Introduction

Biomolecular condensates have gained vast attention in recent biological studies. Among them, biomolecular condensates composed of multiple components often separate into two or even more phases since the components within the condensate are also immiscible with each other (1–3). Condensates with multiple phases can have a variety of morphologies (4–7), such as core-shell or nested structures in which one phase is completely enveloped by the other, as that formed by the FC, DFC and GC components of the eukaryotic nucleolus (4), or structures in which droplets share boundaries and contact each other partially, as in the relationship between P-bodies and stress granules (5). These multiphase morphologies have been found to serve as a rather complex partitioning structure within the cell and are closely related to the condensate’s function as an organelle, as they can control biochemical reactions through the compartment by regulating or buffering the concentration of biomolecules (1, 8–10).

Apart from macroscopic factors such as temperature, pH, and salt concentration, the morphology of the condensate is innately programmed into its constituent molecules themselves. Molecular interactions can be divided into attraction and repulsion, which results in associative and segregative phase separations, respectively (11, 12). The major driving force of liquid-liquid phase separation (LLPS) is the attraction between molecules, which can be often achieved by interactions between intrinsically disordered regions or by multivalent domain-wise interactions (13, 14). Condensates caused by associative phase separation are known to move more inward into the phase as the constituent molecules have a higher valency (15–17). On the other hand, repulsive forces according to the volume of the molecule can lead to a segregative phase separation. Even if there are no direct attractive forces between molecules, under crowded environment, volume interaction alone can separate the phase with respect to the volume; the smaller ones move inwards and the larger ones move outwards (18, 19). The elaborate fractionated structure inside the cell can be realized by the competition of these attractions by multivalent moduli and repulsions by excluded volume (20, 21). However, it is not yet clear how the combination of attraction and repulsion contributes to the changes of the condensate shape and how it is coupled with local environment.

Postsynaptic density (PSD), a protein condensate that forms beneath the postsynaptic membrane, is an ideal model system to examine the competing attraction and repulsion in multi-phase morphology. First, PSD proteins contain a wide variety of repulsive volume and attractive multivalent interactions. Second, recent studies reported a puzzling multiphase morphology (22, 23), which is reversed between the solubilized constructs and the system on and beneath the membrane (as explained below), suggesting subtle competing interactions. PSD is one example of biomolecular condensate that can yield multiple compartment structures (24–27). Condensation of PSD proteins is closely related to synaptic function and elicits the formation of a receptor-assembled structure called nanodomain during long-term potentiation (LTP) (28–30). A nanodomain on the postsynaptic membrane enables AMPA receptors (AMPARs) and NMDA receptors (NMDARs) to receive neurotransmitters released from the presynapse more efficiently and to cause firing with a higher probability, compared to the scattered configuration (30–32). Thus, PSD regulates synaptic plasticity via the receptor nanodomain formation. The mechanism can be regarded as the fundamental process of our memory and learning (33, 34).

The formation of such nanodomain structures is triggered by the activation of calcium calmodulin-dependent protein kinase II (CaMKII) (22). CaMKII is one of the most abundant molecules in PSD and is estimated to contain as many as 5600 monomers per synaptic spine (35). Other than its abundance, CaMKII has notable features that contribute to phase separation, namely in its high valency and gigantism. First, CaMKII is a highly multivalent molecule consisting of a dodecamer or a tetradecamer (36, 37). When activated, each of 12 or 14 CaMKII subunits has a strong interaction with the C-terminus of GluN2B (a subunit of NMDAR) near its kinase domain (38, 39). This multivalent attraction allows a two-component mixture of GluN2Bc and active CaMKIIα to form a monophasic condensate (22, 40). Second, the dodecameric CaMKII is a very bulky molecule with a diameter of 15-35 nm; a recent study has reported that extension of CaMKII linker disrupts the formation of the condensate with GluN2Bc and exhibits the homogenous phase (40). This result suggests that phase separation is disrupted by repulsion due to the volume of CaMKII itself. Taken together, CaMKII has two properties with opposite effects on condensation: a high valency, which promotes LLPS, and gigantism, which represses LLPS.

Recent studies reported that CaMKII activation induces multiphase separation into a phase containing AMPAR and another containing NMDAR, presenting a puzzling multiphase morphology (22, 23). In vivo observation of the postsynaptic membrane after LTP elucidated that NMDAR forms nanodomains near the center, whereas the AMPAR domain is distributed at the periphery (22, 23). On the other hand, in vitro experiments with soluble counterparts of AMPAR and NMDAR, together with PSD-95 and the active CaMKII showed a core-shell two-phase droplet with AMPAR and PSD-95 at the core and NMDAR and CaMKII in the shell (22). Thus, the mixture of four components in solution apparently shows a reversed multiphase morphology to what is observed on the postsynaptic membrane. What interaction causes such an apparently contradictory morphology is currently unknown. Since CaMKII activation induces the multiphase separation, the above-mentioned opposite properties of CaMKII toward LLPS may contribute to the changes in the multiphase morphology.

