Introduction

Phase separation is a fundamental physicochemical process by which biomolecules—such as proteins and nucleic acids—spontaneously demix from the surrounding solution (the dilute phase) to form concentrated, mesoscale condensates (the dense phase) within cells1,2,3,4,5. These biomolecular condensates are dynamic, membraneless compartments composed of multiple components and are formed independently of lipid bilayers. Their assembly is primarily driven by two mechanisms: interactions mediated by intrinsically disordered regions (IDRs) of proteins6, and multivalent interactions among modular macromolecules7. Phase separation-driven condensates exhibit spatiotemporal self-organization and coarsening dynamics[8](https://www.nature.com/articles/s41467-025-65069-4#ref-CR8 “Li, C., Guo, M.-T., He, X., Liu, Q.-X. & Qi, Z. Modeling phase separation of biomolecular condensates with data-driven mass-conserving reaction-diffusion systems. Structure https://doi.org/10.1016/j.str.2025.05.018

(2025).“) and are integral to numerous biological processes1,9,10. Furthermore, dysregulation of condensate formation has been linked to the pathogenesis of various human diseases, including neurodegenerative disorders and cancer3.

Multicomponent condensates exhibit a range of complex architectures, including “pearl chain”-like structures11,12,13,14, nested “Russian doll” structures5,15, and hollow condensate architectures16,17,18. Among these, hollow condensates represent the simplest model that provides an entry point for studies of complex architectures. For example, Banerjee and colleagues17 demonstrated the formation of hollow condensates consisting of an arginine-rich disordered nucleoprotein, protamine (PRM), in combination with RNA. They proposed that PRM and RNA assemble in a manner similar to a lipid-like diblock copolymer, where these copolymers, along with high concentrations of RNA or protein, lead to vesicle-like hollow condensates.

In this work, we report a new type of hollow condensates. One important FUS/EWS/TAF15 (FET) family fusion oncoprotein, FUS-ERG, can form hollow co-condensates with 25-base pair (bp) double-stranded DNA (dsDNA) containing a 4 × GGAA microsatellite sequence (25-bp 4 × GGAA dsDNA). FUS-ERG is formed by the fusion of the low-complexity domain (LCD) and the RGG domain of FUS with the DNA-binding domain (DBD) of the E26 transformation-specific (ETS) family transcription factor ERG. The DBD endowed FUS-ERG with in vivo binding specificity to a genomic microsatellite sequence characterized by GGAA repeats19,20.

To interrogate the biophysical mechanism of FET family fusion oncoprotein hollow co-condensates and their practical implications, we employed a multidisciplinary approach. Remarkably, super-resolution imaging experiments combined with mathematical modeling illuminated a molecular mechanism driving the formation of these hollow co-condensates, which is distinctly different from that of vesicle-like hollow co-condensates described in a previous study17.

Biomolecular condensate offers significant advantages in facilitating rapid biomolecular self-assembly, which has been leveraged in the design of various synthetic functional structures21,[22](https://www.nature.com/articles/s41467-025-65069-4#ref-CR22 “Ji, B.-T., Pan, H.-T., Qian, Z.-G. & Xia, X.-X. Programming biological communication between distinct membraneless compartments. Nat. Chem. Biol. https://doi.org/10.1038/s41589-025-01840-4

(2025).“). Because the newly discovered FET family fusion oncoprotein hollow co-condensates involve protein-DNA interactions, we asked whether they can provide a compelling opportunity for dynamic information manipulation in DNA-based data storage. While studies have shown that in-storage DNA encapsulation within abiotic polymers enhances data longevity and PCR uniformity, DNA encapsulation has so far been produced by pre-loading in a non-selective manner and lacked dynamic storage capacity23,24. In this study, we leverage protein-DNA self-assembly based on sequence-specific FET fusion oncoprotein-DNA interactions and the hollow co-condensate architecture as a dynamic and selective encapsulation medium. We demonstrated in-storage precise information selection and deletion, as well as hierarchical information sorting by programming DBD-dsDNA interactions. These results underscore the potential of using multi-layer biomolecular condensate for the dynamic spatial regulation of molecular information, providing a distinctive route for in-storage molecular storage and computation.

