- Python 99.3%
- Shell 0.7%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
|
|
||
| __pycache__ | ||
| backbones/sdnn | ||
| eval_results | ||
| utils | ||
| config_train.yaml | ||
| losses.py | ||
| main.py | ||
| metric_recorder.py | ||
| ppi_dataset.py | ||
| ppi_learner.py | ||
| README.md | ||
| train.sh | ||
eSIG-Net
Accurate prediction of single-mutation induced perturbations on protein interactions using a language model.
Overview
eSIG-Net (edgetic mutation Sequence-based Interaction Grammar Network) is a novel sequence-based "Interaction Language Model" for predicting protein interaction alterations caused by single mutations. eSIG-Net combines various protein sequence embeddings, introduces a mutation-encoding module with syntax and evolutionary insights, and employs contrastive learning to evaluate mutation-induced interaction changes. eSIG-Net significantly outperforms current state-of-the-art sequence-based and structure-based prediction methods at predicting mutational impact on protein interactions.
Installation
Requirements
- Python 3.7+
- PyTorch
- CUDA (for GPU acceleration, recommended)
Dependencies
Install required packages:
pip install torch torchvision torchaudio
pip install numpy pandas scikit-learn
pip install h5py pyyaml optuna
Or create a conda environment:
conda create -n esignet python=3.8
conda activate esignet
conda install pytorch torchvision torchaudio -c pytorch
pip install numpy pandas scikit-learn h5py pyyaml optuna
Data Setup
Download Datasets
The datasets for eSIG-Net model training and evaluation can be downloaded from: Zenodo Repository
Data Structure
After downloading and extracting the datasets, organize your data as follows:
datasets/
├── csvs/
│ ├── germline.csv
│ ├── random_fold_0.csv
│ ├── random_fold_1.csv
│ ├── random_fold_2.csv
│ ├── random_fold_3.csv
│ └── random_fold_4.csv
└── embeddings/
├── sdnn_corrected_ppi.h5
└── 650M_1/
├── smile.10015.pt
├── smile.10015_A309T.pt
├── smile.10015_G429S.pt
└── ...
Data Format
-
CSV files: Each CSV file should contain columns (no header) for:
id_A: Original protein A identifierid_B: Protein B identifierlabel: Interaction label for original pairid_A_2: Mutated protein A identifierlabel_2: Interaction label for mutated pairmute_idx: Mutation index/position
-
H5 file: SDNN embeddings in HDF5 format with protein IDs as keys
-
PT files: ESM embeddings stored as PyTorch tensors with structure:
['representations'][33]: ESM layer 33 embeddings
Configuration
The training configuration is specified in config_train.yaml. Key parameters include:
train_task: Model task type (default:sdnn)exec_mode: Execution mode (trainoreval)k_fold: Number of folds for cross-validation (default: 5)batch_size_trn: Training batch size (default: 32)n_epochs: Number of training epochs (default: 8)lr_init: Initial learning rate (default: 0.001)dropout: Dropout probability (default: 0.5)- Model architecture parameters (
fc_hidden,pred_layer_num, etc.)
Usage
Training
Basic Training (Single Fold)
Train a single fold:
python main.py --config_fname config_train.yaml --fold 0 --save_model --save_results --mdl_dpath tmp_results/exp01
K-Fold Cross-Validation
Train all folds (specified in config):
python main.py --config_fname config_train.yaml --save_model --save_results --mdl_dpath tmp_results/exp01
The model will automatically train on all folds specified by k_fold in the config file.
Evaluation
Evaluate a trained model:
python main.py --config_fname config_train.yaml \
--eval \
--eval_csv datasets/ppi_case/case_reports_mutation.csv \
--eval_model_fpath tmp_results/ODrd01_5fold/exp0/model_pred_opt_fold_0.pth \
--feat_h5_fpath datasets/embeddings/sdnn/sdnn_corrected_ppi.h5 \
--save_result_fpath eval_results/ODrd01_5fold_exp0_fold0.csv \
--save_feature_fpath eval_results/feat_ODrd01_5fold_exp0_fold0.pt \
--fold 0
Using the Training Script
Alternatively, use the provided shell script:
bash train.sh
Note: The script includes GLIBC path configurations that may need adjustment for your system.
Output
Training Outputs
- Model checkpoints: Saved in
mdl_dpathdirectory asmodel_pred_opt_fold_{fold}.pth - Results CSV: Contains training metrics and predictions (if
save_results=True)
Evaluation Outputs
- Results CSV: Contains predictions and metrics for each sample
- Feature files: Extracted features saved as PyTorch tensors (.pt format)
Project Structure
eSIG-Net/
├── main.py # Main entry point
├── config_train.yaml # Training configuration
├── ppi_learner.py # Model training and evaluation logic
├── ppi_dataset.py # Dataset loading and preprocessing
├── losses.py # Loss functions
├── metric_recorder.py # Metrics tracking
├── backbones/
│ └── sdnn/
│ └── sdnn_model.py # SDNN model architecture
├── utils/
│ └── sdnn_utils.py # Utility functions
└── datasets/ # Data directory (to be created)
├── csvs/
└── embeddings/
Notes
- Ensure CUDA is available if using GPU acceleration
- The model requires both SDNN embeddings (H5 format) and ESM embeddings (PT format)
- For k-fold cross-validation, ensure all fold CSV files (
random_fold_0.csvthroughrandom_fold_{k-1}.csv) are present - ESM embeddings should be stored in
datasets/embeddings/650M_1/directory
Citation
If you use eSIG-Net in your research, please cite:
eSIG-Net: Accurate prediction of single-mutation induced perturbations on
protein interactions using a language model
Xingxin Pan, Aditya Shrawat, Sidharth Raghavan, Yuntao Yang, Zhao Li,
W. Jim Zheng, S. Gail Eckhardt, Erxi Wu, Juan I. Fuxman Bass, Daniel F. Jarosz,
Sidi Chen, Jason H. Huang, Daniel J. McGrail, Nidhi Sahni, and S. Stephen Yi
Correspondence: Xingxin Pan (xingxin.pan@bswhealth.org), Nidhi Sahni (nidhi.sahni.2025@gmail.com), Daniel J. McGrail (mcgraid@ccf.org), or S. Stephen Yi (song.yi@bcm.edu)