Small-molecule properties define partitioning into biomolecular condensates

1Biomolecular condensates are cellular compartments that concentrate proteins, nucleic acids and likely small molecules in the absence of an encapsulating membrane,. Many condensates appear to form through liquid–liquid phase separation of multivalent macromolecules–. Condensates are involved in numerous biological processes, including signal transduction, intermediary metabolism, RNA metabolism and gene regulation. Aberrant condensates are associated with human diseases including neurodegeneration, cancer and viral infections–, and bioinformatic analyses have recently implicated condensate dysregulation in over 1,000 genetic disorders. Condensates have thus emerged as therapeutic targets,,. Drug development efforts seek to identify small molecules that disrupt condensates or modulate their material properties, or that alter the activity of specific macromolecular components.

2Condensates are believed to segregate, concentrate and modulate biochemical processes in vivo, affording specificity and efficiency to individual reactions and pathways,,. Understanding the mechanisms that determine condensate composition is thus important for defining condensate function. In models of macromolecular composition, many proteins and RNAs are recruited through specific binding to other condensate components–. In addition, proteins can be enriched in or excluded from condensates without apparent site-specific binding, in a manner that depends on a protein’s overall charge–. Less is understood about the factors that govern the composition of small organic molecules in condensates, although recent efforts have begun to characterize such factors–. Multiple mechanisms are likely to be important and can be conceptualized on a spectrum of binding affinities and specificities. At one end, high-affinity, stereospecifically defined interactions can occur between a compound and a complementary element in a macromolecular condensate component. These will drive recruitment of the compound into the condensate, as has been observed for many macromolecules. By contrast, low-affinity, low-specificity interactions with solvent and/or other molecules can govern the distribution of a compound between the distinct biochemical environments inside and outside a condensate. Such factors can lead to the enrichment or exclusion of a compound from the condensate phase. It remains unknown which types of interaction play dominant roles in dictating the recruitment of small molecules into biomolecular condensates.

3Here we examined the partitioning of a large collection of small-molecule metabolites and pharmaceuticals into four different condensates composed of unrelated proteins and protein–DNA mixtures, both purified and in cell extracts. The degree of partitioning spans nearly six orders of magnitude, from ~100-fold exclusion to ~10,000-fold enrichment. Surprisingly, partitioning is correlated among the different condensates, indicating a similarity in their underlying physical properties. Several highly partitioning compounds do not show measurable interactions with one model protein system under non-phase-separating conditions, suggesting that enrichment is not due to specific binding to condensate components. We developed a machine learning model, whose only inputs are physicochemical properties of the compounds (and not the molecular structures), that predicts values of the partition coefficient (PC) with quantitative accuracy in both the pure protein and extract systems, providing potential routes to drugs that target condensates. Furthermore, the model indicates that the strongest predictors of partitioning behaviour are aqueous solubility and other measures related to hydrophobicity. Whereas high-affinity binding to condensate macromolecules will also contribute to enrichment, and may account for some inaccuracies of the model, such effects are not dominant for the systems under study. The experimental data and modelling suggest that the chemical properties of small molecules largely define their partitioning behaviours, such that compounds with enhanced hydrophobic character favour concentration into the condensates. This enrichment of organic compounds in the condensate is an emergent biochemical property of the phase-separated compartment, which is not found in the individual macromolecules.

4We examined the partitioning of small molecules into four disparate condensates composed of proteins and nucleic acids. We characterized two synthetic condensates, one composed of polySH3 and polyproline-rich motif (PRM) proteins (that is, SH3PRM) and the other composed of polySUMO and polySUMO-interaction motif (SIM) proteins (that is, SUMOSIM). In addition, we examined two condensates formed by naturally occurring macromolecules, one formed by the yeast P-body protein Dhh1 (ref. ) (that is, Dhh1) and the other formed by the innate immune signalling protein cyclic GMP-AMP synthase (cGAS) and double-stranded DNA (that is, cGASDNA).

