A Molecular Grammar for Programmable Multiphase Protein–RNA Vesicles

Abstract

Protein–RNA phase separation gives rise to biomolecular condensates with rich internal organization, yet the molecular rules that connect sequence-encoded interactions to the emergent condensate spatial organization remain poorly understood. Here, using large-scale residue-level coarse-grained simulations, we identify a molecular grammar that governs the formation of multiphase protein–RNA condensates. We show that asymmetries in protein–protein and protein–RNA interactions, together with protein stoichiometry, chain length, and condensate density, collectively determine whether condensates adopt homogeneous, layered, biphasic, or vesicle-like morphologies. Vesicular condensates form spontaneously from well-mixed initial conditions without requiring flux-driven oversaturation or extreme charge imbalance, distinguishing this mechanism from previously proposed routes to condensate hollowing. We rationalize the full morphological progression as sequence-encoded amphiphile self-assembly: the protein–RNA complex behaves as a single-chain amphiphile whose effective packing parameter is set by Domain H/L stoichiometry, spanning micelles, cylinders, hollow vesicles, and inverse phases.

Introduction

Biomolecular condensates emerge from networks of weak, multivalent interactions that partition macromolecules into distinct mesoscale phases. − Their assembly and dissolution are highly responsive to changes in physicochemical conditions, including temperature, pH, pressure, and acute environmental stress, enabling cells to adapt and maintain homeostasis. ,

The internal organization of membraneless organelles is often far from uniform. , Stress granules, for example, are heterogeneous assemblies in which relatively stable, denser subregions are dispersed throughout a more dynamic and dilute matrix, an organization often described as “raisins in pudding” rather than a strict core–shell architecture. − In contrast, P granules in C. elegans embryos display two coexisting subphases with markedly different material properties: a less viscous, dynamically exchanging interior surrounded by a more viscous, slowly exchanging outer subphase, with this contrast in dynamics arising from preferential localization of distinct protein components to the two regions. , When anchored to the nuclear pore, C. elegans P granules further assemble into a characteristic “tripartite sandwich” structure. ,

RNA–protein mixtures composed of low-complexity peptides and RNA have emerged as powerful in vitro model systems for probing the formation of multiphase condensates. − In ternary mixtures containing a prion-like polypeptide (PLP), an arginine-rich polypeptide (RRP), and RNA, PLP and RNA compete for binding to RRP, driving RNA-induced demixing into coexisting homotypic PLP-rich droplets and heterotypic RRP–RNA condensates; depending on the RNA-to-RRP ratio and interaction strengths, the resulting biphasic structures adopt nonengulfing, partially engulfing, or fully engulfing morphologies. Recent work has further shown that mixtures of DNA, RNA, and proteins generate multiphasic structures driven by preferential RNA–protein interactions, with DNA localizing at the interface. Despite these advances, it remains unclear how sequence-encoded interactions and composition collectively select among distinct multiphase morphologies.

Although such behavior was historically associated with amphiphilic assemblies, recent studies have shown that RNA–protein mixtures can also form dynamic, membrane-free vesicles with well-defined hollow architectures under nonequilibrium conditions. − Similarly, hollow condensates can emerge through an abrupt dynamic transition induced by the addition of p21 DNA to protein–DNA condensates, and mixtures of cholesterol-modified DNA and histones exhibit three distinct phase states depending on composition: a homogeneous solution, coacervate droplets, and coacervate vesicles containing low-density internal cavities. Table S1 summarizes representative experimental studies of vesicular and multiphasic condensates.

However, a simple, general framework linking sequence-encoded interaction patterns and stoichiometry to the emergence of vesicular, layered, or biphasic condensates remains lacking. Here, we address this gap using a minimal coarse-grained model of protein–RNA condensates. We design a low-complexity protein composed of two domains, H and L, with distinct propensities to interact with RNA and with each other (Figure A), inspired by amphiphilic architectures that combine segments with contrasting affinities. A central organizing idea of this work is that, once Domain H engages RNA, the protein–RNA complex behaves as a single-chain amphiphile, with the charged H–RNA region serving as a structural “head” and Domain L as a “tail” whose chemistry can be tuned independently. This abstraction lets us anticipate, and later confirm, that varying the H/L stoichiometry traces out the classical hierarchy of amphiphile self-assembly morphologies: spherical micelles, cylinders, hollow vesicles, and inverse phases, predicted by the classical packing-parameter framework. By systematically tuning the relative strengths of Domain H–RNA, Domain L–RNA, and Domain L–L interactions, we define three classes of protein–RNA systems spanning regimes from weakly interacting auxiliary domains to strongly self-associating ones (Figure B). Within each regime, we further vary the stoichiometric ratio of Domains H and L along the protein chain, while matching RNA length to the RNA-binding domain. Large-scale, residue-level coarse-grained simulations reveal a rich spectrum of condensate morphologies, including hollow multiphase vesicles, multilayered droplets with compositionally distinct shells, and biphasic condensates with dense cores (Figure C). By analyzing domain-resolved radial distributions and effective interaction energies, we extract simple design rules that link molecular grammar, stoichiometry, density, and chain length to condensate architecture.

1.

(A) Schematic representation of the simulation setup. (B) Interaction matrix of domains. (C) Sequence space of the simulated condensates as a function of Domain L–L and Domain L–RNA interactions.

