Robust mapping of chromatin loops from sparse or single cell Hi-C data with DeepLoop

Project Abstract Mapping the gene-regulatory chromatin interactions within topologically associated domains (sub- TAD) remains a major challenge in 3D genome research. It is generally believed that multibillion-read sequencing depth are required for Hi-C analysis at kilobase-resolution due to the complex bias structure and severe data sparsity. However, we recently discovered that this is problem can be largely solved computationally without the need for ultradeep-sequencing. We developed a new pipeline named DeepLoop that can robustly identify high-resolution chromatin interactions from low-depth Hi-C data. The conceptual innovation of DeepLoop is to handle systematic biases and random noises separately: we used HiCorr to improve the rigor of bias correction, and then applied deep-learning techniques for noise reduction and loop signal enhancement. Preliminary results showed that DeepLoop significantly improves the sensitivity, robustness, and quantitation of Hi-C loop analyses, and can be used to reanalyze most published low-depth Hi-C datasets. Remarkably, DeepLoop can identify chromatin loops with Hi-C data from a few dozen single cells. These successes motivate us to further optimize, benchmark, simplify and upgrade DeepLoop into a versatile tool for the 3D genome field. Aim 1 will optimize and benchmark DeepLoop performance, improve its compatibility with a variety of different Hi-C protocols, and expand its utility to ultra-resolution analysis. Aim 2 will develop new DeepLoop-based pipelines to enable robust mapping of dynamic chromatin loops at high-resolution, including the identification of homolog-specific loops and loops affected by structure variants. Aim 3 will develop a full-package solution for high- resolution loop analysis of complex tissues with single cell Hi-C, a significant amount of data will be generated in this project as a resource for the scientific community.