Environmental structuring of the bacterial resistome

The bacterial resistome—the collection of antimicrobial-resistance (AMR) genes in a population—varies among ecological settings in abundance, diversity, core/accessory composition, resistance mechanism, and association with mobile genomic structure. These conclusions are conditional on catalog scope: the amrfinderplus reference catalog includes stress-response genes alongside classical antibiotic-resistance genes, so catalog-defined AMR is partly a broader stress-resistance landscape rather than a narrow count of antibiotic resistance. ₁ The central analyses examined 82,908 AMR gene clusters across 14,723 bacterial species and 293K genomes, while a uniform Bakta–AMRFinderPlus analysis identified 83,008 hits across 14,723 species; 53.2% of species carried at least one hit. _2,1

The five most abundant AMRFinderPlus families were bla beta-lactamases with 6,115 hits, merA mercury reductase with 4,506, arsD arsenic metallochaperone with 2,611, merP mercury-binding protein with 2,222, and vanR vancomycin response regulator with 1,929. Mercury-resistance families contributed approximately 15,000 hits and 18% of annotations, and arsenic-resistance families another approximately 6,000. ₁ Thus environment-associated AMR abundance and mechanism profiles are catalog-dependent measurements, not direct counts of clinically defined antibiotic resistance. ₁ The atlas also interprets heavy-metal and antibiotic resistance as genomically intertwined and often co-located on mobile elements, citing the COG P enrichment; that co-location is an interpretation and is not separately quantified. ₁

The evidence establishes environment-associated structure, not environmental causation: exposure, phylogeny, sampling, isolation practices, annotation, and correlated host-associated traits remain alternative explanations. ₂ AMRFinderPlus-focused annotation may underestimate novel environmental resistance and favor clinical enrichment because clinical resistance genes are better characterized. ₂ The 15,550 mechanism-unassigned clusters, 18.7% of all AMR clusters, are especially important because their environmental distribution is unknown; because they were excluded from mechanism fractions, a non-random distribution across environments could bias mechanism comparisons. ₂

Conservation compartments: intrinsic and acquired resistance

The Pan-Bacterial AMR Gene Landscape supports a genomic distinction between intrinsic resistance—resistance functions maintained as part of a lineage’s conserved genome—and acquired resistance—resistance genes found in accessory, minority, or singleton compartments. This distinction is operational and population-genomic rather than a gene-name rule: vertically inherited intrinsic functions can be stable core residents, whereas acquired genes are often present in a minority of genomes, accessory clusters, or singleton clusters, but the same broad mechanism category can contain both conserved and accessory genes. ₁

Across 27,690 pangenome species, the atlas identified 83,008 AMRFinderPlus hits spanning 82,908 distinct gene clusters, 1,939 AMR gene families, 2,079 AMR products, and 14,723 species (53.2% of the 27,690 pangenome species); detection methods were HMM (51.5%), BLASTP (22.7%), EXACTP (13.0%), PARTIALP (9.7%), and ALLELEP (3.0%). ₁ Only 30.3% of AMR genes were core, compared with 46.8% for the pangenome baseline (odds ratio 0.49, chi-squared=23,117, p≈0); the auxiliary genome was 2.2x enriched for AMR, with 33.6% of AMR genes classified as auxiliary versus 15.3% of the baseline. ₁ Among 4,252 species containing at least 5 AMR clusters, 63.7% had AMR less core than their species baseline (Wilcoxon p=1.1e-130; mean difference −0.102). ₁ This supports the broader environmental-resistome result that AMR is disproportionately allocated to accessory genomic compartments rather than merely mirroring each species’ general pangenome structure. ₁

The conservation pattern differs among classes. Beta-lactamases were 54.9% core and significantly enriched relative to the species baseline (p=7.7e-74), regulatory genes were 6.5% core, and the acquired-resistance examples blaTEM, tet(C), and ant(2”)-Ia were each 0% core. Intrinsic efflux pumps such as emhABC were >95% core, whereas acquired efflux genes were accessory. ₁ The atlas describes blaTEM, tet cassettes, and aminoglycoside-modifying enzymes such as ant(2”)-Ia as fully accessory, often 0% core, appearing as singletons or in a minority of genomes, and calls them horizontally transferred elements that spread under antibiotic selection pressure; transfer and selection are interpretations rather than measurements in this analysis. ₁ In the within-species analysis across 1,305 species and 180,025 genomes, 51.3% of AMR gene-species occurrences were rare, occurring in <=5% of strains; 41.3% were variable, occurring in 5–95%; and 7.5% were fixed, occurring in >=95%. The median variability index was 0.526 and median pairwise Jaccard distance—a distance based on shared versus differing gene repertoires—was 0.435. ₃ The discoveries digest reports the same 51.3% rare, 41.3% variable, and 7.5% fixed split across 1,305 species and 180,025 genomes, with median within-species pairwise Jaccard distance of 0.435. ₄ These findings refine the intrinsic/acquired distinction: conservation is a graded, sampling-dependent property, not an absolute classification.

The environmental-resistome analysis similarly found clinical species with 68% accessory AMR versus 43% in soil. Its environment effect was significant (Kruskal–Wallis H = 506.0, p = 4×10⁻¹⁰⁷, η² = 0.036); mean core/accessory fractions were 57.1%/42.9% in soil, 54.4%/45.6% aquatic, 49.7%/50.3% host-associated, 40.4%/59.6% other environmental, 32.4%/67.6% clinical, and 19.7%/80.3% human-gut. ₂ The atlas likewise reported clinical AMR as 30.8% core versus 58.1% in soil and 63.1% in plant species. ₁ Together these results support the hypothesis that clinical datasets contain more acquired or mobile resistance and that analyzed environmental datasets contain a larger intrinsic component, but they do not establish that every environmental resistance gene is intrinsic or every clinical gene acquired. The environmental-resistome report reads core AMR as chromosomally encoded (intrinsic) and accessory AMR as mobile (acquired) by analogy to the river-metagenome pattern of Jiang et al. (2024). That mechanism was not established for each cluster in this analysis. ₂

The atlas comparison included only 7,838 of 14,723 AMR-carrying species with non-Other/Unknown classifications; 46.8% remained Other/Unknown because of sparse and inconsistent free-text isolation_source metadata. ₁ This limits environmental interpretation. Among 2,684 species with at least 3 genomes and AlphaEarth embeddings, environmental diversity predicted AMR count (Spearman rho=0.466, p=1.6e-144), while environmental diversity and AMR core fraction were negatively correlated (rho=−0.173, p=1.8e-19). ₁ Because AlphaEarth coverage was only 28% of genomes and biased toward genomes with geographic metadata, these results suggest a hypothesis about niche breadth and resistance accumulation rather than proving a causal route from environmental diversity to acquired resistance. The atlas authors themselves propose, from the rho=0.466 result, that niche breadth enables resistance accumulation because species encountering diverse environments and microbial communities have more opportunities to acquire resistance genes by horizontal transfer, aligning this with the environmental-reservoir hypothesis of Larsson & Flach (2022); this mechanism was proposed, not tested. ₁

AMR clusters were enriched in COG V (Defense mechanisms), 7.05x relative to baseline and representing 14.9% of AMR clusters versus 2.1%, and COG P (Inorganic ion transport), 1.93x and representing 10.7% versus 5.6%. COG means the Cluster of Orthologous Groups functional classification. ₁ The comparison used 77K AMR clusters with eggNOG annotations against an 86M baseline; COG J (Translation) was 1.50x enriched, which the atlas links to ribosomal protection proteins such as erm and helR, while replication (L: 0.12x), lipid metabolism (I: 0.10x), and cell motility (N: 0.08x) were strongly depleted. ₁

