MG-GeneNet: A Multi-Relational Graph Convolutional Network with Adaptive Genetic Feature Selection for Myasthenia Gravis Classification and Biomarker Discovery

Abstract

Myasthenia gravis (MG) is an autoimmune disease and characterized by defective communication between nerves and muscles, thus causing fluctuating muscle weakness. Accurate disease classification and relevant molecular biomarker detection remain a difficult task owing to the complexity of gene-expression data with high dimensionality. This paper presents MG-Gene Net, a multi-relational graph convolutional network with adaptive genetic feature selection for MG classification and biomarker detection. MG-Gene Net is evaluated on the GSE85452 gene-expression dataset, which is publicly available and comprised of molecular data from MG patients and comparison subjects. The gene-expression data are preprocessed and used to construct a multi-relational gene network. A relational graph convolutional network is used to learn the relation features, and adaptive genetic algorithms are employed to select promising gene subsets for classification and candidate biomarker detection. The approach is evaluated in terms of classification accuracy, feature-selection stability, and candidate biomarker semantics.