Abstract 104: A machine learning approach to predict and mitigate artifact variants for molecular residual disease detection.

Key Points

• The aim is to improve the accuracy of tissue-informed molecular residual disease detection by mitigating false positive calls due to artifacts.
• Developed a machine learning framework to score tissue variant reliability.
• Utilized retrospective whole-genome sequencing data from tumor and matched normal samples.
• Engineered features for strand bias and variant characteristics for better prediction.
• Classified variants using their performance in matched patient plasma and healthy donor plasma.
• Achieved an ROC AUC of 0.94 and average precision of 0.99 for distinguishing true tumor variants.
• Identified key predictive features such as tumor variant allele frequency and mutant allele count.
• Showed better performance than prior methods with a higher median Spearman correlation coefficient.