Association between telomere length and Plasmodium falciparum malaria endemicity in sub-Saharan Africans

Summary

Leukocyte telomere length (LTL) varies significantly across human populations, with individuals of African ancestry having longer LTL than non-Africans. However, the genetic and environmental drivers of LTL variation in Africans remain largely unknown. We report here on the relationship between LTL, genetics, and a variety of environmental and climatic factors in ethnically diverse African adults (n = 1,818) originating from Botswana, Tanzania, Ethiopia, and Cameroon. We observe significant variation in LTL among populations, finding that the San hunter-gatherers from Botswana have the longest leukocyte telomeres and that the Fulani pastoralists from Cameroon have the shortest telomeres. Genetic factors explain ∼50% of LTL variation among individuals. Moreover, we observe a significant negative association between Plasmodium falciparum malaria endemicity and LTL while adjusting for age, sex, and genetics. Within Africa, adults from populations indigenous to areas with high malaria exposure have shorter LTL than those in populations indigenous to areas with low malaria exposure. Finally, we explore to what degree the genetic architecture underlying LTL in Africa covaries with malaria exposure.

We report on the relationship between leukocyte telomere length (LTL), genetics, and environmental factors in ethnically diverse Africans. LTL is longer in African ancestry individuals than in non-Africans and we observe a significant negative relationship between malaria endemicity and LTL, while adjusting for age, sex, and genetic ancestry.

Introduction

Leukocyte telomere length (LTL), which reflects telomere length in hematopoietic cells and other tissues, shows vast person-to-person variation (SD 700 base pairs [bps]) from birth onwards.¹ It shortens with age^2,3 and is a predictor of a number of aging-related diseases and mortality.^4,5 LTL is a highly heritable human trait^6,7 that is nominally affected by environmental factors in middle-upper-income populations.^8,9,10 LTL variation at birth largely determines LTL variation throughout the life course. However, the majority of large-scale studies examining LTL variation among humans have focused primarily on populations of European ancestry.⁵ This under-representation of diverse populations hampers our ability to understand the genetic and environmental drivers of LTL variation and their effects on telomere-related disease risk.

Little is known about the genetic, environmental, and evolutionary forces that have shaped the vast LTL variation across sub-Saharan African populations. LTL is longer by about 200 bps in African Americans (AfAms) than in Europeans,^11,12 and it is about 500 bps longer in sub-Saharan Africans (SSAfrs) than in AfAms.^13,14 While the underlying cause of longer LTL in individuals of African ancestry is unknown, we previously showed that LTL varies significantly across diverse ethnic groups in sub-Saharan Africa.¹³ This variation in LTL is largely explained by genetic factors, but environmental factors could also play a role. Exposure to Plasmodium falciparum malaria is one environmental factor of particular interest in impacting LTL, due to recent studies demonstrating a link between malaria infection and LTL. For example, European individuals who were infected by P. falciparum displayed shortened LTL upon returning from their travels.¹⁵ Volunteers experimentally infected with P. falciparum also showed a temporary shortening in LTL.¹⁶ In addition, birds experimentally injected with avian malaria (Plasmodium ashfordi) showed shorter telomeres in different tissues compared to non-infected controls.¹⁷

While these studies suggest a link between malaria infection and telomere shortening, they rely on single, acute infection events where participants received rapid medical treatment. These previous studies also measured LTL using a quantitative PCR (q-PCR) methodology, which is associated with higher error rates compared to the gold-standard measurements performed by Southern blotting that we used in this research.¹² It remains unknown, however, whether repeated malaria exposures throughout life in populations living in endemic regions has a lasting effect on LTL. It is also unclear whether having longer leukocyte telomeres at birth in malaria endemic regions or regions with a high pathogen burden could be selectively advantageous.

In this paper, we examine LTL from diverse environmental contexts across Africa, including those where malaria is highly endemic. We further analyze the relationship between LTL and malaria endemicity, as well as with a variety of genetic, environmental, and demographic factors. Finally, we examine the degree to which the genetic architecture underling LTL in Africa covaries with malaria endemicity.

Material and methods

Sample acquisition, genotyping, and LTL measurements

We obtained Institutional Review Board approval from the University of Pennsylvania and written informed consent from all study participants. Ethics approval was granted from the following African institutions: the National Institute of Medical Research in Dar es Salaam, Tanzania; COSTECH (the Tanzania Commission for Science and Technology); the Ministry of Public Health and National Ethics Committee for Research on Human Health, Cameroon; the University of Addis Ababa, Ethiopia; the Federal Democratic Republic of Ethiopia Ministry of Science and Technology National Health Research Ethics Review Committee, Ethiopia; and the University of Botswana and the Ministry of Health in Gaborone, Botswana. Samples were collected across the years 2010 (n = 464), 2011 (n = 296), 2012 (n = 226), 2013 (n = 373), and 2015 (n = 459).

