Limits of Composite Resistance Scores for Mechanistic Inference

Composite resistance scores summarize tolerance across multiple metals or conditions, but they do not by themselves identify the metal-specific mechanisms producing that tolerance. The BacDive analysis illustrates this limitation by comparing classical phenotypes with genome-based predictions from the metal-fitness-atlas, rather than with direct metal-tolerance measurements. ₁ Gene-resolution analysis of 452 metal experiments likewise found a nearly universal positive cross-metal fitness direction—311/317 correlations were positive—while the magnitude of association varied by metal pair; this supports interpreting an aggregate score as a mixture of general stress response and chemistry-dependent effects, not as a single metal-specific mechanism. ₂

The newer atlas evidence supports and sharpens this interpretation. Across 22 organisms and 14 metals, metal-important genes were 87.4% core versus 76.9% core for baseline genes (OR=2.08, p=4.3e-162), and 21 of 22 organisms showed greater core representation than baseline. ₃ This result is consistent with a two-tier model: broadly required core cellular and stress functions dominate genome-wide metal-fitness defects, while specialized accessory resistance mechanisms remain a distinct component. ₃ It therefore refines the implication of a composite score: high aggregate tolerance may reflect conserved cellular robustness rather than a transferable, metal-specific resistance mechanism. The atlas also found that essential-metal tolerance genes had a mean core-fraction delta of +0.148 versus +0.081 for toxic metals (one-sided Mann-Whitney U=39, p=0.015), further indicating that the composition of an aggregate score depends on the chemistry and biological role of the metals included. ₃

The newer metal-specificity analysis supports and sharpens this interpretation: among 7,609 metal-important gene records, 4,177 (54.9%) were metal-specific by showing a metal fitness defect but a <5% sick rate across non-metal experiments, whereas 2,888 (38.0%) were generally sick and 544 (7.2%) were metal+stress. ₄ Thus, a composite score can combine genuinely metal-specific determinants with broadly pleiotropic stress genes; its aggregate value cannot identify which component explains tolerance to a particular metal. The source classified 6,504 experiments, including 559 metal experiments and 5,945 non-metal experiments, and excluded 5,229 of 12,838 metal-important gene records because of locusId or fitness-matrix coverage problems. ₄

The atlas directly refines the interpretation of repertoire-based scores: a 1,286-KO metal functional signature scored 27,702 pangenome species, but the simple gene-presence repertoire score failed to predict metal fitness. Genome-size normalization was necessary because large open pangenomes otherwise dominated scores, and bioleaching genera were not significantly enriched after normalization (Mann-Whitney p=0.17). ₃ This supports treating a composite or repertoire score as a prioritization feature rather than a mechanistic assay; the atlas specifically identifies regulatory or expression-based models as the needed extension. ₃

The atlas derived a metal functional signature of 1,286 KEGG KO terms from conserved metal gene families and used it to score all 27,702 pangenome species. Genome-size normalization (metal clusters / total KEGG-annotated clusters) changes which organisms stand out. After normalization, individual specialist bioleaching genera rank high: Leptospirillum at the 91st percentile, Acidithiobacillus at the 77th, Marinobacter at the 75th, and Sulfobacillus at the 71st. Without normalization, species with large open pangenomes (K. pneumoniae, P. aeruginosa) dominate, which reflects genome size rather than metal biology. ₃ However, bioleaching genera as a group are not significantly enriched over background after normalization (Mann-Whitney p=0.17). The report reads this as evidence that metal tolerance genes are broadly distributed across bacteria rather than concentrated in specialists, consistent with its core-genome robustness model. High individual percentiles therefore do not amount to a group-level specialist signal. ₃ The report proposes a remedy for the count-based design: replace the simple count-based metal score with a per-species hypergeometric enrichment test that controls for total functional annotation content. ₃

