Decoding the condensate grammar necessitates connecting protein sequences with collective physical properties. ELPs stand out because of their sequence simplicity and amenability for biological engineering, allowing systematic exploration and precise attribution of amino acid contributions. We focus on diblock ELPs of sequence (V-P-G-V-G)_n-(V-P-G-X-G)_n, which we abbreviate as V_nX_n (Figure 1A). The guest residue in the first block is set as valine to promote phase separation, and we explore 20 systems in which every natural amino acid is substituted into the X position in the second block.

We adopt a multiscale approach to balance accuracy and efficiency for simulating ELP condensates (Figure 1B). The first stage of this strategy is to simulate condensate formation using a coarse-grained force field, MOFF (Latham and Zhang, 2021; Latham and Zhang, 2019; Latham and Zhang, 2020; Latham and Zhang, 2022c), with one bead per amino acid and implicit solvation. These simulations are highly efficient for system relaxation and promoting large-scale conformational rearrangements that occur over slow timescales. We further converted these α-carbon-based models to approximately four heavy atoms per bead resolution and introduced explicit water and ions for simulations with the MARTINI force field (Marrink et al., 2007). The latest version of this force field (MARTINI3) provides a more balanced set of parameters for protein-protein interactions and has been applied to study biological condensates (Souza et al., 2021; Thomasen et al., 2022; Benayad et al., 2021; Tsanai et al., 2021). These simulations further relax the protein configurations and the partition of water and counterions in condensed and dilute phases. Finally, we carried out explicit solvent all-atom simulations for 250 ns to produce an accurate characterization of the chemical environment of the condensates.

Utilizing the simulated condensate conformations, we computed various quantities to benchmark against experimental measurements. While the critical temperature has been widely used as a measure for condensate stability, determining it computationally is expensive. As an alternative, we computed the surface tension, τ, using 100-μs-long MARTINI simulations performed with the NP_NAT ensemble (Zhang et al., 1995). As detailed in the ‘Supplemental theory’ in Appendix 1, an inverse relationship is expected between τ and the critical temperature, Tc, for systems exhibiting LCSTs. We further approximate Tc with the transition temperatures (Tt) of ELP sequences (Urry, 1997), which are the temperatures at which ELPs undergo an LCST transition at a specified solution condition. Tt was shown to be linearly proportional to TC (McDaniel et al., 2013; Meyer and Chilkoti, 2004). As expected, a negative correlation can be readily seen between computed surface tension and experimental Tt (Figure 1C). This observed negative correlation between Tt and τ supports the simulation approach’s accuracy in reproducing the sequence-dependent changes in ELP phase behavior.

We note that the experimental values were determined using different ELP sequences from those simulated here, and the relation between the surface tension and Tc is likely nonlinear (Roe, 1975), both contributing to the imperfect agreement. In another study (Ye et al., 2024), we adopted the same multiscale approach to determine the dielectric constant for several ELP condensates. The simulated values agree well with those determined using FLIM experiments.

Upon validating the accuracy of the simulation protocol for reproducing collective properties of ELP condensates, we examined their structural organization. Hassouneh et al. argued that, in dilute solutions, ELP diblocks self-assemble into the so-called weak spherical micelles with dense cores and almost unstretched coronas (Hassouneh et al., 2015). At phase separation conditions, whether similar structural features are preserved in condensates remains unclear.

We found that most ELP condensates undergo a microphase separation. Representative configurations for four typical systems from MARTINI simulations are shown in the top panels of Figure 2, with the V blocks in blue and X in red. Overall, the microphase separation manages to orient the more hydrophobic blocks toward the condensate interior as opposed to the interface. The relative density of guest blocks near the center of the condensate positively correlates with their hydrophobicity, as can be seen in both MARTINI (Figure 2—figure supplement 1) and all-atom (Figure 2—figure supplement 2) simulations.

Surprisingly, microphase separation did not produce lamellar morphology as expected for block copolymers with equal volume fraction of the two blocks (Figure 2—figure supplement 3; Bates and Fredrickson, 1999; Mai and Eisenberg, 2012; Zhulina et al., 2005; Leibler, 1980; Bates and Fredrickson, 1990; Matsen and Schick, 1994; Matsen and Bates, 1996; Grason, 2006; Swann and Topham, 2010; Shi, 2021). In particular, the condensates appear to form gyroid-like structures, in which the V and X blocks form two interpenetrating networks (Figure 2—figure supplement 4). This morphology also differs from micelle-like structures seen in simplified hydrophobic-polar (HP) polymers (Statt et al., 2020; Wessén et al., 2022). It promotes interfacial contacts while maintaining substantial self-interactions as well. Weak interfacial tension between different ELP blocks has also been noted by Hassouneh et al., 2015. These qualitative observations are insensitive to the polymer length and system size in our simulations (Figure 2—figure supplement 5). While longer diblock copolymers drive more prominent microphase separation, similar gyroid structures can also be observed. Notably, for more hydrophilic X blocks, the condensates begin to dissolve into a collection of micelle-like structures, consistent with the predictions by Hassouneh et al. in dilute solutions (Hassouneh et al., 2015).

