Inferring Potential CircRNA-Disease Associations via Deep Autoencoder-Based Classification

Circular RNAs (circRNA) are endogenous non-coding RNA molecules with a stable circular conformation. Growing evidence from recent experiments reveals that dysregulations and abnormal expressions of circRNAs are correlated with complex diseases. Therefore, identifying the causal circRNAs behind diseases is invaluable in explaining the disease pathogenesis. Since biological experiments are difficult, slow-progressing, and prohibitively expensive, computational approaches are necessary for identifying the relationships between circRNAs and diseases.

We propose an ensemble method called AE-RF, based on a deep autoencoder and random forest classifier, to predict potential circRNA-disease associations. The method first integrates circRNA and disease similarities to construct features. The integrated features are sent to the deep autoencoder, to extract hidden biological patterns. With the extracted deep features, the random forest classifier is trained for association prediction.

AE-RF achieved AUC scores of 0.9486 and 0.9522, in fivefold and tenfold cross-validation experiments, respectively. We conducted case studies on the top-most predicted results and three common human cancers. We compared the method with state-of-the-art classifiers and related methods. The experimental results and case studies demonstrate the prediction power of the model, and it outperforms previous methods with high degree of robustness. Training the classifier with the unique features retrieved by the autoencoder enhanced the model's predictive performance. The top predicted circRNAs are promising candidates for further biological tests.