These rankings remain untested predictions. The atlas scores 27,702 species, but it cannot validate predictions for organisms without RB-TnSeq libraries (random barcode transposon sequencing, a pooled mutant-fitness assay). ₃ Acidithiobacillus ferrooxidans sits at the 77th percentile after normalization, yet the report states that no TnSeq data exist for any Acidithiobacillus species. It proposes RB-TnSeq under Fe, Cu, Co, Ni, and Zn as the first genome-wide functional map of bioleaching gene requirements. ₃ Even among Fitness Browser organisms, the repertoire score is a prediction rather than a measurement. P. putida has the 5th-highest metal gene repertoire score among all FB organisms (0.707), and a full Co, Ni, Cu, Zn, Al, Cr, U panel is proposed as the test that would validate or refute this prediction. ₃

The study used 97,334 bacdive strains, of which 37,368 (38.4%) matched a pangenome species and metal score, covering 5,647 unique GTDB species. ₁ Among these, 3,994 species had at least 5 phenotype features and were used for the principal predictive comparisons. ₁

Evidence that the composite score is not mechanistically specific

The strongest univariate association was between Gram stain and the composite metal-tolerance score: Gram-negative species scored higher than Gram-positive species with Cohen’s d = -0.61, p < 1e-60, and n = 3,272 species. ₁ This supports environmental-resistome at the level of broad genome-associated resistance patterns, but it does not establish that the Gram-negative envelope causes tolerance to any particular metal. ₁ The report calls the association mechanistically plausible, because the Gram-negative outer membrane can restrict metal cation uptake. It also says the association is statistically confounded with phylogeny: it is almost entirely a between-lineage signal that could not be tested within taxonomic classes. ₁

The Gram-stain association was almost entirely between lineages, especially between Gram-positive Actinomycetes and Gram-negative Proteobacteria, and could not be tested within taxonomic classes because of this structure. ₁ Thus, a strong association with a composite score can reflect lineage composition rather than a transferable mechanism. The atlas’s cross-organism core enrichment supports this caution because it can capture conserved cellular processes—including envelope functions, DNA repair, central metabolism, protein quality control, and general stress response—rather than specialized resistance alone. ₃

Urease-positive species had lower composite metal-tolerance scores, with d = -0.18 and p < 1e-5, reversing the prediction that urease-associated nickel handling would confer nickel tolerance. ₁ The effect was driven by Actinomycetes, where d = -0.59 and p < 1e-16, but disappeared within Gammaproteobacteria, where d = +0.08 and was not significant, and Bacilli, where d = +0.06 and was not significant. ₁ This refines the interpretation of the composite score: an apparent phenotype–resistance relationship may fail when tested within the lineage in which the proposed mechanism should operate. ₁ The report reads the urease reversal as a lineage-composition effect rather than evidence about nickel tolerance. Urease positivity correlates with Actinomycetes and other lineages that carry fewer metal resistance genes overall. Urease-associated nickel handling is narrowly specific to nickel and tightly coupled to nickel efflux, so it is a homeostatic balance rather than a broad tolerance mechanism and need not confer general metal tolerance. ₁

Catalase provides a second example of non-mechanistic aggregation. ₁ Its overall association with the composite score was positive, d = +0.10, but catalase-negative species scored higher within Actinomycetes, with d = -0.62 and p < 1e-5, Gammaproteobacteria, with d = -0.49 and p = 0.004, and Betaproteobacteria, with d = -0.51 and p = 0.006. ₁ The overall result therefore reflected between-class composition rather than a consistent within-class relationship. ₁

Other phenotype associations also demonstrate why composite scores should not be read as metal-specific evidence. ₁ Seven of 10 tested features passed FDR (false discovery rate) correction at q < 0.05, including Gram stain, oxidase, motility, urease, enzyme breadth, nitrate reduction, and catalase, while metabolite breadth and acetate utilization were not significant. ₁ H₂S production did not pass correction, with d = -0.867 and q = 7.3e-02, and only 8 H₂S-negative species were present, making that estimate unreliable and likely inflated by small-sample bias. ₁

