HiCDiffusion - diffusion-enhanced, transformer-based prediction of chromatin interactions from DNA sequences

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Predicting chromatin interactions from DNA sequence is a key challenge.
• Existing encoder-decoder models generate blurred, artificial Hi-C matrices.
• Deep learning architectures often struggle with the resolution and realism of predicted Hi-C data.

Purpose of the Study:

• To develop a sequence-only model for predicting high-resolution chromatin interaction (Hi-C) matrices.
• To address the blurring and artificiality issues in current deep learning-based Hi-C prediction methods.
• To improve the resemblance of predicted Hi-C matrices to experimental results.

Main Methods:

• Proposed HiCDiffusion, a novel sequence-only diffusion model.
• Utilized an encoder-decoder neural network as a component within the diffusion model.
• Guided the diffusion process using sequence latent representation and encoder-decoder output.

Main Results:

• Achieved high-resolution Hi-C matrices with improved resemblance to experimental data.
• Demonstrated an average 11-fold improvement in Fréchet Inception Distance (FID), with a maximum of 56-fold.
• Obtained classic performance metrics comparable to state-of-the-art encoder-decoder architectures.

Conclusions:

• HiCDiffusion effectively generates realistic and high-resolution Hi-C matrices from DNA sequence.
• The diffusion model approach significantly enhances the fidelity of predicted chromatin interaction data.
• This method offers a promising advancement in computational genomics for predicting 3D genome structure.