Cells control their proteins not only by turning them on and off but by attaching small chemical tags — phosphate, methyl, acetyl groups and dozens of others — to specific amino acids after translation. These post-translational modifications (PTMs) do not act in isolation: a lysine residue can carry only one tag at a time, so acetylation and ubiquitination compete for the same spot, and one tag can recruit or repel the enzymes that write another. This PTM crosstalk is almost always read out biochemically — as binding affinity, degradation rate, or transcriptional output. A parallel literature shows that many tagged, disordered proteins spontaneously cluster into droplets — biomolecular condensates — whose interiors are viscoelastic, halfway between a liquid and a solid, and that the mechanical regime of these droplets matters: in decellularized kidney matrix, stiffness alone flips endothelial cells between quiescence and proliferation (Decellularized Human Kidney Cortex Hydrogels Enhance Kidney Microvascular Endothelial Cell Maturation and Quiescence). What nobody has connected is whether crosstalk between PTMs is also a mechanical control layer: whether the combinatorial tag pattern on a disordered protein decides whether the assembly it forms flows like a fluid or resists shear like a solid.
The core obstacle is that the two halves of this question are measured by communities that do not share instruments. Condensate mechanics has been characterized only in simulation, on unmodified protein chains: coarse-grained dynamics of FUS and DDX4 disordered regions shows interiors with sub-linear diffusion ($\Delta^{0.85}$) and fractional Brownian motion, explicitly excluding methylation, acetylation, and phosphorylation (Diffusion of intrinsically disordered proteins within viscoelastic membraneless droplets). Meanwhile, co-immunoprecipitation — pulling a chosen "bait" protein out of a cell lysate with an antibody so that its physical partners and their PTM states can be catalogued by mass spectrometry — has been shown to co-capture complexes and their PTM marks in one experiment (17 spliceosome proteins plus site-specific histone acetylation in the H7N9 nucleoprotein interactome), but only as a static catalogue (Proteomics Analysis of Cellular Proteins Co-Immunoprecipitated with Nucleoprotein of Influenza A Virus (H7N9)). The proposal here is to close both gaps at once: use Co-IP-MS to determine which combinatorial PTM states actually assemble in cells, reconstitute those states on semisynthetic disordered proteins, and measure their condensates with oscillatory shear rheometry — the technique that applies a small sinusoidal deformation and reads out the storage modulus $G'$ (elastic stiffness), the loss modulus $G''$ (viscous dissipation), and their ratio $\tan \delta$.
Core Claim
PTM crosstalk on intrinsically disordered proteins functions as a mechanical rheostat: the combinatorial tag pattern — not any single tag — determines whether the condensate a protein nucleates sits in a fluid regime ($G'' > G'$), a solid regime ($G' \gg G''$), or at the gel point where they cross. This claim is testable because the two marks on one chain need not add: define the mechanical crosstalk coefficient
$$ C \;=\; \log \frac{G'_{AB}}{G'_0} \;-\; \log \frac{G'_{A}}{G'_0} \;-\; \log \frac{G'_{B}}{G'_0}, $$
where $G'_0$ is the storage modulus of the unmodified condensate and $G'_A$, $G'_B$, $G'_{AB}$ are the moduli with mark A alone, mark B alone, and both marks together. If PTM crosstalk has a mechanical signature, $C$ is significantly nonzero — the pair is more (or less) stiffening than the sum of its parts — and the sign of $C$ should track whether the marks compete for the same lysine or recruit a common multivalent reader. If $C \approx 0$ everywhere, the biochemical crosstalk literature has no mechanical counterpart, and that negative result would itself sharply constrain how condensate mechanics can be regulated.
