Resistance to modulation of in vivo gene expression by genomic editing

Abstract

Enhancer editing is widely proposed as a strategy to modulate gene dosage therapeutically. EKLF/KLF1, a master erythroid transcription factor, is an attractive target because haploinsufficiency is benign and associated with hereditary persistence of fetal hemoglobin (HPFH). Here we test whether enhancer disruption can recapitulate this state by targeting two KLF1 regulatory elements, EHS1 and INT1. EHS1 editing reduces KLF1 expression and increases γ-globin, and combined INT1/EHS1 targeting produces additional increases in some samples, although responses are variable. In xenotransplantation models, edited cells engraft and retain indels but fail to sustain consistent γ-globin induction. Epigenetic profiling across the KLF1 genetic locus reveals a remarkable resistance to loss of hypersensitive sites. These results establish a key principle: the transcriptional consequences of enhancer editing are strongly constrained by local genomic architecture. Attempts to mimic KLF1 haploinsufficiency by non-coding indels are blunted by dense regulatory architecture, enhancer redundancy, and long-range interaction. Thus, it may be difficult to develop KLF1 as a target for gene therapy for hemoglobinopathies until tools for recapitulating natural mutations via heterozygous edits are developed. Our findings underscore the observation that disruption of non-coding elements may not always recapitulate the biology of haploinsufficiency from heterozygous loss-of-function mutations.

Structural and epigenetic properties of the human KLF1 locus constrain the in vivo effects of efficient editing at its non-coding regulatory elements, providing a cautionary tale for the design of functional enhancer modifications.

Introduction

Hematopoietic stem cells generate all mammalian blood lineages, each arising following a process of gradual restriction of downstream cellular potencies^1,2. Functional integration of transcription factor expression^3,4 is important for cell expansion and the terminal differentiation steps of erythropoiesis. In this context, Erythroid Krüppel-like Factor (EKLF/KLF1⁵) is an essential zinc finger transcription factor that plays a critical role in erythroid transcriptional activation and chromatin/higher order assembly^6,7, and is recognized as a global regulator of all aspects of erythropoiesis^8–13. Its function in hemoglobin switching is both direct (activation of the adult ß-globin gene) and indirect (regulation of repressors of γ-globin gene expression)^6,14,15.

Genetic studies underscore the critical importance of KLF1 expression. Ablation of Klf1 in mice leads to embryonic lethality precisely at the time of the switch to adult ß-globin^16,17 due to a profound ß-thalassemia and the low to virtually nonexistent expression of erythroid genes of all categories. KLF1 haploinsufficiency in humans is benign and leads to hereditary persistence of fetal hemoglobin (HPFH) (originally mapped in a Maltese family¹⁸) and also to persistence of embryonic globin expression¹⁹. Over 100 heterozygous human mutations in KLF1 have been identified^10,12,13. Allelic distribution of KLF1 mutations is of dramatically higher prevalence in regions of endemic ß-thalassemia²⁰, where the dysregulation of γ-globin expression is clinically advantageous. In this study, a population analysis from over 5000 consecutive patients showed that there is a selective pressure for haploinsufficient KLF1 expression to decrease KLF1 levels so that, ultimately, HbF expression can increase and thus lead to alleviation of symptoms. A case report of ß^o/ß^o thalassemic twins highlights the remarkable clinical difference in anemia and transfusion dependence that follows from loss of one KLF1 allele²¹. KLF1 haploinsufficiency has no effect on erythroid differentiation²², these individuals have no pathology²³, and the “fitness” of HUDEP2 cells targeted at the KLF1 locus is increased²⁴, indicating that its removal is not negative for cell viability. These data suggest that modulation of its expression level could be clinically useful for alleviation of hemoglobinopathies, particularly as this would also result in lower levels of γ-globin repressors such as BCL11A and ZBTB7A^18,25,26.

Based on these observations, our work has focused on deciphering and defining KLF1 controlling elements, aiming for a fundamental understanding of KLF1 promoter regulatory mechanisms to enable therapeutic modulation of its levels in thalassemic and sickle cell patients. We recently analyzed the KLF1 transcription unit to more precisely investigate the role of a conserved sequence in intron 1 (INT1), which binds multiple regulatory factors^27,28, in modulating expression of KLF1, and used a dual Cpf1 (Cas12a) approach to delete this important region²⁸. Unexpectedly, the overall effect on KLF1 expression of deleting the INT1 site remained relatively minor in primary cells, despite prior evidence from reporter transfections and cell line editing showing that even a directed mutation/small deletion exerts a significant effect on reporter or endogenous KLF1 transcription. However, these observations parallel those seen in pedigrees of families with KLF1 haploinsufficiency, where downstream effects on expression are variable, but universally benign^18,22,29.

We thus hypothesized that the overall architecture of the KLF1 genomic region may be buffered against localized changes having a substantive effect on its expression²⁸. Consistent with this idea, the KLF1 transcription unit is located in the midst of multiple interacting domains^30–32. Based on these insights, we postulated that reproducible and precise modulation of KLF1 levels may require combinatorial manipulation(s) in coordination with directed changes at INT1. Here, we investigate this possibility and define additional mechanisms contributing to robust control of KLF1 expression.

