TIS granules have a mesh-like morphology and form through assembly of the RNA-binding protein TIS11B (Ma and Mayr, 2018). When we initially discovered TIS granules, we expressed mCherry-tagged TIS11B in HeLa cells and observed that the transfected TIS granules largely recapitulated the mesh-like three-dimensional structure of endogenous TIS granules (Figure 1A). At the time, we performed fluorescence recovery after photobleaching (FRAP) after 16 hr of transfection and observed 20% of fluorescence recovery in 120 s (Ma and Mayr, 2018). It is an accepted fact that surface tension drives liquid-like phase-separated condensates to adopt sphere-like shapes as spheres have a minimal surface area for a given volume (Hyman et al., 2014). Therefore, we initially thought that TIS granules are mesh-like because they are gel-like (Ma and Mayr, 2018).

Figure 1 with 2 supplements see all

Download asset Open asset

More recently, we repeated the FRAP experiment at 5 hr after transfection and observed 43% fluorescent recovery of GFP-tagged TIS11B in 10 s (Figure 1—figure supplement 1A). This observation suggested that at this time point TIS granules generated from transfected constructs have somewhat dynamic protein components despite being irregularly shaped. This observation is reminiscent of L-bodies in frog oocytes, which have a mesh-like morphology but dynamic protein components (Neil et al., 2020). Our FRAP experiments suggested that the mesh-like shape of TIS granules obtained by transfection cannot simply be explained by gel-like biophysical properties. It is important to point out that the material properties of endogenous TIS granules are currently unknown as all experiments so far have been performed with fluorescently tagged TIS11B constructs (Ma and Mayr, 2018).

As TIS11B binds to AU-rich elements and the localization element of the Vg1 mRNA is highly AU-rich, we hypothesized that mRNAs play a role in determining the mesh-like morphology of condensates. TIS11B contains a double zinc finger RNA-binding domain (RBD). When we introduced different point mutations to disrupt RNA binding (Figure 1—figure supplement 1B–C; Lai et al., 2000), the mesh-like assemblies were turned into sphere-like condensates that are no longer intertwined with the ER (Figure 1B, Figure 1—figure supplement 1D, E). This suggested that the mesh-like organization of TIS granules requires the recruitment of mRNAs.

Expression of the TIS11B RBD alone is not sufficient for condensate formation (Figure 1—figure supplement 2A). To assess the importance of the TIS11B RBD for mesh-like condensate formation, we tested if it results in network formation in the context of different multivalent domains. Expression of SUMO-SIM generates sphere-like condensates in the cytoplasm (Figure 1C; Banani et al., 2016). However, when we fused SUMO-SIM to the TIS11B RBD (SUMO-SIM-TIS), we observed a filamentous condensate that is not intertwined with the ER (Figure 1C, Figure 1—figure supplement 2B). This was an important result as it demonstrates that intertwinement with the ER is not necessary for mesh-like condensate formation.

Another well-studied multivalent domain is the intrinsically disordered region (IDR) of FUS (Kato et al., 2012). Expression of the FUS-IDR alone did not generate condensates (Figure 1D). When we fused the RBD of TIS11B to the IDR of FUS to generate FUS TIS, we observed cytoplasmic mesh-like condensates that look very similar to wild-type TIS granules (Figure 1D). In the context of FUS-TIS, the RBD of TIS11B is functional as it recruits the same mRNAs to the condensates (Figure 1—figure supplement 2C).

Importantly, both FUS-TIS and SUMO-SIM-TIS show fast fluorescence recovery in FRAP experiments (Figure 1E), indicating that the filamentous networks have highly mobile protein components and are not aggregates. These experiments reveal that in the context of various multivalent domains the TIS11B RBD generates mesh-like condensates in vivo. Based on the current knowledge, this observation represents a paradox as it is thought that biomolecular condensates with dynamic protein components should be sphere-like because of surface tension (Hyman et al., 2014). However, the seemingly liquid-like condensates generated by FUS-TIS and SUMO-SIM-TIS have mesh-like morphologies in cells. This observation motivated us to investigate how filamentous but liquid-like condensates are generated. To do so, we chose to perform in vitro reconstitution experiments with FUS-TIS as a model system as it formed mesh-like condensates and showed a highly dynamic behavior in the FRAP experiments in cells.

We recombinantly expressed and purified FUS-TIS (Figure 2—figure supplement 1A–C). FUS-TIS phase separates into sphere-like condensates that are liquid-like (Figure 2—figure supplement 1D, E). We then added to FUS-TIS in vitro transcribed 3′UTRs of mRNAs, which were recruited to FUS-TIS condensates (Figure 2A). The addition of the FUS 3′UTR did not change the morphology of the sphere-like condensates formed by FUS-TIS (Figure 2A). However, when we added 3′UTRs of TIS11B target mRNAs, including CD47, CD274 (PD-L1) or ELAVL1 (HuR) (Ma and Mayr, 2018), we observed formation of mesh-like FUS-TIS condensates (Figure 2A). Importantly, the mesh-like condensates do not represent aggregates as FUS-TIS protein showed fast fluorescence recovery in FRAP experiments performed at 2 and 16 hr after induction of phase separation (Figure 2B, Figure 2—figure supplement 1F).

Figure 2 with 2 supplements see all

Download asset Open asset

All phase separation experiments were performed at two time points (after 2 and 16 hr of incubation) in the presence of 5% dextran and at RNA concentrations spanning three orders of magnitude. Network formation was already observed at the early time point, but longer incubation led to formation of a more extensive network (Figure 2A, C, Figure 2—figure supplements 1G and 2A–D). Although the minimum RNA concentration required to induce network formation varied, these experiments revealed that the capacity for network formation is an intrinsic property of the RNA as sphere-forming RNAs did not form networks even at high concentrations. Instead, at high RNA concentrations, we often observed inhibition of phase separation, as was observed previously (Figure 2C, Figure 2—figure supplement 1G; Maharana et al., 2018).

The three network-forming RNAs are longer than the sphere-forming RNA (Figure 2A). To examine if network formation is only accomplished by long RNAs, we tested 19 additional RNAs with a length spanning 500–3000 nt. All longer RNAs formed networks, but we observed both network and sphere formation for RNAs shorter than 2000 nt, indicating that network formation is not only determined by the length of the RNA (Figures 2D and 3A, Figure 3—figure supplements 1A–C and 2A, Figure 3—source data 1). We did not observe phase separation when using high concentrations of the RNAs alone without the addition of FUS-TIS protein (Figure 3—figure supplement 2B).

Figure 3 with 5 supplements see all

Download asset Open asset

Figure 3—source data 1

Length, number of AU-rich elements, GC-content, and normalized ensemble diversity values of the 47 experimentally tested 3′UTRs.

https://cdn.elifesciences.org/articles/64252/elife-64252-fig3-data1-v3.xlsx

Download elife-64252-fig3-data1-v3.xlsx

To identify the responsible determinants for network formation, we focused on the 18 3′UTRs that were shorter than 2000 nt and correlated their ability for network formation with several parameters. Within this size-restricted cohort, the number of AU-rich elements or the GC-content of the RNA had no influence on network formation (Figure 3B, C). We then used RNAfold to predict the secondary structure of the RNAs. RNAfold predicts the minimum free energy secondary structure and the centroid structure, which is the RNA structure that contains a minimal base-pair distance to all structures in the thermodynamic ensemble (Gruber et al., 2008). Every RNA is predicted to have unstructured regions (indicated by the green and blue colors) and regions of strong local structure (red color code; Figure 3D, E, Figure 3—figure supplement 3A–D; Aw et al., 2016). We observed a strong association between the propensity of an RNA to induce network formation and the predicted ‘unstructured-ness’ of the RNA. For sphere-inducing RNAs, the majority of their nucleotides are predicted to form strong local structures, whereas RNAs with a high propensity for network formation are predicted to contain large, unstructured regions (Figure 3D, E, Figure 3—figure supplement 3A–D, Figure 3—source data 1). We call the unstructured regions ‘large disordered regions’ (LDRs) of mRNAs.

