Phylogenetic Confounding of Pangenome Associations

Phylogenetic confounding occurs when differences among taxonomic groups dominate an aggregate association, potentially hiding relationships that exist within lineages or creating apparent relationships because two traits share the same taxonomic distribution. ₁

The Pan-Bacterial AMR Gene Landscape provides a direct example using antimicrobial-resistance (AMR) burden and pangenome openness across bacterial species. ₁ Pangenome openness positively correlated with AMR count in 8/10 tested phyla, but the overall correlation across the dataset was near zero (rho=0.006), indicating that phylogeny dominated the aggregate signal. ₁ The strongest within-phylum associations were observed in Bacillota (rho=0.219, p=1.0e-16) and Bacillota_C (rho=0.374, p=3.1e-5). ₁

This pattern supports the interpretation that a near-zero global association does not demonstrate the absence of a relationship between openness and AMR burden. ₁ Instead, associations may be positive within several phyla while differing in magnitude, direction, or baseline across phyla, causing them to cancel when species are pooled. ₁ The result is an association rather than a causal inference, because the analysis describes correlations across pangenome species and does not by itself establish that openness causes AMR accumulation. ₁

The Pangenome Openness Analysis refines this interpretation by testing whether openness itself predicts the relative influence of environment or phylogeny on gene-content variation. In the matched species set, openness was not significantly related to the environment effect (Spearman rho = -0.05, p-value = 0.54) or the phylogeny effect (Spearman rho = 0.03, p-value = 0.73). ₂ Thus, the AMR example supports lineage-aware analysis, but the broader result does not support treating open or closed pangenome status as a general predictor of whether environmental or phylogenetic structure dominates. ₂ This does not remove phylogenetic confounding from individual associations; it limits the claim that a single openness metric predicts that confounding across species. ₂ The report also cautions that its environment and phylogeny effects come from partial correlations that may not fully disentangle confounded variables. ₂

Within-species AMR evidence supports the need for this lineage-aware interpretation at a finer scale. Across 1,305 species and 180,025 genomes, 55.6% of 1,261 species tested showed significant correlation between ANI (average nucleotide identity) distance and AMR Jaccard distance at FDR < 0.05, and 87.8% had positive correlations; the median Mantel r for all AMR was 0.247. ₃ Non-core AMR genes had a higher median Mantel r than core genes (0.222 versus 0.117; paired t-test t = -8.35, p = 7.0e-16, n = 489), consistent with acquired resistance elements becoming lineage-associated, although near-universal core genes have little Jaccard-distance variance and can therefore produce attenuated distance-based correlations. ₃ This refines the cross-species argument: shared ancestry can structure AMR repertoires among strains within the same named species, but the contrast between non-core and core signal is partly a property of prevalence and metric variance rather than proof of a distinct biological process. ₃ The significant species numbered 701 of the 1,261 tested (55.6%, FDR < 0.05). ₃ Core AMR genes are defined by >=95% prevalence, so they are nearly universal and have near-zero Jaccard distances with little variance. The report notes that this inherently suppresses distance-based metrics such as the Mantel r, a statistical artifact that partly explains the lower core signal, independent of biology. ₃ The central discoveries digest reports the same contrast across 1,261 species: non-core AMR genes had a median r=0.222 versus 0.117 for core genes, with paired t-test p=7.0e-16. ₄ The digest reads this as stable vertical inheritance, but that is an interpretation of a distance correlation, not a direct measurement of inheritance. ₄

Phylogeny can also generate an aggregate association when both variables are concentrated in the same clades, even if the variables are not mechanistically related within those clades. ₁ In this dataset, Gammaproteobacteria contained 45% of all AMR clusters (37,752/83,008), while the taxonomic census recorded 4,992 Pseudomonadota species with AMR, 42,904 total AMR hits, 8.6 mean AMR hits per species, and 67.0% of species carrying AMR. ₁ These taxonomic concentrations make lineage composition a necessary consideration when interpreting pan-bacterial AMR summaries. ₁

The Prophage-AMR Co-mobilization Atlas supports this lineage-aware interpretation while showing that it is not the only relevant structure. Across 4,770 species, prophage marker density was positively associated with AMR repertoire breadth (Spearman rho=0.572, p<10^-300, n=4,770 species); after controlling for genome count, the partial Spearman correlation remained rho=0.464 with p=1.0×10^-253. ₅ The association was significant across all five reported major phyla—Pseudomonadota (rho=0.54), Bacillota_A (rho=0.55), Bacillota (rho=0.40), Bacteroidota (rho=0.59), and Actinomycetota (rho=0.29)—which supports a relationship extending beyond a single clade, while not proving phage-mediated AMR transfer. ₅ The report also found only a modest and threshold-dependent enrichment of accessory AMR genes near prophage markers, with an odds ratio of 1.10 at 10 genes and a median species-level odds ratio of 0.85, so species-level covariance should not be treated as evidence that local prophage proximity uniformly drives AMR acquisition. ₅

The same issue applies to comparisons of environmental AMR patterns. ₁ Human/Clinical species carried 10.6 AMR clusters per species (n=2,248), compared with 4.6 for Soil/Terrestrial (n=2,469), 3.9 for Aquatic (n=1,827), and 3.0 for Animal (n=959), with a significant Kruskal-Wallis result (H=440, p=7.0e-93). ₁ Clinical AMR was less core (30.8%) than soil AMR (58.1%) or plant AMR (63.1%), but these contrasts may combine environmental effects with differences in taxonomic composition, sampling, and genome representation. ₁ The report identifies database sampling bias because genome collections over-represent clinical pathogens, potentially inflating AMR counts for human-associated species. ₁

The Environmental Resistome at Pangenome Scale supports and refines this caution with a larger analysis of 14,723 bacterial species and 293K genomes: environment was associated with AMR burden and mechanism composition, but the effects persisted only incompletely across phylogenetic strata. ₆ Five of six tested major phyla showed significant within-phylum environment effects, with Bacteroidota having the largest effect (eta-squared = 0.130); at the family level, 20 of 141 testable families showed significant effects after FDR correction, while most families lacked sufficient environmental breadth. ₆ Thus, the new result supports a genuine ecological component beyond taxonomic composition, while refining the claim by showing that phylogenetic control remains limited by uneven sampling and incomplete within-lineage coverage. ₆ The larger analysis also found that clinical species had a median of 5 AMR gene clusters versus 2 for soil, aquatic, and host-associated species (Kruskal–Wallis H = 781.9, p = 9.4×10⁻¹⁶⁷, eta-squared = 0.056), but it remained observational and subject to sampling and annotation bias. ₆

In that analysis, Pseudomonadota, Bacillota_A, Actinomycetota, Bacteroidota and Bacillota showed significant within-phylum environment effects, and Chloroflexota was the only non-significant major phylum. Bacteroidota showed the largest within-phylum effect (η² = 0.130). ₆ The 20 significant families were 14% of the 141 testable families. ₆ The strongest were Enterobacteriaceae (q = 3×10⁻²¹), Bacteroidaceae (q = 1.3×10⁻¹⁷), Lachnospiraceae (q = 2.5×10⁻¹²) and Pseudomonadaceae (q = 3.9×10⁻¹²). ₆ The report states that environment and phylogeny remain deeply entangled despite family-level control, because some families, such as many Enterobacteriaceae, are almost entirely clinical. ₆ It treats the within-family effects as its most convincing evidence, although most families do not span enough environments to test. ₆ All tests were highly significant (p ~ 0), but effect sizes were modest (η² = 0.02–0.13), so environment explained 2-13% of variance in AMR composition. ₆ The report suggests the hypothesis that phylogeny explains much more of this variance, but it did not quantify that share. ₆