Results

FUS-ERG can form hollow co-condensates with dsDNA containing GGAA microsatellite sequence

We first conducted in vitro droplet assays to determine whether FET fusion oncoprotein FUS-ERG can form biomolecular condensates. We purified a GFP-tag-labeled FUS-ERG (GFP-FUS-ERG, Supplementary Fig. 1a(i)–(ii)), and performed electrophoretic mobility shift assays (EMSAs) to validate the protein activity in vitro (Supplementary Fig. 1a(iii)). Next, biochemical assays revealed that GFP-FUS-ERG protein can form homogeneous condensate at a concentration as low as 1 μM in vitro (Fig. 1a), consistent with previous findings25. Mixing 0.6 μM of 25 bp dsDNA containing a random sequence (referred to as 25 bp random dsDNA) and 5 μM GFP-FUS-ERG resulted in the co-localization of both components within homogeneous droplets (Fig. 1b(i)–(ii)). However, when dsDNA substrates of the same length were designed to contain GGAA microsatellites, such as 2 × or 4 × GGAA, FUS-ERG formed hollow co-condensates with these dsDNAs (Fig. 1b(iii)–(iv)). The three-dimensional hollow architecture was confirmed by confocal microscopy (Supplementary Movie 1). Fluorescence analysis revealed that 25 bp 4 × GGAA dsDNA and GFP-FUS-ERG co-localized on the shell of hollow co-condensates (Fig. 1b(v)). Additionally, we observed two interesting phenomena. First, the homogeneous condensates formed by 25-bp random dsDNA and GFP-FUS-ERG had the dsDNA enveloped by FUS-ERG (Fig. 1b(ii)), suggesting that the protein itself has stronger interactions with the solvent. Second, when the length of random dsDNA increased from 25 to 306 bp, hollow co-condensates can also be formed, albeit with a larger diameter (Supplementary Fig. 2a, b), suggesting that complex protein-DNA interactions govern the condensate architecture.

Fig. 1: FUS-ERG can form hollow co-condensates with dsDNA containing GGAA microsatellite sequence.

a GFP-FUS-ERG can form biomolecular condensate in the concentration of 1, 2, 5, 10 and 20 μM. b 5 μM GFP-FUS-ERG mixed with 0.6 μM 25-bp random dsDNA (i), 0.6 μM 25 bp 2 × GGAA dsDNA (iii); 0.6 μM 25 bp 4 × GGAA dsDNA (iv). dsDNA was labeled with AlexaFluor647. (ii) and (v) Normalized intensity profiles in (i) and (iv). (c) (i) GFP-FUS-ERG mixed with 25 bp 4 × GGAA dsDNA; (ii) Phase diagram in (i). (d) (i) GFP-FUS-ERG (9YS) can form biomolecular condensate in the concentration of 1, 2, 5, 10, and 20 μM; (ii) 5 μM GFP-FUS-ERG (9YS) mixed with 0.6 μM 25-bp 4 × GGAA dsDNA. (iii) GFP-FUS-ERG (27YS) cannot form biomolecular condensate in 20 μM; (iv) 5 μM GFP-FUS-ERG (27YS) mixed with 0.6 μM 25-bp 4 × GGAA dsDNA. (e) (i) GFP-FUS-ERG (5RA) can form biomolecular condensate in the concentration of 1, 2, 5, 10, and 20 μM; (ii) 5 μM GFP-FUS-ERG (5RA) mixed with 0.6 μM 25 bp 4 × GGAA dsDNA. (iii) GFP-FUS-ERG (9RA) can form biomolecular condensate in the concentration of 5, 10, and 20 μM; (iv) 5 μM GFP-FUS-ERG (9RA) mixed with 0.6 μM 25-bp 4 × GGAA dsDNA. All in vitro droplet assays were executed under physiological conditions, specifically 40 mM Tris-HCl (pH 7.5), 150 mM KCl, 2 mM MgCl2, 1 mM DTT and 0.2 mg/mL BSA, with thorough mixing and a 30-minute incubation period prior to imaging, unless otherwise indicated. Scale bar: 5 μm in (a, b(i), (iii), (iv), c(i), d, and e). Scale bar: 2 μm in b(i) insert and b(iv) insert. Source data are provided as a Source Data file.

