MATR3 regulates the spatial organization of nuclear chromatin

To investigate the fine‐scale localization pattern of MATR3 in the nucleus, we performed super‐resolution fluorescence microscopy and immuno‐electron microscopy in mouse hepatocytes (AML12 cells), which had been applied in studying nuclear architecture previously (Fan et al, ²⁰¹⁸; Huo et al, ²⁰²⁰). The imaging data showed that MATR3 proteins organize into network‐like structures with some concentrated puncta, which were adjacent to chromatin fibers in the nucleus (Figs 1A and EV1A). To determine the chromatin types associated with MATR3, we measured the correlation coefficient (r) between MATR3 and histone marks on randomly selected regions of interest (ROI) in the nuclei. The nuclear distribution of MATR3 showed the highest correlation with H3K27me3, comparing to other histone marks, including H3K9me3, H3K9me2, H3K27ac, or H3K4me3 (Figs 1B and EV1B).

To reveal the biological function of MATR3 proteins in chromatin regulation, we established a doxycycline (Dox)‐inducible, short hairpin RNA (shRNA)‐based RNAi system (Wiederschain et al, ²⁰⁰⁹) in AML12 cells. Using this system, 3 days of Dox treatment can generate a knockdown of MATR3, and 3 days after Dox removal can restore the expression level consistent to the control group (Fig 1C). We first investigated the chromatin changes within regions of different chromatin types by immunofluorescence staining. To quantify the dispersion degree of fluorescence signal, we measured the standard deviation (SD) values of pixel intensity in the whole nuclear region. After MATR3 depletion, SD values of H3K9me2 and H3K27me3 increased, and that of H3K27ac decreased. Among these histone marks, H3K27me3 showed the most significant redistribution (P < 0.0001; Fig 1D). For the following investigation, therefore, we took H3K27me3 as the representative mark to probe chromatin organization changes. H3K27me3 staining in the control group showed a relatively diffused distribution pattern with some irregular foci in the nucleus. After depletion of MATR3, H3K27me3 staining presented larger and brighter foci in the inner nucleus as well as near the nuclear periphery and the nucleolus (Figs 1E and EV1C). To quantify the redistribution of H3K27me3‐modified chromatin toward nuclear periphery or nucleolus, we analyzed the pixel signal distribution on randomly selected ROIs in Lamin A/C or C23 (a nucleolus marker) co‐stained cells. The data indicated that H3K27me3‐modified chromatin became significantly closer to the nuclear periphery and the nucleolus after MATR3 depletion (Figs 1F and G, and EV1D and E). Importantly, Dox removal for 3 days restored the H3K27me3 pixel intensity SD values and spatial distribution relative to nuclear periphery and nucleolus (Figs 1E–G and EV1D–F), suggesting a direct regulatory role of MATR3 on chromatin organization. We confirmed that the changes in H3K27me3 staining were not a result of the Dox treatment or the cellular response to the shRNA, as no H3K27me3 alteration was observed in Dox‐treated control cells that transfected with the scrambled shRNA (Appendix Fig S1A). Next, we asked whether the changes in H3K27me3 staining were due to the alteration of H3K27me3 modification levels. However, the total H3K27me3 level and the genome‐wide profiles were largely unchanged upon MATR3 depletion, as indicated by western blotting (Fig 1H) and ChIP‐seq (Figs 1I and EV1G) in AML12 cells. Therefore, the changes in H3K27me3 staining reflected the spatial redistribution of chromatin rather than the alteration of histone modification levels.

To rule out the potential off‐target of RNAi methods and other secondary effects after knockdown, we established an auxin (IAA)‐inducible rapid protein degradation system (Natsume et al, ²⁰¹⁶) in mouse embryonic stem cells (ESCs), which can generate MATR3 degradation within 6 h (Figs 1J and EV1H). As AML12 cells cannot be cultured in clones, the acute MATR3 degradation is technically difficult in this cell line. In ES cells, we first evaluated colocalization between MATR3 and different types of chromatins. H3K27me3‐modified chromatin showed positive correlation with MATR3, although the correlation in ES cells was weaker than that in AML12 cells (Appendix Fig S1B and C). After rapid degradation of MATR3, H3K27me3 in ES Cells showed larger and brighter staining near the nuclear periphery (Fig 1K and L), and the SD of H3K27me3 pixel intensity significantly increased accordingly (Fig 1M). Collectively, the acute degradation of MATR3 in ES Cells and RNAi‐based MATR3 knockdown in AML12 cells resulted in similar spatial changes in chromatin as probed by H3K27me3 staining.

MATR3 interacts with nuclear RNAs including antisense LINE1 RNAs

Nuclear RNAs were shown to participate in chromatin spatial organization (Rinn & Guttman, 2014; Nozawa & Gilbert, ²⁰¹⁹). As MATR3 acts as an RNA‐binding protein (Salton et al, ²⁰¹¹; Uemura et al, ²⁰¹⁷; Attig et al, ²⁰¹⁸), we wondered whether MATR3's functional role in chromatin organization is RNA‐dependent. According to cell fractionation assay (Bhatt et al, ²⁰¹²) followed by western blotting detection, most MATR3 proteins present in chromatin‐associated fractions in control cells, but greatly delocalized from chromatin after RNA pol II inhibitor dichloro‐ribofuranosyl benzimidazole (DRB) treatment (75 μM, 12 h) or RNase A treatment (10 μg/ml, 1 h; Fig 2A). Interestingly, immunofluorescent images showed that treating cells with RNA transcription inhibitor (75 μM DRB or 50 μg/ml α‐amanitin for 24 h) or RNase A (10 μg/ml for 1 h) disrupted network of MATR3 proteins and led them forming into spheroidal puncta (Fig 2B; Appendix Fig S2A and B). Pixel signal of MATR3 and H3K27me3 is positively correlated (r = 0.66) in the control cell, but the correlation turned negative in the DRB (r = −0.36) or RNase A (r = −0.13) treated cells (Fig 2B; Appendix Fig S2B). We observed that 29% of cells showed the spheroidal MATR3 puncta after 1 h treating with 75 μM DRB, and the percentage of cells containing MATR3 puncta increased with treating time (Fig EV2A and B). The overall RNA decay and the secondary consequences caused by protein synthesis inhibition (Tamm, 1984) could contribute to this phenomenon. Therefore, we looked into the nuclear distribution of MATR3 truncations, which lack the RNA‐binding ability. Of note, RNA recognition motifs (RRM1/2) deletion made MATR3 distribute as spheroidal foci, and the impact of RRM2 deletion was more prominent (Appendix Fig S2C). These data suggested that the interaction between MATR3 and RNAs is necessary for MATR3 meshwork formation and MATR3‐chromatin association.

To further identify MATR3‐associated RNAs in chromatin organization, we performed strand‐specific RNA immunoprecipitation sequencing (RIP‐seq) in AML12 and ES cells with an anti‐MATR3 antibody. In total, 42.6% of RIP‐seq peaks were associated with intronic regions of annotated genes (Fig 2C), which is in accordance with MATR3 PAR‐CLIP data from human neuronal cells (Uemura et al, ²⁰¹⁷) and eCLIP data from HepG2 cells (Van Nostrand et al, ²⁰²⁰). Notably, 17.2% of MATR3 RIP‐seq peaks distributed at repeat sequences in AML12 cells (Fig 2C). Among a total of 17,190 repeat sequences associated MATR3 RIP‐seq peaks, 40.99% are localized in L1 loci (Fig 2D). The degree of MATR3 RIP‐seq signals was strongly enriched at antisense (AS) L1 loci (Fig 2E; Appendix Fig S2D). Furthermore, for most of L1 subfamilies, the interactions between MATR3 and AS L1 RNAs were stronger than those between MATR3 and sense L1 RNAs (Fig 2F). Our data in ES cells (Fig EV2C; Appendix Fig S2E) and previous MATR3 eCLIP conducted in HepG2 cells (Van Nostrand et al, ²⁰²⁰) indicated the same results.