AMR density is also taxonomically concentrated, which the atlas describes as highly phylogenetically structured. At genus level, klebsiella led with 206 AMR clusters per species, followed by Salmonella (198), Citrobacter (134), and Enterobacter (93); Gammaproteobacteria contained 45% of all AMR clusters (37,752/83,008), and the top families were Enterobacteriaceae (37.5 AMR/species) and Staphylococcaceae (37.3 AMR/species), both dominated by clinical pathogens. By phylum, the atlas table lists species count, total AMR, mean AMR per species, and percentage of species with AMR as 4,992, 42,904, 8.6, and 67.0% for Pseudomonadota; 1,401, 12,614, 9.0, and 65.3% for Bacillota; 2,036, 9,027, 4.4, and 64.2% for Actinomycetota; 2,254, 8,776, 3.9, and 57.5% for Bacillota_A; and 1,835, 5,458, 3.0, and 50.6% for Bacteroidota. ₁ Within-species AMR variability differed modestly among phyla, with median variability index 0.487 in Bacillota (248 species), 0.500 in Bacillota_C (21), 0.509 in Bacteroidota (100), 0.533 in Pseudomonadota (591), 0.533 in Actinomycetota (133), and 0.600 in Bacillota_A (184). ₃ Because clinically dominated taxa lead these rankings, this concentration supports the caveat that environment contrasts are entangled with phylogeny rather than independent of it. _1,2

Catalog scope and detection

AMRFinderPlus detection methods were HMM at 51.5%, BLASTP at 22.7%, EXACTP at 13.0%, PARTIALP at 9.7%, and ALLELEP at 3.0% of hits. Bakta product annotations and eggNOG hits were present for 93.0% of AMR clusters, while 7.0% had Bakta annotations only. Among sparsely annotated one-source clusters, 55.4% were singletons compared with 34.6% of two-source clusters. Zero AMR clusters had Pfam hits in the bakta_pfam_domains table. ₁ These results support structural-annotation-gap: singleton calls may be true species-specific genes or annotation artifacts, and detection modality affects observed resistome composition.

Mechanism classification in the atlas used keyword matching against AMRFinderPlus product descriptions rather than CARD Antibiotic Resistance Ontology mapping. Its table contained 18,448 Other/Unclassified hits (22.2%), 15,755 enzymatic-inactivation hits (19.0%), 14,338 efflux hits (17.3%), 12,824 beta-lactamase hits (15.4%), 6,955 target-modification hits (8.4%), 6,852 oxidoreductase hits (8.3%), 4,964 cell-wall-modification hits (6.0%), and 2,848 regulatory hits (3.4%). ₁ The same table gives core percentages of 24.0% for Other/Unclassified, 28.2% for enzymatic inactivation, 30.9% for efflux, 54.9% for beta-lactamase, 19.6% for target modification, 15.4% for oxidoreductase, 45.2% for cell-wall modification, and 6.5% for regulatory hits. ₁ The environmental-resistome analysis instead classified 15,550 of 82,908 clusters (18.7%) as mechanism-unassigned and excluded them from mechanism fractions; these percentages arise from different datasets and classification tables, not a numerical contradiction. _2,1 In the environmental-resistome scheme, tetracycline-resistance genes were split between efflux (tet(A-E,K,L)) and target modification (tet(M,O,Q,W) ribosomal protection proteins). ₂

The broad atlas covered 27,690 pangenome species and found AMR diversity associated with environment, prevalence class, phylogeny, lineage-associated ecotypes, and tightly co-inherited resistance islands. _1,3

Environmental burden and metadata resolution

Clinical-source species had a median of 5 AMR gene clusters, compared with 2 for soil, aquatic, and host-associated species. Human-gut species resembled clinical species, also with a median of 5. The environment effect was significant (Kruskal–Wallis H = 781.9, p = 9.4×10⁻¹⁶⁷, η² = 0.056), with 13 of 15 pairwise comparisons significant after false-discovery-rate (FDR), a correction for multiple testing; the largest effect was clinical versus aquatic (rank-biserial r = −0.49). Robustness to 50%, 60%, 75%, and 90% majority-vote thresholds produced η² values from 0.044 to 0.056. ₂ The report presents this as extending Gibson et al. (2015), who found that resistomes cluster by ecology in ~6,000 genomes, to 14,723 species across 293K genomes. ₂ This concerns annotated and classifiable AMR, not total resistance diversity, and does not show that individual genomes within a species share the majority environment or AMR profile.

The atlas supports clinical enrichment at broader scale: Human/Clinical species (n=2,248) averaged 10.6 AMR clusters per species, compared with 4.6 for Soil/Terrestrial (n=2,469), 3.9 for Aquatic (n=1,827), and 3.0 for Animal species (n=959) (Kruskal–Wallis H = 440, p = 7.0e-93); the test was restricted to the 7,838 (53.2%) of 14,723 AMR-carrying species that received a non-Other/Unknown classification across 6 categories. In this classified subset, clinical AMR was also less core (30.8%) than soil (58.1%) or plant (63.1%) AMR, which the atlas reads as clinical selection for acquired or mobile resistance. The atlas headline calls the clinical excess 2.7x; that multiplier is the report’s own label and is recorded alongside the displayed 10.6 and 4.6 means rather than reconciled with them. ₁ The environmental-resistome dataset contained 280,337 genomes with NCBI environment metadata; 93.5% were classified per genome and species-level majority voting covered 95% of species, or 13,981 species. Majority voting can erase heterogeneity when genomes from one species occur in multiple source categories. ₂

In the within-species resource, NCBI and rule-based BacDive classifiers found more AMR genes per genome in host-associated species than terrestrial or aquatic environments, with human-clinical isolates highest. The NCBI classifier assigned environments to 1,190/1,307 species (91%), the BacDive approximation classified 459/1,307 (35%), and 52.7% of genomes had no classifiable isolation_source. ₃ Both classifiers agreed in direction (Kruskal–Wallis, p < 0.05), but the report describes both as approximate and states that dedicated metadata curation would improve ecotype analyses. ₃

Among 823 species with genomes in at least 2 environments and at least 5 genomes per environment, clinical genome fraction predicted total AMR cluster count (Spearman rho = 0.465, p = 2.2×10⁻⁴⁵). Clinical-dominated species had a mean of 72.9 AMR clusters versus 16.4 in environmental-dominated species, a 4.4× difference (Mann–Whitney U p = 2×10⁻¹⁸), and higher grouped accessory fraction (93.6% versus 81.7%, p = 0.004); the continuous clinical-fraction/accessory-percentage correlation was borderline (rho = 0.065, p = 0.064). ₂ These are species-level summaries rather than direct genome-level environment-by-AMR comparisons; the planned per-genome Fisher’s exact test was replaced by this proxy because billion-row joins were computationally costly. ₂ The discoveries digest repeats the 823-species correlation (rho=0.465, p=2.2e-45) and notes that klebsiella-pneumoniae has 1,115 AMR clusters, only 7 core, across 13,637 genomes; the correlation does not establish causation. ₄

Among 2,684 species with at least 3 genomes and AlphaEarth embeddings, environmental diversity predicted AMR count (Spearman rho = 0.466, p = 1.6e-144), while environmental diversity and AMR core fraction were negatively correlated (rho = −0.173, p = 1.8e-19). ₁ AlphaEarth coverage was only 28.4% of genomes in the ecotype analysis. Among 13,381 genomes spanning 224 species, only 172 species produced correlation results; its median partial correlation for environment was 0.0025 versus 0.0143 for phylogeny, and no significant environmental effect was detected for 156 of 172 species (90.7%). It found no significant difference between environmental and host-associated bacteria (p=0.66), while noting that host-associated coordinates may describe collection sites rather than organisms’ microenvironments. ₅

The evidence therefore supports an association between environmental composition and AMR burden, but does not identify which individual genomes acquired resistance, whether environmental transitions drive acquisition, or how much AMR variation exists among genomes sharing a species name. ₂ Only 2 species had sufficient within-species environmental diversity for chi-squared testing after strict expected-frequency criteria. ₃ Nevertheless, 190 of 974 species with at least 15 genomes formed at least 2 distinct AMR ecotypes using UMAP, a nonlinear dimensional-reduction method, followed by DBSCAN; median silhouette score was 0.620. _2,3 Case-study plots for klebsiella-pneumoniae, staphylococcus-aureus, and salmonella-enterica showed visible environmental structuring, but sparse metadata make this a hypothesis rather than a general environment–ecotype finding. In the environmental-resistome case-species table, K. pneumoniae had 13,637 genomes, 1,115 AMR clusters, 7 core and 1,108 accessory (99%), with clinical as the dominant environment (80%). S. aureus had 13,274 genomes, 642 clusters, 9 core and 633 accessory (99%), with clinical dominant (85%). S. enterica had 10,097 genomes, 836 clusters, 11 core and 825 accessory (99%), with host-associated dominant (31%). streptococcus-pneumoniae had 7,944 genomes, 59 clusters, 1 core and 58 accessory (98%), with clinical dominant (99.6%). mycobacterium-tuberculosis had 6,673 genomes, 44 clusters, 4 core and 40 accessory (91%), with clinical dominant (98%). ₂ These genome counts come from the environmental-resistome dataset and are not reconciled with the counts used in the strain-variation ANI analysis.

