CERDA-HOSR

CERDA-HOSR is a novel computational method that leverages higher-order graph attention networks and graph convolutional networks to predict ceRNA-disease associations.

Requirements

• python==3.8
• networkx==3.0
• numpy==1.24.1
• scikit-learn==1.3.0
• tqdm==4.65.0
• matplotlib==3.7.1
• pandas==2.0.3
• torch==2.0.0+cu118, torchaudio==2.0.1+cu118, torchvision==0.15.1+cu118

Files

Data

data.xlsx: Dataset1 comprises 218 lncRNAs, 605 miRNAs, 1,051 mRNAs, and 314 human diseases, encompassing a total of 2,115 ceRNA network–disease regulatory associations. The dataset is publicly available and can be accessed through the LncACTdb database at http://bio-bigdata.hrbmu.edu.cn/LncACTdb.

The data files needed to run the model, which contain CircR2Disease v1.0, CircR2Disease v2.0, HMDD v2.0, HMDD v3.0 and HMDD v4.0.

• disease semantic similarity matrix 1.txt and disease semantic similarity matrix 2.txt: Two kinds of disease semantic similarity.
• miRNA functional similarity matrix.txt: miRNA functional similarity.
• known disease-miRNA association number.txt: Validated miRNA-disease associations.
• disease number.txt: Disease ids and names.
• miRNA number.txt: miRNA ids and names.

Code

• HOGAT.py: Example of Higher-Order attention process
• training.py: Train the model
• negative-cerna.py: High-order negative sampling process
• main.py: running the complete training and evaluation pipeline
• predict_horda.py: predict scores

Usage

• Download code and data, take the CircR2Disease v1.0 dataset as an example: python predict_horda.py
  --model_file best.pkl
  --train_graph graph_train.txt
  --full_graph CircR2Disease.txt
  --rna_num 585
  --disease_num 88
  --topk 15