Quantitative analysis of multilayer organization of proteins and RNA in nuclear speckles at super resolution

Eukaryotic cells contain both membrane-bound and non-membranous organelles. Non-membranous organelles are proposed to form through non-stoichiometric assembly of proteins and RNAs (Banani et al., 2017). These RNA/protein granules are pervasive in both the cytoplasm (e.g. stress granules and P-bodies) (Buchan and Parker, 2009; Decker and Parker, 2012) and the nucleus (e.g. nucleoli, nuclear speckles and paraspeckles) (Dundr, 2012; Mao et al., 2011a; Zhu and Brangwynne, 2015), and are believed to influence the cellular activities. In the mammalian nucleus, nuclear speckles form one of the most conspicuous RNA/protein granules. Fluorescence imaging of nuclear speckle marker proteins (Fu and Maniatis, 1990) typically reveals 20–50 irregular punctate structures per cell, with sizes ranging from 0.5 µm up to a few microns (Spector and Lamond, 2011). In addition to poly(A)-positive RNAs and noncoding RNAs, nuclear speckles contain ∼200 proteins involved in pre-mRNA metabolism, including pre-mRNA splicing, surveillance and RNA export, as well as non-pre-mRNA processing factors such as transcription factors, translation initiation factors and subunits of RNA polymerase II (Hall et al., 2006; Spector and Lamond, 2011).

Many biological processes have been linked to nuclear speckles, although their exact physiological function remains under debate (Spector and Lamond, 2011). Earlier studies suggested that nuclear speckles act as storage and/or assembly sites for pre-mRNA processing factors. These conclusions are supported by the lack of DNA and low level of transcriptional activity inside the speckles (Fakan and van Driel, 2007; Spector and Lamond, 2011). However, other studies have suggested that speckles play an active role in coordinating gene expression (Hall et al., 2006). A subset of genes tend to cluster around nuclear speckles upon transcriptional activation, and mRNAs of several of the speckle-associated genes colocalize with speckles (Brown et al., 2008; Hall et al., 2006; Han et al., 2011; Hu et al., 2010; Khanna et al., 2014; Moen et al., 2004; Shopland et al., 2002; Smith et al., 1999; Yang et al., 2011), implying that nuclear speckles may serve as an active hub, coordinating transcription of these genes with the processing of their pre-mRNAs (Hall et al., 2006). In addition, nuclear speckles may also be involved in post-transcriptional mRNA splicing, mRNA export and quality control for specific splice-defective pre-mRNAs (Ando et al., 2016; Dias et al., 2010; Girard et al., 2012; Hall et al., 2006; Han et al., 2011; Mor et al., 2016).

The internal organization of protein and RNA components within nuclear speckles has been a topic of investigation. Fluorescence and electron microscopy studies have demonstrated the existence of fine structures within speckles. Several constituents appear in the form of sub-speckles (0.2–0.5 µm) within a speckle (Mintz and Spector, 2000) and are composed of clusters of electron-dense granules of 20–25 nm in size, connected by thin fibrils (Thiry, 1995a,b). In addition, a few components have been reported to occupy different locations within nuclear speckles. For example, poly(A)-positive RNA within the nuclear speckles define a broader area compared to the region defined by speckle marker proteins, such as phosphorylated SR-family splicing regulators [as stained with an anti-SC35 (also known as SRSF2) antibody that preferentially recognizes phosphorylated SR proteins (Fu and Maniatis, 1990)] or splicing co-activator Srm300 (also know as SRRM2) (Carter et al., 1993; Hall et al., 2006). This data suggests that the SC35- or SRm300-stained region within the speckle could form a core domain that is surrounded by a larger shell decorated by poly(A)-positive RNA. In addition, RNA synthesized from several of the speckle-associated genes localized either in speckle periphery or interior (Hall et al., 2006). For example, collagen 1A1-encoding (COL1A1) mRNA was shown to localize to the interior core of the speckle, whereas fibronectin (FN1) or actin B (ACTB) transcripts were predominantly enriched at speckle periphery (Johnson et al., 2000; Shopland et al., 2002). All these observations indicate a non-random organization of at least a subset of the speckle components. Based on these observations, it has been proposed that nuclear speckles are functionally sub-compartmentalized, with transcription and most of the splicing being executed in the speckle periphery, whereas the inner part of the domain plays a vital role in modulating other steps of RNA maturation (Hall et al., 2006). However, a systematic analysis to quantify the relative distribution of various protein and RNA components within the speckle is needed to develop models of the order of operations that might be the hallmarks of speckle functions.

Here, we used multi-color structured illumination microscopy (SIM) to confirm a distinctive multilayer organization of protein and RNA components within nuclear speckles (Gustafsson et al., 2008; Schermelleh et al., 2008). The SIM data reveal a multilayered organization of all labeled components at a higher resolution than in previous studies, with several spliceosome components defining a broader territory compared to splicing regulators and scaffold proteins. A lattice-based computational model is able to reproduce the observed multilayer organization. In this model, spatially organized architectures result from spontaneous self-organization of nuclear speckle components through the interplay of sequence-encoded and solvent-mediated physical interactions between protein and RNA modules. In addition, our data also suggest a role for RNAs in influencing the size of nuclear speckle.

In order to determine the spatial organization of speckle constituents, we analyzed the distribution of the protein SC35, together with an essential spliceosome component (U2 snRNA), and a speckle-enriched long non-coding (lnc)RNA (MALAT1) (Hutchinson et al., 2007; Ji et al., 2003) in human lung diploid fibroblasts (WI-38 cells). The intra-speckle distribution of these three components were measured by performing 3D SIM using a combination of fluorescence in situ hybridization (FISH) (Raj et al., 2008) and immunofluorescence (IF) staining (Fig. 1A; Movie 1). Each of the three components formed small foci within the speckles, potentially corresponding to the sub-speckles. SC35 resided within the core of nuclear speckles, as reported previously (Hall et al., 2006). U2 and MALAT1 defined a broader territory (Fig. 1A). In a subset of cells (34±12%, mean±s.d., with variation among biological replicates), this trend was more obvious, and the majority of the speckles in those cells showed a stronger peripheral distribution of U2 and MALAT1, while SC35 occupied the interior region of the speckles (Fig. 1A). We refer to these speckles as the ‘demixed’ population. In another subset of the cellular population (66±12%) the speckle components did not show a dramatic distinction between the interior and periphery, but U2 and MALAT1 still demonstrated a broader radial distribution compared to SC35 (Fig. S1A). We refer to these speckles as the ‘mixed’ population.