Can we use RNA-fold-based structure prediction as a tool to identify network-forming RNAs? Ensemble diversity is the number of potential RNA structures that are predicted for a given RNA (Lorenz et al., 2011). As ensemble diversity increases with RNA length (Ding et al., 2005), we are using a length-normalized value (NED). RNAs with low NED values have predominantly strong local structures, whereas RNAs predicted to have high NED values often have LDRs (Figure 3D, E, Figure 3—figure supplement 3A–D, Figure 3—source data 1). The NED values correlated strongly with the ability of an RNA to induce network formation and clearly separated the two groups of RNAs with respect to network formation (Figure 3F). The majority of network-forming RNAs had NED values that were higher than 0.280, whereas the majority of sphere-forming RNAs had NED values lower than 0.265. This is consistent with their stronger secondary structure as, for example, highly structured RNAs such as tRNAs and six MS2 repeats have NED values of 0.04 and 0.17, respectively (Figure 3—figure supplement 3E, F). We want to emphasize that we do not use RNAfold to predict the correct secondary structure of an RNA molecule, but we are using the NED value that estimates the conformational heterogeneity of an RNA as a tool to predict RNAs that induce formation of sphere-like or mesh-like condensates.

To test the predictive value of NED, we chose a new set of 24 AU-rich element-containing 3′UTRs purely based on their NED values and tested their network-forming abilities. We found that 19/24 (79%) of the tested RNAs were predicted correctly with respect to their sphere- or network-forming abilities (Figure 3G, Figure 3—figure supplement 4A, B, Figure 3—source data 1). As the number of AU-rich elements in both groups is comparable (Figure 3—figure supplement 5A), the high success rate strongly suggests that LDRs of 3′UTRs determine network formation. To test this prediction experimentally, we performed a loss-of-function experiment. We used the TNFSF11 3′UTR that contains several LDRs (Figure 3E) and introduced strong local base-pairing by the addition of two oligonucleotides that were perfectly complementary to upstream regions and did not disrupt AU-rich elements (Figure 3—figure supplement 5B). The TNFSF11 3′UTR mutant has stronger local structures indicated by increased base-pairing and a lower NED value, and it has largely lost the ability for network formation (Figure 3H, I). Taken together, these results support a model wherein a high diversity of predicted structural conformations correlates with the extent of LDRs in RNAs and is associated with the formation of mesh-like condensates.

The minimum RNA concentration for network formation was 20 nM. It was observed for the CD47 and ELAVL1 3′UTRs and corresponds to 27 and 32 ng/µl, respectively (Figure 2C, Figure 2—figure supplement 1G). This is higher than the mRNA concentration in the cytoplasm of mammalian cells, which was estimated to be 8 pM to 8 nM (9.5 pg/µl to 9.5 ng/µl; see Materials and methods) (Chen et al., 2015; Chen et al., 2016; Maharana et al., 2018). As TIS granules and L-bodies contain many mRNAs (Ma and Mayr, 2018; Neil et al., 2020), we hypothesized that multiple RNAs together may contribute to network formation. Therefore, we tested whether two RNAs co-localize in the network. Labeling of several pairs of RNAs with two different fluorescent dyes showed that they co-localize (Figure 4A). Furthermore, the mixing of suboptimal amounts of four network-forming RNAs, together with FUS-TIS, resulted in network formation, indicating that the different RNAs have an additive effect (Figure 4B). Importantly, mixing 10 mRNAs together with FUS-TIS only required 2 nM (0.7–3.2 ng/µl) of each mRNA which is substantially lower than the average mRNA expression in cells (Figure 4B). These data indicate that the RNA concentrations used for the in vitro experiments are in a range that is physiologically relevant.

Figure 4

Download asset Open asset

Next, we set out to address how RNAs with LDRs induce networks. We had observed that RNAs that are unable to induce networks are predicted to form strong local structures, meaning that they have a high propensity for intramolecular interactions (Figure 5A). This led us to hypothesize that network formation is caused by intermolecular RNA–RNA interactions mediated by the LDRs of mRNAs (Figure 5B). To test this, we performed native gel electrophoresis with sphere-forming and network-forming RNAs. We observed the appearance of diverse RNA species with high molecular weight only with the network-forming RNAs (Figure 5C). The observed smear is not due to degradation as the denaturing gel demonstrates that the used RNAs are intact (Figure 5C). This suggested that mRNAs with LDRs form higher-order RNA interactions in vitro.

Figure 5 with 1 supplement see all

Download asset Open asset

To investigate if intermolecular RNA–RNA interactions are indeed the cause of mesh-like condensates, we performed in vitro reconstitution experiments with RNAs that were designed to form multivalent RNA–RNA interactions (Figure 5D). We selected two 3′UTRs (TLR8 and MYC) that are unable to induce network formation (Figure 2D). However, they are able to dimerize, and this feature provides one degree of multivalency (Figure 5D, E, Figure 5—figure supplement 1A). We added two different RNA dimerization elements to their 5′ and 3′ ends to increase RNA multivalency (Figure 5D; Figure 5—figure supplement 1B, C). Adding RNA dimerization elements did not substantially change the NED values of TLR8 and MYC 3′UTRs. (Figure 5—figure supplement 1D, E), but this strategy enables intermolecular RNA–RNA interactions and allows the formation of a complex RNA network, demonstrated by native gel electrophoresis (Figure 5D, E). The RNA dimerization elements were derived from tracrRNA/crRNA (D1) and from HIV (D2) (Figure 5—figure supplement 1B, C; Skripkin et al., 1994; Jinek et al., 2012; Paillart et al., 2004; Khan et al., 2019).

A phase separation experiment with FUS-TIS confirmed that the addition of the two predominantly structured 3′UTRs (TLR8 and MYC) generates sphere-like condensates, whereas the addition of RNAs capable of forming a crosslinked RNA network (D1-TLR8-D2 and D1-MYC-D2), which we call an RNA matrix, induces formation of mesh-like FUS-TIS condensates (Figure 5F). Taken together, this in vitro reconstitution experiment demonstrated that an extensive, multivalent RNA interaction network can drive the formation of mesh-like condensates.

To investigate if mesh-like condensates can also be reconstituted in vivo, we used the same RNAs (TLR8 and MYC 3′UTRs with and without dimerization elements) together with SUMO-SIM as protein component. SUMO-SIM is especially suitable for this approach as all SUMO-SIM condensates are sphere-like. Our goal was to recruit mRNAs that are able to form a pervasive RNA interaction network into SUMO-SIM condensates using the MS2 system (Bertrand et al., 1998; Berkovits and Mayr, 2015) as this should turn the sphere-like condensates into mesh-like condensates (Figure 6A).