The purpose of this study is to explore the interplay of molecular interactions for the multiphase condensate morphology and the roles of membrane geometry in the PSD proteins. Using four components of PSD (AMPAR complex, NMDAR, PSD-95, and CaMKII) as a model system, we explored the determinants causing the difference of the multiphase morphology between a soluble construct (3D system) and system on and beneath the membrane (2D system, for brevity). We built a mesoscale model based on available experimental data and performed its simulations with these four nanodomain components of PSD. We first validated the model by reproducing the multiphase morphology observed in experiments in the 3D and 2D systems, and found that the CaMKII activation induced the multiphase separation. The interaction network analysis elucidated that, while non-specific interactions dominate in determining the phase behaviors in the 3D system, specific interactions play major roles in the 2D system. Then, we examine the roles of CaMKII in the multiphase behaviors by artificial modulation of the interactions. When we decrease the non-specific volume interaction of the CaMKII in the 3D system, it induces a reversal of the multiphase morphology. Then, in the 2D system, a decrease in the valence of specific interactions of CaMKII induces a reversal of the multiphase morphology. These results suggest that the membrane geometry alleviates the non-specific excluded volume interaction due to the layered organization in the 2D systems, resulting in the dominance of specific multivalent interactions, compared to the 3D system.

Results

Mesoscale model for multivalent protein LLPS

We investigate multiphase separation of four components of the PSD upper layer, 1) the glutamate receptor AMPAR (or its soluble counterpart) in complex with its regulatory protein called stargazin (TARP γ-2 subunit, or its soluble C-terminal fragment TARPc), 2) the glutamate receptor NMDAR (or its soluble fragment), 3) the scaffold protein PSD-95, and 4) the giant dodecameric kinase complex CaMKIIα, which may be either active or inactive. As in the previous studies (41–43), we represent these four protein complexes with a mesoscale model (Fig. 1), and perform comparative molecular dynamics (MD) simulations for the mixture of them. The mesoscale model has domain-resolution; each globular domain (or segment) is represented by a single spherical particle (Fig. 1ACD). AMPAR(TARP)₄ is represented by five particles; one for AMPAR and four for TARPs. NMDAR is modeled as three particles; one for a tetrameric receptor core and two for the C-terminus of GluN2B (termed GluN2Bc) that contains the PDZ-binding motif (PBM). The PSD-95 model contains six particles; representing the N-terminal segment (which can be palmitoylated), three PDZ domains, the SH3 domain, and the GK domain. CaMKII is represented by 13 particles, 12 kinase domains and one central hub domain, which are mutually connected throughout the simulations.

Figure 1.

One-to-one domain-domain binding is treated as a virtual reaction to rigorously maintain the interaction valency (41). Interaction strengths between domains are carefully calibrated based on a set of in vitro experiments. Previous in vitro experiments (22, 24, 44) showed that soluble counterparts of these proteins exhibit core-shell (or phase-in-phase) phase separation at an average concentration of 10 μM. The dissociation constant between TARPc and the full length of PSD-95, that between GluN2Bc and the full length of PSD-95, and that between GluN2Bc and the kinase domain of active CaMKII are known to be 2. 66 μM (45), 1. 19 μM (45) and 0. 10 μM (46), respectively (Fig. 1B). Note that PBMs in both of TARPc and GluN2Bc interact with the PDZ domains of PSD-95, and thus the binding of two PBMs are mutually exclusive for the binding to the PDZ domain. GluN2Bc interacts with PSD-95 through the PBM, which is the last 4 residues of GluN2B C-terminus (47), and with CaMKII at 1290-1309 residues (39, 48), allowing GluN2Bc to bind to both PSD-95 and CaMKII simultaneously (see Material and Methods for more details).

To validate the simulation method and compare the results with in vitro and in vivo experiments unambiguously, we set up the two simulation systems corresponding to each system. The first system is a mixture containing four soluble protein complexes in solutions, soluble counterparts of AMPAR(TARP)₄ and NMDAR, PSD-95 and CaMKII, dissolved in a cubic simulation box (Fig. 1C). In this system, based on a previous experiment (22) AMPAR and TARP are replaced with DsRed2 (PDB ID: 1ZGO, N = 936 residues) and a soluble C-terminal part of TARP (TARPc), respectively. NMDAR is also replaced with the soluble fluorescent protein eqFP670 (PDB ID: 4EDS, N = 482 residues) and the soluble C-terminal portion of two GluN2B subunits connected to eqFP670 (GluN2Bc). CaMKII can be either in active or inactive state; the active, but not the inactive, CaMKII binds to GluN2Bc of NMDAR. This system is denoted as the 3D system hereafter. The second simulation system contains the AMPAR(TARP)₄ complex and NMDAR both embedded into the membrane plane by force fields, the palmitoylated PSD-95 anchored to the membrane via the palmitoyl group by the same potential, and CaMKII dissolved in the solution under the membrane (Fig. 1D). Domains of PSD-95 other than palmitoylated N-terminus and CaMKII are not bound to the membrane and diffuse freely below the membrane. This system is denoted as the 2D system for brevity.

