The study protocol was approved by the Institutional Review Boards at the California Health and Human Services Agency, University of California (San Francisco and Berkeley), Yale University, and of all participating hospitals. Sample acquisition for the California-based Childhood Cancer Record Linkage Project (CCRLP) GWAS of ALL is described in detail in Wiemels et al. ⁹. In brief, newborn dried bloodspots (DBS) for cases and controls were obtained from the California Biobank Program, California Department of Public Health (CDPH), Genetic Disease Screening Program. Childhood ALL cases were identified through linkage between the CDPH statewide birth records (years 1982-2009) and the California Cancer Registry (CCR, diagnosis years 1988-2011), with controls randomly selected and matched on year and month of birth, sex, and race/ethnicity (non-Latino white, non-Latino black, Latino (any race), Asian/Pacific Islander, other). Additional controls were included from the Genetic Epidemiology Research on Aging (GERA) study ¹⁰. In this study, analyses were limited to Latino and non-Latino white subjects.

ALL subtype-stratified analyses were carried out in the California Childhood Leukemia Study (CCLS), an independent case-control study of childhood leukemia that is described in detail elsewhere ¹¹. In brief, patients with childhood leukemia under 15 years of age were identified and rapidly ascertained into the CCLS from California-based pediatric hospitals from 1995 to 2015. One or two controls were matched to each case on child’s date of birth, sex, Latino ethnicity, and maternal race as indicated on the birth certificate record. DNA samples for cases and controls were obtained from newborn DBS from the California Biobank Program Genetic Disease Screening Program, or from saliva or buccal swab samples. In the current study, we limited analyses to Latino and non-Latino white subjects with available genome-wide SNP array data (n=927 cases and 750 controls), and these were largely representative of the CCLS population as a whole.

DNA was extracted from newborn DBS and genotyped on genome-wide SNP arrays as previously described ⁹. Briefly, CCRLP ALL cases and controls were genotyped using Affymetrix Axiom World LAT arrays. Following quality control filtering, 757,935 polymorphic autosomal SNPs were included in analyses. Genotype data for additional GERA controls, genotyped on the same Affymetrix arrays, were downloaded from dbGAP (Study Accession: phs000788.v1.p2). In total, 3133 childhood ALL cases (1949 Latino and 1184 non-Latino white) and 12,135 healthy controls (8584 Latino and 3551 non-Latino white) were included in association analyses. Case demographic data are included in Table S1. Case-control analyses were stratified by ethnicity, with SNP associations calculated using logistic regression and adjusted for the first ten ancestry-informative principal components, calculated using Eigenstrat ¹², to control for population stratification. After adjusting for the first ten principal components, genomic inflation factors were calculated for the Latino and non-Latino white association analyses, which revealed minimal inflation of test statistics (λ_Latinos = 1.034; λ_{Non-Latino whites} = 1.004). A fixed-effects meta-analysis was used to combine results from the ethnicity stratified analyses.

Imputation was carried out across the chromosome 10p12 locus encompassing the two association peaks upstream of BMI1 (“BMI1 peak”) and at PIP4K2A (“PIP4K2A peak”). BMI1 is centered between the two association peaks, therefore we took coordinates 500Kb upstream and downstream of this gene. Imputation of this 1Mb region was performed using the Impute2 v2.3.1 software and its standard Markov chain Monte Carlo algorithm, with default settings for targeted imputation ¹³ and using 1,000 Genomes Phase 3 haplotypes for the imputation reference panel ¹⁴. Poorly imputed SNPs were removed, i.e. those with imputation quality (info) scores <0.60 or posterior probabilities <0.90. Association statistics for imputed and directly-genotyped SNPs were calculated using logistic regression in SNPTESTv2, using an allelic additive model and probabilistic genotype dosages ¹⁵. The effect of individual SNPs on ALL risk was calculated while adjusting for the first 10 principal components from Eigenstrat. Analyses were carried out separately in Latino cases and controls and in non-Latino white subjects. Meta-analysis was carried out using the program META ¹⁶.

Following identification of two physically separate association peaks at chromosome 10p12.31 and 10p12.2, we investigated whether they were independently associations with ALL or whether this was due to linkage disequilibrium (LD) between SNPs in the two peaks. Thus, association analysis across 10p12 was repeated using the same logistic regression model adjusted for 10 principal components and also adjusting for additive effects of the lead PIP4K2A peak SNP rs10741006, and then again adjusting for the lead BMI1 peak SNP rs12769953.

