Introduction

Approximately 25% of disease-associated missense variants are located in intrinsically disordered regions (IDRs)1, present in ∼63% of human proteins. However, studying variants in IDRs is challenging due to the lack of a fixed tertiary structure and limited evolutionary conservation2, as only a small proportion (∼15%) of IDRs exhibit high conservation and high pLDDT scores (AlphaFold23) which denote prediction confidence4. Numerous IDR variants are, therefore, annotated as variants of uncertain significance (VUSs), making it difficult to evaluate and predict their functional impact on diseases.

IDRs, particularly low-complexity IDRs, are crucial for the formation of membrane-less biomolecular condensates through phase separation, a complex and not fully understood physicochemical process in which molecules aggregate and segregate into distinct fluid phases5,6. Phase separation leads to the formation of membrane-less organelles called condensates which play essential roles in various biological processes. An increasing number of proteins are recognized for their roles via the formation of phase separation condensates. These condensates include the nucleolus and transcription factories in the nucleus7, stress granules8, and protein densities at neuronal synapses9 and inner-ear hair cells10,11,12.

Accumulating studies in IDRs reveal that missense variants in IDRs can perturb protein functions by altering phase separation13,14,15. Missense mutation impacts phase separation by altering specific residues, such as key residues for polar interaction16, pi-contact17,18, or other multivalent interaction19,20,21. These mutations can affect the IDR conformations, intra-molecular interactions, and inter-molecular protein-protein interactions, leading to abnormal condensate formations7. This can cause either loss22,23 or gain13,14,24 of natural phase separation, leading to misplacement and related gain or loss-of-function outcomes. Notably, the gain of unnecessary phase separation can contribute to disease phenotypes like neurodegenerative disorders such as Alzheimer’s and Parkinson’s diseases25,26,27. These pioneering works have confirmed the association between phase separation and disease variants in several cases.

Existing missense variant prediction algorithms typically rely on protein structural features or evolutionary features derived from multiple sequence alignments (MSA). Using such features, prediction algorithms of disease variants have made significant progress in predicting protein clinical relevance28,29,30,31,32, bridging the variant interpretation gap left by experiments due to the cost and labor constraints. However, for numerous disease variants located in IDRs which often lack a fixed structure7 and show poor evolutionary conservation2, encoding variants using these traditional features becomes unsuitable, making accurate prediction challenging. To address this challenge, we propose to employ phase separation as a promising feature for improving the prediction of IDR disease variants, given the role of numerous IDRs in phase separation associated with various diseases.

Machine learning algorithms have shown remarkable performance in predicting the propensity of proteins to undergo phase separation. Several sequence-related features have been identified that made phase separation prediction effective33,34,35,36,37,38,39 including pi-pi35, cation-pi interactions36,37,40, electrostatic interactions39,41, hydrophobic interactions38,40, and the valency and patterning of the low-complexity region8,36,42. Protein-protein interactions (PPIs) and post-translational modifications (PTMs) were also found to be promising features recently43,44,45. Missense mutations can cause pathogenic changes by altering phase separation. However, current sequence-based phase separation prediction algorithms are trained on a limited set of known phase-separating proteins, their ability to predict the impact of missense mutations on phase separation remains unclear and has not been systematically evaluated.

To enhance comprehension of the correlation between IDR missense mutations and diseases, we approached the issue from the phase separation standpoint and devised features to represent the underlying phase separation alterations resulting from the mutation. Using experimentally validated missense mutations that alter phase separation propensity in proteins naturally undergoing this process, we constructed a computational tool named PSMutPred to predict the effect of missense mutation on phase separation (Fig. 1). We demonstrated that missense mutations that impact the normal phase separation propensity can be predicted, and the direction of the shifts in the PS threshold caused by these mutations are also predictable.

Fig. 1: Overview of the study.
figure 1

The upper green panel illustrates PSMutPred, a machine learning approach designed to predict the effect of missense mutations on natural phase separation. Each mutation is converted into a feature vector and distinct models were employed for two main tasks: Identifying mutations that impact PS (termed ‘Impact Prediction’) and determining whether a mutation strengthens or weakens the PS threshold (labeled as ‘Strengthen/Weaken Prediction’). Additionally, PS features, including the output from PSMutPred, were evaluated for their utility in predicting the pathogenicity of missense variants (lower orange panel). dim. dimension.

Full size image

We next explored whether PS can be integrated into the pathogenicity prediction of disease variants. By analyzing PSMutPred scores for 520,000 + missense variants, we observed a positive correlation between the variant’s tendency to impact phase separation and its pathogenicity, especially in proteins prone to phase separation. We also observed that in proteins related to neurodegenerative diseases, disease variants that were predicted to enhance phase separation were proportionally more prevalent compared to those that might weaken phase separation. By integrating PSMutPred scores and other PS-related features into representative unsupervised pathogenicity prediction methods for missense variants (Fig. 1), such as EVE31 and ESM1b46, we observed significant improvement in prediction accuracy, especially for variants within low-conservation IDRs, with an ∼10% increase in AUPR. These findings demonstrate that PSMutPred not only can serve as a tool for predicting the impact of mutations on phase separation, but also can provide informative encoding for mutation impacts on IDRs. Additionally, phase separation offers fresh perspectives and opportunities for pre-screening studies on the pathogenicity of numerous VUSs.

Results

Collection of phase separation-related missense variants

To investigate relationships between missense mutations and protein phase separation (PS) properties, we reviewed missense variants with altered PS propensity that were documented in PhaSepDB47 and LLPSDB48,49 databases. To minimize the noise effect caused by multiple mutations within a single sequence, we narrowed our selection to mutation records with a limited number of mutations in the sequence that alter the normal PS threshold (annotated as ‘Impact’ mutations). Examples include P22L in Ape1, which solidifies semi-liquid Ape1 Droplets50, and S48E in TDP-43, which disrupts PS51. We limited our analysis to variants influencing proteins’ spontaneous PS, excluding partner-dependent PS.

Our compilation yielded a list of 307 experimentally validated ‘Impact’ mutation records from 70 proteins (Supplementary Fig. 1a and Supplementary Data 1), including 79 that strengthened the PS properties (annotated as ‘Strengthen’) and 228 that weakened or disabled them (annotated as ‘Weaken/Disable’). The PScore35 and PhaSePred43 scores indicate that the proteins from which these missense variants originate are predominantly proteins undergoing PS spontaneously, as the PS propensity scores predicted for these proteins are significantly higher compared to those of the human proteome (Fig. 2a).

Fig. 2: Analyses of mutations that impact phase separation (PS).
figure 2

a Comparison of the PS propensity of proteins corresponding to collected mutations (70 proteins) with that of the human proteome. (PScore35, Left; PhaSePred-SaPS43, Right; ****P < 0.0001, two-sided Mann–Whitney U test, p = 4.6e-11 and 3.3e-14, respectively; the boxplot components within each violin, from top to bottom are maxima, upper quartile, median, lower quartile, and minima.). b The proportion of ‘Impact’ mutations (Left) located in IDRs and Domains, compared with the total proportion of IDRs and Domains (Right). c The top 30 high-frequency mutations among collected ‘Impact’ mutations. d Distribution of amino acid (AA) distances from each mutation site to the nearest domain boundary. Distances of ‘Impact’ mutations and random ‘Background’ positions were compared within Domains (Left) and within IDRs (Right) (The number of data points were 139, 1000, 202, and 1000, respectively; ****P < 0.0001, two-sided Mann–Whitney U test, p = 4.4e-40 and 1.4e-30, respectively; the boxplot components within each violin plot from top to bottom are maxima, upper quartile, median, lower quartile, and minima). e Distribution of eight pi-contact prediction values (PPVs) for mutation sites. Values of ‘Impact’ mutations (in red) and ‘Background’ mutations (in gray) were compared. The dot in each violin represents the average of values. (NS not significant, **P < 0.01, ****P < 0.0001, two-sample Kolmogorov–Smirnov test; P-values are 0.0029, 0.140, 5.9e-11, 1.2e-7, 0.106, 3.3e-8, 4.5e-6, and 4.1e-14, respectively). f Statistical comparison of the changes of AA property index before and after mutation between collected ‘Strengthen’ (n = 79, orange) and ‘Weaken/Disable’ groups (n = 228, blue) under two-sample Kolmogorov–Smirnov D test (WT wild-type AA, MT mutant AA). The direction of the D statistic was set as positive when the mean value of the ‘Strengthen’ group was higher and as negative when that of the ‘Weaken’ group was higher. Source data are provided as a Source Data file.