Multiphasic structures of PSD are induced by CaMKII activation

We first performed MD simulations for the mixture of four kinds of PSD proteins starting from random configurations both in the 3D and 2D systems and monitored phase behaviors in comparison with experiments. Based on the previous experiment, the 3D simulation system consists of 135 AMPAR(TARP)₄ complexes, 90 NMDARs, 240 PSD-95s, and 60 active CaMKII complexes in a cubic box with 155 nm per side, which corresponds to the concentrations above the estimated critical concentration for the phase separation (Table S1). Starting from a random configuration, we observe step-wise growths of the cluster that contains four proteins over tens of milliseconds (Fig. 2A). Each cluster is nearly spherical, supporting the fluidity of the protein condensates. The step-wise growth is due to the merge of two small droplets into a larger one. In about 20 milliseconds, nearly all AMPAR(TARP)₄, NMDAR, and PSD-95s molecules together with ∼37 CaMKIIs out of 60 are in one large droplet (Movie S1). The rest of the space, the dilute phase, contains ∼23 CaMKIIs and less than one other proteins, on average. Thus, the 3D system exhibits clear phase separation into a high-density droplet and a dilute phase. In general, when the LLPS contains more than one component, its phase transition can be quantified by the solubility product of the components in the dilute phase (43). For the current four component PSD system, the product of concentrations of each molecule in the dilute phase is in good agreement with that of the experimental concentrations (Table S2).

Figure 2.

Figure 2C shows the final snapshot of the simulation in the 3D system, which clearly shows a core-shell droplet configuration; a phase containing AMPAR and PSD-95 is at the core, while a phase containing NMDAR, PSD-95, and CaMKII is at the outer layer (Fig. 2D). These are all consistent with the in vitro experiments (22, 44), thus validating our simulation model.

When we substituted the active CaMKIIs into the inactive ones which do not have specific interactions to NMDARs, we observe that AMPAR(TARP)₄, NMDAR, and PSD-95, but not CaMKII, coalesce into a large droplet, while nearly all the CaMKII molecules are outside of the droplet forming a dilute phase (Movie S2, Fig. 2E). Notably, molecules inside the droplet exhibit a different configuration from the case of the active CaMKII; AMPAR(TARP)₄, NMDAR, and PSD-95 are well mixed in the droplet with no demixing (Fig. 2F). Therefore, the CaMKII activation induces multiphase separation into the AMPAR-containing phase and the NMDAR-containing phase. The selective binding of the active CaMKII to GluN2Bc in NMDAR, but not to AMPAR, induces the multiphase separation. This reproduces the experimental behavior observed in the in vitro solution system reported by Hosokawa and coworkers (22, 44).

Next, we performed a simulation mimicking the membrane geometry (2D system) and checked how PSD proteins behave and assemble on and beneath the membrane, using the same interaction parameters and the number of molecules as that in the 3D system. Here, AMPAR(TARP)₄, NMDAR(GluN2Bc)₂, and the N-terminal palmitoyl group of PSD-95 are assumed to be embedded to membrane. The length of the cubic box is 500 nm and the receptors diffuse over a 500 × 500 nm membrane plane located at z=0. The number of receptors and the area of membrane are based on the typical areal densities of AMPARs in postsynaptic membranes. In the 2D system, starting from a random configuration, we see modest-sized and dynamic clusters, which repeatedly grow and collapse (Fig. 2B). As in our previous study that contained only two components (41), the current 2D system that consists of four components does not show clear separation into macroscopic phases, but transient and dynamic clustering. The two types of clusters, one consisting of AMPAR(TARP)₄ and PSD-95, and the other consisting of NMDAR, PSD-95 and CaMKII, form nanoscale domains each of which contains ∼50 copies of the two receptors, but does not grow into one macroscopic phase. Both nanodomains with AMPARs and those with NMDARs are unstable; clusters with NMDARs seem to be more stable than that with AMPARs, though both are repeatedly merged and collapsed throughout the simulation time (Movie S3). Clusters of both receptors shape each distinct area, though those interfaces are in contact with each other (Fig. 2G). From the molecular distribution of the cross sections, it appears that NMDAR clusters more on the inner side and AMPAR clusters more on the outer side. (Fig. 2H).

When we replaced the active CaMKIIs into inactive CaMKIIs and performed a simulation under the same condition, a mixture of AMPARs, NMDARs, and PSD-95s take shape into a slightly larger cluster in about 60 ms (Movie S4, Fig. 2I). The distribution of these three types of molecules from the cluster center illustrates that NMDARs reside slightly outside of the cluster and AMPAR and PSD-95 form a cluster inside. Inactive CaMKIIs, on the other hand, are all located away from the membrane since they are unable to bind to GluN2Bc (Fig. 2J). From the results, it is found that with the four PSD components in the 2D system, NMDAR, the active CaMKII, and PSD-95 are clustered on the inner side, while AMPAR(TARP)₄ and PSD-95 are clustered on the outer side in the 2D system. The results are consistent with the nanodomain-like structure of the receptor observed in real synapses in several previous studies (22, 23, 49).