We extracted DNA from whole blood and genotyped individuals on two different genotype arrays. Samples from Tanzania, Botswana, and Ethiopia were genotyped on the Illumina Omni5 array^13,18 (∼4.2 million SNPs), and samples from Cameroon were genotyped on the Illumina H3 Africa Consortium array¹⁹ (∼2 million SNPs). We merged both datasets together, keeping only variants that were genotyped on both arrays (∼1.1 million variants), and we removed all A/T and C/G SNPs. We then phased this merged dataset using Eagle²⁰ (v.2.4.1) and imputed missing genotypes using a reference panel of 180 diverse African whole-genome sequences²¹ and haplotypes from the 1000 Genomes database²² using the program Impute2.²³ We filtered this imputed dataset to retain only variants with an INFO imputation quality score >0.95, resulting in a final dataset containing 11,001,050 variants for 1,758 people with LTL measurements. We measured LTL by Southern blotting of the terminal restriction fragments.²⁴ All samples were run in duplicate and passed DNA quality and integrity tests. The intra-class correlation coefficient (ICC, otherwise known as repeatability) from duplicate LTL measurements was 0.987 (CI 0.985–0.988).

Population grouping

We grouped all individuals with LTL measurements into 18 population groupings, based on self-identified ethno-linguistic affiliation (Figure 1A; Table S1). In some areas, we grouped multiple ethnic groups together into a single population grouping. For example, in Botswana, individuals self-identifying as any of multiple Khoesan-speaking hunter-gatherer ethnic groups were combined into a single “Botswana, Khoesan, San” grouping. Individuals in Botswana self-identifying as both a Khoesan-speaking group and a Bantu-speaking ethnic group were combined into a single “Botswana, San/Bantu, Admixed” grouping. The Bantu-speaking Herero and the Bantu-speaking Mbukushu ethnic groups in Botswana received their own groupings, respectively, while individuals identifying with any additional Bantu-speaking ethnic group in Botswana were combined into a “Botswana, Bantu, Other” grouping. In Ethiopia, we combined all Afroasiatic Semitic, Cushitic, and Omotic-speakers into single population groupings, respectively. All other ethnic groups comprise their own population grouping, except the Nilo-Saharan speaking Mursi and Surma ethnic groups in Ethiopia, who were grouped together based on shared language family and genetic similarity (Table S1; Figure 1B).

Figure 1.

LTL variation across sub-Saharan Africa

(A) Variation in LTL (adjusted for age and sex) across sub-Saharan Africa.

(B) Principal component analysis (PCA) of all samples, showing the first two principal components.

Genetic analyses

We performed principal component analysis using the smartpca program from the EIGENSOFT package.²⁵ We used PLINK²⁶ (v.1.9) to calculate pairwise linkage disequilibrium among sites within 100 kb windows and excluded sites with pairwise R² > 0.1 and a minor allele frequency (MAF) <1%, resulting in a dataset of 264,471 SNPs for PCA analysis. We performed ADMIXTURE analysis on the same dataset used for PCA analysis. We ran ADMIXTURE 10 times at a k-value of 7.

To estimate the SNP-based heritability of LTL, we constructed a genetic relatedness matrix (GRM) using the same set of 264,471 LD-pruned SNPs used in the PCA analysis. We fit this GRM as a random effect in a restricted maximum likelihood (REML) analysis in the program GCTA²⁷ with age and sex as covariates, and we report the SNP-based heritability estimate from this model (Table S2). We performed a genome-wide association study (GWAS) analysis using linear mixed models in the program GCTA (--mlma option), with age and sex as covariates and the GRM as a random effect. The GWAS only included variants with a MAF >1%. We calculated polygenic scores (PGS) using the ‘--score’ option in Plink (v.1.9) based on a recent GWAS of LTL in 472,174 UK Biobank participants⁵ (hereafter referred to as the “UK Biobank GWAS”; summary statistics obtained from https://figshare.com/s/caa99dc0f76d62990195). Of the 197 independent LTL-associated variants discovered in the UK Biobank GWAS, 73 were present in our final, filtered genetic dataset, which we used to construct the LTL PGS (Table S3). We excluded A/T and C/G SNPs from the polygenic score calculation and discarded the LTL-associated SNP in HBB, as this was shown to be an artifact.⁵

Malaria/environmental data acquisition and analyses

We downloaded global Plasmodium falciparum incidence data in raster format at a 5 km² resolution globally from the malaria atlas project (malariaatlas.org; version 2020; accessed October 2020). We downloaded one layer for each year from 2000 to 2019, and we extracted the mean value across this time frame from the geographic coordinates of each sample. We downloaded bioclimatic and altitude variables (Table 1) in raster format from the worldclim v2 database²⁸ (worldclim.org) at a 2.5 arc-minute resolution globally. We downloaded annual mean UV-B raster data at 15 arc-minute resolution globally from the dataset generated by Beckman et al.²⁹ and publicly available at https://www.ufz.de/gluv/. All raster layer manipulation and analysis was performed using the raster package in R (see web resources).

Table 1.