Figure 6 with 1 supplement see all

Download asset Open asset

Figure 6—source data 1

Transcriptome-wide analysis on normalized ensemble diversity values of 3′UTRs.

https://cdn.elifesciences.org/articles/64252/elife-64252-fig6-data1-v3.xlsx

Download elife-64252-fig6-data1-v3.xlsx

We fused SUMO-SIM to the MS2 coat protein, which can be considered as a selective RBD. The MS2 coat protein binds to RNA stem loops that were introduced in both TLR8 and MYC 3′UTRs. This system allows phase separation as SUMO-SIM is a multivalent protein and at the same time recruits the two structured 3′UTRs into the condensate (Figure 6B). In the presence of all the required elements (SUMO-SIM fused to MS2 coat protein, TLR8 and MYC 3′UTRs fused to MS2-binding sites, and the presence of the dimerization elements D1 and D2 in the RNAs), we observed mesh-like condensate formation in vivo, indicating that we are able to reconstitute mesh-like condensates in living cells (Figure 6A). Omission of MS2-binding sites or the MS2 coat protein prevents recruitment of the mRNAs to the multivalent protein and completely prevents the formation of mesh-like condensates, indicating that RNA is absolutely required for mesh-like condensate formation (Figure 6C, D). Omission of the two RNA dimerization elements from the system still allows recruitment of the RNAs to the multivalent protein. As the RNAs are able to dimerize, they generate mesh-like condensates in a few cases, but form sphere-like condensates in the majority of cases (Figure 6B, E). However, the presence of RNA dimerization elements that allow the generation of an extensive RNA interaction network that interacts with a multivalent protein substantially increases formation of mesh-like condensates in living cells (Figure 6A, E). These in vivo reconstitution experiments confirm that phase separation of an RNA-binding protein, together with mostly structured RNAs, predominantly generates sphere-like condensates, whereas phase separation of the same RNA-binding protein bound to RNAs with an ability to generate an extensive interaction network form mesh-like condensates in living cells.

In the in vivo reconstitution experiments, we did not use the TIS11B RBD, but we functionally replaced it by the recruitment of RNA matrix-forming RNAs. This suggests that other RBDs may also be able to generate mesh-like condensates in cells. TIS11B binds to AU-rich elements (Peng et al., 1998). Transcriptome-wide analyses showed that mRNAs with several AU-rich elements in their 3′UTRs have longer 3′UTRs and higher NED values (Figure 6—figure supplement 1A, B, Figure 6—source data 1). HuR is an RNA-binding protein that also binds to U- or AU-rich elements (Lebedeva et al., 2011; Mukherjee et al., 2011; Uren et al., 2011). Analyzing HuR PAR-CLIP data showed that HuR targets also have longer 3′UTRs and higher NED values (Figure 6—figure supplement 1C, D). Therefore, we examined if the HuR RBD is capable of forming mesh-like condensates in cells. In the context of two multivalent domains (TIS11B N/C-terminus or FUS-IDR), the RRM1/2 of HuR was sufficient for the generation of mesh-like condensates (Figure 6F, G, Figure 6—figure supplement 1E). Taken together, we showed that several chimeric proteins that consist of multivalent domains that are paired with RBDs that bind to RNA matrix-forming RNAs can form mesh-like condensates in cells (Table 1).

To start to get at the mechanism of mesh-like condensate formation, we imaged the early phases of condensate fusion with sphere- and network-forming mRNAs. As expected, the condensates that contain the predominantly structured 3′UTR of TLR8 fully fuse within 90 s (Figure 7A). In contrast, when two condensates that contain RNAs with LDRs, such as the 3′UTR of CD47, come into contact, they largely retain their shape and do not fully mix (Figure 7B). However, the two condensates connect at the contact sites; this enables linkage of individual condensates and allows them to grow into large assemblies with irregular shapes (Figure 7C). It is worth pointing out that despite the filamentous condensate shape the protein components are still highly mobile as was shown by FRAP experiments performed at 2 and 16 hr after mixing (Figure 2B, Figure 2—figure supplement 1F).

Figure 7 with 1 supplement see all

Download asset Open asset

To get a better understanding of the morphology of the mRNAs within sphere- and mesh-like condensates, we performed high-resolution imaging. We observed that the distribution of sphere-forming mRNAs is relatively uniform in the FUS-TIS condensates (Figure 7—figure supplement 1A). In contrast, the network-forming mRNAs are unevenly distributed and generate an underlying scaffold for the phase-separated protein that resembles a skeleton (Figure 7D).

We propose the following speculative model (Figure 7E). Fusion of condensates that contain structured RNAs allows the mixing of components, thus resulting in the formation of larger spheres. Highly structured RNAs can be viewed to be globule-like, and their movement within condensates is only somewhat restricted due to very weak interactions (Hyman et al., 2014; Ranganathan and Shakhnovich, 2020). In contrast, the LDRs within RNAs (depicted as loops and tails) form pervasive interactions (Figure 5E, F), thus forming an RNA matrix. When two condensates that contain such an RNA skeleton come into contact, the semi-rigidity of the crosslinked RNA network prevents full fusion and only allows mixing of the components at the contact sites. This arranges the condensates as beads-on-a-string and results in formation of filamentous mesh-like condensates whose protein components are dynamic. These data suggest that the fusion force provided by the surface tension of liquid condensates is counteracted by an anti-fusion effect provided by the underlying RNA matrix, thus forming a mesh-like condensate.

This model predicts that sphere-forming mRNAs are mobile and exchange with their environment, whereas network-forming mRNAs are static within the mesh-like condensates. We used FRAP on several RNAs to test the prediction (Figure 7F, Figure 7—figure supplement 1B). We observed indeed fast fluorescence recovery of sphere-forming mRNAs, including TLR8 and MYC, suggesting that they are highly mobile within the condensates (Figure 7F, Figure 7—figure supplement 1B). In contrast, network-forming mRNAs of various lengths showed no fluorescence recovery (Figure 7F, Figure 7—figure supplement 1B), which supports the notion that these mRNAs act as skeleton or underlying RNA matrix for mesh-like condensates.

Our data suggest that, in addition to acting as information templates for protein synthesis, mRNAs and, in particular, 3′UTRs may further fulfill roles as structural elements for cytoplasmic RNA granules. This is reminiscent of the scaffolding roles of lncRNAs in nuclear bodies, including paraspeckles (Chujo et al., 2016). Both 3′UTRs and lncRNAs have a similar AU-content and hexamer composition that differs substantially from the coding region. This finding provides further support for the scaffolding role of 3′UTRs because a higher AU-content is associated with decreased RNA structure and higher RNA flexibility (Niazi and Valadkhan, 2012). Here, we identified two RBDs (TIS11B and HuR) that both bind to longer 3′UTRs with high NED values that are able to form mesh-like condensates in cells. Therefore, it is likely that interacting 3′UTRs do not only act as RNA matrix for TIS granules and L-bodies, but may also scaffold additional cytoplasmic membraneless compartments, possibly on other membrane surfaces or on the cytoskeleton (Smith et al., 2020; Béthune et al., 2019).

