GATLncLoc+C&S: Prediction of LncRNA subcellular localization based on corrective graph attention network

Abstract

Long non-coding RNAs (LncRNAs) have a wide range of regulatory roles in gene expression, and the subcellular localization identification of LncRNAs is of great value in understanding their biological functions. Graph neural networks can not only utilize sequence characteristics, but also learn hidden features from non-Euclidean data structures to obtain features with powerful characterization capabilities. To learn more fully from the limited LncRNA localization samples and efficiently exploit easily ignored label features, we propose a corrective graph attention network prediction model GATLncLoc+C&S in this paper. Compared with previous methods, the similarity of optimal features is first used to construct the graph. Then, a re-weighted graph attention network R-GAT is constructed and the soft labels obtained from it are used to correct the graph. Finally, the predicted localization label is further obtained by label propagation. Based on the combination of R-GAT and label propagation, GATLncLoc+C&S effectively solves the problems of few samples and data imbalance in LncRNA subcellular localization. The accuracy of GATLncLoc+C&S reached 95.8% and 96.8% in the experiments of 5- and 4-localization benchmark datasets, which reflects the great potential of our proposed method in predicting LncRNA subcellular localization. The source code and data of GATLncLoc+C&S are available at https://github.com/GATLncLoc-C-S/GATLncLoc-C-S.