A replication study of novel fetal hemoglobin-associated genetic variants in sickle cell disease-only cohorts

Abstract

Sickle cell disease (SCD) is the most common monogenic disease in the world and is caused by mutations in the β-globin gene (HBB). Notably, SCD is characterized by extreme clinical heterogeneity. Inter-individual variation in fetal hemoglobin (HbF) levels strongly contributes to this patient-to-patient variability, with high HbF levels associated with decreased morbidity and mortality. Genetic association studies have identified and replicated HbF levels-associated variants at three loci: BCL11A, HBS1L-MYB, and HBB. In SCD patients, genetic variation at these three loci accounts for ~ 50% of HbF heritability. Genome-wide association studies (GWAS) in non-anemic and SCD patients of multiple ancestries have identified 20 new HbF-associated variants. However, these genetic associations have yet to be replicated in independent SCD cohorts. Here, we validated the association between HbF levels and variants at five of these new loci (ASB3, BACH2, PFAS, ZBTB7A, and KLF1) in up to 3740 SCD patients. By combining CRISPR inhibition and single-cell transcriptomics, we also showed that sequences near non-coding genetic variants at BACH2 (rs4707609) and KLF1 (rs2242514, rs10404876) can control the production of the β-globin genes in erythroid HUDEP-2 cells. Finally, we analyzed whole-exome sequence data from 1354 SCD patients but could not identify rare genetic variants of large effect on HbF levels. Together, our results confirm five new HbF-associated loci that can be functionally studied to develop new strategies to induce HbF expression in SCD patients.

Graphical Abstract

Graphical Abstract.

Introduction

Sickle cell disease (SCD) is the most common monogenic disorder in the world, affecting nearly 7.7 million individuals [1]. It is caused by a recessive mutation in the HBB gene (HbS) that encodes the β-globin subunit of adult hemoglobin (HbA). In hypoxic conditions, HbS polymerizes and forms long fibers that distort red blood cells (RBCs), giving them their hallmark crescent-like shape. While it is primarily a disorder of the blood, SCD affects multiple organs (e.g. brain, heart, lungs, bones, kidneys, eyes) and is characterized by a wide range of clinical manifestations, from very severe and life-threatening to almost asymptomatic [2]. Understanding the causes of this clinical heterogeneity is the key to predict, prevent, and treat SCD-related complications.

Fetal hemoglobin (HbF) is the main hemoglobin produced in humans in utero. In contrast to HbA, its β-globin subunits are encoded by the γ-globin genes HBG1 and HBG2 (which are not mutated in SCD). Just before birth, a molecular switch occurs in erythroid precursor cells, such that HbA production increases while HbF production decreases [3]. Interestingly, it was noted that newborn babies with SCD who still produce high HbF levels do not suffer from SCD-related complications [4]. That connection between high HbF and mild SCD was further supported by clinicians who noticed that SCD patients who constitutively express high HbF levels due to hereditary persistence of fetal hemoglobin (HPFH) mutations are also relatively protected from the main SCD complications [5]. With the advent of large prospective SCD cohorts, in particular the Cooperative Study of Sickle Cell Disease (CSSCD) in the United States, HbF was established as a major modifier of severity: SCD patients with high HbF levels suffer fewer complications (e.g. painful crises [6], acute chest syndrome [7], and stroke [8, 9]) and live longer [10]. HbF works by preventing the polymerization of HbS, the root cause of all SCD complications [11]. Hydroxyurea, one of only four drugs approved to treat SCD patients, works at least in part by inducing HbF production [12].

Human genetics has emerged as a powerful strategy to prioritize targets for drug development [13]. Previous genetic association studies in non-anemic participants and SCD patients have identified variants at three loci—BCL11A, HBS1L-MYB, and HBB—that are robustly associated with HbF levels [14–17]. One of these genes, BCL11A, is an approved gene therapy target to cure SCD [18, 19]. Genetic variation at BCL11A, HBS1L-MYB, and HBB explains ~ 50% of the heritable phenotypic variation in HbF levels in SCD patients [20]. Therefore, it is expected that larger genetic studies of HbF levels could yield new loci, and potentially new drug targets, for SCD. Four genome-wide association studies (GWAS) in non-anemic European-ancestry participants [21, 22], in Nigerian SCD patients [23], and in a cross-ancestry meta-analysis that included non-anemic and SCD individuals [24] have together identified 20 new variants associated with HbF levels.

Because these variants have not been systematically identified in SCD cohorts nor replicated in independent samples, we tested their association with HbF levels in a dataset totaling up to 3740 SCD patients. We also piloted a screening method that combines CRISPR inhibition (CRISPRi) with single-cell RNA-sequencing (scRNAseq) in erythroid cells to test the effect of the genomic regions where these new HbF variants map on the expression of the β-globin genes. Finally, we queried whole-exome DNA sequencing (WES) from 1354 SCD patients to identify rare coding variants associated with HbF levels.

Results

Genetic diversity among SCD participants

The SCD participants (only with HbSS or HbSβ⁰ genotypes) originated from seven studies based in the USA, Jamaica, France, and Tanzania (Table S1). To characterize the genetic diversity of these individuals, we combined their genotype information at 9 8176 common SNPs with data from the 1 000 Genomes Project and performed dimension reduction analyses using principal component (PC) and uniform manifold approximation and projection (UMAP) analyses. As expected, the PC analysis revealed that most SCD participants in our experiment are aligned along an African-European axis of genetic variation on PC1 (Fig. S1A). The UMAP representation, generated using the first five PCs, provided more resolution and allowed us to identify participants from the East-African Tanzania SCD cohort, who overlap with the Luhya in Webuye (Kenya) population from the 1000 Genomes Project (Fig. S1B and C). We were also able to identify a small number of SCD participants that cluster with South Asians or admixed Americans from the 1000 Genomes Project (Fig. S1C).

SCD-only meta-analysis for HbF levels

For each study, we corrected raw HbF levels for age, sex and β-globin genotype, and then applied inverse normal transformation (Methods). This approach minimizes the impact of between-study heterogeneity while maximizing our sample size for genetic discovery. To identify novel genetic regulators of HbF levels, we combined association results at 16.4 M autosomal and X-linked variants (minor allele frequency [MAF] ≥1%) from 3740 SCD participants, adding 1541 individuals from four cohorts to our previous HbF genetic study (Table S1). Because of the genetic heterogeneity described above, we used PCs as covariates and opted to analyze each study individually, before combining results using a fixed-effect meta-analysis method [25].

