Research Progress and Frontier Prospects of Chromatin A/B Compartment Classification Methods

Abstract

Accurate classification of chromatin A/B compartments is one of the core issues in three-dimensional genomics, which is critical for understanding how genome spatial organization affects gene regulation, cell fate determination, and disease pathogenesis. Since Hi C technology first revealed in 2009 that the genome is spatially segregated into transcriptionally active A compartments and inactive B compartments, principal component analysis (PCA) based on Hi C contact matrices has long served as the standard strategy for compartment identification. However, with the in-depth advancement of research toward single-cell resolution, cross-cell-type prediction, and the integration of multimodal epigenomic data, the inherent limitations of conventional PCA have become increasingly prominent. In recent years, the introduction of machine learning and deep learning has brought paradigm innovation to chromatin compartment classification. Methodological evolution ranges from early stacked artificial neural networks based on sequence-derived features, to convolutional neural networks utilizing raw DNA sequences, and further to advanced approaches integrating recurrent neural networks, Transformer architectures, graph-theoretic optimization, and interpretable learning, which have substantially expanded the theoretical framework and application scope of compartment analysis. This review systematically summarizes the research progress of chromatin A/B compartment classification methods and covers more than ten representative approaches, including SACSANN, ABCNet, CoRNN, TECSAS, MaxComp, DeepExDC, HiC-SCA, ABCRNet, SCI and CDACHIE. We comprehensively summarize and comment on existing advances from three dimensions: biological principles and experimental detection techniques, machine learning-based methods, and deep learning-driven strategies. Furthermore, we discuss the major challenges currently restricting this field, such as limited cross-cell-type generalization, insufficient model interpretability, the absence of unified benchmarking criteria, and widespread spatial heterogeneity in compartment annotation. Finally, we highlight future research directions, including multimodal data integration, cross-domain adaptation of foundation models, and the translational application of compartment prediction tools to clinical research scenarios such as cancer genomics and developmental biology.