Role of desolvation on biomolecular liquid–liquid phase separation Kai Zhang, Zhiyu Peng, Wenfei Li, Wei Wang, Bin Zhang, Qiang Cui Abstract Biomolecular condensates play essential roles in cellular organization and are implicated in diverse pathological processes. Their formation is driven by liquid–liquid phase separation (LLPS), a process that requires coordinated multistep desolvation of biomolecular chains and multivalent inter-chain interactions. Although coarse-grained (CG) models with implicit solvent are widely used to probe LLPS thermodynamics and kinetics, they typically neglect water-mediated desolvation effects, limiting their accuracy and mechanistic interpretability. Here, guided by all-atom simulations and experimental measurements, we develop a desolvation-aware implicit-solvent CG model by incorporating residue-level desolvation terms directly into the pairwise energy function, and apply it to investigate LLPS of intrinsically disordered proteins. Incorporating these desolvation interactions reshapes the phase diagram, alleviating dense-phase overcompaction. Notably, we observe an approximately linear correlation between the temperature gap (simulation temperature relative to the critical point) and the extent of conformational expansion accompanying the dilute-to-dense phase transition, a result further supported by theoretical analysis. We also find that desolvation barriers slow early density-fluctuation growth and shorten transient kinetic arrest, whereas solvent-separated contact interactions exert the opposite effects. Both terms further modulate chain mobility within mature condensates through competing packing and energy-landscape effects. Together, this framework enables an efficient representation of desolvation in CG simulations and reveals how desolvation energetics shape both the thermodynamic landscape and kinetic properties of biomolecular LLPS. Introduction Biomolecular condensates formed through liquid–liquid phase separation (LLPS) are increasingly recognized as fundamental regulators of diverse cellular processes, including stress responses (; ; ; ; ; ; ), cellular signaling (; ), gene expression regulation (; ), and chromatin organization (). Dysregulation of condensate formation, in contrast, has been implicated in a wide range of human diseases, most notably cancer and numerous neurodegenerative disorders (; ; ; ; ). Intrinsically disordered proteins (IDPs), which lack stable tertiary structures and often contain multivalent interaction motifs, are key molecular drivers of LLPS (; ; ; ; ). Understanding the molecular principles governing IDP-mediated LLPS is therefore essential for elucidating the mechanisms underlying both physiological condensate function and pathological phase transitions. A central yet incompletely understood aspect of LLPS is the multistep desolvation of biomolecular chains as they transition from a homogeneously dispersed solution to a condensed phase. How desolvation energetics contribute to the thermodynamics and kinetics of LLPS remains a major open question. Because experimentally characterizing the microscopic molecular events of LLPS at the submolecule level is still challenging for most experimental assays, a broad spectrum of computational approaches has been developed to characterize biomolecular phase separation. These range from mean-field (; ) and random-phase-approximation polymer theories (; ; ) to diverse molecular simulation frameworks based on atomistic or coarse-grained (CG) force fields (; ; ; ; ). Although all-atom molecular dynamics (MD) simulations can capture the microscopic processes of biomolecules with unprecedented temporal and spatial resolution, they become computationally prohibitive when applied to biomolecular phase separation involving a large number of protein chains. This challenge has motivated the development of CG models that reduce computational cost while retaining the essential physicochemical features governing phase behavior. These CG frameworks include early experiment-parameterized hydrophobicity-based models for IDPs (), the sticker-and-spacer paradigm and its lattice Monte Carlo implementation LASSI (; ; ), the hydrophobicity-scale (HPS) family and its refinements (; ; ; ), and Mpipi, which explicitly incorporates cation-ππ interactions (). Advanced multiscale models have also been pioneered to predict the coupled phase behavior of IDPs and chromatin with remarkable accuracy (). Higher-resolution or experimentally fine-tuned CG force fields (e.g., AWSEM-IDP [], Martini3 [; ], COCOMO [; ], AICG2+ [; ], and SIRAH []) further capture hydrogen bonding, secondary-structure propensity, and sequence-dependent packing. Notably, the maximum entropy optimized force field integrates maximum entropy biasing and energy landscape theory, successfully predicting the experimental dimensions and macroscopic phase separation of both ordered and disordered proteins (; ). More recently, Tesei et al. introduced the CALVADOS model, which optimizes residue-specific interaction parameters within the HPS framework using SAXS and paramagnetic relaxation enhancement (PRE) data across a large IDP dataset (; ; ), thereby improving the accuracy and transferability of CG models for predicting IDP conformations and LLPS behavior. Despite these advances, most residue-level CG models rely on implicit solvent representations, in which individual water molecules are not explicitly represented. This simplification can neglect the contribution of water molecules in mediating inter-residue interactions, which is crucial for the thermodynamics and kinetics of biomolecular folding and phase transitions (; ; ; ; ). In particular, the desolvation penalty associated with electrostatic and hydrophobic interactions exhibits marked differences between implicit and explicit solvent representations (; ). More importantly, conventional implicit-solvent CG models used for LLPS usually do not directly account for the multi-step desolvation process that accompanies the transition from a dilute solution to a dense condensate. Recent theoretical and experimental studies highlight that water reorganization and solvent-mediated entropy changes are not merely passive consequences of LLPS but active driving forces that shape condensate stability, internal structure, and dynamics (; ; ). Related studies on hydrostatic pressure effects have further suggested that desolvation barriers and solvent-separated minima can help rationalize pressure-modulated LLPS behaviors, including the pressure-dependent reentrant phase separation of α-elastin (; ). Nevertheless, implicit-solvent CG models remain widely used because they strike an effective balance between computational efficiency and residue-level resolution. Incorporating desolvation effects into such implicit-solvent models thus represents a particularly promising strategy for enhancing their physical realism without sacrificing tractability in LLPS simulations. Notably, residue-level desolvation terms have already been successfully employed in protein folding studies (; ; ; ; ; ; ; ). Within a structure-based modeling framework, Cheung et al. introduced a desolvation barrier separating the direct-contact minimum from the solvent-separated minimum in pairwise residue-residue potentials (). Simulations demonstrated that this minimalist desolvation model captures water-expulsion events during the final stages of folding. Chan and colleagues further showed that incorporating a desolvation barrier markedly enhances folding cooperativity (; ; ; ). These previous successes provide a strong foundation for incorporating desolvation effects into coarse-grained models to more realistically investigate the critical contribution of desolvation to LLPS. In this study, we advance the coarse-grained modeling of biomolecular LLPS by incorporating desolvation terms into the pairwise energy function of a CG model. By performing molecular simulations, we systematically investigate how desolvation influences phase behavior, thermodynamic properties, and the microscopic dynamics underlying condensate formation as well as the internal dynamics of the equilibrated dense phase. We further parameterize the desolvation terms using all-atom MD simulations and experimental Rg\begin{document}$R_{g}$\end{document} measurements, yielding a refined framework for investigating how generic desolvation features modulate the sequence-dependent phase behavior encoded by the underlying residue-level models. Results Effective representation of desolvation effects in coarse-grained models In order to elucidate the role of water molecules in mediating inter-residue interactions, we first performed all-atom MD simulations with explicit solvent for a set of amino acid analogues and computed the potential of mean force (PMF) along inter-residue distances (rr). Methane, methanol, acetate ion, ammonium ion, and methanamide were chosen to represent nonpolar, polar, negatively charged, positively charged, and backbone-like residues, respectively. These amino acid analogues cover the key features of most amino acid types. The simulations were performed under ambient conditions (298K\begin{document}$298\,\mathrm{K}$\end{document} and 1atm\begin{document}$1\,\mathrm{a}\mathrm{t}\mathrm{m}$\end{document}), and the sampled snapshots were then analyzed to determine the inter-molecule distances and the corresponding PMF profiles. Additional simulation details are provided in the ‘Methods’ section. shows the PMF profile for the methane system. In addition to the deep potential well corresponding to the direct contact between two methane molecules (r ~ 4Å), there exists a shallow potential well at a larger distance (r ~ 7Å). This shallow minimum arises from transient, solvent-separated pseudo contacts between the solute residues, hereafter referred to as solvent-separated contact. To form a direct contact between the solutes, the water molecules mediating the solvent-separated contacts need to be excluded, which leads to a desolvation barrier between the two potential wells. Similar features have been reported in previous all-atom MD simulations (). The PMF profiles extracted from the simulations for other amino acid analogues exhibit the same overall shape (, ), but the positions of the direct-contact minimum (rdc\begin{document}$r_{\mathrm{dc}}$\end{document}), the solvent-separated minimum (rss\begin{document}$r_{\mathrm{ss}}$\end{document}), and the desolvation barrier (rb\begin{document}$r_{\mathrm{b}}$\end{document}) vary depending on the analogue molecules. The depths of the two potential wells (εε and εss\begin{document}$\epsilon_{\mathrm{ss}}$\end{document}) and the height of the desolvation barrier (εb\begin{document}$\epsilon_{\mathrm{b}}$\end{document}) likewise differ across analogue pairs. The desolvation effect can also be modulated by solvent polarity and ionic conditions, which can be tuned experimentally by adding cosolvents and changing ionic strength, respectively. To examine the role of solvent polarity, we performed additional simulations of the methane system using water models with systematically varied polarity. For simplicity, the solvent polarity was tuned by uniformly scaling the partial charges of the hydrogen and oxygen atoms in the water molecules. This approach directly modulates the strength of solvent-solvent electrostatic interactions, which in turn dictates the stability of the hydration shell surrounding the solute. We compared the standard water model (1.0× charges) with reduced-polarity models (0.9× and 0.8× charges) (). The results reveal that a reduction in solvent polarity leads to a simultaneous decrease in the height of the desolvation barrier and the depth of the solvent-separated minimum. These observations underscore the importance of incorporating desolvation-related effective terms and exploring the effects of different desolvation strengths on the thermodynamics and kinetics of protein LLPS. In this work, we introduce desolvation terms into the HPS model developed by . In this framework, each residue is simplified to a single spherical bead. The HPS energy function includes a bonded term and a nonbonded term (see ‘Methods’). The nonbonded term is described by the short-range pairwise interactions and long-range electrostatic interactions. To represent the desolvation features, we added two Gaussian terms centered at rb\begin{document}$r_{\mathrm{b}}$\end{document} and rss\begin{document}$r_{\mathrm{ss}}$\end{document}, corresponding to the desolvation barrier and the solvent-separated minimum, respectively. The resulting nonbonded effective potential is given by(1)U(r)=Φ(r,ε,λ,σ)+Gb(r,εb,rb,σb)−Gss(r,εss,rss,σss)=Φ(r,ε,λ,σ)+εbexp⁡[−(r−rb)22σb2]−εssexp⁡[−(r−rss)22σss2].\begin{document}$$\displaystyle U(r)&= \Phi(r, \epsilon, \lambda, \sigma) + G_{\mathrm{b}}(r, \epsilon_{\mathrm{b}}, r_{\mathrm{b}}, \sigma_{\mathrm{b}}) - G_{\mathrm{ss}}(r, \epsilon_{\mathrm{ss}}, r_{\mathrm{ss}}, \sigma_{\mathrm{ss}}) \\&= \Phi(r, \epsilon, \lambda, \sigma) + \epsilon_{\mathrm{b}}\exp\left[-\frac{(r - r_{\mathrm{b}})^2}{2\sigma_{\mathrm{b}}^2}\right] - \epsilon_{\mathrm{ss}}\exp\left[-\frac{(r - r_{\mathrm{ss}})^2}{2\sigma_{\mathrm{ss}}^2}\right]. $$\end{document} Here, Φ(r,ε,λ,σ)\begin{document}$\Phi(r,\epsilon,\lambda,\sigma)$\end{document} denotes the nonbonded term in the original HPS model energy function (). The parameters λλ and σσ correspond to the hydrophobicity scale (HPS) values and effective bead sizes, respectively, for the 20 amino acids. The widths of the Gaussian terms were fixed at σb=σss=0.05nm\begin{document}$\sigma_{\mathrm{b}}=\sigma_{\mathrm{ss}}=0.05\,\mathrm{n}\mathrm{m}$\end{document}, while their centers were positioned at rb=rdc+0.15nm\begin{document}$r_{\mathrm{b}}=r_{\mathrm{dc}}+0.15\,\mathrm{n}\mathrm{m}$\end{document} and rss=rdc+0.30nm\begin{document}$r_{\mathrm{ss}}=r_{\mathrm{dc}}+0.30\,\mathrm{n}\mathrm{m}$\end{document}. These offsets approximately correspond to the radius and diameter of a water molecule, respectively. We first employed a homopolymer protein model with a chain length of 50 residues to illustrate the desolvation effect. For each residue, we use the parameter set extracted by averaging over the IDPs in the DisProt dataset (), which gives an amino acid mass of mresidue=110amu\begin{document}$m_{\mathrm{residue}}=110\,\mathrm{a}\mathrm{m}\mathrm{u}$\end{document} (atomic mass unit), a hydrophobicity parameter λ=0.640λ=0.640, and an effective bead size σ=0.536nm\begin{document}$\sigma=0.536\,\mathrm{n}\mathrm{m}$\end{document}. The parameter εb\begin{document}$\epsilon_{\mathrm{b}}$\end{document} controls the height of the desolvation barrier associated with the transition from solvent-separated to direct-contact configurations, whereas εss\begin{document}$\epsilon_{\mathrm{ss}}$\end{document} determines the depth of the local minimum corresponding to a solvent-separated contact. Together, these parameters shape the desolvation-inspired effective potential and modulate the statistical balance between direct-contact and solvent-separated configurations. By tuning the parameter values of εss\begin{document}$\epsilon_{\mathrm{ss}}$\end{document} and εb\begin{document}$\epsilon_{\mathrm{b}}$\end{document}, we generate energy functions with varying desolvation contributions (), enabling systematic investigation of how desolvation influences the phase behavior and thermodynamics of IDPs. Because the effective interaction parameters used in the model are temperature-independent, the absolute simulation temperature should not be directly interpreted as an experimental temperature. We therefore report all thermodynamic quantities, unless explicitly stated otherwise, in terms of a reduced temperature, defined as T∗=kBT/ε\begin{document}$T^{*}=k_{\mathrm{B}}T/\epsilon$\end{document}. Here, εε denotes the characteristic interaction energy scale of the system, which is set to ε=0.2kcal/mol\begin{document}$\epsilon=0.2\,\mathrm{k}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{/} \mathrm{m}\mathrm{o}\mathrm{l}$\end{document} following the convention of the HPS model (). Accordingly, T∗\begin{document}$T^{*}$\end{document} should be interpreted primarily as a model temperature that controls the relative strength of thermal fluctuations, rather than as having a direct quantitative correspondence with experimental temperature. Thermodynamic impact of desolvation on the phase diagram To capture the essential physics of IDP phase separation, we employed a coarse-grained model of 125 homopolymer chains using a slab simulation protocol (see ‘Methods’). As a control, we first performed molecular dynamics simulations within the HPS model framework without considering the desolvation terms across a range of temperatures (). Representative snapshots clearly illustrate the transition from a homogeneous solution to a stable condensate as the temperature decreases (). We extracted time-averaged density profiles along the zz-axis to identify coexisting dense and dilute phases and thereby constructed the binodal curve (), which reproduces phase behavior reported in previous molecular simulation studies (). We then repeated the slab simulations using desolvation-inspired effective potentials with varying εb\begin{document}$\epsilon_{\mathrm{b}}$\end{document} and εss\begin{document}$\epsilon_{\mathrm{ss}}$\end{document} to investigate how desolvation reshapes the phase diagram and condensate structure. To elucidate how desolvation regulates the thermodynamics of LLPS, we systematically mapped phase diagrams under varying desolvation parameters (). Increasing the desolvation barrier εb\begin{document}$\epsilon_{\mathrm{b}}$\end{document} monotonically lowers the critical temperature Tc∗\begin{document}$T_{c}^{*}$\end{document} (), suggesting a reduced phase separation propensity. Analysis of residue-residue radial distribution functions showed that higher εb\begin{document}$\epsilon_{\mathrm{b}}$\end{document} suppresses the population of configurations near the barrier region (). This reduction in the statistical weight of barrier-region configurations can be interpreted as an entropy-related configurational restriction and thus disfavors phase separation. At the pair-potential level, the desolvation barrier modifies the equilibrium Boltzmann weight and thereby alters the integrated effective attraction, as quantified by the bead-level second virial coefficient B2\begin{document}$B_{2}$\end{document}. Specifically, increasing εb\begin{document}$\epsilon_{\mathrm{b}}$\end{document} makes B2/σ3\begin{document}$B_{2}/\sigma^{3}$\end{document} larger (), indicating a weaker integrated effective attraction and providing a thermodynamic basis for the lower Tc∗\begin{document}$T_{c}^{*}$\end{document}. By contrast, deepening the solvent-separated well εss\begin{document}$\epsilon_{\mathrm{ss}}$\end{document} elevates Tc∗\begin{document}$T_{c}^{*}$\end{document} (), which