The non-conditional and conditional analyses (Methods) only identified the three known HbF loci at genome-wide significance (P < 5 × 10⁻⁸): BCL11A, HBS1L-MYB, and β-globin (Table 1, Fig. S2, and Tables S2 and S3). The recovery of these known positive controls validate our HbF normalization method and our genetic association framework. Recent genetic analyses in a non-anemic White British cohort (N = 11 004)(medRxiv preprint [22]), a multi-ancestry dataset that included healthy participants but also SCD patients (N = 28 279)(medRxiv preprint [24]), and a Nigerian SCD cohort (N = 1006) [23] identified 19 new variants associated with HbF levels (P < 5 × 10⁻⁸ or false discovery rate [FDR] < 5%). To this list, we added an HbF-associated variant at the NFIX locus reported in the Sardinian population (N = 6602) [21]. We attempted replication of these 20 variants in our SCD-only meta-analysis. When appropriate, we excluded the Tanzania SCD cohort from the meta-analysis, as it was also used in one of the discovery studies. While the sample size of our SCD-only meta-analysis is modest, we calculated that we have reasonable power (≥70%) to replicate 12 of the 20 HbF associations using a liberal one-tailed P-value < 0.05 (Table 1). For these well-powered loci, we could replicate the association between HbF levels and genetic variation at ASB3, BACH2, PFAS, ZBTB7A, and KLF1 (Table 1). We obtained similar results when meta-analyzing results using MR-MEGA, which allows for effect size heterogeneity across studies (Table S4) [26]. In the African-American CSSCD (N = 1279 participants), the six replicated variants at these five loci increased the HbF phenotypic variance from 26.6% (for the three known loci: BCL11A, HBS1L-MYB, and β-globin) to 27.3%, a small increment due to the small effect size of the new HbF variants (Methods). Many reasons may explain the lack of replication for many loci: (1) false positive associations, (2) inflated effect sizes in the discovery studies (i.e. Winner’s curse), (3) differences in allele frequencies, or (4) epistatic interactions with the SCD genotype for HbF variants initially discovered in non-anemic participants (Discussion).

Table 1.

Replication results in sickle cell disease (SCD) cohorts of newly identified variants associated with fetal hemoglobin (HbF) levels. For variants identified by Cato et al., we excluded the Tanzania SCD cohort because it was already used in the discovery phase. At all loci, we considered sentinel and linkage disequilibrium (LD) proxies (r² ≥ 0.8). We underlined LD proxies and provided the LD r² value with the sentinel variant. We calculated statistical power to replicate the original association for a one-tailed P-value <  0.05 using the original effect size, and the sample size and allele frequency in the SCD-only meta-analysis (this study). ^*for SLC28A3_rs115555854, the direction of the effect is inconsistent with the original report. We defined as replicated variants with a consistent direction of effect on HbF levels and a meta-analytic nominal P-value < 0.05. For each new locus, we provide the reference in which the initial association was detected. EA, effect allele; OA, other allele; EAF, effect allele frequency; BETA, effect size in standard deviation units; SE, standard error; P-value, two-tailed P-value; corrected P-value, Bonferroni correction for the 20 variants tested in the replication analysis; HetPVal, heterogeneity P-value of effect sizes across studies.

Locus	rsID	Position (hg38)	EA/OA	EAF	BETA (SE)	P-value	Corrected P-value	HetPVal	Sample size	Power
Known (already replicated) HbF-associated loci
BCL11A	rs1427407	2 (60490908)	T/G	0.25	0.604 (0.027)	2.4 × 10−111	4.8x10−110	0.076	3740	>0.99
HBS1L-MYB	rs9389269	6 (135106021)	T/C	0.94	−0.586 (0.051)	8.1 × 10−31	1.6x10−29	0.39	3740	>0.99
HBB	rs3759074	11 (5236548)	A/G	0.11	0.365 (0.045)	6.8 × 10−16	1.4x10−14	0.19	3740	>0.99
New (unreplicated) HbF-associated loci
ASB3	rs111607747	2 (53759561)	T/G	0.050	−0.129 (0.054)	0.017	0.34	0.30	3740	0.78
TMEM161B	rs1997323	5 (87924839)	A/G	0.052	0.082 (0.054)	0.12	1	0.63	3740	>0.99
BACH2	rs2325259	6 (90140445)	T/C	0.94	0.027 (0.058)	0.65	1	0.36	2527	0.47
BACH2	rs4707609	6 (90236760)	T/C	0.89	−0.050 (0.037)	0.18	1	0.63	3740	0.73
	rs62408218 (r2_EUR = 0.98)	6 (90222139)	T/C	0.13	0.078 (0.035)	0.026	0.52	0.0061	3740
GRIK2	rs190118557	6 (102999561)	T/C	0.0005	0.080 (0.701)	0.91	1	0.19	2211	0
SLC28A3	rs115555854	9 (84339933)	A/C	0.0091	0.245 (0.119)	0.039*	0.78	0.12	3740	>0.99
TICRR	rs140496989	15 (89619296)	A/G	0.035	0.005 (0.064)	0.94	1	0.38	3740	>0.99
ABCC1	rs246232	16 (16036667)	C/G	0.15	−0.040 (0.034)	0.24	1	0.025	3740	0.92
ABCC1	rs60782127	16 (16048222)	T/G	0.0024	0.418 (0.377)	0.27	1	0.36	2211	0.32
ABCC1	rs165975	16 (16040137)	T/C	0.56	−0.022 (0.023)	0.34	1	0.044	3740	0.93
PFAS	rs6503094	17 (8254955)	A/G	0.52	0.040 (0.023)	0.080	1	0.24	3740	>0.99
PFAS	rs62637606	17 (8269188)	T/G	0.99	0.176 (0.142)	0.22	1	0.13	3740	0.39
PFAS	rs9891699	17 (8253992)	T/C	0.47	−0.065 (0.027)	0.016	0.32	0.44	2527	0.85
PIEZO2	rs58817161	18 (10818373)	T/C	0.99	−0.115 (0.100)	0.25	1	0.74	3740	>0.99
ZBTB7A	rs7259699	19 (4072148)	A/G	0.30	0.044 (0.030)	0.15	1	0.61	2527	0.7
	rs10423017 (r2_EUR = 0.84)	19 (4059166)	T/G	0.38	0.087 (0.028)	0.0021	0.042	0.38	2527
DNASE2; GCDH; KLF1	rs4804210	19 (12879166)	A/G	0.56	−0.061 (0.028)	0.032	0.64	0.17	2527	0.76
DNASE2; GCDH; KLF1	rs3817621	19 (12887391)	C/G	0.21	0.057 (0.035)	0.10	1	0.90	2527	0.56
DNASE2; GCDH; KLF1	rs11085824	19 (12890733)	A/G	0.80	0.062 (0.034)	0.072	1	0.13	2527	0.64
	rs7508226 (r2_EUR = 0.91)	19 (12922053)	A/G	0.38	−0.083 (0.029)	0.0041	0.082	0.37	2527
FARSA; CALR	rs2965220	19 (12942311)	T/C	0.20	−0.052 (0.035)	0.14	1	0.17	2527	0.62
NFIX	rs183437571	19 (13011085)	T/C	0.0051	−0.081 (0.171)	0.63	1	0.43	3740	0.88

