RS repeats are often found in SR proteins and SR-related proteins. The serine residues in RS repeats are frequently phosphorylated, a process which regulates the functions of RS repeats. To quantify the abundance of RS repeats in the human proteome, we systematically searched for uninterrupted repeats that were 2–8 amino acids in length (Figure 1A) and tested whether the length of RS repeats was correlated with protein condensates. Proteins were defined as being in condensates if they were listed in any of three available phase separation databases (PhaSepDB, LLPSDB, and DrLLPS) (Ning et al., 2020; Li et al., 2020; You et al., 2020). Because both RRM domains and RS repeats have been associated with phase separation in previous studies, we separated proteins containing RRM domains in this analysis from those that did not (Wang et al., 2018; Murthy et al., 2019). We found that the chance that a protein was found in condensates increased with the length of RS repeats it harbored regardless of whether the protein contained an RRM domain (Figure 1A). In our analysis, we did not distinguish RS repeats from SR repeats, and counting started with the first residue, whether it was R or S. We analyzed the correlation between the RS repeat length and the percentage of proteins found in condensates using correlation analysis (Supplementary file 1) and contingency tables (Supplementary file 3). Using correlation analysis, we found that the two-tailed Pearson’s p-value is 0.02, and the correlation coefficient is 0.93 (Supplementary file 1). We further extended the correlation analysis to all other possible dipeptide motifs, assuming that the order of the two amino acids in a repeat is interchangeable. As the protein population of some dipeptide motifs is low, we estimated a population-based error of 1/Nps , where N_ps is the number of proteins containing 8-mer peptides found in condensates (as described more completely in the methods section). Applying a criterion of p-value <0.05 and fraction of proteins in condensates greater than twice the population-based error, we found six dipeptide motifs showed significant correlation with phase separation: GG, KK, QQ, PP, RG, and RS (Supplementary file 1). Except for the KK motif, the five other dipeptide repeats have been shown to directly drive phase separation for some proteins (Lafontaine et al., 2021; Brangwynne et al., 2009; Molliex et al., 2015; Fei et al., 2017). It is noteworthy that among the six dipeptide motifs, RS-containing proteins have the highest percentage of 6-mer and 8-mer-containing proteins in condensates. As the sample sizes across these datasets vary widely, we performed the same analysis on 50 randomly selected size-matched subsets from each category (Supplementary file 2). We obtained similar results when size-matched datasets were used.

In addition to the effect of increased RS repeat length, we also found that proteins possessing both RS repeats and RRM domains were especially likely to be found in condensates (Figure 1A, blue bars). Among proteins with an RRM and at least one 4-mer RS repeat, the likelihood of appearance in condensates was 89%, and this trend became more pronounced as the repeat length was increased (Figure 1A). Using contingency tables, we analyzed the correlation between the presence of RS dipeptides and the occurrence of RRM domains (Supplementary file 3). For proteins found in condensates, occurrence of RS dipeptides and RRM domains is clearly correlated, as shown by a p-value of 0.0013 (Supplementary file 3). In contrast, no significant correlation was found when the same analysis was performed on proteins not in condensates (p-value = 0.1202). As repeat length increased, the correlation between RS repeats and RRM domains also increased.

Many proteins contain multiple copies of short RS repeats. In fact, most SR and SR-related proteins have several short RS repeats instead of a few long, continuous ones (Boucher et al., 2001). One extreme example is nuclear speckle scaffolding protein SRRM2, which has 56 4-mer RS repeats, more than any other protein (Figure 1—figure supplement 1A). Therefore, we also analyzed how the number of short RS repeats is correlated with phase separation. We found that the number of RS repeats that a protein harbors also affects its likelihood of being found in condensates (Figure 1B and C). In the absence of an RRM domain, on average, proteins in condensates have more copies of 2-mer or 4-mer RS repeats than those not in condensates (Figure 1B). Further, proteins with several RS repeats and an RRM domain are particularly likely to be found in condensates (Figure 1C, Supplementary file 4). Definitions of RS domains usually specify either a threshold repeat number (Boucher et al., 2001; Manley and Krainer, 2010) or a threshold percentage R/S composition (Cascarina and Ross, 2022; Manley and Krainer, 2010). To test the effect of percent R/S composition on the likelihood of phase separation, we used LCD composer (Cascarina et al., 2021). Similar to the results we found for the effect of short repeats, we found that a 20-amino acid sequence of at least 40% RS composition increased the likelihood of a protein being found in condensates from 31% to 36% (Figure 1—figure supplement 1B and C, Supplementary file 4). Further addition of at least one RRM domain increased the likelihood of phase separation to 89% (Figure 1—figure supplement 1C). In summary, we found a correlation between RS repeats and phase separation whether it was analyzed by length, number, or composition of RS repeats.

