A systematic evaluation of Hi-C data enhancement methods for enhancing PLAC-seq and HiChIP data

Area of Science:

• Genomics
• Epigenetics
• Computational Biology

Background:

• Chromatin's 3D organization is crucial for gene regulation.
• HiChIP and PLAC-seq (HP) offer cost-effective insights into protein-mediated chromatin interactions.
• Current HP data resolution is limited by sequencing depth, necessitating enhancement methods.

Purpose of the Study:

• To benchmark deep learning models, originally developed for Hi-C data, for enhancing the sequencing depth of HP data.
• To evaluate the performance and transferability of these models across different cell types and experimental conditions.

Main Methods:

• Comprehensive evaluation of existing deep learning models (HiCPlus, HiCNN, etc.) on diverse HP datasets (Smc1a HiChIP, H3K4me3 PLAC-seq, CTCF PLAC-seq).
• Analysis of Pearson correlation and statistical power for detecting chromatin interactions.
• Benchmarking models trained on HP data versus those trained on Hi-C data.

Main Results:

• Most Hi-C models show reasonable performance on HP data, improving statistical power for long-range interaction detection (Pearson correlation 0.76-0.95 within 300 Kb).
• Models specifically trained on HP data outperform those trained on Hi-C data.
• Model transferability across different cell types was observed.

Conclusions:

• Deep learning models can effectively enhance HP data, providing valuable insights into 3D genome organization.
• HP-trained models offer superior performance for sequencing depth enhancement.
• These findings offer guidelines for utilizing existing deep learning tools for HP data analysis.