MVASA-HGN: multi-view adaptive semantic-aware heterogeneous graph network for KRAS mutation status prediction

In the treatment of advanced non-small cell lung cancer (NSCLC), the mutation status of the Kirsten rat sarcoma virus oncogene homolog (KRAS) gene has been shown to be a key factor affecting the efficacy of immune checkpoint inhibitors (ICIs), which is an important guideline for physicians to develop personalized treatment strategies. However, existing mutation prediction studies have primarily focused on the feature representation of individual patient medical data, ignoring the complex semantic relationships among patients in diverse clinical features. This study aimed to accurately identify KRAS gene status, which will not only assist physicians in accurately screening the patient population most likely to benefit from immunotherapy, but also reduce patient burden by avoiding unnecessary treatment attempts.

A multi-view adaptive semantics-aware heterogeneous graph framework (MVASA-HGN) based on multimodal medical data was developed to accurately predict KRAS mutation status in NSCLC patients. The framework first parses the relational semantics through clinical feature clustering and constructs a heterogeneous graph by combining computed tomography (CT) image and clinical features. In the second step, the heterogeneous graph is split into relational subgraphs under multiple views, and the node representations are constructed and updated gradually through a two-stage strategy of single-view graph representation learning and multi-view heterogeneous information fusion. In the single-view phase, we enhance the node self-embedding and construct the adjacency embedding of neighbors with the same type of relationship to ensure that the relational subgraph under each semantic preserves the complete local structure. Two attention mechanisms are introduced in the multi-view fusion phase to capture the enriched semantics preserved in nodes and heterogeneous relations, respectively. Finally, a comprehensive node representation is obtained through adaptive aggregation of different view neighborhood information and enhanced node embedding without predefined meta-paths.

The classification results were evaluated on cooperative hospitals and The Cancer Imaging Archive (TCIA) datasets, and ablation experiments and comparison experiments were performed on the components of the framework, while exploring the framework's rationality and interpretability. Accuracy reached 85.29% and specificity reached 89.67% on the test set, indicating that our framework has significant advantages in deeply modeling complex heterogeneous semantics in local structures and fully exploiting and utilizing the rich semantic information preserved in heterogeneous relationships. The source code of MVASA-HGN is available at https://github.com/Yangwanter37/MVASA-HGN.

Our proposed MVASA-HGN framework provides a new perspective for multimodal information fusion and creates a new avenue to explore the potential link between images and genes, and the framework provides a non-invasive and cost-effective solution for identifying KRAS mutation status, which has a broad application prospect.