Highly dynamic assemblies of glycolytic enzymes by quinary determinants Authors: Blanca Poquet, Galal Yahya, Laia Ferrer, Enrique Marcos, Carme Gallego, Martí Aldea Published: 2026 ABSTRACT AbstractMolecular channelling has long since been proposed as an amenable solution to efficiently drive sequential enzymatic reactions in the cell. Nonetheless, metabolic channelling in central pathways such as glycolysis is still a matter of debate and plausible underlying mechanisms are not known. Here we use correlation spectroscopy and proximity labelling methods to uncover extremely dynamic interactions among enzymes of the glycolysis pathway in yeast and human cells. The resulting transient protein assemblies, tentatively named as protein flocks, form in the cytoplasm with a lifetime shorter than 10 ms. Compared to other metabolic pathways, glycolytic enzymes display a greater similarity in the electrochemical properties of surface atoms, pointing to quinary structural features as molecular determinants of protein flocks. We observe a remarkable negative correlation between the spatiotemporal coincidence by correlation spectroscopy and the Euclidean distance in surface atom configuration within multiple protein pairs, including a collection of surface variants generated from a de novo designed protein framework. By underlying protein flock formation from yeast to human, quinary determinants would provide enzymes with higher efficiency and specificity in glycolysis but, perhaps more important to human cells, would limit intermediate concentrations and prevent their inherent toxic effects.Subject terms: Enzymes, Bioenergetics, Molecular modelling Introduction Our current understanding of cellular physiology is still fragmentary, and it is not surprising that our initial descriptions are simplistic. For instance, the long-standing idea of metabolism as a complex set of independent reactions carried out by enzymes dispersed in a uniform aqueous milieu has been widely assumed when modelling enzymatic reactions and supported by in vitro biochemical experiments. However, accumulated evidence1,2 points to the opposing notion that metabolic pathways would be spatially organised to channel metabolites through sequential enzymes, such that the product of one enzyme is efficiently transferred to the next enzyme as substrate, minimising diffusion and improving flux. The expression metabolic channelling was first introduced3 to refer to amino acids being specifically directed to boost protein synthesis during human nephrosis. But it was not until twenty-five years later that the concept of metabolic channelling was associated with proteins forming dynamic complexes for a functional purpose4. Some years later these assemblies were named metabolons by Paul Srere5 while working on the mitochondrial tricarboxylic acid (TCA) cycle. Proteins constitute different types of assemblies via electrostatic forces, hydrogen bonding, or hydrophobic effects, ranging from stable to transient6,7. In stable protein complexes, interactions among monomers are strong and long-lived, reflecting either necessities for conformation stability or functionality. These interactions generally depend on extensive protein-protein contacts and usually involve an important number of hydrophobic and aromatic residues. On the other hand, transient interactions are temporary, reversible and typically modulated by specific cellular conditions like phosphorylation, conformational changes, or local concentration shifts. Transient interactions range from weak to strong depending on the dissociation constant, in the micromolar or nanomolar scale, respectively7. Whereas strong transient complexes typically change their composition as a result of signalling effects, weak transient complexes exhibit a dynamic diversity of multimeric states and are facilitated by high concentrations of the constituent monomers, which promote interactions despite their weak affinity8. Owing to their intrinsic instability, weak protein-protein interactions have been technically difficult to study and harder to detect than more stable interactions6. However, relatively recent developments based on molecular fluctuation analysis9,10 and proximity labelling11,12 have opened avenues to study these highly dynamic interactions in vivo. Transient assemblies of enzymes have been demonstrated in a few instances13–15 and proposed in a wide variety of metabolic pathways, including glycolysis, oxidative phosphorylation, and fatty acid and amino acid biosynthesis, but also in more complex macromolecular processes such as DNA replication and ribosome biogenesis16,17. Hypoxia induces the formation of biomolecular condensates containing RNA and glycolytic enzymes18, and about 5% of the cell glycolytic activity has been found associated with mitochondria in plants19 and human20,21. However, no evidence of physical interactions between enzymes of glycolysis has been demonstrated in non-stressed cells. In all, whereas molecular channelling is increasingly thought of as a key property of multi-step biological processes, we have very little direct evidence of its existence or functional relevance and, far less, the underlying molecular determinants. In this study, we combine correlation spectroscopy and proximity labelling methods to uncover the highly dynamic assemblies of glycolytic enzymes in yeast and human cells, supporting a functional model of metabolic channelling in glycolysis. Results Glycolytic enzymes display spatiotemporal coincidence in yeast Molecular-fluctuation analysis methods such as fluorescence correlation spectroscopy (FCS) and raster-image correlation spectroscopy (RICS) exploit high-temporal-resolution photon-count data to determine concentration and diffusion properties of fluorescent proteins within the cell22–24. Data collection in FCS is restricted to the focal volume at a given time and thus cannot deliver the spatiotemporal information needed to understand molecular processes that simultaneously occur in larger cell regions. In contrast, RICS provides spatial information at high temporal (ca. 10 μs) resolution and throughout the cell on the number of molecules per moving particle, a parameter that reports on protein complex or assembly status10. Taking advantage of this property, we developed a method based on RICS to determine the spatiotemporal coincidence of protein pairs that physically and functionally interact in a highly dynamic manner such as Hsp70 and Hsp40 chaperones25,26. Briefly, pixels along the raster line produced by laser-scanning confocal microscopy are used to obtain autocorrelation functions and high spatial resolution maps of brightness (B), i.e. the number of fluorescent molecules per moving particle (Fig. 1a). As previously shown25,26, B maps of mEGFP are very heterogeneous and rapidly change with time (Fig. 1b and movie S1), indicating that mEGFP participates in dynamic assemblies that display a high variability in size. These maps can also be obtained from 3-ms time-lapse stacks covering a short linear cytoplasmic region to build a brightness carpet (Fig. 1c) in which most mEGFP assemblies lasted less than three consecutive linear frames, suggesting that their lifespan