We investigated the conditions required for hollow co-condensate formation through in vitro droplet assays, combining 25 bp 4 × GGAA dsDNA at concentrations of 0, 0.15, 0.3, 0.6, 1.2, and 2.4 μM with GFP-FUS-ERG protein at concentrations of 0.25, 0.5, 2, 5, and 10 μM (Fig. 1c(i)). When the DNA concentration was fixed at 0.15 μM, hollow condensates formed only at a protein concentration of approximately 2 μM, with no hollow condensates observed at either lower or higher protein concentrations. At protein concentrations below 2 μM, no condensates were detected, whereas at concentrations above 2 μM, the hollow structures became filled, yielding homogeneous condensates. This non-monotonic dependence on protein concentration about the hollow structure formation is a hallmark of reentrant phase behavior26,27.

The resulting phase diagram (Fig. 1c(ii)) reveals that the DNA-to-protein molar ratio beyond a threshold number ([DNA] / [protein] ~ 0.075) drives the hollow co-condensate formation. This threshold arises because the protein alone can form homogeneous condensates at concentrations above ~2 μM. Notably, increasing the DNA concentration from 0.15 to 2.4 μM did not alter the threshold of protein concentration for hollow structure formation, which remained fixed at ~2 μM. The molecular basis of this reentrant phase behavior in these hollow condensates remains to be elucidated and will be the subject of future investigation.

Both hollow and homogeneous condensates exhibited slow fusion kinetics (Supplementary Fig. 3a). Following this, we conducted fluorescence recovery after photobleaching (FRAP) experiments (Supplementary Fig. 3b(i)–(ii) and Methods) using GFP-FUS-ERG alone or mixed with 25 bp 4 × GGAA dsDNA. The molecular dynamics of GFP-FUS-ERG in hollow co-condensates were slower compared to homogeneous condensates, suggesting slower internal dynamics inside the shell region of hollow co-condensates.

While modifying the dsDNA sequence and length has been shown to regulate the formation of hollow condensates, we next asked how specific FUS-ERG domains affect the condensate morphology. The FUS LCD domain contains 27 [G/S]Y[G/S] repeats, where the tyrosines were shown to be important for FUS’s threshold concentration for condensation28,29. When nine relevant tyrosine residues were changed to serine in GFP-FUS-ERG (9YS) (Supplementary Fig. 1c), the threshold concentration for condensation remained unchanged (Fig. 1d(i)). However, 5 μM GFP-FUS-ERG (9YS) cannot form hollow co-condensates with 25 bp 4 × GGAA dsDNA (Fig. 1d(ii)). When all 27 tyrosines were mutated to serine in GFP-FUS-ERG (27YS) (Supplementary Fig. 1d), this mutant cannot form condensates even at 20 μM (Fig. 1d(iii)), and 5 μM GFP-FUS-ERG (27YS) with dsDNA also failed to form condensates (Fig. 1d(iv)).

For the FUS RGG motif, we observed a similar trend. Mutation of five key arginine residues to alanine (GFP-FUS-ERG (5RA), Supplementary Fig. 1e) did not affect the threshold concentration for condensation (Fig. 1e(i)). In contrast, mutating nine arginine residues to alanine (GFP-FUS-ERG (9RA), Supplementary Fig. 1f) significantly increased the threshold concentration for condensation (Fig. 1e(iii)). GFP-FUS-ERG (5RA) formed both homogeneous and hollow co-condensates with 25 bp 4 × GGAA dsDNA (Fig. 1e(ii)). Conversely, GFP-FUS-ERG (9RA) only formed smaller homogeneous condensates with 25-bp 4 × GGAA dsDNA (Fig. 1e(iv)). Lastly, GFP-FUS-DDIT3, another FET fusion oncoprotein, has been reported to form similar spherical shell in vivo30. We repeated this experiment in U2OS cells (Supplementary Fig. 4a). Interestingly, when all arginine residues in the RGG motif were mutated to alanine, FUS-DDIT3-GFP (9RA) also cannot form this architecture in vivo (Supplementary Fig. 4b). Taken together, these results indicate that the LCD and RGG motifs in FUS-ERG strongly regulate hollow co-condensate formation.

