More than 50% of the human genome is made up of transposable elements (TEs)1,2,3. Three families of TEs are still active in humans: the autonomous long interspersed nuclear element 1 (LINE-1), the nonautonomous Alu short interspersed element and the composite element SVA (SINE (short interspersed nuclear element)–VNTR (variable number of tandem repeats)–Alu)4,5,6,7,8,9. The mobilization of these elements represents a notable source of genomic variation in the human population and is the underlying cause of some genetic diseases10.

SVAs are a class of hominoid-specific TEs. They are nonautonomous, depending on the LINE-1 machinery for retrotransposition, and consist of a fusion of two TE fragments separated by a VNTR11. On the basis of their evolutionary age, SVAs are divided into different subfamilies (A–F), of which SVA-E and SVA-F are human-specific and make up about half of the approximately 3,800 fixed SVAs annotated in the human genome5,12. In addition to the annotated SVAs, there are thousands of polymorphic SVA alleles in the human population. Current estimates suggest about one new germline SVA insertion in every 60 births13. The individual genetic variation caused by polymorphic SVA insertions is thought to contribute to phenotypic variation in the human population and contribute to, or cause, disease10,14,15. However, SVAs have been notoriously challenging to study because of their highly repetitive nature and little is known about how polymorphic SVA insertions are regulated by the human genome or how they influence phenotypic traits and disease.

SVAs harbor strong gene regulatory sequences that can function as both transcriptional activators and repressors, influencing the expression of genes in the vicinity of their integration site14,16,17,18,19,20,21,22. Notably, SVAs appear to be particularly potent as cis-regulatory elements in the human brain, where they have been linked to enhancer-like activities19,22. In line with this, polymorphic SVAs have been linked to several genetic neurological disorders23,24,25,26,27,28. The most well-characterized of these is X-linked dystonia parkinsonism (XDP), a recessive adult-onset autosomal genetic neurodegenerative disorder29,30,31. XDP is caused by a germline SVA retrotransposition event in intron 32 of TAF1, a gene that encodes TATA box-binding protein-associated factor 1, an essential part of the transcriptional machinery30,32.

It is unclear how the SVA insertion in TAF1 leads to XDP pathology. The presence of the SVA has been linked to alternative splicing and reduced transcript levels of TAF1, which may be a direct consequence of intron retention of the 32nd intron30,31,33. In addition, the XDP SVA contains an unstable hexameric repeat that exhibits polymorphic variation associated with age of onset and has the potential to be expanded in somatic tissues. Such expanded hexameric repeats have been speculated to be the source of toxic transcripts or to induce transcriptional interference of TAF1 (ref. 34).

The example of XDP illustrates the important role of polymorphic SVA insertions in human brain disorders. However, although the SVA insertion is the underlying genetic cause of XDP, the molecular mechanism behind how the SVA interferes with TAF1 expression is unknown and there is still no mechanistic insight into why certain SVA insertions cause brain disorders. For example, there are hundreds of intronic SVA insertions in the human genome that do not cause disease. How is the human brain protected against the strong regulatory impact of SVAs in these cases and what makes the disease-causing SVA insertion in TAF1 unique?

In this study, we demonstrate that the DNA-binding KRAB zinc finger protein (KZFP) ZNF91 has a key role in protecting the human genome against the cis-regulatory impact of SVAs by establishing a dual layer of repressive epigenetic modifications over SVAs in neural cells, including new polymorphic alleles such as the disease-causing XDP SVA. The resulting mini-heterochromatin domains are characterized by the presence of both DNA methylation and H3K9me3. Notably, the presence of ZNF91-mediated heterochromatin on the polymorphic XDP SVA is highly relevant for XDP pathology, as the removal of this heterochromatin domain aggravates the molecular XDP phenotype, resulting in increased intron retention and reduced TAF1 expression. In summary, our results provide unique mechanistic insights into how human polymorphic TE insertions are recognized and how their potential regulatory impact in neural cells is minimized by an innate epigenetic defense system based on a KZFP.


XDP neural progenitor cells (NPCs) to study the epigenetic regulation of SVAs

To investigate the molecular mechanisms controlling SVAs in human neural cells, including the polymorphic XDP SVA, we established an NPC model system using induced pluripotent stem cell (iPS cell) lines derived from three persons with XDP and three control individuals (Fig. 1a and Table 1)35. The XDP SVA carriers presented initially with dystonia at a mean age at onset of 42.6 ± 13.6 years, similar to what has previously been reported for persons with XDP (42.3 ± 8.3 years) (Table 1)30,36. The controls used were unaffected sons of two of the XDP SVA carriers (Table 1). The six iPS cell lines were converted into stable NPC lines (XDP and control NPCs) that could be extensively expanded or differentiated into different neural cell types37. The XDP and control NPCs exhibited NPC morphology and expressed NPC markers such as SOX2 and Nestin, monitored with immunocytochemistry (Fig. 1b and Extended Data Fig. 1a). The expression of NPC markers, as well as the lack of expression of pluripotency markers, was also confirmed by RNA sequencing (RNA-seq; Fig. 1c).

Fig. 1: Characterization of the XDP NPC model system.
figure 1

a, Schematic of the generation of XDP NPCs. b, Bright-field images of control (CNPC1) and XDP (XNPC1) NPCs (top). Immunocytochemistry (bottom) of SOX2 (green) and Nestin (red) in control and XDP NPCs. Scale bars, 200 µm. c, Heatmap of NPC marker gene expression in control (n = 3) and XDP (n = 3) NPCs measured using RNA-seq. d, Schematic of the TAF1 gene locus. The polymorphic XDP SVA is depicted in red. e, PCR analysis of genomic DNA identifying the XDP SVA. f, Left, genome browser tracks showing gene expression of the TAF1 gene and a magnification of intron 32 of TAF1, highlighting the characteristic intron retention in XDP NPCs. The XDP SVA and its direction relative to the TAF1 gene are depicted in red. Right, quantification of TAF1 intron 32 retention in control (n = 12) and XDP (n = 12) NPCs. Bars show the normalized mean expression of the group (adjusted P value (Benjamini–Hochberg correction) as calculated by DESeq2 (Wald test, two-sided)). Error bars show the s.e.m. g, Quantification of TAF1 exon 38 expression in control (n = 12) and XDP (n = 12) NPCs. Bars show the normalized mean expression of the group (adjusted P value (Benjamini–Hochberg correction) as calculated by DESeq2 (Wald test, two-sided)). Error bars show the s.e.m.

Source data

Table 1 Description of XDP and control cell lines used in this study

The presence of the XDP SVA insertion, which is ~2.6 kbp long and located in the antisense direction in intron 32 of the TAF1 gene, was confirmed using PCR (Fig. 1d,e). RNA-seq analysis confirmed that XDP NPCs displayed a characteristic retention of intron 32 of TAF1 (P < 0.001, DESeq2) and lower expression of downstream TAF1 exons (P = 0.025, DESeq2), such as exon 38, when compared to control NPCs (Fig. 1f,g). The retention of intron 32 in XDP NPCs appeared to terminate at the 3′ end of the XDP SVA, although mappability issues with this polymorphic SVA did not allow us to determine exactly where transcription ended (Extended Data Fig. 1b). These observations are similar to those previously described for XDP iPS cell and NPC lines30.

