Negative Results and Caveats

What each project reports as limitations, null results, or abandoned analyses — read before repeating an analysis.

Acinetobacter baylyi ADP1 Data Explorer

No gene has measurements across all 6 data modalities, and FBA flux data cover only 15% of genes; consequently, the FBA-TnSeq concordance analysis is limited to 866 genes. ₁

The pangenome cluster mapping is indirect and passes through 3 tables. Although it is 100% complete at the reported gene level, the indirection could introduce edge cases where clusters split or merge between the KBase Data Lakehouse and ADP1 pangenome analyses. ₁

The 87% dependence of growth phenotype predictions on gapfilled reactions limits interpretation of those predictions, and the 243 missing functions represent gaps in genomic evidence that affect prediction reliability. ₁

The database covers only A. baylyi, so cross-species comparisons require comparable databases for other organisms. ₁

The observed FBA-TnSeq discordance, urea-specific fitness pattern, core-metabolism conservation, and gapfilling dependence are primarily results from this ADP1-centered dataset. The report presents testing whether the 227 FBA-TnSeq-discordant genes are enriched for specific pathways or regulatory functions, to guide metabolic-model refinement, cross-species fitness comparison, PhageFoundry cross-referencing, urea-specific gene identification, and pangenome-informed gapfill confidence assessment as future analyses rather than completed findings. ₁

ADP1 Deletion Collection Phenotype Analysis

• Growth ratios are single-timepoint measurements with unknown technical noise, so condition-specificity scores may reflect measurement error as well as biology. ₂
• The complete matrix contains 2,034 genes and excludes 499 essential genes plus 316 genes with incomplete data, biasing the analysis toward dispensable genes with successful deletion mutants. ₂
• Only 8 carbon sources were tested, so the approximately 5 independent dimensions may increase when additional conditions are measured. ₂
• Pangenome core/accessory status comes from the KBase Data Lakehouse’s species-level pangenome for A. baylyi and may have limited resolution compared with a population-level analysis. ₂
• The low clustering silhouette score and absent FDR-significant enrichments support a gradient interpretation, but the report’s proposed independent component analysis (ICA) and expanded condition panel remain future work rather than completed analyses. ₂

Triple Essentiality Concordance Analysis for *Acinetobacter baylyi* ADP1

The original analysis is restricted by design to 478 TnSeq-dispensable genes because TnSeq-essential genes do not provide viable deletion mutants for the growth-rate analysis. Therefore, its null result concerns growth variation among dispensable genes and cannot be interpreted as a whole-genome test of FBA lethality prediction. The original analysis also had no direct lethality measurements and used TnSeq as a proxy. The report states that, among dispensable genes, growth rates vary from 10% to 100% of wild-type. ₃

The Q25 threshold flags the bottom 25% of growth rates separately in each condition, so the reported “any defect” rate across eight conditions is an aggregation-dependent quantity. The report notes an expected rate of 1 − 0.75⁸ = 90% under independent conditions, compared with the observed 72% of 478 genes with a defect on at least one of the 8 conditions; the report attributes this difference to positive inter-condition correlation. ₃

The refined knockout comparison merges minimal-media and rich-media calls by prioritizing minimal-media data and using rich-media data as fallback. The knockout labels come from the de Berardinis et al. (2008) complete collection of single-gene deletion mutants of Acinetobacter baylyi ADP1 (Molecular Systems Biology 4:174; adp1-single-gene-deletion-collection). The report treats this collection as biological ground truth for essentiality in both studies. It supplied mutant growth rates on 8 carbon sources for the original study and essentiality calls for rich (LB) and minimal media for the refined study. Rich-media (LB) knockout data covered 2,953 genes, including 346 essential and 2,607 dispensable genes; the merged minimal-media set covered 3,092 genes, including 499 essential genes. ₃

Proteomics was not part of the original study and was added in the refined analysis. Its input was average log2 expression across 7 Acinetobacter strains for 2,383 genes, with locus tags parsed from complex-format feature IDs. It is not an ADP1-only measurement and was not measured only in the exact knockout assay condition, so its association with essentiality is independent supporting evidence rather than a condition-matched causal test. ₃

The five RB-TnSeq thresholds were tested without formal multiple-testing correction. The report states that because every threshold produced negative Cohen’s kappa, correction would not change the conclusion of systematic disagreement across the tested thresholds. The report is internally inconsistent about the optimal threshold. Its connection-to-original-study passage notes that the original study analyzed only TnSeq-dispensable genes (n=478). It then says only 7.9% (18/229) of KO-essential genes are caught by TnSeq “at the optimal threshold”. However, 7.9% is the recall reported at the 0.05 threshold. The report’s threshold table instead labels 0.025 as the optimal-F1 threshold, with recall of 15.3% there. This summary records both statements and does not resolve the discrepancy. ₃

The proposed explanations for TnSeq/knockout discordance—including partial or truncated protein production, read-through transcription, retained functional domains, and condition aggregation—are mechanistic hypotheses rather than directly resolved demonstrations in this report. The report proposes that the 211 KO-essential/TnSeq-dispensable genes (mean essentiality_fraction = 0.0034) likely tolerate partial function from transposon-disrupted genes. It proposes that the 293 TnSeq-essential/KO-dispensable genes (mean essentiality_fraction = 0.10) may be fitness-important but not lethal, because transposon insertion is not equivalent to gene deletion. On condition aggregation, the Fitness Browser TnSeq data span diverse conditions across many studies, and essentiality_fraction mixes those conditions. The report states that condition-matched comparisons are more informative. Similarly, the proposed explanation that trace aromatics in experimental media account for aromatic-pathway discordance requires testing with measured media composition and revised FBA constraints. ₃

The report recommends condition-matched TnSeq and knockout experiments (running TnSeq in identical media to separate condition effects from method effects), domain- and insertion-position analysis of the 211 knockout-essential/TnSeq-dispensable genes, addition of trace aromatic compounds to the minimal-media FBA definition, which would test whether aromatic-degradation gene discordance resolves and match in silico environmental assumptions to experimental reality. It also recommends using continuous flux values instead of the binary essentiality class, through linear or rank regression between flux and growth rate, which may capture quantitative growth effects. Further recommendations are condition-specific flux simulations, and combined FBA-plus-fitness-plus-proteomics prediction. For dispensable genes, the report advises growth assays on specific carbon sources, because FBA class shows no association with growth-defect rate (p=0.63). ₃

AlphaFold MSA Depth as a Lens on the Bacterial Annotation Gap

• The coverage figures are internally inconsistent. The limitations section labels bridge coverage as 29.3%, but the dataset-coverage section gives 29.3% as the share with a real UniProt accession and 28.7% as the share that successfully bridged to AlphaFold MSA depths. The analysis included only clusters with a non-UPI UniProt accession. The dataset-coverage section gives 70.7% as excluded, whereas the limitations section says 71% lack AlphaFold entries. A larger annotation gap in these excluded, poorly characterised organisms remains a hypothesis and was not directly measured. ₄
• MSA depth was looked up for each gene cluster’s representative sequence, so within-cluster sequence diversity was ignored and the representative may have higher or lower MSA depth than typical cluster members. ₄
• The 293K genomes were not phylogenetically balanced; common taxa such as Pseudomonas and E. coli were over-represented, influencing core-gene counts and MSA-depth distributions. ₄
• Spearman ρ = 0.7563 was computed on the full 38,051,842-pair dataset without subgroup stratification, so its value may differ among core, auxiliary, and singleton clusters and among organisms with different annotation gaps. ₄
• The analysis used a static version-6 KBase Data Lakehouse AlphaFold snapshot, and later UniProt deposits may change MSA depths. ₄

AMR Co-Fitness Support Networks

• The flagellar and biosynthesis enrichment may reflect shared “useless under laboratory conditions” status rather than mechanistic co-regulation. A permutation matched on mean fitness level, rather than only conservation class, is required to distinguish these explanations. ₅
• Cofitness is not equivalent to co-regulation: high cofitness indicates shared fitness phenotypes, not direct transcriptional control. Missing fitness values were treated as zero in z-score space through np.nan_to_num, which approximates but does not equal pairwise-complete Pearson correlation; the report considers the dense Fitness Browser matrices unlikely to substantially alter conclusions. ₅
• GO-term granularity may obscure specific signals because broad categories such as transmembrane transport and membrane functions appear in nearly all support networks and genomes. ₅
• At |r| > 0.3, the mean support network contains 233 genes and therefore includes many weak associations; confirmation at |r| > 0.4 is needed. ₅
• The null result for the relationship between network size and fitness cost may reflect insufficient variance in fitness cost across genes. ₅
• The 28 organisms are lab-adapted, phylogenetically biased, include many Pseudomonas organisms, and have limited ecological diversity. ₅
• The operon-exclusion heuristic uses a 5-ORF exclusion zone based on matrix row index position as a proxy for genomic proximity. Row position may not reliably reflect chromosomal gene order. Genomic begin, end, and strand columns were loaded in the annotation data and could support a proper coordinate-based exclusion in future work. Only 0.6% of pairs were excluded by the current heuristic. ₅
• The most important follow-up analyses are fitness-matched permutations; cofitness computed separately for antibiotic and standard-growth conditions; direct assessment of mean fitness for flagellar knockouts; Pfam-domain enrichment; and testing other conditionally dispensable gene classes such as phage-defense and secondary-metabolite genes. ₅

Environmental Resistome at Pangenome Scale

Environment and phylogeny remain deeply entangled despite family-level controls. Some families are nearly entirely clinical, only 14% of tested families showed significant within-family effects, and most families do not span enough environments to be tested. NCBI sampling strongly overrepresents clinical isolates, while soil and aquatic species are undersampled, potentially inflating the clinical-versus-environmental contrast. Majority-vote species classifications collapse within-species variation, although sensitivity analyses at 60–90% thresholds partly mitigate this issue. ₆

AMRFinderPlus-focused annotation may underestimate novel environmental resistance and bias results toward clinical enrichment, where resistance genes are better characterized. The 95% prevalence threshold for core AMR is dependent on species genome count, so core/accessory labels are imprecise for sparsely sampled species. The association is correlational: clinical exposure could increase AMR, AMR could contribute to clinical isolation, or both could reflect virulence-resistance co-selection. ₆

Effect sizes were modest despite very small p-values: environment explained 2–13% of variance in AMR composition (η² = 0.02–0.13), while phylogeny likely explains more. The 18.7% of clusters (15,550) lacking mechanism assignments were excluded from the H3 mechanism fractions, which may bias mechanism comparisons if unclassified clusters are distributed non-randomly across environments. ₆

Planned but unperformed analyses included PCoA ordination, PERMANOVA, environment-specific gene identification, metagenome-assembled genome/isolate assessment, and an archaea-specific analysis. The originally planned per-genome Fisher’s exact test was replaced by the species-level within-species proxy because billion-row joins were computationally costly. ₆

Fitness Cost of Antimicrobial Resistance Genes

• All 25 tested organisms are lab-adapted strains. Laboratory compensation may have reduced measurable costs relative to wild strains, so the estimate may underestimate costs in natural populations. ₇
• Tier 2 contains 86% of the AMR genes and is based on keyword matching from Bakta annotations, which may include non-AMR genes such as general efflux transporters; the Tier 1 sensitivity analysis gives consistent results. ₇
• Matched antibiotic validation covers only four AMR classes: beta-lactam, aminoglycoside, chloramphenicol, and tetracycline. Macrolide, glycopeptide, and polymyxin resistance could not be validated. ₇
• Approximately 4.6% of AMR genes are putatively essential and absent from fitness matrices. If these are the most costly AMR genes, +0.086 is a lower bound. ₇
• RB-TnSeq measures fitness relative to the pool average. The +0.086 value is the difference between AMR knockout fitness (−0.024) and non-AMR knockout fitness (approximately −0.11), not an absolute selection coefficient; comparison with isogenic-strain literature is not a direct equivalence. ₇
• Transposon insertions can have polar effects on downstream genes in operons, potentially confounding AMR-gene fitness measurements. ₇
• The product classifier does not handle fosfomycin or tellurite resistance annotations, leaving approximately 25 genes in the unknown mechanism category; reclassification would move them to enzymatic inactivation and metal resistance, respectively. ₇
• Core/accessory labels use a ≥95% prevalence threshold, but most Fitness Browser species have few GTDB genomes: the median is 9, with a range of 2–399. A gene present in all 9 sampled genomes may be mislabeled core at larger sampling depth, so the core-versus-accessory null result is especially cautious for species with fewer than 20 genomes. ₇
• The class-matched analysis has only 157 pairs and is non-significant (p = 0.14) despite a mean flip of +0.113; the any-antibiotic analysis has greater power (N = 797, p = 0.0001). ₇

Pan-Bacterial AMR Gene Landscape

The report identifies sampling bias as a limitation because genome databases over-represent clinical pathogens, potentially inflating AMR counts for human-associated species. AMRFinderPlus also includes stress-response genes such as mercury- and arsenic-resistance genes, so the 83K hits are not all antibiotic resistance in the narrow sense. ₈

AlphaEarth embeddings covered only 28% of genomes and were biased toward genomes with geographic metadata. The Fitness Browser analysis covered only 37/48 Fitness Browser organisms with AMR genes, and these were predominantly environmental strains; consequently, it did not capture the cost of recently acquired mobile resistance in pathogens. ₈

In its limitations, the report says keyword-based mechanism classification left 22% of hits in Other/Unclassified. The mechanism section gives this category as 22.2%. It includes genes whose product descriptions do not match any keyword set, such as ribosomal protection proteins with non-standard names and novel resistance mechanisms. The report proposes reducing this fraction by mapping bakta_db_xrefs cross-references to CARD Antibiotic Resistance Ontology (ARO) terms, which would give a systematic, ontology-based classification. Singleton inflation may also cause some singleton AMR clusters to reflect annotation artifacts rather than true species-specific resistance genes. ₈

The six hypotheses involved many individual tests, including per-mechanism binomial tests, per-phylum correlations, and per-environment comparisons. Formal Bonferroni or FDR correction was not applied because the primary p-values were extreme, with many < 1e-100, and the report states that correction would not change the conclusions; per-gene and per-phylum tests remain exploratory and hypothesis-generating. FDR means false discovery rate. ₈

The 100% DIAMOND identity threshold used for Fitness Browser pangenome linking is conservative: it avoids paralog confusion but may miss closely related variants, including alleles differing by a single synonymous substitution, and may therefore undercount fitness effects. The Fitness Browser organisms are also mostly environmental isolates in which intrinsic resistance predominates. As a result, the fitness cost of recently acquired mobile resistance elements in clinical pathogens may differ substantially from these measurements. ₈

Within-Species AMR Strain Variation

• Collection bias. The GTDB/NCBI genome collection is heavily biased toward clinical and human-associated isolates, particularly for Klebsiella pneumoniae, Staphylococcus aureus and Escherichia coli. Environmental species are underrepresented. ₉
• Metadata sparsity. This limits interpretation of the temporal and ecological results. Only 70% of genomes had parseable collection dates, and 52.7% had no classifiable isolation_source. ₉
• AMR detection method. Detection relies on the AMRFinderPlus database, so novel resistance mechanisms absent from that database are missed. ₉
• Mantel size cap. ANI extraction for Mantel tests was limited to species with <=500 genomes. This excluded mega-species such as Escherichia coli (15,388 genomes) and Klebsiella pneumoniae (14,240 genomes) from the phylogenetic-signal analysis. ₉
• Environment classification. The BacDive and NCBI keyword classifiers are approximate, and dedicated metadata curation would improve ecotype analyses. ₉
• Causality. Resistance-gene co-occurrence in islands does not prove co-selection. The genes may simply be linked on the same mobile genetic element without functional synergy. ₉
• Core versus non-core signal. The stronger phylogenetic signal of non-core than core AMR genes may be partly statistical. Core genes are nearly universal by definition, which produces little Jaccard-distance variance and suppresses Mantel r. ₉

Annotation-Gap Discovery via Phenotype-Fitness-Pangenome-Gapfilling Integration

The draft models were based on automated RAST annotations and contained systematic errors; the 42.5% baseline FBA accuracy, dominated by false positives, reflected overly permissive models that predicted growth on carbon sources the organisms could not use. ₁₀

Gapfilling is non-unique: multiple valid solutions may exist for each false-negative case. The study used default ModelSEED gapfilling, which minimizes the number of added reactions but does not guarantee biological optimality. ₁₀

The mapping between Fitness Browser experiment names and ModelSEED exchange reactions required manual curation of 109 carbon sources to compound IDs, so mapping errors could generate spurious false negatives. ₁₀

Fitness significance depends on the selected absolute-fitness threshold and the number of experiments; organisms with fewer carbon-source experiments have less statistical power. ₁₀

GapMind covers ~80 carbon and amino acid pathways rather than full metabolism, so many gapfilled reactions fall outside its coverage. ₁₀

The FBA knockout validation was inconclusive because the models could not grow on carbon-source minimal media without the gapfilled reactions; consequently, the reactions being tested were themselves required for growth, making single-gene knockout analysis circular in this setting. ₁₀

The dataset was phylogenetically biased: 12 of 14 organisms were Proteobacteria. The sole Bacteroidetes organism, B. thetaiotaomicron, had the lowest resolution rate, suggesting that the approach may be less effective for phylogenetically distant clades with divergent metabolism. Because this rests on a single organism, it should be read as a hypothesis rather than an established finding. ₁₀

The report labels baseline FBA across 574 combinations as 42.5% overall accuracy. It also gives precision as 42.5% (244 of 574 growth predictions correct), along with recall of 86.5% (244 of 282) and 330 false positives. The report gives the same value for accuracy and precision, so the two should not be treated as independently verified metrics. ₁₀

The report states that conditional gapfilling of 38 false-negative cases added 219 reactions, averaging 5.8 reactions per case. However, its listed categories (201 enzymatic, 14 transport, 12 exchange) do not add up to the stated 219 total, and the report does not resolve this inconsistency. ₁₀

The report is internally inconsistent about BLAST search direction. Its Key Findings describe DIAMOND v2.1.16 blastp searches against the Swiss-Prot exemplar sequences (328 reviewed bacterial sequences for 75/84 unique ECs). The NB06 pipeline description instead says DIAMOND BLAST was run against concatenated proteomes using the 328 downloaded exemplars. Both descriptions report 154 hits, and the report does not reconcile them. ₁₀

Aromatic Catabolism Support Network in ADP1

The PQQ dependency is not exclusively aromatic: PQQ-biosynthesis genes also appear as glucose-specific in the adp1_deletion_phenotypes project, where they are associated with PQQ-dependent glucose dehydrogenase. ₁₁

The report’s summary states that aromatic catabolism requires 51 genes, “over 8×” the core pathway, and that the support requirement exceeds the pathway itself “by 7:1”. Neither ratio is consistent with the report’s own counts of 51 total genes and 8 core-pathway genes. The intended denominators and the meaning of “support requirement” are unresolved, so these ratios should not be relied on. ₁₁

The 8-condition co-fitness matrix provides approximately 5 independent dimensions, so additional conditions are needed to sharpen subsystem boundaries. The ortholog-transferred cross-species data is confounded by organism-specific respiratory architectures, and the non-core Complex I assignments rely on correlation rather than direct physical evidence. ₁₁

The proposed NDH-2 compensation mechanism remains untested. The report proposes searching the ADP1 genome for NDH-2 and testing its deletion on quinate versus glucose; experimentally validating ACIAD3137 (UPF0234) and ACIAD2176 (DUF2280), which show r > 0.98 co-fitness with Complex I, by protein-protein interaction or co-purification studies; expanding the condition panel with benzoate, catechol, vanillate, iron limitation, and respiratory inhibitors; comparing Complex I retention across aromatic-degrading species using pangenome data; and adding PQQ biosynthesis, iron homeostasis, and respiratory-chain capacity constraints to the ADP1 FBA model. ₁₁

BacDive Isolation Environment × Metal Tolerance Prediction

The heavy-metal group contains only n=10 matched isolates and was at the detection limit for d=1.00, so a larger dataset is needed to estimate the effect precisely. ₁₂

BacDive represents culturable, described strains rather than the full diversity of environmental bacteria, and culture-collection bias may under-represent metal-tolerant extremophiles. ₁₂

Species-level matching is lossy: 56.6% of BacDive strains did not match a GTDB species, primarily because GTDB uses different species boundaries from LPSN/DSMZ. Genome-accession matching through GCA→pangenome genome_id could improve coverage but requires a Spark query. ₁₂

The metal tolerance score is genome-size-normalized as metal clusters divided by annotated clusters. This controls for genome size, which is important because Pseudomonadota tend to have larger genomes, but normalization does not eliminate the possibility that metal-tolerance functions correlate with total metabolic complexity. Without normalization, the environmental signal could partly reflect genome size rather than metal gene content. ₁₂

The metal-utilization validation is underpowered because only 24 records matched strains with metal scores; the negative direction for positive utilizers should not be over-interpreted. ₁₂

The absence of a significant signal in Bacillota and Bacteroidota cannot distinguish real biological differences from limited power because contamination-isolate sample sizes were small. ₁₂

BacDive Phenotype Signatures of Metal Tolerance

The Metal Fitness Atlas scores are genome-based predictions rather than direct metal-tolerance measurements. Controlling for n_metal_clusters in partial correlations only mitigates circular reasoning. The tested associations remain phenotype-to-genome correlations rather than phenotype-to-phenotype measurements. ₁₃

The report says BacDive species-name matching achieved 38.4%, citing 5,647 of 27,702 GTDB species. That percentage is inconsistent with the stated species numerator and denominator. The report uses the same 38.4% elsewhere for 37,368 matched strains out of 97,334 BacDive strains, which suggests that the strain-level rate was carried over to the species-level count. The report does not give a correct species-level rate. GCA accession matching was not implemented. The report proposes that it could recover 10–30% more species, but that gain is prospective, not measured. ₁₃

Erratum. The report gives 38.4% for two different denominators. 37,368 of 97,334 matched strains is 38.4%; 5,647 of 27,702 GTDB species, the figure this sentence uses, is about one fifth, roughly half what the percentage suggests. ₁₃

The 12-organism direct validation was underpowered: all Gram-typed organisms were Gram-negative, which prevented within-set testing of H1a. ₁₃

BacDive testing is biased toward well-studied organisms such as Pseudomonas and Escherichia coli, which have many phenotype tests. Poorly studied species have sparse data. ₁₃

The H₂S result is underpowered because only 8 H₂S-negative species were present in the matched set. Its d = -0.87 is therefore unreliable and likely inflated by small-sample bias. The H₂S–chalcophilic-metal link remains a hypothesis; the report proposes testing it experimentally with sulfate-reducing bacteria under zinc/copper challenge. ₁₃

The composite-score analysis does not establish metal-specific mechanisms. Testing whether particular phenotypes predict tolerance to particular metals (for example catalase and copper, urease and nickel, or H₂S and zinc, copper and cadmium) would require per-metal scores for the 27K species. ₁₃

The analysis used only measured BacDive phenotype values. Adding BacDive’s machine-learning-predicted Gram stain, motility and oxygen tolerance could increase coverage substantially, but would introduce model-dependent bias. ₁₃

Whether urease-positive organisms are specifically nickel-tolerant, rather than generally metal-tolerant, remains unresolved. The proposed test is per-metal fitness profiling of urease-positive versus urease-negative bacteria from the same taxonomic class. Gammaproteobacteria are the preferred class, because there the pangenome-scale urease effect is near zero. RB-TnSeq (random barcode transposon sequencing, a pooled mutant-fitness assay; see tnseq) of matched urease+/- strains under nickel versus other metals would test this directly. ₁₃

Other suggested follow-up includes GCA accession matching. It also includes phylogenetic-generalized-least-squares (PGLS, a regression that models phylogenetic non-independence) or phylogenetic PCA, to formally remove phylogenetic signal before testing phenotype–metal associations. ₁₃

Subsurface Bacillota_B Specialization — What Distinguishes Deep-Clay Lineages from Soil Congeners?

