Cultivation and collection bias constrain ecological genomic inference

Ecological genomic inference is constrained when available genomes come disproportionately from culturable, described strains rather than the full diversity of environmental bacteria. ₁ The BacDive metal-validation project directly illustrates this limitation by linking curated isolation environments to genome-derived metal-tolerance scores through a species-level pangenome-integration bridge. ₁ The underlying analysis is documented in BacDive Isolation Environment × Metal Tolerance Prediction. ₁

The MicrobeAtlas metal-ecology analysis supports this broader concern by showing that ecological-range inference is itself sensitive to representation: MicrobeAtlas niche breadth is a sequencing-effort proxy rather than confirmed ecological range, and is affected by sampling intensity, primer bias, geographic and temporal non-uniformity, and missing environments. ₂ Its genus-level analysis also found that sequenced-genome coverage correlated with inferred metal-type diversity (r = 0.35, p = 2.6×10⁻¹⁹), while single-genome genera had lower metal-type counts than multi-genome genera (Mann–Whitney p = 5.1×10⁻¹⁰). ₂ Thus, both cultivated genome collections and environmental occurrence atlases can select which organisms and traits appear ecologically informative. ₂ The report states the remaining gap plainly: primer bias and missing environments cannot be corrected without a multi-primer, multi-region survey design. ₂

The plant microbiome analysis supports and extends this concern to host-associated collections: plant-compartment labels were inferred primarily from NCBI isolation_source metadata, only 7,995 of 293,059 genomes (2.7%) had plant-associated annotations, and endophytes comprised only 29 species. ₃ BacDive identified 2,482 plant-related strains but contributed zero additional genome upgrades beyond ncbi_env. ₃ These figures show that both sparse metadata and the cultivation history of available genomes can constrain ecological classification before genomic comparisons begin. ₃

The plant project’s own source accounting refines that sparsity figure. From 293,059 GTDB r214 genomes, plant association was determined via three complementary sources: NCBI isolation_source regex matching as the primary route, ncbi_env EAV (entity–attribute–value) table cross-validation, which upgraded 695 genomes, and BacDive isolation records, which identified 2,482 plant strains but produced 0 new upgrades. ₃ The retained panel was 1,136 species and 7,995 genomes. ₃

Sparsity then decides which compartments are testable at all. Only 7,995 of 293,059 genomes (2.7%) carried plant-associated annotations, and endophyte species (n=29) fell below the 30-species threshold. ₃ The report states that with more targeted isolation efforts the endophyte compartment, currently 29 species, could be expanded to enable a full four-way compartment comparison, so the missing compartment is a cultivation gap rather than a demonstrated ecological absence. ₃ Host labels are similarly uneven: 11,852 genomes had identifiable plant hosts parsed from NCBI isolation_source, and species-level aggregation yielded 1,307 species across 19 testable hosts with ≥10 species each — rice (286), Arabidopsis (184), wheat (122), maize (98), soybean (84), potato (58), tomato (57), sugarcane (47), lettuce (34), banana (30), pepper (25), cotton (24), grape (22), strawberry (18), peanut (17), citrus (14), tea (13), tobacco (12), apple (10). ₃

The ecotype analysis refines this representation concern by showing that environmental metadata may be too incomplete or spatially indirect to explain genome-wide gene-content similarity: AlphaEarth embeddings covered only 28.4% of genomes, and geographic coordinates for host-associated bacteria may describe collection sites rather than their actual microenvironments. ₄ Across 172 species with sufficient data, phylogeny generally dominated environmental similarity as a predictor of gene-content similarity, while significant environmental effects were detected in only 12 species (7.0%) and no significant effect was detected in 156 species (90.7%). ₄ This does not contradict the evidence that collection and sequencing bias shape ecological inference; it indicates that weak environmental effects can reflect limited environmental coverage and metadata resolution, as well as genuinely lineage-dominated genome structure. ₄ The analysis is documented in Ecotype Correlation Analysis. ₄

The plant analysis further refines this interpretation: across 607 root, rhizosphere, and phyllosphere species, compartment explained only approximately 6% of functional-marker variance after the refined analysis, and the original much larger effect lost 86% of its R² after excluding three genome-rich species per compartment. ₃ The report attributes the initial effect mainly to taxonomic sampling and dominant rhizobial or Pseudomonas clades rather than a community-wide compartment signal. ₃ This supports the view that ecological labels and uneven taxonomic representation can make apparently strong environment–genome associations look broader than they are, while not proving that the residual compartment signal is purely an artifact. ₃

The environmental reanalysis supports the clinical-sampling-bias observation but refines its interpretation: environmental species did not have stronger environment–gene-content partial correlations than human-associated species (one-sided Mann–Whitney U=1536, p=0.83), and the environmental-genome fraction was not associated with correlation strength (Spearman rho=-0.085, p=0.25). ₅ Thus, the confirmed clinical bias does not account for the weak environment–gene-content signal under the reanalysis methodology. ₅ The report used genome-level harmonized classifications and majority voting across genome isolation sources, with 106 of 224 species majority human-associated, 47 majority environmental, and 71 Mixed/Other. ₅ This result is documented in Ecotype Reanalysis — Environmental vs Human-Associated Species. ₅

