HIF1α-AS1 is a DNA:DNA:RNA triplex-forming lncRNA interacting with the HUSH complex

DNA:DNA:RNA triplexes that are formed through Hoogsteen base-pairing of the RNA in the major groove of the DNA duplex have been observed in vitro, but the extent to which these interactions occur in cells and how they impact cellular functions remains elusive. Using a combination of bioinformatic techniques, RNA/DNA pulldown and biophysical studies, we set out to identify functionally important DNA:DNA:RNA triplex-forming long non-coding RNAs (lncRNA) in human endothelial cells. The lncRNA HIF1α-AS1 was retrieved as a top hit. Endogenous HIF1α-AS1 reduces the expression of numerous genes, including EPH Receptor A2 and Adrenomedullin through DNA:DNA:RNA triplex formation by acting as an adapter for the repressive human silencing hub complex (HUSH). Moreover, the oxygen-sensitive HIF1α-AS1 is down-regulated in pulmonary hypertension and loss-of-function approaches not only result in gene de-repression but also enhance angiogenic capacity. As exemplified here with HIF1α-AS1, DNA:DNA:RNA triplex formation is a functionally important mechanism of trans-acting gene expression control.

RNA strand binds in a parallel or antiparallel manner. Hoogsteen bonds are weaker than Watson-Crick bonds, resulting in Hoogsteen pairing rules being more flexible 3 .
Ex vivo characterization of triplex formation relies on a variety of different biophysical methods including circular dichroism-(CD) and nuclear magnetic resonance-spectroscopy (NMR) [4][5][6] . Even with these techniques it can be challenging to discriminate DNA-RNA heteroduplexes from triplexes and analyses are usually restricted to oligonucleotides of a limited length. Nevertheless, a few lncRNAs have been suggested to form triplexes with dsDNA, however, triplex studies using living cells are still in early development 4,[6][7][8][9][10][11][12][13] . In silico analyses of RNA-DNA triplex formation predicted several genomic loci and lncRNAs to form triplexes 14 . In line with this, a global approach in HeLa S3 and U2OS cells to isolate triplex-forming RNAs on a genome-wide scale yielded several RNA:DNA triplex-forming lncRNAs 15 .
In addition to the sparse initial findings of triplex formation within cells, several other open questions remain: What is the physiological relevance of triplex-forming lncRNAs and are these cell-and tissuetype specific? What is the mechanism of action of triplex-forming lncRNAs? Do they disturb transcription in a similar way to R-loops 16 or recruit certain protein complexes to DNA in a site-specific manner? Regarding the latter aspect, Polycomb Repressive Complex 2 (PRC2) has been identified as a target of the lncRNAs HOX Transcript Antisense RNA (HOTAIR), FOXF1 Adjacent Non-Coding Developmental Regulatory RNA (FENDRR) and Maternally Expressed 3 (MEG3) 4,12,13 , but, given the highly promiscuous nature of PRC2, this function remains controversial. Other examples of protein interactors involve e.g. E2F1 and p300, which are recruited by the triplex-forming antisense lncRNA KHPS1 to activate gene expression of the proto-oncogene sphingosine kinase 1 (SPHK1) in cis 7,10 .
Much of today's in vitro RNA research heavily relies on immortalized cell lines. Although such model systems are well suited for transfection or genomic manipulation, they are highly dedifferentiated and exhibit reaction patterns such as unlimited growth and immortalization -characteristics not observed in primary cells 17 . Considering that lncRNAs are expressed in a species-, tissue-and differentiation-specific manner 1 , biological evidence for lncRNA functions in primary cells is limited. Among such cells, endothelial cells stand out due to their well documented importance in regeneration, angiogenesis and tissue vascularization. Indeed, endothelial cell dysfunction is one of the main drivers of systemic diseases like diabetes and inflammation 18 .
Here, we combined molecular biology and biophysics, bioinformatics and physiology to systematically uncover the role of triplexforming lncRNAs in endothelial cells. This approach identified HIF1α-AS1 as a trans-acting triplex-forming lncRNA that controls vascular gene expression in endothelial cells with implications for vascular disease.

HIF1α-AS1 is a triplex-associated lncRNA
To identify triplex-associated lncRNAs, we used Triplex-Seq data from U2OS and HeLa S3 cells 15 . Triplex-Seq relies on the isolation of RNase H-resistant RNA-DNA complexes from cells followed by DNA-and RNA-Seq 15 . RNase H cleaves the RNA in DNA-RNA heteroduplexes as present in R-loops 19 and has previously been used to distinguish between heteroduplexes and triplexes 20 . The Triplex-Seq data comprised all RNA entities and was filtered for the number of individual lncRNA genes, resulting in 989 (for HeLa S3, Supplementary Data 1) and 1363 (for U2OS, Supplementary Data 2) different lncRNAs associated with triplexes, with an overlap of 280 lncRNA genes between the two cell lines (Fig. 1a). To further narrow down this set of enriched triplexassociated lncRNAs, parameters for specificity (fold enrichment >10, -log10(P value peak enrichment)) were increased so that 11 lncRNA candidates with high confidence remained. Subsequently, these were correlated to ENCODE and FANTOM5 Cap Analysis of Gene Expression (CAGE) [21][22][23] data. Of the 11 candidates, only 5 (RMRP, HIF1α-AS1, RP5-857K21.4, SCARNA2 and SNHG8) were expressed in endothelial cells. All 5 candidates were predicted as non-coding by the online tools Coding Potential Assessment Tool (CPAT 3.0.0) and coding potential calculator 2 (CPC2) and at least partially nuclear localized by ENCODE CAGE (Fig. 1a). To further analyze these candidates, the Triplex-Seq enriched regions were manually inspected in the IGV browser. This led to the exclusion of SNHG8 as the triplex-associated regions within this lncRNA were exclusively within the overlapping small nucleolar RNA 24 (SNORA24) gene. In the case of the other candidates, triplexassociation was within the individual lncRNA gene body. The cumulative fold enrichment of the remaining lncRNAs in the Triplex-Seq dataset illustrated strong triplex-association ( Supplementary Fig. 1a). To verify the candidates experimentally, RNA immunoprecipitation (RIP) with antibodies against dsDNA with or without RNase H treatment in human endothelial cells was performed. Cleavage of the RNA in DNA-RNA heteroduplexes by RNase H 19 revealed that HIF1α-AS1 was the strongest triplex-associated lncRNA (Fig. 1b).
As HIF1α is a central regulator of oxygen-dependent gene expression 18 , we decided to measure the expression of HIF1α-AS1 in endothelial cells under altered oxygen and disease conditions. Hypoxia led to a decrease in HIF1α-AS1 expression in endothelial and pulmonary artery smooth muscle cells (paSMC) (Fig. 1f, Supplementary Fig. 1b), which was restored in endothelial cells after 4 h and even surpassed basal levels after 24 h of normoxic conditions (Fig. 1g). Importantly, HIF1α-AS1 was downregulated in endothelial cells isolated from human glioblastoma ( Supplementary Fig. 1c) and in lungs from patients with end stage idiopathic pulmonary arterial hypertension (IPAH) or chronic thromboembolic pulmonary hypertension (CTEPH) (Fig. 1h). In paSMCs isolated from pulmonary arteries of patients with IPAH, HIF1α-AS1 was strongly decreased (Supplementary Fig. 1d). Together, these data demonstrate that HIF1α-AS1 is an oxygendependent and disease-relevant lncRNA.

