Fig. 2

Gene embedding method based on the CGE bipartite graph with the signed, weighted adjacency matrix W. The rows of W represent regulating genes (for which embeddings are constructed), and columns are the genes of the downstream expression signature. The edge signs \(s_{ij}\) are positive for upregulation, negative for downregulation, and \(N_i\) is the total number of genes that are regulated by gene i. It is \(s_{ij}=0\) if there is no edge. Gene embeddings are computed using a low-rank approximation of the singular value decomposition of W, \(W\approx U \Sigma V^T\), where the matrix U projects one-hot encoded vectors representing single genes onto K-dimensional embedding vectors