Clinical over-representation in public genome collections

Several antimicrobial-resistance (AMR, resistance of bacteria to antibiotic compounds) projects support the cultivation-bias argument from outside the metal domain. The GTDB/NCBI genome collection is heavily biased toward clinical and human-associated isolates, particularly for species like K. pneumoniae, S. aureus, and E. coli, and environmental species are underrepresented. ₆ Clinical-pathogen over-representation in genome databases may inflate AMR counts for human-associated species. ₇ The resistome project states the circularity directly: clinical isolates are massively overrepresented in NCBI, so the “clinical species carry more AMR” finding partly reflects that clinical species are the ones that get sequenced because they have AMR, while soil and aquatic species are undersampled. ₈ This is a selection-on-outcome-leakage pattern, so the raw clinical-versus-environmental resistome contrast must be read as conditioned on which organisms entered the database. ₈ The bias does not, however, dismiss the phylogenetically controlled evidence reported by the same project: 20 of 141 testable families (14%) show significant within-family environment effects after FDR (false discovery rate, the expected share of false positives among called hits) correction, with top families Enterobacteriaceae (q = 3×10⁻²¹), Bacteroidaceae (q = 1.3×10⁻¹⁷), Lachnospiraceae (q = 2.5×10⁻¹²), and Pseudomonadaceae (q = 3.9×10⁻¹²). ₈

Sparse labels compound the composition problem, because the metadata used to assign environments is itself thin. In the AMR pangenome atlas, the large “Other/Unknown” bin covering 46.8% of species reflects the sparsity and inconsistency of free-text isolation_source metadata in NCBI BioSample records. ₇ In the strain-variation project, 52.7% of genomes had no classifiable isolation_source, which left only 2 species with sufficient within-species environmental diversity for chi-squared testing after strict expected-frequency criteria were applied. ₆ This refines the matching-loss evidence below: ecological inference can fail at the metadata step even when genomes are plentiful. ₆

Classifier choice determines how much of a collection is eligible for ecological analysis at all: an NCBI keyword classifier assigned environments to 91% of species (1,190/1,307), while a BacDive approximation classified 35% (459/1,307). ₆ Both classifiers are described as approximate, and the report concludes that a dedicated metadata curation effort would improve ecotype analyses. ₆ Higher label coverage is therefore not automatically better ecological labelling, and the two classifiers are not interchangeable inputs to downstream environment tests. ₆

The central pitfalls digest records the same failure at per-genome resolution: parsing isolation_source and host from ncbi_env at the per-genome level produces 52.7% “Unknown” labels (94,957/180,025 genomes), because most BioSample records lack structured isolation metadata, and it recommends species-level majority-vote environment labels instead, which reach 91% coverage via keyword classification of NCBI metadata. ₉

The large unknown bin recurs in unrelated projects, which supports reading it as a property of the metadata rather than of any one analysis. The prophage-ecology project classified 293,050 genomes by NCBI isolation_source into 10 environment categories with ≥30 species and found significant environment × module associations for B_head (chi2=1,051, p=1.72e-220), C_tail (chi2=1,749, p≈0), and G_anti_defense (chi2=818, p=2.58e-170). ₁₀ It nonetheless warns that this metadata is sparse and inconsistently labelled, that collapsing it into 10 categories hides substantial within-category variation, and that the “other_unknown” category of 9,659 species (35%) limits statistical power. ₁₀

The PHB (polyhydroxybutyrate) granule project reports the same bin at similar size: 34.9% of species have “other_unknown” as their primary environment. ₁₁ Its family-level environmental interpretation is limited accordingly, because other_unknown accounts for 30 of 41 enriched and 26 of 62 depleted families. ₁₁

Within a single well-sequenced genus the imbalance is severe: P. aeruginosa comprises 53% of all genomes (6,760/12,732) because of its clinical importance, while many environmental species have fewer than 10 sequenced genomes, an imbalance the report says affects the power of species-level comparisons. ₁₂ Isolation-source metadata was available for 64.2% of genomes (8,171/12,732 with classifiable sources), assigned by keyword classification into 10 environment categories. ₁₂

The plant-growth-promotion survey shows how this skew makes enrichment estimates conservative rather than merely noisy. Of its genomes, 291,279 had ncbi_isolation_source metadata and 93.5% were classifiable. ₁₃ Only 1,637 species (5.9% of species with an environment label) were classified as soil/rhizosphere-dominant, which the report attributes to NCBI’s clinical and host-associated sampling bias; it therefore treats the acdS and pqqC enrichment odds ratios of 7.0 and 2.9 as likely underestimates of true rhizosphere enrichment, since many soil isolates are deposited under broader “environmental” labels. ₁₃ It restates the limitation as a methodological caveat: ncbi_isolation_source is noisy, NCBI is heavily biased toward clinical and host-associated isolates, only 5.9% of species are soil/rhizosphere-dominant, and the soil enrichment effects are likely underestimates. ₁₃

The same composition also shapes how that project reads one of its depletion results: the overall soil depletion of nifH is attributed partly to database composition, because most nitrogen fixers represented are aquatic or marine — cyanobacteria, Azotobacter — or host-associated rhizobia. ₁₃ A depletion measured against such a reference set reports where sequenced nitrogen fixers came from as much as where the gene is ecologically rare. ₁₃

The ENIGMA genotype-to-phenotype project draws a transfer-learning consequence from genus-level composition. Across all GTDB genomes, Pseudomonas is 37.8% clinical (driven by P. aeruginosa), 12.9% soil/plant, and 9.4% aquatic, whereas all ENIGMA Pseudomonas belong to Pseudomonas_E, the fluorescens/protegens clade, which is environmental rather than clinical, and Rhodanobacter is purely aquatic/contaminated (55% aquatic, 10% contaminated, 0% clinical). ₁₄ The report concludes that ENIGMA’s subsurface field isolates occupy environmental niches under-represented in the NCBI genome collection and that transfer learning from clinical phenotype databases may therefore be biased. ₁₄

