김민지 (울산대학교 )

유지영 (서울아산병원)

김경곤 (울산대학교 )

김민지 (울산대학교 )

김보경 (울산대학교 )

이예린 (울산대학교 )

이광선 (울산대학교 )

김민중 (울산대학교 )

김가빈 (울산대학교 )

장귀주 (울산대학교 산학협력단)

우종규 (서울대학교 )

성제경 (서울대학교 )

유지영 (서울아산병원)

김경곤 (울산대학교 )

Metaproteomics enables simultaneous profiling of host physiology and microbial function, but the large search space of microbiome gene catalogs increases protein inference errors and false-positive microbial identifications that cannot be distinguished from true microbial signals by computational target–decoy approaches alone.

We therefore introduced a biological decoy. Germ-free (GF) mice harbor no microbiota, so every microbial protein reported from GF feces is by definition a false positive — allowing direct empirical estimation of false discovery. Feces from GF and specific pathogen-free (SPF) mice were extracted in a urea/SDS/TEAB buffer, digested with S-Trap, and analyzed on an Orbitrap Eclipse Tribrid MS in both data-dependent (DDA) and data-independent (DIA) acquisition. Database searches were performed against Mouse Swiss-Prot and the Mouse Gut Gene Catalog (2.6 million genes) and filtered at 1% FDR, followed by GO, KEGG, and COG annotation.

In SPF feces, DIA-NN identified 19,421 microbial proteins (14,329 assigned to 99 genera) compared with 19,368 by Proteome Discoverer-based DDA (14,221 assigned to 95 genera). The two methods shared 8,202 microbial proteins, with 11,219 and 11,166 proteins uniquely identified by DIA and DDA, respectively. Quantitative abundances of shared proteins were moderately correlated (Pearson r = 0.578). In GF feces, only 88 and 0 microbial proteins were retained by DIA and DDA, bounding false-positive contribution at 0.45% and 0% of SPF identifications, respectively. Although microbial proteome coverage was comparable, DDA identified more host proteins than DIA (1,966 vs. 1,579).

By replacing an estimated error rate with a measured one, this workflow establishes GF mice as a general validation anchor for host–microbiome metaproteomics and yields acquisition-mode guidance grounded in ground-truth negatives rather than identification counts alone. Despite near-identical counts, the two sets shared only 8,202 proteins (42%) with moderately correlated abundances (r = 0.578), showing that identification number conceals substantial divergence in protein identity.