Mobile, accessory, and co-inherited resistance

Resistance islands provide a structural explanation for some accessory organization. They were detected in 705 of 1,305 species (54%), with 1,517 islands overall, mean size 6.2 genes, median size 4, maximum size 43, and mean phi=0.827; phi is a pairwise co-occurrence coefficient, so co-occurrence indicates linkage rather than co-selection or functional synergy. _4,3 Of the islands, 1,343 (88%) contained multiple resistance mechanisms. Efflux pumps occurred in 954 and enzymatic inactivation in 698; complete mechanism counts were Other/Unclassified 1,026, Efflux 954, Enzymatic inactivation 698, Oxidoreductase 694, Regulatory 502, Beta-lactamase 341, Target modification 293, and Cell wall modification 137. ₃ These modules support cofitness-network-architecture and complement resistance-island-coinheritance, but do not establish shared selection.

Within-species analysis found AMR less conserved than baseline: AMR was 30.3% core versus 46.8% for the pangenome baseline, the auxiliary genome was 2.2x enriched for AMR (33.6% versus 15.3%), and 63.7% of 4,252 species with at least 5 AMR clusters had AMR less core than their species baseline (Wilcoxon p = 1.1e-130; mean difference −0.102). ₁ It also found 77.3% of Core AMR genes fixed, 78.7% of Singletons rare, and Auxiliary genes 57.3% variable and 42.7% rare. AMR variability was weakly anti-correlated with pangenome openness (Spearman rho = −0.193, p = 2.2e-12). ₃ The discoveries digest describes this result, in which species with more open pangenomes have slightly lower AMR variability, as unexpected. ₄

Prophage-marker density was positively associated with AMR repertoire breadth across 4,770 species (Spearman rho=0.572, p<10^-300), while across 27,702 species, 14,669 (52.9%) carried both AMR and prophage markers. ₆ Proximity was weaker: AMR genes within 10 genes of a prophage marker were 67.6% accessory versus 65.5% for distal genes (OR=1.10, one-sided p=0.005, bootstrap 95% CI [1.024, 1.185]); ORs were 0.78 at 3 genes, 0.92 at 5, 1.10 at 10, 1.19 at 15, and 1.28 at 50, with only 33 of 74 testable species above 1 and median species-level OR=0.85. ₆ The report presents its reference-genome analysis as a genomic counterpart to Chen et al. (2018), who found resistance-gene and mobile-element co-occurrence on assembled contigs from river metagenomes, including associations with prophages. It describes this as confirming that the resistance-gene–prophage association is encoded in reference genomes rather than only in environmental assemblies; the modest and heterogeneous local-proximity results above qualify that reading. ₆

The Gene Function Ecological Agora found, across 18,989 GTDB-r214 species representatives, 17,073,194 Sankoff-parsimony gain events and recent-to-ancient acquisition ratios of 24.5× for CRISPR-Cas, 10.3× for TCS histidine kinases, 9.0× for β-lactamases, and 2.3× for clean tRNA-synthetase controls. ₇ Its MGE-machinery baseline was 1.37% of KO-bearing gene clusters, versus 0.57% for Mycobacteriaceae mycolic-acid KOs and 0.00% for both Cyanobacteriia PSII and Bacteroidota PUL KOs; PSII neighborhoods had 10.91% of focal features with an MGE neighbor within ±5 kb versus a Poisson expectation of 10.6%. PUL and mycolic-acid scans were deferred because of memory failures. ₇ The Agora cites Forsberg 2012 as context for the intra-phylum antibiotic-resistance horizontal-gene-transfer pattern that motivated its M1 analysis. ₇

The T4SS–CAZy analysis extended the mobility framework beyond prophages. Among 30,497 high-quality environmental MAGs, 6,652 (21.8%) carried T4SS or conjugative machinery under a multi-marker definition. Ninety-two CAZy families showed elevated co-occurrence with T4SS loci within <=10 kb; gt2 glycosyltransferases was the top hit and occurred in 767 genomes with average length 5,041 bp. Associations were enriched in marine sediment (OR=5.5, q<10⁻⁹⁸), barley rhizosphere (OR=10.4), and maize rhizosphere (OR=4.1). ₈ CAZy genes were not detected on plasmids by ICEfinder, while 12 integrative mobilizable elements occurred among the top 100 accumulators; T4SS-positive genomes had 10× higher mobile genetic element density (p<0.001). GT2 trees contained 77 detected HGT events, including 32 normalized high-confidence cross-phylum events; Node_4915 spanned 8 phyla at maximum divergence 4.843 and contained 35 genes, 82.9% syntenic. Sixty-five events spanned 2 phyla and 12 spanned at least 3; divergence and synteny were negatively correlated (Spearman ρ = −0.615, p<0.001). ₈ These findings support environmental organization of mobility-related content but do not demonstrate T4SS-mediated transfer.

Environment-associated mechanisms and metal responses

The environmental-resistome dataset contained 26,220 clusters assigned to enzymatic inactivation, 22,223 to metal resistance, 11,067 to target modification, and 5,224 to efflux; 15,550 clusters (18.7%) were unassigned and excluded from mechanism fractions. ₂ All four tested mechanism classes differed significantly among environments after Benjamini–Hochberg false-discovery-rate (BH-FDR) correction (all q ~ 0). Metal resistance accounted for 45.0% of AMR in aquatic species and 44.0% in soil but 6.1% in human-gut species (η² = 0.107, the largest reported effect); target modification accounted for 43.6% in human-gut and 27.5% in clinical species but 6.2% in aquatic species (η² = 0.100). Statistics were H = 1498.4, p ~ 0, η² = 0.107 for metal resistance; H = 1394.9, p = 1.7×10⁻²⁹⁹, η² = 0.100 for target modification; H = 768.0, p = 9.8×10⁻¹⁶⁴, η² = 0.055 for efflux; and H = 266.0, p = 2.0×10⁻⁵⁵, η² = 0.019 for enzymatic inactivation. ₂

Across clinical, human gut, host-associated, soil, aquatic, and other environmental species, efflux fractions were 6.1%, 7.0%, 3.7%, 2.0%, 1.1%, and 2.7%; enzymatic inactivation 28.2%, 24.1%, 44.1%, 34.3%, 41.2%, and 33.6%; metal resistance 19.7%, 6.1%, 12.7%, 44.0%, 45.0%, and 24.9%; and target modification 27.5%, 43.6%, 22.8%, 10.1%, 6.2%, and 25.1%. ₂ These patterns are consistent with distinct selection regimes, but exposure and selection were not directly measured and incomplete annotation could distort mechanism profiles. The report proposes that niches select for different resistance strategies. In its account, soil and aquatic bacteria rely on metal resistance, while clinical and gut bacteria rely on target modification and, to a lesser extent, efflux. It links this to Fitness Cost of Antimicrobial Resistance Genes, citing that project as showing metal 44% accessory versus efflux 13% core. This ecological cause is an interpretation and was not directly tested. ₂ The attribution itself conflicts with its source: the fitness-cost report states that metal-resistance genes are 44% accessory, while efflux (13%) and enzymatic-inactivation (16%) genes are overwhelmingly core, so its efflux 13% figure is presented as the contrast with an accessory fraction, not as a core fraction. The environmental report’s ‘13% core’ label therefore contradicts the fitness-cost report’s description of efflux as overwhelmingly core; this label conflict is recorded rather than resolved. _9,2 The discoveries digest reports that mechanism strongly predicts conservation status, with metal resistance 44% accessory versus efflux 13% accessory (χ²=69.3, p=1.4e-13). This accessory-fraction contrast supports the fitness-cost report’s reading and, like it, conflicts with the environmental report’s ‘13% core’ wording. _4,9,2