This differs from the closest existing work in both directions. Sun et al. catalogue PTM marks on Co-IP-purified complexes but never test crosstalk or measure a single mechanical quantity (Proteomics Analysis of Cellular Proteins Co-Immunoprecipitated with Nucleoprotein of Influenza A Virus (H7N9)); we use their Co-IP-MS design as the cellular readout, not the endpoint. Watanabe et al. establish that condensate interiors are viscoelastic but on unmodified chains, in silico, with a 20 µs simulation standing in for roughly 20 ms of physical time — far too short to see any enzyme-driven tag turnover (Diffusion of intrinsically disordered proteins within viscoelastic membraneless droplets); we bring PTM states and experimental rheology to exactly the system they modeled. And the effective-medium theory of disordered networks near the onset of rigidity derives universal scaling — correlation-length exponent $\nu = 1$ for jamming, $\nu = 1/2$ for rigidity percolation, with a frequency crossover from elastic to dissipative response — but only for quenched networks whose bonds never re-form (Universal scaling for disordered viscoelastic matter near the onset of rigidity); PTM-remodeled condensates are annealed networks, and measuring whether their exponents match the quenched predictions is a direct, unresolved test of that theory's own stated limitation.
Research Program
The program has three coupled components built around one model system, the FUS low-complexity domain (FUS-LCD), chosen because it is the substrate of the only existing condensate-viscoelasticity study in the file (Diffusion of intrinsically disordered proteins within viscoelastic membraneless droplets) and because its lysines fall squarely inside the competitive-lysine problem that the strongest PTM predictors address.
First, computational triage. We run MethylSight 2.0 — a model that couples protein language model embeddings to multitask learning across competing lysine modifications (methylation versus acetylation, ubiquitination, and sumoylation) — over FUS-LCD and DDX4-DR sequences to rank which lysines are predicted sites of multi-class competition. This tool reached an AUPRC of 0.672 on its lysine-methylation benchmark, more than doubling the previous best of 0.267, and 68 of its 100 predicted novel sites were confirmed by targeted mass spectrometry in HeLa cells (Leveraging learned representations and multitask learning for lysine methylation site discovery), so its site calls are a credible shortlist. We deliberately do not improve its architecture; we consume it. Its own authors note it is cell-type agnostic, which is precisely the gap our cellular experiments fill.
Second, defined-state rheology. We produce semisynthetic FUS-LCD variants — unmodified, single-mark (one acetylated or methylated lysine), and combinatorial (two marks on the same chain: competing pairs sharing a lysine are rendered as mutually exclusive variants, and hierarchical pairs as both-marked chains). Semisynthetic installation of defined PTM patterns is established practice for disordered proteins: the α-synuclein review assembles its N-terminal acetylation, Ser129 phosphorylation, Y39 nitration, and O-GlcNAcylation findings from such constructs (Modulation of the Interactions Between α-Synuclein and Lipid Membranes by Post-translational Modifications). Each variant is concentrated to form bulk condensates, and we measure $G'$, $G''$, $\tan \delta$, the gel point, and the full frequency sweep by parallel-plate oscillatory rheometry — the same instrument class used to compare decellularized kidney matrix against collagen gels (Decellularized Human Kidney Cortex Hydrogels Enhance Kidney Microvascular Endothelial Cell Maturation and Quiescence) and to track gelation of co-assembled small-molecule hydrogels, where a binary gel reached a yield strain 7-fold higher than its single-component parent at a lower critical gel concentration of 0.9% versus 1.4% (The Inhibitory Impact of a Co-Assembly Gel with Natural Carrier-Free Binary Small Molecules, as Used in Traditional Chinese Medicine, on the Viability of SW1990 Cells). That binary-gel result is the closest mechanical analogue to what we expect from PTM co-assembly, and its metrics — yield strain, critical concentration — become ours. Where a multivalent reader exists for a mark pair, we add it at fixed stoichiometry; the Spindlin1 precedent shows bivalent readout can bind 1 to 3 orders of magnitude tighter than single-mark reading ($K_d = 16$ nM for the doubly methylated H3 tail), so reader-mediated crosslinks are the likeliest route to a large positive $C$ (Molecular basis for histone H3 "K4me3-K9me3/2" methylation pattern readout by Spindlin1).