The correlation of RS-repeats with phase separation is consistent with observations that many RS- containing proteins have low solubility in vitro. For example, up to this point, none of the full-length SR proteins have been obtained in concentrations suitable for biophysical/biochemical or structural characterization, although the founding member of the family, SRSF1, was identified more than three decades ago (Krainer et al., 1990a; Ge and Manley, 1990; Krainer et al., 1990b).

To overcome this obstacle in investigating SR and SR-related proteins, we aimed to develop a new purification and solubilization method using full-length SRSF1. The high Arg composition in these proteins inspired us to use high concentrations of Arg amino acid in our protocol to purify and solubilize SRSF1. An Arg/Glu mixture of 50 mM has been used to increase solubility of some RNA-binding proteins (Golovanov et al., 2004). We found that 0.8–1 M of Arg was able to solubilize all SRSF1 constructs during the purification procedure (details in the Materials and methods section). However, the high ionic strength of Arg at this concentration range is unsuitable for many analytical methods, such as NMR and binding assays.

We predicted that we could solubilize phase-separating proteins using peptide co-solutes that mimic these repeats to compete with inter- and intra- molecular interactions (Figure 2A). We tested this concept on SRSF1, which contains RS repeats of 16, 5, and 6 amino acids, respectively (Figure 2B). Serine residues in these regions can be phosphorylated, resulting in an alternation of positive and negative charges (Figure 2B). To solubilize SRSF1 in its unphosphorylated and phosphorylated forms, we therefore tested peptides of varying lengths to mimic unphosphorylated RS (RS), and phosphorylated RS (DR, and ER) repeats (Figure 2B). Here, using the purified protein, we found that SRSF1 phase separated at concentrations lower than 300 nM in a phosphate buffer (Figure 2C and D). This was also the case when the protein was diluted into 90 mM KCl (Figure 2E).

To quantify protein solubility, we used ammonium sulfate precipitation followed by resuspension of the proteins in peptide-containing buffers (Figure 2F). This approach to measuring protein solubility has been used in many studies (Burgess, 2009; Trevino et al., 2008). If mimicking the repetitive sequences with the peptides helps resolve phase separation, we expect to see a more dramatic solubility increase when the peptide co-solutes most closely mimic the repetitive sequences of the proteins. To this end, we measured solubility of unphosphorylated full-length SRSF1, hyper-phosphorylated SRSF1 (pi-SRSF1), and RS-truncated SRSF1 (ΔRS) (Figure 2F). We found that whether phosphorylated or not, full-length SRSF1 was essentially insoluble in the 50 mM and 100 mM KCl control buffers (Figure 2F). Previous studies have found that truncation of the RS domain increases protein solubility (Tacke and Manley, 1995). This suggests that the RS domain is responsible for the protein’s low solubility. To verify this, we measured the solubility of ΔRS and found that it was overall more soluble than full-length SRSF1 in all tested buffers (Figure 2F). Although an Arg/Glu mixture has been reported to promote protein solubility (Golovanov et al., 2004), Arg/Glu at 100 mM provided only a limited solubilizing effect for full-length SRSF1 (Figure 2F). Among the peptides we tested, only RS8 dramatically increased solubility of unphosphorylated SRSF1 (from 0.6±0.29 μM in 100 mM KCl to 120±12 μM with 100 mM RS8). Consistent with our hypothesis, all tested peptides had a less dramatic solubilizing effect on ΔRS, likely because it does not contain the repeat sequences that the peptides are designed to mimic. Hyper-phosphorylation of an RS domain converts the region into a basic acidic repeat resembling ER and DR repeats. To mimic a phosphorylated RS domain, we tested the solubilizing effects of ER and DR peptides (Cho et al., 2011; Feng et al., 2012). ER8 increased the solubility of hyper-phosphorylated SRSF1 (pi-SRSF1) more than other peptides. DR8 and ER4 also had substantial, albeit less pronounced, solubilizing effects (Figure 2F). The preference for ER8 may be due to the fact that glutamic acid resembles phosphoserine more than aspartic acid in size. This confirmed our hypothesis that the solubilizing effect was more notable with peptides most closely resembling the proteins’ own repeats.

These trends are clearer when the effect of the peptides on solubility is normalized by the solubility in 100 mM Arg/Glu (Figure 2G). Whereas peptides designed to mimic the repetitive constructs produce as much as a 30-fold increase in solubility relative to Arg/Glu, the solubility increase of ΔRS only reaches about a twofold difference (Figure 2G). In accordance with our solubility tests, we found that using increasing concentrations of the RS8 peptide reduced the number of liquid-like droplets in solution (Figure 2H).

We also tested the solubilizing effect of these peptide co-solutes on two other RNA-binding proteins, Nob1 and Nop9. These peptides increased solubility of Nob1 and Nop9 by 50–60% and 2–10%, respectively (Figure 2—figure supplement 1A). Nob1 contains unstructured regions rich in basic and acidic-basic residues (Figure 2—figure supplement 1B) and has an increased solubility despite its unstructured regions having a lower homology to the tested peptides. Nop9 does not have such unstructured sequence regions (Figure 2—figure supplement 1C). Consistent with our hypothesis, the tested peptides have a moderate solubilizing effect on Nob1, whereas they have limited or no effect on Nop9. These results for SRSF1 constructs, Nob1, and Nop9 suggest that repetitive peptides improve solubility for proteins that have similar sequences.

