Sequence-encoded and Composition-dependent Protein-RNA Interactions Control Multiphasic Condensate Topologies

Multivalent protein-protein and protein-RNA interactions are the drivers of biological phase separation. Bio-condensates typically contain a dense network of multiple proteins and RNAs, yet the role of overlapping molecular interactions in regulating the condensate composition and structure is not well understood. Employing a ternary system comprising of a prion-like polypeptide (PLD), arginine-rich polypeptide (RLD), and RNA, here we show that competition between the PLD and RNA for a single shared partner, the RLD, leads to PLD–RLD demixing and spontaneous formation of biphasic condensates. Combining experiments with simulations, we show that the topology of coexisting condensates is regulated via mixture composition and the nature of protein-protein and protein-RNA interactions, giving rise to a diverse set of multiphasic patterns including completely separated, partially and completely engulfed droplet morphologies, and Janus droplets. Our findings provide a minimal set of physical rules that govern the composition and spatial organization of multicomponent and multiphasic bio-condensates.

In biological cells, many multivalent ribonucleoproteins (RNPs) form biomolecular condensates (BMCs) that act as active or repressive hubs for intracellular storage and signaling 1,2 . These condensates can rapidly assemble and dissolve in response to cellular stimuli via a physical process known as liquid-liquid phase transition [3][4][5] . The functional specificity of BMCs as subcellular organelles is linked to selective enrichment of specific enzymes/signaling factors within 1,6 , whereas altered compositions of signaling condensates are associated with disease pathologies 2,4,[7][8][9] . Mounting evidence now suggest that the spatial organization of biomolecules into distinct sub-compartments within BMCs [e.g., nuclei 10,11 , nuclear speckles 12 , paraspeckles 13 and stress granules 14 ] adds another layer of internal regulation of composition and plays a fundamental role in facilitating their complex biological functions. These mesoscopic multilayered structures can be qualitatively understood on the basis of a multi-phasic condensate model, where two or more distinct types of partially immiscible condensed phases are formed by spontaneous phase separation of individual components in a multi-component mixture 11,15,16 . However, due to the complexities generated by the presence of a large number of distinct protein and RNA components in a typical intracellular BMC, the underlying molecular mechanisms that regulate the multiphasic condensate composition and spatial organization remain largely unexplored.
The composition and spatial organization of BMCs are ultimately controlled by the nature of intermolecular interactions between RNPs and RNAs as well as their interactions with the solvent molecules 2,11,17,18 . Analysis of sequence features of eukaryotic BMC proteins revealed that intrinsically disordered low complexity domains (LCDs) are common drivers and/or regulators of RNP phase separation with and without RNAs [19][20][21][22][23] . The LCD sequence composition and patterning provide programmable modules for dynamic multivalent protein-protein and protein-RNA interactions [24][25][26][27][28][29] . In multi-component mixtures, these dynamic inter-chain interactions are ubiquitous and can either cooperate or compete within a dense network of LC proteins and RNAs 30,31 . The presence of multiple RNPs and RNAs within an intracellular BMC highlights the relevance of understanding how networks of overlapping interactions control the condensate composition and structure 32 , given these very properties are intricately linked to their functional output in the cell 2 .
To systematically explore the regulatory principles of multicomponent RNP condensation with overlapping protein-protein and protein-RNA interactions, here we employ a minimalistic three-component system composed of two LC disordered polypeptides: a prion-like LCD (PLD) and an Arg-rich LCD (RLD), and RNA. PLDs are typically characterized by the presence of  electron-rich amino acids (Y/P/Q/G/S; examples: hnRNPA1, TDP43, FUS) 23 , whereas R-rich LCDs are promiscuous RNA binders and commonly occur as intrinsically disordered RGG domains 22,33 (examples: G3BP1, LSM14A, hnRNPDL, EWSR1, FUS, TAF15). Both LCD types are highly abundant in stress granule and processing body proteins 34 -the two major cytoplasmic BMCs in eukaryotes. From a pathological point of view, multivalent R-rich repeat polypeptides, such as poly(GR) and poly(PR), are potent neurotoxins and are directly linked to c9orf72-derived repeat expansion disorder [35][36][37][38][39][40][41] . These R-rich repeat polymers can invade SGs and impair their fluid dynamics by aberrantly interacting with SG components, including PLDs and RNAs 29,[42][43][44] . Within the PLD-RLD-RNA system, three interactions have the capacity to drive independent phase separation processes: PLD-PLD interactions can drive homotypic PLD phase separation 23,45 , PLD-RLD interactions can drive the co-phase separation of PLD-RLD into heterotypic condensates 46 , and RLD-RNA interactions can drive the formation of RLD-RNA condensates 28,47,48 . Therefore, the RLDs within the PLD-RLD-RNA system represent promiscuous modules that can interact with both PLDs and RNAs 42,43,46 . As such, the PLD-RLD-RNA ternary system represents a suitable and biologically relevant triad for dissecting how networks of overlapping biomolecular interactions control the condensate composition and structure.