HIF1α-AS1-triplex binding suppresses target gene expression
Triplex-Seq can provide evidence for existing triplex forming regions of the RNA (TFR) and triplex target sites (TTS) within the DNA but the details of exactly which TFR and TTS interact cannot be derived from Triplex-Seq. To identify the TFRs within HIF1α-AS1 as well as HIF1α-AS1dependent TTS, a combination of bioinformatics and wet lab approaches were used: An Assay for Transposase-Accessible Chromatin with high-throughput sequencing (ATAC-Seq) was performed after HIF1α-AS1 knockdown to identify DNA target sites in human endothelial cells. LNA-GapmeRs targeting HIF1α-AS1 led to a strong knockdown of the lncRNA (Supplementary Fig. 1e). Triplex Domain Finder (TDF), a computational tool for the prediction of RNA and DNA triplexforming potential 14 , predicted the TFRs within HIF1α-AS1 to target DNA regions around genes that displayed altered ATAC-Seq peaks after HIF1α-AS1 silencing (Fig. 2a). The software identified three statistically significant TFRs (TFR1-3) within the pre-processed HIF1α-AS1 RNA ( Fig. 2b). There was also a high incidence of triplex-prone motifs predicted in regions whose chromatin state was altered in the ATAC-Seq data after HIF1α-AS1 knockdown (Fig. 2c, Supplementary . Of these TTS, 38 overlapped within all three TFRs (Fig. 2d). To identify which TFR is most strongly associated with triplexes, RIP with S9.6 antibodies recognizing RNA-DNA association was performed. RNA-DNA associations remaining after RNase H treatment excluded the possibility that these were RNA-DNA heteroduplexes. Of the three HIF1α-AS1 TFRs, TFR2 was identified as the TFR most resistant to RNase H (Fig. 2e). TFR2 is located intronically 478 nucleotides (nt) downstream of Exon1 and was detected by RT-PCR within nuclear isolated RNA with primers covering the first 714 nt (E1-I) of the pre-processed HIF1α-AS1 ( Supplementary Fig. 1f). Triplex-prone motifs in the TFR1-3overlapping target regions yielded more than 20 different associated genes, some of which displayed a high number of DNA binding sites (Fig. 2f). If this binding of the lncRNA is relevant for the individual target gene, then a change in target gene expression would be expected. Importantly, in response to the downregulation of HIF1α-AS1 with LNA-GapmeRs the expression of the following triplex target genes increased: ADM, PLEC, RP11-276H7.2, EPHA2, MIDN and EGR1 (Fig. 2g). Interestingly, as exemplified by the target genes HIF1A, EPHA2 and ADM, the triplex target sites are often located close to the 5ʹ end of the gene. In this region, histone modifications, transcription factor binding and chromatin conformation often have the greatest effect on promoter function and gene expression (Fig. 2h, Supplementary  Fig. 1g). In order to prove that the triplexes also exist in vivo, Chromatin Immunoprecipitation (ChIP) was performed with S9.6 antibodies. After RNase H treatment, the TTS of EPHA2 and ADM were both more resistent to RNase H treatment compared to DNA regions upstream or downstream of both TTS (Fig. 2i).
These data indicate that HIF1α-AS1 contains triplex forming regions and target sites important for the regulation of gene expression.
HIF1α-AS1 TFR2 RNA forms triplexes with EPHA2 and ADM Our analysis identified HIF1α-AS1 TFR2 as the best suited candidate for verification of triplex formation of the lncRNA using biophysical and biochemical techniques. To monitor triplex formation of HIF1α-AS1, EPHA2 was chosen as the target gene due to its abundance of triplex target sites (Figs. 2f, h), its regulatory potential (Fig. 2g) and its importance for vascularization 29 . Triplex domain finder predicted not the complete TFR2 to bind EPHA2 TTS, but rather a core TFR2 sequence that binds the TTS. The formation of DNA:DNA:RNA triplexes between lncRNA HIF1α-AS1 TFR2 and its proposed DNA target site within intron 1 of EPHA2 was characterized by electrophoretic mobility shift assay (EMSA), CD-and solution NMR-spectroscopy. From electrophoretic mobility shift assay (EMSA) the HIF1α-AS1 TFR2 RNA was found to form a low-mobility DNA-RNA complex with the  EPHA2 DNA target sequence (Fig. 3a). We also used CD-spectroscopy to confirm triplex formation of HIF1α-AS1 TFR2 on EPHA2. The CD spectrum indicated typical features for triplex formation, such as a positive small peak at ∼220 nm, two negative peaks at ∼210 nm and ∼240 nm and a blue-shift of the peak at ∼270 nm 30,31 , which was distinct from the EPHA2 DNA duplex or the heteroduplex spectra (Fig. 3b) . 3d). Using 10 eq HIF1α-AS1 TFR2 RNA, triplex 1 H NMR imino signals were observed in a spectral region between 9 and 12 ppm providing further evidence that HIF1α-AS1 was associated with EPHA2 through Hoogsteen base pairing (Fig. 3e, Supplementary Fig. 2a). Further, we conducted NMRspectroscopic analysis of the triplex: we first measured a 1 H, 1 H-NOESY spectrum for EPHA2 DNA duplex and assigned cross peaks in this spectrum of the DNA duplex. We identified 11 G and 12 T imino proton signals (Fig. 3f) semi-quantitatively analyzed the change in the DNA duplex spectrum. For 7 G-and 6 T-imino protons either a strong or medium attenuation of cross peak intensities in the imino-imino region of the NOESY spectrum was observed ( Supplementary Fig. 2b). We rationalize this attenuation as to arise from weakening of the Watson-Crick base pairing induced by the Hoogsteen interaction with the RNA strand. From this analysis, we compared predicted Hoogsteen interactions in the triplex with the detected changes in the NOESY spectrum for different positions i of RNA relative to DNA duplex strand. Interestingly, the previously predicted position (i = 0) is supported by the observed attenuations in the NOESY, where 3 G-and 2 T-imino sites disappear completely and 3 G-and 1 T-imino sites are significantly attenuated. In total, 12 sites remain unaffected in the DNA duplex ( Supplementary Fig. 2c). Further, based on our interpretation that Hoogsteen interactions can be mapped from the analysis of cross peak attenuation in the NOESY, we generated structural models for the c Thermal melting assay of the EPHA2 DNA duplex (black), the heteroduplex (dark gray) and EPHA2:HIF1α-AS1-TFR2 (red). d Sequence of EPHA2 DNA (black) and HIF1α-AS1-TFR2 RNA (red). Watson-Crick base pairing is indicated with | and the Hoogsteen base pairing is indicated with * . Changes in the DNA duplex were quantitatively analyzed using NOESY spectra of duplex and triplex. Imino protons with strong attenuation (dark blue arrows) or medium attenuation (light blue arrows) of cross peak intensities in the imino-imino region were observed. e 1 H-1D NMR spectra of the EPHA2 DNA duplex (black), HIF1α-AS1 TFR2 RNA (blue), heteroduplex (dark gray) and EPHA2:HIF1α-AS1-TFR2 triplex (red) at 288 K.   EPHA2:HIF1α-AS1_TFR2 triplex. The ensemble of the 20 top-ranked structures for the triplex are displayed as cartoon (Fig. 3g).
To confirm the formation of triplexes with lower equivalents, stabilized triplex formation was investigated: the intermolecular dsDNA formed by two complementary antiparallel DNA strands was changed into a hairpin construct, where both DNA strands were linked with a 5 nt thymidine-linker and duplex formation thus became intramolecular. With this approach, triplex formation was obtained with 3 eq RNA, indicating that triplex formation is favored under those conditions as expected. 1 H-1D NMR spectra of hairpin EPHA2_CTGA and 15 N-labeled HIF1α-AS1 TFR2:EPHA2_CTGA triplex indicated changes in the Hoogsteen region (9-12 ppm) and the spectral region of imino (12-14 ppm) and amino signals (7-8.5 ppm) ( Supplementary Fig. 3a, b). In addition to EPHA2, we also tested ADM, a preprohormone involved in endothelial cell function 32 . For ADM_CTGA:HIF1α-AS1 TFR2 triplex, the new imino protons in the Hoogsteen region arose at lower temperatures (Supplementary Fig. 3c and d). For both ADM_CTGA and EPHA2_CTGA triplex constructs the CD spectra showed an increased negative ellipticity at ∼240 nm and positive ellipticity at ∼270 nm ( Supplementary Fig. 3e Supplementary Fig. 3h). The data demonstrate that HIF1α-AS1 TFR2 forms triplexes with EPHA2 and ADM dsDNA under regular and triplex-stabilized conditions upon DNA hairpin formation.

