NMR and EPR reveal a compaction of the RNA-binding protein FUS upon droplet formation

Abstract

Many RNA-binding proteins undergo liquid-liquid phase separation, which underlies the formation of membrane-less organelles, such as stress granules and P-bodies. Studies of the molecular mechanism of phase separation in vitro is hampered by coalescence and sedimentation of organelle-sized droplets that interact with glass surfaces. Here we demonstrate that liquid droplets of FUS, which is a protein found in cytoplasmic aggregates of ALS and FTD patients, can be stabilized in vitro using an agarose hydrogel that acts as cytoskeleton mimic. This allows their spectroscopic characterization by liquid phase NMR and electron paramagnetic resonance (EPR) spectroscopy. Protein signals from both dispersed and condensed phases can be observed simultaneously and their respective proportions be quantified precisely. Furthermore, the agarose hydrogel acts as a cryoprotectant during shock freezing which facilitates pulsed EPR measurements at cryogenic temperatures. Surprisingly, Double Electron-Electron Resonance (DEER) measurements revealed a compaction of FUS in the condensed phase.

Introduction

Cells enclose thousands of different biomolecules within their few picoliters of volume (1). Experiments reveal that medium-sized HeLa cells may contain as much as hundreds of milligrams of proteins per milliliter, which translates into a few billion protein molecules packed within the cell (2). While the role of chaperones in preventing aggregation of proteins is well understood, the formation of membraneless organelles by liquid-liquid phase separation (LLPS) has only recently emerged as an additional strategy. Such organelles exhibit distinct compartmentalization, composition and function (3, 4). Stress granules, for example, are cytoplasmic assemblies formed upon stress and have been proposed to form in order to protect proteins and RNA when they are not functionally required (5, 6).

The RNA binding protein Fused in Sarcoma (FUS) is a component of such stress granules (7, 8). The unstructured amino-terminal half of the protein is considered responsible for its phase separation behavior (9). This QGSY- and RGG-rich segment forms a network of intermolecular interactions favoring demixing from aqueous solutions. Cumulative evidence points to cation-π interactions between tyrosines and arginines as the main driving force (10, 11).

The separation behavior of FUS can be recapitulated in vitro by formation of liquid droplets (12, 13). Upon formation they gradually rigidify during a maturation or aging period (13–15). FUS mutations have been identified in human neurodegenerative diseases like Amyotrophic Lateral Sclerosis (ALS) and Frontotemporal Dementia (FTD) that could accelerate this aging, yet it remains unclear how this correlates with the diseases (16).

Recently, an increasing number of studies attempted to reveal the structure and dynamics of FUS in phase separated form (11, 14, 17). Considering the disordered nature of a large part of the protein, solution NMR and EPR are ideal methods to structurally characterize the droplet formation of FUS. Previously, solution NMR was used to study the QGSY-rich domain of FUS in a large phase-separated compartment that resulted from the complete fusion of many droplets; highly similar spectra in the dispersed solution state and in the phase separated form were observed arguing that the global structure of FUS does not change upon transition to the condensed phase (14). However, the probability for a molecule to encounter the phase boundary and thus change environment is by orders of magnitude greater in the droplets of a biphasic sample. Hence, droplets are expected to be a better model for cellular stress granules than a bulk phase is, as exchange of constituents between phases is crucial for function of membraneless organelles and related to their ageing (18).

Consequently, there is a need for methods to study the in vitro structure of droplets under conditions which are closer to the ones found in physiological stress granules. To this end, we developed a method that limits protein droplet size to a near physiological size by using a cytoskeleton-mimicking agarose hydrogel. This method permits NMR and EPR measurement where signals from both the dispersed and the condensed phase of the protein can be measured simultaneously. We can thus quantify the fraction of protein present in the droplets using NMR diffusion experiments and measure intramolecular distance distributions in both phases using EPR spectroscopy.

Results

FUS droplets coat the NMR tube

We are studying the first half of human FUS protein (1-267) which comprises the QGSY-rich segment and the first arginine-glycine repeat (RGG1). This protein construct (N-terminal Domain, NTD) was kept soluble throughout the purification with 6 M urea. Upon dilution of the denaturant, FUS NTD forms droplets as observed by light microscopy (Extended Data Fig. 1) and ¹H-¹⁵N HSQC spectra resulted in additional peaks, which are absent in the presence of urea (Fig. 1a). Murthy and coworkers recently also observed such an additional set of peaks using the QGSY fragment of FUS. (17). Strikingly, this new set of peaks appears to correspond to the peaks seen in the NMR spectrum collected in 6 M urea with systematic upfield shifts of 0.15 ppm and 0.4 ppm in the ¹H and ¹⁵N dimensions, respectively (Fig. 1b). Using NMR diffusion-ordered spectroscopy (DOSY), we assessed the translational diffusion of molecules originating from the two groups of peaks. The peaks corresponding to the soluble form of the protein (dispersed phase) disappeared in a DOSY experiment with a strong magnetic field gradient applied due to the protein’s rapid diffusion during the long (0.08 s) diffusion period (Fig. 1c). In contrast, the systematically “shifted” peaks were still observed, due to a much slower diffusion rate and indicating that they correspond to the droplet state, henceforth condensed phase of FUS. Similar slow diffusion was reported previously by Kay and coworkers in the bulk solutions of other phase-separated proteins (19).

Figure 1. Non-specific interaction of FUS NTD droplets with glass surfaces.

a) ¹H-¹⁵N HSQC comparison of 200 μM FUS NTD dispersed (black - (30mM HEPES, 200mM KCl, 3M urea, pH 7.3) and biphasic (red - 30mM HEPES, 200mM KCl, pH 7.3). b) Magnified regions of the spectra in (a) with the black spectrum copied and shifted (grey) to match the positions of additional signals in the red spectrum. c) ¹H 1D spectra of 200 μM dispersed and biphasic sample and the corresponding overlay of DOSY spectra. Increasing gradient strength in DOSY is visualized by a color gradient (magenta‘blue’red). d) ¹H 1D spectra of 500 μM FUS NTD droplets in 30mM HEPES, 200mM KCl, pH 7.3 (bottom), when liquid is transferred to another NMR tube (middle) and initial tube with fresh buffer added (top). For this experiment protein stock concentration of 5 mM was used. Asterisks and red bars denote fast and slow diffusion species, respectively. e) Comparison of physical properties of different sized particles. Stress granules share the same surface-area-to-volume ratio (SA:V) as similar sized liquid droplets (20). Increasing the radius of a continuous-phase sphere by four orders of magnitude (under the assumption of a 200 μl bulk phase) results in a four orders of magnitude lower surface-to-volume ratio.

