Missing-Modality-Aware Graph Neural Network for Cancer Classification

--A key challenge in learning from multimodal biological data is missing modalities, where all data from some modalities are missing for some patients. Current fusion methods address this by excluding patients with missing modalities, imputing missing modalities, or making predictions directly with partial modalities. T o address these limitations, we propose MAGNET (M issing-modality-A ware G raph neural NET work) for direct prediction with partial modalities, which introduces a patient-modality multi-head attention mechanism to fuse lower-dimensional modality em-beddings based on their importance and missingness. T o generate predictions, MAGNET further constructs a patient graph with fused multi-modal embeddings as node features and the connectivity determined by the modality missingness, followed by a conventional graph neural network. Experiments on three public multiomics datasets for cancer classification, with real-world instead of artificial missingness, show that MAGNET outperforms the state-of-the-art fusion methods. The data and code are available at https://github.com/SinaT ANCER development is a complex process driven by interactions across multiple molecular layers [1]-[3]. To unravel this complexity, cancer research increasingly profiles patients using these molecular modalities, known as multi-omics. Each omics modality provides unique value individually while multimodal fusion can offer complementary insights [4], [5]. Multimodal machine learning approaches integrate these biological modalities to construct a comprehensive patient profile for improving downstream predictive tasks, such as cancer classification and subtyping [6]-[8]. Despite the effectiveness of multimodal biological data fusion, conventional approaches often assume that all omics modalities are available for each patient [9], [10]. However, missing modalities, characterized by structured missingness where all data from some modalities are missing for some patients, are an unavoidable challenge in biomedical applications [11]. For example, some patients may have missing tran-scriptomic profiles due to sample degradation or insufficient RNA quality, while others may lack proteomic data because of cost constraints or technical limitations [12], [13].