The impact of Indigenous American-like ancestry on the risk of acute lymphoblastic leukemia in Hispanic/Latino children

Summary

Acute lymphoblastic leukemia (ALL) is the most common childhood cancer, with Hispanic/Latino children having a higher incidence of ALL than other racial/ethnic groups. Among the genetic variants previously implicated in ALL risk, a number of them were found to be enriched in Indigenous American (IA)-like ancestries and inherited by many Hispanic/Latino individuals. However, due to potential confounding from environmental factors, the association between IA-like ancestry and risk for ALL has remained unclear. In this study, we characterized the impact of IA-like ancestry on overall ALL risk and on the frequency and effect size of known risk alleles, while accounting for non-genetic correlates of ancestry. Contrary to previous findings, we found that global IA-like ancestry was not significantly associated with ALL risk after adjusting for socioeconomic indicators. However, locally at known ALL risk regions, we uncovered that increasing copies of the IA-like haplotype were positively and significantly associated with ALL risk (e.g., the IA-like haplotype had ∼1.33 times the odds of harboring the risk allele compared to non-IA-like haplotypes), but we found no evidence of interaction between genotype and ancestry in relation to ALL. Admixture mapping identified replicable association signals at chr7p12.2 and chr10q21.2, consistent with the benefit of leveraging genetic ancestry in identifying genetic risk loci. Our results suggest that increased risk of ALL in Hispanic/Latino children may be conferred by the higher frequency of risk alleles within IA-like ancestry and that local ancestry-based analyses are robust strategies to elucidate genetic etiology of disease.

Through leveraging local ancestry and sociocultural factor data, we clarify the role of Indigenous American (IA)-like ancestries in acute lymphoblastic leukemia (ALL) risk. We show that alleles present at higher frequencies in IA-like ancestries confer an increased risk for ALL, rather than exerting larger effects on IA haplotypes compared to non-IA haplotypes.

Introduction

Acute lymphoblastic leukemia (ALL) is the most common childhood cancer and accounts for 25% of malignancies before the age of 20 years.¹ The long-term survival rate of the disease in children is upwards of 85%–90%,² but the survival rate decreases to approximately 40% after relapse^3,4,5,6 and to 30%–40% in individuals older than age 35. Despite decades of investigation into ALL risk, the etiology of the disease is not well understood, and disparities in risk persist. Children of Hispanic/Latino ethnicity have 1.4 and 2 times the risk of developing ALL compared to their Non-Latino White and Black counterparts, respectively.⁷ This disparity in risk widens in margin for Hispanic/Latino adults⁷ and is specific to the B cell immunophenotype, which comprises approximately 85% of cases.⁸ While environmental factors are involved in this increased risk, genetic factors, which have been historically understudied in this population,^9,10 likely contribute in part to this disparity.^11,12

Genomic, migrational, and ecological data strongly suggest a complex genetic etiology of ALL, both broadly across humans and within diverse Latin American populations. Most directly, previous genome-wide association studies (GWASs) have estimated the heritability of ALL to be approximately 18%–21% in both Hispanic/Latino and Non-Latino White populations, and approximately 24% of familial relative risk has been attributed to known ALL risk variants.^13,14,15 Consistent with these estimates, GWAS have identified a number of loci associated with ALL susceptibility, including risk alleles located near genes such as IKZF1, ARID5B, GATA3, and ERG that appear to have higher frequency or stronger effects in Hispanic/Latino populations.^12,16,17,18 Previous migration studies have indicated that non-native and US-born Hispanic/Latino individuals have comparable rates of B cell ALL to each other, and both have higher (∼2.13 times) incidence rates compared to their Non-Latino White counterparts.^19,20 Furthermore, ecological data have indicated that countries with the highest age-standardized incidence rates of ALL were in Latin America (e.g., Honduras and Mexico), while countries with the lowest incidence rates were in Africa, with European countries exhibiting intermediate incidence.^21,22

Notable among populations with differing incidence rates of ALL is the presence or dominance of a particular ancestry component. Due to the unique demographic history of Latin America, Hispanic/Latino populations are admixed, drawing ancestral components from the American, European, and African continents.^23,24 Previous literature has alluded to associations between Indigenous American (IA)-like ancestry and ALL, suggesting that every 20% increase in IA-like ancestry was associated with a 20% increase in odds of developing B cell ALL.²⁵ Because Hispanics/Latinos inherited a significant proportion of their genetic ancestries from Indigenous Americans (although there is geographical heterogeneity across the Americas), this association was presumed to explain the migration and ecological findings above. However, global ancestry can be confounded by non-genetic factors such as socioeconomic status (SES),²⁶ which was not considered in previous analyses.²⁵ Further exploration of the association between IA-like ancestry and ALL risk would need to account for this potential confounding.

Given the important role of genetic factors and genetic ancestry in ALL risk, we undertook a systematic admixture analysis in the largest ALL cohort of self-reported Hispanic/Latino children to date. Here, we sought to clarify the association between IA-like ancestry and ALL risk through a study design that would better account for non-genetic confounders and leveraged local ancestry analyses, which are less susceptible to confounding by environmental factors and have previously yielded promising results in examining ancestry-ALL relationships.²⁵ We investigated both inferred global ancestry across the genome and local ancestry at known ALL loci for associations with risk. Furthermore, we employed admixture mapping,²⁷ which is best powered to detect loci that exhibit frequency differences between ancestries and may underlie the risk differences between populations.^28,29

We note that given our goal to further understand the disparity in ALL risk epidemiologically observed among Hispanic/Latino children, we inferred the genetic ancestries and characterized their associations with ALL risk in self-reported Hispanic/Latino individuals. We defined genetic ancestries at the continental (e.g., Africa, Europe, and the Americas) level, which represents patterns of genetic similarities across geographical space over the last ∼10,000–50,000 years due to human migration and isolation by distance. These relatively distinct components of ancestries, although arbitrarily defined, are only recently reunited in admixed populations, such as in individuals of Hispanic/Latino origin. These components are inferred and defined relative to reference populations available to us and are leveraged to uncover associations with traits in this research setting, but they are invariably a simplified representation of the complex nature of genetic ancestries. We thus append the suffix “-like” to refer to these ancestries to remind the reader of this simplification and uncertainty.³⁰