We reasoned that the systematic chemical shift change of 0.4 ppm observed for the ¹⁵N signals might be due to a change in magnetic susceptibility caused by the tendency of the droplets to stick to the glass wall of the NMR tube, with the susceptibility effect on the amide protons being reduced due to their exchange with solvent protons. To test this, we carefully transferred the solution of the sample to a new NMR tube and then added fresh buffer without protein to the initial tube. ¹H 1D NMR spectra were then recorded for both samples immediately. The sample transferred to the new tube showed exclusively the peaks corresponding to the “fast” diffusing species of the soluble protein, whereas the initial tube with fresh buffer contained primarily the condensed phase corresponding to the less mobile species sticking to the glass of the NMR tube (Fig. 1d). This clearly demonstrated that the protein droplets adhered to the glass surface of the tube, resulting in a systematically shifted spectrum due to the different local magnetic field. The nearly identical chemical shifts of the two species after correction of the constant offset (Fig. 1b), suggest that their overall conformation is very similar in these two phases, namely random coil-like.

Quantification of the dispersed and condensed phases

The presence of these two sets of peaks reports on the two populations of protein molecules in two different phases. Yet, the slowly diffusing proteins residing either in one layer or in many droplets adhering to the glass wall of the NMR tube exist in a state that differs from the one found in biological systems such as stress granules. Moreover, we estimate that the liquid droplets are at least four orders of magnitude smaller than a bulk monophasic phase and thus have an approximately four orders of magnitude larger surface-area-to-volume ratio (Fig. 1e). We reasoned that, if we could provide an environment mimicking the eukaryotic cytoskeleton, we might obtain FUS in the form of small droplets and prevent their fusion into a large bulk phase as well as their non-specific interaction with the glass wall. It has previously been shown that nuclear actin network supports the nucleoli in X. laevis oocytes and its disruption leads to sedimentation and fusion of the nucleoli, supporting this hypothesis (21). A potentially good mimic of the cytoskeleton is a low percentage agarose hydrogel. Agarose hydrogels have been used previously in solution NMR and, in addition to being NMR silent, they are known not to influence the conformation of either disordered or globular proteins (22–24). We could indeed achieve the stabilization of FUS droplets using a 0.5% agarose hydrogel which is formed in the NMR tube using the following procedure: a stock of FUS NTD in 6 M urea was diluted into a buffer containing 0.5% agarose at 55°C and then transferred to an NMR tube while the sample was still liquid. Gradually, as the temperature equilibrated to room temperature, the gel formed and the sample became turbid. High turbidity in such an agarose hydrogel was still visible by visual inspection three hours after the preparation, while, in the absence of agarose, samples became clear due to sedimentation of the droplets (Fig. 2a). These turbidity changes were quantified by turbidity measurements using the buffer with agarose as reference (Fig. 2b). When the same process was repeated in the presence of 0.5% agarose and 3 M urea final concentration, although the gel still formed, the sample remained transparent (Fig. 2a).

Figure 2. FUS NTD droplets in 0.5% agarose hydrogel.

a) Light microscope images of stabilization in agarose hydrogel. Representative images from four independent experiments. Scale bar: 15 μm. b) Time progression of sample turbidity in the absence (black) and presence (red) of agarose hydrogel. Error bars indicate standard deviation and dots the mean from three independent experiments. c) Overlay of ¹H 1D DOSY spectra of dispersed (top, 3M urea) and biphasic (bottom) sample in agarose. Increasing gradient strength is indicated with the gradient magenta→blue→red. d) Integral of the spectral region shown in (c) as a function of the gradient strength. Black and red denote dispersed (3M urea) and biphasic sample in agarose, respectively. e) Integrals of ¹H 1D DOSY spectra as function of gradient strength at different FUS NTD concentrations. f) Final percentage of unattenuated DOSY signal (red dots) as reported in (e) and droplet fraction measured by absorbance concentration of centrifuged droplets (rectangles) and fluorescence (triangles) versus total protein concentration. Error bars indicate standard deviation and dots the mean from three independent experiments.

NMR spectra in agarose hydrogels of FUS NTD showed only the protein signals and not significantly broader lines (Extended Data Fig. 2a,b). Additionally, a change in agarose concentration did not result in a change in the fraction of slowly diffusing protein (Extended Data Fig. 2c). Hence, FUS NTD phase separation behavior is independent from the amount of agarose hydrogel at least between 0.25 and 1%. Importantly, NMR spectra of FUS droplets obtained in 0.5% agarose hydrogel did not show the second set of shifted signals. This supports our initial hypothesis that such a gel prevents significant interaction of most of the droplets with the glass surface of the tube (Fig. 2c). NMR diffusion experiments in the presence of 0.5% agarose hydrogel allow to distinguish between resonances belonging to either the condensed phase or the dispersed phase and confirm strongly reduced mobility of FUS in micrometer-sized droplets. Indeed, with increasing gradient strength, peaks belonging to the condensed phase persist while peaks belonging to the dispersed phase disappear (Fig. 2c). Integration of the Asn, Gln and Tyr side chain peaks of the biphasic sample at 120 μM total protein concentration reveals that 20% of FUS resides in the condensed phase (Fig. 2d). The same quantitative analysis using non-exchangeable aliphatic protons yields a similar percentage (Extended Data Fig. 3a,b). Notably, as protein concentration was increased, we could quantify that an increasing fraction of the protein population forms the condensed phase, 48% is reached at 200 μM (Fig. 2e). Similar values result from both concentration measurement in separated phases obtained by centrifugation and fluorescence analysis (Fig. 2f). Interestingly, we find FUS droplet size distribution determined by fluorescence microscopy to be independent of the total protein concentration (Extended Data Fig. 3c) and, accordingly, the fraction of protein in droplets determined by NMR to be proportional to turbidity (Extended Data Fig. 3d).

Diffusion time and protein fold impact quantification

Our quantification of protein in the condensed phase only assumes that no significant fraction of the protein molecules exchanges between phases during the time of the NMR experiment. Based on previously published FRAP data on the FUS QGSY-rich region, the time required for half of the molecules in a micrometer-sized droplet to exchange between the two phases is estimated to be 1.2 s (14), while the total duration of the diffusion block of our pulse sequence is 0.08 s. Thus, using a simple exponential model we calculated that from the protein molecules initially in a droplet only ~5% have exchanged during our diffusion experiment, suggesting a slim underestimation of the droplet fraction (Extended Data Fig. 4a). Note that half-life times published for most systems are considerably longer than 1 s (25), indicating that the fraction of molecules exchanging during the diffusion times accessed by conventional DOSY experiments is insignificant, so that this quantification method might be more generally applicable (Extended Data Fig. 4b).