Next, we examined the genomic distribution of the L1 elements that transcribed MATR3‐associated AS L1 RNAs. Of note, 88% of them located in introns of annotated genes (Fig EV2D). It raised the question whether MATR3‐associated AS L1 RNAs are independent RNA fragments or parts of host gene transcripts. We thus conducted RT–PCR with MATR3 RIP samples at two intronic L1 loci transcribed AS L1 RNAs; for each locus, we designed two pairs of primers: one targeted the borders of RIP peaks (primer 1 or 3), and the other targeted ~ 200 bp flanking regions (primer 2 or 4). The former can detect RNA fragments about 300 nt with both input and MATR3 RIP products, while the latter cannot detect RNA with MATR3 RIP products (Fig EV2E). Besides, the enrichment degree of MATR3‐associated AS L1 in RIP‐seq is not correlative to the expression level of their host genes (Fig EV2F). Additionally, we compared the half‐life of AS L1 RNAs with those of Actb mRNA and Neat1, the architectural lncRNA for paraspeckles (Fig EV2G). In AML12 cells, the half‐life of the AS L1 RNAs is 2–3 h (Fig EV2G), similar to that of Neat1 (2.54 h), although much shorter than that of Actb mRNA (35.22 h). Therefore, MATR3‐associated AS L1 RNAs could be mainly independent RNA fragments; however, we cannot exclude the possibility that MATR3 interacts with some AS L1‐containing unspliced pre‐mRNAs or other long transcripts.

In order to visualize nuclear distribution of AS L1 RNAs, we performed fluorescent in situ hybridization (FISH) in AML12 cells. Probes for RNA FISH were designed according to the consensus sequence of antisense transcripts from L1_Mus 1 subfamily, which ranked high on MATR3‐associated AS L1 subfamilies (Fig 2F). Remarkably, nuclear localization of AS L1 RNAs was highly correlated with that of MATR3 in both AML12 (r = 0.78; Fig 2G) and ES (r = 0.62) cells (Fig EV2H), displaying a meshwork‐like structure in the nucleus. Imaging data in AML12 cells showed that AS L1 RNAs were positively correlated with H3K27me3‐modified chromatin (r = 0.52), but negatively correlated with H3K27ac‐modified chromatin (r = −0.43; Fig 2G). Integrated analyses of RIP‐seq and ChIP‐seq data showed that MATR3‐associated ASL1 elements are enriched for H3K27me3 and H3K9me3, but depleted of H3K27ac (Fig 2H), which suggested the cis‐acting role of AS L1 RNAs. Of note, H3K27me3 and H3K9me3 are not fully overlapped but enriched at the end and start boundaries of ASL1 elements, respectively (Fig 2H, top panel).

Together, our genomic and imaging data revealed an extensive interaction between MATR3 and AS L1 RNAs.

MATR3‐AS L1 meshwork shapes the nuclear chromatin architecture

Next, we investigated the functional roles of the MATR3‐AS L1 meshwork in higher‐order chromatin organization. We began to test whether AS L1 RNAs affect MATR3's function in chromatin organization by treating AML12 cells with antisense oligonucleotides (ASOs) to knockdown AS L1 RNAs. We designed five ASOs, targeting the scrambled sequence, the sense or antisense transcripts of a MATR3‐associated L1 subfamily (L1_Mus1) and a MATR3‐non‐associated L1 subfamily (L1_MA7), respectively (Figs 2F and EV3A). Sense and antisense L1_Mus1 ASOs treatment for 24 h caused MATR3 forming into spheroidal foci in 10 and 69% of the cells (Fig EV3A and B), in accordance with the results that MATR3 is more associated with the antisense than the sense L1_Mus1 RNAs (Fig 2F). As additional negative controls, both sense and antisense L1_MA7 ASO treatment had no effects on MATR3 localization (Fig EV3A and B). Thus, we took the ASO targeting the antisense L1_Mus1 RNA (hereafter referred to as ASL1 ASO) for the following investigation. After 6 h of AS L1 ASO treatment, 22% of cells showed MATR3 forming into spheroidal foci in the nucleus; after 12 and 24 h, this kind of cells increased to 53 and 65%, respectively. Moreover, a small portion of cells appeared spheroidal foci in both nucleus and cytoplasm (Fig EV3C and D).

To determine the specificity of MATR3 redistribution upon AS L1 RNA knockdown, we tested SAF‐A/HNRNPU and SAFB, which are also nuclear matrix proteins. Unlike MATR3, they did not form the spherical foci after ASL1 depletion (Fig EV3E and F). We also examined the RNA‐binding protein FUS, which was reported to condensate into droplets after RNase A treatment (Maharana et al, ²⁰¹⁸). After treating with AS L1 ASO for 24 h, FUS proteins kept the dispersed distribution in most nuclear areas while some of them assembled around the MATR3 foci (Fig EV3G). The interaction between FUS and MATR3 (Kamelgarn et al, ²⁰¹⁶) may contribute to this phenomenon.

As most AS L1 elements are located on introns, we conducted rRNA‐off and strand‐specific RNA‐seq to examine whether the host genes that containing AS L1 in introns were knocked down after AS L1 ASO treatment. The expression changes in these genes were subtle and not correlative to the changes in the intronic AS L1 RNAs (PCC = 0.04; Fig EV3H). Therefore, the consequence of AS L1 ASO treatment was not the secondary effects by the degradation of host genes.

Subsequently, we looked into the effects on chromatin structure after AS L1 ASO treatment. Immunofluorescent staining revealed that after AS L1 knockdown, MATR3 foci disassociated from H3K27me3‐modified chromatin (r = 0.66 in Ctrl ASO; r = −0.11 in AS L1 ASO), resulted in a redistribution of H3K27me3 which shown as larger foci in the nucleus (Fig 3A). Furthermore, SD of H3K27me3 pixel intensity significantly increased in AS L1 ASO‐treated cells (Fig 3B). The spatial alteration of H3K27me3 upon AS L1 depletion is similar as the phenomenon caused by MATR3 knockdown.

We further investigated whether loss of MATR3 could affect AS L1 RNAs. Compared with the meshwork‐like organization in control cells, the distribution of AS L1 RNAs appeared to be more dispersed in MATR3‐depleted cells (Fig 3C). To quantify these changes, we analyzed the Gaussian fit distribution curve of the fluorescence signal. Distribution of AS L1 RNAs in control cells is right‐skewed with a mean value greater than mode value and a skewness greater than zero, which suggested the presence of concentrated distribution in partial areas. After MATR3 depletion; however, the fluorescence signals of AS L1 RNAs nearly followed a standard Gaussian distribution with a mean value nearly equal to mode value and a skewness close to zero, which suggested a random distribution (Fig 3C and D).

MATR3 protein has two zinc finger (ZF) domains and two RNA recognition motifs (RRMs; Hibino et al, ²⁰⁰⁶). We wondered which domains are necessary for MATR3 to interact with AS L1 RNAs. We expressed GFP‐tagged MATR3 truncations (▵ZF1, ▵ZF2, ▵RRM1, and ▵RRM2) in endogenous MATR3‐depleted AML12 cells and examined their colocalization with AS L1 RNAs. The deletion of ZF1 or ZF2 did not affect the MATR3 distribution and its association with AS L1 RNAs. The deletion of RRM1 and RRM2 both affect MATR3 distribution, but only RRM2 deletion abolished MATR3's association with AS L1 RNAs (Fig 3E and F). Therefore, RRM2 domain is essential for MATR3 to interact with AS L1 RNAs. To examine the effects of these truncations on chromatin structure, we depleted the endogenous MATR3 with shRNA and re‐expressed WT or truncated MATR3 to rescue the phenotype. The WT MATR3 completely rescued changes in H3K27me3 that caused by MATR3 knockdown, while ▵ZF1, ▵ZF2, and ▵RRM1 truncations partially rescued the phenotype. However, H3K27me3 in ▵RRM2‐expressed cells showed no significant difference compared with the MATR3‐depleted cells (Fig 3G and H), indicating that RRM2 is essential for the chromatin organization function of MATR3.