is associated with the enhanced population of solvent-separated configurations and a smaller B2/σ3\begin{document}$B_{2}/\sigma^{3}$\end{document} (). This solvent-separated minimum provides effective free-energy stabilization for water-mediated configurations and thereby promotes phase separation. These opposing effects suggest that the desolvation potential regulates macroscopic phase behavior by redistributing residue-pair configurations between direct-contact, barrier-region, and solvent-separated states (lower panels of ). Because the critical temperature is an intrinsic, sequence-specific property of a given system, it is also important to examine how desolvation strength alters the overall shape of the phase diagram when the temperature is normalized by the corresponding Tc∗\begin{document}$T_{c}^{*}$\end{document}. To do so, we rescaled the temperature axis by Tc∗\begin{document}$T_{c}^{*}$\end{document}, which effectively normalizes overall energetic strength and isolates desolvation-specific effects on the phase boundary. For the desolvation models with different barrier heights εb\begin{document}$\epsilon_{\mathrm{b}}$\end{document}, the normalized binodal curves nearly overlap (). By contrast, varying the solvent-separated well depth εss\begin{document}$\epsilon_{\mathrm{ss}}$\end{document} results in distinct deviations in the density of the condensed phase (). Specifically, stronger water-mediated interactions lower equilibrium density by biasing the ensemble toward configurations with larger inter-residue separations (). Thus, solvent-separated interactions directly regulate molecular packing and, by introducing solvent-separated configurations, alleviate the overcompaction typically encountered in coarse-grained models with implicit solvent. These results suggest that incorporating desolvation effects may enable the model to better capture the physicochemical structure of biological condensates. Conformational modulation by desolvation in dense and dilute phases We next investigated how desolvation modulates the conformational distribution of IDPs across dilute and condensed phases. Previous studies have shown that IDP condensation can reorganize chain conformations by redistributing the balance between intra-chain and inter-chain interactions (; ; ; ; ). Related work has further linked dilute-phase conformations to condensate material properties (). To quantify global chain dimensions, we calculated the radius of gyration (Rg\begin{document}$R_{g}$\end{document}) as a primary structural metric. Consistent with this picture, the resulting Rg\begin{document}$R_{g}$\end{document} distributions show that protein chains adopt more compact conformations in the dilute phase, whereas chains in the condensed phase exhibit more extended conformations (). This conformational expansion upon condensation reflects a transition from intra-chain-dominated interactions in the dilute phase to inter-chain-dominated interactions in the dense phase (). Within the condensed phase, protein chains preferentially form inter-chain contacts that lower the free energy of the system, compensating for the entropic cost associated with reduced translational and conformational freedom. In addition to the difference between dilute- and dense-phase conformations, varying the desolvation parameters further reveals a phase-dependent conformational response (). Increasing εb\begin{document}$\epsilon_{\mathrm{b}}$\end{document} or decreasing εss\begin{document}$\epsilon_{\mathrm{ss}}$\end{document} shifts the dilute-phase Rg\begin{document}$R_{g}$\end{document} distributions toward larger values, whereas the dense-phase Rg\begin{document}$R_{g}$\end{document} remains comparatively insensitive to these parameter changes. This contrast suggests that desolvation more strongly affects chain dimensions in the dilute phase, while conformations in the dense phase are less responsive, likely reflecting the constraints imposed by intermolecular packing in the condensate. As a result, the desolvation-dependent variation in ΔRg=Rgdense−Rgdilute\begin{document}$\Delta R_{g}=R_{g}^{\mathrm{dense}}-R_{g}^{\mathrm{dilute}}$\end{document} arises predominantly from the conformational changes of isolated chains in the dilute phase. Physically, ΔRg\begin{document}$\Delta R_{g}$\end{document} captures the transition from intramolecular-dominated collapse in the dilute phase to intermolecular-mediated expansion in the dense phase, thus serving as a metric for the relative strength of inter-chain versus intra-chain interactions. To further elucidate the thermodynamic implications of desolvation-modulated conformations, we analyzed the relationship between the conformational shift ΔRg\begin{document}$\Delta R_{g}$\end{document} and the thermal distance to the critical point, defined as Tc∗−Tsim∗\begin{document}$T_{c}^{*}-T_{\mathrm{sim}}^{*}$\end{document}. Notably, data from the simulated systems approximately follow a common trend, revealing a strong correlation between the magnitude of conformational change and the thermal distance to the phase transition point (R2=0.942\begin{document}$R^{2}=0.942$\end{document}, ). This result suggests that the conformational response to phase separation is closely associated with how far the system resides thermally from the critical point. We rationalize this observation using the Flory–Huggins theory (; ) as a bridge between the macroscopic thermal distance and the microscopic interaction strength. In this theoretical framework, the thermodynamic state is described by the interaction parameter χ=εeff/(kBT)\begin{document}$\chi=\epsilon_{\mathrm{eff}}/(k_{B}T)$\end{document}, where εeff\begin{document}$\epsilon_{\mathrm{eff}}$\end{document} represents the effective pairwise energy scale specific to each desolvation model (). At the critical point, χ(Tc)=χc\begin{document}$\chi(T_{c})=\chi_{c}$\end{document}, which gives εeff=kBTcχc\begin{document}$\epsilon_{\mathrm{eff}}=k_{B}T_{c}\chi_{c}$\end{document}. Substituting this relation into the expression for χ(Tsim)\begin{document}$\chi(T_{\mathrm{sim}})$\end{document} yields χ(Tsim)=χcTc/Tsim\begin{document}$\chi(T_{\mathrm{sim}})=\chi_{c}T_{c}/T_{\mathrm{sim}}$\end{document}. Subtracting χc\begin{document}$\chi_{c}$\end{document} from both sides gives(2)χ(Tsim)−χc=χcTc−TsimTsim.\begin{document}$$\displaystyle \chi(T_{\mathrm{sim}})-\chi_{c}=\chi_{c}\frac{T_{c}-T_{\mathrm{sim}}}{T_{\mathrm{sim}}}.$$\end{document} For systems with the same chain length, χc\begin{document}$\chi_{c}$\end{document} is a fixed constant. Thus, the deviation from the critical interaction parameter is directly related to the rescaled thermal distance (Tc−Tsim)/Tsim\begin{document}$(T_{c}-T_{\mathrm{sim}})/T_{\mathrm{sim}}$\end{document}. The thermodynamic driving force χ(Tsim)−χc\begin{document}$\chi(T_{\mathrm{sim}})-\chi_{c}$\end{document} can then be related to the structural observable ΔRg\begin{document}$\Delta R_{g}$\end{document}. Since ΔRg\begin{document}$\Delta R_{g}$\end{document} captures the structural transition from an intrachain-interaction-dominated state in the dilute phase to an interchain-interaction-dominated state in the dense phase, we assume, as a first-order approximation, that this conformational shift responds approximately linearly to the excess interaction strength, expressed as ΔRg∝[χ(Tsim)−χc]\begin{document}$\Delta R_{g}\propto\left[\chi(T_{\mathrm{sim}})-\chi_{c}\right]$\end{document}. Combining these relations gives(3)Tc−Tsim∝TsimΔRg.\begin{document}$$\displaystyle T_{c}-T_{\mathrm{sim}}\propto T_{\mathrm{sim}}\Delta R_{g}. $$\end{document} This relation explains the approximate linearity observed in , while also accounting for the systematic deviations caused by the temperature factor Tsim\begin{document}$T_{\mathrm{sim}}$\end{document}, which is evident from the varying slopes at different simulation temperatures. To test this theoretical argument and the underlying linear-response assumption, we replotted the data against the rescaled thermal distance (Tc∗−Tsim∗)/Tsim∗\begin{document}$(T_{c}^{*}-T_{\mathrm{sim}}^{*})/T_{\mathrm{sim}}^{*}$\end{document}. As shown in the inset of , this normalization yields a stronger linear relationship, supporting the above argument that ΔRg\begin{document}$\Delta R_{g}$\end{document} scales with the excess interaction strength χ−χc\begin{document}$\chi-\chi_{c}$\end{document}, and further suggesting that the conformational response is closely associated with the rescaled thermal distance from the critical point. We further show the relationship between the phase density difference (Δρ=ρdense−ρdilute\begin{document}$\Delta\rho=\rho_{\mathrm{dense}}-\rho_{\mathrm{dilute}}$\end{document}) and the conformational difference (ΔRg\begin{document}$\Delta R_{g}$\end{document}) between coexisting phases in . Based on the empirical correlation between ΔRg\begin{document}$\Delta R_{g}$\end{document} and the thermal distance from the critical point, one can relate the density contrast Δρ\begin{document}$\Delta\rho$\end{document} to ΔRg\begin{document}$\Delta R_{g}$\end{document} through the critical-scaling relation by Δρ=A(Tc−T)β\begin{document}$\Delta\rho=A(T_{c}-T)^{\beta}$\end{document} (), where ββ is the critical exponent. Although the complete relation in contains an additional Tsim\begin{document}$T_{\mathrm{sim}}$\end{document} factor, the unscaled quantities remain strongly correlated over the simulated range. We therefore use