS risk with imprecise confidence intervals (Supplementary Material, Table S3; Figs S7–S10).

Discussion

In this analysis of genetically predicted telomere length and myeloid malignancy, we identified a statistically significant association between longer predicted leukocyte telomere length and AML with three genetic instruments of LTL. Subset analyses by sex indicated that telomere length may play a stronger role in risk of AML for females, although the interaction term for sex was not significant and the confidence limits in females were wide due to the small sample size. In contrast, no significant association was observed for MDS. The use of genetic predictors for telomere length reduces the possibility that the association is explained by reverse causation and/or confounding, both of which are likely relevant in case–control studies in which telomere length is measured directly. This is because genotypes in the general population are expected to be randomly distributed. Additionally, genetically predicted telomere length has the added advantage of representing the effects of telomere length over the lifetime of the individual, while measuring LTL directly represents only one instance in an individual’s life (39).

Studies that have measured telomere length directly to determine association with cancer risk have provided mixed results. Shorter telomere length has been reported to be associated with multiple cancer types; however, recent meta-analyses have noted that these associations are typically observed only in retrospective studies (12,14,15). There is a considerable body of literature reporting on measured telomere length and myeloid malignancy. Several studies have reported telomere erosion in AML (47,48) and MDS (49). Further, shortened telomere length at diagnosis has been correlated with increased genomic complexity and chromosomal instability (50–53), risk of disease progression (26,27,50,52) and death (51,53,54). Among AML and MDS patients, telomere length at diagnosis has been found to be longer in older patients (>60 years old) as compared to younger patients (53,55,56) and is similar at diagnosis and relapse but increases after chemotherapy-induced remission (47,55,57). While these studies may provide helpful information on the role of telomere length in disease progression and outcomes, the lack of available pre-diagnostic measurements limits our ability to draw conclusions about a potential role in disease risk.

In contrast to studies that have reported measured telomere length, studies evaluating relationships between SNPs that predict telomere length and cancer have produced consistent results. To date, longer predicted LTL has been reported to increase the risk of multiple types of cancer, including B-cell lymphoma, leukemias, lung cancer, melanoma, glioma, neuroblastoma, bladder cancer, testicular cancer, kidney cancer and endometrial cancer (36–46,58,59). In this study, we observed a consistent and statistically significant association between longer predicted LTL and higher AML risk. Combined with the knowledge that individuals with short telomere length syndromes are at increased risk of MDS and AML (23), this finding suggests that the association between telomere length and AML may be bidirectional with both long and short telomeres leading to increased risk.

Given the overlapping etiology of MDS and AML, it was somewhat surprising that we saw an association between genetically predicted LTL and risk of AML but did not see a similar association with MDS. One possible explanation for this difference is the heterogeneity of MDS, where we see differences in cytogenic abnormalities (60,61); recurrent somatic mutations (62); and variability in disease duration, progression and death (63) by disease subtype. The percentage of MDS cases that progress to AML varies widely by subtype (64), and the majority of AML cases are de novo rather than arising secondary to MDS (62). Studies with sufficient sample size to evaluate associations by disease subgroup would be especially valuable to determine whether subtype specific differences could explain the difference in our findings for MDS and AML. We also note that while our findings for MDS were null overall, the wide confidence intervals highlight the imprecision of the estimates and suggest that further studies are warranted.

The biological mechanism for a relationship between longer telomere length and risk of cancer is not completely understood at the current time; however, one hypothesis is that longer telomere length may increase the risk of cancer due to the increased potential for replication (65,66). Recently, a large study using UK Biobank data reported an association between longer genetically predicted telomere length and an increased number of clonal somatic copy number alterations in peripheral blood leukocytes (67). Clonal somatic copy number alterations have also been associated with hematologic cancers in previous studies (68–70). Further, the genomic rearrangements that are common in hematological malignancies share similarities with the error-prone non-homologous end joining observed following end-to-end fusion of chromosomes with dysfunctional telomeres (25). While future studies will be required to confirm potential mechanistic links, these data provide biologic plausibility for the role of longer telomere length in the etiology of myeloid malignancy.

Studies of high-risk families with multiple affected members with AML and MDS have identified germline mutations in TERC and TERT (71). One previous study evaluated telomerase mutations in sporadic AML and suggested that inherited genetic variation in TERT may be responsible for up to 8% of sporadic cases (19). Genetic variants in TERT have also been associated with multiple cancer types (72). In this analysis, we saw nominal associations between common variants in TERC and TERT and the risk of both AML and MDS. We also observed a statistically significant association between a variant in SMC4 and both AML and MDS and an association between a variant in ACYP2 and MDS. However, evidence that germline genetic variation in these genes plays a role in the etiology of myeloid leukemias is limited. A more comprehensive analysis of common genetic variants in telomere biology genes using available GWAS data may be warranted.