The global biogeography analysis found that among 30,497 host-associated-filtered environmental MAGs, 22,356 had usable coordinates and 2.8% carried at least one metal-resistance type. The report cautions that this geolocated-set prevalence is substantially lower than implied by reporting on the 30,497-MAG filtered set, and concludes that metal resistance is rare in the global environmental MAG pool. In a 5° grid restricted to 289 cells with ≥5 MAGs, 11 hotspots (OR>2, q<0.05) and 3 coldspots (OR<0.5, q<0.05) were significant after Fisher testing and BH-FDR correction. The top hotspot was the Atacama/Andean region of Chile (atacama desert) at lat=-25°, lon=-70° (21.8% prevalence, OR=9.83, q=7.6e-12, n=101 MAGs). Eastern and central USA clusters at lat=40°, lon=-80° and -90° had OR=7.9 and 6.3, and an East/Southeast Asia cluster at lat=25-30°, lon=105-120° had OR=4.4–5.9. Biome-stratified prevalence was tested against the global baseline with Fisher’s exact test and BH-FDR. Soil was enriched (7,939 MAGs, 5.8%, OR=5.05, q=1.8e-82), marine samples were depleted (13,995 MAGs, 1.2%, OR=0.20, q=9.2e-80), and rhizosphere prevalence was 1.9% (422 MAGs, OR=0.66, q=0.30, NS). The report contrasts the 2.8% metal-resistance prevalence with 21.8% T4SS prevalence and concludes that AMRFinderPlus-detected metal-resistance genes and T4SS machinery are distinct features that are not uniformly co-distributed. ₁₀

Across 51,748 soil samples and nine metals—copper, cobalt, chromium, nickel, zinc, lead, arsenic, cadmium, and mercury—soil functional genomics identified 2,355 significant COG–metal associations at FDR < 0.05. Chromium and lead produced the strongest signals, while ABC and RND transporters and biosynthesis genes dominated top hits. ₁₁ A db-RDA, distance-based redundancy analysis, yielded R² = 0.799 and p = 0.005 using 999 permutations; after conditioning on batch and project effects, metal concentrations explained 80% of community COG profiles. Unconditional metal-only R² was not reported and may be substantially lower. Biome-stratified PGLS found distinct metal–COG relationships in soil, marine, and wastewater environments. ₁₁

Across 452 metal experiments involving 37 organisms and 14 metals, 98.1% of 317 organism–metal-pair gene-level correlations were positive (311/317), and 99.1% statistically significant (p < 0.05); all 15 pairs tested in at least 5 organisms had greater than 90% sign consistency, with no systematically negative pair. ₁₂ A separate classification found 4,177 (54.9%) of 7,609 metal-important records metal-specific, 2,888 (38.0%) general sick, and 544 (7.2%) metal+stress. Specificity fractions were 60.6% manganese (20/33), 60.5% molybdenum (185/306), 56.5% tungsten (173/306), 46.4% selenium (64/138), 55.9% cadmium (52/93), 51.9% copper (1,346/2,594), 50.2% cobalt (1,167/2,324), 49.3% chromium (132/268), 48.6% uranium (88/181), 47.2% zinc (843/1,786), 43.7% nickel (993/2,271), 42.4% aluminum (752/1,772), 32.7% mercury (35/107), and 21.9% iron (144/659). Seven of 31 metal-tested organisms could not be processed because locusIds did not match the fitness-matrix index; the 24 included organisms accounted for 7,609 of 12,838 records (59.3%), leaving 5,229 excluded. ₁₃

The discoveries digest supports a graded rather than binary interpretation of conservation: across 7,609 metal-important gene records, 4,177 (55%) were genuinely metal-specific, 38% general sick, and 7% metal-plus-stress; metal-specific genes were 84.8% core versus 90.2% for general-sick genes (Cochran–Mantel–Haenszel p=0.011). The detailed analysis found pooled core fractions of 84.8% (2,969/3,500), 90.2% (2,183/2,420), and 94.3% (467/495), respectively. _4,13 All categories were core-enriched, but metal-specific genes were less enriched than general-sick genes. Metal-specific genes matched metal-resistance keywords at 12.2% versus 7.8% for general-sick genes (Fisher exact OR=1.64, p=2.4e-8). Approximately 14% of protein-coding genes, estimated at approximately 82% core, were putatively essential and absent from fitness data, biasing all categories toward core enrichment and making reported deltas conservative. ₁₃

Candidate prioritization identified UCP030820 (OG01015) as metal-specific in 2/3 organisms across 7 metals (67%) with mean sick rate 0.021; YebC (OG01383) in 7/12 organisms across 6 metals (58%) with mean sick rate 0.056; and DUF1043/YhcB (OG03264) in 3/6 organisms across 5 metals (50%) with mean sick rate 0.054. YebC’s proposed translation-bottleneck mechanism is a hypothesis. Among 149 novel metal-candidate families, 45.6% had dominant specificity of metal-specific versus 58.2% of annotated families (Fisher exact OR=0.60, p=0.003); ICA, independent component analysis, identified 0 metal-specific modules, attributed to scale mismatch and sparse experiments rather than absence. ₁₃

Phylogeny, ecotypes, and integration

Environment effects persisted within 5 of 6 tested major phyla—Pseudomonadota, Bacillota_A, Actinomycetota, Bacteroidota, and Bacillota—while Chloroflexota was non-significant; Bacteroidota had the largest within-phylum effect (η² = 0.130). At family level, 20 of 141 testable families (14%) showed significant within-family effects after FDR correction, including Enterobacteriaceae (q = 3×10⁻²¹), Bacteroidaceae (q = 1.3×10⁻¹⁷), Lachnospiraceae (q = 2.5×10⁻¹²), and Pseudomonadaceae (q = 3.9×10⁻¹²). ₂ These controls refine rather than eliminate confounding. The report states that environment and phylogeny remain deeply entangled and that some families, such as many Enterobacteriaceae, are almost entirely clinical. Most families do not span enough environments to test, so the 14% of families with significant within-family effects are the most convincing evidence. ₂

Within-species phylogenetic structure was assessed by comparing average nucleotide identity (ANI) distance matrices with AMR Jaccard distance matrices using a Mantel test, which evaluates correlation between two distance matrices. Across 1,261 species, 701 (55.6%) showed significant signal at FDR < 0.05; the median Mantel r was 0.247 and 87.8% were positive. Non-core putatively acquired AMR had stronger median Mantel r than core genes (0.222 versus 0.117; paired t-test t = −8.35, p = 7.0e-16, n = 489). ₃ This supports the hypothesis that acquired resistance elements can become stably maintained and vertically transmitted within particular lineages, but core genes are nearly universal by definition and therefore have little Jaccard-distance variance, which can suppress distance-based correlations independently of biology. The pattern establishes structure, not whether it reflects vertical inheritance, repeated acquisition, loss, or lineage-specific sampling. ₃ The discoveries digest reports the same contrast across 1,261 species (median r=0.222 for non-core versus 0.117 for core genes, paired t-test p=7.0e-16), labelling non-core genes acquired and core genes intrinsic; stable vertical inheritance is its interpretation. ₄

ANI extraction was limited to species with <=500 genomes, excluding escherichia-coli, with 15,388 genomes, and klebsiella-pneumoniae, with 14,240 genomes. Consequently, the phylogenetic proportions do not describe all species or genomes. ₃ The result is relevant to phylogenetic-confounding-of-pangenome-associations, because environment or functional associations may partly reflect shared ancestry.