Tc−Tsim∝ΔRg\begin{document}$T_{c}-T_{\mathrm{sim}}\propto\Delta R_{g}$\end{document} as an empirical scaling approximation. Using this approximation in the critical-scaling relation yields Δρ∝(ΔRg)β\begin{document}$\Delta\rho\propto(\Delta R_{g})^{\beta}$\end{document}. On logarithmic scales, a clear power-law dependence emerges (), with Δρ∝ΔRg0.4\begin{document}$\Delta\rho\propto\Delta R_{g}^{0.4}$\end{document}, which is consistent with our theoretical expectation. The fitted exponent (≈ 0.4) is in reasonable agreement with the theoretical critical exponent β=0.325β=0.325 based on the 3D Ising model (). Collectively, these results reveal a fundamental coupling between microscopic conformational reorganization and macroscopic packing density during biomolecular condensation, highlighting conformational reorganization as a structural signature of the phase transition. Dynamic consequences of desolvation in phase separation Beyond governing thermodynamic phase boundaries and conformational distribution, the desolvation potential fundamentally reshapes the dynamical landscape of IDP condensates. The interplay between inter-chain interactions and solvent exclusion not only dictates the stability of the dense phase but also modulates the transport properties within the condensate and the macroscopic kinetics of phase separation. To probe chain mobility within the dense phase, we employed a strategy analogous to fluorescence recovery after photobleaching (FRAP) experiments following prior molecular dynamics simulation work (), tracking the spatiotemporal evolution of specific chains initially localized at the slab center (). The systems were compared at the same simulation temperature Tsim∗\begin{document}$T_{\mathrm{sim}}^{*}$\end{document}. For the standard HPS model without desolvation, the tagged chains achieve a uniform distribution that matches the equilibrium density profile within 50ns\begin{document}$50\,\mathrm{n}\mathrm{s}$\end{document} (), consistent with the liquid-like nature of the condensate. For the representative desolvation model, the spatial distribution of tracked chains remains biased toward the initial center position after 50ns\begin{document}$50\,\mathrm{n}\mathrm{s}$\end{document} and approaches equilibrium on a prolonged timescale (∼150ns\begin{document}$\sim150\,\mathrm{n}\mathrm{s}$\end{document}). These contrasting relaxation profiles demonstrate that desolvation alters chain mobility within the condensate. To more quantitatively assess the influence of desolvation on chain mobility, we computed the time-averaged mean squared displacement (TAMSD) () for chains within the slab. The chain motion exhibits sub-diffusive scaling, TAMSD=K⋅Δtα\begin{document}$\mathrm{TAMSD}=K\cdot\Delta t^{\alpha}$\end{document}, with a characteristic exponent α≈0.91\begin{document}$\alpha\approx 0.91$\end{document} (, bottom-right panel). This deviation from normal diffusion (α=1α=1) likely reflects the combined effects of molecular crowding, multivalent dynamic interaction networks, and finite-size confinement within the dense phase. At extended lag times, the TAMSD curves naturally plateau as they approach the geometric limit imposed by the slab thickness. Notably, the observed exponent is comparable to simulation results reported for protein diffusion in three-dimensional condensates (α≈0.85\begin{document}$\alpha\approx 0.85$\end{document}) (), consistently highlighting the restricted nature of chain dynamics within the dense phase. Given the relatively consistent values of αα across different interaction parameters, we use KK as a quantitative measure of the chain diffusion rate. The diffusion prefactor KK exhibits a clear dependence on the interaction parameters (). At a fixed temperature, KK increases with the desolvation barrier εb\begin{document}$\epsilon_{\mathrm{b}}$\end{document} and decreases with the solvent-separated minimum depth εss\begin{document}$\epsilon_{\mathrm{ss}}$\end{document} (, left panels). These opposing trends can be largely understood from the distinct effects of the two desolvation parameters on condensate packing. As illustrated in (right panel), KK is predominantly correlated with the bulk density (ρdense\begin{document}$\rho_{\mathrm{dense}}$\end{document}), with further modulation by temperature. As expected, higher-temperature systems generally exhibit slightly larger KK values at comparable densities. Consistent with the density trends established in , a higher εss\begin{document}$\epsilon_{\mathrm{ss}}$\end{document} promotes tighter packing and increases the effective friction, whereas a higher εb\begin{document}$\epsilon_{\mathrm{b}}$\end{document} prevents close contacts and creates greater free volume that facilitates chain motion. These results indicate that, at a fixed absolute temperature, changes in dense-phase packing serve as the primary determinant of chain mobility. However, this density-dominated view is incomplete because it convolutes intrinsic energy-landscape features with thermodynamic state shifts. Since variations in εb\begin{document}$\epsilon_{\mathrm{b}}$\end{document} and εss\begin{document}$\epsilon_{\mathrm{ss}}$\end{document} significantly alter the critical temperature, comparisons made at a constant simulation temperature inherently involve systems positioned at different thermodynamic quench depths, defined as the normalized temperature deviation from the critical point. This highlights the distinct thermodynamic driving forces for different systems. To reduce this thermodynamic-state dependence, we performed a more carefully controlled comparison. First, we scaled the simulation temperature by the critical temperature of each system and used Tsim∗/Tc∗\begin{document}$T^{*}_{\mathrm{sim}}/T_{\mathrm{c}}^{*}$\end{document} as the control parameter for quench depth, with a lower value corresponding to a deeper quench. This provides more comparable thermodynamic conditions across the different interaction parameters. Second, to account for the leading linear temperature dependence associated with Langevin dynamics, we defined a reduced diffusion prefactor K~=K/T\begin{document}$\widetilde{K}=K/T$\end{document}, as detailed in the ‘Methods’ section. Under these controlled conditions, a clearer parameter-dependent kinetic trend emerges (). At the same quench depth, a higher desolvation barrier εb\begin{document}$\epsilon_{\mathrm{b}}$\end{document} (darker blue points) lowers K~\begin{document}$\widetilde{K}$\end{document}, consistent with a rougher energy landscape that hinders the configurational rearrangements required for chain diffusion. This kinetic divergence becomes more pronounced at deeper quenches but largely vanishes near the critical point, indicating that the kinetic influence of the desolvation barrier depends on the thermodynamic state. Notably, the standard HPS model exhibits much faster temperature-normalized mobility (, red triangles), suggesting that omitting desolvation energetics may lead to an overestimation of chain mobility due to reduced roughness of the effective energy landscape. Together, the fixed-temperature and renormalized analyses in reveal two distinct and opposing contributions of desolvation to condensate dynamics. At fixed temperature, increasing εb\begin{document}$\epsilon_{\mathrm{b}}$\end{document} loosens dense-phase packing and thereby increases the measured diffusion coefficient, whereas at matched thermodynamic quench depth, the same parameter change suppresses chain mobility by roughening the microscopic energy landscape and slowing local rearrangements. Condensate dynamics therefore emerge from a balance between density-regulated mobility and energy-landscape-regulated mobility, with macroscopic packing determining the dominant trend and microscopic barrier roughness imposing an additional kinetic modulation. This interplay highlights how desolvation reshapes condensate dynamics across multiple physical scales. Desolvation-modulated kinetic arrest of coarsening dynamics Elucidating the molecular events during the early stage of phase separation is challenging experimentally. We therefore examined droplet-growth dynamics at the early stage of phase separation by performing molecular simulations with desolvation-aware effective potentials. Initially, the system was equilibrated at a supercritical temperature (T∗=2.98\begin{document}$T^{*}=2.98$\end{document}) for 10ns\begin{document}$10\,\mathrm{n}\mathrm{s}$\end{document} to ensure a homogeneous distribution. The temperature was then quenched to T∗=1.79\begin{document}$T^{*}=1.79$\end{document}, located deep within the two-phase coexisting region, to initiate spontaneous phase separation (). Here, tsim=0ns\begin{document}$t_{\mathrm{sim}}=0\,\mathrm{n}\mathrm{s}$\end{document} denotes the first snapshot immediately after the temperature quench, corresponding to a homogeneous but thermodynamically unstable non-equilibrium state. During the subsequent time evolution, the initially uniform distribution rapidly develops interconnected density fluctuations within 1−2 ns, which then break up into discrete high-density domains along the zz-axis. Over time, these domains gradually coalesce into larger slabs, exhibiting the characteristic morphological evolution of spinodal decomposition (). To quantify the phase separation kinetics, we adopted a two-stage metric strategy. In the early stage, where interfaces are diffusive, we monitored the normalized density variance (⟨σ^ρ2⟩\begin{document}$\langle\hat{\sigma}^{2}_{\rho}\rangle$\end{document}, see Methods) derived from the instantaneous density profile ρ(z)ρ(z) to track the growth rate of density fluctuations. As clear interfaces emerged, we switched to tracking the average slab size (⟨ξ⟩\begin{document}$\langle\xi\rangle$\end{document}), defining dense domains as continuous regions where ρ(z)ρ(z) exceeds a threshold value (here the average system density was used as the threshold). To ensure statistical robustness, all kinetic metrics were averaged over six independent slab simulation replicas. The resulting kinetic profiles () resolve a three-stage dynamical pathway. In the first stage, immediately after the temperature quench, the system undergoes a rapid exponential growth in density variance (top panels). This behavior arises from the growth of initial density fluctuations in a thermodynamically unstable state and is characteristic of spinodal decomposition (). Stronger effective attraction, corresponding to higher εss\begin{document}$\epsilon_{\mathrm{ss}}$\end{document} or lower εb\begin{document}$\epsilon_{\mathrm{b}}$\end{document}, accelerates these initial fluctuations, as quantified by the monotonic decrease in the characteristic time scale τ1/2\begin{document}$\tau_{1/2}$\end{document}, defined as the time required to reach 50% of the maximum normalized variance. Intriguingly, after dense domains have formed, the system enters an intermediate plateau regime in which domain growth is temporarily arrested before late-stage coarsening begins. This kinetic arrest is evident from the slab-size evolution in , where ⟨ξ⟩\begin{document}$\langle\xi\rangle$\end{document} remains nearly unchanged for a distinct period, and is visually corroborated by the density profiles at 5ns\begin{document}$5\,\mathrm{n}\mathrm{s}$\end{document} in , where distinct domains are observed to form but fail to fuse immediately. This behavior is consistent with the characteristics of viscoelastic phase separation (). In this regime, interfacial tension favors domain fusion, whereas transient inter-chain connectivity within the dense domains generates viscoelastic resistance to the shape deformation required for coalescence, thereby delaying domain fusion. Similar viscoelastic effects have recently been quantified in molecular simulations of biomolecular condensates using rheological analyses of time-dependent material properties (). Molecular simulations of condensate aging have further highlighted the roles of chain flexibility, sticker lifetime, and desolvation-associated rigidification in promoting more solid-like states (). In the zoom-in schematic in , this transient network is represented by connections between residues involved in inter-chain contacts, illustrating how multivalent interactions can resist domain deformation during the plateau regime. Quantitative analysis in further reveals that the duration of this plateau regime depends strongly on the desolvation parameters. Systems with higher εss\begin{document}$\epsilon_{\mathrm{ss}}$\end{document} or lower εb\begin{document}$\epsilon_{\mathrm{b}}$\end{document} exhibit a longer plateau, indicating that stronger inter-chain attractions reinforce the transient network and prolong kinetic arrest. This trend corroborates our viscoelastic interpretation, as enhanced intermolecular connectivity increases the resistance to domain deformation and fusion. This kinetic trap is particularly pronounced in the standard HPS model (red trajectories in ), consistent with its relatively stronger pairwise attractions. At later times (t>10ns\begin{document}$t > 10\,\mathrm{n}\mathrm{s}$\end{document}), the system leaves the plateau regime and enters a coarsening-and-growth stage, where domain fusion proceeds through interfacial-tension-driven relaxation and Brownian aggregation. The average slab size follows a power-law scaling ⟨ξ⟩∼tα\begin{document}$\langle\xi\rangle\sim t^{\alpha}$\end{document}, with α≈0.26\begin{document}$\alpha\approx 0.26$\end{document} in our simulations. This exponent is slightly lower than the asymptotic prediction of Brownian motion coalescence, α=1/3α=1/3 (), likely because the viscoelastic nature of the polymer-rich phase and the sub-diffusive motion of individual chains (TAMSD∼Δt0.91\begin{document}$\mathrm{TAMSD}\sim\Delta t^{0.91}$\end{document}) constrain macroscopic domain diffusion (). Collectively, these analyses reveal a desolvation-modulated kinetic pathway that links early density fluctuations, transient domain arrest, and late-stage coarsening into a unified dynamical picture, as summarized by the vertical schematic in and the mechanistic illustration in . This picture shows that desolvation does more than slow down local chain rearrangements through an added barrier. It also regulates the balance between fluctuation growth, transient arrest, and domain coarsening, thereby shaping the evolution of phase-separated domains. The sensitivity of kinetic arrest and coarsening dynamics to desolvation parameters underscores the importance of incorporating desolvation features into coarse-grained potentials for more physically plausible molecular simulations of LLPS, especially when connecting microscopic interaction lifetimes to emergent viscoelastic or aging-like material behavior. Beyond modulating LLPS thermodynamics and conformational ensembles as discussed above, desolvation therefore acts as a critical regulator of phase-separation dynamics. Parameterization of desolvation model for IDPs The results from the homopolymer model discussed above underscore the critical role of desolvation in regulating both the thermodynamics and dynamics of LLPS. Consequently, incorporating water-mediated interactions into residue-level coarse-grained models is essential for capturing the physics of protein phase separation more realistically. To this end, we developed a transferable parameterization strategy that combines information derived from all-atom simulations with calibration against experimental observables of IDPs. Directly determining residue-specific desolvation parameters, specifically the barrier height εb\begin{document}$\epsilon_{\mathrm{b}}$\end{document} and the solvent-separated minimum depth εss\begin{document}$\epsilon_{\mathrm{ss}}$\end{document}, remains challenging. To bridge this gap, we leveraged all-atom molecular dynamics simulations of amino acid analogues. Using the global interaction energy scale εε of the underlying HPS-type potential as a reference, we expressed the desolvation amplitudes as εb=αbε\begin{document}$\epsilon_{\mathrm{b}}=\alpha_{\mathrm{b}}\epsilon$\end{document} and εss=αssε\begin{document}$\epsilon_{\mathrm{ss}}=\alpha_{\mathrm{ss}}\epsilon$\end{document}. By fitting the PMF profiles obtained from all-atom simulations (), we obtained coefficients of comparable magnitude across the limited set of analogue pairs examined. For simplicity and transferability, we therefore adopted the averaged values αb=0.33\begin{document}$\alpha_{\mathrm{b}}=0.33$\end{document} and αss=0.06\begin{document}$\alpha_{\mathrm{ss}}=0.06$\end{document} as a global baseline for our coarse-grained model. This uniform parameterization captures the generic desolvation features of the PMFs but does not resolve residue-pair-specific variations in desolvation energetics. To evaluate the effectiveness of these physics-derived parameters, we first integrated them into the framework of the standard HPS model. While the HPS model performs remarkably well in capturing sequence-dependent properties of LLPS, previous studies have noted that parameters optimized solely against single-chain Rg\begin{document}$R_{g}$\end{document} may have limited transferability to the condensed phase, resulting in an overestimation of phase-separation propensity, as reflected by higher critical temperatures (Tc\begin{document}$T_{c}$\end{document}) (). We assessed the model performance on a dataset of 47 IDPs taken from , which was subsequently used to calibrate the CALVADOS2-desolvation parameters. The protein sequences and experimental Rg\begin{document}$R_{g}$\end{document} measurements used for this calibration were reused from the published CALVADOS2 resource (). The deviation of simulated single-chain Rg\begin{document}$R_{g}$\end{document} from experimental SAXS data was quantified by χRg2=[(Rg,exp−Rg,sim)/σexp]2\begin{document}$\chi_{R_{g}}^{2}=[(R_{g,\mathrm{exp}}-R_{g,\mathrm{sim}})/\sigma_{\mathrm{exp} }]^{2}$\end{document}, where σexp\begin{document}$\sigma_{\mathrm{exp}}$\end{document} represents the experimental uncertainty. When the desolvation parameters derived from all-atom simulations (αb=0.33,αss=0.06\begin{document}$\alpha_{\mathrm{b}}=0.33,\alpha_{\mathrm{ss}}=0.06$\end{document}) were added to the HPS model without modifying its original εε, we observed a substantial improvement in conformational accuracy. The average discrepancy ⟨χRg2⟩\begin{document}$\langle\chi_{R_{g}}^{2}\rangle$\end{document} decreased drastically from 1,082.27 (original HPS) to 85.88 (HPS-desolvation) (). This improvement indicates that introducing the desolvation terms effectively mitigates the over-compaction tendency of the original HPS model, likely because of competition between direct and water-separated contacts. We also incorporated the desolvation terms into CALVADOS2 (), a force field whose hydrophobicity parameters (λλ) were reparameterized using Bayesian optimization to reproduce experimental Rg\begin{document}$R_{g}$\end{document} and