Formation of filamentous and mesh-like condensates is not restricted to the chimeric RNA-binding proteins used here but was also observed by others (Lin et al., 2015; Boeynaems et al., 2019; Smith et al., 2020; Neil et al., 2020). For example, the mixing of a PR30 peptide with two homopolymeric RNA species with perfect base-pairing capabilities (poly-rA plus poly-rU) induced formation of filamentous condensates (Boeynaems et al., 2019). The PR30 condensates showed little fluorescent recovery upon FRAP, thus, suggesting that physical cross-linking of the base-paired RNA structures and the PR molecules arrests phase separation (Boeynaems et al., 2019). It is likely that this represents a pathological feature of PR molecules that were shown previously to reduce FRAP recovery in many phase separation systems (Lee et al., 2016). In contrast, in our system, we used the TIS11B and HuR RBDs that bind to 3′UTRs that contain mixtures of structured and unstructured regions under physiological conditions. These observations indicate that some filamentous condensates are solid, whereas others are dynamic, and their material properties are regulated by the extent and strength of RNA–protein and RNA–RNA interactions (Ferrandon et al., 1997; Jambor et al., 2011; Trcek et al., 2015; Van Treeck et al., 2018; Trcek et al., 2020).

Although the physiological relevance of mesh-like condensates is currently largely unknown, our study indicates that mesh-like condensates cannot simply be viewed as aggregation. Surface tension enforces a low surface to volume ratio for liquid-like spheres (Hyman et al., 2014). In contrast, mesh-like condensates have a larger surface to volume ratio. This gives them an advantage for reactions that occur on their surface. As TIS granules are intertwined with the ER, they share a large interface (Ma and Mayr, 2018). Our previous data suggests that the mesh-like shape of the TIS granule network is necessary for its function during translation of membrane proteins. We showed previously that translation in the TIGER domain enables protein complex formation of membrane proteins and promotes their trafficking to the plasma membrane (Ma and Mayr, 2018).

These observations may indicate that the mesh-like shape of TIS granules is especially important for reactions that occur in the interface with the ER. However, also mRNAs that encode non-membrane proteins are translated in TIS granules. As translation of these proteins does not require them to be incorporated into membranes, it seems that for these reactions the shape of TIS granules is irrelevant. Rather, these reactions may take advantage of the local environment generated by TIS granules where a connected membraneless compartment may enable the enriched protein components to exchange relatively freely, thus allowing the sharing of RNA-binding proteins and chaperones (Ma and Mayr, 2018).

One potential biological implication for intermolecular mRNA–mRNA interactions in TIS granules could be to facilitate co-translational protein complex assembly. Although the majority of protein complexes form co-translationally in yeast (Shiber et al., 2018), it is unclear how the protein subunits come into proximity. It is possible that the physical interaction between LDRs in 3′UTRs provides the necessary proximity of two translating ribosomes and their nascent peptide chains to promote complex assembly (Shiber et al., 2018; Mayr, 2018).

We noticed that the generation of mesh-like networks – either lipid membrane-enclosed or membraneless – is conceptually similar. Our model proposes that mesh-like condensates with dynamic protein components are generated from the interplay of two forces (Figure 7E). One force promotes fusion and is driven by the surface tension of the liquid condensates. The other force prevents full fusion and is provided by the underlying RNA matrix that forms a semirigid skeleton and only allows fusion of the condensates at the contact sites. Such interplay of two antagonistic forces was also observed upon in vitro reconstitution of the membrane-enclosed tubular ER network (Powers et al., 2017). Sphere-like liposomes were mixed with proteins that exert two opposing forces: One of the proteins promotes liposome fusion and network assembly, whereas the other promotes fragmentation and disassembly. Taken together, these data indicate that membrane-enclosed and membraneless network structures result from the balance of two opposing forces.

RNA is probably the most versatile molecule in cells (Mayr, 2017). Through RNA mimicry, viral RNAs can mimic the shape of cellular tRNAs (Colussi et al., 2014). RNA can also mimic the shape of proteins and protein interaction surfaces (Athanassiou et al., 2004; Shao and Hegde, 2016; Mizrak and Morgan, 2019). Here, we found that the network structure and morphogenesis of the tubular ER that is generated by proteins and lipids can be mimicked by phase-separated condensates formed by protein and RNA. Our work showed that dynamic subcellular structures with complex shapes can be generated through phase separation without the need for lipid membranes.

The human cervical cancer cell line, HeLa, was a gift from the lab of Jonathan S. Weissman (UCSF), provided by Calvin H. Jan. Cells were maintained at 37°C with 5% CO₂ in Dulbecco’s Modified Eagle Medium containing 4500 mg/l glucose, 10% heat-inactivated fetal bovine serum, 100 U/ml penicillin, and 100 mg/ml streptomycin. The cell line has not been authenticated. The cell line is free of mycoplasma. Mycoplasma detection was performed by DAPI staining.

All primers are reported in Table 1. All PCR reactions were performed using Q5 High Fidelity DNA polymerase (NEB). The basis for all mammalian expression vectors was pc-DNA-puro described previously (Ma and Mayr, 2018). The following inserts were also described previously mCherry-TIS11B, mCherry-SEC61B, eGFP-SEC61B, eGFP-CD47-3′UTR, eGFP-ELAVL1-3′UTR, eGFP-FUS-3′UTR, and eGFP-CD274-3′UTR (Ma and Mayr, 2018).

Point mutations were generated using QuikChange Lightning Multi Site-Directed Mutagenesis Kit (Agilent Technologies, #210513) if not otherwise stated. pcDNA-puro-mGFP (monomeric GFP, A207K) was generated from pcDNA-puro-eGFP using the primer eGFP A207K. For TIS11B RBD mutants, the following primers were used: TIS11B C135H, TIS11B C173H, TIS11B F137N, TIS11B F175N, TIS11B K116L, TIS11B K154L, TIS11B R116L, and TIS11B K152L. We called the RBD mutants CC, FF, KK, and RK because the mutated amino acids are C135H/C173H, F137N/F175N, K116L/K154L, and R114L/K152L, respectively.

TIS-HuR RBD contains the N-terminus of TIS11B (TIS11B N; aa 1–113, based on uniprot ID Q07352-1) fused to RRM1/2 of HuR (aa 19–189) fused to the C-terminus of TIS11B (TIS11B C); (aa 182–338). This construct was generated using PCR amplification of three overlapping fragments. Fragment 1: TIS11B N was PCR-amplified from the mCherry-TIS11B construct with primers TIS-HuR 1F and TIS-HuR 1R. Fragment 2: RRM1/2 of HuR was PCR-amplified from the eGFP-HuR-3′UTR construct with primers TIS-HuR 2F and TIS-HuR 2R. Fragment 3: TIS11B C was PCR-amplified from the mCherry-TIS11B construct with primers TIS-HuR 3F and TIS-HuR 3R. A ligation PCR was performed to ligate Fragment 1 and Fragment 2 to generate Fragment 1–2 with primers TIS-HuR 1F and TIS-HuR 2 R-2. Then Fragment 1–2 was digested with HindIII and ApaI; Fragment 3 was digested with ApaI and EcoRI. To generate full-length TIS-HuR chimera, Fragment 1–2 and Fragment 3 were cloned into the pcDNA3.1-puro-mCherry vector with HindIII and EcoRI restriction sites.