Subjects and methods

Study subjects and quality control

We use the term “Hispanics/Latinos” as a population descriptor of our study subjects to align with both the self-reported race/ethnicity categories and the definitions of the US Office of Management and Budget,³¹ while acknowledging that we focus on individuals of Latin American origin. Our analysis utilized the following cohorts of self-reported Hispanic/Latino study participants: the California Cancer Record Linkage Project (CCRLP), the Kaiser Resource for Genetic Epidemiology Research on Aging cohort (Kaiser-GERA), the California Childhood Leukemia Study (CCLS), Children’s Oncology Group (COG), Multi-Ethnic Study of Atherosclerosis (MESA), the Adolescent and Childhood Cancer Epidemiology and Susceptibility Service for Texas and Reducing Ethnic Disparities in Acute Leukemia (ACCESS/REDIAL), and the Hispanic Community Health Study/Study of Latinos (HCHS/SOL). All relevant information on sample size, arrays, study design, and references/Database of Genotypes and Phenotypes dbGaP accession numbers can be found in Table 1. Following the practice in previous studies,^12,13,32 cases from CCRLP were augmented with external controls from Kaiser-GERA genotyped on a similar array. Cases from COG were paired with external controls from MESA, and ACCESS/REDIAL cases were paired with external controls from HCHS/SOL. HCHS/SOL was a multi-center study, with study participants being ascertained from Bronx, New York; Chicago, Illinois; Miami, Florida; and San Diego, California. The SOL San Diego (SOL-SD) center was chosen as the controls for ACCESS/REDIAL due to similarity in the first 2 principal components and estimated ancestry distributions (Figures S1 and S2).

Table 1.

All cohorts included in the analysis

Study	In-text name/acronym	Study design	N	Genotype array	Reference or accession number	Usage
California Cancer Report Linkage Project	CCRLP	population-based case-control study	4,033	Affymetrix Axiom World Array	Wiemels et al.32	Global and local ancestry analyses; Discovery Admixture Mapping cohort (combined with Kaiser-GERA)
Kaiser Resource for Genetic Epidemiology Research on Aging cohort	Kaiser-GERA	cohort study (controls only)	6,417	Affymetrix Axiom World Array	phs000788.v1.p2	Discovery Admixture Mapping cohort (combined with CCRLP)
California Childhood Leukemia Study	CCLS	population-based case-control study (cases and controls)	1,120	Illumina OmniExpress version 1	Ma et al.33	Global and local ancestry analysis; replication cohort
Children’s Oncology Group (AALL0232, COG9904/9905)	COG	clinical trial (cases only)	836	Genome-Wide Human SNP Array 6.0	Yang et al., 201134	Replication cohort (combined with MESA)
Multi-Ethnic Study of Atherosclerosis	MESA	cohort study (controls only)	1,118	Genome-Wide Human SNP Array 6.0	dbGap: phs000209.v13.p3	Replication cohort (combined with COG)
Adolescent and Childhood Cancer Epidemiology and Susceptibility Service for Texas and Reducing Ethnic Disparities in Acute Leukemia	ACCESS/REDIAL	hospital cases (cases only)	510	Illumina Global Screening Array	Harris et al.35	Replication cohort (combined with HCHS/SOL)
Hispanic Community Health Study/Study of Latinos	HCHS/SOL	cohort study (controls only)	2,285	Illumina Omni2.5M array	dbGap: phs000810.v2.p2.c2, phs000810.v2.p2.c1	Replication cohort (combined with cases from Texas)

Information includes title, design, sample size, genotype array, dbGaP accession number or reference, and how each study was used in our analysis.

All analyses were limited to autosomes. Within each study, SNPs and individuals were removed based on the following pre-defined quality control thresholds: SNP call rate < 0.98, minor-allele frequency < 0.01, Hardy-Weinberg Equilibrium p value < 1 × 10⁻⁵, high relatedness ($\hat{\pi }>0.2$), genome heterozygosity rate > 6 standard deviations from the mean, or sample call rate < 0.95. Palindromic SNPs were also excluded as they could impact imputation accuracy due to unidentified strand flips or inversions between genome builds.³⁶ The number of individuals and SNPs remaining can be found in Table S1. Local ancestry was inferred using RFMIX, which was selected for its high accuracy and suitability for array data;³⁷ however, alternative algorithms, such as FLARE,³⁸ may also be suitable, particularly when dealing with sequencing data. Reference haplotypes were obtained from the genome aggregation database (gnomAD)³⁹ version 3, including 708 African, 671 non-Finnish European, and 94 IA individuals previously chosen for ancestry inference in Hispanics/Latinos.⁴⁰ After restricting to the intersection of SNPs across each paired dataset, each dataset was imputed separately using the TopMed Imputation Server.⁴¹ We observed systemic differences in imputation quality as measured by the imputation R² between cases and external controls, presumably due to batch effects in data generation or subtle differences in ancestry compositions between cohorts and differential imputation accuracies across populations.⁴² Consequently, for all comparisons of cases to controls, we imposed filters such that only SNPs with an imputation R² > 0.30 in both cases and controls and an R² difference <0.15 between cases and controls were included in the analysis.