In order to exclude potential artifacts due to the specific fluorophores used to label speckle components, we switched the combination of the fluorophores and components and observed the same phenomena (Fig. S1B). SIM imaging of three additional speckle components [SON protein, U1 snRNA and U2B″ (also known as SNRPB2) protein] showed that scaffold proteins such as SON localized to the speckle interior when compared to U1 snRNA and U2 snRNA-associated U2B″ (Fig. 1B–D).

To obtain a quantitative comparison between different speckle components, we developed an automated approach to analyze the compositional distribution of speckle constituents in thousands of speckles. We first selected individual speckles in 3D by applying an intensity threshold based on the summed intensities from all three channels (Fig. S1C). Since the resolution along the z-axis is worse than the resolution in the xy-plane in SIM images, we identified the middle z-plane of individual speckles based on selected 3D volumes, and only considered the distribution in the associated 2D plane. We determined the geometric center of each speckle and plotted the radial intensity distribution of each component (Fig. S1D). There are two factors that complicated the distribution analysis: (1) the relatively irregular shape of nuclear speckles, and (2) the large variation in the size of speckles compared to other nuclear bodies such as Cajal bodies and paraspeckles, which have more spherical shapes and relatively uniform sizes. We therefore applied two additional thresholds based on 2D area and ellipticity. Thresholding on 2D area excluded oversized speckles that may exist due to the fusion of multiple speckles, while thresholding on ellipticity excluded the speckles that strongly deviated from a spherical shape (Fig. S1E,F).

We superimposed the normalized radial distribution from all selected speckles (demixed and mixed populations) to generate the average radial distribution (Fig. 1E), which revealed a layered distribution of all of the tested speckle components. For comparison, we manually selected cells with obvious demixed and mixed organizations, and plotted the average radial distribution in each category (Fig. S1G). Both the demixed and mixed populations of cells showed layered distribution of speckle components, although the layered distribution was more evident for the demixed case and less evident for the mixed case as compared to the total population-averaged distribution shown in Fig. 1E. The protein SON, a member of the SR-family-like proteins, which has been implicated as a scaffold component in the nuclear speckle (Saitoh et al., 2004; Sharma et al., 2010), exhibited the same radial distribution as SC35. The localization of U1, U2 snRNAs and MALAT1 were indistinguishable from each other (Fig. 1E). U2B″ (Price et al., 1998; Scherly et al., 1990), a component of U2 snRNP complex, was mainly present near the peripheral regions of nuclear speckles (Fig. 1E). Considering the radius at which the density of each component accumulated to 50% of the total (Fig. 1F), the outer layer decorated by U1, U2 and MALAT1, was on average ∼10–15% larger than the adjacent inner layer occupied by SC35 and SON. Considering the average radius of nuclear speckle is ∼200–500 nm, this 10–15% increase (which corresponds to 20–50 nm) is difficult to resolve using conventional fluorescence microscopy with diffraction-limited resolution. The observed multilayer distribution of speckle components was unlikely to be due to differential accessibility of FISH probes or antibodies because U2B″, labeled by immunostaining, displayed the outermost localization, whereas COL1A1 mRNA transcripts (described below), labeled by RNA-FISH, displayed interior speckle localization.

In order to test whether this layered organization shows cell cycle dependence, we performed the same analysis in HeLa cells, which could be synchronized into specific cell cycle stages (Fig. S2A). We imaged cells at the G1/S, S and G2 cell cycle phases (Fig. S2B). We picked G1/S over G1 phase because MALAT1 is largely dispersed in the nucleoplasm during the G1 phase and is enriched in speckles in the G1/S phase (Tripathi et al., 2013). SC35, MALAT1 and U2 displayed similar organizations in speckles in HeLa cells in all tested phases as well as in WI-38 cells (Fig. S2C), suggesting that the layered distribution of speckle components is not limited to a particular cell type or to a specific stage in the cell cycle. In order to determine the stage where proteins such as SON and SC35 define the core of the speckle, we performed co-immunostaining for SON and SC35 in early G1 cells that had just exited mitosis. We found that SON and SC35 had already assembled into nuclear speckles, even in early G1 phase cells (Fig. S2D). The radial distributions of SC35 and SON were found to be the same in all of the cell cycle stages, including early G1, supporting the hypothesis that SC35 and SON define the structural core of nuclear speckles (Fig. S2E).

Next, we investigated whether active transcription contributes to the morphology and organization of nuclear speckles. We inhibited RNA polymerase II-mediated transcription by incubating cells with 5,6-dichloro-1-bold β-D-ribofuranosylbenzimidazole (DRB). As reported above (Fig. S2F), MALAT1 was delocalized from nuclear speckles to the nucleoplasm in transcription-inhibited cells (Bernard et al., 2010). However, the layered organization of other speckle components such as SC35 and U2 remained approximately the same in control versus transcription-inhibited cells (Fig. S2G), except for a slight decrease in the overall speckle size in transcription-inhibited cells (Fig. 2).