Collectively, these results indicated that AS L1 RNAs and MATR3 affect cellular localization of each other and loss of either of them could lead to the redistribution of chromatin. MATR3 interplay with AS L1 RNAs by the RRM2 domain, and they act together to form a meshwork‐like structure that helps to shape the chromatin architecture in cell nuclei.

Phase separation facilitates the formation of MATR3‐AS L1 meshwork

To investigate the physical characters of the meshwork formed by MATR3 and AS L1 RNAs, we first tested the dynamic feature in GFP‐MATR3‐expressed AML12 cells with fluorescence recovery after photobleaching (FRAP) assay. Both the wild‐type MATR3 and the RRM2‐depleted MATR3 showed partial recovery of fluorescence intensity, while the mobile fraction of RRM2‐depleted MATR3 (87.33%) was much higher than that of the wild‐type protein (49.11%) (Fig 4A and B).

Recent studies have revealed dynamic nuclear compartments driven by the liquid–liquid phase separation (LLPS) of proteins and/or RNAs (Stanek & Fox, 2017; Lyon et al, ²⁰²¹). To test the phase separation potential of MATR3 proteins, we expressed and purified GFP‐tagged MATR3 proteins and performed the in vitro droplet formation assay. As expected, GFP‐MATR3 proteins formed spherical assemblies at room temperature; area of these assemblies increased with higher protein concentration and decreased with higher NaCl concentration (Fig 4C–F). When incubating 3 μM of GFP‐MATR3 proteins with total RNAs extracted from AML12 cells, 5–10 ng/μl of total RNAs facilitated GFP‐MATR3 proteins to form larger irregular particles, while more than 20 ng/μl of total RNAs buffered these particles (Appendix Fig S3A). These results are in agreement with reported features of phase‐separated proteins (Maharana et al, ²⁰¹⁸).

To examine the roles of repeat RNAs in shaping MATR3 condensates in vitro, we incubated GFP‐MATR3 proteins with various in vitro transcribed, Cy5‐modified repeat RNAs and observed their droplet formation behaviors. The B1 and major satellites (MajSAT) RNAs were transcribed from full‐length B1 or MajSAT elements, respectively. As the full‐length L1 element is too long (~ 6 kb), the L1 RNAs were transcribed from a 318 bp consensus sequence of L1 Md_F2 elements (Appendix Fig S3B). Intriguingly, AS L1 RNAs facilitated MATR3 proteins forming into mesh‐like assemblies in vitro. While under the same molecule concentration, other repeat RNAs only made MATR3 droplets to undergo a slight deformation (Fig 4G).

To further interpret the action mechanisms of each repeat RNAs, we analyzed the numbers of MATR3‐binding motifs (Uemura et al, ²⁰¹⁷) on them. As expected, only the AS L1 sequence contains MATR3‐binding motifs. Furthermore, motif density on the 318 nt‐antisense L1 Md_F2 RNA fragment (4/318) is comparable to that on full‐length antisense L1 Md_F2 RNA (89/5948) and is much higher than that on full‐length sense L1 Md_F2 RNA (4/5948) (Appendix Fig S3B). Hence, we suggest that AS L1 RNA has a higher affinity to bind MATR3 proteins. In cell nuclei, MATR3 could stochastically interact with different RNAs due to electrostatic forces. Those nonspecifically interacting RNAs would lower MATR3′s saturation concentration and buffer the MATR3 liquids at high RNA concentration (Maharana et al, ²⁰¹⁸; Wang et al, ^2018a). While the specifically interacting RNAs like AS L1 RNAs could form multivalent interaction with MATR3 proteins, that may enhance the overall avidity (Bhat et al, ²⁰²¹) and promote the meshwork‐like assembly formation.

Furthermore, when incubating 3 μM of MATR3 proteins with different concentrations of AS L1 RNAs, 5–200 nM AS L1 RNAs could help the formation of the meshwork‐like structure (Fig 4H). According to a quantitative proteomics analysis, there are on average 3,240,673 MATR3 protein molecules in each mammalian cell (Hein et al, ²⁰¹⁵). the estimated nuclear concentration of MATR3 protein is about 5.6 μM (close to 3 μM MATR3 concentration within in vitro system). In addition, based on RNA‐seq data combined with experimental measurements, the estimated molar concentration of the 318 nt‐AS L1 RNA is about 469 nM (calculation details described in Materials and Methods). We suggest the in vitro system containing 200 nM 318 nt‐AS L1 RNA fragment (Fig 4H) could nearly mimic the average stoichiometric ratio between MATR3 and AS L1 RNAs in vivo.

In summary, we suggest that MATR3 proteins have the potential to undergo liquid–liquid phase separation and the weak interaction between proteins' disorder regions contributes to this behavior. When MATR3 proteins interact with AS L1 RNAs via the RRM2 domain, the affinity between molecules increased and the mobility decreased, thereby facilitating the higher‐order MATR3‐AS L1 meshwork formation (Fig 4I).

3D genome organization changes upon MATR3 depletion

To study the role of MATR3 on 3D genome organization, we performed Hi‐C in control and Matr3‐depleted AML12 cells. We used DpnII and obtained over 140 million uniquely aligned read pairs per replicate (Appendix Fig S4A). Biological replicates from the same condition were highly correlated (Appendix Fig S4B), and then, we merged replicates for further analysis. We next examined intrachromosomal interactions at different resolutions and found that the Hi‐C contact maps in these two samples were similar (Appendix Fig S4C). Most of the compartments (97.7%) were unswitched after MATR3 knockdown (Fig 5A). However, the degree of genome compartmentalization, as visualized by heatmaps of average contacts, showed compartment‐specific alteration: increased in AA but decreased in BB compartments (Fig 5B). Consistently, the interactions between compartment A regions increased but decreased between compartment B regions (Fig 5C), and the interactions within compartment regions showed opposite trends (Fig 5D).

We also observed the chromatin interaction changes within TADs changed upon MATR3 knockdown (Fig 5E). We then investigated the Hi‐C data at TAD level. After MATR3 depletion, 78.8% TAD boundaries were overlapped between Ctrl and shMATR3 samples (Fig 5F). By calculating the insulation score, TAD boundaries located in compartment A regions increased in boundary strength, while TAD boundaries located in compartment B regions and AB boundaries hardly changed (Fig 5G). Furthermore, we categorized TADs into six groups according to the intensity of MATR3‐associated AS L1 and H3K27me3 (Fig 5H). All the six groups of TADs altered the local contacts, and the degree of change showed no clear correlation with the extent of H3K27me3 enrichment. However, the intra‐TAD interaction decreased in AS L1‐high TADs (groups 1 and 4), but increased in AS L1‐low TADs (groups 3 and 6); in the intermediate groups (2 and 5), intra‐TAD interaction increased with the decrease in MATR3‐associated AS L1 intensity (Fig 5H).

To further examine the roles of MATR3‐AS L1 meshwork in chromatin interactions at the TAD level, we compared TADs with or without MATR3‐associated AS L1 RNAs, as indicated by MATR3 RIP‐seq data. After MATR3 knockdown, the intra‐TAD contacts within no‐AS L1 TADs increased while decreased within with‐AS L1 TADs (Fig 5I); TADs with the higher AS L1 density exhibited a greater degree of reduction in terms of intra‐TAD contacts (Fig 5J). Furthermore, TADs with the higher AS L1 density were more enriched in compartment A regions, while TADs with no‐AS L1 density were highly enriched in compartment B regions (Fig 5K). The overall gene expression changes exhibited no trend along with AS L1 density in TADs (Fig 5L).