Methods

Residue-Resolution Coarse-Grained Model

We employ a residue-level coarse-grained (CG) model to study protein–RNA condensates. , This framework has been validated for IDP–nucleic acid phase behavior across sequence variation, chain length, and macromolecular crowding regimes. ,,

In this model, a harmonic potential is used for bonded interactions between adjacent beads,

$$U_{\mathrm{bond}}(r)=\frac{1}{2}{k}_{\mathrm{bond}}{(r-r_0)}^2$$ 1

with a bond force constant k _bond = 1000 kJ mol^–1 nm^–2. The equilibrium bond length is set to r ₀ = 0.5 nm for RNA and r ₀ = 0.38 nm for peptides.

Nonbonded interactions are modeled using the Ashbaugh–Hatch (AH) potential for short-range interactions, combined with a Debye–Hückel (DH) potential to account for screened electrostatics. The AH potential is given by

$$U_{ij}^{\mathrm{AH}}(r)=\{\begin{array}{ll}{U}^{\mathrm{LJ}}(r)-{\lambda }_{ij}{U}^{\mathrm{LJ}}(r_c)+\epsilon (1-{\lambda }_{ij}),& r\le 2^{1/6}{\sigma }_{ij}\\ {\lambda }_{ij}[U^{\mathrm{LJ}}(r)-U^{\mathrm{LJ}}(r_c)],& 2^{1/6}{\sigma }_{ij}<r\le r_c\\ 0,& r>r_c,\end{array}$$ 2

where U ^LJ(r) is the Lennard–Jones potential, σ _ij = (σ _i + σ _j )/2 is the arithmetic mean of bead diameters, and λ _ij = (λ _i + λ _j )/2 is the mean hydrophobicity parameter. When λ = 0, the AH potential reduces to a purely repulsive Weeks–Chandler–Andersen (WCA) interaction.

Amino acid bead diameters σ _i are assigned based on van der Waals radii, and protein hydrophobicity parameters λ _i are taken from the CALVADOS 2 force field. All nucleotides are represented by a single bead type with fixed diameter σ = 0.817 nm, hydrophobicity λ = −0.027, and unit negative charge. The λ values and charges of all residues used in our simulations are listed in Table S2.

Electrostatic interactions are modeled using the Debye–Hückel potential,

$$U_{ij}^{\mathrm{DH}}(r)=\frac{q_i{q}_j{e}^2}{4\pi \epsilon }\frac{\exp (-\kappa r)}{r}$$ 3

where q _i and q _j are the bead charges. The inverse Debye screening length κ is determined from the ionic strength I as

$$\kappa =\sqrt{\frac{2e^2{N}_{\mathrm{A}}I}{{\epsilon }_0{\epsilon }_r{k}_{\mathrm{B}}T}}$$ 4

All simulations were performed at an ionic strength of 100 mM, corresponding to a screening length κ^–1 ≈ 0.96 nm at T = 300 K. The temperature dependence of the dielectric constant was incorporated using the empirical relation

$$\frac{\epsilon (T)}{{\epsilon }_0}=\frac{5321}{T}+233.76-0.9297T+1.417\times {10}^{-3}{T}^2-8.292\times {10}^{-7}{T}^3$$

Molecular Simulations

Initial configurations were generated by randomly placing protein and RNA chains in cubic simulation boxes using Packmol. The systems were subsequently energy-minimized and propagated in the NVT ensemble using a Langevin thermostat at T = 290 K, with a friction coefficient γ = 0.01 and a time step of 0.01 ps on graphics processing units (GPUs). All simulations were performed at an ionic strength of 100 mM.

The simulation length was system-dependent and chosen to ensure adequate equilibration, convergence, and the formation of stable condensate morphologies. In most cases, trajectories exceeded 2.5 μs, while certain systems required longer simulation times to achieve stable structural and dynamical properties. Convergence of the simulations was evaluated by monitoring the time evolution of the radial distribution functions (RDFs) of the condensate components relative to the condensate center of mass (COM). The persistence and stability of the RDF profiles over extended simulation times were used as indicators of equilibration and convergence (Figure S1).

The cubic simulation box lengths ranged from 120 to 160 nm and contained between 800 and 1600 chains of each component. A complete summary of all simulation parameters is provided in Table S4.

Analysis

To characterize the spatial organization of proteins and RNA within condensates, we computed radial distribution functions (RDFs) of selected molecular species with respect to the condensate center of mass (COM). The RDF is denoted g _αβ (r), where α is the condensate COM reference and β is the molecular species of interest. In our analysis, β corresponds either to Domain H residues together with RNA beads, or to Domain L residues. Here, r is the radial distance from the condensate COM.