The within-species study further supports and qualifies these environmental comparisons. Host-associated species carried more AMR genes per genome than terrestrial or aquatic species under both a BacDive keyword approximation and an NCBI keyword classifier, with Kruskal-Wallis p < 0.05; however, only 2 species had sufficient within-species environmental diversity for strict environment–AMR-ecotype testing, while 52.7% of genomes had no classifiable isolation_source. ₃ The direction of the host-associated contrast was consistent across classifiers, but the report cautions that clinical and human-associated overrepresentation, approximate environment labels, and sparse metadata prevent a general environment-to-ecotype inference. ₃

The concept connects directly to pangenome-integration, because interpreting pangenome openness requires integrating taxonomy, gene-family distributions, and environmental metadata rather than treating species as exchangeable observations. ₁ It also qualifies interpretations of environmental-resistome: environmental and clinical AMR differences should be tested with lineage-aware models before being attributed to environment alone. ₁ The newer within-phylum, within-family, and within-species results support this qualification but do not eliminate it. _6,3 Associations between AMR burden and condition-specific-fitness should likewise be evaluated across comparable lineages and organisms, because cross-organism fitness measurements can inherit taxonomic structure. ₁

Lineage Confounding Beyond AMR

The BacDive Phenotype Signatures of Metal Tolerance extends the pattern to metal tolerance. Its metal tolerance scores are Metal Fitness Atlas genome-based predictions, not direct measurements of metal tolerance, so its phenotype associations are phenotype-to-genome correlations rather than phenotype-to-phenotype ones. ₇ Gram-negative species had higher metal tolerance scores than Gram-positive species (Cohen’s d = -0.61, p < 1e-60, n = 3,272 species). ₇ The report’s table gives this as d = -0.610, p = 4.0e-61, with false discovery rate (FDR, the expected share of false positives among significant calls) q = 4.0e-60. ₇ This was the largest effect among the phenotype features tested. ₇ However, class-stratified analysis showed that the effect could not be tested within taxonomic classes, because it was almost entirely a between-lineage signal: Gram-positive Actinomycetes versus Gram-negative Proteobacteria. ₇ The dataset contained no Gram-positive Proteobacteria and no Gram-negative Actinomycetes. ₇ As a result, the Gram-stain effect could not be separated from lineage in prediction, even though an outer-membrane permeability barrier is a plausible mechanism. ₇

Urease-positive species had lower metal tolerance scores (d = -0.18, p < 1e-5). ₇ The table gives d = -0.175, n = 3,035, p = 3.7e-06, FDR q = 9.1e-06. ₇ This runs opposite to the prediction that nickel-import machinery would confer tolerance. ₇ Actinomycetes drove the association (d = -0.59, p < 1e-16), and it was non-significant within Gammaproteobacteria (d = +0.08) and Bacilli (d = +0.06). ₇ The report interprets the reversal as a lineage-composition effect: urease positivity co-occurs with lineages that carry fewer metal resistance genes, and urease-associated nickel handling need not confer general metal tolerance. ₇ Whether urease-positive organisms are specifically nickel-tolerant remains unresolved. ₇ The report proposes per-metal fitness profiling of urease-positive versus urease-negative bacteria within one taxonomic class, ideally Gammaproteobacteria. ₇

Catalase shows a sign reversal (Simpson’s paradox). The overall association was marginally positive (d = +0.104, n = 2,930, p = 2.8e-02, FDR q = 4.1e-02). ₇ Yet catalase-negative species scored higher within every major class: Actinomycetes (d = -0.62, p < 1e-5), Gammaproteobacteria (d = -0.49, p = 0.004) and Betaproteobacteria (d = -0.51, p = 0.006). ₇ The report attributes the pooled positive effect to between-class composition. ₇ This refines the openness–AMR example: there, pooling hid positive within-phylum associations, whereas here pooling reversed the within-class direction. _7,1

In the report’s model comparison, taxonomy alone (phylum/class/order) explained 35.4% of metal tolerance variance and phenotype features alone explained 16.3%. ₇ Combining taxonomy and phenotype gave R² = 34.5%, slightly worse than taxonomy alone. ₇ Adding the number of metal resistance gene clusters raised the full model to R² = 0.63. ₇ SHAP (Shapley additive explanation) feature importance from the full XGBoost model ranked taxonomic class/order codes and metal-resistance-cluster count as the top predictors. ₇ Phenotype features contributed minimally once taxonomy was included. ₇ This supports the page’s central caution that cross-species phenotype associations can act as phylogenetic proxies. ₇ The predictive value of gene-cluster count needs a caveat, though: the target score is itself a genome-based prediction, not a measured tolerance. The report says controlling for the cluster count in partial correlations only mitigates this circularity. The result also comes from a single dataset. ₇

Phylum stratification in the BacDive Isolation Environment × Metal Tolerance Prediction gives a mixed result for the same kind of genome-based metal tolerance score. ₈ Contamination isolates still had significantly higher metal scores than environmental isolates within Pseudomonadota (85 contamination versus 5,655 environmental isolates, delta=+0.040, p<0.001). ₈ The same held within Actinomycetota (62 versus 772 isolates, delta=+0.035, p<0.001). ₈ The signal was not significant within Bacillota (19 versus 492 isolates, delta=-0.012, p=0.285) or Bacteroidota (5 versus 156 isolates, delta=-0.008, p=0.456), where contamination-isolate samples were small. ₈ This supports a within-lineage signal in the two significant phyla. ₈ The report gives two possible explanations for the Bacillota and Bacteroidota nulls, and neither is established. One is real biology: these phyla are primarily gut-associated and rarely encounter metal contamination. The other is limited statistical power, because contamination isolates from these phyla are rare in BacDive. ₈

The Self-Sufficiency, Anaerobic Toolkit, and Cultivation Bias in Clay-Confined Cultured Bacterial Genomes gives a habitat example in which within-phylum control removed most of an apparent signal. ₉ At cohort level, deep-confined clay isolates jointly carried the Wood–Ljungdahl, group 1 [NiFe]-hydrogenase and dissimilatory sulfate-reduction toolkit at higher rates than other cohorts. ₉ On a scale of 3 modules, their mean toolkit score was 1.89, versus 0.39 for soil baseline and 0.03 for shallow clay. ₉ Wood–Ljungdahl completeness occurred in 5 of 9 deep versus 15 of 140 soil-baseline genomes (odds ratio OR=10.4, Benjamini–Hochberg-adjusted p=0.004). ₉ [NiFe]-hydrogenase completeness occurred in 7 of 9 versus 35 of 140 genomes (OR=10.5, adjusted p=0.004). ₉ Neither enrichment survived within-phylum control. ₉

The Bacillota_B phylum contains Desulfosporosinus, BRH-c4a and BRH-c8a, the lineages that dominate the anchor cohort. _9,4 It carries Wood–Ljungdahl and [NiFe]-hydrogenase at high background rates even in soil samples, with a Bacillota_B baseline mean toolkit score of 1.65. _9,4 Within Bacillota_B, deep-versus-baseline comparisons were not significant for Wood–Ljungdahl (5/5 vs 15/19, p=0.54) or [NiFe]-hydrogenase (5/5 vs 14/19, p=0.54). ₉ GapMind means also did not establish a difference: 16.50 for 4 deep anchors versus 16.79 for 19 baseline genomes (d=−0.13, p=0.073). ₉ Only sulfate reduction (dsrAB-aprAB-sat) survived the phylogenetic control, present in 5/5 deep versus 4/19 soil-baseline Bacillota_B isolates (OR=∞, p=0.003). ₉ The report gives the adjusted p-value as p_BH=0.04 in its text and 0.044 in its table. ₉