A failed enrichment test shows collection composition redirecting an ecological hypothesis rather than merely weakening it. In the ecological-agora project, Mycobacteriaceae (n=459) was not enriched in soil: 12.9% versus a 14.5% background, 0.88× fold, Fisher p = 0.87. ₁₅ The report attributes the null to a clinical-isolate skew among the species representatives held in the KBase Data Lakehouse — the MGnify breakdown shows 53 human-skin, 16 human-vaginal, 15 human-gut, and 11 chicken-gut species — and treats the 7.88× host-pathogen fold-enrichment as the clean anchor, refining the mycolic-acid signature toward host-pathogen-niche mycobacteria rather than soil mycobacteria. ₁₅

The same project refines the substrate question in the opposite direction from the usual bias argument, and the distinction between effect size and detection matters. Alm 2006’s r ≈ 0.74 in 207 genomes was recorded as NOT REPRODUCED on a substrate of 18,989 species representatives, but that verdict concerns the point estimate rather than the existence of the association: all correlations tested were statistically significant at n = 18,989 while biologically modest (r = 0.10–0.29). ₁₅ The report attributes part of the dilution to substrate composition — Alm 2006 used 207 genomes, mostly cultivated isolates with high annotation density and broad taxonomic spread, whereas the 18,989 species reps span the full GTDB tree including DPANN, CPR, and weakly annotated lineages where histidine protein kinase (HPK) detection is genuinely sparse, so a taxonomically heterogeneous original signal is diluted by aggregation across the tree. ₁₅ On this reading, a smaller effect size on a broader substrate is not by itself evidence against a cultivated-isolate finding, which is the mirror image of the usual concern that cultivated panels overstate ecological generality. ₁₅

Environmental metadata and embedding coverage limit what can be tested

Coverage audits across projects refine the representation argument by showing that the binding constraint is often not genome count but which genomes carry usable environmental context. In the ecological-agora project, per-genome BioSample mapping reached 100% of genomes, while the BioSample environmental metadata fields were far patchier. Its substrate audit reports these figures as coverage of the 18,989-species substrate rather than of all genomes — 76.8% isolation_source, 88.8% geo_loc_name, 48.6% env_broad_scale, 47.3% lat_lon — with AlphaEarth 64-dimensional environmental vectors at 27.2% (5,157 species). ₁₅ Categorical biome labels from MGnify covered 28.4% of species (5,392 species) across 18 biome categories. ₁₅

Occurrence-side resources carry their own depth limits. The ecological-agora audit used MGnify biome categorical labels at 28% species coverage across 18 biome categories but did not query MGnify’s per-sample taxa-coverage matrices, which it names as a substantive future direction for finer-grained ecological inference. ₁₅ Its NMDC sample-level taxa enrichment reached 9.4% species coverage. ₁₅ NMDC studies are themselves biased toward terrestrial and soil environments, under-representing the full environmental gradient. ₁₁

Satellite-derived AlphaEarth coverage is both partial and non-random. The embedding explorer reports AlphaEarth coverage of only 28.4% of all genomes, biased toward those with valid lat/lon metadata. ₁₆ The dark-matter project reports embeddings for 28% of genomes (83K/293K) together with inconsistent NCBI isolation-source metadata as joint limits on the power of biogeographic tests. ₁₇ At the species level, only 6.8% of species (1,872/27,690) have sufficient AlphaEarth coverage, defined as ≥5 genomes with embeddings, because most NCBI genomes lack lat/lon metadata, especially clinical isolates. ₁₈ Geographic coordinates from NCBI metadata are also often missing or imprecise, which reduces the quality of environmental associations. ₄

Three further projects report the same coverage ceiling but describe the covered subset differently. The SNIPE defense-system analysis gives AlphaEarth coverage as 28.4% of genomes and describes the covered subset as biased toward environmental isolates with lat/lon data. ₁₉ The prophage-ecology analysis gives 28% of genomes and describes the resulting subsample as biased toward clinically and environmentally well-sampled lineages. ₁₀ The pitfalls digest supplies a mechanism: embeddings require valid lat/lon coordinates, which are often missing in NCBI metadata, especially for clinical isolates. ₉ The composition reported below for the embedded set describes what is inside it, not how that composition compares with the unembedded genomes, so these differing characterizations are recorded as an unresolved comparability question rather than a demonstrated contradiction. _19,16

Coverage can also exceed the baseline and still bind. AlphaEarth coverage of xoxF genomes is 39.5% (1,457/3,690), better than the 28% pangenome baseline, but still leaves 60.5% of xoxF carriers without environmental coordinates. ₂₀ The same project notes that its ncbi_env environmental classification is text-mining-derived with hierarchical regex priorities and that the “host_associated” class is broad — gut, skin, oral, infections — so misclassifications are possible, while reporting that the strong signals of soil/sediment enrichment and host depletion survive. ₂₀

Bias is measurable along environmental axes, not only in counts: frontier areas (genomic discovery index, GDI, above 1000) have mean pH = 6.74 against mean pH = 5.94 for mapped areas, a +0.8 pH unit gap that the soil-frontier analysis reads as systematic under-sampling of alkaline soil microbiomes in public genomic databases. ₂₁

Within the embedded subset, composition is skewed toward human hosts: of the 83,287 genomes with AlphaEarth embeddings, 38% are human-associated — Human clinical 16,390 (20%), Human gut 13,466 (16%), Human other 1,669 (2%) — while Soil (6,073; 7%), Marine (5,850; 7%), and Freshwater (5,840; 7%) are much smaller, which the report attributes to pathogen-sequencing bias and to the good geographic metadata that epidemiological tracking gives clinical isolates. ₁₆ The central digest records the same pattern: 38% of the 83,287 embedded genomes are human-associated (clinical 20%, gut 16%, other 2%), with soil, marine, and freshwater at 7% each. ₁₈ The explorer cautions that this clinical/human bias of 38% may not represent the overall NCBI or pangenome population. ₁₆ It also notes, citing Saha et al. (2023), that public genome databases are heavily skewed toward pathogens and clinical isolates, which biases any analysis of environment–genome relationships. ₁₆