The simulations reproduce experimental features and clearly substantiates that the CaMKII activation induces the multiphasic structure both in the 3D and 2D systems. With inactive CaMKII, which does not attract other molecules, only PSD-95 bridges AMPAR and NMDAR via multivalent specific and non-specific interactions, thus forming a mixed phase. On the other hand, the active CaMKII binds to NMDAR alone. The large volume of CaMKII effectively repels AMPAR from the NMDAR-containing phase to induce the multiphasic architecture in the 3D system.

Specific and non-specific interactions in 3D/2D systems

In the four-component PSD systems, the 3D and 2D systems exhibit the CaMKII-dependent multiphase separation in an apparently opposite arrangement of AMPAR-containing phase and NMDAR-containing phase, which is consistent with experiments. The immediate question is why does morphology differ between the 3D and 2D systems. What properties of CaMKII are responsible for such results? Generally, multiphase morphology is ascribed to be shaped by molecular interactions in the condensates, which we explored based on our simulations.

We first investigate multivalent specific interactions between the four protein complexes. To our surprise, the overall type and the number of specific interactions are rather similar between the 3D and 2D systems (Fig. 3A top, Fig. S1A). Especially, NMDARs keep their interactions with CaMKIIs and PSD-95 nearly the same. This is surprising since the 3D system resulted in a large droplet, whereas the 2D system resulted in relatively small unstable clusters. We note some minor differences. First, the number of PSD-95s that bridge AMPARs and NMDARs is decreased in the 2D system. Second, the number of AMPAR-PSD-95 specific interactions is reduced by about 20 % in the 2D system compared to the 3D system.

Figure 3.

We then analyze non-specific interactions between four protein complexes in the simulations. In sharp contrast to the specific interactions, we find markedly reduced non-specific interactions in the 2D system compared to the 3D system (Fig. 3A bottom, Fig. S1B). For example, the number of non-specific contacts between AMPARs and PSD-95, and that between NMDARs and PSD-95 in 2D system are up to 10 times less than those in the 3D system. The non-specific interaction between NMDARs and CaMKIIs are further reduced. Inherently, the 3D system establishes many more non-specific contacts, whereas the 2D system has more extended configurations with much less contacts.

These two together suggest that the multivalent specific interactions in the 2D system are comparable, while the non-specific interactions in the 2D system are much weaker, compared to the 3D system.

Roles of CaMKII in 3D/2D systems

The dodecameric CaMKII has a rather high valency for the specific interaction with NMDARs. We investigate how many NMDARs are bound per one CaMKII complex in the simulations (termed “effective valency”, Fig. 3B). Even though the 2D system organizes smaller clusters compared to the 3D system, we find that each CaMKII in the cluster in the 2D system can bind more NMDARs than in the 3D system. Thus, the effective valency of CaMKII is higher in the 2D system. We note that this increase in the effective valency can be cause or effect of the different morphology in the 3D and 2D systems.

To understand how the high effective valency of CaMKII can be realized in the 2D system, we delve into the localization of four molecular complexes along the normal direction to the membrane both for the cases of the active and inactive CaMKIIs (Fig. 3C). Both AMPA and NMDA receptors are embedded in the membrane (-2. 5 nm ≲ z ≲ 2. 5 nm). The PDZ2 and GK domains of PSD-95 are located at around -2 nm ≲ z ≲ -5 nm and -2 nm ≲ z ≲ -8 nm below the membrane, respectively. The active CaMKIIs interacting to NMDARs are distributed around -8 nm ≲ z ≲ -10 nm (Fig. 3C left). On the other hand, the inactive CaMKIIs are uniformly distributed 12 nm away from the membrane (Fig. 3C right). These suggest that, in the 2D system, four molecular complexes are distributed to have a layered organization. More importantly, the active CaMKII occupies ∼ 9 nm below the membrane, which is separated from, but not too far from, the distribution of the two receptors or PSD-95s. The layered arrangements may enable more NMDARs to access one CaMKII complex.

In the presence of active CaMKII, the distribution of CaMKII appears to have two peaks at approximately -9 nm and -15 nm. Visual inspection suggests that the former and the latter peaks correspond to the horizontal and the vertical arrangement of CaMKII. Quantitatively, the angle between a hexagonal surface composed by the six kinase domains of CaMKII and the membrane (Fig. 3D) shows a maximum frequency at 0°. Therefore, the majority of CaMKII bound to NMDARs is likely to have hexagonal kinase domains attached horizontally to the membrane (Fig. 3E left). The other peak of CaMKII around -15 nm is widely observed with the angle ranging from 30° to 90°, but not around 0°. This can be regarded as a different binding mode from the previous one, in which the hexagonal plane is perpendicular to the membrane (Fig. 3E middle). These two different binding modes of CaMKII, horizontal and vertical, correspond to the two peaks of CaMKII below the membrane. PSD-95 is a flexible molecule with a string-like structure, whereas CaMKII 12-mer is a molecule with a nearly spherical shape and, moreover, an extremely bulky structure. Although both can interact directly with NMDARs, these differences in flexibility, size, and binding mode are thought to create a hierarchical structure under the membrane and thus provide a firm scaffold for NMDARs (Fig. 3E right).

