32. Clin-MSI: A machine learning-based approach to MSI detection in endometrial cancers

Abstract

Microsatellite instability (MSI) is a hallmark of cancers with defective DNA mismatch repair (MMR) including Lynch Syndrome and POLE defects, wherein these repeats are prone to strand slippage replication errors that fail to repair, creating insertion/deletion mutations. Clinically, PCR assays have been used historically to classify microsatellite status, although these are prone to errors based on subjective comparisons to control tissues and the vagaries of Sanger sequencing data. We recently developed a hybrid capture-based gene panel assay to survey Lynch Syndrome genes, POLE and 20 microsatellite sites commonly surveyed for MSI by PCR methods. Using data from this panel to study patients with known MSI status, we extracted NGS reads aligned to the 20 microsatellite sites and trained a gradient boosted decision tree model using xgBoost. Using this model, called Clin-MSI, we were able to classify novel endometrial tumor samples with >97% accuracy. As an analytical tool for use in clinical reporting, we utilized SHAP (SHapley Additive exPlanations), a software package that provides granular information to interpret each feature of the model's decision-making process. The combined NGS assay and Clin-MSI predictor were clinically validated in our CAP-CLIA laboratory and have been utilized to clinically characterize over 700 endometrial cancer samples to-date, as an NCI-funded initiative that tests patient samples submitted from across the state of Ohio. These results can provide important information for patients and providers regarding treatment options as well as the need for family-based germline testing.