Testing HbF-associated SNPs using CRISPR inhibition and single-cell RNA-sequencing

The main challenge of GWAS is to connect robust genetic associations with causal variants and genes. We reasoned that HbF is an ideal human trait for variant-to-gene functional strategies, since it is largely cell-autonomous and often controlled at the gene expression level through the coordinated regulation of the γ (HBG1/2) and β (HBB) globin genes. In a pilot experiment, we designed a small library of 54 guide RNA (gRNA) to target dCas9^KRAB (to enable CRISPR inhibition [CRISPRi]) in erythroid HUDEP-2 cells near 14 SNPs associated with HbF levels (Fig. 1A and Table S5). The gRNA library included two safe target controls, a negative control (GYPA/B promoter), a positive control (BCL11A enhancer) [27], gRNA against each of the sentinel HbF variants at the ASB3, ARHGAP39 and BACH2 loci, gRNA against three variants near CECR2 on chromosome 22 that show strong evidence of association with HbF in our meta-analysis (P-value <6 × 10⁻⁵), and gRNA against each of eight variants in a region of high LD at the KLF1 locus (Table S6 and Fig. S3). After infection with the lentiviral gRNA library and HUDEP-2 differentiation, we collected single-cell RNA-sequencing (scRNAseq) data to measure the impact of each gRNA on the (HBG1 + HBG2)/HBB ratio (trans-effect, as a proxy for a potential effect on HbF levels) and the expression of nearby genes (<100 kb, cis-effect)(Fig. 1A).

Figure 1.

Identification of genomic sequences near two non-coding variants at the KLF1 locus that modulate the (HBG1 + HBG2)/HBB expression ratio in HUDEP-2 cells. (A) Schematic of the single-cell CRISPR inhibition (CRISPRi) protocol used to validate fetal hemoglobin (HbF)-associated variants. On the left, a transcription factor binds a motif and increases the expression of a HbF repressor (like BCL11A) in cis that leads to decrease HbF production in trans. On the right, the dCas9^KRAB complex interferes with the binding of the transcription factor, leading to less repressor (in cis) and more HbF expression (in trans). Created with BioRender.com. (B) CRISPRi near rs2242514 and rs10404876 (stars at the bottom of the figure) increases the (HBG1 + HBG2)/HBB ratio in HUDEP-2 cells. We annotated the locus with ATAC-sequencing peaks generated in human erythroid cells, ENCODE candidate cis-regulatory elements (cCREs) and predicted enhancer-to-gene (rE2G) links in the K562 leukemic cell line.

After quality-control steps, we obtained sufficient scRNAseq data for 40 gRNA (Table S6). We compared the (HBG1 + HBG2)/HBB ratio (or the expression of genes of interest located in cis) between cells that had received safe target gRNA or an HbF-specific gRNA using quasi-Poisson regression (Methods). Our controls confirmed the validity of this perturbation system: a gRNA targeting dCas9^KRAB to the promoters of the paralogous genes GYPA/B reduced their expression but had no effect on the (HBG1 + HBG2)/HBB ratio, whereas a gRNA that targeted the BCL11A erythroid enhancer previously shown to modulate the γ-to-β globin switch reduced BCL11A expression and increased the (HBG1 + HBG2)/HBB ratio as expected (Table 2).

Table 2.

Results for the single-cell CRISPRi experiment in HUDEP-2. We present results for the negative (GYPA/B) and positive (BCL11A) controls, as well as for the four significant gRNA that target variants located at fetal hemoglobin (HbF)-associated loci (P-value for (HBG1 + HBG2)/HBB < 0.0036). For each of these top four gRNA, we provide results for the CRISPRi effect on the genes located within 100-kb when available in the single-cell dataset (cis-effect). Results for all 40 gRNA and tested genes in cis are available in Table S5.

Targeted variant (CHR:POS_hg38)	Rationale for inclusion (SCD-only HbF GWAS meta-analysis P-value)	gRNA	(HBG1 + HBG2)/HBB ratio or target gene	quasi-Poisson coefficient	quasi-Poisson P-value
	Negative control	sg.GYPAprom	(HBG1 + HBG2)/HBB	0.312	0.26
			GYPB	−0.940	5.0 × 10−11
			GYPA	−1.031	2.1 × 10−17
	Positive control	sg.BCL11a	(HBG1 + HBG2)/HBB	1.997	2.0 × 10−5
			BCL11A	−0.757	1.1 × 10−6
rs4707609 (6:90236760)	BACH2 sentinel variant (P = 0.18)	gRNA_ld_6_90 236 760_b_10f	(HBG1 + HBG2)/HBB	1.976	0.00042
		gRNA_ld_6_90 236 760_d_2f	(HBG1 + HBG2)/HBB	1.316	0.0017
			BACH2	20.572	0.99
rs2242514 (19:12871984)	KLF1 locus (P = 0.0050)	gRNA_19_12 871 984_A_G_8f	(HBG1 + HBG2)/HBB	0.856	0.0034
			PRDX2	−0.044	0.37
			KLF1	−0.005	0.90
			JUNB	−0.070	0.53
			FARSA	−0.100	0.098
			RNASEH2A	−0.109	0.19
			GCDH	−0.084	0.49
			DNASE2	−0.194	0.27
			SYCE2	−0.061	0.81
			MAST1	0.130	0.80
			THSD8	0.822	0.14
rs10404876 (19:12876791)	KLF1 locus (P = 0.0054)	gRNA_19_12 876 791_T_C_17r	(HBG1 + HBG2)/HBB	1.028	0.0015
			PRDX2	−0.042	0.43
			KLF1	0.002	0.97
			JUNB	−0.127	0.27
			FARSA	−0.100	0.11
			RNASEH2A	−0.106	0.22
			GCDH	−0.049	0.69
			DNASE2	−0.184	0.28
			SYCE2	0.117	0.67
			THSD8	1.092	0.054
			MAST1	0.115	0.82