Global ancestry analyses

Global ancestry for each study participant was calculated by averaging local ancestry estimates across the genome. We performed logistic regression of ALL case-control status on global IA-like ancestry in CCRLP and CCLS, studies in which indicators of SES variables were available. More important, both of these studies have study-ascertained matched internal controls to alleviate concerns of different ancestry proportions across externally ascertained datasets. Covariates included African-like ancestry, sex, and SES indicators (highest education obtained by a parent and annual household income). Income was categorized in 6 strata, in accordance with its treatment in CCLS⁴³: <$15,000, $15,000–$29,999, $30,000–$44,999, $45,000–$59,999, $60,000–$74,999, and >$75,000. In CCLS, the highest educational attainment of either parent was divided into 4 categories: less than elementary school, high school, some college, and a bachelor’s degree or higher. In CCRLP, the strata were less than 9^th grade, 9^th–12^th grade, and some college or more. Both SES indicators were treated as ordinal categorical variables, in which the lowest category was used as a reference. Linearity assumptions were assessed using residual and Q-Q plots. Analysis was carried out using R version 4.3.1.⁴⁴

Enrichment of risk allele on IA-like ancestry haplotypes

We performed local ancestry analyses at known ALL loci, based on the list provided in de Smith et al.¹¹. We excluded the T cell ALL-specific association locus at USP7⁴⁵ given that the increased risk of ALL in Hispanics/Latinos is limited to B cell ALL. We restricted analyses to remaining SNPs that were independent at the local ancestry level (e.g., choosing SNPs that were not in admixture linkage disequilibrium [LD]; N = 20) using PLINK’s clumping algorithm.⁴⁶ SNPs were clumped if the diploid local IA-like ancestry assignments covering the SNPs were 90% correlated and within 10 Mb of one another.

A total of 3 analyses were performed using local ancestry estimates. First, across known ALL loci,¹¹ we cross-tabulated local genetic ancestry (number of copies of the IA-like, African-like, or European-like haplotypes) and risk allele dosage to identify the association of risk alleles with ancestry. To establish a null distribution to which we would compare our observed ancestry-risk allele association, we randomly selected 1,000 sets of 20 frequency-matched alleles from across the genome, without replacement. Differences between the null and known loci distributions were evaluated via a chi-squared goodness of fit test. We restricted this analysis to CCRLP and CCLS controls to approximate the frequency of the risk allele in the general population. Second, we performed a logistic regression of ALL case-control status on IA-like ancestry dosage at each locus, including as covariates global African-like and European-like ancestry, and sex to evaluate whether the IA-like ancestry was associated with increasing ALL risk. Third, to test for the presence of heterogeneous effect sizes across ancestries, we performed logistic regression of ALL case-control status on risk allele genotype, IA-like ancestry dosage, sex, the first 20 principal components, and an interaction term of local ancestry and risk allele dosage. Interaction was modeled using Equation 1, where β_int(X_genoX_IA)represents interaction between local IA ancestry and genotype. Risk allele dosages were exported using PLINK version 2.0,⁴⁷ and interaction effect sizes were estimated using R version 4.3.1.⁴⁴

$$Logit\left(\mathrm{Pr}\left[Y_{ALL}\right]\right)={\beta }_{geno}{X}_{geno}+{\beta }_{IA}{X}_{IA}+{\beta }_{\mathit{int}}\left(X_{geno}{X}_{IA}\right)+{\beta }_{sex}{X}_{sex}+{{\beta }_{PCs}X}_{PCs}$$ (Equation 1)

Admixture mapping, regional association testing, and replication

We performed a genome-wide scan for loci in which IA-like ancestry was associated with ALL risk. We regressed ALL case-control status on the local IA ancestry dosage, global African-like and European-like ancestry, sex, and the first 20 principal components. We performed this analysis in CCRLP/Kaiser as our discovery dataset. The significance threshold for genome-wide admixture mapping was 5.7 × 10⁻⁵, as suggested in the literature previously for Hispanic/Latino populations.^48,49

We fine-mapped putative admixture association regions through single-variant association testing in the same dataset. The admixture peak region.²⁹ was defined as a range of positions that contained the lead SNP in the region and all SNPs with a p value within the same order of magnitude of that of the lead SNP. For example, if the lead SNP has a p value of 5 × 10⁻⁶, then all SNPs with a p value of at least 1 × 10⁻⁵ were considered. Genotypes of all imputed SNPs within this region passing QC filtering were tested via logistic regression, with sex and the first 20 principal components as covariates. The significance threshold for each region was computed using 1,000 rounds of permutation of case-control labels, choosing the threshold that controls for the 5% false discovery rate. We attempted to replicate the association of the lead variants from regional association testing and conditional analysis in additional Hispanic/Latino cohorts: CCLS, COG/MESA, and SOL-SD/REDIAL. The replication cohorts were meta-analyzed together by inverse variance weighting in the software METAL.⁵⁰

Power calculations

We assessed the power of our study to detect an interaction effect between risk alleles and genetic ancestry across known ALL loci, and the power to replicate putative associations identified through admixture mapping in additional Hispanic/Latino ALL cohorts. We used Quanto⁵¹ to carry out the power analysis. We assumed a disease prevalence of 0.0003, based on estimates from the Surveillance, Epidemiology, and End Results Program.⁵² For the local ancestry-genotype interaction power analysis across the 20 known ALL loci, we used the following parameters: sample size of 2,535 cases and 2,618 controls, risk allele frequency estimated in CCRLP for each of the known ALL risk SNPs, odds ratios for the effect of risk genotype, estimated local ancestry dosage, the standard deviation of local ancestry dosage (range: 0.72–0.76 across all known loci), and a Bonferroni-corrected significance threshold of 0.0025 (0.05/20). The odds ratios at which we had 80% power to detect were reported in Table S11. For the replication of regional associations on chr2, chr7, chr10, and chr15, the parameters included a sample size of 1,951 cases and 3,918 controls among the 3 replication cohorts, mean risk allele frequency estimated in the replication cohorts, odds ratio for the effect of genotype, and a permutation-based significance threshold based on the number of SNPs under the admixture mapping peak.

