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

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Subcellular localization of long non-coding RNAs (lncRNAs) is crucial for their biological functions and involvement in disease mechanisms.
• Current methods for lncRNA localization identification face challenges with imbalanced data and complex sequence structures.

Purpose of the Study:

• To propose a novel Graph Neural Network (GNN)-based method, lncGATSagePre, for enhanced identification of lncRNA subcellular localization.
• To address data imbalance and effectively model complex sequence relationships in lncRNA localization prediction.

Main Methods:

• lncRNA sequences were converted into graph structures using de Bruijn graphs with k-mer nodes initialized by Word2vec.
• The Synthetic Minority Oversampling Technique (SMOTE) was employed to mitigate data imbalance.
• A two-layer Graph Attention (GAT) Network and Graph Sample and Aggregate (GraphSAGE) Network architecture was utilized for adaptive feature aggregation.

Main Results:

• The lncGATSagePre model achieved a weighted F1-score of 0.549 on a four-class classification task (cytoplasm, nucleus, ribosome, exosome) on an independent test set.
• lncGATSagePre significantly outperformed existing methods like lncLocator 2.0, DeepLncLoc, and GraphLncLoc.
• Ablation studies confirmed the synergistic benefits of GAT for local feature extraction and GraphSAGE for large-scale graph processing.

Conclusions:

• The proposed lncGATSagePre model offers a novel and effective approach for lncRNA subcellular localization research by integrating sequence structure and semantic information via GNNs.
• This method holds significant potential for advancing our understanding of lncRNA functional mechanisms and identifying disease targets, despite needing further optimization for minority sample classification.