The anchor cohort contains 10 genomes and the baseline contains 62. Fisher’s exact testing is considered adequate for the large 547-OG effect, but marginal effects supported by anchor counts of 3–5 should be treated descriptively rather than inferentially. Spanning four orders (Desulfotomaculales, Desulfitobacteriales, Ammonifexales, Thermacetogeniales) only partly mitigates genus-level phylogenetic confounding. ₁₄

The anchor cohort is borehole- and porewater-dominated by construction, reflecting cultivation bias toward porewater isolates. The comparison therefore cannot test whether rock-attached Bacillota_B differ in gene content. The 10-genome anchor was assembled ~entirely from ncbi_env keyword search. BacDive expansion through the kescience_bacdive isolation and taxonomy tables yielded zero additional Bacillota_B genomes in this run. ₁₄

Genus-level phylogenetic confounding is only partly mitigated: the cohort spans four orders, but the 10-genome anchor is clumped among 3 BRH-c8a genomes, 2 BRH-c4a genomes, 2 Desulfosporosinus genomes, Desulforudis, Ch130, and 1 other genome. Some enriched OGs may therefore be lineage markers rather than recurrent subsurface-specialization features. The report states that a per-genus analysis is needed to separate genus-specific lineage-marker OGs from OGs that recur across multiple subsurface-specialist genera. It also states that with n=10 anchor genomes such a decomposition is at the limit of statistical resolution. ₁₄

The OG hierarchy is imperfect. The analysis used Firmicutes-level OGs (eggNOG tax 1239) where available and Bacteria/root fallbacks otherwise; some enriched OGs are at root level, which is coarse. A Bacillota_B-specific eggNOG tier was unavailable because Bacillota_B is GTDB-defined while eggNOG uses NCBI taxonomy, where Bacillota_B genomes are spread across Firmicutes / Negativicutes / Tissierellia at the legacy class level. ₁₄

Keyword-based functional categorization undercounted relevant functions: the 462-OG “other_or_unannotated” group contained substantial anaerobic-respiration and electron-transfer signal, including COG1977 molybdopterin metabolism, DsrEFH-like proteins, and 2-oxoglutarate:ferredoxin oxidoreductase. Manual reclassification or an LLM-based extractor would be needed to refine category counts. ₁₄

The corrected iron-reduction analysis is considered robust for the multi-heme cytochrome signal but is a Phase 1 correction. It does not fully determine whether the original clay-project comparison with the Bagnoud porewater pattern should be retained; the SR side remains robust, whereas the iron-reduction side loses force. A planned amendment to the clay-project report would clarify that the iron-reduction side of the clay project’s H3 finding is artefactual and that the SR side stands. The report presents this amendment as pending, not as already applied. ₁₄

The report proposes applying the correction to the clay-project branch, refining the 462-OG category with an LLM-based scan, decomposing the H1 signal by genus, and seeking BacDive linkage to additional clay-isolated Bacillota_B. It also proposes localizing the functions responsible for the larger anchor genomes through a COG / KEGG-pathway breakdown of the “extra” gene fraction. Finally, it proposes testing the same enrichment framework in other phylum-matched subsurface comparisons. Whether the H1 functional-category signal generalizes beyond Bacillota_B or is Bacillota_B-specific therefore remains untested. ₁₄

BERDL Data Atlas — Inventory, Topic Map, and Cross-Reference Synergies

• Join-key presence demonstrates schema-level compatibility, not valid value-space overlap. UC2–UC5 require live-cluster execution; UC1 is the only sample-validated use case. ₁₅
• Two tenant-to-agency mappings, evaluation and lambda, remain unverified by program documentation; together they account for 4 tables. The phagefoundry and msyscolo mappings were first inferred as “likely” Defense/HHS and NSF/USDA. They were then user-corrected to DOE BRaVE and DOE/NSF, respectively. ₁₅
• The realized-use audit mined project README files and may have missed data-source mentions in research plans or notebook source. The reported tenant breadth is therefore a lower bound. ₁₅
• Most NB05 depth counts are COUNT(*) row totals; only the canonical KBase genome count uses COUNT(DISTINCT genome_id). A pangenome gene row represents one genome-gene pair, so 1.01B gene rows correspond approximately to 293K genomes multiplied by approximately 3.4K genes per genome. ₁₅
• Across-tenant deduplication was not performed. Refdata and KBase may contain the same UniProt entries through different cluster indices. ENIGMA and genome-depot tables share genome records with the ENIGMA SDT layer. ₁₅
• The inventory-audit notebook (00_inventory_audit.ipynb) walks the Spark catalog, tags every table by tenant, agency and biological topic, deduplicates dotted-namespace duplicates, and audits the unclassified residual (0.9 %). This catalog-level deduplication is distinct from the across-tenant record deduplication that was not performed. ₁₅
• The validated UC1 cohort lacks per-residue pLDDT (AlphaFold’s per-residue confidence score) and structural-feature data in kescience_alphafold.alphafold_entries. These features must be ingested or computed from PDB files before downstream structure-function analysis. ₁₅
• The report says, citing agent memory rather than a live-cluster count in this study, that NMDC metabolomics (3.1M), proteomics (346K), and lipidomics (1.4M) layers are largely untapped despite being a primary cross-validation source. Both the layer sizes and the underuse assessment therefore carry that provenance. UC4 and UC5 both depend on this layer. ₁₅

Regulatory and proteomic architecture of Δfur-permitted lipid A loss in *Caulobacter crescentus*

The ΔsspB-buffered cbb3/fix respiratory program is a clean transcript-level observation, but its fitness phenotype-bearing rate was 34.6%, versus 33.25% for the genome background, with hypergeometric p = 0.515 and fold = 1.04×. Consequently, the claim that respiratory ATP is required for envelope remodeling remains a working hypothesis, not an established mechanism, and the fitness data provide no independent enrichment support for the SspB-buffered respiratory set. Read as a “dual-release switch”, the Δfur arm is supported, whereas the ΔsspB-arm respiratory protection is a plausible hypothesis that is not statistically distinguished from genome background. The pre-registered ≥10% threshold sat below the 33.25% background rate — a methodological miscalibration that the adversarial review surfaced and that the report records as a learned pattern: pre-registered fitness-phenotype thresholds must be calibrated against the genome-wide background rate using fold-enrichment rather than a fixed percentage, since a ≥10% threshold against a 33% background passes any random gene set. ₁₆

The fitness compendium contained zero iron-limitation experiments among 198 Caulobacter experiments, so the iron-limitation arm of H2 was descoped to envelope-only analysis. Testing that arm requires additional RB-TnSeq experiments under bipyridyl chelation, ferric supplementation, and hemin, or a cross-walk to published Δfur phenotypes. ₁₆

CtpA was rejected at the preregistered bar: the BORDERLINE criterion required both 0.05 < pvalue < 0.15 and protein detection, but the observed pvalue = 0.048 failed the lower bound, FDR = 0.109 failed the PASS criterion, and CtpA was not detected in the outer-membrane proteome. The cumulative 4599-vs-4580 result (logFC +0.76, FDR 0.035) is significant but cannot isolate the Δlpxc-specific response because it conflates Δfur, ΔsspB, and Δlpxc effects; it is reported as a supplemental observation consistent with CtpA being Fur-regulated, not as evidence for lipid-A-loss-driven CtpA induction. CtpA’s specific role as the LpxF-equivalent processing step, which Zik 2022 predicted, is therefore not supported at the pre-registered bar, and that LpxF-substitute hypothesis remains untested. ₁₆

The OM proteome had a single replicate per strain, so no per-protein statistics were available. The lptC2 protein induction, Pal-Tol upregulation, LptD/LptE decline, and CCNA_01217 increase require replicated proteomics before publication-level claims can be made. ₁₆

The transcriptome used a single PYE rich-medium growth condition, and the observed Fur signal represents constitutive Δfur derepression rather than a direct iron-limitation response. Caulobacter SigU lacks a characterized published regulon, and the late ChvI cohort did not pass the relaxed coherence criterion. ₁₆

The comparative PaperBLAST analysis was vulnerable to naming-convention false negatives, and M. catarrhalis was under-annotated in PaperBLAST with 162 genes total, so the report directs that its presence/absence rows be treated as advisory only. NCBI annotation strengthened the headline absence claims, but annotation-based search misses unannotated paralogs, so a deeper Pfam (protein-family hidden Markov model) search is still needed: the report specifies hmmsearch or pyhmmer queries with PF00155 serine palmitoyltransferase, PF03331 LpxC and PF01475 Fur against the named RefSeq proteomes GCF_000196795.1 (A. baumannii ATCC 17978), GCF_000008805.1 (N. meningitidis MC58) and GCF_000092265.1 (M. catarrhalis BBH18), which would close the residual concern that an unannotated spt paralog in one of the comparators was missed. ₁₆

The peptidoglycan gene set included two likely regex false positives: CCNA_00565, a γ-glutamyltranspeptidase matched through “transpeptidase,” and CCNA_01833, a glucosylceramidase matched because “ceramidase” contains “amidase.” Their inclusion did not change the H4 verdict, which remained well above the ≥3 threshold. ₁₆

Two internal labeling corrections are recorded. An earlier draft conflated the 20 unique-to-early genes with the 10 both-phase genes, summing to 30, whereas the disjoint partition is 20 + 10 + 49 = 79 ChvI-induced genes engaged in these data; the same draft mis-labeled ChvI as an “early cohort” gene when it belongs to the 10-gene both-phases subset. ₁₆

The dual-release-switch model’s genetic predictions remain untested here: the model predicts that Δlpxc should not be viable on a Δfur-alone background with sspB intact, because the cbb3/fix respiratory chain would shut down, and should not be viable on a ΔsspB-alone background with fur intact, because Fur would repress the transport machinery needed for sphingolipid trafficking; the report states that both predictions are testable in the existing strain background. ₁₆

The post-transcriptional flux model is likewise untested by this project. It predicts that in strain 4599 (Δlpxc Δfur ΔsspB) the sphingolipid pool size should be maintained or increased despite the absence of biosynthesis induction, because LpxC competition for shared substrates is gone, and that CtpA-mediated processing should produce a specific CPG headgroup as the LpxF substitute step; targeted lipidomics with intact lipid A precursor quantification is proposed to test both predictions. ₁₆

A cross-species engineering test is proposed rather than performed: the “structural unavailability” model predicts that introducing the Caulobacter sphingolipid biosynthesis pathway into A. baumannii, which lacks it, should be neutral or modestly protective under colistin selection, because the PBP1A/Ldt route in A. baumannii works independently of sphingolipids, so a sphingolipid-engineered A. baumannii would have two alternative routes to colistin resistance — predicting increased rates of colistin-resistant escapers. ₁₆

The revised Pal interpretation also generates an untested prediction: following Tan & Chng 2025, the Tol-Pal complex in the Δlpxc state should show increased retrograde phospholipid-transport activity to maintain outer-membrane lipid homeostasis, which pulse-chase lipidomics following exogenous fluorescent phospholipid loading would test directly. Tol-Pal essentiality in Caulobacter (Yeh 2010) complicates depletion experiments, but the test remains feasible in conditional-depletion strains. ₁₆

The report identifies replicated OM proteomics, targeted lptC2 and Pal assays, SigU-induction RNA-seq, genetic tests of the dual-release model, lipidomics, iron-axis fitness experiments, Tol-Pal phospholipid-transport assays, cross-species engineering, and sequence-based homology searches as specific resolving work. ₁₆

The report lists Adam Arkin (University of California, Berkeley; ORCID 0000-0002-4999-2931) as lead analyst and Kathleen R. Ryan (University of California, Berkeley) as data provider and scientific collaborator, with co-authorship pending publication discussion. ₁₆

Rational Design of Protective Microbiome Formulations for Competitive Exclusion of *Pseudomonas aeruginosa* in Cystic Fibrosis Airways

The inhibition assays were planktonic and used PA14, whereas PA in cystic-fibrosis lungs primarily occupies structured biofilms. The carbon panel contained 22 tested substrates and omitted mucins, lipids, iron, polyamines, and the sugar alcohols identified genomically. ₁₇

The metabolic model was based on 142 isolates covering 62 of 211 species (29%), and growth kinetics were available for only 32 isolates; the core cohort was enriched for deeply characterized taxa, including Rothia, Streptococcus, Neisseria, and Gemella. ₁₇

Pairwise interaction data covered only 3 A × 3 B isolate combinations, and the complete 10-pair interaction matrix for the five-species core has not been measured. Moreover, fact_pairwise_interaction was identical to fact_carbon_utilization, with correlation = 1.0 and mean difference = 0.0, so endpoint OD data cannot assess per-substrate co-culture effects; current interaction conclusions rely on the RFU-based competition assay. ₁₇

Engraftability was inferred from patient prevalence and transcriptional activity rather than measured after administration. Only 21 lung genomes across 5 species were available for lung-adaptation comparisons, and M. luteus had zero lung genomes and zero detected patient engraftability despite its central role in achieving 100% niche coverage. ₁₇

PA14-based inhibition measurements have not been validated against PAO1 or ExoS+ clinical strains, although the pangenome analysis found no amino-acid pathway differences between ExoU+ and ExoS+ PA. The report therefore prioritizes testing PAO1 and 3–5 mucoid clinical PA isolates. ₁₇

The primary N. mucosa conservation analysis used a 15-genome clade even though the PROTECT reference mapped to an 8-genome clade; the 8-genome sensitivity check showed stronger conservation, with 18/18 amino-acid pathways at >95% versus 16/18 in the 15-genome clade, 37/62 carbon pathways versus 27/62, and 1 respiratory genome. ₁₇

Self-Sufficiency, Anaerobic Toolkit, and Cultivation Bias in Clay-Confined Cultured Bacterial Genomes

• The deep anchor cohort is small (n = 9 before quality filtering), so only large effects, described as Cohen’s d > 0.7 for unfiltered comparisons, can be reliably detected. Marginal results, such as the within-Bacillota_B self-sufficiency comparison at p = 0.07, are descriptive only. ₁₈
• Compartment annotations were inferred from keywords in isolation-source strings, so some bentonite or “rock” entries could be porewater or rock-attached. The report states that the H3 sulfate-reduction result held under two stricter compartment definitions. ₁₈
• The GapMind metric covers only an 18-pathway amino-acid universe (gapmind_pathways), which saturates near 18 for most cultivable bacteria, so the metric has limited resolving power at its upper end. The report proposes a finer-grained sensitivity analysis that checks all standard amino-acid-biosynthesis EC numbers in eggNOG. ₁₈
• eggNOG cluster-level annotations propagate within clusters of ≥90% AAI, so strain-level marker variants may be missed, including a single non-functional dsrA in an otherwise complete operon. ₁₈
• The report recommends five follow-ups. The strongest next step is to ingest deep-subsurface MAGs from Mont Terri, Olkiluoto, MX-80 bentonite, and Oak Ridge into the pangenome, then re-run the H1/H2/H3 framework. This would test whether an H1 self-sufficiency signal emerges once rock-attached or uncultivated lineages are present, a question that remains untested. Any rerun of the H3 iron-reduction side would need the corrected detector. ₁₈
• An exploratory direct comparison of BRC-3 porewater versus BIC-A1 borehole genomes, with 5 + 3 genomes per borehole, could test whether the Bagnoud-paradigm sulfate-reduction-rich pattern is borehole-specific or site-wide within Mont Terri. ₁₈
• The sulfate-reduction/iron-reduction diagnostic could be applied to other subsurface settings. ₁₈
• The 5/5 Bacillota_B sulfate-reduction enrichment is striking, but its “n=5 vs 19” structure may hide genus-level differences: Desulfosporosinus, BRH-c8a, and BRH-c4a may differ in which non-SR features they carry. The report proposes a within-Bacillota_B genus-level pangenome analysis to surface a possible deep-clay-specific accessory genome. ₁₈
• Genome presence should be linked to Bagnoud’s metaproteomic evidence. Bagnoud (2016) reported protein-level expression of the toolkit modules, but this genome-content analysis does not establish their expression in the genomes analyzed here. ₁₈

Co-fitness and Co-inheritance in Bacterial Pangenomes

The prevalence ceiling attenuates the pairwise signal. Most Fitness Browser genes map to core clusters (>95% prevalence), where phi approaches 0 for both cofit and random pairs because there is almost no variance to correlate. The analysis is therefore most informative for species with substantial auxiliary gene content. ₁₉

The two Ralstonia organisms were excluded even though they were the most phylogenetically diverse and lowest-ANI organisms in the target set. This removed the species likely to have been most informative. ₁₉

Phylogenetic control was limited. Stratification was available for 7 of 9 organisms, and most species lacked genomes in the far stratum (>0.05 branch distance). The near-versus-medium difference (mean phi=0.102 versus 0.067) was consistent with shared ancestry, but the missing far stratum prevents full disentanglement of functional and phylogenetic effects. ₁₉

Pairwise co-fitness captures gene-pair relationships. ICA modules capture multi-gene coordinated regulation and may better represent the selective units that constrain co-inheritance. ₁₉

Near-clonal species behaved differently: Ddia6719 (ANI 99.47%) had delta=+0.093 and pseudo3_N2E3 (ANI 99.66%) had delta=+0.026. Both retained enough accessory variation to detect co-inheritance, but their high baseline phi makes absolute phi values less interpretable. ₁₉

Proposed next analyses:

• Restrict comparisons to auxiliary-only pairs in which both clusters are below 95% prevalence, to maximize presence/absence variance.
• For Ralstonia and other organisms lacking precomputed values, calculate co-fitness directly as pairwise Pearson correlations from raw genefitness data.
• Resolve reference-genome mapping to improve phylogenetic control.
• Build module co-transfer networks and test cross-module prediction.
• Expand to species with >30% auxiliary genes and existing co-fitness data. ₁₉

COG Functional Category Analysis

COG annotations covered approximately 70% of genes, so unassigned genes may skew the reported category distributions. ₂₀

The analysis used 32 species. A larger sample could reveal phylum-specific patterns that are not visible in the current comparison, which qualifies the report’s claim of universality across bacterial phyla. ₂₀

Composite, multi-letter COG categories were counted once per gene rather than split among their component functions. The analysis also relied on eggnog v6 annotations, which may differ from original COG assignments. ₂₀

The report identifies several future analytical directions ₂₀:

• comparisons across additional taxonomic groups;
• detailed examination of COG V defense and COG L recombination functions;
• correlation of novel-gene functions with environmental metadata, to test whether patterns vary by habitat. ₂₀

Gene Conservation, Fitness, and the Architecture of Bacterial Genomes

Module-family breadth did not predict conservation: families spanning more organisms did not have higher core fractions, with rho=-0.01 and p=0.91. The report attributes the absence of a gradient in part to an already high baseline that leaves little room for one. ₂₁

Accessory genes were not systematically burdensome. Contrary to the streamlining hypothesis, they were less costly than core genes. ₂₁

Condition-specific fitness did not mean niche-specific fitness. Genes with strong condition-specific effects were more likely to be core rather than accessory, indicating that core genes simply had more detectable effects under the tested conditions. ₂₁

The findings do not directly establish fitness in natural environments: the laboratory measurements capture the cost of maintaining genes, while the pangenome captures evolutionary pressure to retain them. The costly-and-conserved category is therefore evidence for, rather than a direct measurement of, purifying selection in nature. ₂₁

Conservation vs Fitness — Linking FB Genes to Pangenome Clusters

The essential-gene definition is an upper bound: genes without fitness data may lack transposon insertions because they are short, occur in low-complexity regions, or lie at scaffold edges, rather than because they are essential. Gene-length validation found that essential genes were slightly shorter on average, indicating possible insertion bias. ₂₂

Pangenome coverage varies among clades, and clades containing only 2 genomes can have trivially high core fractions because a gene present in both genomes equals 100% core, reducing the discriminative power of the core-versus-auxiliary classification. ₂₂

The main Escherichia coli clade was absent from the pangenome because it contained too many genomes; Keio, an E. coli BW25113 strain, mapped to the small s__Escherichia_coli_E clade at only 26.1% coverage. ₂₂

Essentiality was measured under a single growth condition represented by the RB-TnSeq library construction conditions, so genes essential only under stress conditions were not captured. ₂₂

Dyella79 was excluded from Phase 2 because the FB gene table used the locus-tag format N515DRAFT_* whereas the protein sequences used ABZR86_RS*, producing a 0% join rate. ₂₂

Ten organisms were excluded from Phase 2 because they had <90% DIAMOND coverage, reducing taxonomic breadth. ₂₂

Core Gene Paradox — Why Are Core Genes More Burdensome?

• Laboratory conditions capture only a fraction of the environmental conditions bacteria face in nature. ₂₃
• “Burden” defined as fit > 1 may reflect trade-offs rather than true dispensability. ₂₃
• The 90% identity threshold used for DIAMOND matching may miss rapidly evolving genes. ₂₃
• Fitness Browser condition types are biased toward experimentally convenient conditions rather than ecologically relevant conditions. ₂₃

The 5,526 Costly + Dispensable Genes

• “Burden” is defined as max_fit > 1 in any experiment, so a single experiment can classify a gene as burdensome and the result is sensitive to noise in the fitness data. ₂₄
• SEED and KEGG annotations cover only 56-79% of genes, and the unannotated fraction may have different functional profiles. ₂₄
• The pangenome core/accessory (auxiliary) classification is binary; using the fraction of genomes carrying each gene would provide more resolution. ₂₄
• The 90% identity DIAMOND threshold used for Fitness Browser–pangenome linking may miss recently acquired genes with low sequence similarity. ₂₄
• The extreme 21.5% costly+dispensable proportion in psRCH2 may reflect strain-specific genomic features rather than a general pattern. ₂₄
• Ortholog groups are assigned using bidirectional best hits (BBH) across 48 organisms, so genes with orthologs outside this set may be misclassified as orphans. ₂₄
• Condition-specific phenotype data is biased toward conditions that can be tested in the laboratory. ₂₄

Counter Ion Effects on Metal Fitness Measurements

The 39.8% overlap depends on the NaCl-importance threshold, defined using fit < -1 or n_sick ≥ 1; stricter or more permissive thresholds could reduce or increase the overlap. The S. elongatus estimate is especially sensitive because it had 12 NaCl experiments spanning 0.5–250 mM, whereas other organisms had 1–6 NaCl experiments. ₂₅

NaCl cannot isolate chloride because it delivers both Na⁺ and Cl⁻ and produces osmotic effects. A KCl or choline chloride control would more specifically test chloride effects, although some Fitness Browser organisms have choline chloride experiments at different concentrations. The report proposes comparing metal–choline chloride (ChCl) overlap with metal–NaCl overlap to separate chloride from osmotic confounding. This proposal rests on the report’s premise that ChCl provides Cl⁻ without Na⁺ or osmotic effects. That premise is not tested here, and the comparison is not reported as completed. ₂₅

Approximately 14.3% of protein-coding genes, classified as putative essential genes, lacked transposon insertions and were absent from both NaCl and metal-fitness data. These genes are 82% core, and their exclusion affects the original and corrected conservation analyses equally. ₂₅

Manganese, cadmium, selenium, mercury, iron, molybdenum, and tungsten were each tested in only 1 organism, DvH, so their overlap statistics lack cross-organism replication. The psRCH2 CuCl₂–CuSO₄ comparison is additionally limited by aerobic/anaerobic confounding. ₂₅

The shared-stress versus metal-specific classification did not include formal functional-enrichment tests; the SEED annotation comparison was descriptive. The proposed mechanistic categories therefore remain hypotheses requiring direct functional testing. ₂₅

The report proposes formal COG (Clusters of Orthologous Groups functional categories), KEGG, and PFAM enrichment tests; comparison with choline chloride; module-level analysis using independent component analysis (ICA); refinement of condition-specific metal-gene analysis; and matched CuCl₂/CuSO₄, ZnCl₂/ZnSO₄, and CoCl₂/CoSO₄ RB-TnSeq experiments under identical conditions. It also proposes DvH NaCl dose-response experiments at 0.1, 1, 10, 100, and 500 mM to match effective chloride doses. ₂₅

The proposed COG, KEGG, and PFAM tests would formally check whether shared-stress genes are enriched for envelope/transport functions and metal-specific genes for metal-binding/efflux; that pattern is not established here. The proposed module-level analysis would apply the metal–NaCl overlap to modules from the fitness_modules project rather than to individual genes. Whether metal-responsive modules overlap NaCl-responsive modules therefore remains open. ₂₅

The report attributes to metal_fitness_atlas the finding that condition-specific genes, important only for metals and not other stresses, are less core. It proposes that this project’s shared-stress versus metal-specific classification could refine that analysis. ₂₅

The report names RB-TnSeq with matched metal chloride and metal sulfate concentrations in a single organism under identical growth conditions as the most impactful follow-up, because it would eliminate the aerobic/anaerobic confound that limits the psRCH2 comparison. It proposes three pairs. The first is CuCl₂ 0.5 mM versus CuSO₄ 0.5 mM in MR-1 or DvH, because copper is the most-tested metal. The second is ZnCl₂ 0.2 mM versus ZnSO₄ 0.2 mM in MR-1: zinc was already tested as sulfate, and adding the chloride form would directly test the null. The third is CoCl₂ 0.1 mM versus CoSO₄ 0.1 mM in DvH: cobalt has the highest Cl⁻ load in current data, and the sulfate form would confirm mechanism. These are proposed conditions, not reported results. ₂₅

Discoveries Log

Several results are explicitly exploratory, coverage-limited, or dependent on analytical specification. The SSO geochemical model lacks ingested measurement values; AlphaEarth analyses had 3,838/83,287 genomes with NaN in at least one of the 64 embedding dimensions — the cause is unclear, possibly missing satellite imagery or coordinates outside satellite coverage, and these records must be filtered before any embedding-based analysis — and genomes clustered at coordinates with more than 50 genomes of more than 10 species; only 6.8% of species had sufficient embedding coverage. ₂₆

That high-density-coordinate screen flags 30,469 genomes (36.6%), but not every such location is an institutional-address artifact: several flagged locations are legitimate DOE or research field sites — Rifle, CO (1,883 genomes, DOE IFRC), Saanich Inlet (1,529, oceanographic time series) and Siberian soda lakes (812, extremophile campaigns) — so the heuristic needs refinement, for example by checking isolation_source homogeneity at each location to distinguish field sites from institutional addresses. ₂₆

The NMDC metabolomics dataset was dominated by one study: 125/131 samples (95%) came from nmdc:sty-11-r2h77870, with the remaining 6 samples from a second study. Only approximately 2% of compounds in nmdc_arkin.metabolomics_gold had KEGG compound IDs populated, so pathway matching had to rely on compound-name substring patterns rather than KEGG-based ontologies; substring matching introduces collision risk ('leucine' matches inside 'isoleucine'), handled by first-match-wins ordering with longer, more specific patterns placed before shorter ones. Chorismate itself is rarely detected, and shikimic acid and 3-dehydroshikimic acid serve only as upstream proxies that reflect precursor availability rather than chorismate pool size. Three amino-acid pathways—cysteine, histidine, and lysine—remained untestable because the corresponding compound detections were absent in the 175-sample overlap cohort. ₂₆

BacDive species matching, by exact species name plus GTDB suffix removal, linked 42,227 strains (43.4% of 97K) to 6,426 GTDB pangenome species, leaving 56.6% unmatched because GTDB species boundaries differ from LPSN/DSMZ taxonomy; the bridge enables cross-referencing BacDive phenotypic data with pangenome-derived predictions without Spark queries. Phenotype-only metal-tolerance models were phylogenetically confounded: classical microbiology phenotypes (Gram stain, oxygen tolerance, enzyme activities) carried real signal on their own (R²=0.16, 7/10 significant after FDR) but added nothing to a taxonomy-based model (delta R²=-0.009), while the genome’s metal-resistance gene count was the predictor that mattered (full model R²=0.63). Urease-positive bacteria surprisingly had lower metal tolerance (d=-0.18), attributed to Actinomycetes lineage composition. ₂₆