TFR2 represses EPHA2 and ADM gene expression
The current data indicate that HIF1α-AS1 forms triplexes with EPHA2 and ADM, however, the mechanistic and functional consequences of this phenomenon are unclear. To investigate these aspects, gain and loss of function approaches were performed. Increasing the expression of HIF1α-AS1 using a dCas9-VP64 CRISPR activation system (CRISPRa) reduced the expression of EPHA2 and ADM (Fig. 4a). Conversely, downregulation of HIF1α-AS1 with a dCas9-KRAB repression system (CRISPRi) increased the expression of EPHA2 and ADM (Fig. 4b). Consistent with HIF1α-AS1 repressing EPHA2 and ADM gene expression, EPHA2 levels increased after knockdown of HIF1α-AS1 (Figs. 2g, 4c). EPHA2 has a multi-faceted role in angiogenesis 29,33,34 . In HUVEC, knockdown of EPHA2 with siRNAs strongly reduced its RNA and protein expression and inhibited angiogenic sprouting ( Fig. 4d and e, Supplementary Fig. 4a-c). Conversely, a knockdown of HIF1α-AS1 with LNA-GapmeRs increased VEGF-A-and bFGF-mediated angiogenic sprouting ( Fig. 4f, g, Supplementary Fig. 4d), confirming the repressive effect of HIF1α-AS1 on EPHA2. Additionally, CRISPRi targeting HIF1α-AS1 or an siRNA-mediated knockdown of the HIF1α-AS1 pre-RNA targeting the intron region next to the TFR2 were performed. Targeting the intron of HIF1α-AS1 not only decreased the expression of TFR2, but also increased EPHA2 and ADM, whereas HIF1α was not significantly altered ( Supplementary Fig. 4e). As expected, both CRISPRi and siRNA against HIF1α-AS1 intron induced VEGF-A-mediated sprouting ( Fig. 4h and i, Supplementary Fig. 4f, g), whereas CRISPRa and an overexpression of the first 1200 nt of the HIF1α-AS1 gene (containing Exon1, the beginning of the intron including TFR2) had the opposite effect ( Fig. 4j and k, Supplementary Fig. 4h, i). The repressive effect of HIF1α-AS1 on EPHA2 was further confirmed by Western analysis, where siRNA-mediated knockdown of the HIF1α-AS1 pre-RNA increased EPHA2 and overexpression of the first 1200 nt of the HIF1α-AS1 gene decreased EPHA2 protein levels ( Fig. 4l and m). The beneficial effect on sprouting is at least partially based on an anti-apoptotic effect as knockdown of HIF1α-AS1 increased caspase 3&7 activity as measured by a cell-permeant fluorescent probe (SR-DEVD-FMK) that bound to active caspase 3 & 7 ( Supplementary Fig. 4j). To demonstrate directly that TFR2 is responsible for the regulation of EPHA2, we replaced TFR2 by genome editing using a recombinant Cas9-eGFP, a gRNA targeting TFR2 and different single-stranded oligodeoxynucleotides (ssODN) harboring either the published MEG3 TFR 4 or a luciferase control sequence ( Supplementary Fig. 4k). Replacement of the TFR2 with the MEG3 TFR, which served as a positive control for a functional TFR repressing TGFBR1 expression 4 , yielded a reduction in TGFBR1 levels compared to the luciferase control (Fig. 4n). More importantly, the loss of TFR2 consequently led to a loss of HIF1α-AS1 TFR2, an upregulation of EPHA2 and partially of ADM ( Fig. 4o, p, Supplementary Fig. 4l), and also ChIP with anti-S9.6 led to a reduced detection of the TTS of EPHA2 and ADM (Fig. 4q, r). These data demonstrate that TFR2 is functional as a TFR and represses EPHA2 and ADM gene expression.