Evaluating the impact of using self-reported race/ethnicity labels versus genetically inferred ancestry

We opted to perform admixture mapping using all self-identified Hispanic/Latino individuals to be inclusive and avoid forming study exclusion criteria based on estimated genetic ancestry. However, alternative approaches could involve choosing to include participants based on inferred genetic ancestries. To examine, in our context, the impact of informing study inclusion based on genetically inferred ancestry (GIA), we implemented the software Harmonizing Ancestry and Race Ethnicity⁵³ (HARE) on genetic data from CCRLP/Kaiser-GERA. HARE is a 2-step machine learning-based algorithm that identifies individuals who may have discordant self-identified race/ethnicity (SIRE) and GIA information, given the mapping of the 2 learned from the rest of the dataset. The first step is a training phase in which GIAs, based on principal components, are used to predict available self-reported race/ethnicity in a support vector machine (SVM). The second step is HARE strata assignment, in which the highest probability of being assigned to a stratum is contrasted with the probability that the individual belongs to their self-reported stratum, using preset thresholds. If the ratio of the probabilities exceeds a pre-defined threshold, then an individual with existing SIRE could be assigned as missing (i.e., removed from analysis), while an individual without SIRE would be assigned into the top-predicted strata.

We applied HARE with the default parameters and thresholds,⁵³ training a multi-class SVM with a radial kernel on the first 30 principal components (calculated using PLINK⁴⁷ version 2.0) of genotype data from individuals across all self-reported race/ethnicity groups from CCRLP/Kaiser-GERA (10,450 Hispanic/Latino, 2,191 non-Latino African, 5,335 non-Latino East Asian, and 58,503 non-Latino European individuals). We first evaluated whether the HARE algorithm would exclude any study participants due to discordance between SIRE and GIA. We also assessed the rate of discordance between SIRE and GIA by performing 10-fold cross-validation, in which we randomly masked the SIRE information of 10% of study participants for each race/ethnicity group in each trial. This assessment simulates a scenario where one may define study exclusion criteria based on comparing the GIA to a known reference panel. Percent discordance was defined as the number of individuals classified as missing or as a stratum not corresponding to the masked SIRE label. We then investigated the impact of this study exclusion criteria on our admixture mapping result.

Institutional review board statement

This study was approved by the institutional review boards of the California Health and Human Services Agency, University of Southern California, COG, Baylor College of Medicine, Yale University, and University of California, Berkeley. All study participants were de-identified, and newborn bloodspots from the CCRLP were obtained via a consent waiver from the California Committee for the Protection of Human Subjects.

Results

The association between global IA-like ancestries, ALL risk, and SES indicators

Previous studies have identified a significant and positive association between global IA-like ancestries and ALL risk,²⁵ but they did not adjust for SES variables that often confound ancestry-disease associations. To quantify the effects of IA-like ancestry on ALL risk, we tested the association between ALL case-control status and global IA-like ancestry, while adjusting for African-like ancestry, sex, as well as covariates that are proxies for SES, such as highest parental education and annual household income. The global ancestry estimates of individuals across all cohorts are shown in Figure 1, with the average proportions of IA-like, African-like, and European-like ancestry being 37.7%, 6.8%, and 55.5%, respectively.

Figure 1.

Ancestry proportions of all cohorts

The ancestry proportions of CCRLP + Kaiser (top left) CCLS (top right), COG + MESA (bottom left), and ACCESS/REDIAL + HCHS/SOL (bottom right) are shown. Case-control status is denoted by CA (cases) or CO (controls). Proportion of Indigenous American-like, European-like, or African-like ancestries are represented by orange, blue, or red coloring, respectively. Global ancestry proportions were calculated by averaging local ancestry assignments, as inferred by RFMIX.³⁹

We performed these assessments for all self-reported Latino individuals in both CCRLP (excluding Kaiser-GERA controls; N = 4,033) and CCLS (N = 605), as both of these cohorts had available SES variables, contained internal controls, and thus were less likely to be confounded due to their study design. In CCLS, IA-like ancestry was significantly associated with ALL risk prior to including SES indicators in the model, as previously reported.²⁵ In the base model, adjusting for sex and African-like ancestry, a 20% increase in IA-like ancestry was associated with a 23% (odds ratio [OR]: 1.23, 95% confidence interval [CI]: 1.07–1.42, p = 0.003) increase in the odds of developing ALL. However, this association appears confounded, as inclusion of SES covariates strongly attenuated the strength of association (OR: 1.12, 95% CI: 0.96–1.31), and case/control status is no longer significantly associated with IA-like ancestry (p = 0.159; Table S2). In CCRLP, IA-like ancestry was not significantly associated with ALL risk in the base model adjusting for only African-like ancestry and sex, and it did not qualitatively change when SES covariates were included (Table S2). Regression analysis between the SES variables and ALL case-control status and ancestry, respectively, further supported the suggestion that sociocultural factors can act as confounders in ancestry-trait relations. For instance, we observed higher income to be associated with lower risk of disease (e.g., 0.30–0.76 times the odds of developing ALL; Tables S3 and S4), and higher European-like ancestry is associated with higher SES (Tables S5 and S6). Because SES measures can correlate with both ancestry and disease risk, the previously observed association between IA-like ancestry and ALL risk²⁵ may have been confounded.