Twelve pangenomes had no matching species clade in the species metadata and are mostly obligate symbionts — Portiera aleyrodidarum, the whitefly symbiont, and Profftella armatura, the psyllid symbiont — plus various uncultivated lineages carrying UBA, TMED and SCGC prefixes; these remain valid pangenomes but were filtered out of species metadata because of their single-genome status. ₂₆

Annotation proxies require source-specific validation. An NB01 calibration across 134K KBase Data Lakehouse pangenome genomes found the two annotation sources for lanmodulin to disagree completely: Bakta product='Lanmodulin' flags 62 genomes, all in α-Proteobacterial methylotroph clades (Beijerinckiaceae Methylobacterium, Acetobacteraceae BOG-930, Hyphomicrobiaceae), whereas eggNOG Preferred_name='lanM' flags 505 additional genomes with zero overlap with the Bakta-validated set, concentrated in unrelated gut Bacillota (Streptococcus pneumoniae 10, Blautia_A wexlerae 9, Enterococcus faecalis 8, Ruminococcus_B gnavus 8). The log suggests the hypothesis that the eggNOG preferred name for lanmodulin is a stale or seed-ortholog mislabeling rather than a true homology call, and advises using the Bakta product exclusively while treating eggNOG Preferred_name='lanM' as zero-information for lanmodulin analyses on the KBase Data Lakehouse pangenome. ₂₆

The same pattern holds for xoxJ: KEGG ortholog K02030 is the generic ‘amino acid ABC transporter substrate-binding protein’ family, and filtering eggNOG on it returns 46,369 hits across nearly every phylum, almost none of which are xoxJ, while the specific Bakta product call (‘rare earth element methanol dehydrogenase accessory protein XoxJ’) matches just 61 genomes. K02030 should therefore never be used as a xoxJ proxy; the Bakta product or eggNOG Preferred_name='xoxJ' is the more specific marker despite the generic KO. ₂₆

For xoxF the two sources disagree in both directions: of 5,092 genomes carrying any xoxF marker, only 418 (8%) hit both eggNOG K00114 and Bakta lanthanide-dependent methanol-dehydrogenase-family products. The eggNOG-only set of 3,272 genomes appears to include Bakta Putative dehydrogenase XoxF entries that the project’s regex did not match plus other generic dehydrogenases matching the KO, and the Bakta-only set of 1,402 includes lanthanide-dependent ethanol dehydrogenase variants that the KEGG KO filter excludes. Results should therefore report marker definitions explicitly, give both an eggNOG-K00114 primary count and an either-source inclusive count while disclosing the gap, and avoid treating generic KOs or stale preferred names as definitive. ₂₆

Ecotype Correlation Analysis

• AlphaEarth embeddings covered only 28.4% of genomes, limiting the environmental signal available for analysis. ₂₇
• Geographic coordinates in NCBI metadata were often missing or imprecise, reducing the quality of environmental associations. ₂₇
• Partial correlations assume linear relationships between distance matrices and may not capture nonlinear ecological effects. ₂₇
• Geographic coordinates are less biologically meaningful for host-associated organisms because they generally describe collection sites rather than the organisms’ actual microenvironments. ₂₇

Ecotype Reanalysis — Environmental vs Human-Associated Species

• No downsampling: The reanalysis produced partial correlations 27x higher overall than the original analysis. The report states that absolute values are not comparable, although the within-method group comparison is valid. ₂₈
• NaN exclusion: Environmental species had a higher NaN rate, 21%, than human-associated species, 7%, so the environmental group was more filtered. The report states that this would bias toward finding a stronger environmental signal, which was not observed. ₂₈
• K. pneumoniae exclusion: Klebsiella pneumoniae was excluded because it exceeded Spark’s maxResultSize during gene-cluster extraction and consequently had no correlation data. ₂₈
• Majority-vote classification: A species with 51% gut genomes would be classified as Human-associated. The continuous Spearman analysis was used to address this limitation and also found no relationship. ₂₈
• Unresolved methodological discrepancy: A specific comparison of downsampled versus full-genome extraction is needed to explain the 27x partial-correlation discrepancy. ₂₈
• Functional specificity: Testing gene subsets, including transport and secondary-metabolism categories, could determine whether environmental effects are masked by whole-genome Jaccard distances. ₂₈
• Environment ontology: Repeating the analysis with structured ENVO terms from env_broad_scale could test whether more precise environmental classification changes the result. ₂₈
• Genome-count control: Adding genome count as a covariate could test whether unequal sampling depth inflates correlations in species with more genomes. ₂₈

Ecotype Functional Differentiation

• The analysis covered 12 species from a 15-species sample (80%), out of 456 eligible species; although the stratified sample represented pangenome-size bins, it may not capture broader phylogenetic or ecological diversity. ₂₉
• KMeans was used because HDBSCAN was unavailable on the cluster; KMeans assumes spherical clusters and requires a selected k, whereas HDBSCAN could better handle variable-density subpopulations. ₂₉
• Approximately 38% of gene clusters had COG annotations, leaving 62% unannotated; unannotated ecotype-specific adaptive genes may therefore be missed or the results may be biased toward better-characterized functions. ₂₉
• Without within-species phylogenetic controls such as core-genome trees, the analysis cannot distinguish ecological adaptation from phylogenetic or demographic substructure. ₂₉
• Effect sizes were small: the largest mean effects were 0.039 for S and 0.034 for L in the report’s rounded narrative, while the detailed table gives 0.0392 and 0.0337; these represent approximately 3–4 percentage-point differences in COG-category proportions, and statistical significance may partly reflect large sample sizes. ₂₉
• Variable Spark query times of 94s–1642s during heavy cluster usage caused two species to be lost to S3 read errors. ₂₉

ENIGMA Carbon Census — Tiered Knowledge Census of 83 Enrichment Compounds

“Organism-dark” means not linkable through the queried the KBase Data Lakehouse and curated resources, not unknown to science. Class-level catabolic literature exists for compounds including monoterpenes and nicotine, while the project’s zero literature rescues resulted from a shallow PubMed-title-only screen. A PaperBLAST or abstract-level search could reclassify part of the dark set. ₃₀

The dark fraction depends on the catabolic-direction filter: the 8 ENIGMA-isolate calls used a genome-prevalence-<10% signature-reaction filter retaining reactions that were catabolic according to KEGG degradation-map membership or a 3-reaction curated allowlist. The filter yielded 846 Tier-2 pathway and 102 Tier-3 signature genome-compound calls. Biosynthetic-only signatures were excluded and landed in the dark set. A different filter could change the callable/dark boundary. Lauric acid was independently callable through measured fitness and was not subject to this filter. ₃₀

Soil-versus-freshwater enrichment statistics were exploratory and not calibrated. Treating each metagenome as independent in rank tests over compositional, zero-inflated relative abundances can inflate significance; all 83 genera reached q<0.05, with many q values near 1e-70. Direction and rank were considered more trustworthy than the p-values, and label-free outlier discovery was the more defensible signal. ₃₀

The environmental atlas is a biome-occupancy proxy, not evidence of compound degradation or activity. Genus occurrence does not validate the upstream compound→pathway→organism inference and cannot show that any genus degrades a census compound in any sampled environment. The H3 contrast was genuinely untestable because only 2 necromass compounds were ENIGMA-isolate-callable and both were phthalate-class aromatics. The marine arm was small and gene-blind, with 302 runs and presence/abundance data only, so it is suitable only as a negative biome contrast. ₃₀

The callable-versus-dark physicochemical analysis was descriptive rather than calibrated inference because n=9 callable compounds were available and the Mann–Whitney p-values were uncorrected. Xanthine remains a category error in the carbon census until downstream tables are regenerated after excluding R02107. ₃₀

The project also identified data-pipeline requirements: NMDC and Planet Microbe inputs were species-level, so genus abundance required species-to-genus aggregation; the NMDC denominator was 3825 taxonomy-bearing covstats files rather than approximately 6700 sample-file-lookup rows (using the lookup count deflates relative abundances by ~1.75×); and sample-level environment labels from biosample_set provided 99% coverage across two ontologies, compared with approximately 13% from study-table GOLD labels. ₃₀

The report’s interpretation section says “75 dark compounds” are not a failure of method, which conflicts with the revised headline count of 74. The report’s generated-data description identifies 75 as the pre-reclassification count. This page uses the headline 74 and records the stale 75 here rather than silently dropping it. ₃₀

Data sources and their roles:

• PubChem PUG-REST supplied identity resolution (name → CID → InChIKey/SMILES) plus KEGG, ChEBI and ModelSEED cross-references.
• ModelSEED biochemistry supplied compound → reaction → enzyme linkage for Tier 2/3 inference.
• Fitness Browser RB-TnSeq carbon-source experiments supplied the Tier-1 measured-utilization channel.
• The ENIGMA SSO field data (Zhou Lab 16S, enigma_coral bricks) supplied local genus occurrence.
• NMDC collections (covstats_taxonomy_rollup, sample_file_lookup, biosample_set) supplied terrestrial/freshwater genus abundance with ENVO/GOLD environment labels.
• Planet Microbe (planetmicrobe.planetmicrobe: run_to_taxonomy, taxonomy, and run/experiment/sample/project/campaign tables) supplied the marine genus-abundance contrast for global deliverable (c).
• The pangenome↔ENIGMA taxonomic bridge is genus-level at best, so the project avoided isolate-specific transfer of pangenome Tier-2 reconstructions in favor of direct genome_depot annotation.
• The report cites these core data sources: PubChem (Kim et al., Nucleic Acids Res. 2023); KBase (Arkin et al., Nat. Biotechnol. 2018); Fitness Browser / RB-TnSeq (Price et al., Nature 2018); GTDB (Parks et al., Nucleic Acids Res. 2022); ModelSEED (Henry et al., Nat. Biotechnol. 2010); NMDC (Eloe-Fadrosh et al., Nat. Microbiol. 2022); and Planet Microbe (Ponsero et al., GigaScience 2021). ₃₀

Contamination Gradient vs Functional Potential in ENIGMA Communities

The ENIGMA taxonomy table ddt_brick0000454 provides labels through Genus but no species or strain labels, so direct species-level bridge testing is not possible for this dataset slice. Genus-level mapping may therefore mask strain-level adaptation. ₃₁

A total of 862 of 1,392 observed genera were unmapped to the current pangenome bridge, and the bridge includes substantial ambiguity, including 380 multi-clade genera and a maximum of 433 species clades per genus. ₃₁

COG-fraction proxies are coarse summaries rather than curated metal-resistance pathways. Exploratory sensitivity significance was concentrated in defense and depended on coverage and covariate specification; broader effects across outcomes are not established. The bridge and annotation stack depends on GTDB and eggNOG conventions, and at genus-level aggregation broad COG-style signals can dilute pathway-specific metal-response effects. ₃₁

Site structure was represented by coarse location_prefix effects rather than full hierarchical or random-effects modeling. The analysis therefore does not yet establish whether the exploratory defense associations persist under richer well- or location-level structure control. ₃₁

The document identifies five next analyses: replace COG-fraction proxies with curated metal-stress gene sets and pathway-level summaries; increase taxonomic resolution using species- or strain-level ENIGMA or metagenomic data; fit models with depth, location cluster, sampling date, and compositional controls; investigate unmapped-genera contributions and bridge expansion; and add mixed-effects or hierarchical site models. ₃₁

SSO Subsurface Community Ecology — Spatial Structure, Functional Gradients, and Hydrogeological Drivers

Direct SSO geochemistry is unavailable in the analyzed dataset: 221 geochemistry sample tubes are registered in CORAL, but metals, ion chromatography/total organic carbon, isotopes, ammonia, and nitrite measurements have not been loaded. Consequently, the proposed northeast-to-southwest plume, the M5 mixing-zone interpretation, the M6 plume-core interpretation, and the inferred redox ladder are environmental hypotheses based on community composition and trait inference. All environmental inferences come from community composition alone. Loading the measurements from the 221 registered geochemistry samples (metals, IC/TOC, isotopes, NH₃/NO₂) would allow direct correlation of community composition with measured environmental parameters, validating or refuting the plume model. ₃₂

Groundwater ASV coverage exists for only 5 of 9 wells—L7, L9, M4, M6, and U2—and excludes the critical inferred hotspot wells M5 and U3. The denitrification and iron-oxidation hotspot interpretations therefore rest on sediment data, and the report states that groundwater validation at these wells would substantially strengthen the plume model. Pump-test ASV data from Brick 460-462 for L8, M5, and U2 remains available for future extraction. The predicted groundwater pattern is that Rhodanobacter will be highest at M5 and lower at L8 and U2. ₃₂

Genus-level functional annotation covered only 21% of total reads, with 65 of 1,038 genera annotated; the report therefore treats process abundance estimates as lower bounds. Genus-level taxonomy covered 44% of sediment reads, species-level classification was approximately 0%, and 56% of reads remained outside the genus-level inference. Class-level traits had 78% coverage and showed consistent redox patterns, but the report recommends sensitivity analysis against the lower-coverage genus-level results. ₃₂

Trait scores at phylum and class levels are consensus estimates rather than empirical measurements of the specific SSO populations, and functional assignments are based on literature-linked taxonomy rather than direct genomic evidence. Metagenomics at the same spatial resolution is proposed to test these assignments and capture the 56% of reads without genus-level classification. ₃₂

The sediment–groundwater comparison is confounded by the 18-month sampling offset (sediment cores Feb–Mar 2023, groundwater Sep 2024), sediment has no within-well temporal replication, and seasonal or plume dynamics could affect the comparison. The report also notes that the single sediment timepoint cannot assess temporal dynamics, while the 9-day groundwater result cannot establish stability across seasons or longer plume fluctuations. ₃₂

The most direct resolving analyses are to load the 221 SSO geochemistry samples into CORAL; test whether nitrate, pH, and metal concentrations form the predicted northeast-to-southwest gradient, with highest contamination at U3/M6 and lowest at M4/U1; examine nearby EU/ED well metals from the 100WS/27WS bricks, 90–120 m NE of SSO, which are predicted to show metal concentrations decreasing toward the SSO if the plume approaches from the northeast; extract pump-test ASVs from Brick 460-462 as a third temporal snapshot (Mar 2024) from L8, M5 and U2 to test the M5 denitrification-hotspot prediction; analyze the 18 M6-C2 isolate genomes, which are predicted to encode anaerobic metabolisms (fermentation, sulfate reduction) consistent with M6’s inferred plume-core position, a genomic check not reported here; perform weighted UniFrac using ASV sequences from Bricks 457/460/477, which the report suggests may be more sensitive to plume effects than Bray–Curtis on ASV counts, though this is not demonstrated; and repeat 16S profiling across seasons and plume dynamics to reveal whether community structure tracks plume fluctuations. All of these are predictions awaiting data, not confirmed results. ₃₂

The SSO sediment and groundwater 16S ASV data, sample metadata and well coordinates were drawn from the enigma_coral collection (enigma-coral), including the sdt_sample, sdt_location, sdt_community, ddt_ndarray, ddt_brick0000457-459 and ddt_brick0000477-479 tables. ₃₂

AlphaEarth Embeddings, Geography & Environment Explorer

AlphaEarth coverage is only 28.4% of all genomes and is biased toward genomes with valid latitude and longitude metadata. The 38% human-associated fraction may not represent the overall NCBI or pangenome population. ₃₃

The coordinate-quality heuristic is a crude first pass that flags legitimate field sites such as Rifle and Saanich Inlet. Refinement should check whether genomes at each location have homogeneous isolation sources (a real site) or diverse unrelated sources (an institutional address). ₃₃

Environment harmonization has a long tail: 17% of genomes fall into Other, and keyword matching may miss site-specific labels and non-English terms. ₃₃

The proposed harmonization improvements are to:

• add clinical body-site terms such as cerebrospinal fluid, lung and throat;
• add underground-laboratory terms such as Aspo and Olkiluoto;
• use generic water-source terms;
• use env_broad_scale as a fallback when isolation_source is ambiguous. ₃₃

UMAP is a nonlinear projection. Its apparent cluster structure depends on parameters including n_neighbors and min_dist and may not represent the true high-dimensional topology. DBSCAN with eps=0.5 produced 320 clusters, which may be too fine-grained; coarser clustering may better match environment categories. ₃₃

Embedding NaN values affect 4.6% of the AlphaEarth records, and their cause is unknown; missing satellite imagery at the corresponding coordinates is one possible explanation. ₃₃

The coordinate-distance relationship is strongest below 2,000 km and plateaus above 5,000 km, so it should not be read as a simple raw-distance effect. The report instead relates it to environmental distance-decay. It also notes that in comparable microbial-ecology studies, environmental variation can explain community differences more strongly than geographic separation. ₃₃

The Pan-Bacterial Essential Genome

The essential-gene definition is an upper bound: genes without transposon insertions may lack insertions because of small size, AT-rich sequence, or scaffold-edge effects rather than true essentiality. ₃₄

RB-TnSeq, or random barcode transposon sequencing, defines essentiality under specific library-construction conditions, typically rich media. Genes essential only under stress may therefore be missed. ₃₄

BBH orthology is conservative and can miss paralogs, gene fusions, and distant homologs; consequently, some apparent orphan essentials may have undetected orthologs with diverged sequences. ₃₄

Connected components can over-merge unrelated genes through transitive connections in the BBH graph, particularly for multi-domain proteins. ₃₄

The 48-organism set is taxonomically limited and biased toward culturable Proteobacteria, so essentiality patterns in uncultured lineages, Actinobacteria, or Firmicutes are underrepresented. ₃₄

Module-transfer predictions are indirect because they derive from non-essential orthologs in other organisms, and the function of an essential gene may have diverged from that of its ortholog. ₃₄

Essential Metabolome: GapMind Pathway Analysis Across Essential-Gene Organisms

The report suggests the hypothesis that amino-acid prototrophy is the ancestral state for free-living bacteria. The basis is that 6 of 7 organisms (85.7%) had complete biosynthesis pathways for all 18 amino acids. With so few organisms, this is a hypothesis, not an established pan-bacterial finding. ₃₅

The report also proposes that the high conservation (17/18 pathways at 100% within the sample) indicates a minimal metabolic repertoire required for independent growth. This extrapolates from a small sample and is not a demonstrated requirement. ₃₅

The report does not establish that complete pathways are essential for viability. The essential-gene experiments used RB-TnSeq (random-barcode transposon sequencing, which measures gene fitness from pooled mutant libraries) in rich media. Nutrient supplementation in rich media can make biosynthetic genes appear non-essential, so the data cannot distinguish “essential for biosynthesis” from “essential for viability.” ₃₅

The study analyzed only 7 organisms, not the 45 originally planned. Its phylogenetic diversity is limited, mostly Proteobacteria plus one Deltaproteobacterium. The results therefore cannot be generalized to a pan-bacterial scale. ₃₅

GapMind predictions are computational, not experimental. The analysis used only the complete and likely_complete categories, so partial pathways may be missed, and non-canonical or poorly characterized pathways are not detected. ₃₅

The source data were:

• 305M GapMind predictions across 293K genomes in kbase-ke-pangenome.
• 859 universally essential gene families across 45 organisms, from the essential_genome project. ₃₅

The generated tables were:

• 80 pathway-completeness records.
• 18 amino-acid pathway records.
• 62 carbon-pathway records.
• 7,389 raw GapMind predictions for the selected organisms.
• 8 manual organism-to-genome mappings. ₃₅

Metadata Correlates of Eukaryotic Contamination in NMDC Prokaryote-Targeted Metagenomes

The cross-study analysis included only approximately 9 studies, and one soil study accounted for approximately 43% of runs. No on-system source provided per-sample eukaryotic read fraction across many studies. MGnify on the system is a MAG (metagenome-assembled genome) catalog, SPIRE/GEM/Tara are MAG collections, and EMP is 16S. The cross-study generalization test was therefore under-powered. The within-study result was shown for one soil study only and may not extend to aquatic or host-associated collections. ₃₆

NMDC did not populate DNA-extraction kit, size fractionation or filtration, host-depletion method, or library-preparation fields. Processing booleans in biosample_to_workflow_run were near-constant, and has_filtration was all false. The strongest literature-linked wet-lab factors, host depletion and extraction kit, therefore remain unmeasured residuals that NMDC omits. ₃₆

Absolute eukaryotic fractions depend on the classifier and database. Only GOTTCHA2 yielded a usable eukaryotic fraction, so its values should be read as relative or ordinal rather than as calibrated absolute contamination. ₃₆

Even within the NEON soil study, env_local_scale and geography may track sub-batches such as sampling campaigns. The within-study analysis is the best available control, not a randomized design. ₃₆

Of the 2,759 runs, 1,067 were pooled from multiple biosamples. Each pooled run inherited environment and collection metadata from a single representative biosample selected by MIN(biosample_id). Where pooled biosamples differed in local metadata, this introduced label noise into the predictors. The report describes this as a conservative bias that can weaken associations but not manufacture them. ₃₆

The NEON soil study had zero non-null depth values, so sampling depth was not measured there. The depth association is therefore a cross-study statistic and is not part of the batch-controlled within-study result. ₃₆

The classifier databases were not interchangeable for eukaryote quantification. In this NMDC deployment, Kraken2 and Centrifuge were prokaryote-restricted, whereas GOTTCHA2 was plastid- and eukaryote-aware. Cross-collection analyses of contamination-QC correlates should therefore control for study or batch, using GroupKFold by study or within-study contrasts. ₃₆

Field vs Lab Gene Importance in *Desulfovibrio vulgaris* Hildenborough

The fitness analysis excludes 678 essential genes, 80.1% of which are core, because they lacked transposon mutants. Including them would raise the overall baseline slightly but would not change condition-class comparisons among non-essential genes. ₃₇

This is a single-organism analysis of DvH, so generalizability is limited, particularly to organisms with larger accessory genomes. The Nitratidesulfovibrio vulgaris pangenome contains relatively few genomes, producing a coarse core/auxiliary classification with a high 76.3% baseline core fraction that compresses effect sizes. ₃₇

Condition classification was manually mapped from condition_1 labels, and edge cases such as zinc sulfate as a metal versus sulfate source required subjective judgment. The primary fitness threshold was < -2, although sensitivity analysis from -1 to -3 supported the reported pattern. ₃₇

Gene length is confounded with both fitness-measurement quality, because short genes receive fewer transposon insertions, and core status, because core genes tend to be longer. The field-specific and lab-specific gene sets were small (n=50-52 per group), limiting power for their comparison. Field and lab fitness effects were correlated (r ~ 0.7 from the scatter plot), so most genes that were sick in one context were also sick in the other. ₃₇

The ENIGMA CORAL field samples and ASVs may enable future community and geochemistry analyses, but the current database survey does not provide DvH gene-level fitness data. The report proposes several extensions but does not report results for any of them. They include applying the classification to other ENIGMA organisms with environmental relevance and Fitness Browser data; replacing binary core/auxiliary status with quantitative conservation (fraction of genomes carrying the gene cluster) to increase statistical power; using the 213K ASVs and 4,346 field samples in ENIGMA CORAL to ask whether Desulfovibrio abundance at Oak Ridge sites correlates with the geochemistry conditions tested in the Fitness Browser; deeper analysis of the accessory-genome antibiotic and heavy-metal resistance genes, asking whether they are on mobile elements, recently acquired, or shared with co-occurring species at Oak Ridge, questions that remain unresolved; and replacing binary important/not-important calls with continuous fitness scores (e.g., mean or minimum fitness per condition class) as logistic-regression predictors. The report’s reference list also names Desulfovibrio vulgaris Hildenborough, Pseudomonas aeruginosa, Mycobacterium abscessus, and Shewanella algae in cited literature. ₃₇

Fitness Effects vs Conservation — Quantitative Analysis

The fitness measurements are biased toward rich media and standard stresses, so many ecological niches are unrepresented. ₃₈

The 16-percentage-point conservation gradient, although statistically robust, means that fitness importance is only a weak predictor of conservation. ₃₈

Fitness measurements were based on single-gene knockouts and therefore did not capture epistatic interactions. ₃₈

The Fitness Browser covered 43 bacteria, primarily Proteobacteria, limiting generalizability to other bacterial lineages. ₃₈

Singleton and novel genes may lack fitness data because of poor transposon coverage rather than true neutrality. ₃₈

Pan-bacterial Fitness Modules via Independent Component Analysis

The report warns that organisms with fewer than ~100 experiments produce weaker modules. Its example is Caulo, with 198 experiments and only 2.9x correlation enrichment. That example does not fall below the stated ~100-experiment threshold, so it is internally inconsistent as written and is not clear support for the threshold. ₃₉

A 40% component cap (components <= 40% of experiments) was needed to avoid FastICA convergence failures. It may cause some modules to be missed in low-experiment organisms. ₃₉

PFam-based annotations gave the best coverage, but they work at the domain level and may overcount functional associations. ₃₉

The module and annotation results depend on analysis choices. The initial D’Agostino K-squared membership approach gave oversized, weakly enriched modules. The reported improvements came only after the switch to absolute weight thresholds and a lower enrichment-overlap threshold. ₃₉

The cross-organism alignment used BBH ortholog pairs and ortholog groups. The 156 module families and their consensus labels therefore show aligned conservation patterns, not proof that every member has an identical molecular function. ₃₉

Functional Dark Matter — Experimentally Prioritized Novel Genetic Systems

Environmental metadata are sparse: AlphaEarth embeddings cover only 28% of genomes (83K/293K), and NCBI isolation-source metadata are inconsistent, limiting the power of biogeographic tests. NMDC validation is genus-level (mapping NMDC taxon columns to pangenome genera via ncbi_taxid) and may miss species-specific dark-gene signals, only 5 of 6 carrier genera were matched, and common genera such as Pseudomonas, Klebsiella, and Bacteroides may generate broad correlations with abiotic variables independent of specific dark-gene functions; the report says the high significance rate (76/105) likely reflects the dominance of such common genera in both datasets. ₄₀

