Motivation: Piwi-interacting RNAs (piRNAs) play a critical role in the progression of various diseases. Accurately identifying the associations between piRNAs and diseases is important for diagnosing and prognosticating diseases. Although some computational methods have been proposed to detect piRNA-disease associations, it is challenging for these methods to effectively capture nonlinear and complex relationships between piRNAs and diseases because of the limited training data and insufficient association representation. Results: With the growth of piRNA-disease association data, it is possible to design a more complex machine learning method to solve this problem. In this study, we propose a computational method called iPiDA-GCN for piRNA-disease association identification based on graph convolutional networks (GCNs). The iPiDA-GCN predictor constructs the graphs based on piRNA sequence information, disease semantic information and known piRNA-disease associations. Two GCNs (Asso-GCN and Sim-GCN) are used to extract the features of both piRNAs and diseases by capturing the association patterns from piRNA-disease interaction network and two similarity networks, respectively. Performing GCNs on these networks can capture complex network structure information, and learn discriminative features. Finally, the full connection networks and inner production are utilized as the output module to predict piRNA-disease association scores.The flowchart of GraLTR-LDA model is shown in Fig.1
Fig.1
The framework of iPiDA-GCN
(a) Heterogeneous network construction (Fig. 1A). Three kinds of edges including piRNA-piRNA similarities, disease-disease similarities and piRNA-disease interactions are collected in the heterogeneous piRNA-disease association network.
(b) GCN-based node feature extraction (Fig. 1B). “Asso-GCN” and “Sim-GCN” modules are designed to continuously learn node features from different subnetworks of piRNA-disease association network. Specifically, “Asso-GCN” captures hidden associated features of heterogeneous nodes from piRNA-disease interaction subnetwork, while “Sim-GCN” captures hidden associated features of homogeneous nodes from two similarity subnetworks.
(c) Association prediction for piRNA and disease (Fig. 1C). Three fully connected layers with 400, 200 and 100 neurons are employed to learn the low-dimensional representations of piRNAs and diseases. Finally, association scores between piRNAs and diseases are computed through inner product operation.