Environmental variables used in linear mixed models

Environmental factor	Environmental factor p value	Environmental factor beta	Model R2 (marginal)	Model AIC	Variable significant after Bonferroni correction
P. falciparum incidence	0.001677	−1.128	0.3776402	3,657.694	yes
Annual precipitation	0.002367	−2.16E−04	0.3781806	3,659.539	no
Precipitation of wettest quarter	0.004654	−4.14E−04	0.376283	3,661.046	no
Mean temperature of driest quarter	0.011818	−2.74E−02	0.3651958	3,663.502	no
Precipitation of wettest month	0.012337	−9.93E−04	0.3716752	3,663.379	no
Min temperature of coldest month	0.013275	−2.59E−02	0.3692669	3,663.149	no
Mean diurnal range	0.037507	3.28E−02	0.3695131	3,665.128	no
Precipitation seasonality	0.03768	4.29E−03	0.3582584	3,666.471	no
Mean temperature of coldest quarter	0.049774	−2.42E−02	0.3582539	3,666.706	no
Precipitation of driest quarter	0.06347	−1.71E−03	0.3562135	3,667.776	no
Precipitation of coldest quarter	0.0896	−1.80E−04	0.3613363	3,668.227	no
Precipitation of driest month	0.10124	−5.65E−03	0.3543116	3,668.62	no
Annual mean temperature	0.107395	−2.03E−02	0.3514084	3,668.548	no
Mean temperature of warmest quarter	0.115193	−1.78E−02	0.349294	3,668.786	no
Altitude	0.155476	1.08E−04	0.3497236	3,669.093	no
Temperature annual range	0.260588	9.52E−03	0.3591513	3,669.442	no
Isothermality	0.316905	6.64E−03	0.3474315	3,670.468	no
Annual mean UV-B	0.339884	8.03E−05	0.3487421	3,670.726	no
Max temperature of warmest month	0.40991	−8.01E−03	0.3459811	3,671.017	no
Mean temperature of wettest quarter	0.461032	−8.17E−03	0.3465636	3,671.12	no
Precipitation of warmest quarter	0.472223	2.69E−04	0.3466803	3,671.332	no
Temperature seasonality	0.62055	2.39E−04	0.3526096	3,671.115	no

List of environmental variables used in linear mixed models, sorted by ascending p value. Each variable was included in a separate linear mixed model with LTL as the response variable, and age, sex, the LTL PGS, and 20 principal components (PCs) included as fixed explanatory variables. Sampling site was included as a random effect. Malaria incidence is the only variable that passes Bonferroni p value correction and has the lowest model AIC.

Statistical analyses

Linear models were run in R using the lm function. Linear mixed models were run using the lme4³⁰ and lmerTest³¹ packages in R. To adjust LTL for covariates (e.g., age and sex), we ran a multiple linear regression with age and sex as predictors, and then extracted residuals from this model. We then added the overall raw mean LTL to each residual to get the adjusted value (e.g., Figure 1A). In linear mixed models containing environmental variables, we used the fixed effects p values generated by the lmerTest R package, using the Satterthwaite method³¹ (e.g., Table 2).

Table 2.

Linear mixed model output

Predictors	Estimates	LTL CI	p
Intercept	8.541	8.365–8.718	<0.001
Age	−0.022	−0.024–−0.020	<0.001
Sex [M]	−0.153	−0.219–−0.088	<0.001
PGS	0.425	0.206–0.645	<0.001
P. falciparum Incidence	−1.128	−1.807–−0.449	0.002
PC1	5.676	3.940–7.411	<0.001
PC2	−2.012	−3.826–−0.197	0.031
PC3	4.942	2.150–7.734	0.001
PC4	−5.65	−7.892–−3.408	<0.001
PC5	0.627	−1.520–2.774	0.562
PC6	−0.585	−2.331–1.161	0.502
PC7	−0.398	−1.792–0.997	0.576
PC8	0.617	−0.739–1.974	0.372
PC9	0.24	−1.285–1.764	0.756
PC10	0.567	−0.807–1.942	0.418
PC11	1.16	−0.179–2.500	0.09
PC12	0.02	−1.640–1.681	0.981
PC13	1.138	−0.210–2.486	0.098
PC14	−1.215	−2.848–0.419	0.142
PC15	−2.177	−3.807–−0.547	0.01
PC16	1.943	0.326–3.560	0.019
PC17	−0.457	−1.878–0.965	0.528
PC18	0.351	−1.002–1.703	0.611
PC19	0.55	−0.808–1.908	0.427
PC20	2.803	1.279–4.327	<0.001
Random Effects
σ2			0.46
τ00Sampling Site			0.01
ICC			0.02
N Sampling Site			55
Observations			1758
Marginal R2/conditional R2			0.378/0.387

Linear mixed model output for a model containing age, sex, 20 principal components, the LTL PGS, and Plasmodium falciparum malaria incidence as explanatory variables. Sampling site included as a random effect. Malaria has a significant effect on shortening LTL, adjusting for all other covariates.