HIF1α-AS1 binds to and recruits HUSH to triplex targets
To elucidate the mechanism by which HIF1α-AS1 represses gene expression, HIF1α-AS1-associated proteins were studied using RNA pulldown experiments. 3'-biotinylated spliced HIF1α-AS1 lncRNA or 3ʹbiotinylated pcDNA3.1+ negative control were incubated in nuclear extracts from HUVECs and RNA-associated proteins were identified by electrospray ionization mass spectrometry, which retrieved M-phase phosphoprotein 8 (MPP8)-a component of the human silencing hub (HUSH) complex 35 -as top hit ( Fig. 5a-b, Supplementary Data 6). The HUSH-complex is a nuclear machinery consisting of the chromodomaincontaining protein MPP8, TASOR (FAM208A) and PPHLN1 (Periphilin), and was originally thought to mediate gene silencing during viral infection by recruiting the SET Domain Bifurcated Histone Lysine Methyltransferase 1 (SETDB1) which methylates H3K9 35 . The HUSH complex has not yet been studied in vascular cells and an interaction of its core protein MPP8 with lncRNAs has not been reported. To support our finding, RIP revealed that HIF1α-AS1 and its TFR2, but not HIF1A mRNA, interact with MPP8 ( Fig. 5c, Supplementary Fig. 5a-b). Furthermore, HIF1α-AS1 was highly enriched with H3K9me3 (Fig. 5d).
To map the RNA binding region of MPP8 on HIF1α-AS1, we used catRAPID fragments 36 , an algorithm involving division of polypeptide and nucleotide sequences into fragments to estimate the interaction propensity of protein-RNA pairs. This highlighted potential binding regions within Exon1 ( Supplementary Fig. 5c). To substantiate these data experimentally, ex vivo bindings assays were performed between fragments of HIF1α-AS1 and recombinant MPP8 (Fig. 5e) as well as with HIF1α-AS1 and in vitro translated MPP8 mutants, among them the mutation in the chromodomain W80A, a chromodomain deletion, an N-or C-terminal half deletion and a deletion of the Ankyrin repeats (ANK) (Fig. 5f). MPP8 interacted directly with HIF1α-AS1 full length and a HIF1α-AS1 mutant lacking Exon2 (Fig. 5g). In contrast and in accordance with the catRAPID prediction, deletion of Exon1 (nucleotides 26-78 in particular) prevented the interaction (Fig. 5g), indicating that this region of HIF1α-AS1 is critical for the interaction of HIF1α-AS1 with MPP8. On the protein side, RIP of the MPP8 mutants with anti-His antibodies followed by RT-qPCR for HIF1α-AS1 revealed that the interaction of HIF1α-AS1 with MPP8 was strongly reduced by a deletion of the C-terminal half of MPP8, but not by deletion or mutation of the chromodomain (Fig. 5h). Further, the Ankyrin repeats in the C-terminus seem to effect the interaction only to a minor extent (Fig. 5h). Uniprot 37 listed three disordered regions, two of them in the N-terminal half and one in the C-terminal half (Fig. 5f), which could potentially be involved in the interaction.
To demonstrate that HIF1α-AS1 acts through HUSH complex recruitment, we first tested whether parts of this complex exist in endothelial cells. Proximity ligation assays with antibodies against MPP8, dsDNA, H3K9me3 and SETDB1 confirmed the association of MPP8 with dsDNA ( Supplementary Fig. 5d), H3K9me3 (Fig. 5i) and SETDB1 (Fig. 5j) in the nuclei of endothelial cells, indicating that parts of the complex are present at endothelial chromatin.
ChIP with and without RNase A revealed that targeting of MPP8 and SETDB1, but not NP220, which is another protein associated with the HUSH complex 38 and interacting with HIF1α-AS1 (Fig. 5b), to the HIF1α-AS1 TTS of EPHA2 and ADM were attenuated after RNA depletion (Fig. 6a, Supplementary Fig. 6a, b). A region 5.7 kb downstream of EPHA2, which harbors different triplex target sites to the one studied here (Fig. 2h), also appeared to be reduced after RNase treatment the binding of MPP8 and SETDB1, but not NP220, indicating that MPP8 and SETDB1 might also act there ( Supplementary Fig. 6c-e). To demonstrate the dependence of the interactions with the TTS on HIF1α-AS1, ChIP experiments with antibodies targeting SETDB1, MPP8 and NP220   (Fig. 6f, Supplementary Fig. 8). An increase in accessibility to the region downstream of the EPHA2 TTS was detected after knockdown of HIF1α-AS1, SETDB1 and MPP8; however, this could not be validated with CRISPRi/a for HIF1α-AS1 or LentiCRISPR-dependent HIF1α-AS1 TFR2, EPHA2 TTS or ADM TTS experiments, suggesting that the regions within the TSS, but not downstream of the EPHA2 TTS, may contain the most essential repressor regions. These data indicate that triplex formation by HIF1α-AS1 is important for fine-tuning chromatin accessibility locally and thereby gene expression of EPHA2 and ADM through SETDB1 and MPP8.