A recent biophysical study showed that features of the well-established spatial organization of nucleoli can be reproduced in vitro using only a small subset of protein and RNA components. Importantly, lattice-based computer simulations also recapitulated this spatial organization (Feric et al., 2016). These simulations used a coarse-grained description of protein and RNA molecules based on the known structural and sequence properties of constituent molecules. Interactions between protein and RNA polymers were modeled by using an effective interaction matrix that captured the balance of polymer–solvent and polymer–polymer interactions. The simulations showed that spatial organization results from the differential affinities of protein and RNA components for one another as compared with their affinity for solvent. We investigated the generality of this approach (Feric et al., 2016) for reproducing the multilayer spatial organization of a five or six component system, in this case a minimalist form of the nuclear speckle. The ability to reproduce the multi-layer architecture of speckles would imply that our experimental observations might be consistent with coexisting demixed phases of SC35, SON, U2B″, MALAT1 and U1/U2.

Following the approach of Feric et al. (2016), protein and RNA molecules were modeled as coarse-grained, lattice-based linear polymers made up of distinct interaction modules, as shown in Fig. 3A. The number and type of modules were determined using knowledge of the sequence architectures and known interaction domains (Fig. S3). The criteria used for mapping the sequences of protein and RNA molecules to lattice interaction modules are summarized in the Materials and Methods section. Each protein module represents a folded domain, an intrinsically disordered region, or a combination of both. The RNA modules represented regions that engage in either module or motif-specific interactions or non-specific interactions with other proteins and RNA molecules (see Fig. 3A for a description of the types of interactions). The effective strengths of pairwise interactions between protein and RNA modules were governed by the parameters of an interaction matrix (see Fig. 3B,D for two examples), which quantified the balance of inter-module and module–solvent interactions. The choice of which elements of the interaction matrix are non-zero was made based on known interaction domains and inferences from the sequence composition of disordered regions. The magnitudes of these energies were chosen based on the expected interaction specificities. Specific binding sites were given the strongest interaction strength, collapsing disordered regions were given the second strongest interactions, and non-specific interactions were given the weakest interactions. As discussed below, given our choice for the structure of the interaction matrix, there are numerous parameters that can reproduce the observed spatial organization of the five- or six-component system. This suggests that the absolute strengths of most interactions require very little fine tuning. Specifically, the pattern of localization for each protein/RNA is mainly determined by the structure of the interaction matrix, while the magnitudes of the interaction terms dictate the extent of segregation.

Fig. 3C and E show the radial distributions of molecules within the speckles that were simulated based on the two different interaction matrices shown in Fig. 3B and D, respectively. These two distributions were consistent with the two different types of organization we observed experimentally, as described above (Fig. 1A; Fig. S1A). Fig. 3E represents the mixed case, in which MALAT1, U1 and U2B″ could mix into the core region defined by SON and SC35, whereas Fig. 3C represents the demixed case, in which MALAT1, U1 and U2B″ were excluded from the SON and SC35 core. The change in the architecture of the interaction matrix that was required to discriminate between demixed versus the mixed cases emerged through the promiscuity of the SR modules (Fig. 3B,D, explained below). Considering the relative population of these two distributions, a population average of the simulated distribution (Fig. 3F) correctly reproduced the population average of the distribution observed experimentally (Fig. 1E).

Overall, our computational results are robust to the choices of the actual magnitudes for the effective pairwise affinities in the interaction matrix. The central determinant of the robustness of our results is the structure we prescribe for the interaction matrix. Here, the structure refers to the choice of coarse-grained modules and the choices made for the inter-module interactions, which can be favorable or null. A null interaction implies that the modules prefer to interact with the surrounding solvent and not with one another. In this context, it is worth emphasizing that every module has a finite excluded volume that corresponds to a single lattice site. Specific features of the interaction matrix that are important for producing the experimentally derived radial distributions are: (1) a favorable interaction between SON and SC35 to ensure that these two components make up the core of a multilayered speckle; (2) setting the parameters of the 5′ and 3′ halves of MALAT1 to have different interaction architectures, which yields robust realization of multilayered speckles in simulations; (3) a favorable heterotypic interaction between the SR modules and MALAT1, which is required to observe penetration of MALAT1 into the core; (4) favorable interactions between MALAT1 and U1/U2 as well as SR modules, which are required for the recruitment of U1/U2 to the shell of the speckle; and finally (5) preferential solvation of the linker modules within U2B″ to target this molecule to the shell of the multilayered speckle. Additionally, the proposed promiscuity of interactions involving SR domains from SC35 and SON offers a mechanism to determine whether mixed or demixed spatial organization is obtained. If the SR domains are less promiscuous and preferentially bind the SC35 RNA recognition motifs (RRMs) they generate a finite interfacial tension between the core and shell of the speckle (Fig. 3B,C), facilitating the demixed architectures. However, if the SR domains are more promiscuous in their interactions, as represented by weaker preferential interactions with SC35 RRMs, but enhanced interactions with RNAs, then the interfacial tension is reduced and the mixed case prevails (Fig. 3D,E).

In order to test the involvement of SR and SR-like proteins in the layered organization of speckle components, we performed SIM analyses in cells that were depleted of SRSF1 or SON proteins (Fig. 4A,B; Fig. S4A). Knockdown of SRSF1 led to the demixed organization between MALAT1/U2 and the SC35 core in almost 100% of the cell population (Fig. 4A and C), with a slight decrease in the speckle size (Fig. 2). In SON-depleted cells, MALAT1 showed a more homogenous nuclear distribution, as observed previously (Tripathi et al., 2010) (Fig. 4B), and U2 and SC35 exhibited a strong demixed organization at the periphery (Fig. 4D). Moreover, speckles became spherical in shape with a significant increase in size in SON-depleted cells as compared to control cells (Fig. 2). An earlier study examining SON-depleted cells reported a donut-shaped localization pattern of several speckle components, including the U1 snRNA-associated protein U1-70K (also known as snRNP70), SRSF1, components of exon-junction complex and poly(A)-positive RNA (Sharma et al., 2010). The organization of U2 observed in our work was consistent with the previous studies of other speckle components, whereas SC35 remained in the speckle core in SON-depleted cells. All of these results strengthen our model that SR domain-mediated interactions could modulate the non-random organization of speckle components.