Ionic co-solutes usually increase the dielectric constant of a sample and complicate NMR data acquisition, producing difficulty in probe tuning/matching, elongation of pulse width, and reduction of sensitivity (Wider and Dreier, 2006; Kelly et al., 2002). This imposes a considerable obstacle, as NMR is one of the few methods that provides an atomic level description of the dynamic interactions of phase-separating proteins. This adverse effect can be quantified by the elongation of the pulse width, which is inversely proportional to the signal sensitivity (Wider and Dreier, 2006). For example, increasing KCl concentration from 100 mM to 400 mM elongates the pulse width by about 50% on the NMR probe used in this study (Figure 3A). In contrast to the effect of KCl, peptide co-solutes did not significantly elongate the pulse width (Figure 3A). Whereas the 800 mM Arginine buffer used to solubilize SRSF1 during purification increased the pulse width to 16.97 µs, a combination of peptide and arginine had a less pronounced effect on the pulse width (Figure 3—figure supplement 1A). This is likely due to the low mobility of short peptides compared with salts or free amino acids (Kelly et al., 2002).

We expect the peptides to compete with homotypic inter-molecular interactions (interactions between SRSF1 molecules) to solubilize the protein, but the competition should not be strong enough to abolish binding or disrupt protein structure. With the peptide co-solutes, we were able to obtain high quality NMR spectra for both unphosphorylated and phosphorylated SRSF1. The TROSY-HSQC overlay in Figure 3B is consistent with the expected presence of both globular domains that show a higher level of dispersion and disordered regions with proton shifts in the 8.0–8.5 ppm range. Although we were able to solubilize unphosphorylated and phosphorylated SRSF1 in these respective buffers (RS8 and ER8), a buffer that solubilizes both proteins was desired to allow direct comparison of the two proteins and facilitate NMR assignment. Therefore, we examined the spectra and pulse width of different combinations of peptides and Arginine (Figure 3—figure supplement 1). We found that a buffer of 100 mM ER4 mixed with 400 mM Arg/Glu, pH 6.4, maintained structure and solubilized both unphosphorylated and phosphorylated SRSF1 constructs at concentrations above 350 µM (Figure 3—figure supplement 1B). It also weakened but did not abolish binding of SRSF1 constructs to an RNA ligand (Figure 3—figure supplement 1C) and resulted in a pulse width of 15.04 µs (Figure 3—figure supplement 1A), significantly shorter than that observed for 800 mM Arg/Glu. This buffer was used for future experiments. Using this buffer, we assigned the backbone for unphosphorylated SRSF1 (Figure 3C and D). This accomplishment enabled us to investigate the mechanism by which repetitive peptides solubilize SRSF1. To this end, we selected RS8 and unphosphorylated SRSF1 for further study.

Mimic peptides were able to provide us with control over the critical point for SRSF1 phase separation, enabling us to obtain a backbone assignment in the solution state (Figure 3). According to our hypothesis, the mimic peptide should provide transient competition for contacts with the protein’s repetitive sequence. Inter- and intra- molecular interactions should still occur under these conditions. However, they should be weakened enough to prevent droplet formation, providing us with a stable sample that can be used to study the intermolecular interactions that lead to the initiation of phase separation. To verify that this was the case, we performed a series of paramagnetic relaxation enhancement NMR experiments to probe peptide, intramolecular, and homotypic intermolecular interactions.

To locate the RS-mimic peptide interacting sites, we labeled RS8 with a paramagnetic probe (MTSL) and mixed it with SRSF1 (Figure 4A and C, and Figure 4—figure supplement 1A). The paramagnetic probe decreases intensities of residue peaks on the NMR spectrum in a distance-dependent manner (Clore and Iwahara, 2009). A higher PRE value indicates that peptides come closer to the residue analyzed. PRE is suitable for probing transient interactions, including the weak interactions between co-solutes and macromolecules (Clore and Iwahara, 2009; Okuno et al., 2021) and the intermolecular interactions that precede phase separation (Murthy and Fawzi, 2020; Ryan et al., 2018). We found that RS8 interacted primarily with RRM1 residues (Figure 4A and C, and Figure 4—figure supplement 1A). This is consistent with the fact that RRM1 (pI = 4.7) is more acidic than RRM2 (pI = 6.9), with regions of high negative charge on its two helices (Figure 4B). Dramatically perturbed sites were clustered on electronegative and aromatic sites, with the sequence D³¹IED on the α₁-Helix and D⁴⁴ID on the neighboring β-sheet being particularly perturbed (Figure 4C). Other hotspots included D⁸⁰GYR and E⁸⁷F on loops neighboring the α₁ and α₂ helices, respectively, D⁶⁶AED on the α₂ helix, and RRM2 residues W134 and A¹⁵⁰DVYR (Figure 4C).