Third, the cellular anchor. Co-IP-MS of endogenous FUS from cells in which we shift the lysine-acetylation/methylation balance (writer- and eraser-directed perturbations, following the probe logic of Chemical biology research on targeting protein post-translational modifications) tells us which combinatorial states exist in cells and which partners co-assemble with them, with occupancy quantified by parallel-reaction-monitoring mass spectrometry of the eluted complexes — the validation assay used by MethylSight 2.0 itself. The design follows Sun et al.'s demonstration that Co-IP-MS captures interactors and co-purified PTM marks simultaneously (Proteomics Analysis of Cellular Proteins Co-Immunoprecipitated with Nucleoprotein of Influenza A Virus (H7N9)). The loop closes when the cellular PTM-state census is compared against the rheology table: the states cells actually produce should be the ones that sit near a mechanically consequential regime boundary.
Proposed Experiments
The decisive experiment is the crosstalk matrix. For three to five top-ranked competitive lysine pairs on FUS-LCD, we measure $G'$, $G''$, $\tan \delta$, critical condensate/gel concentration, and the crossover frequency $\omega^*$ (where $G'$ and $G''$ cross) for five states: unmodified, A only, B only, AB, and PTM-dead mutants (lysine-to-arginine for deacetylated/demethylated mimics, since no faithful "dead" mimic exists for lysine acylation). Success is a $C$ value significantly different from zero — judged against propagation of the replicate-to-replicate spread in $\log G'$ — for at least two independent site pairs, with the sign of $C$ explained by the crosstalk class (shared-lysine competition versus reader-recruited hierarchy). A secondary endpoint tests the theory: the measured frequency crossover and concentration scaling of the PTM-remodeled condensates are fit against the quenched-disorder predictions ($\nu = 1/2$ rigidity percolation, $\nu = 1$ jamming) from Universal scaling for disordered viscoelastic matter near the onset of rigidity; deviation of the exponents, or a relaxation time set by enzyme turnover rather than bond stiffness, would be the first experimental evidence that annealed PTM networks break the quenched universality classes. Third, the Co-IP-MS arm reports, per perturbation, the occupancy of each combinatorial state and the co-purified partner list, benchmarked against the specificity controls standard in the field (isotype IgG pulldowns, as in the H7N9 study). Throughout, the FUS-LCD simulation dataset provides the in silico reference point — subdiffusion exponent $\alpha \approx 0.85$, interior-to-interface first-passage scaling of $t^{-1.5}$ (Diffusion of intrinsically disordered proteins within viscoelastic membraneless droplets) — against which we check whether PTM states move the measured micromechanics away from the unmodified-chain prediction. No reported result found on Lacuna for an experimental $G'$ or $G''$ of FUS-LCD condensates, so establishing that first number is itself a contribution.
Baselines
The do-nothing reference is the unmodified FUS-LCD condensate, the state the entire existing literature implicitly assumes: its only mechanical characterization is the simulation of Diffusion of intrinsically disordered proteins within viscoelastic membraneless droplets, which reports a subdiffusion exponent of approximately 0.85 and no experimental modulus. We will measure this state experimentally ourselves, with the same rheometer, concentration series, and temperature as all PTM variants, so every comparison is internal to one instrument and protocol. Beating this baseline simply means showing that any PTM state moves $G'$, the gel point, or $\omega^*$ measurably beyond its replicate spread — if no state does, the mechanical-rheostat claim is dead on arrival.
The strongest existing method on the prediction side is MethylSight 2.0, which reported an AUPRC of 0.672 against a prior best of 0.267 and 68 of 100 predicted sites confirmed by PRM mass spectrometry (Leveraging learned representations and multitask learning for lysine methylation site discovery). It is the strongest competitor because it already solves the "which lysines compete" half of the problem better than anything else in the file, including MeToken, whose uniform micro-environment tokenization reached an F1 of 50.40% versus 43.94% for fine-tuned ESM-2 but which predicts PTM classes from static structures rather than modeling lysine-level competition (MeToken: Uniform Micro-environment Token Boosts Post-Translational Modification Prediction). We run MethylSight 2.0 exactly as published, with its released weights and no extra tuning budget, over our sequences; our pipeline gets the same access, nothing more. We do not attempt to beat its AUPRC — the fair comparison is downstream: success means that its top-ranked competitive sites are the ones that carry the largest nonzero $C$ in the rheology matrix and the highest combinatorial occupancy in the cellular Co-IP-MS census, that is, that its chemically capable sites prove mechanically consequential, which it explicitly cannot test because it is cell-type agnostic.
