PTM-Blocker
1 Introduction
Post-translational modifications (PTMs) are key regulators of protein function, stability, and interactions, and are critical in cellular signaling, localization, and disease mechanisms. However, experimental identification of PTMs (e.g., mass spectrometry, western blotting, radioactive labeling) is costly and time-consuming, making computational approaches attractive alternatives. Traditional computational models rely only on local sequence features around PTM sites. Many existing pretrained protein language models (PLMs) are sequence-only, lack structural information, and are often single-task, preventing feature sharing across PTM types and limiting knowledge transfer and prediction performance.
MTPrompt-PTM addresses these limitations by using a structure-aware PLM (S-PLM) combined with prompt tuning, multi-task learning, and knowledge distillation. It predicts 13 types of PTMs and outperforms mainstream tools such as MusiteDeep and PTMGPT2 on modifications including phosphorylation and glycosylation, providing a reliable solution for efficient PTM site localization.
Figure 1. Schematic architecture of the MTPrompt-PTM model
Figure 2. Performance of MTPrompt-PTM model
PTM blocking requires not only precise localization of modification sites but also assessment of how mutations at those sites affect overall protein function. VespaG is an ultra-fast missense variant effect predictor that uses protein language model (pLM) embeddings as input and trains a lightweight deep learning model with GEMME evolutionary scores as pseudo-labels, achieving prediction performance comparable to mainstream methods (e.g., GEMME, AlphaMissense).
Figure 3. Schematic architecture of the VespaG model
PTM-Blocker links MTPrompt-PTM and VespaG to perform one-click prediction of PTM risk sites and the functional impact of mutation-based modifications. We also summarize common PTM types, involved amino acids or sequence motifs, and commonly used site-directed replacement mutation schemes (see table below).
Note that the recommended mutation schemes are not foolproof. Therefore, PTM-Blocker performs saturation mutagenesis for all potential PTM risk sites and clearly marks in the result files whether each mutation belongs to the recommended scheme; see section 3 Results Explanation.
| PTM | AA/Motif | Mutation |
|---|---|---|
| Phosphorylation | S | A |
| Phosphorylation | T | A |
| Phosphorylation | Y | F |
| N-Linked Glycosylation | N | Q |
| N-Linked Glycosylation | NXS/T | NXA |
| Ubiquitination | K | R |
| Acetylation | K | R |
| Acetylation | K | A |
| SUMOylation | K | R |
| Methylation | K | A |
| Methylation | K | Q |
| Methylation | R | A |
| Methylation | R | G |
| Succinylation | K | R |
| Succinylation | K | A |
| O-Linked Glycosylation | S | A |
| O-Linked Glycosylation | T | A |
| Palmitoylation | C | S |
| Palmitoylation | C | A |
2 Parameters
| Name | Description |
|---|---|
| Sequences | Amino acid sequence(s) or a FASTA file |
| PTM Type | Select the PTM type(s) to predict |
3 Results Explanation
The results are provided as two .csv files: ptm_site_predictions.csv and ptm_reform_evaluation.csv
In ptm_site_predictions.csv, each column means:
| Name | Description |
|---|---|
| prot_id | Sequence ID (protein identifier) |
| position | Predicted PTM risk position |
| probability | Score predicted by MTPrompt-PTM, range [0,1]; higher means more likely to be modified |
| ptm_type | Predicted PTM type |
| wildtype | The amino acid at the site |
In ptm_reform_evaluation.csv, each column means:
| Name | Description |
|---|---|
| prot_id | Sequence ID (protein identifier) |
| ptm_type | Predicted PTM type |
| ptm_position | Predicted PTM risk position |
| ptm_wildtype | Amino acid type before modification |
| ptm_probability | Score predicted by MTPrompt-PTM, range [0,1]; higher means more likely to be modified |
| mutation | Mutation notation, e.g., S35A indicates residue 35 mutated from S to A |
| is_rule_mutation | Whether the mutation belongs to our recommended replacement schemes |
| vespag_score | Score predicted by VespaG, range [0,1]; higher means greater predicted functional impact after mutation |
4 References
[1] Han Y, He F, Shao Q, Wang D, Xu D. MTPrompt-PTM: A Multi-Task Method for Post-Translational Modification Prediction Using Prompt Tuning on a Structure-Aware Protein Language Model. Biomolecules. 2025; 15(6):843. https://doi.org/10.3390/biom15060843
[2] Marquet C, Schlensok J, Abakarova M, Rost B, Laine E. Expert-guided protein language models enable accurate and blazingly fast fitness prediction. Bioinformatics. 2024; 40(11):btae621. https://doi.org/10.1093/bioinformatics/btae621