This study has a number of strengths, including the population-based design and the rigorous pathology, cytogenetics and clinical review. Moreover, the use of an MR approach and PRSs is also a strength as it reduces the biases associated with measuring LTL directly in affected individuals. A number of limitations must also be considered. Given the poor outcomes associated with AML and MDS (73), there is a potential for survival bias. While rapid case ascertainment was used to reduce this possibility, there is a possibility of bias if genetic variants predicting telomere length are associated with prognosis. Selection bias is also possible given the response rates, although we observed no difference between participating and non-participating cases or controls with respect to education or income in either study. Given that AML and MDS are rare diseases, our power was limited, especially for the subgroup analyses stratified by sex and we could not look by clinical stage or subtype of disease. Finally, MR requires certain core assumptions to be met in order for the genotype to serve as an instrumental variable (74). These assumptions may be violated in situations of pleiotropy, population stratification or genetic heterogeneity (74). We have guarded against population stratification by limiting our analyses to individuals genetically inferred as European ancestry and incorporating principal components that capture ancestry into our analyses. We have also explored the possibility of pleiotropy using multiple sensitivity analyses.

In conclusion, our results suggest an association between longer predicted LTL and the risk of AML but not MDS. The finding for AML is consistent with prior studies demonstrating an association with risk of multiple cancer types and longer predicted telomere length. Further analysis of genetic variation in genes relevant for telomere function and maintenance may be warranted given the known increase in the risk of myeloid malignancy in individuals with telomere biology disorders.

Materials and Methods

Study participants

Data for the current study came from two sources. The Discovery sample consisted of case–control samples from three independent studies: the Adults in Minnesota with Myelodysplastic Syndromes (AIMMS) Study, the Predictors of Adult Leukemia in Minnesota (PALM) Study and case-only samples from the Moffitt Cancer Center. The Replication sample included MDS and AML cases and unaffected controls from the UK Biobank. To improve our power to detect associations, we meta-analyzed the results of the Discovery and Replication samples together.

This study was approved by the Institutional Review Boards (IRBs) of the University of Minnesota, the Mayo Clinic, the Minnesota Department of Health and participating area hospitals.

Discovery sample data collection and genotyping

Discovery case ascertainment

Cases for the Myeloid Leukemia (PALM) and MDS studies (AIMMS) were identified by the Minnesota Cancer Reporting System (MCRS), which is a population-based registry that uses a rapid case ascertainment system to collect information on all cancers diagnosed in Minnesota. Detailed information regarding the case and control recruitment and response rates has been described (75,76). Cases were eligible for the study if they were a Minnesota resident, were diagnosed between the ages of 20 and 79 years (up to 85 years for MDS) and could understand English or Spanish. Proxy interviews were not conducted. We included eligible cases who were diagnosed with acute myeloid leukemia (AML; ICD-O-3 codes: 9840, 9861, 9866–9867, 9871–9874, 9891–9897, 9910, 9920) between June 1, 2005 and November 30, 2009. MDS cases were eligible if they had a diagnosis of MDS (ICD-O-3 codes: 9980, 9982–9987, 9989) between April 1, 2010 and October 31, 2014. Centralized pathology and cytogenetics review were conducted to confirm the classification as AML or MDS.

Control recruitment

Controls were identified through the Minnesota State driver’s license/identification card list and were eligible if they were alive at the time of contact; resided in Minnesota; and were between the ages of 20 and 80 years (up to 85 years for MDS controls), could understand English or Spanish and had no prior diagnosis of myeloid leukemia or MDS. Controls were frequency matched to cases based on decile of age.

Discovery sample collection and DNA extraction

Genomic DNA was collected from cases and controls using Oragene DNA collection kits (DNA Genotek, Ontario, Canada; MDS cases and controls) or mouthwash collection kits (leukemia cases and controls). DNA was extracted via Autopure LS Instrument according to the manufacturer’s instructions (Qiagen, Inc., Valencia, CA). DNA yield was quantified using a 1:10 dilution tested in triplicate by quantitative real-time PCR on an ABI Prism 7900HT Sequence Detection System using Sequence Detection Software v2.1 (Life Technologies, Grand Island, NY). Extracted DNA was stored at −20°C until further analysis.

Moffitt study

Case-only MDS samples were collected from patients seen at Moffitt Cancer Center (Tampa, Florida) between 2005 and 2016 who provided written consent on IRB-approved protocols. Samples from cases diagnosed between the ages of 41 and 94 years old and diagnosed with MDS by the 2008 World Health Organization (WHO) classification. Centralized pathology and cytogenetics review were conducted to confirm the classification as AML or MDS. DNA was isolated from unfractionated bone marrow mononuclear cells using QIAamp DNA Mini Kit (Qiagen, Germantown, MD) following the manufacturer’s protocol.