The strongest biophysical competitor for measuring what PTM states do to assemblies is isothermal titration calorimetry of reader–mark binding, the method that established Spindlin1's 16 nM bivalent affinity (Molecular basis for histone H3 "K4me3-K9me3/2" methylation pattern readout by Spindlin1). Where a reader exists for our mark pairs, we run ITC on the identical semisynthetic constructs under identical buffer conditions, with matched protein amounts. The rheology beats this baseline only if network-level mechanics predict something pair-affinity does not — concretely, if variants with near-identical $K_d$ differ in gel point or $G'$ by more than experimental spread, or if $G'$ changes non-monotonically across the affinity series. If $K_d$ and $G'$ track each other perfectly, the mechanical measurement is redundant for that pair and we will say so.
The key ablation isolating the proposal's ingredient is the single-mark column of the matrix: the same protein, the same instrument, one mark instead of two. The crosstalk coefficient $C$ is defined so that this ablation is exactly what it subtracts; a skeptic who believes marks act independently predicts $C = 0$, and the single-mark measurements are what make that prediction falsifiable rather than rhetorical.
Why not just…?
Why not just use ITC, NMR, circular dichroism, or FRAP instead of building a rheology pipeline? These are genuinely simpler, and for single-chain questions they suffice — Spindlin1's bivalent readout was fully resolved by ITC and crystallography (Molecular basis for histone H3 "K4me3-K9me3/2" methylation pattern readout by Spindlin1). But the quantities this proposal targets are emergent at the network level, and the file is explicit that bulk biophysics misses them: the condensate simulations note that FRAP-style bulk averages miss internal viscoelasticity, transient subdiffusion, and the mechanical split between droplet interior and interface (Diffusion of intrinsically disordered proteins within viscoelastic membraneless droplets); the effective-medium theory shows the floppy-to-rigid transition is intrinsically a frequency-dependent modulus question, not a single-molecule thermodynamic one (Universal scaling for disordered viscoelastic matter near the onset of rigidity); and the kidney-matrix study shows mechanical regime, not composition or affinity, is what flips cell phenotype (Decellularized Human Kidney Cortex Hydrogels Enhance Kidney Microvascular Endothelial Cell Maturation and Quiescence). Still, the honest test is built in: the ITC-versus-rheology comparison above is exactly the experiment that would reveal the simpler assay to be sufficient, and we have committed to reporting it if that is what it shows.
Why not just use proximity ligation assays or split-fluorescent reporters in intact cells instead of Co-IP, avoiding lysis artifacts? This is a fair criticism of Co-IP's known weaknesses — lysis mixes compartments and washes strip weak interactions — and we adopt in situ PLA for one purpose: confirming that the key FUS partner interactions we infer from Co-IP are genuinely proximal in intact cells, as the Tax–Scribble work did with puncta visualization (HPF-MPI: A Programming System Supporting Task and Data Parallelism). But proximity assays yield only optical puncta: they cannot harvest material, cannot quantify stoichiometry, and cannot be coupled to mass spectrometry for PTM mapping, whereas the H7N9 study proved Co-IP-MS delivers the partner list and site-specific PTM marks in a single experiment (Proteomics Analysis of Cellular Proteins Co-Immunoprecipitated with Nucleoprotein of Influenza A Virus (H7N9)). Our claim structure also limits exposure: we use Co-IP-MS as a composition-and-occupancy census, never as an affinity measurement, so the class of artifact it is most vulnerable to is the class of claim we avoid making.