Here, employing a multi-scale approach, we systematically study the role of competitive protein-protein and protein-RNA interactions in controlling the topology of ternary PLD-RLD-RNA condensates. For two-component systems (PLD-RLD and RLD-RNA), we show that the mixture composition is a key factor in controlling the phase behavior, condensate spatial organization, and client recruitment in a context-dependent manner. Specifically, we show that for PLD-RLD mixtures, RLD monotonically enhances PLD condensation. On the contrary, RLD-RNA mixtures display a reentrant phase behavior in which the RNA-to-RLD mixing ratio determines the surface organization of RLD-RNA binary condensates in a highly non-monotonic fashion. Within the PLD-RLD-RNA ternary system, RNA-RLD interactions dominate over PLD-RLD interactions, leading to an RNA-induced demixing of PLD-RLD condensates into stable coexisting phases (homotypic PLD condensates and heterotypic RLD-RNA condensates). The topology of these biphasic condensates (non-engulfing/ partial engulfing/ complete engulfing) is determined by the RNA-to-RLD stoichiometry and the hierarchy of intermolecular interactions, providing a glimpse of the broad range of multiphasic patterns that are accessible to these condensates. Mechanistically, the multiphasic structuring of PLD-RLD-RNA condensates is governed by the molecular interactions at the liquid-liquid interfaces, which are encoded in the amino-acid sequence of the proteins and regulated by the composition of the mixture. This multi-scale regulation of inter-condensate surface interactions controls the relative interfacial tensions between the three liquid phases (PLD condensed phase, RLD-RNA condensed phase, and the dispersed phase). Together, our findings reveal that overlapping intermolecular interactions in a multicomponent system represent a regulatory paradigm for controling the composition and structure of multiphasic bio-condensates.

RESULTS AND DISCUSSION
Mixture composition controls the structure and dynamics of binary condensates Before probing the ternary PLD-RLD-RNA phase separation, we studied the phase behavior of the three corresponding binary mixtures: PLD-RLD; PLD-RNA; and RLD-RNA. The presence of R-rich RNA binding domains has been reported to enhance PLD phase separation through intermolecular Arg-Tyr interactions 42,46 . To quantify such an effect in our PLD-RLD binary system, we first determined the isothermal state diagram of the FUS PLD in the presence of arginine/glycine-rich polypeptides (RLD: [RGRGG]5 and FUS RGG3 ; Table S2). Our analysis revealed that PLD undergoes heterotypic phase separation with RLD in a composition-dependent manner. Specifically, the LLPS concentration threshold for PLD ( ) decreases monotonically with increasing RLD concentration [in the absence of the RLD, = ~ 240 µM; at RLD-to-PLD ratio of 5, ~ 120 µM] ( Fig. 1a; Fig. S1). Furthermore, confocal microscopy imaging and fluorescence recovery after photobleaching (FRAP) assays indicate that the PLD partition coefficient and mobility within PLD-RLD condensates scale monotonically with increasing RLD concentration (Fig. 1b&c, Figs. S2&S3). Taken together, our experiments show that with increasing RLD concentration, PLD apparent diffusion coefficient ( ) along with the LLPS threshold concentration ( ) decrease while PLD partition coefficient (K) increases (Fig. 1a-c and Fig. S3). These results are consistent with previous literature reports 42,46 and indicate that RLDs significantly enhance PLD phase separation and impact the condensate dynamics ( Fig. 1a-c). However, in contrast to PLD-RLD mixtures, the isothermal state diagram of PLD-RNA mixtures [utilizing a homopolymeric RNA, poly(U)] showed that RNA does not have any significant impact on PLD phase separation (Fig. S4a). We independently confirmed that poly(U) RNA does not significantly interact with the PLD by using fluorescence correlation spectroscopy (FCS), which revealed identical PLD autocorrelation curves in the absence and presence of poly(U) RNA (Fig. S4b).
