An AI model developed by researchers at Yonsei University, Mayo Clinic, and Vanderbilt University Medical Center may help uncover cancer patients who are likely to benefit from immunotherapy but would otherwise be overlooked by traditional diagnostics.

The deep-learning tool, called MSI-SEER, analyzes routine pathology slides to predict whether a gastric or colorectal tumor displays microsatellite instability-high (MSI-H), a genetic signature that makes tumors more responsive to immune checkpoint inhibitors. MSI-H tumors typically have numerous mutations, which make them more recognizable to the immune system.

The results were published in the journal npj Digital Medicine on May 19.

In several cases, the model identified MSI-H regions that standard lab tests, such as PCR (a DNA-based test) or IHC (a protein-based staining technique), failed to detect. These findings enabled patients to receive immunotherapy and respond.

“Treatment strategies can vary significantly depending on how accurately we analyze a patient’s tumor,” said corresponding author Professor Cheong Jae-ho of Yonsei’s Department of Surgery in a Tuesday release from Severance Hospital. “This AI model clearly presents evidence to support the use of immunotherapy, helping physicians make more precise treatment decisions.”

The study evaluated more than 3,000 pathology slides from patients in Asia, Europe, and North America, using a statistical AI approach known as deep Gaussian processes within a weakly supervised learning framework.

MSI-SEER not only matched the performance of widely used AI models like ResNet and EfficientNet, it also outperformed them in patient groups where tissue staining quality or racial variation had previously thrown off algorithms.

The model goes beyond simply classifying tumors. It assigns a confidence score to each prediction and maps the spatial distribution of MSI-H regions within tumors, offering insights that conventional tests cannot.

In a real-world gastric cancer cohort treated with immunotherapy, MSI-SEER reclassified five patients who had been labeled as microsatellite stable (MSS) and found that three of them had 85 percent MSI-H tumor content and responded to treatment. It also flagged one patient labeled as MSI-H who did not respond and had very few MSI-H regions.

To further refine predictions, the model incorporated analysis of stromal cells, which form part of the tumor’s surrounding tissue. Non-responders tended to have MSI-H regions filled with these cells, suggesting that the local tumor environment may blunt the immune response.

A rule-based classifier using both MSI-H proportion and the stroma-to-tumor cell ratio was able to separate responders from non-responders with 94.1 percent accuracy.

While still investigational, MSI-SEER’s ability to deliver slide-level predictions and quantify uncertainty may help reduce the need for expensive and labor-intensive molecular tests, especially in settings where PCR and IHC are not readily available.

“It marks the beginning of an era where physicians’ clinical expertise and AI’s computational power work hand-in-hand,” said corresponding author Professor Hwang Tae-hyun of Vanderbilt University Medical Center in a Tuesday statement.