Abstract
Chromatin immunoprecipitation sequencing (ChIP-seq) experiments targeting histone modifications are commonly used to characterize the dynamic epigenomes of diverse cell types and tissues. However, suboptimal experimental parameters such as poor ChIP enrichment, low cell input, low library complexity, and low sequencing depth can significantly affect the quality and sensitivity of histone ChIP-seq experiments. We show that a convolutional neural network trained to learn a mapping between suboptimal and high-quality histone ChIP-seq data in reference cell types can overcome various sources of noise and substantially enhance signal when applied to low-quality samples across individuals, cell types, and species. This approach allows us to reduce cost and increase data quality. More broadly, our approach -- using a high-dimensional discriminative model to encode a generative noise process -- is generally applicable to biological problems where it is easy to generate noisy data but difficult to analytically characterize the noise or underlying data distribution.
