KS-CMI: A circRNA-miRNA interaction prediction method based on the signed graph neural network and denoising autoencoder

Summary

Circular RNA (circRNA) plays an important role in the diagnosis, treatment, and prognosis of human diseases. The discovery of potential circRNA-miRNA interactions (CMI) is of guiding significance for subsequent biological experiments. Limited by the small amount of experimentally supported data and high randomness, existing models are difficult to accomplish the CMI prediction task based on real cases. In this paper, we propose KS-CMI, a novel method for effectively accomplishing CMI prediction in real cases. KS-CMI enriches the ‘behavior relationships’ of molecules by constructing circRNA-miRNA-cancer (CMCI) networks and extracts the behavior relationship attribute of molecules based on balance theory. Next, the denoising autoencoder (DAE) is used to enhance the feature representation of molecules. Finally, the CatBoost classifier was used for prediction. KS-CMI achieved the most reliable prediction results in real cases and achieved competitive performance in all datasets in the CMI prediction.

Graphical abstract

Highlights

• •

  We propose a circRNA-miRNA interaction prediction method based on the balance theory

• •

  The model enriches association information by adding cancer molecules

• •

  KS-CMI allows expansion in heterogeneous graphs of binary and triplet groups

Abstract

Gene network; Neural networks