Chromatin Capture Upsampling Toolbox - CCUT: A Versatile and unified Framework to Train Your Chromatin Capture Deep Learning Models

Chromatin Capture Experiments such as Hi-C and Micro-C have become popular methods for genome architecture exploration. Recently, also a protocol for long read sequencing, Pore-C, was introduced, allowing the characterization of three-dimensional chromatin structures using Oxford Nanopore Sequencing Technology. Here, we present a framework that focuses on the efficient reconstruction of low-resolution Pore-C data but can also process all other 3C data, such as Hi-C and Micro-C matrices, using models that can be trained on a consumer GPU. Furthermore, we integrate building blocks of popular super-resolution methods such as SWIN-Transformer or residual-in-residual-blocks to modify or build customized networks on the fly. Pre-built models were trained and evaluated on multiple publicly available gold-standard Micro-C and Pore-C datasets, allowing for fine-scale structure prediction. Our work aims to overcome the drawback of high sequencing costs to construct high resolution contact matrices, as well as the problem of mapping low-coverage libraries to high-resolution structures in the genome. Although there have been major breakthroughs regarding NGS-based methods for the reconstruction of high-resolution chromatin interaction matrices from low-resolution data, for data obtained by long-read sequencing, there is currently no solution to reconstruct missing and sparse information and to improve the quality. Availability The tool is available at ( https://github.com/stasys-hub/CCUT )