Results

Localization of an indel target sequence within an upstream DNAse hypersensitive site, EHS1

Building on prior studies, we focused on a 950 base pair region, located just upstream of the KLF1 transcription initiation site, that is sufficient to generate erythroid-specific expression in transient assays³³. This region exhibits the most significant homology upon a seven species alignment of 30 kB of surrounding genomic DNA²⁷, and harbors erythroid-restricted DNAse hypersensitive sites³³ that show a regulated onset during hematopoiesis, development, and erythroid differentiation as judged by ATAC-seq analyses²⁸. One of these sites (erythroid hypersensitive site 1; EHS1) (Fig. 1) behaves as a very strong enhancer, which in conjunction with the proximal promoter³⁴ accounts for KLF1’s tissue-specific expression³³. The importance of this short region has been verified in vivo^25,35,36. EHS1 is considered a well-characterized erythroid enhancer where a complex of proteins (the “enhanceosome”) provides high level expression^27,37,38 (Fig. 1a).

Fig. 1. Design and testing of gRNAs directed at the human KLF1 upstream enhancer, EHS1.

a Schematic of the mouse Klf1 genomic locus (not drawn to scale), showing the relative locations of erythroid hypersensitive 1 (EHS1)³³ and the INT1²⁷ elements. The EHS1 region is expanded to show a detailed layout of conserved KLF1 cis-regulatory elements^27,37. Blocks of conserved sequence homology between seven mammalian species are color-coded for their transcription factor binding sites, highlighting locations of known interacting enhanceosome components at EHS1 that includes the Gata/Tal1/Lmo2/Ldb1 multi-protein complex, Smad5, Riok2, P300/CBP, DEK, and DDX5 proteins^{37,38,71–77}. A similar analysis for INT1 is published²⁷. Below are indicated the locations of the Cas9 gRNA target sites within EHS1 tested in the present study. b Extent of indels was determined by ICE analysis of genomic DNA from Cas9 RNP-treated human CD34+ cells. n = 2–3 independent samples. c Expression of KLF1, ß-, and γ- globin RNAs were monitored by RT-qPCR at the indicated time points (days) before (D0) and after expansion (D7) and differentiation (D18) of CD34+ cells following a 3-step protocol. n = 9–12 independent samples. d KLF1 expression was monitored after RNP treatment with scrambled (scr; black) or the best EHS1 gRNA (sg63; green). Also shown are the results from deep sequencing of the sg63-treated population, with the gRNA sequence highlighted in yellow and the PAM sequence in turqoise; the total of all deletions was 67%. n = 9 independent samples; p value is shown and from a two-tailed, unpaired t-test. e KLF1 expression was monitored by RT-qPCR after concurrent treatment of hCD34+ cells with both EHS1/INT1-directed RNPs (blue), compared to scr control (black). Indel percentages (ICE analysis) were 92 ± 1.5% at INT1 and 63 ± 4.5% at EHS1. n = 9 independent samples; p value is shown and from a two-tailed, unpaired t-test.

A critical aspect from the extended studies of KLF1 gene regulation is that removal of the complete EHS1 site leads to a dramatic 50-fold reduction in KLF1 expression²⁵; however, removal of a smaller subregion gives a more nuanced effect in edited mouse cells^37,38. This led us to consider whether small changes in the human EHS1 region could achieve reduced hKLF1 expression without complete loss of EHS1 function.

We tested whether Crispr/Cas9-directed deletion of small sequences within EHS1 could yield quantitative, rather than total, ablation of activity. Based on the Cas9 PAM sequence (NGG), we designed five sgRNAs based on location and CRISPOR/off-target scores (Supplementary Table S1) and tested these in primary human CD34+ cells (Fig. 1a), using gRNA/Cas9 protein RNP protocols and transfection conditions previously used in our studies of INT1²⁸. These five gRNAs were directed at regions that span over 100 bp of EHS1 (Fig. 1a). We evaluated genomic DNA within one day after transfection and find that only the “63rev” design achieves greater than 50% efficacy in changing the EHS1 region, leading to 60% indels (Fig. 1b). The “21rev” guide is also moderately effective, while none of the other three give appreciable levels of indels within the EHS1 region. These data support the idea that only a proper gRNA design at the KLF1 locus will lead to significant editing, even when various gRNAs are tested within a small region of interest.

Effects of EHS1 deletion on KLF1 expression

We next tested the effects on KLF1 expression levels after the full 18-day, 3-step expansion/differentiation protocol, analyzing time points at d0, d7 (end of phase 1, start of expansion phase 2), and d18 (end of differentiation phase 3). This procedure induces robust levels of KLF1 and its target ß-globin expression by the end of the protocol, with only a small and transient increase in γ-globin expression (Fig. 1c). Editing with 63rev results in a reproducible decrease in KLF1 expression compared to a ‘scrambled’ (scr) negative control (with no homology) (Fig. 1d). Deep sequence analysis indicates that ~67% indels are generated following 63rev Crispr RNP delivery. For clarity, 63rev is hereafter designated “EHS1”. These findings indicate that EHS1 editing achieves a partial reduction in KLF1 expression, resembling the effects of haploinsufficiency rather than a complete loss of expression.