The per-feature statistics show how strong and how marginal the significant associations were. Gram stain (+) had d = -0.610, n = 3,272, p = 4.0e-61, q = 4.0e-60. Oxidase (+) had d = +0.530, n = 1,799, p = 2.7e-25, q = 1.3e-24. Motility had d = +0.345, n = 3,138, p = 2.2e-23, q = 7.2e-23. Urease (+) had d = -0.175, n = 3,035, p = 3.7e-06, q = 9.1e-06. Enzyme breadth had rho = -0.058, n = 3,746, p = 4.1e-04, q = 8.2e-04. Nitrate reduction had d = +0.100, n = 3,088, p = 4.4e-03, q = 7.4e-03. Catalase (+) had d = +0.104, n = 2,930, p = 2.8e-02, q = 4.1e-02. ₁ The catalase association only marginally passed correction, and it reverses within major classes (see above). Passing FDR on the composite score is therefore weak evidence of a consistent biological relationship. ₁

The gene-resolution analysis supports this limitation mechanistically: among 8,162 metal-important genes, general-stress, metal-shared, and metal-specific tiers had mean pangenome core fractions of 92.0%, 91.0%, and 89.8%, respectively, with fully core fractions of 57.2%, 50.4%, and 45.7%. ₂ The modest but reproducible conservation gradient is consistent with aggregate scores combining broadly conserved stress functions with more specialized, accessory resistance functions; it does not identify which function explains tolerance to any one metal. ₂

The atlas supports this two-tier interpretation at larger scale: among 2,891 ortholog groups with metal phenotypes, 1,182 were conserved across at least two organisms and 601 across at least three, while 149 conserved candidates lacked full functional annotation. ₃ These cross-species fitness candidates are function predictions, not demonstrations that any candidate explains a composite score or a particular metal phenotype. ₃

The metal-specificity analysis refines this conservation interpretation: metal-specific genes had a pooled core fraction of 84.8% (2,969/3,500), compared with 90.2% (2,183/2,420) for general-sick genes and a baseline of 79.8% (73,957/92,650); the comparison between metal-specific and general-sick categories was significant by a Cochran-Mantel-Haenszel test (p=0.011). ₄ Both classes remain core-enriched, but the difference supports a two-tier model in which general stress functions are more conserved while specialized metal determinants are relatively less core-associated. This still does not map a composite score to a particular mechanism. Metal-specific genes also matched metal-resistance keywords at 12.2% versus 7.8% for general-sick genes (Fisher exact OR=1.64, p=2.4e-8), which supports their biological distinction without proving causation for any individual metal. ₄

Predictive models and circularity risk

Taxonomy alone explained 35.4% of composite metal-tolerance variance, phenotype features alone explained 16.3%, and taxonomy plus phenotype yielded R² = 0.345, slightly below taxonomy alone at R² = 0.354. ₁ The phenotype-plus-taxonomy change was delta R² = -0.009, indicating that the measured phenotypes added little independent predictive information after taxonomy was included. ₁ The report states that adding the metal-resistance-cluster count boosted the full model to R² = 0.63. ₁

Adding the number of metal resistance gene clusters, n_metal_clusters, increased the full model to R² = 0.633. ₁ In 5-fold phylogenetic-blocked cross-validation, the gene-count-only model had R² = 0.063 and RMSE = 0.045; phenotype only had R² = 0.163 and RMSE = 0.043; taxonomy only had R² = 0.354 and RMSE = 0.038; taxonomy plus phenotype had R² = 0.345 and RMSE = 0.038; and the full model had R² = 0.633 and RMSE = 0.028. ₁

These results support the interpretation that, in the combined model, genome-encoded resistance repertoire adds predictive information about the composite score beyond taxonomy and the tested classical phenotypes. The gene-count-only model was nonetheless weaker than the phenotype-only model (see Tensions). The results also do not prove that the resistance-gene count captures the causal mechanism for each metal. ₁ The score itself is a genome-based prediction rather than a direct measurement of metal tolerance, so correlations between phenotype and score remain phenotype-to-genome correlations rather than phenotype-to-phenotype measurements. ₁ Controlling for n_metal_clusters reduces one form of circular reasoning in partial correlations, but it does not convert the composite score into a mechanistic assay. ₁