Similar to RLD-PLD mixtures, RLD-RNA mixtures display a compositiondependent phase behavior. However, unlike RLD-PLD mixtures, their phase behavior is highly non-monotonic, wherein the two-phase regime is only stabilized within a small window of mixture compositions (Fig. 1d). Such phase behavior is usually referred to as reentrant phase transition 28,47,49 . Within the reentrant phase separation window (Fig. 1d), we previously predicted that disproportionate mixture compositions may lead to the formation of "micellar" condensates, in which the condensates' surfaces can be either enriched in RLDs or RNAs depending on the mixture stoichiometry 49 . To validate the micellar condensate formation in our system, we performed molecular dynamics (MD) simulations using a single-residue resolution coarse-grained model of an RLD from an archetypal ribonucleoprotein FUS (FUS RGG3 ) and a homopolymeric RNA, poly(U) 49 ( Table S2). The representative equilibrium structures of the RLD-RNA condensates revealed two distinct condensate architectures: (a) condensates with RLD-enriched surfaces at < , and (b) condensates with RNA-enriched surfaces at > (Fig. 1d, Fig. S5). Therefore, these condensates appear to be spatially organized 49 and their surface composition is dynamically varied in an RNA dose-dependent manner.
Since RLD chains, but not RNA chains, multivalently interact with the PLD chains (Figs. 1a-c and S4), we next considered the potential of RLD-RNA condensates to deferentially recruit prion-like clients based on their surface architecture. We hypothesize that RLD-rich condensates, but not the RNA-rich condensates, would positively recruit PLDs. Furthermore, the recruited PLDs are expected to be localized on the surface of RLD-rich condensates due to the availability of free RLD sites (Fig. 1e). Performing MD simulations similar to those in Figure 1d but now including -rich FUS PLD as a client (Table S2), we observed enhanced surface recruitment of PLD chains into RLD-RNA droplets at low-RNA conditions, while high-RNA conditions resulted in no recruitment of PLD clients into RLD-RNA condensates (Fig. 1f, Fig. S6). Experimentally, we utilized PLDs from two different RNPs, EWS and FUS, and quantified their partitioning in FUS RGG3 -poly(U) condensates (Table S2). Briefly, we formed RNA-RLD condensates at variable RNA-to-RLD mixing ratios in a buffer that contains the desired PLD clients (labeled with Alexa488 dye). Confocal fluorescence microscopy assays showed that both PLDs are preferentially recruited into RLD-RNA condensates at > while the same clients showed no preferential patitioning into the RLD-RNA condensates at < (Fig. 1g&h, Figs. S7&S8). Analysis of these fluorescence micrographs reveals a ~10-fold increase in the partition coefficient of PLDs in RLD-rich condensates as compared to RNA-rich condensates ( Fig. 1g-h&k). Furthermore, inspecting the fluorescence intensity profiles across RLD-rich condensates revealed that PLDs are preferentially recruited on the condensates' surface while being relatively depleted from the condensates' core (Fig. 1j). These observations confirm that the PLD recruitment in RLD-rich condensates is mediated by molecules on the surface which are predominantly, as our simulation suggests, unbound segments of RLDs (Fig. 1d). To test the generality of this phenomenon, we next performed similar analysis with two additional client polypeptides with similar sequence features as FUS and EWS PLDs: the P/S/G/Q-rich LCD of a transcription activator BRG1 (AA: 1-340) and the C-terminal LCD of RNA polymerase II (Pol II CTD) which contains 30 repeats of YSPTSPS (Table S2). Both BRG1 LCD and Pol II CTD are enriched in amino acid residues which have exposed πcontaining peptide backbones 21 . In addition, Pol II CTD is also enriched in Tyrosine residues (Table S2). Thus, both BRG1 LCD and Pol II CTD are expected to interact with RLDs through Arg-π contacts and therefore display RNA-dependent recruitment into spatially organized RLD-RNA condensates. This prediction was verified experimentally using our confocal microscopy assay (Fig 1i-k, Fig. S9&10).
Collectively, our analysis of the phase separation behavior of different binaries in the PLD-RLD-RNA ternary system reveals two significant pairwise interactions: PLD-RLD and RLD-RNA. While the mixture composition in the PLD-RLD system monotonically impacts PLD-RLD condensate dynamics, in the RLD-RNA binary system, it controls the RLD-RNA condensate architecture. The stoichiometry-dependent recruitment of PLD clients and their spatial localization within RLD-RNA condensates highlights that RNA is capable of regulating PLD-RLD interactions by controlling the availability of free RLD chains on the surface of these condensates (Fig. 1e). To investigate this regulatory effect of RNA on PLD-RLD interactions further, we next examined the impact of RNA on the phase behavior of heterotypic PLD-RLD condensates.