Full size image

Properties of variants impacting phase separation

We observed that mutations impacting PS (annotated as ‘Impact’ mutations) are predominantly located in intrinsically disordered regions (IDRs) rather than structured domain areas (Domains) (Fig. 2b). Existing variant effect prediction models primarily rely on evolutionary features28,30,31,52 and structural features32,53,54,55. However, these features are inadequate for representing variants in IDRs and their effects on PS. We found that current advanced pathogenicity prediction models, such as AlphaMissense32 and EVE31, exhibit higher uncertainty of predictions for variants in IDRs than in Domains (Supplementary Fig. 1d). Moreover, they both fail to distinguish ‘Impact’ mutations from random background mutations (Supplementary Fig. 1c). This underscores the urgent need for developing specialized models to predict the effects of mutations on PS.

Among these ‘Impact’ mutations, those involving serine, tyrosine, arginine, lysine, and glutamine are the most prevalent (Fig. 2c and Supplementary Fig. 5a). These residues play key roles in PS. Specifically, tyrosine and arginine residues are important PS drivers, with cation-pi interactions between them serving as crucial forces for PS37. Although lysine has a lower ability to form pi-contacts compared to arginine, its interaction with nucleotide is essential for the PS of specific proteins such as tau and DDX3X56,57,58. Additionally, glutamine and serine are vital for cross-beta sheet interaction, which can enhance the propensity of PS37.

Interestingly, our analysis revealed that ‘Impact’ mutations tend to occur near domain boundaries. Specifically, the amino acid (AA) distance from each mutation site to its closest domain boundary was calculated. We noted that, whether within Domains or IDRs, ‘Impact’ mutations are significantly closer to these boundaries compared to random mutation sites (P ≤ 0.0001, Fig. 2d and Supplementary Fig. 5b). Furthermore, we observed that pathogenic mutants predicted by AlphaMissense32 also demonstrated shorter distances to domain boundaries compared to those predicted as benign in both Domains and IDRs (P ≤ 0.0001, Supplementary Fig. 1e).

In addition, we assessed pi-contact values at mutation sites using the prediction function provided by PScore35, as pi-pi interactions are vital in determining PS. For 6 out of the 8 predicted pi-contact values, ‘Impact’ mutations have higher predicted values than random mutation sites (Fig. 2e). This suggests that mutations affecting PS are likely to occur at sites with high pi-contact frequencies.

Next, we compared the changes in AA properties before and after mutation between the ‘Strengthen’ and ‘Weaken/Disable’ groups. We observed that missense mutations that strengthen PS propensity are prone to have higher AA mass (P = 0.0097), hydrophobicity (P = 0.0024), and decreased polarity (P < 0.0001) compared to ‘Weaken/Disable’ mutations (Fig. 2f and Supplementary Fig. 1f). Notably, the increase in hydrophobicity supports the previous discovery that hydrophobicity serves as an important PS driver38,40.

Together, these findings show the distinct properties of missense mutations impacting PS and the varying characteristics between mutations that either strengthen or weaken PS. Based on these properties, we have developed tools to predict the impact of mutations on PS, which will be discussed in the following section.

PSMutPred for predicting the impact of missense mutation on phase separation tendency

To predict the impact of missense mutations on protein PS properties, we developed PSMutPred composed of two machine learning (ML) approaches. The first approach termed the ‘Impact Prediction’ (IP) task, trained ML models to predict missense mutations that impact PS. In the second approach termed the ‘Strengthen/Weaken Prediction’ (SP) task, ML models were trained to predict the direction of the shifts in the normal cellular PS threshold induced by ‘Impact’ mutations (Methods). The process is depicted in Fig. 1, indicated by the green section.

To encode such changes caused by missense mutations for quantification and model learning, we considered the physicochemical properties of both the wild-type AA and mutant AA at each mutation site, as well as the properties of AAs within IDRs. We applied pi-pi contact frequency35 to encode the mutation site’s underlying significance for PS. Additionally, features such as IUPred59 score and mutation distance to domain boundary were used to encode the position of these mutations. Each mutation sample was converted into a 39-dimension feature vector (“Methods” section). Neither MSA features nor structural features widely adopted in existing mutation-related machine learning prediction28,29,30,31,52,53,60 were considered due to the nature of IDRs.

Due to the limited availability of variants that have no PS effect, for each of the 70 proteins, we randomly generated 500 single AA variants, resulting in a total of 35,000 random ‘Background’ mutations, which were further used as negative samples (i.e., ‘Background’ mutations) for analyses (“Methods” section). The dataset of missense mutations was divided into a cross-validation dataset (246 ‘Impact’ samples and 23,500 ‘Background’ samples from 47 proteins) and an independent test set (61 ‘Impact’ samples and 11,500 ‘Background’ samples from 23 proteins) (Methods). We trained prediction models including Logistic Regression (LR), Random Forest (RF), and Support Vector Regression (SVR) for both tasks to explore the discriminative power of predicting missense mutations’ effect on PS.

Performance evaluation of PSMutPred

Although there is currently no computational algorithm for missense mutations on predicting the effect on PS, we attempted to determine if existing PS prediction methods are capable of discerning alterations in PS propensity caused by missense mutations. Five high-performing PS methods were selected, including DeePhase39, PSAP61, PScore35, catGRANULE38, and FuzDrop62. We found that the prediction score differences of ‘Impact’ mutations were higher than those of ‘Background’ mutations, by comparing the absolute differences of prediction scores (Fig. 3a). However, the area under the curve of the receiving operating characteristics (AUROCs) were unsatisfactory (Fig. 3b). Except for FuzDrop62, none of the methods could differentiate between ‘Strengthen’ mutations and ‘Weaken’ mutations (Supplementary Fig. 2c, d), as their prediction scores did not accurately reflect the increase or decrease of PS propensity caused by mutations.

Fig. 3: Evaluation of methods’ performance on predicting missense mutations that impact natural phase separation (PS) (‘Impact’ mutation).
figure 3

Methods with abbreviations include DeePha (DeePhase), FuzDro (FuzDrop), and catGRA (catGRANULE). a Discriminative power evaluation of representative PS prediction methods for ‘Impact’ mutations against random ‘Background’ mutations, comparing absolute score changes pre- and post-mutation (P-values computed by a two-sided Mann–Whitney U test, left for ‘Impact’ mutations, n = 307 and right for ‘Background’ mutations n = 35,000; the boxplot components within each violin plot, from top to bottom are maxima, upper quartile, median, lower quartile, and minima.). b Performance evaluation of representative PS prediction methods on discerning ‘Impact’ mutations against random ‘Background’ mutations (IP task). AUROC is based on the absolute score changes. c–h Model performance evaluation in identifying ‘Impact’ mutations. For LOSO, 50 replicates of subset sampling from the background dataset were used to evaluate performance, and the average AUROC and the area under the curve of the precision-recall curve (PRC) (AUPR) were computed and visualized. For LOSO AUPR, data are presented as mean values ± SD (Standard Deviation), and the scatter points represent the distribution of background dataset sampling repeats. c, d Model performance in identifying ‘Impact’ mutations evaluated using leave-one-source-out (LOSO, Left) and an independent test set (Right), measured by AUROC (c) and AUPR (d). e, f A parallel evaluation similar to (c and d) but the ‘Background’ mutations were generated following the same IDRs: Domains ratio as the collected ‘Impact’ samples (weighted sampling). (g, h) A parallel evaluation similar to (c, d) but the ‘Background’ mutations were generated by aligning the frequency of different mutations with their frequency in the impact dataset (AA weighted sampling). We assigned weights to each type of mutation based on the number of occurrences in the impact dataset, with a minimum weight of 1 to ensure all mutation types are considered. Source data are provided as a Source Data file.