5For each condensate, we assessed the partitioning of two libraries of molecules, one of ~200 metabolite compounds and one with ~1,500 drug compounds approved by the US Food and Drug Administration (FDA). To assess the diversity of the compounds, we combined them with 200,000 randomly chosen biologically active small molecules from the ChEMBL database, and calculated ~50 ADME (absorption, distribution, metabolic and excretion) and chemical properties for each using the QikProp software package. The multidimensional descriptor set was embedded into a two-dimensional chemical space using the Uniform Manifold Approximation and Projection (UMAP) algorithm. As illustrated in Fig. 1a, the compounds used in our study distribute relatively evenly across the chemical space representation mapped by the ChEMBL compounds, illustrating the diversity of our library. When UMAP was applied solely to our library, clustering analysis using an unsupervised learning algorithm (HDBSCAN) identified 11 groups of small molecules with similar chemical attributes, suggesting that QikProp descriptors convey enough chemical information to discriminate between classes of molecules (Extended Data Fig. 1—an interactive map can be accessed in Supplementary Data 1). There are well-defined isolated clusters for steroidal compounds, guanidine-containing compounds and β-lactam antibiotics. The other clusters share boundaries but have distinct characteristics. The largest of these clusters contains small biomolecules, multi-chlorinated aryl compounds and common non-steroidal anti-inflammatory drugs with a variety of structures. Overall, the UMAP and clustering indicate a diverse but relatively sparsely populated set of molecules, with a few classes (for example, steroids and β-lactams) that are more densely sampled.

6We developed two assays to measure the partitioning of the compound libraries (Fig. 1b). In the first, we incubated sublibraries of ~300 compounds, each at 1 µM concentration, with the macromolecules under phase-separating conditions (2–10 µM concentrations). As described in Extended Data Fig. 2, the compound libraries did not substantially alter the volume, PC or dynamics of the condensates (Supplementary Fig. 1). After equilibration (1–24 h), the condensates were collected via centrifugation followed by manual separation of the pellet and supernatant. Compounds were subsequently extracted from the two samples and quantified using mass spectrometry, yielding the PC as the ratio of concentrations in the pellet and supernatant. Analysis of replicate measurements revealed that the average percentage error in the measured PC values was less than 20% for all of the condensates (SUMOSIM = 19.5%, SH3PRM = 13.6%, Dhh1 = 10.8%, cGASDNA = 10.0%), and that the percentage error is consistent across the range of PC values.

7Our pooled-library approach relies on each compound behaving independently, with the compounds not substantially modulating the exclusion or enrichment of others. To examine this, we analysed the partitioning into SUMOSIM condensates of a sublibrary that contained 240 drug molecules and the same molecules divided into three subgroups of 80 compounds each. Apart from a few outliers (five molecules), the PCs measured in each case were similar (coefficient of determination (R2) = 0.86; Extended Data Fig. 3a). Thus, whereas we cannot entirely rule out cross-talk between certain molecules, the compounds generally behave independently from each other in this assay.

8To validate the mass spectrometry results, we used confocal fluorescence microscopy with the 30 fluorescent molecules in the drug library individually incubated with macromolecules under phase-separating conditions. PCs were determined from the ratio of fluorescence intensity values within the condensate and the surrounding solution. In each condensate, PCs measured using mass spectrometry and microscopy were reasonably correlated (R2 = 0.52–0.83, root mean squared deviation (RMSD) = 0.41–0.71; Extended Data Fig. 3b). The modest number of outliers was probably due to changes in the fluorescence intensity of certain compounds in the dense and dilute phases or to cross-talk between the pooled molecules in the mass spectrometry assay.

9Figure 1c plots the PC values of the metabolite and drug libraries in the SUMOSIM condensate measured via the mass spectrometry assay. The values span a nearly 1,000,000-fold range, from ~0.01 to 10,000, with the strongest partitioning corresponding to a free energy of ~5.6 kcal mol−1 for transfer of a compound into the condensate from the dilute phase. Of the 1,037 compounds that could be reliably measured (Supplementary Methods), 247 had PC values of <0.8, indicating that they are substantially excluded from the condensates; 224 compounds had PC values in the range of 0.8–1.5, indicating little preference for condensates versus surroundings; and 566 had PC values of 1.5–10,000, indicating modest to strong enrichment in the condensates. Given the droplet volume fraction of 1.2 ± 0.2% and the 1 µM total compound concentration, the highest enrichment corresponds to a concentration of ~90 µM in the condensate. The metabolites populate the lower end of the PC range, spanning 0.14–35, whereas the drugs sample the entire range of values.