Among the 606 genera with AMR data, MicrobeAtlas found Pagel’s λ = 0.260 for total AMR cluster count, λ = 0.441 for core AMR fraction, and λ = 0.335 for metal-type diversity, compared with λ = 0.787 for niche breadth (Levins’ B_std), a value estimated on a different denominator, across 1,264 bacterial genera with ≥ 3 OTUs; nitrification showed stronger conservation (λ = 0.939 in bacteria and λ = 1.000 in archaea). ₁₄ That project infers genus-level metal-type diversity, spanning mercury, arsenic, copper, zinc, cadmium, chromium, and nickel, from AMRFinderPlus annotations across 6,789 GTDB species. It links that diversity to niche breadth from a 464,000-sample 16S amplicon atlas, partitioning phylogenetic signal with Pagel’s λ and PGLS (phylogenetic generalized least squares, a regression that models shared ancestry). Broader metal repertoires were associated with broader ecological ranges beyond phylogenetic expectation. The authors call this association robust to genome size, species richness, prevalence filtering, OTU abundance filtering, and all 13 environment categories, but their own sensitivity analyses qualify that claim. Under strict (≥5%) prevalence filtering, PGLS n fell from 606 to 379 and β fell from +0.0215 to +0.0166, and the effect lost significance (p = 0.092) without changing direction. In all seven leave-one-metal-out models (excluding Hg, As, Cu, Zn, Cd, Cr, or Ni), β stayed positive but p > 0.10. The report says this loss of significance is consistent either with a genuinely distributed signal or with an original finding near the significance threshold that the compressed predictor range pushed below it, and that these alternatives cannot be definitively distinguished. The raw OLS correlation between mean metal-type diversity and Levins’ B_std was modest (r ≈ 0.21). The ordering core fraction λ > type diversity λ > cluster count λ is consistent with, but does not establish, a model in which core metal resistance is mainly vertically inherited while horizontal transfer and local metal exposure shape accessory accumulation and metal-type expansion. ₁₄ Among 2,659 species with AMR data and AlphaEarth embeddings, 52 of 64 dimensions correlated with AMR diversity at FDR < 0.05, with A34 top (rho = +0.24). Environmental cosine distance predicted AMR-profile Bray–Curtis distance (Mantel r = 0.098, p = 0.001), strongest for clinical species (r = 0.177), moderate for soil (r = 0.129), and weakest for aquatic species (r = 0.061); embedding coverage was 28% of genomes and dimensions were opaque. ₂

Within the MicrobeAtlas analysis, phylogenetic signal for metal AMR traits among 606 bacterial genera (GTDB r214) was intermediate and significantly non-zero for all three metrics: total AMR cluster count λ = 0.260 (p = 6.1×10⁻²⁸), core AMR fraction λ = 0.441 (p = 1.8×10⁻⁸), and metal type diversity λ = 0.335 (p = 1.1×10⁻²³). Niche breadth and habitat range were more phylogenetically conserved than metal AMR traits (λ_niche ≈ 0.79–0.91 versus λ_AMR ≈ 0.26–0.44); the niche-breadth estimate came from 1,264 bacterial genera with ≥ 3 OTUs rather than the 606-genus AMR subset. The report reads ecological range as deeply canalized by lineage identity and metal AMR as more labile, likely because AMR genes reside on mobile genetic elements and are exchanged among distantly related taxa; that attribution is an interpretation, not a direct measurement of horizontal gene transfer in this analysis. It cites Hemme et al. (2016) for recombinational hotspots for metal-resistance gene acquisition in contaminated groundwater and Gillieatt & Coleman (2024) for mobile-genetic-element transfer of AMR genes under metal co-selection. ₁₄

After phylogenetic control via PGLS in 606 genera, the number of distinct metal types resisted per genus was the only metal AMR predictor of Levins’ B_std that survived Bonferroni correction (β = +0.021, SE = 0.0056, p = 1.5×10⁻⁴, ΔAIC = −12.4, λ = 0.708; threshold p < 0.0083 for 6 models); the robustness table lists this simple model as β = +0.0215 with BH-FDR q = 0.003. In the multi-predictor model, metal types retained β = +0.023 (table value +0.0231), SE = 0.0067, p = 5.5×10⁻⁴, q = 0.004, while total AMR gene burden and core AMR fraction were non-significant. Neither AMR clusters (β = −0.071, SE = 0.067, p = 0.290) nor metal types (β = −0.007, SE = 0.071, p = 0.923) predicted the number of environments, so the metal-type association is specific to Levins’ B_std rather than to a simple environment count. ₁₄

Total AMR gene cluster count and core AMR fraction did not predict niche breadth in either the simple or multi-predictor PGLS models. The report interprets this to mean that the breadth of the metal resistance repertoire, rather than its depth (many genes for few metals), is associated with ecological versatility: genera with many AMR genes for a single metal type are not ecological generalists, whereas genera spanning multiple metal resistance families are. This refines the use of AMR burden elsewhere on this page: in this genus-level analysis, gene count and core fraction did not carry the niche-breadth signal that metal-type diversity did. The report frames the positive association (β = +0.021, Bonferroni-corrected) as consistent with a metabolic versatility hypothesis, which remains a hypothesis. That passage describes β as “SD per SD,” but elsewhere the report describes β per SD increase in metal diversity and specifies z-scored predictors rather than a z-scored response, so the outcome units are unresolved. It cites Qi et al. (2022) as independent soil evidence that multiple heavy metal contamination favors microbial generalists as network keystones. ₁₄

Genome size did not remove the metal-type effect. With log genome size as a covariate (R5; n = 527), metal types had β = +0.0218, SE = 0.0061, p = 3.6×10⁻⁴, λ = 0.629, q = 0.003, while log genome size independently predicted B_std (β = +0.0263, SE = 0.0069, p = 1.5×10⁻⁴). The genome-size-only baseline (β=+0.029, p=2.7×10⁻⁵) had AIC=−630.3 versus AIC=−641.1 with metal types added, ΔAIC = 10.8 in favour of inclusion. Controlling simultaneously for species richness and genome size (R6; n = 527), metal types remained significant (β = +0.0218, SE = 0.0061, t = 3.60, p = 3.5×10⁻⁴, q = 0.003 BH-FDR), as did genome size (β = +0.0255, SE = 0.0069, t = 3.68, p = 2.6×10⁻⁴). The report’s prose and reviewer response round the R5 coefficient to β = +0.022 (p = 3.6×10⁻⁴) and state that this rules out genome size as the primary confounder. Within-genus SD of metal types was not a significant predictor (β = +0.0074, SE = 0.0055, p = 0.181), while the mean metal-type effect persisted (β = +0.0189, SE = 0.0060, p = 0.0016), so a few outlier species do not appear to inflate the association. All 13 leave-one-environment-out models were nominally significant (p < 0.05). Excluding aquatic, the most heterogeneous and over-represented category (40,353 OTUs; 40.8% of the dataset), reduced β from +0.021 to +0.0085 (−59%), with p = 0.031. ₁₄

The report’s limitations section summarizes the strict 5% within-environment prevalence filter (R4) as confirming signal direction (β = +0.017 vs +0.022 original), with significance reduced because 37% fewer genera passed, and states that the effect is not explained by rare transient detections. Because the R4 coefficient was nonsignificant (p = 0.092), direction alone does not establish robustness under that filter. Archaeal PGLS was positive but nonsignificant for metal types (n = 48, β = +0.0145, SE = 0.0198, t = 0.73, p = 0.467, λ = 0.567, BH-FDR q = 0.548), AMR clusters (β = +0.0146, SE = 0.0209, t = 0.70, p = 0.488, λ = 0.567), and core fraction (β = +0.0002, SE = 0.0211, t = 0.01, p = 0.991, λ = 0.553). The bacterial association has therefore not been shown to extend to archaea in this small sample. ₁₄

The report describes the PGLS coefficient as a partial correlation after phylogenetic correction. The reported β is in fact a phylogenetic regression coefficient rather than a partial-correlation coefficient, but on either reading it is not a causal estimate. Both variables are also inferred proxies, not confirmed ecological range or physiology. Genus-level aggregation of OTU niche scores can make a genus appear broad-niched because different species within it occupy different habitats. Levins’ B_std also conflates habitat use with sampling intensity and primer bias (V3–V4 or V4), and it cannot correct for missing environments or detection-probability differences. The report therefore states that β reflects covariation between inferred niche breadth and inferred metal tolerance capacity. Three directional scenarios fit the data: metal tolerance enabling habitat expansion; ecological generalism enabling gene acquisition because broad-niche genera encounter more environments, more HGT donors, and higher metal concentrations; or a shared driver such as genome size, metabolic versatility, or biofilm capacity. The cross-sectional, observational design cannot distinguish them. The report argues that a nonsignificant total AMR cluster count alongside significant metal-type diversity slightly favours scenario 1 or 3 over a simple genome-size-drives-everything model. It also states that the association remains correlational and that time-series genomic data or longitudinal metal-enrichment experiments on defined communities would be needed to establish direction. ₁₄ This directional ambiguity refines the atlas proposal that niche breadth enables resistance accumulation through horizontal transfer: the MicrobeAtlas data are equally consistent with the reverse direction, and neither direction has been tested. _14,1

