Exploring noncoding RNAs in thyroid cancer using a graph convolutional network approach

Noncoding RNAs (ncRNAs) are crucial regulators in initiating and promoting thyroid cancer. Exploring the relationship between ncRNAs and thyroid cancer is essential for the diagnosis and treatment of thyroid cancer. Wet-lab experiments are costly and are difficult to conduct on a large scale. Although there are several ncRNA and cancer-related databases, there are few data related to thyroid cancer. There is a lack of computational approaches for predicting ncRNA-thyroid cancer associations. This work describes TCGCN, a linear residual graph convolution network to predict ncRNA-thyroid cancer associations. We collected a large amount of ncRNA-disease association data and constructed a bipartite graph. We use a simple linear embedding propagation at each convolutional layer and use the weighted sum of the embeddings on all graph convolutional layers to make the final prediction. In 5-fold cross-validation on the ncRNA-thyroid cancer dataset, TCGCN obtained signifi-cantly better performances with an AUC of 0.8162 and an AUPR of 0.8049, which are considerably better than those of other state-of-the-art approaches. We also demonstrate the usability of our method in the case studies.

Automated summary

This study presents TCGCN, a linear residual graph convolution network, for predicting ncRNA-thyroid cancer associations. By collecting a large amount of ncRNA-disease association data and constructing a bipartite graph, TCGCN combines linear embedding propagation and graph convolutional layers to achieve significantly better performance in prediction.