Known ALL risk loci are enriched among IA-like ancestral haplotypes

While global ancestry association analyses can be confounded by socioeconomic indicators, use of local ancestry to identify associations at risk loci may be more robust. Furthermore, the absence of an association between IA-like ancestry and ALL risk globally would not necessarily preclude an association locally at known risk loci. Thus, we performed 3 analyses leveraging the pattern of local ancestry at known B cell ALL risk loci¹¹ (Table S7) to examine whether the risk alleles are found more frequently or show larger effects on IA-like ancestry haplotypes. First, we tabulated the proportion of times the risk allele at known ALL loci is associated with the African-like, IA-like, or European-like ancestry haplotypes, and compared these to the proportions found across 1,000 sets of SNPs randomly selected across the genome and matched by the risk allele frequencies. Across a total of 20 quasi-independent ALL SNPs (subjects and methods), the risk alleles were found on the African-like, IA-like, and European-like ancestry haplotypes 5.76%, 45.35%, and 48.88% of the time, respectively, in CCRLP. This significantly deviates from the genome-wide expectation (4.83%, 43.33%, and 51.85% for African-like, IA-like, and European-like ancestry haplotypes; p = 1.04 × 10⁻³⁶ by a chi-square goodness-of-fit test), with IA-like haplotypes showing a significant enrichment in association with the risk alleles. We also replicated this observation in CCLS, with IA-like haplotypes enriched with ALL risk alleles (p = 1.629 × 10⁻⁹). Consistent with this finding, the IA-like haplotypes have 1.32–1.33 (95% CI: 1.29–1.36 in CCRLP and 1.26–1.41 in CCLS) times the odds of carrying the risk allele compared with non-IA-like haplotypes (p = 4.098 × 10⁻⁸⁸ and 2.614 × 10⁻²³ in CCRLP and CCLS, respectively; Table S8).

In the second local ancestry analysis, we regressed ALL case-control status on the number of copies of IA-like ancestry segments locally at each known ALL locus while adjusting for sex and global Africa-like and European-like ancestry proportions. Because we anticipate that risk alleles for ALL are found disproportionately on IA-like haplotypes across these loci, we would expect that IA-like ancestries show an enrichment of association increasing the risk of ALL in the setting of admixture mapping studies. Indeed, we found a positive association of IA-like ancestry with ALL at 14 out of 20 known B cell ALL loci. At all loci in which IA-like ancestry was nominally significantly (p < 0.05) associated with ALL, IA-like ancestry was positively associated with risk (binomial test, p = 0.008; Table 2). To confirm that the known risk alleles were in fact driving the ancestry association, we included risk genotype as an additional covariate in each model. At each locus, upon adjustment for risk genotype, the association between IA-like ancestry and ALL risk reduced in significance (Table S9). Our results suggest that IA-like haplotypes are enriched for association with increased ALL risk at known loci, driven by the increased frequency of risk alleles.

Table 2.

Indigenous American-like ancestry is positively associated with risk at known ALL loci

Gene	Chr	Location	OR	L95	U95	pa
BCL11A	2	2p16.1	0.94	0.86	1.02	0.138
RPL6P5	2	2q22.3	1.01	0.93	1.10	0.751
C5orf56	5	5q31.1	1.07	0.98	1.17	0.113
BAK1	6	6p21.31	0.946	0.87	1.03	0.208
MYB/HBS1L	6	6q23	0.99	0.91	1.07	0.749
IKZF1	7	7p12.2	1.32	1.21	1.43	2.990 × 10−10
CCDC26	8	8q24.21	0.93	0.85	1.01	0.084
CDKN2A/2B	9	9p21.3	0.93	0.86	1.02	0.115
TLE1	9	9q21.31	1.04	0.96	1.13	0.358
GATA3	10	10p14	1.14	1.04	1.24	0.003
BMI1/PIP4K2A	10	10p12.2/10p12.31	1.16	1.07	1.27	5.422 × 10−4
ARID5B	10	10q21.2	1.29	1.19	1.41	4.924 × 10−9
JMJD1C	10	10q21.3	1.27	1.16	1.38	5.383 × 10−8
TET1	10	10q21.3	1.15	1.06	1.25	0.001
LHPP	10	10q26.13	0.97	0.89	1.06	0.508
ELK3	12	12q23.1	1.10	1.01	1.19	0.038
CEBPE	14	14q11.2	1.10	1.01	1.20	0.022
IKZF3	17	17q21.1	1.03	0.95	1.11	0.515
IGF2BP1	17	17q21.32	1.02	0.93	1.11	0.728
ERG	21	21q22.2	1.02	0.93	1.11	0.711

Loci that reached nominal significance (p < 0.05) are in boldface type. The analysis was restricted to CCRLP and CCLS cohorts.

^aALL case-control status was regressed on local Indigenous American ancestry dosage, African-like and European-like global ancestries, and sex.

In addition to ALL risk alleles being found more prevalently on IA-like haplotypes, risk alleles may exert larger (or smaller) effects on ALL when they are found on the IA-like background. Therefore, for the third local ancestry analysis, we tested for effect size heterogeneity by including a SNP × ancestry interaction term in the standard association testing model. After Bonferroni correction, no significant interactions between genotype and local IA-like ancestry were observed in either CCRLP or CCLS or when the 2 cohorts were combined (Table S10). Across all loci, we had greater than 80% power⁵¹ to detect a mean interaction effect of OR = 1.40 (range = 1.24–2.89; Table S11). Therefore, we found no evidence that ALL risk alleles exert effects with ORs greater than ∼1.40 on IA haplotypes, compared with other haplotypic backgrounds. We additionally found no evidence of interaction between global IA ancestry and genotype (Table S12).

Admixture mapping proposes putative novel ALL loci associated with IA-like ancestry

Although results from our global ancestry analysis showed minimal association between IA-like ancestry and ALL risk after adjusting for SES indicators, this does not preclude the possibility that some risk alleles show drastic differences in frequency between ancestries and thus are amenable to being discovered through admixture mapping, particularly since we observed that IA-like ancestries are enriched for association with risk alleles for ALL (Table 2). We thus performed admixture mapping using the CCRLP and Kaiser-GERA Hispanic/Latino cohort as our discovery panel (1,930 cases and 6,417 controls). Our discovery admixture mapping analyses identified genome-wide significant (threshold = 5.7 × 10⁻⁵; see subjects and methods) signals at 2q21.2, 7p12.2, 10q21.2, and 15q22.31 (Figure S3). The loci on chr7 (IKZF1) and chr10 (ARID5B) are known, while those on chr2 and chr15 are potentially novel (Figure 2).

Figure 2.

