Differential Abundance Analysis of the Human Tumor Microbiome

Benchmarking of four DA tools (DESeq2, ANCOM-BC2, ALDEx2, MaAsLin2) applied to SMURF-normalized 16S abundance data from Nejman et al. (Science, 2020) (https://doi.org/10.1126/science.aay9189), comparing normal adjacent tissue (NAT) vs. tumor across three cancer types using paired patient samples. Data from Table S2 was rounded to integers and analyzed at the genus level across three data conditions (raw, mbImpute, BMDD) with 5% and 10% prevalence filters.

---

Figure 1. Analysis pipeline: three cancer types × three data conditions × two prevalence filters × four DA tools. Significance threshold: q < 0.05 (Benjamini-Hochberg). Paired patient design accounted for in all tools.

---

Key findings

Tool sensitivity varied dramatically across conditions, consistent with Nearing et al. (Nature Communications, 2022). DESeq2 identified the most genera (up to 71 in lung, raw 5% filter), while ALDEx2 and ANCOM-BC2 were most conservative. Discordance was amplified by the pre-normalized input data, which violates the raw count assumptions of count-based tools. ANCOM-BC2 findings were significant under the standard test but not the robust test across all cancer types, reflecting the low-biomass nature of tumor microbiome data.

The table below shows genera found significant by more than one tool across any condition.

Genus	Cancer	Direction	Tools
Staphylococcus	Breast	Tumor-enriched	DESeq2, ALDEx2, MaAsLin2
Prevotella	Breast	Tumor-enriched	ALDEx2, MaAsLin2
Afipia	Breast	NAT-enriched	DESeq2, MaAsLin2
Kocuria	Breast	NAT-enriched	DESeq2, MaAsLin2
Sphaerobacter	Breast	NAT-enriched	DESeq2, MaAsLin2
Bacillus	Lung	NAT-enriched	DESeq2, MaAsLin2
Devosia	Lung	NAT-enriched	DESeq2, MaAsLin2
Friedmanniella	Lung	Tumor-enriched	ALDEx2, MaAsLin2
Staphylococcus	Ovary	Tumor-enriched	DESeq2, ANCOM-BC2
Escherichia/Shigella	Ovary	Tumor-enriched	DESeq2, ANCOM-BC2

Staphylococcus is the most robust finding — tumor-enriched in both breast and ovary across the most conservative tools, corroborating prevalence findings in Nejman et al. and published literature (Urbaniak et al. 2016; Hieken et al. 2016). Full per-tool, per-condition results are in Summary.txt and Detailed_Interpretation_Report.txt.

---

Repository structure

Differential-Abundance-Analysis/
├── Breast_new/
│   ├── Raw_data/          # R scripts + outputs for raw integer data
│   └── Imputed_data/      # R scripts + outputs for mbImpute and BMDD
├── Lung_new/              # Same structure as Breast_new
├── Ovary new/             # Same structure (no BMDD — sample size too small)
├── Supplementary_Tables/  # Original supplementary tables from Nejman et al.
├── abundance_tumor.csv    # Integer-rounded genus-level abundance matrix
├── clinical_tumor.csv     # Paired clinical metadata (NAT vs Tumor)
├── Summary.txt            # Significant genera per tool per condition
└── Detailed_Interpretation_Report.txt  # Full biological interpretation