Fig. 6
From: Semi-supervised learning improves regulatory sequence prediction with unlabeled sequences
SNP effect prediction with our semi-supervised model trained from human labeled genome and 22 unlabeled genomes, compared to the baseline model. Allelic imbalance ChIP-seq SNPs were downloaded from ADASTRA database [22]. A Scatter plot of predicted CTCF binding SNP effect versus observed effect with a deep CNN. B Scatter plot of predicted CTCF binding SNP effect versus observed effect with a deep CNN–GNN. C Comparison of Pearson correlation between predicted and observed CTCF binding SNP effects for CNN and CNN–GNN, for 10 different model trainings. D Scatter plot of predicted ESR1 binding SNP effect versus observed effect with a deep CNN. E Scatter plot of predicted ESR1 binding SNP effect versus observed effect with a deep CNN–GNN. F Comparison of Pearson correlation between predicted and observed CTCF binding SNP effects for CNN and CNN–GNN, for 10 different model trainings