Fig. 2

ANN-SCGP model with selective connection showed stronger generalization ability than models with full connection. A Constructing Test_set as discovery set using WGCNA. B & C The correlation of gene similarity of L1 weight matrix and the input matrix increased with it-erations in 1 of the 82 leave-one-out training. D The correlation of gene similarity of L1 weight matrix and the input matrix increased with iterations in 73 of the 82 leave-one-out training. E & F Illustration of constructing SCM via NMF. G RMSE of the 4 models in training. H RMSE of the 4 models in testing. WGCNA, weighted gene co-expression network analysis; RMSE, root mean squared error; NMF, non-negative matrix factorization; SCM, selectively connected matrix