Nuclear speckle proteins form intrinsic and MALAT1-dependent microphases

SUMMARY

Pre-mRNA processing components in nuclear speckles encompass one or more folded RNA recognition motifs (RRMs) and disordered regions with specific sequence grammars. Such proteins include serine/arginine-rich splicing factors (SRSFs) and transactive response DNA binding protein (TDP)-43. The SRSFs and TDP-43 are unique archetypes of block copolymers encoding specific patterns of inter-domain homotypic and heterotypic attractions and repulsions. The interplay of these interactions drives microphase separation and the formation of ordered, size-limited assemblies. Microphases of SRSFs and TDP-43 are 23–45 nm in diameter, each comprising tens of molecules. Sub-micron-scale assemblies of SRSFs in cells are consistent with being clusters of microphases. The speckle-associated regulatory long non-coding RNA (lncRNA) metastasis-associated lung adenocarcinoma transcript 1 (MALAT1) binds specifically and preferentially to SRSF1 microphases, while destabilizing TDP-43 microphases. In protein mixtures, the interactions between microphases drive the formation of micron-scale double-emulsion structures with core-shell organization. Our findings show how interactions involving copolymers featuring folded domains and disordered regions drive the formation of microphases.

Graphical Abstract

In brief

In nuclear speckles, sequence-encoded inter-domain attractive and repulsive interactions drive proteins to form hierarchical structures, including size-limited microphases, clusters of microphases, micron-scale double emulsions in protein mixtures, and microphases stabilized by preferential and protein-specific binding to the lncRNA MALAT1.

INTRODUCTION

In metazoan cells, nuclear speckles function as hubs that regulate key steps of gene expression.^1-8 Speckles are enriched in RNA-binding proteins that contribute to pre-mRNA splicing.^8-13 Splicing factor activities are regulated by polyadenylated RNAs and regulatory long non-coding RNAs (lncRNAs) that are enriched in the transcriptomes of speckles.^7,14-16

It has been proposed that nuclear speckles form via macrophase separation or liquid-liquid phase separation (LLPS) of individual components.¹⁷ Drivers of macrophase separation feature uniformly attractive interactions^18,19 (Figure 1A), and macrophase sizes are orders of magnitude larger than the nanoscale dimensions of the underlying macromolecules, encompassing millions of molecules. At equilibrium, a macrophase would be a single dense phase that coexists with the dilute phase.²⁰ However, equilibrium macrophases are not typically observed in cells. Control over size can arise through active processes,^21,22 modulation by solid particles,²³ confinement effects,^24,25 or dynamical arrest.²⁶ Super-resolution structured illumination microscopy has shown that nuclear speckle components are organized inhomogeneously into distinct territories, forming size-limited sub-micron-scale structures.^27-33 Based on this, we conjectured that nuclear speckles might be emulsions of microphases formed by block copolymer-like molecules.

Figure 1. SRSF1 is an unconventional block copolymer.

(A) Schematic of a linear flexible homopolymer that undergoes macrophase separation or LLPS featuring uniformity of attractions (blue lines). Above c_sat, a single micron-scale dense phase should form that coexists with the dilute phase. Inset shows a schematic of how millions of molecules would be homogeneously mixed within dense phases.

(B) Schematic of a conventional di-block copolymer featuring disordered A- (purple) and B-blocks (green). Homotypic inter-block interactions are attractive, whereas heterotypic interactions are repulsive.

(C) Sequence architecture of hnRNP-A1, which forms macrophases in vitro and in cells.

(D) Space-filling model of hnRNP-A1. Residues are colored based on residue type, with blue, red, green, and white spheres signifying basic, acidic, polar, and hydrophobic residues, respectively.

(E) Sequence architecture of SRSF1.

(F) Space-filling model of SRSF1.

(G) Computed ${\Delta }_{\mathrm{X}\mathrm{Y}}$ values for pairs of domains within hnRNP-A1.

(H) Computed ${\Delta }_{\mathrm{X}\mathrm{Y}}$ values for pairs of SRSF1 domains.

(I) Sequence architecture of SRSF1Δ213–247.

(J) Space-filling model of SRSF1Δ213–247.

(K) Computed ${\Delta }_{\mathrm{X}\mathrm{Y}}$ values for pairs of SRSF1Δ213–247 domains.

See also Figure S1.

Conventional block copolymers comprise disordered blocks that are mutually incompatible with one another^34-42 (Figure 1B). The homotypic inter-block interactions are attractive, and heterotypic inter-block interactions are repulsive^36,40-43 (Figure 1B). At finite concentrations, the homotypic inter-block attractions can drive phase separation. These attractions compete against heterotypic inter-block repulsions that oppose growth into macrophases. The result is the formation of size-stabilized assemblies known as microphases that are defined by system-specific nanoscale morphologies and internal ordering.^36,41,44

Here, we show that speckle-associated serine/arginine-rich splicing factor (SRSF)1, SRSF3, SRSF5, SRSF7, and TDP (transactive response DNA binding protein)-43 are unconventional block copolymers that undergo microphase separation. The regulatory lncRNA MALAT1 (metastasis-associated lung adenocarcinoma transcript 1) binds specifically and preferentially to remodel SRSF1 microphases.⁴⁵

RESULTS

Establishing the block copolymeric nature of SRSF1

Multi-domain RNA recognition motif (RRM)-containing proteins, such as hnRNP-A1 (Figures 1C and 1D) and SRSF1 (Figures 1D and 1F), have distinct RRMs and intrinsically disordered regions (IDRs). Folded RRMs are pre-organized by intra-domain microphase separation⁴⁶ and have protein-specific physicochemical properties (Figures S1A-S1F). The sequence grammars of the IDRs are also different.^47-49 IDR1 in SRSF1 and the low complexity domain (LCD) of hnRNP-A1 (A1-LCD) are enriched in Gly-rich patches, high non-random Gly content, and the presence of arginine-glycine (RG)-motifs (Figure S1G). However, the two regions differ in the Phe, Asn, and aromatic contents (enriched in A1-LCD), and the extent of linear segregation of Gly residues (high in IDR1 of SRSF1). The IDR2 of SRSF1 shows non-random enrichments of Ser/Arg residues (Frac S, Frac R, R patch, S patch).

Next, we used atomistic simulations to characterize homotypic and heterotypic inter-domain interactions for hnRNP-A1 (Figure 1G) and SRSF1 (Figure 1H). For each pair of domains X and Y, we extracted an inter-domain interaction parameter, ${\Delta }_{\text{XY}}$, where $-1\le {\Delta }_{\text{XY}}\le +1$ (see STAR Methods). Negative values of ${\Delta }_{\text{XY}}$ signify attractions, whereas positive values reflect repulsions. The magnitudes of ${\Delta }_{\text{XY}}$ quantify the strengths of interactions. If $\mid {\Delta }_{\text{XY}}\mid \approx 0$, then interactions between X and Y are nearly ideal. Note that each element in the inter-domain interaction matrix is a grand sum over the collection of inter-residue attractions and repulsions, and changes at the residue level directly affect the calculated ${\Delta }_{\text{XY}}$.⁵⁰

For hnRNP-A1, all inter-domain interactions are either attractive or nearly ideal (${\Delta }_{\text{XY}}<0$ or ${\Delta }_{\text{XY}}\approx 0$) (Figure 1G). This explains why it drives macrophase separation.^51,52 For SRSF1, the homotypic inter-domain interactions are either strongly repulsive (RRM1-RRM1, IDR1-IDR1, IDR2-IDR2) or nearly ideal (RRM2-RRM2) (Figure 1H). Heterotypic IDR-RRM interactions (IDR1-RRM1, IDR1-RRM2, IDR2-RRM1, and IDR2-RRM2) are attractive, whereas heterotypic IDR1-IDR2, and RRM1-RRM2 interactions are repulsive. The pattern of homotypic repulsions combined with heterotypic attractions and repulsions makes SRSF1 an unconventional tetra-block copolymer (compare Figure 1H to Figure 1B). Cellular studies have used SRSF1Δ213–247, featuring an internal deletion within IDR2^45,53,54 (Figures 1I and 1J). The interactions for SRSF1Δ213–247 are qualitatively similar to those of SRSF1, although the inter-IDR repulsions are weaker. The interactions for SRSF1Δ213–247 may be viewed as an interpolation of interaction patterns of SRSF1 and SRSF9, which is another tetrablock SRSF (Figures 1H and 1K versus Figures S1H-S1J).

Threshold concentrations for phase separation

We measured right-angle light scattering (RALS) in vitro to assess the existence of a threshold concentration (${\mathrm{c}}_{\mu }$) for phase separation of SRSF1 and SRSF1Δ213–247.⁵⁵ For microphase-forming systems, the existence of a threshold concentration signifies a system in the strong segregation limit.⁴⁴ We performed RALS measurements over a 100-fold concentration range in aqueous buffers (10% glycerol, 20 mM HEPES, pH 7.4, 50 mM KCl, 5 mM MgCl₂) and observed discontinuities at 0.57 ± 0.03 μM for SRSF1 and 0.45 ± 0.05 μM for SRSF1Δ213–247 (Figures 2A, S2A, and S2B).⁴⁵ The measured ${\mathrm{c}}_{\mu }$ increased with increasing [KCl] (Figure S2C). To compare sequence-intrinsic interactions, all subsequent measurements for all systems were performed using identical solution conditions. Measurements of time-dependent fluorescence intensities of pyrene, a probe that partitions into dense phases if the microenvironments are distinct from those of dilute phases,⁵⁶ yielded estimates of ${\mathrm{c}}_{\mu }$ that were consistent with RALS (Figures S2D and S2E).

Figure 2. SRSF1 and SRSF1Δ213–247 form microphases.

(A) c_μ of SRSF1 = 0.57 ± 0.03 μM and for SRSF1Δ213–247 = 0.45 ± 0.05. Legends show the mean ($\overline{\mathrm{d}}$) ± standard deviation (SD) and median ($\tilde{\mathrm{d}}$) values.

(B) QF-DEEM images at three different magnifications of the microphases formed by SRSF1.

(C) QF-DEEM images at three different magnifications of SRSF1Δ213–247 microphases.

(D and E) Histograms of diameters of microphases constructed from 1,516 images for SRSF1 and (E) 753 images for SRSF1Δ213–247 (mean ± SD).

(F and G) Show results from TRFQ experiments (mean ± SD, see STAR Methods). Estimates of n_μ, values of slopes, and unreliability index (Φ) are shown.

(H) Histogram of R_g values for SRSF1 from atomistic simulations.

(I) Histogram of R_g,μ values from atomistic simulations of microphases comprising 24 SRSF1 molecules (mean ± SD).

(J) Profiles of g_D(r). Legend provides annotation of color coding, and the inset is a zoomed-in version.

(K) Representative snapshot from simulations showing 24 SRSF1 molecules within a microphase (see Video S1).

(L) Histogram of R_g values for SRSF1Δ213–247 extracted from atomistic simulations.

(M) Histogram of R_g,μ values from simulations of microphases comprising 48 SRSF1Δ213–247 molecules (mean ± SD).

(N) Profiles of g_D(r).

(O) Representative snapshot from simulations showing 48 SRSF1Δ213–247 molecules within a microphase.

See also Figure S2.

Structural characterization reveals the formation of microphases

We used quick freeze deep-etch electron microscopy (QF-DEEM) for structural characterization of microphases. Protein samples were flash-frozen against a polished copper block cooled by liquid He at 4 K⁵⁷ and fractured at 170 K to optimize etching, deposition, and imaging. This avoided drying, wicking, or exposure to air-water interfaces.⁵⁸ SRSF1 formed spheroidal nanoscale structures (Figure 2B) whereas SRSF1Δ213–247 formed biconcave-disc-like structures (Figure 2C). We analyzed images spanning a large field of view to quantify the diameters of 1,516 SRSF1 (Figure S2F; see STAR Methods) and 753 SRSF1Δ213–247 microphases. The mean and median values for the diameters are 23.6 ± 6.0 and 22.4 nm for SRSF1 and 41.0 ± 8.7 and 42.5 nm for SRSF1Δ213–247 (Figure 2D).

Next, we estimated the number of protein molecules (${\mathrm{n}}_{\mu }$) per microphase by measuring time-resolved fluorescence quenching (TRFQ) of pyrene in the presence of cetylpyridinium chloride as a quencher.⁵⁹ We estimated ${\mathrm{n}}_{\mu }$ of ~23 for SRSF1 and ~50 for SRSF1Δ213–247 microphases (Figures 2F and 2G), corresponding to apparent packing fractions of p ≈ 0.29 and p ≈ 0.1 for SRSF1 and SRSF1Δ213–247 microphases, respectively.

We performed atomistic simulations, using the ABSINTH implicit solvation model and forcefield paradigm,^60,61 of individual SRSF1 molecules to obtain a mean radius of gyration $({\mathrm{R}}_{\mathrm{g},\text{SRSF}1})$ of 2.7 ± 0.2 nm (Figure 2H). We also performed atomistic simulations of 24 SRSF1 molecules to characterize the internal organization within microphases. The mean radius of gyration of the simulated microphase $({\mathrm{R}}_{\mathrm{g},\mu })$ was 9.8 ± 0.2 nm (Figure 2I). The spatial organization of domain sites within simulated microphases was analyzed in terms of radial distribution functions, ${\mathrm{g}}_{\mathrm{D}}(\mathrm{r})$ (Figure 2J). For a distance $\mathrm{r}\approx {\mathrm{R}}_{\mathrm{g},\text{SRSF}1}$ of an SRSF1 molecule, the relative density of IDR2 within the interior is the lowest, and that of IDR1 is the highest of the four domains. The densities of the two RRMs proximal to the center-of-mass lie between those of the two IDRs. Beyond a length scale of 8 nm, there is a crossover, and the density of IDR2 is higher than that of IDR1 (Figure 2J, inset). Due to confinement imposed by the spheroidal envelope of the microphase, all ${\mathrm{g}}_{\mathrm{D}}(\mathrm{r})$ profiles converge and decay to zero at the edge of the envelope (Figure 2K; Video S1).

