Background: Blood-derived circulating cell-free microbial DNA (cfmDNA) has emerged as a potential non-invasive biomarker source for cancer detection. However, low biomass and high susceptibility to analytical variability raise concerns regarding the stability and interpretability of inferred microbial signatures. This study aimed to evaluate how different taxonomic profiling strategies influence downstream machine learning-based classification and feature interpretation in lung cancer. Methods: cfDNA sequencing data from 168 individuals (80 lung cancer patients and 88 non-cancer controls) were analyzed using two taxonomic profiling workflows: a Bracken-based abundance estimation approach and a BLAST-refined alignment-based strategy. Microbial profiles derived from each pipeline were evaluated using supervised machine learning models within a nested cross-validation framework. Feature stability and fold-change trends were compared across profiling strategies. Results: A Random Forest model achieved robust classification performance under both workflows (AUC 0.852 for Bracken-derived data and 0.906 for BLAST-derived data). However, substantial pipeline-dependent variation was observed in feature selection patterns and quantitative fold-change directionality. Although 13 genera were consistently selected across cross-validation folds in both workflows, the magnitude and direction of abundance differences were not uniformly concordant. Conclusions: Blood-derived microbial DNA profiles can support machine learning-based lung cancer classification; however, feature-level interpretation remains sensitive to taxonomic assignment strategy. These findings underscore the importance of pipeline-aware interpretation and methodological transparency in low-biomass blood microbiome research.
Goh et al. (Thu,) studied this question.