The pmCherry-SUMO10-SIM5 construct was a gift from the lab of Liam J. Holt (NYU). To generate the SUMO-SIM-TIS (TIS11B RBD fused to the N-terminus of SUMO10-SIM5) fusion protein, two overlapping fragments were PCR-amplified. Fragment 1, mCherry, was PCR-amplified from the mCherry-TIS11B construct with primers TIS-SUMO-SIM 1F and TIS-SUMO-SIM 1R. Fragment 2, the RBD (aa 114–181) of TIS11B, was PCR-amplified from the mCherry-TIS11B construct with primers TIS-SUMO-SIM 2F and TIS-SUMO-SIM 2R. A ligation PCR was performed to generate mCherry-TIS11B RBD with primers SUMO-SIM 1F and TIS-SUMO-SIM 2R. mCherry-TIS11B RBD was cloned into the SUMO10-SIM5 construct with AgeI and BsrGI restriction sites.

For the FUS-TIS (FUS IDR fused to the N-terminus of TIS11B RBD) fusion protein, two overlapping fragments were PCR-amplified. Fragment 1: the IDR (aa 1–214) of FUS was PCR-amplified from the pcDNA-puro-eGFP-FUS-3′UTR vector with primers FUS-TIS 1F and FUS-TIS 1R. Fragment 2: the RBD (aa 114–181) of TIS11B was PCR-amplified with primers FUS-TIS 2F and FUS-TIS 2R. A final ligation PCR was performed to ligate two PCR fragments to the full-length FUS-TIS chimera with primers FUS-TIS 1F and FUS-TIS 2R. The full-length FUS-TIS was cloned into pcDNA3.1-puro-mGFP vector with BsrGI and EcoRI restriction sites.

For the pcDNA3.1-puro-BFP-FUS-TIS construct, a nuclear export signal (nes) was added upstream of the FUS IDR to increase cytoplasmic localization. It was obtained from pcDNA3.1-puro-mCherry-nes-FUS-TIS (provided by Neil Robertson, MSKCC) using BsrGI and EcoRI restriction sites and cloned into the pcDNA3.1-puro-BFP vector with the same restriction sites.

The RBD of HuR (aa 19–189) was PCR-amplified from the eGFP-HuR-3′UTR construct with primers HuR RBD F and FUS-HuR 2R. HuR RBD was cloned into pcDNA3.1-puro-mGFP vector with HindIII and EcoRI restriction sites.

To generate the FUS-HuR RBD (FUS IDR fused to the N-terminus of HuR RBD) fusion protein, two overlapping fragments were PCR-amplified. Fragment 1: the IDR (aa 1–214) of FUS was PCR-amplified from the pcDNA-puro-eGFP-FUS-3′UTR vector with primers FUS-HuR 1F and FUS-HuR 1R. Fragment 2: the RBD (aa 19–189) of HuR was PCR-amplified from the eGFP-HuR-3′UTR construct with primers FUS-HuR 2F and FUS-HuR 2R. A final ligation PCR was performed to ligate two PCR fragments to obtain FUS-HuR chimera with primers FUS-HuR 1F and FUS-HuR 2R. The full-length FUS-TIS was cloned into pcDNA3.1-puro-mGFP vector with HindIII and EcoRI restriction sites.

Contructs for in vivo reconstitution. To generate the eGFP-3ʹUTR of TLR8 construct, the TLR8 3ʹUTR was PCR-amplified from HeLa genomic DNA with primers TLR8-MS2 F and TLR8-MS2 R. The TLR8 3ʹUTR was cloned into the pcDNA-puro-eGFP vector with BsrGI and EcoRI restriction sites.

To generate the eGFP-3ʹUTR of MYC construct, the MYC 3ʹUTR was PCR-amplified from HeLa genomic DNA with primers MYC-MS2 F and MYC-MS2 R. The MYC 3ʹUTR was cloned into the pcDNA-puro-eGFP vector with BsrGI and EcoRI restriction sites.

To generate the eGFP-D1-3ʹUTR of TLR8-D2 construct (TLR8 3ʹUTR with RNA dimerization elements D1 and D2), the D1-TLR8 3ʹUTR-D2 was PCR-amplified from the TLR8 3ʹUTR PCR product with primers D1D2-MS2 F1 and D1D2-MS2 R. D1-TLR8 3ʹUTR-D2 was cloned into the pcDNA-puro-eGFP vector with BsrGI and EcoRI restriction sites.

To generate the eGFP-D1-3ʹUTR of MYC-D2 construct (MYC 3ʹUTR with RNA dimerization elements D1 and D2), the D1-MYC 3ʹUTR-D2 was PCR-amplified from the MYC 3ʹUTR PCR product with primers D1D2-MS2 F2 and D1D2-MS2 R. The D1-MYC 3ʹUTR-D2 was cloned into the pcDNA-puro-eGFP vector with BsrGI and EcoRI restriction sites.

For the eGFP-6xMS2 binding site construct, 6xMS2 binding site was cloned into pcDNA-puro-eGFP vector with EcoRI and XhoI restriction sites.

To generate the eGFP-3ʹUTR of TLR8-6xMS2 construct, the TLR8 3ʹUTR was PCR-amplified from HeLa genomic DNA with primers TLR8-MS2 F and TLR8-MS2 R. The TLR8 3ʹUTR was cloned into the pcDNA-puro-eGFP-6xMS2 vector with BsrGI and EcoRI restriction sites.

To generate the eGFP-3ʹUTR of MYC-6xMS2 construct, the MYC 3ʹUTR was PCR-amplified from HeLa genomic DNA with primers MYC-MS2 F and MYC-MS2 R. The MYC 3ʹUTR was cloned into the pcDNA-puro-eGFP-6xMS2 vector with BsrGI and EcoRI restriction sites.

To generate the eGFP-D1-3ʹUTR of TLR8-D2-6xMS2 construct (TLR8 3ʹUTR with RNA dimerization elements D1 and D2), the D1-TLR8 3ʹUTR-D2 was PCR-amplified from the TLR8 3ʹUTR PCR product with primers D1D2-MS2 F1 and D1D2-MS2 R. The D1-TLR8 3ʹUTR-D2 was cloned into the pcDNA-puro-eGFP-6xMS2 vector with BsrGI and EcoRI restriction sites.

To generate the eGFP-D1-3ʹUTR of MYC-D2-6xMS2 construct (MYC 3ʹUTR with RNA dimerization elements D1 and D2), the D1-MYC 3ʹUTR-D2 was PCR-amplified from the MYC 3ʹUTR PCR product with primers D1D2-MS2 F2 and D1D2-MS2 R. The D1-MYC 3ʹUTR-D2 was cloned into the pcDNA-puro-eGFP-6xMS2 vector with BsrGI and EcoRI restriction sites.

To generate the MS2-mCherry-SUMO10-SIM5 (MS2 coat protein fused to the N-terminus of mCherry-SUMO10-SIM5), MS2-mCherry was PCR-amplified from the pcDNA-MS2-mCherry-HuR vector (Berkovits and Mayr, 2015) with primers MS2-SUMO-SIM F and MS2-SUMO-SIM R. pmCherry-SUMO10-SIM5 vector was digested with AgeI and BsrGI to release the mCherry fragment. The MS2-mCherry was cloned into the digested pmCherry-SUMO10-SIM5 vector with AgeI and BsrGI restriction sites.

Lipofectamine 2000 (Invitrogen) was used for all transfections.