We next tested whether using gRNAs directed at both INT1 and EHS1 would yield a greater additive, or perhaps synergistic, effect on KLF1 levels. Optimal editing at INT1, to 99%, requires concurrent addition of Cas12a (Cpf1) CRISPR proteins with two gRNAs directed at the INT1 region. Here, dual editing (at both EHS1 and INT1) is complicated by the requirement for two different CRISPR proteins (Cas9 for EHS1 and Cas12a (Cpf1) for INT1²⁸), which are thus presented to the cells at lower concentrations than that used at full strength. Under these conditions, 58% of the cells are calculated to carry indels at both sites, but unexpectedly, KLF1 expression after dual editing is no greater than that generated with EHS1 alone (Fig. 1e).

Testing effects of EHS1 deletion on γ-globin expression

To assess the biological consequences of the targeted deletions at INT1 and EHS1, we focused on the anticipated effects of KLF1 knock-down on increasing γ-globin gene expression for two reasons: one that it is a sensitive readout for downstream effects^10,20,28; and two, that it would provide a clinical basis for future studies, if effective. We performed the full 18-day erythroid differentiation/expansion protocol and analyzed globin expression after treatment with the various gRNAs designed in Fig. 1. Unsurprisingly, the gRNAs that gave rise to minimal edits have little effect; this is also true for the 21rev gRNA. In contrast, 63rev/EHS1 induces a significant increase in γ-globin expression (Fig. 2a). These data show that only a proper editing design at the KLF1 locus can attain a measurable downstream biological outcome. Directly comparing the effects of INT1 or EHS1 (63rev) deletion within the same experiment demonstrates that an indel directed at EHS1 is more effective than at INT1 (Fig. 2a).

Fig. 2. Monitoring downstream effects of single and combination indels at the KLF1 genetic locus.

a γ-globin expression of differentiated CD34+ cells treated with each of the RNPs described in Fig. 1 (green) or INT1 (red) was monitored by RT-qPCR and compared to scr control (black). n = 9 independent samples. b γ-globin expression was monitored after concurrent treatment of hCD34+ cells with both EHS1/INT1-directed RNPs (blue) or scr control (black). Indel percentages were as in Fig. 1E. n = 9–12 independent samples. c Direct comparison of single EHS1 (green) vs INT1/EHS1 dual (blue) treatment effects on γ-globin expression within the same experiment. Indel percentages were 33% for EHS1 alone, or 75% (INT1) and 18% (EHS1) in dual. n = 9 independent samples; p values are shown and from two-tailed, unpaired t-tests.

We next tested whether combination editing at both EHS1 and INT1 would yield an even more significant increase in γ-globin. Use of the dual editing approach at EHS1 and INT1 increases γ-globin expression effectively (Fig. 2b). To provide a baseline for comparison to other experiments, we repeated this experiment with inclusion of EHS1 alone; again, the dual approach is better (Fig. 2c). However, it is noteworthy that the effective range is quite large, with some samples showing little effect beyond that of INT1 alone, but others showing a significant increase. This parallels our earlier findings of variability after INT1-directed indels²⁸. HbF protein expression is significantly but modestly induced, while HbA2 and HbA levels remain unchanged (Supplementary Fig. S1).

Xenotransplantation studies

Collectively, the studies of Figs. 1 and 2 show a significant level of resistance to modifying KLF1 transcription in cultured cells. A true test for an effective and stable effect requires in vivo testing. We therefore performed xenotransplantation experiments of control and gene-edited human CD34+ cells into NBSGW mice and harvested/analyzed bone marrow and other tissues 17 weeks after transplant.

We initially tested INT1-directed indels and compared the cells to those derived from the scr negative control. We also included tests of indels directed at the +58 enhancer within the BCL11A gene as a positive control. Human/mouse chimerism is adequate based on levels of human CD45+ cells, and the +58/BCL11A positive control yields increased γ-globin mRNA in the two hCD235+ samples with the highest number of indels and chimerism of that set (Fig. 3a; transplant #1). In contrast, INT1-directed editing failed to increase γ-globin expression in any of the samples, despite 98-99% indel frequency in three of the four CD235a+ samples (Fig. 3a; transplant experiment #1).