$$g_{\alpha \beta }(r)=\frac{⟨N_{\beta }(r)⟩}{{\rho }_{\beta }{V}_{\mathrm{shell}}(r)}$$

where ⟨N _β (r)⟩ is the average number of particles of species β within a spherical shell at distance r from the COM, ρ _β is the bulk density of component β, and V _shell(r) is the shell volume between adjacent radial bins. Statistical uncertainties in the RDFs were estimated by block-averaging over independent, equilibrated trajectory segments and reporting the standard deviation across blocks. The resulting error bars are very small owing to the large system size and extensive sampling, indicating that the RDFs are well converged. From each RDF we define a local, distance-resolved potential of mean force:

$$W_{\alpha \beta }(r)=-k_{\mathrm{B}}T\ln g_{\alpha \beta }(r)$$ 5

To compress this distance-resolved profile into a single scalar that can be compared across systems and stoichiometries, we compute an RDF-weighted mean of W _αβ (r) over a finite radial window [r _min, r _max] that brackets the principal correlation shells of the condensate,

$${\mathrm{W̅}}_{\alpha \beta }=\frac{{\int }_{r_{\min }}^{r_{\max }}{W}_{\alpha \beta }(r)g_{\alpha \beta }(r)dr}{{\int }_{r_{\min }}^{r_{\max }}{g}_{\alpha \beta }(r)dr}$$ 6

The lower bound r _min excludes the depleted innermost region (where g _αβ → 0 would dominate the logarithm), and r _max is set to the outer condensate radius; identical limits are used when comparing different systems. We emphasize that W̅ _αβ is an effective pair-interaction descriptor: a g-weighted average of the local PMF that yields an energy-valued scalar summarizing the strength of the αβ correlation within the condensate. By construction, its sign tracks the sign of the local pair enrichment in the integration window: g _αβ (r) > 1 over the window yields W̅ _αβ < 0, while net depletion (g _αβ (r) < 1) yields W̅ _αβ > 0. Comparisons of W̅ _αβ across systems are therefore meaningful only when the same species pair (α, β) and the same integration window [r _min, r _max] are used, and the reported trends should be interpreted as a comparative ranking of effective pair correlation strength rather than as a free-energy landscape of assembly. This usage parallels recent sequence-based effective-interaction frameworks for disordered regions.

Local interaction patterns were further quantified by computing the average number of RNA or protein neighbors within a cutoff distance of 11 Å around each protein residue. These contact numbers were averaged over all residues and chains, providing a direct measure of local protein–RNA and protein–protein interaction strengths across different systems. All other structural analyses have been carried out using the MDAnalysis package.

Results

Protein Sequence Grammar Dictates Multiphase Condensate Architecture

Protein composition plays a central role in determining the internal organization of multiphase condensates. ,− For instance, mixtures of two distinct proteins and RNA can give rise to rich multiphase architectures as a function of protein composition. ,, To dissect how specific protein–RNA and protein–protein interactions encode the emergent multiphase architectures of condensates, we engineer a low-complexity protein composed of two distinct regions: Domain H (high RNA affinity) and Domain L (low RNA affinity). The interaction matrices for the sequences in each system are summarized in Table S3. While both domain composition and stoichiometry influence multiphasic morphology, here we focus exclusively on the role of protein grammar, defined by the amino acid composition of Domains H and L.

By design, Domain H–L interactions are kept comparatively weak while Domain H–RNA interactions are strongly attractive (Figure B). We systematically tune Domain L–RNA and L–L interaction strengths by altering the amino acid composition of Domain L (Figure C).

Based on these interaction patterns, we define three classes of protein–RNA systems. System A is characterized by weak Domain L–L interactions, with Domain L–RNA interactions ranging from repulsive to strongly attractive (Figure C). System B is characterized by attractive Domain L–L interactions combined with weak Domain L–RNA affinity (Figure C). System C is characterized by strong Domain L–L interactions with Domain L–RNA interactions that are weaker than those of Domain H–RNA (Figure C). For each system, we select amino acid compositions for Domain H and L that realize the targeted interaction profiles (Table S2). We then perform long-time scale molecular dynamics simulations using the CALVADOS-2 single-bead coarse-grained model. Simulation details, including chain numbers and ionic conditions, are summarized in Table S4.

In our notation, protein sequences are written as A _m B _n , where the first term A _m denotes Domain H composed of m residues of amino acid A, and the second term B _n denotes Domain L composed of n residues of amino acid B. In System A, Domain L–L interactions are designed to be repulsive or weak, while Domain L–RNA interactions span a range from repulsive to weak to attractive (Figure B). To realize this interaction landscape, we construct four protein compositions: (i) R ₁₅₀(KD)₂₅, in which Domain L contains alternating lysine (K, positively charged) and aspartic acid (D, negatively charged) residues; although the full protein remains net positive and therefore associates favorably with RNA overall, the embedded aspartic acid residues introduce local unfavorable Coulombic interactions both with negatively charged RNA and between neighboring Domain L segments; (ii) R ₁₅₀(K ₃ A ₂)₁₀, where Domain L–L interactions remain electrostatically repulsive because of lysine (K), while Domain L–RNA interactions become attractive due to alanine (A); (iii) R ₁₅₀ K ₃, which exhibits repulsive Domain L–L interactions and attractive Domain L–RNA interactions, albeit weaker than those of Domain H–RNA, such that RNA preferentially associates with Domain H; and (iv) R ₁₅₀(N ₃ G ₂)₁₀, in which both Domain L–L and Domain L–RNA interactions are uniformly weak.

Across all four compositions in System A, we observe the robust formation of multiphasic vesicles with a characteristic architecture: an empty core, an inner Domain L–rich shell, a middle Domain H–RNA shell, and an outer Domain L–rich shell, as resolved by the radial distribution functions about the condensate center of mass (Figure B–E). To our knowledge, this specific multiphase vesicle architecture has not been reported previously, and Domain L segments together with Domain H–RNA complexes act as composite monomers analogous to amphiphilic molecules driving liposome formation (Figure A).