Together, we hypothesize that CaMKII has two competing effects on the morphology of multiphase separation; the multi-valency in the specific interactions tends to make the NMDAR-containing phase inside, whereas the large volume of CaMKII that contributes to non-specific interactions tends to put the NMDAR-containing phase outside. The relative balance between the two competing effects are different between the 3D and 2D systems; the non-specific volume effect dominates in the 3D system, whereas the specific multivalent interaction overwhelms the volume interaction in the 2D system.

Reduction of CaMKII volume can invert multiphasic structure in 3D system

We have shown that i) the CaMKII activation induces the multiphase separation into AMPAR-containing and NMDAR-containing phases both in the 3D and 2D systems, ii) the morphology of multiphases in the 3D and 2D systems is apparently opposite, and iii) relative strengths of specific and nonspecific interactions are sharply different between the 3D and 2D systems. Importantly, CaMKII has strong impact both on the specific interactions via its multi-valency and on the non-specific interactions via its extraordinary large volume. These together led us to hypothesize that multi-valency and large volume of CaMKII have competing effects to modulate morphology of multiphases in the 3D and 2D systems. To test our hypothesis, we design a set of artificial systems that modulate multivalent specific interactions or large volume nonspecific interactions of CaMKII.

In the 3D system, non-specific interactions overwhelm the specific ones to modulate the multiphase morphology. Non-specific interactions of CaMKII are dominated by its large excluded volume. Thus, we design artificial systems with reduced volume of CaMKII keeping other properties unchanged. We carry out simulations by replacing the molecular radii of CaMKII to 2/3 and 1/2 of their original size (Fig. 4A). With this modification, the volume of CaMKII is reduced by 8/27 and 1/8, respectively. The simulation containing CaMKII with r=2/3 of the original radius demonstrate that the AMPAR-containing phase (denoted as A phase) and the NMDAR-containing phase (denoted as N phase) are both separated from the dilute phase to form their respective hemispheres with a shared boundary, which resembles structures called Janus droplets (Fig. 4BC center). When the radius is halved, CaMKII (r=1/2), the N phase is positioned inside of the cluster and nearly completely engulfed by the A phase (Fig. 4BC right). Namely, reducing the non-specific excluded volume interactions of CaMKII results in the inversions of multiphase morphology in the 3D system. The violin plot in Fig. 4D illustrates how each of the two types of receptors are distributed starting from the center of mass of the cluster. As the radius is decreased, the peak of AMPARs moved outwards and that of NMDARs moved inwards.

Figure 4.

Next, in the 3D system we reduced specific interactions of CaMKII to see how the phase morphology changes. Specifically, we designed another set of systems with reduced valency of CaMKII. With reduced specific interaction of CaMKII, we originally expected not to observe the opposite morphology, but actually found this did not change the phase morphology (Fig. S2).

Reduction of CaMKII valency can invert multiphasic structure in 2D system

Continuing the examination of our hypothesis, we next modulated the setup of 2D simulations to weaken the contribution of specific interactions of CaMKII. We reduced the valency of active kinase domains that can bind to NMDAR; the valency is reduced from the original 12 to six (CaMKII (V=6)) and three (CaMKII (V=3)) (Fig. 5A). In the CaMKII with reduced valency, only 6 or 3 kinase domains are active, while others are kept as inactive. In the case of CaMKII (V=6), in 60 ms, clusters are separated into AMPAR-containing and NMDAR-containing domains (Fig. 5B center). The mean cluster size is smaller than that of the original setup. Some clusters in this condition have a morphology with NMDARs inside and AMPARs outside (dotted arrows, clusters in the middle and bottom right), similar to those obtained in the original CaMKII simulations, whereas others in contrast, have opposite morphology with NMDARs outside and AMPARs inside (solid arrows, clusters at the bottom). For the case of CaMKII (V=3), there is no clear domain segregation of AMPARs and NMDARs as in the case of 12 or 6 valences, and homogenous clusters are shaped, almost analogous to the condition with inactive CaMKII (Fig. 5B right). To gain a more precise understanding of the segregation of receptor domains and their relative positions within clusters in more detail, we examined the distribution of the two types of receptors in clusters above a certain size for each condition, in cross sections at the X = 0 line. The receptor distribution of each cluster is normalized by the radius of gyration of the cluster and plotted on the horizontal axis as a relative distance to obtain an accurate estimate of the distribution between clusters of different sizes. In the section with valency V=12, a distinct AMPAR-outward and NMDAR-inward structure are observed. In the case of CaMKII with V=6, the inside-outside separation almost disappears and the NMDARs are evenly distributed both inside and outside the cluster; for CaMKII with V=3, AMPARs are distributed closer to the center and NMDARs are slightly distributed outside (Fig. 5C). The violin plot also shows that as the valence decreases, AMPARs are distributed closer to the center of the cluster and NMDARs are distributed further away from the center (Fig. 5D).

Figure 5.

On the other hand, when we reduced the volume of CaMKII in the 2D system, the NMDAR phase becomes more rigid as the volume is reduced from the original CaMKII. In addition, the AMPAR and NMDAR clusters does not share a boundary and are completely separated into their own clusters (Fig. S3).

