AI model detects mutations and predicts biomarkers across 32 cancer types

A single artificial intelligence model has demonstrated that routine cancer biopsy images may contain far more biological information than pathologists can readily see under a microscope. In a study published in The American Journal of Pathology, researchers report that a Vision Transformer trained on whole slide images could simultaneously predict tumor type, the mutation status and RNA expression of the cancer-associated gene TP53, and survival-related outcomes across 32 types of solid cancer. The model’s performance suggests that standard hematoxylin and eosin, or H&E, stained slides could one day serve not only as diagnostic records but also as computational windows into a tumor’s molecular identity.

The findings address a central challenge in modern oncology. Histopathology remains the foundation of cancer diagnosis: tissue is stained, scanned at high resolution, and examined for cellular and structural features that reveal whether a tumor is present and what type it may be. Yet many clinically important characteristics cannot be established reliably from appearance alone. Physicians typically need separate molecular tests to determine whether a tumor carries alterations in genes such as TP53, to measure gene activity through RNA sequencing, or to estimate how aggressively a disease may behave. These assays can be expensive, slow, and difficult to obtain in hospitals with limited laboratory infrastructure.

TP53 is especially important because it encodes p53, a tumor-suppressor protein that helps prevent damaged cells from multiplying. When the gene is mutated, this protective system can fail, allowing abnormal cells to survive, divide, and accumulate further changes. TP53 alterations occur across a wide range of malignancies, although their biological consequences vary by cancer type. Identifying these mutations can assist with disease characterization and research into treatment resistance, but conventional testing requires additional tissue, specialized equipment, and technical expertise. The new model was designed to determine whether visual patterns embedded in an H&E slide could provide an indirect signal of the tumor’s molecular state.

The researchers trained the system using more than 11,000 primary tumor cases from The Cancer Genome Atlas Pan-Cancer Atlas. The dataset paired whole slide images with somatic mutation data, RNA-sequencing measurements, and clinical information. Rather than building an individual algorithm for each tumor or prediction task, the team developed a multi-output model capable of generating seven predictions from the same slide. These included cancer classification, TP53 mutation status, TP53 RNA expression, and clinical outcomes associated with overall survival and progression-free interval. The model therefore attempted to link microscopic tissue architecture with both molecular biology and the course of disease.

At the center of the approach was a Vision Transformer, a deep-learning architecture that analyzes images by dividing them into smaller regions, or patches, and learning how information from those regions relates to the image as a whole. Whole slide images can contain billions of pixels, making it impractical to process an entire slide at full resolution in a single operation. The researchers therefore extracted image patches at sixfold downsampling, corresponding to an approximate magnification of 6.7 times relative to the original slide resolution of about 40 times. The model then used attention mechanisms to estimate which regions contributed most strongly to each slide-level prediction.

This attention-based design also offered a way to inspect what the algorithm considered important. In representative examples, the researchers displayed slide thumbnails, attention overlays, and patches receiving the highest and lowest attention scores. These visualizations did not amount to a definitive explanation of the model’s reasoning, but they provided an indication of where the network was detecting informative morphological patterns. The selected areas could include combinations of tissue organization, nuclear appearance, stromal features, necrosis, or other subtle characteristics that are difficult to translate into a single visual rule. Importantly, the model was not told exactly which pixels contained a mutation-associated feature.

Instead, the team used weakly supervised learning. In a conventional supervised pathology system, experts might label individual cells, tumor regions, or tissue structures at the pixel or patch level. Such annotation is time-consuming, expensive, and vulnerable to disagreement between observers. In this study, the molecular and clinical labels were available at the slide level, while the exact regions containing relevant morphological information were not manually marked. The model learned by comparing many patches within each slide and determining which combinations were associated with the known outcome. This strategy allowed it to search for signals across enormous images without requiring exhaustive expert annotation.

In an independent validation set containing 1,729 slides, the model achieved an area under the receiver operating characteristic curve of 0.766 for detecting TP53 mutations across the 32 tumor types. The AUROC measures how well a classifier separates two categories across different decision thresholds; a value of 0.5 represents chance-level discrimination, while 1.0 represents perfect separation. A score of 0.766 indicates meaningful predictive ability, although it is not sufficient to replace a validated molecular assay. The researchers also found that the system could estimate TP53 RNA expression and identify tumor taxonomy from the same whole slide images, suggesting that tissue morphology carries overlapping signals about genetics, gene activity, and disease classification.

The potential clinical use is therefore more likely to involve screening and prioritization than autonomous diagnosis. In a hospital without immediate access to genomic testing, an algorithm could flag cases that appear likely to harbor a TP53 alteration and help determine which patients should receive confirmatory testing first. It might also help identify slides requiring additional review or provide a preliminary molecular profile when tissue is limited. Such applications could be particularly valuable in remote or under-resourced settings, where sending samples to specialized laboratories can delay treatment decisions. However, the model’s predictions would need rigorous external validation across hospitals, scanners, patient populations, ethnic groups, and variations in tissue preparation before being considered for routine care.

The study also illustrates both the promise and the limits of extracting molecular information from images. A computational association between a visual pattern and a mutation does not necessarily reveal the biological mechanism behind that association. Performance may be influenced by differences in staining, image quality, tumor purity, sampling, or clinical factors that are unevenly distributed across datasets. Survival predictions are especially complex because outcomes depend on treatment, stage, comorbidities, and healthcare access, in addition to tumor biology. The researchers emphasize that their system should complement, rather than replace, molecular testing and clinical judgment. Even so, the work points toward a future in which one digitized pathology slide can support a broad set of diagnostic, molecular, and prognostic assessments, potentially making precision oncology more accessible without requiring a separate test for every question.

News Publication Date: August 13, 2026

Web References: The American Journal of Pathology: https://ajp.amjpathol.org/ ; DOI: https://doi.org/10.1016/j.ajpath.2026.05.008

References: Chaurasia et al., “Predicting TP53 Biomarkers from Whole Slide Images across Human Solid Tumors Using Weakly Supervised Learning,” The American Journal of Pathology, DOI: 10.1016/j.ajpath.2026.05.008