A Dresden-led research team has developed an artificial intelligence (AI) model that simultaneously predicts multiple genetic alterations in colorectal cancer from standard pathology slides, offering a potential new tool for accelerating diagnostics and guiding treatment strategies.
The multicenter study, published recently in The Lancet Digital Health, analyzed nearly 2,000 digitized tissue slides from colon cancer patients across 10 independent cohorts in Europe and the United States.
Researchers at Dresden University of Technology (TU Dresden) and their collaborators created a multi-target transformer model capable of detecting a broad range of genetic alterations from hematoxylin and eosin-stained tissue samples. Unlike earlier AI systems, which were typically trained to identify single mutations in isolation, this model recognized shared morphological patterns that signal multiple, co-occurring mutations.
“Earlier deep learning models and analyses of the underlying tissue alterations have generally focused on only a single mutation at a time,” said Marco Gustav, first author of the study and researcher at TU Dresden. “Our new model can identify many biomarkers simultaneously, including some not yet considered clinically relevant.”
One important insight from the study was that several mutations occur more frequently in microsatellite-instable (MSI) tumors. MSI is a key biomarker in colorectal cancer and is used to identify patients who may benefit from immunotherapy. The team’s findings suggest that groups of mutations collectively shape tissue morphology in ways that can be detected by AI.
For diagnostics, this approach could help reduce dependence on sequencing-based testing, which is accurate but costly and time-consuming. Whole-exome sequencing typically costs between US $1,292 and $3,594 per patient, and whole-genome sequencing can reach over $10,000 depending on depth and coverage.
Currently, turnaround times for molecular biomarker testing are often seven to ten days, with delays common in multi-site clinical settings. By comparison, an AI model that extracts biomarker signatures directly from routine pathology images could serve as a rapid pre-screening tool, identifying which patients most need molecular confirmation and potentially shortening time to treatment decisions.
The implications extend to drug discovery and development. Broad molecular testing has been shown to improve clinical trial enrollment in colorectal cancer, with one study reporting trial participation of eight percent for patients who received next-generation sequencing (NGS) versus 1 percent for those tested with limited panels. AI systems that link tumor morphology with underlying mutations could further refine trial design by uncovering subgroups with distinct mutational profiles, improving patient stratification, and helping identify new biomarker candidates for targeted therapies and immunotherapies.
The new model was benchmarked against established single-target approaches and performed as well or better at predicting several alterations, including BRAF (B-Raf Proto-Oncogene) and RNF43 mutations and MSI status, directly from pathology slides.
According to Jakob N. Kather, a co-author on the work at TU Dresden, the work highlights the dual role of AI in both research and clinical care. “Our research shows that AI models can significantly accelerate diagnostic workflows,” he said. “At the same time, these methods provide new insights into the relationship between molecular and morphological changes in colorectal cancer.”