Harmonized per-genome classification quantifies that skew inside the ecotype panel and supports the clinical-bias claim while contradicting the assumption that the bias drives the weak environment signal. Using isolation_source harmonization into 12 categories from 5,774 values, 106 of 224 ecotype species are majority human-associated (gut plus clinical) and only 47 of 224 (21%) are majority environmental (Soil, Marine, Freshwater, Extreme, Plant); the digest describes this as more systematic than the original manual species-level categorization, confirming the clinical bias while showing it does not confound the ecotype results. ₁₈ The reanalysis reports the same panel definition — 224 species with ≥20 genomes carrying AlphaEarth embeddings and ≥30% coverage, of which 106 (47%) are majority human-associated, 47 (21%) majority environmental, and 71 (32%) mixed/other. ₅ Its authors read this composition as confirming the strong clinical sampling bias in the AlphaEarth subset while showing that the bias does not account for the weak environment signal. ₅

The digest also records the specific bias prediction that failed. Human-associated samples were observed to dampen the AlphaEarth geographic signal (2.0x versus 3.4x for environmental samples), which motivated the hypothesis that clinical over-representation explained the weak environment effect in the ecotype analysis (median partial correlation 0.003). ₁₈ Stratifying 213 species by dominant environment type (47% human-associated, 21% environmental) showed the opposite direction: human-associated species had higher partial correlations (median 0.084) than environmental species (median 0.051), Mann-Whitney p=0.83, so the geographic signal in embeddings and the environment–gene-content relationship behave as distinct phenomena. ₁₈ This contradicts a sampling-bias explanation for the weak signal while leaving the underlying coverage limits in place. ₁₈

Evidence from BacDive–pangenome matching

Of 97,334 BacDive strains, 42,227 strains (43.4%) matched 6,426 GTDB species, while 55,107 strains (56.6%) remained unmatched. ₁ Exact species-name agreement produced 33,535 matches (34.5%), and removing GTDB suffixes produced 8,692 additional matches (8.9%). ₁ Only 25,089 matched strains had isolation-source metadata, so both genome-score linkage and ecological metadata availability reduced the usable sample. ₁

This matching loss limits which cultivated strains can contribute to ecological validation and can make the retained sample unrepresentative of environmental diversity. ₁ The report specifically notes that BacDive represents culturable, described strains and may under-represent metal-tolerant extremophiles because of culture-collection bias. ₁ The unresolved 56.6% could partly reflect different species boundaries between GTDB and LPSN/DSMZ rather than absence of relevant genomes, making accession-level matching an important remedy. ₁

The MicrobeAtlas coverage result refines this interpretation: increasing the number of available genomes can reveal more metal types, but does not by itself establish that the added traits are distributed across the environment rather than concentrated in better-sampled lineages. ₂ In a 200-iteration rarefaction analysis representing each genus by one randomly sampled species, the metal-type association remained directionally consistent, but the median coefficient was +0.0147 rather than the full-data estimate of +0.0215; 89.5% of iterations had p < 0.05 and 57.5% were Bonferroni-significant at p < 0.0083. ₂ This supports coverage sensitivity as a measurement issue while showing that the observed association was not eliminated by that particular subsampling procedure. ₂

The plant analysis supports the same caution through a second representation mismatch: only 11.7% Jaccard overlap existed between 487 pangenome-derived plant genera and 438 MGnify rhizosphere genera. ₃ The two resources measure different phenomena—mostly isolation metadata versus metagenomic detection—so their disagreement cannot be interpreted simply as ecological absence or presence. ₃ This strengthens the case for treating collection-based and occurrence-based catalogs as complementary, rather than interchangeable, samples of environmental diversity.

The plant project’s follow-up accounting refines that mismatch rather than resolving it. It attributes the low concordance to different data sources — isolation metadata versus metagenomic detection — and to sampling biases rather than classification error, noting that 93 pangenome plant genera were found in MGnify soil rather than rhizosphere, which suggests broader environmental associations. ₃ The asymmetry runs both ways, with 390 pangenome-only plant genera that the report tentatively suggests may be opportunistic soil bacteria isolated from plants but not rhizosphere specialists, and 341 MGnify-only rhizosphere genera that may include uncultured lineages absent from GTDB. ₃ Because MGnify contributes metagenomic classifications of plant-associated genera while the pangenome side contributes isolation metadata, the 11.7% Jaccard overlap limits the strength of cross-validation conclusions rather than adjudicating either catalog. ₃

A third representation loss occurs at the pangenome end rather than the metadata end: 12 pangenomes without matching species clades are mostly obligate symbionts, including Portiera aleyrodidarum, a whitefly symbiont, and Profftella armatura, a psyllid symbiont, alongside various uncultivated lineages carrying UBA, TMED, and SCGC prefixes. ₁₈ These are valid pangenomes but were filtered from species metadata due to single-genome status, so organisms least represented by cultivation can also be removed by downstream filters. ₁₈

Collection bias affects environmental validation

Among matched strains, heavy-metal contamination isolates had a median metal score of 0.240 and a mean of 0.236, compared with an environmental-baseline median of 0.187 and mean of 0.195. ₁ Their median delta was +0.053, Cohen’s d was +1.00, and the Mann–Whitney p-value was 0.006 for n=10 isolates. ₁ Cohen’s d was calculated as (mean_group - mean_baseline) / pooled_SD, with pooled_SD = sqrt((SD_group² + SD_baseline²) / 2). ₁