RNA induces condensate switching from PLD-RLD to RLD-RNA droplets
To probe for the effect of RNA on the heterotypic PLD-RLD condensates, we first generated PLD-RLD condensates at a PLD concentration lower than the homotypic PLD LLPS concentration (CPLD < Csat; the green region in the state diagram in Fig 1a). Twocolor fluorescence time-lapse imaging showed that the addition of poly(U) RNA to PLD-RLD phase separated mixture leads to the dissolution of PLD-RLD droplets and subsequent formation of RLD-RNA droplets (Fig 2a&b, Movie S1). The dissolution of PLD-RLD droplets is preceded by a change in their color from yellow (RLD+PLD) to green (PLD), indicating that RLD is leaving these droplets and hence weakening the condensate network and leading to their dissolution (Fig 2b). This condensate switching behavior (from PLD-RLD to RLD-RNA condensates) in the presence of RNA signifies a competition between RLD-PLD and RLD-RNA interactions, with RLD-RNA interactions being stronger than RLD-PLD interactions (Fig. 2a). To investigate whether this effect is specific to poly(U) RNA, we repeated the same assay using poly(A) RNA. Confocal video microscopy revealed that poly(A) RNA is also able to induce the observed condensate switching by sequestering RLDs out of PLD condensates and subsequently forming RLD-RNA condensates (Fig. 2c). Collectively, these results suggest that RNA can induce a condensate-switching transition from PLD-RLD condensates to RLD-RNA condensates due to its superior interactions with the RLD.

RNA triggers a de-mixing transition of RLD and PLD
Our observations in Figure 2 indicate that RNA can sequester RLD out of PLD condensates, which leads to the dissolution of PLD condensates. We therefore asked whether forming PLD-RLD condensates at PLD concentration (CPLD) greater than PLD saturation concentration (CPLD > Csat; the pink region in the state diagram in Fig 1a; Fig  S1), would alter the observed RNA-triggered condensate switching effect (Fig. 2). Repeating our measurements under such conditions, we observed a de-mixing transition, where PLD-RLD condensates reorganized into homotypic PLD condensates and heterotypic RLD-RNA condensates in response to RNA addition (Fig 3a&b, Movie S2, Fig. S11). The time-lapse images reveal that RNA sequesters RLD (red) from PLD-RLD droplets, reaffirming the apparent dominance of RLD-RNA interactions over RLD-PLD interactions. Following the sample equilibration upon RNA addition, we observed that PLD condensates and the newly formed RLD-RNA condensates coexist in a multiphasic pattern where RLD-RNA droplets are distributed on the surfaces of PLD droplets (Fig 3c).
To confirm that the multiphasic condensate formation is not an artifact due to the order of RNA addition, we mixed pre-formed PLD condensates (CPLD>Csaturation) with RLD-RNA condensates (prepared independently) and imaged them using a confocal microscope. We observed that PLD condensates coexist with RLD-RNA condensates in a multiphasic pattern that is reproducible irrespective of the method of sample preparation (Fig. S12) and is stable for more than 24 hours (Fig. S13). We also confirmed that our results are not specific to poly(U) RNA by repeating these experiments with a considerably shorter RNA chain U40 and poly(A) RNA (Fig. 3e&f).
The co-existing condensates (e.g., PLD and RLD-RNA), although quite dynamic as indicated by their respective coalescence-induced condensate growth (Fig. S14), do not exchange their components. Based on these observations, we hypothesized that these coexisting condensates can support distinct microenvironments and differentially recruit specific client biomolecules accordingly, and that the RNA-induced condensate demixing may represent a pathway for sorting those client biomolecules in different condensates (Fig. 3g). Indeed, we observed that a short fluorescently-labeled RNA client (5'-FAM-UGAAGGAC-3') and a protein client (Pol II CTD) that simultaneously partition into PLD-RLD droplets in absence of any RNA are differentially sorted into coexisting condensates in the presence of RNA. More specifically, we observed that the RNA client partitions into the RLD-RNA droplet and the Pol II CTD is uptaken by the PLD droplets (Fig. 3h). Taken together, these results suggest that the competition between RLD-RNA and RLD-PLD intermolecular interactions in a ternary PLD-RLD-RNA mixture can give rise to a rich multiphasic behavior that constitutes condensate switching, condensate demixing, and client sorting (Figs. 2&3).

Mixture composition tunes the topology of coexisting PLD and RNA-RLD condensates
Multi-phasic structures, which stem from the coexistence of multiple immiscible liquid phases [50][51][52] , are hallmarks of several subcellular biomolecular condensates such as the nucleolus and stress granules 11,14,53 . For a three-phase-system (such as A-droplets, Bdroplets, and C-dispersed liquid phase), the equilibrium spatial configuration is determined by the relative interfacial tensions (AC, BC, and AB) of the three liquid phases 11,15,16 . Based on the rank order of AC, BC, and AB, three configurations are possible ( Fig. 4a): (i) condensates do not share any interface and remain separated ( >  +  ; non-engulfment); (ii) condensates partially merge in a way that they share a common interface but are still partly exposed to the dispersed liquid phase ( ~  ~  ; partial engulfment); and (iii) condensate A resides within the condensate B and remains completely separated from the dispersed liquid phase ( >  +  or vice-versa; complete engulfment).
