Research map / Probabilistic and scientific modeling / Molecules and proteins
SAGEPhos: Sage Bio-Coupled and Augmented Fusion for Phosphorylation Site Detection
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 SAGEPhos, a structure-aware network that fuses kinase sequence, substrate local sequence (ESM2), and AlphaFold-derived substrate structure (R-GCN) via gated-residual fusion to predict individual kinase-specific phosphorylation sites, and releases a new structure-augmented dataset.
Problem
Predicting whether a given substrate site is phosphorylated by a specific individual kinase; existing individual-kinase methods use only sequence and ignore structural information.
Approach
A Dynamic fusiOn meChanism (DOC) combining a gated and a residual branch, applied in two stages: intra-modal Bio-Augmented fusion (substrate sequence primary, predicted structure auxiliary) and inter-modal Bio-Coupled fusion (substrate primary, kinase sequence and physicochemical properties auxiliary), with an emphasis embedding at the phosphorylation site.
Key results
- On the warm-start partition SAGEPhos reaches 80.6% accuracy and 88.3 AUC-ROC, versus 73.5% and 78.5 for the best baseline MusiteDeep, roughly 10% and 12% gains as claimed.
- On the zero-shot CDK17 test set SAGEPhos achieves 91.6% accuracy and 97.4 AUC-ROC, versus 69.3 accuracy and 94.2 AUC-ROC for MusiteDeep (its best baseline on accuracy); other baselines are much lower.
- Ablation shows removing both fusion modules drops accuracy from 80.6% to 65.8%, and removing the phosphorylation-site emphasis reduces AUC-PRC from 86.2 to 82.6.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add substrate sequence–structure intra-fusion (ablation) | SAGEPhos · Warm-start | Accuracy: 65.8±0.7 -> 78.2±0.5 | Helped |
| add substrate sequence–structure intra-fusion (ablation) | SAGEPhos · Warm-start | AUC-ROC: 73.1±0.8 -> 85.8±0.3 | Helped |
| add substrate–kinase inter-fusion (ablation) | SAGEPhos · Warm-start | Accuracy: 65.8±0.7 -> 77.6±0.2 | Helped |
| add substrate–kinase inter-fusion (ablation) | SAGEPhos · Warm-start | AUC-ROC: 73.1±0.8 -> 86.0±0.2 | Helped |
| add phosphorylation-site emphasis (ablation) | SAGEPhos · Warm-start | Accuracy: 76.8±0.2 -> 80.6±0.2 | Helped |
| add phosphorylation-site emphasis (ablation) | SAGEPhos · Warm-start | AUC-ROC: 84.5±0.3 -> 88.3± 0.1 | Helped |
| use R-GCN instead of GCN as the structure encoder (ablation) | SAGEPhos · Warm-start | Accuracy: 78.2 -> 80.6 | Helped |
| use R-GCN instead of GIN as the structure encoder (ablation) | SAGEPhos · Warm-start | Accuracy: 78.6 -> 80.6 | Helped |
| use R-GCN instead of GAT as the structure encoder (ablation) | SAGEPhos · Warm-start | Accuracy: 77.6 -> 80.6 | Helped |
| vary initial fusion weights α1, α2, β1, and β2 (ablation) | SAGEPhos · Warm-start | Accuracy: 79.1 to 80.2 | No clear effect |
| apply the model to an unseen kinase–substrate dataset | SAGEPhos · CDK17 zero-shot | Accuracy: 69.3 -> 91.6 | Helped |
Surprising findings
- Phosformer has a very low FPR (1.4%) but only 11.3% TPR, so its high-looking metrics suggest a tendency to predict negatives; the authors note Phosformer and Phosformer-ST training code is not open-sourced.
- MusiteDeep achieves comparable AUC-ROC on zero-shot kinases but very low AUC-PRC (4.0-23.8%), showing AUC-ROC alone can mislead on imbalanced settings.
Limitations
- Kinase-cold-start performance drops (68.6% accuracy) because entire kinase families are absent from training and kinase data is sparse.
- Truly unseen kinase and substrate cold-start test sets were not feasible due to 99.9% interconnectivity of the dataset.
- Structures are AlphaFold2 predictions that may be noisy; baselines were evaluated on the authors' own dataset and some baselines' training code was unavailable.
Experimental setup
Datasets: Phospho.ELM PhosphoNetworks PhosphoSitePlus AlphaFoldDB Johnson et al. 2023 kinase ranking dataset (CDK17, MYLK4, CDKL1, PHKG2, SRPK3 zero-shot)
Benchmarks: SAGEPhos warm-start dataset kinase-cold-start partition substrate-cold-start partition CDK17 zero-shot MYLK4/CDKL1/PHKG2/SRPK3 zero-shot
Metrics: Accuracy AUC-ROC AUC-PRC FPR
Baselines: MusiteDeep DeepPhos PhosIDN PhosIDNSeq EMBER Phosformer Phosformer-ST
Builds on: ESM2 AlphaFold2 R-GCN Phosformer
Closest papers by meaning
- Towards precision protein-ligand affinity prediction benchmark: A Complete and Modification-Aware DAVIS Dataset (NeurIPS 2025 · similarity 0.83)
- UniSite: The First Cross-Structure Dataset and Learning Framework for End-to-End Ligand Binding Site Detection (NeurIPS 2025 · similarity 0.83)
- PPDiff: Diffusing in Hybrid Sequence-Structure Space for Protein-Protein Complex Design (ICML 2025 · similarity 0.82)
- PSC-CPI: Multi-Scale Protein Sequence-Structure Contrasting for Efficient and Generalizable Compound-Protein Interaction Prediction (AAAI 2024 · similarity 0.82)
- MeToken: Uniform Micro-environment Token Boosts Post-Translational Modification Prediction (ICLR 2025 · similarity 0.82)
- MMSite: A Multi-modal Framework for the Identification of Active Sites in Proteins (NeurIPS 2024 · similarity 0.82)
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.