Additionally, we examined the 3D genome changes upon acute MATR3 degradation in mouse ES cells. We conducted Hi‐C with IAA‐treated (6 h) and control cells (Appendix Fig S5). The fast degradation allowed us to detect the direct effects on the 3D genome caused by MATR3 loss in ES cells. Most of (77.9%) TAD boundaries were overlapped between control and MATR3‐AID samples (Fig EV4A). Of note, after degradation of MATR3 in ES cells, TADs with higher MATR3‐associated AS L1 density presented a more loss of local contacts (Fig EV4B and C) and a greater enrichment in A compartments (Fig EV4D); these results are highly consistent with the those observed in AML12 cells. However, TAD boundaries strength decreased in compartment A (Fig EV4E). Furthermore, 98.2% compartments were unswitched (Fig EV4F), while the intercompartmental interactions increased in both A and B compartments (Fig EV4G), and the intracompartmental interactions increased in B compartments (Fig EV4H). There are some differences between ES and AML12 cells on TAD boundary strength and intercompartmental interactions, which could be attributed to the secondary effects by long‐time shRNA treatment in AML12 cells, or the cell type‐specific interactions within chromosomes (Jin et al, ²⁰¹³; Rao et al, ²⁰¹⁴). We have reported previously that AML12 and ES cells showed different patterns on chromatin compartmentalization (Yan et al, ²⁰²⁰).

Collectively, our Hi‐C data in both AML12 and ES cells suggested that MATR3‐AS L1 meshwork facilitates clustering of chromatin within TADs highly transcribed MATR3‐associated AS L1 RNAs.

MATR3‐AS L1 meshwork maintains the expression stability of essential genes

We further asked whether the interplay between MATR3 and AS L1 RNAs contributes to gene regulation. As most of the MATR3‐associated AS L1 loci locate in annotated genes (Fig EV2D), we focused on the genes which contain MATR3‐associated AS L1 on gene body. After MATR3 knockdown, the genes containing MATR3‐associated AS L1 were more susceptible to change their expression level (Appendix Fig S6A).

Subsequently, to investigate the functional relevance of MATR3‐AS L1 meshwork, we examined genes that contain MATR3‐associated AS L1 in intronic regions. MATR3 RIP‐seq data in AML12 cells and ES cells were used for these analyses. Genes with AS L1 signals were highly overlapped between the two types of cells (Appendix Fig S6B). Gene Oncology (GO) and KEGG enrichment analyses indicated that the common genes were significantly enriched in survival‐related pathways including cell cycle, cellular response to DNA damage stimulus and chromatin organization (Appendix Fig S6C). This partially explained the observation that cell growth was greatly impeded after MATR3 knockdown (Appendix Fig S6D). Furthermore, the AS L1‐associated differentially expressed genes (DEGs) were significantly enriched in terms related to CNS development (neuron projection morphogenesis) and liver function (Cholesterol metabolism with Bloch and Kandutsch−Russell pathways; Appendix Fig S6E).

MATR3 depletion alters the structure of H3K27me3‐modified chromatin

To obtain a higher resolution, genome‐wide view about the dynamics of H3K27me3‐modified chromatin, we conducted SAMMY‐seq (Sequential Analysis of Macro‐Molecules accessibilitY), which was reported to be more sensitive than Hi‐C in detecting structure and accessibility changes in heterochromatin (Sebestyen et al, ²⁰²⁰). DNA of control and MATR3‐depleted AML12 cells was extracted from four fractions: the CSK buffer‐soluble fraction (S1), DNase I‐sensitive fraction (S2), DNase I‐resistant fraction (S3), and the high‐salt resistant fraction (S4) (Fig 6A). We examined the S4 vs S2 domains (refer to as SAMMY domain thereafter), which are robust to detect the dynamics of heterochromatin, including H3K9me3 and H3K27me3‐marked chromatin; the higher the SAMMY‐seq signal, the lower the accessibility (Sebestyen et al, ²⁰²⁰). The SAMMY‐seq signals of control and MATR3 knockdown samples were consistent at the larger scale, while showing noticeable changes at some smaller regions (Fig 6B). Overall, 672 SAMMY domains were common between the two conditions and accounted for 43.08% of the genome; 856 domains (7.01% of the genome) were lost, while 397 domains (4.21% of the genome) were gained in MATR3‐depleted AML12 cells (Fig 6C). As indicated by the average profiles, after MATR3 knockdown, lost or gained domains showed the strong reduction or increase in SAMMY‐seq signals, and the common domains also displayed a mild decrease in the domain borders (Fig 6D). We then examined average profiles of H3K27me3 on these SAMMY domains. The H3K27me3 level was the highest in the lost SAMMY domains, and for both common and gained domains, H3K27me3 level was lower in the center but higher on the borders (Fig 6E). Of note, for all the three types of domains, H3K27me3 level was unchanged after MATR3 knockdown (Fig 6E).

Next, we took a closer look at H3K27me3 regions overlapped with SAMMY domains. H3K27me3 level was consistent between control and MATR3 knockdown (Fig 6F). However, the SAMMY‐seq signals were reduced in H3K27me3 regions overlapped with lost and common domains (Fig 6G), indicating an increase in accessibility at these regions. Furthermore, MATR3‐associated AS L1 elements were highly enriched near the boundaries of these H3K27me3 regions (Fig 6H). Taken together, our SAMMY‐seq data indicated that H3K27me3‐modified chromatin domains adjacent to AS L1 elements increase their accessibility upon MATR3 knockdown, without changing the H3K27me3 modifications. It should be noted that SAMMY‐seq is more sensitive to detect heterochromatin structure, and those H3K27me3 regions located at A compartments may not be detected by this assay.

To investigate the biological consequences by the structure changes in SAMMY domains, we counted numbers of DEGs in these regions. The number of DEGs associated with lost domains was significantly higher than that of common domains (Fig 6I; P < 0.001, Chi‐Square test). Furthermore, DEGs tended to be enriched at the borders of lost and gained, but not the common SAMMY domains (Fig 6J). GO analysis of these structure‐associated genes indicated that some terms were related to hepatocytes functions, such as lipid transport, cholesterol metabolism, and P450 pathway (Appendix Fig S7A–C).

As H3K27me3 is deposited and recognized by polycomb repressive complex 1 and 2 (PRC1 and PRC2; Margueron & Reinberg, ²⁰¹¹; Schuettengruber et al, ²⁰¹⁷; Blackledge & Klose, ²⁰²¹), we studied the regulatory relationship between MATR3 and PRC1 component BMI‐1 and PRC2 component EZH2. The co‐staining and the co‐immunoprecipitation results indicated that MATR3 interacted with BMI‐1 and EZH2 in cell nuclei (Fig EV5A and B). After depletion of MATR3, both BMI‐1 and EZH2 exhibited larger and brighter foci (Fig EV5C and D), and the significant SD increase in pixel intensity can be detected (Fig EV5E). However, MATR3 depletion did not change the protein level of BMI‐1 and EZH2 or their association with chromatin (Fig EV5F). Consistently, our imaging data showed that the redistributed BMI‐1 and EZH2 maintained their colocalization with H3K27me3 (Fig EV5G and H). As previous studies reported that the chromatin occupancy of PRC2 could be dependent on RNA‐PRC2 interaction (Long et al, ²⁰²⁰), we conducted the RIP–qPCR to examine whether MATR3 affect the RNA‐binding ability of EZH2. Besides the well‐known EZH2‐binding lncRNAs Malat1 and Neat1, we also examined the liver‐specific EZH2‐binding RNA Hnf1aos1 (Chen et al, ²⁰¹⁸; Wang et al, ^2018b). The RNAs were enriched in anti‐EZH2 RIP samples compared with IgG, while MATR3 depletion did not affect the enrichment (Fig EV5I). On the contrary, knockdown of BMI‐1 and EZH2 did not affect the protein level of MATR3 (Fig EV5J) and did not abolish the MATR3‐AS L1 RNA association (Fig EV5K and L). These data suggested that MATR3 regulated the localization of PcG proteins, but the latter did not modulate MATR3.

