A graph neural network-based method to identify lncRNA subcellular localizations.

The subcellular localizations of long non-coding RNAs (lncRNAs) are closely related to their biological functions and disease mechanisms. Existing methods for their identification have limitations in handling data imbalance and complex sequence structures. In this paper, a Graph Neural Network-based method to identify lncRNA subcellular localizations, lncGATSagePre, is proposed, which effectively captures the local dependencies and long-range structural relationships of sequences by integrating Graph Attention (GAT) Network and Graph Sample and Aggregate (GraphSAGE) Network. Specifically, the de Bruijn graph is used to transform lncRNA sequences into graph structures, with k-mer nodes extracted via sliding windows and their semantics initialized using Word2vec. The Synthetic Minority Oversampling Technique (SMOTE) is introduced to address the data imbalance problem. A two-layer GAT-GraphSAGE network is employed to realize adaptive aggregation and weight assignment of node features, enhancing the model's ability to identify rare categories. Experimental results show that in the four-class classification task involving cytoplasm, nucleus, ribosome, and exosome, the lncGATSagePre model achieves a weighted F1-score of 0.549 on the independent test set, significantly outperforming existing methods such as lncLocator 2.0, DeepLncLoc, and GraphLncLoc. Ablation experiments verify the synergistic advantages of GAT in local feature extraction and GraphSAGE in large-scale graph data processing, and the hierarchical structure of GraphSAGE followed by GAT is more conducive to feature aggregation and relationship modeling. Although the model still has room for optimization in classifying minority samples (e.g., exosomes), its strategy of fusing sequence structure and semantic information through Graph Neural Networks provides a new and effective way for lncRNA subcellular localization research, which is of great significance for revealing lncRNA functional mechanisms and mining disease targets.