Why not just predict everything in silico with protein language models and graph networks? Because every error bar in the file points the other way. MethylSight 2.0 is cell-type agnostic by its authors' own statement; MeToken's 50.40% F1 means roughly half its predictions are wrong, and it depends on static AlphaFold structures that omit disordered dynamics (Leveraging learned representations and multitask learning for lysine methylation site discovery; MeToken: Uniform Micro-environment Token Boosts Post-Translational Modification Prediction); the phase-separation graph model reaches AUC 0.92 but encodes no PTMs and predicts no mechanics (A Variational Graph Partitioning Approach to Modeling Protein Liquid-liquid Phase Separation); and PepBenchmark shows computational scores can be illusions of data leakage, with near-duplicate removal dropping performance by up to 17% and simple random forests beating graph networks (ROC-AUC 79.2% versus 69.5%) (PepBenchmark: A Standardized Benchmark for Peptide Machine Learning). Above all, the simulation speed gap — 20 µs of molecular dynamics standing for roughly 20 ms of physical time (Diffusion of intrinsically disordered proteins within viscoelastic membraneless droplets) — means no current simulation can reach the timescale on which enzymes rewrite PTM states. Our design already uses the in silico tools for the one job they are good at, shortlisting sites, and the cancer-mutation study whose authors explicitly call for exactly this wet-lab validation of computational PTM predictions is precedent that the community regards the remaining step as undone (Frequent mutations in acetylation and ubiquitination sites suggest novel driver mechanisms of cancer).
Decisive Checks
Three results would settle the proposal. First, a crosstalk coefficient $C$ significantly nonzero for at least two independent lysine pairs on FUS-LCD, with sign matching the crosstalk class — this establishes the core claim. Second, cellular Co-IP-MS occupancy showing that cells preferentially populate combinatorial states near the measured mechanical regime boundary rather than distributing uniformly across states — this establishes biological relevance, the gap left open by every cataloguing study in the file. Third, a measured frequency crossover or concentration exponent that disagrees with the quenched-disorder predictions of Universal scaling for disordered viscoelastic matter near the onset of rigidity — this would be the first experimental handle on annealed, enzyme-remodeled networks, a problem that theory has flagged and no experiment in the file has touched. Conversely, if $C$ is indistinguishable from zero across all pairs and all exponents match the quenched classes, the mechanical-rheostat hypothesis fails cleanly, and the field learns that PTM crosstalk, however rich biochemically, is mechanically additive — a constraint that would redirect condensate-mechanics research toward concentration and valence alone.
References
- Decellularized Human Kidney Cortex Hydrogels Enhance Kidney Microvascular Endothelial Cell Maturation and Quiescence
- Diffusion of intrinsically disordered proteins within viscoelastic membraneless droplets
- Proteomics Analysis of Cellular Proteins Co-Immunoprecipitated with Nucleoprotein of Influenza A Virus (H7N9)
- Universal scaling for disordered viscoelastic matter near the onset of rigidity
- Leveraging learned representations and multitask learning for lysine methylation site discovery
- Modulation of the Interactions Between α-Synuclein and Lipid Membranes by Post-translational Modifications
- The Inhibitory Impact of a Co-Assembly Gel with Natural Carrier-Free Binary Small Molecules, as Used in Traditional Chinese Medicine, on the Viability of SW1990 Cells
- Molecular basis for histone H3 “K4me3-K9me3/2” methylation pattern readout by Spindlin1
- Chemical biology research on targeting protein post-translational modifications
- MeToken: Uniform Micro-environment Token Boosts Post-Translational Modification Prediction
- HPF-MPI: A Programming System Supporting Task and Data Parallelism
- A variational graph-partitioning approach to modeling protein liquid-liquid phase separation
- PepBenchmark: A Standardized Benchmark for Peptide Machine Learning
- Frequent mutations in acetylation and ubiquitination sites suggest novel driver mechanisms of cancer