Predicting Mirna-Disease Associations Based on Neighbor Selection Graph Attention Networks

Area of Science:

• Biomedical informatics
• Computational biology
• Genomics

Background:

• Complex human diseases are linked to abnormal microRNA (miRNA) expression.
• Identifying miRNA-disease associations is crucial for clinical medicine.
• Traditional experimental methods for this identification are inefficient and time-consuming.

Purpose of the Study:

• To develop an efficient deep learning method for predicting miRNA-disease associations.
• To improve the accuracy and reliability of miRNA-disease association identification.

Main Methods:

• Proposed a novel deep learning model, NSAMDA (Neighbor Selection Attention model for MiRNA-Disease Association).
• Fused miRNA sequence and integrated similarity information to enrich miRNA features.
• Constructed a heterogeneous miRNA-disease graph using fused miRNA features and disease integrated similarity.
• Employed a neighbor selection mechanism within graph attention networks for feature aggregation.
• Utilized an inner product decoder for scoring potential miRNA-disease pairs.

Main Results:

• Achieved a mean Area Under the Curve (AUC) of 93.69% on the HMDD v2.0 dataset via five-fold cross-validation.
• Demonstrated high performance in predicting miRNA-disease associations.
• Case studies on esophageal neoplasm, lung neoplasm, and lymphoma confirmed the model's effectiveness.
• NSAMDA outperformed state-of-the-art models in predicting miRNA-disease associations.

Conclusions:

• NSAMDA offers a powerful and accurate deep learning approach for predicting miRNA-disease associations.
• The model's performance suggests significant potential for clinical applications in disease diagnosis and treatment.
• NSAMDA represents an advancement over existing computational methods in the field.