To examine the RNA-binding activity of TIS11B WT and TIS11B RBD mutants, RNA oligonucleotide pulldown was performed as described previously (Ma and Mayr, 2018). A 3′-biotinylated RNA oligonucleotide of the TNFα ARE-1 (AU-rich element) was purchased from Dharmacon. mCherry-tagged constructs were transfected into HeLa cells with or without 3′-biotinylated RNA oligonucleotides. Twenty-four hours after transfection, HeLa cells were lysed with 200 µl ice-cold NP-40 lysis buffer (25 mM Tris-HCl pH 7.5, 150 mM NaCl, 1% NP-40, 1 mM EDTA) for 30 min. Then, cell lysates were spun down at 20,000 g for 10 min at 4°C. The supernatant was transferred to a pre-cooled tube and diluted with 300 µl ice-cold dilution buffer (10 mM Tris-HCl pH 7.5, 150 mM NaCl, 0.5 mM EDTA). Streptavidin C1 beads (Invitrogen) were added to each tube and rotated for 1 hr at 4°C. Beads were washed three times with wash buffer (10 mM Tris-HCl pH 7.5, 150 mM NaCl, 0.5 mM EDTA). Lastly, 2× Laemmli sample buffer was added to the beads, boiled at 95°C for 10 min, and cooled on ice before loading on SDS page gels. This was followed by western blotting.

Western blots were performed as described previously (Ma and Mayr, 2018). Imaging was captured on the Odyssey CLx imaging system (Li-Cor). The antibodies used are mouse anti-α-tubulin (Sigma-Aldrich, T9026, RRID:AB_477593), mouse anti-mCherry (Abcam, ab125096, RRID:AB_11133266), rabbit anti-HuR (Millipore, 07-1735, RRID:AB_1977173), IRDye 680RD donkey anti-rabbit IgG secondary antibody (Li-COR Biosciences, 926-68073, RRID:AB_10954442), and IRDye 800CW donkey anti-mouse IgG secondary antibody (Li-COR Biosciences, 926-32212, RRID:AB_621847).

mGFP-FUS-TIS was cloned into the bacterial expression vector pET28a, which was a gift from the lab of Dirk Remus (MSKCC). At the N-terminus of mGFP, we added a 6xHis-MBP tag, followed by a Tev protease cleavage site. At the C-terminus of TIS11B, we added a Strep-Tag II (SAWSHPQFEK). The 6xHis-MBP tag was PCR-amplified from pDZ2087 construct (Addgene, #92414) with primers MBP F and MBP R. Full-length 6xHis-MBP was cloned into pET28a backbone with XbaI and EcoRI restriction sites. mGFP-FUS-TIS-Strep-tag II was PCR-amplified from pcDNA-mGFP-FUS-TIS construct with primers mGFP F and TIS RBD-Strep-tag R. The Strep-tag II sequence was incorporated into primer TIS RBD-Strep-tag R. Full-length mGFP-FUS-TIS-Strep-tag II was cloned into pET28a-6xHis-MBP backbone with NheI and EcoRI restriction sites.

To purify high-quality FUS-TIS protein, we used three steps of purification. Step 1: His-Ni purification; step 2: Strep-Tag II purification; and step 3: size exclusion chromatography. pET28a-6xHis-MBP-Tev cleavage site-mGFP-FUS-TIS-Strep-tag II was transformed into BL21 Escherichia coli (New England Biolabs). Two fresh colonies were cultivated overnight in 2 × 50 ml SOB medium at 37°C, and then 4 × 25 ml bacteria were transferred to 4 × 1 liter SOB medium to grow at 37°C until OD600 reached 0.6. Bacteria were then kept in a 4°C cold room until 18:00. Protein expression was induced by addition of 1 mM IPTG, and bacteria were incubated at 16°C overnight.

Bacteria were centrifuged at 6000 g for 10 min, and the pellet was resuspended in 100 ml cold lysis buffer. High-salt lysis buffer (1 M NaCl, 25 mM Tris-Cl, pH 8.0, 20 mM imidazole, 1 mM DTT, 1× PMSF) was used to remove nucleic acid contamination. Bacteria were sonicated on ice for 60 min with on/off interval of 1 and 2 s. The lysate was centrifuged at 13,000 g for 30 min.

A 6 ml Ni-NTA (Qiagen) was washed with five column volumes of wash buffer 1 (150 mM NaCl, 25 mM Tris-Cl, pH 8.0, and 1 mM DTT). After centrifugation, the supernatant of the bacteria lysate was transferred into new 50 ml Falcon tubes and incubated with Ni-NTA (Qiagen) at 4°C for 30 min. Then, the sample was transferred into three gravity columns and washed respectively with 40 ml wash buffer 2 (1 M NaCl, 25 mM Tris-Cl, pH 8.0, 20 mM imidazole, and 1 mM DTT), followed with 10 ml wash buffer 3 (600 mM NaCl, 25 mM Tris-Cl, pH 8.0, 20 mM imidazole, and 1 mM DTT). Then, the sample was eluted with 30 ml elution buffer 1 (600 mM NaCl, 25 mM Tris-Cl, pH 8.0, 200 mM imidazole, and 1 mM DTT).

After Ni-NTA purification, the eluted sample was transferred to a 5 ml StrepTrap column (GE Healthcare, cat. no. 28907547), which was pre-equilibrated with elution buffer 2 (600 mM NaCl, 20 mM Tris-HCl, pH 7.4, and 1 mM DTT) using the AKTA Purifier system (GE Healthcare). The target protein was eluted with 20 ml elution buffer 3 (600 mM NaCl, 20 mM Tris-HCl, pH 7.4, 2.5 mM desthiobiotin [Sigma-Aldrich], and 1 mM DTT).

The eluted protein was concentrated using Amicon Ultra-centrifugal filters-50K (Millipore). The concentrated sample was further purified by gel filtration on HiLoad 16/600 Superdex200 column (GE Healthcare) in elution buffer 2 (600 mM NaCl, 20 mM Tris-HCl, pH 7.4, and 1 mM DTT) using the AKTA Purifier system (GE Healthcare).

The fractions representing the monomeric protein were collected and concentrated with Amicon Ultra-centrifugal filters-50K (Millipore). The quality of the final protein product was examined by SDS PAGE. The OD260/280 ratio of the final protein product was 0.52, measured with NanoDrop. The protein concentration was measured by Bradford assay (Bio-Rad). The protein was aliquoted into PCR tubes and flash-frozen in liquid nitrogen and then stored at −80°C.

All RNAs were in vitro transcribed using the T7 MEGAscript kit (Ambion by Life Technologies). All DNA templates used for in vitro transcription were PCR-amplified and purified with a gel extraction kit (Qiagen). The T7 promoter (TAATACGACTCACTATAGGG) was incorporated into the forward primers used to amplify the DNA templates. After in vitro transcription, products were DNase-treated and run on agarose gels to evaluate the integrity and size of the RNA.

The DNA sequences from the full-length 3′UTRs of CD47, ELAVL1, CD274, and FUS were PCR-amplified from pcDNA-eGFP-CD47-3′UTR, eGFP-ELAVL1-3′UTR, eGFP-CD274-3′UTR, and eGFP-FUS-3′UTR constructs. The DNA sequences of the 3′UTRs of CD44, VSIG10, IL10, TNFSF11, GPR39, TLR8, GPR34, TNFAIP6, MYC, PLA2G4A, HEATR5B, PPP1R3F, DRD1, FAM72B, MCOLN2, TSPAN13, LHFPL6, FAM174A, VPS29, ADPGK, ASPN, CASP8, CLCA2, EOMES, ESCO1, GLYATL3, HNRNPH3, HOGA1, LPAR4, LRBA, LYPLAL1, ODF2, PRKDC, RHOA, SHQ1, SLC39A6, SLC5A9, SMIM3, SNTN, SOSTDC1, STBD1, TP53TG3, and TTC17 were PCR-amplified from HeLa genomic DNA.

