PTM-AnchorFM

A lysine in a protein can be acetylated, ubiquitinated, methylated, or sumoylated, and which of those happens depends on the sequence context around it. The conventional response is to train a separate predictor per modification, so each independently relearns what an ordinary lysine looks like before learning what distinguishes a modified one. PTM-AnchorFM — the preprint also calls it an anchor-guided rectified flow matching framework — tests whether that shared part can be learned once, leaving each modification type to contribute only its offset from the common residue background.

The mechanism is a pair of anchor sets and a flow field connecting them. For every PTM–residue pair, the model clusters unannotated but chemically compatible residues into background anchors and experimentally annotated modified residues into positive anchors, both in the embedding space of a frozen ProtT5 encoder. A conditional rectified-flow field, trained once across all PTM–residue pairs and then frozen, estimates the displacement from background to positive conditioned on the target modification. A candidate residue is scored by how it compares to both anchor sets and where that flow-derived offset places it.

The framework comes from Suresh Pokharel at the Rochester Institute of Technology and Bishnu Bhusal at the University of Missouri, posted to bioRxiv in August 2026. It sits alongside catalog entries attacking the same annotation problem from other angles — AstraPTM, which scores 25 PTM types in one transformer pass, Prosit-PTM on the mass-spectrometry side, and PTM-dCN on the design side. Its own result is candid: the shared formulation does not beat per-modification training, and the paper states that independent PTM-specific training remains the strongest reference.

#Key Features

• One model for six modification types: A single gated network answers any candidate site–modification query across phosphorylation, acetylation, ubiquitination, methylation, sumoylation, and N-linked glycosylation.
• Explicit anchor sets: Mini-batch k-means fits 32 background anchors per residue type and 16 positive anchors per PTM–residue pair. Anchors and normalization statistics are estimated from the training split alone, with no validation or test information.
• Rectified flow as a frozen feature extractor: The conditional flow field is trained once, then evaluated at one timepoint (t = 0.5) to yield two 1024-dimensional vectors, each projected to 128 dimensions, plus five scalar summaries.
• Modification-gated scoring head: The 261 flow dimensions join 8,198 anchor-contrast dimensions in a shared two-layer feedforward network (512, 128, ReLU, dropout) modulated by a 32-dimensional PTM-embedding gate, emitting a binary site–modification logit.
• Untouched language model backbone: The ProtT5 encoder is never fine-tuned; it supplies 1024-dimensional residue embeddings, standardized residue-wise before anchoring.

#Technical Details

Training uses a dbPTM 2025-derived corpus of 1,717,154 annotated sites across roughly 250,000 proteins, split 80:10:10 at the protein level so no protein appears in two partitions. The six types are steeply imbalanced, from 1,393,764 phosphorylation sites over 187,168 proteins down to 5,773 sumoylation sites over 1,681, with ubiquitination (154,147), acetylation (121,734), N-linked glycosylation (26,550), and methylation (15,186) in between. Negatives are undersampled 1:1 during training only; evaluation uses the full natural class distribution. Optimization is AdamW (lr 1e-3, weight decay 1e-4, batch 1024) for at most 20 epochs with early stopping on validation macro AUPRC, over five seeds.

The full model reaches a macro AUPRC of 0.4195 ± 0.0040, above the authors' gated multi-anchor ablation at 0.4154 ± 0.0024 but below independently trained per-modification models at 0.4353 ± 0.0021. Per-type scores are dominated by the sequon-constrained case: N-linked glycosylation reaches 0.8737, while ubiquitination scores 0.3874, sumoylation 0.3327, phosphorylation 0.3204, acetylation 0.3166, and methylation 0.2864. All comparisons are internal baselines; the authors decline to benchmark against published tools such as PTMGPT2, DeepMVP, MeToken, and MTPrompt-PTM, citing incompatible datasets and evaluation protocols.

#Applications

The intended use is proteome-scale triage: scoring candidate residues so mass-spectrometry or mutagenesis effort goes to the sites most likely to be real, across several modification types through one interface rather than a stack of per-type tools. Lysine is the natural case, since one query residue can be scored against four competing fates under a common representation. Adoption is limited for now — no weights, package, or hosted inference have been released, the repository cited in the paper's Code Availability section is not publicly accessible, and a web server is described as planned future work.

#Impact

The contribution is a framing question answered honestly rather than a performance win. Sharing residue-background structure across modification types is an appealing way to let rare PTMs borrow statistical strength from abundant ones, and the anchor-plus-offset construction gives that intuition a concrete form; the measured outcome is that it does not yet recover what independent per-modification training achieves, with the flow-derived features occupying 261 of 8,459 input dimensions. Evaluation rests on one dbPTM-derived benchmark, and the authors note that broader validation on independent and time-split datasets is needed before the shared-model framing can be judged properly.