From atomistic simulations, the mean ${\mathrm{R}}_{\mathrm{g},\text{SRSF}1\Delta 213-247}$ is 2.6 ± 0.2 nm (Figure 2L), and the ${\mathrm{R}}_{\mathrm{g},\mu }$ of SRSF1Δ213–247 assemblies was found to be 12.2 ± 0.3 nm (Figure 2M; Video S2). The ${\mathrm{g}}_{\mathrm{D}}(\mathrm{r})$ profiles show three distinct layers. First, there is preferential accumulation of all domains at short-range $\mathrm{r}<{\mathrm{R}}_{\mathrm{g},\text{SRSF}1\Delta 213-247}$ (Figure 2N) with the highest density of IDR1 and IDR2 and depletion of RRM2 near the center-of-mass. Second, there is a hierarchical accumulation of domains on the intermediate length scale, with the highest density being for RRM1, followed by IDR1, RRM2, and IDR2 in that order. Third, the densities converge at a radial distance of 15 nm and decay to zero at the edge of the microphase.

We also characterized the phase behavior of SRSF1ΔIDR2, a construct that lacks the entirety of IDR2. The threshold concentration was ${\mathrm{c}}_{\mu }=0.44\pm 0.08\mu \mathrm{M}$ (Figure S2G). QF-DEEM revealed a polydisperse distribution of spheroidal assemblies defined by mean and median diameters of 33.6 ± 13.7 and 28.8 nm, respectively (Figures S2H and S2I).

Microphases associate to form sub-micron-scale clusters

As protein concentrations increase, the abundance of microphases also increases. This enables short-range attractions that drive clustering of microphases, such that condensed counterions are shared among microphases within a cluster⁶² (Figure 3A). The buildup of charge within clusters of microphases engenders long-range repulsions,^63,64 thereby stabilizing the sizes of clusters.

Figure 3. Microphases associate to form sub-micron-scale clusters.

(A) An accumulation of surface charge shown by blue + marks will draw a counterion cloud (dashed red envelope) around each microphase (solid blue circles) to enable clustering via short-range attractions and long-range repulsions.

(B) Distributions of cluster sizes extracted from MRPS measurements of SRSF1Δ213–247 above c_μ. Solid lines and shaded regions show mean and SD, respectively.

(C) Confocal microscopy showing the positive staining with the ANEPPS dye.

(D) Images from confocal microscopy, collected at concentrations above c_μ. Schematic depicts how microphases form clusters and networks of clusters.

(E) Box and whisker plot from analysis of QF-DEEM images yielded estimates for the number of microphases within a ~380 nm sized cluster. The box extends from the 25^th to 75^th percentiles. The line in the middle of the box is at the median value of 23 microphases per cluster. The whiskers extend from the minimum to the maximum number of microphases detected in the probe spheres.

(F) Single-molecule microscopy with photoactivatable Janelia Fluor 549-labeled SRSF1 showing that clusters are composed of multiple microphases. The image is produced as a maximum projection of 200 frames collected, where the color bar shows the fluorescence intensity.

(G) Histograms of fluorophores with trajectories of less than 50 nm (slow) or more than 50 nm (fast) over each 20 ms frame for 20,000 frames each, with an inset for a zoomed-in field of view of the outlined area. The histograms are represented as heatmaps mapped onto the image.

(H) Excess variance computed over bins that quantify the displacements of slow-versus fast-moving molecules.

(I) Confocal microscopy images of AF488-SRSF1Δ213–247 showing destabilization and loss of clusters and microphases with increasing concentrations of L-arginine.

(J) RALS data, shown at two concentrations of L-arginine.

(K) Confocal microscopy images AF488-SRSF1Δ213–247 showing loss of clustering in the presence of increased v/v% of PEG8000.

(L) Microphase separation of SRSF1Δ213–247 is preserved in the presence of 10% v/v of PEG8000.

(M) Image of GFP-tagged SRSF1Δ213–247 bodies in X. laevis GVs.

(N) Merge of images collected simultaneously via bright-field microscopy and the intensity in the GFP channel shows that the SRSF1Δ213–247 bodies are proximal to Cajal bodies in the GVs.

(O) CDF of the diameters of SRSF1Δ213–247 positive bodies remains invariant at different levels of the injected mRNA. Latrunculin A (Lat-A) treatments (orange dotted) to oocytes injected with 100 ng/μL (orange) mRNA did not result in an increase in the size of the intra-GV SRSF1Δ213–247 positive bodies.

See also Figure S3.

We measured the size distributions of SRSF1Δ213–247 at concentrations above ${\mathrm{c}}_{\mu }$ using microfluidic resistive pulse sensing (MRPS) (see STAR Methods).^50,65 Over a 5-fold range of concentrations, the MRPS measurements reveal the presence of clusters of sizes between 358.9 ± 20.4 and 386.1 ± 26.3 nm. While the sizes of clusters stayed essentially the same, their abundance increased with SRSF1Δ213–247 concentration (Figure 3B).

The buildup of charge at interfaces should be measurable as an electric field. We assessed this using the di-4-ANEPPS (di-4-aminoaphthylethenylpyridinium)⁶⁶ dye, which fluoresces in response to a change in the presence of a local electric field. Using confocal microscopy, we detected strong fluorescence signals when we incubated the dye with SRSF1Δ213–247 at concentrations above ${\mathrm{c}}_{\mu }$ (Figure 3C).

We used confocal microscopy for direct imaging of the clusters (Figures 3D and S3A). SRSF1Δ213–247 molecules formed clusters of ~400 nm in diameter, equivalent to sizes measured via MRPS. Cluster sizes did not change as SRSF1Δ213–247 concentrations were increased; the abundance of clusters increased with increasing SRSF1Δ213–247 concentration (Figures 3D and S3A). We estimated a median number of ~23 microphases within a cluster of diameter ~380 nm (Figures 3E and S3B).

To study the internal dynamics of molecules within microphases and clusters, we used single-molecule localization microscopy⁶⁷ and measured the blinking of SRSF1Δ213–247 labeled with the photoactivatable Janelia Fluor 549. Labeled molecules were mixed with unlabeled molecules at a 1:50 molar ratio. The maximum projection of 2 × 10⁴ frames (20 ms per frame) collected over 6.7 min onto a single image showed the presence of clusters, mirroring observations from confocal imaging (Figure 3F).

We utilized single-particle tracking photoactivated localization microscopy (sptPALM)⁶⁸ to track motions of SRSF1Δ213–247 molecules (Figure S3D). The cumulative distribution function (CDF) of displacements showed that the probability of finding molecules with displacements that are less than or equal to 50 nm is 0.5 (Figure S3E). The CDF allowed for the classification of molecules as being slow-moving (Video S3) within a given frame. The displacements of slow-moving molecules were less than 50 nm. This length scale is equivalent to the size of a single SRSF1Δ213–247 microphase (Figure S3F). Molecules were defined to be fast-moving if the displacements were larger than 50 nm (Video S4). Comparing the histograms for the two sets of displacements, we observed that there were more slow molecules, implying that SRSF1Δ213–247 molecules were mostly confined within microphases (Figure 3G). Analysis of the excess variance of displacements across all bins within a mask showed prominent outliers, with some bins featuring a strong localization of fluorophores within a ~50 nm region, corresponding to the sizes of individual microphases (Figures 3H and S3G). In fluorescence recovery after photobleaching (FRAP) of Alexa Fluor 488-labeled SRSF1Δ213–247 clusters, the fluorescence intensity did not recover for at least 120 s after photobleaching (Figure S3C).

Disruption of microphases and clusters

Addition of 500 mM L-arginine^69-71 abrogated microphases (Figures 3I and 3J). These results are concordant with previous observations⁷² showing an increase in solubility in the presence of an octameric peptide of RS repeats (RSRSRSRS) or a mixture of arginine and glutamate (Arg/Glu) in solution.

The addition of high-molecular-weight polyethylene glycol (PEG8000) destabilized clustering without disrupting microphase separation (Figures 3K and 3L). Low-molecular-weight excipients such as ethylene glycol were ineffective at disrupting clustering (Figure S3H). PEG polymers exhibit long-range repulsions from positively charged surfaces⁷³ and shielding happens on the length scale of microphases.

SRSF1Δ213–247 forms size-limited clusters in cells

We leveraged the large sizes of the Xenopus laevis oocytes to quantify the sizes of SRSF1Δ213–247 assemblies in germinal vesicles (GVs). We expressed GFP-tagged SRSF1Δ213–247 by microinjection of different mRNA concentrations into oocytes. Following the expression, the GVs were extracted and imaged using confocal microscopy.⁷⁴ We observed more than 200 SRSF1-positive bodies in each GV (Figure 3M). Some of the SRSF1Δ213–247-positive bodies were attached to a larger non-fluorescent body as observed by bright-field microscopy (Figures 3N and S3I). This is consistent with nuclear speckles being attached to Cajal bodies.⁷⁵ Irrespective of the injected mRNA concentration, the sizes of SRSF1Δ213–247-positive bodies were narrowly distributed with a mean diameter of ~300–400 nm (Figure 3O). The sizes of SRSF1Δ213–247-positive bodies are consistent with the clusters of microphases that were observed in vitro (Figures 3B and 3D).

Nucleoli and histone locus bodies fuse and coarsen upon incubation of X. laevis oocytes with latrunculin A (Lat-A), which disrupts nuclear actin.⁷⁴ In contrast, the size distributions of SRSF1Δ213–247-positive bodies did not change with Lat-A treatment (Figure 3O). Therefore, the uniform size distributions observed for SRSF1Δ213–247-positive bodies in GVs appear to be intrinsic properties of clusters of SRSF1Δ213–247 microphases.

SRSF3, SRSF5, and SRSF7 also form microphases

Next, we studied SRSF3, SRSF5, and SRSF7 as representatives of different architectures⁷⁶ (Figures 4A, S4A, and S4B). In HeLa cells, SRSF1 is the most abundant of the SRSFs, followed by SRSF3 and SRSF7. SRSF5 is of relatively low abundance.⁷⁷ The C-terminal IDRs are enriched in arginine and serine (red box in Figure S4C). SRSF7 has the longest SR-rich IDR. Additionally, the IDR1 of SRSF1 and the 38-residue linker between the RRM and C-terminal IDR of SRSF7 have sequence grammars that are shared with prion-like domains.¹⁸

Figure 4. Microphase separation is a shared feature of SRSFs.

(A) Sequence architectures of SRSF1, SRSF3, SRSF5, and SRSF7.

(B–D) Inter-domain _ΔXY coefficients for SRSF3 (B), SRSF5 (C), and SRSF7 (D).

(E) QF-DEEM images of 5 μM SRSF3 (top), 10 μM SRSF (middle), and 2 μM SRSF7 at two different magnifications (left) versus (right).

(F) Histograms of diameters (in nm) constructed from 1,446 images for SRSF3 (top), 1,656 images for SRSF5 (middle), and 656 images for SRSF7 (bottom) (mean ($\overline{\mathrm{d}}$) ± SD and median ($\tilde{\mathrm{d}}$)).

(G) Sub-micron-scale clusters formed by SRSF1 (top left), SRSF7 (top right), SRSF3 (bottom left), and SRSF5 (bottom left).

(H–J) Results from microinjections and expression of 200 ng/μL mRNAs of mCherry-tagged SRSF3 (H), SRSF5 (I), and SRSF7 (J).

(K) CDFs of size distributions of bodies formed by SRSF3, SRSF5, and SRSF7.

(L) Bright-field (top), fluorescence microscopy (middle), and merged (bottom) images obtained from injections of different amounts (shown across the columns) of mCherry-tagged mRNAs of SRSF7.

(M) CDF showing the evolution of sizes formed by SRSF7 bodies at different injection levels.

See also Figure S4.

Inter-domain homotypic interactions in SRSF3 are nearly ideal (RRM-RRM) or repulsive (IDR-IDR), whereas IDR-RRM interactions are attractive (Figure 4B). The inter-domain interaction patterns for SRSF5 are qualitatively similar to those of SRSF1. However, the homotypic inter-IDR2 repulsions are stronger in SRSF5, and IDR2-RRM2 attractions are weaker (Figure 4C). The presence of the linker introduces a new degree of freedom into SRSF7 when compared with SRSF3 (Figure 4D). Thus, SRSF5 is a tetra-block system, whereas SRSF3 and SRSF7 are di- and triblock systems, respectively.

The measured values of ${\mathrm{c}}_{\mu }$ were 1.0 ± 0.14 μM for SRSF7, 2.0 ± 0.14 μM for SRSF3, and 9.5 ± 0.71 μM for SRSF5 (Figures S4D-S4G). The fluorescence lifetimes of pyrene corroborated the ${\mathrm{c}}_{\mu }$ values, with the rate constants plateauing to protein-specific values of 1.01 ± 0.1, 0.93 ± 0.17, and 1.08 ± 0.17 ns⁻¹ for SRSF7, SRSF3, and SRSF5, respectively (Figure S4H). QF-DEEM imaging showed that all three systems formed spheroidal assemblies (Figure 4E). The mean and median diameters were found to be 39.4 ± 8.0 nm and 39.4 nm for SRSF3, 26.6 ± 7.8 nm and 24.5 nm for SRSF5, and 29.1 ± 11.1 nm and 26.0 nm for SRSF7 (Figure 4F).

The TRFQ assay yielded estimates of ${\mathrm{n}}_{\mu }$ that are highly unreliable (Figures S4I-S4K). Hence, we used the distribution of ${\mathrm{R}}_{\mathrm{g}}$ values of individual molecules, the size distributions of the spheroidal assemblies obtained from QF-DEEM, and an assumption of equivalent packing fractions to that of SRSF1 assemblies to obtain estimates of ${\mathrm{n}}_{\mu }$, which were ~93, ~14, and ~8 for microphases of SRSF3, SRSF5, and SRSF7, respectively.

