Deep Gaussian process with uncertainty estimation for microsatellite instability and immunotherapy response...

Area of Science:

• Computational pathology
• Oncology
• Artificial intelligence in medicine

Background:

• Tumor microsatellite instability (MSI) status is crucial for predicting response to immune checkpoint inhibitors (ICIs) and chemotherapeutics.
• Identifying MSI-H/dMMR tumors guides treatment decisions in gastric and colorectal cancers.

Purpose of the Study:

• To develop and validate MSI-SEER, a deep Gaussian process-based Bayesian model for predicting MSI status using H&E whole-slide images.
• To assess MSI-SEER's accuracy in predicting ICI responsiveness by integrating MSI status and stroma-to-tumor ratio.
• To explore spatial distribution patterns of MSI-H regions and their impact on ICI response.

Main Methods:

• Weakly-supervised learning on H&E whole-slide images to predict MSI status.
• Bayesian deep Gaussian process modeling (MSI-SEER).
• Validation across multiple large datasets with diverse patient populations.

Main Results:

• MSI-SEER achieved state-of-the-art performance in MSI prediction, incorporating uncertainty quantification.
• High accuracy was obtained in predicting ICI responsiveness by combining MSI status with stroma-to-tumor ratio.
• Tile-level predictions provided novel insights into the spatial distribution of MSI-H regions and their association with ICI response.

Conclusions:

• MSI-SEER offers a powerful computational tool for predicting MSI status from histopathology images.
• The model enhances prediction of ICI responsiveness, potentially optimizing cancer treatment selection.
• Analysis of spatial MSI patterns reveals new understanding of the tumor microenvironment and treatment outcomes.