PRE data. Directly adding desolvation terms to CALVADOS2 led to substantial chain over-expansion (⟨χRg2⟩=643.30\begin{document}$\langle\chi_{R_{g}}^{2}\rangle=643.30$\end{document}), reflecting the fact that CALVADOS2 is already finely tuned to reproduce experimental data through a top-down optimization procedure, and the inclusion of additional desolvation terms is incompatible with its parameterization. To address this, we implemented a systematic optimization strategy in which the desolvation barrier height (εb\begin{document}$\epsilon_{\mathrm{b}}$\end{document}), solvent-separated minimum depth (εss\begin{document}$\epsilon_{\mathrm{ss}}$\end{document}), and the overall energy factor εε were scanned over broad parameter ranges to identify the parameter set that minimizes deviations between simulated and experimental Rg\begin{document}$R_{g}$\end{document} values. The full optimization workflow is summarized in the flowchart in , and the resulting error landscapes are shown in and . The optimized parameter set () reduced ⟨χRg2⟩\begin{document}$\langle\chi_{R_{g}}^{2}\rangle$\end{document} to 9.54 on the calibration dataset (), yielding agreement with experimental Rg\begin{document}$R_{g}$\end{document} values comparable to that of the original CALVADOS2 model. After parameterization of the energy scale, we further evaluated the calibrated model by probing the phase behavior of the FUS low-complexity domain (FUS LC) as a test case. Experimental observations indicate that the critical temperature is around 25 °C (). While the original HPS model yields a critical temperature of Tc∼340K\begin{document}$T_{c}\sim340\,\mathrm{K}$\end{document}, incorporating the desolvation terms into the HPS framework (HPS-desolvation) lowers the model Tc\begin{document}$T_{c}$\end{document} to ∼290K\begin{document}$\sim290\,\mathrm{K}$\end{document} (), shifting it toward the experimentally observed room-temperature regime. It is worth noting that the residue-level coarse-grained models used here employ temperature-independent effective interaction parameters. As a result, the temperature values reported here cannot be interpreted as quantitatively equivalent to experimental temperatures, particularly when they deviate substantially from room-temperature conditions. Beyond this reduction in the excessive phase-separation propensity, we observed that the desolvation model mitigates the overestimation of condensate density common in implicit solvent coarse-grained simulations (; ). In the CALVADOS2 framework, the inclusion of desolvation reduces the dense phase density while maintaining an overall phase boundary comparable to the original model (). Similarly, for the HPS model, when we rescaled the interaction strength εε to match the single-chain Rg\begin{document}$R_{g}$\end{document} and thereby align the phase boundary with the HPS-desolvation system (), we observed the same behavior. A similar trend was observed in the simulations of the LAF-1 RGG domain, where CALVADOS2-desolvation preserved a comparable coexistence curve while modestly reducing the dense-phase density relative to the original CALVADOS2 model (). Quantitatively, across both frameworks, the desolvation-modulated models consistently yield a packing density approximately 5% lower than the original models. This distinction underscores a limitation of standard Lennard–Jones potentials where a single parameter εε simultaneously controls chain dimension, overall phase-separation propensity, and condensate density. In contrast, our model introduces tunable parameters εb\begin{document}$\epsilon_{\mathrm{b}}$\end{document} and εss\begin{document}$\epsilon_{\mathrm{ss}}$\end{document} that provide an additional physical handle for modulating dense-phase packing without simply rescaling the overall interaction strength. The observed density reduction can be attributed to an increased population of solvent-separated configurations relative to direct contacts. This effect provides an effective representation of hydration-associated packing in protein condensates, thereby offering a more physically realistic description of IDP phase separation. Discussion Coarse-grained molecular simulations have become indispensable for elucidating the molecular principles underlying LLPS, complementing both theoretical frameworks and experimental observations. While residue-level implicit-solvent models have achieved remarkable success in capturing sequence-dependent phase behavior, they usually do not separately account for desolvation effects associated with water-mediated residue-residue association. As a result, they often face a trade-off when simultaneously describing single-chain dimensions, phase-separation propensity, and dense-phase packing. Specifically, interaction parameters calibrated to match single-chain dimensions or phase-separation propensity frequently result in an overestimation of the condensed phase density. This limitation suggests that standard Lennard–Jones potentials, which couple overall phase-separation propensity and dense-phase packing through a single energetic parameter, may lack the physical dimensionality required to fully describe the thermodynamics of hydrated protein condensates. The present results therefore suggest that desolvation is not merely a microscopic correction to residue-residue contacts, but an additional physically motivated component in coarse-grained interaction models that can modulate LLPS thermodynamics, dense-phase packing, conformational reorganization, and condensate dynamics in partially separable ways. To address this challenge, we developed a desolvation-aware implicit-solvent CG framework that incorporates desolvation barrier and solvent-separated terms into the pairwise potential. Application of this model to homopolymer systems demonstrates that desolvation substantially reshapes phase behavior. Incorporating desolvation modulates the critical temperature and alters the phase diagram by regulating intermolecular packing within the dense phase. Moreover, the model reveals a tight coupling between desolvation-controlled chain compaction during the formation of the condensate phase and the thermal distance from criticality (temperature gap), providing a theoretical basis for linking single-chain conformational changes (ΔRg\begin{document}$\Delta R_{g}$\end{document}) to macroscopic phase behavior. Dynamic analysis further captures the three characteristic stages of early LLPS, including spinodal decomposition, kinetic arrest, and coarsening. Crucially, it reveals how the desolvation effect regulates the kinetics of these stages by modulating the effective inter-chain interactions and the roughness of the energy landscape. These results underscore the importance of accounting for desolvation energetics in governing both the thermodynamics and kinetics of LLPS. These findings may also provide a useful physical basis for future studies of pressure-dependent condensate behavior, as pressure-induced changes in hydration, solvent-separated states, and desolvation barriers have been proposed to contribute to pressure-modulated and reentrant LLPS (; ; ). Motivated by these insights, we further developed a transferable parameterization strategy for residue-level models. The resulting desolvation-aware CG model improves the accuracy of IDP conformational ensembles and LLPS thermodynamics while preserving the computational efficiency inherent to CG methodologies. By retaining key water-mediated features while preserving the computational efficiency of implicit-solvent representations, this framework provides a mechanistic means to decouple overall phase-separation propensity from condensed-phase packing. As illustrated in , incorporating desolvation into CALVADOS2 reduces ρdense\begin{document}$\rho_{\mathrm{dense}}$\end{document} while preserving an overall phase boundary comparable to that of the original model. This decoupling effectively mitigates the overcompaction of the dense phase commonly observed in implicit solvent simulations with CG models, thereby making the simulations qualitatively more consistent with the physical nature of biological condensates. Despite these advances, several limitations remain in the present implementation. First, the current model introduces desolvation terms on top of residue-level CG frameworks with simplified interaction parameterizations. In particular, HPS-type models use a 20-parameter hydropathy-scale representation, which is useful for capturing generic IDP phase behavior but is not flexible enough to resolve residue-pair-specific chemical effects, such as the distinct interaction patterns of arginine and lysine residues (). More general 210-parameter pairwise interaction schemes, such as KH-type and related residue-pair-specific models, provide greater flexibility for encoding amino-acid-pair preferences and capturing sequence-specific interaction heterogeneity (; ; ). Second, the present desolvation model employs a single set of averaged parameters (αb,αss\begin{document}$\alpha_{\mathrm{b}},\alpha_{\mathrm{ss}}$\end{document}) for all residue pairs. While this simplification is effective for isolating the generic physical consequences of desolvation, it has limitations in describing the pair-specific variations in the desolvation barrier and the solvent-separated minimum. Explicit-solvent PMF analyses have shown that desolvation barrier heights and solvent-separated minima can differ substantially among residue pairs and may also exhibit temperature dependence (; ; ). In addition, effective interactions themselves can be temperature dependent, as demonstrated in analytical theories and coarse-grained simulations of biomolecular condensates (; ; ). Future extensions of the model could therefore incorporate residue-specific and temperature-dependent desolvation parameters derived from bottom-up parameterization or expanded experimental datasets, thereby enhancing predictive accuracy for sequence-dependent LLPS. An additional direction would be to combine these potentials with simulation-based rheological analyses to quantify how desolvation reshapes condensate viscoelasticity, aging-like maturation, and long-time material relaxation (; ). In summary, this work provides a physically interpretable route for incorporating solvent-mediated interaction features into implicit-solvent CG models, providing a basis for future residue-pair-specific and condition-dependent models of biomolecular condensates. By enabling an efficient representation of key desolvation-related features, the model provides a robust framework for simulating biomolecular condensates with improved thermodynamic, structural, and kinetic fidelity. Future developments, such as residue-pair-specific desolvation energetics and extensions to multicomponent systems involving nucleic acids, hold strong potential for deepening our mechanistic understanding of the complex phase behaviors that organize cellular biochemistry. Methods All-atom MD simulations The all-atom MD simulations were performed using GROMACS. Ten amino acid analog molecules were solvated in a cubic water box under periodic boundary conditions. The OPLS-AA force field was used to model the solute molecules, and the TIP4P water model was employed to represent solvent. Na+ and Cl- ions were added to neutralize the system, yielding a total system size of 1,010 atoms. After an initial energy minimization of 5×104\begin{document}$5\times 10^{4}$\end{document} steps, the system was equilibrated in the NVT ensemble at 298K\begin{document}$298\,\mathrm{K}$\end{document} for 100ps\begin{document}$100\,\mathrm{p}\mathrm{s}$\end{document}, followed by an NPT equilibration at 298K\begin{document}$298\,\mathrm{K}$\end{document} and 1atm\begin{document}$1\,\mathrm{a}\mathrm{t}\mathrm{m}$\end{document} for 1ns\begin{document}$1\,\mathrm{n}\mathrm{s}$\end{document}. Temperature and pressure were controlled using the Nosé–Hoover thermostat and Parrinello–Rahman barostat (), respectively. A time step of 2fs\begin{document}$2\,\mathrm{f}\mathrm{s}$\end{document} was used for all MD steps. The simulation details are provided in . Subsequently, a production simulation of 100ns\begin{document}$100\,\mathrm{n}\mathrm{s}$\end{document} was performed under the same NPT conditions for structural and thermodynamic analysis. The intermolecular separation rr between two amino acid analogues was defined as the distance between their C atoms. The PMF was computed as PMF(r)=−kBTln⁡P(r)\begin{document}$\mathrm{PMF}(r)=-k_{\mathrm{B}}T\ln P(r)$\end{document}, where P(r)P(r) is the radial probability density obtained from the production trajectory. Coarse-grained simulations We utilized the slab method () to extract the densities of the coexisting dilute and dense phases. In these simulations, 125 protein chains were placed in a rectangular box of dimensions 15 nm x 15 nm x 280 nm. The density profile along the zz-axis was computed to identify the condensed and dilute regions, from which thermodynamic quantities such as the critical temperature were subsequently determined. Simulations were performed over a range of different temperatures. To incorporate the desolvation terms into the pairwise interaction potentials of the HPS model, we used the azplugins package (https://github.com/mphowardlab/azplugins; ) together with HOOMD-blue v2.9.7 (). We also implemented the desolvation terms for both the HPS and CALVADOS models based on the OpenMM open-source toolkit (). GPU-accelerated simulations were performed on NVIDIA GeForce RTX 4090D and Tesla V100 GPUs, and CPU-based simulations were carried out on Kunpeng 920B. All simulations used a time step of 0.01 ps. Each trajectory was first equilibrated for 3 x 107 steps, followed by a production run of 1 x 108 steps. Additional details of the simulation framework are provided in . Hydrophobicity scale model Our desolvation model was developed based on the HPS coarse-grained model by , in which each residue is represented by a single bead. The potential energy comprised a bonded term, an electrostatic term, and a short-range pairwise term. The bonded interaction term is given by a harmonic potential Vb=kb(ri,i+1−r0)2\begin{document}$V^{b}=k_{b}(r_{i,i+1}-r^{0})^{2}$\end{document} with the spring constant kb=10kcal/mol/A ̊2\begin{document}$k_{b}=10\, \mathrm{kcal}/\mathrm{mol}/\mathring{A}^{2}$\end{document} and the bonded length r0=3.8A ̊\begin{document}$r^{0}=3.8 \, \mathring{A}$\end{document}. ri,i+1\begin{document}$r_{i,i+1}$\end{document} represents the distance between the consecutive residues ii and i+1i+1 along the protein chain. The electrostatic interaction term is modeled by a Coulombic potential with electrostatic screening caused by salt ion concentration. The function is given by(4)Vijele(r)=qiqj4πDrexp⁡(−r/κ),\begin{document}$$\displaystyle V^{ele}_{ij}(r) = \frac{q_{i}q_{j}}{4\pi Dr}\exp(-r/\kappa),$$\end{document} where κκ is the screening length and D=80D=80, which is the relative dielectric constant of water. For the short-range pairwise interaction, the HPS model uses a 6–12 Lennard–Jones potential with the interaction strength given by HPS (), which can describe the effective interaction between residues. The hydrophobicity value λλ, ranging from 0 to 1, varies for different residues. The arithmetic average is used for the interaction parameter between two residues, i.e. hydrophobicity value λi,j=(λi+λj)/2\begin{document}$\lambda_{i,j}=(\lambda_{i}+\lambda_{j})/2$\end{document} and amino acid size σi,j=(σi+σj)/2\begin{document}$\sigma_{i,j}=(\sigma_{i}+\sigma_{j})/2$\end{document}. The energy function of the short-range pairwise interaction is given by(5)Φ(r)={ΦLJ+(1−λ)ε,ifr⩽21/6σλΦLJ,otherwise\begin{document}$$\displaystyle \Phi(r) = \left\{ &\Phi_{LJ}+ (1 - \lambda)\epsilon,&&\mathrm{if }r \leqslant 2^{1/6}\sigma \\&\lambda\Phi_{LJ},&&\mathrm{otherwise} \right.$$\end{document} where ΦLJ\begin{document}$\Phi_{LJ}$\end{document} is standard 6–12 Lennard–Jones potential(6)ΦLJ=4ε[(σr)12−(σr)6].\begin{document}$$\displaystyle \Phi_{LJ}= 4\epsilon\left[\left(\frac{\sigma}{r}\right)^{12}- \left(\frac{\sigma}{r}\right)^{6}\right].$$\end{document} The free parameter εε is set as 0.2kcal/mol\begin{document}$0.2\,\mathrm{k}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{/}\mathrm{m} \mathrm{o}\mathrm{l}$\end{document}, with which the molecular simulations can best reproduce the experimental results of Rg\begin{document}$R_{g}$\end{document} values for a set of IDPs. Simulation framework We used a slab simulation scheme in the coexistence simulations. Initially, 125 chains were randomly distributed in a cubic box with periodic boundary conditions at 200 K. The box length is first scaled to 15 nm, which can effectively prevent protein chains from interacting with their periodic images generated by periodic boundary conditions. Then, the xx- and yy- dimensions were held fixed and the zz-dimension is extended to 280 nm. The temperature is then gradually increased to the target temperature, and the simulations are conducted at different temperatures for 1μs\begin{document}$1\,\mu\mathrm{s}$\end{document} in the NVT ensemble maintained by a Langevin thermostat, with a mass-normalized damping rate of 0.01ps−1\begin{document}$0.01\,\mathrm{p}\mathrm{s}^{-1}$\end{document}. A time step of 0.01 ps was used for all the simulations. We used HOOMD-Blue v2.9.7 packages () and OpenMM v8.2.0 () to conduct the simulations. The slab density profile along the zz-axis is then determined after the system had equilibrated. In order to avoid the influence of periodic boundary conditions, we first move the center of mass of the system along the zz-axis direction to the position of z=0z=0. We then determined the distribution of dense phase and dilute phase of the phase-separated system using the defined thresholds. The area with density higher than 0.95 of the maximum density (ρmax\begin{document}$\rho_{\max}$\end{document}) is treated as dense phase, and their average density is regarded as the phase density of the dense phase (ρdense\begin{document}$\rho_{\mathrm{dense}}$\end{document}). For the dilute phase, the threshold is taken as ρmin+50mg/mL\begin{document}$\rho_{\min}+50\,\mathrm{m}\mathrm{g}\mathrm{/}\mathrm{m}\mathrm{L}$\end{document}, where ρmin\begin{document}$\rho_{\min}$\end{document} is the minimum density. The density of dilute phase (ρdilute\begin{document}$\rho_{\mathrm{dilute}}$\end{document}) is then determined by the same method. If ρmax−ρmin<50mg/mL\begin{document}$\rho_{\max}-\rho_{\min} < 50\,\mathrm{m}\mathrm{g}\mathrm{/}\mathrm{m} \mathrm{L}$\end{document}, the system is considered not to undergo phase separation. Critical temperature The critical temperature was determined based on the vapor–liquid interfacial properties of Lennard–Jones chains (). The critical temperature Tc\begin{document}$T_{c}$\end{document} was obtained from ρdense\begin{document}$\rho_{\mathrm{dense}}$\end{document} and