MRPS measurements (Figures S4M-S4P) showed that SRSF7, SRSF3, and SRSF5 formed clusters that were ~275–300 nm in diameter. These sizes remained invariant over a range of protein concentrations, although the abundance of clusters increased with increasing concentration. Confocal imaging using fluorescently labeled SRSF3, SRSF7, and SRSF5 showed the formation of sub-micron-scale clusters for each protein (Figure 4G).

We injected mRNAs expressing mCherry-tagged SRSF3, SRSF5, and SRSF7 into the X. laevis oocytes. At 200 ng/μL mRNA, SRSF3 and SRSF5 formed de novo nuclear bodies that were not adjacent to nuclear speckles or Cajal bodies (Figures 4H and 4I). At the same mRNA concentration, SRSF7 formed bodies that are adjacent to nuclear speckles (Figure 4J). The CDFs of the size distributions of bodies formed by SRSF3 and SRSF5 show a sharp rise (Figure 4K). This is consistent with the bodies having uniform size distributions. The sizes of the bodies are also consistent with estimates from MRPS measurements in vitro (Figure S4L).

We quantified how the sizes of SRSF7 bodies varied with mRNA concentration. At 50 ng/μL, SRSF7 forms bodies that are adjacent to Cajal bodies (Figure 4L). However, above 1,000 ng/μL, SRSF7 forms macrophases that are distributed across the nucleoplasm (Figures 4L and 4M). These data are consistent with high protein levels and possibility of interactions with other nuclear components affecting the overall phase behavior of SRSF7 in GVs.

Double emulsions form in mixtures of microphases

Given the results observed for SRSF7 expression in GVs, we investigated the possibility of microphases being remodeled via heterotypic protein interactions. We computed the normalized interaction parameters ${\Delta }_{\text{XY}}$ for pairs of domains from SRSF1Δ213–247 and SRSF3 (Figure 5A). The inter-RRM interactions are nearly ideal; the interactions between SRSF1-IDR1 and SRSF3-RRM and that of the SRSF3-IDR with SRSF1-RRM1 are equivalently attractive; the interactions between SRSF1Δ213–247-IDR2 and SRSF3-RRM are also attractive; the inter-IDR interactions are different, albeit strongly repulsive. Equivalent interaction matrices were computed for mixtures of SRSF1Δ213–247 with SRSF5 and SRSF7 (Figures S5A and S5B).

Figure 5. Microphase modulation in mixtures and by chimeras.

(A) Interaction matrices for pairs of domains drawn from SRSF1Δ213–247 and SRSF3.

(B) QF-DEEM images of nanoscale structures of microphases formed in mixtures of SRSF1Δ213–247 with SRSF3.

(C) Histogram of diameters (in nm) constructed from the analysis of 1,252 distinct images (mean ($\overline{\mathrm{d}}$) ± SD and median ($\tilde{\mathrm{d}}$)).

(D–F) Fluorescence microscopy images of sub-micron-scale structures formed in mixtures of SRSF1Δ213–247 and SRSF3 (D), SRSF5 (E), and SRSF7 (F). In (D)–(F), yellow lines represent line scans, and magenta arrows point to spherical caps at interfaces. Rightmost column shows normalized fluorescence intensities in SRSF1Δ213–247 (green) and SRSFx (cyan) channels.

(G) Impact of each of SRSF3, SRSF5, and SRSF7 on c_μ of SRSF1Δ213–247.

(H) Matrix of ${\Delta }_{\mathrm{X}\mathrm{Y}}$ for ch-SRSF1.

(I) Data from three independent RALS measurements for ch-SRSF1 yield a c_th value (mean ± SD) shown in the inset. Data shown are from a single measurement.

(J) Fluorescence microscopy images of macrophases formed at different concentrations of ch-SRSF1.

(K) ${\Delta }_{\mathrm{X}\mathrm{Y}}$ for ch-hnRNP-A1.

(L) Data from a single RALS measurement for ch-hnRNP-A1.

(M) Fluorescence microscopy images of ch-hnRNP-A1 collected using samples that comprise 1% of molecules with an AF488 label at the N terminus.

(N and O) Results of the injection of mRNAs of mCherry-tagged ch-SRSF1 (N) and ch-hnRNP-A1 (O) into X. laevis. The images on the left are from bright-field microscopy, those in the middle are from fluorescence microscopy, and the rightmost image shows the merge of the two. ch-hnRNP-A1 forms size-limited assemblies that co-localize with Cajal bodies (O), see CDFs in (P), ch-SRSF1 forms bodies of heterogeneous sizes (N).

See also Figure S5.

We used QF-DEEM to analyze the nanoscale structures formed in mixtures of 2 μM SRSF1Δ213–247, which is above its intrinsic ${\mathrm{c}}_{\mu }$, and 1 μM SRSF3, which is below its ${\mathrm{c}}_{\mu }$ (Figure 4E). The images show biconcave discs and spheroids (Figure 5B). The mean and median sizes were 33.7 ± 13.3 nm and 37.9 nm, respectively (Figure 5C). The biconcave structures were akin to those of the SRSF1Δ213–247 microphases. The spheroids are likely a result of SRSF3 molecules adhering to the interface of pre-formed SRSF1Δ213–247 microphases.

The size-limiting behavior of SRSF1Δ213–247 microphases implies that their interfacial free energies are likely to be low or vanishing.²² Adhesive interactions with an excipient can alter the interfacial free energy and remodel microphases.⁷⁸ We explored this possibility using confocal imaging in binary mixtures of SRSF1Δ213–247 with each of SRSF3 (Figure 5D), SRSF5 (Figure 5E), or SRSF7 (Figure 5F). In all cases, the concentration of SRSF1Δ213–247 was above its ${\mathrm{c}}_{\mu }$, whereas those of the other SRSFs were below their ${\mathrm{c}}_{\mu }$. The micron-scale organizations observed in all three binary mixtures were double emulsions with core-shell architectures (Figure S5C). Line scans show that the core is enriched in SRSF1Δ213–247 and the shell is enriched in the other SRSF (Figures 5D-5F). Sub-micron-scale caps that become generators of the double emulsion structures were also visible (see arrows in Figures 5D-5F).

We used RALS to measure the threshold concentration $({\mathrm{c}}_{\text{th}\cdot })$ of SRSF1Δ213–247 for phase separation in the presence of different amounts of SRSF3, SRSF5, or SRSF7. The threshold concentration rises to ~1 μM at a 0.05:1 molar ratio of SRSF7-to-SRSF1Δ213–247, a 0.5:1 ratio of SRSF3-to-SRSF1Δ213–247, and a 4.5:1 ratio of SRSF5-to-SRSF1Δ213–247 (Figure 5G).

Investigations of chimeric proteins

IDRs such as A1-LCD drive macrophase separation.^18,79 The driving forces for macrophase separation can be enhanced by replacing all aromatic residues within this sequence with tyrosine (allY).⁷⁹ We used QF-DEEM to investigate structures of allY macrophases. We observed micron-scale droplets defined by interiors featuring nanoscale inhomogeneities and percolated hub-like entities⁸⁰ (Figure S5F). Injection of mRNA for mCherry-tagged full-length hnRNP-A1 into X. laevis oocytes showed that the expressed protein localizes to the dense fibrillar and granular components of nucleoli (Figure S5G), which show characteristics of macrophases.^81,82

Next, we designed two chimeric proteins, ch-SRSF1 and ch-hnRNP-A1. In ch-SRSF1, the IDR2 of SRSF1 was replaced with the 129-residue-long A1-LCD(Δhexa).^52,83 In ch-SRSF1, the repulsions of IDR2 are replaced by attractions, the repulsions involving IDR1 are diminished, and so are the attractions involving RRM1 and RRM2 (Figure 5H). Using RALS, we observed a sharp discontinuity at a concentration $({\mathrm{c}}_{\text{th}\cdot })$ of 1.5 ± 0.4 μM (Figure 5I). Above ${\mathrm{c}}_{\text{th}}$, there is a concentration-dependent increase in the abundance, fusion, and growth of micron-scale puncta (Figure 5J). Thus, ch-SRSF1 drives macrophase separation. Imaging disordered, fluid-like macrophases using QF-DEEM is challenging due to their fragility. However, in one field of view, we observed a micron-scale body (Figure S5H). The interiors within this one-off image of a micron-scale punctum show finer, nanoscale structures.

To generate ch-hnRNP-A1, we replaced the A1-LCD with IDR2Δ, which is IDR2 of SRSF1Δ213–247. This replacement enhances the attractions of the C-terminal IDR with RRM1 and RRM2 when compared with the parent protein (Figure 5K), and C-terminal attractions are replaced by repulsions. RALS measurements show a non-linear, albeit monotonic increase in scattering intensity. Confocal microscopy showed the presence of sparse sub-micron-scale assemblies. The ch-hnRNP-A1 chimera exhibits behaviors that are concordant with the weak segregation limit⁴⁴ whereby only small, size-limited clusters are formed.

We also injected mRNAs of mCherry-tagged ch-SRSF1 and ch-hnRNP-A1 into X. laevis oocytes. ch-SRSF1 formed a range of de novo micron-scale bodies that are distinct from nuclear speckles or Cajal bodies (Figure 5N), whereas ch-hnRNP-A1 co-localized to Cajal bodies (Figure 5O). Thus, while ch-SRSF1 forms bodies with a heterogeneous distribution of sizes consistent with macrophase separation, ch-hnRNP-A1 forms submicron-scale clusters that are size-limited (Figure 5P).

TDP-43 forms microphases

TDP-43 colocalizes with sites of active splicing in neurons.⁸⁴ In addition to two RRMs, TDP-43 comprises a folded N-terminal oligomerization domain (NTD),^85,86 and a prion-like C-terminal LCD⁸⁷ (Figure 6A). The two RRMs (Figures 6B, S6A, and S6B) are different from those of RRMs in SRSFs and hnRNP-A1. The computed inter-domain interaction coefficients (Figure 6C) show that prominent attractions are homotypic, involving the LCD and NTD, and heterotypic, involving the LCD and each of the other domains (RRM1, RRM2, and NTD). All other interactions are either repulsive (RRM2-RRM2, RRM2-NTD) or nearly ideal (RRM1-RRM1, RRM1-NTD, RRM1-RRM2). Thus, TDP-43 has the architecture of an “inverse bolaamphiphile,” similar to Matrin-3, which forms spherical and wormlike microphases.⁵⁰ Flexible bolaamphiphiles comprise hydrophilic ends connected by hydrophobic chains.⁸⁸ An inverse bolaamphiphile has the opposite orientation.

Figure 6. TDP-43 forms microphases and double emulsions in mixtures.

(A) Sequence architecture of TDP-43.

(B) Space-filling model of TDP-43. Color coding is identical to that of space-filling models in Figure 1.

(C) Computed ${\Delta }_{\mathrm{X}\mathrm{Y}}$ coefficients for TDP-43.

(D) Representative RALS trace of TDP-43. Results are shown from a single representative measurement (solid circles) and spread over three technical replicates across different protein preps (pale envelope). c_μ value is shown as mean ± SD.

(E) QF-DEEM images 0.7 μM TDP-43.

(F) Histogram of diameters from analysis of 1,326 microphases (mean ($\overline{\mathrm{d}}$) ± SD and median ($\tilde{\mathrm{d}}$)).

(G) Results from TRFQ measurements performed over a range of TDP-43 concentrations.

(H) Histogram of R_g,μ values of microphases comprising 12 TDP-43 molecules.

(I) Profiles of g_D(r).

(J) Representative snapshot from simulations showing organization of 12 TDP-43 molecules within a microphase.

(K) Representative fluorescence microscopy image of a double emulsion formed in a binary mixture of 1 μM TDP-43 and 1 μM SRSF1Δ213–247. Rightmost panel shows the line scan of normalized fluorescence intensities in the TDP-43 (cyan) and SRSF1Δ213–247 (green) channels. Magenta arrows point to spherical caps.

See also Figure S6.

RALS measurements yielded a ${\mathrm{c}}_{\mu }$ of 0.18 ± 0.08 μM for TDP-43 (Figure 6D). The rate constants determined from pyrene fluorescence lifetime measurements decreased with increasing TDP-43 concentration until ${\mathrm{c}}_{\mu }$, beyond which the values of the rate constants plateaued at ~0.93 ± 0.10 ns⁻ (Figure S6C). QF-DEEM imaging at 0.7 μM TDP-43 showed biconcave structures (Figure 6E) with mean and median diameters of 33.7 ± 9.5 nm and 33.7 nm, respectively (Figure 6F). Each TDP-43 microphase accommodates ~13 TDP-43 molecules per microphase as inferred from TRFQ experiments (Figure 6G). Note that previous reports suggested that TDP-43 forms oligomers comprising ~12 molecules.⁸⁹ The packing fraction of TDP-43 in the biconcave structures is 0.07.

Atomistic simulations using 12 TDP-43 molecules that were restrained to be within an envelope of radius 15 nm showed assemblies with a mean ${\mathrm{R}}_{\mathrm{g},\mu }$ of 8.3 ± 0.2 nm (Figure 6H). The radial distribution profiles show accumulation of the NTD within the microphase interior (Figures 6I and 6J). This is followed by RRM2, RRM1, and the LCD. At intermediate length scales, between 6 and 8 nm, the density of the LCD becomes the highest, followed by the NTD, and the two RRMs (Video S5). Confocal imaging showed weak clustering at concentrations above ${\mathrm{c}}_{\mu }$ (Figure S6D).

In binary mixtures with SRSFΔ213–247, all domains of TDP-43 make attractive interactions with IDR1 and IDR2 of SRSFΔ213–247 (Figure S6E). Interactions between RRM2 of TDP-43 and RRM1 of SRSFΔ213–247 are repulsive, as are interactions between RRM2 of SRSFΔ213–247 and both the NTD and RRM2 of TDP-43 (Figure S6E). The attractions enable adhesive interactions at interfaces between microphases. However, the domains cannot mix and form homogeneous structures because of the repulsions. Accordingly, the interface expands, and the spreading effect leads to core-and-shell-like double emulsions with TDP-43 predominantly localizing to the shell (Figure 6K).

