Detecting piRNA-disease associations holds significant importance in understanding disease mechanisms and biomarker discovery. In this study, we propose a new computational method named iPiDA-LGE to detect piRNA-disease associations. In contrast to competing methods, it predominantly possesses the following advantages: (i) The global graph learning module incorporates side information like piRNA sequence and disease ontology, and learns various basic predictors to construct supplementary heterogeneous association network. Therefore, the apparent sparsity issue of original association can be alleviated by enriching biological semantics. (ii) Diverging from general pair features obtained by global graph learning, the local graph learning module in iPiDA-LGE considers the specific functional mechanism of piRNAs in different diseases, and encodes each target piRNA-disease pair as a local graph. As a result, it can learn a more discriminative summary representation by capturing specific contextual information. (iii) Lastly, iPiDA-LGE integrates the local and global graph representation learning, which can simultaneously achieve refined inferences based on local graphs and overarching judgments derived from global graphs, leading to an enhancement in predictive performance. It is anticipated that the proposed local and global graph ensemble learning framework can provide innovative computational insights for other bio-entity link prediction problems.

There are three primary processes in the framework of iPiDA-LGE: (i) Graph construction. The global graph is constructed and supplemented with piRNA sequences, disease ontology knowledge and validated piRNA-disease relationships. In addition, the local context graph for each target pair is extracted from original bipartite graph. (ii) Local/Global graph representation. The representations of piRNA and disease nodes are captured by global-level GCN, while the pair representations are obtained by local-level GCN. (iii) Association prediction. The dense layer and multi-layer perceptron are applied to reduce feature dimensionality and calculate association scores. Finally, the global-level and local-level association scores are integrated with different weight coefficients to predict the relationships between piRNAs and diseases.

RNADisease http://www.rnadisease.org/

piRBase http://www.regulatoryrna.org/database/piRNA/

piRNA-eQTL http://njmu-edu.cn:3838/piRNA-eQTL/

Disease Ontology https://disease-ontology.org/

piRTarBase http://cosbi6.ee.ncku.edu.tw/piRTarBase/

Please cite the following paper when using the web server or dataset at this website:

Wei H, Hou J, Liu Y, Liu B. iPiDA-LGE: A Local and Global Graph Ensemble Learning Framework for Identifying PiRNA-Disease Associations. (under revision) .