A machine learning approach to predict and mitigate artifact variants for molecular residual disease detection

Tumor-informed molecular residual disease (MRD) detection requires removal of artifacts from tissue whole-genome sequencing (WGS), which can cause false positive calls in plasma, even in healthy donors (HDs). To address this, we developed a machine-learning framework to score the reliability of tissue variants, thereby improving the specificity and sensitivity of tissue-informed MRD detection. We used retrospective WGS data from 47 tumor and matched normal (T&MN) samples across a diverse set of cancer types. Somatic single nucleotide variants (SNV) were identified via a consensus calling strategy from multiple variant callers. We extracted variant features from each caller relating to signal amount and quality, and subsequently engineered additional features to qualify strand bias, variant- and locus-specific characteristics. To generate training labels, SNV were evaluated in matched patient plasma and 107 HD plasmas. Variants with high mean VAF across the HD plasma samples were labeled as likely artifacts (~16k), while those with high VAF in the matched patient plasma were labeled as true tumor variants (~187k). An extremely randomized trees model was then trained using 17 informative, uncorrelated features. Model performance was measured by calculating the area under the curve (AUC) from a receiver operating characteristic (ROC) analysis. Our model distinguished true tumor variants from likely artifacts with a cross-validated ROC AUC of 0.94 and an average precision of 0.99. Key predictive features from the tissue included tumor VAF, mutant allele count, and a read orientation quality metric. In each patient, the model-derived confidence score demonstrated a strong and negative correlation with the mean VAF in the HD plasma samples (median Spearman ρ = -0.30), outperforming a previous method (median ρ = -0.17). This work provides a data-driven framework that systematically ranks and filters tissue variants by their likelihood of being artifacts. By enhancing the reliability of mutation calls from the tissue sample, this method can improve the accuracy of tissue-informed MRD detection.