According to our hypothesis, the RS8 peptide should provide transient competition with the RS domain without abolishing inter- and intra-molecular interactions (Figure 2A). To locate the intra-molecular interacting sites, we separately introduced the probe to the center of the RS domain (N220C, Figure 4D, Figure 4—figure supplement 1B, D) and the C-terminal end (T248C, Figure 4—figure supplement 1C, E). We found that labeling at the center of the RS domain produced the most notable perturbations (Figure 4D). To estimate the background PRE resulting from stochastic collisions, we also collected PRE data for SRSF1 mixed with the same concentration of probe alone as a control (Figure 4—figure supplement 2A). Subtracting the background PRE does not significantly change the perturbation pattern (Figure 4—figure supplement 2D and E). To verify that intermolecular interactions were not contributing to the measured intramolecular PRE, we performed a control PRE measurement by mixing an equal amount of probe-labeled (¹⁴N, N220C-MTSL) SRSF1 and ¹⁵N-SRSF1 with no probe labeling (Figure 4—figure supplement 2C). Because ¹⁴N SRSF1 cannot be detected by NMR HSQC, in this experiment, observed PRE can only happen through intermolecular interactions. At this total protein concentration of 200 µM, we did not see a significant intermolecular contribution to the PRE signal. The relative strengths of the spectra are readily observed when the intermolecular interactions under these conditions are subtracted from the intramolecular interactions (Figure 4—figure supplement 2F).

To locate the inter-molecular interacting sites, we placed the probe on the C-terminal end of the protein and doubled the concentration to a total of 400 µM protein, maintaining a 1:1 ratio of HSQC-undetectable SRSF1 with the probe attached and ¹⁵N-labeled SRSF1 possessing no cysteines (Figure 4E, Figure 4—figure supplement 1C and E). With this experimental design, only inter-molecular interactions resulted in PRE on the ¹⁵N-labeled SRSF1. Consistent with our hypothesis, the intra- and inter-molecular PRE patterns are similar to the perturbations from paramagnetic RS8.

To gain an atomic-level picture of the interactions between RS8 and SRSF1, we constructed models with the program Xplor-NIH using the PRE values as restraints (Schwieters et al., 2006; Schwieters et al., 2003) as illustrated in Figure 5—figure supplement 1. We further optimized these models using molecular dynamics simulations (Figure 5, Figure 5—figure supplement 2). The top 25% of initial Xplor-NIH structures agreed with the observed PRE data, with Pearson’s correlation coefficients of 0.916–0.941 (Figure 5—figure supplement 1). MD simulations produced structures in which a paramagnetic center was within the expected 12–15 Å from bleached residues (Supplementary file 5; Iwahara et al., 2007). Representative images are displayed in Figure 5. The hotspot D³¹IED on the α₁ helix was found to be able to provide electrostatic contacts for multiple interactions, including hydrogen bonding with bleached isoleucine residues I42 and I45 (Figure 5A). In the region surrounding W134, an electrostatic interaction with D151 was found to enable a peptide arginine to orient parallel to the aromatic face of W134, forming cation-pi stacking interactions (Figure 5B). Simultaneous cation-pi stacking interactions were also observed in the regions surrounding D80 and Y79 (Figure 5C) as well as F88 (Figure 5D).

We were curious whether interactions found in SRSF1 were conserved across the SR protein family. To this end, we used the program ClustalX (Larkin et al., 2007) to align the RRM1 sequences of SR proteins and found that interaction hotspots on α1 and α2 helices of RRM1 domains were conserved in most SR proteins (Figure 6). The electronegative charge was most highly conserved in the α2 helix and the neighboring loop, while in the α1 helix, acidic residues were replaced by arginine residues in some SR proteins. RRM1 domains have a conserved RNA binding site (Figure 6, sticks). Interestingly, these regions are distal to the RNA-binding sites of the RRM domains, suggesting the conservation may be due to a role other than RNA recognition (Figure 6). Considering the conservation of these sites involved in phase separation, the phase-separating mechanism we revealed for SRSF1 could be applicable to many other members of the SR family.

We demonstrate that electronegative α₁ and α₂ helices along with neighboring aromatic residues serve as interacting sites for unphosphorylated RS repetitive sequences. This finding has implications for how phosphorylation might change interactions within the speckles. If negatively charged residues are important for maintaining protein-protein interactions, charge repulsion between a hyperphosphorylated tail and acidic residues might be one reason that SR proteins leave the speckles upon hyperphosphorylation (Gui et al., 1994). Phosphoserines of RS repeats have been proposed to form salt bridges with neighboring arginine residues (Hamelberg et al., 2007), although any such contacts are likely short-lived and do not result in a stable secondary structure (Ngo et al., 2008; Xiang et al., 2013). Temporary phosphoserine-arginine contacts may be sufficient to compete with the highly transient cation-pi stacking interactions that we observe here.