Results

Effects of age, sex, and genetics on LTL

We measured LTL by Southern blotting of the terminal restriction fragments²⁴ in 1,818 ethnically diverse SSAfr adults from Tanzania, Botswana, Ethiopia, and Cameroon (Figure 1A). Samples collected from Tanzania, Botswana, and Ethiopia were analyzed by Hunt et al.,¹³ and we add here a set of 459 individuals from Cameroon that have not previously been analyzed for LTL and extend the analyses to test for associations with genetic variation, malaria endemicity, and other environmental factors. The mean age of participants was 40.7 years (range 18–100 years) and 54% were female. Study participants reflected a large portion of the linguistic and genetic diversity found within Africa,³² representing the four major African language families: Afroasiatic, Khoesan, Nilo-Saharan, and Niger-Congo. Niger-Congo is the largest language family in Africa, representing ∼1,500 languages, of which the Bantu subfamily of languages are the most widespread.³² Study participants consist of agriculturalists, pastoralists, as well as hunter-gatherers, and live in areas where falciparum malaria endemicity ranges from low (western Botswana)³³ to very high (Cameroon).³⁴ For genetic analyses, we integrated a dataset containing genotypes at ∼11 million single-nucleotide polymorphisms (SNPs) for 1,758 people with LTL measurements (see material and methods).

We first used linear models to determine the effects of age and sex on LTL across the African cohort by performing a multiple linear regression with LTL as the response variable and age and sex as predictors. We found that LTL was shorter in males than females (β = −0.127 kilobases [kb], p = 4.9 × 10⁻⁴; Figure S1) and that it shortened by 24 bp/year (β = −0.024 kb, p < 2 × 10⁻¹⁶; Figure S2). Together, age and sex explained 21% of the LTL variation across individuals (linear model R² = 0.21, p < 2.2 × 10⁻¹⁶). We used this linear model to adjust LTL for age and sex across the cohort by extracting model residuals for each individual and adding the overall raw mean LTL to each residual (Figure 1A). Mean LTL in this cohort was 7.6 ± 0.87 kb (SD), i.e., the estimates of LTL at mean age (40.7 years) and sex (54% female) in the cohort.

We next classified individuals into 18 population groupings based on self-identified ethno-linguistic affiliation, which we outline in detail in Table S1 and describe in the material and methods section (Figure 1A). After adjusting LTL for age and sex, we found that LTL varied significantly across these population groupings (one-way ANOVA population grouping: F(17,1800) = 29.79, p < 2 × 10⁻¹⁶) (Figure 1A). The San from Botswana, who traditionally are hunter-gatherers, had the longest age-and-sex adjusted LTL (8.2 ± 0.72 kb), while the Fulani pastoralists from Cameroon had the shortest LTL (6.9 ± 0.71 kb). The Botswana San had significantly longer LTL than every other African population grouping (all pairwise p < 0.005), except for the “Botswana, San/Bantu, Admixed” group (p = 0.08) and the Afroasiatic-speaking Burunge in Tanzania (p = 0.06). Together, age, sex, and population grouping explained 38% of the LTL variation across individuals (linear model R² = 0.38, p value <2.2 × 10⁻¹⁶).

We examined samples from Cameroon in more detail, as these individuals were not included in Hunt et al.¹³ and inhabit the highest malaria endemic regions in our dataset (see below). Here, age- and sex-adjusted LTL varied significantly across population groupings (one-way ANOVA population grouping: F(2,456) = 37.88, p = 6.02 × 10⁻¹⁶). The Baka and Bagyeli, two hunter-gatherer groups commonly referred to as “Central African Rain Forest Hunter-Gatherers” (hereafter referred to as rainforest hunter-gatherers or RHG), had significantly longer LTL than the neighboring Bantu-speaking Tikari agriculturalists (Baka and Bagyeli mean ± SD = 7.5 ± 0.64 kb; Tikari = 7.0 ± 0.67 kb; Tukey’s HSD p = 3.9 × 10⁻¹¹) and Fulani pastoralists (6.9 ± 0.71 kb; Tukey’s HSD p < 2.2 × 10⁻¹⁶). LTL in the Fulani was not significantly different than LTL in the Bantu-speaking Tikari agriculturalists (Tukey’s HSD p = 0.44).

To examine the influence of genetics on LTL, we estimated the proportion of variance in LTL explained by all SNPs genome-wide, referred to as the “SNP-based heritability,” by performing a restricted maximum likelihood (REML) analysis with the program GCTA.²⁷ This analysis uses autosomal SNPs to estimate pairwise genetic relationships between all individuals in the form of a genetic relatedness matrix (GRM), and then fits this matrix as a random effect in a linear mixed model. We found that the SNP-based heritability of LTL was 66% ± 0.06% (SE) when including age and sex as covariates in the model (Table S2). When including the above-described population groupings (Figure 1A) as an additional covariate, the heritability estimate dropped to 51%, which was similar to the 52% heritability of LTL shown by Hunt et al.¹³