Manhattan plot of the admixture mapping analysis

The analysis identified signals at 2q21.2, 7p12.2, 10q21.2, and 15q22.31. The significance threshold is 5.7 x 10−⁵.

We then fine-mapped each of the 4 admixture loci via single-variant regional association testing (subjects and methods), which revealed the putative lead SNPs in each region to be 2:133605881 (p = 4.209 × 10⁻⁵), 7:50395237 (p = 2.262 × 10⁻¹²), 10:61959980 (p = 2.245 × 10⁻⁴⁴), and 15:66111790 (p = 3.011 × 10⁻⁴; Table 3). SNPs on chr2, chr7, and chr10 surpassed the regional significance threshold after accounting for the number of tests performed in each region via permutation (subjects and methods), and the chr15 SNP was slightly below the threshold. These SNPs exhibit allele frequency differences by ancestry group (Table 3), and conditioning on these SNPs also generally attenuated the admixture mapping association (Table 3). We sought replication of these associations in additional Hispanic/Latino cohorts, including cases from CCLS, COG, and REDIAL (see subjects and methods), totaling 1,951 cases and 3,918 controls. In the meta-analyzed replication cohorts, only the lead SNPs or other SNPs in LD with the lead SNPs on chr7 and chr10 were statistically significant (Tables S13 and S14). At these 2 known loci, conditioning on previously known SNPs (7:50409515, 7:50295636, 7:50384350, and 10:61961417) also eliminated the association signals at 7:50395237 and 10:61959980 (Table S15), suggesting that the association observed in admixture mapping are likely driven by the previously known allelic associations or their correlated proxies. Through power calculations⁵¹ we found that given observed ORs in the discovery cohort and combined allele frequencies in the replication datasets, our replication analysis had 65% and 49% power to detect the associations on chr2 and chr15, respectively; in contrast, we had 100% power to identify both the signals on chr7 and chr10.

Table 3.

ORs, 95% CIs, and p values of top SNPs from regional association testing, under the admixture peak, and the crude and adjusted p values of the admixture mapping tag SNP

Chr	Pos (Hg38)	rsID	Locus	Reference	Risk	ORTopSNP (95%CI)	PTopSNP	RAFEUR	RAFAFR	RAFIA	Gene	Padmixture	Padj	Peak region
2	134116620	rs117345013	2q21.2	T	C	1.22 (1.11–1.72)	4.209 × 10−5	0.010	0.000	0.390	unknown	4.156 × 10−6	0.005	133605881– 134236147
7	50390453	rs6969942	7p12.2	C	T	1.35 (1.25–1.47)	2.262 × 10−12	0.507	0.773	0.870	IKZFI	7.315 × 10−8	0.002	50352094–50457722
10	61959980	rs4948492	10q21.2	T	C	1.85 (1.59–2.13)	2.245 × 10−44	0.295	0.162	0.441	ARID5B	3.210 × 10−5	0.202	61928110–62306779
15	65877726	rs12438510	15q22.31	G	A	1.23 (1.10–1.39)	3.011 × 10−4	0.000	0.000	0.340	RAB11A	3.916 × 10−5	6.718 × 10−5	65540837–66146213

The regional significance thresholds, determined via permutation to account for the regional multiple testing burden, are 7.906 × 10⁻⁵, 6.256 × 10⁻⁴, 1.905 × 10⁻³, and 1.270 × 10⁻⁴ for 2q21.2, 7p12.2, 10q21.2, and 15q22.31, respectively. P_admixture denotes the p values from testing the local ancestry at each SNP position for association with ALL in admixture mapping, and P_adj denotes the same association but with the genotype dosage of each SNP added as covariate. Risk allele frequency (RAF) per ancestry group was obtained from European (Orcadian, Russian, Tuscan, French, Basque, Bergamo, Sardinian, and Adygei), African (Bantu, Mbuti, Biaka, San, Yoruba, and Mandenka), and Indigenous American (Surui, Karitiana, Maya, Pima, and Colombian) superpopulation data from the Human Genome Diversity Project (accessed from version 4 of gnomAD³⁹).

While local ancestry is not likely to be associated with sociocultural variables, such analyses can still be confounded by population stratification.⁵⁴ In our study, we opted to include all self-reported Hispanic/Latino individuals in our admixture mapping analysis and adjusted for potential confounding by including global ancestries as covariates in our admixture mapping model. Nevertheless, the use of GIA is pervasive and often heralded as a convenient, albeit arbitrary, way of identifying racial/ethnic strata to better control for confounding by population structure.⁵³ To evaluate whether our use of self-reported data may lead to starkly different results from using GIA, we employed the HARE algorithm to identify individuals whose self-reported race/ethnicity label may not be consistent with GIA compared to a reference. Consistent with the goals of our study and best practices in using population labels in human genetics studies,³⁰ HARE does not attempt to re-assign individuals to strata; rather, individuals with discordant self-report and genetically inferred data are indicated and thus can be excluded from analysis.

Using all available self-reported race/ethnicity labels in the multi-ethnic CCRLP + Kaiser cohort, we observed very low rates of discordance between HARE-predicted labels and the self-reported labels (percent discordance = 0.47% overall, and 0.68%, 3.37%, 0.18%, and 0.54% for African, East Asian, European, and Latino study participants, respectively; Table S16), suggesting that in general there is not enough genetic evidence to exclude participants from analysis (nor do we advocate for removing individuals from analysis on the basis of genetic makeup alone). In a scenario where individuals might be missing self-reported race/ethnicity labels, HARE can be used to predict the label based on a reference. We simulated such a scenario using 10-fold cross-validations, randomly masking in turn 10% of the cohort and using the remaining 90% as the reference for the HARE algorithm to learn the relationship between self-reported labels and GIA (subjects and methods). In this case, we observed the algorithm to be less confident, recalling 61.5% of Latino individuals, compared to 77.5%, 82.6%, and 86.9% of non-Latino White, African, and East Asian individuals, respectively. Of those not recalled back as Latino, 63.5% were assigned as missing, 35.3% were assigned as non-Latino White, 0.4% were assigned as non-Latino East Asian, and 0.8% were assigned as non-Latino Black (Figure S4; Table S17). Because this was a sensitivity analysis, we restricted our admixture mapping analysis to the set of Hispanic/Latino individuals inferred by HARE across 10 cross-validation trials, which produced a panel of 1,729 cases and 4,696 controls (Table S18). Overall, we observed high correlation in ORs (Pearson’s correlation coefficient, r = 0. 913) and log p values (r = 0.853), with a modest decrease in significance among variants with p < 1 × 10⁻³, including the 4 putative loci (Figures S5 and S6; Table S18), likely reflecting the smaller sample size.