The 57,011 dark-gene count likely overestimates true functional darkness because annotations may exist in databases or releases not checked, such as UniProt, InterPro or recent NCBI updates. Module predictions, which cover 6,142 dark genes in the integrated census and are reported as 6,691 in a later project-specific accounting, are guilt-by-association inferences rather than direct experimental validation; the “high confidence” label in prioritization reflects evidence convergence, not experimental proof. ₄₀

Fitness Browser condition coverage is uneven, and MR-1 has 121 historical conditions; organisms with deeper condition coverage can therefore produce more specific phenotypes and receive higher prioritization scores. The GapMind analysis is limited to amino-acid biosynthesis and carbon-utilization pathways, identifies organism-level co-occurrence rather than direct gene-to-step assignments, and does not address signaling, regulation, or structural functions comprehensively. Curated pathway–enzyme domain matching (EC prefix, Pfam family and keyword compatibility) partially addresses the gene-to-step gap, identifying 42,239 domain-compatible gene–pathway candidates including 5,398 high-confidence EC matches, but full gene-to-step validation would require AlphaFold structure prediction or experimental enzymology. ₄₀

Essential genes are penalized in fitness-centric scoring because they lack viable transposon mutants, no genefitness rows, and no differential fitness magnitudes; the separate essential-gene prioritization partly corrects this bias but relies on neighborhood and domain inference. None appear in the NB05 top 100 because dimension 1 (fitness importance) scores them at the essentiality bonus floor and dimension 6 (experimental tractability) penalizes them for being non-knockable. The neighborhood heuristic uses a five-gene window, same-strand orientation, and gaps of no more than 300 bp, whereas tools such as DOOR, STRING, and EFI-GNT use broader taxonomic and multimodal evidence. Cross-species synteny validation across 48 Fitness Browser genomes and co-fitness confirmation add to the positional heuristic, but the analysis still lacks gene-fusion detection, regulatory-element analysis and broad taxonomic sampling. ₄₀

The NMDC trait correlations are vulnerable to compositional coupling: 441/449 exploratory tests were significant despite only 7/7 pre-registered trait directions being the meaningful directional metric. A full sample-label permutation test remains future work requiring Spark access to raw per-sample matrices. Partial null approximations instead give a sign test of p = 0.0078 for the 7/7 pre-registered trait directions and p = 0.0625 for the 4/4 pre-registered abiotic directions, which the report says confirms non-random directionality; Fisher’s combined probability across all 11 pre-registered tests of p ≈ 0; and an inflation factor of ~20× for exploratory tests (441/449 significant versus 22.5 expected), which quantifies the compositional coupling. The report adds that these NMDC correlations demonstrate ecological co-occurrence rather than causal gene–phenotype links. Likewise, the 29/47 lab–field concordance rate has a one-sided binomial p = 0.072, and equivalent annotated-accessory-gene controls were not run through the complete biogeographic pipeline. ₄₀

The prioritization weights are expert-assigned and exact ranks are sensitive to them. Overall rank correlations remained ρ > 0.93 across six alternative configurations, but only 32/50 (64%) of the original fitness-active top 50 remained under conservation-dominant or drop-tractability settings, while essential-gene top-50 retention was 18/50 (36%) when tractability was dropped and 24/50 (48%) when neighbor context was dropped. Across all 6 fitness-active weight configurations, 18 genes remain in the top 50 and 35 in the top 100 regardless of weight choice; for essential genes, 6 remain always-top-50 and 19 always-top-100. The report calls these always-top candidates the most defensible targets for experimental follow-up. Robust rank indicators, namely rank-range distributions, a min/max rank scatter and always-top-50 candidates by organism, are plotted in fig29_robust_ranks.png. ₄₀

The Fitness Browser’s 77% Pseudomonadota composition (37/48 organisms) limits cross-phylum inference. The extended covering set, which incorporates 25 non-Fitness-Browser organisms from Bacillota, Actinomycetota and Campylobacterota, improves coverage from 4 to 6 phyla, but non-Fitness-Browser ortholog-group coverage is estimated at the genus level and may overestimate individual-organism coverage; these organisms also lack Fitness Browser condition profiles and therefore support broad screens rather than condition-specific experiments. Bacillota organisms (B. subtilis, S. aureus) were not selected by the covering-set algorithm because their ortholog groups are subsets of coverage provided by Pseudomonadota, despite their value for studying genes in native Gram-positive contexts. ₄₀

Metabolic Consistency of Pseudomonas FW300-N2E3 Across Four KBase Data Lakehouse Databases

• Only 21/58 WoM metabolites (36%) could be tested against any other database, and only 3 metabolites had four-way coverage. The untested 64% may contain additional discordances. ₄₁
• WoM exometabolomics was measured on R2A rich medium, whereas Fitness Browser fitness was measured on minimal medium with single carbon or nitrogen sources; condition-dependent metabolism limits direct comparison. ₄₁
• BacDive aggregates P. fluorescens strains and, under GTDB reclassification, spans a broad clade identified as Pseudomonas_E fluorescens_E. Per-strain consensus deduplication, a majority vote among duplicate records per strain, was applied before species-level utilization rates were computed. However, strain-level variation (e.g., trehalose: 1/6 positive) means the species consensus may not reflect FW300-N2E3’s specific capabilities. ₄₁
• The report states in its literature-context discussion that only tryptophan, with n=52, has sufficient data for a confident conclusion, while its primary tryptophan result reports 0+/50- with n=50. This difference is retained rather than reconciled because the document gives both values. ₄₁
• Manual name matching identified 28 WoM–Fitness Browser and 8 WoM–BacDive matches; two Fitness Browser matches were approximate base-to-nucleoside mappings. Additional matches may have been missed because of nomenclature differences, although excluding the two approximate matches changed mean concordance only from 0.937 to 0.930. ₄₁
• Fitness Browser coverage includes pleiotropic housekeeping genes: the top 18 genes were significant across all 21 conditions because of amino acid biosynthesis requirements rather than substrate-specific catabolism. Separating housekeeping and substrate-specific signals is needed to sharpen WoM–Fitness Browser integration. ₄₁
• The planned NB04 pathway-level analysis, which would map fitness-important genes to specific GapMind pathway steps, was deferred. This limits mechanistic interpretation of the tryptophan overflow hypothesis. ₄₁
• BacDive compound coverage varies from 1 to 51 strains, and raw records can be higher because of duplicate entries per strain; sample size must therefore accompany consensus utilization values. ₄₁

Gene Function Ecological Agora

Pilot scope, substrates and readiness

The Phase 1A pilot subset was statistically informative but biologically narrow: 1,000 species across 110 phyla and 1,200 UniRef50s in 6 classes. The report said Phase 1B at full GTDB scale (27,690 species, all UniRef50s) was needed for headline atlas verdicts. At the pilot close, the substantive hypotheses were untested: Bacteroidota PUL Innovator-Exchange, Mycobacteriota mycolic-acid Innovator-Isolated, Cyanobacteria PSII Broker, and Alm 2006 reproduction at higher resolutions all required the corresponding phase to run. The report also stated that the atlas was not yet a hypothesis-generating resource, because per-clade × function quadrant verdicts needed the full GTDB UniRef50 atlas of Phase 1B. The methodology revisions M1–M4 were pre-Phase-1B, so their effect was untested until Phase 1B execution. Phase 1 made no direction inference: it was acquisition-only by plan design, and direction at genus rank was reserved for the Phase 3 architectural deep-dive. ₄₂

The pilot drew on kbase_ke_pangenome tables. genome, gtdb_species_clade, gtdb_taxonomy_r214v1, gtdb_metadata and pangenome formed the core scaffold for species, taxonomy, quality and pangenome size. eggnog_mapper_annotations supplied KO, COG (clusters of orthologous groups), KEGG_Pathway, BRITE and PFAMs, the last serving as a domain-name fallback for control detection. interproscan_domains was the authoritative Pfam annotation source for accession-based control detection. The paralog fallback (n_gene_clusters when UniRef90 is absent) covered 21.5% of presence rows, and the report asked Phase 1B to report with-and-without-fallback sensitivity. CPR / DPANN under-representation is an acknowledged scope limit per RESEARCH_PLAN.md, which the report calls a substrate constraint rather than a Phase 1A failure. ₄₂

Phase 1A pilot data artifacts (file; rows; description):

• data/p1a_pilot_species.tsv: 1,000; phylum-stratified species sample with quality + annotation density.
• data/p1a_pilot_uniref50.tsv: 1,200; UniRef50 pilot pool with control_class + IPR/GO/pathway enrichment.
• data/p1a_pilot_extract.parquet: 6,638; (species, UniRef50) presence + paralog count.
• data/p1a_null_producer_lookup.parquet: 876; per-(rank, clade, prevalence-bin) cohort moments.
• data/p1a_null_consumer_lookup.parquet: 4,800; per-(rank, UniRef50) consumer z-score.
• data/p1a_uniref_prevalence_bin.tsv: 6,000; per-(rank, UniRef50) prevalence bin.
• data/p1a_pilot_scores.parquet: 9,201; per-(rank, clade, UniRef50) producer + consumer z-scores.
• data/p1a_control_validation.tsv: 30; per-(rank, control_class) score summary with pass/fail verdicts.
• data/p1a_alm_2006_pilot_backtest.tsv: 5; per-rank Alm 2006 TCS HK reproduction summary.
• data/p1a_extraction_log.json, data/p1a_null_diagnostics.json and data/p1a_pilot_atlas_diagnostics.json: 1 each; NB01, NB02 and NB03 audit logs.
• data/p1a_phase_gate_decision.json and data/p1a_phase_gate_summary.md: 1 each; the NB04 formal gate verdict and human-readable gate document. ₄₂

The pilot figure table describes p1a_null_producer_distribution.png as the producer null cohort distribution by prevalence bin (sanity check). It describes p1a_null_consumer_distribution.png as consumer null permutation distribution shapes (single-rank version, superseded) and p1a_paralog_count_distribution.png as the paralog count distribution by control class (sanity check). p1a_null_per_rank_distributions.png shows per-rank producer cohort sizes + consumer z-distributions (the multi-rank story). p1a_scores_by_class_per_rank.png holds violin plots of producer + consumer z by control class per rank (the headline distribution figure). figures/p1a_producer_consumer_per_rank.png is the Producer × Consumer scatter per rank, colored by control class (the atlas-style view). ₄₂

Two further computational caveats come from the Reproduction section. JupyterHub idle-timeout runs ~17–25 min, and the workaround is .ipynb → .py + nohup + intermediate parquets (used for NB10, 10b, 26). The driver heap ceiling defaults to -Xmx4g and is raised via environment for heavy coalesce writes. The report offers these as reusable patterns for future BERIL projects working at similar scale (1B-row table joins + 28M-row species-level data + 17M event attribution). ₄₂

Pilot producer-null calibration and the TCS HK power analysis

The pilot formulated a multi-rank Producer × Participation null-model framework for clade-level innovation atlases on UniRef50-scale substrates. It paired a clade-matched neutral-family producer null with a parent-rank dispersion permutation consumer null, and the report says both were formulated, calibrated and validated at pilot scale. The consumer null was later replaced by Sankoff parsimony, as described elsewhere on this page. The natural_expansion class served as the producer positive control: 200 UniRef50 clusters with documented within-species paralog count ≥ 3 and cross-species presence in ≥ 5 pilot species. It showed positive producer z-scores at all five ranks, with effect size growing monotonically with rank. Producer z mean (95% CI; n):

• genus: +0.13 σ [0.08, 0.18]; 975.
• family: +0.19 σ [0.13, 0.26]; 802.
• order: +0.31 σ [0.22, 0.39]; 654.
• class: +0.50 σ [0.38, 0.61]; 482.
• phylum: +0.55 σ [0.43, 0.67]; 457. The report reads this as validating that the clade-matched neutral-family null detects real paralog expansion above cohort baseline. Without that signal, all subsequent Phase 1A scoring would rest on a null model unable to discriminate signal from noise. ₄₂

Pilot paralog comparisons by rank and control class (n; observed mean paralogs; cohort mean; raw difference; percent difference; producer z):

• phylum natural_expansion: 457; 1.77; 1.27; +0.50; +39.5%; +0.55.
• phylum pos_tcs_hk: 186; 1.07; 1.30; −0.24; −18.4%; −0.23.
• phylum neg_ribosomal: 315; 1.04; 1.18; −0.13; −11.4%; −0.15.
• phylum neg_rnap_core: 284; 1.01; 1.20; −0.18; −15.2%; −0.24.
• phylum neg_trna_synth: 248; 1.02; 1.20; −0.18; −15.0%; −0.19.
• phylum pos_amr: 272; 1.17; 1.34; −0.17; −12.6%; −0.16.
• class natural_expansion: 482; 1.74; 1.28; +0.46; +36.0%; +0.50.
• order natural_expansion: 654; 1.56; 1.28; +0.28; +22.3%; +0.31.
• family natural_expansion: 802; 1.45; 1.27; +0.18; +14.3%; +0.19.
• genus natural_expansion: 975; 1.31; 1.20; +0.11; +9.3%; +0.13. ₄₂

At phylum rank, natural_expansion UniRefs had 39.5% more paralogs than the cohort baseline (1.77 vs 1.27 paralogs per phylum). The negative controls (ribosomal, tRNA-synth, RNAP) sat ~12–15% below cohort baseline, which the report calls the dosage-constrained signature. TCS HK at UniRef50 sat 18% below cohort, the opposite direction from the Alm 2006 family-level expansion claim. ₄₂

The report gives a power analysis for detecting a positive TCS HK effect at α=0.05 with a one-sided t-test (rank; n; minimum detectable d at 80% power; observed z; power if the Alm 2006 effect were d=0.3; power if d=0.5):

• genus: 73; 0.29; −0.10; 0.81; 0.99.
• family: 124; 0.23; −0.16; 0.95; 1.00.
• order: 157; 0.20; −0.19; 0.98; 1.00.
• class: 174; 0.19; −0.23; 0.99; 1.00.
• phylum: 186; 0.18; −0.23; 0.99; 1.00. ₄₂

On this analysis, HK paralog expansion at any biologically interesting effect size (d ≥ 0.3) at UniRef50 resolution would have been detected with 81–99% probability at all ranks. It was not: the observed z was consistently negative, indicating slight under-expansion. The report says this rules out underpowering as the explanation. In its account, TCS HK paralog signal is genuinely absent (and slightly negative) at UniRef50, the original Alm 2006 effect lives at family resolution where a single HK family aggregates many UniRef50s, and the substrate-hierarchy argument is the right interpretation. That resolution-dependence is the report’s interpretation of a null result. ₄₂

Pre-registration discipline and planned Phase 1B additions

The adversarial reviewer noted that revising the negative-control criterion (“near zero” → ”≤ 0 with CI not strongly positive”) after seeing data weakens pre-registration discipline, and the report calls this a fair pre-registration concern. Its framing is that the dosage-constraint biology on housekeeping genes (Andersson 2009; Bratlie et al 2010 for ribosomal proteins) was anticipated during plan v2 design, but its quantitative consequence (negative producer z, not zero) was not encoded into the pre-registered criterion. On that account M2 corrects a pre-registration omission rather than redefining a target. The prior expectation of no paralog expansion is preserved and tightened to “should not show paralog expansion AND should show dosage-constraint suppression”. The report says a future Phase 1B / Phase 2 plan should pre-register this corrected expectation, not the original one. The M2 revision remains a data-driven, post-hoc change. ₄₂

The report called the reviewer’s C1 (“no concrete multiple-testing strategy”) partially overcalled, because plan v2.1 has a hierarchical Tier-1 / Tier-2 / Tier-3 strategy. It accepted that implementation details remained TBD at that stage. These were the BH-FDR (Benjamini–Hochberg false discovery rate) variant and the effective-N for KEGG-BRITE × GTDB-family clusters, which it called a real Phase 1B/2 deliverable. Phase 1B NB02 was assigned to specify the BH-FDR variant + effective-N within KEGG BRITE × GTDB family. The report called the “render meaningful discoveries impossible” framing hyperbole given the strategy. ₄₂

Two additions were planned for Phase 1B. The first was a known-HGT positive control set for the consumer null (addressing C2). AMR was the closest current control, but the parent-phylum anchor masks intra-phylum HGT (M1). Phase 1B therefore added specific β-lactamase families with documented cross-phylum spread (CARD bla group) and class-I CRISPR-Cas systems (Pfam family), per Metcalf et al 2014’s documented cross-tree-of-life HGT. The second promoted PIC (phylogenetic independent contrasts) from an optional Phase 2 sensitivity in plan v2 to mandatory at Phase 1B for consumer-null and producer-null score reporting (addressing I1). This plan is in tension with the final review closure recorded below, which lists I3 as an open limit because no PIC correction was implemented. The report does not reconcile the two. ₄₂

The pilot’s next-step plan named the later tests as pre-registered hypotheses, not results:

• Phase 1B: Bacteroidota → Innovator-Exchange on PUL CAZymes (deep-rank category).
• Reproduce Alm 2006 TCS HK paralog expansion at KO level (one of two canonical reproductions per the v2 plan).
• Phase 2: Mycobacteriota → Innovator-Isolated on mycolic-acid pathway.
• Phase 3: Cyanobacteria → Broker on PSII architectures (genus rank, with donor inference). The NB16 test recorded below was instead run donor-undistinguished, with support at class rank. ₄₂

Phase 1B full-scale results

At this report stage Phase 1B was complete with gate verdict PASS_REFRAMED, and Phases 2–4 were in planning. Phase 1B applied methodology revisions M1–M4 at full GTDB scale: 18,989 bacterial GTDB representatives × 100,192 targeted UniRef50s (10K-per-class cap) → 1.54 M (species, UniRef50) presence rows → 1.29 M (rank, clade, UniRef) producer scores. Wall time was ≈ 1 hour total across NB05–NB08. The 1.54 M presence-row count differs from the 28M (species, UniRef50) presence rows listed for data/p1b_full_extract_local.parquet in the data inventory, and the report does not reconcile the two. ₄₂

The producer null was responsive at GTDB scale. The natural_expansion class (UniRefs with documented within-species paralog count ≥ 3 in ≥ 5 species) showed monotone-growing positive producer z across all 5 ranks, which the report calls stronger than at pilot scale. Producer z mean; raw paralog above cohort:

• genus: +0.13 σ; +9.3 %.
• family: +0.19 σ; +14.3 %.
• order: +0.31 σ; +22.3 %.
• class: +0.77 σ; +55.2 %.
• phylum: +0.89 σ; +64.5 %. Negative controls (ribosomal / tRNA-synth / RNAP core) sat ~12 % below cohort with z ≈ −0.15, the dosage-constrained signature M2 anticipates (Notebooks 06, 07). ₄₂

The M1 fix stratified the consumer-null parent rank per child rank rather than always using phylum. It produced the rank gradient that Phase 1A’s parent-phylum-only anchor had masked (child rank; M1 parent rank; informative consumer z mean):

• genus: family; −10.48.
• family: order; −6.40.
• order: class; −4.18.
• class: phylum; −1.84. The Phase 1A parent-phylum anchor had produced a flat-then-jump pattern (≈−4 σ at deep ranks, −1.36 at class), whereas M1 shows the gradient is monotone (Notebook 06). This parent-rank dispersion metric was later replaced by Sankoff parsimony (M14). ₄₂

The pre-registered Phase 1B hypothesis was that Bacteroidota is Innovator-Exchange (high producer + high participation) for PUL CAZyme UniRef50s. Innovator-Exchange was operationalised as “both producer and consumer 95 % CI lower bound > 0”. By that criterion the hypothesis was falsified at all 4 deep ranks (0/4 supported). The report calls the absolute-zero criterion over-stringent at UniRef50 resolution. A relative threshold revealed a graded HGT signal: at family→order parent rank, CAZymes (hyp_cazyme) showed +0.78 σ less clumping than ribosomal proteins (median consumer z = −6.21 vs −6.99; Mann-Whitney U one-sided p = 1 × 10⁻⁴³). Bacteroidota PUL CAZymes specifically showed producer z ∈ [−0.07, −0.10] across deep ranks (95 % CIs entirely below zero) and consumer z ∈ [−9.3, −2.8] at parent ranks. Both legs of Innovator-Exchange therefore failed by the absolute criterion. By the relative criterion, CAZymes still discriminated from housekeeping genes, but Bacteroidota CAZymes showed no additional lift relative to other phyla’s CAZymes. figures/p1b_bacteroidota_pul_position.png shows Bacteroidota PUL position against other phyla’s CAZymes per rank. This parent-rank relative diagnostic is separate from the later Sankoff relative signal (Cohen’s d = 0.15) recorded under the hypothesis verdicts. ₄₂

The pre-registered cross-phylum HGT positive controls (β-lactamase + class-I CRISPR-Cas, plan v2.3 HIGH 1) all sat at strongly negative absolute consumer z, meaning they were clumped at parent ranks. A relative-threshold diagnostic at family rank gave a different picture (class; median consumer z; Δ vs neg_ribosomal; one-sided MW p):

• pos_amr: −7.37; −0.37; 0.94.
• pos_tcs_hk: −6.57; +0.43; 5×10⁻⁴.
• pos_betalac: −6.37; +0.62; 5×10⁻⁸.
• pos_crispr_cas: −5.81; +1.18; 1×10⁻⁸⁰. β-lactamase and class-I CRISPR-Cas discriminated from housekeeping genes by 0.6–1.2 σ. The AMR control (the bakta_amr set, dominated by clade-specific resistance variants) did not, which the report attributes to bakta_amr’s inherent narrowness rather than to methodology failure. The report says the gate document’s “HIGH 1 known-HGT controls fail” framing overstates: the controls fail the absolute Innovator-Exchange threshold but pass the relative-discrimination diagnostic. Its Phase 2 framing was that HGT signal is real at UniRef50 but small in magnitude, and KO aggregation was expected, not yet shown, to amplify it. The diagnostic (Notebook 08b) was captured at user request after the original Phase 1B narrative proved over-pessimistic. ₄₂

The report calls one diagnostic pattern genuinely unexpected. At order rank (parent = class), positive HGT controls had median consumer z = −4.51 versus −3.69 for negative controls. Positive controls were thus more clumped than negatives, the opposite of expectation (Mann-Whitney U one-sided p = 1.0, no discrimination in the expected direction). One proposed explanation was that order rank has only 689 clades and 245 parent classes, and that this small parent-clade count may inflate null variance for the consumer-null permutation; this was a possibility, not a demonstrated cause. The report records the substrate-hierarchy claim as a partial confirmation. A tree-aware metric (Sankoff parsimony, NB08c, which the report calls the canonical Alm-2006-style approach) recovered the expected direction (positive HGT > negative housekeeping at p = 2.1×10⁻⁵), and the report treats the order-rank anomaly as a metric artifact. However, 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 to validate the substrate-hierarchy reasoning; otherwise an M11 reconciliation-based redesign (gene-tree vs species-tree, AleRax sub-sampling) would trigger. The later M18 outcome is recorded under Methodological contributions. ₄₂

Phase 1B diagnostic resolution (NB08b/NB08c) and the first Phase 2 KO gate

At its close Phase 1B carried the verdict PASS_REFRAMED + qualified. The report judged the methodology not broken, but the substrate-hierarchy framing became an empirical bet on Phase 2 amplification rather than an established conclusion, and the project entered Phase 2 with a sharp falsifiable test. The closure’s methodology summary reported three results. Multi-rank null-model construction works at full GTDB scale. M1 rank-stratified parent ranks give a clean monotone clumping gradient. The producer null is responsive: the natural_expansion class showed +64.5 % paralog above cohort at phylum, p ≈ 10⁻⁸⁰. Full-scale per-rank null distributions and per-class scores are shown in figures/p1b_null_per_rank_distributions.png and figures/p1b_scores_by_class_per_rank.png. ₄₂

Reviewer item I8 set this +64.5% against a recomputed +39.4%. The report calls that a phase-mixing error. The +39.4% figure uses the Phase 1A pilot means (obs_mean = 1.77, cohort_mean = 1.27), whereas +64.5% is the Phase 1B full-scale figure (obs_mean = 2.09, cohort_mean = 1.27). Both are correctly computed at their own scale. The report reads the pilot-to-full-scale increase, which it writes as +39.5% → +64.5%, as methodology validation. The reviewer’s +39.4% and the report’s +39.5% for the same pilot comparison are both recorded as written. ₄₂

The Phase 1B closure stated that the pre-registered hypothesis (Bacteroidota → Innovator-Exchange on PUL CAZymes) was falsified, failing at all 4 deep ranks under the absolute-zero criterion. The NB08b relative-threshold diagnostic showed that CAZymes do discriminate from housekeeping (+0.78 σ at family rank, p = 1×10⁻⁴³). In Cohen’s d terms, however, the discrimination is small (≈ 0.07–0.09), below biologically meaningful effect sizes. The report states that the hypothesis is genuinely not supported at UniRef50 resolution and that this is not a methodology save. It adds that the PUL hypothesis remains falsified by either criterion, because the relative-threshold framing changes only how control-class behaviour is described, not the hypothesis-test outcome. This Phase 1B wording differs from the final synthesis’s “qualified pass” label recorded under the hypothesis verdicts. ₄₂

A reviewer cited literature on rapid CAZyme HGT (horizontal gene transfer). The report replies that this counter-evidence describes transfer at the PUL/family aggregation level, not at the UniRef50 sequence-cluster level. On its reading, the UniRef50-level falsification is consistent with that literature: family-level CAZyme HGT exists but operates on aggregated functional units, so the two differ in substrate resolution rather than necessarily in biology. ₄₂

NB08b’s broader reading was that HGT signal is detectable at UniRef50. CAZymes, β-lactamases and class-I CRISPR-Cas all showed +0.6 to +1.2 σ less clumping than housekeeping at family rank, with p << 10⁻⁷. The signal magnitude is small, well below the absolute Innovator-Exchange threshold that the Phase 1B hypothesis pre-registered. The report kept the substrate-hierarchy reasoning in a softer form. It expected KO aggregation to produce much larger effect sizes by collapsing sequence-cluster-specific clade restriction. If amplification failed, the consumer-null methodology itself might need replacing (M11: direct phyletic-incongruence on KO presence/absence instead of parent-rank dispersion). ₄₂