Several lncRNAs are known to function as scaffold components of RNA–protein assemblies (Mao et al., 2011b; Nakagawa et al., 2012; Souquere et al., 2010). In order to test whether speckle-enriched MALAT1 influences the organization of other speckle constituents, we imaged SON, SC35 and U2 in MALAT1-depleted cells (Fig. 5A,B; Fig. S4D). Our data, along with that from previous studies (Clemson et al., 2009; Tripathi et al., 2010), showed that MALAT1 is dispensable for the formation of nuclear speckles even though MALAT1 can facilitate the recruitment of several of the speckle components to a specific sub-nuclear region (Tripathi et al., 2012). However, MALAT1-depleted cells displayed an overall decrease in the size of the SC35-stained core area of the speckle (Fig. 2) and a reduced localization of SON to nuclear speckles (Fig. 5A,C). Analysis of SON nuclear localization revealed that the density of SON within the speckle was fivefold higher than in the rest of the nucleus (Fig. 5C). MALAT1 depletion resulted in an ∼50% reduction in the localization of SON to the speckles. However, this reduction in SON density was also observed across the whole nucleus upon MALAT1 depletion (Fig. 5C), indicating that cellular expression level of SON may be affected by MALAT1 depletion. As SON is reported to be a nuclear speckle scaffold protein (Sharma et al., 2010), the reduction in the size of nuclear speckles in MALAT1-depleted cells may partially be a result of the overall reduction in cellular SON levels.

The layered organization of SC35 and U2 became slightly more demixed in the absence of MALAT1 (Fig. 5D). However, we speculate that it is not likely that the more demixed organization of speckle components was due to a direct effect of MALAT1 depletion because in transcription-inhibited cells, although MALAT1 was delocalized from the speckles, the relative distribution between SC35 and U2 remained identical that in to normal cells (Fig. S2G). Given the fact that MALAT1 knockdown affects the expression and activity of many genes, including that of SR proteins, the minor change in the speckle organization that we observed in MALAT1-depleted cells could be due to changes in overall gene expression (Malakar et al., 2017; Tripathi et al., 2010, 2013).

In order to determine whether speckle-accumulated RNA molecules (i.e. RNA transcripts from speckle-associated genes) preferentially localize to a specific layer within nuclear speckles, we quantified the distribution of RNAs of several speckle-associated genes, including those encoding Collagen, type I, α1 (COL1A1), fibronectin (FN1), β-actin (ACTB) and lamin A/C (LMNA). Quantification was achieved by performing RNA-FISH with exon-specific probes along with SC35 staining (Fig. 6A; Fig. S5). These probes hybridize to both spliced and unspliced transcripts. We selected speckles based on the SC35 signal as described above, and calculated the distance (D) between the center of each RNA accumulation site (referred to as the RNA site hereafter) to the center of its nearest speckle. If D was smaller than the maximum radial distance of that speckle (R_Max), we considered the RNA site to be associated with that speckle. Consistent with previous reports (Smith et al., 1999), COL1A1 and ACTB RNAs showed a higher percentage of speckle association than FN1 and LMNA RNAs (Fig. S5D). Among the four genes tested, only the COL1A1 RNA site was within the interior region of the speckle, with an average relative distance to the center of the associated speckle (defined as D_Rel=D/R_Max) of ∼0.2, whereas the other three RNA sites had an D_Rel of ∼0.6, residing within the RNA-rich region (Fig. S5E). As reported previously (Hall et al., 2006), COL1A1 RNA accumulated within the interior region of the speckle while the COL1A1 gene remained at the periphery, suggesting that RNA is imported into the speckle interior after being transcribed at the periphery (Fig. S5F). For the four genes tested, the speckles containing RNA sites were larger than the average speckle in the same nucleus (Xing et al., 1995), but to different degrees (Fig. 6B). The increase in speckle size did not correlate with the total cellular levels of a particular RNA determined by using real-time quantitative RT-PCR (RT-qPCR) (Fig. S4B).

In order to test the correlation between the speckle size and the amount of poly(A)-positive RNA that was present within a speckle, we co-stained poly(A)-positive RNA together with SC35 (Fig. 6C). Consistent with the previous results (Carter et al., 1993; Hall et al., 2006), poly(A)-positive RNA defined a broader region than the SC35-decorated inner core of the speckle (Fig. 6D). When plotting the signal from poly(A)-positive RNA as a function of the SC35-defined area, we observed a strong positive correlation between the size of the speckle and the amount of poly(A)-positive RNA present in the speckle (correlation coefficient ∼0.56) (Fig. 6E), again supporting the idea that accumulation of RNA inside the speckle positively correlates with speckle size.

Finally, to determine whether RNA accumulation within a speckle contributes to changes in speckle size, we examined the localization of COL1A1 RNA in splicing-defective cells and MALAT1-depleted cells. Depletion of U2 snRNA reduced pre-mRNA splicing, as revealed by RT-qPCR analysis performed with intron–exon junction-specific primers against COL1A1 pre-mRNA (Fig. S4C), and resulted in the accumulation of unspliced transcripts. We labeled U2 along with COL1A1 RNA and SC35 in both control and U2-depleted cells (Fig. 6F). Because U2 depletion efficiency showed a large cell-to-cell variation, we selected cells with a knockdown efficiency that was greater than 75% for further analysis (Fig. S6A). Compared to control cells, the speckle size increased in U2 snRNA-depleted cells (Fig. 2). In addition, U2-depleted cells showed an increase in the number of COL1A1 RNA sites per cell (∼3 RNA sites per cell, compared to ∼1.8 sites per cell in the mock case) along with a concomitant increase in the amount of COL1A1 RNA accumulated within each speckle (Fig. 6G). Such a dramatic increase in the size of COL1A1-containing speckles required a near complete depletion of U2 snRNA (Fig. 6F) because cells with a partial depletion of U2 displayed neither an aberrant accumulation of COL1A1 RNA nor a dramatic change in speckle size (Fig. S6B). In line with our observations under U2 knockdown, cells depleted for SF3b1 (also known as Sap155; a core component of U2 snRNP) showed a very similar size expansion for nuclear speckles (Fig. 2; Fig. S6C). Moreover, both in the control and U2 or SF3b1-depleted cells, EXOSC10 (also known as RRP6 or PM/Scl-100), an essential component of the nuclear exosome, was predominantly localized in the nucleolus and was excluded from nuclear speckles (Fig. S7), suggesting that speckles do not act as sites for the degradation of aberrantly spliced RNAs. MALAT1-depleted cells, on the other hand, showed a decrease in (1) the number of COL1A1 RNA sites per cell (∼1 per cell, compared to ∼1.6 sites per cell in the mock case), (2) the total amount of COL1A1 RNA at the accumulation site (Fig. 6G), and (3) the cellular levels of un-spliced and spliced COL1A1 transcripts (Tripathi et al., 2013) (Fig. S4D). All of these results support the hypothesis that the accumulation of RNA contributes to the growth of nuclear speckles.