Discussion

Motivated by the increased risk of ALL in Hispanic/Latino children, we carried out a systematic investigation into the impact of IA-like ancestry on childhood ALL risk, at both the local and global levels. We found that SES indicators confound the relationship between global IA-like ancestry and ALL risk, and adjustment for these variables attenuates the association. The lack of global ancestry association, however, did not preclude an association at the local level, as IA-like ancestry was significantly associated with ALL risk across known loci and risk alleles segregated with the IA haplotype more often than expected. However, we found no interaction between IA-like ancestry and the risk genotype at either the local or global levels, suggesting that there is little evidence that the risk alleles exert different effects in different ancestry contexts. These results suggest that in addition to environmental factors, risk alleles present in higher frequencies in IA populations may contribute to Hispanic/Latino individuals’ excess risk of B cell ALL.

Furthermore, our in-depth investigation into inferred genetic ancestry adds to and clarifies previous findings. A previous study found a positive, statistically significant association between global IA-like ancestry and ALL risk.²⁵ This result may have been affected by selection bias, due to the lower participation of controls than cases in lower SES areas.⁵⁵ Associations between global ancestries and disease risks are often reported, but these associations are often attenuated once non-genetic factors are included in the model or when a family-based design is used.^56,57 Indeed, observed associations between ancestry and type 2 diabetes risk in Hispanic/Latinos²⁶ and between ancestry and hypertension and lung cancer in African Americans^58,59 were attenuated once socioeconomic covariates were accounted for. In our analysis of the CCLS dataset, we found that the association of ALL risk with global IA-like ancestry was attenuated and no longer statistically significant (although trending in the positive direction) after adjusting for income and education level. We acknowledge that disentangling the association between global genetic ancestry and ALL risk is complex, especially given the observed correlations between SES indicators, ALL risk, and genetic ancestry reported in this study and elsewhere.^60,61 In the CCRLP, however, in which controls were ascertained through statewide birth records³² and thus should not be impacted by selection bias, we found no significant association between genetic ancestry and ALL risk.

Our admixture mapping identified associations on chr2, chr7, chr10, and chr15, but SNPs at the putatively novel loci on chr2 and chr15 did not replicate in additional Hispanic/Latino cohorts. The chr2 locus (2q21.2) is a relative gene desert but overlaps MGAT5, which has been reported to be involved in cancer metastasis and tumor suppression in mice.⁶² Furthermore, transcriptional regulation of the gene is driven by the Ras-Raf-Ets pathway, which is oncogenic.^63,64 On chr15, the gene closest to the putative risk SNP is RAB11A, a member of the RAS oncogene family, which has been associated with the risk of lung squamous cell carcinoma⁶⁵ and ovarian cancer.⁶⁶ In addition, this region contains multiple SNP associations with blood cell phenotypes,^67,68 suggesting a possible connection to the etiology of childhood ALL.¹⁵ Adjusting for the lead SNP did not strongly attenuate the chr15 association identified by admixture mapping, suggesting that more complex genomic variation such as structural variants, which may be in low LD with imputed SNPs,⁶⁹ could underlie this association and would require further validation. While there may be biological plausibility for these potentially novel loci, the lack of replication could also be due to lack of statistical power, heterogeneity of Hispanic/Latino populations across the United States, or that our initial findings were statistical artifacts.

Within the genetic epidemiology literature, careful consideration has been given to the use of self-reported race/ethnicity or GIA as strata in performing GWASs. For instance, it is common to define “White British” based on demographic data complemented by principal components of the genotype data when analyzing UK Biobank data to remove concerns of stratification.^70,71,72 However, analysis in a relatively homogenous cohort such as the UK Biobank is not completely free from the effect of stratification,⁷³ and the notion of defining a homogenous subset of the cohort becomes even more difficult when performing analysis in multi-ancestry cohorts or admixed populations. In those cases, one can pool samples across all ancestries for a “mega-analysis,”⁷⁴ or perform stratified analysis based on either (1) self-reported demographic labels^13,32,75 or (2) some arbitrary heuristics using GIA.^14,18,76,77 The downside of using GIA is the prescription of arbitrary thresholds to define ancestry strata. Across studies, the inclusion criteria vary, with individuals of self-reported White, Black, and Asian race/ethnicity being defined as study participants with over 75%–95% European-like, 50%–90% African-like, and 75%–90% Asian-like ancestries, respectively.^{14,18,76,77,78,79}

Previous genetic studies involving Hispanic/Latino participants have defined their strata as individuals whose IA ancestry was greater than or equal to 10% and greater than the proportion of African ancestry.^77,80,81 Given the heterogeneity of Hispanics/Latinos as a population, this method of categorization may prove untenable. The admixture proportions of Hispanic/Latino individuals vary by country of origin with Caribbean and mainland Hispanics/Latinos exhibiting the highest levels of African-like and IA-like ancestry, respectively.^82,83 Given these nuances, we have tended to favor relying on self-reported demographic labels as strata when performing multi-ancestry GWAS^13,32 or when limiting our analyses to Hispanics/Latinos.¹² Moreover, while recognizing that the HARE algorithm may be perpetuating the practice of imposing discrete labels to an inherently heterogeneous and continuous distribution of ancestries, the results of our HARE analysis showed that the algorithm rarely has enough evidence to exclude an individual from analysis when self-reported labels are present, and imputing missing labels have a relatively low recall rate for a heterogeneous population like the Latinos (Tables S16 and S17). Our observations thus suggest that in cases where self-reported labels are missing at high proportions, it may be preferable to conduct a mega-analysis instead of a stratified analysis. Indeed, even if overt categorizations impact little of the statistical power of the study design and may provide a bit more robustness to the study, the underlying message conveyed by explicit categorization of people based on estimated genetic ancestry may cause more harm than benefits.