SVAs are covered by H3K9me3 in NPCs

TEs, including SVAs, are associated with heterochromatin in somatic tissues, which correlates with their transcriptional silencing and may impact their regulatory potential38. We chose to characterize the repressive histone mark H3K9me3, which is linked to heterochromatin, in fetal human forebrain tissue, two XDP NPCs and two control NPCs using CUT&RUN analysis (Fig. 2a). The computational analysis of histone marks on SVAs using CUT&RUN data is challenging because of their repetitive nature. This results in a large proportion of ambiguous reads. To avoid false conclusions because of multimapping artifacts, we used a strict unique mapping approach to investigate individual SVA elements (Fig. 2a). With this bioinformatic approach, it is only possible to investigate the epigenetic status of the flanking regions of the SVAs where the unique genomic context allows us to discriminate reads without ambiguity and the epigenetic modification can be traced to unique loci in the human genome. The boundaries of nearly all SVAs (>1 kbp in length) of the different subfamilies (A–F), both in the developing human forebrain and in NPCs, were enriched with H3K9me3 (Fig. 2b,c and Extended Data Fig. 2a). However, the genomic context did not enable us to analyze the XDP SVA with this approach. To resolve this issue, we developed a qPCR-based technique, in combination with CUT&RUN (Fig. 2d; see Methods for more details). This analysis, performed with two different primer pairs, revealed a significant enrichment of H3K9me3 at the border of the XDP SVA in XDP NPCs (Fig. 2d and Extended Data Fig. 2b–d).

Fig. 2: ZNF91 is required for H3K9me3 maintenance at SVAs in NPCs.
figure 2

a, Schematic of the CUT&RUN approach to profile H3K9me3 at SVAs in NPCs and human fetal forebrain tissue. b, Heatmap showing enrichment of H3K9me3 over SVAs in human fetal forebrain tissue. Genomic regions ±10 kbp upstream and downstream of the element are shown. c, Heatmap showing H3K9me3 enrichment in NPCs. The genomic regions spanning ±10 kbp upstream and downstream of the element are displayed. d, Schematic of the CUT&RUN qPCR approach. Bar plots demonstrating the enrichment of H3K9me3 over the XDP SVA in XDP NPCs (n = 4) and the lack of enrichment in control NPCs (n = 4; two-tailed t-test). Bars show the H3K9me3 enrichment in each replicate and error bars represent the s.d. e, Schematic of RNA-seq and snRNA-seq experiments in NPCs and human fetal forebrain tissue. f, RNA-seq tracks of ZNF91 expression in NPCs and fetal forebrain. g, Top, uniform manifold approximation and projection (UMAP) showing characterized cell types. Bottom, UMAP representing ZNF91 expression in different cell types in the fetal brain. h, Schematic of the CRISPRi approach including the lentiviral construct and experimental design. i, RNA-seq tracks (left) and quantification (right) of ZNF91 expression in control CRISPRi and ZNF91 CRISPRi in control (n = 4) and XDP (n = 4) NPCs. Bars show the normalized mean expression of the group (adjusted P value (Benjamini–Hochberg correction) as calculated by DESeq2 (Wald test, two-sided)). Error bars show the s.e.m. j, Heatmap showing H3K9me3 over SVAs in control CRISPRi and ZNF91 CRISPRi in control and XDP NPCs. k, Bar graphs showing the effect of ZNF91 CRISPRi on H3K9me3 over the XDP SVA in XDP and control NPCs (n = 4 in each group; two-tailed t-test). Bars show the H3K9me3 enrichment in each replicate and error bars represent the s.d.

Source data

H3K9me3 deposition at SVAs is dependent on the KZFP ZNF91

To protect genomic integrity against TE insertions, organisms have evolved cellular defense mechanisms39,40. Genes encoding KZFPs have amplified and diversified in mammalian species in response to transposon colonization41,42; recent profiling efforts have identified several KZFPs that bind to SVAs, including ZNF91 and ZNF611 (refs. 19,42,43). We noted that ZNF91, in contrast to ZNF611, is highly expressed in human fetal forebrain tissue, as well as XDP and control NPC cultures, as monitored by bulk and single nuclei RNA (snRNA)-seq (Fig. 2e–g and Extended Data Fig. 2e)44. Thus, we hypothesized that ZNF91 could be a KZFP that binds SVAs and recruits the epigenetic machinery that deposits H3K9me3 at these sites in NPCs.

To investigate a role for ZNF91 in SVA repression in NPCs, we designed a lentiviral clustered regularly interspaced short palindromic repeats (CRISPR) inhibition (CRISPRi) strategy to silence ZNF91 expression. We targeted two guide RNAs (gRNAs) to a genomic region located next to the ZNF91 transcription start site (TSS) and coexpressed gRNAs with a KRAB transcriptional repressor domain fused to catalytically dead Cas9 (dCas9) (Fig. 2h). As a control, we used a gRNA targeting lacZ, representing a sequence not found in the human genome. The transduction of XDP NPCs and control NPCs resulted in efficient silencing of ZNF91 expression, monitored with RNA-seq (Fig. 2i and Extended Data Fig. 2f). CUT&RUN analysis of ZNF91 CRISPRi NPCs (XDP and control NPCs) revealed almost complete loss of H3K9me3 around SVAs (Fig. 2j). This finding was reproduced in one additional XDP NPC line and one additional control NPC line (Extended Data Fig. 2g). CUT&RUN qPCR confirmed that the XDP SVA also lost H3K9me3 in a ZNF91-dependent manner in XDP NPCs (Fig. 2k and Extended Data Fig. 2h). Using a similar CRISPRi strategy, we also confirmed that the H3K9me3 at SVAs, including the XDP SVA, also depends on TRIM28, an epigenetic corepressor protein that is essential for the repressive action of KZFPs, in control and XDP NPCs (Extended Data Fig. 2i–k)41,45. Together, these results demonstrate that a ZNF91–TRIM28-dependent mechanism establishes local H3K9me3 heterochromatin over SVAs in human NPCs, including the polymorphic disease-causing XDP SVA.

SVAs are covered by DNA methylation in human NPCs

In addition to H3K9me3, TE silencing in somatic tissues has been extensively linked to DNA CpG methylation40,46,47,48,49. To investigate the presence of DNA methylation on SVAs in NPCs, we performed genome-wide methylation profiling using Oxford Nanopore Technologies (ONT) long-read sequencing (Fig. 3a)50,51 on one XDP NPC line (XNPC1) and one control NPC line (CNPC1). The long-read DNA methylation analysis revealed that the SVA elements of different subfamilies (A–F), which are GC-rich sequences, were all heavily methylated in human NPCs (Fig. 3b). In addition, the polymorphic XDP SVA was fully covered by DNA methylation (Fig. 3c). Furthermore, we performed Cas9-targeted ONT sequencing over the XDP SVA on the ZNF91 CRISPRi and control CRISPRi XDP NPCs (Fig. 3d). These results demonstrated that the XDP SVA was fully methylated in both the ZNF91 CRISPRi and the control CRISPRi XDP NPCs (Fig. 3e). Thus, SVAs in human NPCs, including the XDP SVA, are covered by both DNA methylation and H3K9me3. Our results also indicate that the presence of DNA methylation at SVAs is not dependent on ZNF91 binding or H3K9me3 in this cell type. Notably, the DNA methylation at the XDP SVA did not appear to spread to the surrounding genome, including the TAF1 TSS, which was located 74.2 kbp upstream of the insertion site (Extended Data Fig. 3a).