The discoveries digest turns this into a methodological warning for subsurface comparative genomics. An apparent habitat signal at cohort level may be a phylum-level signature picked up by biased habitat sampling. ₄

A companion accessory-genome comparison in the Subsurface Bacillota_B Specialization — What Distinguishes Deep-Clay Lineages from Soil Congeners? faced the same constraint with 10 anchor versus 62 baseline genomes. ₁₀ The 10-genome anchor is genus-clumped: 3 BRH-c8a, 2 BRH-c4a, 2 Desulfosporosinus, Desulforudis, Ch130 and 1 other. ₁₀ Spanning four orders (Desulfotomaculales, Desulfitobacteriales, Ammonifexales, Thermacetogeniales) only partly mitigates genus-level phylogenetic confounding. ₁₀ Marginal ortholog-group (OG) effects with anchor counts of 3–5 should be treated descriptively rather than inferentially. ₁₀ Fisher’s exact test detects large effects robustly, including the 547 enriched OGs. ₁₀ Without per-genus analysis, however, those OGs cannot be confidently separated into genus-specific lineage markers versus a recurrent deep-clay signature. ₁₀

Phylogenetic structure also complicates co-inheritance tests of functional coupling in the Co-fitness and Co-inheritance in Bacterial Pangenomes. ₁₁ Phylogenetic distance stratification covered 7 of 9 organisms. ₁₁ Cofit pair phi (the phi coefficient of co-occurrence) was higher among near genomes (mean=0.102) than medium genomes (mean=0.067), as expected from shared ancestry. ₁₁ Most species lacked genomes in the far stratum (>0.05 branch distance), which limits separation of functional coupling from phylogenetic signal; the report’s figures/fig3_phylo_control.png shows this stratification. ₁₁ The two Ralstonia organisms were excluded because they had zero co-fitness data in the Fitness Browser, even though they were the most phylogenetically diverse, lowest-ANI cases. ₁₁ Near-clonal species retained detectable signals: Ddia6719 (ANI 99.47%, delta=+0.093) and pseudo3_N2E3 (ANI 99.66%, delta=+0.026). ₁₁ Their high baseline phi, however, makes absolute phi values less interpretable. ₁₁ The report sets its question against Coinfinder, which Whelan et al. (2020) developed to detect pangenome gene associations while controlling for phylogeny (PMID: 32100706). ₁₁ It asks whether fitness-measured coupling predicts such associations. ₁₁

The Rational Design of Protective Microbiome Formulations for Competitive Exclusion of *Pseudomonas aeruginosa* in Cystic Fibrosis Airways shows lineage structure in a different predictive setting. ₁₂ Adding genus-level taxonomy raised training R² to 0.360, an additional 8.6% of variance. ₁₂ The report attributes this to likely species-level mechanisms such as direct antagonism, which it did not directly prove. ₁₂ Virulence classification was strongly correlated with strain group: entire groups were uniformly PAO1-like or PA14-like. ₁₂ The report reads this as clonal inheritance of type III secretion system (T3SS) effector type rather than horizontal transfer. ₁₂

Marker-gene enrichments need the same scrutiny. ₄ The discoveries digest reports root-associated enrichment of ACC deaminase (acdS) with an odds ratio of 69.3, but does not describe that estimate as phylogenetically controlled. ₄ After phylogenetic control, 50 eggNOG ortholog groups distinguished plant-associated from non-plant species. ₄ COG3569 (a cluster-of-orthologous-groups family) was the top hit, with OR=6.01 after phylogenetic control. ₄ Plant-interaction functions for these groups remain candidate interpretations. ₄

Within-species gene content shows the same lineage dominance. The Ecotype Correlation Analysis analysed 172 species with sufficient data. _13,4 The median partial correlation was 0.0025 for environment and 0.0143 for phylogeny. _13,4 Phylogeny dominated in 60.5% of species and environment in 39.5%. _13,4 Environmental and host-associated bacteria did not differ significantly in this comparison (p=0.66). ₄ The report suggests that vertical inheritance dominates the gene-content signal, but this is an interpretation of small partial correlations rather than a direct demonstration. ₁₃ The report states the general conclusion that, for most bacterial species, phylogenetic history is a stronger predictor of gene content than environmental similarity. ₁₃

Related ecotype analyses leave alternative lineage explanations open. The Ecotype Reanalysis — Environmental vs Human-Associated Species raises an untested alternative for human-associated species such as Klebsiella or Enterococcus: different lineages dominate different regions, so embedding-based environment–gene-content associations could be real but epidemiological rather than ecological. ₁₄ The Ecotype Functional Differentiation cautions that gene-content ecotypes may reflect phylogenetic substructure within species; without within-species phylogenetic controls such as core-genome trees, adaptive ecotypes cannot be distinguished from demographic subpopulations. ₁₅ The Metabolic Capability vs Metabolic Dependency likewise treats environment-cluster correlations in Salmonella and Phenylobacterium as observational associations that could be confounded by phylogenetic structure without explicit phylogenetic correction. ₁₆ These caveats support the page’s central caution at the within-species scale. _14,15,16

Formal Phylogenetic Controls and Their Limits

The Metabolic Capability vs Dependency used genus-level grouping from GTDB species clade names as its phylogenetic control and did not perform full phylum-level stratification. ₁₇ Restricted to genera with ≥20 species, the pathway–openness signal was positive in 13 of 18 genera and significant at p<0.05 in 5 of 18 (Clostridium, Eubacterium, Mesorhizobium, Pseudomonas, Streptomyces), strongest within Clostridium (rho=0.534) and Eubacterium (rho=0.506); small per-genus samples limit statistical power. ₁₇ The report suggests the hypothesis that the relationship reflects a general evolutionary dynamic rather than a clade-specific artifact, but this is not established. ₁₇ Correlations were controlled for genome count and checked within taxonomic groups, yet full phylogenetic independent contrasts (PIC, a method that converts trait values into evolutionarily independent comparisons along a tree) were not computed; citing the ecotype finding that phylogeny dominates gene content in 60.5% of species, the report notes that metabolic ecotypes may partly reflect intra-species phylogenetic structure. ₁₇ This supports the within-species ecotype caution above. _17,13

The PGP Gene Distribution Across Environments & Pangenomes refines the acdS marker example above. _18,4 ACC deaminase (acdS) enrichment in soil/rhizosphere species survived phylum-level fixed effects in logistic regression (OR = 6.98, p = 4.8e-61) and held in a strict rhizosphere-only sensitivity analysis restricted to genomes with “rhizosphere” or “root nodule” in their isolation source (acdS OR = 10.6, q = 7.6e-38). ₁₈ These estimates differ from the digest’s odds ratio of 69.3, which the digest does not describe as phylogenetically controlled. _18,4