Genotyping and imputation

The University of Minnesota Genomics Center (UMGC) performed genotyping for the AIMMS, PALM and Moffitt samples using the Illumina Infinium Global Diversity Array v1.0 (Illumina, San Diego) according to the manufacturer’s specified protocol. Allele cluster definitions for each variant were determined via Illumina’s GenomeStudio Genotyping Module in two stages: (1) markers were initially genotyped using all samples in order to compute initial call rate and log R ratio standard deviation estimates, and (2) final genotyping was performed by recomputing allele clusters for each variant using only the intensity data from those study samples with high-quality data (call rate > 0.98; log R ratio standard deviation <0.50). HapMap samples and blind duplicate samples were distributed among the plates to assess genotyping concordance and detect plate effects. There was no evidence of plate-specific genotype effects. Genvisis (http://www.genvisis.org) was used to identify low-quality/contaminated samples (n = 76), samples with sex aneuploidy (n = 8) and trisomy 21 (n = 1). These samples were excluded from all analyses.

Ancestry was genetically inferred by performing a principal components analysis (PCA) as implemented in EIGENSOFT (77,78) including HapMap samples as anchors. The first two principal components from the analysis were plotted and samples were assigned an ancestry of European, African American, Asian or Hispanic based on where they clustered relative to the three HapMap anchor populations. A second PCA was run within each of these ancestry designations to obtain principal components to use as covariates. Supplementary Material, Figure S1 shows principal component scatter plots (PC1 versus PC2) by case control status for the Discovery sample.

To extend our genotype analysis, imputation for the AML and MDS cases and controls was performed using 194 512 haplotypes from the Trans-Omics for Precision Medicine (TOPMed) Imputation Reference Panel (version TOPMed-r2; Eagle v2.49 and minimac3.7) accessed through the TOPMed Imputation Server (79–81). Prior to imputation, variants were removed if (1) the call rate was <98%; (2) missingness differed between cases and controls (P < 1E−7); (3) allele frequencies or missingness differed between males and females (P < 1E−7); (4) significant deviation from Hardy–Weinberg equilibrium was observed in the European founders (P < 1E−7); or (5) significant deviation from expected when imputing the variant from nearby markers using PLINK’s—mishap test (P < 1E−7). Additionally, monomorphic variants and non-SNPs were dropped, leaving 1 642 187 variants for imputation.

Replication sample and genotyping

UK biobank cases and controls

The UK Biobank (https://www.ukbiobank.ac.uk/) is a population-based prospective cohort study that recruited more than 500 000 individuals aged 40–69 years from across the UK as detailed previously (82,83). We identified unrelated males and females who were diagnosed with AML (ICD-10 code C92.0) or MDS (ICD-10 codes D46.0–D46.2, D46.4, D46.7 and D46.9) and had existing genetic data available by the end of 2016. Detailed quality control information pertaining to the genetic data was described previously by Bycroft et al. (83).

Matching algorithm

For the UK Biobank data, genetically inferred European ancestry cases were matched to four controls each based on age (birth year and month), sex (male or female) and 10 genetic principal components (PCs; data field 22 009, https://biobank.ctsu.ox.ac.uk/crystal/field.cgi?id=22009), with the goal of finding a set of controls with the most similar genetic background that was the same sex and had similar ages to that of the cases (Supplementary Material, Fig. S2). This matching was performed using the software MatchSamples (https://github.com/PankratzLab/MatchSamples), which is described in more detail in Cigan et al. (46).

To identify our UK Biobank samples for matching, we used participants who self-identified and were genetically verified as ‘White British’ (data field 22 006; https://biobank.ctsu.ox.ac.uk/crystal/field.cgi?id=22006) and selected additional participants with similar ancestry using Mahalanobis distance methods (84). The means and covariance of the top two ancestry PCs were computed within 409 616 ‘White British’ participants. We then computed the Mahalanobis distance from each participant’s top two PC values to the center of this two-dimensional ‘White British’ distribution. We used the maximum Mahalanobis distance of any of the self-identified ‘White British’ samples as a threshold to select individuals with similar ancestry (i.e. all participants with a value ≤ 15.548). An additional 29 208 participants with a Mahalanobis distance less than this maximum value were selected as individuals with similar ancestry to the ‘White British’ subset. Of these, an additional 23 cases and 9982 controls were identified for inclusion. All calculations were performed in R version 4.1.2. (85).