The atlas refines this result: across 559 metal-related experiments and 383,349 gene × metal fitness records, 12,838 genes were broadly metal-important (3.3%), but the simple presence-based repertoire score did not predict metal fitness. ₃ Thus, even when a resistance-gene count predicts a derived composite score, gene presence alone need not predict measured fitness; context-dependent essentiality, regulation, expression, dose, and phylogenetic structure remain possible explanations. ₃

The metal cross-resistance analysis refines this interpretation: all 15 metal pairs tested in at least 5 organisms showed greater than 90% sign consistency, but association magnitudes differed among pairs and no pair was systematically negative. ₂ This pattern is compatible with a universal general-stress component layered with chemistry-specific responses, so a high multi-metal score cannot be assumed to represent a shared molecular mechanism or resistance to every component metal. ₂

The metal-specificity classification provides a direct condition-resolved counterpoint: specificity varied substantially by metal, from 60.6% for manganese and 60.5% for molybdenum to 21.9% for iron, with copper at 51.9% and nickel at 43.7%. ₄ This supports replacing composite scores with per-metal measurements when making mechanistic claims. Candidate prioritization identified UCP030820, YebC, and DUF1043/YhcB as the strongest cross-organism candidates, but the proposed YebC translation-bottleneck mechanism remains explicitly a hypothesis rather than an established explanation. ₄

SHAP (Shapley additive explanation) importance from the full XGBoost model placed taxonomic class/order codes and n_metal_clusters among the top predictors, while phenotype features contributed minimally after taxonomy was included. ₁ This supports phenotype-database-coverage-bias and refines composite-resistance-score-limitations by showing that predictive importance does not establish biological specificity. ₁

Limits of direct and ecological validation

The direct Fitness Browser–BacDive validation contained 12 organisms representing 6 unique species: Cupriavidus basilensis, Methanococcus maripaludis, Ralstonia solanacearum, Pseudomonas simiae, Azospirillum brasilense, and Pseudomonas fluorescens. ₁ All Gram-typed organisms were Gram-negative, preventing within-set testing of the strongest Gram-stain association. ₁ All urease-typed organisms were urease-negative despite being routinely tested against nickel, and the single anaerobe, Methanococcus maripaludis, had only 1 metal tested versus 4–5 for aerobes, so the anaerobe comparison was not interpretable. ₁

The newer BacDive isolation-environment analysis supports the use of the composite score as an ecological prioritization signal: among 42,227 of 97,334 strains matched to 6,426 GTDB species, heavy-metal contamination isolates had a score effect of Cohen’s d = +1.00, while waste/sludge, all-contamination, and industrial groups had d = +0.57, +0.43, and +0.20, respectively. ₅ The heavy-metal comparison used only n=10 matched isolates and barely exceeded an approximately d=0.93 minimum detectable effect at 80% power, so it does not establish a metal-specific mechanism. ₅ Contamination signals remained significant within Pseudomonadota and Actinomycetota but not within Bacillota or Bacteroidota, where contamination groups were small; this refines the interpretation by showing that ecological validation can persist beyond broad phylum composition in some lineages without resolving mechanism. ₅

The isolation-environment report describes this result as validating the metal-fitness-atlas genome-based prediction method against real-world isolation ecology. The scored quantity was derived from cross-species RB-TnSeq fitness data, where RB-TnSeq (random barcode transposon sequencing) is a pooled mutant-fitness assay. It was projected onto 27,702 pangenome species via KEGG functional annotations. ₅ The central digest reports the heavy-metal comparison as Cohen’s d = +1.00, Mann-Whitney p=0.006, n=10. The reported ordering is heavy metal (+1.00) > waste/sludge (+0.57) > all contamination (+0.43) > industrial (+0.20), and the digest states that the signal holds within Pseudomonadota and Actinomycetota after phylogenetic stratification. ₆ Two features limit this result: the projection is itself built from laboratory fitness data, and the heavy-metal group is small. This page therefore grades the reported ‘validation’ as an ecological association that supports the score as a prioritization signal. It is not mechanistic validation of any metal-specific determinant. ₅