Full size image

To test whether PSMutPred-IP can identify mutations that impact PS, we implemented leave-one-source-out cross-validation (LOSO CV) (Method). LOSO is used to evaluate the models’ predictive capabilities for variants from unseen proteins, ensuring their generalizability. AUROCs and AUPRs on LOSO (Fig. 3c, d) showed the accuracy of our models, especially IP-SVR and IP-RF. Evaluation results on the independent test set showed that our models have stable prediction performance when facing mutations from different PS protein categories (Fig. 3c, d).

To ensure that the superior performance of the algorithm was not caused by the distribution bias of ‘Impact’ mutations, which tend to be located in IDRs (Fig. 2b, d), we conducted an additional LOSO CV and an independent validation for the IP task. Here we generated ‘Background’ samples maintaining the same IDRs: Domains ratio as observed in the ‘Impact’ samples. LOSO CV results and independent validation results (Fig. 3e, f) were as promising as those from the previous dataset (Fig. 3c, d). We also created another ‘Background’ dataset by aligning the AA substitution frequencies with those in the impact dataset. This was done to test whether the predictive power was due to the high ratio of specific AA types in ‘Impact’ samples (Fig. 3g, h).

Moreover, a small proportion of ‘Impact’ mutations are multi-point mutations (Supplementary Fig. 1b), the proposed models still exhibited efficient predictive power and outperformed representative PS models when analyzing the results without considering multi-point mutations (Supplementary Fig. 2a, b). We also assess the performance of PSMutPred-IP by only including naturally occurring mutations in the database, and the model allowed efficient predictive power on these mutations (Supplementary Fig. 5d, e).

We next investigated the performance of PSMutPred-SP models. The significant discriminative power to distinguish between ‘Strengthen’ samples and ‘Weaken’ mutation samples from unseen proteins under LOSO CV (SP-LR, P < 0.0001) (Supplementary Fig. 2e, f) and AUROCs on the independent test set (Supplementary Fig. 2g) indicated that SP-LR, SP-RF models can identify the direction of the shifts in PS caused by missense mutations.

Feature importance indices were calculated for random forest models of both tasks and grouped into feature types to discover potentially key features (Supplementary Fig. 2h). In both tasks, the pi-contact frequency of the mutation site ranked first, suggesting that the pi-pi interaction at the mutation’s specific location significantly influences the effect of mutations on protein phase separation. Beyond Pi-contact, no single feature was identified as a dominating one, suggesting the non-linear nature among features and a multi-factor causal relationship between features and the PS outcome.

Overall, PSMutPred models not only predict missense mutations that impact PS but also predict the direction of the PS-threshold shift. Our final model generates predictions made by the IP-SVR, IP-RF, IP-LR, SP-LR, and SP-RF models as well as their corresponding rank scores (Methods). These models can be employed as effective tools for assessing the tendency of missense mutations to affect PS, enhancing the interpretation of disease variants’ pathogenicity.

Experimental validation on the PS-related mutations identified by PSMutPred

Aberrant phase separation (PS) or aggregation processes have been implicated in the pathogenesis of various diseases, including neurodegenerative disorders, autism7, and hearing loss10,11,12. In this study, we employed PSMutPred to predict the PS impact (‘Impact prediction’ and ‘Strengthen/Weaken prediction’) of missense variants from genes associated with PS-related diseases. Specifically, we selected EPS8, known for its association with deafness, and analyzed the PS impact of its ‘Uncertain’ missense mutations from the ClinVar63,64 database to assess the accuracy of our prediction model.

Epidermal growth factor receptor pathway substrate 8 (EPS8) is a multifunctional protein involved in cell mitosis and differentiation65,66,67,68, in capping proteins through side-binding, in bundling of actin filaments69,70,71 as well as in the elongation of actin in hair cell stereocilia65. Prior research has shown that EPS8 localizes to the tips of stereocilia and contributes to the formation of PS-mediated condensates at the stereocilia tip complex11,72. These findings suggest that EPS8 has the capacity for self-phase separation and to interact with other molecules.

To validate our predictions, we selected missense mutations, including R265C, D586G, and K676R, from the 8 candidate mutations. These candidates were predicted by PSMutPred-IP to impact PS, either by strengthening or weakening it, with rank scores above 0.5 across all models (IP-SVR, IP-RF, and IP-LR). Among these candidates, D586G had the highest scores for both SP-LR and SP-RF, suggesting a stronger propensity to enhance PS, while K676R scored the lowest for both metrics implying an impairment of PS capacity. As negative controls, we selected R702W and E728V from the 15 candidate mutations (all with IP rank scores below 0.5 for IP-LR, IP-SVR, and IP-RF; at least one had a score below 0.1).

For experimental validation, we overexpressed the wild-type and mutants of mouse Eps8 (which is highly conserved with human EPS8, as shown in Supplementary Fig. 3d) fused with a GFP tag in HEK293 cells. Observations made using Olympus fluorescence microscopy highlighted distinct changes in puncta formation quantity to evaluate PS capacity. Specifically, R265C and D585G exhibited a notable increase in the number of puncta, while K675R showed a significant reduction (Fig. 4a, b and Supplementary Fig. 3a, b). Notably, D586G (equivalent to D585G in mice) demonstrated enhanced PS capacity, aligning with its high SP-RF and SP-LR scores. In contrast, K676R (equivalent to K675R in mice) showed diminished PS ability, supported by its low SP-RF and SP-LR scores. To confirm that the observed puncta are indeed a result of liquid-liquid phase separation (LLPS), we also conducted a fluorescence recovery experiment after a photobleaching (FRAP) experiment to validate the dynamic and rapid formation of droplets observed of both wild-type and mutant (D585G) of mouse EPS8 in HEK293 cells (Supplementary Fig. 4h–j).

Fig. 4: Experimental validation of Eps8 missense mutations predicted by PSMutPred to impact PS.
figure 4

a Representative images of overexpressed GFP-Eps8 and its mutants in HEK293 cells (scale bars: 10 μm; n = 10 randomly picked cells). WT denotes wild type. b Quantification of puncta within the wild type and mutants of Eps8 in HEK293 cells (n = 10 randomly picked cells; ****P < 0.0001 by two-tailed Student’s t test, p = 8.1e-9 and 4.0e-10, respectively). Error bars represent SD, and center lines represent mean values. c Ribbon diagram representation of mouse EPS8 structure predicted by AlphaFold2, showing both front (left) and back (right) views. d–f Detailed regions involving missense mutations with their neighboring residues (d, f), and interaction analysis (e). The mutations are shown with the stick mode in red while hydrogen bonds are shown as blue dashed lines. Sequence alignments within critical residues are shown in bold. Source data are provided as a Source Data file.

