Chrombus is generalizable among different cell linage and species. (A) Pearson correlation coefficients between Hi-C scores and Chrombus prediction scores from 5-fold cross-validation crossing six cell lines, GM12878, K562, IMR90, HeLa-S3, HCT116, and CH12. For each round of validation, the model parameters are trained by one cell line and tested by another. (B) ROC curves for cross-cell-line classification of within- and between-TAD chromatin interactions using predicted scores and the true Hi-C scores by models trained on different cell lines. The predictive power of Chrombus retained in different cell lines. (C) Comparison of the interaction scores by Chrombus and Hi-C for segment pairs encompassing eQTL interaction and those do not in GM12878 cell line (upper) and fold-of-enrichment for known eQTL–eGENE pairs corresponding to different thresholding of the predicted scores (lower). In both cases, the model parameters were trained by data from a different cell line. (D) Identifying the enhancer–gene interactions that are specific to the GM12878 and K562 cell lines. These interactions are used for the analysis in (E) and (F). (E) The scores based on Chrombus prediction and Hi-C for CTCF-segment pairs encompassing known enhancer–gene interactions and those that do not. (F) Fold-of-enrichment for known enhancer–gene interactions corresponding to different thresholding of the predicted scores. In all cases, the predictions for each given cell line were based on a model in which the parameters were trained by a different one.