Single-slide histology-based deep learning model for mismatch repair deficiency prediction in colorectal cancer.

3567 Background: Mismatch repair deficiency (dMMR) is a critical predictive biomarker for determining eligibility and response to immunotherapy in colorectal cancer (CRC). The current gold standards for detecting dMMR include next-generation sequencing (NGS) for microsatellite instability (MSI) detection and immunohistochemistry (IHC) for mismatch repair protein expression. However, discrepancies between these techniques have been observed, potentially impacting clinical decisions and patient outcomes. To address this, we developed a histology-based deep learning (DL) model to predict MMR status, with a specific focus on resolving cases of discordance. Methods: Paired hematoxylin and eosin (H&E) slides from 974 CRC tumors were retrospectively collected from the Dana-Farber Cancer Institute, all of which had NGS Oncopanel and IHC MMR reports available. Using NGS-determined MMR status as the training reference, we developed a multi-instance deep learning model to predict MMR status from single H&E slides. Feature extraction employed various pathology foundation models (FMs). The dataset was split into training and tuning sets. A hold-out test set (n = 52, 65% dMMR) was curated including patients treated with immune checkpoint inhibitors or those with NGS-dMMR/IHC-proficient discordance. Results: Among the overall cohort, NGS/IHC concordance identified 82 dMMR patients (9%), 881 proficient MMR (pMMR) patients (90%), and 11 cases (1%) with NGS-dMMR/IHC-proficient discordance. In the hold-out test set, the fine-tuned CTransPath FM demonstrated the highest performance, achieving an area under the curve (AUC) of 0.88 (95% CI 0.77–0.98), a positive predictive value of 0.93, and correctly classifying 8 of 11 discordant cases (73%) as dMMR. Comparative FMs, CONCH and UNI, exhibited slightly lower AUCs (0.86 and 0.85, respectively) and lower accuracy in classifying discordant cases (Table). Conclusions: Our histology-based DL model shows promise as a complementary tool for IHC in predicting dMMR status in CRC. The single-slide approach offers a rapid, robust and cost-effective method to prioritize IHC-proficient cases for further validation by NGS. Cohort expansion and validation in an external dataset are underway. Test set (n=52, 65% dMMR). Model AUC (CI,95%) Sensitivity Specificity PPV NPV Accuracy NGS+/IHC- CTransPath 0.88 (0.77-0.98) 0.85 0.89 0.93 0.76 8/11 CONCH 0.86 (0.76-0.96) 0.88 0.76 0.67 0.93 5/11 UNI 0.85 (0.74-0.96) 0.78 0.82 0.70 0.87 6/11