We therefore applied our method to other phase-separating proteins by performing DOSY experiments in agarose-hydrogel containing droplets of the intrinsically disordered region of the DEAD box protein DDX4 and the first two folded RRM domains of the RNA binding proteins PTBP1 and SRSF1 without their unstructured regions (26–28). While DDX4 phase separation was triggered by urea dilution, PTBP1 and SRSF1 formed droplets only in the presence of polyU RNA and at low salt concentration, respectively. As in the case of FUS, droplets of all three proteins were stabilized in agarose hydrogel for hours after preparation (Extended Data Fig. 5). Interestingly, while we could quantify the slowly diffusing species of the unstructured DDX4 with similar success as for FUS NTD (Extended Data Fig. 6), for the two folded proteins, comparison of DOSY experiments in the presence and in the absence of droplets yielded identical curves (Extended Data Fig. 7a,b). This indicates that, although droplets are present in the sample, the folded proteins cannot be detected by our NMR measurements in the droplets due to their slow tumbling in the viscous environment (29). Notably, ¹H and ¹H-¹⁵N HSQC spectra of both PTBP1 and SRSF1 in the presence of droplets in agarose hydrogel show that the proteins tolerated the heat treatment and retained their fold (Extended Data Fig. 7c-f).

CW EPR spectroscopy is sensitive to phase separation

Next, we investigated if liquid-droplet stabilization in an agarose hydrogel would be compatible with characterization by continuous-wave (CW) and pulse EPR methods. EPR measurements require a paramagnetic center which was introduced via site-specific attachment of a methanethiosulfonate spin label (MTSL) to an engineered cysteine at position 10 in the NTD of FUS. In CW EPR spectroscopy the first derivative of the absorption spectrum is detected. The lineshape depends on the rotational mobility of the spin label and is a good reporter of local crowding in the vicinity of the labeled residue (30). For rapidly rotating nitroxides, three narrow lines of equal width and intensity are observed. As the rotational mobility decreases, line broadening becomes apparent, which is most pronounced for the high-field line.

The CW EPR linewidth of spin-labeled A10C (A10C R1) in the dispersed state is rather narrow (Fig. 3a). This can be explained by the NTD being disordered and exhibiting substantial backbone dynamics. The lineshape, and hence the rotational mobility, in both the dispersed and the condensed state is unaffected when repeating the same measurement with a 0.5% agarose hydrogel (Fig. 3a), confirming again that interaction of agarose with the NTD of FUS does not perceptibly change protein dynamics. Interestingly, the CW spectra of the biphasic sample reveals the presence of an additional, broader component absent in the dispersed sample (Fig. 3b). This component is indicative of a second population of protein molecules that experience slower molecular tumbling or higher local protein concentrations consistent with increased protein-protein interactions in the condensed phase. Lowering the temperature to 5°C provides a stronger linewidth contrast which points to an increase in phase separation and slower molecular tumbling of the droplet fraction. This is in agreement with previous work on the FUS QGSY-rich region (14), where the authors observed that decreasing the temperatures enhances LLPS. These observations establish CW EPR as a sensitive tool for studying phase separation.

Figure 3. EPR measurements on singly spin-labeled A10C.

a) CW spectra acquired with and without agarose hydrogel at ambient temperature. b) CW EPR lineshape as a function of temperature. c) Pulse EPR measurements at 50 K after incubation at either room temperature or on ice followed by shock-freezing detect differences in the liquid phases before shock-freezing.

Agarose as cryoprotectant to study phase separation

EPR can provide information on the distribution of distances between different molecular regions by introducing spin labels at multiple sites. This requires cryogenic temperatures and samples which are shock frozen in a glassy state, achieved by addition of a cryoprotectant. Unfortunately, addition of PEG or glycerol, in sufficient amounts to enable glass formation, hinders LLPS of FUS (Extended Data Fig. 8). However, agarose hydrogel and urea have cryoprotecting properties (31). We reasoned that this should permit us to capture these systems in a frozen glassy state if samples were frozen rapidly by immersion of the tube into precooled isopentane. To test this, we prepared the following samples: a sample of NTD in the dispersed state and two biphasic samples, one shock frozen directly after incubation at room temperature and one after incubation on ice. If the freezing process is fast compared to LLPS and potential restructuring of the droplets, spectra acquired at cryogenic temperatures should reflect the observations made at the incubation temperatures before freezing. Indeed, some of these observations are directly reflected in the lineshape of the echo-detected EPR spectra (EDEPR) acquired at 50 K (Fig. 3c). The high local concentration of spin-labeled protein in the condensed phase affects the spectral shape of the EDEPR spectrum and leads to a significant broadening for both biphasic samples.

More detailed information can be gained from Double Electron-Electron Resonance (DEER) experiments, which report on the distance distribution between two paramagnetic centers in a range of 15 Å to approximately 100 Å (32). The DEER signal, known as the primary DEER data, is the product of a background function and the form factor, with the latter encoding information on the distance distribution. The background is a concentration-dependent contribution of remote spins from other labeled molecules in the vicinity up to about 400 Å. With singly spin-labeled protein, we can qualitatively probe the inter-protein label pairs that contribute to this background. A slight decay is visible for A10C R1 in the monophasic dispersed state but no dipolar modulation is observed (Fig. 3c), as expected for well isolated singly spin-labeled proteins. In contrast, the primary DEER data of both biphasic samples exhibit a much faster decay, corresponding to higher local protein concentrations and thus stronger protein-protein interactions. Importantly, differences in the biphasic samples are apparent for shock-freezing from different initial temperatures, with the sample incubated on ice displaying the strongest background decay corresponding to the on average highest local protein concentration.

These measurements at cryogenic temperatures reflect the properties observed in the liquid state. Namely, proteins in the monophasic dispersed state are mainly isolated and liquid droplet formation leads to an increase in protein-protein interactions, which is most pronounced for the biphasic sample incubated on ice. These results confirm that the freezing process is indeed fast enough to arrest the state of these protein systems.

Conformational changes upon droplet formation

We can assess intra-protein distance distributions by using the double-cysteine mutants A10C S29C R1 and A105C G128C R1. To avoid inter-protein distance contributions, we employed spin dilution by mixing spin-labeled protein with unlabeled WT protein (Extended Data Fig. 9). Fitting the primary DEER data using either a non-parametric fit or a single Gaussian resulted in distance distributions of very similar shape, thus we assumed a Gaussian distribution model in our analysis (Extended Data Fig. 9).

In a first step, we measured the distance distribution in the monophasic dispersed state at 50 μM protein concentration with a spin dilution of 1:10 and using D₂O to prolong the phase memory time. The distance distribution of A10C S29C R1 (Fig. 4a) is somewhat narrower and has a shorter mean distance than that of A105C G128C R1 (Fig 4b), which can be rationalized by the smaller sequence separation between the cysteines. The large width of the distributions is in line with the disordered nature of this protein region.

Figure 4. DEER measurements on A10C S29C R1 and A105C G128C R1.