LD structure across 10p12 was assessed separately in Latinos and non-Latino whites using Haploview v4.2. Haplotypes were constructed using all SNPs with minor allele frequencies >0.05 among control subjects, and blocks were plotted using the default block definition of Gabriel, et al. ¹⁷.

Pairwise interaction analyses were carried out between three BMI1 peak SNPs and three PIP4K2A peak SNPs for a total of 9 interaction analyses. The three BMI1 peak SNPs included the most significant SNP in Latinos (rs12769953), the most significant SNP in non-Latino whites (rs4397732), and the candidate functional SNP rs11591377 that is in high LD with the former two SNPs (r²=0.93 and 0.97 respectively). The three PIP4K2A peak SNPs included the most significant SNP in Latinos (rs10741006), the most significant SNP in non-Latino whites (rs7912551), and the candidate functional SNP rs4748812 that is in LD with the former two SNPs (r²=0.82 and 0.80 respectively). Each interaction model contained the main effects for each SNP and the interaction term between each pair of SNPs, as well as the first 10 principal components. Analyses were performed separately in Latino subjects and non-Latino white subjects. P-values are reported for the interaction term of each logistic regression model performed using the glm command in the R statistical environment.

To identify putatively causal variants in the BMI1 peak and the PIP4K2A peak, at each locus we sought SNPs that demonstrated strong LD with both the lead SNP in Latinos and the lead SNP in non-Latino whites in the corresponding reference populations, i.e. in Admixed Americans (AMR) and Europeans, respectively. Thus, for the BMI1 peak we looked for SNPs in strong LD with rs12769953 in AMR and with rs4397732 in European populations, and for the PIP4K2A peak we sought SNPs in strong LD with rs10741006 in AMR and rs7912551 in Europeans. The LDlink “LDproxy” tool, a publicly available application for SNP LD analysis in 1000Genomes Project Phase 3 genotype data ¹⁸, was used to calculate LD values (r² and D’) for all SNPs within +/−500Kb of queried variants. For AMR populations, we included the “MXL” (Mexican ancestry from Los Angeles), “PUR” (Puerto Ricans from Puerto Rico), “CLM” (Colombians from Medellin, Colombia), and “PEL” (Peruvians from Lima, Peru) datasets. For European populations we included “CEU” (Utah residents from North and West Europe), “TSI” (Toscani in Italia), “FIN” (Finnish in Finland), “GBR” (British in England and Scotland), and “IBS” (Iberian population in Spain).

Scatter plots were generated for the BMI1 peak and PIP4K2A peak SNPs, plotting the r² of SNPs with the lead Latino SNP in AMR populations against r² of the same SNPs with the lead non-Latino white SNP in Europeans. Triallelic SNPs were excluded. We included only those SNPs with P < 5.0 × 10⁻⁶ in the adjusted meta-analyses results for both the BMI1 peak and PIP4K2A peak.

The Roadmap Epigenome Browser ¹⁹ was used to identify enhancer elements overlapping the 10p12.31 and 10p12.2 association peaks. We assessed tracks for DNase I hypersensitive sites (DHS), histone 3 lysine 4 monomethylation (H3K4me1), and H3K27 acetylation (H3K27ac) in HSCs (“Mobilized CD34 primary cells”), lymphoblastoid cell line (LCL) GM12878, and in the HSC-like myelogenous leukemia cell line K562. Predicted gene targets of enhancer elements were investigated using EnhancerAtlas (www.enhanceratlas.org), a recently developed database containing a consensus of human genomic enhancers derived from data on 76 cell lines and 29 different tissue types ²⁰. In brief, enhancers are predicted based on at least three independent experimental tracks (e.g. DNase hypersensitivity, transcription factor binding, and histone modification), and enhancer gene targets are predicted using the Integrated Method for Predicting Enhancer Targets (IM-PET) algorithm.

H3K27ac HiChIP is a recently developed protein-targeting chromatin conformation method for generating high-resolution contact maps between active enhancers and their target genes ²¹. H3K27ac HiChIP data was downloaded for GM12878 cells and for the K562 cell line (Gene Expression Omnibus, GEO dataset GSE101498), and uploaded into Juicebox software for visualization of 3D chromatin interactions across the chr10p12.31-p12.2 region. Knight-Ruiz (balanced) normalization was used to remove Hi-C matrix biases as recommended ²². Tracks for the same data were also uploaded into the Roadmap Epigenome Browser to assess locus interactions.

