Predicting non-coding RNA-based regulatory networks in cancer metastasis using a heterogeneous hierarchical graph transformer.

Abstract

e13663 Background: Advancements in high-throughput sequencing have enabled large-scale studies on the roles of non-coding RNAs (ncRNAs), including microRNAs (miRNAs) and circular RNAs (circRNAs) in cancer progression. Despite proposed involvement of ncRNAs in cancers, their regulatory roles in cancer metastatic events (CMEs) remains poorly understood, limiting the potentials for clinical applications. Methods: We developed a novel hierarchical graph neural network (GNN) to model complex regulatory interactions between miRNA and circRNA with CMEs (e.g. cancer cell invasion), across various cancers. Association of ncRNAs with CMEs were extracted from the ncR2Met database that included 32 CMEs spanning 53 human cancers. miRNA-circRNA interaction data was obtained from CircBank. miRNA and circRNA expression in different cancers were retrieved from dbDEMC and CircAtlas, respectively. The hierarchical GNN developed in this study employed two distinct convolutional layers: Primary layers focused on the associations of various cancer types and distinct CMEs, contextualizing each cancer's progression within the model. Secondary layers focused on ncRNA-level interactions, such as relations between ncRNA and CMEs, providing detailed insights into their regulatory role in metastasis. Training phase utilized relation-specific embeddings, topology-aware negative sampling, and TuckER-based scoring to enhance the model’s ability to learn multi-relational dynamics. Cross-validation performed and model performance was assessed on the held out test set. Results: The model achieved strong predictive performance across link prediction between ncRNAs and CMEs in different cancers, with an average area under receiver operating characteristic curve of 0.85 on the test set. Predictions showed high accuracy (0.83) and precision (0.81), and recall (0.9). As a practical application, we extracted high scoring interactions between ncRNA and Gastrointestinal cancers from the model that revealed 596 novel ncRNA-CME interactions. Among them, circ_0000745, circ_0000711, circ_0005615, miR-137, and miR-122-5p were the most frequently observed ncRNAs, contributing to multiple CMEs, like invasion and lymph node metastasis across various GI cancers, including gastric, pancreatic and colon cancers. Conclusions: The GNN framework shows the potential of hierarchical modeling, unraveling ncRNA-mediated regulatory networks in cancer progression. By combining advanced deep learning approaches and hierarchicaldata integration, this study bridges critical gaps in studying complex role of ncRNAs in cancer metastasis. The model's top-ranked predicted ncRNA-metastasis interactions offer preliminary data for future experimental validation to identify ncRNAs with potential clinical utility in precision oncology in terms of prognostication and treatment.