2.

Structural characterization of multiphase vesicles in System A. (A–D) Top: vesicle slice snapshots; middle: corresponding average number of contacts surrounding the protein, including protein–protein (denoted as PRO) and protein–RNA contacts; bottom: corresponding radial distribution functions (RDFs) of Domain L and Domain H–RNA complexes around the condensate center of mass for four sequences: (A) R₁₅₀(KD)₂₅ with U₁₅₀, (B) R₁₅₀(K₃A₂)₁₀ with U₁₄₀, (C) R₁₅₀K₅₀ with U₁₃₅, and (D) R₁₅₀(N₃G₂)₁₀ with U₁₅₀. (E) Size dependence of vesicle formation for the R₁₅₀(N₃G₂)₁₀ with U₁₅₀ system. Left: snapshot of cross section and slice of a vesicle formed by 1600 chains of each component, showing a pronounced hollow core. Right: RDF around the center of mass (COM) of the condensate as a function of the number of chains, demonstrating that the core size increases with system size. Color scheme: green denotes arginine (R) of Domain H; pink denotes uracil (U); red denotes lysine (K) or asparagine (N); and orange denotes glycine (G) or alanine (A). Red and orange residues together constitute Domain L.

This interaction hierarchy is quantified by the average number of neighboring RNA or protein residues within 11 Å of each protein residue. It differentiates domain-specific protein–protein and protein–RNA interaction patterns across compositions spanning repulsive to attractive regimes (Figure A–D).

To assess the robustness of vesicle formation, we further simulate the R ₁₅₀(N ₃ G ₂)₁₀ system across a range of system sizes by varying the number of chains from 800 to 1600 (corresponding to 281,600–563,200 particles) and adjusting the simulation box dimensions accordingly (Figure E; Figure S2; Table S2). In all cases, stable vesicles form, with the size of the empty core increasing systematically with system size (Figure E). Due to the substantial computational cost, system-size scaling was examined only for this representative composition. Additionally, we have calculated the average number of protein–protein and protein–RNA contacts formed per protein chain across all four system sizes (Figure S3). These quantities are nearly identical irrespective of system size, indicating that our observations are robust and are not significantly affected by finite-size effects.

RNA concentration plays a decisive role when Domain L interacts attractively with RNA. At U ₁₅₀, the R ₁₅₀(KD)₂₅ and R ₁₅₀(N ₃ G ₂)₁₀ systems readily form multiphasic vesicles, whereas the R ₁₅₀(K ₃ A ₂)₁₀ and R ₁₅₀ K ₅₀ systems produce nonvesicular condensates under identical conditions (Figure A; Figure S4). Reducing the RNA content to U ₁₃₅ for R ₁₅₀(K ₃ A ₂)₁₀ and to U ₁₄₀ for R ₁₅₀ K ₅₀ induces vesicle formation in both cases (Figure B,C; Figure S2). When Domain L–RNA interactions are attractive, limiting RNA availability enhances competition between protein–protein and protein–RNA interactions, promoting vesicular architectures.

3.

Influence of RNA concentration on vesicle formation in the R₁₅₀K₅₀ sequence. (A–C) Top: condensate slice snapshots; middle: corresponding average number of contacts surrounding the protein components; bottom: radial distribution functions (RDFs) around the center of mass (COM) for Domain L and Domain H–RNA complexes, of R₁₅₀K₅₀ with (A) U₁₅₀ (B) U₁₄₀, and (C) U₁₃₅.

Consistent with this interpretation, contact analysis at U ₁₅₀ reveals diminished preferential binding of Domain H to RNA relative to Domain L in the R ₁₅₀(K ₃ A ₂)₁₀ and R ₁₅₀ K ₅₀ systems. Upon RNA reduction, strong Domain H–RNA preference is restored (Figure A–C). Together, these trends show that within System A, vesicle formation is driven either by strong repulsion in both Domain L–L and Domain L–RNA interactions or by attractive Domain L–RNA interactions under conditions of limited RNA availability.

To transition from System A to System B, we introduce attractive Domain L–L interactions while maintaining weak Domain L–RNA affinity. Specifically, we replace asparagine (N) with glutamine (Q) in Domain L, yielding the composition R ₆₀(Q ₃ G ₂)₂₈, which promotes favorable Domain L–L contacts. Contact analysis confirms the emergence of attractive Domain L–L interactions, as reflected by an increased number of neighboring Domain L residues within 11 Å of each protein (Figure ).

4.

Structural characterization of multiphase condensates in Systems B and C. (A–C) Top: cross sections of the condensates; middle: corresponding average number of contacts surrounding the protein components; bottom: corresponding radial distribution functions (RDFs) of Domain L and Domain H–RNA complexes around the condensate center of mass for System B and System C: (A) System B: R₆₀(Q₃G₂)₂₈ with U₆₀, (B) System C: R₄₀(F₃G₂)₃₂ with U₄₀, and (C) System C: K₄₀(F₃G₂)₃₂ with U₄₀. Color scheme: green denotes arginine (R), pink denotes uracil (U), red represents glutamine (Q) or lysine (K), cyan represents phenylalanine (F), and orange represents glycine (G).