After correcting for the 14 tested variants, we found four hits in our single-cell CRISPRi experiment (P-value < 0.0036, Table 2). The first two hits involved two different gRNA that target rs4707609, the sentinel BACH2 variant (Table 2). This variant maps to an intron of BACH2, suggesting that the down-regulation of BACH2 expression through CRISPRi is associated with increased HbF levels, a result that is consistent with recent functional experiments. However, we were unable to detect a CRISPRi effect on genes located in cis to validate this hypothesis (Table 2). We queried various databases and could not find an overlap between the genomic position of this variant and an open chromatin site in human erythroid cells or a predicted ENCODE candidate cis-regulatory element (cCRE). The two remaining hits were gRNA that targeted two different SNPs (rs2242514 and rs10404876) at the KLF1 locus and that increased the (HBG1 + HBG2)/HBB ratio (Table 2 and Fig. S3). The two variants lie downstream of KLF1, in intronic sequences of MAST1 and DNASE2, respectively (Fig. 1B). While there was no relevant epigenomic annotation for rs10404876, we found that rs2242514 maps to an ENCODE cCRE, although that predicted enhancer was not linked to any genes through enhancer-to-gene predictions (no rE2G links) (Fig. 1B). Again, we were unable to determine if the effect was solely due to the down-regulation of KLF1, or a more pleiotropic effect on many genes located near these variants (e.g. DNASE2, GCDH, FARSA, CALR). Thus, while this single-cell CRISPRi approach is effective at capturing transcriptional effects on highly expressed genes (such as HBG1, HBG2 and HBB), the sparsity of the scRNA-seq data prevented robust analyses of differential expression for genes located in cis in erythroid cells (Discussion).

Whole-exome DNA sequencing (WES) in SCD participants

To determine if rare coding variants also modulate HbF levels, we analyzed available WES data from 1354 SCD participants (Methods). We focused on variants with a MAF ≤ 1% and performed gene-based testing using two methods (VT and SKAT) and three variant selection strategies (broad, strict, loss-of-function [LoF] as defined in Methods). Across these analyses, we found no genes that reached statistical significance after accounting for the number of genes tested (α = 3.1 × 10⁻⁶, Bonferroni correction for 15 913 genes)(Table S7 and Fig. S4 and S5). Even when we restricted our analyses to genes expressed in human erythroblasts or mature erythrocytes based on transcriptomic or proteomic experiments [28–30], we did not detect an enrichment of statistical signals (Table S7 and Fig. S6). This negative result suggested limited statistical power given our sample size.

Rare mutations that cause the hereditary persistence of fetal hemoglobin (HPFH) condition have been identified at the β-globin locus as well as in the KLF1 and DNMT1 genes [31–33]. We inspected rare (MAF ≤ 1%) coding variants in 65 genes that have been implicated in the γ- to β-globin gene switch during development (Fig. 2 and Table S8). We found 22 variants in 16 genes that are carried by SCD patients with mean HbF levels that are at least two standard deviations away from the mean HbF after correction for age, sex, and β-globin genotypes (Fig. 2 and Table S8). Several of these rare coding variants are associated with a HbF phenotype more extreme than expected when compared with a null distribution generated using rare synonymous variants in the same 65 genes (Fig. S7).

Figure 2.

Rare coding variants identified in 1354 sickle cell disease (SCD) patients by whole-exome sequencing in genes implicated in the γ-to-β globin switch. We only consider missense, nonsense, frameshift and essential splice site variants with a minor allele frequency ≤ 1% (corresponding to a minor allele count ≤ 32 [x-axis]). Average fetal hemoglobin (HbF) levels per variant are on the y-axis (in standard deviation [SD] units after correction for sex, age, and β-globin genotypes). For each variant found in SCD patients with available genome-wide genotyping data, we averaged the normalized HbF polygenic score (in standard deviation units) calculated using known HbF common variants at BCL11A, HBS1L-MYB and β-globin (grey indicates missing genotyping data as not all sequenced individuals were also genotyped).

Given the recent report of HPFH mutations in DNMT1 [33], we were particularly intrigued by the discovery of a novel missense variant in this gene (p.Gly95Ser) in a SCD participant with baseline HbF levels of 28.9% (2.36 standard deviations above the mean and outside of the 90% confidence interval generated with rare synonymous variants [Fig. S7]). Our review of the patient’s medical records indicated that the high HbF levels were stable (additional HbF values of 27.7% and 28.8% taken three years apart), that the patient was not treated with hydroxyurea and that the patient was largely clinically asymptomatic. Furthermore, the high HbF levels phenotype was not explained by the inheritance of common alleles at BCL11A, HBS1L-MYB and β-globin associated with HbF levels (the patient’s normalized HbF polygenic score was 0.86 standard deviation below the mean). These observations are consistent with the clinical benefits associated with the inheritance of a rare HPFH mutation.

To confirm these genetic and clinical observations, we attempted to model the p.Gly95Ser mutation in HUDEP-2 cells. Despite several attempts, we could not successfully install this precise mutation using CBE base editing. As an alternative, we over-expressed the DNMT1 open reading frame with either the Gly95 or 95Ser amino acid in HUDEP-2 cells that carry DNMT1 frameshift mutations introduced by CRISPR-Cas9 genome editing. While disrupting DNMT1 increased the HBG2/HBB ratio (as expected), we could not rescue this phenotype by over-expressing DNMT1-Gly95 (Fig. S8). Therefore, our functional experiment does not allow us to conclude that the DNMT1 p.Gly95Ser variant is responsible for the high HbF and mild disease phenotype noted in one of our SCD participants.

Discussion

HbF is the strongest non-environmental modifier of severity in SCD, and genetic investigation of its regulation has led to the development of therapies targeting the BCL11A transcription factor. Recent GWAS have identified new variants associated with HbF levels, but these discoveries were made in largely non-SCD cohorts. Independent replication of genetic associations remains critical, especially for phenotypes like HbF that are rarely measured in very large cohorts. We attempted to replicate the new HbF genetic associations in well-characterized SCD cohorts totaling up to 3740 patients. We could find evidence of association (meta-analysis P < 0.05) for six of the 12 variants for which we had sufficient statistical power (Table 1). It is possible that some of these published associations are false positives, or that their published HbF effect sizes are in fact smaller because of the Winner’s curse, such that our power calculations are overly optimistic. Finally, we acknowledge that our genotype imputation approach is not ideal to replicate the associations with the four variants in Table 1 with MAF < 1%.