The page records above one explanation for the order-rank inversion: the small number of parent clades. The report also proposed a second possibility. Intra-phylum HGT-active genes may cluster within a few orders inside a single phylum, which would make them more clumped at parent = class than housekeeping genes, which spread to many classes. The report offered this as a possible explanation and did not test it. ₄₂

Revision M12 (new at plan v2.4) made HGT-class-versus-housekeeping deltas a primary atlas metric alongside absolute consumer z. At deep ranks it operationalised Innovator-Exchange as ”≥ X σ less clumped than housekeeping at the rank’s parent” rather than “absolute consumer z > 0”. X was to be calibrated empirically from Phase 2 data. ₄₂

The NB08c diagnostic followed a close-reading of Alm, Huang & Arkin 2006 (docs/alm_2006_methodology_comparison.md), which the report says revealed three load-bearing misreadings:

• Alm 2006 never named “producer/consumer” or four-quadrant categories. It reported a correlation r = 0.74 between HPK count and recent-LSE fraction. The four-quadrant Open/Broker/Sink/Closed framework is the project’s own construction.
• It worked at the single-domain level (IPR005467 / COG4582), which is comparable to UniRef50, not at “family aggregation”.
• It used phylogenetic-tree-aware reconciliation, not parent-rank dispersion. The substrate-hierarchy claim was therefore reformulated from “UniRef50 is too narrow, aggregate to KO” to “we used the wrong metric on the right substrate”. M14 reframed the project’s relationship to Alm 2006 from “generalizing” to “Alm-2006-inspired”, and M15 made Sankoff parsimony (a lightweight tree-aware metric) mandatory. The report’s lesson is that anchor papers must be close-read for exact methodology before infrastructure is built on extrapolations from them. This correction refines the pilot TCS HK reasoning recorded above, which had placed the Alm 2006 effect at family resolution. ₄₂

The report documents the anchor’s prominence in its rebuttal of REVIEW_4 C1, which claimed that Alm 2006 was uncited. Alm 2006 is the first reference in references.md (line 9, “Methodological anchor” section) and is referenced ≥30 times across DESIGN_NOTES.md. It has a dedicated 460-line memo at docs/alm_2006_methodology_comparison.md, linked from the README. It is also cited in REPORT.md’s Phase 1B Diagnostic Resolution section (v1.3) under M14. ₄₂

Diagnostic NB08c ran sankoff parsimony on the GTDB-r214 tree. It tested whether the order-rank anomaly was a metric artifact of parent-rank dispersion or a methodology failure. The result was a metric artifact: Sankoff recovers the expected direction, with positive HGT above negative housekeeping at p = 2.1×10⁻⁵. Headline panel C plots Sankoff parsimony score / n_present_leaves per UniRef. In it, positive HGT classes (TCS HK, β-lactamase, CAZymes) sit above housekeeping classes (ribosomal, tRNA-synth, RNAP), with Cohen’s d = 0.15 (small effect; p = 2.1×10⁻⁵). The four panels are in figures/p1b_metric_diagnostic_panels.png. ₄₂

The other NB08c panels were diagnostic:

• Panel A shows K = n_clades_with at order rank per class. Both positive HGT and negative housekeeping classes have median K = 1, so UniRefs are clade-restricted regardless of class. The distributions differ statistically (p < 10⁻¹⁹), but the medians are the same, so K bias does not explain the order-rank anomaly.
• Panel B shows within-pseudomonadota dispersion (n_pseudo_families / n_pseudo_species). Positive controls have lower dispersion than negative housekeeping (pos median 0.25, neg median 0.33). The report judged the metric biased by class coverage and dropped it. Its reasoning was that housekeeping genes are pan-Pseudomonadota, whereas HGT-active classes are clade-specific within the phylum.
• Panel D shows the Pseudomonadota fraction per UniRef. Only AMR is strongly Pseudomonadota-biased (median 50 %; 49 % of AMR UniRefs >50 % Pseudomonadota), and all other classes are ≤ 4 % median. The report calls this an AMR detection-bias confound, because bakta_amr is sourced from the Pseudomonadota-heavy reference set of amrfinderplus. ₄₂

NB08c median Sankoff/n_present by class, with the report’s annotation:

• none (random atlas baseline): 24.5; top, because random sparse-presence UniRefs require many gains.
• pos_tcs_hk: 19.75; above housekeeping.
• pos_betalac: 19.0; above housekeeping.
• hyp_cazyme: 19.0; above housekeeping.
• neg_rnap_core: 17.5; an inversion, with RNAP ranking above CRISPR-Cas (small).
• pos_crispr_cas: 17.0.
• neg_trna_synth: 14.0.
• neg_ribosomal: 12.8; correctly the lowest housekeeping class.
• pos_amr: 12.0; anomalously low, which the report attributes to Pseudomonadota detection bias.
• natural_expansion: 4.5; correctly the lowest (broad conservation by selection). ₄₂

The report drew four conclusions from NB08c:

• The effect size at UniRef50 remains small. Cohen’s d = 0.15 is well below the d ≥ 0.3 threshold pre-registered for atlas-grade discrimination. The “+0.6 to +1.2 σ less clumped than housekeeping” framing in REPORT.md v1.2 was in parent-rank z-units, not Cohen’s d, and overstated the effect magnitude.
• The softened substrate-hierarchy reading still holds: HGT signal at UniRef50 is real but small. KO amplification is now an empirical claim that Phase 2 must demonstrate, not a substrate-hierarchy guarantee.
• The within-phylum dispersion metric (Diagnostic B) was wrong, because it tracked class coverage rather than HGT activity. The right within-phylum analog is per-phylum Sankoff parsimony.
• AMR is a confounded positive control and should be excluded from the Phase 2 positive-control panel, leaving β-lactamase, CRISPR-Cas and TCS HK. This Pseudomonadota-detection explanation for AMR differs from the Phase 1B relative-threshold passage above, which attributed AMR’s failure to the narrowness of bakta_amr. The report does not reconcile the two. ₄₂

Two revisions followed. M16 made Sankoff parsimony on the GTDB-r214 tree topology the primary atlas metric for Phase 2 and demoted parent-rank dispersion to diagnostic only. As noted elsewhere on this page, other passages credit this replacement to M14 or M15. M18 (plan v2.6) set a hard amplification gate: the Phase 2 KO atlas had to show Cohen’s d ≥ 0.3 on Sankoff parsimony for at least one positive HGT control class versus negative housekeeping. The UniRef50 baseline is stated as d = 0.15 in one passage and as d = 0.146 from NB08c in another. If the gate failed, the substrate-hierarchy claim would be falsified and the M11 redesign would trigger. That redesign is a switch from Sankoff to gene-tree-versus-species-tree reconciliation, per Phase 3’s original sub-sample reservation, with AleRax sub-sampling named as the route. The report calls this a falsification gate, not a validated prediction. ₄₂

The report records several lessons as pre-registration omissions:

• Negative-control criterion. The v2 criterion specified “producer mean within ±1σ of zero” for housekeeping classes. The pilot showed all three negative controls at producer z ≈ −0.15 (CIs entirely below zero), and the criterion was revised post-hoc to “CI upper bound ≤ 0.5”. The lesson drawn is that a known biological prior should be encoded in the criterion text.
• Absolute-zero Innovator-Exchange definition. Plan v2 defined Innovator-Exchange as “both producer and consumer 95 % CI lower bound > 0”. All positive controls (β-lactamase, CRISPR-Cas, AMR, TCS HK) sat at strongly negative consumer z and failed it. The report says a post-gate diagnostic (NB08b) nonetheless showed positive HGT classes +0.6–1.2 σ less clumped than housekeeping at family rank (p << 10⁻⁷). By the family-rank table above, that range holds for β-lactamase and class-I CRISPR-Cas; TCS HK separated by only +0.43 (p = 5×10⁻⁴), and AMR did not separate in the expected direction (−0.37; p = 0.94). M12 recalibrated the criterion against an empirical housekeeping baseline, post hoc.
• Origin of the revisions. The report says most revisions originated in genuine biology engagement (M2 dosage cost, M14 Alm 2006 close-reading, M17 Pseudomonadota AMR bias) rather than data accommodation. It concedes that the cumulative count shows pre-registration was insufficiently calibrated against substrate-resolution behavior.
• Corrective discipline. Before invoking substrate, methodology or “we need more data” explanations, run a focused diagnostic first. NB04b, NB08b and NB08c together took ~3 hours to write and ran in seconds-to-minutes. They dispositively distinguished metric/criterion problems from methodology failures. ₄₂

The first Phase 2 step pulled KO assignments for Phase 1B’s 18,989 species from kbase_ke_pangenome.eggnog_mapper_annotations.KEGG_ko in the KBase Data Lakehouse:

• 103,629,867 gene clusters across the 18,989 species.
• 74,450,297 eggNOG-annotated clusters (71.8% annotation rate).
• 43,803,890 (gene_cluster, KO) edges after parsing the comma- / ko:-prefixed KEGG_ko field.
• 13,062 unique KOs.
• 28,008,764 (species, KO) presence rows after aggregation, with paralog count = distinct UniRef90s per (species, KO).
• 3,592,556 UniRef50s with a dominant KO (median dominant-fraction = 1.0; q25 = 0.97, which the report calls a clean projection).
• 748 control-panel KOs by description-match: 310 TCS HK, 110 β-lactamase, 6 CRISPR-Cas, 249 ribosomal, 54 tRNA-synth and 19 RNAP core. The full assignment parquet (1.4 GB) went to MinIO, and a 2.69 MB control-panel subset (p2_ko_assignments_panel.parquet) was kept locally. On this JupyterHub setup Spark Connect does not share a filesystem between workers and driver, so atlas-scale parquets are MinIO-only. ₄₂

The first-pass M18 test is shown in figures/p2_m18_amplification_panel.png. Sankoff parsimony was scored on the GTDB-r214 species tree, pruned to 18,989 P1B leaves, for all 748 panel KOs in 11 seconds. Pairwise Cohen’s d came with what the report labels M19 cluster-bootstrap CIs, implemented as B = 200 with KO-level resampling within class. That implementation differs from M19 as bound in plan v2.7, which specified cluster-bootstrap on UniRefs at Pfam-family granularity (B = 200 resamples per function class); the report does not reconcile the two. The result was 0/9 pairs PASS on the d ≥ 0.3 + CI lower-bound > 0 threshold. The best pair reported, pos_betalac vs neg_trna_synth, had d = +0.2118 (95% CI [−0.138, +0.531]) and a one-sided (greater) Mann–Whitney p = 1.1×10⁻⁵. The signed direction reversed against expectation for the ribosomal and RNAP core comparisons (d = −0.66 to −2.97). ₄₂

First-pass class medians (n KOs; median Sankoff/n_present; median n_present_leaves):

• neg_rnap_core: 19; 14.50; —.
• neg_ribosomal: 249; 11.35; ~29.
• pos_betalac: 110; 2.64; 390.
• pos_crispr_cas: 6; 2.04; 1499.
• pos_tcs_hk: 310; 1.47; 918.
• neg_trna_synth: 54; 0.30; —. ₄₂

The report calls the neg_ribosomal median n_present ≈ 29 leaves (out of 18,989) the smoking gun, because genuine universal r-proteins should be present in essentially all species. The description-match rule LOWER(ipr_desc) LIKE '%ribosomal protein%' had filled the class mostly with clade-restricted accessory r-proteins. These included RimM, RbfA, processing factors, silencing factors and false matches to mitochondrial ribosomal proteins. This contamination motivated the strict-class retest shown in figures/p2_m18c_strict_class_panel.png, whose PASS outcome under M21 strict housekeeping is recorded under Methodological contributions. ₄₂

Null-model and control diagnostics

The Phase 1A pilot null producer distribution (figures/p1a_null_producer_distribution.png, notebook 02_p1a_null_model_construction.ipynb) visualizes the producer-z null. That null is a per-rank, prevalence-bin, clade-matched neutral-family permutation null (M5/M6). Cohort moments, the mean and standard deviation of paralog count, were computed per (rank, prevalence_bin) from neutral-family UniRef50s in the same cohort. All bins have ≥30 cohort members at all 5 ranks, which the report takes as validating that producer-z scoring against this null is statistically tractable. ₄₂

The Phase 1A pilot null consumer distribution (figures/p1a_null_consumer_distribution.png, same notebook) shows the single-rank consumer-z permutation null on UniRef50 prevalence patterns. It has a unimodal null shape per rank, with distinct shapes per rank, which the report reads as making consumer-z comparable across ranks. The figure is kept for provenance only. Its description says it was superseded by the per-rank null in NB02 Phase 1A v1.1 plus the Phase 1B M16 Sankoff diagnostic. The same description also says it was superseded by the NB08c Sankoff parsimony diagnostic (M14), which uses tree-aware reconciliation. The source names M16 in one clause and M14 in the other, and the Phase 1A figure list later on this page credits M14. ₄₂

The Phase 1A pilot paralog count distribution by class (figures/p1a_paralog_count_distribution.png, notebook 01_p1a_pilot_data_extraction.ipynb) is a sanity check run before scoring, on the 1K species × 1.2K UniRef50 pilot. natural_expansion showed the wide distribution expected of paralog-rich families. The positive controls (AMR, CRISPR-Cas) showed modest paralog inflation. The negative controls (ribosomal, tRNA-synthetase, RNAP core) clustered tightly at 1 paralog, consistent with single-copy housekeeping. ₄₂

Ribosomal proteins, tRNA synthetases and RNAP core subunits all showed negative producer z (−0.15 to −0.24 σ across ranks; 95% CIs entirely below zero), so the pre-registered “near zero” criterion failed. The report calls the result biologically correct, because dosage-constrained genes carry fewer paralogs than typical genes at matched prevalence (Andersson 2009; Bratlie et al 2010 for ribosomal proteins). It treats this as methodology revision M2, not a methodology failure: the producer null distinguishes dosage-constrained genes from typical paralog-expansion patterns. The dosage explanation is the report’s literature-based interpretation. ₄₂

In the pilot, the consumer z-score compared parent-phylum dispersion with a permutation null, for UniRef50s in ≥ 3 clades (the informative subset). Mean consumer z by rank (n_informative; report’s interpretation):

• genus: −3.77 (423); strong vertical inheritance.
• family: −3.95 (329); strong vertical inheritance.
• order: −4.10 (267); peak vertical clumping.
• class: −1.36 (172); clumping weakens. ₄₂

The report calls this the first interpretable biological pattern in the atlas. At class rank, UniRef50 dispersion approaches the random-permutation null. The report reads this as consistent with cross-class HGT becoming detectable above noise; it is consistency, not direct evidence of transfer. From genus through order, vertical inheritance dominates. The report aligns the pattern with literature that most prokaryotic HGT happens within phylum boundaries (Smillie et al 2011; Soucy et al 2015 review) and that cross-phylum HGT is rarer and biased toward specific function classes (Hooper et al 2007). That alignment is literature context, not an additional pilot measurement. ₄₂

AMR showed strongly negative consumer z (−4.4 to −4.8) under parent-phylum dispersion at genus, family and order ranks. The report does not read this as absence of HGT. AMR is a documented intra-phylum HGT phenomenon (Forsberg et al 2012 for the soil resistome; Smillie et al 2011 for the gut microbiome). In the report’s account it is intense within pseudomonadota or Bacillota and rare across phyla. A null that measures cross-phylum dispersion therefore makes intra-phylum HGT look clumped, so the parent-phylum anchor masks the signal. The report cites Forsberg 2012 as context for the AMR intra-phylum HGT pattern that motivated M1. ₄₂

The pilot TCS HK producer back-test was null. Two-component-system histidine kinase UniRef50 clusters had mean producer z ≈ −0.2 across all ranks, with only 4–11% above zero (rank: UniRef50s scored; positive z; above 2σ; mean producer z):

• genus: 73; 8 (11%); 2; −0.10.
• family: 124; 8 (6%); 2; −0.16.
• order: 157; 7 (4%); 3; −0.19.
• class: 174; 9 (5%); 2; −0.23.
• phylum: 186; 9 (5%); 2; −0.23. ₄₂

The report calls this a negative result for Alm 2006 reproduction at sequence-cluster resolution, not a methodology failure. Alm, Huang & Arkin (2006) measured paralog expansion at the HK family level, as the number of distinct HK genes per genome. A UniRef50 typically corresponds to a single HK protein variant and cannot aggregate to the family. The report says the pilot validates the plan’s substrate hierarchy. Under that hierarchy, Alm 2006 reproduction was pre-registered at Phase 2 (KO-level aggregation; KEGG ko02020) and Phase 3 (Pfam multidomain architecture census, the resolution Alm 2006 actually used). The null rules out the simpler reading that any sequence-cluster-level methodology will reproduce Alm 2006. Phase 1A thus used Alm 2006 as a back-test target deferred to Phase 2/3, and Smillie 2011 as literature context for the cross-rank consumer-z trend. ₄₂

A mid-pilot audit found that kbase_ke_pangenome.interproscan_domains had not been used in v1. The table holds 146 M Pfam hits across 132.5 M cluster representatives (83.8 % cluster coverage). The data-sources section below gives InterProScan Pfam as 833M hits at the same 83.8% cluster coverage, and the report does not reconcile the two hit counts. v1 control detection had relied on eggnog_mapper_annotations.PFAMs, which stores domain names (HisKA) rather than accessions (PF00512). That made it unsuitable as the sole accession-based control substrate. ₄₂

Taking the union with InterProScan accessions increased the detected controls (v1 eggNOG name match → v2 union; gain):

• Ribosomal proteins: 9,005 → 19,389 UniRef50s; 2.2×.
• tRNA synthetases: 5,640 → 8,514; 1.5×.
• TCS HKs: 38,488 → 43,217; 1.1×. ₄₂

At the biological level, 4 of 5 controls reached 100% pool coverage across pilot species. Only AMR failed the 80% threshold, at 68.8%, which the report calls biologically expected because environmental and uncultivated lineages carry less AMR. The v1 implementation went through 8 iterations of bug-fixing before the substrate audit. Results stabilized once the pipeline switched to InterProScan. The report draws the lesson to audit substrate before implementing, not after debugging. ₄₂

The Phase 1A gate recorded four revisions:

• M1 made the consumer-null parent ranks rank-stratified (genus → family parent, family → order parent, order → class parent, class → phylum parent), so that the null is sensitive to HGT events at the rank where they occur.
• M2 set the negative-control criterion as CI upper ≤ 0.5, not “near zero”.
• M3 confirmed that Alm 2006 reproduction was deferred to Phase 2/3.
• M4 accepted the paralog fallback (option a), with sensitivity reporting. ₄₂

The M2 gate threshold is more specific than the pilot prose, which gives the corrected criterion as ”≤ 0 with CI not strongly positive”. The report does not reconcile the two wordings. With M1–M4 applied, Phase 1B could proceed at full GTDB scale (27,690 species, all UniRef50s). That species count matches the core-scaffold total under data sources, whereas the atlas itself spans 18,989 species representatives. ₄₂

Other pilot figures show per-rank score distributions by control class (figures/p1a_scores_by_class_per_rank.png) and per-rank null-model distributions (figures/p1a_null_per_rank_distributions.png). ₄₂

Run-time and computational caveats

End-to-end, the project needs ~6–8 hours of cold-start wall time (~3h Spark-side + ~5h pandas-side). With cached intermediate parquets, the synthesis pass alone (NB28a → 28f) takes ~5 minutes. Three of the five documented performance caveats are:

• The Spark-Connect driver result-size cap is 1 GB serialized. The workaround is MinIO staging (coalesce.write.parquet → read locally).
• The pandas spatial merge runs out of memory at ~10M+ rows. This README caveat places the failure at NB26h (Bacteroidota PUL gene-neighborhood, 723K × 210K contigs) and proposes batched processing or a full-Spark Stage 6. The run log described under Caveats gives a different account: it attributes the full-scale 723K × 210K failure to the Spark driver cap and the pandas failure to a sampled run. The report does not reconcile the two accounts.
• Setting Spark autoBroadcastJoinThreshold = -1 is harmful: NB10 hung 17+ min until the setting was removed. ₄₂

NB30 test families and revision bookkeeping

The NB30 Family 3 rows give residualized effects for the focal clades. These are mycobacteriaceae (family) and Mycobacteriales (order) in the NB12 mycolic analysis, and cyanobacteriia (class) in the NB16 PSII analysis (test; d_residualized; p_residualized; status):

• NB12 Mycobacteriaceae × mycolic family consumer_z: −0.16; 1.6×10⁻⁵; survives.
• NB12 Mycobacteriales × mycolic order producer_z: +0.29; 5×10⁻⁶; survives.
• NB16 Cyanobacteriia × PSII producer_z (class rank): +1.50; 2.1×10⁻⁵; survives.
• NB16 Cyanobacteriia × PSII consumer_z (class rank): +0.63; 2.3×10⁻⁵; survives. ₄₂

Five of the six Family 3 tests survive. The non-surviving one is the NB11 regulatory-versus-metabolic null, which the project explicitly reframed as a small effect in the complexity-hypothesis direction. The report states that this test is not a false positive. ₄₂

NB30 survival by test family (tests; surviving Bonferroni; share):

• Family 1, pre-registered hypotheses: 4; 4; 100%.
• Family 2, P4-D1 biome enrichment: 6; 5; 83%.
• Family 3, P4-D5 residualization replication: 6; 5; 83%.
• Total formal tests: 16; 14; 88%. ₄₂

The report states that the actual hypothesis-test family is the 4 pre-registered hypotheses (H1–H4). Biome enrichment and residualization replication are supporting test families. A family-wise error rate (FWER) correction across the 16 formal tests in these 3 families leaves 14 of 16 surviving Bonferroni at family-specific α. The 2 non-surviving tests are correctly null results. ₄₂

The report rejects the reviewer’s treatment of M1–M26 as 26 independent hypothesis tests on the same data. In its account, most M-numbers are pre-registration corrections documented for transparency: M1 positive controls, M2 dosage biology, M3 substrate hierarchy validation, M4 power analysis, M11 reconciliation contingency, M12 absolute-zero criterion, M13 cohort threshold tightening, M14 misreading-of-Alm-2006 correction, M16 Sankoff-parsimony metric, M19/M20 review-driven analyses, M22 acquisition-depth attribution, M23 minimum n_neg, M24 canonical effect-size reporting, M25 composition-donor-inference deferral and M26 tree-based donor inference. Three are post-hoc metric corrections on already-collected data: M14 Sankoff replacing parent-rank dispersion, M21 strict-housekeeping and M23 n_neg ≥ 20. Each was triggered by a specific failed diagnostic and is reported with that diagnostic. Earlier passages count 25 revisions (M1–M25). This is not a conflict: the report says its status updates layer chronologically, and M26 tree-based donor inference was added later, at the synthesis stage. Two descriptions do conflict, and the report does not reconcile them. M14 and M23 appear in both the pre-registration-correction list and the post-hoc list. M1 is described here as positive controls, but in the Phase 1A gate table as rank-stratified parent ranks. ₄₂

Two synthesis layers are descriptive classifications outside the test families, so no multiple-testing correction applies to them. Figure S7 shows distribution shapes (focal versus atlas reference), and its medians (mycolic 0.15, PSII 0.88, PUL 0.41) are reported as observations, not p-values. The per-(genus × KO) Open/Broker/Sink/Closed assignment uses parsimony rules plus a minimum-event-count threshold (≥5). It is not a null-hypothesis test, and plan v2.16 makes it reportable as an exploratory layer. ₄₂

The report’s methods references include four tool and dataset papers:

• Parks et al. (2022) for gtdb, a genome-based bacterial and archaeal taxonomy: “GTDB: an ongoing census of bacterial and archaeal diversity through a phylogenetically consistent, rank normalized and complete genome-based taxonomy”, Nucleic Acids Research 50(D1):D785–D794, doi:10.1093/nar/gkab776.
• Jones et al. (2014) for interproscan 5, a genome-scale protein-function classification method: “InterProScan 5: genome-scale protein function classification”, Bioinformatics 30(9):1236–1240, doi:10.1093/bioinformatics/btu031.
• Cantalapiedra et al. (2021) for eggnog-mapper v2, a method for functional annotation, orthology assignment and domain prediction: “eggNOG-mapper v2: functional annotation, orthology assignments, and domain prediction at the metagenomic scale”, Molecular Biology and Evolution 38(12):5825–5829, doi:10.1093/molbev/msab293.
• Schwengers et al. (2021) for bakta, a bacterial-genome annotation method: “Bakta: rapid and standardized annotation of bacterial genomes via alignment-free sequence identification”, Microbial Genomics 7(11):000685, doi:10.1099/mgen.0.000685. ₄₂

The project did not identify donors at deep ranks, because per-CDS sequence data (codon usage, k-mers) were not available in queryable KBase Data Lakehouse schemas. M26 tree-based donor inference at genus rank (NB28b) uses the existing M22 output, the species tree and the presence matrix. It produces Open/Broker/Sink/Closed labels with explicit ambiguity bookkeeping. It shipped as an exploratory layer, separate from the deferred M25 composition-based donor inference. M26 algebraically counts potential family-mate donors, which biases it toward Open-Innovator classifications. According to the report, composition-based inference would require external FASTA and codon profiling that breaks the Phase 3 budget. Phase 3 therefore reports at the joint Innovator-Exchange label (Broker-or-Open donor-undistinguished), the same level at which Alm 2006 made claims. Per-family DTL reconciliation (AleRax / ALE) does not run at full GTDB scale. The four-quadrant labels therefore exist only at genus rank on the Phase 3 candidate set (~10K KOs), not atlas-wide. Composition-based confirmation and full DTL reconciliation cross-validation (Liu 2021 DTLOR / Bansal 2013) remain future work. ₄₂

