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

Area of Science:

• Oncology
• Genomics
• Medical Imaging

Background:

• Kirsten rat sarcoma virus oncogene homolog (KRAS) gene status is crucial for predicting non-small cell lung cancer (NSCLC) treatment response to immune checkpoint inhibitors (ICIs).
• Existing prediction models overlook complex semantic relationships within diverse patient clinical features.

Purpose of the Study:

• To develop an accurate method for identifying KRAS gene status in NSCLC patients.
• To assist physicians in selecting patients likely to benefit from immunotherapy and reduce unnecessary treatments.

Main Methods:

• A multi-view adaptive semantics-aware heterogeneous graph framework (MVASA-HGN) was developed using multimodal medical data (CT images and clinical features).
• The framework employs a two-stage strategy involving single-view graph representation learning and multi-view heterogeneous information fusion with attention mechanisms.
• Node representations are constructed and updated adaptively without predefined meta-paths.

Main Results:

• The MVASA-HGN framework achieved 85.29% accuracy and 89.67% specificity on test datasets.
• The framework demonstrates significant advantages in modeling complex heterogeneous semantics and exploiting rich semantic information from heterogeneous relationships.
• Performance was validated on cooperative hospital and TCIA datasets with ablation and comparison experiments.

Conclusions:

• The MVASA-HGN framework offers a novel approach for multimodal information fusion, linking medical images and genetic information.
• It provides a non-invasive, cost-effective solution for identifying KRAS mutation status in NSCLC.
• The framework has broad application prospects in personalized cancer treatment.