Amyotrophic lateral sclerosis‐associated MATR3 mutations lead to chromatin redistribution

Multiple mutations of MATR3 were reported to be associated with neurodegenerative diseases including amyotrophic lateral sclerosis (ALS), frontotemporal dementia (FTD), vocal cord and pharyngeal weakness with distal myopathy (VCPDM), and early onset neurodegeneration (EON; Johnson et al, ²⁰¹⁴; Marangi et al, ²⁰¹⁷; Kao et al, ²⁰²⁰; Fig 7A). Nevertheless, how MATR3 dysfunction contributed to the pathology remained unclear. As the S85C and F115C mutations were reported to have more supportive pedigree data (Johnson et al, ²⁰¹⁴) and had been studied the pathogenicity in knock‐in mice models (Kao et al, ²⁰²⁰; Van Bruggen et al, ²⁰²¹), we focused on the action mechanism of these two MATR3 mutants for subsequent investigation. Aberrant LLPS behavior of neurodegenerative‐disease‐associated proteins was found to be pathogenic (Zbinden et al, ²⁰²⁰). Therefore, we tested whether mutated MATR3 proteins lead to aberrant phase separation. Predictor of natural disordered regions (PONDR) algorithm suggested that most ALS‐associated mutations could change the PONOR score of MATR3 (Appendix Fig S8A). Particularly, the S85C mutation leads a disorder region shifting to an ordered one (Fig 7B).

We established a Dox‐inducible MATR3 knockdown system in mouse neuroblastoma N2A cells, a common cell model for neurodegenerative disease studies, and then transfected GFP‐tagged wild‐type or the mutants of MATR3 (S85C and F115C) to replace the endogenous MATR3 protein (Fig 7C). FRAP assay was used to investigate their dynamic features. We showed that 20 s after photobleaching, the fluorescence intensity of GFP‐WT, GFP‐F115C, and GFP‐S85C proteins recovered for 57, 43, and 20%, respectively, suggesting that these mutations make MATR3 proteins less dynamic in the nucleus (Fig 7D and E). Furthermore, we purified GFP‐tagged human MATR3 WT and S85C proteins and performed in vitro droplet formation assay. The WT MATR3 formed into highly dynamic droplets, while the S85C MATR3 proteins were much less dynamic and finally assembled into the fiber‐like structure (Fig 7F; Movies EV1 and EV2). To determine whether the mutation on MATR3 would alter its capacity to interact with AS L1 RNAs, we co‐stained MATR3 proteins (WT/S85C/F115C) with AS L1 RNAs in N2A cells. Both wild‐type and mutated MATR3 are well‐colocalized with AS L1 RNAs in N2A cells (Appendix Fig S8B and C).

Finally, we asked whether the MATR3 mutants could change nuclear chromatin distribution. As H3K27me3‐modified chromatin showed the highest correlation with MATR3 in N2A cells (Appendix Fig S8D and E), we also examined changes in H3K27me3 in this cell line. After MATR3 depletion, H3K27me3 greatly redistributed in the nucleus with brighter foci appeared; GFP‐WT overexpression completely rescued this phenotype. However, neither GFP‐S85C nor GFP‐F115C rescued the redistribution of H3K27me3 (Fig 7G and H). These data suggested that ALS‐associated MATR3 mutants have changes in their physical state, which could further lead to an abnormal chromatin organization.

shRNA system and stable express system in AML12 cells

The mouse hepatocyte cell line alpha mouse liver 12 (AML12; CRL‐2254, ATCC, Manassas, VA) were cultured in DMEM/F12 (11320033, Thermo Fisher Scientific, Waltham, MA) supplemented with 10% fetal bovine serum (16140071, Gibco, Grand Island, NY), ITS Liquid Media Supplement (100×, 41400045, Gibco), and 40 ng/ml dexamethasone (D4902, Sigma, Darmstadt, Germany) at 37°C and 5% CO₂. For Dox‐inducible MATR3 knockdown system (and scramble), plasmids were constructed by cloning the target sequences (the Matr3 shRNA target: GAGACCGATCTTGCTAATTTA; the Scramble shRNA target: ACTCGACACTATAGTATCTCA) to the pLKO‐Tet‐On vector (Wiederschain et al, ²⁰⁰⁹) and then transfect AML12 cells by lentivirus‐based system. To generate stable cell lines, AML12 cells were selected in the presence of 1 μg/ml puromycin for 1 day. One μg/ml dox was used to induce MATR3 depletion. For BMI‐1/EZH2 knockdown system, the 21‐mer sequences (the Ctrl shRNA target: ACTCGACACTATAGTATCTCA; the BMI‐1 shRNA target: CCAGATTGACGTCATGTATGA; the EZH2 shRNA target: GTGTGATAGCACAAAATTTTT) were cloned into pLKO.1‐TRC vector. For stable express cell lines, the GFP‐tagged MATR3 cDNA was cloned into Lv‐ef1a‐blasticidin‐tre‐MCS plasmid as described (Biscotti et al, ²⁰¹⁵). GFP‐tagged truncations (▵ZF1, ▵ZF2, ▵RRM1 and ▵RRM2) were further constructed based on this plasmid using TOYOBO KOD‐401 Kit, using primers as described (Malik et al, ²⁰¹⁸). Plasmid transfection in AML12 cells was accomplished by lentivirus‐based system. To generate stable cell lines, AML12 cells were selected in the presence of 1 μg/ml puromycin and 4 μg/ml blasticidin for 2 days. One μg/ml dox was used to induce GFP‐MATR3 (or its truncations) expression. For exogenous truncations rescue assay, the shMatr3‐resistant mutation was first introduced to Lv‐ef1a‐blasticidin‐tre‐GFP‐MATR3 plasmid. And then GFP‐tagged truncations (▵ZF1, ▵ZF2, ▵RRM1, and ▵RRM2) were further constructed based on this plasmid. Then, plasmids were transfected with Lipofectamine™ 3000 Transfection Reagent (L3000150, Thermo Fisher Scientific, Waltham, MA) into shMatr3 AML12 cells.

Rapid protein degradation system in mouse ES cells

The mouse ES cell line (E14TG2a) was cultured in 2i/LIF conditions as described (Li et al, ²⁰⁰⁸). The auxin (IAA)‐inducible MATR3 degradation system was developed in mouse ESCs according to the rapid protein depletion methods established by Natsume et al (2016). First, the parental cells were generated by introducing the vector encoding constitutive cytomegalovirus‐controlled auxin responsive F‐box protein (CMV‐OsTIR1) at the safe harbor ROSA26 using CRISPR/Cas (Nakao et al, ²⁰¹⁶). Cells were further selected in the presence of 1 μg/ml puromycin. After 7 days, colonies were picked for further selection in a 96‐well plate and the genotype was checked by genomic PCR. Based on this parental cell line, an in‐frame mAID cassette was introduced after the last codon of MATR3 gene by CRISPR/Cas (sgRNA target on Matr3 gene: ATAAATTGGCAGAAGAACGG). Plasmids were transfected with Lipofectamine™ 3000 Transfection Reagent (L3000150, Thermo Fisher Scientific, Waltham, MA). To ensure AID‐tagging on both alleles, two short homology donor vectors containing neomycin and hygromycin resistance markers were transfected simultaneously into cells. Cells were further selected in the presence of 2 mg/ml neomycin and 200 μg/ml hygromycin. After 7 days, colonies were picked for further selection in a 96‐well plate and the genotype was checked by genomic PCR. 0.5 mM IAA (dissolved in alcohol) was used to induce MATR3 degradation.

Dox‐inducible shRNA system and mutant transfection in N2A cells

Mouse Neuro 2a cells (N2A; CCL‐131, ATCC, Manassas, VA) were cultured in DMEM (Dulbecco's Modified Eagle's medium, Hyclone) supplemented with 10% fetal bovine serum, nonessential amino acids (11140050, Thermo Fisher Scientific) at 37°C and 5% CO₂. Dox‐inducible MATR3 knockdown system in N2A cells was established in the same way as in AML12 cells. To generate the shMATR3‐resistant MATR3 cDNA, we introduced a synonymous mutation of shMATR3 targeting sites into the Lv‐ef1a‐GFP‐MATR3 plasmid. And based on this plasmid, we further introduced MATR3 mutants (S85C/F115C) using TOYOBO KOD‐401 Kit. Lv‐ef1a‐GFP‐MATR3 WT/S85C/F115C (resistant) plasmids were separately transfected into the dox‐inducible shMATR3 N2A cells, using standard Polyethylenimine (PEI)‐based transfection approach.