10We initially anticipated that the individual condensate systems would show different distributions of partitioning as they are composed of unrelated molecules with different physical properties (Supplementary Tables 1 and 7). Surprisingly, however, the different condensates showed similar ranges of PC values for the small molecules (Fig. 1d and Extended Data Fig. 4), and the distributions are strongly correlated between the different condensates over the full range of PC values (Fig. 1e and Extended Data Fig. 4). Correlations are weaker when considering smaller ranges of PC values (less than ~1–2 log units; Extended Data Fig. 5), and may indicate that effects specific to each macromolecule govern minor differences in PC (see the Discussion). We also analysed outliers in these correlations to identify molecules that showed specificity for an individual condensate over the three others (Extended Data Figs. 5 and 6). Within each specificity group, members did not share common physical or structural features (Extended Data Fig. 6). Together, these results indicate that, despite the differences in macromolecular components, there is an underlying physical similarity between the condensates such that the same small molecules tend to be enriched or excluded by each condensate system.

11Several features of our PC dataset suggest that small-molecule partitioning is not governed by specific binding to the condensate macromolecular components. First, for enrichment due to stereospecific binding to a discrete site, PCcompound is bounded by 1 and PCscaffold, assuming that the macromolecular scaffolds have equal affinity for a compound in the dense and dilute phases. Yet here, each condensate has numerous compounds that are either substantially excluded (PC < 1) or have PC values that far exceed those of the macromolecular components (Fig. 1c and Extended Data Fig. 4a). Second, whereas PCcompound values can lie outside these limits if the macromolecular components have different affinities for compounds inside and outside the condensate, the similarity of distributions between the different condensates argues against this possibility (Fig. 1e and Extended Data Fig. 4b); it is unlikely that two independent scaffolds would both possess stereospecific binding sites for a compound that are differentially populated to similar degrees in the dense and dilute phases. Finally, the highly partitioning compounds for each condensate, as well as the specifically enriched compounds for each condensate, do not possess common structural features, as would be expected if they recognized a specific binding site in the scaffold (Extended Data Figs. 6 and 7). Thus, the data suggest that partitioning is not driven by stereospecific binding to condensate scaffolds but rather by the general physical properties of the compounds, which differentially favour the dense or dilute phases.

12To test this hypothesis, we used isothermal titration calorimetry (ITC) to examine the binding of 13 compounds to the polySUMO–polySIM complex at a concentration below the phase separation threshold (20 µM module concentration). These compounds had PC values ranging from 50 to 4,000 and were titrated up to 200 µM, a tenfold excess over the scaffolds. As illustrated in Fig. 2 and Supplementary Fig. 2, at both 25 and 35 °C, nine of the compounds showed essentially flat titration profiles, with an enthalpy of binding (ΔH) near zero. The absence of measurable heat at two different temperatures strongly suggests the absence of binding rather than a small heat of binding. Three of the compounds showed relatively flat profiles with non-zero enthalpies, suggesting a weak interaction that is far from saturation even at a tenfold excess of ligand. Only miconazole showed saturable binding, with a dissociation constant (KD) value in the low micromolar range. Thus, most of the compounds examined do not have a measurable affinity for the polySUMO–polySIM complex in the absence of phase separation. Even for miconazole, its PC value is well above that of the scaffolds (107 versus 60), suggesting that binding alone cannot account for its enrichment into the SUMOSIM condensate. These experiments further support the hypothesis that, for most compounds, partitioning is driven primarily by the physical properties of the compounds and condensates rather than stereospecific binding to scaffolds.

13We turned to statistical modelling to determine which chemical features of small molecules determine partitioning, with a related goal of exploiting the models to predict the partitioning behaviour of new compounds. We used the chemical space map and clustering to define balanced training and test sets for modelling and validation purposes. We withheld 20% of the molecules from each cluster for external validation and used an additional 20% as a test set to compare machine learning algorithms. The remaining 60% of the molecules were selected as the training set to develop models. For SUMOSIM condensates, confidence intervals were computed via tenfold re-randomization of the train–test split. Model exploration included the evaluation of different data structures and machine learning algorithms through comparison of the test set R2 and mean absolute error (MAE) in the predicted log PC of the resultant models (see Supplementary Methods for modelling details).