The PSII result is rank-dependent. At genus rank, n = 2,350 with d = 0.08 was STABLE. At class rank, n = 21 PSII KOs with d = 1.50 was Innovator-Exchange. At phylum rank, d = 1.60 was classified Innovator-Isolated. The report considers class rank biologically appropriate per Cardona 2018, since PSII evolved before Cyanobacteria diversified, and so describes the genus null as expected. It attributes the phylum reversal to a substrate artifact: insufficient reference consumer data left the phylum consumer statistic unavailable. Genus, family and order results were STABLE. The report reads the d ≈ 0.08–0.20 producer signals at finer ranks as the expected near-uniformity of a near-universal feature with deep ancestry: genera should not be innovating PSII because it is already shared. It notes that the phylum producer signal (d=1.60) is consistent with the class signal (d=1.50). It says the verdict shifts to ISOLATED solely because consumer data thin out at phylum rank. The class-rank interpretation fits PSII being a class-defining, ancient innovation, but it should not be generalized to all taxonomic ranks. The report itself preserves this caveat: PSII Innovator-Exchange is specifically a class-rank phenomenon, not a generic Cyanobacteria-wide HGT signature at every resolution. The supported claim is narrower: Cyanobacteriia as a class exchange PSII KOs more than reference clades do, with 2.77× photic enrichment (p<10⁻⁵²) and 2.05% ancient gains versus 14.9% atlas-wide. ₄₂

The Alm 2006 quantitative correlation was not reproduced. The project’s 18,989-species substrate differs from the original 207-genome substrate. M22 measures tree-attributed gain events, not all per-genome paralog expansion. The two analyses also define “recent” differently. The report names three mechanisms: substrate-scale heterogeneity, tree-aware versus paralog-count operationalization, and tree-rank granularity. On substrate, Alm 2006 used 207 genomes, mostly cultivated isolates with high annotation density and broad taxonomic spread, whereas this substrate spans the full GTDB tree, including DPANN, CPR and weakly annotated lineages where HPK detection is genuinely sparse; the report argues the original signal is taxonomically heterogeneous and aggregating across the full tree dilutes it. On LSE detection, Alm 2006 used per-genome paralog counts at the HK family level, whereas M22 Sankoff recent-rank gain attribution is tree-aware and would miss expansions occurring entirely within a single recent species, so the two metrics measure overlapping but non-identical phenomena. On granularity, “recent” in M22 is a leaf-adjacent gain on the species tree that captures within-genus diversification, while Alm 2006’s LSE captured all expansion regardless of phylogenetic depth; the definitions converge for very-recent expansions and diverge for older ones. These are proposed explanations, not separately tested. The methodology generalizes, but the point estimate does not survive scaling from 207 to 18,989 genomes. ₄₂

PUL and mycolic gene-neighborhood analyses were limited by scale. The report’s limitations list attributes the deferral to a pandas spatial-merge out-of-memory failure at Bacteroidota scale. Its run log, described below, gives a different account: the full-scale scan of 723K Bacteroidota focal features × 210K contigs failed on the Spark driver result-size cap, and the pandas failure came in a sampled run. The report does not reconcile these two accounts. Only the PSII neighborhood analysis shipped. For PUL and mycolic systems, the non-phage-borne conclusion therefore rests on per-cluster MGE-machinery rates (0% / 0.57%) and literature context, not on complete cargo-neighborhood scans. The report judges these rates sufficient for the headline conclusion. Its run log is more specific. Bacteroidota has 2,581 species, which the report gives as 8.4× Cyanobacteriia’s 309; the limitations list gives the ratio as 8×. Even the SusC/SusD canonical-only scan (2 KOs) yielded 723K focal features × 210K contigs. Its Stage 5 result, 1.3GB serialized, exceeded spark.driver.maxResultSize on toPandas. A run sampled to 309 random Bacteroidota species (80K focal features × 21K contigs, ~105 features/contig) then ran out of memory in the Stage 6 pandas merge (~24M rows × 9 cols, a 9GB DataFrame). The Mycobacteriaceae sampled run never started. The report’s decision was to ship without batched PUL/mycolic neighborhoods, stating that those scans would refine the picture but not change the headline. It nonetheless states a mechanism conclusion: mycolic-acid, PSII and PUL genes flow via chromosomal recombination and ICEs rather than phage-mediated transduction. It adds that the at-baseline PSII neighborhood rules out PSII gains travelling as phage/MGE cargo. For PUL and mycolic acid that mechanism claim rests on per-cluster rates and literature, not on direct cargo tests. ₄₂

Cyanobacteria BacDive coverage was only n = 4. The P4-D4 pangenome-openness cross-validation was null at atlas scale, with Spearman r = −0.011 across 894 genera; the two P4-D4 outputs (data/p4d4_pangenome_openness_per_genus.tsv and data/p4d4_recent_acquisition_vs_openness.tsv) have 3,539 and 894 rows. The report reads this informative null as evidence that M22 and pangenome openness measure distinct evolutionary phenomena. The targeted Mycobacteriaceae and Cyanobacteriia openness tests were underpowered, with n = 10 and n = 83 genera, respectively. ₄₂

Cross-phase uncertainty was not propagated into a unified atlas confidence interval. Bootstrap confidence intervals for individual M22 events were deferred. Sankoff results were not comprehensively cross-validated against DTLOR or other modern reconciliation methods. The ecology results show association or consistency with expected environments, not a causal effect of environment on gene innovation. ₄₂

At the Phase 2 close, the verdict table recorded several items as not yet established. Bacteroidota PUL → Innovator-Exchange (Phase 1B) was falsified at the UniRef50 absolute-zero criterion (NB07) and reframed at the v2.5 close-reading, with a KO retest deferred. Cyanobacteria × PSII → Broker (Phase 3) was not yet tested, pending an architectural deep-dive at genus rank. The Alm 2006 TCS HK reproduction at r ≈ 0.74 (P4-D3) was not yet tested and was deferred to Phase 4. The later outcomes of these items appear in the hypothesis verdicts above. In the verdict table that includes the NB16 result, the report summarized: 2 of 4 currently-testable pre-registered hypotheses confirmed; 1 falsified at UniRef50; 1 reframed at small effect size in the literature-predicted direction. That count labels the PUL hypothesis falsified, whereas the final synthesis calls it a qualified pass/reframed finding; both labels are recorded as written. ₄₂

External validation was staged across phases. The CRISPR-Cas 24.5× and tRNA-synth 2.3× recent-to-ancient ratios align qualitatively with the HGT literature (Smillie 2011, Forsberg 2012, Metcalf 2014, Phillips 2012). At Phase 2 they had not been quantitatively cross-validated against an independent dataset. All Phase 2 tests use atlas-derived scores. Tests against independent experimental HGT datasets were deferred to Phase 4. That work centred on P4-D1 phenotype/ecology grounding (NMDC, MGnify, Fitness Browser, BacDive, Web of Microbes) and the explicit Alm r ≈ 0.74 reproduction (P4-D3). ₄₂

Effect-size reporting was inconsistent across phases. In response, plan v2.10 M24 made Cohen’s d with a 95% bootstrap CI the standard format for all primary statistical comparisons going forward. The report counts 25 methodology revisions (M1–M25), each documented in the research plan revision history with its rationale and triggering observation. It lists three pre-registration omissions that surfaced and were corrected as project-discipline lessons: M2 dosage biology, the M12 absolute-zero criterion and the M14 misreading-of-Alm-2006. ₄₂

The report contradicts itself on D2 residualization. The headline says both confirmed hypotheses (2 of 4 confirmed) survived D2 annotation-density residualization. The P4-D5 passage instead says 3 of 4 confirmed hypotheses survive D2 residualization. Both statements are recorded here as written; they are not reconciled. The final synthesis repeats the inconsistency. It closes with 2 of 4 pre-registered weak-prior hypotheses confirmed, yet says all 3 confirmed hypotheses are ecologically and phenotypically grounded and that all atlas claims survive annotation-density residualization. Its closing assessment likewise says 3 of 3 confirmed hypotheses are anchored in expected biomes and phenotypes. The same assessment describes the pre-registered KOs as not phage-borne and says ICEs and chromosomal recombination are implicated by literature, not measured. ₄₂

Review integrity was uneven across rounds. REVIEW_4, REVIEW_5 and REVIEW_7 contained fabricated citations (verified via PubMed); REVIEW_8 had 0 fabrications and 7/7 DOIs auto-verified. It raised 3 critical, 4 important and 2 suggested issues against the v3.0 Final Synthesis. The report checked its numerical claim C9 (n=2,350 d=0.08 at PSII genus rank) against data/p3_phase_gate_decision.json and found it accurate. It said C7 was already addressed in v3.0. It judged C8 partially valid but said C8 conflates methodology-revision-as-transparency with multiple-hypothesis testing. The project summary counts 7 integrated adversarial reviews (REVIEW_1 through REVIEW_7). It documents three reviewer-side fabrications or errors: the Mendoza 2020 PMID hijack at REVIEW_5, Koech 2025 at REVIEW_7 and a Burch 2023 PMID error at REVIEW_7. That list omits the REVIEW_4 fabrications noted above, and the report does not reconcile the two accounts. In REVIEW_5 I1, the cited “Mendoza 2020 Hologenome evolution” paper was fabricated. Its PMID 32160912 actually points to Rezaeipandari et al. (2020), a Persian-language WHOQOL-OLD quality-of-life validation study in Health and Quality of Life Outcomes, unrelated to HGT. PubMed searches for the purported Mendoza HGT work returned 0 results. The attached claim that regulatory genes show 50-fold lower transfer rates than metabolic genes is therefore treated as fabricated, not as evidence. ₄₂

Reviewer errors were not limited to citations. REVIEW_4 C1 falsely claimed that Alm 2006 was not cited; the report rebuts this by pointing to references.md line 9. ₄₂

Genotype × Condition to Phenotype Prediction from ENIGMA Growth Curves

• Condition alignment was based on normalized names and produced 42 molecular matches; the report estimates that ChEBI-ID-based canonicalization could expand this to 60–80 matches through programmatic ChEBI ID resolution. That would enlarge the cross-dataset anchor frame for Fitness Browser concordance checks, but the expansion is proposed, not demonstrated. ₄₃

• Only 35.7% of growth curves were fit-ok. Although the 55.1% no-growth fraction is interpreted as biological, approximately 9% of curves failed fitting for technical reasons including monotone violations, short duration, and edge-well effects. ₄₃

• Genus-level biogeography used Microbial Atlas 16S data. Species-level biogeography was available for only 20 pangenome-linked strains with verified GTDB matches. ₄₃

• Spearman co-occurrence measures correlation rather than causation. Spearman co-occurrence suggests shared or exclusive niches but does not prove direct interaction. The report recommends compositionally aware SparCC analysis on the full 100WS ASV matrix to strengthen the niche-partition result. The report’s expectations that Cluster B strains should show different pH-dependent growth profiles than Cluster A strains, and that cluster-distinguishing genomic features should predict pH tolerance, are proposals rather than validated results. ₄₃

• Uranium, nitrate, and metal concentrations are available in CORAL bricks 10/80, but sample-to-location name resolution is incomplete, preventing a completed local test of contamination levels for the acidic Cluster B strains. The proposed test would link strain-isolation-well coordinates, already in coral_strain_locations.tsv, to per-well geochemistry. It would ask whether Cluster B (acid-tolerant) strains come from higher-contamination wells locally, not just globally. ₄₃

• GC%, codon usage bias (CUB), and Morgan molecular fingerprints were not used: GC% was available for only 32 of 727 genomes, CUB required nucleotide sequences inaccessible from JupyterHub, and Morgan fingerprints required RDKit. CUB computation from GenBank files is proposed as the next test of continuous growth-rate prediction. The proposed measure is gRodon S-value or ENC computed from CDS nucleotide sequences. It would test whether ribosomal efficiency explains the µmax variation that KO presence cannot. It requires GenBank files on the CGCMS server (/data/www/CGCMS/static/enigma1/genomes/gbff/), which are not currently accessible from JupyterHub. The report cites Xu et al. (2025, Phydon) as showing that this is the natural next feature. ₄₃

• The Fitness Browser concordance analysis produced 1.19× enrichment after correlation-group expansion, but full validation with KEGG-module expansion remains to be done. That pathway-level expansion would include all KOs in the same KEGG module as a top-SHAP KO. It may recover additional mechanistic signal, distinguishing “wrong feature, right pathway” from “wrong pathway,” but remains untested. The H3 hypothesis-outcomes section instead states that full Fitness Browser concordance validation with correlation-group expansion remains to be done. That sits uneasily with the 1.19× result being described as following correlation-group expansion. Both statements are retained as reported, not reconciled. The report also notes that the n = 7 initial analysis—4 Pseudomonas_E strains and one each of Cupriavidus, Acidovorax, and Pedobacter, evaluated by leave-one-strain-out—did not test cross-genus generalization, per-condition prediction quality, or condition-specific features. ₄₃

• The active-learning framework is actionable but H6 is only partially supported. The report links the proposed 50 experiments to 7,844 prediction failures in the current corpus. It suggests that resolving even a fraction of them would move overall accuracy from 65.1% toward the per-class ceiling of ~78% seen on amino acids and nucleosides. This is a projection, not a measured gain. A formal retrospective subsampling test comparing AL-ranked additions with random selection, followed by wet-lab execution of the 50-condition proposal, remains necessary. The H6 passage calls Prescottella (16%) and Microbacterium (23%) the two genera with the least reliable current predictions. It does not define those percentages there, so they should not be read as prediction-accuracy figures. The formal H6 test, called AL retrospective subsampling, would measure accuracy improvement per labeled pair when experiments are added in AL-ranked order versus random order, using held-out current data. The NB10 ranking framework for it is already in place, but the test has not been run. The 50-experiment list (fumaric acid, melibionic acid, nitrate, etc. × Prescottella/Microbacterium) is the report’s deliverable. Executing it in the wet lab is an ENIGMA experimental decision, not an analytical one. ₄₃

• The report states the interaction-feature gain as +0.032, which does not exactly match the difference between its stated mean AUC endpoints of 0.620 and 0.653. Both values are retained as reported, not reconciled. ₄₃

• The stated concordance rates of 18.7% and 16.3% do not yield exactly the stated 1.19× enrichment ratio. All three reported values are retained without reconciliation. ₄₃

• The ENIGMA growth-curve dimensions stated as 123 strains x 194 conditions are not consistent with the stated 13,632 pairs. The reported dimensions and the pair count are preserved separately. ₄₃

• The report states that its Act I–III backbone is complete. Remaining work is either extensions that require data outside JupyterHub reach or formal validations of already-delivered frameworks. ₄₃

Harvard Forest Long-Term Warming — DNA vs RNA Functional Response

• The study used a single sampling date, 2017-05-24, so it cannot detect seasonal effects, which the report notes Shinfuku et al. 2024 show are substantial at this site. ₄₄

• Omics-rich layers had limited sample sizes (n=28 metagenomes and n=39 metatranscriptomes), reducing per-KO FDR power across 12–14K KOs. ₄₄

• Metatranscriptome KO counts came from contig annotations and represent transcript-pool composition rather than TPM-quantified expression; contig-level annotation counts approximately represent relative transcript abundance but are biased by assembly quality. ₄₄

• The paired DNA/RNA comparison is affected by the horizon × incubation confound, and organic-horizon DNA lacks direct samples, preventing incubation from being factored cleanly out of that DNA analysis. ₄₄

• Because all samples were collected on a single date (2017-05-24), the single-timepoint RNA pool may include diurnal and microspatial variation from moisture pulses, root-exudate availability, and temperature conditions that are unrelated to chronic warming. The report therefore treats comparable long-term DNA and RNA response magnitude as compatible with, but not excluding, an earlier transient in which RNA could lead DNA. ₄₄

• abiotic_features was all zeros for these samples because of an NMDC parsing artifact, so the analysis lacked in-lakehouse soil temperature, pH, and nitrogen measurements; the +5°C treatment label was the only environmental contrast. ₄₄

• The project excluded nmdc_arkin, so it had no quantitative NOM (natural organic matter), metabolomics, or proteomics layers, although equivalent quantitative layers exist in nmdc_arkin.{nom_gold, metabolomics_gold, proteomics_gold} and could be added if scope expands. The ChEBI labels for the differential metabolite hits were not resolved through external ontologies. ₄₄

• The pmoA/pmoB and glyoxylate-cycle findings are directional gene-level signals, and the individual RNA signals did not survive FDR across 14K KOs. The report interprets the heated-mineral metabolite-richness decrease as consistent with faster substrate turnover, but it is not quantitative metabolomics evidence. ₄₄

Metagenome-Prioritized Phage Cocktails for Crohn’s Disease and IBD

The ecotype framework is operationally useful but not uniformly reproducible: LOSO ARI averaged 0.113, the E3 Tier-A list came from a single eligible cMD study, E0 and E2 lacked viable within-ecotype CD-versus-nonIBD contrasts, and hard ecotype assignments discretize an underlying continuum (Costea 2018), with soft probabilities preserved in data/ecotype_assignments.tsv so the hard calls are to be treated as operational labels rather than biology. The report’s own summary of this is that Pillar 2 cross-cohort replication is weak: single-study E3 evidence combined with marginal ecotype stability means the Tier-A list, particularly for E3, should be treated as hypotheses for further experimental validation rather than established targets, and the Pillar 5 UC Davis cocktail drafts inherit that limitation, with per-candidate cross-cohort support to be annotated explicitly. The UC Davis Kuehl projection onto the K = 4 framework accordingly used LDA as primary and GMM as advisory: the Kaiju-to-MetaPhlAn3 projection showed method asymmetry, with LDA more robust, whereas GMM assigned all 26 UC Davis samples to E3 with confidence >0.97 under sparse feature overlap, an interpretation judged artifactual. ₄₅

The project used a partial multi-method differential-abundance consensus because LinDA was implemented in pure Python and ANCOM-BC/MaAsLin2 were not completed; the plan had called for ≥ 2 / 3 of {ANCOM-BC, MaAsLin2, LinDA} to agree, and the report calls the bootstrap-CI and within-substudy-meta outputs additional independent evidence streams that are not drop-in ANCOM-BC / MaAsLin2 substitutes, leaving a full three-method R-native consensus as a publication-grade follow-up. That independence framing sits against the report’s own statement, recorded above, that bootstrap-stable within-ecotype differential abundance and LinDA run on the same ecotype-defined subsamples and so share the selection-on-outcome bias, leaving only the within-substudy CD-versus-nonIBD contrast as an independent evidence source; the report does not reconcile the two framings. The NB08a species × BGC interaction test was not performed — the analysis instead used the precomputed ref_cborf_enrichment table, which carries a CD-versus-HC main effect rather than a species × BGC interaction, so the per-species CB-ORF enrichment is consistent with H3c but cannot formally distinguish a species × BGC interaction from a species main effect plus a BGC main effect, and the interaction test was dropped under plan v1.9 as a structural project-scope limitation rather than a deferred follow-up. Its background comparator is the full BGC catalog of 10,060 BGCs across all species rather than a matched comparator, so the iron_siderophore Fisher OR of 44× is partly inflated by the catalog-wide rarity of iron MIBiG matches (51 / 9,774 = 0.5 %) and should not be treated as a precise effect size — a more conservative comparator would be a matched-niche pathobiont set of gut Proteobacteria plus CD-associated Firmicutes at similar genome-assembly depth — while the qualitative finding that E. coli alone among the actionable Tier-A core carries the iron-siderophore signature is described as robust across choice of comparator; raw-read strain-resolution analyses were dropped under the no-raw-reads scope; AIEC-specific targeting therefore depends on a proposed kbase_genomes pks plus iron-BGC genome-content query to identify Yersiniabactin/Enterobactin/Colibactin-positive strains rather than on strain resolution from curatedMetagenomicData raw reads; plan v1.9 dropped the raw-reads-based cMD strain-resolution test as not feasible, leaving a genome-content survey — a kbase_genomes pks plus Yersiniabactin plus Enterobactin co-occurrence query cross-referenced with the Kumbhari fact_strain_competition strain-frequency data — as the no-raw-reads alternative. ₄₅

The strain-adaptation result rests on one cohort (Kumbhari/LSS-PRISM), whose cross-cohort sensitivity test was acknowledged-weak by design, and about 95 % of FDR-passing genes are unclassified by the gene-symbol regex of the curated 13-category classifier, so the functional-enrichment signal belongs to the category-classified subset rather than to the full gene set. The planned cross-cohort species-level sensitivity test on curatedMetagenomicData was specified as sensitivity-only and then dropped unexecuted because raw reads were unavailable. ₄₅

The serology analysis is structurally low-powered rather than conclusive: n = 67 across 3 unevenly distributed sites is, in the report’s terms, structurally low-power for |r| > 0.40 detection at α = 0.05 with correction for 48 tests, and the plan acknowledged this test as weak by design. Subject-level aggregation discards the multiple serology and metagenomics visits each subject contributed, and while mixed-effects regression would have more power, unbalanced visit counts complicate interpretation; OmpC and ANCA also cross-react with related antigens beyond their named targets, so per-species attributions are correlative rather than exclusive. The single-cohort caveat is structural — there is no project-internal replication path, and multi-cohort meta-analysis would be needed to firm the signals up. ₄₅

The metabolomics ecotype clustering result is specifically limited by uncorrected batch effects: pooled HMP2 and FRANZOSA m/z-bridge clustering separated cohorts, with PC1 explaining 79 % of variance and cross-cohort LOSO ARI = 0.000, below the taxonomic baseline of 0.113. The 3-modality MOFA+ plan was reduced to a 2-modality CCA pilot because HMP2 pathway abundance was unavailable in the mart, and the polyamine signature remains single-cohort because no m/z bridge was found. The m/z-bridge feature space is also small at 111 unique features; the report notes that a larger feature space of 500-1000 metabolites with name annotations across cohorts might give different results, but that this requires external Franzosa supplementary annotations placed out of scope by the plan. ₄₅

PhageFoundry coverage applies to its 188 tested E. coli strains, not UC Davis patient isolates — the report states that coverage as 95 % of the 188 PhageFoundry-tested strains, a rounded label distinct from the 94.7 % its set-cover table reaches, and per-patient applicability requires an AIEC strain-resolution diagnostic on UC Davis patient stool isolates, queued in the near-term roadmap — and the dataset lacks explicit AIEC-versus-commensal labels, burst-size measurements, titer data, and in-vivo delivery validation. Those 188 strains do include canonical AIEC strains named as phage isolation hosts — LF82, LF73, 536, NIC06 and BCH953 — but accurate AIEC coverage estimation would require cross-reference to literature isolate annotations, and the cocktail covers broad E. coli strain diversity without by itself distinguishing AIEC from commensal E. coli, so the report makes a strain-level diagnostic (per-patient pks-island / Yersiniabactin / Enterobactin gene-presence detection) a requirement before cocktail selection. The susceptibility matrix is also binary, with no titer or burst size, while real-world cocktail dosing depends on phage burst size plus receptor-binding kinetics plus gastric-passage survival — all out of project scope. HMP2 viromics had an 80 % family-classification Unknown fraction, so the Tier-A × phage-family correlations — given in that caveat as E. coli × Podoviridae +0.18 and × Myoviridae +0.13 — are restricted to the classifiable 20 %, and improving family resolution would require re-running VirMAP/MARVEL against newer phage reference databases, which the report places out of scope; endogenous phage correlations were modest: the report states a maximum absolute ρ ≤ 0.18, a bound inconsistent with its own tabulated E. coli × Podoviridae ρ = +0.183 and left unreconciled, so it should be read as the report’s statement rather than an exact ceiling, with the substantive point being that no correlation reached its |ρ| > 0.30 strong-candidate threshold. The three gut-anaerobe coverage gaps — H. hathewayi, M. gnavus lytic alternatives and F. plautii — remain unresolved pending external-database queries of INPHARED (~25K phages) and IMG/VR (~3M UViGs), and those queries, multi-cohort prospective validation of the state-dependent dosing rule derived from the n = 1 patient 6967 trajectory, and a per-patient AIEC strain-resolution diagnostic are all flagged as required clinical-translation prerequisites in the NB17 roadmap’s near-term tier, out of scope here rather than performed. ₄₅

The UC Davis per-patient framework is based on only 23 patients, so its outputs are exemplars and templates rather than statistically robust per-patient validation and generalization requires multi-center validation on external CD cohorts; it has no patient-specific bile-acid measurements — the bile-acid coupling cost is an ecotype-level, not a per-patient, quantity — no AIEC strain-resolution diagnostics, and it uses Kaiju-derived target presence, which the report flags as lower-confidence than the MetaPhlAn3-based curatedMetagenomicData analyses given the NB02 classifier-mismatch asymmetry. Patient 6967 is the only biological-replicate longitudinal trajectory, so the E1→E3 drift, cocktail Jaccard = 0.60, 3–6-month reassessment interval, five-fold qPCR trigger, and state-dependent dosing rules are hypotheses rather than clinically validated rules. There is also no patient-specific AIEC strain-resolution diagnostic for the 8 E. coli-positive patients, so the cocktail recommendation assumes an AIEC subset prevalence taken from Dogan 2014 and Dubinsky 2022, with per-patient AIEC genotyping (pks-island plus Yersiniabactin plus Enterobactin) recorded as a near-term clinical-translation prerequisite; and the F. plautii deprioritization is precautionary, based on the NB09c mechanism, with clinical validation requiring per-patient bile-acid panels. The cocktail-strategy distribution on this cohort — 12 reserve, 4 E1 hybrid, 4 E0 limited, 1 E1 full, 1 E3 and 1 state-dependent — is the operational stratification the report says should be revisited on a larger CD cohort. ₄₅