Interfacial tensions between each phase substantiate the tendency of multiphase morphologies

In general, the multiphase morphology is determined by the balance of interfacial tensions. When two components, A and B, can form droplets in solution and are mutually immiscible, their mixture can have several morphologies (50). 1) The segregated droplets where two components, A and B, form segregated droplets, which is stable when γ_AB > γ_A + γ_B, where γ_A is the interfacial tension between A and solution, γ_B is that between B and solution, and γ_AB is that between A and B. 2) The core-shell droplet where an A-containing core droplet is completely engulfed by a shell containing B, when γ_A > γ_B + γ_AB. 3) The core-shell droplet where a B-containing core droplet is completely engulfed by a shell containing A, when γ_B > γ_A + γ_AB. 4) The Janus droplet (partial engulfing) where two droplets are coalesced to have partial A/B interface, for the rest of the cases.

To deepen our understanding of the multiphase morphologies in the current four-component of PSD proteins, we calculated the three interfacial tensions among the three phases (AMPAR-containing phase, NMDAR-containing phase and dilute phase) in the soluble 3D system. Using the canonical ensemble, we prepared a three-dimensional slab system (Fig. 6A) and estimate the interfacial tension of each interface based on the Kirkwood-Buff equation (51) with varying size of CaMKIIs. In the system containing CaMKII (WT), the interfacial tension between the AMPAR-containing phase and the dilute phase (γ_A ) is larger than the sum of the other two, the interfacial tension γ_N between the NMDAR-containing phase and the dilute phase and γ_AN between the A and N phases, γ_A > γ_N + γ_AN (Fig. 6BC). Thus, the core-shell morphology with the A phase at the core and N phase in the shell should be most stable. Indeed, this is what we observe in the simulation. As the volume of CaMKII decreases, we find both γ_N and γ_AN increase whereas γ_A is unaffected (Fig. 6BC). With the CaMKII (r=2/3), our estimate satisfies γ_A < γ_N + γ_AN and γ_AN < γ_A + γ_N as the Janus droplet appears (Fig. 6BC). With further reduced volume of CaMKII, γ_N increases to approach the border to the reversed core-shell morphology with the N phase at the core and the A phase in the shell. However, while we observe nearly core-shell arrangement with the N phase at the core with CaMKII (r=1/2), the estimated tensions do not satisfy γ_N > γ_A + γ_AN. This apparent discrepancy may be partly due to the relatively small size of the simulated system, whereas the theory rests on a macroscopic argument. In addition, the numerical estimate of the interfacial tension, especially γ_AN, is inherently difficult and may contain a non-negligible error. Yet, the overall trend of the interfacial tensions explains qualitatively the observed changes in the multiphase morphologies, supporting our interpretation of the underlying physical mechanism.

Figure 6.

Discussion

In this study, we investigate the multiphase morphology of biomolecular condensates induced by multivalent interactions and volume interactions in 3D solution and 2D membrane bound systems, using PSD proteins as a model system. Utilizing a mesoscale model which can simulate molecules on a comparatively large scale in both time and space, we first show that the CaMKII activation induces the multiphase separation and that four components of PSDs form reversed multiphase morphologies between the 3D solution system and the 2D system on and under the membrane. Analyzing the molecular interactions, we identify that, while the 3D system contains vast number of non-specific volume interactions, the 2D system is dominated by specific multivalent interactions. Especially in the PSD protein system, the volume interaction of CaMKII drives the NMDAR/CaMKII-containing phase in the shell of the core-shell morphology in the 3D system, whereas multivalent specific interaction of the active CaMKII stabilizes the NMDAR/CaMKII-containing domain at the core of the 2D system. These results suggest that the layered structure of the PSD created by the membrane geometry reduces the effective volume interaction and thus increases the relative contribution of multivalent interactions in 2D system.

The segregation of nanodomains containing NMDAR or AMPAR can hold significant functional significance in synaptic functions. The clustering of receptors is generally considered essential for efficient neurotransmitter reception from presynaptic terminals (29, 30). Nanodomains of NMDARs and AMPARs are separately aligned with synaptic vesicle release sites in presynaptic cell via different assembly proteins in synaptic clefts and are used in distinct context. The current study shows that the NMDAR-containing nanodomains tend to be located near the center of the postsynaptic membrane, whereas the AMPAR-containing nanodomains are at the periphery, which is consistent with recent in vivo observations of receptor locations (22, 23, 49). Notably, our simulations contain only proteins in the PSD, but neither presynaptic or synaptic cleft proteins. Therefore, the results imply that the postsynaptic proteins alone may be sufficient to form the nanodomain morphology observed in vivo.

