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

In a recent study, researchers at the University of Zurich and the University of Oxford developed a new artificial intelligence (AI) tool to determine DNA mismatch repair (MMR) status in colorectal cancer. The tool achieved 98% accuracy in identifying MMR deficiency across two large clinical trial cohorts of patients with colorectal cancer and could reduce pathologist workload by correctly classifying over three-quarters of cases without human review.

The report was published in Cell Reports Medicine.

Need for Efficient MMR Testing Methods

Colorectal cancer is the third most common cancer worldwide. Approximately 10%–15% of cases display MMR deficiency (MMRd), which leads to increased mutations and is associated with enhanced responsiveness to immunotherapy.¹ Current clinical guidelines mandate testing all patients with newly diagnosed colorectal cancers for MMRd, but this process requires specialized pathologists to examine tissue samples stained for several MMR proteins — a time-consuming procedure that can delay critical treatment decisions.

“Most diagnostic pathology labs determine MMR status in colorectal cancer by pathologist review of immunostained slides. This is costly and subject to inter-observer variability,” explained Dr. David Church, clinician scientist fellow at the University of Oxford and the corresponding author of the study. “AI-based methods offer a powerful alternative, which we explored in this study.”

Solution: Training an AI-powered MMR

The research team developed an AI-based method to detect MMRd — termed AIMMeR — to analyze immunohistochemically stained tumor tissue samples at the single-cell level. The system employs deep learning to classify individual cells and determine their expression of four key DNA repair proteins: MLH1, MSH2, MSH6, and PMS2.¹

Dr. Church highlighted the advantages of this method:

“Existing AI-based methods for detecting MMRd in colorectal cancer have mainly applied deep learning approaches to hematoxylin and eosin-stained sections. While these have shown impressive performance, their use at a sensitivity of 95%, which arguably is below the ideal for clinical application, is only able to exclude 50%–60% of cases from the need for immunohistochemistry and pathologist review. While AIMMeR requires immunostaining of all samples, its performance means that the proportion of tumors requiring pathologist review is overall lower.”

The research team used tissue samples from two large clinical trials: the SCOT trial, comprising 2,352 colorectal cancer cases, and the QUASAR2 trial, including 1,195 cases.¹ The researchers first developed a method to classify cells by their nuclear morphology, distinguishing between epithelial cells, stromal cells, and lymphocytes with 92% accuracy against pathologist ground truth. According to the authors, this step prevents the misclassification of non-cancerous cells that retain MMR protein expression.

AIMMeR was then trained to identify and quantify protein expression in individual cells, calculate the percentage of MMR-expressing cells in tissue samples, and determine overall MMR status based on protein expression patterns.¹

Model Validation

In the training phase, the model was used to analyze 38,113,216 individual cells across 2,015 tumors from patients in the SCOT trial.¹ Examination of MMR protein expression patterns in tumors revealed a strong correlation between related protein pairs (MLH1-PMS2 and MSH2-MSH6). The model achieved a high accuracy against pathologist consensus in determining MMR protein expression, providing an area under the receiver-operator curve (AUROC) of 0.98, sensitivity of 98%, and specificity of 75%.¹

The research team then validated the model in an independent cohort of 965 patients in the QUASAR2 trial. In the validation cohort, the model achieved 98% accuracy against pathologist consensus, with 95% sensitivity and 91% specificity.¹ When used to assess microsatellite instability (MSI), the model achieved an AUROC of 0.86 in the whole cohort.

In addition, the model demonstrated high accuracy in identifying combined MLH1-PMS2 loss (98% positive predictive value) and correctly classified 83% of cases as MMR-proficient without a pathologist review.¹

“AIMMeR AUROC equals or betters other methods for identification of MMRd, and more importantly translates to higher positive and negative predictive value for MMR deficiency, at least in the clinical trials cohorts we have used for training and validation thus far,” noted Dr. Church. “Real-world diagnostic accuracy and correlation with patient outcomes will be the focus of future studies, which we are planning.”

Prognostic Power

The study also revealed insights about the prognostic value of MMR deficiency in patients treated in the SCOT trial. Patients with MMRd tumors, as determined using AIMMeR, showed favorable survival outcomes.¹ The prognostic value was independent of patient gender and disease stage. A stronger prognostic value was observed in younger patients and those with right-sided tumors.

Despite the significant prognostic value of MMR status, MMRd did not predict differential benefit from longer versus shorter chemotherapy duration. However, the team found an interesting correlation between MMRd and outcomes based on chemotherapy type (CAPOX vs. FOLFOX).¹

According to Dr. Church, these findings have potential implications for personalizing treatment approaches.

“Our analysis of the SCOT trial confirmed that the favorable prognosis associated with MMR deficiency in localized colorectal cancer is detectable in oxaliplatin-treated patients,” said Dr. Church. “While this will not change management, it may provide reassurance to clinicians and patients who are receiving 3 months of adjuvant chemotherapy for MMR deficient disease.”

Future Implementation and Next Steps

The researchers acknowledged that the study has limitations, including the use of tissue microarrays rather than whole tumor slides, the potential impact of immunostaining quality on accuracy, and the absence of genetic sequencing data to confirm the causes of single protein loss. Future studies are needed to further validate the model in diagnostic biopsies.

Looking ahead, Dr. Church outlined the path to clinical implementation:

“Our hope is that AIMMeR will find a role in the clinical workup of colorectal cancer. This could be as a standalone test for MMR deficiency, or following first-line deep learning based analysis of H&E sections.” However, he emphasized the need for further validation: “Validation is required before AIMMeR can be implemented in practice. As well as undertaking prospective evaluation in real-world cohorts, which we are doing in our center, we aim to determine AIMMeR performance in other tumor types, for example, endometrial and gastric cancer.”

The research team is actively seeking collaborations to advance this promising technology.

“We are currently working to this end and would be delighted to hear from possible collaborators with similar interests,” added Dr. Church.