Deep Learning to Identify Transcription Start Sites from CAGE Data

Gene transcription start site (TSS) identification is important to understanding transcriptional gene regulation. Cap Analysis Gene Expression (CAGE) experiments have recently become common practice for direct measurement of TSSs. Currently, CAGE data available in public databases created unprecedented opportunities to study gene transcriptional initiation mechanisms under various cellular conditions. However, due to potential transcriptional noises inherent in CAGE data, in-silico methods are required to identify bonafide TSSs from noises further. Here we present a computational approach dlCAGE, an end-to-end deep neural network to identify TSSs from CAGE data. dlCAGE incorporate de-novo DNA regulatory motif features discovered by DeepBind model architecture, as well as existing sequence and structural features. Testing results of dlCAGE in several cell lines in comparison with current state-of-the-art approaches showed its superior performance and promise in TSS identification from CAGE experiments.