Missing-Modality-Aware Graph Neural Network for Cancer Classification

Abstract

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. However, they often struggle with diverse missing-modality patterns and the exponential growth of the number of such patterns as the number of modalities increases. To address these limitations, we propose MAGNET (Missing-modality-Aware Graph neural NETwork) for direct prediction with partial modalities, which introduces a patient-modality multi-head attention mechanism to fuse lower-dimensional modality embeddings based on their importance and missingness. MAGNET's complexity increases linearly with the number of modalities while adapting to missing-pattern variability. To generate predictions, MAGNET further constructs a patient graph with fused multimodal 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/SinaTabakhi/MAGNET.