Fig. 3. In vivo transplants of treated hCD34+ cells into NBSGW mice.

a Detailed results from “Transplant #1” with scrambled (scr), BCL11A-directed (+58), or INT1-directed RNPs. Analysis of γ-globin expression in hCD235a cells was performed 17 weeks after transplant by RT-qPCR. The extent of chimerism for each mouse is shown below, with the percentage indels as indicated. Input indel percent at INT1 was 56%. n = 4–5 independent mice. b Results of all transplant studies (Transplants 1–4), each set grouped together, showing extent of chimerism and levels of CD235a cells recovered, along with each test (scrambled (SC), INT1, EHS1, or both (DG)) as indicated. n = 4–5 independent mice [also see Supplementary Fig. S2]. c γ-globin expression was monitored for transplants #2-4, following RNP treatments as indicated (scr (black), EHS1 (green), dual INT1/EHS1 (blue). Transplant #2, indel percentages were 26 ± 6.7% for EHS1 alone, or 86 ± 2.0% (INT1) and 8.0 ± 4.0% (EHS1) in dual; inputs were 28% for EHS1 alone, or 61% (INT1) and 14% (EHS1) in dual. Transplant #3, indel percentages were 42 ± 17% for EHS1 alone, or 90 ± 5.8% (INT1) and 37 ± 13% (EHS1) in dual; inputs were 52% for EHS1 alone, or 72% (INT1) and 18% (EHS1) in dual. Transplant #4, indel percentage was 22 ± 8.2% for EHS1 alone. n = 12–15 independent samples; significant (≤0.5) p values are shown and from two-tailed, unpaired t-tests [also see Supplementary Fig. S3]. d. EHS1 indel percentages in hCD45+ and CD235a+ cells from individual mice from Transplant #4.

We next tested the effects of indels directed at EHS1 alone or EHS1 together with INT1 in three additional xenotransplantation studies. Although chimerism (CD45+) and CD235a+ levels are significant in all cases (Fig. 3b; transplant experiments #2–4) and the human cells also contribute to other hematopoietic cell types (Supplementary Fig. S2), some things are noteworthy about the data and analysis (Fig. 3c; transplant experiments 2–4). First, even though transplant #2 suggests a positive effect of the dual targeting, this is not reproduced in transplant #3. Second, as noted with the cells in culture, the range of effects within individual samples is quite variable within the same experiment. Third, the scrambled control is not consistently low in all samples (or perhaps better to say it is aberrantly high in transplant #3). We do not see a consistent increase in γ-globin levels beyond that from the scrambled negative control in the erythroid CD235a+ cells. The bottom line is that any effect, if present, of combining deletions within important KLF1 genetic loci is not robust nor consistent across transplant samples. The lack of effect isn’t due to normalization variability, as the same results are apparent when RPL13A is used for normalization instead of catalase (Supplementary Fig. S3). In addition, the lack of effect is not due to any selection for/against the modified cells, as the level of indels in the erythroid CD235a+ cells is the same as in the global CD45+ cells in four of the five cases directly analyzed from the same individual cell harvest (Fig. 3d).

We must conclude that, although indel-modified CD34+ cells increase γ-globin expression in culture, this effect is not robust or reproducible after xenotransplantation, as the cells do not dominate the in vivo environment. This suggests that the current tools are not sufficient to achieve a durable effect at the KLF1 locus.

Epigenetic analyses

Given that our editing strategy was directed at disrupting sites of open chromatin, to understand why the transplants did not yield the anticipated effect, we performed ATAC seq profiling of CD34+ cells after generating KLF1 indels derived from each test condition. We harvested cells for analysis at d11, which is the end of the expansion phase 2, to maximize cell number and minimize enucleated cells that arise during phase 3 of the protocol (d11-d18). The KLF1 transcription unit resides within a gene-dense region of chromosome 19^31,32, and consistent with published studies that mapped HS regions along with intrachromosomal interactions across a ~600 kB region, we find there are numerous open chromatin domains in erythroid cells following the 18-day differentiation protocol (Fig. 4a; scr control).

Fig. 4. ATAC seq analysis.

a ATAC-seq results across a 610 kB region surrounding the KLF1 gene comparing results from scrambled (scr), INT1, EHS1, or INT1/EHS1 dual, staggered treatment of CD34+ cells. Expanded views surrounding TNPO2, RAD23A, and NFIX, which are sites that interact with KLF1^30–32, are shown, highlighting altered ATAC sites in each. The expanded view of the KLF1 gene is in the center and indicates the locations of the INT1 and EHS1 sites relative to the transcription unit. Plots are autoscaled to the same y-axis. Each display is merged from triplicate samples. b Percent indels of each sample in (a) are indicated. c The effect of the INT1/EHS1 dual, staggered treatment on KLF1 expression. n = 9 independent samples. d The total number of differential ATAC peaks (sum of increased and decreased) from individual comparisons between scrambled control (scr) and INT1, EHS1, or INT1/EHS1 were tabulated from the data in (a). Differential peaks with a log2 fold change above 0.5 and below −0.5 with a q-value less than 0.01 were considered for this analysis. e The percent distribution of ATAC peaks across genomic regions was tabulated from the data in (a); TSS transcription start site, TTS transcription termination site.