Next, we sought to track the temporal dynamic of condensate formation by a two-step experiment. Initially, we used GFP-FUS-ERG to form homogeneous condensates (Fig. 1a). Subsequently, we introduced 25 bp 4 × GGAA dsDNA (Fig. 2a(i), time 0). Remarkably, the dsDNA substrates promptly enveloped the exterior surface of the homogeneous GFP-FUS-ERG condensates (Fig. 2a(v)). This process was monitored continuously for 90 min (Fig. 2a and Supplementary Movie 2). By 30 min (Fig. 2a(ii)), the dsDNA substrates had completely infiltrated the condensate, coinciding with a reduction in protein intensity within the central region (Fig. 2a(vi)). By 90 min (Fig. 2a(iv)), hollow co-condensates had fully formed, with both dsDNA and protein concentrated on the shell of the hollow co-condensates (Fig. 2a(vii)).

Fig. 2: 25-bp dsDNA substrates containing GGAA microsatellites can transfer into the homogeneous FUS-ERG condensates, inducing the hollow co-condensate formation.

a Time course of hollow co-condensate formation of 5 μM GFP-FUS-ERG mixed with 0.6 μM AlexaFluor647-labeled 25 bp 4 × GGAA dsDNA at 0 min (i), 30 min (ii), 60 min (iii), and 90 min (iv). At the 0 min time point, we injected dsDNA. (v), (vi), and (vii) are representative events of hollow co-condensate formation and normalized intensity profiles from (i), (ii), and (iv). b (i) Time course of hollow co-condensate formation of 5 μM GFP-FUS-ERG mixed with 0.6 μM AlexaFluor647-labeled 25-bp random dsDNA at 0, 30, and 90 min. At the 0 min time point, we injected dsDNA. (ii) is the normalized intensity profile at the 90 min time point. c Boxplot of the mean intensity of dsDNA inside the condensates for GFP-FUS-ERG with 25 bp random dsDNA and 25 bp 4 × GGAA dsDNA. The total number N examined over one-time in vitro droplet experiments. For the boxplot, the red bar represents median. The bottom edge of the box represents 25th percentiles, and the top is 75th percentiles. Most extreme data points are covered by the whiskers except outliers. The ‘+’ symbol is used to represent the outliers. Statistical significance was analyzed using unpaired t test for two groups. P value: two-tailed; p value style: GP: 0.1234 (ns), 0.0332 (*), 0.0021 (**), 0.0002 (***), < 0.0001 (****). Exact P values are as follows: P < 0.0001 for all conditions: 4 × GGAA dsDNA 0 vs. 30 min; 4 × GGAA dsDNA 0-min vs. 90 min; 4 × GGAA dsDNA 90 min vs. random dsDNA 90 min; random dsDNA 0 min vs. 30 min; random dsDNA 0 vs. 90 min. Confidence level: 95%. Scale bar: 5 μm in a(i)–(iv). Scale bar: 2 μm in a(v)–(vii) and b(i). Source data are provided as a Source Data file.

We conducted two control experiments. First, when 25 bp random dsDNA was used, no hollow co-condensates formed even after 90 min (Fig. 2b), and significantly fewer dsDNA molecules were transferred into the GFP-FUS-ERG condensates (Fig. 2c). Second, we repeated the experiment shown in Fig. 2a, substituting the 25 bp 4 × GGAA dsDNA with 306 bp random dsDNA (Supplementary Fig. 2c). Interestingly, even after 90 min (Supplementary Fig. 2c(iv)), the dsDNA infiltrated the condensates and colocalized with protein, but no hollow condensate formation was observed—contrasting sharply with the outcome in Fig. 2a. These findings demonstrate the spontaneous transfer of dsDNA containing GGAA microsatellites into the condensates accompanying hollow co-condensate formation. These findings also highlight a key mechanistic distinction between site-specific binding by microsatellite DNA and non-specific adsorption or wetting by long random DNA31, which may differentially stabilize hollow condensates. Exploring this distinction represents an interesting avenue for future research.

The formation of hollow co-condensates is driven by nested asymmetric phase separation

To uncover the molecular mechanism driving the formation of hollow co-condensates, we focused on the processes that define the development of the external and internal surfaces. GFP-FUS-ERG alone forms biomolecular condensate with the working buffer (Fig. 1a), suggesting that the stable external surface arises from interactions between free proteins and the buffer. Furthermore, in co-condensates formed by GFP-FUS-ERG and 25 bp random dsDNA, the GFP-FUS-ERG boundary extended beyond the dsDNA boundary (Fig. 1b(i)–(ii)). This observation strongly indicates that phase separation between free proteins and the surrounding buffer could drive the formation of a stable external surface.

