scCAPReSE: detection of large-scale genomic rearrangements from single-cell Hi-C based on few-shot learning

Abstract Large-scale genomic rearrangements are prevalent in cancer genomes and can profoundly rewire three-dimensional (3D) genome architecture, leading to aberrant oncogene activation through enhancer hijacking. The rewired 3D organization generates unique chromatin contact signatures, which can be detected using deep learning-based approaches. However, extending such analyses to single-cell resolution, which is critical to delineate clonal heterogeneity in cancer, remains a major challenge, due to the limited number of training sets as single-cell Hi-C techniques are not standardized and only limited datasets are available across different methods. Here, we introduce scCAPReSE, a few-shot learning-based framework that adopts representations from a pre-trained image foundation model, CLIP, to enable robust classification of structural variation (SV) patterns in single-cell Hi-C data. By extracting and fine-tuning base weights from the foundation model, scCAPReSE enables effective training of deep learning classifiers using only a few hundred large-scale SV examples derived from a single cancer cell line while adapting classification tasks to heterogeneous single-cell Hi-C libraries. scCAPReSE achieved over 90% classification accuracy when evaluated on sci-Hi-C datasets. When further applied to scNanoHi-C data from the K562 chronic myeloid leukemia cell line, scCAPReSE correctly identified the Philadelphia chromosome translocation but also revealed substantial cell-to-cell variability in the contribution of SV-mediated chromatin interactions, highlighting previously inaccessible heterogeneity in cancer 3D genome organization. In summary, scCAPReSE provides a broadly applicable and data-efficient framework for detecting SV-driven 3D genome reorganization at single-cell resolution, enabling quantitative dissection of cancer-specific chromatin architecture and clonal heterogeneity. The developed method is freely available at https://github.com/kaistcbfg/CAPReSE.