Cancer genomes are frequently scrambled by large-scale structural variations — deletions, inversions, duplications and wholesale chromosomal translocations — that do far more than rearrange the linear sequence of DNA. These rearrangements can profoundly rewire the three-dimensional architecture of the genome, rerouting regulatory contacts and, through enhancer hijacking, switch on oncogenes in ways that sequence-level readouts may miss. A new computational method called scCAPReSE, published in Genomics & Informatics, now aims to detect such rearrangements in single-cell Hi-C data using a strategy deliberately built for a situation where training data is scarce.
Hi-C has long been an attractive complement to whole-genome sequencing for spotting structural variations. Rather than reading breakpoints directly, Hi-C captures the physical contacts that chromatin makes across the genome, so a rearrangement shows up both as an aberrant contact pattern and as a rewired three-dimensional organization that can be visualized directly. Early Hi-C-based detection tools relied on statistical frameworks for identifying aberrant contact patterns; more recent approaches have turned to deep learning. But the deep learning route has a bottleneck: these classifiers are data-hungry, and cancer Hi-C data matched to whole-genome sequencing — needed to validate structural variation calls — remains limited. A large cancer cohort yields on the order of only a thousand large-scale structural variation patterns, a pittance compared with the roughly 60,000 training images in classic benchmarks such as MNIST and CIFAR10, or the more than one million images in ImageNet.
The shortage becomes acute at single-cell resolution, which is precisely the level where cancer researchers most want to look. Delineating clonal heterogeneity — the coexistence of genetically and epigenetically distinct tumor cell populations shaped by sequential subclonal selection — requires examining individual cells, yet single-cell Hi-C protocols produce limited data and lack a unified standard, making conventional deep learning training effectively infeasible. Kyukwang Kim, Chul-Hwan Lee and Inkyung Jung at KAIST addressed this by borrowing from the world of image foundation models. Their framework, named single-cell chromatin anomaly pattern recognition and size estimation (scCAPReSE), adopts representations from CLIP, a pre-trained image-text model, and fine-tunes them for the specific task of classifying structural variation patterns in sparse single-cell contact maps.
The technique at the core of scCAPReSE is the Tip-Adapter method, a few-shot learning approach that converts CLIP image embeddings into the weights of a small two-layer neural network. Raw CLIP, which normally performs zero-shot classification by comparing image embeddings with text embeddings, could not be used as-is: the model has no pre-training for terms like “structural variation” or “Hi-C contact map,” so zero-shot classification would fail. The authors therefore removed the text encoder step entirely and instead used embeddings of a small set of example images as the first layer’s weights, with one-hot class labels as the second layer’s weights. Because structural variation detection is a binary classification problem, they also modified the method’s normalization to allow unbalanced numbers of positive and negative examples, and dropped the usual hyperparameter search. An image-processing pipeline, adapted from a previously established Hi-C structural variation detector, first scans contact maps for candidate regions, extracting breakpoint-centered crops that are converted into grayscale images with additional channels encoding low-coverage regions and high-intensity pixels.
To build the base weights, the team assembled bulk Hi-C data from nine cancer cell lines with whole-genome-sequencing-confirmed structural variations, including A549, K562 and MCF7, plus negative examples from normal cell lines and peripheral blood T-cells, all processed at 40-kilobase resolution. The resulting weight matrix was trained on approximately 600 bulk Hi-C structural variation patterns. Fine-tuning then used single-cell combinatorial indexed Hi-C (sci-Hi-C) data: crops from one library with 8,184 cells served as training data — expanded to 284 positive and 304 negative examples through 90-degree rotation-based image augmentation — while crops from another library with 3,182 cells provided an independent test set of 33 positive and 112 negative examples. After 25 epochs of optimization with stochastic gradient descent, cross-entropy loss fell roughly 3.5-fold, and embeddings of the two classes, previously overlapping in UMAP space, separated into distinct, distal clusters. On the held-out test set, the method achieved over 90 percent classification accuracy, recovering nearly all true structural variation patterns with only about 20 false positives.