Inspection of this large region after generation of indels at EHS1 or INT1 shows that there are a small but significant number of chromosomal configuration changes (of both increased and decreased ATAC access) (Fig. 4a; INT1 or EHS1) that include gene promotors known to interact with the KLF1 gene region in mouse (Tnpo2, Rad23a, Nfix; assessed via micro capture C³⁰) and human (RAD23A and NFIX; assessed via HiChIP³¹) studies. Within the KLF1 genetic unit (Fig. 4a, bottom), chromatin accessibility is decreased overall after generation of indels at either INT1 or EHS1; however, ATAC access at each indel target site (INT1 or EHS1) is not more dramatically disrupted by either treatment.

To obtain the maximal effect of simultaneous gRNA directed at both sites, we staggered the introduction of Cas12a and Cas9 gRNAs to CD34+ cells, directed at INT1 and EHS1 respectively, by 24 h; this enabled us to introduce the full amount of each at each transfection rather than being limited to a 50:50 mix. The level of indels at EHS1 and INT1 are higher than we have usually seen under the previous co-introduced dual target conditions, enabling us to achieve 69% double indel frequency (Fig. 4b). Surprisingly, KLF1 expression levels are not downregulated any lower than before (Fig. 4c), and the ATAC seq data shows that any chromosomal conformation change is even less apparent compared with single site editing (Fig. 4a; INT1/EHS1 (staggered), also noted at the TNPO2, RAD23A, and NFIX loci).

KLF1 interacts with histone modifiers^39–41 and chromatin remodelers^42–44, and thus coordinates transcriptional and epigenetic status at its DNA targets⁷, including activity as a pioneer factor⁴⁵. As an independent (albeit indirect) measure of the effect of changes in KLF1 levels, we evaluated global changes in ATAC seq after individual and dual editing by differential comparisons of control (scr) vs INT1, EHS1, or INT1/EHS1. We find the total number of differential peaks affected by INT1 or EHS1 edits is no different from each other; however, the differential peak number after dual editing is significantly lower (Fig. 4d). Notably, sites at the BCL11A internal enhancer locus, a gene that is regulated by KLF1^18,25, fall into this pattern (Supplementary Fig. S4). In spite of the difference in peak number, the genomic distribution of differential peaks is similar across all three conditions (Fig. 4e).

Collectively, these data demonstrate that the KLF1 locus is remarkably buffered against loss of hypersensitive sites. Chromatin accessibility within the gene, although diminished after introduction of indels, is never fully abolished under any editing condition, likely explaining the variable and limited effect on KLF1 expression.

Discussion

Genome editing of non-coding regulatory elements offers therapeutic promise, yet outcomes remain unpredictable and locus-dependent. In this study, we establish the KLF1 locus as a paradigm for enhancer resilience, revealing that even high-efficiency editing of conserved erythroid enhancers fails to reproduce haploinsufficiency due to intrinsic architectural buffering. Our findings highlight that the three-dimensional genomic architecture surrounding a gene can impose intrinsic constraints that buffer transcriptional output against enhancer perturbation.

Building on prior work in which INT1 editing yielded modest, but variable effects on KLF1 expression²⁸, we targeted an additional upstream erythroid-specific enhancer (EHS1) to test whether combining sites of indels could achieve quantitative down regulation. EHS1 editing produced a decrease in KLF1 and a corresponding increase in γ-globin expression in primary human cells, but these effects did not persist in vivo in xenotransplant models, despite confirmation of engraftment and retention of indels. Rather than reflecting experimental failure, these results point to an underlying architectural robustness that preserves transcriptional homeostasis even after enhancer edits are made.

This contrasts with the well-characterized BCL11A locus, where targeted disruption of a single intronic enhancer consistently downregulated BCL11A expression and thereby increased γ-globin production, forming the basis for current gene editing therapies in beta-hemoglobinopathies^46–49. Two key features distinguish the KLF1 and BCL11A loci. First, BCL11A resides in a gene poor region of chromosome 2, while KLF1 resides in a gene-dense domain on chromosome 19 that makes extensive long-range contacts with multiple promoters and enhancers^30–32. Second, removal of the erythroid enhancer element in BCL11A disrupts its chromatin ATAC accessibility and epigenetic insulation, leading to the spread of repressive histone chromatin marks and loss of intrachromosomal connections⁵⁰, while ATAC-seq profiling in our study shows that the KLF1 locus maintains open chromatin even after the dual INT1/EHS1 edits.

In this context, four conclusions can be made. First, use of the staggered approach leads to efficient dual editing as monitored by genomic analysis at d1. Second, dual-edited cells must no longer be well-represented when harvested at d11 (end of expansion phase 2), as their ATAC seq data (by inspection at the KLF1 genetic locus and by global differential analysis) is not as robust as that seen with individual edits at that time point. Third, the near identical number of differential peaks, as well as their distribution, after individual INT1 or EHS1 edits indicates there are virtually no off-target effects by their gRNAs. Finally, in spite of significant changes to chromatin access after individual edits, effects on γ-globin expression remain variable, and published RNA seq studies show only a circumscribed effect on global expression after the INT1 edit²⁸. These observations converge on those made with haploinsufficient individuals, where extensive variation in HbF levels in people containing the same monoallelic KLF1 mutation, even within the same family pedigree, is thought to be due to variable penetrance and microvariation of KLF1 expression^18,29,51.