is shorter than 10 ms. With special relevance to this work, B maps of different proteins can be used to obtain a coincidence score as a Pearson correlation coefficient, i.e. a value that reports on the level of spatial overlap of the brightness of two proteins25. Glycolytic enzymes display spatiotemporal coincidence in yeast. a Raster Image Correlation Spectroscopy (RICS) essentials to derive brightness maps from fluorescence images obtained by laser scanning confocal microscopy. Briefly, photon counts from consecutive pixels in the fluorescence image of a yeast cell expressing a fluorescent protein such as mEGFP are used to obtain an autocorrelation curve. The autocorrelation intersect is inversely proportional to the number of moving particles in the focal volume and, together with the corresponding photon counts, can be used to estimate brightness (B), the number of fluorescent molecules per moving particle (see Methods). b B maps from cell in panel a obtained at the indicated time points. c Carpet corresponding to 300 linear B maps with length 1 μm taken at intervals of 3 ms. A brightness profile of the top pixels is shown. d,e Representative B maps obtained from yeast cells expressing either mNeonGreen and mCherry (d) or Pgi1-mNeonGreen and Cdc19-mCherry (e). The corresponding merged images are shown on the right. Bar is 2 µm. f Coincidence scores from B maps as in (e) of Cdc19-mCherry and mNeonGreen fusions to the indicated glycolytic proteins in downstream order in yeast cells. Individual cell scores (closed grey circles, n = 30) and the corresponding quartiles (boxes) are plotted for each pair of proteins. The p values of Mann-Whitney tests against mNeonGreen and mCherry (Ctrl) are also indicated. As we had previously shown for mEGFP, B maps from mNeonGreen displayed very little overlap with those offered by mCherry (Fig. 1d). However, B maps produced by mNeonGreen and mCherry fusions to two glycolytic enzymes (Pgi1-mNeonGreen and Cdc19-mCherry) were strikingly similar (Fig. 1e), showing a much higher level of spatiotemporal coincidence compared to control cells expressing mNeonGreen and mCherry (Fig. 1f). Cdc19-mCherry also displayed high coincidence scores with mNeonGreen fusions to all other enzymes of glycolysis (Fig. 1f), but not to key enzymatic components of the pyrimidine biosynthesis pathway (Supplementary Fig. 1a) that are also expressed at high levels. We obtained similar results when pairing mCherry fusions of Pgi1 or Fba1, upstream or intermediate components of glycolysis, respectively, to mNeonGreen fusions of all other glycolytic enzymes (Supplementary Fig. 1b, c). The spatiotemporal coincidence of Fba1 with Pgi1 or Cdc19 was unaffected in the presence of carbonyl cyanide m-chlorophenyl hydrazone (CCCP), which inhibits ATP synthesis by disrupting the mitochondrial proton gradient, or after glucose starvation (Supplementary Fig. 1d), indicating that the observed enzymatic assemblies do not depend on the metabolic status of the cell or the glycolytic flux. Finally, fluorescent reporter exchange did not affect the coincidence score of representative glycolytic pairs (Supplementary Fig. 1e) and, as expected, pixel permutation completely abolished the coincidence score of representative glycolytic pairs (Supplementary Fig. 1f). In summary, our RICS analysis suggests that yeast glycolytic enzymes constitute extremely dynamic assemblies that we propose to name as protein flocks. A coarse-grain model predicts spatiotemporal coincidence analysis by RICS We envisaged that a coarse-grain model could give support to the spatiotemporal coincidence data obtained by RICS. First, protein flocks were simulated by single spherical “beads” containing varying molecule numbers of two different fluorescent proteins (P1 and P2) in a cytoplasmic volume (Fig. 2a) at concentrations comparable to those exhibited by glycolytic enzymes in yeast cells (see Methods for details). Second, protein flocks were provided with Brownian movement (movie S2) based on their estimated Stoke’s radius and experimental diffusion coefficients in the eukaryotic cytoplasm27. Third, fluorescence snapshots at different times were generated in a pixel-wise manner by simulating photons emitted by the two proteins from the experimental focal volume with realistic emission rates28. Finally, pixels from consecutive snapshots were used to generate a raster image with a given dwell time (Fig. 2b), which were then used to perform RICS analysis as for experimental images from yeast cells. Whereas fluorescence levels displayed similar pixel-to-pixel variability across the image, the corresponding B maps were very heterogeneous (Fig. 2c), similarly to those obtained with monomeric fluorescent proteins such as mEGFP or mNeonGreen. In the course-grain model, B map heterogeneity depends on the assumption that identical protein monomers interact with each other in flocks. Thus, the mean brightness was a function of the number of molecules per flock (Fig. 2d) and strongly depended on the dwell time reaching a maximum at 10 μs (Fig. 2e), in the usual range of experimental RICS10. Moreover, the coincidence score depended on the number of molecules per flock (Supplementary Fig. 2a) and increased with the ratio of the two fluorescent proteins in flocks (Fig. 2f). Finally, as a control also used in experimental RICS analysis, pixel permutation caused a strong decrease in brightness and a total loss of spatiotemporal coincidence (Supplementary Fig. 2b–d). The coarse-grain model simulated the Brownian movement of an enormous number of protein molecules but, for practical reasons, could not incorporate binding dynamics. Regardless of this limitation, our simulations give support to the intrinsic suitability of RICS as a method to detect the existence of protein flocks in vivo. A coarse-grain model of protein flocks supports coincidence analysis by RICS. a Protein flocks are represented by single spherical “beads” containing varying molecule numbers of two different fluorescent proteins P1 (green) and P2 (red) in a cytoplasmic volume of 1 μm3 in the yeast cell at concentrations comparable to those exhibited by glycolytic enzymes (see Methods). Brownian movement of protein flocks was based on their estimated Stoke’s radius and experimental diffusion coefficients. b Fluorescence snapshots at different times were generated in a pixel-wise manner by simulating photons emitted by the two proteins from the experimental focal volume applying realistic emission rates. To simulate a raster image, pixels from successive snapshots with a dwell time of 10 μs were consecutively assembled. c Representative fluorescence images and the corresponding B maps obtained assuming N1 + N2 = 10 and N1/N2 = 0.7, being N1 and N2 the number of molecules per flock of proteins P1 and P2, respectively. d Brightness of protein P1 as a function of the number of molecules per flock (N1) from simulated images as in panel c. Mean frame values (closed grey circles, n = 20) and the corresponding quartiles (boxes) are plotted. e Brightness of protein P1 as a function of the dwell time (μs). Mean frame values (closed grey circles, n = 20) and the corresponding quartiles (boxes) are plotted. f Coincidence