The DNA encoding human SRSF1 was sub-cloned into pSMT3 using BamH I and Hind III. The ΔRS construct and mutants SRSF1 C16S C148S N220C (N220C), SRSF1 C16S C148S T248C (T248C), and SRSF1 C16S C148S (NoC) were prepared using mutagenesis PCR. All these mutants maintain the folded structure according to NMR spectra, and bind with SRSF1 cognate RNA UCAGAGGA. All proteins were expressed by BL21-CodonPlus (DE3) cells in LB media or minimal media supplemented with proper isotopes for NMR experiments. Hyperphosphorylated SRSF1 was prepared by co-transformation of BL21-CodonPlus (DE3) cells using pSMT3/SRSF1 and CDC2-like kinase 1 (CLK1) cloned in pETDuet-1. Cells were cultured at 37 °C to reach an OD600 of 0.6, and 0.5 mM IPTG was added to induce protein expression. Cells were further cultured 16 hours at 22 °C. The cells were harvested by centrifugation (4000 RCF, 15 min). The cell pellet was re-suspended in 20 mM HEPES, pH 7.5, 150 mM Arg/Glu, 2 M NaCl, 25 mM imidazole, 0.2 mM TCEP supplemented with 1 mM PMSF, 1 mg/mL lysozyme, 1 tablet of Pierce protease inhibitor, and 1 mM NaVO4 for the hyperphosphorylated construct. After three freeze-thaw cycles, the sample was sonicated and centrifuged at 23,710 g for 40 min using a Beckman Coulter Avanti JXN26/JA20 centrifuge. The supernatant was loaded onto 5 mL of HisPur Nickel-NTA resin and then eluted with 60 mL of 20 mM MES pH 6.5, 300 mM imidazole, 600 mM Arg/Glu, and 0.2 mM TCEP. The eluted sample was cleaved with 2 µg/mL Ulp1 for 2 hr at 37 °C. The four unphosphorylated SRSF1 constructs (WT, N220C, T248C, NoC) were further purified by a 5 mL HiTrap Heparin column. The hyperphosphorylated SRSF1 was further purified by a 5 mL Cytiva Fast Flow Q column. The eluted samples from the ion exchange step were further purified by a HiLoad 16/60 Superdex 75 pg size exclusion column equilibrated with 800 mM Arg/Glu, pH 6.5, 1 mM TCEP, 0.02% NaN₃. The protein identities were confirmed by mass spectrometry. As reported in previous study, 18-phosphates were added on the RS region of SRSF1 (Aubol et al., 2013). The protein purities were judged to be >95% based on SDS-PAGE.

Between purification and NMR experiments, the protein was transferred to peptide buffer in one of three ways: (1) It was concentrated in 800 mM Arg/Glu pH 6.5, 1 mM TCEP, 0.02% NaN₃ and diluted with a peptide buffer to the final concentration (For Figures 3D and 4D–E, Figure 3—figure supplement 1, Figure 4—figure supplement 1, Figure 4—figure supplement 2). (2) It was precipitated and re-suspended in the peptide buffer (for Figure 3B). (3) For the peptide titrations (Figure 4C), because a low concentration was needed, the initial spectrum was taken in 200 mM Arg/Glu, and peptide was titrated into the NMR tube.

Nop9 and Nob1 expression and purification are detailed in published papers (Zhang et al., 2016; Lamanna and Karbstein, 2009). SUMO-tagged proteins were induced by 0.4 mM IPTG and expressed at 22 °C overnight in E. coli strain BL21-CodonPlus (DE3). The LB miller medium was supplemented with 0.1 mM ZnSO₄ for Nob1 expression. Cell pellets were re-suspended in 25 mM HEPES, pH 7.5, 1 M NaCl, 1 mM TCEP, 25 mM imidazole, 1 mg/mL lysozyme and lysed by sonication, followed by centrifugation. The supernatant was applied to HisPur Ni-NTA resin, washed with 200 mL of loading buffer, and eluted with 25 mM HEPES, pH 7.5, 500 mM NaCl, 1 mM TCEP, 500 mM imidazole. The SUMO tag was cleaved overnight with 2 µg/mL of Ulp1 at 4 °C. The cleaved sample was purified by a 5 mL HiTrap Heparin column (GE Healthcare), and polished using a HiLoad 16/60 Superdex 200 column (GE Healthcare) equilibrated in 25 mM HEPES, pH 7.5, 500 mM NaCl, and 1 mM TCEP. The protein purities were >95% based on SDS-PAGE.

