PDMDA: predicting deep-level miRNA-disease associations with graph neural networks and sequence features

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• MicroRNAs (miRNAs) are crucial in human diseases, but experimental identification of their associations is costly and time-consuming.
• Existing computational methods predict miRNA-disease associations but not their specific types.
• There is a need for advanced computational approaches to predict deep-level miRNA-disease association types.

Purpose of the Study:

• To develop an end-to-end deep learning method, PDMDA, for predicting deep-level miRNA-disease associations.
• To leverage graph neural networks (GNNs) and miRNA sequence features for enhanced prediction accuracy.
• To differentiate and predict various types of miRNA-disease associations beyond simple existence.

Main Methods:

• PDMDA utilizes a fully connected network (FCN) to extract miRNA feature representations from sequence and structural data.
• Disease feature representations are derived from gene networks using a GNN model.
• A multilayer network with fully connected and softmax layers integrates miRNA and disease features for association scoring.

Main Results:

• PDMDA accurately predicts deep-level miRNA-disease associations, as validated by fivefold cross-validation.
• The model achieved high performance using the area under the receiver operating characteristic curve (AUC) metric.
• Experiments were conducted on six types of association samples, demonstrating the model's versatility.

Conclusions:

• PDMDA offers an effective computational approach for predicting deep-level miRNA-disease association types.
• The method provides a valuable tool for understanding the complex relationships between miRNAs and diseases.
• The developed model advances the field of miRNA-disease association prediction beyond binary classification.