The broader contamination gradient was heavy metal (+1.00) > waste/sludge (+0.57) > all contamination (+0.43) > industrial (+0.20). ₁ Waste/sludge isolates had n=305 and d=+0.57, all-contamination isolates had n=176 and d=+0.43, and industrial isolates had n=796 and d=+0.20; each comparison had p<0.0001. ₁ Host-associated isolates had n=12,086, d=+0.14, and p<0.0001, indicating that collection category can also track biological or genomic composition unrelated to environmental contamination. ₁

These results support environmental-resistome by showing that curated isolation environments can validate genome-based metal-tolerance predictions, but the strength of that validation depends on which environments and organisms enter the culture collection. ₁ The heavy-metal result is especially uncertain because n=10 matched isolates produced an observed d=1.00, only slightly above the approximately d=0.93 minimum detectable effect size for 80% power against an environmental baseline of approximately 5,000 strains. ₁

The MicrobeAtlas analysis supports but qualifies this interpretation: across 606 genera, metal-type diversity was associated with broader inferred niche breadth after phylogenetic adjustment, but the association was correlational and could reflect genome size, metabolic versatility, biofilm capacity, or gene acquisition rather than a causal effect of metal resistance. ₂ Controlling for genome size retained the metal-type effect (β = +0.0218, SE = 0.0061, p = 3.6×10⁻⁴), while log(genome size) also remained associated (β = +0.0263, SE = 0.0069, p = 1.5×10⁻⁴). ₂ This reinforces the need to treat contamination-category differences as ecological validation signals rather than direct causal demonstrations.

The plant study refines the interpretation of marker-based ecological validation by showing that genome presence does not establish activity or phenotype. ₃ In its refined plant-associated panel, 878 of 1,115 species (78.7%) were classified as dual-nature because they carried both beneficial and pathogenic marker types, while all four neutral controls in an 18-organism validation panel were misclassified. ₃ The report therefore treats categorical marker cohorts as a coarse screen and emphasizes continuous scores and experimental or transcriptomic validation. ₃ This is directly analogous to the BacDive limitation: a curated environmental label or genomic marker can provide validation evidence without independently demonstrating realized ecological function.

Taxonomic composition and residual bias

Within Pseudomonadota, contamination isolates had metal scores delta=+0.040 above environmental isolates with p<0.001, using n=85 contamination and n=5,655 environmental isolates. ₁ Within Actinomycetota, the corresponding delta was +0.035 with p<0.001, using n=62 contamination and n=772 environmental isolates. ₁ These results support a contamination-associated signal beyond broad phylum composition in the two most-sampled phyla. ₁

The signal was not significant within Bacillota, where delta=-0.012, p=0.285, n=19 contamination and n=492 environmental isolates, or Bacteroidota, where delta=-0.008, p=0.456, n=5 contamination and n=156 environmental isolates. ₁ These null results cannot distinguish biological differences from limited statistical power because contamination-isolate sample sizes were small. ₁

Host-associated strains had a positive but small deviation, with d=+0.14 and p<0.0001. ₁ The report attributes this pattern potentially to genome-size and taxonomic confounding because host-associated records were dominated by Pseudomonas, Klebsiella, and Acinetobacter, which have large genomes and more KEGG-annotated gene clusters. ₁ Genome-size normalization reduced but did not eliminate this possible confounding, so normalization alone does not establish that isolation environment caused the score difference. ₁

The MicrobeAtlas study refines this taxonomic caution by showing that genus aggregation can make a broad genus score represent different species occupying different habitats rather than one organism being a generalist. ₂ Its strict 5% within-environment prevalence analysis reduced the PGLS (phylogenetic generalized least squares, regression that adjusts for shared ancestry) sample from 606 to 379 genera and changed the metal-type coefficient from +0.0215 (p = 1.5×10⁻⁴) to +0.0166 (p = 0.092), although the direction was preserved. ₂ Ecological conclusions therefore depend on both taxonomic coverage and the detection threshold used to define environmental occurrence. ₂

The MicrobeAtlas archaeal subset turns the same bias into an explicit power limit. At n = 48 genera, archaeal PGLS found no significant effects; the metal-types coefficient was positive at +0.0145, the same direction as the bacterial +0.0215, but the standard error was nearly as large as the estimate (t = 0.73), which the report calls severe underpowering. ₂ The 48 archaeal genera are biased toward cultured, well-studied lineages — methanogens, halophiles, thermoacidophiles — and miss the environmentally dominant archaea in MicrobeAtlas, Thaumarchaeota and Woesearchaeota. ₂ This null therefore cannot distinguish absent biology from absent organisms, and supports genomic-under-representation. ₂

The plant study supports this warning about taxonomic composition: its initial compartment effect was largely attributable to genome-rich species and dominant rhizobial or Pseudomonas clades, while the refined compartment signal was small. ₃ Likewise, of 5,671 plant-enriched eggNOG groups, 5,341 (94.2%) were significant before phylum control, but only 50 retained significance after phylum-level control; the report interprets the near-saturation of raw associations mainly as taxonomic compositional divergence rather than plant-specific enrichment. ₃ This demonstrates that large collections can increase statistical certainty without removing lineage-composition bias.