Another possibility to explain this apparent lack of replication at several HbF loci is an epistatic interaction with the SCD genotype. To draw a parallel, we previously showed that the Duffy null allele, which is associated with benign neutropenia in African-ancestry individuals, only has a weak effect in SCD patients [34]. We hypothesized that this is due to the overall inflammatory state of SCD patients that can ‘mask’ the effect of large effect genetic variants on neutrophil production. Many of the new HbF loci have been identified in non-SCD participants. While they may contribute to HbF phenotypic variation in these populations, they may have no impact on HbF in SCD patients because of their disease-specific pathophysiology. A recent HbF GWAS that only included individuals of European ancestry highlighted variants near ABCC1, which encodes the multidrug transporter MRP1 [22]. Experiments in erythroid cells showed that disrupting MRP1 increases oxidative stress, induces HbF production through the activation of an anti-oxidative stress response that involves the transcription factor NRF2, and reduces sickling in vitro [22]. Given these compelling functional data, it is puzzling why we could not replicate the association between ABCC1 variants and HbF levels despite good statistical power (P-value > 0.2, Table 1). Cells from SCD patients are in a very high oxidative stress, hence the cellular (transcriptional) response to counteract this oxidative state is probably fully deployed already [35]. Thus, reducing MRP1 activity through genetic variants or small molecules may not have the same effect on HbF production in SCD patients as in cells in culture (or non-anemic individuals). This might also explain why several molecules known to trigger NRF2-dependent HbF production have not had successful outcomes in SCD clinical trials [36]. Therefore, while we encourage further work to validate MRP1 as a potential new HbF inducer for SCD patients, we caution that ABCC1 is currently not supported by human genetics in this patient population.

To support our genetic association results with functional data in human erythroid cells, we combined CRISPRi with scRNAseq to test the role of genomic sequences near HbF-associated SNPs on the (HBG1 + HBG2)/HBB ratio and the expression of genes in cis. We adapted existing protocols, designed appropriate positive and negative controls, and implemented a simple statistical model for this HbF-focused experiment. We could detect single-cell CRISPRi effects on the (HBG1 + HBG2)/HBB ratio at variants near the transcription factors BACH2 on chromosome 6 and KLF1 on chromosome 19 (Table 2). We designed the experiment to test multiple variants at the same time, but recognize its limitations. Our results should be further validated using orthogonal methods (e.g. base editing in in vitro differentiated primary human CD34⁺ cells) to confirm the functionality of these sequence variants and determine whether their effect is specifically on the regulation of the β-globin genes as opposed to a more general impact on cell survival/proliferation. It would also be important to validate experimentally that a transcriptional effect (in trans) on the (HBG1 + HBG2)/HBB ratio leads to changes in HbF protein levels. Another limitation of the single-cell CRISPRi experiment is that we did not measure the down-regulation of genes in cis to these gRNA-mediated perturbations. Many reasons could explain this observation: (1) statistical analyses might be under-powered given the number of cells and/or the magnitude of the CRISPRi effect, (2) genes might be expressed at levels too low to be accurately quantified by scRNAseq, or (3) mRNA and protein levels might be discrepant due to compensatory mechanisms. Our data are publicly available, and therefore we hope that they will be used by others to improve the statistical framework and guide the development of better single-cell CRISPRi experimental protocols.

While the focus of our study was on the replication of recently identified common HbF-associated variants, we did attempt to find novel genetic variation associated with HbF levels in SCD patients by querying GWAS and WES data. Unfortunately, this effort was unsuccessful, highlighting the need for larger studies of well-phenotyped SCD patients. Larger SCD cohorts will not only enable the discovery of new HbF variants but also variants associated with SCD-specific complications (e.g. painful crises, acute chest syndrome, leg ulcers) independently of HbF levels. It is by integrating this genetic information with classical (e.g. complete blood counts) and newer (e.g. metabolites) biomarkers that we can envision developing precision medicine strategies to predict SCD complications or tailor therapeutic interventions.

Methods

Study participants

The goal of the study was to identify and characterize DNA polymorphisms and genes associated with HbF levels in SCD patients. This meta-analysis comprised 3740 participants with the HbSS or HbSβ⁰ genotype: detailed description of the participating cohorts is provided in Table S1. Sample collections and procedures were in accordance with the institutional and national ethical standards of the responsible committees and proper informed consent was obtained. The project was approved by the Montreal Heart Institute Ethics Committee (project #09–1137).

SNP array genotyping and quality-control steps

Participants from the different studies were genotyped on different genotyping arrays and at different locations. We carried out all quality-control steps and genotype imputation at the study level. See Table S9 for the number of samples and variants pre-QC, post-QC and post-imputation. We performed genetic association testing at the study level and combined association results using a meta-analysis method. The Cooperative Study of Sickle Cell Disease (CSSCD, N = 1279) and the Tanzania cohort (N = 1213) have been described elsewhere [21, 37, 38]. DNA samples of participants from GEN-MOD (N = 398), Mondor/Lyon (N = 321), the Multicenter Study of Hydroxyurea in Sickle Cell Anemia (MSH, N = 57), the Adult Sickle Cell Center at Georgia Health Sciences University (GHSU, N = 182), and the Jamaica Sickle Cell Cohort Study (JSCCS, N = 89) were genotyped on the Illumina Infinium HumanOmni2.5Exome-8v1.1 array at the Montreal Heart Institute Pharmacogenomics Center. We performed quality control using PLINK [39], removing SNPs with Hardy–Weinberg P < 1 × 10⁻⁷ and genotyping rate < 90%. After quality-control steps, we imputed genotypes using reference haplotypes from TOPMed Freeze5 GRCh38/hg38 and Minimac4 (v1.2.4) as implemented on the TOPMed imputation server [40]. For downstream analyses, we only considered variants with imputation quality Rsq > 0.3. Furthermore, unless variants were previously reported to be associated with HbF levels, we only kept variants with a minor allele frequency (MAF) ≥1% for the GWAS discovery effort because rarer variants are more difficult to impute, especially in non-European-ancestry populations. Rare coding variants are the focus of the whole-exome sequencing experiment described below. For principal component (PC) and uniform manifold approximation and projection (UMAP) analyses, we first created a subset of common (MAF > 5%) variants that are shared between our datasets and the 1000 Genomes Project phase 3 dataset. The first 10 PCs were computed on that subset using PLINK’s —pca function. UMAP values were computed from the first 5 PCs using the UMAP package in R.