SRSF1 cultured in ²H,¹³C, ¹⁵N M9 media was concentrated to 370 µM in 100 mM ER4, 400 mM Arg/Glu, pH 6.4, 1 mM TCEP, 10% D₂O, and 0.02% NaN₃. Triple resonance assignment experiments HNCA, HNCACB, HN(CO)CA, HN(CO)CACB, HNCO, and HN(CA)CO were collected at 37 °C on a Bruker Avance III-HD 850 MHz spectrometer installed with a cryo-probe. Approximately 85% of the protein backbone region was assigned using this method. Another approximately 13% of backbone exists in the disordered state with highly degenerate sequences, which leads to heavy peak overlap. These RS and G-rich regions were grouped into clusters. Multiplicity selective in-phase coherence transfer (MUSIC) experiments were collected to further characterize the clusters and verify the assignment of the rest of the protein. MUSIC was performed on SRSF1 for the following amino acids: Ser, Arg, Thr, Asn, Ala, Tyr/His/Phe, Pro, Asn/Gln, Met, and Gly. When used in combination with analysis of the effect of paramagnetic tag placement, peak clusters were able to be assigned to locations on the disordered regions. The NMR data was processed using NMRPipe (Delaglio et al., 1995), and assignment was performed using NMRViewJ (Johnson, 2004). The assignment of the well dispersed regions (85% of the protein) has been submitted to BMRB (ID: 51299).

RS peptide with a sequence ‘SRSRSRSRC’ was synthesized and purified by GenScript with a purity >98%. The cysteine residue at the C-terminus was introduced for MTSL labeling. RS peptide was mixed with MTSL in a molar concentration ratio of 1:4. The pH was adjusted to 7.0 before a 2 hr labeling at room temperature. To remove unreacted MTSL, 10 mL of ether was added to the sample, and the mixture was vortexed and spun at 4000 rpm for 5 min. The extraction process was repeated twice. After purification, the pH of the peptide was adjusted to 6.5 using KOH, and the peptide was lyophilized. ¹H paramagnetic relaxation enhancement (PRE) data was gathered at 37 °C on a Bruker Avance III-HD 850 MHz spectrometer installed with a cryo-probe. Titrations were performed by adding solid peptide to a ¹⁵N-labeled SRSF1 construct without cysteine (the NoC construct) in 200 mM Arg/Glu, pH 6.3, 1 mM TCEP. PRE spectra were obtained for the protein in 200 mM Arg/Glu alone, 200 mM Arg/Glu with 2.5 mM peptide, 200 mM Arg/Glu with 25 mM peptide, and 200 mM Arg/Glu with 50 mM peptide. After the spectrum with 50 mM peptide was collected, MTSL was quenched using 10 mM sodium ascorbate, and a PRE experiment was run on the quenched sample.

To prepare the sample for inter-molecular PRE, an SRSF1 construct with one cysteine at the C-terminal end of the RS tail (T248C) was obtained using mutagenesis PCR, and the mutated protein was expressed by BL21-CodonPlus (DE3) cells in LB media. The protein was exchanged into an MTSL labeling buffer of 0.8 M Arg/Glu, 100 mM NaCl, 50 mM Tris-HCl pH 7.0 using a HiPrep 26/10 desalting column. The sample was diluted to a concentration of 20 µM, and MTSL was added to a concentration of 0.4 mM. The sample was incubated in the dark at 37 °C for 12 hr, after which a desalting column was used to remove unreacted MTSL. The MTSL-labeled, NMR-inactive SRSF1 was mixed in a 1:1 ratio with a ¹⁵N SRSF1 construct with no cysteine (NoC) in a buffer of 100 mM ER4, 5 mM MES, pH 6.4, 400 mM Arg/Glu, and 5% D2O. The final concentration of protein was 420 µM (210 µM SRSF1 C16/148 S T248C-MTSL and 210 µM ¹⁵N SRSF1 C16/148 S).

To measure intramolecular PRE, an SRSF1 construct with one cysteine at the center of the first RS domain (SRSF1 C16S/C148S/N220C) was obtained using mutagenesis PCR and the purification method described above with the growth performed in M9 media containing ¹⁵N isotopes. MTSL was labeling was performed as described above. The final concentration of protein was 220 µM.

The low concentration intermolecular PRE was collected between SRSF1 C16/148 S N220C-MTSL and ¹⁵N SRSF1 C16/148 S at a total concentration of 185 µM (93 µM of each construct) as described above. A control PRE experiment with free MTSL was conducted by adding 220 µM MTSL to 220 µM ¹⁵N SRSF1 C16/C148S (NoC) in the NMR buffer described above.

All PRE measurements were carried out using a pulse sequence developed by Junji Iwahara (Iwahara et al., 2007). Diamagnetic data were collected after adding 2 mM ascorbic acid. The NMR data was processed using NMRPipe (Delaglio et al., 1995) and analyzed using NMRViewJ (Johnson, 2004). PRE values and errors were estimated as described previously (Iwahara et al., 2007).