Several non-membranous organelles are known to be internally organized into distinct sub-compartments. For example, nucleoli are organized into at least three distinct subdomains that are responsible for specific stages of ribosome biogenesis and maturation (Boisvert et al., 2007; Pederson, 2011). Recent studies have also revealed non-random organization of components within the paraspeckle sub-nuclear domains (West et al., 2016). The layered structure, characterized by a core-shell architecture, has also been found in cytoplasmic ribonucleoprotein (RNP) bodies (Jain et al., 2016; Wheeler et al., 2016). Our results demonstrate that the layered organization also exists in nuclear speckles, in which SON and SC35 assemble into the core region, U1, U2 and MALAT forms a second layer, and U2B″ loosely associates with the peripheral regions of nuclear speckles. Previous studies have revealed that regions of chromatin that contain highly transcribed genes tend to associate with the periphery of nuclear speckles, leading to a model in which pre-mRNA splicing of the speckle-associated genes occurs at the speckle periphery (Brown et al., 2008; Hall et al., 2006; Han et al., 2011; Hu et al., 2010; Khanna et al., 2014; Yang et al., 2011). Our findings that snRNAs and proteins in snRNPs tend to concentrate at the periphery of nuclear speckles further support this model.

Computational simulations suggest that the spatial organization of nuclear speckles observed in vivo might emerge from the interplay of favorable sequence-encoded intermolecular interactions and the preferential solvation among the various protein and RNA modules. The interaction matrix used to recapitulate this organization is in accordance with a variety of observations that have been reported in the literature. SC35 and SON are members of SR and SR-like protein families respectively; both proteins possess SR domains that are enriched in arginine and serine repeats and in addition possess an RRM (Sharma et al., 2010; Shepard and Hertel, 2009; Zhong et al., 2009). In addition, SON has distinctive repetitive sequence elements (Fig. S3) that are critical for speckle formation (Sharma et al., 2010). Therefore, the multivalence of SON and SR repeats appear to provide a crucial driving force for the formation of the speckle core region. The multivalence of MALAT1 enables its interactions with core and shell components through interactions with SR proteins (Änkö et al., 2012; Miyagawa et al., 2012; Sanford et al., 2009; Tripathi et al., 2010) and base pairing with RNAs (Engreitz et al., 2014; Lu et al., 2016; Nguyen et al., 2016), thus promoting the observed multilayer structure. The SR domains are known to undergo extensive and reversible phosphorylation (Lin et al., 2005; Misteli and Spector, 1997; Misteli et al., 1998; Tripathi et al., 2010; Xie et al., 2006), providing a putative mechanism through which the cell could dynamically modulate the ‘promiscuity’ of the SR domain. Finally, MALAT1 and snRNAs undergo different levels of modifications (Bringmann and Lührmann, 1987; Liu and Pan, 2016; Liu et al., 2015; Roundtree and He, 2016), which could lead to changes in RNA–protein interactions and therefore contribute to the heterogeneity of the compositional organizations.

Our knockdown experiments have provided hints on the involvement of different speckle constituents in the formation and maintenance of nuclear speckles. SON and SC35 are important for forming the core region of nuclear speckles, which is supported by two lines of evidence: (1) that SON and SC35 assemble into the speckle at early G1 phase, whereas MALAT1 is dispersed in the nucleoplasm; (2) and that SC35 was localized in the core region of nuclear speckles under all tested perturbation conditions, including upon speckle-resident protein and RNA knockdown, and transcription inhibition. Based on our results, we speculate that SR proteins play an important role in maintaining the ratio of mixed (in which MALAT1 and U2 mix into the core region of the nuclear speckle), and demixed (in which MALAT1 and U2 are excluded from the core region of the nuclear speckles) populations of nuclear speckles. Knockdown of SRSF1 or SON protein changes the populations to being almost 100% demixed. These data support our predictions from computational simulations in that interactions between SR domains, RRMs and RNAs tune the balance between the mixed and demixed modes of organization. SON depletion has an even larger effect on speckle morphology, including inducing an increase in speckle size and a more spherical shape compared to the normally irregular nature of nuclear speckles in control cells. We interpret this observation to indicate that the multivalence of SON may promote changes in the material properties of nuclear speckles, such as the underlying dynamics of the speckles (e.g. liquid-like versus solid-like) (Banani et al., 2016, 2017), an effect which remains to be further investigated. The effect of MALAT1 depletion on nuclear speckle structure is likely to be an indirect effect, given MALAT1 is one of the RNA molecules that is most frequently observed to globally interact with nascent transcripts (Engreitz et al., 2014; Lu et al., 2016; Nguyen et al., 2016) and chromosomes (Li et al., 2017). Indeed, we observe that depletion of MALAT1 affects the cellular level of SON (at the protein level), and COL1A1 (at the RNA level). We therefore speculate that the changes in the overall structure of nuclear speckles caused by MALAT1 depletion, including the overall size reduction and the shift towards the more demixed state, are more likely due to the changes in the abundance of nuclear speckle components rather than being a direct effect from MALAT1.