In System B, we observe a distinct, layered, multiphase architecture: a Domain L–rich core, a shell of Domain H–RNA complexes, and an outer Domain L–rich layer (Figure A). The central core exhibits lower density than the surrounding layers, consistent with a filled, loosely packed interior rather than a hollow lumen. Attractive Domain L–L interactions thus drive core formation, altering the condensate architecture relative to System A.

In System C, where Domain L–L interactions are strongly attractive and Domain L–RNA interactions are also attractive but weaker than Domain H–RNA affinity, we examine two compositions: _K40(F ₃ G ₂)₃₂ and R ₄₀(F ₃ G ₂)₃₂. Within the CALVADOS-2 model, phenylalanine–phenylalanine interactions are strongly attractive, as reflected by elevated numbers of neighboring Domain L residues within 11 Å of each protein (Figure B,C). In both systems, this interaction hierarchy produces biphasic condensates composed of a dense Domain L–rich core surrounded by an outer shell in which Domain H (lysine- or arginine-rich) is complexed with RNA, as evidenced by their RDF profiles (Figure B,C).

Stoichiometry of Protein Governs Multiphase Condensate Formation

The progression from uniform droplets to multiphase condensates and hollow vesicles is driven not only by the “protein grammar of domains” that encodes interaction motifs, but also by the stoichiometric ratio of Domains H and L within a protein chain. To examine this, we systematically vary the H:L ratio in 200-residue proteins while matching RNA length to that of Domain H, across all three model systems (Systems A, B, and C).

Starting with System A, we use the baseline composition R _m (N₃G₂) _n with U _m , in which both Domain L–L and Domain L–RNA interactions are weak. By varying m and n, we generate nine distinct protein sequences (Figure ). In all cases, the RNA length is matched to the number of arginine residues in Domain H.

5.

Stoichiometric evolution of condensate morphology in System A. (A) Representative simulation snapshots illustrating how condensate structure changes with varying ratios of Domain H (R _m ) to Domain L ((N ₃ G ₂) _n ). The sequences (from left to right) correspond to R _m (N₃G₂)n with (m, n) values of (20, 36), (50, 50), (75, 25), (100, 20), (110, 18), (120, 16), (135, 13), (150, 10), and (180, 4). (B) Radial distribution functions (RDFs) of Domain L and Domain H–RNA around the condensate center of mass, comparing partially formed vesicle sequences (top) and fully developed vesicle sequences (bottom). (C) RDF-weighted mean local-PMF descriptor W̅ (eq ) as a function of the Domain H–to–Domain L ratio. More negative values of W̅ indicate stronger local pair enrichment within the integration window and correlate with the vesicle-forming compositions. Color scheme: green denotes arginine (R) (Domain H), pink denotes uracil (U), red denotes asparagine (N), and orange denotes glycine (G); red and orange residues collectively form Domain L.

The internal organization of these condensates evolves systematically with m, i.e., with the H:L ratio of the chain. Sequences with the shortest Domain H (lowest m) fail to condense, reflecting insufficient favorable interactions to drive assembly (Figure A).

As the Domain H content (R _m ) increases, the system forms multiple micelle-like droplets characterized by a Domain H–RNA-rich core surrounded by a Domain L ((N₃G₂) _n ) shell (Figure A). Further increases yield doughnut-shaped condensates that preserve this core–shell organization, and then partially formed vesicles bearing multiple openings – a morphology that, to our knowledge, has not been previously reported (Figure A). Radial distribution functions about the condensate center of mass confirm this progression, revealing incomplete, vesicle-like multiphase structures in this regime (Figure B).

At still higher Domain H content, the system transitions to fully developed multiphasic vesicles featuring a hollow core, an inner Domain L shell, a middle layer of Domain H–RNA complexes, and an outer Domain L shell. This multilayered architecture is clearly reflected in the radial distribution functions, which show well-resolved, concentric domain-specific layers surrounding an empty core (Figure B).

At the highest Domain H ratios examined, the vesicular structure is lost. The system instead forms a multiphase condensate with a Domain H–RNA-rich interior surrounded by a peripheral Domain L ((N₃G₂) _n ) layer (Figure A).

To compress the distance-resolved local PMF (eq ) into a single scalar, we evaluate the RDF-weighted mean descriptor W̅ defined in eq . As shown in Figure C, W̅ varies systematically with the Domain H–to–Domain L ratio: it becomes most negative within the vesicle-forming stoichiometric window – reflecting strong local pair enrichment of Domain H–RNA contacts in this regime – whereas it remains positive for the lowest-Domain H sequences, consistent with the absence of clustering or vesicle formation under those conditions.

In System B, where Domain L–L, Domain L–RNA, and Domain H–RNA interactions are all attractive, the baseline composition is R _m (Q₃G₂) _n with U _m . We generate six protein sequences by varying m and n (Figure A). In contrast to System A, the multiple favorable interactions yield stable condensates across all stoichiometries.

6.