Full size image

Eps8 comprises six domains, encompassing both well-structured regions and intrinsically disordered regions (IDRs) (Supplementary Fig. 3c). D585 in mice (equivalent to D586 in humans) is situated in the SH3 domain. It forms stable hydrogen bonds with the sidechains or backbones of K536 and K538 in mice (equivalent to K537 and K539 in humans), maintaining structural stability through either a β-sheet-loop conformation in mice or β-sheet-β-sheet conformation in humans (Fig. 4d, e and Supplementary Fig. 3e, f). Additionally, D585 and its neighboring residues, as well as interaction residues, exhibit high conservation. However, the substitution of glycine for aspartate disrupts the formation of these hydrogen bonds, leading to a destabilized structure. This change might enhance PS by increasing flexibility and modifying the β-sheet-loop structure. Conversely, K675 in mice (equivalent to K676 in humans) is located in an IDR at the C-terminus of EPS8 (Fig. 4c and Supplementary Fig. 3d). Sequence alignment underscores the high conservation of K675 and its adjacent residues (Fig. 4f and Supplementary Fig. 3g). Lysine’s characteristic long and hydrophobic side chain increases the likelihood of extending and ‘capturing’ polar residues from self-proteins or other proteins. This substitution may alter these interactions, potentially impairing PS.

In summary, our study demonstrates that the changes in PS propensity, as predicted by our model, accurately correspond to a distinct number of puncta formed in live cells. This serves to underscore the utility and accuracy of PSMutPred.

Phase separation effect prediction of disease variants by PSMutPred

Phase separation turns out to be a general mechanism for protein condensate assembly forming as membrane-less organelles in various physiological processes9,73,74,75,76,77,78,79,80,81,82. Mutations that change phase separation (PS) are likely causes of disease13,22,23,83. To exam this theory, we divided missense variants corresponding to 8611 human genes from ClinVar63,64 (522,016 variants in total, downloaded in 2022.12) into PS-prone and low-PS-prone propensity groups for between-group analyses (Methods).

By comparing the PS-prone group defined by 83 PS ClinVar proteins with the low-PS-prone group (other proteins), we found that in the PS-prone group, the PSMutPred-IP scores of variants (n = 1451) were skewed to pathogenicity compared to those from low-PS-prone group (n = 86,291) (Fig. 5a). To make the sizes of the two groups more comparable and improve the efficiency of the comparison, we also defined the PS-prone group by collecting PS proteins predicted by PScore35 and a meta-predictor PhaSePred43 (1276 proteins, 30,889 variants) and grouped variants from other proteins as low-PS-prone group (56,853 variants). The between-group analysis showed consistent results (Fig. 5b). However, when conducting the between-group analysis using representative PS prediction methods trained on phase-separation proteins, the outcomes for DeePhase39 and catGRANULE38 displayed inconsistencies across the analyses (Supplementary Fig. 6a, b). For the other three methods, which include PSAP61, PScore35, and FuzDrop62, the PS-prone group did not demonstrate significantly higher Pearson correlation coefficients than the low-PS-prone group (Supplementary Fig. 6a, b).

Fig. 5: Evaluation of PSMutPred scores across ClinVar64 variants.
figure 5

a–d Comparison of variants’ Pearson correlation between groups. Groups include a PS-prone group (83 known PS proteins, 1451 variants) and a low-PS-prone group (8528 proteins, 84,840 variants) defined by PS proteins, and a predicted PS-prone group (1276 proteins, 30,889 variants) and a predicted low-PS-prone group (7335 proteins, 56,853 variants). (two-tailed P-values computed by sci-kit learn pearsonr package; *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001; NS = no significance). a Comparison between the PS-prone group and the low-PS-prone group. P-values are 1.7e-8, 5.8e-8, 2.6e-4, 3.4e-5, 2.4e-14, and 1.2e-275 respectively. b Comparison between the predicted PS-prone group and the predicted low-PS-prone group. P-values are 9.9e-82, 0.01, 4.9e-53, 0.02, 3.6e-198, and 7.7e-223 respectively. c Comparison between variants located in IDRs (n = 15,427) and Domains (n = 15,462) within the predicted PS-prone group. P-values are 6.4e-61, 2.8e-13, 1.7e-23, 6.0e-37, 1.5e-149, and 1.4e-107 respectively. d Comparison between variants from neurodegenerative disease (ND) related proteins (19 proteins, n = 252) and variants from other proteins (non-ND) (within the predicted PS-prone group). P-values are 0.88, 3.1e-8, 0.005, and 2.6e-95 respectively. e AUROC scores of PSMutPred-IP models on pathogenicity prediction of IDR missense variants from the PS-prone group (n = 489 variants). f A parallel evaluation of (e) but focuses on the predicted PS-prone group (n = 8188). g Comparison of the proportion values defined by different PSMutPred-IP models, including IP-RF (top), IP-LR (middle), and IP-SVR (bottom). Comparison of the PS-prone group and the low-PS-prone group on the left (PS proteins), and comparison between the predicted PS-prone group and the predicted low-PS-prone group (Predicted-PS proteins). Differences are based on 2-sample Kolmogorov’s D statistic, with positive values indicating higher proportions in the PS-prone group and negative values indicating higher proportions in another. Source data are provided as a Source Data file.

Full size image

IDRs and structured domains are both crucial for phase separation6. When analyzing the PS-prone group defined by algorithms, we found that PSMutPred-IP, along with other sequence-based PS metrics (except PSAP), showed a significant positive correlation with the pathogenicity of variants in both IDRs (n = 15,427) and domains (n = 15,462) (Fig. 5c and supplementary Fig. 6c).

Additionally, we isolated variants likely in IDRs (unmapped by PfamScan84,85, with IUPred359 score > 0.5) from the PS-prone group and directly used PhaSePred-IP scores to predict pathogenicity ((Likely) pathogenic as 1 and (Likely) benign as 0). Separate analyses were conducted for the PS-prone group as defined by PS proteins and by algorithms (Fig. 5e, f). The resulting AUROC scores indicated that PSMutPred-IP can identify disease variants that lead to PS alterations.

Studies indicate that neurodegenerative lesions may be associated with excessive PS25,86. Within the predicted PS-prone group, 19 proteins were identified to be highly associated with neurodegenerative disease (Methods). SP prediction scores for the variants (n = 252) of these proteins showed different patterns compared to other variants (n = 19,266). Specifically, among these 252 variants, those predicted to strengthen PS were more inclined towards pathogenicity compared to those predicted to weaken PS. This observation was reflected by the positive Pearson coefficient values for both SP-LR and SP-RF models, contrasting with the negative coefficient values for variants from other PS proteins (Fig. 5d). catGRANULE38 and FuzDrop62 also demonstrate a higher Pearson correlation coefficient than the overall pattern (Supplementary Fig. 6d).

Additionally, we observed that mutations within IDRs of PS proteins are more likely to affect PS than those in proteins less likely to undergo PS. This conclusion was based on evaluating the proportion of variants of ‘Uncertain Significance’ (VUSs) predicted to impact PS for each protein (see Methods). Using Kolmogorov’s D statistic to compare these proportions, we observed that the PS-prone group had a higher incidence of PS-affecting variants than the low-PS-prone group, as shown in Fig. 5g. We found that only for PSAP61 and PScore35, PS-prone proteins have a higher proportion of missense mutations predicted to ‘Impact’ PS (Supplementary Fig. 6e).

In summary, our comprehensive analyses revealed a notable clustering of pathogenic missense mutations with an impact on PS propensity in proteins with inherently higher PS propensity. Additionally, our findings indicate that compared to mutations that might weaken PS, gain-of-PS mutations tend to aggregate specifically in disease mutations associated with neurodegenerative-related genes. These results not only provide valuable insights into the relationship between PS and disease but also underscore the reliability and validity of PSMutPred.