Effects of MALAT1 on microphases

MALAT1 is a key lncRNA that regulates splicing activity within nuclear speckles.^45,90 It is upregulated in cancer cells^91-93 where it accumulates in the nucleus and localizes to nuclear speckles.^27,90,94 MALAT1 serves as a hub that interacts with splicing factors, such as U1 small nuclear ribonucleoprotein particle (snRNP)⁹⁵ and transcriptionally active chromatin.⁹⁶ Human MALAT1 is ~7,000 nt long,⁹⁴ and undergoes further processing to produce a 6,900 nt-long transcript and a tRNA-like small RNA of 61 nt.⁹⁷ MALAT1 is predicted to comprise 200 evolutionarily conserved helices, including a 3′ triple helix,⁹⁸ pseudoknots, tetraloops, internal loops, and long-range intramolecular interactions.⁹⁹

We produced MALAT1 using in vitro transcription and generated 3′-labeled MALAT1 with Alexa Fluor 647 for fluorescence assays. Using fluorescence correlation spectroscopy (FCS), the hydrodynamic radius $({\mathrm{R}}_{\mathrm{h}})$ of MALAT1 was found to be 6.8 ± 0.7 nm (Figure S7A). The reported ${\mathrm{R}}_{\mathrm{h}}$ of $BunVL$, a single-stranded RNA of similar length, is ~9.2 nm in solutions where the ionic strength is more than two times lower than what we use in our measurements.¹⁰⁰ Our ${\mathrm{R}}_{\mathrm{h}}$ measurements are consistent with studies showing that compaction is a universal property of mRNAs and lncRNAs.¹⁰¹

MALAT1 can bind to proteins as dispersed monomers or within microphases. In the polyphasic linkage formalism,^102-104 if MALAT1 is the ligand, and SRSF1 or SRSF1Δ213–247 is the macromolecule, then the threshold concentration for microphase separation in the presence of the ligand $({\mathrm{c}}_{\mu ,\mathrm{L}})$ can be written as: ${\mathrm{c}}_{\mu ,\mathrm{L}}={\mathrm{c}}_{\mu ,0}({\mathrm{P}}_{\mu }∕{\mathrm{P}}_0)$. Here, ${\mathrm{c}}_{\mu ,0}$ is measured in the absence of ligand, ${\mathrm{P}}_{\mu }$ is the binding polynomial of the ligand to macromolecules in the microphase, and ${\mathrm{P}}_0$ is the binding polynomial outside the microphase. If ${\mathrm{P}}_{\mu }={\mathrm{P}}_0$, then ${\mathrm{c}}_{\mu ,\mathrm{L}}={\mathrm{c}}_{\mu ,0}$; if ${\mathrm{P}}_{\mu }<{\mathrm{P}}_0$, then the affinity of the ligand is higher for the macromolecule in the microphase, and ${\mathrm{c}}_{\mu ,\mathrm{L}}<{\mathrm{c}}_{\mu ,0}$; finally, if ${\mathrm{P}}_{\mu }>{\mathrm{P}}_0$, then the affinity of the ligand is higher for the macromolecule in the dispersed phase, and ${\mathrm{c}}_{\mu ,\mathrm{L}}>{\mathrm{c}}_{\mu ,0}$.

We measured ${\mathrm{c}}_{\mu ,\mathrm{L}}$ for SRSF1 and SRSF1Δ213–247 as a function of MALAT1 concentration. Increasing amounts of MALAT1 lead to a lowering of ${\mathrm{c}}_{\mu }$ (Figure 7A). Thus, MALAT1 binds preferentially to SRSF1 and SRSF1Δ213–247 microphases. Changes to ${\mathrm{c}}_{\mu }$ saturate above a 5:1 molar ratio of MALAT1(in nucleotides)-to-protein. The effects of MALAT1 on lowering the ${\mathrm{c}}_{\mu }$ values are similar for SRSF1 and SRSF1Δ213–247. Thus, the region spanning residues 213–247 does not contribute to preferential binding of MALAT1 to SRSF1 microphases. The number of SRSF1 molecules per microphase $({\mathrm{n}}_{\mu })$ decreased from ~24 in the absence of MALAT1 to ~12 in the presence of MALAT1 (Figure 7B). Likewise, for SRSF1Δ213–247, ${\mathrm{n}}_{\mu }$ decreased from ~50 in the absence of MALAT1 to ~12 in the presence of MALAT1. Thus, upon preferential binding to microphases, MALAT1 enables the release of either ~12 SRSF1 molecules or ~37 SRSFΔ213–247 molecules. This release of positively charged proteins provides an entropic driving force that enhances microphase separation. The release of protein molecules also creates room for the incorporation of MALAT1 into microphases.

Figure 7. MALAT1 binds specifically and preferentially to SRSF1 microphases.

(A) In the presence of MALAT1, the c_μ of SRSF1 and SRSF1Δ213–247 decreases in a dose-dependent manner.

(B) TRFQ experiments, performed over a range of SRSF1 and SRSFΔ213–247 concentrations at a molar ratio of SRSF1-to-MALAT1 at 1-to-5 × 10 ⁻⁴, yielded an estimate of ~12 protein molecules per microphase (mean ± SD for slope and n_μ).

(C) MALAT1 colocalizes with sub-micron-scale clusters of SRSF1 (top row) and SRSF1Δ213–247 (bottom row).

(D) QF-DEEM images of SRSF1Δ213–247-MALAT1 microphases.

(E) Histograms collected across 1,510 MALAT1 bound microphases of SRSF1Δ213–247 show their diameters in nm.

(F) Comparative impact of MALAT1 versus MALAT1^scr on the cμ values of SRSF1Δ213–247.

(G) MALAT1^scr colocalizes with clusters of microphases formed by SRSF1Δ213–247.

(H) MALAT1 does not affect c_μ of TDP-43 at a molar ratio of TDP43-to-MALAT1 at 1-to-5 × 10⁻⁴, but at high RNA-to-protein ratios, it binds preferentially to the dilute phase.

(I) QF-DEEM images of TDP-43 + MALAT1 microphases.

(J) Histograms of microphase sizes in nm of 818 images of TDP-43 in the presence of MALAT1. Mean ($\overline{\mathrm{d}}$) ± SD and median ($\tilde{\mathrm{d}}$).

See also Figure S7.

Confocal images show colocalization of MALAT1 with the clusters of microphases formed by SRSF1 and SRSF1Δ213–247 (Figure 7C). In the presence of MALAT1, the QF-DEEM imaging shows spheroidal SRSF1Δ213–247 microphases (Figure 7D). From the size distributions (Figure 7E), the mean and median values are 36.4 ± 12.0 nm and 34.0 nm, respectively. Thus, preferential binding of MALAT1 remodels SRSF1Δ213–247 microphases.

MALAT1 interacts with SRSF1 through hepta-nucleotide AGAAGAA motifs.¹⁰⁵ There are five such motifs within a 500-nt stretch. Using computations, we generated a distribution of 10⁴ sequences by randomly scrambling the RNA sequence within the 500-nt stretch. Regions that are 5′ and 3′ to this stretch were fixed to be those of wild-type MALAT1. Within the ensemble of scrambled sequences, the occurrence of GAA, AGAA, and AAG sub-motifs within MALAT1 is highly non-random (Figure S7B). This is suggestive of a selection bias for these motifs. To test the specificity of MALAT1 effects, we selected a scrambled sequence, MALAT1^scr. In MALAT1^scr, the probabilities of finding GAA, AGAA, and AAG motifs were significantly lower than for the wild-type MALAT1. The nucleotide composition and length of the sequence of the MALAT1^scr are identical to those of the wild-type. Across a range of MALAT1^scr concentrations, excepting large molar excesses, MALAT1^scr has a negligible effect on the measured ${\mathrm{c}}_{\mu }$ of SRSF1Δ213–247 (Figure 7F). However, confocal imaging shows that MALAT1^scr colocalizes with SRSF1Δ213–247 clusters (Figure 7G). Together, these data show that, beyond colocalization, cognate binding sites need to be present in MALAT1 to enable preferential binding.

To further assess the binding specificity, we measured the impact of MALAT1 on the microphase separation of TDP-43 (Figure 7H). The ${\mathrm{c}}_{\mu }$ of TDP-43 is not affected for RNA-to-protein ratios that are below 10-to-1. However, in the presence of more than a10-fold excess of MALAT1, the ${\mathrm{c}}_{\mu }$ of TDP-43 shifts upward. Thus, excess MALAT1 destabilizes microphase separation of TDP-43 (Figure 7H). This is reminiscent of reentrant phase behaviors that result from excesses of RNA versus protein.^106,107 QF-DEEM showed that the structures formed by TDP-43 were preserved in the presence of MALAT1 (Figure 7I). This contrasts with the remodeling of biconcave structures formed by SRSF1Δ213–247. The effect of MALAT1 is evident in the increased sizes of the TDP-43 microphases, with the mean and median sizes being 43.6 ± 11.7 nm and 42.9 nm, respectively.

DISCUSSION

SRSFs and TDP-43 are functional copolymers of ordered and disordered domains (FRODOs) that undergo microphase separation. For individual FRODOs in solution, the structures observed in vitro constitute a well-defined hierarchy. Individual proteins that are 4–6 nm in diameter form microphases that are 25–40 nm in diameter. The sizes of microphases are stabilized by the interplay of intermolecular attractions and repulsions. As concentrations increase beyond ${\mathrm{c}}_{\mu }$, microphases associate to form sub-micron-scale clusters that are ~300–400 nm in diameter. The sizes of clusters are stabilized by the interplay of inter-microphase short-range attractions and long-range repulsions.⁶⁴ Microphase separation of FRODOs involves a combination of site-specific interactions of folded domains, distributed multivalent interactions among IDRs and surfaces of folded domains, and overall solubility considerations.

In mixtures of microphase-separating proteins, there is an interplay between two types of inter-domain attractions and repulsions: those involving proteins of the same kind and those involving different proteins. Adhesive interactions at microphase interfaces generate spherical caps with both proteins. However, repulsions drive spreading over mixing, thus giving rise to double emulsion structures through polymer-polymer phase separation^36,40,78 that maximizes inter-protein and inter-block attractions while minimizing repulsions. Thus, we conjecture that the demixing phenomenon reported recently for TDP-43 in stress granules¹⁰⁸ might be a generic consequence of interactions at the level of microphases. The broader implication would be that stress granules are also emulsions of microphases.

Chromatin is also a multi-block copolymer, and organization into distinct territories might be a manifestation of microphase separation.^37,39 The binding of transcription factors, such as TFEB, to target genomic loci lowers the interfacial tensions of condensates, thereby suppressing coarsening and driving binding-induced microphase separation.^109,110 Bridging-induced phase separation also generates size-limited DNA-cohesin clusters.¹¹¹ Binding of different RRM-containing proteins to the 5′, middle, and 3′ regions of NEAT1 (nuclear enriched abundant transcript 1) has been proposed to generate regions of different hydrophobicity along the lncRNA, thereby converting NEAT1 into a triblock copolymer.¹¹² An intriguing, albeit uncorroborated possibility is that the binding of RRM-containing proteins drives an unfolding of the lncRNA. If so, the intrinsic microphase separation of paraspeckle proteins would be destabilized, and microphases would be realized by the generation of a block copolymer that forms on a very different length scale.¹¹² This scenario contrasts with our observations showing that preferential and specific binding of MALAT1 to SRSF1 microphases releases the proteins from pre-formed microphases. Within nucleoli, Yao et al. showed that fibrillarin (FBL) forms distinct territories within dense fibrillar components (DFCs).¹¹³ These FBL-dense regions are likely to be clusters of microphases. As a preliminary assessment, we computed interaction matrices for the DFC proteins FBL (Figure S7C) and nucleolin (NCL) (Figure S7D). The computations suggest that both proteins are bolaamphiphiles,⁸⁸ thereby making a case for their microphase separation within nucleoli.

What is the relevance of microphase separation in nuclear speckles? On average, there are likely to be ~10⁶ copies of SRSF1 in a cell.⁷⁷ The estimated nuclear concentration would be 1–10 μM, which is above the ${\mathrm{c}}_{\mu }$ value that we measured in vitro. Thus, microphase separation seems unavoidable within nuclei. Also, the effective molarity $({\mathrm{c}}_{\text{eff}})$ within SRSF1 microphases, estimated as ${\mathrm{c}}_{\text{eff}}=(\mathrm{p}\phi {\mathrm{n}}_{\mu }∕{\mathrm{v}}_{\mu })\times (1∕{\mathrm{N}}_{\mathrm{A}})$, would be ~9.6 mM. Here, $p=0.29$ is the packing parameter, $\phi =0.005$ is the volume fraction of each monomer in the microphase, ${\mathrm{n}}_{\mu }=24$ is the number of molecules, ${\mathrm{v}}_{\mu }=6.86\times {10}^{-24}{\mathrm{m}}^3$ is the volume of each microphase, and ${\mathrm{N}}_{\mathrm{A}}=6.02\times {10}^{23}$ is Avogadro’s number. This high local concentration will help increase the splicing efficiency by driving efficient spliceosome assembly.^6,15,114 MALAT1 binds specifically and preferentially to SRSF1 microphases, enhancing the driving forces for microphase separation. This likely contributes to the interplay between regulation of splicing and RNA polymerase II (RNA Pol II) mediated transcription. Indeed, Guo et al.¹¹⁵ showed that phosphorylation of the C-terminal domain (CTD) switches RNA Pol II association from Mediator condensates to nuclear speckles. Specifically, the phosphorylated CTD (pCTD) is incorporated into condensates formed by splicing factors such as SRSF1 and SRSF2.¹¹⁵

Song et al.⁴⁵ showed that the pCTD binds preferentially to SRSF1 microphases. In ternary mixtures with MALAT1 and SRSF1, pCTD displaces MALAT1 from SRSF1 microphases.⁴⁵ Thus, the recruitment of SRSF1 microphases by MALAT1, which stabilizes the microphases while releasing a fraction of the SRSF1 molecules from microphases, can lead to the assembly of microphases near transcription sites of speckle-associated genes. The stronger preferential binding of the pCTD likely displaces MALAT1 from SRSF1 microphases.