Discussion
The present study combined molecular biology, bioinformatics, physiology and structural analysis to identify and establish the lncRNA HIF1α-AS1 as a triplex-forming lncRNA in human endothelial cells.
Through trans-acting triplex formation by a specific region within HIF1α-AS1, EPHA2 and ADM DNA target sites are primed for their interaction with the HUSH complex members MPP8 and SETDB1 to mediate gene repression through control of chromatin accessibility. Physiologically, the anti-angiogenic lncRNA HIF1α-AS1 is dysregulated in hypoxia and severe angiogenic and pulmonary diseases like CTEPH, IPAH and GBM. Thus, the present work establishes a putative link of a disease-relevant lncRNA and the HUSH complex by triplex formation resulting in the inhibition of endothelial gene expression. The interaction of chromatin modifying complexes with lncRNAs suggests that lncRNAs have targeting or scaffolding functions within these complexes to modulate chromatin structure and thereby gene expression. Most studied lncRNAs have been identified to interact with complexes such as PRC2, SWI/SNF, E2F1 and p300, e.g. MEG3 4 , FENDRR 12 , MANTIS 39 , and KHPS1 7,10 . In the present work, we identified another silencing complex that can be targeted by lncRNAs: We demonstrated that HIF1α-AS1 interacts with proteins of the HUSH complex, which mediates gene silencing. HUSH is also involved in silencing extrachromosomal retroviral DNA 38 . Recently it has been shown that the HUSH complex, particularly MPP8, which is downregulated in many cancer types and whose depletion caused overexpression of long interspersed element-1 (LINE-1s) and Long Terminal Repeats, controls type I Interferon signaling involving a mechanism with dsRNA sensing by MDA5 and RIG-I 40 . Here we report a direct interaction of the HUSH complex member MPP8 with HIF1α-AS1. Moreover, we identified Exon1 of HIF1α-AS1 as being critical for this function. The HUSH complex has not yet been studied in vascular cells; it is not known whether its published composition with MPP8, TASOR and PPHLN1 35 is valid for endothelial cells. Our data propose that, in endothelial cells, the HUSH complex member MPP8 interacts with H3K9me3 and DNA and that SETDB1 and MPP8, but not NP220, repress gene expression of HIF1α-AS1-specific target genes.
The finding that HIF1α-AS1 interacts with the C-terminal domain of MPP8, and not with the chromodomain, was unexpected. The N-terminus of MPP8 was reported to interact with H3K9me3 41 . Substitution of Trp80 to alanine (W80A) within the chromodomain showed that it is important for H3K9me3 binding 41 , but also for other interactions, such as with DNMT3A 42 . Douse et al. removed the first 499 aa of MPP8 without impairing HUSH function; the function of HUSH was, however, only affected by the deletion of the amino acids 500-860, which also contain the predicted ankyrin repeats 43 . A similar finding was made for the maintenance of the self-renewal of groundstate murine embryonic stem cells, where not the chromodomain, but a C-terminal region of MPP8 was required for function 44 . The function of the IDRs within MPP8 were not investigated so far, but disordered regions are discussed to be potential linkers or binding partners of RNA 45 .
In our unbiased approach, a large number of DNA binding sites were identified for HIF1α-AS1 with triplex domain finder analysis. The  large number is not unusual as many of these binding sites overlap and are not identical. Also for other lncRNAs, such as GATA6-AS, FENDRR, HOTAIR and PARTICLE, many DNA binding sites have been predicted within their target genes 9,14 . EPHA2 and ADM, as well as PLEC, RP11-276H7.2, MIDN and EGR1 contained a large number of DNA binding sites for HIF1α-AS1 and were upregulated after HIF1α-AS1 knockdown. It is therefore tempting to speculate that similar regulatory mechanisms may play a role in the regulation of these genes. Despite containing triplex target sites, several genes were unaffected by changing the expression level of the triplex-forming RNA. Given the large number of target sites, this could be a consequence of redundancy with respect to target sites or lncRNAs, steric hindering or additional local factors so far unknown. In fact, beyond Hoogsteen base pairing, the local factors required for triplex formation have not yet been identified. For example, it is possible that large protein complexes, like those involved in splicing interfere with binding. Also, the binding of transcription factors could compete with the lncRNA binding. Obviously, also the local epigenetic landscape and chromatin state impacts on triplex binding. On several occasions, our study takes advantage on the fact that RNase H cleaves the RNA in DNA-RNA heteroduplexes 19 , and therefore enriches triplex-forming RNAs within the pool of DNA-interacting RNAs. Although this approach has been widely used in the field 4,6,7,13,15,20,51 , it is indirect and therefore not perfect. Proteins or specific local factors may shield RNAs resulting in false positive results and dynamic triplexes, with weak RNA interactions might also be digested. This is why additional methods, in particular bioinformatics prediction of Hoogsteen base pairing and ex vivo demonstration of triplex forming potential are needed to confirm the data obtained with the aid of differential RNase H-digestion.
The evidence for triplex formation by HIF1α-AS1 is substantiated by a number of findings: Firstly, target recognition by HIF1α-AS1 occurs via triplex formation involving GA-rich sequences of the DNA targets and GA-rich sequences within HIF1α-AS1 lncRNA. This has also been observed for other lncRNAs such as HOTAIR 52 and MEG3 4 . Secondly, the 1 H-1D NMR and CD spectroscopy data for HIF1α-AS1 provided similar but more detailed characteristics for triplex formation, compared with other studies 4,5 . Thorough NMR analysis of attenuations of the individual DNA Watson-Crick base-paired nucleotides allows delineation of those base pairs that are markedly affected by triplex formation. From a total of 25 base pairs in the EPHA2_DNA duplex target, only 13 base-pairs are affected. This observation in turn implies that not the entire HIF1α-AS1 (TFO2-23) RNA is engaged in interaction with the DNA target duplex within the major groove of the DNA duplex, but substantial parts of the RNA strand retain dynamic flexibility which was further assured by our structural modeling. Through the use of heteroduplex samples, measurements at different temperatures, a reduction of equivalents of RNA and triplex analysis with stabilized DNA hairpin sequences, our study allowed an improved and extended analysis of triplex formation. Thirdly, in agreement with previous work 5 , most of the triplex target sites were located in the promoter region or introns of the DNA target genes. Fourthly, the triplex formation of HIF1α-AS1 resulted in gene repression, a finding also observed for other triplex forming RNAs 3 . We could extend this finding by replacing the TFR2 of HIF1α-AS1 with other sequences, which abolished the repressive effects.
HIF1α-AS1 was downregulated in the lungs of patients with specific forms of pulmonary arterial hypertension (PAH). PAH is characterized by several structural changes, remodeling and lesion development in the pulmonary arteries. A study by Masri et al. demonstrated the impairment of pulmonary artery endothelial cells from IPAH patients to form tube-like structures 53 . CTEPH, a complex disorder with major vessel remodeling and small vessel arteriopathy, is characterized by medial hypertrophy, microthrombi formation and plexiform lesions 54 . It has been further shown that TGF-ß-induced angiogenesis was increased by circulating CTEPH microparticles co-cultured with pulmonary endothelial cells, indicating a pro-angiogenic feedback of endothelial injury 55 . Since HIF1α-AS1 knockdown led to an increase in sprouting, we assume that the loss of HIF1α-AS1 is a compensatory mechanism, which could be putatively included in the above mentioned pro-angiogenic feedback loop. HIF1α-AS1 was also reduced in endothelial cells isolated from glioblastoma. Typically this pathology represents a highly angiogenic situation with defective endothelium and abnormal morphology 56 . Additionally, HIF1α-AS1 is pro-apoptotic 27 and so the reduction of HIF1α-AS1 could explain the observed sprouting phenotype by the inhibition of apoptosis. Therefore, it is tempting to speculate that HIF1α-AS1 harbors atheroprotective roles, which could be exploited to alter angiogenesis in patients. Strategies to design such therapeutics require data in other species and in different tissues. HIF1α-AS1 is not endothelial-specific according to CAGE analysis. A comprehensive analysis on HIF1α-AS1 conservation, especially of TFR2, is lacking. Initial attempts with BLAT showed that the first 1000 nt of the pre-processed HIF1α-AS1 including TFR2 were conserved in primates and pigs, but not in rodents (data not shown). A potential application could be the promotion of vascular regeneration after an ischemia damage to promote early blood supply. Indeed, the post-ischemic healing response is not solely dependent on cardiomyocyte loss and adaptation but also on the damage response of the stroma-vascular compartment 57 . HIF1α-AS1 was downregulated in hypoxia, but upregulated in the damage-relevant re-oxygenation phase. This suggests that specifically in that phase where endothelial proliferation is most needed, HIF1α-AS1 limits the angiogenic response and therefore advocates itself as a target. Therefore, we propose an anti-HIF1α-AS1 approach to promote the early angiogenic response to promote post-ischemia regeneration.
Additionally, the data indicate that triplex formation could have therapeutic potential. The single nucleotide polymorphism (SNP) rs5002 (chr11:10326521 (hg19)) was found within the triplex target site of ADM with phenoscanner, which lists an association with hemoglobin concentration, red blood cell count and hematocrit 58 . Another link between a triplex forming lncRNA and PAH was reported by a massive upregulation of MEG3 in paSMCs from IPAH patients. This prevented hyperproliferation after MEG3 knockdown and a reduced apoptosis phenotype of IPAH-paSMCs involving a mechanism with miR-328-3p and IGF1R 59 . Although triplex formation was not studied, another study provided evidence that a ribonucleotide sequence can be used to form a potential triple helix to inhibit gene expression of the IGF1R gene in rat glioblastoma cells 60 . MEG3 is known to impair cell proliferation and to promote apoptosis in glioma cells 61 . This argues that the binding of a lncRNA to DNA is potentially involved in PAH and GBM.
Taken together, the findings presented here highlight an important pathway of a scaffolding lncRNA within an epigenetic-silencer complex that has a crucial role in the regulation of endothelial genes.

Materials
The following chemicals and concentrations were used

Analyses of Triplex-Seq data to identify candidate lncRNAs
Triplex-Seq data of U2OS and HeLa S3 was used from 15 , aligned using STAR 62 and peak-calling was performed with MACS2 63 . Peaks were intersected with Ensembl hg38 gene coordinates to produce a list of gene-associated peaks, which was filtered for lncRNAs. The overlap of U2OS and HeLa S3 lncRNAs was filtered for high confidence candidates by applying two cut-off filters for fold enrichment (>10) and -log10(P value peak enrichment) (>20). The P value for the enrichment of the peak (i-log10(P value peak enrichment)) was calculated using a Poisson distribution to estimate the expected number of reads which should lie within the peak region. The enrichments (fold enrichment, P) are then calculated based on the ratio of observed reads at the peak location (i.e. the real peak) vs. the expected peak. Next, the candidates were filtered for the presence of a nuclear value (>0) in ENCODE and for the presence of a signal (>0) in aorta, artery, lymphatic, microvascular, thoracic, umbilical vein and vein in FANTOM5 CAGE data [21][22][23] . Subsequently, the remaining candidates (RMRP, HIF1α-AS1, RP5-857K21.4, SCARNA2 and SNHG8) were tested for their non-coding probability with the online tools CPAT 64 and CPC2 65 . Lastly, regions enriched in the Triplex-Seq were manually inspected in the IGV browser to rule out the possibility that the signals belong to overlapping genes.