To generate the TNFSF11 3′UTR mutant carrying two 15-nt oligo insertions, two overlapping fragments were PCR-amplified. Fragment 1: TNFSF11 3′UTR with oligo 1 using primers TNFSF11 3′UTR T7 F and TNFSF11 mutant R1. Fragment 2: TNFSF11 3′UTR with oligo 1 and oligo 2 using primers TNFSF11 mutant F2 and TNFSF11 mutant R2. The inserted 15-nt oligo 2 sequence was incorporated into the primer TNFSF11 mutant R2. A final ligation PCR was performed to ligate two PCR fragments to generate the full-length TNFSF11 3′UTR mutant with primers TNFSF11 3′UTR T7 F and TNFSF11 mutant R2.

RNAs with exogenous dimerization elements (D1a, crRNA, D1b, tracrRNA, D2, HIV dimerization motif) were generated as follows. For D1a-TLR8-D2, two rounds of PCR were performed. Round 1: primers D1a-TLR8-D2 1F and D1a-TLR8-D2 1R; round 2: primers D1a-TLR8-D2 2F and D2 R. For D1b-MYC-D2, three rounds of PCR were performed. Round 1: primers D1b-MYC-D2 1F and MYC 3′UTR R; round 2: primers D1b-MYC-D2 2F and D1b-MYC-D2 2R; and round 3: primers D1b-MYC-D2 3F and D2 R.

In vitro transcription was performed in a 20 µl volume according to the manufacturer’s guidelines. To generate Cy3- or Cy5-labeled RNA, 0.2 µl of 2.5 mM Cy3-UTP or Cy5-UTP (Enzo Life Sciences) was added into the in vitro transcription reaction.

The transcription reaction was incubated 3 hr at 37°C in a PCR machine. All transcribed RNAs were digested with DNase for 30 min at 37°C, then precipitated with LiCl for 4 hr to overnight at −20°C. RNAs were centrifuged at 13,000 for 15 min, and the RNA pellets were washed with 70% ethanol three times. RNAs were dissolved in nuclease-free water and stored at −20°C. The concentration of RNAs was measured by NanoDrop One.

To allow phase separation, purified 6xHis-MBP-mGFP-FUS-TIS-Strep tag II protein stock was incubated with Tev protease for 1 hr at room temperature (RT) to cleave off the 6xHis-MBP tag. mGFP and Strep tag II were not cleaved off. All phase separation assays were performed in 20 µl phase separation buffer (150 mM NaCl, 200 µM ZnCl₂, 25 mM Tris-Cl, pH 7.4, 1 mM DTT, 2.5% glycerol, 5% dextran T500 [Pharmacosmos]). Without 5% dextran, there is no phase separation of FUS-TIS protein at the concentration of 10 µM. ZnCl₂ was added as the RBD of TIS11B has two zinc finger motifs. Only in the phase separation assay shown in Figure 2—figure supplement 1D, E ZnCl2 was omitted.

After 1 hr of Tev protease digestion, the FUS-TIS protein stock was diluted into the desired concentrations with protein stock buffer (600 mM NaCl, 25 mM Tris-Cl, pH 7.4, 1 mM DTT) and centrifuged at 13,000 g for 2 min to remove small protein aggregates. The supernatant was transferred into a new Eppendorf tube. The phase separation assay was mixed in PCR tubes. Dextran buffer and RNAs with desired concentrations were first mixed in PCR tubes, then FUS-TIS protein was added into the PCR tube and immediately mixed thoroughly. The final concentrations of FUS-TIS and RNAs are indicated in the figures. The mixture (20 µl) was then transferred into a 384-well glass-bottom microplate (Greiner Bio-One). The chambers of the microplate were pre-treated with 1 mg/ml bovine serum albumin (BSA) (NEB) for 30 min before aspirating the BSA. The microplate was kept in the dark at RT for 2 or 16 hr, followed by imaging of the condensates using confocal microscopy. All phase separation experiments were performed at least three times.

For RNA-only phase separation experiments, Cy5-labeled RNA was diluted with the same phase separation buffer into desired concentrations in a 20 µl volume. The mixture (20 µl) was then transferred into a 384-well glass-bottom microplate. The chambers of the microplate were pre-treated with 1 mg/ml BSA for 30 min before aspirating the BSA. The microplate was kept in the dark at RT for 16 hr, followed by imaging of the Cy5-labeled RNA using confocal microscopy.

Confocal imaging was performed using ZEISS LSM 880 with Airyscan super-resolution mode. Z stack images were captured with an interval size of 487 nm. A Plan-Apochromat 63x/1.4 Oil objective (Zeiss) was used. For live-cell imaging, HeLa cells were plated on 3.5 cm glass-bottom dishes (Cellvis) and transfected with the indicated constructs. Fourteen hours after transfection, cells were imaged in cell culture medium while incubating in a LiveCell imaging chamber (Zeiss) at 37°C and 5% CO₂. Images were prepared with the commercial ZEN software black edition (Zeiss).

The morphology of the condensates was scored by two independent scientists who agreed on the classification. To examine the ability of a specific RNA to induce network formation, several RNA concentrations were tested, for instance, for RNAs with a length of ~1000 nt, we tested concentrations from 50 nM (~16 ng/µl) to 750 nM (~250 ng/µl). If an RNA induced network formation within the concentration range, we considered it as network-forming RNA. If an RNA did not induce network formation even at 750 nM concentration, it was considered as a sphere-forming RNA.

FRAP experiments were performed with a ZEISS LSM 880 confocal microscope. A Plan-Apochromat 63x/1.4 Oil objective (Zeiss) was used.

For FRAP of phase separation experiments, 10 µM mGFP-FUS-TIS was mixed with specific RNAs to induce condensate formation. Two hours after mixing, an area of diameter = 1 µm was bleached with a 405 nm and 633 nm laser. GFP or Cy5 (in the case of RNA) fluorescence signal was collected over time.

For FRAP in live cells, HeLa cells were plated on 3.5 cm glass-bottom dishes (Cellvis) and transfected with the indicated constructs. Five or sixteen hours after transfection, cells were imaged in cell culture medium while incubating in a LiveCell imaging chamber (Zeiss) at 37°C and 5% CO₂. An area of diameter = 1 µm was bleached with a 405 nm laser. GFP or mCherry fluorescence signal was collected over time. For FRAP of TIS granules at 5 hr after transfection, GFP fluorescence signal was only collected for 10 s after bleaching as the TIS granules are highly mobile.

The prebleached fluorescence intensity was normalized to 1, and the signal after bleaching was normalized to the prebleach level.