The longitudinal and dosing conclusions rest on a single trajectory: patient 6967 is the only multi-timepoint patient with biological-replicate samples in the UC Davis cohort, so the ecotype-drift conclusions are based on an n = 1 longitudinal trajectory. Both ecotype calls are moderate-confidence (visit 1 = 0.64, visit 2 = 0.41), so the label shift could partly reflect classifier uncertainty, although the 14× M. gnavus and 3× E. lenta expansions are quantitative biological observations independent of the ecotype label. Patient 1112’s technical-replicate Spearman ρ = 1.000 validates the technical-noise floor for this longitudinal contrast but is not itself a biological-replicate longitudinal trajectory, so patient 6967 remains the only such trajectory in the cohort. There is no timing information between the two visits, so the duration of the E1→E3 drift is unknown, which limits clinical-workflow timing recommendations, the 3-6 month re-test interval being a clinical convention default rather than a value derived from these data; the state-dependent dosing rule is theoretical and not yet validated in clinical practice; and the cocktail Jaccard of 0.60 is moderate, with whether a 3-of-5 component overlap implies clinically meaningful cocktail re-design depending on patient-response heterogeneity, which is out of project scope. ₄₅

Lab Fitness Predicts Field Ecology at Oak Ridge

The report identifies several limitations: 16S amplicon sequencing resolves only to genus level; the 108 overlapping samples may not capture the full Oak Ridge geochemical range; geochemistry measurements are point-in-time observations; the aggregate metal-tolerance score is crude and may obscure condition-specific responses, so condition-specific fitness scores such as uranium-only scores would be more informative; multiple communities per sample, including different filter sizes and replicates, were aggregated; only 12 Fitness Browser genera had sufficient data for the metal-tolerance correlation; and pH, dissolved oxygen, carbon sources, and other confounders were not controlled. ₄₆

The report proposes several next analyses. The first is species- or strain-level matching using metagenomic data instead of 16S amplicons, if such data are available in ENIGMA CORAL genome or assembly tables. The second is multivariate CCA or RDA (canonical correspondence analysis or redundancy analysis, constrained ordinations that model community composition against several geochemical variables at once), controlling for pH, redox, and carbon sources. The report also proposes temporal analysis across sampling dates and metal-specific fitness scores matched to corresponding site concentrations. Finally, it proposes adding Rhodanobacter to the Fitness Browser as a high-impact target; this genus dominates contaminated Oak Ridge wells (Carlson et al. 2019) but is not in the Fitness Browser. ₄₆

Lanthanide Methylotrophy Atlas — Summary

The report addresses phylogenetic non-independence for H1 using pooled, family-equal-weight, and Bayesian GLMM analyses, all of which support H1; however, H2 still relies on within-phylum stratified analyses rather than a fully phylogeny-aware mixed model. ₄₇

Marker calls vary substantially between eggNOG and Bakta, so headline statistics use a marker-specific source of truth and secondary union-of-sources analyses. Sequence-level evidence is out of scope: the study does not screen for pseudogenes, truncated ORFs, or assembly fragmentation. Genuinely novel REE-handling enzymes lacking KEGG or RefSeq homologs would also be missed. ₄₇

The REE-AMD anchor contains only 37 MAGs from a single bioproject and is descriptive only; larger, independent collections from REE-mining tailings, leachate, and bioreactors are needed for inferential testing. The 897 xoxF genomes with no PQQ evidence require ORF-integrity and genome-completeness analysis, such as CheckM2, to distinguish fragmentation, pseudogenization, and community-PQQ acquisition. ₄₇

AlphaEarth coverage is 1,457 / 3,690 = 39.5 % of xoxF genomes, compared with a 28 % pangenome baseline, leaving 60.5 % without environmental coordinates. Embedding-based niche analysis was scoped out of the report, but the report considers a coverage-restricted supplementary analysis feasible: coverage-restricted PCA (principal component analysis) / UMAP (uniform manifold approximation and projection) analysis could test whether xoxF carriers form distinct environmental or biogeographic clusters, stratified by phylum. ₄₇

The ncbi_env environmental classification is text-mining-derived and uses hierarchical regex priorities, so broad classes such as host_associated may contain misclassifications. The host_associated class spans gut, skin, oral, infection and other samples. The report states that the strong signals, soil/sediment enrichment and host depletion, survive despite possible misclassification. The study also proposes characterizing f__REEB76, examining lanmodulin sequence diversity in 22 Methylobacterium extorquens genomes with 1 copy each, testing lanthanum and cerium chloride with targeted RB-TnSeq (random barcode transposon sequencing) on the FitnessBrowser organism panel, and reporting the 505 eggNOG lanM false-positive pattern upstream. ₄₇

Lignin Enrichment and Ecological Memory in Microbial Communities

The report identifies the n=3 permutation floor as its primary statistical limitation, particularly for pairwise tests. It proposes n>=5 per group for future experiments so that pairwise significance testing becomes possible. Environment constraints led the analysis to use 97% vsearch OTUs rather than DADA2 ASVs (amplicon sequence variants, exact error-corrected sequences), which may merge closely related species. The standard UNITE ITS database could not be downloaded. ITS taxonomy was therefore assigned by BLAST against NCBI ITS_RefSeq_Fungi, with 19,375 reference sequences. Some assignments may be imprecise, and the database may underrepresent environmental taxa and miss rare or novel taxa. The report considers genus-level assignments more reliable than finer ones. ₄₈

The ITS findings are preliminary for several reasons. Replicate consistency was poor, with within-group Bray–Curtis distances approaching 1.0 for Round-2 groups. Sequencing depth varied from 6,953–200,902 reads per sample before rarefaction, and some samples had very low diversity. ITS was rarefied to 5,900 reads, versus 24,000 for 16S. The report lists low depth in some samples as contributing to replicate variability even after rarefaction. Enriched ITS groups had only 3–8 OTUs, making composition highly sensitive to stochastic variation. Sample LL_1 was excluded after retaining 74 reads, reducing L-L to n=2. The fungal memory effect was therefore not statistically detectable, and the notable Pleurotus emergence in LC-LC requires confirmation. ₄₈

PERMDISP was significant for both markers, so the PERMANOVA results reflect differences in group dispersion as well as group location. No technical metadata on extraction or library-preparation batches were available, so batch effects could not be formally assessed. The functional interpretation of lignin degradation rests on taxonomic associations and literature context; gene-level pathway enrichment was not directly measured. ₄₈

The planned Procrustes comparison of bacterial and fungal ordinations was not completed, because the near-random ITS replicate structure would make the fit difficult to interpret. The report proposes several lines of future work: larger replication, DADA2 ASV analysis, UNITE-based ITS taxonomy, and phylogenetic diversity and UniFrac analyses. It also proposes cross-referencing Pseudomonas, Acinetobacter and Comamonas in the KBase Data Lakehouse through kbase_ke_pangenome, functional inference with PICRUSt2 or Tax4Fun to test whether the beta-ketoadipate and protocatechuate lignin-catabolism pathways are enriched (a question this report leaves untested), resequencing ITS libraries at higher depth to improve Round-2 replicate consistency, and intermediate time points to resolve restructuring kinetics. ₄₈

Metabolic Capability vs Metabolic Dependency

The study states that SEED-proxy mapping can introduce false positives because related subsystem annotations may not represent direct GapMind pathway membership; direct GapMind per-step gene assignments would improve precision. These false-positive genes from related subsystems dilute the fitness signal and could misclassify pathways that would otherwise differ in conservation. The large intermediate zone (32.3%) makes the latent fraction moderately sensitive to threshold tightening, as shown by the 4.7%–21.1% sensitivity range. ₄₉

Fitness experiments are biased toward laboratory conditions, so pathways that are important in environments not represented by the tested media may appear latent, inflating the latent fraction. Pathway-level conservation cannot detect partial erosion or progressive loss of individual genes within otherwise complete pathways, and only 48 organisms were fitness-tested among 293,000 genomes with pathway predictions, limiting generalization across taxa. ₄₉

The conservation comparison may be affected by species-level averaging: many genomes in a species clade were not fitness-tested, potentially inflating apparent conservation of pathways classified from a limited set of organisms. The reported Black Queen signal may also be difficult to detect because gene loss can require longer evolutionary timescales than current sampling captures. As a proposed explanation of this evolutionary lag, the report suggests that latent capabilities may be mid-loss in some lineages but still appear conserved at the species level. ₄₉

The ecotype analysis is observational. Environment-cluster associations, including those in Salmonella and Phenylobacterium, may be confounded by phylogenetic structure because explicit phylogenetic correction was not performed. Marine non-associations may reflect coarse NCBI isolation-source metadata or pathway-level annotations that miss ecologically relevant gene-content and expression differences. ₄₉

An attempted organism-to-clade linkage using NCBI taxonomy IDs from kbase_ke_pangenome.gtdb_metadata returned zero matches because the relevant column contained boolean strings rather than numeric taxids. Downstream analyses therefore used organism-level fitness aggregates without an explicit clade-level linkage, which may reduce the precision of the H2a conservation comparison. ₄₉

Gene-Resolution Metal Cross-Resistance Across Diverse Bacteria

Metal concentrations differed among experiments, so dose-response effects could influence cross-resistance estimates. Organisms also ranged from 3 to 112 metal experiments, affecting the reliability of per-organism matrices. ₅₀

The 28 organisms are not phylogenetically independent; formal phylogenetic comparative methods such as PGLS (phylogenetic generalized least squares) or independent contrasts would strengthen the conservation claim. ₅₀

The BacDive validation is underpowered: the limitations section reports n = 26 at Fitness Browser organism scale, while the validation results report 20 independent species after matching and collapsing. The report emphasizes that this scale lacks the statistical power achieved by the Metal Fitness Atlas at pangenome scale with 42K strains. ₅₀

No negative controls were tested. Because all tested metal pairs were positive, the study cannot distinguish universal cross-resistance from a general-stress response to all metals without non-metal stress controls; the report states that the counter_ion_effects project partially addresses this issue. ₅₀

The report lists the following as proposed future directions, not as results it reports. (1) Pangenome-scale H3 validation: apply the cross-resistance gene signatures to predict multi-metal tolerance across 27K species using the KEGG/PFAM mapping approach from the metal-fitness-atlas (kegg, pfam), then validate against bacdive polymetallic isolation environments. (2) Phylogenetic independent contrasts: formally control for phylogenetic non-independence in the cross-resistance conservation analysis using PGLS (phylogenetic generalized least squares) or phylogenetic PCA (principal component analysis). (3) Metal dose-response normalization: normalize fitness effects by metal concentration relative to MIC (minimum inhibitory concentration) to control for dose-response confounds. (4) ICA module decomposition: apply ICA (independent component analysis; independent-component-analysis), taken from the Pan-bacterial Fitness Modules via Independent Component Analysis project, specifically to metal conditions. The goal is to identify co-regulated metal-response modules and test whether cross-resistance genes cluster into coherent regulatory units. (5) Structural biology: use AlphaFold structures to investigate whether metal-shared proteins have structural features, such as metal binding sites or membrane interfaces, that explain multi-metal sensitivity. These structural explanations are proposals, not established findings. ₅₀

Pan-Bacterial Metal Fitness Atlas

Metal coverage was uneven: cobalt and nickel were tested in 27 organisms, whereas uranium, chromium, mercury, cadmium, selenium, and manganese were tested in only one or two organisms. Consequently, cross-species patterns for rare metals largely reflect DvH and psRCH2 biology. The limitations section gives a 27-organism count for nickel, which conflicts with the 26-organism nickel count in the atlas-scale findings. The report does not reconcile the two. ₅₁

Metal concentrations were not normalized by dose or tolerance threshold; for example, nickel concentrations ranged from 0.01-2.0 mM across organisms. Fitness effects therefore cannot be compared directly without accounting for each organism’s relative exposure. ₅₁

Phylogenetic non-independence remains a limitation because multiple Pseudomonas fluorescens strains can inflate apparent cross-species conservation, although exclusion of four duplicate FW300 strains left the principal core-enrichment result robust. ₅₁

The broad metal-important definition, fit < -1 OR n_sick ≥ 1, captures approximately 3.3% of genes and includes many general stress genes rather than metal-specific resistance genes. This wording differs from the findings section, which describes significant fitness defects as fit < -1, |t| > 4. The report therefore recommends repeating the analysis with genes important for metals but not for other stresses. ₅₁

Putatively essential genes, estimated at approximately 14.3% of protein-coding genes and approximately 82% core, lack transposon insertions and are absent from the fitness data. Because these genes are overwhelmingly core, their exclusion makes the observed core enrichment a conservative estimate. ₅₁

The conservation analysis covered 22 of 48 organisms because only 22 had pangenome links; the report also states that 22 of 31 metal-tested organisms were covered in the primary conservation analysis. The nine excluded metal-tested organisms lacked Fitness Browser pangenome links, and 26 organisms lacked Fitness Browser pangenome mappings in the broader mapping context. ₅₁

The repertoire prediction did not validate as a predictor of metal fitness. The report proposes several next steps. These are regulatory or expression-based modeling and normalizing fitness effects by metal concentration relative to MIC (minimum inhibitory concentration) to enable fair cross-species comparison. They also include phylogenetic independent contrasts to control for phylogenetic non-independence when comparing conservation patterns across organisms. The report proposes replacing the simple count-based metal score with a per-species hypergeometric enrichment test that controls for total functional annotation content. Finally, it proposes functional annotation of the 149 novel candidates using PaperBLAST, InterPro, and structural prediction (AlphaFold) to characterize the 89 truly unknown and 43 DUF-domain families with conserved metal phenotypes, whose functions remain uncharacterized. ₅₁

Global Biogeography of Environmental Bacterial Metal Resistance

• The report is preliminary. Its status section says that data extraction and coordinate retrieval were complete at NB01 and that NB02 spatial analysis and NB03 figures are pending. Later sections of the same report nevertheless present NB02 results and reference NB02 figures. Production figure completion remains pending. ₅₂
• The report’s interpretation section warns that, because of the 30.8% ENA coordinate gap, any global map of metal resistance based on public archives is subject to substantial geographic blind spots. It also names the question of which biomes are most underrepresented as a key open question. The later NB02 coverage check (0 of 532 MAG grid cells lacking ENA coordinate coverage) narrows this warning: the gap is per-sample, not missing coverage in the grid cells the data represent. Biome underrepresentation remains unresolved. ₅₂
• Expedition-level clustering remains untested. The report’s pending-analysis list proposes testing whether the Atacama cluster and the USA clusters are single-study artefacts by checking distinct sample_accession prefixes within each hotspot. That list gives the Atacama cluster as OR=9.8, while the NB02 hotspot results report OR=9.83. ₅₂
• Sampling-effort correction remains necessary; the report proposes testing whether hotspots persist after normalising by log(n_MAGs). ₅₂
• The matplotlib import must be fixed in the NB01 map cell before the production figure is generated. ₅₂

Metal-Specific vs General Stress Genes

The analysis had 40.7% gene attrition: 7 organisms and 5,229 gene records were excluded because of locusId format mismatches between the metal atlas and fitness matrices. The excluded organisms were ANA3, Dino, Keio, MR1, Miya, PV4, and SB2B, including important model organisms such as Keio (E. coli), MR1 (Shewanella), and ANA3. The report cautions that excluded genes may have different specificity profiles. Elsewhere it states that the excluded organisms are taxonomically diverse and that their absence is not expected to introduce systematic bias. The report asserts this expectation but does not test it empirically. ₅₃

The 5% sick-rate threshold is arbitrary: results were qualitatively stable across 1–20%, but exact fractions varied. The planned validation against the Fitness Browser’s built-in specificphenotype annotations was not performed. ₅₃

The ICA module analysis failed to identify metal-specific modules; the proposed correction is to use precomputed z-scored module activities from Metal Atlas NB05. Further work should also resolve the integer-versus-string locusId mismatch for ANA3, Dino, Keio, MR1, Miya, PV4, and SB2B; the report says this would recover the remaining 40% of gene records. It should also use the Fitness Browser’s built-in specificphenotype condition-specificity annotations as an independent validation. The counter_ion_effects methodology should be replicated exactly (fit < -1 without |t| > 4) to validate the osmotic overlap. Finally, AlphaFold predictions should be used to identify metal-binding sites in the top metal-specific candidates (UCP030820, YebC, DUF1043). This structural analysis is proposed work; the report gives no structural result. ₅₃

Metal Resistance Ecology — Phylogenetic Conservation vs. Environmental Selection

MicrobeAtlas niche breadth is a sequencing-effort proxy rather than confirmed ecological range. Detection across 13 environment categories is affected by sampling intensity, primer bias, geographic and temporal non-uniformity, and missing environments; Levins’ B_std corrects for unequal environment sizes but cannot correct for missing environments or detection-probability differences. Genus aggregation also means that a broad genus score can reflect different species occupying different habitats rather than one organism being a generalist. The report specifies that the primer bias involves V3–V4 or V4 regions and that the non-uniformity concerns SRA/ENA sample contributions. Because of genus aggregation, the PGLS β reflects covariation between inferred niche breadth and inferred metal tolerance capacity, not confirmed physiology. ₅₄

The strict 5% prevalence analysis preserved the direction of the association but reduced the PGLS sample size by 37%, from 606 to 379, and produced p = 0.092. Primer bias, missing environments, and the ecological meaning of sparse detections remain unresolved without a multi-primer, multi-region survey design. The R4 strict-filter coefficient was β = +0.0166 (p = 0.092, n = 379), compared with the original β = +0.0215, p = 1.5×10⁻⁴ (n = 606). In this limitations section, the report rounds these coefficients to β = +0.017 versus +0.022 original. It says the filter confirms the signal direction but reduces significance because 37% fewer genera pass, and concludes that the effect is not explained by rare transient detections. Direction alone does not establish robustness under the filter, because the filtered result was non-significant (p = 0.092). ₅₄

The 48-genus archaeal analysis is biased toward cultured methanogens, halophiles, and thermoacidophiles and misses environmentally dominant groups including Thaumarchaeota and Woesearchaeota. The report states that ≥200 archaeal genera with AMR data are needed for adequate power. Its own formal power analysis, however, estimates n ≥ 702 for 80% power at α = 0.05 under the observed effect and error assumptions, so the ≥200 adequacy figure is in tension with the formal requirement. The report’s own power table projects only 32% power at α = 0.05 for n = 200. Reaching ≥200 archaeal genera with AMR data would require expanded archaeal pangenome sequencing and annotation in kbase_ke_pangenome, which the report identifies as new data. ₅₄

Pangenome coverage is uneven: the Spearman correlation between sequenced genomes per genus and inferred metal type diversity was r = 0.35 with p = 2.6×10⁻¹⁹, and single-genome genera had lower metal type counts than multi-genome genera with Mann–Whitney p = 5.1×10⁻¹⁰. Coverage and rarefaction analyses substantially addressed this concern, but a formal rarefaction curve across multiple species-per-genus thresholds remains outstanding. ₅₄

Environment categories are heterogeneous, especially aquatic, which combines marine, freshwater, estuarine, hydrothermal-vent, and pond environments. Aquatic represented 40,353 OTUs, or 40.8% of the dataset; excluding it reduced β from +0.021 to +0.0085 but retained nominal significance at p = 0.031. A taxon detected in both a marine coastal sample and a freshwater lake sample counts as broad-niched, although this could reflect closely related but ecologically distinct ecotypes rather than a single generalist population. Excluding any environment other than aquatic gave p < 0.01. The report states that the result is not contingent on the heterogeneous aquatic category. That statement rests on nominal significance, because the aquatic exclusion does not survive false discovery rate (FDR) correction (q = 0.061), while 12 of 13 exclusions survive at q < 0.05. Finer subdivision into marine, freshwater, and saline categories requires re-analysis of MicrobeAtlas sample metadata, which is not available in the data extraction used for this project. Because MicrobeAtlas does not provide a finer environment classification, the analysis cannot test sub-category heterogeneity such as marine versus freshwater within aquatic. Splitting categories would require re-running the niche breadth computation with a revised environment taxonomy. The report’s quantitative bias note calls the over-representation of aquatic environments a potential source of downward bias when the category is included. Yet the weaker β after excluding aquatic (+0.0085 vs +0.021) suggests, without establishing, that aquatic OTUs contribute disproportionately to the association; the two readings are in apparent tension. ₅₄

The Gaussian PGLS treats metal type diversity, an integer count from 1–7, as a continuous z-scored predictor; the report considers this standard when the count is an independent variable but notes that the Gaussian assumption applies to the residuals of B_std, not to the distribution of metal types. The report says inspecting PGLS residuals for normality is warranted but does not report that this check was performed. It adds that the main concern would arise only if metal types were modelled as a response (e.g., what predicts metal type count?), where a Poisson or negative binomial PGLS (phylolm with Poisson family, or MCMCglmm) would be appropriate; these are proposed for that reverse regression, not as replacements for the current B_std-response PGLS. Pagel’s λ for the count trait also assumes Brownian-motion evolution (Gaussian increments), an approximation for an integer trait. The report names discrete-state (Mk) or threshold models as alternatives for this λ analysis, while noting they are not standard for this type of continuous-trait PGLS analysis. ₅₄

The GTDB r214 genus-representative tree, constructed from 120 concatenated marker genes, was used for all phylogenetic analyses, but sensitivity to NCBI, SILVA-based 16S, or other GTDB trees was not tested. Genus-level pruning discards within-genus phylogenetic structure, although the within-genus metal-type standard-deviation test was non-significant at p = 0.18. ₅₄

AMRFinderPlus annotations use hidden Markov model (HMM)-based detection against the NCBI Bacterial AMR Reference Gene Database. For metal resistance, AMRFinderPlus uses curated HMMs built from experimentally validated reference sequences (e.g., merA/merB for mercury, arsABCDR for arsenate, copABCD for copper). HMM gathering thresholds are set during model calibration, and no additional identity threshold is applied because HMM scores are not identity values. Cross-reactive metalloproteases or ABC transporters with structural similarity to metal resistance proteins could be misclassified, although AMRFinderPlus applies taxon-specific cutoffs to reduce this. The “other” metal category (17% of annotated clusters) captures metal-associated genes not classified into the 7 primary types and may include weaker hits. The report says excluding it would produce a 6-metal analysis yet lists seven metals (Hg, As, Cu, Zn, Cd, Cr, Ni), so that description is internally inconsistent. No subset was manually validated. Spot checks of highly prevalent gene families (arsD, merP) showed biologically expected distributions, high in contaminated-environment genera and lower in strict aerobes, which the report treats as only informal support. The recommended follow-up is manual BLAST verification of 100 randomly sampled annotated gene clusters. ₅₄

Track B community-weighted means cover only 16.8% of reads because the remaining approximately 83% could not be joined to genus-level GTDB AMR annotations. If uncovered genera have systematically different resistance profiles, the community-weighted means could be biased upward; correlation between per-sample covered-read fraction and community-weighted mean is required as a diagnostic. The 16.8% reflects OTUs whose genus-level SILVA taxonomy could be matched to GTDB AMR annotations. The report argues that the between-sample CWM variance (1.01–1.83) is consistent with a real biological signal. It makes the same argument for the correlation with known contamination status (FW215/FW216 in the plume). It states, however, that robustness to variation in coverage fraction has not been formally tested. ₅₄

The study’s robustness and sensitivity analyses were post-hoc exploratory analyses rather than confirmatory tests, and the analysis plan was not preregistered. The report says the robustness analyses were conducted after the main analysis in response to anticipated reviewer concerns, and should be read as reassurance rather than confirmatory evidence. It describes the primary metal-types-to-B_std result as “Bonferroni-corrected” with p = 1.5×10⁻⁴. That wording does not establish whether 1.5×10⁻⁴ is an adjusted p-value or an unadjusted p-value assessed against the Bonferroni threshold. The report also identifies unresolved needs for HGT-burden predictors, site-level contamination metadata, co-resistance clustering, finer environment classification, alternative phylogenetic mixed models, AMRFinderPlus validation, and a public Zenodo archive with a citable DOI. ₅₄

The future-directions list does not fully match the completed analyses. Future Direction #1 proposes pangenome rarefaction: resampling genera to equal genome coverage (e.g., n = 1 or 2 species) and repeating PGLS to formally control for the genome-sampling depth confound. It says this would require access to the species-level AMR table stratified by genome count. The report does not reconcile this proposal with the R2 one-species-per-genus rarefaction reported above, so the proposal should not be read as evidence that no rarefaction was performed. Future Direction #2 proposes expanding archaeal pangenome coverage in kbase_ke_pangenome to environmental metagenome-assembled genomes (MAGs, genomes reconstructed from metagenomes), particularly Thaumarchaeota. It would repeat the archaeal PGLS once ≥ 100 archaeal genera with AMR data are available. This ≥ 100 threshold differs from both the report’s ≥200 adequacy figure and its formal n ≥ 702 power requirement. Future Direction #3, the genome size covariate, is marked as completed as Robustness Analysis R5. Genus-level mean genome size from kbase_ke_pangenome.gtdb_metadata was added as a PGLS covariate, and the metal type diversity effect is described as independent of genome size (ΔAIC = 10.8). ₅₄

Proposed follow-up analyses and review responses

• Future Direction #4 proposes genus-level pangenome openness, the proportion of genes in the accessory genome, as an HGT-burden predictor of niche breadth alongside metal type diversity. The rationale is that a broader accessory genome may be the underlying driver of both traits. This was not tested. ₅₄
• Future Direction #5 proposes site-level validation. It would test whether MicrobeAtlas genera with high metal type diversity are enriched in metal-contaminated sample metadata (where available: mining-impacted, biosolid-amended, estuarine) versus clean reference environments. The report says this would directly validate the environmental co-selection hypothesis. It remains proposed. ₅₄
• Future Direction #6 proposes grouping metal types by genetic co-occurrence (e.g., Cu-Zn co-resistance vs Hg-As). It would then test whether specific co-resistance combinations predict niche breadth more than the raw count. ₅₄
• Future Direction #7 proposes MCMCglmm with a Poisson or negative binomial family to test the reverse regression direction: whether metal type count, as a response variable, is predicted by niche breadth with phylogenetic structure accounted for. It was not performed. ₅₄
• Future Direction #8 proposes re-running the niche breadth computation with a finer environment taxonomy built from MicrobeAtlas sample metadata, for example splitting “aquatic” into marine, freshwater and estuarine. This would test whether category heterogeneity inflates B_std for certain genera. ₅₄
• Future Direction #9 proposes manual BLAST verification of 100 randomly sampled annotated metal AMR gene clusters against reference sequences. This would quantify the hidden Markov model (HMM)-based false-positive rate for each metal type category. No such validation is reported. ₅₄
• Future Direction #11 proposes depositing notebooks, derived data CSVs and a conda environment lockfile in a Zenodo repository, so that all PGLS and Pagel’s λ results can be verified independently. The derived files (data/*.csv, figures/*.png) would be archived under a citable DOI before manuscript submission. No deposit is reported. ₅₄