Stoichiometric evolution of condensate morphology in System B. (A) Representative simulation snapshots showing condensate evolution with the ratio of Domain H (R _m ) to Domain L ((Q₃G₂) _n ) in System B. The sequences shown (from left to right) correspond to R _m (Q₃G₂) _n with (m,n) values of (50, 30), (60, 28), (120, 16), (135, 13), and (150, 10). (B) Radial distribution functions (RDFs) of Domain L and Domain H–RNA complexes around the condensate center of mass. (C) RDF-weighted mean local-PMF descriptor W̅ (eq ) as a function of the Domain H–to–Domain L ratio, used to compare the relative strength of local pair correlation across compositions that produce layered architectures. Color scheme: green denotes arginine (R) of Domain H, pink denotes uracil (U), red represents glutamine (Q), and orange represents glycine (G); red and orange residues collectively form Domain L.

The internal organization of these condensates evolves systematically with the H:L ratio. At low Domain H content, condensates adopt a multilayered architecture consisting of a Domain L–rich core, a Domain H–RNA-enriched intermediate layer, and an outer Domain L ((Q₃G₂) _n ) shell (Figure A). Importantly, the intermediate Domain H–RNA layer does not fully encapsulate the Domain L–rich core, as reflected in the radial distribution functions (Figure A,B).

As Domain H content (R _m ) increases, the intermediate Domain H–RNA layer becomes thicker, more continuous, and more compositionally enriched, ultimately yielding a fully developed multilayered structure in which Domain H–RNA entirely covers the Domain L–rich interior (Figure A,B). With further increases in Domain H, the Domain L–rich regions fragment into smaller, irregular clusters embedded within a continuous Domain H–RNA matrix, while a thin Domain L–rich rim persists at the condensate periphery (Figure A).

At the highest Domain H ratios simulated, Domain L–rich regions are greatly diminished and appear only as small puncta dispersed throughout the Domain H–RNA network, with the outer surface retaining only a faint Domain L enrichment (Figure A,B).

As in System A, we evaluate W̅ (eq ) as a function of the Domain H–to–Domain L ratio. Unlike System A, W̅ remains negative across all compositions, consistent with the persistent presence of favorable pair correlations in every condensate. The multilayered, vesicle-forming compositions exhibit the most negative W̅, indicating the strongest local pair enrichment among the layered architectures sampled (Figure C).

Finally, in System C with baseline composition R _m (F₃G₂) _n and U _m , we constructed five protein sequences by varying m and n (Figure A). As in System B, Domain L–L, Domain L–RNA, and Domain H–RNA interactions are all attractive, supporting condensate formation across all stoichiometries.

7.

Stoichiometric evolution of condensate morphology in System C. (A) Representative simulation snapshots illustrating condensate morphology at different stoichiometries. The sequences shown (from left to right) correspond to R _m (F₃G₂) _n with (m, n) values of (30, 30), (40, 32), (50, 30), (65, 29), and (80, 24). (B) Radial distribution functions (RDFs) of Domain L and Domain H–RNA around the condensate center of mass, demonstrating progressive changes in internal structure with increasing Domain H content. (C) RDF-weighted mean local-PMF descriptor W̅ (eq ) as a function of the Domain H–to–Domain L ratio, used to rank the relative strength of local pair correlation across compositions that produce layered architectures. Color scheme: green denotes arginine (R), pink denotes uracil (U), red denotes glutamine (Q) or lysine (K), cyan denotes phenylalanine (F), and orange denotes glycine (G); the latter three collectively form Domain L.

At the lowest Domain H content, condensates adopt a multilayered architecture consisting of a Domain H–RNA-rich core, a Domain L intermediate layer, and an outer Domain H–RNA-rich shell (Figure A,B). This arrangement resembles the multilayered organization of System B, but with the layering order reversed.

As Domain H content increases, the system transitions to a biphasic condensate consisting of a Domain L–rich core enveloped by a continuous Domain H–RNA shell, a structure clearly reflected in the radial distribution functions (Figure A,B). Further enrichment of Domain H drives penetration of Domain H–RNA complexes into the central Domain L–rich region, producing a partially mixed core (Figure A,B).

At the highest Domain H ratios examined, condensates become predominantly Domain H–RNA-rich, with only small, isolated Domain L–rich clusters dispersed within the continuous Domain H–RNA matrix (Figure A). This progression illustrates how increasing Domain H abundance transforms the condensate from a Domain L-centered, layered architecture into one dominated by Domain H–RNA interactions, with Domain L–rich components reduced to scattered inclusions.

As in Systems A and B, we evaluate W̅ (eq ) as a function of the Domain H–to–Domain L ratio. As in System B, W̅ remains negative across all compositions, consistent with the persistent presence of favorable pair correlations in every condensate. The multilayered compositions exhibit the most negative W̅, indicating the strongest local pair enrichment among the architectures sampled (Figure C).

Overall, both protein grammar and Domain H:L stoichiometry act in concert to govern condensate morphology, enabling the fine-tuning of multiphase architectures.

Influence of Density and Chain Length on Multiphase Vesicle Formation

Density and chain length govern both the kinetics and thermodynamics of liquid–liquid phase separation (LLPS). ,,− In simulations, the density is chosen pragmatically: below the typical condensed-phase density (∼500 mg/mL) but well above experimental dilute-phase densities (∼1 mg/mL), ensuring rapid droplet formation (<500 ns) and tractable sampling.