Quantitative characterization of the condensate interior supports the presence of interpenetrating networks with substantial interfacial contacts as well. For example, significant contacts between V and X blocks can be readily seen in the inter-chain contact maps (Figure 2, Figure 2—figure supplement 6). We further computed the radial distribution functions for amino acids in various blocks (Figure 2, Figure 2—figure supplement 7). For all systems, we found that the probability of finding cross-block contacts V-X is often comparable to the intra-block contacts between X-X or V-V, depending on the relative hydrophobicity of V and X. For example, the more hydrophobic F blocks exhibit strong self-clustering, much more prominent than V-V blocks that are comparable to V-F blocks. On the other hand, clustering among V-V blocks and V-A blocks is more substantial than among A-A blocks. These results are again robust with respect to system setups (Figure 2—figure supplement 5).

To experimentally test the microphase separation behavior uncovered in simulations, we studied the micro-physicochemical properties of the V-end and X-end of the peptides. We constructed diblock peptides with the combination of 30 pentameric repeats of V block and X (A or G) block, namely V₃₀A₃₀ and V₃₀G₃₀ (‘Experimental sequences’ in Appendix 1). The amino-termini of V₃₀A₃₀ and V₃₀G₃₀ sequences were subsequently labeled with environmentally sensitive BODIPY or SBD fluorophores (Liu et al., 2017; Shen et al., 2023), whose lifetime could be measured to quantify the viscosity or polarity of the V-end (Figure 3A, left panel) (Chambers et al., 2018). These probes have been reported to be only sensitive to single physicochemical properties (Ye et al., 2024; Shen et al., 2023). To avoid artifacts induced by fluorophore labeling, we usually used ELPs labeled with a low fraction of dyes. We also constructed A₃₀V₃₀ and G₃₀V₃₀ diblock peptides, wherein the viscosity or polarity of the A-end or the G-end could be measured by fluorophores that are attached at the amino-terminus (Figure 3A, right panel).

Using FLIM, we found that the lifetime of BODIPY for the V-end (5.43 ns) was longer than that for the A-end (4.35 ns), suggesting that the V-end indeed has a higher microviscosity than the A-end (η_V = 2233.54 cp vs η_A = 969.57 cp). Accordingly, the lifetime of SBD was longer for the V-end (8.75 ns) than the A-end (7.00 ns), indicating that the micropolarity of the V-end was lower than the A-end (ϵ_V=13.25 vs ϵ_A = 18.97). These observations could be largely attributed to the greater extent of dehydration at the V-end due to its higher local peptide density. We further showed that the observed differences are not results of possible artifacts arising from any subtle distinctions between the two sequences V₃₀A₃₀ and A₃₀V₃₀ (‘Experimental characterization of ELP condensates’ in Appendix 1, Figure 3—figure supplements 1 and 2). Similar results were observed using the V-G sequences. FLIM experiments revealed that the V-end was more viscous than the G-end (η_V=2972.72 cp vs η_G = 1958.60 cp) and the V-end was less polar than the G-end (ϵ_V=9.14 vs ϵ_G = 27.50). These experimental observations provided the first line of evidence to support the microphase separation, as suggested by the simulation results. It is worth noting that the reported values, although related, may not quantitatively represent the steady-state viscosity. This discrepancy arises from the slow relaxation timescale inherent in ELP condensates with viscoelastic properties.

To further investigate these phenomena in real time in motion, optical tweezers were used to stretch the droplet to observe the viscosity changes of different blocks. A single diblock condensate was captured by two optical traps (traps 1 and 2) at the same position, and it was stretched as trap 1 moved away from the original position (Figure 3C and D). We used the peptides V₃₀A₃₀ and A₃₀V₃₀, with BODIPY labeled at their amino-terminus, to study the microviscosity changes where the fluorescence intensity decreases as the viscosity decrease. We found that the normalized fluorescence intensity of the V-end decreases greater than the A-end, indicating that the viscosity change of the V-end is larger (Figure 3E). We anticipate that during stretching condensates deform and molecules may deviate from the preferred chemical environment established with microphase separation. Such deviations impact the hydrophobic segments more, leading to more significant reduction in their interactions with surrounding molecules and correspondingly their viscosity.