Primary DEER data (left) and corresponding distance distributions obtained via a one-Gaussian fit (right) of A10C S29C R1 and A105C G128C R1 in the dispersed monophasic state with 3 M urea (a,b). Primary 5-pulse DEER data (left) and distance distributions with each Gaussian weighted by their corresponding scaling factors (right) of biphasic A10C S29C R1 (c) and A105C G128C R1 (d) after incubation at ambient temperature and shock freezing. Primary DEER data (left) and corresponding distance distributions obtained via a one-Gaussian fit (right) of A10C S29C R1 (e) and A105C G128C R1 (f) in the bulk phase. The experimental data in a-f) are displayed as black dots, with the time-domain fit in black. The distance distribution of the dispersed fraction is displayed in red, and that of the condensed fraction in blue with the bootstrapped 95% confidence interval (using 1000 bootstrap samples) as a shaded area.

Next, we characterized the phase-separated state at 200 μM protein concentration and 3:80 spin dilution. We recorded a 5-pulse DEER experiment, a variation of the classical 4-pulse DEER experiment that allows measurement of longer traces at the cost of introducing a partial excitation artefact (33–34). We recorded two 5-pulse DEER traces with shifted positions of the artefact and fitted them globally. A strong background decay is visible in the primary DEER data (Fig. 4c-d), which would not be expected for fully isolated spin pairs in the concentration range used. We expect two distinct background contributions: a slow background decay stemming from protein in the dilute dispersed phase and a fast background decay originating from the protein-rich condensed phase. The interference of distance distributions poses a challenge due to the presence of two fractions with different background decays.

Using DeerLab, a software for the analysis of dipolar EPR spectroscopy data (35), we developed a full model of the primary DEER data, composed of a combination of two different DEER signals arising from the dispersed and condensed fractions (for more details see Methods, Extended Data Fig. 10 and Supplementary Fig. 1). This allowed for unraveling the distance distributions of the two fractions, using information on the distance distribution of the monophasic dispersed state in the same buffer to constrain the fit (Extended Data Fig. 9c-d). Our data reveal that protein in the dispersed phase of the biphasic sample is more compact with respect to the monophasic dispersed state, probably due to reduced urea concentration. Interestingly, further compaction is observed in the condensed phase for, both, A10C S29C R1 (Fig. 4c) and A105C G128C R1 (Fig. 4d). Although these protein regions do not correspond to the fibril core of the NTD (36), they still experience compaction upon LLPS. This observation is consistent with observation by Murthy et. al, who showed that the interactions in the condensed fraction of FUS QGSY-rich region are not localized to a single protein region (17).

Lastly, we characterized the bulk phase of both double mutants with 1:1000 spin dilution (Fig. 4e,f). We observe that, similarly to the biphasic measurements, these regions compact when entering a condensed state. However, direct comparison between the distance distributions in the bulk phase and the biphasic condensed state (Table 1) reveals that the mean distance and width differ slightly. These differences between the monophasic and the biphasic condensed states are too subtle to safely assign them to differences in mean protein conformation; they might be within uncertainty of our distance distribution fitting.

Table 1. Gaussian fit parameters of the distance distributions and their corresponding 95% confidence interval.

Summary of mean 〈r〉 and the full width at half maximum Γ of the distance distributions shown in Figure 4 and their bootstrapped 95% confidence intervals. The corresponding lower and upper bounds of the bootstrapped 95% confidence intervals are given in brackets next to the fitted values.

	A10C S29C				A105C G128C
	dispersed fraction		condensed fraction		dispersed fraction		condensed fraction
	〈r〉 [Å]	Γ [Å]	〈r〉 [Å]	Γ [Å]	〈r〉 [Å]	Γ [Å]	〈r〉 [Å]	Γ [Å]
dispersed, 3 M urea	36.7 [36.4, 36.9]	22.2 [21.4, 23.0]			39.7 [39.5, 39.9]	26.3 [25.6, 27.0]
biphasic	32.1 [31.5, 32.1]	24.4 [22.5, 24.4]	28.2 [28.0, 29.3]	19.6 [18.9, 22.0]	36.0 [35.7, 36.0]	26.4 [26.4, 27.6]	33.4 [32.6, 33.9]	23.7 [21.6, 25.7]
bulk			29.7 [29.7, 29.7]	25.6 [25.6, 25.6]			31.6 [31.6, 31.6]	27.6 [27.6, 27.6]

Discussion

NMR has been applied to obtain the structure of FUS fibrils (36–38) and to characterize the monophasic condensed phase (14, 17, 39–41). However, work on the monophasic condensed phase requires large amounts of material which limits application to a broad range of biomolecules.

Importantly, non-specific interactions with glass surfaces are a known issue in the field; therefore, passivation of surfaces is now established (42, 43). All previous NMR studies on protein liquid droplets did not eliminate these interactions, hence the additional NMR signals observed for the droplet form of FUS were misinterpreted as a distinct chemical environment of the condensed phase (44), while, as we showed here, they in fact correspond to the artificial droplet-glass interaction.

Moreover, it is unclear if the condensed phases in monophasic and biphasic samples age in the same way. The surface-area-to-volume ratio of the micrometer-sized droplets differs vastly from that of the bulk condensed phase. It is plausible to assume that frequent exchange of molecules between the condensed and dispersed phase affects ageing, if so, droplets may age differently from a bulk condensed phase. Whether the surface-to-volume ratio of the FUS condensed phase indeed affects the liquid-to-solid phase transition remains to be seen. Our method preserves the dynamic processes at the droplet interface, enables time-dependent studies of droplet ageing and thus to address this open question.

Further, the protein regions bracketed by the residues A10C and S29C or by the residues A105C and G128C adopt extended conformations in the dispersed phase and we observe for both biphasic and bulk samples that these protein regions compact upon entering the condensed phase. The slight differences that we observe between the distributions for the condensed states in biphasic and monophasic samples may not be significant. From a practical point of view, the much smaller protein quantity required here is a distinct advantage of using biphasic over monophasic samples. In our case, we used 100 times less material. This translates into a significant time and cost difference, especially when considering several protein mutants and isotope labeled protein for the characterization by NMR. Additionally, not all proteins can be concentrated to the high protein concentration (10 mM) and volume required to generate the bulk phase.

As reported previously, if the diffusion time used in the DOSY experiments is significantly shorter than the residence times of the two species in their respective phases, then a clearly biexponential decay is expected with two population-weighted resolved diffusion coefficients (39, 45). Indeed, we observe this for, both, unstructured FUS and DDX4. In contrast, for SRSF1 and PTBP1, both containing mainly folded regions, our approach is unable to detect the protein in the condensed phase. We hypothesize that fast backbone dynamics of the disordered domains is responsible for keeping lines sufficiently narrow for NMR observation of FUS and DDX4.

In conclusion, we have reported here a new method for studying the protein LLPS phenomena in vitro using NMR and EPR spectroscopies. With this approach, we can observe protein signals in the dispersed and condensed phases simultaneously. Using diffusion-based NMR experiments, we can quantify partitioning of intrinsically disordered proteins between the dispersed and condensed phase, while with DEER EPR measurements we can determine the intramolecular distance distribution in both forms. This methodology should be easily and generally applicable to most unstructured proteins undergoing LLPS. We expect it to stimulate further structural and dynamics studies of proteins and their complexes in the liquid-droplet form using NMR and EPR spectroscopies.