Methylation levels at CpGs across the BMI1 peak enhancer locus (overlapping rs11591377) and the PIP4K2A peak enhancer locus (overlapping rs4748812) were assessed using whole genome shotgun bisulfite sequencing (MethylC-Seq) data tracks in the Epigenome Browser ²³. MethylC-Seq data was assessed for HSCs (Mobilized CD34 primary cells) and a selection of different tissue types including fetal thymus, spleen, brain hippocampus, small intestine, liver, lung, gastric, adipose tissue, and ovary. At the rs11591377 (BMI1) enhancer, we assessed the proportion of methylated cytosine at 21 CpGs spanning a 2,215bp region, excluding analysis of CpGs with <10 sequenced reads in any tissue types. At the rs4748812 (PIP4K2A) enhancer, we assessed proportion of methylated cytosines at 7 CpGs spanning a 422bp region (again excluding CpGs with <10 reads).

UCSC Genome Browser was used to identify TFs found to bind to the BMI1 peak and PIP4K2A peak regulatory elements in ENCODE ChIP-seq analyses. To assess whether candidate SNPs altered any TF binding sites, we used the “TFBIND” software (http://tfbind.hgc.jp) ²⁴, inputting an 11bp sequence that included the SNP itself (either risk or protective allele) plus 5bp up- and 5bp downstream sequence.

For TFs predicted to bind to the BMI1 peak locus (at rs11591377), we downloaded available BAM files for corresponding ENCODE ChIP-seq experiments in K562 cells (from http://hgdownload.cse.ucsc.edu/goldenPath/hg19/encodeDCC/wgEncodeSydhTfbs/), including for ARID3A, BHLHE40, CEBPB, CTCF, c-Jun, c-Myc, p300, GATA2, JunD, MAFF, MAFK, MAX, TAL1, and ZNF143. BAM files for MYBL2 ChIP-seq experiments in K562 cells were downloaded separately (from https://www.encodeproject.org/experiments/ENCSR162IEM/), as these data were only released in August 2017. For each TF, separate BAM files were downloaded for each of two ChIP-seq biological replicates. BAM files were imported into the SNP & Variation Suite (SVS) software (Golden Helix), and visualized using the GenomeBrowse tool. For each TF, the number of sequenced reads was counted for the rs11591377 risk allele (G) and protective allele (A) in each biological replicate, and a binomial significance test used to calculate P values for the distribution of risk vs. protective allele reads compared with the expected distribution of 1:1. Fisher’s method was used for meta-analysis of P-values, using the R statistical package ‘metap’. To confirm that K562 cells are heterozygous for SNP rs11591377, we extracted DNA from an aliquot of cells (provided by Dr. Neil Shah, UCSF) using the Qiagen DNA Blood and Mini Kit, and genotyped using a Taqman assay for rs11591377 (Assay ID: C_____85910_10) on a Droplet Digital PCR machine (BioRad).

Cytogenetic subtype information was not available for CCRLP cases. Therefore, to assess subtype-specific associations of the lead chromosome 10p12 SNPs, we utilized data from the CCLS, an independent case-control study of childhood leukemia ¹¹. SNP genotype data was available for 927 ALL cases (589 Latino, 338 non-Latino white, Table S1) including 238 with high hyperdiploidy (HD) (156 Latino, 82 non-Latino white), as confirmed by FISH or G-banding ²⁵, and 750 controls (506 Latino, 244 non-Latino white). DNA was extracted from neonatal DBS, saliva, or buccal cells, and genotyped on Illumina HumanOmniExpress or HumanOmniExpressExome genome-wide SNP arrays ²⁶. SNP imputation was performed using Impute2, as described earlier for analysis in CCRLP. Multidimensional scaling (MDS) components were calculated using PLINK1.9 ²⁷ to control for population stratification. For this analysis, we included six SNPs in total: the lead Latino and non-Latino white SNPs and putative functional variants at both the BMI1 and PIP4K2A peaks (i.e. rs12769953, rs4397732, rs11591377, rs10741006, rs7912551, and rs4748812). Association of each SNP with ALL risk (overall and with HD-ALL subtype) was tested separately in Latinos and non-Latino whites using logistic regressions assuming an allelic additive model, adjusting for the first seven MDS components. A fixed-effects meta-analysis was used to combine results from the ethnicity stratified analyses, separately in overall ALL cases versus controls, in HD cases versus controls, and in non-HD cases versus controls. Cochran’s Q-test was used to test for heterogeneity between Latino and non-Latino white studies.