To elucidate the formation of the internal surface, we firstly employed RNA probes to discern the internal characteristics of the hollow co-condensates. Specifically, we combined SNAP-tag-labeled FUS-ERG (SNAP-FUS-ERG, Supplementary Fig. 1b(i)–(iii)) with 25 bp 4 × GGAA dsDNA to induce the formation of hollow co-condensates (Supplementary Fig. 1b(iv)). After a 30 min incubation with Poly-U RNA, we observed the penetration of the RNA through the shell, resulting in their enrichment within the lumens of the hollow co-condensates (Fig. 3a, d). In contrast, RNA encountered difficulty in penetrating the homogeneous condensates formed by SNAP-FUS-ERG alone (Fig. 3c, d) or in conjunction with 25-bp random dsDNA (Fig. 3b, d). As the RGG motif (amino acids 485–539) (Supplementary Fig. 1a(i)) is the only domain within GFP-FUS-ERG capable of tightly binding RNA, these findings strongly suggest that the internal surface of the hollow condensates was aligned with RGG motifs whereas in homogeneous condensates the RGG motifs were likely sequestered within the protein.

Fig. 3: RNA probe and STED imaging reveal the molecular mechanism of hollow co-condensate formation.

a–c 5 μM SNAP-FUS-ERG first mixed with 0.6 μM ATTO425-labeled 25-bp 4 × GGAA dsDNA (a(i)), random dsDNA (b(i)), or no dsDNA (c). Next, 5 ng/μL 50-nt Cy5-labeled PolyU RNA was injected. a(ii) and b(ii) are normalized intensity profiles from a(i) and b(i). d Boxplot of the mean intensity of RNA inside the condensates in a(i), b(i) and c. The total number N examined over three-time in vitro droplet experiments. Exact P values are as follows: 4 × GGAA dsDNA vs. random dsDNA, P < 0.0001; 4 × GGAA dsDNA vs. no dsDNA, P < 0.0001; random dsDNA vs. no dsDNA, P = 0.0016. e Schematics of hollow co-condensate formation. f (i) Representative fluorescence heatmap of Stimulated Emission Depletion (STED) microscopy image for 5 μM SNAP-FUS-ERG mixed with 25 bp 4 × GGAA dsDNA: 0.6 μM dark dsDNA and 0.006 μM dsDNA labeled with Atto647N (100:1). (ii) Boxplot of the radial intensity distribution of dsDNA labeled with Atto647N within the hollow co-condensates. The total number N examined over three-time in vitro droplet experiments. g (i) Representative fluorescence heatmap of STED image for 4 mg/mL FUS-RGG3 (FUS No. 471-504) mixed with Poly-U RNA: 20,000 ng/μL dark RNA and 100 ng/μL RNA labeled with AlexaFluor 647 (200:1). (ii) Boxplot of the radial intensity distribution of RNA labeled with AlexaFluor 647 within the hollow co-condensates. The total number N examined over three-time in vitro droplet experiments. For the boxplot in d, f(ii) and g(ii), the red bar represents median. The bottom edge of the box represents 25th percentiles, and the top is 75th percentiles. Most extreme data points are covered by the whiskers except outliers. The ‘+’ symbol is used to represent the outliers. Statistical significance was analyzed using unpaired t test for two groups. P value: two-tailed; p value style: GP: 0.1234 (ns), 0.0332 (*), 0.0021 (**), 0.0002 (***), < 0.0001 (****). Confidence level: 95%. Scale bar: 1 μm in a(i), b(i), c; 0.5 μm in f(i); 2 μm in g(i). Source data are provided as a Source Data file.

Integrating the RNA probe experiments (Fig. 3a–d) with earlier observations of the translocation of GGAA motif-containing dsDNA into the condensates (Fig. 2a), we propose an interesting hypothesis for the formation of the internal surface. When a dsDNA substrate containing GGAA motifs binds to a free FUS-ERG molecule near the condensate’s exterior boundary (Fig. 2a(i)), a protein–dsDNA complex is formed, inducing a conformational change in the protein from a Closed state, where the RGG motif is buried and only display week affinity to RNA, to an Open state where the RGG motif is exposed (Fig. 3e(i)). The results shown in Fig. 2 suggest that this protein–dsDNA complex can translocate into the condensate’s inner region, where the exposed hydrophilic RGG motif (Supplementary Fig. 5a) accumulates, thereby establishing the internal surface (Fig. 3e(ii)).