Methods

All reagents were purchased from Merck (Sigma-Aldrich) unless otherwise specified.

Protein expression and purification

E.coli strain BL21 (DE3) was used to express histidine- and Gb1-tagged FUS NTD (1-267), histidine- and Gb1-tagged DDX4 (1-233) and histidine- and Gb1-tagged SRSF1 (1-196). Two point mutations Y37S and Y72S in RRM1 of SRSF1 protein were used to improve solubility of the protein. Cells were grown either in M9 media containing ¹⁵N ammonium chloride (Cambridge Isotopes Laboratories) for uniformly ¹⁵N labelled samples or in LB (BD Difco) for non-isotopically labeled protein. Once OD600 reached 0.6, expression was induced with 200 μM IPTG (PanReac AppliChem) at 20 °C for 16 hours. Cell pellets of FUS NTD or DDX4 were harvested and suspended in suspension buffer (50 mM HEPES, 150 mM NaCl, pH 7.5) followed by sonication. While the soluble supernatant was discarded, the insoluble pellet was resuspended in urea buffer A (50 mM HEPES, 0.5 M NaCl, 8 M urea (PanReac AppliChem), pH 7.5). Sample was loaded onto a nickel column (QIAGEN) and eluted with imidazole buffer (50 mM HEPES, 150 mM NaCl, 1 M urea, 250 mM imidazole, pH 7.5). Subsequently, imidazole was removed and His-tag TEV cleaved simultaneously during two dialysis steps, first against buffer B (50 mM HEPES, 150 mM NaCl, 1 M urea, 5 mM β-mercaptoethanol (PanReac AppliChem), pH 7.5) and then against a second buffer (50 mM HEPES, 150 mM NaCl, 6 M urea, pH 7.5). The sample was loaded onto a second nickel column to remove the cleaved tag and any non-specifically binding molecules. Protein was collected from the flow-through at 80 μM and subsequently concentrated up to 2 mM, unless otherwise specified. SRSF1 was purified similarly to FUS NTD and DDX4 but under native conditions and was finally dialyzed in a buffer containing 50mM Tris pH 8, 1M NaCl, 3mM DTT (PanReac AppliChem), concentrated to 0.2 mM and stored after addition of 10% of glycerol (PanReac AppliChem). PTBP1 (40-316) was prepared as described previously (46).

Light microscopy

Olympus CKX41 microscope and Eclipse Ti Nikon microscope with 40X air objective were used to visually inspect samples. Samples were loaded in a 384-well glass bottom plate CORNING 4581.

Sample preparation in agarose gel

FUS and DDX4

To form stabilized liquid droplets, agarose buffer (30 mM HEPES, 200 mM KCl, 0.5% (w/v) agarose (ThermoFisher), pH 7.3) was boiled to solubilize agarose powder and then cooled in a room temperature water bath. Agarose buffer was still liquid at approximately 55 °C. To form protein droplets inside the hydrogel, protein stock was diluted in this warm agarose buffer in a 1.5 ml pre-warmed Eppendorf and quickly transferred to either the sample tube, or glass-bottom multi well plate. Due to the small volume used, the temperature drops quickly leading to liquid-droplet formation and agarose gelation. The residual urea concentration was 0.6 M. The same procedure was used for the dispersed-state sample, with the exception that the agarose gel contains, in addition, 3 M urea to keep the protein soluble.

A dispersed-phase sample measured with the same urea concentration as for the biphasic sample was prepared with protein concentration of only 5 μM to ensure that the protein does not undergo phase separation.

All samples were prepared with water of natural isotope abundance, except for the dispersed-state sample for the 4-pulse DEER measurement, which was prepared with D₂O to prolong the phase memory time. It is not advisable to use D₂O in preparing samples that undergo LLPS.

PTBP1 and SRSF1

Biphasic samples in agarose were prepared similarly to FUS and DDX4, with the exception of the phase separation trigger factor. PTBP1 droplets were prepared by adding poly-U RNA at a concentration of 1 mg/ml in the warm agarose mixture prior to the addition to 100 μM protein. SRSF1 protein was dispersed at 50 μM concentration in the presence of high salt concentration. Hence, in this case the agarose mixture for the dispersed sample contained 800 mM NaCl, while the biphasic sample contained 200 mM NaCl.

Bulk phase

The bulk phase sample was prepared similarly to Ref. (14). 40 μl of a 10 mM FUS NTD stock (50 mM HEPES, 150 mM NaCl, 6 M urea, pH 7.5) was diluted to a final concentration of 1 mM into a buffer (30 mM HEPES, 200 mM KCl, pH 7.3) preheated to 50°C that contained 0.4 nmol of the double cysteine mutant (1:1000 spin dilution). The sample was kept at 50°C to minimize phase separation and promote mixing. The material was then transferred into a 3 mm tube and centrifuged at 2000 g for 10 min at 25°C. The supernatant above the dense phase was carefully removed using a syringe and the process was repeated a second time. A final bulk phase volume of roughly 40 μl was reached using a total of 80 μl of 10 mM NTD stock.

Turbidity measurements

For turbidity measurements, 20 μl samples of liquid droplets in agarose were prepared as described above and pipetted while liquid in 384-well plates (CORNING 4581). Turbidity at 600 nm was measured using a Synergy 2 (Biotek) microplate reader at 25°C controlled temperature. Agarose buffer was used as blank. All samples were examined in triplicate.

Nanodrop measurements to estimate fraction of FUS within droplets

The fraction of protein in the condensed phase was determined by dilution of a 2 mM WT NTD stock into a buffer containing 30 mM HEPES, 200 mM KCl, pH 7.3. The urea concentration was adjusted to 0.6 M urea across all protein concentrations tested. The sample was centrifuged at 11000 g for 10 min at 25 °C. The supernatant was then carefully transferred to a new Eppendorf tube and the concentration determined by nanodrop (NanoDrop OneC ThermoFisher) measurement. The remaining pellet was resuspended in a buffer containing 50 mM HEPES, 150 mM NaCl, 6 M urea, pH 7.5 and the concentration determined by nanodrop measurement. The volumes of the supernatant and the resuspended pellet was determined by weighting of the Eppendorf tube, subtracting the weight of the empty tube (AT261 DeltaRange, Mettler Toledo) and assuming that the density does not deviate strongly from the one of water. The protein fraction in the condensed phase was then determined by diving the amount of substance in both fractions. All samples were done in triplicate.