Total and nuclear RNA isolation, Reverse transcription and RT-qPCR
Total RNA isolation was performed with the RNA Mini Kit (Bio&Sell). Reverse transcription was performed with SuperScript III Reverse Transcriptase (Thermo Fisher) and oligo(dT)23 together with random hexamer primers (Sigma). CopyDNA amplification was measured with RT-qPCR using ITaq Universal SYBR Green Supermix and ROX as reference dye (Bio-Rad, 1725125) in an AriaMX cycler (Agilent). Relative expression of target genes was normalized to ß-Actin or 18 S ribosomal RNA. Expression levels were analyzed by the delta-delta Ct method with the Agilent Aria 1.7 qPCR software. Oligonucleotides used for amplification are listed in Table 1.

Knockdown procedures
For small interfering RNA (siRNA) treatments, endothelial cells (80-90% confluent) were transfected with GeneTrans II according to the instructions provided by MoBiTec (Göttingen, Germany). The following siRNAs were used: siEPHA2 (Thermo Fisher Scientific, HSS176396), siSETDB1 (Thermo Fisher Scientific, s19112) and siMPP8 (Thermo Fisher Scientific, HSS123184). The stealth siRNA targeting the intron of HIF1α-AS1 (approx. 100 nt downstream of TFR2) was designed with the Invitrogen BLOCK-iT RNAi designer (Thermo Fisher) and had the following sequence: 5ʹ-GCC TGG TCC CAA ACA TGC ATC ATA T-3ʹ. As negative control, scrambled Stealth RNAi™ Med GC (Life technologies) was used. All siRNA experiments were performed for 48 h.

Protein isolation and western blot analyses
HUVECs were washed in Hanks solution (Applichem) and afterwards lysed with Triton X-100 buffer (20 mM Tris/HCl pH 7.5, 150 mM NaCl, 10 mM NaPPi, 20 mM NaF, 1% Triton, 2 mM Orthovanadat (OV), 10 nM Okadaic Acid, protein-inhibitor mix (PIM), 40 µg/mL Phenylmethylsulfonylfluorid (PMSF)). The cells were centrifuged (10 min, 16,000 × g) and protein concentration of the supernatant was determined with the Bradford assay. The cell extract was boiled in Laemmli buffer and equal amounts of protein were separated with SDS-PAGE. The gels were blotted onto a nitrocellulose membrane and blocked in Rotiblock (Carl Roth, Germany). After incubation with the first antibody, infrared-fluorescent-dye-conjugated secondary antibodies (Licor, Bad Homburg, Germany) were used and signals detected with an infrared-based laser scanning detection system (Odyssey Classic, Licor, Bad Homburg, Germany). Images were acquired with Image

Human lung samples
The study protocol for tissue donation from human idiopathic pulmonary hypertension patients was approved by the ethics committee (Ethik Kommission am Fachbereich Humanmedizin der Justus Liebig Universität Giessen) of the University Hospital Giessen (Giessen, Germany) in accordance with national law and with Good Clinical Practice/ International Conference on Harmonisation guidelines. Written informed consent was obtained from each individual patient or the patient's next of kin (AZ 31/93, 10/06, 58/15) 66 .
Human explanted lung tissues from subjects with IPAH, CTEPH or control donors were obtained during lung transplantation. Samples of donor lung tissue were taken from the lung that was not transplanted. All lungs were reviewed for pathology and the IPAH lungs were classified as grade III or IV.

PASMC isolation and culture
Pulmonary arterial smooth muscle cells (PASMCs) were handled and treated as described before 67 . Briefly, segments of PASMCs, which were derived from human pulmonary arteries (<2 mm in diameter) of patients with IPAH or from control donors, were cut to expose them to the luminal surface. Gentle scraping with a scalpel blade was used to remove the endothelium. The media was peeled away from the underlying adventitial layer. 1-2 mm 2 sections of medial explants were cultured in Promocell smooth Muscle Cell Growth Medium 2 (Promocell, Heidelberg, Germany). For each experiment, cells from passage 4-6 were used. A primary culture of human PASMCs was obtained from Lonza (CC-2581, Basel, Switzerland), grown in SmGM-2 Bulletkit medium (Lonza) and cultured in a humidified atmosphere of 5% CO 2 at 37°C. Cells from passages 4-6 were used for experiments. For hypoxia experiments, PASMCs were incubated in hypoxia or normoxia chambers for 24 h in hypoxic medium (basal medium containing 1% FCS for human PASMCs). Hypoxia chambers were equilibrated with a watersaturated gas mixture of 1% O 2 , 5% CO 2 , and 94% N 2 at 37°C.

Brain microvessel isolation from glioblastoma (GBM) patients
Studies for human glioblastoma were covered by an ethics statement according to the guidelines of the University of Frankfurt, whose approval number for autopsy material is GS-249/11 and for resection material GS-04/09. Human Brain microvessel (HMBV) isolation from GBM patients was performed exactly as described before 39 . Within 3 h post surgery, fresh brain specimens were obtained from GBM patients.
For patients without available normal appearing healthy tissue, healthy material was obtained from epilepsy or dementia patients or autopsy material within a day postmortem. To isolate HMBV, specimens obtained in ice-cold MVB (15 mM HEPES, 147 mM NaCl, 4 mM KCl, 3 mM CaCl 2 , 1.2 mM MgCl 2 , 5 mM glucose and 0.5% BSA, pH 7.4) were used. These were cleared using forceps and the tissue was homogenized in 3-fold ice-cold MVB buffer by 15 up and down strokes in a tight-fitting douncer (0.25 mm clearance, 10 mL Wheaton) attached to an electrical overhead stirrer (2000 rpm, VOS 14, VWR). The homogenate was centrifuged (400 × g, 10 min, 4°C) and the pellet was resuspended in fourfold 25% BSA (in PBS). After an additional centrifugation (2000 × g, 30 min, 4°C), myelin fat in the top layer was aspirated. Next, the pellet containing the microvessels was resuspended in 3 mL ice-cold MVB/ gram starting material. To remove large vessels and tissue aggregates, the sample was filtered through 100micron sterile nylon mesh cell strainer (BD) and the microvessels were trapped onto a 40-micron sterile nylon mesh (BD). Afterwards, the mesh was washed once with ice-cold MVB and the microvessels were lysed directly with ice-cold RLT-Plus RNA lysis buffer (Qiagen), vortexed and stored at −80°C until use.