LTL polygenic score in Africa

We next sought to understand how the genetic architecture underlying LTL varies between Europeans and sub-Saharan Africans. To do this, we constructed a polygenic score (PGS) for LTL in the African samples, based on a recent GWAS of LTL in 472,174 UK Biobank participants.⁵ This PGS aggregates information across many trait-associated variants to summarize an individual’s genetic prediction for the trait of interest and is simply the effect size (beta) of a trait-associated allele weighted by the number of alleles a person carries, summed across all trait-associated SNPs. Of the 197 independent LTL-associated variants discovered in the UK Biobank GWAS, 73 were present in our final, filtered genetic dataset (see material and methods), which we used to construct the LTL PGS (Table S3). We found that this PGS explained 3.2% of the inter-individual variation in LTL across Africa, after adjustment for age and sex (Figure 2A). However, after an additional adjustment of 20 genotyping PCs from principal component analysis (PCA) (Figure 1B), this PGS explained only 0.8% of the inter-individual variation in LTL across Africa (model R² = 0.008). For comparison, the 197 LTL-associated variants explained 4.54% of the variation in LTL in the UK Biobank.⁵ This result suggests not only that there is still substantial unaccounted for genetic variation explaining LTL in Europeans and Africans, but that there is some level of PGS transferability between Europe and Africa for this trait, albeit low. This PGS also varies significantly across population groupings in Africa (one-way ANOVA population grouping: F(17,1740) = 10.16, p < 2 × 10⁻¹⁶), with the Botswana San having the “longest” genetically predicted LTL and the Fulani pastoralists having the “shortest” predicted LTL (Figure 2B). In pairwise population comparisons, the Botswana San had a significantly larger LTL PGS than all other population groupings (all p < 0.001), except the “Botswana, San/Bantu, Admixed” grouping (p = 0.57) and the “Botswana, Bantu, Other” grouping (p = 0.24). These population differences in the LTL PGS broadly mirror observed LTL differences between these groups (Figures 1A and 2B).

Figure 2.

LTL polygenic score variation across sub-Saharan Africa

(A) Significant relationship between the LTL polygenic score (PGS) and LTL (adjusted for age and sex) using 73 LTL-associated variants from the UK Biobank.

(B) Variation in the LTL PGS across African population groupings.

We further examined allele frequency differences between the UK Biobank GWAS cohort and the African cohort at the 73 loci used to construct the LTL PGS. To do this, we calculated frequencies of the LTL-increasing allele (as defined by the UK Biobank GWAS) at each of the 73 GWAS loci, in each of the African populations. We then compared them to the frequencies in the UK Biobank GWAS cohort (Figure S3). The Botswana San hunter-gatherers have the greatest allele frequency difference on average, with the trait-increasing allele frequency 2.5% higher than in the UK Biobank GWAS cohort (averaged across the 73 variants; Figure S3A). In contrast, the Cameroonian Fulani show the largest negative allele frequency difference, with the trait-increasing allele frequency 2.4% lower than in the UK Biobank GWAS cohort, on average (Figure S3A). We further show that these allele frequency differences are not distributed evenly across the 73 variants (Figure S3B). For example, the trait-increasing allele at rs28711261 (in CTCF; 16:67617186 [hg19]) is 66% higher in the San than in the UK Biobank GWAS cohort. In contrast, the trait-increasing allele at rs752720994 (in POLN; chr4:2191750) is 57% lower in the Fulani than in the UK Biobank GWAS cohort (Figure S3B).

We then performed a GWAS analysis for LTL in the African cohort. Although we did not find any loci reaching genome-wide significance (p ≤ 5 × 10⁻⁸), likely due to our relatively small sample size, we did find some variants (n = 55) reaching “suggestive” significance (p ≤ 1 × 10⁻⁵ as defined by previous work³⁵) (Figure S4; Table S4). However, none of the SNPs reaching suggestive significance in Africa overlap within 500 kb of a top LTL-associated variant identified in the well-powered UK Biobank GWAS. In addition, gene enrichment analysis using a list of genes within 100 kb of our suggestive GWAS variants did not return any biological processes obviously involved in telomere biology. The Q-Q plot from the GWAS also suggests a null result (Figure S4 inset). Finally, we examined the level of trait association at the 73 variants used to create the PGS within our GWAS results. We found that of these 73 variants, only 7 (9.6%) had p values <0.05 in our African GWAS (Table S3) and 42 (57.5%) had beta values consistent in direction. Together, these results suggest that our African GWAS is underpowered and that much larger sample sizes will be needed to power informative LTL GWASs in Africa, due to the highly polygenic nature of this trait.⁵

Association of LTL with malaria endemicity and environmental variables

To explore the relationship between LTL and malaria endemicity, we first used the geographic coordinates of the African samples to extract publicly available high-spatial-resolution malaria endemicity data from the Malaria Atlas Project³⁴ (https://malariaatlas.org/). Specifically, we examined malaria incidence at a 5 km² resolution across the African continent (Figure 3A). This measure of malaria endemicity is based on cross-sectional survey data from malaria endemic countries and equates to the number of malaria cases per person, per year. We examined estimates for each year from 2000 to 2019 and extracted the average value across this period from the geographic coordinates of each sample. We found that age- and sex-adjusted LTL was inversely correlated with malaria incidence (p = 1.4 × 10⁻⁴, R² = 0.24; Figure 3B). In other words, mean LTL at a sampling site shortens as malaria incidence increases.

Figure 3.

Correlation of LTL with malaria endemicity

(A) P. falciparum incidence in Africa averaged across the years 2000–2019, with sampling sites shown as black squares (n = 55).

(B) Mean LTL (adjusted for age and sex) at a sampling site shortens with increased malaria exposure. Each point is a sampling site, with the size of the point corresponding to the number of individuals sampled at that site, and colored by country.