Fluorescence microscopy to estimate droplet size and fraction of FUS within droplets

FUS droplet fluorescence measurements were done, using a Nikon Eclipse TE2000-U microscope with a 100× (1.46 NA Oil) objective and Andor EMCCD897 camera. An adequate filter set for Cy3 was used (600-50) in combination with a diode laser (561nm). Imaging conditions (EM gain of the camera, integration time and LASER power) were kept constant between the different concentrations. Purified FUS NTD G128C mutant was fluorescently cysteine labeled using Cy3-Maleimide. For that, 2.6 mg/ml protein was incubated for 1 hour at RT with 2 mM TCEP, before adding 10 mM Cy3-Maleimide to get a final concentration of 200 μM Cy3-Maleimide. After overnight incubation at 4°C, 10 % CHAPS was added to a final concentration of 0.1 % CHAPS in solution, before purifying by PD10 column. Fractions containing labeled FUS were then dialysed in protein buffer (6M Urea, 150 mM NaCl, 50 mM HEPES, pH 7.5) without CHAPS over night at room temperature.

For imaging, to exclude saturation and minimize influence of Cy3 fluorophore, labeled FUS NTD G128C protein was mixed with unlabeled FUS in a ratio of 1:100, before incorporating it into a 0.5 percent agarose gel, as described above. Size of droplets was estimated using the number of voxels within a FUS droplet, furthermore calculating the diameter of the assumed sphere using a custom written ImageJ script, based on 3D Objects Counter plugin (47). At least three different fields of view were recorded as a z-stack. The fraction of FUS within droplets was estimated by integrating the fluorescence for all voxels within FUS droplets and comparing it to fluorescence outside. Data was summarized and plotted using Python.

NMR experiments

All NMR experiments were recorded with a 3 mm NMR sample tube at 25 °C using the following Bruker spectrometers with z-axis pulsed field gradients: Avance III at 750 MHz proton frequency with a PATXI room temperature probe, Avance NEO at 500 MHz equipped with a CPQCI cryogenic probe, Avance III HD at 600 MHz, Avance III HD at 900 MHz proton frequency equipped with a CPTCI cryogenic probes. All spectra were processed with Topspin 3.2 (Bruker).

A standard pulse sequence (stebpgp1s19 from Topspin 3.2, Bruker) was used for diffusion experiments on the 750 MHz instrument. In total 4096 points with 32 scans were recorded in the proton dimension for each 1D with variable diffusion gradient strength ranging between 2 to 95% in various steps. The following parameters were used: diffusion time (Δ) 0.08 s, gradient pulse (δ) 12 ms, smoothed rectangular shaped gradients SMSQ10.100, relaxation delay (d1) 5 s.

2D ¹H-¹⁵N HSQC spectra were measured as follows: FUS NTD at the 900 MHz instrument ¹H/¹⁵N: spectral width (SW) 16/21 ppm, acquisition time (AQ) 0.122/0.035 s, d1 1 s, number of scans (NS) 32), PTBP1 at the 600 MHz instrument (¹H/¹⁵N: SW 16/26 ppm, AQ 0.122/0.063 s, d1 1 s, NS 32) and SRSF1 at the 500 MHz instrument (¹H/¹⁵N: SW 16/32 ppm, AQ 0.122/0.078 s, d1 1 s, NS 128).

Site-directed spin labelling

The protein solution was diluted to a protein concentration of 30 μM and incubated for 1 h with 5 mM DTT (Dithiothreitol, PanReac AppliChem). After washing out DTT with PD10 desalting columns (GE Healthcare Life Sciences) the sample was incubated for 2 h with a tenfold excess of the nitroxide spin label MTSL (((1-oxyl-2,2,5,5-tetramethylpyrroline-3-methyl)methanethiosulfonate, Toronto Research Chemicals). Subsequently, free spin label was washed out using PD10 desalting columns (GE Healthcare Life Sciences). The protein was then concentrated in Amicon Ultra-4 Centrifugal Filter Units (Merck & Cie) and the labelling efficiency was determined by CW EPR spectroscopy. These steps were performed in urea-containing buffers to keep the protein soluble.

Continuous-wave EPR spectroscopy

CW EPR spectra were acquired on a Bruker Elexsys E500 spectrometer equipped with a super-high Q resonator at X band (9.5 GHz). 20 μl of the protein samples with a concentration of 50 μM were loaded into a microcapillary tube. All measurements were recorded with 100 kHz field modulation.

Spectra for the determination of spin labelling efficiency were recorded with 1.5 G modulation amplitude. The labelling efficiencies were calculated by double integration of the spectrum and comparison with a sample of known concentration.

The temperature stabilized measurements were performed using a liquid nitrogen cryostat and a temperature controller (ER 4111 VT, Bruker). The spectra were recorded with 0.5 G modulation amplitude for measurements at 25°C and 1.5 G modulation amplitude for measurements at 5°C.

Pulsed EPR measurements

Pulsed EPR measurements were performed on a homebuilt Q-band spectrometer or a Bruker ElexsysE680 spectrometer extended by an incoherent AWG pulse channel (48). A temperature of 50 K was achieved by liquid helium cooling and was controlled by a He flow cryostat and a temperature controller. A sample tube loaded with 35 μl sample was shock frozen in precooled isopentane and subsequently inserted into a homebuilt Q-band resonator for 3 mm o.d. sample tubes (49).

An echo-detected field-swept EPR spectrum was acquired using a Hahn-echo sequence. The pulse sequence for the 4-pulse DEER experiment was π/2_obs-τ₁-π_obs-t₁-π_pump-(τ₁+τ₂−t₁)-π_obs-τ₂. The pump pulse was applied on the spectral maximum and the observer pulses were applied at a frequency offset of 100 MHz. Measurements on the Bruker Elexsys E680 spectrometer were acquired with a pulse delay τ₁ of 200 ns and a dead time delay of 80 ns. Measurements on the homebuilt Q-band spectrometer were acquired with a pulse delay τ₁ of 400 ns and a dead time delay of 280 ns. All traces were acquired using 16 ns pulses and eight step nuclear modulation averaging with an averaging time step of 16 ns.

For the 5-pulse version, the experiment was recorded with the sequence π/2_obs-(τ/2−t₀)-π_pump-t₀-π_obs-t′-π_pump-(τ−t′+δ)-π_obs-(τ/2+δ) with δ = 120 ns to separate the stimulated echo from the refocused echo. 5-pulse DEER data were recorded with HS{1,6} pump pulses of 150 MHz in width and 32 ns observer pulses with a frequency offset of 70 MHz. Two traces with shifted artefact were recorded alternately. For the first trace, t₀ was set to 300 ns. For the second trace, t₀ and the initial value of the delay t′ before the moving pump pulse were increased by a multiple of the time increment, i.e. 192 ns. For details on the nuclear modulation averaging see the Supplementary Information in Ref (34).

