Abstract Proper control of replication timing (RT) is of vital importance to maintain genome and epigenome integrity. However, the genome-wide sequence determinants regulating RT remain unclear. Here, we develop a new machine learning method, named C oncert , to simultaneously predict RT from sequence features and identify RT-modulating sequence elements in a genome-wide manner. C oncert integrates two functionally cooperative modules, a selector, which performs importance estimationbased sampling to detect predictive sequence elements, and a predictor, which incorporates bidirectional recurrent neural networks and self-attention mechanism to achieve selective learning of longrange spatial dependencies across genomic loci. We apply C oncert to predict RT in mouse embryonic stem cells and multiple human cell types with high accuracy. The identified RT-modulating sequence elements show novel connections with genomic and epigenomic features such as 3D chromatin interactions. In particular, C oncert reveals a class of RT-modulating elements that are not transcriptional regulatory elements but are enriched with specific repetitive sequences. As a generic interpretable machine learning framework for predicting large-scale functional genomic profiles based on sequence features, C oncert provides new insights into the potential sequence determinants of RT.

Concert: Genome-wide prediction of sequence elements that modulate DNA replication timing