Limitations of the study

Currently, we lack knowledge about how post-translational modifications such as serine phosphorylation ^90,116,117 impact the intrinsic and MALAT1-mediated microphase separation of SRSFs. We also lack direct probes of microphase separation in live cells that have the resolution of QF-DEEM to investigate microphase functions in cells, although recent methods show promise.^118,119

STAR★METHODS

METHOD DETAILS

Generation of DNA constructs

pGEMT-MALAT1 was cloned by amplifying the MALAT1 cDNA from pCMV-MALAT1 and ligating it into the pGEM-T easy vector (Promega, USA) according to the manufacturer’s protocol. pGEMT-MALAT1scr was generated by Gibson cloning with a synthesized gene block (Genscript, USA) for the 500-nucleotide region.¹²⁸ For Gibson assembly reactions, the insert DNA and plasmid DNA were incubated at a 3:1 molar ratio with the latter fixed at 50 ng and were combined with 2X Gibson Master Mix (final volume: 20 μL). The mixture was incubated at 50°C for one hour then cooled on ice. This mixture containing the assembled plasmid was used to transform into DH5a E. coli cells. DNA concentration was measured by absorbance at 260 nm. pET28-SRSF1Δ213-247 was cloned by subcloning SRSF1 cDNA from YFP-SRSF1 into the pET28 vector by restriction digestion and ligation. pET28-SRSF1 was cloned by sub-cloning SRSF1 from pSMT3-SRSF1. pET28-SRSF3 (Uniprot: P84103), pET28-SRSF5 (Uniprot: Q13243), pET28-SRSF7 (Uniprot: Q16629), pET28-ch-SRSF1, and pET28-ch-hnRNPA1 were generated by gene synthesis (GenScript, USA). pET28-SRSF1ΔIDR2 was generated by site-directed mutagenesis. The listed pET28 constructs were transformed into the DH5α and BL21-CodonPlus (DE3)-RIPL cells.

Protein purification and labeling

SRSF1, SRSF1Δ213-247, SRSF1ΔIDR2, SRSF3, SRSF5, SRSF7, ch-SRSF1, and ch-hnRNPA1

The His₆-SUMO N-terminally tagged constructs were expressed in BL21-CodonPlus(DE3)-RIPL E. coli cells (New England Biolabs, USA) grown in LB broth (Millipore Sigma, USA), supplemented with kanamycin, spectinomycin, and chloramphenicol. Cultures were incubated at 37°C in a shaker incubator at 220 rpm until OD₆₀₀ of 0.4 was reached. Cultures were then induced with 0.35 mM Isopropyl b-D-1-thiogalactopyranoside (IPTG) and incubated for 6-8 additional hours at 23°C. Cells were harvested by centrifugation at 4000 rpm for 30 minutes, washed in Lysis Buffer (10% glycerol, 20 mM HEPES (pH 8.0), 1.25 M NaCl), and stored as pellets at −80°C.

For lysis, the cell pellet was gently resuspended in 5 mL of cold Lysis Buffer per gram of pellet with one cOmplete protease inhibitor tablet (Roche, Switzerland). 1 mg/mL of lysozyme (Roche, Switzerland) was added and incubated at 4°C for 1 hour with gentle rocking. The solution was then sonicated (Branson 550 with an L102C horn attachment) using five sets of the following 20-round cycle: 1 second on / 2 second off at 30% power. The solution was then centrifuged for 30 minutes at 10,000 rpm.

Supernatant was incubated with equilibrated 4 mL HisPur Ni-NTA resin (Thermofisher Scientific, USA) at 4°C for one hour. The solution was then poured into a manual column, and the flow-through fractions were collected. The resin was then washed with 30 mL lysis buffer and then eluted by Elution buffer (10% glycerol, 20 mM HEPES (pH 8.0), 1.25 M NaCl, 300 mM Imidazole) in 2 mL fractions. The fractions were confirmed by SDS-PAGE. Fractions containing His₆-SUMO-SRSF1 were pooled and the SUMO tags were cleaved during an overnight dialysis in the presence of 1:50 molar ratio of His₆-tagged Ulp1 protease to His₆-SUMO-SRSF in Cleavage Buffer (10% glycerol, 20 mM HEPES (pH 8.0), 1.25 M NaCl, 1 mM DTT (dithiothreitol)). The dialysate was incubated with 4 mL of Ni-NTA resin equilibrated with lysis buffer. Flow-through was collected, and the resin was eluted with elution buffer in 5 mL fractions. The presence of SRSF was confirmed in the flow-through and His₆-SUMO and His₆-Ulp1 in the elution fractions. Flow-through fractions were pooled and diluted 25-fold in dilution buffer (10% glycerol, 20 mM HEPES (pH 8.0)) to lower the total NaCl concentration. This solution was further purified using a HiTrap Heparin HP 5 mL column (Cytiva, USA) on a continuous gradient with Buffer A (10% glycerol, 20 mM HEPES (pH 8.0), 50 mM NaCl), and Buffer B (10% glycerol, 20 mM HEPES (pH 8.0), 1.25 M NaCl) by FPLC (ÄKTA Pure, Cytiva, USA). Peak fractions were confirmed by SDS-PAGE, pooled, and concentrated in Amicon Ultra 3 kDa MWCO (molecular weight cut-off) concentrator columns (Millipore Sigma, USA). Concentrated protein was aliquoted into 100 μL, flash frozen in liquid N₂, and stored at −80°C. Concentration of protein was determined by absorbance at 280 nm. Predicted extinction coefficients¹²⁹ for the constructs used to determine protein concentration are as follows: 27850 M⁻¹cm⁻¹ (SRSF1), 24870 M⁻¹cm⁻¹ (SRSF1Δ213-247), 23505 M⁻¹cm⁻¹ (SRSF1ΔIDR2), 16960 M⁻¹cm⁻¹ (SRSF3), 12950 M⁻¹cm⁻¹ (SRSF5), 21890 M⁻¹cm⁻¹ (SRSF7), 33935 M⁻¹cm⁻¹ (ch-SRSF1), 13075 M⁻¹cm⁻¹ (ch-hnRNPA1). Table S1 includes data for the ratio of absorbance measured at 260 nm to 280 nm. Values close to 0.6 confirm the presence of minimal nucleic acid contaminants (Figure S7F).

SDS-PAGE was performed after each purification step for all constructs. Gel images of the purified products are shown in Figure S7G.

TDP-43

The His₆-SUMO N-terminally tagged construct was expressed in BL21(DE3) E. coli cells (New England Biolabs, USA) from the SUMO-TDP43 plasmid (pET28b(+)-SUMO-TDP43). grown in LB broth (Millipore Sigma, USA), supplemented with kanamycin. Cultures were incubated at 37°C in a shaker incubator at 220 rpm until OD₆₀₀ of 0.4 was reached. Cultures were then induced with 0.35 mM Isopropyl b-D-1-thiogalactopyranoside (IPTG) for 6-8 additional hours at 23°C. Cells were harvested by centrifugation at 4000 rpm for 30 minutes, washed with Lysis Buffer (10% glycerol, 20 mM HEPES (pH 8.0), 1.25 M NaCl), and stored as pellets at −80°C.

For lysis, the cell pellet was gently resuspended in 5 mL of cold Lysis Buffer per gram of pellet with one cOmplete protease inhibitor tablet (Roche, Switzerland). 1 mg/mL of lysozyme (Roche, Switzerland) and 10 μg/mL DNase I were added and incubated at 4°C for 1 hour with gentle rocking. The solution was then sonicated (Branson 550 with an L102C horn attachment) for 2 minutes using 30 seconds on/off cycle at 50% power. The solution was centrifuged for 30 minutes at 10,000 rpm.

Supernatant was incubated with equilibrated 4 mL HisPur Ni-NTA resin (Thermofisher Scientific, USA) at 4°C for one hour. Then, the solution was poured into a manual column, and flow-through fractions were collected. The resin was then washed with 30 mL Wash buffer (10% glycerol, 20 mM HEPES, (pH 8.0), 1.25 M NaCl, 50 mM Imidazole) and the eluted by Elution buffer (10% glycerol, 20 mM HEPES, (pH 8.0), 1.25 M NaCl, 300 mM Imidazole) in 2 mL fractions. The fractions were confirmed by SDS-PAGE. Fractions containing His₆-SUMO-TDP-43 were pooled and cleaved of His₆-SUMO tags during an overnight dialysis in the presence of 1:50 molar ratio of His₆-tagged Ulp1 protease to His₆-SUMO-TDP-43 in Cleavage Buffer (10% glycerol, 20 mM HEPES (pH 8.0), 1.25 M NaCl, 1 mM DTT (dithiothreitol)). The dialysate was incubated with 4 mL of Ni-NTA resin equilibrated with lysis buffer. Flow-through was collected, and the resin was eluted with elution buffer in 5 mL fractions. The presence of TDP-43 was confirmed in the flow-through and His₆-SUMO and His₆-Ulp1 in the elution fractions. Flow-through fractions were pooled and diluted 25-fold in Dilution buffer (10% glycerol, 20 mM HEPES (pH 8.0)) to lower the total NaCl concentration. This solution was further purified using a HiTrap Heparin HP 5 mL column (Cytiva, USA) on a continuous gradient with Buffer A (10% glycerol, 20 mM HEPES (pH 8.0), 50 mM NaCl), and Buffer B (10% glycerol, 20 mM HEPES (pH 8.0), 1.25 M NaCl) by FPLC (ÄKTA Pure, Cytiva, USA). Peak fractions were confirmed by SDS-PAGE, pooled, and concentrated in Amicon Ultra 3 kDa MWCO concentrator columns (Millipore Sigma, USA). Concentrated protein was aliquoted into 100 μL, flash frozen in liquid N₂, and stored at −80°C. Concentration of protein was determined by absorbance at 280 nm. Predicted extinction coefficients¹²⁹ for the constructs used to determine protein concentration are as follows: TDP-43 is 44920 M⁻¹cm⁻¹.

Labeling proteins with fluorophores

For covalent conjugation of all protein constructs, including the SRSFs, TDP-43, ch-SRSF1, and ch-hnRNPA1, with NHS (h-hydroxysuccinimide ester) AlexaFluor 488 or AlexaFluor 405 (Thermofisher Scientific, USA), 1 mg of protein was dialyzed into Labeling Buffer (10% v/v glycerol, 20 mM HEPES (pH 8.3), 1.25 M NaCl) using Pierce Microdialysis Plates (ThermoFisher Scientific, USA). The dialysate was combined with NHS-Alexa Fluor dye (dissolved in DMSO) at a dye:protein molar ratio of 4:1. This mixture was incubated under gentle rocking for one hour at 4°C in the absence of light. The mixture was then dialyzed extensively into experimental buffer (10% (v/v) glycerol, 20 mM HEPES (pH 7.4), 50 mM KCl, 5 mM MgCl₂) to remove unincorporated dye. Concentrations of protein and AlexaFluor 488 were determined by absorbance measurements (${\epsilon }_{280}=27850{\mathrm{M}}^{\text{-}1}{\text{cm}}^{\text{-}1}$, ${\epsilon }_{495}=73,000{\mathrm{M}}^{\text{-}1}{\text{cm}}^{\text{-}1}$). Labeling efficiency was determined as the molar ratio of dye to protein and ranged between 0.4–0.6. Labeled protein was aliquoted into single-use volumes (2 μL), flash frozen in liquid N₂, and stored at −80°C.

In vitro transcription and 3’ RNA labelling

For all RNA reagents, plasmids containing the desired RNA transcript were linearized via incubation with 5% v/v restriction enzymes (MALAT1 and MALAT1scr – SalI) and 10% v/v CutSmart Buffer (New England Biolabs, USA) at 37°C for 4 hours. Linearization was confirmed by gel electrophoresis of DNA samples in a 1% agarose gel with non-digested DNA as controls. Linearized plasmids were purified using a PCR clean-up kit (IBI Scientific, USA) as per the manufacturer’s manual. In vitro transcription was performed for each transcript using the mMESSAGE mMACHINE Transcription Kit (ThermoFisher, USA) with either SP6 for the mRNAs that were microinjected into X. laevis oocytes or Hi-T7 polymerase for the in vitro experiments performed using MALAT1 and MALAT1^scr . The purity and molecular weight of RNA were confirmed using gel electrophoresis (Figure 7G). Confirmed RNA transcripts were aliquoted into single use volumes (2 μL), flash frozen with liquid N₂, and stored at −80°C. MALAT1 and MALAT1scr transcripts were labelled with AlexaFluor 647 hydrazide as described previously.⁴⁹

QUANTIFICATION AND STATISTICAL ANALYSIS

In each of the sections that follow, we describe the methods obtaining and analyzing data from in vitro measurements based on right angle light scattering, time-resolved fluorescence quenching, single molecule imaging and tracking, fluorescence correlation spectroscopy, microfluidic resistive pulse sensing, and analysis of QF-DEEM images. The extraction of quantitative information from each of these measurements and the statistical analyses are customized for each method and are described in the paragraphs that describe the methods used to deploy the measurements and analyze the data. In general, for each measurement, our data are reported for at least three biological replicates and at least three technical replicates. In most cases, our data exceed these minimal thresholds. The sections that follow also describe the methods and analysis of data from measurements performed by imaging GVs of Xenopus oocytes following injections and incubation of mRNAs. The sections below also describe the methods used for computing various parameters and the details of the atomistic simulations. Each of these sections have self-contained details regarding the methods of quantification of the information provided in the main text and supplemental figures, and the statistical analyses used in the quantifications.

Right angle light scattering (RALS)

All constructs were dialyzed against the experimental buffer (10% v/v glycerol, 20 mM HEPES (pH 7.4), 50 mM KCl, 5 mM MgCl₂) using Pierce Microdialysis Plates (ThermoFisher Scientific, USA). The intensity of static right-angle light scattering at 320 nm (I) was measured at each titration point using PTI Quantamaster (Horiba Scientific, Japan). The intersection of linear fits to I versus [protein] (log-log plot) identified discontinuity points in the concentration dependence of the scattering intensity, which is indicative of an assembly formation and/or phase transition. For each experiment, we report results using at least two biological and three technical replicates. Note that the RALS data collected at a series of different concentrations are proof of reversibility of the microphases because the intensities are recovered by increasing or decreasing the concentration.