score of proteins P1 and P2 as a function of the ratio (N1/N2) of molecule numbers per flock. Mean frame scores (closed grey circles, n = 20) and the corresponding quartiles (boxes) are plotted. APEX2-proximity analysis of glycolytic enzymes To confirm the spatiotemporal coincidence of glycolytic enzymes we used a proximity labelling approach based on APEX2 (Fig. 3a), an engineered enzyme that catalyses the addition of biotin to proteins within a very small diffusion radius in a native context11,12. To facilitate APEX2 substrate biotin-phenol delivery across the cell wall, yeast cells must be semipermeabilized by a quick freezing step at −80 °C29. Thus, we first tested whether this treatment had any effect on the coincidence score by RICS. Although cell-to-cell variability displayed a modest decrease, the coincidence scores of Cdc19-mCherry and Pgi1-mNeonGreen from semipermeabilized and control cells were not significantly different (Supplementary Fig. 3a). Moreover, addition of H2O2 at the concentration used to trigger biotinylation did not affect the coincidence score of these two glycolytic enzymes either (Supplementary Fig. 3b). A representative gel showing proteins biotinylated by the different APEX2 fusions used in this study (Supplementary Fig. 3c) and a comparative analysis of protein levels in streptavidin-pulldowns from two independent experiments (Supplementary Fig. 3d) are shown. To obtain proximity scores, protein levels in streptavidin-pulldowns from the corresponding APEX2 fusion were made relative to those obtained from cells expressing free cytoplasmic APEX2 as control. Notably, the proximity score of glycolytic enzymes to Pgk1-APEX2 was significantly higher compared to proteins involved in lipid, amino acid or pyrimidine biosynthesis (Fig. 3b). Similar significant differences in proximity were observed with APEX2 fusions to other enzymes of glycolysis such as Pgi1, Fba1, Eno2 and Cdc19, but not with Thr4, a highly-expressed enzyme that catalyses formation of threonine in the cytoplasm (Fig. 3c, d). Finally, we observed a consistent enrichment of glycolysis-related GO terms compared to those related to lipid, amino acid or pyrimidine biosynthesis in datasets produced from APEX2 fusions to glycolytic enzymes but not from Thr4 (Fig. 3e). As a notable exception, among the GO terms corresponding to amino acid anabolism pathways, the glutamine biosynthesis proteins (Fig. 3e, blue squares) also displayed positive enrichment scores by their proximity to all glycolytic enzymes tested. APEX2-proximity analysis of glycolytic enzymes. a Schematic of the experimental approach to analyse protein proximity by biotinylation among glycolytic enzymes in yeast cells. b Proximity scores of Pgk1 to proteins belonging to glycolysis (orange), lipid (magenta), amino acid (blue) or pyrimidine (green) biosynthesis. Individual protein scores (closed grey circles) and the corresponding quartiles (boxes) are plotted. The p values of Mann-Whitney tests for the indicated pairs are also shown. c Proximity scores of the indicated glycolytic enzymes to proteins belonging to glycolysis (orange) or other pathways (blue), namely lipid, amino acid and pyrimidine biosynthesis. Thr4, an enzyme of threonine biosynthesis, was used as control. Individual protein scores (closed grey circles) and the corresponding quartiles (boxes) are plotted. The p values of Mann-Whitney tests for the indicated pairs are also shown. d Heatmap of proximity scores as in panel c of the indicated enzymes in columns to proteins belonging to glycolysis and the 20 most abundant proteins in lipid, amino acid and pyrimidine biosynthesis (other) in rows. e GSEA results by biological process of proximity datasets as in (c). Nominal p-values were obtained as described in Methods. GO terms related to glycolysis (orange), lipid (magenta), amino acid (blue) and pyrimidine (green) biosynthesis are indicated. The glutamine biosynthesis protein set is highlighted (blue squares). Quinary determinants underlie the spatiotemporal coincidence of proteins in yeast and human cells To investigate the molecular determinants responsible for the observed spatiotemporal coincidence and proximity among glycolytic proteins we first estimated the structural similarity between proteins using the template modelling score (TM)30. Although there were significant differences, most enzymes offered low TM values (below 0.5) when compared within glycolysis and/or other metabolic pathways (Supplementary Fig. 4a), indicating a low degree of structural conservation among glycolytic proteins and functionally unrelated enzymes. On the other hand, the degree of intrinsic disorder31 in all these proteins was very low (Supplementary Fig. 4b), ruling out an important contribution of disordered regions in the spatiotemporal coincidence of glycolytic enzymes. We then focused on the specific chemical properties of surface atoms and grouped them by their interaction potential in ionic bridges (negative or positive) and hydrogen bonds (acceptor or donor) (Fig. 4a). The percentage of surface atoms lying in all these categories for glycolytic enzymes was 12.72%, which is very similar to that for enzymes involved in other metabolic pathways (12.75%), and no significant differences were observed in the mean percentage of surface atoms in each group (Fig. 4b). However, we observed a significant decrease in variability in the percentage of hydrogen-bond acceptor and donor surface atoms in enzymes of glycolysis compared to other biosynthetic pathways (Fig. 4b). This reduced variability displayed by glycolytic enzymes was not observed if amino acid groups, and not atom groups, were considered (Supplementary Fig. 4c). Quinary determinants underlie the spatiotemporal coincidence of proteins in yeast and human cells. a Surface representation of Pgi1 depicting atoms by their interaction potential in ionic bridges (negative or positive) and hydrogen bonds (acceptor or donor). b Percentage of surface atoms by their interaction potential in ionic bridges (negative or positive) and hydrogen bonds (acceptor or donor) in enzymes of glycolysis (orange) or other pathways (green), namely lipid, amino acid and pyrimidine biosynthesis. Individual protein percentages (closed grey circles) and the corresponding quartiles (boxes) are plotted. The p values of Mann-Whitney (black) and Levene’s (red) tests for the median rank and variance, respectively, are also shown. c Euclidean distances in surface atom configuration between enzymes in glycolysis (orange) or other pathways (green). Euclidean distances considering all atom groups collectively are plotted on the right. Individual protein pair data (closed grey areas) and quartiles (boxes) are plotted with p values from Mann-Whitney tests. d Surface representation of the indicated fluorescent proteins depicting atoms by their interaction potential in ionic bridges (negative or positive) and hydrogen bonds (acceptor or donor). Euclidean distances in surface atom configuration of the indicated fluorescent protein pairs are indicated. e Coincidence scores from B maps of the indicated fluorescent proteins co-expressed in HEK293T cells. Individual frame data (closed grey circles, n = 200) and