Aleksandra Galitsyna*
Independent Researcher, previous position: Massachusetts Institute of Technology
galitsyn [at] mit.edu
Abstract
Chromatin in the eukaryotic nucleus is organized into regulatory structures that are visible in chromosome-conformation maps, including domains, loops, and other patterns that remain less systematically characterized. Although 3D genome organization is closely linked to gene regulation, development, and disease-associated misregulation, current approaches often rely on predefined structural classes, limiting the discovery of new chromatin patterns and their connection to underlying regulatory determinants.
Here, we present a general deep-learning framework for unbiased discovery and comparison of chromatin structural patterns in Hi-C and Micro-C data. Rather than searching only for known features, our approach identifies recurrent spatial patterns directly from contact maps and enables their systematic analysis across cell types, developmental stages, and species.
Using this framework, we identify "fountains," a previously undercharacterized chromatin contact pattern visible as a signal extending perpendicular to the main diagonal of Hi-C maps. Fountains are detected independently in zebrafish early embryos at the onset of zygotic genome activation and in mouse dendritic cells, and are reproducible across biological replicates and datasets generated by different laboratories.
By integrating multi-omics data, we show that fountains are associated with active enhancer chromatin rather than promoters. They coincide with enhancer marks, pioneer transcription factor binding, cohesin enrichment, increased chromatin accessibility, and nearby early-transcribed genes. Fountains weaken or disappear after cohesin perturbation, and polymer simulations of loop extrusion reproduce the observed contact patterns, supporting a model in which fountains reflect facilitated cohesin loading at active enhancer regions.
In mouse dendritic cells, fountain-associated cohesin activity is linked to immune-cell activation and transcriptional responses to external stimuli. Loss of cohesin impairs enhancer-gene communication and reduces dendritic cell capacity to respond to pathogens and tumor-related challenges, suggesting that enhancer-associated 3D genome structures contribute to context-specific regulatory programs.
Finally, we apply the pattern-discovery framework to 17 available Hi-C and Micro-C maps of diverse species and identify conserved chromatin structural patterns across evolutionary scales. Together, our results establish a generalizable computational approach for exploratory analysis of 3D genome organization and reveal conserved enhancer-associated chromatin structures that connect genome folding, regulatory activity, and biological function.
Keywords: 3D-genome, Deep learning, Enhancer regulation
Acknowledgement: We thank the laboratory of Prof. Leonid Mirny at MIT for valuable guidance and discussions on the polymer simulations. We are grateful to the laboratory of Prof. Boris Reizis at the University of Chicago, and especially Nick Adams, for providing and supporting the dendritic-cell datasets. We thank the laboratory of Dr. Daria Onichtchouk at the University of Freiburg for providing the zebrafish datasets. We also acknowledge the laboratory of Prof. Mikhail Gelfand, and especially Aleksei Shkolikov, for implementing the neural-network approach central to this study.