Residues were considered above noise level if their intensities at the second time point on the diamagnetic spectrum were greater than five times the standard deviation of the spectrum. Peaks below this noise level threshold were excluded from analysis on all spectra. Peaks in regions of high spectral overlap were also excluded. If at the first time point of collection (approximately 12ms after the first 90^o pulse), the intensity of the paramagnetic peak was less than or equal to half of the intensity of the diamagnetic peak at that time point, the peaks were was defined as bleached. In these circumstances, the relaxation occurred too quickly to allow fitting of the exponential decay curve. Residues G52 and R154 met the definition of both bleached and noisy. They were excluded from analysis.

Purified protein aliquots (40 μL) were incubated with 3.2 M ammonium sulfate on ice for 30 min before 10-min centrifugation at 14,000 RCF at 4 °C. After confirming that no protein was present in the supernatant, the supernatant was discarded. The pellets were re-suspended in 20 µL of corresponding buffers and shaken at room temperature for 30 min. The re-suspensions were further centrifuged at room temperature at 14,000 RCF for 5 min. The protein concentrations in supernatants were measured using UV absorbance at 280 nm. Error bars represent standard deviation from three technical repeats. The initial concentration of full-length SRSF1 constructs was 250 μM. As RS-deleted SRSF1 has a higher solubility, the initial protein concentration used for this construct was 400 μM.

An Alphafold structure was downloaded from the Uniprot website for SRSF1 ΔRS and refined using Xplor-NIH. Restraints used for refinement included dihedral angles obtained from the assignment, RDC values obtained in a previous study (Fargason et al., 2020), and chemical shift perturbations. PRE values were projected onto to the structure in PyMOL by reassigning B-factors and coloring on a ramp scale.

An Alphafold structure was downloaded from the Uniprot website for SRSF1 ΔRS and refined using Xplor-NIH. Docking of peptides was accomplished with Xplor-NIH using a restrained rigid-body simulated annealing protocol refined against the PRE, CSP, RDC, and dihedral angle data. In total, 100 Xplor-NIH structures were calculated using an ensemble size of 10. Each ensemble member had a single peptide, resulting in 10 total peptides in the model. The RRM1 domain (residues 16–90) was held in place while the N-terminus and linker were allowed full flexibility. RRM2 residues (residues 121–196) were allowed to move as a group. Linker and N-terminal residues were allowed full flexibility.

The general distance relationship for PRE is defined as Iwahara et al., 2007:(1)Γ2=1r6(μ04π)2115γI2g2μB2S(S+1)(4τc+3τc1+(ωHτc)2)

where Γ₂ is the PRE value, r is the distance between the paramagnetic center and the observed nucleus, µ₀ is the vacuum permeability constant, γ_I is the nuclear gyromacnetic ratio, g is the electron g-factor, µ_B is the electron Bohr magneton, S is the electron spin quantum number, ω_H/2π is the nuclear Larmor frequency, and τ_c is the PRE correlation time (where τ_c^–1 = τ_r^–1+ τ_s^–1, τ_r = nuclear rotational correlation time, τ_s = electron relaxation time).

Because of the flexible nature of the protein, the structures were described using multiple ensemble members. For each residue (h), the PRE was determined by the average distance across the ensemble members:(2)⟨r−6⟩h=1N∑iNri−6

where N is the number of ensemble members, and r_i is the distance between the paramagnetic center (the nitroxy oxygen of MTSL) and the nucleus under observation (the amide proton) in a single ensemble. Angle brackets (<>) indicate ensemble averages. The PRE values were back-calculated using the SBMF mode described in Iwahara et al., 2004:

Agreement between experimentally observed PRE and back-calculated PRE was assessed using the Q-factor (Q) and Pearson Correlation coefficients (R):(3)Q=∑hn{Γ2obs(h)−Γ2calc(h)}2∑hnΓ2obs(h)2(4)R=∑hn(Γhobs−Γobs¯)(Γhcalc−Γcalc¯)∑hn(Γhobs−Γobs¯)2∑hn(Γhcalc−Γcalc¯)2

where n is the number of residues for which PRE values were obtained.

The top 25% of structures possessed Pearson correlation coefficients between 0.916 and 0.941 and Q-factors between 0.454 and 0.548.

These Xplor-NIH structures were used to produce an MD-simulation starting structure with 4 peptides and 1 SRSF1 ΔRS structure. The structure was further refined using AMBER20 with the ff19SB forcefield. Solvation was performed with explicit TIP3P water molecules with 0.15 M NaCl used to balance the charges. The simulation temperature was set to 300 K, and the cutoff distance of nonbonded interactions was set to 10 Å. A simulation in which no restraints were applied was run for 201 ns. This simulation accounted for bleached residues, which, with the exception of residue 88, remained within the expected 12–15 Å from the paramagnetic probe. For residue 88, a separate simulation was run for 17 ns in which a distance restraint was used. The distance restraint maintained an interaction between F88 and peptide 2 that was generated by Xplor-NIH and consistent with the bleaching on the PRE spectrum. Distance restraints were not applied to other sites. Whereas the 10 peptides in the Xplor-NIH models accounted for all PRE data, the four peptides were only sufficient to cover the hotspot regions. For the remaining residues on the α₁ and α₂ helices, four peptides only partially accounted for PRE data (R=0.681 for residues 29–38 and 64–74), and four peptides and was not sufficient to account for the PRE values across the molecule as a whole (R=0.136).