Fig. 3: SVAs are covered by DNA methylation in NPCs.
figure 3

a, Schematic of ONT sequencing experiment to monitor DNA methylation over SVAs. b, Methylation coverage over SVAs in control and XDP NPCs. The different SVA families (A–F) are shown. Box plot centers correspond to the median, hinges correspond to the first or third quartile and whiskers stretch from the first or third quartile + 1.5 interquertile range (IQR; n = 1). c, Methylation coverage over the XDP SVA in control and XDP NPCs. d, Schematic of Cas9-targeted ONT sequencing. e, Targeted ONT sequencing in control CRISPRi and ZNF91 CRISPRi NPCs. The TAF1 XDP SVA locus is shown. f, Schematic of DNA methylation patterns during development in iPS cells and NPCs. ZNF91 CRISPRi in iPS cells and their conversion to NPCs are also shown. g, Violin plot showing DNA methylation over the first quarter (from their TSS) of the differentially expressed SVAs (P value as calculated by Student’s t-test (two-sided)). Box plot centers correspond to the median, hinges correspond to the first or third quartile and whiskers stretch from the first or third quartile + 1.5 IQR (wild-type iPS cell n = 1; wild-type NPC, n = 2; iPS cell and iPS cell to NPC ZNF91 KD, n = 1). h, DNA methylation pattern over an SVA element near the HORMAD1 gene.

ZNF91 establishes DNA methylation during early development

Because the presence of DNA methylation on SVAs in NPCs did not depend on ZNF91, we wondered how and when DNA methylation on SVAs is established. DNA methylation is reprogrammed during the first few days of early development where global DNA methylation patterns, including those of many TEs, are erased and reinstated52,53,54,55,56. During this process, TEs are initially silenced by dynamic epigenetic mechanisms, which are then gradually replaced by other57, more stable epigenetic mechanisms in somatic cell types, such as NPCs45,58,59,60,61. We and others have implicated TRIM28–KZFP complexes in this process45,59,61,62,63,64,65. Thus, we hypothesized that ZNF91–TRIM28 may be involved in dynamically establishing DNA methylation of SVAs during earlier phases of human embryonic development (Fig. 3f).

To test this hypothesis, we used iPS cells that resemble the epiblast stage of early human development66, where DNA methylation patterns are more dynamically regulated (Fig. 3f). In contrast, NPCs are somatic cells with a stably methylated genome49,56,67. We found clear evidence of dynamic ZNF91-mediated DNA methylation patterning of SVAs when we generated ZNF91 CRISPRi iPS cells. By performing genome-wide ONT analysis, we found numerous SVAs (n = 39) where DNA methylation was lost upon inhibition of ZNF91 (Fig. 3g). In contrast, the same SVAs were covered by DNA methylation in control iPS cells and NPCs (Fig. 3g). Notably, when we differentiated the ZNF91 CRISPRi iPS cells into NPCs, we found that these SVAs remained hypomethylated (Fig. 3g). Thus, without ZNF91 expression, DNA methylation could not be established on these SVAs upon differentiation. For example, an SVA-F element located upstream of the HORMAD1 gene was fully methylated in control NPCs and iPS cells. The DNA methylation over this SVA was completely lost in ZNF91 CRISPRi iPS cells and remained absent when the ZNF91 CRISPRi iPS cells were differentiated into NPCs (Fig. 3h). These results demonstrate that the DNA methylation patterns over some SVAs are dynamic in iPS cells and depend on ZNF91. In addition, ZNF91 is essential for establishing the stable layer of DNA methylation found over these SVAs in NPCs. Thus, cellular context is important for the downstream consequence of ZNF91 binding to SVAs. In early development, ZNF91 mediates the establishment of both H3K9me3 and DNA methylation. On the other hand, only H3K9me3 depends on ZNF91 in somatic cells, whereas DNA methylation is propagated through other mechanisms.

DNA methylation and H3K9me3 cooperate to silence SVAs

To investigate the role of H3K9me3 and DNA methylation in the transcriptional silencing of SVAs in NPCs, we combined loss-of-function experiments with RNA-seq analysis. To remove H3K9me3, we used the ZNF91 CRISPRi NPCs. To remove DNA methylation, we deleted DNA methyltransferase 1 (DNMT1), which is the gene encoding the enzyme that maintains DNA methylation during cell division68. We used a previously described CRISPR-cut approach, resulting in a global loss of DNA methylation, including over SVAs49,69, as well as a CRISPRi combination strategy targeting the expression of both DNMT1 and ZNF91 in XDP and control NPCs (Fig. 4a). Both approaches resulted in a global loss of DNA methylation, as monitored with 5-methylcytosine (m5C) immunocytochemistry (Fig. 4b and Extended Data Fig. 3b), and a loss of DNA methylation over the XDP SVA, as demonstrated by targeted ONT long-read sequencing methylation analysis (Fig. 4c). CUT&RUN analysis on NPCs with knockout (KO) of DNMT1 revealed that loss of DNA methylation did not affect the presence of H3K9me3 at SVAs (Extended Data Fig. 3c).

Fig. 4: DNA methylation and H3K9me3 cooperate to silence SVAs in NPCs.
figure 4

a, Schematic of CRISPR-cut and double CRISPRi experiment in NPCs. b, m5C immunostaining showing the global loss of DNA methylation upon DNMT1 CRISPRi. Scale bars, 100 µm. c, DNA methylation coverage over the XDP SVA in control CRISPR and DNMT1 CRISPR-cut conditions. d, Left, heatmap) showing upregulated SVAs in ZNF91DNMT1 double CRISPRi. The same SVAs are also shown in ZNF91 CRISPRi and DNMT1 KO experiments. Right, box plot showing SVA expression in ZNF91 CRISPRi, DNMT1 KO and ZNF91DNMT1 double CRISPRi. Box plot centers correspond to the median, hinges correspond to the first or third quartile and whiskers stretch from the first or third quartile + 1.5 IQR; outliers are indicated by points (n = 4, except for ZNF91DNMT1 double CRISPRi, where n = 3). e, Heatmap showing H3K4me3 enrichment in control NPCs and NPCs with double KD of ZNF91 and DNMT1. Genomic regions spanning ±10 kbp upstream and downstream of the element are shown. f, Left, heatmap showing the expression level of differentially methylated SVAs. Right, box plot showing the expression of differentially expressed SVAs. Box plot centers correspond to the median, hinges correspond to the first or third quartile and whiskers stretch from the first or third quartile + 1.5 IQR; outliers are indicated by points (n = 4, except for ZNF91DNMT1 double CRISPRi, where n = 3). g, Genome browser tracks showing gene expression (top left), H3K4me3 (bottom left) and DNA methylation (right) pattern over an SVA element near to the RNF24 gene.

We used an in-house 2 × 150-bp poly(A)-enriched stranded library preparation for bulk RNA-seq using a reduced fragmentation step to optimize the read length for SVA analysis. Such reads can be uniquely assigned to many SVA loci. We obtained 40 million reads per sample. To quantify SVA expression, we discarded all ambiguously mapping reads and only quantified those that mapped uniquely to a single location (unique mapping)70. We found that ZNF91 CRISPRi in NPCs, which removed H3K9me3 at SVAs, did not result in activation of SVA expression (Fig. 4d). When DNMT1 was deleted in NPCs, which removed DNA methylation, we also found only a small number of SVAs transcriptionally upregulated (Fig. 4d). However, in the ZNF91DNMT1 double CRISPRi NPCs, where both H3K9me3 and DNA methylation over SVAs were lost, we found a massive transcriptional activation of hundreds of SVAs. To confirm that the SVAs were transcriptionally activated, we performed CUT&RUN analysis for the histone mark H3K4me3, which is associated with active promoters. As the signal of this histone modification spreads to the unique flanking genomic context, this approach allows for an accurate identification of transcriptionally active individual transposon loci44. This analysis confirmed that most full-length SVAs gained H3K4me3 upon ZNF91DNMT1 double CRISPRi (Fig. 4d,e and Extended Data Fig. 3d).