Introducing a feature for the pathogenicity prediction of disease variants

Current variant interpretation methods heavily rely on evolutionary features28,30,31,52,87 generated by multiple sequence alignment (MSA). As IDRs, especially those with poor evolutionary conservation2, challenge the effectiveness of traditional features7, using two representative pathogenicity prediction models, including EVE31 and ESM1b46,88, we aim to test whether PS-related features can address the IDR gaps left by evolutionary features in pathogenicity prediction (Fig. 1, orange section). EVE predicts pathogenic variants by thoroughly leveraging MSA information, demonstrating that this feature alone can predict the impact of most known disease mutations, proving the dominant role of the evolutionary feature in this field. ESM1b utilizes large language models (LLM) to learn protein information across species, modeling the space of known protein sequences selected throughout evolution, and can thus be considered an advanced representation of evolutionary features.

We first selected EVE and applied a straightforward approach, that combined the unsupervised EVE score with a simple feature group, including PSMutPred scores as the variant feature (Methods). We then trained three models including RF, SVR, and LR, using ClinVar64 significances as labels. Their performances were evaluated using both blocked 3-fold cross-validation and an independent test set (Supplementary Data 3) (Methods). We observed that all three combined models demonstrated improved pathogenicity prediction, with RF showing the best performance in terms of AUROC and AUPR scores (Supplementary Fig. 4a, b, and Supplementary Table 1). We subsequently selected the RF model as the combined model for further analysis, based on validation on the independent test set.

Given that IDRs typically exhibit poorer evolutionary conservation than Domains7, it is unsurprising that we found EVE to be less effective in predicting IDR variants compared to Domain variants (Fig. 6a, b and Supplementary Table 1). As expected, the combined model led to a more pronounced improvement in identifying IDR disease variants (n = 5656) than those within Domains (n = 9738). Specifically, the combined model showed a 4.3% improvement in AUROC and a 7.6% improvement in AUPR for IDR variants, compared to a 2.6% AUROC improvement and only a 1.7% AUPR rise for Domain variants (Fig. 6a, b and Supplementary Table 1). The Mann–Whitney test further indicated a significant improvement in the prediction of both pathogenic variants as well as benign variants in IDRs (Fig. 6c, d). Additionally, we consider that IDRs include a small subset of potentially conditionally folded IDRs with high evolutionary conservation4, characterized by high AlphaFold23-predicted confidence scores (pLDDT scores). For the 5656 IDR mutations in the test set, we mapped their pLDDT scores and categorized them into a high pLDDT group (pLDDT ≥ 70) and a low pLDDT group (pLDDT < 50). We found a more pronounced improvement in AUPR for IDR variants with low pLDDT scores compared to those with high pLDDT scores (9.8% compared to 3.9% AUPR improvement, Fig. 6e). This indicates that current PS-related features can supplement the evolutionary feature’s weakness in IDRs, especially for low conservation IDRs.

Fig. 6: Analysis of phase separation-related feature contributions to pathogenicity prediction.
figure 6

a–e Pathogenicity prediction performance evaluation of the model combining EVE with PS-related features. a AUROC (Left) and AUPR (Right) evaluations on the independent test set (n = 15,394). The purple line represents the model trained with both EVE and PS features; the green line represents the EVE score alone. b AUROC (Left) and AUPR (Right) evaluations specifically on variants within IDRs from the data set analyzed in (a) (n = 5656). c, d The divergence of predicted scores distributions between the standalone EVE (green) and the combined model (purple), quantified using a two-sided Mann–Whitney U test on the independent test set (****P < 0.0001; P-values are 2.4e-27; 1.7e-15, 9.6e-59; and 7.2e-293 respectively, the boxplot components within each violin plot, from top to bottom are maxima, upper quartile, median, lower quartile, and minima.). c Score distributions for pathogenic-prone variants (pathogenic and likely pathogenic, n = 2044, left graph) and benign-prone variants (benign and likely benign, n = 3612, right graph) with a focus on variants located in IDRs. d A parallel evaluation of (c) but focusing on variants located in Domains (6665 pathogenic or likely pathogenic and 3073 benign or likely benign). e Evaluation of IDRs variants with high AlphaFold2 pLDDT scores (pLDDT ≥ 70, n = 2763) and low pLDDT scores (pLDDT < 50, n = 2407). f–i Pathogenicity prediction performance evaluation of the model combining ESM1b with PS-related features. f Evaluation of the model trained with ESM1b and PS features using 5-fold cross-validation under the ClinVar dataset (n = 140,321). g Evaluation of IDRs variants with high AlphaFold2 pLDDT scores (pLDDT ≥ 70, n = 36,032) and low pLDDT scores (pLDDT < 50, n = 25,755). h, i Pathogenicity prediction for 1,015,769 ClinVar VUSs by combining PS features with ESM1b scores. Source data are provided as a Source Data file.

Full size image

To investigate the role of PSMutPred in pathogenicity prediction, we examined the effect of including PSMutPred scores within various feature combinations, and their performance on the test set was quantified by AUROC and AUPR (Supplementary Fig. 4c). The comparative analysis highlighted a significant enhancement in the prediction accuracy when PSMutPred scores were incorporated. Next, we analyzed ClinVar mutations predicted by EVE with opposite outcomes (pathogenic/likely pathogenic mutations with EVE scores < 0.5 and benign/likely benign mutations with EVE scores ≥ 0.5) to evaluate the PSMutPred’s discriminative power on these mutations. We focused on mutations within disordered regions of potential phase-separating proteins (n = 600, Supplementary Fig. 4d). PSMutPred scores for false negatives of EVE are significantly higher than those for false positives of EVE, with P-values of 1.3e-4, 9.7e-5, and 9e-3 for IP-RF, IP-SVR and IP-LR, respectively, as evaluated by the Mann–Whitney test. This shows that PSMutPred can capture information of mutants where evolutionary features failed.

Subsequently, similar to the process of appending features to EVE, we combined PS-related features with the ESM1b score, which can be considered as an advanced representation of evolutionary features. We mapped ESM1b scores to 140,320 ClinVar variants (Supplementary Data 3) and tested the combined RF model using a blocked 5-fold cross-validation (Methods). The combined model also achieved improved prediction accuracy (Fig. 6f), especially for variants located in IDRs with low pLDDT scores (Fig. 6g). This consistent result indicates that PS-related features can address the weaknesses of evolutionary features in predicting IDR variants, particularly those in IDRs with low conservation. Using the PS features, combined with ESM1b scores, we predicted pathogenicity scores for 1,015,769 ClinVar VUSs (Supplementary Data 4) (“Methods” section). Among them, 527,524 are IDR variants (Fig. 6h), 9.3% of them were predicted pathogenic, and 78.4% were predicted benign (Fig. 6i).

We chose EVE, ESM1b due to their unsupervised nature, which can offer an unbiased baseline, making testing with our features straightforward. These findings reveal that the PS-related features including variants’ impact on PS serve as a valuable encoding for IDR mutations and can be integrated into pathogenicity prediction models in the future to provide a better interpretation of pathogenicity variants.

Discussion

Missense mutations in IDRs are overlooked compared to those in folded regions. Pathogenicity prediction models, such as AlphaMissense32, typically rely on evolutionary information and protein structural features. However, since IDRs often lack a consistent structure and show limited evolutionary conservation, using these traditional features to encode variants in IDRs turns out to be unsuitable, making interpreting the numerous VUSs in IDRs difficult2. To address the challenge, we turn to phase separation (PS), a widely acknowledged property of IDRs linked to various diseases. As recent studies emphasize, mutations within IDRs can disrupt regular PS, which can lead to disease22,23, highlighting its potential as a predictive feature.

