Cracking Colon Cancer
How a new clinically approved AI-based tool enables rapid microsatellite instability detection
Microsatellite instability (MSI) – a tumor genotype characterized by mismatch errors of repetitive DNA – is found in 15 percent of colorectal cancer (CRC) patients and plays an important role in diagnostic, prognostic, and therapeutic outcomes. MSI also acts as a key biomarker for Lynch Syndrome – the most common cause of hereditary CRC. Universal screening of CRC patients for MSI is highly recommended, so a team of researchers from biotech company Owkin has recently developed MSIntuit – an artificial intelligence (AI)-based pre-screening tool for rapid MSI detection from H&E stained slides (1).
“Conventionally, MSI is diagnosed with immunohistochemistry or polymerase chain reaction,” says Charlie Saillard, lead data scientist at Owkin. “These tests contribute to an ever-increasing workload for pathologists and technicians.” In the study, researchers trained MSIntuit on samples from the Cancer Genome Atlas. A blind validation was then performed on an independent dataset of 600 consecutive CRC patients. The AI model either outputs “MSS-AI” (no further testing needed) or “undetermined,” where a standard MSI test is then required.
MSIntuit yielded a sensitivity of 96 percent and a specificity of 46 percent – rivaling the sensitivity of gold standard techniques and ruling out almost half of the non-MSI population. “Given the global shortage of pathologists worldwide, early rule-out of non-MSI patients can reduce the workload of pathologists and accelerate MSI screening in clinical practice,” says Saillard.
“This multidisciplinary approach is key to better understand the needs of pathologists and build a relevant tool that can help them in their daily practice,” says Saillard. He hopes the new AI-based tool will lead the way for other AI platforms to be deployed in clinical practice to help pathologists identify actionable biomarkers from routine histology slides. Saillard also believes that MSI pre-screening tools for other tumor types will be developed in the near future.
- C Saillard et al., Nat Commun, 14 (2023). PMID: 37932267
Associate Editor for the Pathologist