quartiles (boxes) are plotted with p values from Mann-Whitney tests. f Coincidence scores as a function of the Euclidean distance in surface atom configuration for protein pairs tested in yeast and human HEK293T cells. Data points from pairs of non-fused fluorescent proteins are highlighted (orange) and a linear regression line with confidence level of 0.99 is plotted. The results of a Spearman rank correlation test are also shown. g Top and side view of DHR10, a de novo designed protein33 used as a scaffold to generate a set of surface variants. Altered amino acids are depicted in orange as a function of the substitution frequency in the whole set of variants. h Euclidean distances in surface atom configuration of all possible pairs from 46 generated DHR10 surface variants. Selected pairs are indicated (blue lines). i Coincidence scores and Euclidean distances in surface atom configuration for selected DHR10 variant pairs fused to mNeonGreen or mCherry and expressed from centromeric vectors in yeast cells. The results of a Spearman rank correlation test are also shown. As a means to analyse the surface properties of these enzymes in a more comprehensive manner, we calculated the Euclidean distance between proteins based on the differences in the percentage of surface atoms in each group, either separately or collectively (see Methods). With the exception of positive surface atoms, Euclidean distances between glycolytic enzymes were significantly shorter when compared to enzymes in amino acid, lipid or pyrimidine biosynthesis (Fig. 4c). These striking differences between glycolytic and other enzymes vanished completely if all atoms were considered in the analysis (Supplementary Fig. 4d), which pointed to a specific role for surface determinants on the spatiotemporal correlation displayed by enzymes of glycolysis. To test whether surface atom configuration underlies spatiotemporal coincidence independently of the enzymatic context, we focused on the recently developed mAzamiRed protein, a red-fluorescent protein derived from mAzamiGreen by replacing 29 amino acids that mainly lie in the interior of the β-barrel fold32. As expected, these two proteins displayed a smaller Euclidean distance by surface atoms compared to mEGFP and mCherry (Fig. 4d). mAzamiRed is a very efficient fluorescent protein in mammalian cells but not in yeast. As the coefficient of diffusion of GFP is similar in the yeast and mammalian cytoplasm27, we assumed similar levels of cytoplasmic crowdedness in both of them. Thus, we performed RICS in the cytoplasm of HEK293T cells. As shown in Fig. 4e, the mAzamiGreen-mAzamiRed pair yielded higher coincidence scores when compared to the mEGFP-mScarlet pair, pointing to surface determinants as key factors for spatiotemporal coincidence. Indeed, a strong negative correlation (rho = −0.806, p = 2.1e-10) was obtained when plotting the coincidence score by RICS as a function of their Euclidean distance by surface atoms (Fig. 4f). The correlation was much less significant (rho = −0.339, p = 0.030) when amino acid groups, and not atom groups, were considered (Supplementary Fig. 4e). The possibility that assembly formation of glycolytic enzymes depends on surface determinants prompted us to take advantage of the de novo designed protein DHR1033 as a completely heterologous system in which to test our hypothesis. First, we altered surface amino acids in DHR10 (Fig. 4g) to produce a collection of protein variants with overall structural predictions comparable to the original DHR10 template. We then chose a sample of these variants (Fig. 4h) covering a range of Euclidean distances similar to that observed in Fig. 4f, and expressed the corresponding mNeonGreen and mCherry fusions in yeast cells to perform RICS analysis. As shown in Fig. 4i, the coincidence score of the DHR10 variants progressively decreased with their Euclidean distance (rho = −0.865, p = 2.2e-16). Human glycolytic enzymes also display quinary similarities and spatiotemporal coincidence At the substructure level, the core of enzymatic proteins has been more conserved during evolution than their surface34,35. Leaving aside catalytic pockets, a larger degree of freedom in chemical groups at the surface would have been important to adapt protein mobility to the complex crowdedness of the intracellular environment through the evolutionary scale. Thus, we analysed the surface atoms of human glycolytic enzymes as above and grouped them by their interaction potential in ionic bridges (negative or positive) and hydrogen bonds (acceptor or donor) (Fig. 5a). The percentage of surface atoms lying in each group was very similar when comparing enzymes involved in glycolysis and other metabolic pathways (Supplementary Fig. 5a). However, as in yeast, we found a significant decrease in variability of acceptor and donor surface atoms in glycolytic enzymes compared to other enzymes (Supplementary Fig. 5a). As aforementioned, we calculated the Euclidean distance in the percentage of surface atoms in each group, either separately or collectively, of every possible enzyme pair. With the exception of negative surface atoms, Euclidean distances among glycolytic enzymes were significantly much shorter when compared to enzymes in amino acid, lipid or pyrimidine biosynthetic pathways (Fig. 5b). Again, if all atoms were considered, Euclidean distances between glycolytic and other enzymes were much shorter and, when all atom types were collectively considered, were not significantly different (Supplementary Fig. 5b). Supporting the notion that spatiotemporal coincidence of glycolytic enzymes is conserved in evolution, Euclidean distances between human and yeast homologues in glycolysis were also much shorter compared to other enzymes (Fig. 5b). We tested this idea using two different experimental approaches. First, we found that human PFKM displayed high coincidence scores in yeast cells with Pgi1 and Fba1, the preceding and subsequent homologous enzymes in glycolysis (Supplementary Fig. 5c). Second, we performed RICS analysis in HEK293T cells expressing glycolytic enzymes from human cDNAs fused to mEGFP or mScarlet (Fig. 5c). As shown in Fig. 5d and Supplementary Fig. 5d, the coincidence score of glycolytic proteins was significantly higher compared to control in diverse experimental configurations. Glycolytic enzymes also display quinary similarities and spatiotemporal coincidence in human cells. a Name correspondences of glycolytic enzymes in yeast and human (left) and surface representation of homologous Fba1 and ALDOA proteins (right) depicting atoms by their interaction potential in ionic bridges (negative or positive) and hydrogen bonds (acceptor or donor). b Euclidean distances in surface atom configuration between human enzymes in glycolysis (orange) or other pathways (green). Euclidean distances considering all atom groups collectively are plotted on the right, either within human (Hs) or comparing yeast and human (Sc vs Hs) enzymes. Individual protein pair data (closed grey areas) and quartiles (boxes) are plotted with p values from Mann-Whitney tests. c Fluorescence images (top) and B maps (bottom) obtained from HEK293T cells expressing the indicated fluorescent