“…On the other hand, as a single histone mark is not a reliable CRM predictor, a great deal of efforts have been made to predict CRMs based on multiple histone marks and chromatin accessibility (CA) data from the same cell/tissue types using various machine-learning methods, including hidden Markov models(Ernst & Kellis, 2012), dynamic Bayesian networks(Hoffman et al, 2013), time-delay neural networks(Firpi, Ucar, & Tan, 2010), random forest(Rajagopal et al, 2013), and support vector machines (SVMs)(Kleftogiannis, Kalnis, & Bajic, 2015). Many enhancer databases have also been created either by combining results of multiple such methods(Ashoor, Kleftogiannis, Radovanovic, & Bajic, 2015; Fishilevich et al, 2017; Zerbino, Wilder, Johnson, Juettemann, & Flicek, 2015), or by identifying overlapping regions of CA and histone mark tracks in the same cell/tissue types(Chen et al, 2020; Cheneby, Gheorghe, Artufel, Mathelier, & Ballester, 2018; Gao & Qian, 2020; Kang et al, 2019; G. Zhang et al, 2018). In particular, most recently, the ENCODE phase 3 consortium(Moore et al, 2020) identified 926,535 candidate cis -regulatory elements (cCREs) based on overlaps between millions of DNase I hypersensitivity sites (DHSs)(Thurman et al, 2012) and transposase accessible sites (TASs)(Buenrostro, Giresi, Zaba, Chang, & Greenleaf, 2013), active promoter histone mark H3K4me3(Aday, Zhu, Lakshmanan, Wang, & Lawson, 2011) peaks, active enhancer mark H3K27ac(Creyghton et al, 2010) peaks, and insulator mark CTCT(Kim et al, 2007) peaks in a large number of cell/tissue types.…”