We also analyzed the expression of SVAs in the iPS cell–NPC conversion experiments (Fig. 3f). RNA-seq revealed that the SVAs that lost DNA methylation after ZNF91 deletion in iPS cells (ZNF91 CRISPRi iPS cells; Fig. 3g) were also transcriptionally upregulated (Fig. 4f). This contrasted with ZNF91 CRISPRi NPCs, where the same SVA elements were not upregulated upon inhibition of ZNF91 (Fig. 4f). When we analyzed the ZNF91 CRISPRi iPS cells that were differentiated to NPCs, we found that the SVAs were expressed in these NPCs (Fig. 4f). These SVAs were also found to be upregulated upon ZNF91DNMT1 double CRISPRi in NPCs (Fig. 4f). One example was an SVA-E element located upstream of the RNF24 gene (Fig. 4g). This SVA was transcriptionally silent in control iPS cells and NPCs. In ZNF91 CRISPRi iPS cells, we detected a robust activation of the expression of this SVA-E element, also shown by the presence of the H3K4me3 peak, which correlated with the loss of DNA methylation. When the ZNF91 CRISPRi iPS cells were differentiated to NPCs, the SVA remained expressed; this also correlated with a lack of DNA methylation. Thus, the loss of DNA methylation patterns over SVAs in iPS cells upon ZNF91 CRISPRi correlated with the transcriptional activation of SVAs, including when these cells were differentiated to NPCs. These experiments demonstrate that ZNF91 dynamically represses the expression of at least some SVAs in iPS cells and is essential for establishing stable transcriptional repression of these SVAs.

DNA methylation and H3K9me3 restrict SVA cis-regulation

SVAs carry regulatory sequences that can mediate cis-acting transcriptional effects on the surrounding genome16,17,18,19,20,21. We, therefore, investigated whether the ZNF91-mediated heterochromatin domains found over SVAs in NPCs influenced this activity. When investigating transcriptional changes of genes monitored by RNA-seq upon removal of H3K9me3 (ZNF91 CRISPRi), removal of DNA methylation (DNMT1 KO) or removal of both repressive marks (ZNF91DNMT1 double CRISPRi), we only found profound effects on nearby gene expression in the ZNF91DNMT1 double CRISPRi NPCs. The expression of genes located in the vicinity of an SVA element were significantly increased upon ZNF91DNMT1 double CRISPRi but not when deleting only one of the factors (Fig. 5a). This effect could be detected when the SVA was located up to 50 kbp from the TSS but was stronger when the SVA was closer to the TSS (Fig. 5a).

Fig. 5: SVAs have a regulatory influence on nearby genes when heterochromatin marks are lost.
figure 5

a, Violin plot showing the effect of SVAs on nearby gene expression (2–50 kbp) in ZNF91 CRISPRi, DNMT1 KO and ZNF91DNMT1 double CRISPRi. P value was calculated by a one-way analysis of variance (two-sided). Box plot centers correspond to the median, hinges correspond to the first or third quartile and whiskers extend until the minimum and maximum values; outliers and data points outside of the plot’s axes are indicated by crosses (ZNF91 CRISPRi, n = 6; DNMT1 CRISPR-cut, n = 2; ZNF91DNMT1 double CRISPRi, n = 2). b, Left, genome browser tracks showing HORMAD1 expression and H3K4me3 in ZNF91 CRISPRi, DNMT1 KO and ZNF91DNMT1 double CRISPRi. Right, bar plots showing normalized mean expression of HORMAD1 in ZNF91 CRISPRi, DNMT1 KO and ZNF91DNMT1 double CRISPRi (n = 4). The adjusted P value (Benjamini–Hochberg correction) was calculated by DESeq2 (Wald test, two-sided). Error bars show the s.e.m. c, Left, genome browser tracks showing TAF1 intron 32 expression in ZNF91 CRISPRi, DNMT1 KO and ZNF91DNMT1 double CRISPRi. Right, bar plots showing normalized mean expression of TAF1 intron 32 in ZNF91 CRISPRi, DNMT1 KO and ZNF91DNMT1 double CRISPRi (n = 4). The XDP SVA is depicted in red. The adjusted P value (Benjamini–Hochberg correction) was calculated by DESeq2 (Wald test, two-sided). Error bars show the s.e.m. d, Bar plots showing normalized mean expression of TAF1 exon 38 in ZNF91 CRISPRi, DNMT1 CRISPR-cut and ZNF91DNMT1 double CRISPRi (n = 4). The adjusted P value (Benjamini–Hochberg correction) was calculated by DESeq2 (Wald test, two-sided). Error bars show the s.e.m.

Notably, the dynamics of the SVA-mediated influence on gene expression was distinct between different loci. Most genes in the vicinity of an SVA were completely unaffected by ZNF91 deletion or DNMT1 deletion alone but transcriptionally upregulated when both factors were removed (Fig. 5a). Thus, in most instances, the presence of one of the heterochromatin marks was sufficient to protect flanking genomic regions from the regulatory impact of SVAs. However, we also found examples where both marks were needed to block the regulatory impact of SVAs. For example, the expression of HORMAD1 was upregulated because of the activation of an upstream SVA-F element acting as an alternative promoter in both ZNF91 CRISPRi and DNMT1 KO NPCs (Fig. 5b). When both ZNF91 and DNMT1 were inhibited, HORMAD1 expression was even more strongly activated, suggesting a cooperative mode of action (Fig. 5b). This demonstrates that both epigenetic marks are necessary at some loci to block the regulatory impact of SVAs.

TAF1 XDP phenotype exacerbated by loss of heterochromatin

We next used the XDP NPCs to investigate whether the presence of H3K9me3 and DNA methylation over the XDP SVA has any impact on TAF1 expression. Removing H3K9me3 alone (ZNF91 CRISPRi) affected neither intron retention in the TAF1 loci in the XDP NPCs nor exon 38 expression of the TAF1 gene (Fig. 5c,d). When we investigated the TAF1 loci in XDP NPCs that lacked DNA methylation (DNMT1 KO), retention of intron 32 was significantly increased and exon 38 expression was reduced (Fig. 5c,d). Removing both DNA methylation and H3K9me3 (ZNF91DNMT1 double CRISPRi) had an even stronger effect on TAF1 expression, including a considerable increase in intron 32 retention of TAF1 and lower expression of exon 38 in XDP NPCs (Fig. 5c,d). We also observed the appearance of antisense transcripts originating from the XDP SVA in the ZNF91DNMT1 double CRISPRi NPCs (Extended Data Fig. 1b). This suggests that the XDP SVA becomes transcriptionally active in ZNF91DNMT1 double CRISPRi NPCs, a finding that was corroborated by the presence of H3K4me3 at the boundaries of the XDP SVA in NPCs, as monitored by our qPCR CUT&RUN approach (Extended Data Fig. 3e). Thus, the loss of both DNA methylation and H3K9me3 exacerbated molecular pathology in XDP NPCs. These data demonstrate that the regulatory impact of the polymorphic XDP SVA is negatively influenced by the presence of a local mini-heterochromatin domain. When this heterochromatin domain is lost, the cis-regulatory effect of the XDP SVA is strongly and significantly enhanced.

Polymorphic SVA insertions are silenced by ZNF91

To investigate whether the local heterochromatin observed over the polymorphic XDP SVA represented a unique event or was a general effect, we extended our analysis to other polymorphic SVAs in the genomes of two of the individuals in this study. We took advantage of the whole-genome ONT long-read sequencing data from the XNPC1 and CNPC1 lines and used the transposons from long DNA reads (TLDR) pipeline to identify non-reference SVA insertions (Fig. 6a)51. We identified 12 high-confidence polymorphic insertions of the SVA-E and SVA-F subfamilies (average length of 1.3 kbp), eight of which were shared between the two genomes (Fig. 6b and Supplementary Table 1). Notably, several of these polymorphic SVAs, which represent recent TE insertions into the germline of these two individuals, displayed clear hallmarks of local heterochromatinization, including the presence of DNA methylation and H3K9me3 (Supplementary Table 1).