proteins. The corresponding merged images are shown on the right. Bar is 5 µm. d Coincidence scores from B maps as in panel c of the indicated proteins co-expressed in HEK293T cells. Individual frame data (closed grey circles, n = 200) and the corresponding quartiles (boxes) are plotted for each pair of proteins. The p value of Mann-Whitney tests against mEGFP and mCherry (Ctrl) are indicated. The fact that spatiotemporal coincidence of glycolytic enzymes was conserved from yeast to human prompted us to test whether proteins formed flocks in vitro and displayed spatiotemporal coincidence depending on surface determinants. Ovalbumin (OVA) and bovine serum albumin (BSA) offered a large Euclidean distance by surface atom configuration. Thus, we obtained these two proteins labelled with Alexa488 or Alexa555 and performed RICS analysis in a buffered solution at a concentration of 4 µM, similar to that estimated for Fba1-mCherry expressed endogenously in yeast cells36. Since Alexa555 and mCherry display similar molecular brightness coefficients (ca. 1.5e4) we expected also similar B maps under identical acquisition parameters. However, although average photon counts per pixel were comparable, brightness values were much lower in the OVA-Alexa555 B maps (Supplementary Fig. 5e), which would be compatible with a marked reduction of flock formation in vitro. Indeed, OVA-Alexa555 exhibited a 4-fold decrease in Gx,y(0) (Supplementary Fig. 5f). Finally, OVA-Alexa555 displayed the same lack of coincidence with OVA-Alexa488 or BSA-Alexa488 when mixed at equal concentrations even in the presence of crowding agents such as Ficoll or PEG (Supplementary Fig. 5g). Quinary determinants and protein abundance demarcate glycolysis in the yeast and human proteomes Although glycolytic enzymes displayed similar quinary determinants, an important fraction of enzymes from other biosynthetic pathways also exhibited short Euclidean distances by surface atom configuration to glycolytic proteins (Supplementary Fig. 6a), making specificity very unlikely in protein flock formation if it only depended on quinary determinants. However, when the binding constant is very small, interacting species require high concentrations to achieve a quantitatively relevant fraction in heteromers8. Since glycolytic enzymes are amongst the most abundant proteins in yeast, we reasoned that intracellular concentration could play a key role in the effectiveness and, hence, the de facto specificity in flock formation. Thus, we measured the Euclidean distance of the yeast proteome to the mean surface atom configuration of glycolytic enzymes. Euclidean distance clearly differentiated glycolytic enzymes from other very abundant proteins such as those belonging to the ribosome and, when considering both surface configuration and abundance, glycolytic proteins became clearly segregated from the yeast proteome, including enzymes involved in lipid, pyrimidine or amino acid biosynthetic pathways (Fig. 6a). As prominent exceptions, many aa-tRNA synthetases and six translation factors (with ATPase or GTPase activity) overlapped with glycolytic enzymes (Fig. 6b and Supplementary Fig. 6b). Accordingly, the coincidence score by RICS between Fba1 and Tef1 or Eft1, two GTPase translation factors that display a short Euclidean distance with glycolytic proteins, was as high as that displayed by Fba1 and Cdc19 (Fig. 6c), both enzymes of glycolysis. Consistent with the fact that non-NTPase translation factors are mostly excluded from the glycolytic patch (Fig. 6b and Supplementary Fig. 6b), the coincidence score of Fba1 and Tif3, a very abundant non-GTPase translation factor, was not significantly different from the mNeonGreen-mCherry control pair (Fig. 6c). Quinary determinants and protein abundance demarcate glycolytic enzymes in the yeast and human proteomes. a Protein abundance and Euclidean distance in surface atom configuration of the yeast proteome to the average of glycolytic enzymes. All yeast proteins (grey), glycolytic enzymes (orange), ribosomal proteins (blue) and enzymes involved in lipid, amino acid and pyrimidine biosynthesis (green) are plotted with the corresponding density curves. b Frequency distribution of the Euclidean distance to glycolysis as in panel a for glycolytic enzymes (orange), ribosomal proteins (blue), translation factors with (red) and without (cyan) GTPase or ATPase activity, and aa-tRNA ligases (magenta). c Coincidence scores from B maps of the indicated mNeonGreen (top line) and mCherry (bottom line) protein fusions in yeast cells. Individual cell scores (closed grey circles, n = 25) and the corresponding quartiles (boxes) are plotted for each pair of proteins. The p values of Mann-Whitney tests against mNeonGreen and mCherry (Ctrl) are also indicated. d Protein abundance and Euclidean distance in surface atom configuration of the human IPSC proteome to the average of glycolytic enzymes as in (a). e Frequency distribution of the Euclidean distance of the human IPSC proteome to the average of glycolytic enzymes as in (b). f Euclidean distance of 2266 yeast vs human homologue proteins (grey) to the corresponding glycolytic proteins. Glycolytic enzymes (orange), ribosomal proteins (blue), translation factors with (red) and without (cyan) GTPase or ATPase activity, and aa-tRNA ligases (magenta) are specifically indicated. A tendency line with confidence limits of 0.95 is also shown. A very similar segregation of glycolytic enzymes by protein abundance and surface atom configuration was observed in the proteome of human IPSCs (Fig. 6d, e and Supplementary Fig. 6c). Suggesting proteome-wide functional implications of the analysed quinary determinants, 2266 yeast and human protein homologue pairs exhibited a strong correlation in their Euclidean distance to the corresponding glycolytic proteins (Fig. 6f). Discussion In summary, our data give support to the existence of highly-dynamic protein assemblies, which we tentatively name as protein flocks, of glycolytic proteins in the cytoplasm of yeast and human cells. Brightness maps have usually been obtained by pixel-centric fluctuation analysis of image stacks taken at 1-10 frames/s37–39, thus lacking the required temporal resolution attained by RICS10,24. On the other hand, circular scanning FCS is based on photon acquisition at a reduced number of focal volumes40 and, hence, is unable to provide brightness maps covering a continuous cellular region. These are the main reasons that explain why the protein flocks described in this work have been considered nebulous. Although our proximity labelling experiments with APEX2 give general support to the findings by RICS, these methods differ in very important aspects. On the one hand, RICS offers an extremely high temporal resolution and can identify formation and dissolution of flocks in the ms scale. As a drawback, our RICS conditions required dwell times as short as 10 μs and a reduced laser power to limit photobleaching during acquisition of at least 20 consecutive frames, but also statistically amenable photon counts per pixel, which would make coincidence analysis unrealistic for proteins below 10000 copies per cell. On the other hand, APEX2-driven labelling