Whole-exome DNA sequencing and quality-control steps

Study samples

We combined together five cohorts (GEN-MOD [N = 406], the Duke University Outcome Modifying Genes [OMG, N = 13], the CSSCD [N = 245], Differential Response to Hydroxyurea and Incidence of Stroke in Sickle Cell Disease [CIP, N = 370, dbGaP Study Accession: phs000691.v2.p1], and Mondor/Lyon [N = 321]) sequenced at different time points and using different sequencing capture methods [41]. We modeled our quality-control steps after the Exome Aggregation Consortium (ExaC) [42].

Alignment and BAM processing

The paired-end sequence reads from exomes were aligned to the human genome reference (hg19) using bwa (v0.7.17) (BWA MEM) [43]. However, all results were lifted-over onto build GRCh38/hg38 of the human genome using the UCSC Browser LiftOver tool.

Base quality recalibration

The base quality scores were then recalibrated using GATK BaseRecalibrator and a list of known variant sites from dbSNP. The sequenced interval came from GENCODE. The new base quality scores were then applied using GATK ApplyBQSR but retaining the original base quality scores within the BAM.

Variant calling

Using GATK HaplotypeCaller, the recalibrated BAM file from the previous step was used to perform variant calling per sample. The output is in GVCF mode, which can be used for joint genotyping with multiple samples.

Variant-quality score recalibration (VQSR)

VQSR (GATK ApplyVQSR) was then applied and the raw VCFs from the previous step were filtered to achieve a high degree of sensitivity and reduce false positives. The SNP VQSR model is trained using HapMap3.3 and 1KG Omni 2.5 SNP sites and a 99.6% sensitivity threshold was applied to filter variants. Recalibration of insertion/deletion sites used Mills et al. 1KG gold standard and Axiom Exome Plus sites with a 95.0% sensitivity threshold [44].

Variant annotation

We employed Variant Effect Predictor (VEP) version 101 to annotate variants [45]. We retrieved several protein prediction consequences info using VEP’s plugin (LOFTEE, SpliceAI, SIFT, Polyphen2, and MaxEnt). We queried Ensembl/GENCODE and RefSeq transcripts databases and restricted results to produce the most severe consequence per variant. Variants mapping to coding regions were kept for downstream analyses.

High-quality (HQ) variants

Variant sites were labeled as high-quality if they met the following criteria: (1) they were given a PASS filter status by VQSR, (2) at least 80% of the individuals in the dataset had a depth (DP) ≥ 10 and genotype quality (GQ) ≥20, (3) there was at least one individual carrying the alternate allele with depth ≥ 10 and GQ ≥20, and (4) the variant was not located in the 10 1-kb regions of the genome with the highest levels of multi-allelic variation. Once we applied the variant filtering criteria, 985 119 variants were left.

HbF normalization

Except for one participant, HbF was measured in SCD patients not taking hydroxyurea and at least three months from the most recent blood transfusion. Within each cohort, we corrected using linear regression HbF levels (continuous) for age (continuous), sex (binary) and β-globin genotypes (binary, HbSS or HbSβ⁰), and then normalized the residuals using inverse normal transformation to create HbF Z-scores that are comparable between studies (mean of 0 and variance of 1). To condition on the known HbF regulators (genetic variants at chr2-BCL11A, chr6-HBS1L-MYB and chr11-β-globin), we regressed out using linear regression genotypes from HbF-associated variants at these loci from the HbF Z-scores (using a stepwise approach with variants within a 1-Mb window centered on the strongest association signal at each locus).

Genetic association analyses (GWAS)

We calculated association statistics using linear regression (imputation dose, additive model) as implemented in rvtests [46] and the first 10 PCs as covariates. Rvtests implements a covariance matrix to account for cryptic relatedness. In our primary analysis, we meta-analyzed study-level association results using a fixed-effect method (inverse variance-weighted) implemented in the software metal [25]. For the 20 novel HbDF variants, we also meta-analyzed results using MR-MEGA [26]. To determine if known HbF variants replicated in our SCD-only meta-analysis, we considered sentinel variants as well as variants in strong linkage disequilibrium (LD, r² ≥ 0.8 in TOPMed European- or African-ancestry participants, as appropriate) [47]. We calculated the HbF phenotypic variance explained by the HbF-associated variants in the CSSCD using an additive genetic model and linear regression as implemented in R 4.3.0. For this analysis, we considered six independent variants at three known HbF loci (rs1427407_BCL11A, rs7606173_BCL11A, rs6940878_HBS1L/MYB, rs9389269_HBS1L/MYB, rs114398597_HBS1L/MYB, rs10128556_HBB) that we identified in our previous study, and six new variants at five loci that replicated in this study (rs111607747_ASB3, rs62408218_BACH2, rs9891699_PFAS, rs10423017_ZBTB7A, rs4804210_KLF1, rs7508226_KLF1). We calculated the HbF polygenic score as described before using six SNPs independently associated with HbF levels in SCD patients (two at BCL11A, three at HBS1L-MYB and one at HBB) [48].

Genetic association analyses (WES)

Single-variant association analyses

We analyzed variants with MAF < 1%, and excluded variants and samples with low genotyping rate (<95%), as well as variants deviating from Hardy–Weinberg equilibrium (P < 1 × 10⁻⁷). HbF levels were normalized as described above. We performed single-variant associations correcting for age, sex, the first 10 PCs, the kinship matrix and the different sequencing captures. All analyses were performed using rvtests (v.20171009) [46].

Gene-based association analyses

We employed three strategies to aggregate variants for gene-level association testing. For a given SNP if at least one out of seven algorithms (PolyPhen2 HumDiv and HumVar, LRT, MutationTaster, LOFTEE, SpliceAI, SIFT, and MaxEnt) predicted it as ‘deleterious’, we labeled the mask as broad. If all seven algorithms predicted the variant as ‘deleterious’, we labelled the mask as ‘strict’. The last strategy considered just loss-of-function variants as predicted by LOFTEE. Two statistical tests were considered for each mask: an adaptive burden test, which aggregates rare variants based on optimal frequency cut-off (VT), and SKAT, a bidirectional approach that includes SNPs with variable effect size and direction. Gene-level associations were conducted using rareMETALS_7.1 [49]. Only variants (MAF < 1%) annotated as missense, nonsense, essential splice site, and frameshift indel were kept for gene-based analyses.

