Fig. 2

Results of state interpretation classifier. a Schematic of machine learning-based automatic classification strategy. b Association of interpretation terms and classifier features. Color indicates mean feature value (standard deviation units). c State classification confusion matrix. Numbers and colors indicate the number of reference states with a particular term assigned to a predicted term by the classifier under leave-one-out cross-validation. Classifications off of the diagonal indicate misclassifications. d Overlap enrichment of reference annotations with our annotations, in the cell types that have a reference annotation. Numbers and colors indicate the enrichment, calculated as the log₂ of the number of bases that overlap between a given reference and new term, divided by the number expected if the states were distributed independently. Note the difference between c and d: c measures whether the interpretation classifier assigns the same term as the reference annotation, for a fixed state, whereas d measures the genomic similarity of two entirely separate genome annotations. That is, the units of c are states and the units of d are base pairs