Lin, MY., Lo, YC. & Hung, JH
The three-dimensional organization of chromatin is essential for gene regulation and cellular function, with epigenome playing a key role. Hi-C methods have expanded our understanding of chromatin interactions, but their high cost and complexity limit their use. Existing models for predicting chromatin interactions rely on limited ChIP-seq inputs, reducing their accuracy and generalizability. In this work, we present a computational approach, EpiVerse, which leverages imputed epigenetic signals and advanced deep learning techniques. EpiVerse significantly improves the accuracy of cross-cell-type Hi-C prediction, while also enhancing model interpretability by incorporating chromatin state prediction within a multitask learning framework. Moreover, EpiVerse predicts Hi-C contact maps across an array of 39 human tissues, which provides a comprehensive view of the complex relationship between chromatin structure and gene regulation. Furthermore, EpiVerse facilitates unprecedented in silico perturbation experiments at the “epigenome-level” to unveil the chromatin architecture under specific conditions. EpiVerse is available on GitHub: https://github.com/jhhung/EpiVerse.
a Epigenetic signal and DNA sequence generation: representation of the imputation process using the Avocado model, which imputes a wide array of epigenetic signals, and their integration with a one-hot encoded DNA sequence to provide detailed epigenome information. b Diagonal extraction algorithm: A specialized approach used by HiConformer to enhance the receptive field of features for improved prediction of long-range interactions in Hi-C data, capturing both individual and contextual interaction features across a 0.5 Mb region. c HiConformer model: The multi-task CNN Transformer framework within EpiVerse, which processes and integrates epigenetic signals and DNA sequence data for the imputation of ChromHMM states and Hi-Ccontact maps, optimizing the prediction of Hi-C and providing biological annotations. d MIRNet framework: The denoising component of EpiVerse that applies real image restoration techniques to Hi-C data diagonals, enhancing detail and clarity, and reducing biases from the data and imputation models.
