scCAPReSE: an AI trained on just hundreds of examples now reads tumor genome rearrangements cell by cell

By Bioengineer.org · October 29, 2026.

A KAIST team (Kyukwang Kim, Chul-Hwan Lee and Inkyung Jung) has unveiled scCAPReSE, a tool that detects large-scale rearrangements of the tumor genome in single-cell Hi-C data. The work has been published in Genomics & Informatics. Its key feature is that it works with very few training examples: a few hundred.

Hi-C is a technique that measures which regions of DNA physically touch inside the nucleus. When a chromosome breaks and reassembles (deletions, inversions, duplications, translocations), those contacts change and leave a visible footprint on the map. This three-dimensional reorganization can activate oncogenes through "enhancer hijacking" (a regulatory switch that ends up next to the wrong gene), something the linear sequence alone may not reveal.

The starting problem is data scarcity. According to the authors, a large cancer cohort yields on the order of a thousand structural variation patterns, compared with about 60,000 examples in MNIST or CIFAR10 and more than a million in ImageNet. With single-cell Hi-C, moreover, the data are sparse and there is no standard protocol.

Their solution is to repurpose CLIP, a pretrained image-text model, using the Tip-Adapter method. This turns the embeddings (numerical representations) of a handful of example images into the weights of a small two-layer neural network. They removed the text encoder, because CLIP does not know terms such as "Hi-C contact map." The base model was trained on about 600 bulk Hi-C patterns from nine cancer cell lines (including A549, K562 and MCF7). Fine-tuning used 8,184 cells from a sci-Hi-C library, expanded to 284 positive and 304 negative examples through image rotation.

The results, according to the authors themselves: over 90% accuracy on an independent test set from another library of 3,182 cells (33 positives and 112 negatives), with about 20 false positives. Against EagleC, a benchmark ensemble of convolutional networks, on 17 colorectal tumors not seen in training, scCAPReSE achieved 1.03 times higher accuracy. It made considerably fewer calls in normal colon tissue and stem cells, improving precision and recall, although false positives in tumors rose slightly. A 3% gain in accuracy is modest; what matters is the behavior in healthy tissue.

To reach the individual cell, they apply two steps: they aggregate the cells into a "pseudo-bulk" map, detect the rearrangements there and then project the result onto each cell. In the K562 line (chronic myeloid leukemia), they identified the t(9;22) translocation, the Philadelphia chromosome that produces the BCR::ABL1 fusion gene. When compared with another translocation between the same chromosomes, the proportion of associated interactions differed about fourfold at the median, even after correcting for the affected region being 2.5 times larger. In other words, even in a relatively homogeneous line, each cell has a distinct 3D configuration.

Our take: the most interesting part is not the score against EagleC but the underlying pattern. A general-purpose vision model, trained on everyday images, is repurposed for genomics with a handful of data, and runs in one minute on a desktop computer with 4 virtual CPUs and 8 GB of RAM, without a GPU. That efficiency matters because in biology labeled data will always be scarce and expensive; models that learn from few examples are the ones that will reach small labs. The code is freely available on GitHub, which fits with democratizing the tool.

Caution is warranted. This is methods-development work validated on cell lines (HeLa S3, HAP1, K562) and a colorectal cohort, not on primary tumors; the authors themselves note that single-cell multi-omics data from real tumors are lacking to test its clinical relevance. In addition, initial detection depends on being able to pool enough cells.

Even so, the direction is clear. Distinguishing subclones with different oncogenic consequences — the authors point out that clonal enhancer hijacking events have stronger effects than subclonal ones — is a step toward treatments tailored to each tumor. In the short term, this will remain a research tool. In the long term, it is the kind of advance that, cumulatively, brings us closer to treating cancer as a disease that is understood and corrected cell by cell, rather than fought blindly.