14Extreme gradient parallel tree boosting (XGBoost) models generally outperformed other algorithms, and QikProp descriptors outperformed structure-based molecular encodings that produce a larger feature set (Extended Data Fig. 1 and Supplementary Information). Combining QikProp with structure-based encodings did not improve the models, whereas adding descriptors of the condensate scaffolds (Supplementary Table 8) led to overfitting; these features were not used in the final models. The XGBoost model of SUMOSIM partitioning had a modest R2test value of 0.56 and a log PC MAEtest of 0.48 (Fig. 3a and Supplementary Table 9). Models for the other condensate systems performed similarly, with R2test in the range of 0.48–0.58 and log PC MAEtest in the range of 0.44–0.52 (Extended Data Fig. 8 and Supplementary Table 9).

15Given the strong correlation of PC values across the condensate systems, we reasoned that the performance of our models should not depend strongly on which condensate PC dataset is used for model training. To address this, we compared the individual models with one based on the combined data for all four condensates using a single condensate descriptor in addition to the QikProp features (Supplementary Table 9). As hypothesized, this combined model performed similarly to the models of each condensate, with R2test and MAEtest values of 0.56 and 0.50, respectively. We also constructed a model based on the average log PC for each molecule across all condensates, a dataset of comparable size to the individual condensate models (Supplementary Table 10, entry 6). The average log PC model had an overall R2test of 0.58 and a MAEtest of 0.44, a notable improvement compared with the individual condensate and combined models. Together, these findings suggest that condensate–small-molecule interactions are mostly conserved with respect to the condensate identity. This emergent quality implies that, among the condensates examined here, the condensate composition plays a secondary role in small-molecule partitioning.

16We validated the average log PC model by predicting the withheld data. The model successfully predicted the partitioning behaviour for the validation set within one-half of a log PC unit (R2validation = 0.44, MAEvalidation = 0.44; Fig. 3b), corresponding to a PC error of less than threefold. Although the validation predictions plateau near log PC = 0 (probably due to the limited experimental range of log PC for excluded molecules compared with enriched molecules), the MAE of the excluded validation points is only 0.33 log PC units. We recognize that the sparsity of data relative to the molecular diversity surveyed resulted in the observed undertraining, and we aim to supplement our database of small-molecule partitioning measurements to explore this limitation in future studies. Prediction errors did not substantially change as a function of which cluster a compound originated from (Extended Data Fig. 1), suggesting that the model effectively covers the diverse compound representation. Furthermore, only 12% of molecules had prediction errors greater than one log PC unit, showing that our model successfully differentiates highly enriched molecules from those that are weakly attracted to or excluded from the condensates. We anticipate that the ability to predict differences in partitioning across nearly six orders of magnitude will have substantial benefit for future applications.

17Z-test comparisons of the top features in the SUMOSIM model demonstrated a significant difference in the distributions of enriched versus excluded small molecules (Extended Data Fig. 8). Building on these differences, we used SHAP (SHapley Additive exPlanations) analysis to decompose each prediction into the contributions of each feature, to reveal the relative importance of the different features and their interactions in determining specific outcomes from the XGBoost model (Fig. 3c,d and Extended Data Fig. 9a,b). Both the two-sample Z-test and SHAP analyses revealed that, for all condensates, features describing the aqueous solubility (CIQPlogS and QPlogS) and oil/water partitioning (QPlogPo/w) were the most influential in predicting small-molecule partitioning, along with the capacity to engage promiscuous, hydrophobic binding sites in proteins (QPlogKhsa),. These features are all generally associated with hydrophobicity. Conservation of feature importance across models of the four individual condensates suggests that the guidelines of small-molecule partitioning are relatively consistent for all condensates (Extended Data Fig. 9). As described in Extended Data Fig. 9, we also generated an XGBoost model for the SUMOSIM condensate using only the three most important QikProp descriptors, and found that this model had slightly poorer statistics than the model trained on all QikProp features (R2test = 0.51 versus 0.56, MAEtest = 0.52 versus 0.48). None of the top features alone is strongly correlated with the PC, indicating the importance of the collection (Extended Data Fig. 10a). Thus, solubility and hydrophobicity effects dominate the predictive models of partitioning in a nonlinear fashion, while additional physicochemical descriptors can offer some fine-tuning of the predictions.