Cultivation bias as a measurable diagnostic in deep-clay isolates

The clay-confined subsurface work turns cultivation bias from a caveat into a measurement, and it extends this concept from collection composition to habitat fraction. At the same clay site, the Bagnoud (2016) Mont Terri Opalinus porewater paradigm and the Mitzscherling (2023) Opalinus rock-attached paradigm describe qualitatively different microbial communities: porewater is dominated by sulfate reducers, rock surfaces by Geobacter/Geothrix iron reducers. ₁₈ The cited rock-attached community profile has sulfate-reducing bacteria (SRB) below 0.2% and iron-reducing bacteria (IRB) at 4.3–10.2%, dominated by Geobacter and Geothrix; it serves as a literature comparator, not a corrected iron-marker result for the cultured cohort. ₂₂

The digest formalizes a diagnostic statistic from that contrast: for any cohort of cultured clay-isolated genomes, compute sulfate-reduction marker (dsrAB-aprAB-sat) and iron-reduction marker (omcS / mtrC / mtrA) prevalences and compare them against the Mitzscherling rock-attached frequencies (SRB ~0.2%, IRB ~7%) via a binomial test. ₁₈

Applied to the cultured deep cohort, the sulfate-reduction side is overwhelming: compared against the Mitzscherling et al. (2023) rock-attached null distribution (SRB ~0.2%, IRB ~7% of community), enrichment gave a binomial p = 4.0×10⁻¹², with 5 observed versus 0.018 expected of 9. ₂₂ Because all 8 Opalinus genomes in the KBase Data Lakehouse trace to BRC-3 or BIC-A1 borehole isolation sources, the report reads this as a direct, quantitative diagnostic of cultivation bias: the collection captures the porewater fraction, not the rock-attached fraction. ₂₂

The digest reports the same test on the 9 anchor_deep clay genomes held in the KBase Data Lakehouse: 5/9 are sulfate-reduction-marker positive, with a binomial test against the rock-attached null giving p=4×10⁻¹². ₁₈ It interprets this as direct, quantitative evidence that the cohort represents the porewater fraction rather than the rock-attached fraction, and proposes the framework as generalizable to any future cultured cohort drawn from comparable pangenome resources. ₁₈

The diagnostic is only half-supported, and the report says so. The porewater-bias headline holds via the sulfate-reduction side alone; the rock-attached versus porewater dichotomy interpretation is half-supported on the SR side and half-unsupported on the IR side after marker correction. ₂₂ Because the cohort is cultured-only, all conclusions apply to “cultivable porewater-cultured deep-clay isolates,” not to the full Mont Terri / bentonite microbial community — a scope limit the report frames as the headline finding rather than a confounder. ₂₂

What is missing from the cohort is as diagnostic as what is present: CPR / DPANN episymbionts (Bell 2022) and rock-attached Geobacter / Geothrix lineages are essentially absent, and the report states that metagenome-assembled-genome (MAG, a genome reconstructed from metagenomic reads) augmented future work is necessary to test whether the genuine clay-confined community matches the literature predictions. ₂₂ This supports genomic-under-representation: an absent lineage in a cultured collection is not evidence of ecological absence. ₂₂

The subsurface Bacillota_B project shows the same ceiling from the recruitment side. A BacDive expansion for the deep-clay anchor yielded zero additional Bacillota_B genomes in that run, and the 10-genome anchor was ~entirely from ncbi_env keyword search. ₂₃ That the anchor cohort was assembled almost entirely through ncbi_env keyword search is itself a sampling caveat: its composition reflects which isolation-source strings exist, and the curated phenotype database added nothing recoverable for this lineage. ₂₃

Phenotype records are an additional bottleneck

Only 24 matched strains had metal-utilization records covering iron, manganese, arsenate, chromate, cobalt, or zinc. ₁ Eight of those results were positive and 16 were negative. ₁ Positive utilizers had lower metal scores in the exploratory comparison, but the result was inconclusive because Mann–Whitney p=0.14 and Cohen’s d=-0.57 with only 24 records. ₁

This underpowered phenotype comparison shows that linking genomes to collection metadata does not guarantee an independent phenotype validation. ₁ The limitation refines multi-omics-integration: genome-derived scores and curated phenotype records can be connected, but sparse and selectively recorded phenotypes may prevent reliable assessment of agreement. ₁

The environmental atlas provides a related validation constraint: in its 133-sample community analysis, only 16.8% of reads were joinable to genus-level GTDB metal-diversity data, so community-weighted metal-type means could be biased if uncovered genera have systematically different resistance profiles. ₂ The study therefore recommends testing the correlation between per-sample covered-read fraction and the community-weighted mean before interpreting site-level patterns. ₂

The plant analysis supports this bottleneck from a different angle: BacDive metabolite-utilization data agreed with GapMind pathway predictions at 83.1% consistency, but a GapMind core-level completeness score was 0% across compartments and may have been too stringent for broad taxonomic comparisons. ₃ Thus, even when independent records are available, apparent disagreement can reflect assay scope, pathway thresholds, and taxonomic coverage rather than ecological biology alone.

The ecological-agora project finds culture bias inside a phenotype database rather than beside it: BacDive Bacteroidota records are 33.2% anaerobe against 22.0% atlas-wide — 1.5× enriched — with 48.7% aerobe, a contrast the report reports as holding despite type-strain culture bias toward aerobes. ₁₅ The direction in which that culture bias shifts a focal-versus-atlas enrichment ratio is not established by the report, so the caveat is retained alongside the figures rather than used to bound the estimate. ₁₅

Deliberate oversampling in curated clinical collections

Curated clinical strain collections carry a bias that is intentional rather than incidental, which refines the concept: not all over-representation is an accident of sequencing effort. The cystic fibrosis isolate collection is dominated by P. aeruginosa (655 isolates), S. aureus (379), and Rothia dentocariosa (318), reflecting the typical CF airway microbiome plus deliberate oversampling of pathogens. ₂₄ Collection frequencies therefore cannot be read as natural prevalence. ₂₄

Downstream subsetting narrows the usable cohort further and shifts its composition. The 142-isolate core cohort covers 62 of 211 species (29%), growth kinetics are available for only 32 isolates, and the subset is enriched for species with multiple assay types, which may over-represent organisms prioritized for deeper characterization. ₂₄ Rothia, Streptococcus, Neisseria, and Gemella account for 22%, 15%, 8%, and 7% of the core cohort versus 6%, 8%, 3%, and 3% in the full collection, so the report expects the metabolic overlap–inhibition relationship to generalize while the absolute R² may differ for under-represented taxa. ₂₄

