X-CNN

X-CNN enhances the resolution of chromatin interaction mapping from Hi-C data by integrating DNase-seq and ChIP-seq signals and applying a convolutional neural network to localize high-resolution interaction sources.

Key Features:

• Hi-C resolution enhancement: Resolves interaction sources at finer scales than typical Hi-C coarse resolutions (5–25 kb).
• Multi-omics integration: Integrates DNase-seq and ChIP-seq signals targeting transcription factors and histone modifications with Hi-C data.
• CNN classification: Trains a convolutional neural network to distinguish Hi-C interactions from non-interactions.
• Feature attribution: Applies a feature attribution method to predict precise high-resolution sources for each identified Hi-C interaction.
• Validation by aggregation: Predicted high-resolution sources recover original Hi-C peaks when aggregated to coarser resolutions.
• Biological enrichment: Predicted sources are significantly enriched for evolutionarily conserved bases, expression quantitative trait loci (eQTLs), and CTCF motifs.

Scientific Applications:

• Fine-mapping chromatin interactions: Pinpoints candidate transcription factor binding sites and open chromatin regions underlying Hi-C contacts.
• Genome architecture and regulation: Enables analysis of chromatin organization and regulatory implications on gene expression at higher resolution.
• Genetic variant interpretation: Supports linking noncoding variants, including eQTLs, to putative interaction sources.

Methodology:

Integrates DNase-seq and ChIP-seq with Hi-C data, trains a convolutional neural network to classify interactions versus non-interactions, and applies a feature attribution method to assign high-resolution sources, with validation by aggregation to coarser Hi-C resolutions and enrichment testing for conserved bases, eQTLs, and CTCF motifs.