Although a comprehensive understanding of the specific mechanisms by which missense mutations alter PS remains elusive7, we considered differences in physicochemical properties, viewing them as features that quantify mutation-induced changes. Additionally, we used the predicted pi-pi contact frequency35 at the mutation site as a position-specific encoding. These considerations were used to develop models predicting the effect of missense mutation on PS. Separate models were constructed to predict the ‘Impact’ and ‘Strengthen/Weaken’ effects of missense variants. The reliability of the model predictions was validated through prediction results on variants from unseen proteins and further experimental validations. Moreover, subsequent analyses conducted on a larger-scale dataset further affirmed the accuracy and effectiveness of our models.

It is crucial to note that, for single or a few key amino acids, changes at the hot spots involved in charged or multivalent interactions can impact phase separation. These key residues vary among different proteins, with some being specific residues responsible for post-translational modifications (PTMs)89,90, polar interaction16, or pi-contact17,18 within low complexity domains (LCD), while others are residues located at domain surfaces or boundaries facilitating multivalent interactions19,20, or contributing to cross-beta structures within LCD21. Given this, we developed machine learning models using both randomly generated mutations and experimentally validated missense mutations that alter PS propensity. These mutations mainly involve single AA mutations, aiming to uncover the special characteristics of ‘hot spots’ mutations that alter phase separation. The discriminative power of our algorithms indicates an inherent pattern among missense variants impacting PS, allowing for their prediction.

We currently focus on predicting single-site mutations because single AA substitutions play a more critical role in diagnosing Mendelian diseases and are more prominently featured in current disease datasets, such as those in ClinVar64, compared to multi-site mutations. Multi-site mutations, which we are unsure how to perfectly encode, are currently treated as supplementary to our dataset to help the model learn information about AA properties or segments that are important for phase separation. The model trained in this manner has shown a good ability to predict the performance of a single AA mutation that alters phase separation. Our final model only predicts the effect of single AA mutations at present. Going forward, as the dataset expands, different approaches to processing multi-site mutation samples can be explored and the development of a more comprehensive multi-site mutation prediction approach can be employed to improve the accuracy and adaptability of PSMutPred.

Through analysis of PSMutPred scores using ClinVar64 variants, we discovered that proteins prone to PS have a higher proportion of disease variants altering PS, compared to proteins less likely to undergo this process. Among these mutations, those likely to enhance PS were proportionally more prevalent in neurodegenerative-disease-related genes than those that weaken PS, which supports the current discoveries25,86. Therefore, we can confidently assert that PSMutPred scores serve as valuable encodings for assessing the potential impact of missense mutations on PS, making them applicable to a broader range of research.

We further combined PS-related features, including PSMutPred scores, with evolutionary scores31 as well as ESM1b46 to predict pathogenicity and assessed the performance using ClinVar labels. PSMutPred targets the ‘IDR gap’ by evaluating variants in IDRs by their potential effects on phase separation. Consequently, this approach showed a more comprehensive interpretation of pathogenicity, especially in IDRs, our intuitive and straightforward collaborative approach offers a more comprehensive interpretation of pathogenicity across both domains and IDRs. It further demonstrates that the impact of phase separation features effectively complements the current focus on pathogenicity interpretations, which predominantly consider ordered regions. We did not build a model from scratch, instead, we highlighted the contribution of PS to improving the accuracy of pathogenicity predictions, offering a fresh perspective in this field. Looking ahead, leveraging diverse data such as weak labels from population frequency and variants from HGMD91 and gnomAD92 may further refine these predictions. Additionally, advanced machine learning methods, including deep learning, hold the potential to boost accuracy.

However, except IDRs, some folded domains also play important roles in driving phase separation, such as multivalent tandem structured domains6. Our feature encoding metrics currently favor mutations that impact IDRs rather than those impacting folded domains. Due to the limited training data available and to avoid overfitting, we chose simpler feature encoding and traditional machine learning models instead of developing a complex encoding method to comprehensively capture the effects of mutations in structured domain areas. In subsequent evaluations of pathogenicity, we also assessed the pathogenicity of mutations within domains. The results show that the improvement in predictive accuracy for mutations in IDRs was significantly greater than for those in domain areas. Including more high-quality, experimentally validated data, especially regarding the effects of missense mutations on structured domains, and integrating more comprehensive features could refine the prediction of mutation impact on PS.

Ultimately, our study not only presents methods to predict the effects of missense mutations on PS but also contributes to the prediction and improved understanding of VUSs occurring in IDRs, enhancing the diagnostic accuracy for rare genetic disorders.

Methods

Data acquisition

Experimentally validated missense mutations that impact phase separation (PS) and their corresponding experimental sequences were curated from the PhaSepDB47 and LLPSDB48,49 databases up to 2022.11. We focused on entries that documented individual proteins undergoing phase separation, rather than multi-protein co-phase separation; no additional filtering criteria were applied. In the corresponding literature of each entry, proteins involved in these entries were reported to undergo phase separation, along with the mutations that resulted in changes to their phase separation propensities. We retained single amino acid (AA) mutation samples as well as multi-AA mutations with a number of mutation sites less than or equal to five. We got 214 single AA mutation samples, 43 two-mutation-sites-samples, 26 three-mutation-sites-samples, 18 four-mutation-sites-samples, and 6 five-mutation-sites-samples (Supplementary Fig. 1b). We finally obtained 307 samples corresponding to 70 proteins including 79 ‘Strengthen’ samples that strengthened PS, and 228 ‘Weaken/Disable’ samples that weakened or diminished PS (Supplementary Data 1).

As a limited number of experimentally validated mutations that do not affect PS are available for reference, we created ‘Background’ samples by generating random single AA missense mutations with the wild-type sequences for each of the 70 proteins. Mutations already included in the ‘Impact’ samples were excluded. For each gene, we generated 500 random mutations, leading to a total of 35,000 mutations.

It should be noted that this does not imply that background mutations cannot alter PS under any conditions, instead, it allowed us to learn missense mutations that impact PS.

Statistical analysis of phase separation-related missense mutations

We used PfamScan84,85 to predict the structural domain regions, abbreviated as ‘Domains’, within the experimental sequences. Any segment of the sequence not identified as a domain by PfamScan was annotated as an IDR. Multi-AA mutation samples were decomposed into single AA mutations (484 mutations) for subsequent statistical analysis. To analyze the distance of mutation points to domain boundaries, mutations within proteins that have both IDRs and domains were chosen (341 mutation sites), and 1000 random mutation sites were selected for each comparison.

We estimated the pi-contact frequency at the mutation site using the pi-pi interaction prediction function in PScore35. These values correspond to 8 sp2 groups. The categorization process divides pi-contacts based on: (1) short-range (4 sequence separation) vs. long-range (>4), (2) sidechain vs. backbone, and (3) absolute predicted frequency vs. relative difference from sp2 groups with the same identity respectively (see PScore35 for more details).

To analyze the changes in the AA properties pre and post-mutation, we calculated the index differences for properties including hydrophobicity93, polarity94, and mass95. The Kolmogorov–Smirnov test was then applied to compare these index differences between the ‘Strengthen’ and ‘Weaken/Disable’ mutation groups.

Development of PSMutPred

Training and testing dataset

We divided the mutation samples into a cross-validation dataset (47 human proteins) and an independent test set (23 non-human proteins). ‘Non-human’ in this context specifically refers to experimental protein sequences originating from species other than humans. Specifically, we classified the proteins based on their HGNC gene names. The 47 human proteins were those found in the Human Uniprot database, while the remaining 23 proteins, which were not found in the Human Uniprot database, were labeled as non-human proteins.