Optimal linear fits were determined using a modified jackknife approach.¹³⁰ Each dataset was subjected to two independent series of linear fits, with one series of fits for the low concentration arm (LCA) and the second series of fits for the high concentration arm (HCA). For each arm, the fitting procedure was initiated with a linear fit to the four lowest or highest concentration data points, respectively, and the root mean square error (RMSE) of each fit was recorded. Following these initial fits, the next highest or lowest concentration data point was added to the respective LCA or HCA data set, the expanded data sets were re-fit, and the new RMSEs were recorded. This process was continued, expanding the fitted dataset by one data point at a time, until all the points in the full dataset were included in both the LCA and HCA linear fits. The intersections of each pair of LCA and HCA best fits were recorded, and the average intersection point for all best fits, for all trials at a given protein concentration, were determined. This is the value reported as the threshold concentration, ${\mathrm{c}}_{\text{th}\cdot }$, or as ${\mathrm{c}}_{\mu }$ for microphase separation. The presence of a discontinuity in light scattering intensity, measured over a range of protein concentrations, indicates the formation of a distinct phase defined by large assemblies.^130-132

Fluorescence lifetime measurements

Samples were prepared with noted concentration of protein and a constant concentration of 20 nM pyrene. The 340 nm LED was used as an excitation source with a 365 nm long pass filter in an EasyLife X Lifetime Fluorescence Spectrometer (Horiba Scientific, Japan). The fluorescence decay curves were recorded using the EasyLife X software and fitted to the following equation in Prism 10 software (GraphPad, USA): $I(t)=I_0{e}^{k_{decay}t}$ where $I(t)$ and $I_0$ are fluorescence intensities at time $t$ and at $t=0$, respectively. Pyrene fluorescence lifetime was quantified as the inverse of the rate constant $k_{\text{decay}}$:

Time-resolved Fluorescence Quenching (TRFQ)

Samples were prepared with noted concentration of protein and a constant concentration of 20 nM pyrene and 0.4 μM cetylpyridinium chloride. The 340 nm LED was used as an excitation source with a 365 nm long pass filter in an EasyLife X Lifetime Fluorescence Spectrometer (Horiba Scientific, Japan).

The fluorescence decay curves were recorded using the EasyLife X software and fitted by Equation 1 in Prism 10 software (GraphPad, USA).

$$I(t)=I(0)\text{exp}\left(-\frac{t}{{\tau }_0}-R\left[1-\text{exp}\left(-\frac{t}{{\tau }_Q}\right)\right]\right)$$ (Equation 1)

$I(t)$ and $I(0)$ are fluorescence intensities at time $t$ and zero, respectively. ${\tau }_0$, ${\tau }_{\mathrm{Q}}$, and $R$ are three parameters representing the results of the nonlinear fit. ${\tau }_0$ is pyrene lifetime in the absence of quencher. ${\tau }_{\mathrm{Q}}$ is pyrene lifetime in the presence of quencher. In the presence of quencher, the fluorescence decay is described by the Equation 1 with $k_{\mathrm{Q}}=1∕{\tau }_{\mathrm{Q}}$ being the rate constant for inter microphase quenching. The extracted parameter $R$, which is equal to $c_{\mathrm{Q}}∕c_{\mu }$, is the ratio of the quencher concentration $c_{\mathrm{Q}}$ to the concentration of microphases, $c_{\mu }$. The number of molecules per microphase, ${\mathrm{n}}_{\mu }$, can be calculated using the following relationship (Equation 2):

$$n_{\mu }=R\frac{[\text{protein}]-c_{\mu }}{c_Q}$$ (Equation 2)

The pyrene concentration was chosen to be less than 0.05 × quencher concentration which avoids the formation of pyrene excimers. Individual solutions at a given protein concentration were prepared by dilution from the solution with the largest protein concentration. For each solution condition, the fluorescence decay curves were recorded and analyzed to determine ${\mathrm{n}}_{\mu }$. The plots of the number of molecules per microphase, ${\mathrm{n}}_{\mu }$, were generated over the protein concentrations and linearly fit to obtain the y-intercept as an estimated number of molecules per microphase at ${\mathrm{c}}_{\mu }$. The reliability index, $\Phi$, of the estimate was determined as a product of the standard error of the mean of the estimate of ${\mathrm{n}}_{\mu }$, extracted from the linear regression, and the slope of the line of best fit. Small values of $\Phi$ signal high reliability in the estimation of ${\mathrm{n}}_{\mu }$, whereas large values of $\Phi$ imply high unreliability. For constructs with high unreliability, ${\mathrm{n}}_{\mu }$ values were estimated assuming the identical packing fraction as SRSF1 of 0.29. The packing fraction was calculated using Equation 3:

$$p=n_{\mu }\left(\frac{R_g^3}{R_{\mu }^3}\right)$$ (Equation 3)

In Equation 3, ${\mathrm{R}}_{\mathrm{g}}$ is the average of radius of gyration of a single molecule obtained from atomistic simulations. ${\mathrm{R}}_{\mu }$ is the median radius of microphases, which is $\overline{d}∕2$, the median diameter of microphases divided by 2, obtained from the size distributions of microphases determined from QF-DEEM images, and ${\mathrm{n}}_{\mu }$ is the number of molecules within a microphase.

Fluorescence Confocal Microscopy

Confocal microscopy imaging performed on the Eclipse Ti2 microscope (Nikon, Japan) with a Yokogawa CSU X1 disk module and a LunF laser launch equipped with 405 nm, 488 nm, 561 nm, and 647 nm lasers, using a 60X, 1.4 NA Apo oil immersion objective (Nikon, Japan) and an Orca Flash 4.0 CMOS camera (Hamamatsu, Japan). All images were captured at room temperature using NIS-Elements software (Nikon, Japan) and saved as 16-bit ‘.nd2’ files. Images within a data set were taken with identical imaging parameters ensuring that signal was not saturated. No averaging, binning, and projecting were used. All images of purified SRSF1 and RNA constructs shown are representative crops of one or a few entities (e.g., a condensate) where the brightness and contrast have been optimized. Line scans were obtained using ImageJ and plotted using Prism 10 (GraphPad, USA). Samples were prepared with 1:250 molar ratio of labeled-to-unlabeled protein and 1:50 labeled-to-unlabeled RNA and imaged in silicone wells mounted on a coverslip.

Single-molecule imaging of SRSF1 with PA-JF-549

We covalently labeled SRSF1 protein with photoactivatable Janelia Fluor^® 549 (PA-JF-549) NHS ester as described above for AlexaFluor labeling. For imaging, we mixed 1 μM of unlabeled SRSF1 with labeled SRSF1 at the molar ratio of 1:50, such that each microphase is labeled, on average, with an individual fluorophore based on the determined number of SRSF1 molecules per microphase. A reduced dilution of 1:100 was applied for larger cluster imaging to mitigate the background signal. For each sample, 15 μL of solution was placed in a silicon well (Grace Bio-Labs, USA) adhered to a cover slip.

Single-molecule microscopy was performed with a home-built wide-field epi-fluorescence microscope, equipped with an oil-immersion objective (UPLSAPO100XO, NA 1.4, Olympus, Japan), a dichroic mirror (Semrock Di01-R488/561), and an emission filter (Semrock FF01-593/46). Activation (405 nm) and excitation (561 nm) lasers shared the same optical path to illuminate the sample or activate photoactivatable fluorophore, respectively. The PA-JF-549 was activated under constant low intensity (~50 W·m⁻²) activation wavelength. Autofluorescence and crosstalk were not observed with 405 nm activation in the emission channel. A high peak intensity (~4088 W·m⁻²) of 561 nm laser was used to excite activated PA-JF-549. Single-molecule images and trajectories were acquired at 50 Hz for 20,000 frames for an average of ~300 photons per molecule in each 20 ms frame. Imaging was performed in triplicate from three independent preparations.

Single-molecule image analysis: localization and tracking

Single-molecule blinking and movements were observed within the acquired image stacks. Direct projections were created by summarizing all pixel information over time on finding either the maximum value (maximum projection) or the standard deviation (std projection) of the temporal intensity profile of each pixel. The maximum projection indicates the spatial distribution and emission intensity of PA-JF-549, and the std projection indicates the switching rate between the “fluorescent” and “dark” states. Both projections were used to evaluate and optimize activation/excitation control and labeled / unlabeled protein mixture ratio for observation of single molecule behaviors.

The localization and tracking of individual molecules were performed as described⁸⁰ by applying the Robust Statistical Estimation algorithm (RoSE).¹³³ All position estimates were subsequently classified and grouped into molecular displacements. This localization algorithm enables accurate position measurements of molecules on each frame, further improving the temporal resolution. Single molecule localization depends on the sequential detection and position estimation of individual molecules in each frame. A maximum likelihood estimator was used to fit a standard point spread function to the images and measure the sub-pixel lateral positions and brightness of each molecule. The localization algorithm has ~13 nm precision in measuring lateral position of each molecule. This theoretical limit on precision is evaluated from the average of the measured fluorescence signal (300 photons) and background (9 photons).

The dim emitters were filtered out and their positions were grouped into emitter displacements by connecting estimates between two consecutive frames. Two estimates on separate frames were classified as the same emitter if: 1) they were the closest pair of estimates between the frames, and 2) their distance, d, was within the confinement radius of 200 nm. All distances were computed in Euclidean form. We classified the speed of movement of each emitter by applying a threshold to the displacement, with displacements of d < 50 nm categorized as slow and those above 50 nm as fast. This length scale was chosen because it provided the clearest segregation of the displacement of the emitters. The 2D histograms of fast- and slow-moving emitters were generated by counting the number of fast and slow molecules originating from each binned 25-by-25 nm² region on the spatial coordinates.

Excess variance of displacements across all bins within a mask were calculated as described in Equation 4:

$$\text{Excess variance}=\frac{\text{Variance of bin values}}{\text{Mean of bin values}}-e$$ (Equation 4)

Fluorescence Correlation Spectroscopy (FCS)

Two samples were prepared in 8-well plates (0.17 ± 0.005 mm thickness) in the experimental buffer, assumed to have the same viscosity as water: (1) a mixture of 5 nM unlabeled MALAT1 and 0.05 nM labeled MALAT1 and (2) 0.05 nM free dye. MALAT1 was labeled with AlexaFluor 647 hydrazide. The free AlexaFluor 647 hydrazide was used to determine the size of the beam waist, and the measured diffusion time, extracted from the autocorrelation function, was used to estimate the hydrodynamic radius (Rh) of MALAT1 per the Stokes-Einstein Equation. The diffusion coefficient of AlexaFluor 647 hydrazide with the instrument setup used here is 3.3 x 10⁻¹⁰ m²s⁻¹. All data were collected on a Confocor II LSM system (Carl Zeiss-Evotec, Jena, Germany) with a 40x water-immersion objective. Samples were excited at 633 nm, and emission was collected with a 650 nm cut-on long pass filter. 10 replicates of 10-second fluorescence intensity autocorrelation function traces were collected and analyzed with Zeiss Confocor II FCS software. To extract the diffusion time, the autocorrelation function $\mathrm{G}(\mathrm{t})$ was fit to Equation 5 assuming a single-component system:

$$G(t)=1+\frac{1}{N}\frac{1}{\left(1+\frac{t}{{\tau }_D}\right){\left(1+\frac{t}{{\tau }_D}\frac{r_0^2}{z_0^2}\right)}^{\frac{1}{2}}}$$ (Equation 5)

where ${\tau }_D$ is the diffusion time, $N$ is the number of fluorophores in the confocal detection volume, $\mathrm{t}$ is the correlation time, ${\mathrm{r}}_0$ is the resolution in the x-y plane, and $z_0$ is the resolution in the z plane.

The Stokes-Einstein equation, as written out in Equations 6 and 7, was used to estimate the particle hydrodynamic radius ${\mathrm{R}}_{\mathrm{h}}$:

$$R_h=\frac{k_{B}T{\tau }_D}{6\pi \eta D_{ref}{\tau }_{D,ref}}$$ (Equation 6)

$${\tau }_D=\frac{r_0^2}{4D}$$ (Equation 7)

Here, $k_B=1.38\times {10}^{-20}{\mathrm{m}}^2{\mathrm{gs}}^{-2}{\mathrm{K}}^{-1}$ is the Boltzmann’s constant, $T=296\mathrm{K}$ is the temperature in Kelvin, $\eta =0.94{\mathrm{gm}}^{-1}{\mathrm{s}}^{-1}$ is the viscosity of the buffer, assumed to be the same as that of water, and $D$ is the diffusion coefficient.

Surface Electrostatic Potentials (SEPs)

The RRM structures were extracted from full-length structures of each protein from AlphaFold^120,134,135 (SRSF1: AF-Q07955-F1-v4; SRSF3: AF-P84103-F1-v4; SRSF5: AF-Q13243-F1-v4; SRSF7: AF-Q16629-F1-v4; TDP-43: AF-Q13148-F1-v4, hnRNP-A1: AF-P09651-F1-v4). Site-specific surface electrostatic potentials (SEPs) of RRMs were determined using “Coulombic Surface Coloring” on the solvent-excluded molecular surfaces generated by UCSF Chimera.¹²⁵ The following default parameters were used in generating the molecular surfaces by a rolling probe sphere: probe radius of 1.4 Å to approximately represent a water molecule, vertex density of 2.0 vertices per Ų, and line width of 1.0 pixel in mesh surface representation. The SEP values from each vertex of the surfaces were recorded and plotted as scatterplots.

Hydrophobicity

Hydrophobicity values were determined along the RRM sequences using a window size of 9 according to the Kyte-Doolittle scale¹³⁶ using ProtParam of ExPASy.¹²⁶ The hydrophobicity values were plotted as scatterplots in Figures S1E, S4B, and S6B.