ρdilute\begin{document}$\rho_{\mathrm{dilute}}$\end{document} at different temperatures by(7)ρdense−ρdilute=A(Tc−T)β,\begin{document}$$\displaystyle \rho_{\mathrm{dense}}- \rho_{\mathrm{dilute}}= A(T_{c} - T)^{\beta}, $$\end{document} where ββ is the critical exponent, which was set to 0.325 () and AA is the fitting parameter. Data with temperature below the critical temperature was used to fit this equation. The phase density at critical temperature was determined by the law of rectilinear diameters ()(8)ρdense+ρdilute2=B+C⋅T,\begin{document}$$\displaystyle \frac{\rho_{\mathrm{dense}}+ \rho_{\mathrm{dilute}}}{2}= B + C\cdot T,$$\end{document} with BB and CC being the free parameters determined from the simulation data. At the critical temperature, the distinction between dense phase and dilute phase disappears, which means ρdense=ρdilute=ρc\begin{document}$\rho_{\mathrm{dense}}=\rho_{\mathrm{dilute}}=\rho_{c}$\end{document}. The phase density was then calculated by ρc=B+C⋅Tc\begin{document}$\rho_{c}=B+C\cdot T_{c}$\end{document}. The phase diagram was then constructed from the dense- and dilute-phase densities as functions of temperature. Bead-level second virial coefficient The bead-level second virial coefficient was calculated from the effective pair potential as(9)B2=2π∫0rc[1−exp⁡(−U(r)kBT)]r2dr,\begin{document}$$\displaystyle B_{2} = 2\pi \int_{0}^{r_c}\left[1 - \exp\left(-\frac{U(r)}{k_{\mathrm{B}}T}\right)\right] r^{2} \mathrm{d}r,$$\end{document} where U(r)U(r) is the desolvation-aware effective pair potential defined in . The upper limit of integration rc\begin{document}$r_{c}$\end{document} is set as 3σ, which is sufficiently large to capture the full range of interactions while ensuring numerical convergence. The reduced value B2/σ3\begin{document}$B_{2}/\sigma^{3}$\end{document} was used to compare the integrated effective attraction under different desolvation parameters. All B2\begin{document}$B_{2}$\end{document} values reported in were evaluated at the same reduced temperature. This bead-level quantity provides a pair-potential-level measure of effective attraction. Temperature normalization of the generalized diffusion coefficient To investigate the intrinsic influence of the interaction potential on chain dynamics, we must isolate the structural contributions from the trivial thermal acceleration. While keeping the quench depth (T/Tc\begin{document}$T/T_{\mathrm{c}}$\end{document}) constant provides a comparison at matched quench depth, it necessitates varying the absolute temperature TT, which inherently alters the base diffusion rate via thermal fluctuations. To resolve this, we adopt the constant quench depth approach to ensure thermodynamic consistency and introduce a normalized metric to factor out the leading linear thermal contribution. Motivated by the Einstein relation for Langevin diffusion, we describe the leading temperature dependence of the generalized diffusion coefficient KK in terms of the thermal energy kBT\begin{document}$k_{\mathrm{B}}T$\end{document} and the effective friction coefficient ζζ:(10)K(T)∝kBTζ(T,φ,ε).\begin{document}$$\displaystyle K(T)\propto\frac{k_{\mathrm{B}}T}{\zeta(T,\phi,\epsilon)}. $$\end{document} This approximation is motivated by the weakly subdiffusive dynamics observed here, for which α≈0.91\begin{document}$\alpha\approx 0.91$\end{document} and remains nearly constant across the simulated conditions. Here, the effective friction coefficient ζζ encapsulates the resistance to chain motion. Following the hybrid theoretical framework proposed by , which combines free-volume theory with activated rate theory, we phenomenologically factorize the effective friction into packing-dependent and energy-barrier-dependent contributions:(11)ζ(T,φ,ε)∝exp⁡(γv∗vf(φ))×exp⁡(Eeff(ε)kBT)\begin{document}$$\displaystyle \zeta(T, \phi, \epsilon) \propto \exp\left(\frac{\gamma v^{*}}{v_{f}(\phi)}\right) \times \exp\left(\frac{E_{\mathrm{eff}}(\epsilon)}{k_{\mathrm{B}}T}\right)$$\end{document} where the first term represents the hindrance due to limited free volume (vf\begin{document}$v_{f}$\end{document}), governed by the packing density φφ, and the second term is the Arrhenius factor describing the activated hopping over an effective energy barrier Eeff\begin{document}$E_{\mathrm{eff}}$\end{document}, which corresponds to the roughness of the energy landscape induced by the desolvation potential. In our analysis, by fixing the quench depth (T/Tc\begin{document}$T/T_{\mathrm{c}}$\end{document}), we expect the macroscopic packing density φφ to remain approximately comparable across different systems (as verified by the phase diagram in the main text). Consequently, the free-volume contribution to friction is effectively invariant. Substituting this back into , the diffusion coefficient scales as(12)K(T)∼T⋅exp⁡(−Eeff(ε)kBT)\begin{document}$$\displaystyle K(T) \sim T \cdot \exp\left(-\frac{E_{\mathrm{eff}}(\epsilon)}{k_{\mathrm{B}}T}\right)$$\end{document} The prefactor TT represents the leading linear thermal acceleration inherent to Langevin dynamics. To better isolate the intrinsic energy-landscape signature, we define the reduced diffusion coefficient K~\begin{document}$\widetilde{K}$\end{document} by normalizing out this explicit linear thermal contribution:(13)K~≡KT∝exp⁡(−Eeff(ε)kBT)\begin{document}$$\displaystyle \widetilde{K}\equiv \frac{K}{T}\propto \exp\left(-\frac{E_{\mathrm{eff}}(\epsilon)}{k_{\mathrm{B}}T}\right)$$\end{document} Thus, K~\begin{document}$\widetilde{K}$\end{document} serves as a physically motivated metric to better isolate the intrinsic influence of the energy function on chain dynamics. By normalizing out the explicit linear thermal contribution inherent to Langevin dynamics, this definition helps decouple the specific kinetic modulation imposed by the desolvation potential from the background thermal acceleration and thermodynamic driving forces. Consequently, K~\begin{document}$\widetilde{K}$\end{document} allows for a more direct assessment of how the interaction potential alters transport properties after removing the leading linear temperature effect required to maintain thermodynamic consistency. Calculation of normalized density variance To quantify the rate of density fluctuation growth during the early stage of phase separation, we first computed the instantaneous spatial variance of the density profile along the zz-axis, σρ2(t)\begin{document}$\sigma^{2}_{\rho}(t)$\end{document}, defined as(14)σρ2(t)=1Lz∫0Lz[ρ(z,t)−ρ ̄]2dz,\begin{document}$$\displaystyle \sigma^{2}_{\rho}(t) = \frac{1}{L_{z}}\int_{0}^{L_z}[\rho(z, t) - \bar{\rho}]^{2} \mathrm{d}z,$$\end{document} where ρ(z,t)ρ(z,t) is the local density at position zz and time tt, and ρ ̄\begin{document}$\bar{\rho}$\end{document} is the mean density of the system. To enable a fair comparison of kinetics across systems with distinct interaction parameters (εss\begin{document}$\epsilon_{\mathrm{ss}}$\end{document} and εb\begin{document}$\epsilon_{\mathrm{b}}$\end{document}), it is crucial to decouple the rate of phase separation from the difference in equilibrium dense-phase densities. Therefore, we normalized the instantaneous variance by its equilibrium baseline:(15)σ^ρ2(t)=σρ2(t)⟨σρ2⟩eq.\begin{document}$$\displaystyle \hat{\sigma}^{2}_{\rho}(t) = \frac{\sigma^{2}_{\rho}(t)}{\langle \sigma^{2}_{\rho}\rangle_{\mathrm{eq}}}.$$\end{document} Here, the normalization factor ⟨σρ2⟩eq\begin{document}$\langle\sigma^{2}_{\rho}\rangle_{\mathrm{eq}}$\end{document} is the time-averaged density variance calculated from the final 30% of the simulation process. This period corresponds to the steady state characterized by the formation of a well-defined slab with stabilized phase densities. By using this normalized metric σ^ρ2(t)\begin{document}$\hat{\sigma}^{2}_{\rho}(t)$\end{document}, we effectively factor out the variations in final density contrast, allowing the observed differences in the time-evolution curves to more directly reflect the acceleration or deceleration of the kinetic process itself. Desolvation parameters Because residue-pair-specific experimental desolvation data were unavailable, the empirical parameters αb\begin{document}$\alpha_{\mathrm{b}}$\end{document} and αss\begin{document}$\alpha_{\mathrm{ss}}$\end{document} were estimated using all-atom molecular dynamics simulations of amino acid analogues with explicit solvent. The PMF obtained from the simulation results is shown in . Methane (CH4), methanol (CH3OH), acetate ion (C2H3O2 ̄\begin{document}${\rm C_2H_3O_{\overline{2}}}$\end{document}), ammonium ion (NH4+\begin{document}$\rm NH_{4}^{+}$\end{document}), and methanamide (CH3NO) were chosen as representative amino acid analogues for nonpolar, polar, negatively charged, positively charged, and backbone-like residues, respectively. For each analogue pair, αb\begin{document}$\alpha_{\mathrm{b}}$\end{document} and αss\begin{document}$\alpha_{\mathrm{ss}}$\end{document} were obtained by fitting the nonbonded energy function with desolvation effects () to the simulation-derived PMF. The λλ values assigned to the analogues were taken from the corresponding amino acids. Specifically, we used λGLY=0.649\begin{document}$\lambda_{\mathrm{GLY}}=0.649$\end{document} for methane, λSER=0.595\begin{document}$\lambda_{\mathrm{SER}}=0.595$\end{document} for methanol, and λCYS=0.595\begin{document}$\lambda_{\mathrm{CYS}}=0.595$\end{document} for methanamide. The results are shown in .