reports the average proximity among the tested proteins during a relatively long labelling period, ca. 1 min, which would allow their detection due to their relative accumulation at particular subcellular environments as, for example, the surface of the mitochondria, where at least 5% of all glycolytic activity has been shown to associate19–21. Second, although detected by APEX-MS labelling in the cytoplasm, the relatively low abundance of most of these targets would make flock formation very unlikely as discussed above8. Third, protein biotinylation takes place in a range of 10 nm, at least an order of magnitude smaller than the protein flocks observed by RICS, and requires sterical availability of tyrosines at the surface. Thus, some proteins could appear enriched in the APEX assays not by proximity to glycolytic enzymes, but owing to lower specific labelling efficiencies by the APEX2-NES construct used as control. Finally, although the inherently different accessibility of proteins in the cytoplasm should be corrected by making biotinylated levels relative to those from cells expressing APEX2 control, the strikingly different behaviour of Eno1,2 and Cdc19 as target proteins could be due to a relatively high exposure in the periphery of flocks compared to other glycolytic enzymes. Quinary structure has been proposed as an additional complexity level in protein structure to consider the features at the surface of proteins that are shaped by evolution to the physiological crowdedness of living cells41,42, in which transient interactions among macromolecules are constantly occurring. Our findings fit perfectly in this conceptual view, and point to the quinary structure as the basis of protein flock formation, a mechanism to facilitate spatiotemporal coincidence of enzymes acting in a pathway. It is important to note that coincidence in protein flocks correlated with the proportion of surface atoms in the different chemical groups (negative, positive, acceptor and donor), regardless of their spatial distribution. Accordingly, in all cases tested, we have found that monomers of the same protein species exhibit high coincidence scores if expressed at high levels. Nonetheless, glycolytic enzymes do not exhibit a distinct pattern of surface atoms for transient interactions. Rather, they are characterised by a very low variability around a common pattern that has likely been optimised by the crowdedness of the cytoplasm. Enzymes involved in other metabolic processes display larger differences in quinary determinants and, hence, would not be incorporated into the same protein flocks as is the case for mNeonGreen and mCherry, which form flocks but do not coincide. Moreover, our data point to the notion that quinary determinants would particularly provide glycolytic flocks with specificity in front of other very abundant putative interactors such as ribosomal proteins. Glucose starvation and ATP synthesis inhibition did not affect the spatiotemporal coincidence of representative enzyme in glycolysis, ruling out (1) a relevant contribution of metabolic intermediates in the transient interactions between glycolytic enzymes and (2) an adaptive mechanism to pathway requirements. In any event, the transient but effective proximity of glycolytic enzymes could facilitate faster transitions to molecular condensates such as those observed under hypoxia conditions18. Protein-protein interactions greatly rely on electrostatic surfaces, which adopt complementary charge distributions for efficient binding43. Using a biomolecular fluorescence complementation assay, a recent study has characterised the complete protein–protein interaction map of the E. coli 1-carbon metabolism pathway, revealing conserved structural interfaces able to interact with multiple complementary partners44. Our data do not point to such surface complementarity or conserved motifs among interacting proteins in flocks. Rather, extremely weak interactions in protein flocks could emanate from a few hydrogen bonds, thus favouring stochastic interactions among proteins with similar surface densities of donor and acceptor groups. Our analysis with pure fluorescently labelled proteins argue against flock formation in vitro, supporting the notion that quinary interactions are stable in vivo but unstable in vitro45, and crowding agents cannot efficiently reproduce the effects observed in the living cells46. Based on computer simulations and scaling theory, it has recently been proposed that the complex interplay between binding and molecular dynamics in a crowded environment creates a memory effect47. Indeed, non-random rebinding kinetics could have been leveraged by biological systems to substantiate the highly dynamic protein flocks described here, which may contribute to explain the strikingly high percentage (ca. 60%) of yeast proteins that displayed homomeric interactions in a reversible proximity complementation assay48. The mechanisms underlying glycolytic flock formation would not necessarily exclude other proteins as long as they share similar quinary determinants and are expressed at high levels to facilitate the establishment of very weak interactions8. In this regard, we have observed a sustained enrichment of glutamine biosynthetic enzymes in biotinylated proteins by APEX2 fusions to all glycolytic enzymes tested. Glutamine fuels the tricarboxylic cycle of intermediates used in many anabolic processes, and spatiotemporal coincidence of glycolytic and glutamine biosynthetic enzymes could facilitate direct mechanisms of metabolic coordination between energy-production and biosynthesis pathways in cell growth and maintenance49. In addition, aa-tRNA synthetases and translation factors with NTPase activity are abundant proteins with similar quinary determinants to glycolytic enzymes, which could contribute to fuel protein synthesis with glycolysis-born ATP. The conservation from yeast to human suggests that glycolytic flocks did not only originate to improve pathway efficiency by metabolic channelling, which would be more plausible in yeast, but to reduce intermediate metabolite concentrations to a minimum and prevent their toxicity50–53. This dual role, enhancing pathway efficiency while protecting the cell from harmful metabolites, positions protein flocks as a potential fundamental organising principle of metabolic enzymes across evolution. Methods Strains and plasmids Yeast strains and plasmid constructs used in this study are listed in Supplementary Tables 1 and 2. Methods for chromosomal gene transplacement and PCR-based directed mutagenesis have been described54. Unless stated otherwise, all gene fusions in yeast were expressed at endogenous levels from their respective loci in homozygous diploid cells obtained by mating haploid cells harbouring self-integrating cassettes55. Gene fusions and specific constructs requiring multiple fragments were obtained by one-step recombination in yeast cells as described56. To ensure cytoplasmic localisation in HEK293T cells (ATCC), protein fusions to mEGFP and mCherry were added a nuclear export sequence (NES) and expressed from the strong CMV promoter in plasmid vectors. Growth conditions Yeast cells were grown at 30 °C in YPD for APEX2 labelling or SC medium with 2% glucose for RICS and, when required, galactose was used at 2%. HEK293T cells were cultured in DMEM medium with 10% FBS at 37 °C with 5% CO2 and transfected with Lipofectamine 2000 (Invitrogen, 11668030) following manufacturer’s instructions. Cells were routinely tested for mycoplasma contamination. Raster image correlation spectroscopy (RICS) Spatiotemporal coincidence of proteins was analysed by RICS essentially as described25. All fluorescent proteins used in this study were monomeric. Yeast cells expressing the indicated fluorescent-protein fusions were grown to the exponential phase and attached to 8-well glass-bottom culture slides (Ibidi 80806), previously treated with 0.1 mg/ml concanavalin A for 30 min and washed twice with water. HEK293T cells were grown in 8-well glass-bottom culture slides (Ibidi 80806) and analysed 24 h after transfection with plasmids expressing the indicated fluorescent-protein fusions. Ovalbumin (OVA) and bovine serum albumin (BSA) labelled with Alexa488 or Alexa555 (Molecular Probes) were analysed in PBS at 2 μM and, when indicated, 20% Ficoll 400 or 10% PEG 6000 were added as crowding agents. Images were obtained at room temperature (yeast) or 37 °C (HEK293T cells) in a Zeiss LSM780 confocal microscope with a 63X/1.4NA oil-immersion objective; specifically, we used a 35 nm pixel size and 12.6 μs dwell time during two-channel simultaneous acquisition, tacking a stack of 20 frames in photon-counting detection mode. Coincidence analysis was carried out with CoinRICSJ25. After filtering out the immobile fraction as described9, the autocorrelation function of each pixel was obtained using a 16-pixel range only in the raster direction as follows: where ξ is the pixel distance in the scanning direction, Gx,y(ξ) is the autocorrelation variable, Ix,y is the pixel intensity at x,y, and angle brackets indicate the average over a 16 square region centred at x,y. The intercept obtained by linear regression of the autocorrelation function (no specific model of diffusion assumed) was taken as an approximation of the inverse of the number of moving particles (Nx,y). Then, the fluorescence intensity (Ix,y) at each pixel was used to calculate the brightness (Bx,y) parameter as Bx,y = Ix,y/Nx,y, i.e. Bx,y = Ix,y. Gx,y(0), and the resulting values were assembled into B maps covering the whole image being analysed. The coincidence score of two proteins was simply defined as the Pearson’s correlation coefficient between the corresponding B maps, using their mean values as threshold. Finally, only to facilitate visual interpretation, B maps were applied a Gaussian filter with a radius of 2 pixels. A schematic of the procedure and representative fittings of the autocorrelation function are shown in Supplementary Fig. 7. Coarse-grain model of protein flocks and their analysis by RICS Protein flocks were simulated by single spherical “beads” that contained two different fluorescent proteins (P1 and P2) (Fig. 2a). Simulations were routinely performed in a cytoplasmic volume of 1 μm3 to produce 1 μm2 images (32×32 pixels) at 0.031 μm/pixel. Representative images (128×128 pixels) in Fig. 2c and Supplementary Fig. 2b were obtained by simulating a larger volume of 16 μm3 (128x128x32 pixels). Independently of their association in flocks, protein molecule numbers were set to 10,000/μm3, very similar to those exhibited by glycolytic enzymes in yeast cells (Saccharomyces cerevisiae Genome Database, SGD). The number of molecules (N1,N2) of the two fluorescent proteins in each flock was sampled from a normal distribution with given mean (N = N1 + N2) and ratio (R = N1/N2) values, which were modulated to assess the output of the model (Fig. 2d, f). To provide protein flocks with Brownian movement (movie S2) we applied the Euler scheme, i.e. the current state (Xi), the diffusion coefficient (D) and random Brownian draws (ξ) with mean 0 and variance 1 predict the next state (Xi+1) at discrete time steps (Δt) as follows: D is inversely dependent on the Stoke’s radius and was approximated as a function of the number of protein molecules in the flock as follows: Where D0 is the coefficient of diffusion for monomeric GFP in the cytoplasm of yeast or mammalian cells ( ≈ 25 μm2/s) (BNID 112266)27. Flocks were provided with pure reflection at the simulated volume boundaries. Fluorescence snapshots at different times were generated in a pixel-wise manner at 31 nm/pixel by simulating photons emitted by the two proteins from the experimental focal volume with emission rates of 2 photon/molecule in a dwell time of 10 μs28. Photons were applied a Gaussian point spread function (PSF) estimated from experimental image acquisition settings. Finally, photon counts in pixels from consecutive snapshots taken at the chosen dwell time were used to generate a raster image (Fig. 2b). Thus, routinely simulated raster images containing 32×32 pixels were built from 1024 snapshots obtained every 10 μS. For each parameter set or condition, we analysed 20 consecutive raster images, covering a total time of about 0.2 s. Larger raster images, containing 128×128 pixels, were built from 16,384 snapshots. In this case, 20 consecutive raster images, covering a total time of about 3.2 s, were analysed. Simulated raster images were used to perform RICS analysis following the same procedures described above for experimental raster images from yeast cells. APEX2 biotinylation assays Exponentially growing yeast cells (ca. 109) were collected by centrifugation, washed in cold water, resuspended in 250 μl cold freezing solution (50% glycerol, 0.5% glucose, 20 mM sodium Hepes pH 7.2, 150 mM potassium acetate, 2 mM magnesium acetate) and semipermeabilized by freezing in liquid nitrogen as described29. Once thawed at RT, cells were incubated with 0.5 mM biotin-phenol for 2 min and 1 mM hydrogen peroxide for 1 min. The biotinylation reaction was immediately stopped by adding 10 ml freshly-prepared quenching buffer (5 mM Trolox, 10 mM sodium ascorbate, 10 mM sodium azide in PBS at pH 7.2). Cells were collected by filtration and washed three times with 5 ml quenching buffer on the collection filter. Cells were resuspended in 1.5 ml quenching buffer, collected by centrifugation, suspended in 200 μl lysis buffer (5% SDS, 50 mM Tris-HCl pH 9, 100 mM DTT, 5 mM Trolox, 10 mM sodium ascorbate, 10 mM sodium azide and protease inhibitors), and incubated at 55 °C for 30 min. Then, cell extracts were added 400 μl RIPA buffer (50 mM Tris-HCl pH 8, 150 mM sodium chloride, 0.1% SDS, 0.5% sodium deoxycholate, 1% NP-40, 5 mM Trolox, 10 mM sodium ascorbate, 10 mM sodium azide) and centrifuged at 16000 g for 10 min at RT. Supernatants were added one volume of 55 % trichloroacetic acid and incubated on ice for 15 min. Proteins were precipitated at 16000 g at 4 °C for 10 min. Pellets were washed with cold acetone three times, air dried and fully resuspended in 200 μl cell solubilisation buffer (8 M urea, 100 mM sodium phosphate pH 8, 1% SDS, 100 mM ammonium bicarbonate and 10 mM TCEP). The resuspended protein solutions were centrifuged at 16000 g at RT for 10 min, and supernatants were transferred to new microcentrifuge tubes. Biotinylated proteins were captured onto streptavidin magnetic beads (Fisher 10150874, 75 µl per sample) previously washed three times in 4 M urea and once in 0.5% SDS. Tubes were rotated overnight at 4 °C and beads were separated by using a magnetic stand, washed once with 4 M urea, once with 0.5% SDS and three times with 50 mM Tris-HCl pH 8. The beads were frozen at −80 °C prior to on-bead digestion for MS analysis. Immunoblotting Yeast protein extracts were analysed by SDS-PAGE, transferred to nitrocellulose membranes and probed with primary and secondary antibodies essentially as described57. APEX2-tagged proteins were detected with a monoclonal αFLAG antibody (clone M2, Merck-Sigma F3165, 1:500). Biotinylation was monitored using streptavidin-HRP (Fisher N100, 1:1000). APEX2 proteomic analysis Proteins were subject to on-bead trypsinisation and the resulting peptides were separated by reversed-phase chromatography for data-independent acquisition quantitative MS analysis in an Orbitrap Fusion LUMOS (BGI Genomics Co.). MaxQuant (version 1.6.3.3) software was used for database search and label-free quantification of mass spectrometry raw files. The search was performed against a FASTA file of the Saccharomyces cerevisiae proteome, extracted from uniprot.org. Matched proteins were filtered for possible contaminants, and relative iBAQ data were obtained. Overall, 2,638 proteins were detected, 938 of them specifically identified and quantified in the APEX2-NES control sample for downstream analysis. For every protein detected, a proximity score was calculated as the log2 ratio of the relative iBAQ in streptavidin pulldowns from the corresponding APEX2-tagged sample to the relative iBAQ from the APEX2-NES control sample. Averaged data from two independent experiments (Supplementary Fig. 3d) are in Supplementary Data 1. Euclidean distance by surface atom configuration Atoms with a high protrusion score (CX > 0.5)58 were selected from the corresponding PDB files and grouped as negative, positive, acceptor or donor depending on their interaction potential in ionic bridges and hydrogen bonds. The Euclidean distance (EdA,B) for a given pair of proteins A and B was determined as follows: Where pA,i and pB,i are the percentages of surface atoms in group i for proteins A and B, respectively. When only one electrochemical group was considered, the Euclidean distance was calculated as the absolute difference between the corresponding percentages in the two proteins. Computational design of surface variants on a de novo-designed protein framework We selected the de novo–designed helical repeat protein DHR1033 as the structural framework for generating surface variants spanning a broad range of Euclidean distances. DHR10 is a 196–amino acid, hyperstable helical repeat protein with a high-resolution X-ray crystal structure available (1.2 Å, PDB ID: 5CWG). Based on this structure, surface variants were generated through Rosetta59 combinatorial sequence design calculations enabling polar amino acids on surface positions with helix secondary structure and a polar amino acid. Amino acid composition on the surface was further diversified using the aa composition Rosetta scoring term to penalise deviations from specified target compositions. Briefly, independent design calculations were carried out using RosettaScripts60, each applying a bias encouraging one of three defined amino acid groups—(1) Asn/Gln, (2) Ser/Thr, or (3) Arg/Lys/Asp/Glu—at target fractional abundances ranging from 0.0 to 1.0 in increments of 0.2. This strategy enabled systematic sampling of surface compositions across a wide range of Euclidean distances. Designs with normalised Rosetta total scores (per residue) below −3.6 Rosetta energy units, as evaluated using the Ref2015 energy function61, were subjected to AlphaFold262 single-sequence structure prediction via our local installation of ColabFold63. From these, 46 designs with AlphaFold2-predicted models closely matching the designed structure (Cα RMSD < 1.3 Å) were retained. From this subset, seven designs were chosen (Supplementary Table 3), yielding 16 variant pairs (Supplementary Table 4) spanning a broad Euclidean distance range that were selected for RICS analysis by expressing the corresponding mNeonGreen and mCherry fusions to their C terminus in yeast cells from the TEF1 promoter in centromeric plasmids. Statistical and reproducibility Pairwise comparisons of the median rank and variance were performed with two-sided Mann-Whitney and Levene’s tests, respectively, and the resulting p-values are shown in the corresponding figure panels. The Spearman rank correlation test was used for non-parametric regression analysis. Gene set enrichment analysis was as described64. No statistical method was used to predetermine sample size and no data were excluded from the analyses. Biological replicas were subject to RICS and proteomics analysis. Pixel randomisation to assess coincidence scores in RICS analysis was done by permutation. Sample details were excluded during RICS and proteomics analysis procedures. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Dataset 1 Movie S1 Movie S2 Reporting Summary Transparent Peer Review file Source data Source Data Acknowledgements We thank L.L. and K.A. for technical assistance and C.R. for editing the manuscript. We also thank J. Bravo for stimulating comments and suggestions. Author contributions B.P., G.Y. and L.F. performed experiments. G.Y., C.G. and M.A. analysed the data. E.M. designed protein variants. M.A. wrote the manuscript with input from all of the authors. Peer review Peer review information Nature Communications thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available. Funding This work was funded by a grant from the Ministry of Science and Innovation of Spain and the European Union (FEDER) (PID2022-141460NB-I00) to M.A. Data availability The mass spectrometry proteomics data have been deposited in the ProteomeXchange Consortium through the PRIDE partner repository under accession code PXD065061 [www.ebi.ac.uk/pride/archive/projects/PXD06061]. All strains and plasmids used are available upon request. Source data are provided with this paper. Code availability The coarse-grain model is available on Github [https://github.com/mambmc/Protein_flock_simulations] Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Galal Yahya, Email: gmebmc@ibmb.csic.es. Martí Aldea, Email: marti.aldea@ibmb.csic.es. Supplementary information The online version contains supplementary material available at 10.1038/s41467-026-73795-6. References Associated Data Supplementary Materials Supplementary Information Description of Additional Supplementary Files Supplementary Dataset 1 Movie S1 Movie S2 Reporting Summary Transparent Peer Review file Source Data Data Availability Statement The mass spectrometry proteomics data have been deposited in the ProteomeXchange Consortium through the PRIDE partner repository under accession code PXD065061 [www.ebi.ac.uk/pride/archive/projects/PXD06061]. All strains and plasmids used are available upon request. Source data are provided with this paper. The coarse-grain model is available on Github [https://github.com/mambmc/Protein_flock_simulations]