Previous studies have demonstrated the localization of the Hsp70 (HSPA1A) gene to the proximity of nuclear speckle upon its activation, as well as transcription-dependent size expansion of HSPA1A-associated speckle size upon heat shock (Hu et al., 2010; Khanna et al., 2014). Consistent with these results, our measurements on multiple genes reveal that the transcription of speckle-associated genes and the accumulation of RNA in speckles contribute to the size heterogeneity of nuclear speckles. We find that, in general, speckles size strongly correlate with the total poly(A)-positive RNA signal within the speckles at individual speckle level. As specific examples, speckles containing accumulated transcripts from highly transcribed genes, including COL1A1, ACTB, FN1 and LMNA are larger than the average speckles. Previous studies have documented that most nuclear speckles are associated with genes and contain RNA transcripts from multiple genes (Hall et al., 2006; Shopland et al., 2002). Our observations here are in agreement with these previous findings, and further reveal a weak positive correlation between the size of speckles containing transcripts from these specific genes and the total signals from the specific transcripts at single speckle level (correlation coefficient ∼0.27) (Fig. 7A). We reason that because speckles contain transcripts from multiple genes, the correlation between the speckle size and the amount of RNA accumulated from a specific set of genes is weakened compared to the strong correlation between the speckle size and the amount of accumulated total poly(A)-positive RNA. Nevertheless, the observation that speckles containing transcripts from any one of the four genes show an increase in average size (albeit to different extents) is likely because they are among the most active genes in fibroblast cells. We sorted the RPKM values from RNA-sequencing for all of the annotated genes in fibroblast cells (Table S1) (Marthandan et al., 2016), and found that COL1A1 is among the top 0.2% abundant, ACTB among the top 0.05%, FN1 among the top 0.2% and LMNA among the top 0.8%. However, it should still be noted that the propensity of speckle accumulation is not strictly correlated with the cellular RNA abundance, as shown in the case of COL1A1 and ACTB from imaging and RT-qPCR results. What drives the RNA transcripts to accumulate in nuclear speckles needs to be further investigated. Post-transcriptional processing events, such as pre-mRNA splicing and RNA export, could play a role in mediating the association of RNA with speckles.

Interestingly, when considering all four tested genes at the single speckle level, we also observe a weak negative correlation (correlation coefficient∼–0.34) between the speckle size and D_Rel (location of the RNA site within the speckle as defined in the Results) (Fig. 7B) and a weak negative correlation (correlation coefficient∼–0.32) between the amount of accumulated RNAs and D_Rel (Fig. 7C). Including the positive correlation between the speckle size and the amount of accumulated RNAs described above, these correlations suggest that more accumulated RNAs tend to localize toward the center of the associated speckle, resulting in an increase in the speckle size, leading to our proposal of a dynamic interaction between the RNA transcript and the nuclear speckles (Fig. 7D). Once a transcriptionally active gene is recruited to the nuclear speckle periphery, nascent pre-mRNA transcribed from this gene accumulates near speckles. The co-residence of nascent pre-mRNA with snRNPs at the periphery of the nuclear speckle could enhance the effective processing of pre-mRNAs. When a significant amount of pre-mRNA or partially processed pre-mRNA accumulates near a particular speckle, it facilitates the recruitment of more speckle components to that speckle, either from the nucleoplasm or from the neighboring nuclear speckles, leading to the further growth of the speckle (Eils et al., 2000; Hu et al., 2009; Misteli et al., 1997). This dynamic reallocation of nuclear resources is reminiscent of how transcription can modulate nucleoli assembly (Berry et al., 2015). Since recruitment of new components to a speckle can occur from all directions, the RNA accumulation site shows a relative migration towards the center of the speckle. This is exemplified in the case of U2 snRNA-depleted cells, where a large amount of COL1A1 transcripts are enriched at the interior region of the associated speckles (Fig. 6F).

In conclusion, by using super-resolution imaging of multiple nuclear speckles, we characterized in detail a multilayered distribution of several of the nuclear speckle components. This multilayered spatial organization of nuclear speckles might emerge from the interplay of favorable sequence-encoded intermolecular interactions between various speckle components. The morphology and organization of speckles can be directly or indirectly tuned by each of the components. Finally, RNA accumulation in the speckles positively regulates the size of the speckle at the single speckle level.

WI-38 cells were cultured in minimal essential medium (MEM) supplemented with 1 mM sodium pyruvate, 100 U/ml penicillin, 100 µg/ml streptomycin, 10% fetal bovine serum (FBS) and 1% non-essential amino acid (NEA). HeLa cells were cultured in Dulbecco's modified Eagle's medium (DMEM) containing high glucose, supplemented with 1 mM sodium pyruvate, 100 U/ml penicillin, 100 µg/ml streptomycin and 10% FBS. For imaging, cells were immobilized on poly-L-lysine-treated borosilicate chambered coverglass (Thermo Scientific Nunc Lab-Tek). These cells were obtained from the ATCC and maintained as per their recommendations.

MALAT1 and U2 snRNA were depleted by use of modified DNA (antisense DNA oligonucleotides; ASOs) (Ionis Pharmaceuticals, USA). The ASO is a phosphorothioate modified 5-10-5 MOE gapmer with five 2′-O-methoxy-ethyl nucleotides on either end and ten deoxy-nucleotides in the gap. Oligonucleotides were transfected into cells in two rounds with a gap of 24 h at a final concentration of 100–200 nM. Lipofectamine RNAiMax reagent was used for transfection (Invitrogen, USA). SRSF1 knockdown was performed by stably transducing WI-38 cells with lentiviral-based shRNAs. shRNA was cloned into plko.1 vector using protocols from Addgene (plasmid #10878). A control shRNA clone was purchased from Addgene (plasmid #10879).

SRSF1 shRNA sequences were: SRSF1-sh-F, 5′-CCGGAAGGATTGTGGAGCACATTTCCTCGAGGAAATGTGCTCCACAATCCTTTTTTTG-3′ and SRSF1-sh-R, 5′-AATTCAAAAAAAGGATTGTGGAGCA CATTTCCTCGAGGAAATGTGCTCCACAATCCTT-3′.