The strong dependence of molecular organization on amino acid hydrophobicity suggests that the solvation environment of individual residues might be a determining factor for condensate stability. Indeed, as shown in the ‘Supplemental theory’ of Appendix 1, the critical temperature is closely related to the free energy cost of transferring polymer beads from a solution state to a polymer-only environment. This transfer free energy is often used to quantify the hydrophobicity of amino acids (Regy et al., 2021; White, 1996; Wimley et al., 1996; Wilce et al., 1995; Zimmerman et al., 1968; Radzicka and Wolfenden, 1988; Lawson et al., 1984). To explore their relationship more quantitatively, we compared the transition temperature for ELP condensates measured by Urry, 1997 to several hydrophobicity scales. Figure 4A shows that water-octanol and water-POPC interface transfer free energies best correlate with the transition temperature. Meanwhile, other hydrophobicity measures, even including those derived specifically for disordered proteins, do not match as well. We note that our conclusion here is not dependent on the limited number of hydrophobicity measures examined in this study, and can be seen in a previous clustering of 98 different hydrophobicity scales (Simm et al., 2016).

The correlation analysis of the Urry transition temperature with various hydrophobicity scales supports a unique chemical environment emerging from the collective behaviors of condensates. To better understand the microenvironment surrounding guest residues in X blocks, we computed the radial distribution functions using MARTINI simulations. As shown in Figure 4B and C, Figure 4—figure supplement 1, the guest residues are immediately surrounded with hydrophilic glycine residues as the nearest neighbors. In the second shell, more hydrophobic valine and proline residues appear. These profiles reveal the interfacial nature of the chemical environment surrounding the guest residues consisting of both hydrophobic and hydrophilic amino acids. This interfacial solvation environment is consistent with the strong correlation of the Urry scale with water-POPC interface transfer free energy. Therefore, ELP condensates do not simply feature a hydrophobic interior that separates from hydrophilic residues as seen for folded proteins, and hence the poor correlation shown in Figure 4A with the burial propensity of amino acids (RSA1, RSA2).

The ELP peptide sequences partly dictate the interfacial environment. For example, the guest residues are juxtaposed between two glycine residues in the pentamer (Figure 1A), creating the immediate shell of hydrophilic residues. However, the presence of valine in the pentamer prevents the movement of GXG motifs to the interface due to the clustering of hydrophobic residues that prefer the condensate interior. Therefore, while microphase separation occurs in ELP condensates, frustration remains in the system. Hydrophilic residues cannot completely separate from hydrophobic ones due to constraints imposed by the amino acid sequence, creating unique microenvironments. Such a microenvironment arises from the collective behavior of many proteins and can deviate from individual chains (Wei et al., 2017), explaining the less-than-ideal correlation between the Urry scale and the hydrophobicity scales optimized to reproduce the conformational ensemble of IDP monomers.

The interfacial environment uncovered in the previous section suggests the presence of significant solvation in the condensate interior. For condensates with lower surface tension, such as V₅G₅, the protein density near the center of the condensate is only around ∼30%. Even for the stable condensates with high surface tension, such as V₅F₅, the protein density near the center only reaches ∼70%. The relative density of protein, water, and ions in different systems generally follows the trend observed for the transition temperature (Appendix 1—figures 1 and 2).

The large water mass fractions in the condensates prompt the question of the extent to which water is engaged in meeting the protein’s hydrogen bonding requirements.We determined the average number of hydrogen bonds per residue with ELP residues or water from extensive atomistic simulations. As shown in Figure 5A, over 75% of the hydrogen bonds are between protein and water in all of the condensates. Prior work on a different ELP condensate (Rauscher and Pomès, 2017) similarly indicated that water-protein hydrogen bonds constitute most of the hydrogen bonds found in the system. This finding highlights that ELPs, lacking noticeable secondary structures (Figure 5—figure supplement 1), are less capable of meeting its own hydrogen bonding needs, compared to folded proteins (Sticke et al., 1992). Therefore, high water concentration within the condensate is crucial to fulfilling the protein’s hydrogen bonding requirements.

Strikingly, given the importance of water molecules, the average number of hydrogen bonds per residue remained relatively constant across condensates, even when the amount of water inside the condensate varied by almost two folds. Therefore, we hypothesize that each water molecule in more hydrophobic condensates must engage more in hydrogen bonding than each water in hydrophilic condensates. To test this hypothesis, we determined the water hydrogen bond density, defined as the ratio of the number of protein-water hydrogen bonds divided by the total number of water molecules near the condensate (Figure 5D and E). As shown in Figure 5B and C, the water hydrogen bond density indeed negatively correlates with the condensate transition temperature, Tt, supporting our hypothesis that more hydrophobic condensates would have a higher hydrogen bond density.