Selection of LTL-associated SNPs as genetic instrument variants

We used four different sets of SNPs that were previously reported to be validated genetic predictors of leukocyte telomere length from large-scale GWAS (32–35). The seven SNP LTL predictor reported by Codd et al. (32) was used to facilitate comparisons with previously published literature. The additional genetic instruments were included because they each explain a larger percentage of variation in predicted LTL (33–35). For the 197 SNP genetic instrument reported by Codd et al. (35), European-specific effect sizes were used for the PRS and MR analysis. For the 59 SNP genetic instrument reported by Taub et al. (34), European-specific effect sizes from the joint model were used for the PRS and MR analysis. Genotyping, imputation information, variance explained and all the relevant information for each variant used in our analyses is provided in Supplementary Material, Table S1.

Statistical analysis

All statistical analyses were conducted using Gen Score Pipeline (https://github.com/PankratzLab/GenScorePipeline). We estimated associations between genetically predicted telomere length and myeloid malignancy risk separately for the Discovery sample and the Replication sample using logistic regression. For the Discovery sample, we adjusted the models for sex, age (only MDS) and the first two ancestry-specific principal components. For AML, since this disease is age-related and the control samples are, on average, 8 years older than the case’s onset age, we did not include age in the model (86). For the Replication sample, cases and controls were matched as described above and we did not include other variables for adjustment in the model. All individuals were genetically inferred as European ancestry.

Polygenic risk score

Leukocyte telomere length was genetically inferred for participants with available genotype data using a PRS containing variants previously associated with LTL (32–35). For each subject, the allele count of the effect allele associated with longer telomere length was multiplied by the beta reported in the appropriate GWAS meta-analysis of telomere length and summed all together (32–35). For both the Codd et al. (35) and Taub et al. (34) LTL genetic predictors, conditional betas reported for the European ancestry population were used. Individual SNPs were evaluated for significance using a Bonferroni correction to account for multiple comparisons (0.05/number of SNPs in each genetic predictor). Association tests were conducted separately for the Discovery and Replication samples using logistic regression and PRS estimates are reported as odds ratios (OR) and 95% confidence intervals (CI). The PRS was calculated using Gen Score Pipeline.

To improve our effect estimates, we meta-analyzed the individual SNP effect estimates from the Discovery sample and the Replication sample using an IVW method in METAL (87). Individual SNP effects were evaluated for heterogeneity across studies using Cochran’s Q-statistic and no SNP reached statistical significance following Bonferroni correction for multiple comparisons (P ≤ 0.007 for Codd et al. (32), or P ≤ 0.003 for Li et al. (33), P ≤ 0.0003 for Codd et al. (35) and P ≤ 0.001 for Taub et al. (34); Supplementary Material, Table S1).

Genetic instruments for MR

For each of the LTL genetic instruments, we removed correlated variants from loci with more than one conditionally independent variant with an r² > 0.1 based on PLINK linkage disequilibrium (LD) calculations in the Discovery sample and variants that had an effect allele count (EAC) or minor allele count (MAC) < 10 within each sample by disease (Supplementary Material, Table S1).

MR

MR is a method that utilizes genetic variants known to be associated with a trait of interest as instrumental variables to proxy an exposure and measures the causal association between a trait (e.g. telomere length) and an outcome (e.g. myeloid malignancy) (88,89). Taub et al. (34) performed a multivariable joint association test that included all SNPs in the regression model at once, which should have the biggest effect on variants that are in linkage disequilibrium with each other. Since we used the European-specific joint SNP-outcome effect estimates to perform our PRS and MR analyses, we also used effect estimates from our joint model that were derived by including all variants in the genetic model at once. All statistical analyses were conducted using the TwoSampleMR package in R following code provided at https://github.com/MRCIEU/TwoSampleMR. We used this package to implement the random-effects IVW, MR-Egger and the weighted median statistical methods but consider IVW our primary model as it is the standard approach. The IVW approach is described in Burgess et al. (90) and assumes variants are valid instruments. The MR-Egger method described in Bowden et al. (91) is designed to correct for bias due to pleiotropy. The weighted median is described in Bowden et al. (92) and uses a weighted empirical distribution function and assumes 50% of the weight comes from valid instruments. To test for potential violations of the MR assumption, we used the MR-Egger regression intercept (91) to detect overall directional pleiotropy, Cochran’s Q test to detect heterogeneity across SNP-specific MR estimates (with respect to IVW and MR-Egger methods) and MR-PRESSO to detect pleiotropy and potential outliers using the global test (93). MR estimates are reported as OR and 95% CI. We generated forest plots and scatter plots with trend lines associated with the different MR methods, where the slopes and directions of the trend lines represent the magnitudes and directions of the causal estimates, respectively (Supplementary Material, Figs S3–S9).

Conflict of Interest statement. The authors declare no conflict of interest.

Funding

National Institutes of Health (R01 CA142714 to J.N.P.)

Data Availability

The Discovery data used in this study are publicly available through dbGaP (Accession phs002962.v1.p1; https://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs002962.v1.p1). The Replication data used in this study are publicly available through the UK Biobank website (https://www.ukbiobank.ac.uk/).