Cell culture

HUDEP-2 cells stably expressing Cas9 or dCas9^KRAB were produced by lentiviral transduction of Addgene lentiCas9-Blast vector (#52966) and pHR-SFFV-dCas9-BFP-KRAB vector (#46911) and cultured as follows. For expansion, cells were cultured in StemSpan™ SFEM (Cedarlane #09650) supplemented with 1 μg/ml doxycycline (Sigma #D9891), 0.4 μg/ml dexamethasone (Sigma #D2915), 0.1 μg/ml stem cell factor (Cedarlane #255-SC-050), 3 units/ml erythropoietin (Cedarlane #287-TC-500), 2 mM L-glutamine (Life technologies #25030), 200 IU penicillin, 200 μg/ml streptomycin (Wisent Biocenter #450–201-EL), and incubated at 37°C and 5% CO₂. Cells were always kept at a density between 20 000 and 1 000 000 cells/ml. For cell differentiation, HUDEP-2 cells were incubated for 4 days at 37°C and 5% CO₂ in Corning™ Iscove’s modification of DMEM (Thermo Fisher Scientific #MT15016CV) supplemented with 3 units/ml recombinant human erythropoietin (R&D System #287-TC-500), 0.1 μg/ml recombinant human stem cell factor (SCF) (R&D System #255-SC-050), 1 μg/ml doxycycline hydrochloride (Sigma #D3447), 2 mM L-glutamine (Thermo Fisher Scientific #25030–081), 200 U/ml penicillin, 200 μg/ml streptomycin (Wisent Bioproducts #450–201-EL), 10 U/ml heparin (Sigma #H3149), 10.5 μg/ml insulin (Sigma #19278-5Ml)), 330 μg/ml human holo-transferrin (Sigma #T0665), and 5% solvent detergent pooled plasma AB (Rhode Island Blood Center #X0004). HEK293FT cells were cultured at 37°C and 5% CO₂ in DMEM glutamax (Thermo Fisher Scientific #10569010) supplemented with FBS premium grade (VWR #97068–085) and always kept below 70% confluence.

Cloning

The pHKO-09_CM vector was digested with FasDigest BsmBI (ESP3I) (Thermo Fisher Scientific #FERFD0454) and treated with FastAP (Thermo Fisher Scientific #EF0654). Digested plasmid was excised from agarose gel and purified with QIAquick gel extraction kit (Qiagen #28704). For single guide cloning, gRNA were cloned in 50 ng of vector using Quick Ligation™ Kit (New England Biolabs #M2200S) in a 11 μL reaction and transformed in One Shot™ Stbl3™ chemically competent E. coli (Thermo Fisher Scientifc #C7373703). For gRNA library cloning, gRNAs were cloned using Gibson assembly master mix (New England Biolabs #E2611L) as described [50]. Plasmids were extracted with Nucleobond Xtra Midi EF Kit (Macherey-Nagel #MN-740420.50).

DNMT1 mutagenesis

The pcDNA3/Myc-DNMT1 vector (containing DNMT1 isoform a) was purchased from Addgene (#36939). Primers used for site-directed mutagenesis are listed in Table S10. The reaction was performed using the Q5® Site-Directed Mutagenesis Kit (NEB #E0554S) following the manufacturer’s instructions. Transformed mutated plasmids were extracted using QIAprep Spin miniprep kit (Qiagen #27106). The mutated region was sequenced by Sanger sequencing using the CMV_fwd primer (Table S10). The mutated DNMT1 open reading frame (ORF) was amplified by PCR using primers listed in Table S10 and the NEBNext High-fidelity 2X PCR Master mix (NEB #M0541L). The pLJM1-EGFP vector (Addgene #19319) was digested with AgeI-HF (NEB #R3552S) and EcoRI-HF (NEB #R3101S). Digested pLJM1-EGFP and amplified mutated DNMT1 ORF were purified from agarose gels using Qiagen’s QIAquick gel extraction kit (#28704). They were cloned together with the Gibson assembly master mix (NEB #E2611L) and following the manufacturer’s instruction. Recombined plasmids were extracted from bacteria using the QIAprep Spin miniprep kit. The entire DNMT1 ORF was sequenced by Sanger sequencing using primers listed in Table S10. After confirmation of the DNMT1 ORF sequence, bacteria were expanded in 50 ml of LB culture. Endotoxin-free pLJM1-DNMT1 plasmid was extracted from 50 ml culture using Nucleobond Xtra Midi EF (D-mark Bioscience #MN-740420.50).

Viral production

HEK293FT cells were seeded at 8 x 10⁴ cell/cm². On the next day, 68 ng/cm² of pMD2.G, 104 ng/cm² of psPAX2, and 136 ng/cm² the target were transfected in HEK293FT cells using 1.32 μl/cm² of the PLUS reagent (Life technologies # 11514015) and 1.2 μl/cm² of Lipofectamine 2000 (Life technologies #11668–019) in Opti-MEM (Life technologies #31985070). After 4 hours of incubation at 37°C and 5% CO₂, medium was changed for DMEM, with 10% FBS and 1% BSA. Medium containing viral particles was collected 2 days post transfection, centrifuged 5 min at 200 × g and filtered using Millipore Steriflip disposable 0.45 μm (Millipore Sigma #SE1M003M00).

Viral infection

For single-cell experiments, HUDEP-2 dCas9^KRAB cells were seeded at a density of 2 × 10⁵ cells/ml with 1 μl/ml of polybrene 1 mg/ml in 0.9% NaCl. Viral suspension was added to achieve a MOI of 0.3. 24 h post infection, 1 μg/ml of puromycin was added. 48 h post infection, cells were transferred in differentiation medium, maintaining the puromycin concentration at 1 μg/ml. Cells were maintained in differentiation for 4 days. For single gRNA experiments, HUDEP-2 Cas9 cells were seeded at a density of 2.5 × 10⁵ cells/ml and 1.25 μl/ml of polybrene 1 mg/ml were added. 3 μl of viral suspension were added for each 8 μl of cell culture. 24 h after infection, 1.200 μg/ml of G418 were added. Cells were maintained under this selection for at least 6 days. For mutated DNMT1 overexpression, cells were infected the same way as for single gRNA, but selected with 1 μg/ml of puromycin for 4 days.

Targeted single-cell CRISPR inhibition for regulators of HBG1/2 expression

gRNA library design

We designed the gRNA library for the single-cell CRISPRi experiment before we had finalized the HbF meta-analysis in SCD patients (Table S5). Therefore, some of the selected variants for targeting did not show strong evidence of association in the final analysis. Nonetheless, we provide all our results in Table S6. Our complete library included 54 gRNA (Table S5). When possible, we attempted to design up to three gRNA per targeted variant using a previously described bioinformatic pipeline [51].

Library construction, cell culture and DNA sequencing.