To compare this negative association between malaria and LTL to other climatic and environmental variables that might play a role in LTL, we used linear mixed models to examine the influence of 21 additional environmental variables (Table 1) on LTL while adjusting for age, sex, the LTL polygenic score (PGS), and genetic ancestry inferred from principal component analysis (PCA) using a set of genome-wide SNPs (Figure 1B). The additional environmental variables include 19 bioclimatic variables from the WorldClim database²⁸ which reflect annual trends in temperature, precipitation, extreme climatic factors, and seasonality. We also examined altitude and annual mean UV-B radiation,²⁹ as previous studies have found associations between LTL and melanoma-related variables.³⁶ For each environmental variable listed in Table 1, we fit a linear mixed model with LTL as the response variable and age, sex, the first 20 ancestry principal components from PCA analysis (Figure 1B), the LTL PGS, and the environmental variable of interest as fixed explanatory variables. We included sampling site as a random effect in each model. We tested 22 different environmental variables (21 environmental variables + malaria incidence) in 22 different models. We chose not to include all environmental variables in a single model in order to avoid interpretation issues caused by the strong correlations between the different environmental variables (multicollinearity). We identify several environmental factors that are nominally significant (p < 0.05) (Table 1). However, of these nominally significant factors, most are highly correlated with malaria (all |Pearson’s R| > 0.43). To account for multiple testing, we set a Bonferroni-corrected significance p value cutoff of ≤0.0023 (p = 0.05/22) for the environmental variable of interest, to conservatively identify environmental factors with a statistically significant effect on LTL. We found that malaria incidence was the only environmental factor that passed this p value cutoff (beta = −1.128 kb, variable p value = 1.67 × 10⁻³, SE = 0.337 kb) (Table 1), indicating that short LTL was significantly associated with increasing malaria exposure. We present the linear mixed model containing malaria incidence in Table 2 (hereafter referred to as the “full model”). We next performed a likelihood ratio test to compare this full model to a similar model that differed only in that it did not include the malaria fixed effect. We found that the model that included malaria showed a significantly better fit to the underlying data than the model without malaria (likelihood ratio test χ²(1) = 14.0, p = 1.8 × 10⁻⁴; Table S5). Further, the full model containing malaria incidence had the lowest Akaike information criterion (AIC) estimate of all the models we examined (Table 1), which suggests malaria was the most important environmental variable we considered. In this instance, the model with the lowest AIC score identifies the one that best fits the underlying data and is often used as a criteria for model selection. Finally, when including an interaction term between malaria and age in the full model, we found that the interaction was not significant (p = 0.33), suggesting malaria exposure does not influence the rate of LTL shortening with age in adults.

To verify the association between malaria and a shorter LTL, we performed an additional analysis using the program GCTA, which uses a slightly different but complementary methodology to the linear mixed models described above. In this analysis, we fit two genetic relationship matrices (GRMs) as random effects in a linear mixed model, with age, sex, and malaria incidence as fixed effect covariates and LTL as the response variable. One GRM was built using all autosomal SNPs, and the other was built from the 73 SNPs used to construct the LTL PGS (see above, Figure 2; Table S3). We again found that short LTL was associated with increasing malaria endemicity (malaria incidence β = −1.53 kb, SE = 0.283 kb, p < 0.0001; Table S6). Given the association between LTL and malaria, we performed another GWAS for LTL with malaria incidence as an added covariate. However, we still find no genome-wide significant signals in this analysis (Figure S5).

Another way to distinguish genetic from environmental factors influencing a trait is to examine populations of similar genetic ancestry that inhabit vastly different environments. If the trait of interest differs greatly between populations that are genetically similar, the effect is more likely to be environmental, as opposed to genetic, in origin. We identify just such a signature of an environmental impact on LTL in Africa, since populations of similar genetic ancestry living in vastly different malarial environments showed large differences in LTL. For example, the Bantu-speaking Herero and Mbukushu ethnic groups in Botswana are genetically similar to the Bantu-speaking Tikari in Cameroon; these groups cluster very closely in PCA analysis (Figure 1B, zoomed in PCA in Figure S6) and share similar admixture proportions in ADMIXTURE³⁷ analysis (Figure S7). In addition, these populations have virtually identical mean LTL PGS values (Figure 2B; Herero mean [±SD] PGS = 0.54 [0.13]; Mbukushu = 0.54 [0.13]; Tikari = 0.54 [0.13]). The genetic similarity of these geographically distant populations is due in part to the migration of Bantu speakers from Central African regions to Eastern and Southern Africa starting ∼5,000 years ago³⁸ (i.e., the “Bantu expansion”). However, despite their genetic affinity, LTL differs significantly between these populations. The Tikari, who are indigenous to high-malaria endemic areas in Cameroon, had a significantly shorter age- and sex-adjusted LTL (mean = 7.0 kb) than the Herero (mean = 7.7 kb; t test, t(199) = 6.2, p = 2.9 × 10⁻⁹) and the Mbukushu (mean = 7.8 kb; t test, t(182) = 5.5, p = 1.3 × 10⁻⁷), who are both indigenous to low-malaria regions in Botswana (Figures 1A and 3A). Because the Mbukushu and Herero share a small amount of ancestry with the San (Figure S7), we further wanted to account for the fact that this San ancestry may influence this difference in LTL between Botswanan and Cameroonian Bantu speakers. To do this, we examined only the individuals from these three populations (Mbukushu, Herero, and Tikari) and constructed a multiple linear regression model with LTL as the response variable, and age, sex, population grouping, and the San admixture component (from ADMIXTURE analysis; Figure S7) as fixed effects. From this model, we find that LTL in the Mbukushu and Herero is still significantly longer than the Tikari (p = 1.6 × 10⁻⁶ and 2.3 × 10⁻⁵, respectively). These results suggest that admixture with the San is not driving the large differences in LTL among these populations. This pattern of genetically similar populations from different malarial environments having very different LTL is consistent with the significant association with malaria that we observe in the full model described above (Table 2).