Depletion of SON and SF3B1 were performed by using double-stranded siRNAs against SON (Sharma et al., 2010) and SF3B1 (hs.Ri.SF3B1.13.1, IDT, USA). siRNAs were transfected into cells two times with a 24 h interval, at a final concentration of 100 nM for SON and 20 nM for SF3B1 using Lipofectamine RNAi MAX (Invitrogen). Cells were collected 48 h after the second transfection for further analyses.

HeLa cell synchronization was carried out by double thymidine block and release. Briefly, cells were cultured in 2 mM thymidine for 24 h as the first block. Then, cells were washed and released for 12 h in fresh medium. Next, the second block was performed by adding 2 mM thymidine for another 24 h to synchronize cells at G1/S. Then, fresh medium was replaced to permit release into the cell cycle; cells were collected at 0 h (G1/S), 4 h (S phase) and 8 h (G2 phase) after release.

Total RNA was extracted using Trizol reagent (Invitrogen) and cDNA was synthesized using a reverse transcriptase kit (Applied Biosystem). Quantitative RT-qPCR was performed with the StepOne plus system to quantify transcript level. Specific primer sets were designed for each gene as follows: MALAT1-F, 5′-GACGGAGGTTGAGATGAAGC-3′ and MALAT1-R, 5′-ATTCGGGGCTCTGTAGTCCT-3′; COL1A1-F, 5′-GTGCGATGACGTGATCTGTGA-3′ and COL1A1-R, 5′-CGGTGGTTTCTTGGTCGGT-3′; COL1A1-Intron24-Exon25-F, 5′-TGACCGGCGGGACCCTAA-3′ and COL1A1-Intron24-Exon25-R, 5′-CATTGTCCTGTCTGCCTCCCTG-3′.

Immunoblotting was performed by following the protocol described previously (Tripathi et al., 2010). Antibodies used for western blotting (wB) include antibodies against SRSF1 (mIgG, mAb96, WB, 1:1000) (Hanamura et al., 1998), SF3B1/SAP155 (Santa Cruz Biotechnology, sc-514655, 1:200) and α-tubulin (1:10000, Sigma-Aldrich, T5168).

Chemically synthesized single-molecule (sm)FISH probes were designed by using the Stellaris Probe Designer and ordered from Biosearch Technologies (http://www.biosearchtech.com). Probes were 20 nucleotides in length with GC content of ∼45% and 3′ amine modification. Each probe was synthesized and dissolved individually in water to a final concentration of 100 μM in a 96-well plate. An equal volume of each probe were combined for labeling in 0.1 M sodium bicarbonate (pH 8.5). Alexa Fluor-conjugated succinimidyl ester (Life Technologies) was first dissolved in 1–5 µl DMSO and then mixed with the probe solution. The molar ratio of dye to probe is ∼25:1. The labeling reaction was incubated in the dark at 37°C overnight. Reactions were quenched by adding a 1/9th reaction volume of 3 M sodium acetate (pH 5). Labeled probes were purified from unconjugated free dye, first by ethanol precipitation and then further purified with a P-6 Micro Bio-Spin Column (Bio-Rad). All probes were labeled with more than 80% labeling efficiency. The number of probes used for each RNA is as follows: MALAT1, 32 probes; U2, 8 probes; U1, 6 probes; COL1A1, 40 probes; ACTB, 36 probes; LMNA, 31 probes; FN1, 39 probes.

Secondary antibodies against mouse and rabbit IgG (Jackson ImmunoResearch) were labeled with Alexa Fluor succinimidyl ester for immunofluorescence staining; 24 μl of antibody (at 1 mg/ml concentration) was mixed with 3 μl of 10× PBS and 3 μl of 1 M sodium bicarbonate (pH 8.5), and added with 0.001–0.005 mg of Alexa dye. The reaction was incubated at room temperature for 30 min, and labeled antibody was purified on a P-6 Micro Bio-Spin column pre-equilibrated with 1× PBS. The antibody to dye ratio was between 0.5 and 2.5 from batch to batch.

FISH and immunofluorescence staining were performed as previously described (Raj et al., 2008). Briefly, cells were fixed with 4% paraformaldehyde (Electron Microscopy Sciences) in 1× PBS for 10 min at room temperature and permeabilized with solution containing 0.5% Triton X-100 and 2 mM vanadyl ribonucleoside complexes (VRC New England Labs) in 1× PBS for 10 min on ice. Cells were then temporarily stored in 70% ethanol until hybridization with FISH probes. Cells were washed once with FISH wash solution [10% formamide (Ambion) in 2× saline-sodium citrate (SSC)]. Labeled probes were mixed with 50 μl hybridization buffer [10% dextran sulfate (Sigma-Aldrich), 1 mg/ml yeast tRNA (Sigma-Aldrich), 0.2 mg/ml BSA (Ambion), 2 mM VRC, 10% formamide in 2× SSC]. Labeled probes for all transcripts and MALAT1 were 4 nM each and for U1 and U2 were 10 nM each in the final reaction. Probe solution was then added to the chamber. An 18×18 mm coverslip was used to cover the bottom of the chamber such that the probe solution was evenly distributed to the entire chamber. Hybridization reactions were incubated in the dark at 37°C overnight. On the second day, the cells were washed twice with FISH wash solution.