The Lanthanide Methylotrophy Atlas — Summary provides a case in which formal correction strengthened rather than weakened a pooled result. ₁₉ Its initial per-phylum binomial framework did not formally control for phylogenetic non-independence, a reviewer point that led to three validation strategies for H1 (xoxF dominance). ₁₉ The naive genome-level analysis used 3,885 units and estimated an xoxF fraction of 0.950 (95% CI [0.942, 0.956]); equal weighting of 271 GTDB families gave 0.960 [0.941, 0.976]; and a Bayesian binomial mixed model with phylum and family random intercepts used 3,885 units and estimated 0.993 [0.992, 0.994]. ₁₉

The variational-Bayes binomial mixed model gave a phylogeny-corrected xoxF:mxaF ratio of ~143 : 1 (95% credible interval [122, 169]); the report explains that the random effects absorb phylum-level heterogeneity and reveal that, within typical phyla, xoxF dominance is more extreme than the global pooled estimate suggests. ₁₉ All three validations had 95% interval lower bounds above the 10:1 H1 threshold, as compared in h1_phylogenetic_validation.png. ₁₉ For environmental hypothesis H2, a within-Acidobacteriota analysis retained soil/sediment enrichment (within-phylum OR=2.16, p_BH=2.2 × 10⁻⁵), which suggests the hypothesis that at least part of the broader soil signal is not pure phylogenetic confounding. ₁₉ However, H2 still relies on within-phylum stratified analyses rather than a fully phylogeny-aware mixed model. ₁₉

Cross-organism fitness conservation analyses face a related sampling problem. The Pan-Bacterial Metal Fitness Atlas notes that multiple Pseudomonas fluorescens strains inflate apparent cross-species conservation. ₂₀ In its sensitivity analysis, excluding 4 duplicate P. fluorescens FW300 strains (keeping only pseudo3_N2E3) reduced the analysis from 22 to 18 organisms, and the core enrichment remained (OR=2.065, p=5.9e-141, versus OR=2.083, p=4.3e-162 with all organisms). ₂₀ This supports the report’s conclusion that the core enrichment is not driven by Pseudomonas overrepresentation, but removing duplicate strains is not a formal phylogenetic correction; the report lists phylogenetic independent contrasts as future work for cross-organism conservation comparisons. ₂₀

The Gene-Resolution Metal Cross-Resistance Across Diverse Bacteria states the same limitation directly: its 28 organisms are not phylogenetically independent, and formal comparative methods such as phylogenetic generalized least squares (PGLS, regression that models covariance among related taxa) or independent contrasts were not applied to strengthen its conservation claim. ₂₁ It proposes formally controlling for this non-independence using PGLS or phylogenetic principal component analysis (PCA). ₂₁

The Metal Resistance Ecology — Phylogenetic Conservation vs. Environmental Selection implements PGLS directly. Metal-type diversity positively predicted niche breadth (B_std) in a simple PGLS (n=606, β=+0.0215, SE=0.0056, p=1.5×10⁻⁴, Benjamini–Hochberg FDR q=0.003) and a multi-predictor PGLS (n=606, β=+0.0231, SE=0.0067, p=5.5×10⁻⁴, q=0.004). ₂₂ The report interprets the PGLS β (+0.021 per SD increase in metal type diversity) as the additional covariation after removing the shared-ancestry component estimated by the λ-transformed variance–covariance matrix. ₂₂

Pagel’s λ (a measure of phylogenetic signal in a trait) for bacterial B_std was 0.787 when estimated standalone and 0.708 within PGLS (Δλ = 0.079), while metal-type diversity itself had λ = 0.335. ₂₂ The report attributes the discrepancy to the predictor’s own phylogenetic signal: including it as a covariate removes some phylogenetically structured variance from the residuals. ₂₂ It reads the drop as a 10% reduction, meaning metal type diversity accounts for approximately 10% of B_std’s phylogenetic covariance, and treats the residual λ = 0.708 (p ≪ 0.001) as strong remaining phylogenetic signal. ₂₂ This refines the BacDive model-comparison result above: there, taxonomy absorbed most of the phenotype signal, whereas here a genome-derived predictor kept a positive coefficient after phylogenetic covariance was modelled, while most phylogenetic signal in niche breadth remained unexplained. _22,7

Archaeal niche breadth also showed significant phylogenetic signal (B_std λ = 0.197, p = 1.1×10⁻⁵; n_envs λ = 0.898, p = 4.6×10⁻¹⁴), with the B_std λ notably lower than the bacterial value of 0.787. ₂₂ At an operational taxonomic unit (OTU, a 16S sequence cluster) relative-abundance threshold of ≥0.01%, the reported Pagel’s λ was 0.750 (n=576, p=2.8×10⁻⁴⁰); standard error and q-value were not reported for that row. ₂₂

Covariate analyses retained the metal-type effect. ₂₂ In robustness analysis R5, adding genus-level mean genome size as a PGLS covariate gave a metal-type effect of β = +0.022 (p = 3.6×10⁻⁴) with ΔAIC = 10.8; the summary table lists R5 as n=527, β=+0.0218, SE=0.0061, p=3.6×10⁻⁴, q=0.003, and R6 (adding log_n_species and log_genome_size) as n=527, β=+0.0218, SE=0.0061, p=3.5×10⁻⁴, q=0.003. ₂₂ The genome-count covariate analyses R1 and S4 used PGLS via gls() + corPagel, not ordinary least squares, and β remained significant after adjustment. ₂₂

The analysis has stated limits. ₂₂ All PGLS and Pagel’s λ analyses used the GTDB r214 genus-representative tree built from 120 concatenated marker genes; sensitivity to NCBI, SILVA-based 16S, or another GTDB release was not tested, although within-genus heterogeneity was tested in sensitivity analysis S3 and was non-significant. ₂₂ Metal-type diversity is an integer count (integers 1–7); the report notes that Gaussian PGLS assumes appropriate residual behaviour of B_std rather than a Gaussian predictor, and that Pagel’s λ on metal types uses a Brownian-motion approximation for an integer trait, with discrete-state (Mk) or threshold models as alternatives. ₂₂ A Poisson or negative-binomial phylogenetic mixed model (MCMCglmm) testing the reverse prediction direction is proposed, not performed. ₂₂

Predictive models show how lineage structure can masquerade as transferable genotype–phenotype signal. In the Genotype × Condition to Phenotype Prediction from ENIGMA Growth Curves, continuous growth rate (mumax), lag time and yield (max_A) had negative R^2 under genus-blocked holdout in both the full KEGG Orthology (KO) model and a dedicated bulk-feature regression. ₂₃ Weak univariate correlations (n_unique_KOs vs mumax: r=+0.42; n_tRNA vs mumax: r=+0.30) were phylogenetically confounded—large-genome genera such as Pseudomonas grow fast and small-genome genera such as Pedobacter grow slowly—and provided zero predictive power under cross-genus holdout. ₂₃ Per-strain fitness enrichment ranged from 1.72x (Cup4G11) to 0.83x (pseudo1_N1B4, no enrichment), and the report cautions that the model’s correct growth predictions (area under the receiver-operating-characteristic curve, AUC, 0.78 for amino acids) come from combinations of prevalence-variable KOs acting as genus-level proxies rather than individually causal genes. ₂₃ This supports the BacDive finding that phenotype associations can act as phylogenetic proxies. _23,7