Our experimental results in Figure 1d-k showed that the surface architecture of RLD-RNA condensates is tuned by RNA-to-RLD stoichiometry, resulting in a switch-like change of PLD interactions with these condensates. This raises an interesting possibility of regulating interfacial interactions between RLD-RNA condensates and PLD condensates, given that RLD, but not RNA, multivalently interacts with PLDs. As such, increasing RNA concentration may lead to a controlled morphological variation in the coexisting PLD and RLD-RNA condensates. To test this possibility, we reconstituted the condensate pair by mixing homotypic PLD condensates with preformed RNA-RLD condensates prepared at a variable RNA-to-RLD stoichiometry. We chose RNA-to-RLD ratios such that they span the entire range of the reentrant RNA-RLD LLPS regime (Fig.  S15). We observed that the RLD-rich condensates ([RNA]:[RLD] = 0.1, 0.2, 0.75) are almost completely engulfed by the PLD homotypic droplets (Fig 4b). On the contrary, RNA-rich condensates ([RNA]:[RLD] = 1.25, 2.5, and 5.0) are only partially engulfed by PLD droplets with a shared interface between the two condensate types that substantially decreases with increasing RNA concentration (Fig 4b). To quantify the variation in the interfacial patterning of the two condensate types with changing RNA-to-RLD ratio, we estimated the contact angles for both PLD ( ) and RLD-RNA ( ) droplets using an image analysis approach (see Materials and Methods; Fig. 4c). The cosine of a contact angle ( ), defined as the dihedral angle formed by a liquid interface at the three-phase boundary, is a function of the interfacial tension of the three liquid phases 50,54 .
A small indicates that PLD phase is engulfing RLD phase partially and the contact area between the two droplets is significant. On the other hand, a large value of indicates that the two phases do not have a significant preference for a shared interface (Fig. c, top panel). Remarkably, we observed a sigmoidal-like transition of PLD contact angle ( ) as the RNA-to-RLD ratio is increased, while the RLD contact angle ( ) remained unchanged (Fig. 4c, bottom panel). Thus, variation in quantifies the transition in the coexistence pattern observed in Figure 4b with increasing RNA concentration. According to equation-1, the observed variation in the PLD contact angle indicates a change in the relative rank order of the interfacial tensions in the three-phase system 50 . To confirm this finding, we performed computer simulations utilizing a fluidinterface modeling tool [Surface Evolver 55 ]. Briefly, we created equal volumes of two immiscible liquids with specified interfacial tensions and minimized the total energy of the system by modifying the shape of each liquid interface (see Materials and Methods; Fig.  S16a). As the relative magnitude of the interfacial tension between the two condensates (i.e., ) increased, we observed that the droplets transitioned from a completely engulfed morphology to a partially engulfed state and subsequently to a non-engulfing state (Fig. 4d), similar to our experimentally observed morphological transition with RNA concentration (Fig. 4b). By comparing the results in Figure 4d with those in Figure 4b, we conclude that increasing the RNA concentration energetically destabilizes the shared interface between PLD droplets and RLD-RNA droplets.
A molecular mechanism that explains the effect of RNA on regulating the coexistence pattern can be deduced from the consideration that the surface organization of RNA-RLD condensates is altered as a function of the RNA-to-RLD mixing ratio (Fig.  1d-e). At low RNA-to-RLD ratio, the surface enrichment of free and/or partially-condensed RLDs (Fig. 1d) confers a high propensity for the RLD-RNA condensates to favorably interact with the homotypic PLD condensates (Fig. 4e) through PLD-RLD interfacial binding (Fig. 1d-f). However, at high RNA-to-RLD concentration ratio, the surface of RLD-RNA condensates is enriched with free/partially-condensed RNA chains (Fig. 1d), rendering the interfacial interactions between PLD and RLD-RNA droplets significantly less favorable ( Fig. 4e; Fig. S4). Consequently, the interfacial tension between the two types of condensates (PLD and RLD-RNA) is expected to increase. Therefore, as the RNA-to-RLD ratio increases, a progressive decrease in the contact area between the two condensates takes place (Fig. 4b-e). To test this idea, we used molecular dynamics simulations utlizing homotypic PLD condensates and RLD (FUS RGG3 )-RNA condensates. The simulated equilibrium structures (Fig 4f, Fig. S16b) show that with increasing RNA concentration, a progressive morphological transition occurs wherein the coexistence pattern of PLD and RLD-RNA droplets transitions from complete engulfment to partial engulfment to completely separated (non-engulfment) topologies. Taken together with the MD simulations shown in Figure 1d&f, these results further validate that relative RNA concentration (i.e., RNA-to-RLD stoichiometry) plays a central role in determining topologies of multiphasic condensates (Figs. 4b-d&f).
