Prediction of Cancer-Related miRNA Targets Using an Integrative Heterogeneous Graph Neural Network-Based Method

• Emanoel Aurelio Vianna Fabiano UFRGS https://orcid.org/0009-0001-7519-209X
• Mariana Recamonde-Mendoza UFRGS / Hospital de Clínicas de Porto Alegre https://orcid.org/0000-0003-2800-1032

Resumo

MicroRNAs (miRNAs) are crucial regulators of gene expression, including in diseases such as cancer. Although machine learning methods have shown promise in predicting miRNA-target interactions, they encounter challenges related to imbalanced classes and false positives. To tackle these issues, this study proposes a GNN-based model, using a variant of GraphSAGE algorithm named HinSAGE, which integrates validated miRNA-mRNA and mRNA-mRNA interactions with cancer-related gene expression data. Results show that our approach effectively learns miRNA-target interaction patterns from the graph structure and node features. The model achieves 77% precision, 80% recall, 78% F1-score, and 86% ROC AUC on the test data. It competes well with related approaches, reaching an F1-score of approximately 90% on a common test set. Thus, GNNs offer a promising avenue for studying miRNA-target interactions, providing balanced predictive power and improved precision through negative interaction sampling from the graph.