RNA pulldown assay and mass spectrometry
The RNA pulldown assay was performed similar to 39 . For proper RNA secondary structure formation, 150 ng of 3ʹend biotinylated HIF1α-AS1 or control RNA was heated for 2 min at 90°C in RNA folding buffer (10 mM Tris pH 7.0, 0.1 M KCl, 10 mM MgCl 2 ), and then put on RT for 20 min. 1 × 10 7 HUVECs were used per sample. Isolation of nuclei was performed with the truCHIP™ Chromatin Shearing Kit (Covaris, USA) according to the manufacturers protocol without shearing the samples. Folded Bait RNA was incubated in nuclear cell extracts for 3 h at 4°C. After incubation, samples were UV crosslinked. Afterwards, Streptavidin M-270 Dynabeads (80 µL Slurry, Thermo Fisher) were incubated with cell complexes for 2 h at 4°C. After 4 washing steps with the lysis buffer of the truCHIP chromatin Shearing Kit (Covaris, USA), beads were put into a new Eppendorf tube. For RNA analysis, RNA was extracted with TRIzol (Thermo Fisher). Afterwards, RNA purification was performed with the RNeasy Mini Kit (Qiagen). If indicated, RT-qPCR was performed. For mass spectrometric measurements in order to reduce complexity, samples were eluted stepwise from the beads. Beads were resuspended in 50 mM ammoniumhydrogencarbonate and 1 µL RNAse A. Supernatant was reduced and alkylated with DTT and chloracetamid, respectively. Remaining Beads were resuspended in 20 µL 6 M Guanidinhydrochlorid (GdmCl), 100 mM Tris/HCl, pH 8.5, 10 mM DTT and incubated at 95°C for 5 min. Reduced thiols were alkylated with 40 mM chloroacetamid and samples were diluted with 25 mM Tris/HCl, pH 8.5, 10% acetonitrile to obtain a final GdmCl concentration of 0.6 M. Proteins of both fractions were digested with 1 µg Trypsin/LysC (sequencing grade, Promega) overnight at 37°C under gentle agitation. Digestion was stopped by adding trifluoroacetic acid to a final concentration of 0.1%. Peptides were loaded on multi-stop-and-go tip (StageTip) containing three a stack of three C18-disks. Both fractions were eluted in wells of microtiter plates and peptides were dried and resolved in 1% acetonitrile, 0.1% formic acid. Liquid chromatography/mass spectrometry (LC/MS) was performed on Thermo Scientific™ Q Exactive Plus equipped with an ultra-high performance liquid chromatography unit (Thermo Scientific Dionex Ultimate 3000) and a Nanospray Flex Ion-Source (Thermo Scientific). Peptides were loaded on a C18 reversed-phase precolumn (Thermo Scientific) followed by separation on a with 2.4 µm Reprosil C18 resin (Dr. Maisch GmbH) in-house packed picotip emitter tip (diameter 100 µm, 15 cm long from New Objectives) using an gradient from mobile phase A (4 % acetonitrile, 0.1% formic acid) to 40% mobile phase B (80% acetonitrile, 0.1% formic acid) for 60 min followed by a second gradient to 80% B for 30 min with a flow rate 400 nL/min. Run was finished by washout with 99% B for 5 min and reequilibration in 1% B. MS data were recorded by data dependent acquisition Top 10 method selecting the most abundant precursor ions in positive mode for HCD fragmentation. The Full MS scan range was 300 to 2000 m/z with resolution of 70000, and an automatic gain control (AGC) value of 3E6 total ion counts with a maximal ion injection time of 160 ms. Only higher charged ions (2 + ) were selected for MS/MS scans with a resolution of 17500, an isolation window of 2 m/z and an automatic gain control value set to E5 ions with a maximal ion injection time of 150 ms. Selected ions were excluded in a time frame of 20 s following fragmentation event. Fullscan data were acquired in profile and Fragments in centroid mode by Xcalibur software. For data analysis MaxQuant 1.5.3.30 and Perseus 1.5.4.1 were used. The enzyme specificity was set to Trypsin, missed cleavages were limited to 2. Following variable modifications were selected: at N-terminus acetylation (+42.01), oxidation of methionine (+15.99), as fixed modification carbamidomethylation (+57.02) on cysteines. Human reference proteome set from Uniprot (Download 4/2015, 68506 entries) was used to identify peptides and proteins. False discovery rate (FDR) was set to 1 %. Protein group file was uploaded to Perseus and data set was cleaned from reverse identifications and common contaminants. Data were Log2 transformed. Identification were filtered for 4 valid values in at least one group. To enable calculation of ratios between sample and control, missing values were replaced from normal distribution. Positive hits from p values (p < 0.05) of students t test between experimental groups were highlighted. The samples were labeled H1-H5 for HIF1α-AS1 and C1-C5 for the negative control RNA. MaxQuant 1.5.3.30 and Perseus 1.5.4.1 were used to analyze the data. , 100 mmol/L NaCl, 2 mmol/L EDTA, 0.5% Triton X-100, 1 µL Superase In (per 100 µL) and protease inhibitors). Prior to elution, beads were put into a new Eppendorf tube. RNA was extracted with TRIzol (Thermo Fisher) followed by RNA purification with the RNeasy Mini Kit (Qiagen), reverse transcription and qRT-PCR.
Assay for transposase accessibility (ATAC)-sequencing ATAC-Seq was performed similar to 39 . 100.000 HUVECs were used for ATAC library preparation using Tn5 Transposase from Nextera DNA Sample Preparation Kit (Illumina). Cell pellets were resuspended in 50 µL PBS and mixed with 25 µL TD-Buffer, 2.5 µL Tn5, 0.5 µL 10% NP-40 and 22 µL H 2 O. The mixture was incubated at 37°C for 30 min followed by 30 min at 50°C together with 500 mM EDTA pH 8.0 for optimal recovery of digested DNA fragments. 100 µL of 50 mM MgCl 2 was added for neutralization. The DNA fragments were purified with the MinElute PCR Purification Kit (Qiagen). Amplification of library together with indexing was performed as described elsewhere 72 . Libraries were mixed in equimolar ratios and sequenced on NextSeq500 platform using V2 chemistry and assessed for quality by FastQC. Reaper version 13-100 was employed to trim reads after a quality drop below a mean of Q20 in a window of 5 nt 73 . Only reads above 15 nt were cleared for further analyses. These were mapped versus the hg19 version of the human genome with STAR 2.5.2b using only unique alignments to exclude reads with uncertain arrangement. Reads were further deduplicated using Picard 2.6.0 (Picard: A set of tools (in Java) 74 for working with next generation sequencing data in the BAM format) to avoid PCR artefacts leading to multiple copies of the same original fragment. The Macs2 peak caller (version 2.1.0) 63 as employed in punctate mode to accommodate for the range of peak widths typically expected for ATAC-seq. The minimum qvalue was set to −4 and FDR was changed to 0.0001. Peaks overlapping ENCODE blacklisted regions (known misassemblies, satellite repeats) were excluded. Peaks were annotated with the promoter (TSS + /− 5000 nt) of the gene most closely located to the center of the peak based on reference data from GENCODE v19. To compare peaks in different samples, significant peaks were overlapped and unified to represent identical regions. The counts per unified peak per sample were computed with BigWigAverageOverBed (UCSC Genome Browser Utilities, http://hgdownload.cse.ucsc.edu/ downloads.html). Raw counts for unified peaks were submitted to DESeq2 (version 1.14.1) for normalization 75 . Spearman correlations were produced to identify the degree of reproducibility between samples using R. To permit a normalized display of samples in IGV, the raw BAM files were normalized for sequencing depth (number of mapped deduplicated reads per sample) and noise level (number of reads inside peaks versus number of reads not inside peaks). Two factors were computed and applied to the original BAM files using bedtools genomecov resulting in normalized BigWig files.
For samples used after siRNA-mediated silencing of MPP8 and SETDB1 as well as the corresponding LNA GapmeR knockdown of HIF1α-AS1 or samples from CRISPRa, CRISPRi and LentiCRISPR experiments, the improved OMNI-ATAC protocol 76 was used and samples were sequenced on a Nextseq2000. The resulting data were trimmed and mapped using Bowtie2 77 . Data were further processed using deepTools 78 . For visualization, the Integrative Genomics Viewer 79 was used.