Sequence-encoded protein-protein and protein-RNA interactions determine multiphasic condensate structuring Our results in Figure 1 and Figure 4 indicate that PLD-RLD interactions (or a lack thereof) at the liquid-liquid interface determines the stability of the interface between PLD condensates and RLD-RNA condensates. These results are consistent with the idea that interfacial tension of a fluid-fluid interface is determined by the intermolecular interactions of the two given fluids at a known thermodynamic state 56 . Combining the results shown in Figure 4 with those in Figure 1, we propose that tuning the molecular interactions between components present on the surfaces of coexisting droplets is sufficient to control the multiphasic coexistence pattern (Fig. 4e). To experimentally test this idea, we designed an all K-variant of RLD (KLD: [KGKGG]5, Table S2) that selectively weakens the PLD-RLD interactions while preserving the ability to phase separate with RNA. The R-to-K substitution is designed based on prior studies showing that (i) lysine residues have a much lesser potency (as compared to arginine residues) to interact with tyrosine and other -rich amino acids/nucleobases 21, 28,29,46,57,58 , and (ii) KLDs can phase separate with RNA via ionic interactions since lysine is expected to carry a similar charge to that of arginine 28 . Indeed, our state diagram analysis indicates that KLD has no impact on PLD phase separation, even at KLD concentrations that are 20 times more than the PLD concentration (Fig 5a). Simultaneously, we confirmed that KLD phase separates with RNA under similar conditions as the RLD (Fig S15) 28 . We further verified the lack of PLD-KLD interactions by confocal microscopy experiments, which reveal that both PLD partition and PLD apparent diffusion rate within PLD condensates remain unchanged with increasing KLD concentrations in solution (Fig 5b, Fig. S17; compare these results with Fig. 1a-c). Consistent with these results, PLD partitioning within KLD-RNA condensates remained very low (< 1.0) and unchanged at variable RNA-to-KLD mixing ratio (Fig. S18), further confirming the absence/insignificance of PLD-KLD (and PLD-RNA) interactions. Therefore, the R-to-K substitution successfully abrogates PLD-RLD interactions by eliminating Arg-Tyr interactions 21 .
Next, we probed for the coexistence pattern of PLD and KLD-RNA condensates. As expected, we observed that the two types of droplets (PLD homotypic and KLD-RNA heterotypic) do not share any interface (non-engulfment) at all the tested RNA-to-KLD stoichiometric ratios (Fig 5c, Fig. S15). These results confirm that the stability of a shared interface is critically contingent on the presence of PLD-RLD interactions at the surfaces of these two condensates. Consistent with this, we further observed that a -rich peptide variant of the KLD, [KGYGG]5, which enhances the PLD phase separation (Fig. S19a), restores the partial engulfment topology of ternary PLD-KLD-RNA condensates (Fig.  S19b). Overall, these findings suggest that sequence-encoded molecular interactions at a liquid-liquid interface between two condensed phases have a direct role in dictating the respective morphological pattern of these coexisting phases (Fig. 5a-d).
According to our model and experimental data (Fig. 4), RNA-rich RLD condensates do not share a significant interface with the PLD condensates due to the absence of PLD-RNA interactions (Fig. 4e). We posit that the addition of an RNA-binding module to the PLD could aid in lowering the interfacial tension between the condensates and thereby forming a shared interface with the RNA-rich RLD condensates. We tested this idea by utilizing the full-length FUS (Fused in Sarcoma), which is composed of a PLD (FUS PLD ) and an RNA binding domain (FUS RBD ) (Fig. 5e). The full-length FUS (FUS FL ) showed enhanced partition into RLD-RNA condensates (as compared to FUS PLD , Fig.  1h) across all RNA-to-RLD mixing ratios (Fig. S20). Additionally, MD simulations indicate that FUS FL is recruited on the surface of RLD-RNA droplets at both low and high RNA concentrations (Fig. S21). These results confirm that FUS interacts with both RLD-rich and RNA-rich condensates. Consistent with our prediction and computational data, confocal microscopy imaging of co-existing homotypic FUS FL droplets and heterotypic RLD-RNA droplets revealed that FUS FL droplets completely engulf the RNA-RLD droplets (i.e. a significant interfacial contact area) across all tested RNA-to-RLD mixing ratios ( Fig.  5f and Fig. S22). Furthermore, weakening the FUS-RLD interactions by replacing RLD with its R-to-K variant (KLD: [KGKGG]5) resulted in a partially engulfed topology (Fig. 5g  and Fig. S23). Interestingly, these FUS-KLD-RNA multiphasic condensates displayed topologies that are reminiscent of Janus spheres 59,60 with two compositionally distinct lobes ( Fig. 5g and Fig. S23). Taken together, these results confirm that modular intermolecular interactions directly govern the coexistence pattern for multiphasic condensates by controlling the dominant interactions at the liquid-liquid interface.