Groundwater field checks are related to the global PGLS result but not equivalent to it. In Track A, using 1,624 groundwater samples, metal-type diversity was associated with groundwater prevalence (Spearman ρ = +0.112, 767 genera, p = 0.0019). Top- versus bottom-quartile metal-type genera (213 vs 429 genera) differed by Mann–Whitney test (p = 0.007, one-sided greater), with median groundwater prevalence 0.81% versus 0.62% (ratio = 1.31×; the reviewer response phrases this as 31% more prevalent). Fold enrichment in groundwater versus non-groundwater samples was not significant (ρ = +0.042, 767 genera, p = 0.242). The report calls this a meaningful null: Track A confirms the prevalence association but not groundwater-specific enrichment. That gap could reflect a general breadth effect rather than groundwater specificity, or an underpowered 1,624-sample subset, so Track A does not uniquely support the metal-contamination mechanism. The report’s description of the groundwater surveys as an independent field-observation cross-validation of the PGLS finding establishes consistency, not mechanism. ₁₄

Track B ran the amplicon pipeline on all 133 PRJNA1084851 ENIGMA samples from the oak-ridge-field-research-center (24,295 OTUs; 133/133 QC pass). Community-weighted mean (CWM) metal-type diversity differed across 8 wells (Kruskal–Wallis H = 29.10, n = 133 samples, p = 0.0001). It was higher in wells FW215 and FW216, which lie within the ORFRC U/NO₃ contamination plume, and increased with time point after carbon amendment (Spearman ρ = +0.383, n = 133, p < 0.0001; ρ = +0.576, n = 40, p = 0.0001 within FW216). CWM is not a whole-community measurement. It used the 1,637 OTUs (16.8% of reads) joinable to genus metal-diversity data, and the remaining ~83% of reads were excluded. If uncovered lineages were predominantly metal-sensitive, CWM would be biased upward, and robustness to coverage-fraction variation has not been formally tested. The report states that Track B and Track A’s 1,624-sample prevalence result both independently confirm the PGLS prediction. However, higher plume-well CWM is consistent with, not proof of, selection for broader metal resistance. The proposed explanation, that carbon amendment creates reducing conditions that shift speciation of redox-active metals (U, Fe, S), is a hypothesis. Sulfurimonas is a highly prevalent groundwater generalist whose abundance the report attributes to electron-acceptor availability rather than metal resistance. It is a counterexample showing that not all ecologically dominant groundwater taxa are metal-diverse, though that attribution is itself an interpretation. ₁₄

The report also proposes, but has not completed, a site-level test of whether MicrobeAtlas genera with high metal-type diversity are enriched in metal-contaminated sample metadata (mining-impacted, biosolid-amended, estuarine) versus clean reference environments. It also lists untested, falsifiable predictions that each of 8 shortlisted OTUs should increase in relative abundance by ≥2-fold vs. metal-free controls after 7–14 days under specified metal stress. ₁₄

The ecotype analysis found phylogeny dominated gene-content similarity in 60.5% of species, while environment dominated in 39.5%; environmental adaptation may act on specific gene subsets rather than whole-genome content. ₅ This is compatible with the AMR findings but limits inference that species-wide gene-content structure directly represents environmental differentiation.

The BERDL Data Atlas inventories 1,011,650,903 KBase pangenome gene rows spanning 293,059 genomes and 27,690 GTDB species clades or pangenomes, alongside 27,410,721 FitnessBrowser measurements, 83,287 AlphaEarth embeddings indexed to KBase genomes, and 97,334 BacDive strain phenotype profiles. It identifies 536 schema-level cross-tenant bridges across 29 canonical join keys, but only the FitnessBrowser–AlphaFold bridge was sample-validated. ₁₅

BacDive–GTDB matching found 42,227 of 97,334 strains (43.4%) mapped to 6,426 GTDB species with metal-tolerance scores; contamination-associated isolates generally had higher scores. A separate analysis matched 37,368 strains (38.4%) to 5,647 GTDB species. The full predictive model reached R² = 0.633 after adding n_metal_clusters, whereas taxonomy alone explained 35.4%, phenotype features alone 16.3%, and taxonomy plus phenotype 34.5%; in 5-fold phylogenetic-blocked cross-validation, the full model had R² = 0.633 and RMSE = 0.028, taxonomy alone R² = 0.354 and RMSE = 0.038, and phenotype alone R² = 0.163 and RMSE = 0.043. _16,17

The PGP pangenome analysis extends environmental gene-distribution evidence beyond AMR. Among 27,702 species, 11,272 carried at least one of 13 PGP markers; acdS, pqqC, and hcnC were enriched in soil/rhizosphere species, whereas nifH was depleted. acdS prevalence was 15.8% versus 2.6% in other environments (OR = 7.02, q = 5.1e-62), pqqC 43.8% versus 21.2% (OR = 2.90, q = 2.8e-53), hcnC 11.3% versus 6.4% (OR = 1.85, q = 6.1e-08), and nifH 12.9% versus 19.7% (OR = 0.60, q = 5.5e-08). All 13 PGP genes had higher core fractions than the 46.8% genome-wide baseline; pqqC was 81.5% core and acdS 70.4% core, while mean PGP accessory fraction was 29.7% versus 53.2% genome-wide. PGP richness correlated negatively with pangenome openness (Spearman ρ = −0.195, p = 2.0e-97, n = 11,272). This contrasts with AMR, for which only 30.3% of genes were core. _18,1

Fitness and functional interpretation

A Fitness Browser cross-reference identified 178 AMR genes across 37 Fitness Browser organisms and 29,386 fitness measurements using a DIAMOND-based pangenome link table at 100% sequence identity. AMR genes had a median fitness of −0.007 versus −0.012 for the non-AMR baseline (Mann–Whitney p=3.7e-6); beta-lactamases had a median fitness of −0.001, while singleton AMR genes had a median of −0.019. ₁ These measurements are consistent with well-integrated intrinsic resistance genes being less burdensome under tested laboratory conditions, but do not establish that recently acquired mobile resistance is cost-free in clinical pathogens. The source separately reports fitness data for 162 AMR genes in 36 Fitness Browser organisms and does not reconcile that count with the 178-gene, 37-organism analysis. The 100% DIAMOND identity threshold is conservative and may miss closely related resistance variants, including alleles differing by a single synonymous substitution. ₁ These limitations connect to antimicrobial-resistance-fitness-cost and condition-specific-fitness.

The conservation pattern is not itself a fitness-cost measurement. Among 801 AMR genes with fitness measurements, core/intrinsic genes (N = 638) and accessory/acquired genes (N = 163) had identical mean fitness values of −0.024, Cohen’s d = 0.002, and Mann–Whitney U p = 0.33. ₉ This contradicts any claim that accessory status alone establishes a larger baseline knockout-fitness cost and refines pangenome-conservation-fitness-decoupling: accessory status and laboratory fitness effect are related dimensions, not interchangeable measures of evolutionary cost.

The field/lab analysis found field-stress and field-core genes enriched in the core genome, but field or lab fitness alone predicted poorly and ICA-module conservation was not associated with field-condition activity (rho=0.071, p=0.62). Fitness importance and gene length were more informative; heavy-metal-important genes were 71.2% core and lab-antibiotic-important genes 73.4%. It also found 678 essential genes at 80.1% core but lacking transposon fitness data, versus a 76.3% core baseline among 2,725 analyzed non-essential genes. ₁₉ The report interprets antibiotic and metal resistance as accessory-enriched and consistent with recently acquired adaptive traits carried on mobile genetic elements, although its class comparisons against baseline were not statistically significant. Mobility and recent acquisition were not demonstrated, and it remains unresolved whether these accessory resistance genes sit on mobile elements, were recently acquired, or are shared with co-occurring species at Oak Ridge. ₁₉

Temporal and sampling constraints

Among 513 species with >=20 genomes spanning >=3 years after 1990, none showed a significant temporal trend in AMR gene count after Benjamini–Hochberg FDR correction. Slopes were approximately symmetrically distributed, with 251 positive and 262 negative slopes. ₃ This null result is not evidence that AMR accumulation lacks temporal trends because collection dates were sparse and noisy and only 70% of genomes had parseable dates.