Western blotting and immunofluorescence

For western blotting, the cell lysates were blotted against primary antibodies and the blots were visualized with peroxidase‐coupled secondary antibodies using ProteinSimple FluorChem M gel imaging system. For immunofluorescence, cells were cultured on the glassy coverslip, fixed by 4% formaldehyde for 10 min, treated by 0.5% Triton X‐100 for 10 min, blocked by 4% BSA for 30 min at room temperature, cultured by diluted primary antibody overnight at 4°C, followed by adding second‐fluorescence antibody for 1 h at room temperature, and stained with DAPI. Most experiments involving MATR3 antibody were performed by Bethyl (A300‐591A), which recognizes carboxy‐terminal of MATR3 protein and has higher antibody titers. There was an exception: for detecting the MATR3 degradation efficiency in mouse ES cells. As mAID‐tag blocked the antigen‐binding sites of the carboxy‐terminal MATR3 protein, which could not be recognized by Bethyl (A300‐591A), we used the amino‐terminal MATR3 antibody (Abcam, ab51081) as an alternative. Other antibody used in this paper as follows: anti‐H3K9me3 (ABclonal, A2360), anti‐H3K9me2 (ABclonal, A2359), anti‐H3K27me3 (ABclonal, A16199), anti‐H3K27ac (Active Motif, 39,133), anti‐H3K4me3 (ABclonal, A2357), anti‐β‐actin (Proteintech, 66,009‐1‐Ig), anti‐H3 (Abcam, ab1791), anti‐BMI‐1 (Santa Cruz, SC‐390443), and anti‐EZH2 (CST, 5246).

RNA fluorescence in situ hybridization (RNA FISH)

RNA fluorescence in situ hybridization was performed using RNAscope® Multiplex Fluorescent Reagent Kit v2 (ACD, 323100). Probes for AS L1 RNA were designed by ACD company (Advanced Cell Diagnostics, Hayward, CA, USA) based on the consensus sequence on antisense L1_Mus1 RNAs. The product code is 1084381‐C1 (Advanced Cell Diagnostics, Hayward, CA, USA). The RNAscope experiment was performed according to the standard RNAscope protocol without protease treating, and then the slides were refixed with 4% PFA followed by the standard protocol of immunofluorescence.

Immuno‐electron microscopy

Cells were fixed with 4% formaldehyde and incubated with the primary antibody. The HRP secondary antibody was used to detect the primary antibody, and then using DAB (P0203, Beyotime) for the HRP staining, the product is EM‐visible. After that, the samples were prepared for conventional transmission electron microscope. In brief, cells were fixed with 2.5% glutaraldehyde, followed by 1% osmium tetroxide treating, gradient dehydration with alcohol, dehydrated with acetone, and embedded with resin. The sections were imaged at the EM facility of ION (Institute of Neuroscience, Shanghai, China).

Chromatin immunoprecipitation sequencing (ChIP‐Seq) assay

ChIP experiments were performed as previously described (Wen et al, ²⁰⁰⁸), antibody against H3K9me3 (Abcam, ab8898), H3K9me2 (Abcam, ab1220), H3K27me3 (ABclonal, A16199), and H3K27ac (Active Motif, 39,133) were used. Cells were fixed with 1% formaldehyde for 10 min at room temperature. The libraries were prepared using the VAHTS Universal DNA Library Prep Kit for Illumina V3 (Vazyme, ND607‐01) followed by next‐generation sequencing (NGS) using the Illumina HiSeq X Ten system.

Assay for transposase‐accessible chromatin with high‐throughput sequencing (ATAC‐seq)

ATAC‐seq experiments were performed as described (Buenrostro et al, ²⁰¹⁵). For each sample, 4 × 10⁴ AML12 cells were used. The transposition reaction was incubated at 37°C for 40 min. The libraries were prepared using the TruePrep DNA Library Prep Kit V2 for Illumina (Vazyme, TD501‐01) followed by next‐generation sequencing (NGS) using the Illumina HiSeq X Ten system.

Analysis of subcellular protein fractions

Cells were washed once with 1× PBS and were processed into chromatin‐non‐associated and chromatin‐associated fractions as described (Bhatt et al, ²⁰¹²). For RNase A treatment, final concentration of 10 μg/ml RNase A (EN0531, Thermo Fisher Scientific) was added to the lysis buffer. The extracts were diluted with equal volume 1× SDS loading buffer, and proteins in each fraction were detected by western blotting.

Co‐immunoprecipitation assay and RNA immunoprecipitation sequencing (RIP‐Seq)

RIP experiments were performed as previously described (Rinn et al, ²⁰⁰⁷). Briefly, AML12 cells were cross‐linked with UV (4,000 J) and the cell nuclei were extracted and sonicated. Then, incubated the supernatant with anti‐IgG (CST, 2729), anti‐MATR3 (Bethyl, A300‐591A) or anti‐EZH2 (CST, 5246) antibodies overnight at 4°C. After wash for 3 times in RIP buffer and elution in 65°C, the RNA samples were extracted with Trizol. Genomic DNA was digested with DNase I (EN0523, Thermo Fisher Scientific) at 37°C for 1 h. The libraries were prepared using the VAHTS Total RNA‐seq (H/M/R) Library Prep Kit for Illumina (Vazyme, NRM603‐01) followed by next‐generation sequencing (NGS) using the Illumina HiSeq X Ten system. For anti‐MATR3 RIP‐qPCR test, the following primers were used: primer 1‐f (CCCAGCCCCTTGTTAAATACC); primer 1‐r (TGGGGCACACATGAACCTT); primer 2‐f (CACTCTCCTGTCTTGCCTCT); primer 2‐r (CACACACACACAAAGCAATCAAA); primer 3‐f (ACTGTTGGTGGGATTGCAAG); primer 3‐r (TGGGTTCTTTCCAGCTTCTG); primer 4‐f (CCCACTTCACTCCAGTCAGA); primer 4‐r (AGTTTCTGGGCTAATATCCACTT). For anti‐EZH2 RIP‐qPCR test, the following primers were used: Actb‐f (GGTCATCACTATTGGCAACG), Actb‐r (ACGGATGTCAACGTCACACT), Gapdh‐f (AACTTTGGCATTGTGGAAGG), Gapdh‐r (GGATGCAGGGATGATGTTCT), Malat1‐f (CTTTGCGGGTGTTGTAGGTT), Malat1‐r (ACAAAACTGGGAGGTTGTGC), Neat1‐f (TGAGTGCTTTGCCACTGAAT), Neat1‐r (TAACAGCTTCCCCTCTGCTC), Hnf1aos1‐f (ACCTGGCAATCATCCTTTCA), Hnf1aos1‐r (GCTCATTCCTGGGTAGCAGA).

RNA decay assay

RNA transcription was blocked with 5 μg/ml actinomycin D. Cells were collected after 0, 2, 4, 6, 8, and 12 h and RNAs were extracted with Trizol. For RT–qPCR, the following primers were used: AS L1 RNA‐1‐f (CTGTTGAGGGACATCTGGGT); AS L1 RNA‐1‐r (CCAATACCTCTCCTGGGCAT); AS L1 RNA‐2‐f (CCAGCTTGCAATCCCATCAG); AS L1 RNA‐2‐r (CCACCTCACACCAGTCAGAA); Neat1‐f (TGAGTGCTTTGCCACTGAAT); Neat1‐r (TAACAGCTTCCCCTCTGCTC).

