Graph Attention Networks with Local and Global Attention Mechanisms for Learning Single-Shot Omics Data Representations

Abstract

Aiming at the high-dimensional " size " problem in biological omics data where the number of genes is much larger than the number of samples pn, a graph attention network GATOr with local and global attention mechanisms is proposed. The model first calculates the correlation between features using the Pearson correlation coefficient on omics data and constructs a single-sample network of omics data. Then, a graph attention network combining local and global attention mechanisms is proposed to learn graph-based omics feature representation from the single-sample network, thereby converting the high-dimensional characteristics of omics data into low-dimensional representation. Experimental results show that GATOr has achieved better performance in classification task accuracy and other indicators than other traditional classification algorithms.