Generation of biconcave disc geometry

The biconcave discs were generated as a continuous surface from a quartic equation (Equation 5) using the Mathematica package (Wolfram, USA) script from Kuchel et al.,¹²⁷ while varying the diameter of the biconcave disc (d), thickness of the dent at the center (b), and maximum height at the rim (h) as noted in each figure caption. The Mathematica notebook is included in supplemental information. In brief, the quartic equation used was:

$${(x^2+y^2+z^2)}^2+P(x^2+y^2)+Q{z}^2+R=0$$ (Equation 8)

where $\mathrm{P}$ and $\mathrm{Q}$ are coefficients, $\mathrm{R}$ is the constant term, and $\mathrm{x}$, $\mathrm{y}$, and $\mathrm{z}$ are coordinates. Here, $\mathrm{P}$, $\mathrm{Q}$, and $\mathrm{R}$, are defined as follows:

$$\begin{gathered} P=-\frac{d^2}{2}+\frac{h^2}{2}\left(\frac{d^2}{b^2}-1\right)-\frac{h^2}{2}\left(\frac{d^2}{b^2}-1\right){\left(1-\frac{b^2}{h^2}\right)}^{\frac{1}{2}} \\ Q=P\frac{d^2}{b^2}+\frac{b^2}{4}\left(\frac{d^4}{b^4}-1\right) \\ R=-P\frac{d^2}{4}-\frac{d^4}{16} \end{gathered}$$

Microfluidics Resistive Pulse Sensing

Microfluidics Resistive Pulse Sensing (MRPS) measurements were performed on an nCS1 instrument (Spectradyne, USA). This technique measures the change in electrical resistance as particles of different size pass through a constriction of defined resistance.⁶⁵ The abundance and the size of particles are determined based on the transit times and the magnitudes of the electrical signals, respectively.

Samples were prepared at indicated concentrations in experimental buffer (10 % v/v glycerol, 20 mM HEPES, pH 7.4, 5 mM MgCl₂, and 50 mM KCl). 5 μL of each sample was measured using C-2000 cartridges, with a measurement range of 250–2000 nm. At least three acquisitions were collected and combined for analysis, with each acquisition being collected over 10 minutes. Control experiments were performed with buffer, in the absence of protein, where no significant abundance of particles was detected. The particle size distributions from the triplicate experiments were averaged and the resulting curve was smoothed using the non-linear least squares locally weighted scatterplot smoothing (LOWESS) method as implemented in Prism (GraphPad, Boston, MA, USA). The algorithm in Prism uses the number of neighboring points, input by the user, to create a smooth curve. We set the number of neighboring points to ten. This was deemed reasonable since each MRPS measurement generates 200-300 independent points.

The particle size distributions were fitted to Lorentzian distributions to accommodate the presence of low abundances away from the center of the peak using nonlinear least squares regression, set to converge when five iterations in a row change the sum-of-squares by less than 0.0001% (Equation 9):

$$L(d)=\frac{A}{1+\frac{{(d-c_d)}^2}{w}}$$ (Equation 9)

where $A$ is the particle abundance at the center of the distribution, $c_d$ is the center of the distribution, $w$ is the width of the distribution, and $d$ is the particle diameter. The $c_d$, which is located at the peak of the Lorentzian, were plotted as the diameter of the clusters (Figure S4L).

Average fitted parameters for Figure 3B are as follows. 0.5 μM: A = 5167 ± 321, c_d = 358.9 ± 20.4, w = 74.9 ± 14.0; 1.0 μM: A = 8011 ± 347, c_d = 345.7 ± 25.8, w = 83.5 ± 14.1; 1.5 μM: A = 7098 ± 269, c_d = 386.1 ± 26.3, w = 79.1 ± 10.6; 2.5 μM: A = 12099 ± 327, c_d = 378.4 ± 43.4, w = 84.6 ± 11.8.

Isolation and image analysis of fluorescent protein-positive bodies from Xenopus laevis oocytes

Harvested oocytes were manually disrupted from clustering, then subjected to collagenase digestion (gentle rocking) for 2 hours at 18°C. Oocytes were stored in ND96 Buffer (5 mM HEPES, 96 mM NaCl, 2 mM KCl, 1.8 mM CaCl₂, 1 mM MgCl₂) that was filter-sterilized and supplemented with 2.5 mM sodium pyruvate (ThermoFisher, USA) and 1X penicillin-streptomycin (MilliporeSigma, USA) at 18°C. Healthy stage VI oocytes were selected and injected using freshly pulled microneedles (Drummond, USA) and an injector (Nanoinject II, Drummond, USA). A total of 23 nL of mRNA in ddH2O (typically at a total mass of 20 ng) encoding for noted constructs were injected into each oocyte. Injected oocytes were stored individually in wells in a 48-well polystyrene SterileTissue Culture Plates (Fisher Scientific, USA) supplemented with ND-96 buffer at 18°C for at least 18 hours to allow expression and localization of exogenous protein. Immediately prior to imaging, the germinal vesicles were manually dissected in mineral oil and mounted on a glass slide with 6 μL of mineral oil. A 22 x 22mm glass coverslip was overlaid onto the sample and then immediately imaged. This procedure was carried out for at least two separate harvests of oocytes and similar results were obtained across these biological replicates. All images shown are of a single Z-slice that is representative and is at least 2 mm above the coverslip; brightness and contrast have been optimized.

The diameters of fluorescent protein-positive bodies expressing the noted mRNA-encoded constructs were determined using Fiji. The recorded.nd2 files were split into each fluorescent and brightfield channels for analysis. The fluorescent channel images was converted to 8-bit from 16-bit and turned into a binary image by adjusting the intensity threshold to 99%. Using the “Analyze particle” and the “Particle manager” tools, the mask for the fluorescent protein-positive bodies was used to obtain the area. The area was assumed to be a perfect circle and used to determine the diameter of the SRSF1-positive bodies as $\mathrm{d}=2^{\ast }\mathrm{sqrt}(\text{area}∕\pi )$. Apparent bodies smaller than 32 pixels (0.10835 mm/pixel) were omitted due to falling below resolution limits. At least 100 fluorescent protein-positive bodies from at least three independent GVs were analyzed for each construct.

Quick-freeze deep etch electron microscopy (QF-DEEM)

Samples for freeze-fracture were prepared in the experimental buffer (10 % v/v glycerol, 20 mM HEPES, pH 7.4, 5 mM MgCl₂, and 50 mM KCl) at a concentration above the ${\mathrm{c}}_{\mu }$, except for allY A1-LCD. The allY A1-LCD was dialyzed against 20 mM HEPES (pH 7.0), 1 M NaCl and diluted to a final concentration at 20 mM HEPES (pH 7.0), 150 mM NaCl final at approximately 20°C. The saturation concentration $({\mathrm{c}}_{\mathrm{sat}})$ for allY A1-LCD under the final solution condition is 50 μM.⁷⁹ 3 μL drops of sample were added onto a glass chip supported on boiled and fixed (2% glutaraldehyde) egg white prior to rapidly pressing the sample onto a liquid helium-cooled copper block polished to a mirror finish. Frozen samples were transferred to liquid nitrogen for storage prior to replica generation.

Replicas of each sample were generated using a Leica EM ACE900 Freeze Fracture System. In brief, quick-frozen samples were transferred from liquid nitrogen to the ACE900 Freeze Fracture System. The surface was fractured at −115 to −120°C using a cold knife held at −185°C. Samples were then etched at −104°C for 3 minutes prior to shadowing with 4-5 nm Pt/C at an angle of 24°. The replica was supported with 8 nm of carbon deposited at an angle of 85° for a total of 12 nm thickness. Replicas were washed in 48 wt% HF to remove the glass support prior to washing several times with deionized water. Cleaned replica were transferred to copper grids for TEM imaging (JEOL JEM-1400 Plus) at 120 kV.

QF-DEEM Image Analysis

Identification of microphases from.tif files was performed as follows:

1. Particle (microphase) identification in a rectangular bounding box using a difference of Gaussians method

2. Fitting to ellipses within bounding boxes

3. Estimating the number of particles within a probe circle of a defined radius.

A Gaussian filter was applied on each image with a standard deviation of σ=1 and σ=3. Both operations smooth the image, which when subtracted, identifies boundaries (“bounding box”) of particles. Images which have even intensity produce consistent detections while images with uneven intensity may require different thresholds to be applied for sections of image. Particles within a defined range of area (e.g., 100–500 pixel² for 10,000x magnification) are selected for further processing. The detected particles within the bounding box are fitted to an ellipse and scaled by ${(ab)}^{1∕2}$, where $a$ is the semimajor and $b$ is the semi-minor axis, respectively, of the fitted ellipse. Then, a circular ansatz is applied to the fitted ellipses. Particles were excluded if: 1) the fitted ellipse eccentricity was greater than 0.6; or the area of the fitted ellipse covered less than 40% of the area of the bounding box. These two criteria in combination allowed us to achieve high sensitivity in detecting areas that were circular and confidently fitted within the bounding boxes.

The average number of particles found within a probe circle of diameter of 380 nm was determined using a sweep approach by generating a circular mask. If a particle was at the edge of the probe circle, it was still considered as detected within the circular mask. Particles are detected within the probe circle as it scans over the image at a 100-pixel interval. To ensure the entirety of the probe circle remains within the image, the scan was constrained to the region of an image that excludes 100 pixels along the edge of the image. A box plot of the number of particles found in all non-empty sweeps are recorded.

Fluorescence Recovery after Photobleaching (FRAP)

FRAP was performed on the LSM 980 with Airyscan 2 (Zeiss, Germany) using a 63X high NA oil immersion objective with 488 nm excitation laser at 0.2% laser power and electron gain of 650. All images were captured at room temperature using ZEN Connect. The image was colored with Fire Lookup Table using ImageJ.¹³⁷ The fluorescence intensity was plotted using Prism 10 (GraphPad, USA). Samples were prepared with 1:250 molar ratio of labeled-to-unlabeled AlexaFluor 488-SRSF1 and imaged in silicone wells mounted on a coverslip.

Atomistic simulations of individual molecules and pairs of molecules

All-atom Metropolis Monte Carlo (MC) simulations were performed using the ABSINTH implicit solvent model and forcefield paradigm as made available in the CAMPARI simulation package (http://campari.sourceforge.net).^60,61 The simulations utilized the abs_3.5_opls.prm parameter set in conjunction with optimized parameters for neutralizing and excess Na+ and Cl− ions.¹³⁸ Each simulation was performed using spherical droplets with a diameter of 200 Å and explicit ions to mimic a concentration of 1.7 mM NaCl. Ensembles corresponding to a simulation temperature of 340 K were used in the analysis reported in this work. This choice reflects recent calibrations, which show that better concordance with experimental data and improved sampling efficiency are achieved at the higher temperature.^18,139

For simulations involving the folded domains, we used the predicted structures from AlphaFold (Uniprot: Q07955 (SRSF1), P84103 (SRSF3), Q13243 (SRSF5), Q16629 (SRSF7), Q13148 (TDP-43), P09651-2 (hnRNP-A1)).^120,134 Conformations of the RRMs were initially equilibrated to ensure consistency of bond lengths and bond angles with the ABSINTH model, as was done in previous work.^140,141 For simulations involving the IDRs, a single simulation was performed at 20 K to build the IDRs. Then the end PDB structures from these simulations were used to generate ten distinct full-length starting structures to seed ten independent simulations, each initiated using a distinct random seed. The backbone flexibility of folded domains was limited by applying restraints on the starting structure,¹³⁹ whereas the sidechain degrees of freedom were unrestrained. To free the simulations of biases from the initial coordinates, a short, high-temperature (500 K) simulation comprising 2×10⁴ MC steps was performed. The final structures from each of the high-temperature simulations were then used as the input structures for ten independent simulations. The temperature for the production runs was set to be 340 K. Each simulation comprised a total of 4.0×10⁷ MC steps that combined the full spectrum of translational, rotational, pivot, local, and concerted moves.^61,139,142 Conformations pertaining to the first 5×10⁶ MC steps were discarded as equilibration. Simulation results were analyzed to extract quantities of interest including ${\mathrm{R}}_{\mathrm{g}}$ values using the MDTraj¹⁴³ and SOURSOP packages.¹²⁴

In simulations with pairs of domains designated as X and Y, the goal was to extract pairwise inter-domain interaction coefficients. These serve as proxies of normalized second virial coefficients. To achieve this, each simulation comprised two domains X and Y that were either a pair of RRMs, a pair of IDRs, or an RRM and an IDR. Simulations were performed in large spherical droplets. Each droplet is over an order of magnitude larger than the dimensions of an individual domain represents an ultra-dilute system for a pair of domains. In a finite unbiased simulation, most of the inter-domain distances sampled would correspond to dissociated molecules. To overcome this problem, we applied an inter-domain distance restraint, modeled using a harmonic potential. Pairs of alpha carbon atoms located close to the centroid of each molecule were restrained to be within a shallow harmonic well, defined by a spring constant of 10 kcal/mol-Å and an equilibrium distance of 75 Å. Note that no restraints were applied on any other inter-residue or inter-atomic distances. From the ensemble of equilibrated configurations for each of the distinct pairs of domains, extracted from ten independent simulations for each pair, we quantified statistics for inter-residue distances using MDTraj¹⁴³ and SOURSOP.¹²⁴ These statistics were used to compute the ensemble-averaged distances $\langle R_{\mathrm{X}\mathrm{Y}}^{(\mathrm{i}\mathrm{j})}\rangle$ between pairs of residues $i$ and $j$, where the former is located on domain X and the latter on domain Y. To calibrate the observed pattern of inter-domain, inter-residue distances against a suitable prior, we performed reference simulations to model the expectations for an ideal, non-interacting pair of the domains of interest. In this non-interacting prior, the domains freely sample the available simulation volume, with the presence of the weak restraint as described above. These reference simulations were performed using the inverse power potentials introduced by Tran et al.¹⁴⁴ In these simulations, the non-bonded interactions are such that all terms, excepting the Lennard-Jones repulsions were switched off. The exponent for the repulsive exponent was set to six, which corresponds to Maxwellian particles, and the $\epsilon$ was set to 0.0001. This choice was based on calibrations, which showed for IDRs, irrespective of the sequence, an exponent of six and $\epsilon$ of 0.0001, with all other non-bonded interactions switched off, the ensemble-averaged intramolecular, internal distance profiles showed fractal behavior whereby the distances between pairs of residues $i$ and $j\langle {\mathrm{R}}_{ij}\rangle \sim {\mid \mathrm{j}-\mathrm{i}\mid }^{0.5}$. This is the scaling expected for a Flory random coil.^140,141,145 Inter-domain, inter-residue distances extracted from simulations using pairs of domains based on the inverse-power potential with an exponent of six and $\epsilon$ of 0.0001 were used to extract ensemble-averaged distances denoted as $\langle R_{\mathrm{X}\mathrm{Y},\text{ideal}}^{(\mathrm{i}\mathrm{j})}\rangle$.