Fig. 6: Polymorphic SVA insertions are repressed by DNA methylation and H3K9me3.
figure 6

a, Schematic of ONT sequencing and annotation of polymorphic SVAs. b, Venn diagram showing annotated polymorphic SVA insertions. c, Left, genome browser tracks showing gene expression, H3K9me3 and H3K4me3 over a polymorphic SVA insertion and the nearby gene SLC12A6. Right, ONT sequencing data showing DNA methylation over the annotated polymorphic insertion. d, Left, genome browser tracks showing gene expression, H3K9me3 and H3K4me3 over a polymorphic SVA insertion and the nearby gene GABPA. Right, ONT sequencing data showing DNA methylation over the annotated polymorphic insertion.

For example, we found a polymorphic SVA insertion present only in XNPC1 in an intron of SLC12A6, which is a gene encoding a potassium–chloride transporter linked to neurological disorders71,72. This SVA insertion site displayed H3K9me3 at its boundaries and was fully covered by DNA methylation (Fig. 6c). ZNF91 CRISPRi led to loss of H3K9me3 over this SVA, while ZNF91DNMT1 double CRISPRi in XNPC1 led to its transcriptional activation, resulting in the expression of an antisense readthrough transcript extending into the SLC12A6 gene (Fig. 6c). Another example was a polymorphic SVA insertion shared between CNPC1 and XNPC1 downstream of GABPA, which is a gene encoding a DNA-binding protein involved in mitochondrial function73,74. Similarly, the SVA insertion resulted in the accumulation of H3K9me3 and DNA methylation; ZNF91 CRISPRi and ZNF91DNMT1 double CRISPRi resulted in the loss of H3K9me3 and transcriptional activation of the SVA, respectively, generating a readthrough antisense transcript (Fig. 6d).

These data confirm that recent, polymorphic germline SVA insertions are recognized by ZNF91 in human NPCs and are covered by a dual layer of repressive epigenetic marks. Loss of this local heterochromatin domain results in transcriptional activation of these elements and the formation of novel transcripts, which are likely to have a regulatory impact on nearby genes. In these cases, this was illustrated by the production of polymorphic antisense transcripts to SLC12A6 and GABPA.


Our data support a model in which the KZNF ZNF91 binds to SVAs in early human development and throughout brain development, modeled herein using iPS cells and NPCs. ZNF91 binding results in the establishment of local heterochromatin over SVAs, characterized by DNA methylation and H3K9me3, in a TRIM28-dependent manner. In early development (represented herein by iPS cells), both modifications are dependent on the binding of ZNF91 to SVAs. In later phases of development, after epigenetic reprogramming of the genome (represented herein by NPCs), only H3K9me3 depends on ZNF91, whereas DNA methylation over SVAs is propagated by DNMT1. These two repressive chromatin modifications work together to limit the influence of SVAs on the host genome. They prevent expression of the SVA elements and restrict the cis-acting influence of SVAs on the surrounding regions in neural cells. It is worth noting that this mechanism not only involves evolutionarily older SVA insertions that are fixed in the human population but also includes recent polymorphic germline SVA insertions, including the disease-causing XDP SVA. There are limitations to these conclusions. Although the iPS cell–NPC model system used in this project recapitulates some aspects of human brain development, it is a simplified model of early brain development. To validate some of the findings in this study, we relied on human fetal tissue. However, such analysis does not provide much mechanistic insight and we cannot exclude the possibility that some of our observations were in vitro artifacts. In addition, the iPS cell–NPC system does not allow us to investigate differences in the epigenetic silencing of SVAs over time and in different cell types. It will be interesting to address such questions using more complex model systems, such as cerebral organoids.

SVAs carry regulatory sequences with the potential to exert strong cis-acting influences on gene regulatory networks16,17,18,19,20,21,23. The ZNF91-mediated mini-heterochromatin domains prevented this cis-acting influence in the NPC model system. Our data demonstrate that, for most SVA loci, only one of the heterochromatin marks was necessary to silence SVA expression and to prevent its regulatory influence on nearby gene expression. However, there are examples where the cooperation of the two mechanisms appeared necessary. The most striking example was the HORMAD1 locus, where an upstream SVA could act as an alternative promoter43. The regulatory effect of this SVA was activated when H3K9me3 and DNA methylation were removed individually, demonstrating the need for both marks to prevent the cis-acting influence of this SVA. Removal of both marks resulted in a massive activation of HORMAD1 expression, indicating that dual removal has a synergistic effect. It is not yet understood why some SVA loci, such as that upstream of HORMAD1, require both mechanisms for their control. However, it is likely that the transcriptional and epigenetic state of the integration sites is important, as well as structural variants within SVAs. For example, it is known that the VNTR region of SVAs is highly variable and has expanded recently in human evolution51,75. It is also evident that ZNF91 heterochromatin domains are not able to prevent the regulatory influence of all new SVA germline insertions. SVA insertions on the sense strand within genes are less abundant than expected by chance, suggesting that such SVA insertions are selected against11,76. In addition, there are a growing number of polymorphic SVA insertions linked to genetic disorders, with XDP being the best-characterized example.

The SVA insertion linked to XDP is in intron 32 of the essential gene TAF1; the molecular phenotype includes intron retention and reduced TAF1 expression30,31. Our data demonstrate that ZNF91 binds to the XDP SVA and establishes a polymorphic mini-heterochromatin domain. The epigenetic status of the XDP SVA is highly relevant for XDP pathology, as this layer of heterochromatin protects against the gene regulatory impact of the SVA. When DNA methylation and H3K9me3 are lost, intron retention is greatly increased, TAF1 expression levels are further reduced and the XDP SVA is transcriptionally activated. Thus, while ZNF91 can limit the impact of the XDP SVA insertion, it cannot entirely remove the regulatory impact over the TAF1 gene. This explains why the XDP SVA causes disease while most other SVA insertions are inert. However, we still do not understand why ZNF91 is unable to fully block the cis-regulatory impact of the XDP SVA.

Our data are limited to cell culture models; the epigenetic status of the XDP SVA in human brain tissue has not been extensively studied. However, long-read ONT data obtained from postmortem brain tissue of one individual suggested that SVAs were highly methylated in the adult brain, although the methylation level appeared to be slightly lower in polymorphic SVA insertions51. In addition, the presence of DNA methylation over the XDP SVA was investigated in postmortem brain tissue from a patient77. This analysis showed that, although the XDP SVA was methylated in brain tissue, the methylation levels in the brain were reduced compared to blood samples from the same individual. It is worth noting that DNA methylation patterns in the human brain change with age78,79,80. It will be interesting to investigate whether DNA methylation over the XDP SVA is stable in the human brain or lost with age. Such phenomena could explain the late-onset phenotype of XDP, where a gradual increase in the loss of TAF1 function ultimately results in cellular dysfunction. Such a scenario would also open new therapeutic possibilities where restoration of DNA methylation on the XDP SVA could block or reverse the pathology. However, further analysis using larger cohorts of postmortem XDP tissue is required to investigate this possibility.

