VEHiCLE: a Variationally Encoded Hi-C Loss Enhancement algorithm for improving and generating Hi-C data.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Chromatin conformation is crucial for genomic processes.
• Hi-C assays provide insights into chromatin structure but are limited by resolution.
• Current Hi-C resolution limits downstream analysis.

Purpose of the Study:

• To introduce VEHiCLE, a deep learning algorithm for enhancing Hi-C contact map resolution.
• To improve the viability of Hi-C data for detailed genomic analysis.
• To develop a generative model for synthetic Hi-C data.

Main Methods:

• Utilized a deep variational autoencoder and adversarial training.
• Implemented four loss functions: adversarial, variational, insulation, and mean square error.
• Developed a user-tunable, full generative model.

Main Results:

• Achieved state-of-the-art results in Hi-C data enhancement across multiple metrics.
• Provided novel insights into biologically meaningful feature extraction.
• Enabled generation of high-quality synthetic Hi-C data.

Conclusions:

• VEHiCLE significantly improves Hi-C data resolution and utility.
• The algorithm offers advanced capabilities for genomic analysis and data generation.
• VEHiCLE represents a significant advancement in Hi-C data processing and interpretation.