To test the hypothesis of conformational switch from closed to opened state of GFP-FUS-ERG in Fig. 3e, we divided the full-length GFP-FUS-ERG into two separate modules: GFP-FUS-LCD and RGG-ERG (comprising the RGG motif and ERG domain). SDS-PAGE and EMSA analyses confirmed that both proteins were purified to high homogeneity and that RGG-ERG retained sequence-specific DNA-binding activity (Supplementary Fig. 6a). Notably, GFP-FUS-LCD did not exhibit any detectable DNA-binding, even at very high concentrations (Supplementary Fig. 6b).

To test whether FUS-ERG can adopt a closed conformation mediated by interactions between the FUS-LCD and the ERG domain, we performed an in vitro droplet assay. GFP-FUS-LCD and RGG-ERG proteins were incubated either individually or in combination at equimolar concentrations for 30 minutes at room temperature prior to imaging (Supplementary Fig. 7a(i)). GFP-FUS-LCD alone did not form condensates, while RGG-ERG readily formed droplets on its own. Upon mixing, the two proteins exhibited clear co-localization of fluorescent signals (Supplementary Fig. 7b(i)), indicating direct interactions between GFP-FUS-LCD and RGG-ERG. These results strongly support the presence of a closed state driven by intramolecular or intermolecular contacts between the FUS-LCD and ERG domain (Fig. 3e(i)).

To determine whether dsDNA containing GGAA motifs can induce a conformational transition of FUS-ERG from a closed to an open state, we performed an in vitro droplet assay in which GFP-FUS-LCD and RGG-ERG were mixed at equimolar concentrations and incubated with increasing amounts of 25 bp 4 × GGAA DNA or control 25 bp random DNA (Supplementary Fig. 7a(ii–iii)). At a low concentration of 4 × GGAA DNA (0.3 μM), GFP-FUS-LCD continued to co-localize with RGG-ERG. However, at or above 0.6 μM, this co-localization was abolished (Supplementary Fig. 7b(ii)). In contrast, random DNA failed to disrupt co-localization even at concentrations as high as 2.4 μM (Supplementary Fig. 7b(iii)). These results indicate that only high concentrations of GGAA-containing dsDNA can displace GFP-FUS-LCD from RGG-ERG condensates, likely by disrupting the interaction between the FUS-LCD and ERG domain. This observation aligns with the previous work32, which showed that high-affinity DNA disrupts the interaction between EWS-LCD and FLI1-DBD within condensates. Collectively, these data strongly support that GGAA motif-containing dsDNA induces a transition of FUS-ERG from a closed to an open state. Moreover, our results demonstrate that control dsDNA lacking GGAA motifs does not elicit this conformational change. Together with the droplet assay results from Supplementary Fig. 7 and Fig. 1b, our findings support that the closed-to-open transition of FUS-ERG is a critical prerequisite for the formation of hollow condensates.

Despite the above experiments, we are still unable to precisely demonstrate the existence of conformational changes at the molecular level. If this hypothesis holds true, the interior surface of the hollow co-condensates is expected to exhibit the highest concentration of dsDNA–protein complexes, with a radial decrease in concentration toward the exterior. To investigate this distribution, we employed Stimulated Emission Depletion (STED) microscopy, which provides sub-100 nm resolution33, to image the dsDNA distribution within hollow condensates. Fluorescent signals from the ATTO647N-labeled (1:100) dsDNA substrates are presented in Fig. 3f(i). Quantitative imaging analysis (Supplementary Methods) revealed that the interior surface exhibited significantly higher dsDNA intensity, with a gradual decrease in intensity along the radial axis toward the exterior (Fig. 3f(ii)), confirming the asymmetric distribution of the FUS-ERG–dsDNA complex.

