mtDeOxoGer: A machine learning-based strategy to effectively filter artifactual mutations induced by oxidative damage...

Mitochondrial DNA (mtDNA) mutations are among the most common genetic alterations in tumors, and next-generation sequencing (NGS) has greatly increased their detection sensitivity. However, the accuracy of mtDNA NGS is often compromised by oxidative artifacts, particularly those introduced during ultrasonic DNA fragmentation in library preparation, which can be misinterpreted as true low-frequency mutations. While such artifacts have been extensively studied in nuclear DNA, their impact on mtDNA remains insufficiently explored. In this study, we systematically evaluated oxidative damage-induced artifacts, focusing on 8-oxoguanine lesions, in mtDNA NGS using both private and public datasets, and developed a machine learning-based filtering strategy to improve detection accuracy. We found that low-frequency C > A and G > T substitutions frequently occurred in mtDNA NGS data, accompanied by strong batch effects and correlations with the overall mutation burden and local GC content. Experimental validation demonstrated that ultrasonic fragmentation, influenced by parameters including temperature, intensity, and cycle number, was a principal source of these artifacts, whereas enzyme-based fragmentation markedly reduced oxidative damage and improved mutation detection fidelity. Our logistic regression model, mtDeOxoGer, which incorporated substitution type, variant allele frequency (VAF), strand orientation bias (SOB) score, local GC content, and sequence context, showed robust performance in distinguishing authentic mtDNA mutations from oxidative artifacts. These findings underscore the need to refine mtDNA sequencing methodologies, and the machine learning-based strategy presented here provides a practical framework for enhancing the reliability of mitochondrial genetics studies in cancer and other disease contexts.