Refining Cell-free DNA Metagenomics for Sepsis in Acute Leukemia: Towards Standardized Computational Strategies for Host Depletion and Microbial Detection from Shallow-Depth, Low-Coverage Profiles.

Introduction: Accurate alignment is critical for metagenomic profiling of plasma cell-free DNA (cfDNA), where microbial fragments are short, low in abundance, and often obscured by abundant human DNA. Yet, no unified guideline exists for selecting alignment algorithms in cfDNA studies. Results: We benchmarked six alignment methods (BWA MEM2, Minimap2, Bowtie2(sensitive), Bowtie2(very sensitive), Bowtie2(very sensitive local), and Kraken2) using in silico simulated datasets and cfDNA from 102 acute leukemia samples with sepsis. Evaluation criteria included bacterial read retention after human read filtering, minimization of false positives in real cfDNA datasets, and detection sensitivity for antimicrobial resistance (AMR) genes. BWA-MEM2 emerged as the most effective for removing stringent human reads and bacterial classification, while bowtie2(very_sensitive_local mode) was optimal for AMR gene detection. Post-filtering, cfDNA profiles exhibited marked heterogeneity in bacterial and AMR gene signatures, reflecting expected clinical variability. Conclusions: BWA-MEM2 provides optimal human read removal and bacterial taxa classification, minimizing false positives, though some true reads may be lost. For AMR gene detection, Bowtie2 (very_sensitive_local mode) demonstrated superior sensitivity in simulated datasets. This study presents the first systematic evaluation of read-based and kmer based aligners in plasma cfDNA, offering practical guidance for balancing human read removal with bacterial and AMR gene profiling to achieve robust and clinically meaningful results.