At the end of the selection time, cells were processed with the Dead Cell Removal Kit (Miltenyi Biotec Inc. #130–090-101) using MS Columns (Miltenyi Biotec Inc. #130–042-201) to remove dead cells from the selection process following the manufacturer’s instructions. Cell density was assessed using Thermo Fisher Scientific’s Countess™ Automated Cell Counter II FL and Countess™ Cell Counting Chamber Slides (Thermo Fisher Scientific #C10228). Cells were processed with Chromium Next GEM Single Cell 5′ GEM kit v2 (10× Genomics #1000266), Chromium Next GEM Single Cell 5’ Gel Bead Kit v2 (10× Genomics #1000267), Chromium Next GEM chip K Single Cell Kit (10× Genomics #1000286), and 5’ CRISPR Kit 10× Genomics #1000451) in a Chromium Controller from 10X Genomics. DNA was purified with SPRIselect Reagent Kit (Beckman Coulter #B23318). The cDNA library fraction was processed with a Library construction kit (10X Genomics #1000196) and the gRNA library was processed using a 5’ CRISPR Kit (10X Genomics #1000451). Samples were indexed using Dual Index Kit TT Set A. The final quality of the libraries was assessed using Agilent’s High Sensitivity DNA kit (Agilent #5067–4626) and Agilent 2100 Bioanalyzer. The concentrations of the librairies were assessed using a Qubit 4 Fluorometer (Life technologies #Q33226) and the Qubit 1X dsDNA High Sensitivity Assay kit (Life technologies #Q33231). All steps from cell processing with the Chromium Controller to the libraries’ quality assessment were done following Chromium Next GEM Single Cell 5′ v2 with Feature Barcode technology CRISPR Screening Rev A. For gRNA library purification, the last purification step using SPRISelect beads (step 6.5) was performed twice. Libraies were sequenced using Illumina NovaSeq 6000 S4 PE100.

Statistical analyses

We used CellRanger from 10X Genomics to generate unique molecular identifier (UMI) count tables and to link gRNA and the transcriptome within each cell. For the statistical analysis of gRNA individually, we divided the datasets between control (no UMI for the tested gRNA and a UMI count ≥1 for the safe gRNAs) and targeted cells (UMI count ≥20 for the tested gRNA). We selected this threshold of 20 based on calibration analyses using the negative (GYPA/B) and positive (BCL11A) controls. We compared gene expression levels between control and targeted cells using quasi-Poisson regression, correcting for: (1) the total number of genes expressed in the cell, (2) the total number of gene UMIs sequenced in the cell, (3) the proportion of sequenced gene transcripts that map to mitochondrial genes, and (4) the sequencing batch. To declare significance, we considered a Bonferroni correction for the number of independent variants targeted (0.05/14 = α = 0.0036). We performed all statistical analyses in R 4.3.0. We vizualized the KLF1 locus in IGV 2.17.4 [52]. We obtained the ATAC-sequencing track for human erythroid cells from ref.[53], the enhancer-to-gene predicted links in K562 cells from ref.[54] (filtered with a score threshold > 0.7), and the ENCODE candidate cis-regulatory elements (cCREs) from ref.[55].

Western blot

The lysis of 1 million cells was performed in 200 μl of RIPA buffer supplemented with phosphatase inhibitor cocktail 2 (Sigma #P5726), phosphatase cocktail 3 (Sigma #P0044), 1 mM PMSF (Sigma #P7626), and protease inhibitor cocktail (Millipore Sigma #P2714). Samples were incubated at 4°C for 5 min with constant shaking, then centrifuged at 17900 × g for 10 min at 4°C. Supernatant was collected, proteins were quantified using the BCA protein assay (Thermo Fisher #PI23225), and protein extracts were denatured in 4X Laemmli sample buffer (BioRad #1610745) supplemented with 50 mM DTT 5 min at 95°C. 15 μg of proteins were loaded onto a 8% polyacrylamide gel. After migration, proteins were transferred onto a nitrocellulose membrane. The DNMT1 protein was detected with a DNMT1 monoclonal antibody (Thermo Fisher #MA5–16169) and the GAPDH protein was detected with GAPDH antibody (NEB #2118S).

RNA extraction and quantification

HUDEP-2 cells were collected and centrifuged to remove media and processed with the RNeasy Plus Mini Kit (Qiagen #74134) following the manufacturer’s instructions. Extracted RNA was quantified using the Biotek’s Cytation 5 imaging reader and a BiotekTake3 plate, and with Agilent RNA 6000 Nano Kit and Agilent 2100 Bioanalyzer following the manufacturer’s instructions. The average of the two methods was used.

cDNA synthesis and qPCR

RNA was reverse-transcribed using the High-Capacity cDNA Reverse Transcription Kit (Thermo Fisher Scientific #4368814) and RNAseOUT™ (Thermo Fisher Scientific #10777019) following the manufacturer’s instructions. cDNA was diluted 1:50 to set up the qPCR reaction. Platinum SYBR™ Green qPCR Master Mix-UDG (Thermo Fisher Scientific #11733046) was used in combination with Precision Blue Real-Time PCR dye (Bio Rad #1725555) and the reaction was carried over in Hard-Shell® 384-Well PCR plates (Bio Rad #HSP3805) in a CFX™ Real-Time System from Bio Rad. Primers are listed in Table S10. Data were analyzed using Bio Rad’s CFX Manager v3.1.

Author contributions

Conceived and designed the analyses: Y.I., N.B., and G.L.; Collected the data: Y.I., N.B., and K.S.L.; Contributed data: S.M., K.M., M.R., A.K., A.E.A.K., M.J.T., P.J., F.G., and P.B.; Performed analyses: Y.I., N.B., K.S.L., and G.L.; Secured funding and supervised the work: G.L.; Wrote the manuscript: Y.I., N.B., and G.L., with contributions from all authors.

 

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

Funding

This work was funded by the Canadian Institutes of Health Research (PJT #186159) and the Canada Research Chair program (Lettre). OMG-SCD was funded by R01HL68959 (Telen) and R01HL079915 (Telen) from the National Heart, Lung, and Blood Institute (NLHBI).

Data availability

The CSSCD genetic dataset is available in the database of Genotypes and Phenotypes (dbGaP: https://www.ncbi.nlm. nih.gov/gap/), accession phs000366.v1.p1. The Tanzania SCD data is available from the European Genome-phenome Archive at the European Bioinformatics Institute (accession number EGAS00001000990). The other datasets have not been deposited in a public repository because data are not public but are available from the corresponding studies on request. The CRISPRi + scRNA-seq UMI count matrices in HUDEP-2 cells are available at: http://www.mhi-humangenetics.org/en/resources/.

Ethics approval statemet

We collected data according to the Helsinki declaration and the study was approved by the Montreal Heart Institute institutional ethics committee, Project #2009–106, (09–1137).