During the simulation, the nature of interactions within the hotspots changed in some cases, but the distance between peptides and bleached residues did not significantly change. The MD trajectory analysis was performed by CPPTRAJ.

Domain annotations and sequences for human proteins were obtained from the Uniprot website. Analysis was restricted to full-length, reviewed, human proteins for which there was evidence at the protein level. A Python script was used to search for consecutive Ser-Arg or Arg-Ser repeats of 4, 6, or 8 amino acids. Identification of proteins in condensates was based on databases PhaSepDB (PhaSepDB2.0 download), DrLLPs, and LLPSDB (natural protein download).

RRM domains were identified by further restricting our Uniprot search to proteins containing RRM domains of any manual assertion. In analysis of the correlation between condensation, RS repeats, and RRM domains, we found 206 proteins that together contained 365 RRM domains. Python scripts and bioinformatic data can be accessed via the github link: https://github.com/taliafargason/Repeats_in_Condensates (Fargason, 2023).

Percent composition values were obtained using the program LCD composer developed by Sean Cascarina in the Ross lab (Cascarina et al., 2021). Search was conducted for sequences 20 amino acids in length with at least 5% composition R/S. Proteins were then crossmatched against the lists of proteins containing RRM domains and proteins found in condensates. Proteins with more than one hit were only counted once (the sequence with the highest percent composition was used).

Five categories of proteins were identified: proteins with no instances of the dipeptide (x=0), proteins with at least one instance of the dipeptide (x>2), proteins with at least one 4-mer dipeptide repeat (x>4), proteins with at least one 6-mer dipeptide repeat (x>6), and proteins with at least one 8-mer dipeptide repeat (x>8). Only proteins in category x=0 were automatically excluded from other categories. For instance, if a protein had a 16-mer RG repeat, it would be counted in the x>2, x>4, x>6, and x>8 categories but not x=0. Likewise, if a protein had neither ‘RG’ nor ‘GR’ anywhere in its sequence, it would be counted in the x=0 category only. Only the length of the longest uninterrupted repeat was considered. The number of repeats in the protein was not a factor in this analysis.

Within each category, the fraction of protein in condensates (f_ps)was calculated as:(5)fps=NpsNps+Nnp

where N_ps is the number of proteins with the repeat that have been found in condensates and N_np is the number of proteins with the repeat that have not been found in condensates.

A population-based error (E_ps) was calculated as:(6)Eps=1Nps

Dipeptide repeats were considered within this error if they existed in the 8-mer form and met the criterion:(7)Eps(x≥8)<2∗fps(x≥8)

A correlation analysis was performed between the number of repeats (x) and the fraction of proteins in condensates (f_ps). Dipeptide repeats were considered to correlate significantly with phase separation if they met the criteria of: R>0 (positive correlation) and p<0.05.

In addition to the correlation analysis described above, Fisher’s exact test and the Mann-Whitney test were used to assess the correlation between RRM domains, RS repeats, and phase separation. Because in each of these analyses, two factors were being compared against a third factor, Bonferroni’s adjustment was used to set the significance threshold to p<0.025. For the cases in which Fisher’s exact test were used, contingency tables are included in Supplementary Files (Supplementary files 3-4).

Because the number of RS repeats and percent R/S composition both occur across a broad distribution, the Mann-Whitney test was employed to determine to what extent these distributions differed between proteins in condensates and proteins outside of condensates. Because RRM domains are associated with both phase separation and an increase in the number of RS repeats, these variables were separated. The Mann-Whitney test is suitable for non-normal distributions with different sample sizes (Widen et al., 2020). It is applicable in cases in which the sample size is greater than 30. It should be noted that the sample size of our smallest category in Figure 1B and Figure 1—figure supplement 1B (Proteins that have RRM domains but do not phase separate).

An SRSF1 construct (SRSF1 C16S/C148S/N220C) was tagged with Maleimide-Alexa488, dissolved into 800 mM Arg/Glu, pH 6.5, 0.2 mM TCEP, and stored at a concentration of 28 µM. Surplus Alexa488 dye was removed by a desalting column. The protein was then diluted to its final concentration in 100 mM KCl, 10 mM MES pH 6, 0.1 mM TCEP, with or without short peptides in a 96-well Cellvis glass bottom plate coated with Pluronics F127. Images are brightfield/GFP channel overlays taken on a Cytation5 imager using the software Gen5 3.10. More than three biological replicates were performed for phase separation experiments.