After adjusting for the lead PIP4K2A SNP (rs10741006), associations at the BMI1 peak remained genome-wide significant for several SNPs, including the lead Latino and lead non-Latino white SNPs (rs12769953 and rs4397732, respectively) (Table 1, Figure 1). Similarly, after adjusting for the lead SNP in the BMI1 peak (rs12769953), rs10741006 in the PIP4K2A peak retained genome-wide statistical significance (P = 2.21 × 10⁻¹⁶), thus confirming independent associations between these two loci and ALL risk (Table 1, Figure 1).

Epistasis analyses between the lead SNPs in the two association peaks revealed no evidence of synergistic or antagonistic interaction between SNPs (P>0.5 for all interactions tested in both Latinos and non-Latino whites).

Given the independence of association and lack of epistasis, we hypothesized that associated variants in the two peaks impart ALL risk via discrete mechanisms and possibly by affecting different gene targets. To explore this hypothesis, we capitalized on our multi-ethnic study design and that the lead SNP at both association peaks differed by ethnicity to investigate putative causal variants underlying the association at each peak. We predicted that underlying causal variant(s) would be in strong LD with lead tag SNPs in both Latino and non-Latino white populations.

At the BMI1 peak, there were 6 successfully imputed SNPs in tight LD (r² >0.90) with both the lead Latino SNP (rs12769953) in AMR reference populations and the lead non-Latino white SNP (rs4397732) in Europeans (Figure 2a). All 6 SNPs were strongly associated with ALL (P_meta < 5 × 10⁻¹⁰) after adjusting for the lead PIP4K2A SNP. Three SNPs – rs11591377, rs12773841, and rs11599410 – were in perfect LD (r² =1) in AMR and in perfect or near-perfect LD (r² >0.99) in Europeans (Table S3). None of the BMI1 peak SNPs were expression quantitative trait loci (eQTLs) for expression of BMI1 or any other genes in the Genotype-Tissue Expression (GTEx) project ²⁸ or in the Genetic European Variation in Health and Disease (GEUVADIS) project RNA sequencing data (in which no significant eQTLs were detected for BMI1).

We next explored the potential regulatory function of candidate SNPs. Both rs11591377 and rs12773841 overlapped a ~2kb region containing multiple transcription factor (TF) binding sites in ENCODE ²⁹ (Figure S2). A DNase I hypersensitivity (DHS) cluster was identified at this locus in 22 cell types in ENCODE, of which CD34⁺ Mobilized cells (i.e. hematopoietic progenitor cells) had by far the highest signal. In EnhancerAtlas ²⁰, this DHS cluster lies within a designated enhancer for BMI1 in 7 different tissues/cell lines, including: CD34⁺ cells, GM12878 LCL, fetal thymus, and the leukemia cell lines K562 and HL-60. SNP rs11591377 is positioned at the maximum height of the enhancer consensus score, whereas rs12773841 lies ~200bp upstream of the 5’ end of the enhancer locus (Figure S3). This locus is the only predicted enhancer for BMI1 located in the BMI1 peak region.

We then used the Roadmap Epigenome browser ¹⁹ to assess signals for DHS, which represents open chromatin regions, and for the histone modifications H3K4me1 and H3K27ac, which are known markers of enhancer elements. In support of the rs11591377 locus as an enhancer, we found strong peaks for DHS, H3K4me1, and H3K27ac in hematopoietic stem cells (HSCs) (Figure 2b) and in K562 cells, with smaller peaks in GM12878 LCLs. Analysis of H3K27ac HiChIP chromatin interaction data supported this locus as an enhancer for BMI1 in K562 cells, with a very strong signal for looping between the rs11591377 SNP locus (chr10:22,420,000-22,425,000) and BMI1 (chr10:22,605,000-22,610,000), with score = 203.96 (Figure 2b). This interaction was detected in GM12878 LCLs with a much lower score (score = 5.84), and no looping was detected in T-cells.

Methylation of cytosines at CpGs is frequently associated with transcriptional silencing at promoters and enhancer regions. Thus, we assessed methylation levels at CpG loci across the BMI1 enhancer region, using whole genome bisulfite sequencing data in different tissue types in the Epigenome Browser. This revealed variation across tissues, with HSCs almost entirely unmethylated across 9 CpGs in a region spanning 758bp (mean methylation = 0.56%) (Figure S4), whereas average methylation levels at other tissues ranged from 6.7% (ovary) to 65.7% (brain hippocampus) (Figure S5). Assessment of other BMI1 enhancers revealed these to be largely unmethylated across different tissues (data not shown), suggesting that the BMI1 enhancer spanning SNP rs11591377 is unique in its specificity to HSCs.