KZFPs have been implicated in an evolutionary arms race with TEs, where expansions and modifications of genes encoding KZFPs limit the activity of newly emerged transposon classes42. This event is followed by mutations in the TEs to avoid repression in an ongoing cycle. One such example is ZNF91, which appeared in the last common ancestor of humans and Old World monkeys and underwent a series of structural changes about 8–12 million years ago that enabled it to bind to SVA elements42,43. However, the presence of the SVA-binding ZNF91 in hominoids has not prevented the expansion of SVAs in their genomes. On the contrary, SVAs are highly active in the human germline, providing a substantial source of genomic variation in the population5,13,14,23,81,82,83. Although ZNF91 is not able to completely prevent new SVA germline insertions, sustained expression during brain development greatly limits the cis-regulatory impact of these insertions. Thus, it appears that, in this case, ZNF91 may facilitate the expansion of SVA insertions by limiting their gene regulatory impact on the human genome. Our data are consistent with a model where KZFPs are not only TE repressors but also facilitators of inert germline transposition events, thereby fueling genome complexity and evolution.

Our results demonstrate how a KZFP prevents the regulatory impact of TEs in human neural cells but it is still not known whether the ZNF91–SVA partnership represents a unique event or whether these results can be extrapolated to other TE families and lineages. In humans, there are three active TE classes: Alu elements, LINE-1s and SVAs. The relationship between LINE-1s and SVAs is of special interest because SVA retrotransposition depends on coexpression of the LINE-1 machinery. We and others previously found that LINE-1s are controlled by DNA methylation in human NPCs and their transcriptional activation is recognized and silenced by the human silencing hub (HUSH) complex by a mechanism that is independent of KZFPs and TRIM28 (ref. 84). Thus, in human brain development, LINE-1 activity appears to be controlled by fundamentally different mechanisms to SVAs. This suggests that the control of TE activity in the human brain is not only multilayered but also highly specialized. Our data are limited to cell models of early development and neural cells, where ZNF91 is particularly highly expressed. We do not know how SVAs are controlled in other human tissues, although high DNA methylation of SVAs has been shown in the heart and liver51. In addition, our data indicate that there are additional KZFPs controlling SVAs in early human development (see one example in Extended Data Fig. 4a and Supplementary Table 1). ZNF91 deletion in iPS cells activated only a fraction of SVAs, whereas all others remained silenced. It is likely that these SVAs contain binding sites for additional KZFPs, such as ZNF611, that cooperate to control SVAs in early development19,43. ZNF91 may then have a unique role in neural tissues. It is the only KZFP that binds SVAs and protects against cis-acting mechanisms from regulatory sequences in SVAs that are highly active in neural cells.

In summary, our results provide unique mechanistic insights into an epigenetic defense system, based on a KZFP, active against the regulatory impact of SVA transposons in the human brain. On one hand, this system protects the genome from any negative impact of SVAs, with SVA insertions resulting in genetic disease only in very rare instances, exemplified herein by XDP. On the other hand, this system has likely contributed to the expansion of SVAs in our genomes, maximizing the potential for TEs to contribute to increased genome complexity and suggesting that SVAs likely had an important role in primate brain evolution.


All iPS cells used in this study were previously described30. All research using patient-derived iPS cells was performed according to national legislation. Before experimental use, all cell lines were confirmed to be free of mycoplasma.

iPS cells

We used iPS cell lines derived from three persons with XDP and three healthy individuals from WiCell (Table 1). iPS cells were maintained on Biolaminin 521-coated (0.7 mg cm2; Biolamina) Nunc multidishes in iPS medium (StemMACS iPS-Brew XF and 0.5% penicillin–streptomycin (Gibco)). Cells were passaged 1:3 every 2–3 days. Briefly, cells were rinsed once with Dulbecco’s PBS (DPBS; Gibco) and dissociated using Accutase (Gibco) at 37 °C for 5 min. Following incubation, Accutase was carefully aspirated from the well and the cells were washed off from the dish using washing medium (9.5 ml of DMEM/F-12 (Gibco) and 0.5 ml of KO serum replacement (Gibco)). The cells were then centrifuged at 400g for 5 min and resuspended in iPS-Brew medium supplemented with 10 mM Y27632 ROCK inhibitor (Miltenyi Biotech) for expansion. The medium was changed daily85.


The NPCs were generated from iPS cells from the three persons with XDP and three unaffected individuals (Table 1). The neural induction was performed as previously described86. The NPCs were cultured in DMEM/F-12 (Thermo Fisher Scientific) supplemented with glutamine (2 mM; Sigma), penicillin–streptomycin (1×; Gibco), N2 supplement (1×; Thermo Fisher Scientific), B27 (0.05×; Thermo Fisher Scientific), epidermal growth factor (EGF) and fibroblast growth factor 2 (FGF2) (both 10 ng ml−1; Thermo Fisher Scientific). Additionally, 10 mM Y27632 ROCK inhibitor (Miltenyi) was used. Cells were grown on Nunc multidishes or in T25 flasks precoated with poly-l-ornithine (15 μg ml−1; Sigma) and laminin (2 μg ml−1; Sigma). Cells were passaged every 2–3 days using TryplLE express enzyme (Gibco) and trypsin inhibitor (Gibco).


First, 24-well Nunc plates were precoated with poly-l-ornithine (15 μg ml−1; Sigma) and laminin (2 μg ml−1; Sigma). Approximately 50,000 cells were plated in the wells and were allowed to expand until they reached 70–80% confluency. At this point, cells were washed three times with DPBS (Gibco), fixed with 4% paraformaldehyde (Merck Millipore) solution for 15 min at room temperature and washed again three times with DPBS. Fixed cells were stored in DPBS at 4 °C for a maximum of 1 month until staining and imaging.

For blocking, cells were incubated for 1 h with 5% normal donkey serum (NDS) in TKBPS (PBS with potassium phosphate (KBPS) with 0.25% Triton X-100; Fisher Scientific). Subsequently, they were incubated overnight at 4 °C with the corresponding primary antibody (m5C from Active Motif, cat. no. 39649, lot 02617020, 1:250; SOX2 from R&D Systems, AF2018, 1:100; Nestin from Abcam, AB176571, 1:100). For a negative control, cells were incubated overnight with TKPBS + 5% NDS. After overnight incubation, cells were washed two times for 5 min in TKBPS, followed by 5 min in TKBPS with NDS. Next, they were incubated at room temperature for 2 h with the secondary antibody (donkey anti-rabbit Alexa fluor 647 from Jackson Lab, 1:200; donkey anti-goat cy3 from Jackson Lab, 1:200) and for 5 min with DAPI (1:1000; Sigma Aldrich) as a nuclear counterstain. This was followed by two 5-min washes with KPBS; then, cells were stored in PBS until imaging.

m5C staining

As described by Jönsson et al.49, cells stained for m5C were pretreated with 0.9% Triton in PBS for 15 min, followed by 2 N HCl for 15 min and then 10 mM Tris-HCl pH 8 for 10 min before incubation with the primary antibody. Cells were imaged using a fluorescence microscope (Leica).


Total RNA was isolated using the RNeasy Mini Kit (Quiagen) with on-column DNAse treatment following the manufacturer’s instructions. The isolated RNA was used for qPCR (see Methods, ‘CUT&RUN qRT–PCR’) and RNA-seq. RNA-seq was performed using four biological replicates. Libraries for RNA-seq were generated using Illumina TruSeq Stranded mRNA library prep kit (poly(A) selection), optimized for long fragments and sequenced on a Novaseq6000 (paired end, ~250 bp), yielding an average of 46 million reads. The reads were mapped to the human reference genome (hg38) using STAR aligner (version 2.7.8a)87 and gene quantification was performed using featureCounts (Subread package version 1.6.3; hg38 GENCODE version 38), setting the flags ‘-p’ and ‘-s 2’ for paired-end and reversely stranded reads, respectively (TruSeq)88.

