HumanFilt: a multi-reference host depletion pipeline improves Fusobacterium detection accuracy in tumor WGS data sets.

内容摘要

The study of tumor-associated microbiomes using whole-genome sequencing (WGS) has attracted considerable attention, but microbial signal detection remains controversial due to host contamination and methodological artifacts. As the necessity of human-read removal becomes increasingly evident, many groups now include this step in their data pre-processing workflows. In this work, we introduce an open-source tool, HumanFilt, designed for rigorous host-read removal and apply it to the re-analysis of WGS data from 10 mucinous rectal adenocarcinoma cases originally published by Reynolds et al. The workflow integrates k-mer-based classification (Kraken2), quality and adapter trimming (Trim Galore), vector filtering (BBDuk/UniVec_Core), and duplicate removal (FastUniq). After reducing data complexity, a multi-aligner, multi-reference approach (BWA-MEM/GRCh38, Bowtie2/T2T-CHM13, and Minimap2/Human Pangenome Reference Consortium v1.1) removes remaining host sequences, collectively eliminating more than 99.9% of human-derived reads. Although the additional alignment steps eliminated only a small fraction of total reads, they consistently removed millions of residual sequences per sample, underscoring the importance of rigorous filtering in data sets where non-human reads are a small minority. Taxonomic profiling with PathSeq and MetaPhlAn revealed reproducible enrichment of Fusobacterium species in tumor versus matched normal tissues, and comparison before and after filtering showed that this tumor-over-normal pattern was preserved despite an overall reduction in RPM values. Simulation analyses further showed that HumanFilt preserved more than 99.7% of true Fusobacterium signals, supporting high specificity without meaningful false-negative loss of microbial reads. In direct comparison with Deacon and NoHuman, HumanFilt achieved the most stringent host-read removal but also removed a greater proportion of PathSeq-classified Fusobacterium reads, highlighting the trade-off between maximal host depletion and preservation of ambiguous microbial signal. Cross-validation with immunofluorescence analysis using pan-Fusobacterium (detecting both Fusobacterium animalis and Fusobacterium nucleatum) and F. nucleatum-specific antibodies showed general consistency with Fusobacterium subspecies detected by WGS. Compared to the unfiltered analysis, host depletion markedly reduced artificial microbial signals in normal samples while preserving tumor-associated Fusobacterium, resulting in a more reliable microbial profile. We developed an open-source tool that enables rapid removal of human-derived sequences and applied it to rectal cancer whole-genome sequencing data. This approach reduced false microbial signals while preserving true tumor-associated Fusobacterium, and simulation analyses showed that it retained more than 99.7% of true Fusobacterium reads. Comparison with Deacon and NoHuman showed that HumanFilt achieved more stringent host depletion but also highlighted the trade-off between aggressive host filtering and preservation of ambiguous microbial signal. We also observed general consistency between the sequencing results and immunofluorescence staining in tissue. Together, these findings provide a more reliable basis for studying tumor-bacteria interactions.