Figure 3

Models trained in one cell type can generalize to others. Each model, trained in one cell type, was applied to the chromatin feature datasets from the other two cell types. (A) The GM12878 model achieved high accuracy when applied to K562 features (PCC =0.76), as did the reciprocal cross (PCC =0.75). (B) In each case, predictive accuracy decreased on cross-application but there remains significant agreement between predicted and empirical values. Acc., accuracy; AUROC, area under the receiver operating characteristic curve; PCC, Pearson correlation coefficient; RMSE, root mean-squared error.