In contrast to the above scenario, one can examine populations of differing genetic ancestry that inhabit similar environments to determine whether a trait is more determined by genetics than environmental factors. In our dataset, the Baka and Bagyeli hunter-gatherers from Cameroon (RHG) have differing genetic ancestries from the neighboring Tikari agriculturalists and Fulani pastoralists (Figures 1B, S6, and S7), but these populations all experience a relatively high level of malaria exposure compared to the rest of our cohort. The RHG have significantly longer LTL than the Tikari (p = 3.9 × 10⁻¹¹) and the Fulani (p < 2.2 × 10⁻¹⁶), as well as a “longer” LTL PGS, though not significantly so (Figure 2B). This result might suggest an underlying genetic or life-history mechanism which may lengthen telomeres in the RHG (or shorten them in the Tikari/Fulani), but more work is needed to fully understand this observation. Similarly, the Botswana San hunter-gatherers have a different genetic ancestry and have significantly longer LTL than the neighboring Bantu-speaking populations in Botswana, despite inhabiting a similar environment (Figure 1A). The fact that the San and the RHG have comparatively longer LTL than neighboring populations might be explained by the ancient shared common ancestry between them.²¹

Covariance of LTL-associated genetic variation with malaria endemicity

Given that a short LTL is associated with increasing malaria exposure, it is possible that selection may favor a genetic predisposition for longer telomeres in endemic regions, as a buffer against this shortening. Such a process might explain in part why individuals of African ancestry have a longer LTL, on average, than non-Africans from non-endemic regions. However, the data do not support this notion. When examining the LTL PGS across the entire African cohort, we found that there was a significant negative relationship with malaria incidence (i.e., genetically “predicted” LTL as determined by the PGS decreases as malaria increases; Pearson’s r = −0.13; beta = −0.15; p = 4.8 × 10⁻⁸; Figure S8). To investigate the potential effect of malaria on the PGS further, we constructed a linear mixed model with the PGS as the response variable, and the first 20 principal components and malaria incidence as predictor variables. We included sampling site as a random effect in this model. We found that the relationship between malaria and the PGS in this model was still negative (beta = −1.22; i.e., the PGS becomes “shorter” as malaria incidence increases), although it was no longer significant (p = 0.07).

Discussion

Our study shows that LTL is shorter in adult SSAfrs indigenous to regions of high malaria endemicity than in those indigenous to regions of low malaria endemicity. Although the LTL PGS predicts shorter telomeres in malaria endemic regions, the association between malaria and LTL remains significant while adjusting for age, sex, and genetic factors (Table 2). The difference in LTL between sampling sites with the lowest malaria endemicity compared to sites with the highest endemicity in our dataset is ∼500 bp (Figure 3). This difference is consistent with the beta estimate we observe for malaria incidence at the individual level in the full model, which equates to an LTL shortening of ∼1.1 kb for every unit increase in malaria incidence (i.e., increasing malaria from zero cases/person/year to 1 case/person/year; Table 2). Mean malaria incidence in our dataset ranges from 0 to 0.45 cases/person/year, which would equate to an LTL shortening of ∼500 bp from one extreme (no malaria) to the other (0.45 cases/person/year).

The potential impact of malaria endemicity on LTL reported in this study appears larger than previously identified environmental factors that impact LTL. For instance, a meta-analysis based on about 400,000 non-African adults revealed a nominal effect of environmental factors on LTL,⁹ with most environmental factors or exposures correlating only weakly (−0.2 < r < 0.2) with telomere length. Similarly, potentially modifiable traits and behaviors such as smoking, diet, and physical activity levels jointly explained less than 0.2% of the inter-individual LTL variation among 400,000 adult participants in the UK Biobank.⁸ What then might be the mechanism to account for the correlation of malaria and LTL? We hypothesize that one potential mechanism may involve malaria-induced bouts of massive erythrocyte destruction and erythropoiesis, by which malaria may shorten LTL (Figure 4). We describe this model below.

Figure 4.

Model of malaria-induced leukocyte telomere length shortening

Malaria-induced red blood cell loss stimulates repeated bouts of erythropoiesis and telomere shortening. Figure credit to Anita Zhang.