The hollow co-condensates formed by PRM and RNA17 are expected to exhibit a symmetric distribution of RNA due to their vesicle-like characteristics. As the RGG motifs of FUS are similar to PRM in RNA binding abilities, we repeated the STED experiment to image the distribution of Poly-U RNA co-condensates with the third RGG motif inside FUS (FUS RGG-3, 471-504) into vesicle-like droplets (Supplementary Fig. 8). The fluorescent signals of RNA substrates are depicted in Fig. 3g(i), and subsequent image analysis confirmed the symmetric distribution of RNA substrates (Fig. 3g(ii)). Collectively, these super-resolution imaging results elucidate a molecular mechanism underlying the internal surface formation of FUS-ERG and dsDNA hollow co-condensates, which we term “nested asymmetric phase separation”.

A mathematical model reproduces the formation of hollow co-condensates and predicts their selective capacity for DNA

To elucidate the molecular mechanism underlying hollow co-condensate formation, we developed a molecularly informed phase-field model34. This model incorporates three key order parameters. The first order parameter, (\eta), represents the concentration of protein–dsDNA complexes (Open state in Fig. 3e(i)) relative to a critical threshold ({\psi }_{C}) for phase separation. The second order parameter, (\phi), indicates local hydrophilicity ((\phi > 0)) or hydrophobicity ((\phi < 0)) within the hollow co-condensates. Using (\eta) and (\phi), we derived the third-order parameter, (\chi), which represents the dsDNA concentration within the hollow co-condensates. The free energy functional of this mesoscopic model was constructed following the classic Ohta–Kawasaki framework35,36. Detailed descriptions of the model development and numerical simulations are provided in the Supplementary Methods.

Figure 2a(i)–(ii) shows that dsDNA molecules containing GGAA microsatellites coat the surface of FUS-ERG condensates before slowly penetrating the interior. These observations were used to set the initial conditions of our model (Fig. 4a(i)), where the concentrations of protein and dsDNA are denoted as ({\zeta }_{{DNA}}) and ({\zeta }_{{protein}}), respectively. The diffusion-driven simulation, which considers only dsDNA diffusion, revealed a gradual inward migration of dsDNA molecules into the condensates (Fig. 4a(ii)), closely aligning with the experimental results shown in Fig. 2a(ii).

Fig. 4: A mesoscopic, molecularly-informed phase field model was developed to reproduce the hollow co-condensate formation.

a Simulations for FUS-ERG-DNA hollow co-condensate formation. (i) FUS-ERG molecules form biomolecular condensates with dsDNA containing GGAA microsatellites on their surface. dsDNA (({\zeta }_{{DNA}})) and protein (({\zeta }_{{protein}})) concentrations are represented. (ii) dsDNA molecules are transited into protein droplets, triggering the formation of hollow co-condensates. Order parameters denoting protein-DNA complex concentration ((\eta)), dsDNA concentration ((\chi)) and hydrophobic and hydrophilic distributions within condensates (({{{\rm{\phi }}}})) are represented. (iii) Hollow architecture is observed in steady states of the model proposed. DNA is coupled with hydrophobic region of the protein-DNA complex. b Distributions of simulated dsDNA concentration ((\chi)) within hollow architectures. c States diagram of simulated structures by varying initial states in a(i). Blue cross: One-phase, which is a homogeneous state; Orange dot: homogeneous condensates; Green circle: hollow co-condensates. d Simulations for PRM-RNA hollow co-condensate formation. (i) RNA first forms tadpole-like diblock copolymer with PRM protein. Order parameter denoting protein-RNA complex concentration ((\eta)), RNA concentration ((\chi)) and hydrophobic and hydrophilic distributions within copolymer ((\phi)) are represented. (ii) Hollow structure is observed in simulation results. RNA is coupled with both hydrophobic and hydrophilic region of protein-RNA complex. e Distributions of simulated RNA concentration ((\chi)) within hollow structures. Source data are provided as a Source Data file.

A central hypothesis of our model is that dsDNA, which binds to the hydrophobic ERG DBD domain (Fig. 3e(i)), preferentially localizes within hydrophobic regions of the condensate based on the Nile-red experiments (Supplementary Fig. 5b). We encoded this preference into the model’s dynamics by assuming dsDNA ((\chi)) preferentially diffuses within the hydrophobic regions of the hollow co-condensates in our case (Supplementary Methods). Simulations incorporating these interactions revealed the spontaneous emergence of a stable internal cavity, defined by a sharp interface in the hydrophobicity parameter (\phi) (Fig. 4a(iii) and Supplementary Movie 3). Additionally, the distribution of (\chi) showed that the simulated dsDNA concentration within the shell region of the condensates (Fig. 4b) closely aligned with the STED experimental data (Fig. 3f(ii)). By varying initial values of protein and DNA concentrations, we constructed a phase diagram (Fig. 4c) that closely matches the experimental observations (Fig. 1c).