Several transcription factors (TFs), including known hematopoietic regulators, were predicted to bind to this enhancer locus and overlap rs11591377 (Table S4). Analysis of ENCODE ChIP-seq data revealed that the K562 cell line was heterozygous for SNP rs11591377, and revealed significant preferential binding of p300 (P=1.55×10⁻³) and JunD (P=2.61×10⁻³) to the ALL risk allele (G), with trends in the same direction for c-Myc (P=0.077), JunB (P=0.194), MAFK (P=0.069), SPI1 (P=0.134) and TAL1 (P=0.14) (Figures S6 and S7, Table S5). Additional TFs did not have available data in K562 cells or showed minimal binding at the rs11591377 enhancer. We assessed whether rs11591377 altered any TF binding motifs using TFBIND software ²⁴, and found the highest motif score (0.88) for the MYB transcription factor when including the risk allele G, and no predicted MYB binding for the non-risk allele A (Figure 3). MYB-like 2 (MYBL2), a MYB paralog with a similar binding motif sequence, showed a high level of ChIP-seq coverage at the enhancer locus and with highly significant preferential binding of the risk allele in K562 cells (P=1.73×10⁻⁵) (Figure 3). Genotyping of the rs11591377 SNP in genomic DNA from K562 cells using droplet digital PCR (ddPCR) confirmed a 1:1 ratio of allele G:A, supporting that the bias in the ChIP-seq data was not due to allelic copy number differences.

At the PIP4K2A peak, there were no variants with r²>0.90 but there were 4 SNPs with r²>0.75 with both the lead Latino SNP rs10741006 and the lead non-Latino white SNP rs7912551 in corresponding populations (Figure S8). All 4 SNPs – rs4747443, rs4748812, rs12146350, and rs746203 – were associated with ALL at P_meta<5×10⁻¹⁴ after adjusting for the lead BMI1 peak SNP (rs12769953), and were significant eQTLs for PIP4K2A expression in GTEx with ALL risk alleles associated with increased expression (Table S6).

We next sought to determine whether any of these SNPs overlapped putative regulatory regions in PIP4K2A, based on DHS, H3K4me1, and H3K27ac peaks in relevant tissues in the Epigenome Browser. Within a ~38Kb region encompassing the lead Latino SNP, the lead non-Latino white SNP, and the 4 SNPs listed above, only rs4748812 overlapped a region with strong evidence of regulatory function, with a large DHS peak found at chr10: 22,839,000-22,840,000 (Figure S9). SNP rs4748812 was also the only variant found to overlap a DHS cluster locus in the UCSC Genome Browser (Table S6).

H3K27ac HiChIP data analysis in the Epigenome Browser revealed evidence of chromatin looping between the locus containing rs4748812 (chr10:22,835,000-22,840,000) and another PIP4K2A regulatory region ~100kb downstream (chr10:22,935,000-22,940,000) in GM12878 cells, with score = 11.2 (Figure S9). HiChIP analysis in Juicebox also supported interaction between these loci in GM12878 cells (observed/expected score = 1.49), and even stronger interaction between the rs4748812 locus and the PIP4K2A promoter region at chr10:23,000,000-23,005,000 (O/E score = 2.22) (Figure S10).

Inspection of ENCODE ChIP-seq data revealed multiple TFs binding at the rs4748812 locus in LCLs and in K562 cells, including TAL1, CTCF, p300, and ARID3A. Analysis of MethylC-seq data across tissues revealed that demethylation at the rs4748812 locus is largely specific to HSCs (Figures S11 and S12). Investigation of potential TF binding motifs revealed that the rs4748812 risk allele T creates a RUNX1 (AML1) binding site (TFbind score = 0.911), which is not predicted with the protective allele C (Figure S13).

Association analyses of lead chromosome 10p12 SNPs in the CCLS revealed that BMI1 SNP rs11591377 was more strongly associated with high hyperdiploid (HD) ALL (OR = 1.56; 95% CI: 1.07-2.27) than with overall ALL risk (OR = 1.22; 95% CI: 1.01-1.49), with no association in non-HD ALL (OR = 1.06; 95% CI: 0.87-1.29) (Table S7, Figure S14). The HD-ALL association with rs11591377 appeared to be stronger in Latinos (OR = 1.71; 95% CI: 1.20-2.50), although the inter-ethnic study difference was not significant (P-value for heterogeneity, P_het = 0.58, Figure S14). The PIP4K2A putative causal SNP rs4748812 also appeared to be more strongly associated with HD-ALL than with overall ALL risk in Latinos. This pattern was not seen in non-Latino whites, although the difference between ethnic groups was again not significant (P_het = 0.32) (Table S7, Figure S14).