The activity of telomerase, the reverse transcriptase that lengthens telomeres,³⁹ is insufficient to prevent telomere shortening resulting from hematopoietic cell replication throughout the hematopoietic hierarchy.^40,41 Consequently, hematopoietic cells experience telomere shortening with age. Population-based studies typically measure telomere length in leukocytes as an indicator of telomere length in hematopoietic cells. This has generated a misconception that LTL shortening with age is largely explained by the turnover of circulating leukocytes. However, hematopoietic cell telomere shortening is largely the outcome of erythropoiesis—the process by which erythrocytes (red blood cells) are produced from hematopoietic stem cells.⁴² Circulating erythrocytes outnumber circulating leukocytes by approximately a thousand to one and comprise 84% of all somatic cells in the body.^42,43 The telomere length reserves of the hematopoietic system are, thus, principally spent on building and maintaining the massive pool of about 25 trillion erythrocytes in the average human adult.⁴³

Malaria causes massive destruction of erythrocytes and their heightened clearance by the spleen.⁴⁴ Numerous biological and physiological mechanisms are triggered in response to acute erythrocyte loss. Ultimately, however, only erythropoiesis can restore the loss of erythrocytes. Children raised in regions of high malaria endemicity experience numerous bouts of erythrocyte destruction that require robust erythropoietic response, which over time might increase shortening of the hematopoietic cell telomeres from the top of the hematopoietic hierarchy to its bottom. Depending on the intensity and frequency of malaria, these children, we propose, are more likely to become adults with short hematopoietic cell telomeres, as expressed in LTL (Figure 4). However, after repeated malaria infections during childhood, adults in regions of high malaria endemicity are largely immune against the blood stage of Plasmodium falciparum.^45,46 In line with this proposal, there was no indication in our study that the rate of LTL shortening with age depended on malaria incidence in adults participating in this research.

Two recent studies in Europeans report massive shortening of LTL, measured by a qPCR-based method, during acute malaria infection. Participants in both studies were treated with anti-malarial drugs. The first study was performed in adults (ages 26–65 years) diagnosed with malaria upon returning from their travels.¹⁵ In these patients, LTL was shortest three months post diagnosis, after which LTL began to slowly recover over the course of a year. However, LTL was never measured at “baseline” (i.e., before malaria infection), so it remains unknown whether LTL truly recovered to its original level. In the second study, young adults (ages 20–29) were subjected to controlled acute malaria infection.¹⁶ In these persons, LTL was the shortest 10–13 days after infection and had recovered by 2 months after infection. The reversal of LTL shortening in both studies suggests a transient effect of malaria on circulating leukocytes rather than an irreversible shortening effect of the infection on telomere length stemming from replication of hematopoietic cells up the hematopoietic hierarchy. However, it is unclear how relevant these data are to populations living in malaria endemic regions, where individuals are infected repeatedly throughout life and often receive no medical treatment. Further, the qPCR methodology used in these studies has a larger measurement error and correlates only modestly with the more accurate Southern blotting technique used here.^12,47 Specifically, a large study comparing Southern blot estimates of LTL with those derived from a qPCR methodology showed that the correlation between the two measures was modest (r = 0.52), while the inter-assay coefficient of variation (CV) was almost four times higher for the qPCR data compared to the Southern blot data (5.8% vs. 1.5%).¹²

We acknowledge the following limitations of our study: First, we propose that the effect of malaria on hematopoietic cell telomere shortening with age primarily unfolds during childhood, yet our LTL data are derived from adults. Clearly, the next step in testing the relationship between malaria and LTL is to characterize LTL dynamics in children born and raised in regions of high malaria endemicity versus those born and raised in regions of low or no malaria endemicity. Second, our study is cross-sectional. A longitudinal study, not only in children but also in adults indigenous to regions of high and low malaria endemicity, would provide more insightful information. Further, we relied on regional estimates of malaria endemicity rather than individual patient histories. However, when malaria effects on LTL arise in childhood (i.e., many years before donors were sampled), regional estimates of historical malaria endemicity may well be the most robust information on population level prevalence of malaria infection during childhood that is available. It is also possible that the geographic measure of malaria endemicity we use here is correlated with some other unmeasured environmental, lifestyle, socio-economic, or pathogenic factor that may drive the association with LTL. Future directed studies will be needed to tease apart the effects of these factors on LTL. Finally, the LTL PGS we constructed was based on GWAS variants discovered in a predominantly European cohort,⁵ which may not transfer well to African populations.⁴⁸ Some potential reasons for lack of PGS transferability include differences in effect allele frequency across populations (Figure S3), differential tagging of causal variants due to differences in linkage disequilibrium (LD) across populations, gene-by-environment interactions that differ by population, and different causal mutations occurring in different populations.⁴⁹ Future work and larger samples will be needed to uncover the exact loci underlying LTL variation in Africa.

In conclusion, we highlight the contributions of genetic and environmental factors influencing LTL, and we have uncovered a potential role of malaria in shortening LTL across sub-Saharan Africa. This association between malaria and LTL appears larger than any other known exposure or behavior that has been investigated in large-scale studies. We also propose a model by which we hypothesize malaria might exert this effect on LTL (Figure 4).

Data and code availability

The genotype and phenotype data underlying this study is available at dbGaP accession:phs003567.v1.p1.

Published: May 2, 2024

Web resources

• raster, https://cran.r-project.org/web/packages/raster/index.html