FISH-labeled cells were then subjected to fixation again at room temperature for 5 min to prevent probe dissociation during immunostaining. Before immunostaining, cells were first incubated with blocking solution [1% BSA (Gibco) in 1× PBS] at room temperature three times with a 10-min incubation each time. Primary antibody was mixed with blocking solution using the following dilution factors: mouse antibody against SC35 (Fu and Maniatis, 1990), 1:500, rabbit antibody against SON (Sharma et al., 2010), 1:500; mouse antibody against U2B″ (clone: 4G3), 1:25 (Habets et al., 1987, 1989; Mattaj et al., 1986), mouse anti-EXOSC10 antibody (sc-374595, Santa Cruz Biotechnology), 1:50. 50 μl antibody solution was used for each chamber, and covered with an 18×18 mm glass coverslip during the 1-h incubation at room temperature. Cells were washed with blocking solution three times with a 10-min incubation each time. Labeled secondary antibody was diluted with the blocking solution by 100-fold. Incubation with secondary antibody is performed in the same manner as for primary antibody for ∼40 min. After staining, the cells were washed three times with 1×PBS with a 10-min incubation each time. Cells were stored in 4× SSC at 4°C until imaging.

Image acquisition was performed on a commercial instrument (SR-SIM Elyra system with Axioobserver Z1 microscope from Zeiss); 488 nm, 561 nm and 632 nm lasers were used for three-color imaging. The exposure time was 100 ms for all three channels. The excitation power was adjusted, but was no larger than 30% of the maximum power, so that signals can be maximized without saturating the camera. For knockdown experiments, the exact same laser power was used for knockdown samples as for corresponding negative controls. SIM images were reconstructed by using commercial software (ZEN 2011 from Zeiss). Fluorescent nanodiamonds (140 nm diameter, Acedemia Sinica, Taipei, Taiwan) were nonspecifically attached to the surface of the chamber and were used for mapping of the different channels.

Custom MATLAB code was developed to automate nuclear speckle selection, processing and analysis (available from the corresponding author upon request). First, grayscale images were constructed from the mean intensity of selected channels. Specifically, for distribution analysis for all components, all three channels were used, and for RNA accumulation site analysis, channels corresponding to RNA and to SC35 were used separately to identify each RNA accumulation site and speckle. An overall fluorescence intensity histogram was produced to assist the determination of a threshold for speckle identification. Nuclear speckles in 2D were identified by a single intensity threshold for each slice in the z-stack. The 2D binary images after thresholding were further processed by filling and opening binary operations to remove internal voids and shot noise. After combining all binary images in the z-stacks into one 3D binary matrix, individual 3D nuclear speckles were identified and indexed by connected component analysis. Each identified object serves as the 3D mask for subsequent quantitative structural analysis of each nuclear speckle. Each speckle was individually saved with all three fluorescence channels. The center of each speckle was determined from the geometric centroid of its mask, this removes the ambiguity of the interpretation of grayscale intensity and potential biasing effect. The normalized radial distribution function of the intensity of each fluorescent channel is calculated based on this center. The normalized radial distribution carries the most meaningful result when the speckle is shaped closest to a sphere.

The design and implementation of the simulations were identical to that of Feric et al. (Feric et al., 2016). The architectures of coarse-grained protein and RNA molecules were designed to mimic the modular nature of these molecules. SC35 is a 221-residue protein. It has an SR-rich domain (92 residues, 80% S/R) and an RNA-binding domain. The RNA domain has specificity for sites on MALAT1.

We modeled SC35 as a linear polymer of seven beads, six of which spanned the SR region and one of which was an RNA-recognition module (RRM) that interacts specifically with sites on MALAT1. SON is a 2426-residue protein. Its sequence encompasses charge tracts, nearly perfect repeat regions and an SR-rich region. The mapping of the sequence features to a 20-bead linear polymer was derived from analysis of the sequence features of different regions as described in Fig. S3. This collection of large repetitive blocks of sequences are observed to collapse in simulations, which motivated the architecture of three blocks of five beads that are attractive toward similar beads. The final five beads model the SR-rich domain.

U2B″ is a 225-residue protein comprising two RRMs and a linker enriched in charged residues. Accordingly, this molecule was modeled by using four beads, two corresponding to the RRMs interspersed by two solvated beads corresponding to the linker residues. U1 and U2 are 100-nucleotide RNA molecules. These molecules are indistinguishable in the coarse-grained simulations. They were modeled as linear polymers, with seven beads that were either RNA interaction modules (RIMs) or protein interaction modules (PIMs). Finally, MALAT1, which is a 6000–7000-nucleotide RNA is known to interact with the SR family of proteins, including SRSF1. Most of the SRSF1 interactions sites are located on the 5′ region of MALAT1. Additionally, sites on MALAT1 also form base pairs with U1/U2, which are located on the 3′ end. Accordingly, MALAT1 was modeled as a linear di-block copolymer comprising 25 beads and interactions of these beads with other beads that corresponded to protein and RNA interaction modules were as described in Fig. 2. MALAT1 also encompasses beads that are non-specific RNA modules (NRMs) that interact weakly with SR-rich modules on SON and SC35.

There were 1600 linear polymers in each of the lattice-based Monte Carlo simulations. Of these, 200 corresponded to SON, 400 to SC35, 200 to MALAT1, 400 to U1/U2, and 400 to U2B″. The simulations were initiated with all the molecules dispersed uniformly across the lattice. The system was allowed to evolve via a collection of Monte Carlo moves that were accepted or rejected based on the Metropolis criterion. A bond can form between pairs of modules that occupy adjacent lattice sites if the interaction energy for the pair of modules is negative. Each module can make only one bond at a time.

The probability densities were calculated using an approach that mimics the calculation of densities from experimental data. We accomplished this by identifying the geometric center of the largest cluster in the simulation and analyzing the modules within eight lattice units along the z-dimension. We calculated the radial density of modules from the geometric center of the speckle in the central 2D plane through the center of the speckle with a finite thickness. As all three dimensions were equivalent in simulation, the results were an average of repeating this over each of the three dimensions.

We thank the core facilities and service at Carl R. Woese Institute for Genomic Biology at University of Illinois Urbana-Champaign for SIM microscopy and imaging reconstruction software, Drs P. Bubulya (Wright State University, USA), Fu X-D (University of San Diego, USA), W. van Venrooij (Radbound University Nijmegen, Netherlands) for the gift of reagents, and Dr. V. Tripathi for technical assistance.