HiCNN

HiCNN enhances the resolution of Hi-C contact matrices using a deep convolutional neural network to enable kilobase-level analysis of three-dimensional genome organization.

Key Features:

• Deep Convolutional Neural Network: A 54-layer CNN that learns the mapping between low-resolution and high-resolution Hi-C contact matrices and incorporates global and local residual learning.
• Performance Evaluation: Performance is assessed using mean squared error and Pearson's correlation coefficients, showing consistent improvement over HiCPlus when training and testing datasets derive from the same cell type or different species.
• Cross-Species and Cross-Cell-Type Application: Demonstrated applicability across datasets, including training on GM12878 and testing on K562 and CH12-LX.
• Consistency with Experimental Data: Enhanced Hi-C matrices produce greater consistency with experimental data in identifying statistically significant chromatin interactions compared with HiCPlus.
• Chromatin Loop Recovery: Recovered specific chromatin loops that were validated by 3D-FISH experiments.

Scientific Applications:

• Kilobase-Level 3D Genome Analysis: Enables analysis of three-dimensional genome organization at kilobase resolution without generating high-depth Hi-C datasets.
• Cross-Cell-Type and Cross-Species Comparative Studies: Facilitates applying models trained on one cell type or species to other cell types or species for comparative Hi-C analyses.
• Chromatin Interaction Detection and Validation: Supports identification of statistically significant chromatin interactions and recovery of loops for validation by 3D-FISH.

Methodology:

HiCNN uses a 54-layer deep convolutional neural network with global and local residual learning to map low-resolution to high-resolution Hi-C contact matrices, with performance evaluated using mean squared error and Pearson correlation and compared to HiCPlus using cross-cell-type/species training/testing (e.g., GM12878 → K562/CH12-LX).