Custom Stellaris EGFP FISH probes were described previously (Berkovits and Mayr, 2015). RNA-FISH was performed as published with slight modifications (Ma and Mayr, 2018). HeLa cells were plated on 4-well Millicell EZ silde and transfected with BFP-FUS-TIS and GFP fusion constructs. Fourteen hours after transfection, cells were washed with PBS, fixed with 4% paraformaldehyde for 15 min at RT, and washed twice for 5 min with PBS. PBS was discarded and 1 ml 70% ethanol was added. The slide was kept at 4°C for 8 hr. The 70% ethanol was aspirated, 1 ml wash buffer was added (2× SSC, 10% formamide in RNase-free water), and incubated at RT for 5 min. Hybridization mix was prepared by mixing 10% dextran sulfate, 10% formamide, 2× SSC, 2 mM ribonucleoside vanadyl complex (NEB), 200 µg/ml yeast tRNA, and FISH probe (1:100). To each well, 200 µl hybridization mix was added and hybridized at 37°C overnight. Slides were washed twice for 30 min each with pre-warmed wash buffer (1 ml, 37°C) in the dark, followed by one quick wash with PBST, and then mounted with mounting solution. Images were captured using confocal ZEISS LSM 880 with Airyscan super-resolution mode (Huff, 2015).

In order to examine whether specific mRNAs are enriched in the FUS-TIS granule, line profile analysis was performed. Line profiles were generated with FIJI (ImageJ). A straight line was drawn across the FUS-TIS granule, indicated by the arrows shown in the figures. Fluorescence signals along the straight line of FUS-TIS protein and the examined mRNAs were calculated with the plot profile tool in FIJI. The Pearson’s correlation coefficient (r) of two fluorescence signals was calculated with Excel.

RNA native agarose gel electrophoresis was performed as described previously with a few modifications (Skripkin et al., 1994). For sphere-forming and network-forming RNAs, RNAs were diluted into 4 µl buffer A (150 mM NaCl, 25 mM Tris-Cl, pH 7.4) to a final concentration of 5 µM (CLCA2 1.7 µg/µl, TLR8 1.6 µg/µl, HNRNPH3 1.9 µg/µl, LHFPL6 1.7 µg/µl, TNFSF11 1.8 µg/µl, TSPAN13 1.7 µg/µl). RNAs were incubated at 95°C for 2 min in a PCR machine and then incubated on ice for 2 min. RNAs were kept at 37°C for 2 hr. Also, 1 µl native agarose gel loading buffer (6× stock: 60% glycerol, 10 mM Tris-Cl, pH 7.4, 0.03% bromophenol blue, and 0.03% xylene cyanol FF) was added into the RNA. A total of 1 µg RNA was loaded into the 1% agarose gel made with the Tris-acetate-EDTA (TAE) buffer for electrophoresis with TAE buffer.

For RNAs containing dimerization elements (TLR8 3′UTR, MYC 3′UTR, D1a-TLR8-D2, D1b-MYC-D2), each RNA was diluted into 4 µl buffer A (150 mM NaCl, 25 mM Tris-Cl, pH 7.4) to a final concentration of 2 µM. RNAs were incubated at 95°C for 2 min in a PCR machine and then incubated on ice for 2 min. Also, 2 µl TLR8 3′UTR or 2 µl MYC 3′UTR were each diluted with 2 µl buffer A. A 2 µl TLR8 3′UTR and 2 µl MYC 3′UTR were mixed together. Then, 2 µl D1a-TLR8-D2 or 2 µl D1b-MYC-D2 were each diluted with 2 µl buffer A. Also, 2 µl D1a-TLR8-D2 and 2 µl D1b-MYC-D2 were mixed together. The final concentration of each RNA was 1 µM (TLR8 326 ng/µl, MYC 154 ng/µl, D1a-TLR8-D2 345 ng/µl, D1b-MYC-D2 193 ng/µl) in 4 µl buffer A. RNAs were kept at 37°C for 2 hr. A 1 µl native agarose gel loading buffer was added to the RNA. A total of 1 µg RNA was loaded onto the 2% agarose gel made with the Tris-borate-EDTA (TBE) buffer for electrophoresis with TBE buffer.

For denaturing agarose gel electrophoresis, glyoxal was used. RNAs were mixed with 10 µl glyoxal and incubated at 55°C for 60 min and then incubated on ice for 10 min. A 2 µl agarose gel loading buffer was added into the RNA. A total of 1 µg RNA was loaded into the 1% agarose gel made with the TAE buffer for electrophoresis with TAE buffer.

We estimated the concentration of a specific mRNA in mammalian cells is between 8 pM and 8 nM (9.5 pg/µl to 9.5 ng/µl) based on the following assumptions: (1) the volume of a HeLa cell is 2000 µm³; (2) the average length of an mRNA is 3500 nt, which corresponds to an average molecular weight of 1155 kDa; and (3) there are between 10 and 10,000 copies of mRNAs per cell (Chen et al., 2016; Chen et al., 2015).

The ensemble diversity of 3′UTR sequences was calculated using the RNAfold software (version: 2.4.14; command line: RNAfold --MEA -d2 -p --infile=<RNA_sequences.fasta> --outfile=<RNA_sequences.RNAfold.summary>) (Hofacker et al., 1994; Lorenz et al., 2011). Only 3′UTRs with a length <7500 nt can be analyzed by RNAfold. As the values for ensemble diversity depend on the sequence length, we calculated the NED by dividing the value of ensemble diversity by the length of the 3′UTR in nucleotides. All values are listed in Figure 6—source data 1. For the 47 experimentally tested 3′UTRs, the NED values range from 0.18 to 0.38. Among the RNAs that induced mesh-like condensates (N = 28), 75% of them had NED values higher than 0.28, whereas 75% of the RNAs that induced sphere-like condensates had NED values lower than 0.265. We used these cut-offs to identify 3′UTRs with high or low NED values transcriptome-wide. The range of transcriptome-wide NED values is 0–0.44.

The 3′UTR length is the full-length 3′UTR length obtained from Refseq. For counting of AU-rich elements, we only considered the canonical sequence AUUUA. We counted the number of AU-rich elements in annotated 3′UTRs of mRNAs expressed in HeLa cells. All values are listed in Figure 6—source data 1.

PAR-CLIP data of HuR were analyzed from two datasets (Lebedeva et al., 2011; Mukherjee et al., 2011). Processed peak files were downloaded from POSTAR2 (Zhu et al., 2019). Peaks were intersected with a bed file containing human 3′UTRs coordinates (hg38) using bedtools (Quinlan and Hall, 2010), and the number of CLIP tags that fall into 3′UTRs was counted. The union of CLIP tags was used to categorize the different groups, meaning that the indicated number of CLIP tags was detected in at least one dataset.

For all pair-wise comparisons, a two-sided Mann–Whitney test was performed. For comparisons containing more than two groups, a Kruskal–Wallis test was performed. The Pearson’s correlation coefficient is reported.

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

We thank all members of the Mayr lab for helpful discussions and critical reading of the manuscript. We thank Juncheng Wang for suggestions for protein purification and Neil Robertson for the mCherry-FUS-TIS construct. We also thank Bede Portz and James Shorter for attempting the purification of TIS11B and for valuable suggestions on phase separation experiments. This work was funded by the NIH Director′s Pioneer Award (DP1-GM123454), the Pershing Square Sohn Cancer Research Alliance, and the NCI Cancer Center Support Grant (P30 CA008748).

1. Received: October 22, 2020
2. Accepted: March 1, 2021
3. Accepted Manuscript published: March 2, 2021 (version 1)
4. Version of Record published: March 17, 2021 (version 2)
5. Version of Record updated: February 2, 2023 (version 3)

© 2021, Ma et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.