ProtSyntax

Post-translational modifications (PTMs) — phosphorylation, acetylation, ubiquitination, glycosylation and dozens of rarer chemistries — expand what a protein can do by encoding context-dependent regulatory states, and their dysregulation contributes to cancer, neurodegeneration and metabolic disease. Most computational PTM predictors score each candidate residue as an independent binary label, which makes it hard to separate residues that are merely chemically plausible from those that are contextually permissible, to model crosstalk between modifications on the same protein, or to connect a modification event to a functional consequence.

ProtSyntax, developed at Changzhou University, reframes the problem as learning a syntax: the hierarchical rules by which residue chemistry, local motif order, long-range sequence context, three-dimensional accessibility and pre-existing modification state jointly determine whether a residue can be modified and which modification type is compatible. A shared sequence–structure encoder feeds task-specific heads for general PTM-site prediction, kinase-specific phosphorylation, PTM crosstalk and enzyme-kinetic regression, so residue-level recognition and protein-level function are learned together rather than in separate pipelines.

General-purpose protein language models such as ESM-2 learn residue representations from canonical, unmodified sequences, keeping modification state outside the vocabulary; PTM-dCN instead works in the generative direction, designing sequences conditioned on chosen PTM sites.

#Key Features

• Protein-aware positional encoding: Bio-RoPE extends rotary position embeddings with structural periodicity and physicochemical phase shifts, encoding position in terms of protein geometry and residue chemistry rather than token index alone.
• Bidirectional state-space propagation: A Bi-Gated DeltaNet layer propagates context in both N-to-C and C-to-N directions, so a candidate site is evaluated against upstream and downstream sequence at long range.
• Geometry-constrained attention: Geometric gated attention applies learnable three-dimensional microenvironment corrections that down-weight residue pairs which are sequence-compatible but spatially implausible.
• Adaptive multi-objective learning: The PACE-Nash objective combines asymmetric focal classification for extreme class imbalance, contrastive alignment, evidential regression for kinetics, and Nash-style task weighting to limit gradient conflict between heads.
• Syntax probing beyond site calling: The model recovers masked PTM types and positions, rejects structurally incompatible motif decoys, transfers to data-scarce modifications, and reconstructs crosstalk relationships between modification events.

#Technical Details

Training data comprises 348,903 positive and 3,902,925 negative site-centred samples spanning 40 PTM classes, plus 170,290 protein records carrying enzyme-kinetic measurements, together with kinase-specific phosphorylation and PTM crosstalk sets; the benchmark corpus is published as a separate Hugging Face dataset under Apache-2.0. Inputs are amino-acid sequences and residue-centred candidate sites, with optional structural features as C-alpha coordinates or residue-frame transforms. Across the 40 PTM-site benchmarks the authors report mean Matthews correlation coefficient and average precision gains of 12.66% and 10.67% over the strongest baselines, PTM-type recovery accuracy of 0.842, and a mean R² improvement of 8.09% on enzyme-kinetic regression. These figures are self-reported and the preprint is under peer review. The code repository is Apache-2.0 and contains the core model modules, a demo prediction script, a model card and dataset documentation; trained checkpoints are distributed separately and are not tracked in the repository. The preprint itself is posted under CC BY-NC-ND, and no parameter count is published.

#Applications

The model is aimed at researchers studying regulation rather than structure: mapping which residues in a protein of interest are plausibly modified and by what chemistry, attributing phosphosites to specific kinases, testing whether two modifications are likely to condition one another, and estimating how a modification perturbs enzyme kinetics. The authors apply it to pathogenic variants, biomolecular condensates and disease-associated PTM landscapes, where the practical question is which of many annotated or predicted sites deserve experimental follow-up. Because the task heads share one encoder, phosphoproteomics groups can ask several of these questions of a single model rather than assembling separate per-modification predictors.

#Impact

ProtSyntax argues that PTM prediction is better treated as a modelling problem over a structured regulatory language than as 40 independent binary classification tasks, and it packages that argument with a public benchmark corpus, an Apache-2.0 code release and a model card documenting intended use and limitations. Its practical reach is still constrained: checkpoints are not distributed through the repository, and the reported margins have neither been peer reviewed nor confirmed independently. The authors are explicit about the underlying data limits — PTM databases are incomplete and skewed toward well-studied organisms and modification types, static or predicted structures do not capture conformational ensembles or condensate dynamics, and rare PTM classes stay data-limited even with transfer learning. The model card is explicit that predictions are computational evidence for experimental follow-up, not clinical decisions.