The Gene Function Ecological Agora shows that the choice of phylogeny-aware metric can change a conclusion. ₂₄ Its consumer z-score compared parent-phylum dispersion with a permutation null for UniRef50s in ≥ 3 clades, and an order-rank anomaly arose under that parent-rank dispersion metric. ₂₄ Diagnostic NB08c, using Sankoff parsimony (a tree-aware method that counts minimum state changes on a phylogeny) on the GTDB-r214 tree, attributed the anomaly to a metric artifact: Sankoff recovered the expected direction (positive horizontal-gene-transfer controls > negative housekeeping controls at p = 2.1×10⁻⁵). ₂₄ At that Phase 1B milestone the report called the substrate-hierarchy claim only a partial confirmation, because the effect size at UniRef50 was small (Cohen’s d = 0.146 on Sankoff); Phase 2 KO aggregation had to amplify it to d ≥ 0.3, and if it did not, an M11 reconciliation-based redesign (gene-tree versus species-tree comparison with AleRax sub-sampling) would be triggered. ₂₄ The later Phase 2 entry M18 amplification gate tested exactly this—whether KO aggregation amplified the Sankoff Cohen’s d from the UniRef50 baseline (d = 0.146) to ≥ 0.3 for at least one positive HGT control class versus negative housekeeping—and, after one round of diagnostic-driven correction, reported a final verdict of PASS, so the M11 redesign was not triggered and the partial confirmation describes the earlier milestone rather than the project’s later standing. ₂₄ Decision M16 made Sankoff parsimony on GTDB-r214 the primary Phase 2 atlas metric and relegated parent-rank dispersion to diagnostics, and Sankoff parsimony with M22 recipient-rank gain attribution used 17M gain events on an 18,989-leaf tree. ₂₄ The four-panel diagnostic figures/p1b_metric_diagnostic_panels.png compares parent-rank dispersion with Sankoff parsimony as discrimination metrics. The earlier single-rank consumer-z null figure (unimodal, with rank-distinct shapes) is kept for provenance but superseded by a per-rank null and the tree-aware Sankoff diagnostic; the source labels that superseding Sankoff diagnostic inconsistently, as M16 in one phrase and M14 in another. ₂₄

Its handling of phylogenetic independent contrasts changed during the project. ₂₄ An earlier plan promoted PIC from an optional Phase 2 sensitivity analysis to a mandatory Phase 1B analysis for producer- and consumer-null score reporting. ₂₄ A later plan argued that classical Felsenstein PIC controls taxon non-independence on per-species traits, whereas Sankoff’s per-leaf parsimony (gain_events / n_present_leaves) already encodes tree topology, so stacking PIC on Sankoff would double-count. ₂₄ It identified the remaining concern as UniRef-cluster non-independence: multiple UniRef50s within a Pfam family inherit related phylogenetic distributions, so Mann–Whitney comparisons across UniRefs treat them as iid and inflate effective N; plan M19 binds a Pfam-family cluster bootstrap with B = 200 resamples per function class, reporting 95% CIs for per-leaf Sankoff Cohen’s d. ₂₄ The final open items nonetheless list phylogenetic non-independence with PIC correction not implemented, alongside sample-size effect inflation (especially for photosystem II, PSII, with n=21), unquantified cross-phase error propagation and mixed two-component-system (TCS) architectural concordance. ₂₄ AMR was excluded from the positive-control panel because the AMR substrate is Pseudomonadota-biased (49% of AMR UniRefs >50% Pseudomonadota); its informational profile was 54.8% recent and 4.1% ancient. ₂₄ This lineage bias supports the taxonomic concentration of AMR noted earlier on this page. _24,1

Its confirmatory results were also tested against residualization. Both confirmed hypotheses survived D2 residualization and were grounded in expected biomes at p < 10⁻¹¹. ₂₄ In a six-test residualization replication family (Bonferroni α=0.00833), five tests survived: Mycobacteriaceae × mycolic family producer_z (d=+0.31, p=2×10⁻⁶) and consumer_z (d=−0.16, p=1.6×10⁻⁵), Mycobacteriales × mycolic order producer_z (d=+0.29, p=5×10⁻⁶), and Cyanobacteriia × PSII class-rank producer_z (d=+1.50, p=2.1×10⁻⁵) and consumer_z (d=+0.63, p=2.3×10⁻⁵); the NB11 regulatory-versus-metabolic producer_z test was null (d=+0.06, p=0.69), consistent with its falsified original hypothesis. ₂₄ The report flags possible sample-size effect inflation for the PSII result, which rests on n=21. ₂₄

Plant-Association, Prophage and Lifestyle Controls

The PGP Gene Distribution Across Environments & Pangenomes also shows phylogenetic control changing the significance, though not the direction, of a raw comparison. The raw ipdC soil comparison was non-significant (OR = 0.79), whereas a phylum-controlled logit found significant soil depletion (OR = 0.56, p = 0.027). ₁₈ The positive trp→ipdC coupling existed only in non-soil species (OR = 3.56) and reversed in soil species (OR = 0.30). ₁₈

This lineage-dependent result rests on sparse data. A phylum + soil logistic model (Model 3) was attempted but failed because of quasi-complete separation caused by ipdC’s rarity across phyla (214 species, 1.9%), so the report’s conclusions rest on Model 1 (trp only) and Model 2 (trp + is_soil). ₁₈ Only 214/11,272 species (1.9%) carry ipdC, which limits statistical power for the stratified analysis; the report treats the soil reversal (OR = 0.30, p = 0.02) as hypothesis-generating rather than conclusive. ₁₈

The Plant Microbiome Ecotypes — Compartment-Specific Functional Guilds and Their Genetic Architecture shows how raw gene-content enrichment can mostly measure taxonomy. Fisher’s exact test with Benjamini–Hochberg false-discovery-rate (BH-FDR) correction found 94.2% of tested eggNOG ortholog groups (OGs) (5,341/5,671) significantly associated with plant status, of which 3,840 showed strong enrichment (OR>2). ₂₅ Top raw hits included COG3569 (OR=8.92), COG1764 (OR=9.29), COG5343 (OR=7.74), COG0654 (OR=12.66) and COG1845 (OR=14.47). ₂₅ The report reads this near-saturation as genome-wide compositional differences between plant-associated and non-plant bacteria (in effect, taxonomic divergence) rather than plant-specific functional enrichment. It treats only the OGs surviving phylogenetic control as candidate plant-interaction genes, not as independently confirmed adjusted fits. ₂₅

After phylum-level logistic regression with 28 large phyla as fixed effects, 50 novel OGs retained significance (all p<0.05, OR>1), and all 50 top-enriched OGs survived. ₂₅ The top hit was COG3569 (Fisher OR=8.92, phylo-controlled OR=6.01, q=7.7e-242), found in 54% of plant-associated species but only 12% of non-plant species. ₂₅ The report reads the attenuation from 8.92 to 6.01 as partial phylogenetic confounding, because plant-associated species cluster taxonomically, with a substantial ecological signal persisting after correction. ₂₅ However, all these models showed convergence warnings, likely from quasi-separation when some phyla are entirely plant-associated or entirely non-plant, so the reported p-values and odds ratios are approximate. ₂₅ This qualifies the marker-gene paragraph above, where the digest’s COG3569 OR=6.01 is reported without that convergence caveat. _25,4

Finer-grained control proved intractable in the same analysis. Family-level logistic regression found insufficient within-family variation for all 10 OGs tested, indicating that plant association is taxonomically clustered at the family level, and the planned genus-level fixed effects were not executed because of computational constraints. ₂₅ The report’s practical interpretation is that plant-associated functional enrichment is real but partially confounded by phylogeny: it is difficult to disentangle whether the genes cause plant association or whether plant-associated clades simply retain them. ₂₅