To quantify TE expression, reads were remapped using STAR aligner and discarded if mapped to more than one location (‘--outFilterMultimapNmax 1’). A maximum of 0.03 mismatches per base were allowed (‘--outFilterMismatchNoverLmax 0.03’). featureCounts (Subread package version 1.6.3) using the hg38 RepeatMasker annotation ‘parsed to filter out low complexity and simple repeats, rRNA, scRNA, snRNA, srpRNA and tRNA’ was used to quantify reads89.

BigWig files for genome browser tracks were generated using bamCoverage (deepTools version 2.5.4), set to ‘-normalizeUsingRPKM’ and ‘-filterRNAstrand’ to split signal between strands. Visualization was performed in the Integrative Genome Browser (IGV)90. Matrices for deepTools heatmaps were generated including only SVAs longer than 1 kbp (grouped by subfamily; individual BED files), using computeMatrix scale-regions setting ‘-regionBodyLength’ to 1 kbp and flanking regions (‘-a’ and ‘-b’) to 10 kbp. Heatmaps were generated using plotHeatmap (version 3.5.1). Profile plots were generated the same way, ungrouping the SVAs (using a single BED file as input) before the matrix computation.

Normalization of counts to visualize the expression of different features on bar plots (genes, TAF1 intron 32 or TAF1 exon 38) was performed as TPM (transcript per million); the length of the feature was used to calculate an approximate TPM value. Statistical tests, however, were performed using DESeq2, which normalizes using the median of ratios91. Intron 32 and exon 38 of the TAF1 gene were added as part of the gene count matrix DESeq2 model correcting for sequencing batch and the individuals’ diagnosis (XDP or control) whenever appropriate.


H3K9me3 CUT&RUN analysis was performed on CNPC1, CNPC2, XNPC1 and XNPC3 in both ZNF91 CRISPRi and TRIM28 CRISPRi, as well as control CRISPRi (lacZ). H3K4me3 CUT&RUN was performed on CNPC1 and XNPC1 in ZNF91DNMT1 double CRISPRi conditions. We followed the protocol described by Skene and Henikoff 92. Briefly, 300,000 cells were washed twice (20 mM HEPES pH 7.5, 150 mM NaCl, 0.5 mM spermidine and 1× Roche cOmplete protease inhibitors) and attached to ten ConA-coated magnetic beads (Bangs Laboratories) that were preactivated in binding buffer (20 mM HEPES pH 7.9, 10 mM KCl, 1 mM CaCl2 and 1 mM MnCl2). Bead-bound cells were resuspended in 50 ml of buffer (20 mM HEPES pH 7.5, 0.15 M NaCl, 0.5 mM spermidine, 1× Roche cOmplete protease inhibitors, 0.02% w/v digitonin and 2 mM EDTA) containing primary antibody (rabbit anti-H3K9me3, Abcam ab8898, RRID: AB_306848; rabbit anti-H3K4me3, Active Motif cat. no. 39159, RRID: AB_2555751; goat anti-rabbit IgG, Abcam ab97047, RRID: AB_10681025) at 1:50 dilution and incubated at 4 °C overnight with gentle shaking. Beads were washed thoroughly with digitonin buffer (20 mM HEPES pH 7.5, 150 mM NaCl, 0.5 mM spermidine, 1× Roche cOmplete protease inhibitors and 0.02% digitonin). After the final wash, pA-MNase (a generous gift from S. Henikoff) was added in digitonin buffer and incubated with the cells at 4 °C for 1 h. Bead-bound cells were washed twice, resuspended in 100 ml of digitonin buffer and chilled to 0–2 °C. Genome cleavage was stimulated by adding 2 mM CaCl2 at 0 °C for 30 min. The reaction was quenched by adding 100 ml of 2× stop buffer (0.35 M NaCl, 20 mM EDTA, 4 mM EGTA, 0.02% digitonin, 50 ng ml−1 glycogen, 50 ng ml−1 Rnase A and 10 fg ml−1 yeast spike-in DNA (a generous gift from S. Henikoff)) and vortexing. After 10-min incubation at 37 °C to release genomic fragments, cells and beads were pelleted by centrifugation (16,000g, 5 min, 4 °C) and fragments from the supernatant were purified. Illumina sequencing libraries were prepared using the Hyperprep kit (KAPA) with unique dual-indexed adaptors (KAPA), pooled and sequenced on a Nextseq500 instrument (Illumina). Paired-end reads (2 × 75) were aligned to the human and yeast genomes (hg38 and R64-1-1, respectively) using Bowtie 2 (‘--local --very-sensitive --local --no-mixed --no-discordant --phred33 -I 10 -X 700’) and converted to BAM files using SAMtools93. Normalized BigWig coverage tracks were made using bamCoverage (deepTools)94, with a scaling factor accounting for the number of reads arising from the spike-in yeast DNA (104 per aligned yeast read number). Tracks were displayed in IGV.

CRISPR approaches


To silence the transcription of ZNF91 and TRIM28, we used dCas9 fused to the transcriptional repressor KRAB95. Single-guide sequences were designed to recognize DNA regions just downstream of the TSS, according to the Genetic Perturbation Platform (GPP) Portal (Broad Institute). ZNF91 sgRNA: GAGTTTCCAGGTCTCGACTT (no protospacer adjacent motif (PAM)). The guides were inserted into a dCas9-KRAB-T2A-GFP lentiviral backbone containing both the guide RNA under the U6 promoter and dCas9–KRAB and GFP under the Ubiquitin C promoter (pLV hU6-sgRNA hUbC-dCas9-KRAB-T2a-GFP, a gift from Charles Gersbach, Addgene plasmid 71237, RRID: Addgene_71237). The guides were inserted into the backbone using annealed oligos and the BsmBI cloning site. Lentiviruses were produced as described below, yielding titers of 108–109 TU per ml, which was determined using qRT–PCR. Control virus with a gRNA sequence absent from the human genome (lacZ) was also produced and used in all experiments. All lentiviral vectors were used with a multiplicity of infection (MOI) of 2.5 unless stated differently. GFP cells were isolated by fluorescence-activated cell sorting (FACS; FACSAria, BD sciences) on day 10 at 10 °C (reanalysis showed >97% purity) and pelleted at 400g for 5 min, snap-frozen on dry ice and stored at −80 °C until RNA isolation. All groups were performed in four biological replicates unless indicated differently. Knockdown (KD) efficiency was validated using RNA-seq.


LV.gRNA.Cas9-GFP vectors were used to target DNMT1 (ref. 69) or lacZ (control) as previously described49. Lentiviral vectors were produced as described previously and had a titer of 108–109 TU per ml, which was determined using qRT–PCR. Human NPCs were transduced with an MOI of 10–15, allowed to expand for 10 days and sorted by FACS as described previously.


We used a double-transduction method to perform a double CRISPRi of DNMT1 and ZNF91. We transduced ZNF91 with the previously mentioned dCas9-KRAB-T2A-GFP lentivirus containing the dCas9 protein and a GFP. At the same time, the cells were transduced with lentivirus containing the pLV.U6BsmBI.EFS-NS.H2b-RFPW lentiviral backbone with the gRNA for DNMT1 and mCherry as a marker but without dCas9 to knock down DNMT1. DNMT1 sgRNA: TGCTGAAGCCTCCGAGATGC (no PAM). Double-positive (mCherry and GFP) cells were sorted by FACS as previously described and stored at −80 °C until RNA extraction.