The GTDB/NCBI collection was heavily biased toward clinical and human-associated isolates, particularly for klebsiella-pneumoniae, staphylococcus-aureus, and escherichia-coli, while environmental species were underrepresented. AMRFinderPlus therefore missed novel mechanisms absent from its database, and collection bias may affect both lineage structure and generalization to environmental populations. ₃ The environmental-resistome analysis reinforces this caveat: NCBI sampling overrepresents clinical isolates, so the clinical-AMR finding partly reflects that clinical species are sequenced because they carry AMR, while soil and aquatic species are undersampled; AMRFinderPlus targets known resistance genes, potentially undercounting novel environmental resistance and favoring well-characterized clinical resistance, and only 14% of tested families showed significant within-family environment effects after phylogenetic control. ₂

Literature context and untested links

The environmental-resistome report frames its results against prior literature, but the causal readings it draws are interpretive. It cites Gibson, Forsberg & Dantas (2015; PMID: 25003965) for ecology-structured resistome clustering in ~6,000 genomes. It cites Forsberg et al. (2014; PMID: 24847883) for phylogenetically structured soil resistomes with rare mobility elements, and aligns its soil-core versus clinical-accessory pattern with that work. It cites Surette & Wright (2017; PMID: 28657887) for the environment as the largest resistance reservoir, and from this proposes an ancient soil baseline plus anthropogenic clinical selection. It cites Van Goethem et al. (2018; PMID: 29471872) for antibiotic-resistance genes in pristine Antarctic soils, which it reads as an ancestral baseline. Forsberg et al. (2012; PMID: 22936781) reported perfect nucleotide identity between soil and clinical-pathogen resistance genes, but this project did not test for that. AMR clusters shared by soil and clinical species are proposed as future work; such clusters would be candidates for recent cross-niche horizontal gene transfer, not demonstrations of it. ₂

Tensions

The strongest tension is between robust ecological association and the absence of a direct exposure-to-selection test. Environment effects explain 2–13% of AMR-composition variance (η² = 0.02–0.13) despite highly significant tests (p ~ 0), and the report suggests phylogeny likely explains much more; phylogeny, sampling, isolation, gene mobility, and annotation coverage may explain part of these effects. The association is correlational. Antibiotic exposure could increase AMR in clinical species, AMR could contribute to clinical status through virulence–resistance co-selection, or both could reflect correlated host-associated traits. Majority-vote classification collapses within-species variation, so a species classified as clinical may include environmental strains; the report states that sensitivity analysis at 60-90% thresholds mitigates this. ₂

The environment–AMR association persisted after majority-vote thresholds and phylum- and family-level controls, supporting an ecological signal beyond a simple species-label artifact. ₂ However, only 20 of 141 testable families (14%) showed significant within-family effects after FDR correction, and whole-genome ecotype analysis found environmental effects significant and positive in 12 species (7.0%), significant and negative in 4 species (2.3%), and absent in 156 species (90.7%). _2,5 The strain analysis found AMR ecotypes in 19.5% of eligible species, but metadata were sparse and only 2 species passed strict testing criteria. ₃ Thus broad ecological association remains compatible with weak or unresolved within-species environmental structure.

The positive relationship between AMR and phylogenetic distance in 701/1,261 species may reflect lineage-associated acquisition or uneven sampling of lineages across environments; available analyses do not separate these explanations. Geographic coordinates were often missing or imprecise, and partial correlations assume linear relationships between distance matrices, so weak environmental signal cannot distinguish absent adaptation from inadequate environmental representation. _3,5

The broad AMRFinderPlus catalog and narrow antibiotic-resistance language remain in tension: the catalog detects metal and stress systems, while the narrow interpretation is weakened by mercury- and arsenic-resistance families and the 22.2% Other/Unclassified category. The environmental-resistome analysis reports 15,550 unassigned clusters (18.7%), not 18,448 (22.2%); differing denominators and vocabularies must not be silently reconciled. _1,2

The intrinsic/acquired interpretation is also limited by conservation and fitness evidence. The atlas links core status to intrinsic examples and accessory status to acquired examples, while the fitness analysis found identical mean fitness values of −0.024 for core/intrinsic and accessory/acquired groups (Cohen’s d = 0.002, p = 0.33). _1,9 Thus genomic compartment supports a mode-of-maintenance hypothesis but does not by itself establish resistance origin, mobility, or cost.

Metal validation remains in tension. BacDive reports higher metal-tolerance scores among contamination-associated isolates, including Cohen’s d = +1.00 for heavy-metal contamination (n = 10), whereas cross-resistance found no species-scale correlation between multi-metal tolerance and metal-associated isolation (Spearman rho approximately −0.02, p > 0.8), leaving only 20 independent species. _16,12 The global MAG result is related but non-equivalent: soil had 5.8% prevalence and OR=5.05, marine samples 1.2% and OR=0.20; MicrobeAtlas found metal-type diversity associated with groundwater prevalence (ρ = +0.112, p = 0.0019) but not groundwater-specific fold enrichment (ρ = +0.042, p = 0.242). _10,14

Soil COG associations versus metal-specific interpretation are unresolved because chromium, copper, lead, and zinc co-vary in many industrial soils. ₁₁ The discoveries digest reports efflux at 21% of AMR in human-gut species and 1% in aquatic species, while the environmental-resistome analysis reports 7.0% and 1.1%; the 21% versus 7.0% and related 13% versus 44% core/accessory values require harmonized entity sets, filters, annotations, and denominators. _4,2,9 The digest attaches η²=0.127 to its efflux contrast, whereas the environmental-resistome report gives efflux η² = 0.055. The digest also reports metal resistance at 45% in soil/aquatic versus 6% in human gut (η²=0.107) and 68% accessory AMR in clinical versus 43% in soil species, without documenting denominators. _4,2

The environmental-resistome report’s interpretive prose also conflicts with its own tables, and these conflicts are recorded here rather than resolved. One interpretive passage says soil and aquatic metal-resistance genes are more often core (intrinsic), whereas an earlier passage calls metal resistance accessory because it is only needed in metal-rich environments. The prose gives clinical and gut efflux fractions of 15-21% and target-modification fractions of 19-29%, while the mechanism table gives clinical/human-gut efflux of 6.1%/7.0% and target modification of 27.5%/43.6%. The prose says enzymatic inactivation is not dominant in any environment, yet the table gives it the largest listed fraction in clinical (28.2%), host-associated (44.1%), and other environmental (33.6%) categories. The literature-context passage attributes ecology-structured resistome clustering in ~6,000 genomes to Gibson, Forsberg & Dantas (2015; PMID: 25003965), claims to replicate it at 50× scale (14,723 species), and calls efflux and metal resistance the most environment-discriminating mechanisms, whereas the reported effect sizes rank metal (η² = 0.107) and target modification (η² = 0.100) above efflux (η² = 0.055). ₂

Prophage density and local proximity provide non-equivalent evidence: density was strongly associated with AMR breadth (rho=0.572; partial rho=0.464 after controlling for genome count), whereas local proximity was modest and heterogeneous, with median species-level OR=0.85. ₆ T4SS–CAZy evidence supports dissemination as a hypothesis rather than demonstrated transfer: it reports 92 elevated co-occurrences, 77 GT2 HGT events, 32 normalized high-confidence cross-phylum events, and 10× higher MGE density in T4SS-positive genomes, but no CAZy genes on plasmids by ICEfinder and only 12 IMEs among the top 100 accumulators. ₈