The same analysis contradicts the expectation that host niches carry lower scores. Host-associated bacteria scored slightly but significantly higher than environmental bacteria (d=+0.14, p<0.0001), and the central digest puts the host-associated group at n=12,086. _{5 6} Both sources propose genome-size confounding as the likely explanation. Host-associated BacDive strains are dominated by Pseudomonadota pathogens such as Pseudomonas, Klebsiella, and Acinetobacter. These have large genomes and correspondingly more KEGG-annotated gene clusters, which would inflate the normalized metal score. The comparison does not establish this explanation. _{5 6} The analyzed metal_score_norm divides metal clusters by annotated clusters to control for genome size. This matters because Pseudomonadota tend to have larger genomes, and without normalization the environmental signal could partly reflect genome size. The report cautions, however, that normalization reduces but does not eliminate the effect, because metal tolerance functions correlate with total metabolic complexity. ₅ This refines the composite-score caution: even a genome-size-normalized score can carry residual metabolic-complexity and taxonomic-composition signal. ₅

The metal cross-resistance study contradicts any stronger interpretation of this ecological signal as a consistently validated composite phenotype: at the Fitness Browser species scale, multi-metal tolerance scores did not correlate with BacDive metal-associated isolation, with Spearman rho approximately -0.02 and p > 0.8. ₂ After matching and collapsing strains, the effective sample size was 20 independent species, which the study considered too small for a meaningful correlation test. ₂

Community-scale ENIGMA evidence concerns a different aggregate: a composite environmental exposure index, not a fitness-derived tolerance score. The contamination index was the row-wise mean of per-metal log1p z-scores for arsenic, cadmium, chromium, copper, lead, nickel, uranium, and zinc. ₇ The confirmatory endpoint was a community defense functional-potential signal built from COG-fraction proxies. COG (clusters of orthologous groups) fractions are coarse functional summaries, not curated resistance pathways. This endpoint was re-tested under four contamination index variants: composite all-metals, uranium-only, top-3 variance metals, and the first component of a PCA (principal component analysis) of metal z-scores. Across both confirmatory mapping modes, all index variants remained non-significant after FDR (false discovery rate) correction (q = 0.546 across the 8 confirmatory-variant tests), including uranium-only. ₇ This null shows only that replacing the multi-metal exposure index with single-metal or alternative indices did not recover an association for this coarse COG-derived endpoint. It does not test uranium-specific laboratory fitness scores. It also cannot show whether aggregating tolerance scores hides metal-specific signal. ₇

A lab–field ecology analysis calls the metal tolerance score a crude aggregate and suggests that condition-specific fitness scores (e.g., uranium-only) would be more informative. ₈ It also cautions that the contamination-associated community shift is not simply “more metal-tolerant organisms”. The shift reflects broader ecological restructuring, including changes in rare taxa, redox conditions, and carbon/energy sources. ₈ The ENIGMA index-sensitivity null does not bear on this proposal, because it varied the exposure index while keeping a COG-derived functional endpoint. Whether condition-specific fitness scores outperform the aggregate therefore remains a hypothesis to test. _{7 8}

The study therefore did not establish that catalase predicts copper tolerance, urease predicts nickel tolerance, or H₂S production predicts zinc, copper, or cadmium tolerance. Testing whether specific phenotypes predict tolerance to specific metals, rather than composite scores, would require per-metal scores for the 27K species. ₁ The composite-score result is consequently best treated as an association and prioritization signal, not as evidence for a metal-specific mechanism. ₁ The isolation-environment result strengthens that conclusion’s ecological side, but the 24 matched metal-utilization records were inconclusive (Mann-Whitney p=0.14; Cohen’s d=-0.57), so curated phenotype data did not independently establish mechanism. ₅

