Evan Y. Yu

Purpose: Microsatellite instability–high/mismatch repair–deficient (MSI-H/dMMR) tumors exhibit distinct biological features and are associated with Lynch syndrome. Current approaches for detecting MSI-H/dMMR status rely on immunohistochemistry or molecular assays. Recent advances in digital pathology have enabled the extraction of clinically relevant molecular tumor features from routine hematoxylin and eosin (H&E)–stained diagnostic slides. Image-based deep learning approaches have the potential to screen for MSI-H/dMMR at scale in tumor types such as localized prostate cancer, in which MSI-H/dMMR is so rare that traditional testing is not routinely performed. Experimental Design: We developed and validated PathStage-MSI, a multistage transfer learning framework to predict MSI-H/dMMR status from H&E-stained slides. The model was trained and validated on 984 standard-of-care whole-slide images (WSI; n = 906 patients) from three international clinical prostate cancer cohorts with matched H&E images and genomic data. To assess model generalizability and stability, we additionally evaluated PathStage-MSI in an independent metastatic prostate cancer cohort comprising 85 WSIs from 29 patients. Results: PathStage-MSI achieved the area under the receiver operating characteristic curve (AUROC) of 0.89 (0.79–0.96) and 0.91 (0.81–0.98) in two internal hold-out validation cohorts and 0.75 (0.57–0.88) and 0.81 (0.66–0.94) in two independent external validation cohorts. Conclusions: PathStage-MSI is a deep learning–based image analysis framework for detecting MSI-H/dMMR prostate cancer directly from routine histopathology. By enabling efficient prescreening for this rare molecular subtype, it has the potential to improve patient identification for precision therapies and appropriate germline genetic testing.

Distribution of KLK2 and STEAP1 single cell mRNA expression in cells from ARPC samples across MSK scRNA dataset.

Supplementary Figure 6 shows representative KLK2 and STEAP1 IHC micrographs with the H-score threshold of 20.

IHC H-scores of UW-RA samples, including phenotypes as defined by IHC.

Supplementary Figure 4 shows representative KLK2 IHC micrographs of LuCaP PDX tumors across phenotypes.

Supplementary Figure 9 contains donut plots showing the distribution of tumor expressing only one marker, both markers or neither marker for each KLK2-STEAP1, STEAP1-PSMA, and KLK2-PSMA pair across molecular phenotypes and across H-score thresholds.

Cell surface targets and their associated limma values from KLK2-high vs KLK2-low and STEAP1-high vs STEAP1-low differential expression analyses.

Results from predictive model integrating somatic alterations and serum PSA to estimate the probability that a sample is positive for the given antigen.

Supplementary Figure 3 shows KLK2 and STEAP1 expression in LuCaP PDX samples.

Tier 1 druggable genome targets and their associated limma values from KLK2-high vs KLK2-low and STEAP1-high vs STEAP1-low differential expression analyses.