To evaluate the influence of biomolecular density and to test the robustness of our findings, we performed simulations at multiple densities using R₁₃₅(N₃G₂)₁₃ with U₁₃₅ (800 chains each, 1600 total), which forms a vesicle-like condensate. We follow a protocol adapted from lipid-vesicle molecular dynamics simulations: − biomolecules are uniformly placed in the simulation box using Packmol, briefly equilibrated under repulsive forces to ensure a well-mixed starting state, after which the box is expanded by 25 nm in all dimensions to suppress periodic-image interactions that would otherwise favor lamellar phases over closed vesicles.

We constructed nine simulation systems with box lengths from 115 nm (60 mg/mL) to 135 nm (37.1 mg/mL) and ran long-time scale simulations for each. Vesicle formation is observed under nearly all conditions (Figure S5). The rate of vesicle formation depends strongly on density: higher-density systems coalesce rapidly into a single large condensate that subsequently reorganizes into a vesicle, whereas lower-density systems proceed through sequential assembly intermediates, indicating a more gradual, hierarchical pathway.

To quantify density-dependent effects, we analyze the temporal evolution of the number of clusters, the largest cluster size, and the radius of gyration of the largest cluster across all systems (Figure B and Figure S6A). At higher densities, these metrics exhibit rapid convergence: the number of clusters drops sharply as chains quickly merge into a dominant aggregate, whose radius of gyration stabilizes before reorganizing into a hollow vesicle-like structure.

8.

Temporal evolution of vesicle formation in the R₁₃₅(N₃G₂)₁₃ system with U₁₃₅. (A) Representative simulation snapshots depicting the sequential structural transitions observed over time (time unit = 10 fs). The system initially forms tadpole-like morphologies, followed by tadpoles with double tails, micelles, and percolated network-like structures. These intermediates subsequently reorganize into partially opened vesicle-like assemblies and eventually mature into fully developed vesicles. (B) Quantification of density-dependent effects through the temporal evolution of key structural metrics, including the number of clusters and the cluster size distribution of the largest cluster across all systems. (C) Effect of chain length on biomolecular condensate formation: quantification of length-dependent effects based on the temporal evolution of key structural metrics, including the number of clusters and the cluster size distribution of the largest cluster across all systems.

At lower densities, RNA molecules preferentially associate with the arginine-rich peptides (Domain H), generating tadpole-like complexes rather than immediately forming a cohesive droplet. These elongated tadpole complexes subsequently coalesce into micelle-like assemblies, which later merge into a percolated network. Over time, these networks evolve into condensates containing highly open, cavity-like regions, which progressively reorganize and close to yield a fully formed multiphase vesicle (Figure A and Figure S5).

To further probe the role of chain length in LLPS and vesicle formation, we examined the same R₁₃₅(N₃G₂)₁₃ system with three RNA chain lengths – U₁₃₅, U₃₀₀, and U₄₀₀ – while keeping the total number of RNA residues constant in a 135 nm box (37.1 mg/mL).

Increasing RNA chain length strongly promotes condensation: longer chains reduce the conformational-entropy penalty associated with binding and enhance multivalent interactions, accelerating nucleation and the progression toward vesicles (Figure C and Figure S6B). The number of clusters decreases more rapidly with longer chains, while the largest cluster grows faster, both reflecting earlier coalescence into a single dominant condensate.

Discussion

Multiphasic condensates arise when distinct classes of intermolecular interactions drive components to partially demix rather than form compositionally uniform droplets. Such compositional structuring, including biphasic, multilayered, or vesicle-like organization, is increasingly recognized as a key determinant of functional specialization in membraneless organelles. − Protein–RNA mixtures are particularly rich examples, yet the physical rules that map sequence-encoded interaction patterns and stoichiometry onto specific internal architectures remain incompletely understood. Using a minimal two-domain protein model and coarse-grained simulations, we identify simple principles governing the formation and stability of multiphase condensates across a broad range of interaction strengths and compositions.

A unifying statistical-mechanical interpretation of these results is that, once Domain H engages RNA, each chain–RNA complex behaves as a sequence-encoded amphiphile, with the charged, RNA-bridged H region serving as a structural “head” and Domain L as a “tail” whose effective solvent quality is set by the chosen L–L and L–RNA interactions. The morphological cascade we observe along the System A stoichiometry sweep – progressing from noncondensed states through spherical micelles, cylindrical and doughnut-shaped clusters, percolated networks, and hollow multiphase vesicles to inverse, biphasic cores, is precisely the morphological hierarchy predicted by the classical packing-parameter framework as the head-to-tail volume ratio is increased, and parallels the spherical, cylindrical, lamellar, and inverse phase sequence familiar from block copolymer microphase separation. Within this picture, the architectural distinctions between Systems A, B, and C correspond to three tail-solvency regimes: athermal (A), weakly cohesive (B), and strongly cohesive (C) Domain L. Stronger L–L cohesion progressively fills the lumen, converting hollow vesicles into multilayered droplets and finally into biphasic, “filled” inverse architectures. Crucially, in this framework the symmetry-breaking responsible for the hollow-condensate formation is encoded at the single-chain level by the H/L block architecture, rather than at the bulk-composition level by charge or stoichiometric imbalance, which explains why the present vesicles emerge spontaneously from charge-balanced, well-mixed initial conditions, in contrast to influx-driven or reentrant-coacervation routes. ,, The minima in W̅ that coincide with vesicle-forming compositions are then naturally read as a measure of how cleanly the bilayer-like midplane is resolved within this self-assembled geometry. We note, however, that the present simulations identify the topology of the assembled state but do not quantify its size selection: the equilibrium balance between bilayer bending energy and edge tension, which dictates the closed-vesicle vs planar-bilayer competition and the preferred vesicle radius, is set by molecular elastic constants we have not measured here, and represents a natural target for follow-up work.