Habitat labels attached to genomes are similarly uneven across species. Of 21 lung/respiratory genomes identified, Rothia dentocariosa contributes 10 (38% of its species) and Neisseria mucosa contributes 5 (33%), while M. luteus has 0 lung genomes and is primarily skin/environmental, which the report flags as a possible engraftment challenge. ₂₄ Lung-adapted S. salivarius genomes show enrichment for L-malate (+0.39 score) and depletion for sorbitol (−0.82), a pattern the report reads as metabolic adaptation to the airway carbon landscape. ₂₄

This representation gap has a design consequence, not only an inferential one: M. luteus is the species whose addition at k=3 achieves 100% P. aeruginosa substrate coverage, yet it has zero engraftability — not detected in any patient metagenome — and zero lung genomes in the pangenome. ₂₄ Engraftability here is inferred rather than measured: the report states that patient prevalence is a proxy, not a direct measure of colonization persistence after probiotic administration. ₂₄ The pattern therefore suggests the hypothesis that the species most important for metabolic coverage is the least likely to persist, which supports treating collection-derived candidate sets as biased toward what has been isolated rather than what has been shown to establish in the target habitat. ₂₄

Where habitat metadata is dense enough, it can support a within-species test rather than only a caveat. Among 6,760 P. aeruginosa genomes, 5,199 had metadata and 1,796 were from lung/respiratory or CF sources; seven GapMind pathways differed significantly at an FDR below 0.05 between lung and non-lung isolates, and all were carbon-source pathways that lung P. aeruginosa is losing — sorbitol (−0.165), mannitol (−0.204), gluconate (−0.185). ₂₄ The report’s reading of this as evolutionary metabolic streamlining under relaxed selection in amino-acid-rich sputum is an interpretation of a cross-sectional contrast between metadata-labelled subsets rather than a measured trajectory. ₂₄

Phage collections carry the analogous bias on the host side. In the IBD phage-targeting project, 65 of 94 phages (69 %) were isolated against B2/D phylogroup hosts, predominantly adherent-invasive E. coli (AIEC) per Dogan 2014 / Dubinsky 2022, and the top phages include those isolated against the canonical AIEC reference strains LF82, LF73, and 536. ₂₅ Because isolation host determines which phages exist to be ranked, a ranking built on this panel inherits the panel’s phylogroup composition — an interpretation of that composition rather than a measured bias in therapeutic efficacy. ₂₅

Interpretation and tensions

The contamination gradient supports the hypothesis that genome-based Metal Fitness Atlas scores capture real ecological metal-adaptation signal, but the heavy-metal estimate is imprecise because only n=10 isolates matched species with scores. ₁ This evidence is stronger at the larger contamination-category sample sizes than for the heavy-metal category itself, where the effect barely exceeded the power threshold. ₁

The result also needs to be distinguished from the finding that metal-tolerance genes were 88% core in the metal-specificity project: within-species conservation can coexist with between-species differences in the total number of metal-tolerance genes. ₁ The proposed explanation that species with larger core genomes encoding more metal-tolerance functions are more likely to occur in contaminated environments is an interpretation rather than a direct causal test. ₁

Compared with the Oak Ridge field-abundance analysis, which found a suggestive but non-significant correlation between laboratory metal tolerance and field abundance of rho=0.50, p=0.095, n=11 genera, the BacDive analysis provides larger-scale support for the same hypothesis through contamination-category effect sizes of d=+0.43 to +1.00. ₁ The comparison does not remove collection bias because the BacDive heavy-metal result still uses only n=10 isolates. ₁

The MicrobeAtlas groundwater analysis supports an environmental-occurrence relationship but does not resolve this collection-bias problem: metal-type diversity correlated with groundwater prevalence (Spearman ρ = +0.112, p = 0.0019), whereas groundwater-specific fold-enrichment was null (ρ = +0.042, p = 0.242). ₂ This distinction argues against treating either curated isolation source or broad environmental detection as uniquely identifying a metal-contamination mechanism. ₂

The ecotype result refines this interpretation by showing that a weak whole-genome environmental signal cannot by itself distinguish absent ecological adaptation from inadequate environmental representation: the analysis used partial correlations, which assume linear relationships between distance matrices, and found no significant environmental effect for 90.7% of species. ₄ The reanalysis supports the robustness of the environmental-versus-human-associated null within one consistent method: Environmental species had median partial correlation 0.051 versus 0.084 for human-associated species, while the Mixed/Other median was 0.109; the environmental group was not stronger despite the confirmed clinical sampling bias. ₅ However, absolute correlations are not comparable across the two analyses: the reanalysis median across 183 species was 0.081 versus 0.003 originally, characterized as a 27x difference, because it used all embedded genomes, including up to 3,505 genomes per species, rather than diversity-maximizing downsampling capped at 250 genomes and also used different genome sets. ₅ Environmental species also had a 21% NaN rate versus 7% for human-associated species, so unequal filtering does not explain the absence of a stronger environmental signal and may, if anything, bias toward finding one. ₅ The original proposed follow-up—testing functional subsets such as COG categories rather than whole-genome gene content—therefore remains relevant because environmental adaptation may be concentrated in specific genes and missed by genome-wide summaries. ₄

The reanalysis further refines the role assigned to clinical sampling bias: it confirms that environmental metadata and genome composition remain difficult to interpret, but indicates that the clinical imbalance alone does not explain the weak correlation. ₅ Possible explanations involving geographic epidemiological structure in human-associated species, greater statistical power in species with more genomes, and heterogeneous Mixed/Other sampling campaigns remain hypotheses rather than demonstrated mechanisms. ₅

The plant study introduces a related tension with interpretations that treat broad plant-associated marker patterns as community-wide ecological specialization. Its refined PERMANOVA found a statistically significant compartment effect (pseudo-F = 23.2, p = 0.001) but R² = 0.071, while excluding three genome-rich species reduced the earlier R² = 0.527 to R² = 0.072. ₃ The first result supports a real compartment-associated signal; the second supports the competing explanation that the original magnitude was largely driven by taxonomic sampling. ₃ These findings should not be reconciled by averaging: the effect is statistically detectable, but its community-wide magnitude remains sensitive to representation and model refinement.

