Cancer’s genetic chaos may soon be readable in the ordinary stained images pathologists already examine under a microscope. A new study published in Nature Biomedical Engineering introduces HisToSpatialCNV, an interpretable deep-learning method designed to predict spatial copy number variations directly from histopathology images. The approach addresses a long-standing challenge in cancer biology: understanding not only which genetic alterations exist inside a tumour, but also where those alterations are located across its complex, heterogeneous tissue landscape. By connecting visual patterns in tissue architecture with genomic abnormalities, the method could help researchers and clinicians study tumour evolution without requiring every region of a specimen to undergo separate molecular testing.
Copy number variations, or CNVs, occur when sections of DNA are gained or lost. Unlike a single-letter mutation, which may affect one position in the genetic code, a CNV can alter large stretches of chromosomes and influence many genes at once. Amplified DNA segments may increase the activity of cancer-promoting genes, while deleted regions can remove genes involved in restraining tumour growth or repairing damaged DNA. These alterations are common across cancer types and can shape how aggressively a tumour behaves, how it responds to treatment and how it evolves over time. Yet CNVs are rarely distributed uniformly. Different patches of the same tumour may carry distinct genomic changes, creating a mosaic of cancer cell populations that is difficult to capture with bulk molecular measurements.
Spatially resolved genomic technologies can reveal this mosaic, but they often require specialised instruments, expensive reagents and substantial quantities of tissue. Histopathology, by contrast, is comparatively accessible and remains central to cancer diagnosis. In a typical workflow, thin sections are stained to reveal nuclei, cytoplasm, connective tissue, blood vessels and other structures, then examined by experts or digitised for computational analysis. The visual information in these slides is not a direct readout of DNA. It reflects the combined effects of cell shape, tissue organisation, necrosis, inflammation, vascular structure and staining variation. HisToSpatialCNV is built around the premise that these features may contain reproducible signals associated with underlying chromosomal gains and losses.
The method uses deep learning, a family of computational models capable of identifying complex patterns across very large images. A digitised pathology slide can contain millions or even billions of pixels, making direct analysis impractical. Systems of this kind generally divide a slide into smaller image regions, extract numerical representations of their visual features and then integrate those representations to produce predictions across the tissue. In the case of HisToSpatialCNV, the objective is not simply to assign one genomic profile to an entire specimen. It is to estimate the distribution of copy number alterations across individual locations, producing a spatial map that can be compared with the tissue’s microscopic appearance.
The word “interpretable” is particularly important. Deep-learning models can be highly accurate while offering little explanation for their decisions, a problem often described as the black-box challenge. For a biomedical tool, a prediction without a biologically meaningful rationale may be difficult to trust or translate into clinical practice. HisToSpatialCNV is presented as an approach that links its predictions to visible regions and structures in the histopathology image. Such interpretability can allow investigators to ask which areas of a tumour drive a predicted amplification or deletion, whether the model is focusing on malignant cells rather than staining artefacts, and how its conclusions correspond to known tissue biology. Explanation does not automatically prove that a model is correct, but it provides a pathway for scrutiny, validation and improvement.
This visual-genomic connection could be especially valuable in tumours with pronounced intratumour heterogeneity. A biopsy samples only a fraction of a cancer, and a bulk assay averages signals from many cells and locations. If one subregion contains a clinically important amplification while another does not, the average result may conceal that difference. A spatial prediction map could help identify genetically distinct territories within a specimen, highlight boundaries between tumour subpopulations and suggest where targeted molecular confirmation should be performed. In research settings, the method could also support studies of how cancer cells interact with their surroundings, including immune infiltration, stromal remodelling and the emergence of treatment-resistant clones.
The proposed strategy does not mean that microscopy can replace genomic testing. Histological appearance is influenced by many factors, and visually similar structures can arise from different molecular causes. Conversely, the same CNV may appear differently depending on tumour type, tissue preparation, staining protocol and imaging platform. A model trained on one collection of slides may therefore encounter difficulties when applied to another hospital, scanner or patient population. Rigorous evaluation, external validation and careful calibration would be essential before predictions could guide treatment. The most realistic near-term role for a system such as HisToSpatialCNV may be as a complementary tool: a way to generate hypotheses, prioritise regions for testing and extend the information extracted from routinely collected tissue.
Interpretability may also make the method useful beyond individual predictions. By revealing associations between chromosomal alterations and microscopic architecture, the system could help generate new biological questions. Researchers might investigate why certain genomic changes coincide with particular patterns of cell density, necrosis or tissue organisation, or whether spatially separated genetic states correspond to different stages of tumour development. Such analyses could turn the pathology slide into more than a diagnostic image: it could become a computationally searchable map of tumour structure and evolution. That possibility reflects a wider shift in biomedical imaging, in which machine-learning systems are being developed not merely to recognise disease, but to infer hidden molecular and cellular states from visual evidence.
HisToSpatialCNV arrives at the intersection of two increasingly connected fields: digital pathology and spatial genomics. Its central promise is technological efficiency combined with biological visibility—using widely available tissue images to estimate where copy number changes may occur, while showing the visual evidence behind those estimates. If validated across cancers, institutions and imaging conditions, this type of model could reduce the molecular blind spots created by limited sampling and help researchers study genomic diversity at the scale of individual tissue regions. For now, the work represents a compelling step toward pathology systems that do more than describe what a tumour looks like. They may also help reveal the genetic geography hidden within it.
Subject of Research: Predicting spatial copy number variations from histopathology images using interpretable deep learning.
Article Title: HisToSpatialCNV: an interpretable deep learning method predicting spatial copy number variations from histopathology images.
Article References: Chen, T., Unjitwattana, T., Guo, X. et al. HisToSpatialCNV: an interpretable deep learning method predicting spatial copy number variations from histopathology images. Nat. Biomed. Eng (2026). https://doi.org/10.1038/s41551-026-01754-z
Image Credits: AI Generated
DOI: https://doi.org/10.1038/s41551-026-01754-z
Keywords: HisToSpatialCNV, spatial copy number variations, CNV, histopathology, digital pathology, deep learning, interpretable artificial intelligence, cancer genomics, spatial genomics, tumour heterogeneity.