18In addition to the QikProp features, we explored the possibility that structural features, such as molecular fingerprints, may outperform physiochemical representations of small molecules. Therefore, we computed two types of molecular fingerprint– for the SUMOSIM dataset. As demonstrated in Supplementary Table 10 and discussed in the Supplementary Methods, QikProp features outperformed both robust (ECFP4 fingerprints) and compact (MACCS keys) representations of small molecules. Thus, physical properties, rather than specific chemical structures, are most predictive of partitioning behaviour (in the absence of stereospecific binding—see the Discussion).

19Together, these observations provide further support for the hypothesis that partitioning in this compound library is largely driven by the physical properties of the molecules, as opposed to stereospecific binding.

20Our results suggest that the chemical environment within a condensate is different from the surrounding solution, and this difference plays an important role in determining whether organic compounds are enriched or excluded. However, the vastly more complex environment of the cell relative to simple buffers may impact partitioning in ways that are not accounted for by our experiments thus far. To address this, we quantified partitioning of the drug library into SUMOSIM condensates under conditions that are designed to mimic the cellular environment, either in a relatively dilute U2OS cell lysate (~4 mg ml−1 total protein) or a highly concentrated Xenopus laevis oocyte extract (80–100 mg ml−1, comparable to the eukaryotic cytoplasm). As shown in Fig. 4a,b, for both systems, the partitioning behaviours remain modestly correlated to those observed for SUMOSIM condensates in simple buffer (with R2 values of 0.39 and 0.34, respectively). However, the total range of PC values decreases by ~100-fold, spanning ~0.1–1,000, with only a few values larger than 100. The reduced range is probably due in part to a roughly tenfold decrease in PC values for the polySUMO and polySIM proteins themselves (60 in buffer, 8 in U2OS lysate, 5 in Xenopus extract). In addition, the much greater complexity of the lysate/extract solution probably dampens the differences between dense and dilute phases. On average, the log PC values decrease around twofold between the buffer and either lysate, as indicated by the slopes of the lines correlating the distributions and the relatively large RMSD in the log PC values between the datasets (1.04 and 0.97, respectively, for buffer versus lysate and buffer versus extract). Although the lysate and extract have disparate total protein concentrations, their effects on compound partitioning are similar, with an R2 of 0.44 and an RMSD of 0.45 log PC units between the datasets (Fig. 4c). Thus, the differences between the buffer and lysate/extract environments appear to stem more from the number of molecular species rather than the concentrations of those species.

21As with the pure condensate systems, we were able to generate a features-based statistical model of compound partitioning into the SUMOSIM condensates in a Xenopus extract system (Fig. 4d). The model was somewhat less predictive in the more complex solution than in the simple buffer (R2validation = 0.39, MAEvalidation = 0.30), although this deficit may be due to the reduced number of compounds that can be analysed due to technical limitations (Supplementary Information and Supplementary Table 6) rather than an inherent limitation of modelling extract systems. The features driving the model in the extract system shared similarities with the buffer model: octanol/water partitioning and features related to hydrophobicity are still dominant, although measures of aqueous solubility are not important in the extract model (Extended Data Fig. 10b,c). Together, the data and modelling demonstrate that, even in highly complex solution environments that mimic the cytoplasm, the drugs show a wide range of PC values, which can be predicted by their physical properties.

22We have demonstrated that the partitioning of small molecules into condensates varies over nearly six orders of magnitude, from modest exclusion to strong enrichment. The combination of biochemical data and statistical modelling indicates that, in most cases, partitioning is largely driven by the aqueous solubility and hydrophobic character of compounds rather than their ability to stereospecifically bind to discrete sites on condensate scaffolds. In general, though, for compounds that do bind scaffolds through specific chemical structures, both effects should contribute to partitioning and could reinforce or oppose each other depending on their relative signs; the net behaviour will depend on both their signs and magnitudes. An unexpected similarity of partitioning profiles among several different condensates suggests that phase separation results in a shared, general physical or chemical property among the macromolecules that we have characterized. This property differentiates the physical environment within condensates from that of the surrounding solution, leading to enrichment or exclusion. The physical basis of the difference in environment remains unclear but may involve the solvent structure or ionic composition within the condensates or perhaps a greater abundance of transient, partially unfolded states of proteins that could create hydrophobic surfaces for recruiting small molecules non-stereospecifically,.