Antisense oligonucleotides (ASOs) treatment

Cells were adherently cultured to about 60% confluency, and then, the antisense oligonucleotides were added to the culture medium using Lipofectamine™ RNAiMAX transfection regent (13778030, Thermo Fisher Scientific). The final concentration of ASOs is 50 mM. ASO target for AS L1 RNA was chosen on the consensus sequence on antisense L1_Mus1 RNAs. The following sequences were used: Scramble ASO (CCUUCCCTGAAGGTTCCUCC; Huo et al, ²⁰²⁰); Antisense L1_Mus1 (AS L1)‐ASO (GAAGGTATAGGGAACUUUCG); Sense L1_Mus1‐ASO (CGAAAGTTCCCTATACCUUC); Antisense L1_MA7‐ASO (CUUAAAAAAGAAGGAAAUCC); Sense L1_MA7‐ASO (GGAUUTCCTTCTTTTUUAAG).

Fluorescent recovery after photobleaching (FRAP)

The FRAP was performed as described before (Huo et al, ²⁰²⁰). Briefly, captured one image at prebleach, and then an approximately 1 mm² region was selected to bleach once with maximum light intensity; following images were captured every 1.3 s on a confocal laser scanning microscopy (Leica TCS SP5, Wetzlar, Germany). Image data analyses were performed using LAS AF Lite software (Leica Microsystems, Wetzlar, Germany). The mobile fraction and the half‐life of recovery was calculated as described before (Day et al, ²⁰¹²).

In vitro droplet formation assay

1. Protein purification. The GFP‐mouse MATR3 (or human MATR3 WT/S85C) cDNA was cloned into the pCAG‐flag‐6 × His plasmid and then transfected into HEK293T cells. The proteins were first purified with Ni Agarose 6 FF (AOGMA) with the AKTA system (GE Healthcare Life Sciences) and further purified with anti‐Flag affinity beads. After concentrated in an Amicon Ultracel‐50K spin concentrator to exchange the storage buffer (50 mM Tris–HCl [pH 7.5], 200 mM NaCl, 1 mM DTT, 10% glycerol), the proteins were stored in −80°C after flash freezing in liquid nitrogen.

2. In vitro RNA transcription. cDNA of L1, B1 and MajSAT containing T7 promoter sequence and restriction sites hanging on both sides (in opposite orientation) were cloned to the PUC57 vector. And then, linearize the vector with endonuclease for sense‐ or antisense‐ oriented transcription, separately. RNAs were transcribed in vitro using TranscriptAid T7 High Yield Transcription Kit (K0441, Thermo Fisher Scientific). UTP in this kit was replaced by Cy5‐modified UTP for RNA labeling. The sequences of B1 and MajSAT were obtained from previous reports (Lehnertz et al, ²⁰⁰³; Vassetzky et al, ²⁰⁰³), and the representative L1 sequence was a 318 bp consensus sequence of L1 Md_F2 elements.

3. In vitro droplet formation. Proteins or the mix of protein and RNAs were diluted in PCR tubes at different concentrations in reaction buffer (50 mM Tris–HCl (pH 7.5), 5 mM DTT, 0.1% Triton x‐100 and 50–800 mM NaCl). Then, 3 μl mixed liquid was dropped on the living cell chamber and the images were captured with confocal laser scanning microscopy (Leica TCS SP5, Wetzlar, Germany). Image data analyses were performed using LAS AF Lite software (Leica Microsystems, Wetzlar, Germany).

In vivo protein and RNA concentration estimation

The radius of cell nuclei was measured on the confocal images using LAS AF Lite software (Leica Microsystems, Wetzlar, Germany). The average radius of AML12 cell nuclei is 6.1 μm. So, the average volume of nuclei was calculated to be 9.5 × 10⁻⁷ μl. Assuming the MATR3 are equally distributed in cell nuclei, the 3,240,673 molecules in one cell nuclei can be converted to the concentration of 5.7 μM. For AS L1 RNA concentration, we first measured the total RNA amount of 500 k AML12 cells to be 12,294 ng. Excluding 80% proportion of the rRNA, the total RNA is 0.005 ng per cell. Then, we analyzed proportion of total AS L1 RNAs by RNA‐seq data (library prepared using Ribo‐off depletion kit). On average, 1.52% of mapped reads are AS L1 RNAs. Therefore, the nuclear concentration of total AS L1 RNAs can be calculated to be 74.7 ng/μl. The molecular weight of the 318 nt‐AS L1 RNA fragment is 159,128.6 g/mol. So, the concentration of AS L1 RNAs like 318 nt‐AS L1 RNA fragment in cell nuclei could be 469.4 nM.

In situ Hi‐C experiment

The in situ Hi‐C libraries were prepared as previously described (Rao et al, ²⁰¹⁴). In brief, cells (2.5 millions/sample) were cross‐linked by 1% formaldehyde at 25°C for 10 min. Then, the cell chromatin was digested with 100 U MboI at 37°C overnight. Subsequently, filled in the restriction fragment overhangs with biotin at 37°C for 1.5 h and performed proximity ligation at room temperature for 4 h. After protein degradation and cross‐link reversal, DNA was purified by ethanol precipitation and was sheared to a size of 300–500 bp. After size selection, the biotinylated DNA was pulled down by streptavidin beads and then prepared for Illumine sequencing. The libraries were sequenced via the Illumina HiSeq X Ten system at Annoroad Gene Technology. Two biological replicates were performed for both control and MATR3‐depleted AML12 cells.

SAMMY‐seq experiment

The SAMMY‐seq experiment was conducted as previously described (Marasca et al, ²⁰¹⁶; Sebestyen et al, ²⁰²⁰). In brief, 4 million AML12 cells were used for one sample. After 5 min treatment of 600 μl CSK buffer at 4°C, the supernatant was collected as S1. The pellets were washed and treated with 10 U Turbo DNase in 200 μl CSK buffer for 1 h at 37°C, the supernatant was collected as S2. Then, the pellets were washed and extracted with 2 M NaCl in CSK buffer for 10 min at 4°C. After centrifugation, the supernatant was collected as S3. The pellets were washed and solubilized in 8 M urea, and this fraction was collected as S4. For DNA sequencing, 50% volume of each fraction was extracted by phenol/chloroform and sonicated into 300–700 bp fragments. The DNA libraries were prepared using the VAHTS Universal DNA Library Prep Kit for Illumina V3 (Vazyme, ND607‐01) followed by next‐generation sequencing (NGS) using the Illumina HiSeq X Ten system. The remaining 50% of each fraction was lysed with SDS loading buffer and prepared for western blotting test.

Image processing and quantification

The images of immunofluorescence or RNA FISH were obtained with the confocal laser scanning microscopy (Leica TCS SP5, Wetzlar, Germany). To achieve the comparable image data between groups, the microscope parameters were kept unchanged in each set of experiments. To obtain the super‐resolution images, we conducted confocal laser scanning microscopy using Leica TCS SP8 STED system, followed by processing the images with a deconvolution software Huygens Ver. Image data analysis was performed using LAS AF Lite software (Leica Microsystems, Wetzlar, Germany) and ImageJ software (ImageJ Software, National Institutes of Health, Bethesda, MD, USA).

ChIP‐seq data and ATAC‐seq data processing

CutAdapt v.1.16 (Martin, 2011) was used to remove adaptor sequences from raw reads of ChIP‐seq and ATAC‐seq. Reads were mapped to mouse genome (mm10) using Bowtie (v1.2.3) (Langmead et al, ²⁰⁰⁹). Duplicate reads were excluded and kept only one read for each genomic site. BedGraph files were normalized for total mapped read counts using genomecov from bedtools v2.29.2 (Quinlan & Hall, 2010). The normalized read density bigwig tracks were used for visualization with Integrative Genomics Viewer (IGV; Robinson et al, ²⁰¹¹; Thorvaldsdottir et al, ²⁰¹³). H3K27me3 peaks were called using SICER (Zang et al, ²⁰⁰⁹). To generate the ChIP‐Seq signal distribution for interested regions, we calculated the average ChIP‐Seq signal across these regions.

