Epiphany: predicting Hi-C contact maps from 1D epigenomic signals

Area of Science:

• Genomics
• Computational Biology
• Deep Learning

Background:

• Deep learning models for Hi-C contact map prediction show promise but struggle with generalization across cell types.
• Existing models often fail to capture cell-type-specific differences in 3D genome organization.

Purpose of the Study:

• To develop a novel neural network, Epiphany, capable of predicting cell-type-specific Hi-C contact maps.
• To leverage widely available epigenomic tracks for accurate Hi-C map prediction.

Main Methods:

• Epiphany utilizes bidirectional long short-term memory (LSTM) layers to capture long-range genomic dependencies.
• An optional generative adversarial network (GAN) architecture is incorporated to enhance the realism of predicted contact maps.

Main Results:

• Epiphany demonstrates excellent generalization to unseen chromosomes within and across different cell types.
• The model accurately predicts topologically associating domains (TADs) and specific DNA-protein interactions.
• Epiphany successfully predicts structural genomic changes resulting from epigenomic signal perturbations.

Conclusions:

• Epiphany offers a robust and generalizable approach for predicting cell-type-specific Hi-C contact maps from epigenomic data.
• The model advances the understanding of 3D genome organization and its regulation across diverse cellular contexts.