The water hydrogen bond density also highlights an interfacial environment of blended hydrophobic and hydrophilic regions. For example, when we limit the analysis to water molecules near the guest residues, X, the hydrogen bond density now becomes positively correlated with Tt. Therefore, locally, protein-water interactions are strongly influenced by the chemistry of side chains, and water molecules have fewer hydrogen bonding opportunities with more hydrophobic residues. However, the opposite trend is observed for the hydrogen bond density near glycine residues that immediately surround the guest residues. Therefore, the exposed backbones retain water molecules even when hydrophobic side chains do not provide hydrogen bonding opportunities, and the inseparability between the two further contributes to the observed interfacial property of condensates.

As a complementary measure of the solvation environment of individual residues, we computed the relative solvent accessibility (RSA) for the guest amino acids using atomistic simulations (Figure 6A and B). The RSA measures the solvent-accessible surface area (SASA) normalized by the maximum possible solvent exposure. RSA values in ELP condensates are much lower than those computed from all-atom simulations of ELP monomers, consistent with the increase of polymer density upon phase separation. However, they are almost always higher than the values estimated for folded proteins due to many water molecules roaming inside the condensates (Tien et al., 2013). Therefore, while ELP condensates provide a mechanism to shield hydrophobic amino acids from solvent, they are less effective in doing so than folded proteins.

While the amino acid RSA values increase upon phase separation, unlike protein folding, the increase does not result from polymer collapse. We observe a significant expansion for almost all the studied peptides upon condensation, supporting the notion that water is a poor solvent for ELP (Figure 6C). The expansion is most evident for more hydrophobic ELPs, the radii of gyration of which are now comparable to that of an ideal Gaussian chain. Notably, the sizes of the two separate peptide blocks are similar in most systems, supporting an interfacial environment that can accommodate both hydrophobic and hydrophilic residues (Figure 6—figure supplement 1).

To better understand the properties of biological condensates, we performed simulations on 40 diblock ELPs with the sequence (V-P-G-V-G)_n-(V-P-G-X-G)_n. We set n=5 to balance computational efficiency and polymer topological effects. As shown in Figure 2—figure supplement 5, the results presented in the main text are qualitatively robust with respect to the system size and peptide length.

The multiscale simulations of ELP condensates started with an implicit solvent model, MOFF (Latham and Zhang, 2021), and a similar homopolymer model, to generate initial configurations for each system. Next, we converted the α-carbon-only configurations into those consistent with the MARTINI force field (Souza et al., 2021) for coarse-grained, explicit solvent simulations. These simulations measured the coexistence between the condensate and solvent, lasted for 100 μs, and were conducted in the NP_NAT ensemble. Finally, we converted the MARTINI configurations into atomistic structures for simulations with CHARMM36m (Huang et al., 2017) force field, with protein-water interactions shifted to better characterize IDPs. Atomistic simulations were conducted in the NP_NAT ensemble and lasted 250 ns. All simulations were conducted using the software GROMACS (Berendsen et al., 1995) and are described in detail in ‘Condensate simulation details’ in Appendix 1.

In addition to the condensates, we performed atomistic simulations of ELP monomers. These simulations began from RoseTTAFold predictions of the peptide structure (Baek et al., 2021; Bonneau et al., 2001) and lasted for 1 μs in the NPT ensemble. Full details are available in ‘Monomer simulation details’ in Appendix 1.

The protocol describes the preparation and manipulation of a sample containing a BODIPY-labeled ELP protein using an optical tweezer, specifically the C-Trap from LULICKS. The goal is to form a single large condensate from small droplets and then stretch it using the optical tweezers while monitoring its fluorescence intensity. The first step is to mix the BODIPY-labeled ELP protein with NaCl solution to obtain a final concentration of 20 μM protein in 2 M NaCl. The sample is then loaded into the flow channel of the optical tweezer. The dual trap mode of the C-Trap is then used to move the two traps to the same position. The sample is flushed until a single large condensate of about 4 μm is caught by the dual trap via the fusion of small droplets. The traps are then moved along with the droplet to another flow channel containing 2 M NaCl buffer. The channel and pressure are turned off to avoid any additional force on the droplet. The next step is to stretch the droplet by moving trap 1 0.2 μm each time. Fluorescent images of each stretch are taken using a 488 nm excitation laser. The mean fluorescence intensity of the whole droplet is then quantified using ImageJ.