<![CDATA[Identifying liver cancer-associated lncRNAs is a classification problem that can leverage various biological characteristics, such as secondary structure, expression levels, and molecular associations with RNA-binding proteins (RBPs). This study develops a VAE-GraphSAGE framework to integrate this information within a heterogeneous graph. A Variational Autoencoder (VAE) is employed for representation learning to extract lncRNA secondary structure embeddings through a reconstruction process that does not rely on class labels. These embeddings are combined with expression features to form node features, while data on lncRNA similarity, lncRNA-RBP interactions, and protein-protein interactions (PPI) are used to construct the heterogeneous graph. Subsequently, GraphSAGE serves as the classifier, utilizing inter-node relationship information to classify the lncRNAs. The dataset comprises 1,000 lncRNAs (400 positive and 600 negative) and 85 RBPs (as protein nodes), with a training-validation-testing split of 70:15:15. Comparative results indicate that VAE-GraphSAGE achieves the highest recall among the tested models (outperforming AE-GraphSAGE, MLP-GraphSAGE, CNN-GraphSAGE, and ResNet-GraphSAGE); meanwhile, MLP-GraphSAGE yields the highest accuracy, precision, specificity, and F1-score, and CNN-GraphSAGE produces the highest ROC-AUC. Furthermore, using the same VAE embeddings, VAE-GraphSAGE demonstrates superior performance compared to VAE-HGT and VAE-RGCN across the evaluated classification metrics. These results demonstrate that the VAE effectively generates secondary structure representations that facilitate the detection of positive lncRNAs, while GraphSAGE successfully leverages biological relationships within the heterogeneous graph for the classification task.]]>
Copyrights © 2026