A later per-species analysis regressed each of the 50 novel OGs on is_plant + log₁₀(genome_size) + top-10 phylum dummies across 22,384 species with complete covariates (1,134 plant-associated, 21,250 non-plant). ₂₅ Fits converged for 48/50 OGs, and 48 of 50 OGs (96%) retained a positive plant-association coefficient that was large (|coef| > 0.2) and significant at BH-FDR q < 0.05. Controlled odds ratios for the top 10 ranged from 4.4 (COG5516) to 6.6 (COG1845, cytochrome oxidase ctaE), with q-values as small as 10⁻⁹⁶ (COG3569). ₂₅ The genome-size coefficient was strongly positive (median 3.4 on log₁₀-scale), but it did not explain the plant effect away. ₂₅ For COG2351 and COG5343 the design matrix was singular. The report attributes this to likely perfect separation by a phylum dummy, given raw prevalence gaps of 62% vs 23% and 64% vs 21%, and calls these failures numerical rather than biological (strong raw signals, 3× higher in plant vs non-plant). Their controlled effects were not estimated, however, so they are not counted here as successful adjusted fits. ₂₅

Genome size (gene cluster count) correlated moderately with total marker count (r=0.44, p<1e-300). ₂₅ An earlier recap in the report says all 50 novel OGs retain positive coefficients with genome size as a covariate. The canonical per-species result supersedes that statement: 48/50 OGs at q<0.05 with top-10 controlled OR = 4.4–6.6, the other two being singular-matrix failures. ₂₅ The dual-nature rate was genome-size-dependent: plant species in the bottom genome-size quartile were 54% dual-nature versus 87% in the top three quartiles (the report’s table abbreviates that group as ‘top’). This indicates that the 78.7% figure masks a real gradient relevant to dual-nature-plant-associated-bacteria. ₂₅

Marker-level phylogenetic control went through several stages. The research plan specified genus-level fixed effects, but Phase 1 execution used phylum-level control because of computational cost. The planned genus-level logistic regression for the H1 compartment enrichments was not completed because of a code error (undefined logit import), so those enrichments are not phylogenetically controlled beyond the PERMANOVA test itself. ₂₅ NB10 then attempted genus-level fixed-effects logistic regression for all 14 refined markers. One passage says all models failed to converge (0/14), and the output table records ‘Phylogenetic control logistic regression — 0/14 converged’. ₂₅ Another passage says the logit ‘failed to converge for 0/14 markers’, which, read literally, conflicts with total non-convergence; that sentence alone does not establish how many markers, if any, failed. The report attributes the failure to plant association being heavily confounded with genus-level taxonomy. ₂₅ The report treats the subsequent L1-bootstrap control as suggestive rather than confirmatory and closed the issue with a later genus-cluster-robust GLM, described below. ₂₅

After the original genus-level logit failed (NB10’s genus-fixed-effect models for the 14 markers did not converge), the report fitted an L1-regularized logit (scikit-learn saga, C=1.0) with top-20 genus dummies, log₁₀(genome_size) and a 100-replicate bootstrap. This model produced convergent fits, and 9/14 markers had bootstrap 95% CIs excluding zero. ₂₅ Two recap passages give this as 9/17 markers; the 9/14 in the report’s canonical section supersedes that denominator. ₂₅ The report cautions that L1 bootstrap CIs are biased near zero (Chatterjee & Lahiri 2011) and that n_bootstrap=100 is small, and its paired adversarial review flagged this version as suggestive-not-confirmatory. The later genus-cluster-robust result is 8/14 significant markers, and a phylogenetic generalized linear mixed model (PGLMM) with the full GTDB tree remains future work. ₂₅

The report closed this item (C1) with an unpenalized binomial GLM using top-10 phylum fixed-effect dummies, log₁₀(genome_size) and cluster-robust covariance with 7,555 genus clusters across 24,554 species. This gave proper Wald p-values and 95% Wald CIs without L1 shrinkage, and the report calls it the pragmatic frequentist analogue of a PGLMM at species scale. ₂₅ 8/14 markers survived at BH-FDR q < 0.05 with |coef| > 0.2, a stricter result than the suggestive L1 bootstrap, and 13 of 14 markers agreed with the L1 verdict. The one disagreement was iaa_biosynthesis (L1 coef = 0.61 sig; cluster-robust coef = −0.09, p = 0.82, not sig), which the report treats as an L1 false positive. ₂₅ The report regards a full PGLMM with GTDB-tree-derived random-effect covariance as theoretically stronger, but the 25K x 25K dense covariance matrix is intractable, so the genus-level cluster-robust GLM is its practical analogue. ₂₅ The report complemented this with within-genus permutation, under which 3/15 markers were supported at species level. ₂₅

• Cluster-robust GLM results per marker (coefficient; OR; 95% CI; BH-FDR q):
• Significant: acc_deaminase 2.08; 8.01; [4.19, 15.32]; 2.1×10⁻⁹. t3ss 1.00; 2.71; [1.94, 3.80]; 2.6×10⁻⁸. nitrogen_fixation 1.00; 2.71; [1.54, 4.78]; 1.3×10⁻³. phenazine 0.75; 2.12; [1.69, 2.64]; 5.9×10⁻¹⁰. cwde_cellulase 0.63; 1.87; [1.46, 2.40]; 2.5×10⁻⁶. phosphate_solubilization 0.52; 1.68; [1.20, 2.35]; 4.8×10⁻³. cwde_pectinase 0.46; 1.59; [1.25, 2.01]; 3.4×10⁻⁴. effector 0.30; 1.35; [1.09, 1.67]; 1.2×10⁻².
• Not significant: dapg_biocontrol −0.02; 0.98; [0.41, 2.33]; 0.96. t4ss −0.02; 0.98; [0.72, 1.34]; 0.96. hydrogen_cyanide −0.04; 0.96; [0.52, 1.76]; 0.96. iaa_biosynthesis −0.09; 0.91; [0.40, 2.06]; 0.96 (L1 false positive). siderophore −0.33; 0.72; [0.44, 1.17]; 0.26. acetoin_butanediol −0.38; 0.69; [0.46, 1.03]; 0.11. ₂₅

Within-genus label shuffling, the strictest species-level null, retained only 3 of 15 testable markers: nitrogen fixation (perm p < 0.001), ACC deaminase (perm p < 0.001) and T3SS (perm p = 0.005). ₂₅ The other 12 markers (phenazine, IAA biosynthesis, CWDE cellulase, CWDE pectinase, effector, DAPG biocontrol, hydrogen cyanide, siderophore, acetoin/butanediol, phosphate solubilization, T4SS and phytotoxin) failed the test. ₂₅ The test used n = 200 permutations, which is under-resolved for p ≤ 0.005 (so T3SS is marginal), and no multiple-testing correction was applied across 15 markers. Bonferroni α = 0.0033 would demote T3SS, while nitrogen fixation and ACC deaminase would remain significant. ₂₅

After the original genus-fixed-effects logit failed to converge, combining the cluster-robust GLM (8/14 significant, with iaa_biosynthesis identified as an L1 false positive) and within-genus shuffling (3/15) gave the report’s three-tier verdict. ₂₅ Tier 1, robust at species level under any test (3 markers): nitrogen fixation, ACC deaminase, T3SS. Tier 2, robust under the cluster-robust GLM but failing the within-genus shuffle (5 markers): phenazine, CWDE cellulase, CWDE pectinase, phosphate solubilization, effector. These are cassette-level signals that a plant-associated genus tends to acquire but that do not distinguish species within the genus. Tier 3, not significant in any test (6 markers): DAPG biocontrol, T4SS, hydrogen cyanide, IAA biosynthesis, siderophore, acetoin/butanediol. ₂₅ This refines the acdS evidence above. ACC deaminase is one of only three markers that survive within-genus shuffling, but its cluster-robust OR of 8.01 and the phylum-fixed-effect OR of 6.98 are far below the root-species Fisher OR of 69.3 that the digest repeats without phylogenetic control. _25,18,4

