This repo has been put on shelf for now due to other projects

ContextPTM

ContextPTM: Context-Aware PTM Site Prediction via Protein Interaction Network Embeddings and Kinase Priors

Justin Li, Ji Yu · Center for Cell Analysis and Modeling, UConn Health

---

Overview

DeepMVP (Wen et al., Nature Methods 2025) predicts PTM sites using only local sequence windows, ignoring protein-level context. This repository proposes ContextPTM, which augments the CNN+BiGRU ensemble with two complementary protein-level feature sets:

1. node2vec PPI embeddings (128-dim): learned from the STRING v12.0 physical interaction network via node2vec
2. Kinase prediction scores (605-dim): aggregated per-protein scores from GPS 5.0, NetworkKIN, and PhosphoPICK
3. Combined (733-dim): both feature types concatenated

All variants are evaluated against a reproduced DeepMVP baseline across all 8 PTM types.

---

Results

All AUROC values on held-out test sets. Ensemble of 10 models with IQR outlier-excluded averaging.

PTM	Original DeepMVP	Baseline	+Kinase	+node2vec PPI	+Combined
Phosphorylation S/T	0.950	0.9510	0.9665	0.9654	0.9672
Phosphorylation Y	0.910	0.8710	0.9321	0.9220	0.9340
Sumoylation K	0.850	0.8615	0.9256	0.9210	0.9269
Ubiquitination K	0.870	0.8814	0.9296	0.9248	0.9307
Acetylation K	0.900	0.9049	0.9615	0.9560	0.9623
N-Glycosylation N	0.980	0.9868	0.9968	0.9954	0.9969
Methylation K	0.950	0.9503	0.9814	0.9785	0.9815
Methylation R	0.960	0.9306	0.9810	0.9738	0.9811

All three feature variants consistently outperform the original DeepMVP across all 8 PTM types. The combined model achieves the best performance on all PTM types, suggesting that PPI network topology and kinase regulatory context provide complementary information.

---

Repository Structure

ContextPTM/
│
├── README.md
│
├── models/                         # PTM prediction models
│   ├── deepmvp_reproduce_v2.py     # Baseline: CNN+BiGRU ensemble, sequence only
│   ├── deepmvp_ppi.py              # +node2vec PPI embeddings (128-dim)
│   ├── deepmvp_kinase.py           # +Kinase prediction scores (605-dim)
│   └── deepmvp_combined.py         # +Combined (node2vec + kinase, 733-dim)
│
├── embeddings/                     # Feature generation scripts
│   ├── node2vec_train.py           # Train node2vec on STRING PPI network
│   └── build_protein_features.py   # Parse node2vec output → protein_features_ppi.npy
│
├── slurm/                          # SLURM job submission scripts
│   ├── submit_deepmvp_noppi.sh
│   ├── submit_deepmvp_ppi.sh
│   ├── submit_deepmvp_ppi_n2v.sh
│   └── submit_deepmvp_kinase.sh
│
└── notebooks/                      # Exploratory analysis (poster work)
    └── exploratory/

---

Requirements

pip install torch numpy pandas scikit-learn tqdm node2vec

Tested with Python 3.10, PyTorch 2.x, CUDA 12.8 (NVIDIA A100).

---

Data

Training and test data are from PTMAtlas (Wen et al. 2025), the same dataset used to train the original DeepMVP. Files follow the format:

protein  aa  pos  x  y

where x is a 31-character amino acid window centered on the candidate site and y is the binary label.

PPI network: STRING v12.0 physical interaction edges, filtered at combined_score ≥ 200.

---

Reproducing Results

Step 1: Generate node2vec PPI embeddings

# Train node2vec on STRING network
python embeddings/node2vec_train.py

# Parse output and map to UniProt IDs
python embeddings/build_protein_features.py

This produces:

• protein_features_ppi.npy — shape (18423, 128)
• protein_ids_ppi.json — corresponding UniProt IDs

Step 2: Run baseline (sequence only)

sbatch slurm/submit_deepmvp_noppi.sh

Or directly:

python models/deepmvp_reproduce_v2.py

Step 3: Run +node2vec PPI

sbatch slurm/submit_deepmvp_ppi_n2v.sh

Or directly:

python models/deepmvp_ppi_n2v.py

Step 4: Run +kinase scores

Requires protein_features.npy and protein_ids.json (kinase prediction scores from GPS/NetworkKIN/PhosphoPICK, generated separately).

sbatch slurm/submit_deepmvp_ppi.sh

---

Model Architecture

The baseline follows the CNN+BiGRU architecture described in DeepMVP (Extended Data Fig. 3c), re-implemented in PyTorch:

Input: one-hot encoded sequence window (23 × 31)
  → Conv1D(512, k=5) + BN + LeakyReLU + Dropout  ×3
  → Bidirectional GRU (50 units) → hidden state concat → (100,)
  → FC(100→64) + BN + LeakyReLU + Dropout
  → FC(64→1)

The +PPI variant adds a parallel branch:

PPI vector (128-dim or 605-dim)
  → FC(dim→128) + LeakyReLU → (128,)
  
Concat([sequence_branch(100,), ppi_branch(128,)]) → (228,)
  → FC(228→64) + BN + LeakyReLU + Dropout
  → FC(64→1)

Ten models are trained per PTM type (different random seeds), with early stopping on validation AUROC. Final predictions use IQR outlier-excluded averaging across the 10 models.

---

Training Details

Hyperparameter	Value
Optimizer	Adam
Learning rate	1e-3
Batch size	64
Max epochs	100
Early stopping patience	10
Val split	10% of train
Loss	BCEWithLogitsLoss + pos_weight
Ensemble size	10 models

---

Citation

If you use this work, please also cite the original DeepMVP paper:

Wen B, Wang C, Li K, et al. DeepMVP: deep learning models trained on
high-quality data accurately predict PTM sites and variant-induced alterations.
Nature Methods, 2025.