Repertoire breadth versus gene count in ecological association

A MicrobeAtlas niche-breadth analysis gives a further null for count-type repertoire scores. It used PGLS (phylogenetic generalized least squares, regression that accounts for shared ancestry). Total AMR gene cluster count and core AMR fraction did not predict niche breadth in either the simple or the multi-predictor PGLS models. The report interprets this as showing that the breadth of the metal resistance repertoire, rather than its depth (many genes for few metals), is associated with ecological versatility. ₉ For Levins’ B_std, AMR cluster count did not pass the Bonferroni threshold (β = +0.010, p = 0.090), and core fraction was also non-significant in simple PGLS (β = +0.003, p = 0.523). ₉ In the multi-predictor model, AMR clusters had β = −0.002, SE = 0.0065, p = 0.789, and core fraction had β = +0.006, SE = 0.0054, p = 0.288. ₉ By contrast, the full seven-metal diversity predictor had β = +0.0215, SE = 0.0056, p = 1.5×10⁻⁴. It remained significant after genome size was formally controlled in PGLS (β = +0.022, p = 3.6×10⁻⁴). ₉ Genera with many AMR genes for a single metal type are not ecological generalists, whereas genera spanning multiple metal resistance families are. ₉ This supports the atlas result that simple gene presence/absence does not predict metal tolerance. Both projects find that a raw repertoire count is a weak summary, although they use different outcomes: niche breadth in one, laboratory metal fitness in the other. _{9 3} It also refines the composite-score caution: how a repertoire is summarized changes the result, so a single aggregate count can hide the component that carries the signal. The association is correlational, however, and the report states that the causal direction cannot be established without time-series data or a natural experiment. ₉

The breadth result is not robust to removing any single metal. Each leave-one-metal-out model (n=606) kept a positive coefficient, but none of the seven exclusions passed BH-FDR (Benjamini-Hochberg false discovery rate) correction: ₉

• excluding mercury: β = +0.0082, SE = 0.0069, p = 0.233, q = 0.313 ₉
• excluding arsenic: β = +0.0039, SE = 0.0061, p = 0.518, q = 0.572 ₉
• excluding copper: β = +0.0102, SE = 0.0065, p = 0.115, q = 0.187 ₉
• excluding zinc: β = +0.0095, SE = 0.0064, p = 0.140, q = 0.212 ₉
• excluding cadmium: β = +0.0084, SE = 0.0064, p = 0.186, q = 0.258 ₉
• excluding chromium: β = +0.0088, SE = 0.0064, p = 0.173, q = 0.255 ₉
• excluding nickel: β = +0.0095, SE = 0.0064, p = 0.140, q = 0.212 ₉

The report concludes that no single metal is the driver. All seven exclusions keep a positive β and all drop to p > 0.10. The report attributes the loss of significance to the maximum possible diversity score falling from 7 to 6, which compresses predictor variance and costs power. ₉ Its own interpretive caveat is more cautious. The pattern is consistent with either (a) a genuinely distributed signal or (b) an original finding near the significance threshold that the compressed predictor range pushes below it, and the two cannot be definitively distinguished without a larger metal type vocabulary. ₉ A later passage of the same report claims that the leave-one-out analysis shows the result requires contributions from multiple metal types and cannot be explained by a single category’s spurious annotations. This page grades that claim as stronger than the leave-one-out results support, because every exclusion was non-significant. ₉ The report’s response table similarly marks single-metal robustness as addressed, but it states that formal AMRFinderPlus validation of the metal-type classification was not performed; only spot checks were done. ₉

The report also notes that its metal-type categories are themselves partly aggregated. The “other” metal category (17% of annotated clusters) captures metal-associated genes outside the seven primary types and may include weaker hits. The report describes excluding it as producing a 6-metal analysis, yet lists seven metals (Hg, As, Cu, Zn, Cd, Cr, Ni); the source leaves this inconsistency unresolved. ₉ The report proposes grouping metal types by genetic co-occurrence (e.g., Cu-Zn co-resistance versus Hg-As) and testing whether specific co-resistance combinations predict niche breadth better than the raw count. That test would replace a breadth count with chemistry-resolved components. ₉

