kruskal-wallis-test

The Kruskal-Wallis test (often reported as “KW” or “Kruskal H”) is a rank-based method for testing whether three or more groups differ. In this corpus it is used to compare antimicrobial-resistance (AMR) gene content across environments, AMR fitness costs across mechanisms, eukaryotic DNA fraction across sample types, sites and vegetation, growth rates across metabolic-model classes, and pathway completeness across ecosystems. _1,2,3,4,5,6

Uses in the corpus

Environmental resistome

Species from clinical sources have a median of 5 AMR gene clusters, compared with 2 for soil, aquatic and host-associated species. Human gut species resemble clinical species, with a median of 5. The Kruskal-Wallis result is H = 781.9, p = 9.4×10⁻¹⁶⁷, η² = 0.056. Of 15 pairwise environment comparisons, 13 remain significant after FDR (false discovery rate) correction. The largest effect is clinical versus aquatic (rank-biserial r = −0.49). These results are reported in Environmental Resistome at Pangenome Scale and feed environmental-resistome. ₁

Resistome composition also differs across environments. In clinical species, 68% of AMR is accessory (acquired); in soil species the figure is 43%. The Kruskal-Wallis result is H = 506.0, p = 4×10⁻¹⁰⁷, η² = 0.036. ₁

The table above lists the project’s per-mechanism H3 tests. Each test compares one mechanism’s fraction across environments, but these fractions cover only the classified clusters. 18.7% of AMR clusters (15,550) could not be assigned a mechanism from gene name or product annotation, and the project excluded them from the mechanism fractions. ₁

The project’s data/mechanism_by_environment.csv output holds these per-mechanism results in 4 rows. ₁

Pangenome-scale AMR atlas

An independent pangenome analysis reports a similar pattern; it supports the clinical-enrichment result above. Human/Clinical species carry 10.6 AMR clusters per species (n=2,248). The comparison groups are Soil/Terrestrial at 4.6 (n=2,469), Aquatic at 3.9 (n=1,827) and Animal at 3.0 (n=959). The Kruskal-Wallis result is H=440, p=7.0e-93. Clinical AMR is also less core (30.8%) than soil AMR (58.1%) or plant AMR (63.1%). Caveat: of the 14,723 AMR-carrying species, only 7,838 (53.2%) received an environment classification other than “Other/Unknown”, and the test is restricted to these species across 6 categories (Pan-Bacterial AMR Gene Landscape). ₃

Eukaryotic DNA admixture in prokaryotic metagenomes

In a univariate test across studies, the eukaryotic fraction differs strongly across sample matrix (Kruskal–Wallis H=77.8, p=1.3×10⁻¹⁷). All pairwise matrix contrasts are significant after BH-FDR, the Benjamini-Hochberg false-discovery-rate correction (benjamini-hochberg-fdr). This result should not be read as an environment effect. The project calls the study/batch confounding its central methodological result: each biome is ~80–100% nested within a single study. When whole studies are held out, the environment model does not generalize (out-of-study detection AUC = 0.56 ≈ chance). The matrix association is therefore largely batch-driven. This feeds eukaryotic-dna-admixture-in-prokaryotic-metagenomes and study-batch-confounding-of-environmental-associations. ₄

The vegetation and geography tests below differ in kind. They are run within a single neon soil metagenome study (1,186 runs, constant protocol/batch), so batch is held constant. Within this study, the eukaryotic fraction differs across 11 local-vegetation levels (env_local_scale; Kruskal H=119.1, p=7.6×10⁻²¹). It is highest in sedge/forb herbaceous soil (median 23%), followed by emergent wetland (14%), dwarf scrub (13%) and evergreen forest (2%). It is ≈0 in deciduous forest, cropland and pasture. ₄

Within the same NEON soil study, the eukaryotic fraction also differs by geography across 47 sites (Kruskal H=310.4, p=2.4×10⁻⁴⁶). It is highest at Arctic tundra sites: Utqiaġvik 30%, Caribou-Poker Creeks 22% and Toolik 17%. Temperate forests range from 2–8%. ₄

Community metabolic ecology

In the NMDC community metabolic ecology project, H1 used spearman correlation with BH-FDR. H2 combined principal-component-analysis with Kruskal-Wallis tests of ecosystem separation. The figure figures/pathway_completeness_boxplot.png shows amino-acid pathway completeness by ecosystem type, with Kruskal-Wallis significance marked. ₆

Null results

In the ADP1 triple-essentiality project, the test covered only the 478 triple-covered genes (genes with TnSeq, FBA and growth data). All of these genes are TnSeq-dispensable on minimal media. Within this restricted set, mean growth rates did not differ significantly across FBA (flux balance analysis; flux-balance-analysis) classes (Kruskal-Wallis H = 1.67, p = 0.43). The null result does not extend to TnSeq-essential genes. ₅

In the AMR fitness-cost project, the test detected no significant difference in AMR-gene fitness cost by resistance mechanism, contrary to that project’s H2. Four testable mechanisms were compared: efflux (N=254), enzymatic inactivation (N=304), metal resistance (N=144) and unknown (N=74). The result was H = 0.65, p = 0.89. A non-significant test does not show that costs are equal across mechanisms. The figure h2_mechanism_stratification.png plots fitness by mechanism with this test. The result bears on antimicrobial-resistance-fitness-cost. ₂

Caveats

Several environment comparisons combine extremely small p-values with modest effect sizes. Examples are η² = 0.056 for AMR cluster counts and η² = 0.036 for accessory fraction. With thousands of species, significance alone overstates the share of variance explained by environment. ₁

In the pangenome atlas, the test is restricted to environment-classified species, so its Kruskal-Wallis result covers only about half of AMR-carrying species. It is not a full census. ₃

In these projects, pairwise follow-ups are reported with FDR correction: 13 of 15 comparisons significant in one project, and all matrix contrasts significant in the other. Pairwise effect sizes such as rank-biserial r are reported alongside them; see mann-whitney-u-test for the two-group analogue. _1,4