23Kilgore et al., recently described a ‘chemical grammar’ model for small-molecule partitioning, wherein distinct chemical environments occur in different condensates, which enhance the partitioning of molecules bearing specific chemical moieties. By contrast, our observations suggest a model of overall chemical similarity between condensates, with partitioning driven by compatibility between the physical properties of compounds and the similar chemical environments of the condensates. Key differences between the studies lie in the compound libraries and corresponding PC ranges. Kilgore et al. use a library based on three intrinsically fluorescent molecules that were diversified via combinatorial synthesis, which afforded a 15-fold variation in PC (0.9–13). By contrast, we used a library of metabolites and drugs that were not constructed around a common molecular core, which produced around a million-fold variation (~0.01–10,000). For data in the PC range of 0.9–13, we also observe substantial differences between condensates, with Pearson correlation coefficients ranging from 0.33 to 0.57, rather than 0.75–0.87 observed with our complete dataset (Extended Data Fig. 5c; see also Extended Data Figs. 6 and 7). Thus, within narrow PC windows, condensates show distinct patterns of partitioning, probably due to subtle variations in chemical environment and/or weak binding of scaffolds to compounds. But when sampling a wide range of chemical space and PC values, overall correlations in partitioning and thus global similarities in chemical environment between condensates are revealed. Our model suggests that physical properties are better predictors of partitioning than specific chemical structures. A limitation of our data is that all condensates we examined are formed by proteins with folded domains, whereas intrinsically disordered proteins and nucleic acids can also form condensates. Thus, it remains possible that other types of condensate have more distinct chemical environments. Future experiments on diverse condensate systems will help to clarify the generality and limitations of our observation for similar chemical environments among the four condensate systems studied here.

24A variety of approaches have been suggested for modulating the activities and/or physical properties of condensates to impact human diseases,,. The machine learning methods that we deployed here could accelerate these approaches by facilitating the rapid identification of highly enriching compounds. In diseases where therapeutic targets reside in condensates, modelling could guide the rational modification of existing inhibitors to increase their concentrations in the compartments. Existing drugs could also be grafted onto strongly partitioning compounds to direct them into condensates; both of these strategies could provide increased potency or selectivity towards targets. It is important to note that enrichment via the site-specific binding of a small molecule to one condensate component does not imply a thermodynamic advantage for binding any other component. Thus, the potency of a drug that targets one condensate component may not improve simply by adding moieties that specifically bind another component. However, our finding that strong enrichment can be achieved without stereospecific binding, potentially due to solvent properties, suggests that drug potency could be enhanced by an overall increase in compound concentration within the condensate not localized at any specific binding sites.

25The physical properties of condensates that form through phase separation, such as viscoelasticity and surface tension, are emergent. That is, they are inherently macroscopic, resulting from interactions of the parts of the system but not manifest in the parts individually. Here we have found that the ability to recruit small molecules appears to be an emergent biochemical property of SUMOSIM condensates. Twelve of the 13 drugs analysed using ITC do not bind the polySUMO–polySIM complex tightly in the absence of phase separation but are recruited strongly into SUMOSIM condensates. Thus, for these compounds the free energy of partitioning emerges only upon phase separation. Given that the compounds analysed via ITC were selected randomly (Supplementary Methods), this emergent behaviour is probably operative for many compounds. It may also occur for other condensates. Moreover, we have previously demonstrated that the membrane dwell time, and consequently the signalling activity, is higher in Nephrin/Nck/N-WASP condensates than in small complexes of the same molecules, again emerging as a function of size. Future work on how the biochemical activity that is resident in condensates varies as a function of the condensate size will reveal the extent to which emergent biochemistry is a general feature of condensates in biology.

26Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

27Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41557-024-01630-w.