A/B Compartment Prediction

A/B compartments reflect the physical separation of chromatin into two mutually exclusive spatial territories within the nucleus. Active 'A' compartments localize to the nuclear interior and are enriched for transcription, while inactive 'B' compartments associate with the nuclear lamina and periphery. This segregation is visible in Hi-C contact maps as a characteristic 'plaid' pattern, where A-A and B-B interactions are enriched and A-B interactions are depleted. The compartment identity of a genomic region is strongly correlated with its gene density, GC content, and replication timing.

Compartment identity is computationally determined using principal component analysis (PCA) on normalized Hi-C contact matrices. The first principal component (PC1), or eigenvector, corresponds to the A/B compartment profile. Key steps include:

• Computing the observed/expected contact matrix to remove distance-dependent background
• Performing PCA on the correlation matrix of this normalized data
• Assigning compartment identity based on the sign of PC1 values
• Using gene density or GC content to orient the eigenvector, ensuring A compartments have positive values This method transforms raw interaction frequencies into a continuous compartment score for every genomic bin.

A and B compartments exhibit diametrically opposed functional landscapes:

• A Compartments: Enriched for housekeeping genes, high transcriptional activity, open chromatin (DNase I hypersensitivity), active histone marks (H3K36me3, H3K27ac), and early replication timing
• B Compartments: Enriched for gene deserts, late replication timing, repressive histone marks (H3K9me3, H3K27me3), and lamina-associated domains (LADs) This functional polarization makes compartment prediction a powerful tool for annotating regulatory potential directly from 3D interaction data, even in the absence of orthogonal epigenomic assays.

A/B compartment identity is not static; it undergoes dynamic reorganization during cellular differentiation and development. Genomic regions can switch from B to A compartments upon gene activation and vice versa during silencing. Notable examples include:

• The Hox gene clusters transitioning from B to A during embryonic body plan specification
• Olfactory receptor loci switching compartments during neuronal maturation
• Pluripotency genes moving to B compartments upon lineage commitment These switches are often accompanied by changes in histone modifications and DNA methylation, making compartment prediction a readout of large-scale epigenomic remodeling.

Modern approaches use convolutional neural networks and graph neural networks to predict A/B compartments directly from DNA sequence and epigenomic features. Models like Akita and DeepC learn to map sequence motifs to compartment identity by:

• Encoding DNA sequence as one-hot vectors or learned embeddings
• Integrating CTCF binding sites, histone modification ChIP-seq signals, and chromatin accessibility data as auxiliary inputs
• Predicting a continuous compartment score for each genomic bin
• Training on Hi-C-derived PC1 values as ground truth This enables compartment annotation for any genome with reference sequence, bypassing the need for expensive Hi-C experiments.

A/B compartments exist at a higher hierarchical level than topologically associating domains (TADs) and chromatin loops. A single compartment typically contains multiple TADs, and TAD boundaries can exist within both A and B compartments. However, the relationship is not purely hierarchical:

• TADs within A compartments tend to have higher internal interaction frequencies and more active regulatory landscapes
• TADs within B compartments are often more compact and silenced
• CTCF binding sites at TAD boundaries can act as barriers that prevent the spreading of one compartment type into another Compartment prediction thus provides the broad organizational context within which finer-scale folding features operate.