EGTrans

EG-TransAttention: Edge-aware multiscale attention network for predicting circRNA-disease associations

Description

EG-TransAttention is a deep learning framework designed to predict potential associations between circular RNAs (circRNAs) and diseases. It integrates edge-aware multiscale attention mechanisms to capture both local and global topological features from heterogeneous biological networks. The model combines graph convolutional operations with edge-guided attention to enhance representation learning and improve prediction accuracy.

---

Dataset Information

The datasets folder contains the processed circRNA–disease association data, similarity matrices, and name mapping files used for training and evaluation.

Files Description

File Name	Description
all_circrna_disease_pairs.csv	All known circRNA–disease associations used as the base dataset.
c_c.csv	circRNA–circRNA similarity matrix (e.g., based on sequence or functional similarity).
c_d.csv	circRNA–disease association matrix (binary: 1 for known association, 0 for unknown).
c_d_name.csv	circRNA–disease pair names corresponding to c_d.csv indices.
circname.txt	List of circRNA names corresponding to the circRNA similarity matrices.
d_d.csv	disease–disease similarity matrix (e.g., based on semantic similarity).
disease semantic similarity.csv	Additional disease semantic similarity scores used for constructing graphs.
disname.txt	List of disease names corresponding to the disease similarity matrices.
funicircRNA.csv	Functional similarity information of circRNAs.
GCC_similarity.csv	Global topological similarity of circRNAs.
GDD_similarity.csv	Global topological similarity of diseases.

Data Source

The data is collected and integrated from the circR2Disease and circRNADisease databases, then processed into:

• circRNA–circRNA similarity matrices
• disease–disease similarity matrices
• circRNA–disease association matrices
• circRNA/disease name mapping files

These files are loaded automatically by load_data.py during training.

Code Structure

• main.py: Entry point for training and evaluation
• train1.py: Training loop and model optimization
• model.py: Defines EG-TransAttention architecture and attention modules
• load_data.py: Loads and preprocesses circRNA-disease data
• param.py: Contains hyperparameters and configuration settings
• evaluation_scores.py: Computes evaluation metrics (AUC, accuracy, F1-score, etc.)
• MA.py / PA.py: Implements multiscale attention and Parnet attention

Usage Instructions

Run main.py directly to do CDA tasks.

• Prepare the Dataset

  Download circRNA–disease association data from circR2Disease and circRNADisease.

• Generate Fusion Similarities

  Before training, run load_data.py to preprocess the data and compute Gaussian-fused similarities:

  python load_data.py

• This will:

  Read raw matrices: c_d.csv, c_c.csv, d_d.csv.

• Compute Gaussian similarity matrices:

  GCC_similarity.csv for circRNAs

  GDD_similarity.csv for diseases

  Integrate original similarity matrices with Gaussian similarities.

  Prepare inputs for graph construction in the model.

⚠️ Make sure the paths in param.py match your dataset location.

• Set Hyperparameters

  Edit param.py to configure model parameters, such as learning rate, epochs, batch size, number of attention heads, etc.

• Run Training python main.py

  Loads preprocessed data from load_data.py

  Constructs heterogeneous graphs

  Trains the EG-TransAttention model

  Outputs prediction scores and learned node features

Requirements

• Python ≥ 3.8
• PyTorch ≥ 1.10
• NumPy
• SciPy
• scikit-learn
• NetworkX
• tqdm