 In this paper, we propose a graph sequence neural network, GSNN, for decoding patterns of motor imagery from electroencephalogram, e.g., data in the presence of distractors. GSNN builds subgraphs based on biologically plausible topologies among brain regions to capture both local and global relationships between channels. We also introduce a node domain attention selection network, NASN, which allows us to adjust the connection and sparsity of the adjacency matrix dynamically according to the e.g. signals acquired from different subjects. Experimental results demonstrate that our model outperforms other state-of-the-art methods on the publicly available Berlin distraction dataset. Furthermore, we find that the NASN plays a key role in improving the sensitivity and adaptability of the GSNN model. This article was authored by Xinyuan Kai, Haren Li, Chang Wu, and others.