Toward Predictive Coarse-Grained Simulations of Biomolecular Condensates

Abstract

Phase separation is a fundamental process that enables cellular organization by forming biomolecular condensates. These assemblies regulate diverse functions by creating distinct environments, influencing reaction kinetics, and facilitating processes such as genome organization, signal transduction, and RNA metabolism. Recent studies highlight the complexity of condensate properties, shaped by intrinsic molecular features and external factors such as temperature and pH. Molecular simulations serve as an effective approach to establishing a comprehensive framework for analyzing these influences, offering high-resolution insights into condensate stability, dynamics, and material properties. This review evaluates recent advancements in biomolecular condensate simulations, with a particular focus on coarse-grained 1-bead-per-amino-acid (1BPA) protein models, and emphasizes OpenABC, a tool designed to simplify and streamline condensate simulations. OpenABC supports the implementation of various coarse-grained force fields, enabling their performance evaluation. Our benchmarking identifies inconsistencies in phase behavior predictions across force fields, even though these models accurately capture single-chain statistics. This finding underscores the need for enhanced force field accuracy, achievable through enriched training data sets, many-body potentials, and advanced optimization techniques. Such refinements could significantly improve the predictive capacity of coarse-grained models, bridging molecular details with emergent condensate behaviors.

Keywords: OpenABC; biomolecular condensates; coarse-grained simulations; force field accuracy; phase separation.

Figure 1:

Atomistic simulations allow detailed characterization of the microenvironments inside condensates. (A) Comparison of protein conformation and solvation of a single peptide chain (red) from atomistic simulations of ELP condensate (top) and monomer (bottom). Atoms within 1 nm of the peptide are shown, including water (cyan), chlorine (green), sodium (orange), and protein atoms from other chains (gray). Figure adapted from Ref. . (B) Snapshot from atomistic simulations reveals the diverse array of contacts that contribute to the phase separation of (GRGDSPYS)₂₅. Figure adapted from Ref. . (C) The condensate formed by HP1α proteins presents a diverse array of binding environments capable of accommodating the small molecule Mitoxantrone. These environments are represented by projecting the simulated binding poses onto two variables derived through the UMAP dimensionality reduction technique. The binding poses organize into distinct clusters, each characterized by varying levels of hydrophobic interactions with the small molecules. Figure adapted from Ref. .

Figure 2:

Comparison of nonbonded contact interaction patterns, excluding electrostatic interactions, among amino acids across various 1BPA protein models. Interaction strength parameters—$\lambda$ for HPS-Urry and CALVADOS2,^– and $ϵ$ for Mpipi and MOFF^,—are unified under the notation $ϵ$ for consistency. For each model, $ϵ$ values were normalized to the range [0, 1] to emphasize relative variations within each model. Normalization was performed using the formula $\overline{ϵ}=\frac{ϵ-{ϵ}_{\text{min}}}{{ϵ}_{\mathrm{m}\mathrm{a}\mathrm{x}}-{ϵ}_{\mathrm{m}\mathrm{i}\mathrm{n}}}$, where ${ϵ}_{\mathrm{m}\mathrm{i}\mathrm{n}}$ and ${ϵ}_{\mathrm{m}\mathrm{a}\mathrm{x}}$ denote the minimum and maximum interaction parameter values within each model. A larger $\overline{ϵ}$ corresponds to stronger attraction. This normalization scheme highlights the intrinsic relative differences in contact strengths within each model.

Figure 3:

OpenABC supports CG and multiscale simulations of biomolecular condensates using various force fields. (A) Illustration of the OpenABC workflow, featuring Python scripting for both simulation setup and analysis. (B) OpenABC offers tools to convert Cα protein condensates into all-atom models, enabling simulations in explicit solvent environments. Figures adapted from Ref. .

Figure 4:

Benchmarking the performance of 1BPA models on single proteins, with a complete list of protein names provided in Table S1. (A) Comparison between simulated and experimental average $R_g$ values for a selection of IDPs, computed using different models. Relative root mean square error (rRMSE, equation S1) values for each model are indicated in the legend. (B) Comparison between simulated and all-atom average $R_g$ values for a variety of folded proteins computed with different models and structure-based terms to stabilize ordered domains. rRMSE values for each model are also indicated in the legend. Reference values were obtained from long-time atomistic simulations ^,, (Table S2). (C) Comparison between simulated and experimental average $R_g$ values for multi-domain proteins with CG beads placed at Cα positions. (D) Same comparison as (C), but with CG beads within ordered domains shifted to amino acid’s center of mass.

Figure 5:

(A)-(C) Benchmarking the performance of 1BPA models on condensates formed by (GRGDSPYS)₂₅ and its variants. The three panels present phase diagrams generated using HPS-Urry, CALVADOS2, and MOFF, respectively. Legends denote the three condensates, each with its corresponding repeat sequence. (D) Experimental measurements of saturation concentrations from Rekhi et al.. Figure adapted from Ref. .

Figure 6:

Illustration of the strategy for developing next-generation CG force fields. (A) This approach utilizes conformational ensembles of both folded and disordered proteins derived from atomistic simulations. (B) Multi-body potentials, formulated analytically based on local density or employing neural networks, are crucial for enhancing force field accuracy. (C) Parameter optimization in these force fields can benefit from newly developed algorithms, such as potential contrasting, which efficiently exploit extensive training datasets. Figure adapted from Ref. .