To further validate our model, we tuned two key parameters (Supplementary Methods). The first parameter(,) ({\psi }_{C}), represents the critical volume fraction of the protein–dsDNA complex required for phase separation and is inversely related to protein-DNA binding affinity. In our primary simulation, setting ({\psi }_{C}) to (0.85) yielded hollow condensates. When ({\psi }_{C}), was increased from (0.85) to (0.95), which mimics a lower DNA-protein binding affinity, the formation of hollow condensates was not observed (Supplementary Fig. 9a and Supplementary Movie 4). Such prediction was experimentally confirmed by substituting dsDNA containing GGAA microsatellites with random dsDNA—which exhibits lower binding affinity— also prevented the formation of hollow co-condensate, corroborating the simulation results.

The second parameter, ({b}_{1}), quantifying the differential interaction of the protein–dsDNA complex’s hydrophilic versus hydrophobic domains with the aqueous buffer, was set to (0.14) in the main simulation (Fig. 4a). When ({b}_{1}) was decreased from (0.14) to (0.04), which indicates enhanced interaction between the hydrophobic component of protein-dsDNA complex and the working buffer compared to its hydrophilic region, the simulations revealed a failure to form hollow condensates (Supplementary Fig. 9b and Supplementary Movie 5). Notably, this result aligns with experiments using the GFP-FUS-ERG (9RA) (Fig. 1e(iv)), which exhibits increased hydrophobicity compared to wild-type FUS-ERG and fails to form hollow condensates (Supplementary Fig. 5a). This tight correspondence between model parameters and experimental perturbations substantiates our model’s physical basis.

We next investigated whether our model could simulate the formation of vesicle-like hollow condensates in a distinct PRM-RNA system17 (Fig. 3g) and sought the divergent molecular mechanisms between this system and FUS-ERG hollow condensates. It is known that RNA interacts with PRM to form a tadpole-like structure. In the RPM-RNA system oversaturated with RNA, one portion of RNA binds to PRM, generating hydrophobic regions, while the remaining RNA contributes to hydrophilic regions within the system. We therefore hypothesized that RNA ((\chi)) preferentially diffuses within both the hydrophobic and hydrophilic regions of the hollow co-condensates for such case (Supplementary Methods). Simulation with this hypothesis yielded the results shown in Fig. 4d(i)–(ii) and Supplementary Movie 6. Our simulation correctly reproduced the experimental phenotype: RNA localized to both the inner and outer surfaces of the hollow condensate (Fig. 4d-e), in contrast to the internal-surface-only localization of dsDNA in the FUS-ERG system. Such distribution closely matched the STED experimental data (Fig. 3g(ii)), revealing a distinct molecular mechanism underlying the formation of PRM-RNA hollow condensates.

Building on the ability of our mathematical model to accurately reproduce hollow co-condensate formation, we examined the outcome when FUS-ERG coexists with two dsDNA substrates differing in binding affinity. In simulations, the high- (red) and low- (blue) affinity dsDNA were assigned distinct binding affinity coefficients, and no interactions were allowed between the two dsDNA types. The results (Supplementary Methods) revealed that the low-affinity dsDNA was selectively excluded from the hollow co-condensates (Supplementary Fig. 10a). We validated this experimentally by forming hollow condensates with high-affinity GGAA-tagged dsDNA and then introducing non-specific dsDNA; as predicted, the non-specific dsDNA was excluded (Supplementary Fig. 10b, c). These findings highlight the selective nature of hollow co-condensates, driven primarily by protein–dsDNA binding affinity. This inherent property further inspired us to seek its potential applications in DNA storage systems, particularly for dynamic data manipulation.

FET fusion protein-DNA hollow co-condensates enabled dynamic data manipulation for DNA-based information system

In DNA storage, data is encoded in a library of short DNA fragments each bearing a unique barcode sequence indexing the encoded content for selective random access. To examine the potential selectivity of hollow co-condensate for specific d