In the case of the ‘Impact Prediction’ (IP) models (‘IP task’ in short), we grouped ‘Strengthen’ and ‘Weaken/Disable’ mutations into one group and set their labels to 1, while the labels of ‘Background’ mutations were set to 0. For the IP task, we got 246 ‘Impact’ mutation samples, and 23,500 ‘Background’ samples in the cross-validation dataset, 61 ‘Impact’ samples, and 11,500 ‘Background’ samples in the independent test set. In the case of the ‘Strengthen/Weaken Prediction’ (SP) models (‘SP task’ in short), labels of ‘Strengthen’ mutations were set to 1, and labels of ‘Weaken/Disable’ mutations were set to 0. For the SP task, we got 174 ‘Weaken/Disable’ samples and 72 ‘Strengthen’ samples in the cross-validation dataset; 7 ‘Strengthen’ samples, and 54 ‘Weaken/Disable’ samples in the independent test set.

Machine learning features

We used a set of simple features (39 dimensions) to encode each mutation entry. The two tasks share a common set of feature encoding. For multi-point mutations, certain features were adjusted to account for the presence of multiple mutation sites. The details of the features for each sample are described below:

IDR-related features of the mutation site(s) (5 dimensions): The structural domains of the experimental sequences corresponding to the sample were predicted using PfamScan84. Segments not predicted as Domains were designated as IDRs. (1) One binary value indicating whether the site is located in IDRs. For multi-point mutation, this value is set to 0 if any of the mutation sites are located within Domains; (2) One value quantifies the residue distance of the mutation site to its nearest domain boundary. For multi-point mutation, this value reflects the average distance of all mutation sites; (3) One value representing the likelihood of the mutation site being within IDRs predicted by IUPred359. For multi-point mutations, we use the average of the predicted values for all mutation sites; (4) One value representing the average of the full sequence IUPred359 scores; and (5) One binary value indicating whether the protein sequence has IDRs or not.

The predicted pi-contact frequency at the mutation site (8 dimensions): We derived 8 predicted values for pi-pi interactions at the mutation site from the wild-type sequence using the pi-pi interaction prediction function in PScore35 (see PScore35 for more details). In the case of multi-point mutation, we computed the 8 values as the mean scores across all mutation positions.

Physicochemical feature encoding for the mutation site (26 dimensions): We targeted the changes in AA characteristics at the mutation site whenever possible by employing five physicochemical indices that were previously used for protein feature encoding96,97,98,99. We selected hydrophobicity93, polarity100, volumes of side chains (VSC) of amino acids101, solvent-accessible surface area (SASA)102, and net charge index (NCI) of side chains of amino acids96,103. For each index, we computed four feature values: (1) Index value for the wild-type AA at the mutation site. For multi-point mutation, we used the mean value; (2) Numerical difference in index value between mutant and wild-type AA, we summed the difference values for multi-point mutations; (3) Average index values for all AAs in the wild-type sequence’ IDRs (defined as areas outside of Domains predicted by PfamScan84,85), and (4) the difference between the index value of mutant AA and the average index value (from point 3), for multi-point mutation, we accumulated the difference values for each mutant AA. Additionally, six binary values (6 dimensions) were used to capture the presence of three properties (positively charged, negatively charged, and hybridized) in both the mutant AA and wild-type AA. For multi-point mutations, an element is set to 1 if at least one AA exhibits the respective property.

Machine learning algorithm and model performance evaluation

For both the IP task and the SP task, to evaluate the generalizability of our models on variants from unseen proteins, we implemented a blind test called ‘leave-one-source-out cross-validation’ (LOSO CV). In this approach, for each validation iteration, we held out variants from a single protein from the total set of proteins (variants from cross-validation dataset; 47 proteins). We iteratively held out all variants (both positive and negative samples) associated with a single protein as the validation dataset from the total set of cross-validation proteins, while variants from the other proteins in the dataset were used for model training, and the trained model predicted the values for the left-out samples. After cycling through all proteins in the dataset, the prediction results of the validation dataset corresponding to each protein were combined, and metrics including AUROC and AUPR were applied to assess predictive performance.

Especially for the IP task, a dataset balancing process was applied before the LOSO CV was initiated. First, before initiating the LOSO CV process (i.e., before traversing samples corresponding to different proteins), we randomly selected a subset of negative samples. This subset, drawn randomly from the total pool of 35,000 mutations (500 for each protein), was twice the size of the positive sample set. These selected negative samples were then combined with the positive samples to form a sub-dataset. Next, the LOSO CV procedure was initiated on this sub-dataset, where we sequentially traverse different proteins, using the mutation samples corresponding to a single protein as the test set and the remaining for training the model. This whole process is repeated 50 times for robustness.

We added representative PS predictors including DeePhase39, PSAP61, PScore35, catGRANULE38, as well as FuzDrop62 for comparison in the IP task. For FuzDrop62, we applied the residue-level PS scores for calculation. Expecting greater score changes for ‘Impact’ mutations than ‘Background’ mutations, we calculated the absolute difference in scores before and after each mutation to represent the predicted values.

We applied the Python Scikit learn (sklearn) package to develop machine learning models. Data was normalized by the MinMaxScaler from sklearn before fitting the models. We trained models including a RandomForestClassifier (with parameters: class_weight = ‘balanced’, and max_depth = 10), an SVR (with default parameters), and a LogisticRregression (with parameters: class_weight = ‘balanced’, kernel = ‘liblinear’, and penalty = ‘l1’) for both the IP and the SP task. We calculated the AUROC with the sklearn roc_auc_score, and roc_curve metric to evaluate the performance. Additionally, the AUPR were calculated with sklearn precision_recall_curve. The discriminative power of models was evaluated using the Mann–Whitney test.

The cross-validation dataset and independent test set were combined to train the final PSMutPred models for both tasks. According to the models’ performance, we produced IP-RF, IP-SVR, and IP-LR scores and SP-RF, SP-LR scores. To prevent the overuse of negative samples in the IP task, for each model, 10 different subsets of samples were randomly sampled from the ‘Background’ mutations with twice the size of the collected ‘Impact’ mutations to train 10 sub-models. The averaged prediction scores of the 10 trained models as well as the ranks for these scores in all ClinVar63,64 variants were used as the final prediction scores for each IP-RF, IP-SVR, and IP-LR model. These models were used to predict PSMutPred scores for each missense variant in ClinVar.

Experimental materials

Plasmid preparation and cell culture

The full-length coding sequences of mouse Eps8 (NM_007945.4) and its mutations were PCR amplified and cloned into the pEGFP-C3 vector. The primers were purchased from biosune, Shanghai (sequences of primers were shown in Supplementary Data 5). These recombinant plasmids were then transiently transfected into HEK293 cells (HEK293T (source: ATCC, CRL-3216; RRID: CVCL_0063)) using the Lipofectamine 3000 Kit (Invitrogen), with 1–2 μg of plasmid used per well in a 12-well plate (Costar, Corning) for each transfection. Before fixation, HEK293 cells were cultured for 16-24 h in DMEM medium (Gibco) supplemented with 10% FBS (Gibco) and 1% Penicillin-Streptomycin Solution (Gibco), under 5% CO2 conditions. Cells were washed 3 times with PBS for 3 min each, then fixed in 4% PFA for 20 min at room temperature. Detailed validation information and references for these products can be found on the respective manufacturers’ websites.

Image analysis and quantification

HEK293 cells were visualized using an IX73 Inverted Fluorescence Microscope (Olympus) equipped with a 60× 1.42 NA Plan Apochromat oil objective. Images were captured with an Iris 9 sCMOS camera (Teledyne Photometrics) and the setup was controlled by the cellSens imaging system (Olympus). After subtracting the background and setting an identical fluorescence threshold, ImageJ software was applied to quantify the number of puncta per cell, with modules including ‘Analysis Particles’, ‘Find Maxima’, and ‘Set Measurement’. The statistical analysis applied a two-tailed Student’s t-test for experimental comparisons of puncta numbers using GraphPad Prism 9. Each assay was performed at least three times.