The genus-cluster-robust GLM is a practical substitute for, not an equivalent of, a full PGLMM on the GTDB tree, whose random-effect covariance would come from tree branch lengths rather than genus clustering. ₂₅ The report regards the full model as the gold standard but intractable, because its 25K × 25K dense covariance matrix cannot be held in memory, and proposes sparse approximations or species subsampling as the practical path. ₂₅

The same report shows a multivariate habitat effect shrinking under a taxonomic sensitivity analysis. The original NB04 PERMANOVA on the Phase 1 25-marker panel reported R² = 0.527 (pseudo-F = 235.1, p = 0.001) among 1,136 plant-associated species classified across four compartments (root 292, rhizosphere 160, phyllosphere 157, endophyte 29). The stated compartment counts do not sum to 1,136, and the source does not reconcile them. ₂₅ Excluding the top-3 species per compartment reduced R² from 0.527 to 0.072 (86% loss; still significant, p = 0.001). Most of the original 0.527 was therefore a taxonomic-sampling artifact of a few genome-rich rhizobial / Pseudomonas clades rather than a community-wide signal. The report’s second safeguard, within-genus label shuffling, retained only 3/15 markers (nitrogen fixation, ACC deaminase, T3SS). ₂₅ The per-marker Fisher patterns persisted: root species had high ACC deaminase (OR = 69.3), T3SS (OR = 65.6), nitrogen fixation (OR = 14.5) and quorum sensing (OR = 24.1) enrichment, and 69 of 96 marker × compartment pairs were significant. Only the headline magnitude of the multivariate R² was inflated. ₂₅

In Phase 2b, the report ran distance-based redundancy analysis (db-RDA, a constrained ordination of a distance matrix) on principal-coordinate scores of Jaccard distance over a refined 17-marker panel (n=607 plant-compartment species). It gave location-only R² = 0.060 (permutation p = 0.001, 999 permutations). ₂₅ PERMDISP confirmed group-dispersion heterogeneity (F = 15.6, p = 0.001), with root tightest (mean centroid distance 0.452), phyllosphere intermediate (0.503) and rhizosphere most variable (0.528). About 84% of total PERMANOVA R² was real centroid shift and 16% dispersion, so compartments are functionally distinct at a small effect size (~6% of variance). ₂₅ A phylogeny-versus-ecology variance decomposition remains future work, and PERMANOVA rather than logit was the primary compartment test. ₂₅

Categorical phylum control is not the same as finer phylogenetic control. In the Pan-Bacterial Anti-Phage Defense Arsenal, partial correlation and a negative-binomial GLM controlled for phylum categorically but not for finer phylogenetic structure. The report states that its results are consistent with, but not a formal test of, a phylogenetically independent arms race, and names a phylogenetic mixed-effects model with the GTDB tree as covariance structure as a stronger test. ₂₆ The Polyhydroxybutyrate Granule Formation Pathways — Distribution Across Clades and Environmental Selection similarly leaves shared ancestry as a potential confound and proposes phylogenetically controlled analyses such as phylogenetic logistic regression to strengthen its findings. ₂₇

The Prophage Gene Modules and Terminase-Defined Lineages Across Bacterial Phylogeny and Environmental Gradients gives a case where phylogeny-preserving nulls removed lineage-level environmental structure. PERMANOVA on Bray–Curtis distances of prophage module composition (n=1,773 species subsample, 99 permutations) found all three predictors significant (p=0.01), with F-statistics of 212.99 for genome-size quartile, 30.04 for environment and 6.17 for phylogeny (family). ₂₈ Constrained permutation null models (500 permutations preserving host family × genome size quartile strata across 18,031 species) identified 8 significant module × environment enrichments (FDR < 0.05) but no significant lineage-level enrichment. The report concludes that the lineage distribution is well explained by host phylogeny alone, in contrast with Paez-Espino et al.’s (2016) finding of strong habitat-type specificity for most viral groups. ₂₈ Four modules (A_packaging, D_lysis, E_integration, F_lysogenic_regulation) were NMDC-significant but not pangenome-enriched beyond phylogenetic expectation, which the authors interpret as near-universal presence masking environment-specific variation. ₂₈ Each predictor was tested independently; no sequential or partial PERMANOVA has yet quantified environment’s unique variance after simultaneous control for phylogeny and genome size. ₂₈

In the Carbon Source Utilization Predicts Ecology and Lifestyle in Pseudomonas, the primary principal component analysis (PCA) axis of pathway profiles separated Pseudomonas s.s. from Pseudomonas_E, driven by sugar-pathway loss, while lifestyle categories (free-living, host-associated, plant-associated) overlapped substantially within Pseudomonas_E. ₂₉ The analyses did not explicitly control for phylogenetic non-independence among species, so whether lifestyle–pathway associations survive phylogenetic control is unresolved. The report proposes PGLS or phylogenetic logistic regression with the GTDB species tree. ₂₉

Tensions

The aggregate result and the within-phylum results point in different directions: the overall openness–AMR correlation was near zero (rho=0.006), whereas openness positively correlated with AMR count in 8/10 tested phyla, including Bacillota and Bacillota_C. ₁ This is a tension between a pooled analysis that suggests little association and stratified analyses that show positive associations in multiple lineages; the report attributes the discrepancy to phylogenetic dominance of the aggregate signal rather than resolving it as a causal effect. ₁

The prophage analysis introduces a related but non-identical tension: prophage marker density remained associated with AMR repertoire breadth across all five reported major phyla and after controlling for genome count, whereas the openness analysis found that openness did not predict the magnitude of environment or phylogeny effects. _5,2 These results concern different predictors and response variables, so they do not establish whether prophage density is independent of shared ancestry or whether the association reflects open-pangenome lineages acquiring multiple mobile elements. ₅

The openness-effects analysis introduces a related scope tension rather than directly contradicting the AMR result: openness did not predict either environment or phylogeny effect sizes in the matched species set, despite the AMR analysis showing lineage-dependent openness associations. _2,1 The two results concern different response variables, so they do not establish whether the within-phylum AMR patterns persist after explicit phylogenetic correction. _2,1

The environmental-resistome analysis qualifies rather than contradicts the earlier environmental comparison: significant effects within five of six phyla and 20 of 141 families support an ecological association, but the majority of families were not testable or significant because of limited environmental breadth. ₆ Its species-level classifications, sampling imbalance, and AMRFinderPlus-focused annotation also leave unresolved how much of the clinical contrast reflects lineage, representation, or unrecognized environmental resistance. ₆ The within-species analysis adds a similar tension: visible environmental structuring in case-study AMR ecotypes was not matched by a general statistical demonstration because metadata and within-species environmental diversity were insufficient. ₃

A within-project tension concerns phylogenetic independent contrasts in Gene Function Ecological Agora: one plan made PIC mandatory at Phase 1B, a later plan argued that PIC stacked on Sankoff parsimony double-counts tree structure and substituted a Pfam-family cluster bootstrap, yet the final limitations still list the missing PIC correction as a real methodological limit. ₂₄ The available evidence does not show whether the cluster bootstrap fully addresses the non-independence that PIC was meant to control. ₂₄