Alternate strategies to attenuate KLF1 activity could involve direct targeting of its coding sequence. Many HPFH-associated KLF1 variants are truncating or frameshift mutations^10–13. However, these are all heterozygous, and any successful design would require that only one allele be selectively targeted to avoid a total loss of function⁵². Base editing could introduce specific variants; however, these missense mutations would likely be hypomorphs that might have an unpredictable penetrance/effect on the red cell if present as a homozygote. In this context, an additional complication is that although adenine base editors directed to the KLF1 gene (ABE-gRNA 19-3, -10, and -11) led to increases in HbF levels in the HUDEP2 cell line, these did not achieve any significant effect when tested in primary cells³¹.

Our results underscore a broader principle. The success of enhancer editing is determined not only by sequence conservation or guide design, but also by the structural and epigenetic context of the gene locus. This likely arises from a number of factors. First, it is consistent with studies showing that not all deletions directed at a target genetic loci lead to changes in expression of the associated gene⁵³. In those studies, less than half (43%) resulted in full or major loss of function, while 31% led to no change, showing substantial differences in “robustness to mutation”. Second, the α-globin locus provides an example of a highly resistant structure whose chromatin accessibility and compartments remain intact in spite of deletion or orientation of individual component enhancer elements^54,55. Third, nested two-layer enhancer networks^56,57 provide multiple controlling elements that enable stable expression in a variety of physiological, developmental, and environmental situations, including those that alter chromosomal conformations. In the present case, this may reflect an evolutionary selective pressure for stable expression of KLF1 by production of a redundant multi-element enhancer architecture, since red blood cells are required throughout life. We demonstrate this robustness experimentally by showing that disruption of one element isn’t sufficient to reduce expression.

Finally, enhancers and especially promoters in gene-dense regions make contacts across a long (MB) range, with strong interactions occurring between dispersed CTCF binding sites. Both of these characteristics are observed along a 400 kB region surrounding the KLF1 gene, which makes at least 15 contact points with other promoters and enhancers³⁰. This may partly explain the extensive inter-individual variability in HbF levels observed among individuals containing the same monoallelic KLF1 mutation, even within the same family pedigree^18,22,29,51. Our observations that non-coding indels across both alleles do not recapitulate haploinsufficiency arising from complete loss of one allele further emphasizes the protective redundancy encoded in the KLF1 regulatory environment.

In summary, our results demonstrate that while targeted disruption of erythroid enhancers can quantitatively decrease KLF1 and modestly increase γ-globin in primary human cells, these effects are not robust when tested in vivo, suggesting strong buffering by the KLF1 locus architecture. While our study was not designed to address the underlying buffering mechanism, preserved gene expression may arise from the dense regulatory and three-dimensional chromatin architecture of the region, which conserves accessibility despite efficient editing. Future studies should investigate this mechanism further. Together with clinical observations showing that heterozygous loss-of-function KLF1 mutations variably elevate HbF, these findings indicate that enhancer disruption across both alleles does not phenocopy haploinsufficiency resulting from loss of a single allele.

We propose that enhancer buffering by locus architecture represents a limiting principle for therapeutic genome editing, emphasizing that gene density, enhancer redundancy, and chromatin compartmentalization may critically constrain the outcomes of regulatory element targeting. It is possible to be optimistic about enhancer deletion therapies while cautioning on the variability that such techniques have on different target genes and the missing understanding of the underlying mechanisms that control the buffering effect. In some cases it might be appropriate to explore enhancers, at other times promoters, or splice sites, or the gene itself. There is no simple rule saying that one element is better than another, but recapitulating natural mutations is probably most likely to succeed. Recognition and consideration of these features will be essential to encourage alternate approaches and for designing durable, predictable strategies to modulate KLF1 or other transcriptional regulators for hemoglobinopathies and related disorders.

Methods

Cell protocols

Human CD34+ cells were purchased from the Yale Cooperative Center of Excellence in Hematology. Post-thaw, these cells were >95% CD34+, exhibited low cell death, and low CD235a+. Cells were obtained from different donors. These were differentiated (as individual samples) ex vivo under the three-phase protocol as previously described^28,58, with the exception that IMDM media was from Sigma (#FG0465). Briefly, CD34+ cells were thawed on day 0 into X-VIVO 15 (Lonza, 04–418Q) supplemented with 100 ng/ml human stem cell factor (SCF), 100 ng/ml human thrombopoietin, and 100 ng/ml recombinant human Flt3-ligand. Transfection was performed before the expansion phase I, and cells were allowed to recover for 24 h in X-VIVO media. Genomic DNA was isolated at this point for indel analysis. If a second transfection was used, cells were allowed to recover a second time prior to expansion. Cell morphology after such staggered transfections was excellent (Supplementary Fig. S5). The base erythroid differentiation medium (EDM) consists of IMDM supplemented with 330 μg/ml holo-human transferrin, 10 μg/ml recombinant human insulin, 2 IU/ml heparin, 5% human Octoplas AB plasma (OctaPharma), 3 IU/ml erythropoietin, 1% L-glutamine, and 1% penicillin/streptomycin. During phase I (days 0-7), EDM was further supplemented with 10E–6 M hydrocortisone, 100 ng/ml human SCF, and 5 ng/ml human IL-3. During phase II (days 7–11), EDM was supplemented with 100 ng/ml human SCF alone. During phase III (days 11–18), EDM had no additional supplements. All analyses were performed after harvesting the cells either on day 7 (end of expansion phase 1), day 11 (end of expansion phase 2), or 18 (end of differentiation phase 3).

