HiCP2GAN: A Plug and Play Foundation Model-based GAN for Hi-C Enhancement

Area of Science:

• Genomics
• Computational Biology
• Epigenetics

Background:

• Chromatin's 3D organization is crucial for gene regulation and cellular function.
• High-throughput chromosome conformation capture (Hi-C) maps these interactions genome-wide.
• Generating high-resolution Hi-C data is expensive, leading to sparse maps that hinder analysis.

Purpose of the Study:

• To introduce HiCP2GAN, a novel generative adversarial network (GAN) framework for enhancing low-resolution Hi-C data.
• To leverage pretrained Hi-C foundation models as discriminators in GANs for more stable and generalizable results.
• To provide a plug-and-play solution for improving Hi-C data resolution without requiring custom discriminator training.

Main Methods:

• Developed HiCP2GAN, a GAN framework utilizing a pretrained Vision Transformer-based Hi-C foundation model as its discriminator.
• Pretrained the discriminator on a large dataset of 118 million Hi-C patches across diverse species and cell types.
• Adapted the foundation model's encoder as the discriminator backbone, experimenting with fine-tuning strategies.

Main Results:

• HiCP2GAN demonstrated consistent improvements in Hi-C data resolution compared to standalone generators and conventional GANs.
• The framework proved generator-agnostic, allowing integration with various Hi-C resolution enhancement architectures.
• Fine-tuning initial layers of the foundation model while freezing deeper layers preserved essential knowledge for task-specific adaptation.

Conclusions:

• HiCP2GAN offers a powerful and flexible approach to enhance low-resolution Hi-C data, overcoming limitations of existing methods.
• The use of pretrained foundation models as discriminators provides stable and effective adversarial supervision.
• This work facilitates more accessible and comprehensive analysis of chromatin organization and its role in gene regulation.