Skip to main content
Fig. 2 | Genome Biology

Fig. 2

From: An interpretable bimodal neural network characterizes the sequence and preexisting chromatin predictors of induced transcription factor binding

Fig. 2

Preexisting chromatin information improves Bichrom predictions of induced Ascl1 binding sites. a Distribution of model performance (auPRC) in predicting Ascl1 binding sites for a neural network trained using only chromatin data (chromatin-only), a neural network trained using only sequence data (sequence-only), Bichrom using sequence and a ChIP input control experiment (Bichrom control), and Bichrom using sequence and 12 preexisting ES chromatin datasets. The boxplots represent data from 9 independent training sets, each consisting of a separate held-out test chromosome. b The precision-recall curves for 9 models, each tested on 9 distinct held-out test chromosomes. The P-R curves for networks that use a training set in which chromosome 10 is held-out for testing are highlighted in solid lines, performance for other training sets is represented with lighter (alpha = 0.2) traces. Precision at a fixed recall of 0.5 is highlighted for both the sequence-only network and Bichrom

Back to article page