AsCpf1 (enhanced⁵⁹) or SpCas9 proteins were purchased from IDT. RNPs were freshly formed by mixing 105 pmol of CRISPR protein with 120 pmol of gRNA (IDT) for 15 min at room temperature and kept on ice until use. As before²⁸, two gRNAs were used for indel formation at INT1, mixed at equal mass. When both CRISPRs were used for one experiment, the total amount of protein and gRNAs were kept at the same level, resulting in 50:50 mix of Cpf1:Cas9 and thus a 50:25:25 mix of EHS1:INT1a:INT1b (Figs. 1–3). In Fig. 4, we staggered the introduction of molecules by 24 h, enabling us to introduce the full amount of CRISPR RNAs for INT1 or EHS1. Scrambled gRNA was used as a negative control²⁸.

Transfection of CD34+ cells (100,000 cells) with RNP used the Amaxa 4D Nucleofector with program EO-100 in 20 μl cuvettes using P3 buffer. Biological replicates (typically three or four) were each ultimately analyzed in triplicate. Pools of CD34+ cells were analyzed for indel percentages by ICE⁶⁰ of genomic DNA using PCR products generated from sequencing primer pairs⁶¹. Alternatively, next generation sequencing (MGH DNA Core Facility) was used²⁸. Scrambled gRNA (negative control) consistently gave indel values ≤ 1% at all sites.

Molecular design/analyses

Predicted off-targets for EHS1 gRNA designs used the CRISPOR program⁶², which showed zero or one predicted off-targets even allowing for up to 2 mismatches (Supplementary Table S1). All gRNA sequences were as described²⁸ or listed in Supplementary Table S2.

RNA isolation and RT-qPCR analysis were as previously described^27,63, and normalized to catalase⁶⁴. Sequences of all primers and oligos used for analyses are as described^28,65 or listed in Supplementary Table S2.

Brightfield photography was performed with a Keyence BZ-X810 microscope.

HPLC

Cell pellets were tested for hemoglobin concentrations using Cation Exchange HPLC (CE-HPLC) on the Bio-Rad Variant II instrument running the β-thalassemia short program (Bio-Rad, California). Cell pellets were treated as a short sample and prediluted as recommended by the manufacturer, using 1.5 mL of Wash/Diluent solution added to the cell pellet. The solution was mixed and added to a sample vial adapter for loading on the instrument. HbF, HbA, and HbA2 percentages were calculated directly by the instrument.

Xenotransplant

All animal studies were conducted under a protocol approved by the Institutional Animal Care and Use Committed. Five to six-week-old NBSGW mice (Strain 026622) were purchased from the Jackson Laboratory and housed in a room designated for immunocompromised mice at the fully accredited animal facility at the Icahn School of Medicine at Mount Sinai. Non-irradiated NBSGW mice (6- to 8-week-old, 4-5 per group and transplant) were injected intravenously into the retro-orbital venous sinus with approximately 1 million ex vivo edited CD34+ cells. The edited cells were prepared as described above except that 5 million cells were transfected in 100 µL cuvettes; this was then split into five aliquots for transplants performed within 24 h. Pre-transplant editing efficiency was assessed on day 2 of in vitro culture. Mice were euthanized 16- to 18-weeks post transplantation and bone marrow cells collected in PBS buffer containing 0.02 mM EDTA and 0.5% BSA for flow cytometric analysis of human chimerism and hematopoietic lineage analysis using antibodies (Supplementary Table S2) against human CD45 (pan-hematopoietic; BD Bioscience), CD19 (B cells; Southern Biotech), CD3 (T cells; Invitrogen), CD33 (myeloid cells; Invitrogen), CD15 (monocytes; Life Technologies), CD41 (megakaryocytes/platelets; Invitrogen) and CD235a (erythrocytes; Invitrogen) on an Attune flow cytometer. All antibodies were used at a concentration of 5 µL per million cells. Representative plots are shown in Supplementary Fig S2. Molecular analysis of indel percentage and γ-globin expression in human CD45+ and human CD235a+ cells followed their enrichment using a Miltenyi magnetic bead cell separation system (130-045-801 and 130-00050-501 microbeads).

