⸻

DeepMethy

DeepMethy: A Deep Learning Model for Protein Methylation Site Prediction

This project provides a deep neural network framework for the prediction of arginine methylation sites in protein sequences using evolutionary features (BLOSUM62 & PSSM), convolutional layers, residual blocks, and attention mechanisms.

⸻

🧠 Requirements

Ensure the following packages are installed before running the code:

keras==2.0.0
numpy>=1.8.0
tensorflow (backend)

⸻

📁 Data Preparation

The input data should be stored in:

/dataset/test_file.csv

Data Format

•	For Prediction: The CSV file should contain the following columns: position, sequence


•	For Training: The CSV file should contain: label, position, sequence

•	label = 1 for methylated sites
•	label = 0 for non-methylated sites

Missing residues should be padded with "0" for sequences shorter than the required length.

⸻

🔍 Prediction

To perform predictions on your test dataset:

1.	Prepare a CSV file with: position, sequence


2.	Run: python predict.py

The output will be a .txt file with prediction scores:

"21" "0.9999963" "21" "0.95067513" "21" "1.0669616e-24" ...

🔧 Customization

You can modify parameters in predict.py under the main() function to adjust:

•	Input file path
•	Output file name
•	Threshold or model settings

⸻

🏋️‍♂️ Training with Your Own Data

To train the DeepMethy model from scratch:

1.	Prepare a CSV file with: label, position, sequence


2.	Run: python train.py

You can fine-tune hyperparameters and model architecture in train.py.

⸻

🧬 Model Architecture

•	Evolutionary Features:
•	BLOSUM62 Matrix
•	PSSM Matrix
•	Deep Learning Backbone:
•	1D Convolutional Layers
•	Residual Blocks
•	Dense Connections
•	Attention Mechanism
•	Multi-window feature extraction
•	Weighted loss for class imbalance

The complete structure is implemented in train.py under the /methods/ directory.

⸻

📂 Project Structure

DeepMethy/
│
├── dataset/
│   └── test_file.csv
│
├── methods/
│   ├── train.py
│   └── predict.py
│
├── README.md
└── requirements.txt (optional)

⸻

📫 Contact

If you have any questions or suggestions, feel free to reach out:

⸻