Generation of computed interaction matrices

For each pair of domains, X and Y, the computed sets of ensemble-averaged, inter-domain, inter-residue distances denoted as $\langle R_{XY}^{(ij)}\rangle$ and $\langle R_{\mathrm{X}\mathrm{Y},\text{ideal}}^{(\mathrm{i}\mathrm{j})}\rangle$, were used to compute excess interaction coefficients as follows:

$${\Delta }_{\text{XY}}^{(\text{ij})}=\frac{\langle R_{XY}^{(ij)}\rangle -\langle R_{\text{XY},\text{ideal}}^{(\text{ij})}\rangle }{\langle R_{\text{XY},\text{ideal}}^{(\text{ij})}\rangle }$$ (Equation 10)

The parameter defined in Equation 10 varies between −1 and +1. Next, we used the excess interaction coefficient for each pair of residues across the domains X and Y, as defined in Equation 10, to compute the excess interaction coefficient ${\Delta }_{\mathrm{X}\mathrm{Y}}$ using:

$${\Delta }_{\text{XY}}=\text{sgn}(\Delta )\frac{\sqrt{\sum _{i=1}^{N_{\mathrm{X}}}\sum _{j=1}^{N_{\mathrm{Y}}}{[{\Delta }_{\text{XY}}^{(\text{ij})}]}^2}}{\sqrt{N_X{N}_{\mathrm{Y}}}}$$ (Equation 11)

Here, ${\Delta }_{\mathrm{X}\mathrm{Y}}$ is the grand sum (defined as the sum over all the elements) of the upper or lower triangular matrix, where each element of the matrix is ${\Delta }_{\mathrm{X}\mathrm{Y}}^{(\mathrm{i}\mathrm{j})}$ as defined in Equation 10. The signum or sign function sgn is such that, the sgn(Δgn(Δ) is set t if Δ if Δ) is if Δ if Δ) is s if Δ if Δ) is set to −1n function sgn is such that,$N_{\mathrm{X}}$ and $N_{\mathrm{Y}}$ refer to the numbers of residues within domains X and Y, respectively. This procedure yields values for ${\Delta }_{\mathrm{X}\mathrm{Y}}$ that are normalized and signed such that $-1\le {\Delta }_{\mathrm{X}\mathrm{Y}}\le +1$.

Simulations of internal organization within microphases of SRSF1, SRSF1Δ213-247, and TDP-43

All-atom Markov Chain Metropolis Monte Carlo (MC) simulations were performed using the ABSINTH implicit solvent model and forcefield paradigm as made available in the CAMPARI simulation package (http://campari.sourceforge.net).^60,61 The simulations utilized the abs_3.5_opls.prm parameter set.¹³⁸ Each simulation was performed using spherical droplets with a diameter of 300 Å (for SRSF1 and TDP-43) or 400 Å (for SRSF1Δ213-247), which are consistent with QF-DEEM observations. Ensembles corresponding to a temperature of 340 K were used in the analysis reported in this work. Copies of a molecule of a resulting conformation from a single molecule simulation (see above) were placed on a cuboid grid using a 4x3x2 lattice for 24 molecules of SRSF1, a 4x4x3 lattice for 48 molecules of SRSF1Δ213-247, and a 4x3x1 lattice for 12 molecules of TDP-43. On each lattice, the spacing between the molecules was 90 Å. The cuboids were then placed in a spherical droplet as a boundary determined by the mean diameter of microphases observed in the QF-DEEM images, which were as follows: 300 Å for SRSF1, 400 Å for SRSF1Δ213-247, and 300 Å for TDP-43. To free the system of biases from the initial coordinates, ten replicates of a short simulation of 1×10⁴ MC steps, allowing only for 3D rigid body moves of the molecule, were performed to generate distinct starting structures for independent simulations. These rigid body moves enable a loss of memory of the original placements of molecules on a grid. The final structures from each of the rigid-body simulations were then used as the input structures for 50 independent simulations for TDP-43, 100 independent simulations for SRSF1, and 200 independent simulations for SRSF1Δ217-243. These numbers account for the total number of atoms in each system. Each simulation was initiated using a different random seed. For the folded domains, the backbone flexibility was limited by applying restraints on the starting structure, whereas the sidechain degrees of freedom were unrestrained. Each simulation comprised a total of 5.0×10⁵ MC steps that combined 10% of rigid body moves and 90% of the full spectrum of translational, rotational, pivot, local, and concerted moves. Conformations pertaining to the first 1×10³ MC steps were discarded as equilibration.

For each microphase, the radius of gyration of the system, ${\mathrm{R}}_{\mathrm{g},\mu }$, was determined from the eigenvalues (${\lambda }_{\mathrm{x}}$, ${\lambda }_{\mathrm{y}}$, ${\lambda }_{\mathrm{z}}$) of the gyration tensor as follows:

$$R_{g,\mu }=\sqrt{{\lambda }_x^2+{\lambda }_y^2+{\lambda }_z^2}$$ (Equation 12)

Generation of ${\mathbf{g}}_{\mathbf{D}}(\mathbf{r})$ profiles

The radial distribution profiles, ${\mathrm{g}}_{\mathrm{D}}(\mathrm{r})$, for each domain D were computed from atomistic simulations of microphases. For each domain D, we chose a fixed number of sites, with the number of sites scaling proportional to the ${\mathrm{R}}_{\mathrm{g}}$ of each domain: 10 each for SRSF1 RRMs, 15 for SRSF1 IDR1, 25 for SRSF1 IDR2, 15 for SRSF1 IDR2Δ213-247, and 5 for each domain in TDP-43. At a 5 Å interval from the center-of-mass of the microphase, the relative probability of observing a site of a specific domain D was determined by referencing to a prior of $4\pi p{\mathrm{r}}^2\Delta$ of obserp is the packing fraction as described in the time-resolved fluorescence quenching (TRFQ) section of STAR Methods. The packing fractions are 0.29 for SRSF1, 0.1 for SRSF1Δ213-247, and 0.07 for TDP-43. Then, the ${\mathrm{g}}_{\mathrm{D}}(\mathrm{r})$ quantifies the density of observing a domain D relative to the prior. The analysis was performed over the last 80 frames of each simulation and repeated for each replicate. Mean values across all frames and replicates are plotted as solid circles and lines. The samples were then pooled and bootstrap analysis across the 800 samples for each system was performed using sampling with replacement, whereby 10² distinct samples were generated by randomly choosing 200 data points from the pool of samples. Each bootstrap trial is shown as a lighter colored line on the same plot.

Sequence feature comparison by NARDINI+

The current version of NARDINI+^48,49 uses the protein sequences from the human proteome from Swissprot Homo sapiens database (May 2015, 20882 entries).¹⁴⁶ All IDR sequences within the SRSFs and TDP-43 were analyzed in terms of 90 sequence features that quantify to compositional biases and the binary patterning of pairs of residue types with respect to one another.^48,49,147

Non-random binary patterns were extracted using the NARDINI algorithm.⁴⁷ This helps quantify if pairs of residue types are non-randomly distributed along the linear sequence. The non-randomness is quantified in terms of z-scores, whereby deviations from z-scores of zero quantify the magnitude and type of non-randomness. A positive z-score implies non-random segregation of the pairs of residue types in question, whereas a negative z-score implies non-random mixing of the pairs of residue types. For this analysis, we followed the approach of Cohan et al.,⁴⁷ and grouped residues as follows: pol (S, T, N, Q, C, H), hyd (I, L, M, V), pos (K, R), neg (E, D), aro (F, W, Y), ala (A), pro (P), and gly (G).

Compositional biases were determined by analyzing 54 parameters, that include the fractions of each amino acid type (20 features); the fractions of positive, negative, polar, aliphatic, and aromatic residues (5 features); the ratios of Arg to Lys and Glu to Asp (2 features); the fraction of residues promoting chain expansion (FCE), the fraction of charged residues (FCR), the net charge per residue (NCPR), the fraction of disorder promoting residues, the hydrophobicity, the isoelectric point, and the polyproline-II propensity (7 features). Most compositional features were quantified using localCIDER.¹²³ The “RG Patch” feature is defined as the fraction of the sequence that is made up of patches of an Arginine-Glycine (RG) pair. Here, a patch had to have at least four occurrences of the given residue or two occurrences of RG and could not extend past two interruptions. For the various compositional features, z-scores for each IDR were calculated using the mean and standard deviation of the human IDRome. Z-scores greater than zero indicate an enrichment of the compositional feature compared to the human IDRome, and z-scores less than zero indicate a depletion of the compositional feature compared to the human IDRome.

Supplementary Material

Supplemental information can be found online at https://doi.org/10.1016/j.cell.2025.11.026.

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Chemicals and other resources
Kanamycin	Sigma	1355006
Spectinomycin	Fisher Scientific	ICN15899302
Chloramphenicol	Gold Biotechnology	C-105
LB broth (Lennox)	Sigma	L3022
NHS AlexaFluor 488	ThermoFisher Sci	A20000
NHS AlexaFluor 405	ThermoFisher Sci	A30100
AlexaFluor 647 Hydrazide	ThermoFisher Sci	A20502
PA JaneliaFluor 549, NHS ester	Tocris	6149
HEPES	Sigma	54457
NaCl	Sigma	S3014
KCl	Sigma	P9541
MgCl2	Sigma	M8266
Glycerol	Fisher Sci	G33-1
Dithiothreitol	Gold Biotechnology	DTT10
HisPur Ni-NTA resin	ThermoFisher Sci	88221
Imidazole	Sigma	I202
Pyrene	Sigma	82648
Cetylpyridinium chloride	Sigma	C0732
Critical Commercial Assays
Fisher Pure link quick plasmid miniprep kit	Fisher	K210011
T7 mMESSAGE mMACHINE Transcription Kit	Fisher	M0255A
Monarch RNA clean up	NEB	T2040L
Biological samples
Stage VI Xenopus laevis oocytes	Jonathan Silva Laboratory at Washington University in St. Louis	N/A
Deposited data
HeLa cell protein abundances	Nagaraj et al.77	N/A
AlphaFold structures	Jumper et al.120	SRSF1: AF-Q07955-F1-v4 TDP43: AF-Q13148-F1-v4 SRSF3: AF-P84103-F1-v4 SRSF5: AF-Q13243-F1-v4 SRSF7: AF-Q16629-F1-v4; SRSF9: AF-Q13242-F1-v4; hnRNPA1: AF-P09651-F1-v4
Bacterial strains
DH5α strain E. coli cells	New England Biolabs	C2987I
BL21-CodonPlus(DE3)-RIPL cells	New England Biolabs	230280
Recombinant DNA
pSMT3-SRSF1	Fargason et al.72	Addgene: 201056
pET28-SRSF1	This study	N/A
pET28-SRSF1Δ213-247	This study	N/A
pET28-SRSF1ΔIDR2	This study	N/A
pET28-SRSF3	This study	Uniprot: P84103 (Human)
pET28-SRSF5	This study	Uniprot: Q13243 (Human)
pET28-SRSF7	This study	Uniprot: Q16629 (Human)
pET28-ch-SRSF1	This study	N/A
pET28-ch-hnRNPA1	This study	N/A
pET28b(+)-SUMO-TDP43	Grese et al.121	N/A
pDEST17-A1LCD_deltahexa	Martin et al.18	Addgene: 169714
pGEMT-MALAT1scr	This study	N/A
Software and Algorithms
Prism	GraphPad	https://www.graphpad.com/
ZEN	Zeiss	https://www.zeiss.com/microscopy/us/products/software/zeiss-zen.html
FIJI	Schindelin et al.122	https://imagej.net/software/fiji/
NARDINI+	Cohan et al.47 and King et al.49	https://www.github.cco/mshinn23/nardini; https://github.com/kierstenruff/RUFF_KING_Grammars_of_IDRs_using_NARDINI-
localCIDER	Holehouse et al.123	http://pappulab.github.io/localCIDER/
SOURSOP	Lalmansingh et al.124	https://github.com/holehouse-lab/soursop
CAMPARI	Vitalis and Pappu60	https://campari.sourceforge.net/
UCSF Chimera	Pettersen et al.125	https://www.rbvi.ucsf.edu/chimera
ProtScale (Expasy)	Gasteiger et al.126	https://web.expasy.org/protscale/
Custom Python 3 scripts for DEEM analysis	This study	https://www.github.com/Pappulab/microphase-assemblies/
Custom Python 3 scripts for gD(r) analysis	This study	http://doi.org/10.5281/zenodo.17096527
Custom Python 3 scripts for interaction matrix construction	This study	http://doi.org/10.5281/zenodo.17096527
Custom Python 3 scripts for single-molecule trajectory analysis	This study	http://doi.org/10.5281/zenodo.17096527
Mathematica notebook	Kuchel et al.127	http://doi.org/10.5281/zenodo.17096527

Highlights.

• Proteins with RRMs and IDRs are defined by inter-domain attractions and repulsions

• Interplay of attractions and repulsions yields 25–40 nm protein-specific microphases

• MALAT1 binds preferentially to microphases by releasing positively charged proteins

• Microphases form clusters in cells and micron-scale core-shell emulsions in mixtures