CLASPP

Train one classifier on all of dbPTM and it learns phosphorylation well and hydroxylation barely at all: the number of cataloged human sites differs between the two by orders of magnitude, and a single loss surface lets the abundant chemistry dominate the gradient. The field's usual answer has been to abandon the unified model — one predictor per modification, or several stitched into an ensemble — which leaves no shared representation across chemistries and no principled way to ask which of several modifications a residue is most likely to carry.

CLASPP — the Contrastively Learned Attention-based Stratified PTM Predictor — attacks the imbalance in the data rather than in the loss. Each of 12 PTM types is broken into sequence-defined sub-clusters by unsupervised clustering of its 21-mer peptides, and positives are then sampled uniformly from every sub-cluster rather than from the PTM type as a whole. Abundant modifications fragment into many sub-clusters, rare ones stay whole, and the balanced corpus that results lets one model score a residue for every chemistry in a single forward pass. It was developed in Natarajan Kannan's group at the University of Georgia, with a collaborator at Georgia State University, and published in PLOS Computational Biology in 2026.

Architecturally CLASPP sits in the same family as other frozen-backbone PTM predictors — PTM-AnchorFM over ProtT5 embeddings, or AstraPTM over ESM-2 — but inserts a supervised contrastive stage between pretraining and the task head, using the sub-cluster labels as the contrastive targets. That stage does more than regularize: the resulting embedding separates serine/threonine phosphosites by the substrate preference of their upstream kinase, a biological structure nobody supervised it to find.

#Key Features

• Fifty-four sub-clusters, twelve PTM types: Spectral or k-means clustering over inverse-Hamming affinity splits each well-populated PTM type into 2–20 sub-clusters, giving 54 stratified classes that are then sampled uniformly.
• Contrastive refinement, then a frozen encoder: A supervised contrastive objective over the sub-cluster labels reshapes the encoder; the encoder is then frozen and only the classification head is fine-tuned, which the authors found outperforms leaving it trainable.
• Kinase specificity as built-in explainability: The two largest S/T-phosphorylation sub-clusters resolve into basophilic and acidophilic kinase substrate profiles, matching known specificity logos, and separate cleanly in UMAP only after the contrastive stage.
• Constraint-based negative sampling: Negatives are graded hard, medium and easy, with easy negatives at least Hamming distance 7 from any known positive, so unannotated but motif-similar sites are not called true negatives.

#Technical Details

The encoder is ESM-2 150M — 30 layers, hidden width 640 — pretrained by masked language modeling on UniRef, refined by supervised contrastive loss, then frozen. On top sits a two-linear-layer head emitting 54 independent binary logits, max-pooled into 19 residue-specific positive labels plus one negative. The head dominates the parameter budget: the final model comprises 1.4 billion parameters against the encoder's 150M. Training data came from dbPTM, restricted to human sites, remapped to current accessions through UniProt and UniParc, and deduplicated by greedy clustering at a Hamming cutoff of 3. Uniform sampling capped each sub-cluster at 2,000 positives, yielding 97,989 21-mer peptides split 70/15/15 with the negative ratio capped at 1/60. The contrastive stage ran 1,000 epochs on four NVIDIA RTX 6000 Ada GPUs in under 72 hours; head fine-tuning ran 500 epochs on a single card in under 24.

Against MusiteDeep, MIND-S, PTMGPT2, MTPrompt-PTM and DeepMVP on a held-out dbPTM test set, CLASPP led on S/T-phosphorylation (F1 0.85, AUC-PRC 0.93), Y-phosphorylation (F1 0.85, AUC-PRC 0.77) and K-hydroxylation (F1 0.61, AUC-PRC 0.56), with competitors ahead on several other chemistries. The curation ablation is sharper: spectral sub-clustering reached F1 0.971 against 0.943 for the unstratified baseline.

#Applications

The immediate use is proteome-scale triage — scoring every candidate residue in a protein or a pathway to decide which sites justify targeted mass spectrometry. The paper's case study is the dark kinase DCLK3, where CLASPP recovered all five MS-validated ubiquitination sites that no model saw in training, though with more false positives than the lower-sensitivity baselines. The curated Zenodo release doubles as a standardized, class-balanced benchmark for anyone building the next multi-PTM model.

#Impact

CLASPP's contribution is as much a data argument as a modeling one: hierarchical curation and uniform sub-cluster sampling recover most of the ground ensembles were built to cover, and the resulting corpus ships so the comparison is reproducible. The limits are stated plainly. The 21-mer window is narrower than the full-length context MIND-S and AstraPTM use, and widening it under a transformer costs compute. Training was human-only; precision and false-positive rate stay stable across species, but recall falls outside Opisthokonta and drops sharply in E. coli, so predictions in distant organisms warrant caution. Inference code, the checkpoint and the curation pipeline are public under MIT, but no training script accompanies them, so the three-stage regimen must be reimplemented to be repeated.