Stability diagram of multiphasic condensates establishes a link between intermolecular interactions, interfacial tensions, and experimentally-observed condensate topologies
Our experimental results and computational modeling presented here collectively suggest a clear relation between the microscopic intermolecular interactions and the mesoscopic multiphase structuring that transcends length-scales. By comparing our experimental and MD simulation data with our fluid-interface modeling results (Fig. 4), we infer that interplay between various intermolecular interactions amongst components (PLD, RLD, and RNA) determine the relative rank order of interfacial tensions between the coexisting liquids. To verify this idea, we consider the equilibrium configurations of two immiscible droplets (PLD droplet and RLD-RNA droplet) in water as a function of the relative values of their interfacial tensions. Since all of our multiphasic condensate analysis was performed at the same conditions for PLD condensates, we chose to fix the surface tension of PLD droplets ( ) and vary the surface tension of RLD-RNA droplets ( ) and the interfacial tension between PLD droplets and RLD-RNA droplets ( ). Employing fluid-interface modeling, we construct a topological stability diagram (Fig. 6a) that marks the boundaries between three distinct morphological states (non-engulfing, partially engulfing, and complete engulfing) in terms of the relative values of these interfacial tensions (i.e. in terms of / and / ). This stability diagram identifies possible transition pathways between different ternary condensate topologies, which were observed in our experiments (Fig. 6a). We note that our stability diagram is able to capture all the variable morphologies observed in experiments, indicating that tuning interfacial tensions may be sufficient to encode diverse multiphasic patterning of the twocondensate system. Subsequently, based on our experimental data and MD simulation results (Figs. 4&5), we propose a mechanistic model that connects the RNA-dependent RLD and PLD interactions with the fluid-fluid interfacial interactions in the mesoscale. The transition between total engulfment to partial and non-engulfment with increasing RNA concentrations (Fig. 4b) can be recapitulated in simulations by varying / while keeping the other interfacial tensions unchanged ( Fig. 6b; Pathway-A). By casting this transition (Pathway-A) in Figure 6b onto the experimental data shown in Figure 4b, we can deduce that increasing RNA-to-RLD ratio increases the interfacial energy between RLD-RNA droplets and PLD droplets. This is expected since RNA-PLD interactions (which dominate the interface at excess RNA conditions) are significantly weaker than RLD-PLD interactions (which dominate the interface at excess RLD conditions, Fig. 1a-c & Fig. S4). However, we note that Pathway-A may not be unique and a more complex pathway (such as Pathway-B) is also possible due to a simultaneous change in , and hence / , as a function of RNA. Replacing arginine with lysine abolishes RLD-PLD interactions, leading to an overall non-engulfment morphology ( Fig.  6a; Pathway-B, Fig. 5c). Next, covalent coupling of PLD with an RNA-binding module creates a bi-valent scaffold (i.e., FUS FL ) with independent sites for RLD and RNA binding, which manifests in completely engulfed ternary topologies in an RNA-independent manner (Fig. 6a-Pathway-C; Fig. 5e-f). Subsequent weakening of RLD-FUS interactions via R-to-K substitutions leads to the formation of partially engulfed Janus-like topologies (Fig 6a-Pathway-D, Figs. 5g and S23). Taken together, these results enable us to reliably correlate sequence-encoded intermolecular interactions with multiphasic behavior, leading to coexistence patterns that can be controlled by sequence perturbations as well as mixture composition (Fig. 6c-f).

CONCLUSION AND OUTLOOK
Intracellular biomolecular condensates encompass a plethora of multivalent proteins and RNAs, which constitute a dense intermolecular interaction network 61 . The fluid structure of these multi-component condensates typically shows coexisting layers of liquid phases as opposed to a well-mixed isotropic liquid condensate. Here, we report experimental and computational evidence of a minimal ternary bio-condensate forming system displaying a rich variety of multiphasic structuring and spatial organization that are primarily governed by composition-dependent and sequence-encoded intermolecular interactions. First, we showed that varying RNA-to-RLD mixing ratio changes the mesoscale organization of the condensed phase at the single condensate level (Fig. 1). The distinction between the core and surface compositions of RNA-RLD condensates can be attributed to differential solvation of non-stoichiometric RLD-RNA complexes at compositionally disproportionate mixtures 49 (Fig. 1d). For example, in RNA-rich condensates, an unbound or partiallycomplexed RNA chain is expected to have a larger effective solvation volume as compared to fully complexed RLD-bound RNA chains, resulting in free RNA chains being preferentially positioned on the condensate surface 49,62 (Fig. 1d-e). A similar argument can also be made for RLD-rich condensates, leading to a composition-dependent tuning of the condensate surface architecture. Such organizational tuning is perceived to be a general phenomenon for heterotypic condensates and is likely to be functionally important in regulating client recruitment and controlling their spatial sub-organelle localization (Fig.  1e-k).