The method was also benchmarked against EagleC, a state-of-the-art convolutional neural network ensemble classifier, on colorectal cancer data that had not been used in training: 17 tumor tissue Hi-C datasets with matched whole-genome-sequencing-derived structural variation calls, plus 10 normal colon and six differentiating stem cell Hi-C samples. scCAPReSE showed 1.03-fold higher accuracy than EagleC, though false positives on the tumor test set rose slightly. Crucially, it made far fewer calls in normal colon tissue and stem cell data, raising overall precision and recall. The authors attribute this to their approach’s resistance to clutter-driven misclassification.
Applying the framework directly to individual cells was not viable, because the sparse, low-depth contact maps of single cells often render structural variation signatures indistinguishable. The team instead designed a two-step strategy: single-cell contact maps are first aggregated into a pseudo-bulk map on which structural variation patterns are robustly detected, and the results are then projected back to individual cells by examining interaction signatures within the identified cancer-specific contact regions. This worked on scNanoHi-C data from the K562 chronic myeloid leukemia cell line, which carries the well-characterized t(9;22) Philadelphia chromosome translocation that generates the BCR::ABL1 fusion gene. scCAPReSE correctly identified the translocation from the pseudo-bulk map, and breakpoint-crossing interaction signals — the newly formed contacts between the rearranged chromosomes — were clearly visible.
The single-cell projection then revealed something bulk analyses could not. Because sequencing depth varied across cells, the authors normalized each cell’s structural variation-associated interactions by its total number of chromosome 9–22 inter-chromosomal contacts. After normalization, cells showed a broad distribution of contribution values, with a median of approximately 0.4. When the team applied the same analysis to a second, WGS-supported translocation between the same two chromosomes, a striking difference emerged: although both rearrangements generate interaction patterns on the same contact map, the ratio of structural variation-involved interactions differed substantially between them — even after accounting for the fact that the Philadelphia chromosome-inducing event altered a genomic area 2.5 times larger, a roughly four-fold difference in median interaction ratio and about two-fold difference in dispersion remained. Individual cells within this relatively homogeneous cancer cell line, in other words, exhibit heterogeneous three-dimensional genome configurations at structural variation-mediated de novo contacts.
The authors emphasize that this cell-to-cell variability was previously inaccessible; their earlier work could infer clonality of structural variation-mediated three-dimensional genome alterations only by integrating bulk Hi-C with whole-genome sequencing. Resolving such patterns at single-cell resolution matters because distinct subclones can carry structural variations with markedly different oncogenic consequences, and clonal enhancer hijacking events have been shown to exert stronger oncogenic effects than subclonal or passenger rearrangements — context that bulk measurements obscure. The method is also practical: the example training and testing code runs in about one minute on a desktop virtual machine with four virtual CPUs and 8 gigabytes of RAM, no GPU required, allowing users to rapidly retrain base weights when data depth or experimental technique differs. The software is freely available on GitHub.
Important limitations remain. The study is a methods development effort validated on cell lines — HeLa S3, HAP1 and K562 — and a colorectal cancer cohort rather than primary tumors, and the authors note that current shortages of multi-omics single-cell Hi-C data from primary tumors restrict immediate validation of the framework’s broader clinical relevance. Accuracy on the tumor benchmark test set came with a slight increase in false positives, and the pseudo-bulk step means the initial detection depends on pooling enough cells. Even so, the demonstration that a few hundred examples suffice to train a robust single-cell structural variation classifier — by repurposing visual representations learned from millions of generic images — offers cancer genomics a data-efficient path toward quantitatively dissecting structural-variation-driven chromatin architecture and clonal heterogeneity, one cell at a time.
Subject of Research: A few-shot learning method for detecting large-scale genomic rearrangements from single-cell Hi-C data
Article Title: scCAPReSE: detection of large-scale genomic rearrangements from single-cell Hi-C based on few-shot learning
Article References: Kim, K., Lee, C.-H., & Jung, I. (2026). scCAPReSE: detection of large-scale genomic rearrangements from single-cell Hi-C based on few-shot learning. Genomics & Informatics, 24(1), Article 8. https://doi.org/10.1186/s44342-026-00069-4
Image Credits: AI Generated
DOI: 10.1186/s44342-026-00069-4
Keywords: single-cell Hi-C, structural variation, few-shot learning, CLIP, cancer genomics, 3D genome organization, Philadelphia chromosome, clonal heterogeneity, enhancer hijacking, K562, deep learning, pseudo-bulk
