Fig. 1

BOM accurately classifies mouse embryonic context-specific CREs.

a Framework of BOM. First, we defined at least two sets of CRE sequences specific to different cell types or conditions. Then, we identify TF binding motif instances within these sequences. Vertebrate motifs from GimmeMotifs were used to annotate CREs. The model is trained to classify cell states via XGBoost for binary or multiclass classification. SHAP values are calculated to explain the importance of TF binding motifs for the classification task. b Overview of the mouse E8.25 snATAC-seq dataset from ref. ²⁵. UMAP shows the cell types identified in the mouse embryos and used in our analyses. Notochord cells were excluded because of the low number of CREs. c Receiver operating characteristic (ROC) for the binary classification of each of the cell types in the mouse E8.25 embryonic dataset. The mean value of AUC is shown. d Summary of the prediction statistics of binary BOM models trained to classify cell-type-specific CREs for 17 cell types (auROC area under the receiver operating characteristic (ROC) curve, Acc accuracy, auPR area under the precision recall (PR) curve, F1 F1 score, MCC Matthews correlation coefficient). The error bars represent standard deviation. e Recall and specificity for the classification of cell type-specific CREs and CRE-flanking regions as the negative class f Summary of the prediction statistics of BOM models trained to classify cell type-specific and pleiotropic CREs across 18 cell types. Error bars are standard deviation. g ROC curves (left) and precision-recall curves (right) for the classification of mouse E8.5 CREs using the models trained on mouse E8.25 CREs. Source Data are provided as a Source Data file.