Open Directions

• Map the 83,008 AMRFinderPlus hits to CARD ARO terms; separate classical antibiotic-resistance genes from mercury-, arsenic-, and other stress-response genes; and repeat mechanism counts, core fractions, environmental contrasts, the 30.3% versus 46.8% core comparison, and the 2.2x auxiliary-genome enrichment. ₁
• Re-annotate the 15,550 mechanism-unassigned clusters with broader homology, profile-HMM, and structure-based searches; compare independent resistance resources; quantify false-negative rates by environment and mechanism; and test whether the 18.7% unassigned fraction changes mechanism effect sizes. ₂
• Validate singleton calls with homology, domain, and synteny analyses, asking whether the 55.4% singleton rate among sparsely annotated one-source clusters represents artifacts or genuine resistance. ₁
• Use the 280,337 genomes with NCBI environment metadata and a scalable per-genome AMR-by-environment contingency analysis, such as the originally planned Fisher’s exact test, to determine whether AMR enrichment remains after conditioning on species. ₂
• Use the 823 species in the multi-environment subset and hierarchical or mixed-effects models to separate within-species environment effects from between-species clinical representation, asking whether the 72.9-versus-16.4 mean AMR-cluster contrast persists within species. ₂
• Use deeply sampled species including klebsiella-pneumoniae, staphylococcus-aureus, and salmonella-enterica, with genome-level environment labels and accessory-cluster matrices, to test whether the reported 99%, 99%, and 99% accessory fractions are concentrated in particular environments. ₂
• Reanalyze the 190 species with at least 2 AMR ecotypes using curated isolation_source labels, phylogeny, and genome-level AMR matrices to test whether ecotypes correspond to environments after lineage composition is controlled. ₃
• Subsample deeply represented species and species with >500 genomes, repeat ANI–AMR Mantel analyses, and use phylogeny-aware models to separate lineage inheritance from repeated acquisition and loss. ₃
• Infer AMR gene gain and loss rates from species phylogenies and gene-family distributions; test whether low-core AMR families show turnover rates consistent with repeated acquisition and loss. ₁
• Test co-localization of accessory AMR genes with genomic islands, insertion sequences, and integrons, and determine whether the least-conserved AMR genes are physically associated with mobility features. ₁
• Curate collection dates and isolation_source metadata, then apply temporal regression and adequately powered within-species association tests to distinguish AMR change through time from sampling composition. ₃
• Use the 1,305 genome-by-AMR matrices and genomic-context mapping to test whether the 1,517 resistance islands reside on plasmids, chromosomes, integron boundaries, or insertion sequences, and whether physical context predicts persistence across lineages. ₃
• Extend Fitness Browser linking beyond 100% DIAMOND identity and measure fitness under antibiotic stress; test whether closely related acquired variants and mobile clinical resistance genes incur costs missed under current laboratory and sequence-matching filters. ₁
• Combine AMR ecotype assignments with virulence-factor profiles and metabolic pathway variation to test whether resistance-defined lineages also differ in pathogenicity-associated or ecological functions, and whether co-inheritance patterns predict future AMR gene co-acquisition. ₃
• Reanalyze AMR repertoires with resistance databases beyond AMRFinderPlus and replace keyword-based mechanism labels with CARD ARO mappings to quantify effects on rare, variable, fixed, core-fraction, and ecotype estimates. _3,1
• Reanalyze the 172 ecotype-analysis species using direct environmental metadata and alternative embedding distances, then test COG functional subsets rather than whole-genome gene content. ₅
• Harmonize core/accessory thresholds, mechanism annotations, genome sampling, the 21% versus 7.0% efflux estimates, and all denominator differences; test whether the clinical 68% versus soil 43% accessory contrast persists under matched sampling. _4,2,9
• Compare catalog-defined gradients against nmdc and mgnify environmental metagenomes after CARD ARO harmonization and metadata-stratified modeling. Ask whether accessory AMR genes seen in isolate genomes are also prevalent in community DNA (proposed by the atlas, not yet tested), and whether the intrinsic/acquired genomic-compartment pattern persists in community-level environmental resistomes rather than only in genome databases. _1,3
• Map resistance islands to plasmids, chromosomes, integrons, ICEs, and insertion sequences, then test whether mechanism composition predicts persistence or acquisition after controlling for phylogeny. _3,4
• Match BacDive accessions directly to pangenome genome_ids, expand MIC and growth-inhibition extraction, and apply PGLS or phylogenetic PCA; resolve the metal-isolation tension with matched metadata, KEGG/PFAM signatures, concentration-normalized fitness, non-metal stress controls, and independent validation. _16,17,12,13
• Fit partial-correlation models such as COG ~ Cr | Cu + Zn + Pb, classify COGs into resistance, stress, membrane, energy, and unknown categories, report conditional and unconditional db-RDA R², and test Moran’s I and spatial residuals. ₁₁
• Validate global biogeography with sample-accession checks, prevalence normalization by log(n_MAGs), biome-specific missingness tests, and sampling-effort correction; reanalyze MicrobeAtlas with finer habitats, alternative phylogenetic or mixed models, primer-detection correction, rarefaction, and experimental validation. _10,14
• Replace broad ENIGMA COG-fraction proxies with curated metal-stress gene sets while controlling for abundance, depth, location, date, and compositional structure; load registered SSO geochemistry samples into CORAL and repeat AMR profiling. _20,21
• Apply field/lab fitness–conservation analyses to additional organisms while controlling for gene length and essential-gene exclusion. ₁₉
• Execute value-space-validated NMDC–PROTECT and ENIGMA–PhageFoundry bridges and harmonize pangenome, AlphaEarth, BacDive, and environmental identifiers. ₁₅
• Test whether PGP soil-enrichment signals and AMR mechanism profiles persist after matched phylum, genome-size, sampling-effort, and environment-label controls; analyze pqqC–acdS and hcnA–hcnC neighborhoods and phylogenies. ₁₈
• Reanalyze genomes from klebsiella-pneumoniae, acinetobacter baumannii, pseudomonas-aeruginosa, and escherichia-coli with dedicated prophage predictors, scaffold-based distances, exhaustive sampling, and explicit plasmid/ICE controls; apply geNomad or PHASTER and test whether density associations persist after controlling for pangenome openness, genome count, phylogeny, plasmids, and ICEs. ₆
• Run the pending T4SS–CAZy synteny-threshold permutation test, validate Node_4915 by BLAST against NCBI nr, establish a housekeeping-gene null, calculate θ = OR(T4SS-CAZy) / [OR(T4SS) × OR(CAZy)], and perform experimental transfer assays with base-pair-resolution neighborhoods. ₈
• For species sampled across decades, such as staphylococcus-aureus and klebsiella-pneumoniae, test whether the accessory AMR fraction has increased over time; no temporal increase has been established. ₂
• Combine environment-AMR profiles with Fitness Cost of Antimicrobial Resistance Genes to test whether clinically enriched AMR genes are costlier because they were recently acquired, or cheaper because they have been under strong selection; this remains untested. ₂
• Use nmdc_arkin metagenome taxonomic profiles to infer community-level AMR load from pangenome-derived species AMR profiles, validating the species-level findings in real environmental communities; this validation is proposed, not completed. ₂
• Test whether accessory AMR clusters in clinical species are preferentially associated with plasmids, transposons, or integrative elements; this is proposed, not reported. ₂

See the project summaries: Environmental Resistome at Pangenome Scale, Fitness Cost of Antimicrobial Resistance Genes, Pan-Bacterial AMR Gene Landscape, Within-Species AMR Strain Variation, BacDive Isolation Environment × Metal Tolerance Prediction, BacDive Phenotype Signatures of Metal Tolerance, BERDL Data Atlas — Inventory, Topic Map, and Cross-Reference Synergies, discoveries, Ecotype Correlation Analysis, Contamination Gradient vs Functional Potential in ENIGMA Communities, SSO Subsurface Community Ecology — Spatial Structure, Functional Gradients, and Hydrogeological Drivers, Field vs Lab Gene Importance in *Desulfovibrio vulgaris* Hildenborough, Gene Function Ecological Agora, Gene-Resolution Metal Cross-Resistance Across Diverse Bacteria, Global Biogeography of Environmental Bacterial Metal Resistance, Metal-Specific vs General Stress Genes, Metal Resistance Ecology — Phylogenetic Conservation vs. Environmental Selection, PGP Gene Distribution Across Environments & Pangenomes, Prophage-AMR Co-mobilization Atlas, Soil Metal Concentrations Drive Functional Gene Shifts, and T4SS–CAZy Environmental HGT.