Research map / Label-efficient and robust learning / Recognition and long-tail
Improving PTM Site Prediction by Coupling of Multi-Granularity Structure and Multi-Scale Sequence Representation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PTM-CMGMS, which predicts PTM sites by combining AlphaFold-based multi-granularity (atom, amino acid, whole protein) structure representations trained with contrastive loss and a multi-scale sequence model (BiLSTM+Transformer+MCNN) plus a motif-similarity feature.
Problem
Predicting post-translational modification (PTM) sites, which existing methods treat mainly as sequence modeling and which ignore structure and atom-level granularity.
Approach
Learns neighborhood structure representations at atom, amino acid, and whole-protein granularity from AlphaFold structures, optimizes them with contrastive loss, and couples them with a multi-scale sequence encoder and a motif (aligned context consensus) similarity feature.
Key results
- On the Crotonylation dataset PTM-CMGMS reaches MCC 0.7005, AUC 0.9245, AUPR 0.9386, versus 0.6426 MCC for the best baseline Adapt-Kcr.
- On Succinylation, MCC 0.3072 vs 0.2990 for LMSuccSite; on Nitrosylation, MCC 0.3398 vs 0.3060 for LMSuccSite.
- Relative to LMSuccSite, MCC improves by 7.83% on Crotonylation, 0.82% on Succinylation and 3.38% on Nitrosylation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PTM-CMGMS instead of Adapt-Kcr | Crotonylation site prediction · Crotonylation test set | MCC: 0.6426 -> 0.7005 | Helped |
| Use PTM-CMGMS instead of LMSuccSite | Succinylation site prediction · Succinylation test set | MCC: 0.2990 -> 0.3072 | Helped |
| Use PTM-CMGMS instead of LMSuccSite | Nitrosylation site prediction · Nitrosylation test set | MCC: 0.3060 -> 0.3398 | Helped |
| Add amino acid granularity structure representations (ablation) | PTM-CMGMS · PTM site prediction, three datasets | predictive performance | Helped |
| Add atom granularity structure representations (ablation) | PTM-CMGMS · PTM site prediction, three datasets | predictive performance | Helped |
| Add whole protein granularity structure representations (ablation) | PTM-CMGMS · PTM site prediction, three datasets | predictive performance | Helped |
| Add contrastive learning (ablation) | PTM-CMGMS · PTM site prediction, three datasets | predictive performance | Helped |
| Add multi-granularity structure information (ablation) | PTM-CMGMS · PTM site prediction, three datasets | predictive performance | Helped |
| Add multi-scale CNN (ablation) | PTM-CMGMS · PTM site prediction, three datasets | predictive performance | Helped |
| Add motif-based information extraction (ablation) | PTM-CMGMS · PTM site prediction, three datasets | predictive performance | Helped |
| Use BiLSTM, Transformer, and MCNN together instead of pairwise architecture combinations (ablation) | Context sequence information extraction · Crotonylation dataset | predictive performance | Helped |
| Use BiLSTM+Transformer+MCNN instead of ESM or ProtTrans (ablation) | Context sequence information extraction · Crotonylation dataset | predictive performance | Helped |
Surprising findings
- Removing structure information gives the largest performance drop in ablations, indicating structure features drive most of the gain.
- The gain over LMSuccSite on Succinylation is marginal (MCC 0.3072 vs 0.2990).
Limitations
- Structure inputs depend on AlphaFold predicted structures rather than experimental ones.
- Gains on Succinylation over the best PLM-based baseline are small.
- Evaluated only on three PTM types (crotonylation, succinylation, nitrosylation), with no crosstalk modeling.
Experimental setup
Datasets: Crotonylation dataset Succinylation dataset Nitrosylation dataset
Benchmarks: Crotonylation dataset Succinylation dataset Nitrosylation dataset
Metrics: MCC AUC AUPR
Baselines: PSuccE LightGBM-CrotSite DeepNitro Deep-Kcr DeepSuccinylSite Adapt-Kcr BERT-Kcr LMSuccSite
Builds on: AlphaFold ESM ProtTrans BLOSUM62
Closest papers by meaning
- MeToken: Uniform Micro-environment Token Boosts Post-Translational Modification Prediction (ICLR 2025 · similarity 0.88)
- PROSPECT PTMs: Rich Labeled Tandem Mass Spectrometry Dataset of Modified Peptides for Machine Learning in Proteomics (NeurIPS 2024 · similarity 0.85)
- MMSite: A Multi-modal Framework for the Identification of Active Sites in Proteins (NeurIPS 2024 · similarity 0.84)
- Multi-Scale Representation Learning for Protein Fitness Prediction (NeurIPS 2024 · similarity 0.83)
- PSC-CPI: Multi-Scale Protein Sequence-Structure Contrasting for Efficient and Generalizable Compound-Protein Interaction Prediction (AAAI 2024 · similarity 0.83)
- From Mechanistic Interpretability to Mechanistic Biology: Training, Evaluating, and Interpreting Sparse Autoencoders on Protein Language Models (ICML 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.