Formal corrections did not move results in a single direction, and the projects used different controls (genus grouping, phylum fixed effects, family equal weighting, mixed models, PGLS, Sankoff parsimony), so their outcomes are not directly comparable. _{17,18,19,22,24} In the lanthanide atlas a phylum-and-family mixed model gave a stronger xoxF dominance estimate than naive pooling, whereas within-phylum control removed the Wood–Ljungdahl and [NiFe]-hydrogenase enrichments in the clay study, and PGLS retained a positive but small metal-type coefficient for niche breadth. _19,9,22

Within Plant Microbiome Ecotypes — Compartment-Specific Functional Guilds and Their Genetic Architecture, the number of plant signals that survive depends on the stringency of control. 50 OGs survived phylum fixed effects with convergence warnings, 48/50 survived per-species phylum plus genome-size regression, 9/14 markers passed the L1 bootstrap, 8/14 passed the genus-cluster-robust GLM and 3/15 survived within-genus shuffling. ₂₅ The report also states some counts inconsistently: the NB10 ‘0/14’ wording, a 9/17 recap denominator and an all-50 genome-size recap. This page follows the report’s later canonical analyses and does not treat the recaps as independent confirmation. ₂₅

Open Directions

• Use the existing species-level openness, AMR-count, and taxonomic data with phylogenetic generalized least squares (PGLS), a regression method that models covariance among related species, to test whether the openness–AMR association remains after accounting for shared ancestry. ₁
• Combine the pangenome species table with a species phylogeny and fit within-phylum hierarchical models to ask whether the positive associations in 8/10 phyla share a common slope or instead arise from lineage-specific processes. ₁
• Reanalyze the six-category environmental comparison using taxonomic matching or lineage-stratified permutation tests to determine whether the clinical, soil, plant, aquatic, and animal contrasts persist after controlling for uneven taxonomic sampling. ₁
• Extend this matching to the 14,723-species environmental-resistome dataset and test whether the five significant within-phylum effects and 20 significant within-family effects remain under balanced sampling and explicit phylogenetic models. ₆
• Add environmental metadata and AlphaEarth embeddings for the currently under-covered genomes, then use phylogeny-aware partial association analyses to test whether environmental diversity predicts AMR count independently of lineage. ₁
• Estimate AMR gene gain and loss rates on a dated or calibrated species phylogeny, and test whether pangenome openness predicts transition rates rather than only present-day AMR counts. ₁
• Stratify the openness-versus-environment and openness-versus-phylogeny analyses by gene function, auxiliary fraction, Heap’s law alpha, and pangenome fluidity to test whether the null result reflects an overly coarse openness metric. ₂
• Recompute the prophage-density–AMR-breadth association using dedicated prophage predictors, base-pair distances, all available genomes, and explicit phylogenetic models to distinguish shared ancestry, broad mobile-element propensity, and phage-mediated transfer. ₅
• Within the 1,305-species AMR dataset, subsample species with more than 500 genomes and fit phylogeny-aware models to test whether within-species AMR Mantel signals and non-core/core contrasts persist without computational-cap and prevalence-variance artifacts. ₃
• Curate collection dates and isolation sources, then repeat AMR ecotype–environment tests with balanced within-species sampling to distinguish ecological structure from lineage and metadata coverage. ₃
• The report proposes applying phylogenetic generalized least squares (PGLS) or phylogenetic principal component analysis (PCA) to the species-level BacDive phenotype and metal-tolerance-score table. Formally removing phylogenetic signal before testing would show whether any phenotype–metal association survives. ₇
• Profile per-metal fitness for urease-positive versus urease-negative bacteria within one taxonomic class, ideally Gammaproteobacteria, to test whether urease-positive organisms are specifically nickel-tolerant rather than generally metal-tolerant. ₇
• Use organisms with many AMR genes (Cup4G11: 77, BFirm: 50) to test whether resistance mechanism predicts fitness cost within a single genetic background, which controls for between-organism phylogenetic differences. This comparison is proposed in the report, not completed. ₃₀
• Run per-genus enrichment on the genus-clumped Bacillota_B anchor cohort to separate genus-specific lineage-marker OGs from OGs that recur across multiple subsurface-specialist genera. ₁₀
• Add genomes from the far stratum (>0.05 branch distance) and generate new co-fitness measurements for the two excluded Ralstonia organisms, which currently have zero co-fitness data in the Fitness Browser. This would test whether cofit-pair phi exceeds the background expected from shared ancestry. ₁₁
• Fit a fully phylogeny-aware mixed model, such as the phylum-and-family binomial mixed model already used for H1, to the lanthanide H2 soil/sediment enrichment, which currently relies on within-phylum stratification. ₁₉
• Repeat the niche-breadth PGLS on NCBI, SILVA-based 16S and alternative GTDB-release trees, and fit the proposed Poisson or negative-binomial MCMCglmm in the reverse direction, to test tree and error-model sensitivity of the metal-type coefficient. ₂₂
• Apply PGLS or phylogenetic PCA to the 28-organism cross-resistance conservation table to test whether conservation survives once phylogenetic non-independence is modelled. ₂₁
• Join GTDB taxonomy to the pathway–openness species table for full phylum-level stratification and compute independent contrasts, to test whether the positive direction in 13 of 18 genera reflects a general evolutionary dynamic. ₁₇
• Report the M19 Pfam-family cluster-bootstrap confidence intervals for per-leaf Sankoff Cohen’s d alongside a PIC-based sensitivity analysis, to test whether the bootstrap captures the non-independence that the unimplemented PIC correction was meant to control and whether conclusions change. ₂₄
• Build within-species core-genome trees for ecotype clusters, including Salmonella and Phenylobacterium, and test environment–cluster associations conditional on that phylogeny to separate adaptive ecotypes from demographic subpopulations. _15,16
• Within human-associated species such as Klebsiella and Enterococcus, condition environment–gene-content tests on lineage and region to test whether embedding associations are epidemiological rather than ecological. ₁₄
• Fit a sparse-approximation or species-subsampled PGLMM on the GTDB tree for the 14 plant markers and the 50 novel OGs. This would test whether the 3 species-level and 5 cassette-level markers keep their tiers under branch-length covariance rather than genus clustering. ₂₅
• Repeat within-genus label shuffling with more than 200 permutations and correction across the 15 markers, to settle whether T3SS is species-level robust or marginal. ₂₅
• Run the planned phylogeny-versus-ecology variance decomposition on the 17-marker compartment profiles to separate the ~6% location effect from lineage composition. ₂₅
• Run a partial PERMANOVA on prophage module composition with phylogeny, genome size and environment fitted together, to estimate environment’s unique variance. ₂₈
• Fit a phylogenetic mixed-effects model with GTDB-tree covariance to the defense-arsenal data, to formally test a phylogenetically independent arms race beyond categorical phylum control. ₂₆
• Apply phylogenetic logistic regression to the PHB granule associations to test whether they persist after accounting for shared ancestry. ₂₇
• Apply PGLS or phylogenetic logistic regression with the GTDB species tree to Pseudomonas_E lifestyle–pathway associations, to test whether lifestyle signal survives the subgenus divide. ₂₉
• Test the ipdC soil reversal with more ipdC carriers or with a model that tolerates quasi-complete separation, so that the failed phylum + soil model (Model 3) can be estimated. ₁₈