RIP‐Seq data processing

RIP‐Seq reads were mapped to mouse reference genome (mm10) using HISAT2 (v2.1.0; Kim et al, ²⁰¹⁹), with parameters: ‐‐rna‐strandness RF ‐‐dta. Peaks were called using MACS (v1.4.2) with parameters: ‐‐nomodel ‐‐nolambda (Zhang et al, ²⁰⁰⁸). Read counts were generated using HTSeq‐count (v0.6.1; Anders et al, ²⁰¹⁵). BedGraph files were generated using genomecov from bedtools (v2.29.2; Salton et al, ²⁰¹¹). Integrative Genomics Viewer (IGV) was used for data visualization (Robinson et al, ²⁰¹¹, Thorvaldsdottir et al, ²⁰¹³). To identify the Matr3‐associated L1 RNAs from RIP‐seq, L1 loci with RIP (MATR3 ‐IgG) count number ≥ 10 are kept (Fig 2F).

Hi‐C data analysis

1. Mapping and matrix generation. For Hi‐C data, paired‐end reads were mapped, processed, and iterative correction (ICE) using HiC‐Pro software (version 2.11.4) (https://github.com/nservant/HiC-Pro; Servant et al, ²⁰¹⁵). Read pairs were mapped to the mouse mm10 reference genome (https://hgdownload.soe.ucsc.edu/downloads.html#mouse) with end‐to‐end algorithm and “‐very‐sensitive” option. Singleton, multimapped, dumped, dangling, self‐circle paired‐end reads, and PCR duplicates were removed after mapping. To eliminate the possible effects on data analyses of variable sequencing depths, we randomly sampled equal numbers read pairs from each sample for downstream analyses. Valid read pairs were used to generate raw contact matrices at 100‐kb, 250‐kb, and 1‐Mb resolutions and applied iterative correction (ICE) on them. We converted .all ValidPairs files to .hic files by the script hicpro2juicebox.sh from HiC‐Pro utilities. HiCExplorer (version 3.5.3) (https://hicexplorer.readthedocs.io/en/3.5.3/; Ramirez et al, ²⁰¹⁸) was used to compute correlations between replicates and plot contact maps for ICE normalized matrices.

2. Identification of A and B Compartments. A and B compartments were identified as described previously (Lieberman‐Aiden et al, ²⁰⁰⁹). For normalized contact matrix at 100‐kb resolution, expected matrix was calculated as the sum of contacts per genomic distance divided by the maximal possible contacts and then converted to a Pearson correlation matrix. Then, principal component analysis was performed on the correlation matrix. The first principal eigenvector (PC1) for each bin was used to calculate the overlap ratio with H3K27ac ChIP‐seq peaks to assign each bin to A or B compartment.

3. TAD calling. TADs were called by the insulation score method using cworld software (https://github.com/dekkerlab/cworld-dekker) as previously described (Crane et al, ²⁰¹⁵). Insulation scores were calculated by the cworld script “matrix2insulation.pl.” at 100‐kb resolution matrix with the parameters “‐‐im iqrMean ‐‐is 500000 ‐‐ids 250000 ‐‐nt 0.3.” The topologically associated domains were identified by the cworld script “insulation2tads.pl,” and the 0.3 min boundary strength was set as the threshold.

4. Aggregate contact frequency to compartments. In order to examine the relationship between PC1 values and contact frequency, we sorted PC1 values for all 100 kb bins in decreasing order and divided into 50 equal quantiles (Fig 5B). We then built the 50 × 50 grid of 2D interval that x‐axis and y‐axis are 50 equal quantiles described above. Then in a 50 × 50 2D interval, we assigned the intrachromosomal contacts to pairs of 100 kb loci in each grid and calculated the average contact frequency. We then plotted heatmap to show average contact frequency enrichment in Ctrl and shMatr3 samples and the fold change of contact frequency between them.

5. Average contact frequency for compartment regions. To quantify the average enrichment of contacts at compartment level, we first connected the continuous bins that have the same state of PC1 values in Ctrl defined as compartment regions. We then calculated average log ratio of the observed and the expected contacts for A/B intracompartment regions and intercompartment region pairs between the same type (AA and BB) and different types (AB) in Ctrl and shMatr3. Finally, changes in compartment regions between Ctrl and shMatr3 were calculated and represented as boxplots (Fig 5C and D).

6. Insulation score was computed by cworld software (https://github.com/dekkerlab/cworld-dekker; Crane et al, ²⁰¹⁵) at 50 kb resolution using function matrix2insulation.pl with parameters ‐‐is 1,000,000 ‐‐ids 200,000 ‐‐nt 0.1.

7. TADs classified into MATR3‐AS L1 RNAs associated TADs and nonassociated TADs according to density of anti‐MATR3 RIP‐seq signal of AS L1 RNAs within TADs (Fig 5H). MATR3‐AS L1 RNAs associated TADs classified into four quantile groups by increasing in MATR3‐AS L1 RNAs density (Fig 5I). GO/KEGG analysis was performed using online software metascape (https://metascape.org/gp/index.html#/main/; Appendix Fig S5D and F).

Strand‐specific RNA‐Seq

The libraries were prepared using the VAHTS Stranded mRNA‐seq Library Prep Kit for Illumina (Vazyme, NR602‐01) followed by next‐generation sequencing (NGS) using the Illumina HiSeq X Ten system. Reads were mapped to mouse reference genome (mm10) using HISAT2 v2.1.0 (Kim et al, ²⁰¹⁹), with parameters: ‐‐rna‐strandness RF. FPKM values were calculated using StringTie v1.3.5 (Pertea, Pertea et al, ²⁰¹⁵) based on Refgene annotation from the UCSC genome browser, with parameters: ‐‐rf. Genes with FPKM ≥ 1 were considered to be expressed. Read counts were generated using HTSeq‐count v0.6.1 (Anders et al, ²⁰¹⁵). Differentially expressed genes (DEGs) were calculated using the R package DESeq2 v1.26.0 (Love et al, ²⁰¹⁴), Expression genes were considered differentially expressed if DESeq2 P‐value < 0.1 and | log₂(fold changes) | ≥ 1. BedGraph format files were calculated using genomecov from bedtools v2.29.2 (Quinlan & Hall, 2010). Integrative Genomics Viewer (IGV) was used for data visualization in genome (Robinson et al, ²⁰¹¹; Thorvaldsdottir et al, ²⁰¹³).

SAMMY‐seq data analysis

SAMMY‐seq reads were mapped to mouse reference genome (mm10) using Bowtie (v1.2.3; Langmead et al, ²⁰⁰⁹). Duplicate reads were excluded and kept only one read for each genomic site. We generated the bigwig and bedgraph files of log2(S2/S4) at 10 kb bin using bedtools (v2.29.2; Quinlan & Hall, ²⁰¹⁰). Integrative Genomics Viewer (IGV) was used for data visualization (Robinson et al, ²⁰¹¹; Thorvaldsdottir et al, ²⁰¹³). SAMMY domains were determined using package DNAcopy (http://bioconductor.org/packages/DNAcopy/). The log₂(S2/S4) profiles were smoothed using the default smooth.CNA process. Smoothed samples were then segmented using the segment command with parameter: alpha = 0.01, min.width = 2, undo.splits = “none.”

Statistical analyses

Image data analyses were performed using LAS AF Lite software (Leica Microsystems, Wetzlar, Germany) and ImageJ software (ImageJ Software, National Institutes of Health, Bethesda, MD, USA). Statistical analyses were performed using GraphPad Prism (8.0) and Microsoft Excel. Representative data with ≥ 3 independent experiments were expressed as the mean ± s.d. or mean ± s.e.m. Significance testing was accomplished using unpaired two‐tailed Student's t‐test. No statistical method was used to predetermine the sample size. No data were excluded from the analyses. The western blots were performed in two or three independent biological replicates with similar results, and a representative blot was shown. The ChIP‐seq, RIP‐seq, ATAC‐seq, and Hi‐C were all performed in two independent biological replicates. The details for each experiment are also provided in the figure legends.