RNA and DNA Hybridization
By hybridization of the RNA strand to the DNA duplex or DNA hairpin DNA:DNA:RNA triplexes were formed. First the complementary DNA single strands were incubated at 95°C for 5 min in hybridization buffer (25 mM HEPES, 50 mM NaCl, 10 mM MgCl 2 (pH 7.4)) and afterwards cooled down to RT. Triplex formation was performed by adding RNA to previously hybridized double stranded DNA for 1 h at 60°C and then cooled down to RT 13 . For the 1 H-1D NMR, CD and melting curve experiments, the HIF1α-AS1-TFR2 (TFO2-23) sequence 5ʹ-GCGGCGGA GGAAAGAGAAAGGAG-3ʹ (length 23 nt, GC = 50.9%) was used in combination with the DNA sequences listed in Table 2.

CD spectroscopy and melting curve analysis
Circular dichroism spectra were acquired on a Jasco J-810 spectropolarimeter. The measurements were recorded from 210 to 320 nm at 25°C using 1 cm path length quartz cuvette. CD spectra were recorded on 8 µM samples of each DNA duplex, DNA:RNA heteroduplex and DNA:DNA:RNA-triplex in 25 mM HEPES, 50 mM NaCl, 10 mM MgCl 2 (pH 7.4). Spectra were acquired with 8 scans and the data was smoothed with Savitzky-Golay filters. Observed ellipticities recorded in millidegree (mdeg) were converted to molar ellipticity [θ] = deg x cm 2 x dmol −1 . Melting curves were acquired at constant wavelength using a temperature rate of 1°C/min in a range from 5°C to 95°C. All melting temperature data was converted to normalized ellipticity and evaluated by the following Eq. (1) using SigmaPlot 12.5:  and mixing times of 150 ms. NMR data were collected, processed and analyzed using TopSpin 3.6.2 (Bruker) and Sparky 3.115 82 .

Structural modeling
Models of the DNA:DNA:RNA triplex were generated by using the ARIA/CNS software packages [83][84][85] . To generate and keep the B-form DNA duplex, ample modeling distances and dihedral angle restraints were used. For flexible docking of the RNA on the B-DNA, a starting structure was generated containing a DNA duplex template and an extended RNA molecule. The docking was solely driven by hydrogenbonds and base-planarity restraints for the triplex. No further restraints were added for the RNA; leaving it fully flexible during the conventional simulated annealing stages with cartesian angle dynamics. In total, 2000 models were generated and the 200 best structures (lowest energy) were used as input for a further refinement in explicit water using the nucleic acid forcefield with OPLS charges and nonbonded parameters 86 . The final ensemble of 20 top-ranked structures was validated and had no violations. Figure production was done by using PyMol 2.5 (Schrödinger, LLC).

Caspase-3/7 activity assay
The Caspase-3/7 activity assays were carried out using 1×10 6 HUVEC. The assay was performed using SR-FLICA Caspase-3/7 assay Kit (ImmunoChemistry Technologies LLC, 931) following the manufacturer's instructions. Briefly, cells were washed and a 1:5 dilution of FLICA was added in a dilution of 1:30 to the cell suspension. After an incubation of 1 h, cells were washed three times with buffer provided by the kit, counted and diluted to 3000 cells/µL before measuring emission at 595 nm in a TECAN infinite M200 Pro plate reader using the TECAN i-control 3.7.3.0 software (Männedorf, Switzerland).

Chromatin Immunoprecipitation
Preparation of HUVEC extracts, crosslinking and isolation of nuclei was performed with the truCHIP™ Chromatin Shearing Kit (Covaris, USA) according to the manufacturers protocol. The procedure was similar to 88 . The lysates were sonified with the Bioruptur Plus (10 cycles, 30 s on, 90 s off, 4°C; Diagenode, Seraing, Belgium). Cell debris was removed by centrifugation and the lysates were diluted 1:3 in dilution buffer (20 mmol/L Tris/HCl pH 7.4, 100 mmol/L NaCl, 2 mmol/L EDTA, 0.5% Triton X-100 and protease inhibitors). Pre-clearing was done with DiaMag protein A and protein G coated magnetic beads (Diagenode, Seraing, Belgium) for 1 h at 4°C. As indicated, the samples were incubated over night at 4°C with the following antibodies and dilutions: Anti RNase A per 100 µL buffer. Elution of the beads was done with elution buffer (0.1 M NaHCO 3 , 1% SDS) containing 1x Proteinase K (Diagenode, Seraing, Belgium) and shaking at 600 rpm for 1 h at 55°C, 1 h at 62°C and 10 min at 95°C. After removal of the beads, the eluate was purified with the QiaQuick PCR purification kit (Qiagen, Hilden, Germany) and subjected to qPCR analysis. As a negative control during qPCR, primer for the promoter of GAPDH were used. The primers are listed in Table 3.

Triplex domain finder analysis
Triplex formation of HIF1α-AS1 was predicted using the Triplex Domain Finder 0.13.2 (TDF) 14 with the human pre-spliced HIF1α-AS1 sequence (NR_047116.1, gene ID 100750246) to target DNA regions around genes with ATAC-Seq peaks upon HIF1α-AS1 silencing. For annotation of HIF1α-AS1 triplex forming regions across DNA triplex target sites, genome version hg19 was used. Boxplots of Fig. 2b show the distribution of triplex prediction from 200 randomizations by shuffling the positions of the same DNA target regions in the genome. Enrichment was given at a p-value <0.05.

Statistics and reproducibility
Unless otherwise indicated, data are given as means ± standard error of mean (SEM). Calculations were performed with Prism 8.0 or BiAS.10.12. The latter was also used to test for normal distribution and similarity of variance. For multiple group comparisons ANOVA followed by post hoc testing was performed and multiplicity adjusted p values were shown, if indicated. Individual statistics of dependent samples were performed by two-tailed Student's t test (paired or unpaired), and if not normally distributed by Mann-Whitney test. P values of <0.05 were considered as significant. Unless otherwise indicated, n indicates the number of individual experiments.

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

Data availability
The ATAC-Seq data generated in this study have been deposited in the Sequence Read Archive (SRA) (https://www.ncbi.nlm.nih.gov/sra) under BioProject ID . The CRISPR ATAC-Seq data generated in this study have been deposited in the NCBI's Gene Expression Omnibus under the GEO Series accession number GSE203252. The mass spectrometry proteomics data about HIF1α-AS1 interaction partners identified in this study have been deposited to the the ProteomeXchange Consortium via the PRIDE partner repository 89 with identifier PXD023512. Triplex-Seq data was used from 15 and is deposited in NCBI GEO under accession number GSE120850. Ensembl hg38 was used for the identification of candidate lncRNAs from the Triplex-Seq data. FANTOM5 ENCODE CAGE expression data was obtained from FAN-TOM5 website (Gencode v19) [21][22][23] . ChIP-Seq datasets were taken from ENCODE 90 and are deposited at NCBI GEO under accession number GSM733673 for HUVEC H3K4me3, for H3K27Ac under accession code GSM733691 and for H3K9Ac under GSM733735. Source data are provided with this paper.