To determine the number of spins that contribute to a distance distribution and ensure that no multi-spin effects are observed, we measured 4-pulse DEER traces as a function of the pump pulse amplitude and hence the flip angle, as described in Ref. (50). This was done by setting the main attenuator to 0, 1.8, 3, 5, 7.2, and 10 dB, and readjusting the observer pulses while keeping the pump channel fixed. The main attenuator setting of 10 dB could not be reached for all samples due to insufficient power on the π_obs channel. In those cases, a main attenuator setting of 9 dB was employed instead. These measurements were performed with a 16 ns pump pulse and 32 ns observer pulses.

DEER data analysis

All DEER data were analysed using an older version of DeerLab based on MATLAB (35) (release 0.9.0, available at github.com/JeschkeLab/DeerLab-Matlab). All data analysis scripts are available upon request.

We used a multi-pathway kernel K(t,r) which allows the fitting of a distance distribution to 4-pulse and 5-pulse primary DEER data V(t) including any additional contributions which might arise from other pathways (35). This kernel has the form

$$K(t,r)=\left[{\Lambda }_0+{\lambda }_1{K}_0(t,r)+{\lambda }_2{K}_0\left(t-T_{0,2},r\right)\right]e^{-k\left({\lambda }_1|t{|}^{d/3}+{\lambda }_2{\left|t-T_{0,2}\right|}^{d/3}\right)},$$

where K₀ is the elementary dipolar kernel, ∧₀ accounts for the contribution of unmodulated dipolar pathways, λ₁ and λ₂ describe the amplitudes of the modulated dipolar pathways and T_0,2 is the refocusing time of the additional modulated dipolar pathway. The exponential term is the background function B(t) with fractal dimension d and decay rate k. While K₀ is fixed, Λ₀, λ₁, λ₂, T_0,2, d and k are fit parameters.

The signal of the dispersed state DEER data V_dis(t) was modelled according to the kernel above with an exponential background function (d = 3). The distance distribution P_dis(r) was fitted by Tikhonov regularization (using either the Bayesian information criterion or the residual method for regularization parameter selection) and by fitting a single Gaussian distribution (see Extended Data Fig. 9). Given the similarities in the shape of the regularization and Gaussian fits, we assumed a Gaussian distribution model for the forthcoming analysis.

Due to the presence of the two different species in the biphasic sample, we modelled the measured primary signal as a linear combination of two dipolar signals corresponding to spins in the dispersed fraction (V_dis) or in the condensed phase (V_cond) weighted by the fraction η of spins in each state

$$V(t)=\eta V_{\text{dis}}(t)+(1-\eta )V_{\text{cond}}(t).$$

The fraction η was constrained to ±10% of the fraction of proteins in the dispersed phase determined by NMR. Each of these signals arises from different distance distributions P_dis(r)

$$P_{\text{dis}}(r)=\frac{1}{{\sigma }_{\text{dis}}\sqrt{2/\pi }}{e}^{-\frac{{\left(r-{\langle r\rangle }_{\text{dis}}\right)}^2}{2{\sigma }_{\text{dis}}}}\text{with}{\sigma }_{\text{dis}}=\frac{{\Gamma }_{\text{dis}}}{2\sqrt{\ln 2}}$$

and P_cond(r)

$$P_{\text{cond}}(r)=\frac{1}{{\sigma }_{\text{cond}}\sqrt{2/\pi }}{e}^{-\frac{{\left(r-\langle r\rangle {)}_{\text{cond}}\right)}^2}{2{\sigma }_{\text{cond}}}}\text{with}{\sigma }_{\text{cond}}=\frac{{\Gamma }_{\text{cond}}}{2\sqrt{\ln 2}},$$

modelled as Gaussian distributions, where 〈r〉_dis and 〈r〉_cond are the mean distances, and Γ_dis and Γ_cond the full-width at half-maximum (FWHM) values of the dispersed and condensed states distributions, respectively.

Due to the difference in spin concentration inside and outside of the liquid droplets, they also exhibit significantly different background functions: a slowly decaying one B_dis(t) and a rapidly decaying one B_cond(t). The model can be rewritten as

$$\begin{array}{c}V(t)=\eta {\int }^{\text{}}{K}_{\text{dis}}(t,r)P_{\text{dis}}(r)\text{d}r+(1-\eta ){\int }^{\text{}}{K}_{\text{cond}}(t,r)P_{\text{cond}}(r)\text{d}r\\ =\eta {\int }^{\text{}}\left[{\Lambda }_0+{\lambda }_1{K}_0(t,r)+{\lambda }_2{K}_0\left(t-T_{0,2},r\right)\right]e^{-k_{\text{dis}}\left({\lambda }_1\left|t\right|+{\lambda }_2\left|t-T_{0,2}\right|\right)}{P}_{\text{dis}}(r)\text{d}r\\ {+(1-\eta ){\int }^{\text{}}\left[{\Lambda }_0+{\lambda }_1{K}_0(t,r)+{\lambda }_2{K}_0\left(t-T_{0,2},r\right)\right]e^{-k_{\text{cond}}({\lambda }_1|t{|}^{d/3}+{\lambda }_2{\left|t-T_{0,2}\right|}^{d/3}})}_{P_{\text{cond}}}(r)\text{d}r,\end{array}$$

where the kernels K_dis(t,r) and K_cond(t,r) differ by their background function. For the condensed phase we employed a stretched exponential function as the basis function for the multi-pathway background

$$B_{\text{cond}}(t)=e^{-k_{\text{cond}}\left({\lambda }_1|t{|}^{d/3}+{\lambda }_2{\left|t-T_{0,2}\right|}^{d/3}\right)},$$

while for the dispersed fraction we employed an exponential background function as the basis function for the multi-pathway background

$$B_{\text{dis}}(t)=e^{-k_{\text{dis}}\left({\lambda }_1\left|t\right|+{\lambda }_2\left|t-T_{0,2}\right|\right)},$$

with a very constrained decay rate k_dis thanks to the information obtained from the fit of the monophasic dispersed state data. An illustrative schematic of this model is shown in Extended Data Fig. 10a).

We performed a global analysis of two 5-pulse DEER signals, whose secondary pathway is shifted in time with respect to each other. All model parameters (η, Λ₀, λ₁, λ₂, d, k_dis, k_cond, 〈r〉_dis, 〈r〉_cond, Γ_dis and Γ_cond) were fitted globally to both datasets, except the refocusing time T_0,2, which was fitted locally to each dataset. All constraints used to fit these parameters are listed in Supplementary Table 1.

Here we assumed that the dispersed state in the biphasic sample retains the same distance distribution as the dispersed state in the monophasic sample. Therefore, we constrained the parameter range to the values obtained for the dispersed state in the monophasic sample with the same urea concentration (0.6 M urea, Extended Data Fig. 9c,d), which stabilizes fitting. As a control experiment we prepared a droplet sample under the same conditions as specified above without the addition of agarose, centrifuged for 10 min at 25 °C and 11,000 g and loaded the supernatant into a sample tube. The primary DEER data of this measurement overlap perfectly with the data for the dispersed state in a monophasic sample at 0.6 M urea (Supplementary Fig. 1), confirming our hypothesis that the distance distribution of the dispersed state is the same in monophasic and biphasic samples.

