Deep-belief network for predicting potential miRNA-disease associations

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• MicroRNAs (miRNAs) are crucial in disease pathogenesis, development, diagnosis, and treatment.
• Computational methods offer a cost-effective alternative to traditional experiments for predicting miRNA-disease associations.

Purpose of the Study:

• To develop an advanced computational model, Deep Belief Network for miRNA-Disease Association prediction (DBNMDA), for identifying potential miRNA-disease links.
• To improve prediction accuracy by leveraging information from all miRNA-disease pairs during model pre-training.

Main Methods:

• Constructed feature vectors for all miRNA-disease pairs to pre-train Restricted Boltzmann Machines.
• Fine-tuned the Deep Belief Network (DBN) using positive and selected negative samples to generate prediction scores.
• Employed global leave-one-out cross-validation (LOOCV), local LOOCV, and 5-fold cross-validation for performance evaluation.

Main Results:

• DBNMDA achieved high AUC values: 0.9104 (global LOOCV), 0.8232 (local LOOCV), and 0.9048 ± 0.0026 (5-fold CV).
• Performance metrics surpassed those of previous computational models.
• Case studies demonstrated high accuracy, with 84% (breast neoplasms), 100% (lung neoplasms), and 88% (esophageal neoplasms) of top predictions validated by recent literature.

Conclusions:

• DBNMDA is a robust and effective computational method for predicting potential miRNA-disease associations.
• The model's innovative pre-training strategy mitigates issues related to limited known associations.
• DBNMDA offers a valuable tool for advancing research in miRNA-related diseases.