MiMSI

MiMSI detects microsatellite instability (MSI) in cancer genomes from next-generation sequencing (NGS) data to improve MSI classification for guiding immune checkpoint inhibitor therapy.

Key Features:

• Deep Neural Network-Based Classifier: Uses a deep learning neural network to classify MSI status from NGS-derived features.
• Multiple Instance Learning Framework: Trains the classifier under a multiple instance learning paradigm to handle sample-level labels arising from heterogeneous locus-level signals.
• Training Inclusion of Low Tumor Purity Cases: Incorporates MSI cases with low tumor purity into the training dataset to increase robustness on routine clinical samples.
• Validation on Targeted NGS Panels: Validated using large panel targeted NGS data from the MSK-IMPACT study.
• Reported Performance Metrics: Achieved sensitivity 0.940 and auROC 0.988 on a challenging real-world case set and outperformed MSISensor (sensitivity 0.57, auROC 0.911).

Scientific Applications:

• MSI detection in clinical samples: Enables sensitive identification of MSI in tumor samples, including those with low tumor purity.
• Biomarker-driven patient stratification: Supports selection of patients for immune checkpoint inhibitor therapy based on MSI status.
• Cancer genomics research: Facilitates discovery and characterization of MSI-high phenotypes in sequencing studies.

Methodology:

MiMSI implements a deep neural network classifier trained using a multiple instance learning framework on targeted NGS data (MSK-IMPACT), with training datasets that include low tumor purity MSI cases.