Our simulations examine three interaction regimes that differ in the relative strengths of protein–RNA and protein–protein interactions. In System A, both Domain L–L and Domain L–RNA interactions are weak or repulsive, while Domain H–RNA affinity remains strong. This regime yields robustly multiphase vesicles composed of an empty core, a Domain L-rich inner layer, a Domain H–RNA intermediate layer, and a Domain L-rich outer shell. This spatial organization is qualitatively consistent with experimental observations of ternary mixtures of RNA, lysine, and arginine, in which arginine preferentially associates with RNA to form a dense core while lysine is enriched in the outer regions. , These vesicles form spontaneously from well-mixed initial conditions and do not require strong charge imbalance or influx-driven oversaturation, distinguishing them from previously studied hollow condensates. Their stability arises from a balance between strong Domain H–RNA binding and weak cohesion among Domain L residues.

In System B, moderate Domain L–L cohesion is introduced while Domain L–RNA interactions remain weak. The dominant architecture shifts to multilayered droplets with a Domain L-enriched core, an intermediate Domain H–RNA layer, and a secondary Domain L-rich outer shell. Increasing the fraction of Domain H drives partially covered multilayers into fully developed multilayers, then into dispersed Domain L-rich microdomains within a Domain H–RNA network, and eventually into nearly homogeneous Domain L–RNA–dominated droplets.

In System C, strong Domain L–L interactions dominate over moderate Domain L–RNA affinity, while Domain H–RNA binding remains strongest. The condensates adopt biphasic architectures in which a dense Domain L-rich core is surrounded by a Domain H–RNA shell, with the layer ordering reversed relative to System B due to the much stronger cohesion among Domain L residues. Increasing the abundance of Domain H gradually disrupts the compact core, yielding scattered Domain L-rich inclusions within a continuous Domain H–RNA phase.

Across Systems A–C, varying the stoichiometric ratio of Domains H and L modulates the internal organization in predictable ways. For each system, the stoichiometries that yield vesicles or well-defined multilayers coincide with the most negative values of W̅ (eq ), indicating that these morphologies are accompanied by the strongest local pair enrichment within the condensate. The observed transitions, from micelles to incomplete vesicles to closed vesicles in System A, or from layered structures to fragmented microdomains in Systems B and C, are consistent with classical behavior in multicomponent polymer–amphiphilic systems but arise here solely from competing protein–RNA interactions.

Additionally, we anticipate that sequences containing multiple interspersed H and L domains, rather than the two-block architectures considered in this study, could significantly influence condensate organization and morphology. In such systems, condensate behavior would likely depend on additional factors, including block length, the number of repeats, and the relative strengths of intermolecular interactions. A systematic investigation of these parameters is beyond the scope of the present work and would require a dedicated future study.

We also investigate the influence of protein–RNA density and RNA chain length. Increasing density accelerates chain coalescence and leads to faster vesicle closure, whereas lower densities reveal sequential intermediates, such as tadpoles, micelles, and network-like aggregates. Similarly, increasing RNA chain length enhances multivalent binding and speeds the emergence of the final condensate architecture. Importantly, these perturbations influence kinetics but not the ultimate structure: the vesicle, multilayered, or biphasic architectures are robust to changes in density and chain length, further highlighting that morphology is governed primarily by interaction grammar and stoichiometry.

A disparity in RNA-binding affinity between the H and L domains is essential to this picture: if both domains possessed similarly high affinity for RNA, the contrast that drives multiphasic organization would be lost, and the system would exhibit homogeneous mixing into a single-phase condensate without clear spatial segregation between domains.

The structural motifs identified here connect naturally to experimental observations. Reconstituted protein–RNA systems frequently display competitive binding, demixing, and layered organization, and our results suggest how specific sequence alterations, such as increasing aromatic content in Domain L or modifying charge patterning, could shift condensates between vesicular, multilayered, or biphasic states. The prediction that vesicles can form spontaneously under charge-balanced conditions is particularly notable, as most reported hollow condensates arise from influx-driven or oversaturated states. Matching the predicted stoichiometric windows or manipulating RNA chain length offers straightforward routes for experimental validation.

Beyond biological relevance, the vesicle-like structures observed here resemble amphiphilic assemblies despite being composed entirely of intrinsically disordered protein segments and RNA. , This points to opportunities for engineering membrane-free vesicles with programmable permeability, encapsulation, and responsiveness, with potential applications in synthetic organelles, microreactors, and stimulus-controlled delivery systems. −

Together, these findings provide a mechanistic map linking sequence grammar, stoichiometry, and intermolecular competition to the emergence of multiphasic protein–RNA condensates, with direct implications for in vitro reconstitution and the design of programmable biomolecular materials.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/jacsau.6c00322.

• Additional simulation snapshots, radial distribution functions, cluster analysis metrics, and complete tables of sequence compositions (Table S1) and simulation parameters (Table S2) (PDF)

The authors declare no competing financial interest.