Innovative Deep Learning Model Enhances Prediction of Microsatellite Instability in Cancer Tumors

Cancer remains a significant health issue worldwide, with approximately one in three individuals expected to develop the disease during their lifetime. A critical factor influencing cancer prognosis and treatment response is the tumor’s microsatellite status—whether it is stable or unstable. Microsatellites are short, repetitive DNA sequences, and their stability reflects the DNA’s overall integrity within tumor cells.

The microsatellite instability-high (MSI-H) status of tumors has notable clinical implications. Patients with MSI-H cancers generally experience better outcomes compared to those with microsatellite stable (MSS) tumors. Moreover, tumors exhibiting deficient mismatch repair (dMMR)—a condition marked by mutations in specific genes responsible for repairing DNA replication errors—are often more responsive to immune checkpoint inhibitors (ICIs), a promising class of immunotherapy drugs. Consequently, testing for MSI status is recommended for newly diagnosed gastric and colorectal cancers to guide personalized treatment strategies.

Recent advancements in artificial intelligence (AI) are transforming how MSI testing is conducted. While previous studies employed deep learning techniques, such as convolutional neural networks and vision transformers, they often lacked the ability to account for prediction uncertainty, which is crucial for clinical decision-making. To address this gap, a team of researchers from the U.S. and Korea, including Jae-Ho Cheong and Jeonghyun Kang from Yonsei University College of Medicine, developed MSI-SEER. This innovative model leverages a deep Gaussian process-based Bayesian approach to analyze hematoxylin and eosin-stained whole-slide images using weakly-supervised learning, offering a highly accurate prediction of microsatellite status in gastric and colorectal cancers.

A key feature of MSI-SEER is its capacity to quantify its confidence in each prediction through uncertainty estimation. The model employs Monte Carlo dropout to assess predictive variance, converting it into a Bayesian Confidence Score (BCS). This score indicates the reliability of each prediction, enabling the system to identify cases where certainty is low. When uncertainty is high, MSI-SEER automatically flags such slides for secondary review by human pathologists, enhancing the safety and precision of diagnoses.

The study, published in npj Digital Medicine on May 19, 2025, highlights the model's robust performance across diverse datasets, including patients from various racial backgrounds. Beyond predicting MSI status, MSI-SEER adeptly forecasts ICI responsiveness by integrating tumor MSI status and stromal characteristics. Additionally, tile-level predictions provide insights into the spatial distribution of MSI-H regions within the tumor microenvironment, further informing treatment strategies.

Prof. Cheong emphasized the potential of this technology to serve as a blueprint for AI-Human collaboration in clinical settings. By 'knowing what it doesn’t know,' the model offers greater reliability and safety, facilitating integration into clinical workflows for precision cancer therapy. The overarching goal is to expand AI's role in multi-modal data analysis, ultimately advancing personalized treatment options for cancer patients.

This breakthrough demonstrates the power of AI in developing clinically applicable models that improve the prediction of immunotherapy response, paving the way for safer, more effective cancer treatments.