Machine Learning Algorithm for the Detection of Tumor Microsatellite Instability Based on Multiomics Biomarkers

Area of Science:

• Oncology
• Genomics
• Bioinformatics

Background:

• Accurate microsatellite instability (MSI) classification is crucial for guiding immunotherapy selection in advanced cancers.
• Current MSI detection methods can exhibit variability, potentially leading to missed diagnoses and delayed treatment.
• Next-generation sequencing (NGS) offers a platform for developing complementary MSI screening approaches.

Purpose of the Study:

• To develop and validate a machine learning (ML) model for predicting MSI status in colorectal cancer (CRC) using multiomics data.
• To assess the model's performance and translatability across different cancer types, including uterine and gastric cancers.
• To investigate the utility of the ML model in identifying MSI-high cases among those with indeterminate status.

Main Methods:

• Analysis of NGS data from 2,756 CRC patients, including single-nucleotide variants (SNVs), copy-number variants (CNVs), and immune-related gene expression.
• Training ML algorithms on 70% of the CRC cohort, utilizing tumor mutation burden (TMB) and Boruta-selected features.
• Testing the trained models on independent CRC cohorts, The Cancer Genome Atlas (TCGA) datasets (COAD/READ), and uterine and gastric cancer cases.

Main Results:

• The developed ML model demonstrated strong predictive performance across multiple cancer cohorts, with high specificity and negative predictive value in CRC.
• Sensitivity ranged from 78% in uterine cancer to near-perfect in other cohorts.
• Among indeterminate cases in CRC and uterine cancer, 15% were reclassified as likely MSI-high, with a significant proportion showing loss of MMR proteins (MLH1/PMS2).

Conclusions:

• The ML model accurately predicts MSI status in colorectal and uterine cancers using multiomics NGS data, independent of direct microsatellite analysis.
• This approach enhances diagnostic accuracy for MSI-high tumors, particularly in indeterminate cases.
• The findings support the potential of ML-driven multiomics analysis to improve patient selection for immunotherapy and ensure timely access to treatment.