Global optimization was employed by means of a multi-start algorithm to ensure that the solution is a global solution. For all fitted signals, distributions, and parameters, bootstrapped confidence intervals were estimated from 1000 bootstrap samples drawn from a Gaussian distribution based on the noise in the measurements.

Analysis of the spin counting experiment

For the spin counting experiment, the nominal inversion efficiency was calculated by

$$\lambda =\frac{1-\cos \beta }{2}$$

with the flip angle

$$\beta =\pi \cdot {10}^{-\frac{A}{20\text{dB}}},$$

where A is the main attenuator setting in dB. The dependence of the total modulation depth on the nominal inversion efficiency, was then fitted using

$$\Delta (\lambda )=\underset{k=1}{\overset{N-1}{{\sum }^{\text{}}}}{d}_k{\lambda }^k.$$

A good fit was obtained for N = 2 for all samples tested (Extended Data Fig. 9g-j) confirming that intermolecular distances do not significantly contribute to the distance distributions.

Calculation of physical properties of differently sized particles

The trends shown in Fig. 1e are calculated under the assumption of a spherical shape using

$$V=\frac{4}{3}\pi r^3$$

for the volume and

$$SA:V=\frac{3}{r}$$

for the surface area to volume ratio. In the above expressions r is the radius of the sphere. We used a volume of 200 μl for the bulk phase, which corresponds approximately to the volume used in NMR experiments.

Extended Data

Extended Data Fig. 1. FUS NTD droplets.

Light microscope image of FUS NTD droplets formed upon dilution of urea. Representative image from four independent experiments. Scale bar: 10 μm.

Extended Data Fig. 2. FUS NTD droplets in presence and absence of agarose.

a,b) Upfield region of ¹H 1D and ¹H-¹⁵N HSQC NMR spectra of FUS NTD in absence and presence of 0.5% agarose hydrogel showing comparable linewidths. c) Unattenuated DOSY signal of FUS NTD biphasic sample at different agarose concentrations.

Extended Data Fig. 3. FUS NTD droplet fraction quantification correlates with turbidity.

a) Aliphatic proton region of ¹H 1D spectra in the presence and in absence of urea. Color scale (magenta -> blue -> red) corresponds to spectra recorded with increasing diffusion gradient strength. b) Integrals of the spectral region in (a) as a function of diffusion gradient strength normalized to the integral at the lowest gradient strength. The fraction of the slowly diffusing component obtained by averaging data points obtained at >30 G/cm is 19%. c) Droplet size comparison among different total protein concentrations determined by fluorescence microscopy. d) Correlation of final percentage of unattenuated signal from the DOSY experiments with sample turbidity as function of protein concentrationror bars indicate standard deviation and dots the mean from three independent experiments.

Extended Data Fig. 4. Protein phase exchange during NMR DOSY experiment.

Estimation of protein molecules remaining in the same phase during the diffusion time of our diffusion NMR experiments (0.08 s). In vitro half time, defined as ln(2)/(exchange rate), of FUS NTD in the droplet phase, as measured by FRAP experiments (14), is marked with a yellow dashed line. The range of accessible experimental diffusion times in NMR experiments is highlighted with diagonal stripes.

Extended Data Fig. 5. Effect of agarose hydrogel on stability of liquid droplet of FUS, DDX4, PTBP1 and SRSF1.

Brightfield microscope images of liquid droplets from different proteins hours post preparation in the presence and absence of agarose hydrogel. Representative images from three independent experiments. (Scale bar: 10 μm)

Extended Data Fig. 6. DDX4 droplets in 0.5% agarose hydrogel.

a) Time progression of sample turbidity in the absence (black) and presence (red) of agarose hydrogel. Error bars indicate standard deviation and dots the mean from three independent experiments. b) Overlay of ¹H 1D DOSY spectra of dispersed (top) and biphasic (bottom) sample in agarose. Increasing gradient strength is visualized by a color gradient magenta ‘blue’ red. c) Integral of the spectral region shown in (b) normalized to the integral at the lowest gradient strength as a function of the gradient strength. Black and red data denote dispersed and biphasic sample in agarose, respectively. d) Integrals of ¹H 1D DOSY spectra as function of gradient strength at different DDX4 concentrations. e) Correlation of final percentage of unattenuated signal from the DOSY experiments shown in (e) with sample turbidity as function of protein concentration. Error bars indicate standard deviation and dots the mean from three independent experiments.

Extended Data Fig. 7. Diffusion NMR experiments on SRSF1 and PTBP1.

a, b) Integrated normalized aliphatic spectral region plotted vs. gradient strength for dispersed (black) and biphasic (red) PTBP1 and SRSF1 respectively. c, d) Amide and aliphatic regions of ¹H 1D NMR spectra of biphasic PTBP1 and SRSF1 in agarose. e, f) ¹H-¹⁵N HSQC spectrum of biphasic PTBP1 and SRSF1 in agarose.

Extended Data Fig. 8. Brightfield microscope images of FUS NTD.

FUS NTD in a) 50% glycerol and in b) 50% PEG shows no phase separation under buffer conditions where liquid phase separation is otherwise observed, as reflected in c). Representative images from three independent experiments. Scale bar: 30 μm

Extended Data Fig. 9. DEER experiments on the NTD of FUS.

Primary DEER data and corresponding distance distributions using a model-free fit with Tikhonov regularization (black) and a Gaussian distribution (red) of A10C S29C R1 and A105C G128C R1 in a,b) the dispersed state with 3 M urea, in c, d) the dispersed state with 0.6 M urea, and e, f) in the bulk phase. The experimental data are displayed as black dots, the fit as a solid line and the 95% confidence interval obtained via 1000 bootstrap samples as shaded area. The Tikhonov regularization and fit with unimodal Gaussian function with variable mean and width lead to very similar distribution shapes. Total modulation depth Δ as a function of nominal inversion efficiency λ_nominal for A10C S29C R1 in g) the biphasic state and h) the bulk phase, and A105C G128C R1 in i) the biphasic state and j) the bulk phase. Good fits of the experimental data (black dots) are obtained using a model for two spins (solid red line), which confirms that the spin dilution employed is sufficient to avoid inter-molecular distance contributions in the DEER experiment.

Extended Data Fig. 10. Schematics of the fitting algorithm employed for the analysis of the biphasic DEER measurements.

Red corresponds to spin-labeled protein in the dispersed phase, blue is spin-labeled protein in the condensed phase, and gray represents unlabeled, and therefore EPR-silent, protein. The total signal of the biphasic sample is displayed in black.

Data availability

The authors declare that the data supporting the findings of this study are available within the paper and its supplementary information files.

Code availability

MATLAB code used for the analysis of the DEER data and ImageJ script used for the analysis of the fluorescence microscopy experiments are available from the authors upon reasonable request.