Omni-ATAC-seq

ATAC libraries were synthesized as previously described⁶⁶ from 100,000 cells from SC, INT1, EHS1, and INT1/EHS1 CD34⁺ cells at Day 11 of the differentiation protocol. Each of these libraries was then subjected to 7–8 cycles of PCR amplification. The libraries were then quantitated and assessed using a bioanalyzer to determine the number of PCR cycles required to get enough libraries for sequencing and then further amplified using NEBNext multiplex oligos for Illumina (NEB #E7335L). The samples were then subjected to double size selection (0.5X, 1.2X) using AMPure XP beads (Beckman Coulter, #A63881). Reads were generated by Azenta (formerly Genewiz) sequencing using the Illumina NovaSeq X Plus platform, 2 × 150 paired-end.

ATAC-seq data mapping, peak calling, and peak processing

Cutadapt⁶⁷ was performed for trimming the paired-end reads and were then aligned to hg38 (bowtie2 version 2.3.2; --local --very-sensitive --no-mixed --no-discordant --phred33 -I 10 -X 2000 -p 8), and the mitochondrial DNA alignments were excluded. Removal of mitochondrial reads and duplicates was performed using samtools⁶⁸ (samtools view -h -b -F 1024) and Picard [“Picard Toolkit.” 2019, Broad Institute, GitHub Repository. https://broadinstitute.github.io/picard/; Broad Institute] (MarkDuplicates INPUT = ./$file/$file”.sorted.rmChrM.bam” QUIET=true OUTPUT = ./$file/$file”.marked.bam” METRICS_FILE = ./$file”.metrics” REMOVE_DUPLICATES=false CREATE_INDEX=true VALIDATION_STRINGENCY = LENIENT). At this point, all the replicates per sample were merged. The read ends were adjusted for Tn5 transposase offsets using alignmentSieve command from deepTools⁶⁹. Peaks were called at q-value less than 0.01 (MACS⁷⁰ version 2.1.1; (callpeak -f BAMPE -g mm -B -q 0.01) on merged replicate and bigwig files created using deepTools (bamCoverage --numberOfProcessors 8 --binSize 10 --normalizeUsing RPGC --effectiveGenomeSize 2913022398--ignoreForNormalization chrX). The scan from each individual track used in Fig. 4 is included in the Supplementary Fig S6.

Differential peak calling

Differential accessibility in ATAC-seq data was performed by following the steps outlined in the Epigenomics workshop 2025 (https://nbis-workshop-epigenomics.readthedocs.io/en/latest/content/tutorials/atac-chip-downstream/PeakDA_tsao2022.fulldata_rtds.12ix2025.html). Peaks called for different conditions were merged accordingly (SC&INT1, SC&EHS1, SC&INT1/EHS1) using bedops --merge and then converted to a saf format using the following script: awk “OFS = “\t” {print $1”.”$2”.”$3, $1, $2, $3, “.”}”. These saf files were then used to generate a counts table (mapped reads in each replicate on merged peaks) using featureCounts (Bioconductor Rsubread package) with the -p -F SAF --fracOverlap 0.2 options. The following R packages were then used to call differential peaks using R: tidyverse, dplyr, kableExtra, ggplot2, wesanderson, GenomicRanges, Hmisc, Biostrings, regioneR, bamsignals, MASS, BSgenome. Hsapiens.UCSC.hg38, edgeR, limma, SummarizedExperiment, EDASeq, cqn. Differential peaks with log2FC above 0.5 and below −0.5 with q-value less than 0.01 were considered for our analysis.

Statistics and reproducibility

Results are shown as mean ± sd, and all p values are from two-tailed, unpaired t-tests. Box plot lines represent 25th, 50th (median), and 75th percentiles; whiskers are 5th and 95th percentiles. The number of replicates is indicated in each Figure legend. With respect to primary cells, biological replicates (typically three or four) were each ultimately analyzed in triplicate. With respect to xenotransplant, 4–5 mice were generated per test, and each was ultimately analyzed in triplicate. As an aid to assess reproducibility and the full range of data, each box plot shows all data points. In addition, individual panels that led to the merged ATAC seq data are shown in the supplement.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

A.P., S.P., and B.M. performed experiments and analysis of data; A.P., C.S., and H.L. performed transplants; S.T.C. carried out analysis; C.S., H.L., R.H., J.A.W., J.A.G., and J.J.B. discussed and designed protocols; J.A.G. and J.J.B. conceptualized and oversaw the study; J.A.G. and J.J.B. provided support; J.J.B. wrote the initial draft, then J.A.W. and all authors contributed to and/or approved the final version.

Peer review

Peer review information

Communications Biology thanks the anonymous reviewers for their contribution to the peer review of this work. Primary Handling Editors: Yuhong Cao and Mengtan Xing.

Funding

S.P. was supported by NIH postdoctoral fellowships TL1-DK136048 and K99-DK140510, and H.L. was supported by K01-DK131401. J.J.B. and J.A.G. were supported by Doris Duke Charitable Foundation awards 2019113 and 20230352, and JJB was supported by PHS grant R01-DK046865.

Data availability

Source data that support the graphs are available within the paper and in the supplement (Supplementary Data 1). Data generated from ATAC-seq is deposited in the Gene Expression Omnibus (GEO) and publicly available as of the date of publication, accession number GSE338680.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s42003-026-10868-8.