AI Predicts Mismatch Repair Deficiency in Colorectal Cancer

AI Model Predicts Mismatch Repair Deficiency From Both Tumor and Non-Tumorous Regions in Colorectal Cancer

Full Article

by Christos Evangelou, MSc, PhD – Medical Writer and Editor

DNA mismatch repair (MMR) deficiency in colorectal cancer guides immunotherapy decisions and flags potential Lynch syndrome, but standard testing is slow and resource-intensive. A new study from the University of Jyväskylä and the University of Oulu in Finland offers evidence that a deep learning model trained on routine hematoxylin and eosin (H&E) slides can predict MMR status, and that non-tumorous tissue carries detectable predictive information.

The study was published in Computer Methods and Programs in Biomedicine.

Study Rationale

Although testing for MMR status using immunohistochemistry (IHC), polymerase chain reaction (PCR), or next-generation sequencing (NGS) is recommended for patients with colorectal cancer, testing is not always performed, explained Liisa Petäinen, PhD researcher at the University of Jyväskylä and the first author of the study. Barriers to testing include long turnaround times and high labor costs.

Artificial intelligence applied to routine pathology slides could act as a pre-screening step, but most existing deep learning models for MMR prediction focus exclusively on tumor regions of whole-slide images (WSIs) and rely on a single magnification level. Petäinen and her colleagues wanted to test whether this approach was leaving predictive information behind.

“I find the molecular-level features predicted from histopathology really interesting,” Petäinen said. “The central question her team set out to answer: does it matter whether you look at tumor regions only, which part of the slide you tile, and at what magnification?”

Methodology

The researchers evaluated 24 modeling configurations with different tissue source (tumor versus non-tumor), magnification (5x and 20x), and tile embedding strategy. For tile-level feature extraction, they tested a supervised lightweight MobileNetV3 architecture trained on their dataset as well as two label-agnostic digital pathology foundation models (UNI-v2 and HOptimus0). For slide-level classification, the team used multiple instance learning (MIL).

For model development, the researchers used 1,228 WSIs from a Finnish multicenter dataset comprising 264 dMMR and 1,018 MMR-proficient (pMMR) cases and combined 5x and 20x tumor embeddings in parallel MIL branches. External validation was conducted on 1,010 WSIs from Oulu University Hospital and 457 WSIs from the TCGA public database.

Key Findings

On internal validation, models trained on 20x tumor regions performed best, and the top configurations achieved weighted F1 scores above 0.99 and area under the precision-recall curve of 0.997 to 0.999. Non-tumorous regions achieved an F1 of 0.896 and a specificity of 0.982.

“This study highlighted the effectiveness of training based on digital pathology foundation models,” Petäinen noted.

The foundation models, pre-trained on large and diverse pathology image datasets without task-specific labels, appear to capture more transferable feature representations than a supervised model trained on a single institution’s data, she explained.

“The predictive ability of 5x non-tumorous regions was surprising, as the non-tumorous regions were treated as one class including all tissue types,” Petäinen said.

She added, however, that whether specific tissue types within non-tumorous regions drive the effect remains an open question.

On external validation, models using MobileNetV3-derived embeddings, which had performed best on the internal dataset, achieved F1 scores of 0.318 to 0.579. By contrast, the three top-performing models on external validation all used foundation model embeddings. On CRC-OYS-FIN, the best foundation model configurations (UNI-v2 clustering-constrained attention MIL TUM20x, HOptimus0 attention-based deep MIL TUM20x, and UNI-v2 multi-scale attention-based deep MIL) achieved F1 scores of 0.916 to 0.919 with specificity of 0.986 to 0.992. On TCGA, F1 scores reached 0.928–0.934. Sensitivity on external cohorts ranged from 0.500 to 0.682.

The authors attributed the limited generalizability of the models to scanner differences, variations in tissue fixation protocols, and potential differences between the Finnish training cohort and the international TCGA cohort. They also hypothesized that retraining or fine-tuning the MIL layer for each dataset would likely improve the model’s predictive performance.

Potential Clinical Implications

Despite the limited generalizability, the model is a step forward in improving computational efficiency to support the clinical adoption of dMMR prediction models.

“To facilitate the practical application of these models in clinical settings, it is important to improve their ability to make predictions from low magnification, which are more computationally efficient to process,” Petäinen said.

She explained that a model that works at 5x rather than 20x tiles requires processing roughly 16 times fewer pixels, reducing infrastructure demands and turnaround time.

The second implication concerns tissue availability. In resection specimens, the tumor is often easy to delineate, but in biopsy material or cases with extensive necrosis, sufficient tumor regions may not be available for tiling. Models that can draw on non-tumorous tissue offer a potential backup.

“The finding that non-tumorous regions can provide predictive information is important for the clinical applicability of these models,” Petäinen said.

Future Directions

The training cohort in the study came exclusively from Finnish hospitals, which may contribute to the limited generalizability of the model. The dMMR class was underrepresented (roughly one case in five in the training data), and despite balancing strategies, sensitivity on external sets remained lower than specificity.

“What is the role of non-tumorous regions, which specific features or tissue types within these regions are most predictive, and could the incorporation of additional clinical data further improve the model?” Petäinen listed as open questions for future work.

The code from this study has been made publicly available on GitHub, and the TCGA data used for external validation can be accessed through the Genomic Data Commons Data Portal.

This research is part of the Central Finland AI Hub I and II projects, which are co-funded by the European Union.

References

Petäinen L, Väyrynen JP, Böhm J, et al. dMMR prediction from colorectal cancer histopathology: Leveraging non-tumor and low-magnification regions. Comput Methods Programs Biomed. 2026;280:109317. doi:10.1016/j.cmpb.2026.109317

Summary

A recent study from Finnish universities demonstrates that deep learning models can predict DNA mismatch repair (MMR) deficiency in colorectal cancer using routine hematoxylin and eosin (H&E)-stained slides, potentially streamlining clinical workflows. MMR status is critical for guiding immunotherapy decisions and identifying Lynch syndrome, but current testing methods like immunohistochemistry or next-generation sequencing are resource-intensive. The researchers explored whether non-tumorous tissue and lower-magnification images could provide predictive value, addressing gaps in existing models that focus solely on high-magnification tumor regions. The study evaluated 24 modeling configurations, comparing tumor versus non-tumor tissue, magnification levels (5x and 20x), and feature extraction methods, including supervised MobileNetV3 and foundation models (UNI-v2 and HOptimus0). Using 1,228 whole-slide images from Finnish hospitals and external validation datasets, the best-performing models achieved high accuracy (F1 scores up to 0.99) on internal validation, with non-tumorous regions also showing predictive capability (F1 0.896). However, model generalizability was limited on external datasets, likely due to scanner variability and tissue processing differences. Foundation models outperformed supervised approaches in external validation, suggesting better adaptability to diverse datasets. Clinically, the findings highlight the potential for AI-assisted MMR prediction as a cost-effective pre-screening tool. Lower-magnification analysis (5x) could reduce computational demands, while leveraging non-tumorous tissue may improve applicability in cases with limited tumor availability. Future research should explore feature contributions from non-tumorous regions and refine model generalizability through multi-center datasets. The study’s code and data are publicly available, facilitating further development and validation. These advances could enhance efficiency in MMR testing, supporting broader adoption in clinical practice.