Breast cancer remains one of the foremost causes of cancer-related mortality among women worldwide, driven by a complex interplay of genetic and epigenetic alterations. Identifying both coding and non-coding cancer driver genes is crucial for understanding tumorigenic mechanisms, developing targeted therapies, and improving patient prognostics. Traditional methodologies predominantly focus on coding regions, often overlooking the regulatory roles of non-coding regions in cancer progression. To address this gap, we propose an innovative attention-based graph neural network framework designed to identify both coding and non-coding breast cancer driver genes by leveraging multi-omics data integration and sophisticated attention mechanisms. Our proposed model integrates diverse genomic datasets from The Cancer Genome Atlas (TCGA), including gene expression profiles, mutation data, copy number variations, methylation patterns, and protein-protein interaction networks. We used these datasets to construct a comprehensive heterogeneous gene network containing various types of nodes and edges. To process this network, we utilized a graph convolutional network (GCN) architecture called the heterogeneous graph transformer (HGT). The model captures the intricate relationships and dependencies among genes within the network. Incorporating a self-attention mechanism enables the model to assign different weights to various nodes and interactions, allowing it to focus on the most influential features and effectively filter out noise and irrelevant data, thereby enhancing the identification of critical driver genes that play pivotal roles in cancer development and progression amidst the vast genomic landscape. The framework operates through a two-stage process: (۱) constructing a condition-specific breast cancer network that encompasses both coding genes and non-coding RNAs, and (۲) applying hierarchical attention layers to prioritize nodes based on their significance within the network. This dual approach not only improves the detection of known coding drivers but also uncovers novel non-coding drivers that regulate key oncogenic pathways. Furthermore, the integration of multi-omics data provides a holistic view of the molecular landscape, facilitating the discovery of driver genes with increased accuracy and biological relevance. Comparative analyses show that our proposed model outperforms state-of-the-art methods like CBNA and NIBNA, achieving superior performance in identifying both coding and non-coding drivers. Notably, it predicted a significant number of novel miRNA and coding drivers, many of which have been validated in recent literature. In conclusion, the attention-based graph neural network offers a robust and scalable solution for the comprehensive identification of coding and non-coding breast cancer driver genes. By leveraging multi-omics data integration and advanced attention mechanisms.
