Prediction of small-molecule partitioning into biomolecular condensates from simulation

Abstract

Predicting small-molecule partitioning into biomolecular condensates is key to developing drugs that selectively target aberrant condensates. However, the molecular mechanisms underlying small-molecule partitioning remain largely unknown. Here, we first exploit atomistic molecular dynamics simulations of model condensates to elucidate physicochemical rules governing small-molecule partitioning. We find that while hydrophobicity is a major determinant, solubility becomes a stronger regulator of partitioning in more polar condensates. Additionally, more polar condensates exhibit selectivity toward certain compounds, suggesting that condensate-specific therapeutics can be engineered. Building on these insights, we develop minimal models (MAPPS) for efficient prediction of small-molecule partitioning into biologically relevant condensates. We demonstrate that this approach reproduces atomistic partition coefficients in both model systems and condensates composed of the low complexity domain (LCD) of FUS. Applying MAPPS to various LCD-based condensates shows that protein sequence can exert a selective pressure, thereby influencing small-molecule partitioning. Collectively, our findings reveal that partitioning is driven by both small-molecule-protein affinity and the complex interplay between the compounds and the condensate chemical environment.

FIG. 1:. Atomistic molecular dynamics simulations were used to probe the partitioning of organic molecules into model condensates.

(a) Chemical structures of the simulated compounds. Colored circles in the titles correspond to the markers’ colors for the data points reported throughout the work. The molecules are ordered alphabetically. (b) Snapshot from the all-atom molecular dynamics simulations (part of the simulation box is shown), illustrating the partitioning of styrene into an FYAFYF model condensate (protein is colored in pink colors, small molecules are in yellow, water is in blue and salt ions are in green and grey).

FIG. 2:. Physicochemical properties of small molecules, as well as chemical environment of condensates, dictate partitioning.

Partitioning coefficient of a set of small molecules into model condensates composed of: NYANYN (first row), YQHQHY (second row), FYAFYF (third row), and FWAFWF (fourth row) peptides, based on all-atom simulations at 300 K, plotted against the compounds (a) $\mathrm{l}\mathrm{o}\mathrm{g}P$ values (partition coefficient between octanol and water [37]), (b) solubility (log₁₀) at 25°C [38], and (c) molecular diameter. The solid black line shows a linear fit with a corresponding Pearson correlation coefficient, and the red solid line shows a linear fit in the region of $0<\mathrm{l}\mathrm{o}\mathrm{g}P<2$. Error bars represent standard errors. The concentration of the small molecules in each system was 100 mM, the protein concentration was 300 mM for FYAFYF, NYANYN, and YQHQHY peptides and 240 mM for FWAFWF peptides.

FIG. 3:. Illustration of the parametrization algorithm used to develop MAPPS.

Results for 2-methylpropane are shown as an example. (a) The potential of mean force for a molecule self-interaction was used to determine molecule self-interaction parameters of the Wang-Frenkel potential. (b) A series of slab simulations with Y₆ peptides at 300 K and salt concentration of 150 mM in coarse-grained resolution was used to optimize the molecule-tyrosine interaction parameter, ${\epsilon }_{\mathrm{S}\mathrm{M}-\mathrm{Y}}^{\text{fitted}}$, to reproduce partitioning behavior observed in the all-atom simulations of the corresponding system. (c) The distribution of the normalized contacts frequencies between the molecule and a single protein chain, obtained from all-atom simulations at 300 K and salt concentration of 150 mM, was used to derive interaction parameters for the molecule with other residues $\left({\epsilon }_{\mathrm{S}\mathrm{M}-\mathrm{X}}\right)$. (d) Validation of the optimization based on the contacts distribution: ${\epsilon }_{\mathrm{S}\mathrm{M}-\mathrm{X}}^{\text{fitted}}$ values corresponding to partitioning in all-atom simulations obtained by extending the optimization procedure through all-atom slab simulation in (YXY)₂ peptides for each residue ($\mathrm{X}$) vs. ${\epsilon }_{\mathrm{S}\mathrm{M}-\mathrm{X}}$ values obtained through Eq. 1. (e) Snapshot from a coarse-grained molecular dynamics simulation illustrating partitioning of the molecules into an LCD condensate, simulated with parameters obtained via steps 1–4.

FIG. 4:. Validation of transferability of MAPPS on systems of short peptides.

Coarse-grained (CG) and all-atom (AA) simulations of a condensate with small molecules were conducted at 300 K and 150 mM salt concentration, with the protein concentration of 300 mM for FYAFYF, NYANYN, and YQHQHY peptides and of 240 mM for FWAFWF peptides, and the small molecules concentration of 100 mM. (a) Partitioning coefficients of a subset of small molecules (cisplatin, phenol, 2-methylpropane, styrene, and nitrobenzene) in the CG simulation versus AA simulations, for a set of peptides of various degrees of polarity FWAFWF, NFWAFS, YQHQHY, FYAFYF, NYANYN; (b) partitioning of a larger set of small molecules in the CG simulation versus AA simulation for FYAFYF and FWAFWF peptides and (c) YQHQHY and NYANYN peptides. The Pearson correlation coefficient $r$ and the root-mean-square deviation (RMSD) values are displayed. The dashed line represents the equity line.

FIG. 5:. Simulations of FUS LCD condensate with small molecules in all-atom (AA) and coarse-grained (CG) resolutions.

Simulations were conducted at 300 K and 150 mM salt concentration. 64 chains of FUS LCD were simulated, at a protein concentration of 12.5 mM and small molecules concentration of 30 mM. (a) Backmapping procedure to obtain the AA configuration of the condensed FUS LCD chains. (b) Partitioning coefficients obtained in CG simulation versus AA simulations. The error bars for the values from the CG simulations are within the markers. (c) Normalized contact frequency between a molecule and a single residue of FUS-LCD in the CG simulation versus AA simulation (each marker corresponds to a single residue). The errors are computed as standard errors for the frequency of contacts across the analyzed time frames.

FIG. 6:. Coarse-grained simulations of small molecules partitioning into condensates of low-complexity domains of proteins.

The following low-complexity domains were simulated: FUS, EWSR1, TIA1, and A1 of 12 mM, 7 mM, 20 mM, 16 mM concentrations, respectively. The concentration of the small molecules for each system was 30 mM. Simulations were conducted at 300 K and 150 mM salt concentration. (a) Compositions of the residues for the simulated LCDs. The following residues classification is used: “hydrophobic” – A, I, L, M, V; “aromatic” – F, W, Y; “polar” – N, Q, S, T; “charged (+)” – R, K, H; “charged (−)” – D, E; “other” – C, G, P. (b) Partitioning coefficients $K$ for FUS, EWSR1, TIA1, and A1 LCDs per compound. Error bars represent standard errors. The inset plot is the box-and-whisker plot of partitioning coefficients $K$ for each protein across all tested molecules. The box represents the interquartile range (IQR), with the lower and upper edges corresponding to the first (Q1) and third (Q3) quartiles, respectively. The horizontal line inside the box indicates the median $K$ value. (c-f) Partitioning coefficients plotted against the compounds $\mathrm{l}\mathrm{o}\mathrm{g}P$ values. The yellow arrow represents the decreasing percentage of the polar residues among the LCDs. Displayed values $r$ is the Pearson correlation coefficient.