While the estimated interfacial tension well explains the multiphase morphology in the 3D system, the same approach cannot be used in 2D because the 2D system does not show stably separated phases, but only nanoscale dynamic clusters (41). For such clusters, the interfacial tension is ill-defined. However, we can still rationalize the cluster morphology on the membrane by focusing on its interfacial properties. There are reasons to speculate that the domain boundary cost between domains containing NMDARs/CaMKII and the dilute phase in 2D systems is greater than that containing AMPARs. First, in the light of the fluctuation in the copy numbers of receptors in the largest cluster, the NMDAR clusters show markedly smaller fluctuations and exhibit step-wise growth similar to that in the 3D system (Fig. 2AB). According to the capillary-wave theory, the larger the interfacial fluctuation is, the smaller the interfacial tension is (52–56). Thus, the NMDAR-containing domain can be regarded to have a larger domain boundary cost than that of the AMPAR domain. Furthermore, the number of multivalent interactions of NMDAR with either PSD-95 or CaMKII was almost the same between 3D and 2D systems, whereas that between AMPAR and PSD-95 was significantly reduced (Fig. 3A). This is in harmony with the view that the decline of domain boundary cost of AMPARs with the dilute phase in 2D system relative to that of NMDARs. At first glance, this seems to be due to differences in the dissociation constants of the specific interactions. However, modulating the K_d of the AMPAR-PSD-95 and NMDAR-CaMKII specific interactions did not change the multiphase morphology, and thus domain boundary cost is not strongly dependent on K_d of the molecules (Fig. S4). Instead, the multivalence of the interaction is the primary determinant of the interfacial cost and thus the domain morphology below the membrane. NMDARs migrate outwards when the multivalent interaction is disturbed by a decrease in the CaMKII valence.

Modulation of the balance between specific and non-specific interactions via the membrane geometry should not be limited to the PSD proteins, but can be a general phenomenon. Domain-wise specific interactions tend to be relatively strong (per domain) and have strict, often one-to-one, stoichiometry. Non-specific, or more broadly less specific, interactions tend to be weak and not to have clear stoichiometry. In the 3D solution systems, each protein can interact with many other proteins via specific or non-specific interactions. On the other hand, membrane-associated protein systems can have less interaction pairs per protein inherently due to their reduced dimensionality. Limited pairs should be dominated by strong specific interactions, thus reducing the contribution of non-specific interactions. Thus, compared to 3D solution systems, multivalency in specific interactions should be more important on and beneath the membrane. In fact, recent observations have revealed that multiphase separation by LLPS is more likely to occur at the two-dimensional solid-liquid interface than in a three-dimensional liquid droplet (57). In a pseudo-2D system, specific interactions may be effective and facilitate multiphase separation.

The current study elucidates the primary role of specific multivalent interactions on and beneath the membrane, which was previously illustrated for the Nephrin/Nck/N-WASP complex by Rosen’s group and others (14, 58). The Nephrin/Nck/N-WASP complex exhibits multivalent interactions through SH3 domains and PRMs. This complex not only induces phase separated macroscopic droplets in 3D in vitro solution systems (14) but also shows 2D in vitro phase separated clusters on membranes driven by increasing valency (58). These two-dimensional clusters exhibit fluid properties, as in the 3D droplets. Our current research further expands on these concepts elucidating the primacy of specific multivalent interactions in the 2D system, compared to the relative importance of non-specific volume interactions in the 3D system. Especially, this difference in the balance between specific and non-specific interactions causes a reversal of the multiphase morphology in the case of the four component PSD system.

Physical mechanisms of demixing into multiphase droplets, especially the role of multivalent interactions therein, have been intensively studied by Brangwynne’s group (59). In vitro 3D solution systems that can dynamically change the multivalency via oligomerization are designed to demonstrate that the increase of valency by protein oligomerization can enhance multiphase immiscibility leading to core-shell morphology (59). Notably, some of the tested systems use extraordinarily high valence (24- or 60 folds) to realize the demixing. The current study reveals that specific multivalent interactions, relative to non-specific volume interactions, operate more efficiently in 2D systems. Therefore, we hypothesize that similar oligomerization-based system may form immiscible multiphase clusters at lower valences on and beneath the membrane.

Biomolecular condensates are ubiquitous in cells, most notably in nucleus and under membrane. Membrane-associated biomolecular condensates are linked to signal transduction, membrane shape manipulation, stepwise transport of molecules and vesicle-membrane tethering by wetting (60–63). Most cases including Nephrin/Nck/N-WASP, Galectin/Saccharide, and Grb2/Sos1 systems are mediated by multidomain proteins, which thus possess specific multivalent interactions for LLPS (64). Condensate formations in the nucleus are often linked to activation/inactivation of local chromatin regions associated with epigenetic marks (65, 66). In contrast to the membrane-associated cases, many of nuclear proteins/RNAs that form condensates in 3D solutions are characterized by full of intrinsically disordered regions (IDRs)/heterogeneous folds including transcription factors, RNA polymerases, and ribosomal RNAs in nucleolus. Interactions therein are dominated by less specific interactions of disordered polymers. Our current study points out that specific multivalent interactions play major roles in the membrane systems, whereas the non-specific interactions that do not have strict stoichiometry are relatively important in the 3D solution system. We emphasize that this difference in the relative importance is the direct outcome of the different dimensionality, and thus is not limited to PSD systems, but can be general. We speculate that membrane-associated condensates and condensates in nuclei may have evolved differential interaction features under distinct pressures resulting from their differences in dimensionality.

