Loading…
Post-translational modification prediction from sequence and 3D structure, quantizing each residue's micro-environment into a per-PTM discrete token.
No providers recorded yet. Browse all providers
N-glycosylation announces itself with a clean Asn-X-Ser/Thr motif that a sliding sequence window can spot reliably. Phosphorylation does not. The kinase reads a surface pocket assembled from residues that may be far apart in the primary sequence, so a predictor restricted to a linear window is working from half the evidence. MeToken's response is to describe every residue by its micro-environment — the neighbors within a sequence cutoff, the residues whose backbone atoms fall inside a spatial radius, and its K nearest neighbors in 3D — and to compress that entire local picture into one discrete token.
The vocabulary those tokens are drawn from is the part that does the work. Modification frequencies are severely long-tailed: phosphorylation and glycosylation account for most annotated sites, while dozens of modifications appear a handful of times each. In a single shared codebook the common classes absorb the capacity and rare ones never get a distinct representation. MeToken instead allocates one equal-size sub-codebook per modification type, so a rare PTM holds exactly as many codes as phosphorylation does, and trains the codebook with a uniform loss that tightens tokens of the same modification while separating tokens of different ones.
MeToken comes from the AI Lab at Westlake University, with collaborators at Zhejiang University and Xi'an Jiaotong University, and was presented at ICLR 2025. Alongside the model the authors compiled the sequence-structure PTM corpus it trains on, which had no equivalent at this scale.
The encoder is a ten-layer message-passing network over a residue graph with 128-dimensional node and edge features built from backbone dihedrals, radial-basis distances, direction vectors and inter-residue orientation quaternions — no language model is involved. Its output is quantized against a codebook of 26 sub-codebooks of 128 codes each, and a three-layer PiGNN head maps the quantized vectors to a modification class. Training data comes from dbPTM annotations joined to Protein Data Bank entries and AlphaFold DB predictions: 1,263,935 annotated sites across 187,812 proteins spanning 72 modification types, split by MMseqs2 clusters at 40% sequence identity. Modifications with fewer than 100 examples are folded into one "rare sites" bucket, leaving 24 distinct PTM classes plus rare sites plus an unmodified class.
On that corpus MeToken reaches 50.40 macro F1, 89.91% accuracy, 76.03 MCC, 92.20 AUROC and 51.26 AUPRC, against 43.94 F1 for a fine-tuned ESM-2 and 16.63 for GearNet adapted to the task. The same checkpoint, applied without retraining, scores 42.07 F1 on PTMint and 61.95 F1 on qPTM, the latter filtered to under 40% identity against the training set. Ablations isolate where the gain comes from: micro-environment encoding without tokenization reaches 40.04 F1, adding the uniform sub-codebook lifts it to 48.04, and temperature-scaled quantization over an unconstrained codebook drops to 37.46.
The practical use is triage. Given a structure — experimental or predicted — MeToken ranks which residues are plausible modification sites and names the chemistry involved, which is what turns a proteome-scale structure set into a shortlist for targeted mass spectrometry. Taking a plain PDB file and a list of residue indices, it slots into annotation pipelines for predicted structures, into interpretation of variants that create or destroy modification sites, and into studies of modifications too sparsely annotated to train a dedicated predictor.
MeToken argues concretely that residue-level PTM prediction is a structural problem, not only a motif problem, and its discrete codebook yields a representation that can be inspected: embedding visualizations separate C-glycosylation from N- and O-glycosylation, and show phosphorylation and acetylation sites interleaved, matching their competition for the same residues. The accompanying corpus is arguably the more durable contribution, since any future method can train against it. The authors are candid about the limits: the structures are predominantly AlphaFold predictions and inherit that model's biases, and the micro-environment excludes the enzymes, ligands and binding partners that determine modification in vivo. Distribution is modest — the checkpoint is a GitHub release asset with no model hub presence, and the repository ships inference and evaluation code rather than the pretraining pipeline.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.