An open comparability question remains over which genomes the AlphaEarth-embedded subset over-represents, and it should not be settled by averaging the accounts. The embedding explorer reports that 38% of the 83,287 embedded genomes are human-associated and attributes that to pathogen-sequencing bias plus the good geographic metadata epidemiological tracking gives clinical isolates. ₁₆ The SNIPE analysis puts coverage at the same 28.4% of genomes but describes the covered subset as biased toward environmental isolates with lat/lon data. ₁₉ These are not strictly opposing measurements: a within-subset composition does not establish how the embedded fraction compares with the unembedded fraction, and the pitfalls digest’s mechanism — embeddings require valid lat/lon, often missing for clinical isolates — is a stated reason rather than a measured comparison, so the question stays open until embedded and unembedded genomes are compared within one genome set. ₉

Open Directions

• Match BacDive GCA accessions directly to pangenome genome_ids through kbase_ke_pangenome.genome, then test whether recovered strains change the contamination effect estimates and the 56.6% unmatched fraction. ₁
• Integrate ENIGMA CORAL community data from the Oak Ridge metal-contaminated site and compare community-level metal-tolerance profiles with BacDive isolation-source predictions to test whether the culture-collection signal generalizes to field communities. ₁
• Stratify matched BacDive strains by metal-tolerance gene families and contamination environment, using metals with sufficient representation including iron and manganese, to test whether specific functions predict specific environments. ₁
• Expand BacDive phenotype extraction beyond the metabolite_utilization table to include MIC and growth-inhibition data, then test whether independently measured tolerance agrees with genome-derived scores. ₁
• Reanalyze contamination effects within taxonomic strata and after genome-size normalization, using the small Bacillota and Bacteroidota groups to determine whether their null results reflect biology or limited power. ₁
• Build multi-primer, multi-region occurrence datasets and perform species-per-genus rarefaction to test whether the metal-type/niche-breadth association persists after detection probability and pangenome coverage are jointly equalized. ₂
• For community-level validation, relate covered-read fraction to weighted metal-diversity scores and compare genus-level predictions with independent metal measurements across sites. ₂
• Recalculate environmental–gene-content associations using direct environmental metadata and alternative embedding distances, then test COG functional subsets and ecotype-cluster gene content to determine whether environmental effects are masked by incomplete AlphaEarth coverage or whole-genome aggregation. ₄
• Compare downsampled and full-genome extraction on the same species, controlling for genome count, to resolve the 27x partial-correlation discrepancy; repeat the test with structured ENVO terms and functional subsets. ₅
• Expand plant-associated phylogenetic-tree coverage beyond 18 of 65 candidate species and repeat within-species host/subclade tests, while matching collection metadata to genome identifiers, to determine whether apparent host specialization generalizes beyond the currently testable lineages. ₃
• Compare pangenome-derived plant association with MGnify detection after genus-level rarefaction and matched taxonomic coverage, then test whether compartment effects persist when genome-rich species and uneven environmental sampling are controlled. ₃
• Validate plant-associated beneficial and pathogenic marker sets with transcriptomics or experiments, and test whether continuous scores outperform categorical labels under matched genome-size and taxonomic controls. ₃
• Ingest deep-subsurface metagenome-assembled genomes from Mont Terri, Olkiluoto, MX-80 bentonite, and Oak Ridge into the pangenome and re-run the full H1/H2/H3 framework on the expanded cohort — including the H1 amino-acid biosynthetic pathway-completeness analysis, not the respiratory SR/IR marker survey alone — to test whether the self-sufficiency signal (H1) emerges once the rock-attached / uncultivated lineages are present. ₂₂
• Apply the SR/IR marker-versus-rock-attached-null binomial diagnostic to other cultured environmental cohorts, to measure which habitat fraction a collection represents before using it for ecological inference. ₁₈
• Curate environmental metadata for the AMR species panel instead of relying on the approximate NCBI keyword and BacDive classifiers, then re-run the within-species environment–ecotype tests that only 2 species currently support. ₆
• Re-test environment–gene-content associations restricted to genomes with env_broad_scale and lat_lon present, and compare against the AlphaEarth-embedded subset, to separate metadata-coverage effects from biological signal. ₁₅
• Compare the CF collection’s isolate frequencies with patient-metagenome detection for the same species, to test whether deliberate pathogen oversampling distorts consortium candidate ranking. ₂₄
• Query MGnify’s per-sample taxa-coverage matrices instead of the 28%-coverage categorical biome labels, and test whether finer-grained biome assignment recovers environment effects lost at the 18-category resolution. ₁₅
• Expand archaeal genus coverage beyond the 48 cultured-lineage genera to include Thaumarchaeota and Woesearchaeota, then re-run the metal-type PGLS to test whether the positive but underpowered +0.0145 coefficient resolves. ₂
• Target isolation and sequencing at alkaline soils, where frontier areas average pH 6.74 against 5.94 for mapped areas, and test whether the discovery frontier closes or the pH gap persists. ₂₁
• Compare the environment composition of embedded versus unembedded genomes within one project’s genome set, to settle whether the AlphaEarth subset over-represents clinical or environmental isolates. _19,16