Our mesoscale model approach not only provides a better understanding of the PSD and synaptic plasticity, but also allows prediction of how the arrangement and structure of multi-component biomolecular condensates, driven by multivalent interactions, affects the LLPS in various systems.

Materials and Methods

Methods are briefly summarized here; a detailed version is provided in Supporting Information.

Molecular Representation

We represented each protein domain as a spherical particle. The radius of each spherical domain R_n and the diffusion coefficient D are calculated by the formula:

The viscosity of the cytoplasmic domain in the 3D system and the 2D system are determined based on the viscosity of water at 300 K (0. 89 cP); the viscosity of the protein domains on membrane in the 2D system is set to 10 times the viscosity of the cytoplasmic fluid (8. 9 cP) to reduce the calculation time required for assembly.

Basic simulation method

All simulations were performed using ReaDDy 2 (67, 68). The overdamped Langevin equation is used for the equation of motion. The calculation timestep and temperature of all the simulations is set to 0. 25 ns and 300 K, respectively, for all the systems.

Every particle is subject to the following total potential energy function:

The bonds between AMPAR and TARP, NMDAR and GluN2Bc, and the hub and kinase domains of CaMKII are set as harmonic bonds with a spring constant of 10 kJ/mol/nm². In contrast, all domain linkers of PSD-95 are represented by modified flexible Gaussian polymer chains, which has a soft spring constant determined by the length of linker amino acids. Other detailed potentials (angle, dihedral, specific, nonspecific, and membrane mimicking potentials) are described in Supporting Information.

Simulation in the 3D system

In the 3D system containing four kinds of protein complexes, as an initial configuration, we placed 135 molecules of DsRed2(TARPc)₄, 240 molecules of PSD-95, 90 molecules of eqFP670(GluN2Bc)₂, and 60 molecules of CaMKII randomly avoiding their mutual overlaps, in a 155. 362 × 155. 362 × 155. 362 nm cubic box with the periodic boundary condition. The amount of PSD-95 is determined so that the total amount of PDZ domains is equal to the amount of its binding clients (sum of TARPc and GluN2Bc). This setup theoretically allows the PSD-95 and the two types of receptors to bind without excess or deficiency, while the amount of CaMKII is in excess relative to the possible binding sites of NMDARs. To investigate the morphology of the two phases, we adopt a cubic box system, rather than an elongated box; the concentrations of the dilute phases may be subject to artefacts from the finite size and thus not be very accurate as an estimate of critical concentration of LLPS. Then, three independent simulations at 300 K with different stochastic forces are repeated for 8 × 10⁷ MD timesteps.

Simulation in the 2D system

In the 2D system containing the same amount of four kind of protein complexes as in 3D setup, AMPAR(TARP)₄ and NMDAR(GluN2Bc)₂ are embedded to the membrane, PSD-95 is tethered to the membrane via the palmitoyl group, and CaMKII is in the cytoplasm beneath the membrane. In a 500 × 500 × 500 nm cubic box (the periodic boundary condition), we set the plane at the center of the box (Z = 0) as a center of the postsynaptic membrane, and apply a potential energy for the membrane. Transmembrane proteins AMPARs and NMDARs and the palmitoyl group of PSD-95 are assumed to diffuse only on the membrane due to this membrane potential. Other domains of PSD-95 and CaMKII can move around in the cytoplasm below the membrane (Z < -2. 5 nm). Starting from the randomly placed initial conditions, simulations at 300 K are performed for 8 × 10⁷ MD time steps.

Measurement of interfacial tension in the 3D slab system

For measurement of the interfacial tension of each phase, we first conducted the simulation in the 50 × 50 × 500 nm periodic box for 4 × 10⁶ MD timestep, confining the molecules written in Table S4 for each condition into a small region of 50 × 50 × 100 nm by applying a confinement potential. The number of molecules in the slab box is chosen to meet the same stoichiometry of each molecule as the simulation in 3D cubic box. Then, we remove the confinement potential and continue the simulation for 2 × 10⁷ MD timestep in an elongated 50 × 50 × 1500 nm box with the periodic boundary condition. From the pressure tensor obtained in the simulations, we estimate interfacial tension value between respective condensed and dilute phases, using Kirkwood-Buff equation (51, 52) as follows:

The amount of CaMKII in each condition is corresponding to the result obtained from simulations in a cubic box. It should be noted that we have confirmed that the value of interfacial tension does not change dramatically even when the CaMKII content was fixed (Fig. S5).

Acknowledgements

We would like to thank Yutaka Murata, Tsuyoshi Terakawa, and Evangelia Pantatosaki for insightful discussions. This work was supported by JSPS KAKENHI grants 24KJ1449 (to R. Y. ), 21H02441 (to S. T. ), and 24K01991 (to S. T. ).

Additional information

Author contributions

R. Y. and S. T. conceived and designed the study. R. Y. constructed a mesoscopic model of proteins in PSD, set up the simulation program ReaDDy 2, performed the molecular simulation, and conducted the analysis. R. Y., G. B. and S. T. discussed the results and wrote the manuscript.

Additional files

Supplemental information

Movie S1

Movie S2

Movie S3

Movie S4