In addition to operating with a discrete view of genetic ancestry, admixture mapping necessitated estimating admixture proportions and local ancestry segments based on the reference samples available. These notions of references are subject to fleeting geographic and temporal definitions of an ancestry component one wishes to estimate. While we utilized publicly available resources to define African-like, European-like, and IA-like ancestries (subjects and methods), the estimation of local ancestry is also complicated by the presence of fine-scale differences within a continental-level ancestry. The 3 major ancestral components of modern Latin American populations—IA, southern European, and west African—are the result of ancient admixture events themselves, with many southern European individuals drawing ancestry from western European hunter-gatherer and Near East sources^84,85,86 and west African populations originating from gene flow between western, central western, and south Niger-Congo peoples.⁸⁷ These ancestral populations may also exhibit geographic structures that are aggregated together (and thus overlooked). For instance, the IA-like ancestry component may be inherited from various previously isolated people-groups within Meso- and South America and the Caribbean, such as the Seri and Maya,⁸⁸ and European-like ancestry in Hispanic/Latinos may be inherited from European populations from Spain or the Mediterranean region.⁸⁹ As such, our conclusions in regard to IA-like ancestry need to be made in light of the available IA references; future investigation taking into account finer-scale sub-continental-level ancestries⁹⁰ may further complement the observations we present here.

Finally, the results of our study have implications for future investigations of the genetic architecture of ALL and designs of ancestry-based analyses. As reflected in our findings, the elevated ALL risk in Hispanic/Latino children is to some extent driven by the higher frequencies of known ALL risk alleles on IA-like haplotypes, which may underlie the shift toward higher mean polygenic risk scores for ALL in Hispanic/Latino individuals compared with in non-Latino White individuals.^13,91 Moreover, these consistent directional associations with risk could be indicative of a signature of past adaptation. There is some evidence that ALL risk loci, such as the IKZF1 locus that we found to be genome-wide significant in admixture mapping, may be under positive selection in Hispanic/Latino populations.¹² Given the overall enrichment of risk alleles on IA-like haplotypes, the selection may be more oligogenic or polygenic, although careful scrutiny is needed to rule out other potential confounding.^70,92 Irrespective of whether selective pressures are acting on ALL risk loci, the increased prevalence of risk alleles on IA-like haplotypes suggests that continued discovery in larger cohorts with IA-like ancestries and/or utilizing mapping approaches that incorporate signals from both ancestry and allelic associations may be fruitful in discovering more ALL risk loci and in understanding the genetic contributions to risk disparities.

Data and code availability

Genetic data from CCRLP and CCLS were obtained from the California Biobank. Because the samples and data are property of the state of California, researchers who would like to request to use the data should establish a protocol with the California Department of Public Health Committee for the Protection of Human Subjects and other relevant agencies to understand whether approval to share data are acceptable. The state has provided guidance on data sharing, as noted in the following statement: “California has determined that researchers requesting the use of California Biobank biospecimens for their studies will need to seek an exemption from NIH or other granting or funder requirements regarding the uploading of study results into an external bank or repository (including into the NIH dbGaP or other bank or repository). This applies to any uploading of genomic data and/or sharing of these biospecimens or individual data derived from these biospecimens. Such activities have been determined to violate the statutory scheme at California Health and Safety Code Section 124980 (j), 124991 (b), (g), (h) and 103850 (a) and (d), which protect the confidential nature of biospecimens and individual data derived from biospecimens. Investigators may agree to share aggregate data on SNP frequency and their associated p values with other investigators and may upload such frequencies into repositories including the NIH dbGaP repository providing: a) the denominator from which the data is derived includes no fewer than 20,000 individuals; b) no cell count is for <5 individuals; and c) no correlations or linkage probabilities between SNPs are provided.” Data that we derived from CCRLP and CCLS comprise less than 20,000 subjects from each study and therefore cannot be uploaded to a repository such as dbGaP. Data from COG and REDIAL were obtained by special request (Yang et al., 2011³⁴ and Harris et al.³⁵). Data from COG (phs001350.v1.p1), MESA (phs000209.v13.p3) and HCHS/SOL (phs000810.v2.p2.c1, phs000810.v2.p2.c2) were obtained from dbGaP. No new software was developed for this paper.

Acknowledgments

We would like to thank the families and patients who participated in each study for ALL, and Soyoung Jeon for helpful conversations contributing to this research. Data used in this study were generated with support from multiple grants from the National Institutes of Health (R01CA155461, R01CA175737, R01ES009137, P42ES004705, P50ES018172, P42ES0470518, R24ES028524, P20CA262733, U24ES028524, and R01CA262263), the Environmental Protection Agency (RD83615901), the Cancer Prevention and Research Institute of Texas (RP160771, RP210064), and St. Baldrick's Foundation (consortium grant 522277 with support from the Micaela's Army Foundation). Analysis performed in this study was supported by F31CA278359 to J.L., R01CA262263 to A.J.D., and R35GM142783 to C.W.K.C. A.J.D. is a Scholar of Blood Cancer United (formerly the Leukemia & Lymphoma Society). The content and conclusions from this work do not reflect the official views of the sponsors. Computation of this work was supported by USC's Center for Advanced Research Computing (https://www.carc.usc.edu/).

The full list of the institutions that supported this work is available in the supplemental acknowledgments.

Declaration of interests

The authors declare no competing interests.