Field-scale functional associations face a parallel aggregation problem on the exposure side. In a soil analysis, all reported associations are observational, and co-contamination is a significant confound because Cr, Cu, Pb, and Zn co-vary in many industrial soils. Whether individual COG-metal associations are metal-specific, or instead reflect a generic stress response to multi-metal contamination, remains unresolved. ₁₀ This supports the page’s distinction between aggregate association and metal-specific mechanism, now applied to exposure as well as tolerance. ₁₀

Tensions

The two BacDive analyses report different species-level bridge totals: the phenotype analysis reports 37,368 matched strains across 5,647 GTDB species, whereas the isolation-environment analysis reports 42,227 matched strains across 6,426 GTDB species. _{1 5} This tension may reflect different matching or filtering pipelines and should be resolved before treating the coverage estimates as interchangeable.

The isolation-environment analysis reports a strong heavy-metal association for the composite score (Cohen’s d = +1.00), whereas the metal cross-resistance analysis reports no species-scale correlation with metal-associated isolation (Spearman rho approximately -0.02, p > 0.8). _{5 2} This tension cannot be resolved from the available summaries because the analyses differ in matching, aggregation, and outcome definition; a common strain-level bridge and preregistered metal-specific outcomes are needed. ₂

The counter-ion comparison introduces a further methodological tension rather than a biological resolution: metal-important genes overlapped with NaCl-stress genes at 39.8% in the counter-ion analysis, whereas 14.7% were sick under osmotic stress in the metal-specificity analysis. ₄ The difference was attributed to stricter fitness and statistical thresholds and partially different organism sets; the directional agreement supports overlap between metal and osmotic stress while showing that composite or thresholded scores are sensitive to operational definitions. ₄

The atlas’s genome-wide core-enrichment result is directionally consistent with the existing two-tier interpretation, but its 87.4% core fraction for broad metal-important genes is not directly interchangeable with the 84.8% pooled core fraction for metal-specific genes or the 92.0%/91.0%/89.8% tiers in the cross-resistance analysis because the sources use different gene sets and definitions. _{3 4 2} This tension should be resolved by applying a common locus set, conservation definition, metal-specificity threshold, and non-metal control set rather than by averaging the estimates.

The phenotype report is internally inconsistent about the standalone value of gene content. Its narrative states that n_metal_clusters explains more variance than all phenotype features combined and substantially more than taxonomy alone. It attributes the lion’s share of the full model’s improvement to gene count (delta R² = +0.28 over taxonomy alone). Yet its 5-fold phylogenetic-blocked CV table (n = 3,994) reports R² = 0.063 for gene count only, 0.163 for phenotype only, 0.354 for taxonomy only, 0.345 for taxonomy plus phenotype, and 0.633 for the full model. ₁ The central digest repeats the ‘true predictor’ framing. It reports that phenotype features alone yield R²=0.16, with 7/10 significant after FDR, but add no predictive value to a taxonomy model (delta R²=-0.009). It reports that the full model with genomic metal-resistance gene count reaches R²=0.63, and that urease-positive bacteria show lower tolerance (d=-0.18), attributed to Actinomycetes lineage composition. ₆ The tabled values are compatible with gene count contributing conditionally on taxonomy rather than acting as a strong standalone predictor. This tension should be resolved by reporting a taxonomy-plus-gene-count model and partial contributions, not by preferring either the narrative or the table. ₁