Lentiviral vector production

Lentiviral vectors were produced according to Zufferey et al.96. Briefly, HEK293T cells were grown to a confluency of 70–90% on the day of transfection for lentiviral production. We used third-generation packaging and envelop vectors (pMDL, psRev and pMD2G), together with polyethyleneimine (PEI; Polysciences PN 23966, in DPBS (Gibco)). The lentivirus was harvested 2 days after transfection. The supernatant was then collected, filtered and centrifuged at 25,000g for 1.5 h at 4 °C. The supernatant was removed from the tubes and the virus was resuspended in PBS and left at 4 °C. The resulting lentivirus was aliquoted and stored at −80 °C.


To identify whether the XDP SVA was surrounded by an H3K9me3 or H3K4me3 mark, we designed a qPCR approach. For the H3K9me3 analysis, we used CNPC1 and XNPC1 control CRISPRi and ZNF91 CRISPRi cells, four replicates each. For the H3K4me3 analysis, we used CNPC1 and XNPC1 control CRISPRi and ZNF91DNMT1 double CRISPRi cells, one replicate each. Briefly, two primer pairs were designed on the 5′ region of the XDP SVA: one in the flanking region and one in the SVA (Fig. 2d). As a positive control, primers were designed for a genomic region (hg38, chromosome 5 (chr5):141253464–141255143) known to be covered by H3K9me3. The CUT&RUN library was used as a template for amplification. The qRT–PCR was performed with SYBR Green I Master (Roche) on a LightCycler 480 (Roche). The primer pairs used were as follows:





Positive control forward (5′–3′): AAATGGGAATTAAAATCAGTGAGG

Positive control reverse (5′–3′): TTGACATATCATTAAGGGGGCA

ONT sequencing

Whole-genome ONT sequencing

DNA was extracted from frozen pellets using the Nanobind HMW DNA Extraction kit (PacBio) following the manufacturer’s instructions. The final product was eluted in 100 μl of elution buffer provided in the kit. DNA concentration and quality were measured using Nanodrop and Qubit from the top, middle and bottom of each tube. Only DNA with a quality of 260/280 1.8–2.0 and 260/230 2.0–2.2 was further processed. Whole-genome sequencing on samples XNPC1 and CNPC1 was performed at the SciLife lab in Uppsala using SQK-LSK109 Ligation Sequencing kit (ONT) and FLO-PRO002 PromethION Flow Cell R9 Version on a PromethION (ONT).

Cas9-targeted ONT sequencing

To target the XDP locus, we used the Cas9 sequencing (SQK-CS9109) kit following the manufacturer’s instructions (ONT). To enrich for the fragment of interest, four previously described gRNAs were used97. Briefly, two guides were designed upstream and two were designed downstream of the XDP SVA insertion. The excision using these guides resulted in a 5.5-kbp product, including the XDP SVA (2.6 kbp). A total of 5 μg of DNA was used. Samples were sequenced on a MinION Mk1Mc using flow cell R9.4.1 (ONT). One flow cell per sample was used. For cas9-targeted enrichment, we obtained 20,770 reads from three samples (XNPC3 control CRISPRi, 7,805 reads; XNPC3 DNMT1 KO, 7,581 reads; XNPC3 ZNF91 CRISPRi, 4,384 reads; SAMtools view ‘-c BAM’). The proportion of reads that mapped to the target was 1.25% on average (control CRISPRi, 52 reads; DNMT1 KO, 171 reads; ZNF91 CRISPRi, 36 reads; primary alignments over TAF1 intron 32 only; SAMtools view ‘-c -F 260 BAM chrX:71424238–71457170’). XNPC3 control CRISPRi, ZNF91 CRISPRi and DNMT1 KO samples were sequenced using the targeted approach.

FASTQ files were indexed using the nanopolish index (version 0.13.3) on default parameters98. To build an index of the XDP genome (with the XDP SVA insertion), a consensus of the SVA sequences (two enrichment methods: PCR and CRISPR) as reported by Reyes et al.97 ( was created using EMBOSS cons (version ( resulting in a 2,638-bp sequence.

A TAF1 FASTA file was generated using grep ‘-w TAF1’ from hg38 GENCODE version 38 and BEDTools getfasta (version 2.30.0)99. The SVA consensus sequence and the TAF1 sequence were then aligned using ClustalW2 (version 2.1). We observed that the breaking point between the sequence extracted by Reyes et al. (2022) and the reference genome’s sequence of TAF1 occurred at nucleotide chrX:71,440,502 (ref. 97).

The FASTA file of chrX was then chopped using BEDTools getfasta using the following breaking points:



The three sequences (chrX:1–71440502, the SVA sequence and chrX:71440503–156040895) were then stitched (concatenated) together. A new genome FASTA file was created by concatenating all hg38 chromosome FASTA files (except for chrX) and the XDP chrX FASTA file. A Minimap2 (version 2.24) index was then created using the ONT preset (‘-x map-ont’)100. Mapping of the reads was performed using Minimap2 (version 2.24) using the ONT preset (‘-a -x map-ont’) with the XDP genome index. BAM files were sorted and indexed using SAMtools (version 1.16.1).

To characterize the XDP SVA hexameric region, reads mapping to the insertion were extracted using samtools view for chrX:71440502–71443153. Consensus sequences of the XDP SVA for each cell line were produced using SAMtools consensus, including insertions and deletions. Hexamer length was defined upon manual curation of their sequence and confirmed using Noise-Cancelling Repeat Finder ( using the samples’ consensus sequence as the input FASTA and CCCTCT as the motif (‘--positionalevents --maxnoise 20% --minlength 0 --stats=events’).

Polymorphic insertions were identified using TLDR (version 1.2.2), using GRCh38.p13 as the reference genome (‘-r’) and a TE library (‘-e’) including the consensus sequences for TE subfamilies: L1Ta, L1preTa, L1PA2, SVA A–F and HERVK (sequences provided by TLDR developers)51. Insertions were required to have an UnmapCover of at least 80% (percentage of insertion with TEsequence), a sequence similarity (TEMatch) of at least 80% to the TE consensus sequence and a minimum of three reads supporting it (span-reads). SVAs were considered for further analysis if their length was greater than 1 kbp.

The local consensus sequences of the polymorphic insertions (as the output from TLDR) were introduced to the reference genome. A custom script ( was used to read the TLDR output table, sort the polymorphic insertions from the end to the start of each chromosome and perform the following operations for each of the chromosomes:

  1. 1.

    Read its FASTA file (chr.fa).

  2. 2.

    Extract its sequence before and after the insertion using BEDTools getfasta (‘-fi chr.fa -bed coordinates.bed’), where coordinates.bed included two coordinates: the first spanning from the beginning of the chromosome to the start of the insertion (as reported by TLDR) and the second spanning from the end of the insertion (as reported by TLDR) to the end of the chromosome.

  3. 3.

    Concatenate the following sequences, before rewriting the chromosome’s FASTA with the output:

    1. a.

      The chromosome before the insertion

    2. b.

      The sequence of the polymorphic insertion

    3. c.

      The chromosome after the insertion

This process was repeated for each polymorphic insertion, introducing them in order from the end to the start of each chromosome. Similarly, an updated gene annotation GTF file (to fit the coordinates including all polymorphic insertions) was created using a custom script following the same logic (

The reads were remapped to the custom genome using an indexed version of the output FASTA (output from using Minimap2 (indexing using ‘-x map-ont’; mapping using ‘-a -x map-ont’).

Methylation for each of the regions of interest was called using Nanopolish call-methylation (version 0.13.3) with the raw reads (‘-r’), the alignment files to the custom genome (‘-b’) and the custom genome’s FASTA file as a reference (‘-g’). Databases for Methylartist were produced using Methylartist db-nanopolish (version 1.2.2), using the methylation call files as input (default parameters). Specific loci were visualized using Methylartist locus (version 1.2.2)50.

Reporting summary

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