Modeling circRNAs expression pattern with integrated sequence and epigenetic features identifies H3K79me2 as regulators for circRNAs expression

Circular RNAs (circRNAs) are an abundant class of noncoding RNAs with widespread, cell/tissue specific pattern. Because of their involvement in the pathogenesis of multiple disease, they are receiving increasing attention. Previous work suggested that epigenetic features might be related to circRNA expression. However, current algorithms for circRNAs prediction neglect these features, leading to constant results across different cells. Here we built a machine learning framework named CIRCScan, to predict expression status and expression levels of circRNAs in various cell lines based on sequence and epigenetic features. Both expression status and expression levels can be accurately predicted by different groups of features. For expression status, the top features were similar in different cells. However, the top features for predicting expression levels were different in different cells. Noteworthy, the importance of H3K79me2 ranked high in predicting both circRNAs expression status and levels across different cells, indicating its important role in regulating circRNAs expression. Further validation experiment in K562 confirmed that knock down of H3K79me2 did result in reduction of circRNA production. Our study offers new insights into the regulation of circRNAs by incorporating epigenetic features in prediction models in different cellular contexts.