Protein structure analysis of EPS8

We performed the sequence alignment of EPS8 with the ClustalX tool. The structures of human and mouse EPS8 were predicted using AlphaFold23,104. The interaction among residues was analyzed using the crystal structure of the SH3 domain of EPS8 (PDB: 7TZK). The protein structure analysis was performed using PyMol version 2.5.0.

Between-group analysis

Missense mutation data was collected from ClinVar up to 2022.12, which contains 522,016 variants corresponding to 8611 human proteins. Of these, 39,166 are pathogenic or likely pathogenic, 48,576 are benign or likely benign and the remainder 434,274 are ‘Uncertain Significance’.

After excluding the training data points for PSMutPred, we divided missense variants corresponding to 8611 proteins into a PS-prone group and a low-PS propensity group in two ways: 1. By grouping variants from experiment-verified PS protein to the PS-prone group and the rest to the low-PS-prone group. Specifically, we referenced 155 human PS proteins (59 PS-Self proteins, 96 PS-Part proteins) from PhaSepDB43,47, among which 83 were found in ClinVar63. This resulted in 1451 variants in the PS-prone group, leaving 86,291 variants for the low-PS-prone group. 2. Variants from predicted PS proteins, based on either a PScore35 higher than 4 or a PhaSePred43 rank (PdPS-10fea_rnk) higher than 0.9, were allocated to the PS-prone group. This resulted in 30,889 variants from 1276 proteins in the PS-prone group, leaving 56,853 variants for the low-PS-prone group.

We first performed analysis excluding variants with ‘Uncertain significance’. For each group, the Pearson correlation score was computed using the Python Scipy package by comparing the PSMutPred-IP ranks of its variants against their respective ClinVar pathogenicity labels (pathogenic or likely pathogenic coded as 1 and benign or likely benign coded as 0). The ranks of the absolute score difference of DeePhase39, PSAP61, PScore35, catGRANULE38, as well as FuzDrop62, were also used to compute the Pearson correlation.

The significance of each computed Pearson value is annotated with corresponding P-values, indicating the robustness of the correlation within each group. We compared the coefficient values between the PS-prone group and the low-PS-prone group. We also computed and compared the coefficient values between IDR variants (n = 15,427) and Domain variants (n = 15,462) within the PS-prone group defined by predictions (a variant was classified by Domain variants if identified in a structured domain by PfamScan84,85 and as an IDR variant if it is unmapped).

In addition, we divided the PS-prone group into a neurodegenerative disease (ND) group and a non-ND group. To do that, we collected disease genes related to common neurodegenerative diseases including Alzheimer’s disease, Amyotrophic lateral sclerosis, Frontotemporal dementia, Huntington’s disease, Multiple Sclerosis, and Parkinson’s disease from DisGeNet105. After removing duplicates, we screened genes with a gene-disease association score (GDA score) higher than 0.5, resulting in a subset of 91 genes (Supplementary Data 2), 19 of which were both matched in ClinVar genes and the PS-prone group (363 variants, 252 being predicted ‘Impact’ mutation by PSMutPred-IP, having a combined IP-LR, IP-SVR, and IP-RF rank score sum above 1.5). We compared the coefficient values of PSMutPred-SP between variants corresponding to the ND genes and those were not within the predicted PS-prone group. The ranks of the score difference of various sequence-based PS methods, were also used to compute the Pearson correlation.

We then performed a between-group analysis to test whether proteins with a higher propensity to undergo PS have a higher proportion of variants predicted to impact PS. For each PSMutPred-IP model (IP-RF, IP-LR, IP-SVR), we calculated the proportion of IDR VUSs predicted as ‘Impact’ (PSMutPred-IP rank score > 0.8) for each protein. The proportion values of the PS-prone group and the low-PS-prone group were compared. We compared the PS-prone group (1276 proteins) with the low-PS-prone group (7335 proteins) defined by algorithm prediction and compared the PS-prone group (83 proteins) with the low-PS-prone group (8528 proteins) defined by experimental-verified PS proteins. Sequence-based PS methods were also added to the comparison, and for each method, a variant is predicted to be an ‘Impact’ variant if its corresponding absolute score difference rank is above 0.8.

Analysis of phase separation feature’s performance in refining pathogenicity prediction

Dataset acquisition

We collected ‘EVE_scores_ASM’ scores from EVE31, which include scoring for all possible missense variants corresponding to 3219 disease-associated genes. Among them, 47,870 ClinVar63 variants (by annotations up to 2021.4.) were used to evaluate the EVE model, 30,758 pathogenic or likely pathogenic were marked as positive samples and 17,112 ‘Benign’ or ‘Likely benign’ variants were marked as negative samples. We applied these samples for model cross-validation (cross-validation set). To construct an independent test dataset, we screened variants (ClinVar63 annotations up to 2022.12) that were not used to evaluate the EVE model and removed variants in the same AA position seen in the cross-validation set to avoid data leakage. This resulted in 15,394 variants including 8,709 pathogenic or likely pathogenic variants and 6685 benign or likely benign variants.

We collected ESM1b46 scores, which encompass the scoring of all possible missense variants for 42,286 proteins and their isoforms. In total, scores for 140,321 variants were mapped to ClinVar63. These variants were then used as a cross-validation dataset to evaluate the combination of ESM1b and PS features.

Pathogenicity prediction combining PS-related features

To test whether the integration of PS-related properties can improve the accuracy of pathogenicity prediction, we built a model based on the following features: PSMutPred-IP model scores predicting the propensity of the missense mutation to impact PS (3 dimensions, predicted values of IP-LR, IP-RF, and IP-SVR); PSMutPred-SP model scores predicting the direction of the shift in the normal PS threshold induced by the mutation (2 dimensions, SP-LR, SP-RF); PScore35 representing the PS tendency of the wild-type protein (1 dimension); IUPred59 score of the mutation site representing the probability of the site to locate in IDRs (1 dimension); binary encoding whether the mutation is in a Domain predicted by PfamScan84,85 (1 dimension); the residue distance of the mutation site to the nearest IDR (1 dimension); one-hot encoding for wild-type and mutant AA (40 dimensions); EVE score or ESM1b score of the variant (1 dimension). We trained models including a RandomForestClassifier (with parameters: n_estimators = 200, class_weight = ‘balanced’, and max_depth = 15), an SVR (with default parameters), and a LogisticRregression (with parameters: class_weight = ‘balanced’, kernel = ‘liblinear’, and penalty = ‘l1’). The dataset was standardized using the MinMaxScaler from Scikit Learn before fitting each model. Based on the performance, we selected the RandomForestClassifier as the optimal model.

Model performance benchmarks

To evaluate the models on the cross-validation set, we applied a blocked n-fold cross-validation where variants from the same gene were strictly assigned to the same group. To evaluate the models on the independent test set, the models were first trained on the cross-validation set and then evaluated using the independent test set.

Pathogenicity prediction

The data points were combined to train a final pathogenicity predictor (140,321 variants; PS feat. + ESM1b predictor). Moreover, to transform the continuous pathogenicity scores into ‘Pathogenic’, ‘Likely benign’, and ‘Uncertain’ categories, we determined 0.5 as the initial threshold to distinguish between ‘Likely pathogenic’ and ‘Likely benign’ based on the test set’s F1 score. To define the uncertainty of prediction, we considered applying an offset around this threshold. For instance, with a 0.1 offset, predictions below 0.4 are categorized as ‘Likely benign’, while those above 0.6 are classified as ‘Likely pathogenic’. We observed that starting from the initial 0.5 threshold, the AUROC and Accuracy improved as the offset increased (Supplementary Fig. 4f).

Reporting summary

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