Relation to the wider wiki

This page supports condition-specific-fitness because the proposed remedy is to replace aggregate scores with measurements resolved by metal and experimental condition. ₁ It also refines circularity-in-metabolic-model-validation by showing that a genome-derived target can be useful for prediction while remaining unsuitable as independent mechanistic validation. ₁ The atlas further supports this distinction because its presence-based pangenome score failed to predict metal fitness and because genome-size normalization changed cross-species rankings. ₃ The ecological validation further supports environmental-resistome while retaining the distinction between ecological association and metal-specific mechanism. ₅ The cross-resistance result supports metal-cross-resistance as a framework for separating universal directional stress effects from chemistry-dependent magnitudes. ₂ The complete source analyses are available at BacDive Phenotype Signatures of Metal Tolerance, BacDive Isolation Environment × Metal Tolerance Prediction, Gene-Resolution Metal Cross-Resistance Across Diverse Bacteria, Metal-Specific vs General Stress Genes, and Pan-Bacterial Metal Fitness Atlas. _{1 5 2 4 3}

Open Directions

• Match BacDive strains to GTDB genomes using GCA accessions and repeat the phenotype-to-score analysis to test whether species-name matching limited mechanistic resolution; reconcile the 37,368/5,647 and 42,227/6,426 bridge totals first. _{1 5}
• Replace the composite score with per-metal scores and use PGLS (phylogenetic generalized least squares) or phylogenetic PCA to test whether catalase predicts copper, urease predicts nickel, and H₂S production predicts zinc, copper, or cadmium tolerance after removing phylogenetic signal. ₁
• Assemble urease-positive and urease-negative organisms within the same taxonomic classes, ideally Gammaproteobacteria where the pangenome-scale urease effect is near zero, and apply tnseq or crispri under nickel and other metals to test nickel-specific mechanisms rather than general composite tolerance. ₁
• Experimentally test the H₂S hypothesis with balanced positive and negative controls under zinc, copper, and cadmium exposure to determine whether the observed d = -0.867 was a small-sample artifact. ₁
• Compare composite-score predictions with direct Fitness Browser growth or fitness measurements across matched organisms and metals, and integrate ENIGMA CORAL community data from the Oak Ridge metal-contaminated site, to determine which resistance-gene features predict measured tolerance rather than a derived genome-based score. _{1 5}
• Reanalyze a common strain-level BacDive/Fitness Browser bridge with per-metal scores, metal-dose normalization, and non-metal stress controls to determine whether the positive cross-metal layer is general stress rather than metal-specific resistance. ₂
• Resolve the 40.7% locusId-related gene attrition, validate classifications against the Fitness Browser’s specificphenotype annotations, and repeat module analysis with precomputed z-scored activities to test whether candidate metal-specific modules were missed by scale and callability effects. ₄
• Refit the atlas using concentration-relative-to-MIC normalization, phylogenetic independent contrasts, and genes important for metals but not other stresses; test whether regulatory or expression features outperform presence-only repertoire scores for direct per-metal fitness. ₃
• Functionally annotate the 149 conserved atlas candidates with PaperBLAST, InterPro, and structural prediction, then test the strongest candidates under matched metal and non-metal conditions to distinguish general stress from metal-specific mechanisms. ₃
• Re-test the host-associated versus environmental contrast within Pseudomonadota, with total annotated-cluster count as a covariate. This would show whether the d=+0.14 host-associated excess reflects residual metabolic-complexity confounding of metal_score_norm. ₅
• Replace the count-based atlas score with the proposed per-species hypergeometric enrichment test that controls for total functional annotation content. Then test whether bioleaching genera become enriched as a group, beyond their high individual percentiles. ₃
• Build RB-TnSeq libraries for Acidithiobacillus ferrooxidans under Fe, Cu, Co, Ni, and Zn. Separately, run the Co, Ni, Cu, Zn, Al, Cr, U panel on P. putida. Together these would test whether normalized and repertoire-score ranks predict measured metal fitness. ₃
• Formally validate the MicrobeAtlas metal-type assignments with AMRFinderPlus, expand the metal-type vocabulary, and test genetic co-resistance groupings (e.g., Cu-Zn versus Hg-As). This would distinguish a distributed breadth signal from a near-threshold result. ₉