%0 Journal Article
%T A High-Accuracy Machine Learning Approach Improves Mitochondrial DNA Variant Fidelity in Archival FFPE Samples
%A Kenji Tanaka
%A Yui Kobayashi
%A Takeshi Nakamura
%J Interdisciplinary Research in Medical Sciences Specialty
%@ 3062-4401
%D 2025
%V 5
%N 1
%R 10.51847/awQVwVWCFL
%P 260-275
%X Archival formalin-fixed paraffin-embedded (FFPE) specimens represent a highly valuable resource for mitochondrial DNA (mtDNA) biomarker research. Nonetheless, artifacts generated during formalin fixation frequently resemble somatic mutations, thereby impeding accurate clinical interpretation. Existing artifact-filtering methods, primarily developed for nuclear DNA, fail to accommodate key mtDNA characteristics such as its elevated copy number, strand-specific GC-content differences, and heteroplasmic properties. These shortcomings call for the creation of specialized computational strategies. To meet this need, we introduced mtFFPECleaner, a machine learning-based system that combines multiple features of mtDNA variants, encompassing variant allele frequency (VAF) patterns, strand orientation bias metrics, surrounding sequence context, and local nucleotide composition. The system employs a random forest classifier trained on 837 verified genuine mutations and 1169 artifacts sourced from 23 matched FFPE–fresh frozen (FF) specimen pairs. Training involved tenfold cross-validation, with performance subsequently confirmed on an independent cohort of 15 additional paired FFPE–FF samples. Our results indicate that formalin-related artifacts in mtDNA next-generation sequencing (NGS) data mainly consist of C > T/G > A transitions, particularly at low VAF levels, and display pronounced strand bias along with clear sequence context preferences. Following optimization with balanced sampling (1:2 ratio of artifacts to genuine mutations), the mtFFPECleaner classifier demonstrated superior performance in the validation cohort, attaining 98.7% specificity and 98.2% sensitivity. It substantially surpassed established nuclear DNA artifact-removal approaches, including SOBDetector and DEEPOMOICS FFPE. After applying mtFFPECleaner for artifact correction, mutational spectra derived from 314 FFPE samples exhibited close alignment with those obtained from FF specimens. Notably, artifact prevalence increased in association with longer FFPE storage times, confirming the tool’s capacity to counteract accumulated formalin-induced damage spanning many decades in preserved biospecimens. mtFFPECleaner is the first tool developed specifically to improve the reliability of mtDNA mutation profiles from FFPE material in NGS analyses. Availability of the open-source R package (https://github.com/AlienLemon117/mtFFPECleaner) supports its broad application in extensive archival investigations and enhances the utility of FFPE biobanks for translational research.
%U https://galaxypub.co/article/a-high-accuracy-machine-learning-approach-improves-mitochondrial-dna-variant-fidelity-in-archival-ff-gga7b0gpixhkidn