Second, we demonstrate that multiphasic condensates can form in a minimal ternary system via molecular competition for a shared binding partner. In the case of PLD-RLD-RNA system, RLD-RNA interactions dominate over PLD-RLD and this competition leads to an RNA-induced demixing of PLD and RLD into two immiscible phases from a single associative well-mixed RLD-PLD condensed phase (Fig. 3). The coexisting droplets offer distinct microenvironments and display selective client partitioning. We envision that the sequence and structure of RNA would strongly regulate the extent of this demixing phenomenon. We also speculate that competitive inhibition of aberrant intermolecular interactions between protein and RNA components can provide an attractive route to target certain bio-condensate microenvironments in human pathologies.
Third, we show that the coexistence patterning between homotypic PLD condensates and RNA-RLD condensates is directly related to the PLD-RLD interactions at the fluid-fluid interface (Figs. 4&5). More specifically, our experiments and simulations suggest that PLD-RLD interactions govern the thermodynamic stability of a shared interface between PLD condensates and RLD-RNA condensates. As such, perturbation of PLD-RLD interactions on the condensate surface affects the inter-condensate interactions and hence, the interfacial tensions of the coexisting liquid phases. We report two mutually exclusive mechanisms to control the stability of a shared interface: (1) RNA dose-dependent regulation of PLD-RLD interactions through molecular competition (Fig.  4), and (2) perturbation of protein-protein intermolecular interaction network via RLD and PLD sequence variations, which in turn eliminates or enhances the interfacial interactions between PLD condensates and RNA-RLD condensates (Fig. 5). We note that our proposed mechanism 1 is unique to heterotypic condensates where the surface composition of the condensate can be distinct from the condensate core (Fig. 1). We further note a distinction between the composition-dependent regulation of interfacial energies in our ternary system, which is exclusively comprised of intrinsically disordered polymers, and a recently suggested mechanism for the coexistence of stress granules (SGs) and P-bodies (PBs) 31 , where a distinct protein with a common preference for both SGs and PBs acts as a "bridge" between the two condensate types and controls the shared interfacial area. In the later case 31 , the relative amount of the bridge protein is an important variable in dictating the multiphase coexistence pattern. However, our results presented here for RNA-RLD condensates' multiphasic patterning with PLD condensates reveal a unique role of the condensates' surface organization, which can be manipulated by varying the mixture composition. This is a direct result of the formation of fuzzy nonstoichiometric protein-RNA complexes at disproportionate mixture compositions 28,49 . In such a case, an RLD-rich protein-RNA complex, but not an RNA-rich protein-RNA complex, can act as an emergent molecular bridge between RLD-RNA condensates and PLD condensates (Fig. 4e). This allows control over the coexistence patterns in our minimal system without additional bridging proteins.
Together, our presented experimental and computational results highlight hitherto unknown physical principles for multi-component phase separation and suggest that overlapping protein-protein and protein-RNA fuzzy interactions are a regulatory paradigm for the organization of multiphasic biomolecular condensates. They also provide simple physical rules to utilize the phase separation of ternary biopolymeric mixtures to create soft Janus-like particles with tunable morphologies in a stimuli-responsive fashion.    . (e) Proposed mechanism of RNA-mediated fluid-fluid interface regulation. At low RNA concentration, RNA-RLD condensates (red) are enriched with RLD chains on their surfaces, thus facilitating RLD-PLD interfacial binding and mediating a wetting behavior. At high RNA concentration, RLD-RNA condensate surfaces (blue) are enriched with RNA chains, limiting the available RLD molecules for PLD binding, which is responsible for minimal wetting behavior with PLD condensates (green). (f) Equilibrium MD snapshots at variable RNA-to-RLD mixing ratios (see Fig. S16b for the corresponding density profiles). (RLD=FUS RGG3 , red; RNA= poly(U), blue; PLD=FUS PLD , green). =5.6 mg/ml, RNA-to-RLD ratio=0.3, 1.8, 3.2 (respectively for the three snapshots shown), =7.22 mg/ml.

MATERIALS AND METHODS
Details and protocols for the expression/purification of proteins, protein/RNA sample preparation, state diagram measurements, turbidity measurements, confocal imaging and partition analyses, FRAP, RNA-induced condensate switching and demixing, FCS, contact angle analysis, molecular dynamics simulation, fluid-interface simulations and other relevant experimental procedures are provided in the supplementary information appendix.