Future Direction #10, ENIGMA field validation, is marked fully addressed, with Tracks A and B complete. The report restates Track A as 1,624 groundwater samples in which metal-diverse genera are significantly more prevalent (Spearman ρ = +0.112, p = 0.0019; Mann–Whitney p = 0.007). It restates Track B as a full amplicon pipeline run on all 133 PRJNA1084851 samples (24,295 OTUs; 133/133 QC pass). In Track B, CWM metal-type diversity is higher in plume wells FW215/FW216 (Kruskal–Wallis p = 0.0001) and increases with time after carbon amendment (Spearman ρ = +0.383, p < 0.0001 across all wells; ρ = +0.576, p = 0.0001 in FW216 alone). The report concludes that both tracks independently confirm the PGLS prediction in field data. That conclusion is stronger than the evidence supports. Track A’s groundwater-specific fold-enrichment test was null, Track B CWM covers 16.8% of reads, and neither track establishes the proposed metal-selection mechanism. ₅₄

The report’s reviewer-response table records the status of each critique. Environment-category validity (Q1a) is only partially addressed: leave-one-out was done (S2), but sub-category splitting requires new metadata. Primer bias and differential sequencing depth (Q1b) are not correctable with available data and are only acknowledged. On tree sensitivity (Q2b), GTDB r214 is called the best available tree. An alternative-tree comparison was not performed, which the report calls a known gap. Within-genus heterogeneity was tested via S3 and was non-significant. The λ discrepancy (Q2c) is described as directly addressed, with Δλ = 0.079 expected given the predictor’s own phylogenetic signal. On the genome-count covariate (Q3a), the response states that R1 and S4 both use gls() + corPagel (PGLS), not OLS, and that β remains significant after the PGLS covariate. On metal-type classification (Q3b), the single-metal question is called addressed because all 7 S1 exclusions were positive and no one metal is the driver. Formal AMRFinderPlus validation was not performed; only spot checks were done. All seven S1 exclusions were also non-significant, so the “addressed” status for the single-metal question rests on direction alone. ₅₄

On genome size (Q4a), the response states that the formal test was performed as R5. Metal types had β = +0.022 (p = 3.6×10⁻⁴) independent of genome size, with ΔAIC = 10.8. On causation and directionality (Q4b), the response states that the 3 scenarios are not distinguishable and that the cross-sectional design limitation is explicit. It adds that causal phrasing was removed from Key Findings and Interpretation. On multiple comparisons (Q5a), it reports that 47 tests were tabulated and BH-FDR was applied across all of them. The primary result survives both Bonferroni (α/47=0.0011) and FDR (q=0.003). On rarefaction and within-genus variance (Q6b), the response states that S3 confirms genus means are not inflated by outlier species and that R2 rarefaction shows a consistent positive direction. This sits uneasily with Future Direction #1’s call for pangenome rarefaction, so it is unclear which sampling-depth control the report considers outstanding. ₅₄

On data archiving (Q7a), the response states that a public Zenodo archive has not yet been prepared. The materials are in the institutional repository only, and Zenodo archiving was added as Future Direction #11. On PGLS ordering (Q7b), the response states that scripts/test_pgls_ordering.py generates synthetic data, scrambles the row order and verifies that the sorted model recovers the correct λ and β. The test passes. On post-hoc analyses (Q8a), the response states that all robustness and sensitivity analyses are labelled exploratory. It says the primary confirmatory result (6 PGLS + Bonferroni) stands independently. On preregistration (Q8b), the report acknowledges that the plan was not preregistered and that the confirmatory/exploratory distinction was documented post hoc. On a concrete ENIGMA prediction (Q8c), it calls the item fully addressed. It cites top-quartile metal-diversity genera as 31% more prevalent in groundwater (Mann–Whitney p=0.007). It also cites CWM metal diversity as significantly higher in contaminated FW wells (KW p=0.0001) and increasing with carbon amendment time (ρ=+0.38, p<0.0001) across 133/133 ENIGMA ORFRC samples. The microcosm predictions themselves remain untested. ₅₄

Fitness Modules × Pangenome Conservation

The baseline core rate is already ~81.5%, limiting the maximum observable enrichment; consequently, the +4.5 percentage-point difference to 86% is statistically significant but represents a modest absolute effect. ₅₅

The pangenome-linked analysis covers a 29/32 organism subset because Cola, Kang, and SB2B lack pangenome links; their species had too few genomes in GTDB for pangenome construction. ₅₅

Upstream ICA module membership uses |Pearson r| >= 0.3 with a maximum of 50 genes per module, so this threshold can influence which genes are classified as module members and therefore the conservation composition. ₅₅

The >90% core and <50% core cutoffs used to classify core, mixed, and accessory modules are convenient classification thresholds rather than biologically motivated boundaries. ₅₅

Essential genes are excluded from all modules because ICA requires fitness data and essential genes lack transposon insertions; therefore, the reported module conservation profiles cannot characterize essential-gene modules. ₅₅

Community Metabolic Ecology via NMDC × Pangenome Integration

The 175 metabolomics samples derive from multiple NMDC studies, so metabolomics technical heterogeneity could affect broader multi-study analyses, although 95% of H1 samples (125/131) came from one NMDC study, and the second study contained only 6 samples. The report recommends study-level random-effects models when broader multi-study coverage becomes available. ₅₆

All 33 Freshwater samples lacked paired metabolomics, so H1 was effectively a soil-only test. Whether BQH dynamics operate at the same scale in freshwater communities remains untested. ₅₆

Abiotic features, including pH, temperature, and total organic carbon, were absent as usable measurements: all were NaN in the 174-sample analysis matrix. Consequently, partial correlations controlling for environmental gradients could not be performed, and abiotic variables may confound H1 and H2. ₅₆

GapMind measures genomic potential rather than expression. The presence of pathway genes does not show that biosynthesis is active; metatranscriptomic data would be needed to test whether expressed pathway completeness correlates more strongly with metabolite pools. ₅₆

Metabolite-to-pathway matching used string-based compound-name matching. An isoleucine substring collision with the leucine pattern was corrected in NB04 cell-14 using first-match-wins, and the reported results use the corrected run. After the fix, isoleucine was included as a 13th testable pathway (r = −0.057, ns). KEGG compound IDs had only a 2% annotation rate; cysteine, histidine, and lysine remained untestable because corresponding compounds were absent from detections. ₅₆

Shikimic acid and 3-dehydroshikimic acid were used as metabolomics proxies for chorismate, but they are upstream intermediates rather than chorismate itself. Their concentrations therefore reflect precursor availability rather than the chorismate pool directly, making the chorismate correlation (r = −0.038) especially uncertain. ₅₆

Approximately 1,352 Centrifuge taxa matched multiple GTDB clades within the same genus. One representative clade was selected by alphabetical tiebreaking on gtdb_species_clade_id; these genus-proxy-ambiguous taxa accounted for approximately 6.5% of mapped abundance. ₅₆

Because of sample-size imbalance, glutamine (n = 4) and proline (n = 9) could not be tested: their metabolomics coverage was insufficient, even though they would be informative given their biological importance. Methionine also had limited power (n = 18, q = 0.117), and the report identifies larger metabolomics datasets as necessary to evaluate these pathways more reliably. ₅₆

The habitat identity of the 61 Unknown-ecosystem samples is unresolved; the report proposes, but does not establish, that they likely represent a mix of soil subtypes or sediment environments. Resolving their habitat identity through NMDC ENVO annotations or study metadata could improve ecosystem comparisons and reveal finer-scale metabolic niche structure. ₅₆

NMDC Context Audit

The provenance classes are inferred from schema, table properties, tenant metadata, and prior project usage rather than an ingestion manifest, because no ingestion manifest is exposed in the catalog. ₅₇

Descriptions for kbase.nmdc_* databases are access-restricted by ForbiddenException, so steward-authored notes, if any, could not be captured. ₅₇

Completeness is assessed relative to snapshot timestamps and was not tested by comparison with live upstream NMDC or NCBI record counts; such comparison would require external API calls and was out of scope. ₅₇

The report does not directly observe users choosing the wrong resource. Its interpretation that label overload drives sub-optimal selection is based on gap analysis and prior-project reuse skew, so that causal claim remains inferred. ₅₇

The audit enumerates 7 real, maintained resources among 20 NMDC-named database entries, but also records aliases, test databases including globalusers.nmdc_core_test*, a phantom kbase_nmdc_neon with 0 tables, and broken user copies including mamillerpa/my.nmdc_flattened_biosamples with a dangling Iceberg pointer. ₅₇

Earlier in-repository knowledge covers nmdc_arkin well but leaves the other six resources thinly documented or undocumented. That knowledge is the nmdc_arkin section of the pitfalls document plus the discoveries document. The report presents this as consistent with its gap analysis. ₅₇

Pangenome Openness Analysis

The sample is limited to species with both pangenome statistics and ecotype-analysis results. The report does not give a sample count. ₅₈

Pangenome openness is a single summary metric and may not represent the full complexity of pangenome structure. ₅₈

Environment and phylogeny effects were derived from partial correlations, which may not fully disentangle confounded variables. ₅₈

The upstream ecotype analysis may have limited statistical power for some species with few genomes. ₅₈

PaperBLAST Data Explorer — Literature Coverage Bias in Protein Sequence Space

• Text mining is not equivalent to functional characterization: a gene mentioned in a paper may be tangential to the study, and the 65.6% of genes with one paper may include incidental mentions. ₅₉
• PaperBLAST mines PubMed Central full-text articles, so papers behind paywalls are missed; this creates systematic bias against fields and journals with lower open-access rates. ₅₉
• Domain classification is approximate. The report states that heuristic organism-to-domain mapping classified 35% of organisms as Unknown, but this conflicts with its domain table. There, Unknown holds 14,798 of 27,718 organisms, and 35.0% is the share of genes, not organisms, classified Unknown. The report proposes a formal taxonomy lookup to improve accuracy. ₅₉
• Clustering identity thresholds are arbitrary: the 50% identity cutoff used for “protein family” is conventional and does not represent a universal biological boundary. ₅₉
• SwissProt coverage is partial: only 19% of SwissProt is in PaperBLAST, likely because many entries lack matching PMC full-text; curated knowledge therefore exists for proteins that PaperBLAST cannot connect to. ₅₉
• The analysis has no negative controls and cannot easily distinguish genuinely unstudied genes from genes whose literature was missed by text mining. ₅₉

Metabolic Capability vs Dependency

• Fitness Browser coverage limited Tier 1 to 7 of 48 organisms with matching GapMind data, and these model organisms had near-complete core genomes that compressed the conservation-validation signal. ₆₀
• GapMind covered 80 pathways—18 amino acid biosynthesis pathways and 62 carbon-source utilization pathways—but did not assess cofactor biosynthesis, lipid metabolism, or secondary metabolism. ₆₀
• Laboratory fitness does not capture the full range of natural selective pressures; the report specifically notes that 28,017 genes were costly in laboratory conditions but conserved in nature. ₆₀
• KEGG-based mapping can miss genes lacking KEGG annotations or carrying incorrect annotations, potentially underestimating the number of genes assigned to a pathway. ₆₀
• Ecotype counts depend on hierarchical clustering with a fixed 50% maximum-distance cut; different thresholds would produce different counts, so the median of 4 should be treated as an order-of-magnitude estimate. ₆₀
• Correlations were controlled for genome count and checked within taxonomic groups, but full phylogenetic independent contrasts were not computed. The related ecotype_analysis project found that phylogeny dominates gene content in 60.5% of species, and the report cautions that metabolic ecotypes may partially reflect intra-species phylogenetic structure. ₆₀
• Species with more sequenced genomes are more likely to show pathway variation and more ecotypes; partial correlations reduce but do not eliminate this sampling bias. ₆₀
• The median-based importance threshold is circular for condition-specific reclassification, because applying a median threshold to each subset can cause pathways to cross it. An independent validation set, such as known essentials from essential_metabolome, would provide a more defensible threshold. ₆₀
• The proposed correlation between metabolic ecotypes and AlphaEarth environmental niche breadth was not executed. AlphaEarth embeddings covered 28% of genomes, or 83K/293K, which would limit the available sample size. ₆₀

PGP Gene Distribution Across Environments & Pangenomes

Environment classification was conservative and noisy: only 1,637 species (5.9% of species with an environment label) were classified as soil/rhizosphere dominant, while 291,279 genomes had isolation-source metadata and 93.5% were classifiable. The report states that NCBI is biased toward clinical and host-associated sampling and that the acdS and pqqC enrichment effects may therefore underestimate true rhizosphere enrichment. ₆₁

PGP detection relied on Bakta gene annotations matching exact gene names such as nifH, acdS, and pqqC. Product-only annotations and variant gene names could be missed, particularly for less-characterized PGP genes. ₆₁

Gene-cluster annotations were not functionally validated: truncations, frameshifts, and pseudogenization were not filtered, so a cluster annotated as pqqC was not necessarily functional. ₆₁

ipdC was rare, occurring in only 214 of 11,272 species (1.9%), which limited statistical power for the stratified H4 analysis; the soil reversal (OR = 0.30, p = 0.02) should therefore be treated as hypothesis-generating rather than conclusive. ₆₁

GapMind tryptophan and tyrosine completeness scores may proxy for overall metabolic pathway completeness. The report states that genome size, COG (Clusters of Orthologous Groups functional-category) coverage, or total pathway count should be controlled to separate aromatic-pathway-specific effects from general metabolic capacity. ₆₁

The report also notes that co-occurrence does not establish physical linkage: whether pqqC and acdS occupy the same genomic island, operon, or separate loci remains unresolved. It proposes operonic-context analysis, deeper nifH ecological stratification, a controlled ipdC model, focused hcnA–hcnC phylogeny, and comparison with commercial inoculant strains as next steps. ₆₁

Pan-Bacterial Anti-Phage Defense Arsenal

CRISPR-Cas prevalence (96 %) is inflated by permissive EggNOG description matching and should be read as an upper bound. The Pfam Cas1 (PF01867) hit is present in only 55 % of species. The combined EggNOG-or-Pfam presence set used for syndrome analysis prioritizes recall, so CRISPR-Cas syndrome specificity may be reduced even though the significance results are not expected to be invalidated. ₆₂

The DISARM accessory-enrichment result is unreliable because PF00176, the DrmB SNF2 helicase Pfam, is a widespread housekeeping-helicase family. Its inclusion inflates the DISARM cluster set with non-DISARM SNF2 hits, which biases the DISARM core/auxiliary/singleton distribution toward core. A PADLOC MacSyFinder-style HMM-plus-context refinement is needed; the report states that arms-race and syndrome results involving DISARM are unaffected because they use species-level presence/absence rather than per-cluster classification. ₆₂

Retron detection uses the broad RVT_1 Pfam PF00078. The stringent call requires at least one other narrow defense system and is therefore a defense-context filter rather than a retron-specificity filter; it reduced candidates from 15,109 to 15,098, dropping 11 species. Results should be interpreted as reverse-transcriptase candidates in defense-syndrome context, not as characterized retron systems. ₆₂

The eggNOG-description prophage classifier deliberately captures broad module candidates including integrase, holin, endolysin, CI-like repressor, and tail proteins, which can also match non-phage bacterial genes. Consequently, n_prophage_modules saturates at 7 for 35% of species; the primary arms-race predictor is instead n_prophage_clusters, an unbounded continuous count of matched clusters. n_prophage_modules serves only as a coarse categorical secondary predictor. ₆₂

The ≥5-genome filter reduced the analysis set to 7,323 of 27,690 pangenome species, or 26%, biasing the analysis toward well-sampled, culturable, high-priority organisms. Extrapolation to the full 293K-genome tree, including many environmental MAGs, remains untested. ₆₂

The arms-race analysis controls for phylum categorically but does not use a phylogenetically corrected regression. Its partial correlations and negative-binomial GLM are therefore consistent with, but do not formally establish, a phylogenetically independent arms race. ₆₂

The negative-binomial GLM fixes dispersion at alpha = 1.0 rather than estimating it. Both focal coefficients remained significant at p < 0.001 with wide margins, but standard errors may be slightly under- or over-stated; refitting with an estimated alpha would sharpen inference. ₆₂

Polyhydroxybutyrate Granule Formation Pathways — Distribution Across Clades and Environmental Selection

NMDC cross-validation correlations were statistically significant but small (|rho| < 0.12). NMDC abiotic measurements are point-in-time values rather than measures of temporal variability, and the studies are biased toward terrestrial and soil environments, limiting direct testing of the feast/famine hypothesis. ₆₃

PHA synthase class analysis failed because Pfam accession IDs were not mapped to eggNOG domain names; mapping PF00561 to Abhydrolase_1 for Class I/II and PF07167 to PhaC_N for Class III/IV is required to classify the 11,792 phaC clusters. ₆₃

Environment metadata are sparse: 34.9% of species have “other_unknown” as their primary environment, so environmental enrichment may be underestimated. AlphaEarth coverage is also limited: 83K/293K genomes, or 28%, have embeddings, and the 2,008 species analyzed represent 7.2% of total species diversity, potentially biasing niche-breadth analyses toward better-sampled lineages. ₆₃

Because phaA and phaB participate in general metabolism beyond PHB biosynthesis, the 46.5% precursors-only category likely overestimates partial PHB capability. The pangenome includes complete genomes and metagenome-assembled genomes (MAGs), with variable genome quality and gene-detection rates. ₆₃

PHB presence is correlated with phylogeny and genome size, and the PHB–niche-breadth association falls from raw rho = 0.106 with p = 1.77e-06 to partial rho = -0.047 with p = 0.037 after controlling for genome size. The report’s interpretation states that PHB environmental enrichment persists within all genome-size quartiles with 1.4–4.6x enrichment and all p < 1e-11; this conflicts with its own quartile table, where Q1 has p = 1.18 x 10^-11, which is not below 1e-11, so both statements are retained here without reconciliation. Shared ancestry also remains a potential confound: phylogenetic logistic regression or phylogenetic independent contrasts are still needed to account for shared ancestry. ₆₃

PHB has functions beyond carbon storage, including stress resistance, redox balance, and cryoprotection, so environmental variability is supported as a selective force but is not necessarily the sole driver of PHB distribution. The HGT inference is likewise based on phylogenetic discordance and core/accessory status; a directly reconstructed phaC gene tree is needed to identify incongruent branches. ₆₃

KBase Data Lakehouse: Common Pitfalls & Gotchas

The document repeatedly warns that table schemas, namespace availability, access permissions, API behavior, database contents, and naming conventions can change. Queries should therefore begin with live catalog and schema discovery rather than copied historical SQL, and archived reports and notebooks should be treated as historical records rather than automatically valid executable instructions. ₆₄

Several recommendations are project-specific safeguards rather than universal thresholds. Examples include the held-out-feature Jaccard boundaries of 0.5 and 0.3, bootstrap sizes of 250–400, the ICA (independent component analysis) limit of at most 40% of the number of experiments, the |r| >= 0.3 module-membership rule, the ANI cap of <=500 genomes, and the 50–100 genome extraction batches for large species. These values should be reported with their analysis context rather than generalized without validation. ₆₄

The document records apparent method- or interface-specific inconsistencies that require care: direct Spark accepts a quoted species ID containing --, whereas the REST API rejects queries containing --; one section describes per-organism Fitness Browser convenience tables as long format while another database-specific note says they may be pre-pivoted and have non-standard schemas; and NMDC schema notes differ on whether abiotic numeric fields arrive as strings or doubles. The stated resolution is to inspect the live schema and query interface before relying on either description. ₆₄

Coverage limitations are not equivalent to biological absence. Missing AlphaEarth coordinates, sparse NCBI environment attributes, BacDive–GTDB naming differences, unmatched species concepts, orphan pangenomes, absent tables, and workflow-specific file identifiers can all produce zero or incomplete joins without indicating that the underlying biology is absent. ₆₄

Plant Microbiome Ecotypes — Compartment-Specific Functional Guilds and Their Genetic Architecture

Plant-compartment labels were inferred primarily from NCBI isolation_source metadata, whose quality and coverage vary; only 7,995 of 293,059 genomes, or 2.7%, had plant-associated annotations, and the endophyte category contained only 29 species, below the 30-species threshold. The report states that more targeted isolation could expand the endophyte compartment to enable a full four-way compartment comparison. Plant association was determined for 293,059 GTDB r214 genomes from NCBI isolation_source regex matching as the primary source, cross-validated against the ncbi_env EAV table (695 genomes upgraded) and BacDive isolation records; BacDive identified 2,482 plant-related strains but contributed 0 new genome upgrades beyond ncbi_env, which the report reads as complete overlap between these metadata sources for the plant-associated genome set. Species-level classification used majority vote, with mean confidence 0.88. ₆₅

The marker set is literature-curated and gene presence does not establish expression or phenotype. T3SS, T6SS, flagella, chemotaxis, biofilm, quorum sensing, and secretion functions can support colonization, interbacterial competition, beneficial symbiosis, or pathogenicity, and negative controls were frequently classified as dual-nature. In the NB08 negative controls, Escherichia, Salmonella and Clostridioides were 100% classified as dual-nature, and Staphylococcus was 98%. Combining this with the genome-size confound, the report concludes that the NB02 dual-nature rate of 60.3% substantially overestimates genuine plant-adapted duality. It adds that the NB07 composite scoring (25.2% dual-nature) only partially addresses this, because it depends on arbitrary scoring weights. The report documents T3SS classification as context-dependent (adversarial-review issue I3). It cautions that classification as beneficial or pathogenic from genomic markers alone is insufficient: the same T3SS that delivers pathogenic effectors in Pseudomonas syringae facilitates beneficial nodulation signaling in Rhizobium, and context (host genotype, environmental conditions, community composition) determines the outcome (Osayande et al. 2025). RNA-seq or experimental validation is required to determine whether PGP and pathogenic marker sets are co-expressed under the same conditions. The report adds that presence of both marker types does not confirm simultaneous expression. Such validation is needed to tell genuine lifestyle flexibility from annotation artifacts (e.g., T6SS serving inter-bacterial competition rather than pathogenicity). ₆₅

Genome size correlates with total marker count at r = 0.44, p < 1e-300. Among plant-associated species, the dual-nature rate was 54% in the lowest genome-size quartile versus 87% in the top three quartiles; approximately 33% of dual-nature species shifted to neutral after genome-size normalization in an earlier sensitivity analysis. The report states that this gradient was measured directly on refined cohort data and is observational, not model-dependent; the 78.7% refined dual-nature value therefore contains a real genome-size gradient and remains dependent on marker definitions and scoring assumptions. The earlier sensitivity analysis described the normalization effect as meaningful but not eliminating the dual-nature pattern. The limitations recap states that all 50 novel OGs retain positive plant-association coefficients with genome size as a covariate (resolved in NB14). That statement is superseded by the report’s per-species analysis, in which 48/50 converged and were significant, while COG2351 and COG5343 had singular design matrices and do not count as successful adjusted fits. ₆₅

Phylogenetic control remains incomplete. The original genus-level fixed-effect logistic regression is described in the report’s results text as having “failed to converge for 0/14 markers”, whereas the report’s data table lists the same NB10 output as “0/14 converged”. The two wordings are inconsistent. The report’s limitations recap (item 13) states it explicitly: NB10 attempted genus-level fixed-effects logistic regression for all 14 refined markers, and all models failed to converge (0/14). That recap and the data table’s “0/14 converged” both say no model converged. The results-text phrasing is the outlier, and the report does not reconcile it. A third passage, in the Phase 1 compartment results, says the planned genus-level logistic regression was not completed because of a code error (an undefined logit import), so the H1 compartment enrichments are not phylogenetically controlled beyond the PERMANOVA test itself. The report explains the failure as plant association being heavily confounded with genus-level taxonomy, so that standard maximum-likelihood logistic regression cannot fit; the cluster-robust GLM is a practical genus-level analogue rather than a full phylogenetic generalized linear mixed model. The limitations section adds that RESEARCH_PLAN.md specified genus-level fixed effects, but Phase 1 used phylum-level control because of computational cost. It says NB10’s genus-level logit “failed for 0/14 markers”. It describes the partial resolution in NB14: a scikit-learn L1-regularized logit with top-20 genus dummies and log₁₀(genome size) as covariates produced convergent fits, with “9/17 markers” having bootstrap 95% CIs excluding zero. The results text reports this result as 9/14, so its denominator is inconsistent within the report. The canonical 9/14 result supersedes the recap’s 9/17 denominator. The 9/14 figure rests on 100-replicate L1-bootstrap CIs. These are biased near zero and only suggestive. The later genus-cluster-robust GLM found 8/14 markers significant. Phylogeny-versus-ecology variance decomposition remains future work, and compartment profiling uses PERMANOVA rather than logit as the primary test. A full GTDB-tree covariance analysis was not feasible because a dense 25K × 25K covariance matrix would be intractable in memory. The report names sparse approximations or species subsampling as the practical path. The within-genus shuffle used only 200 permutations, which is under-resolved for p ≤ 0.005, and a full permulation-style analysis remains future work. ₆₅

The raw OG enrichment was saturated at 5,341/5,671 significant tests (94.2%). The report cautions that this mostly reflects taxonomic divergence, meaning genome-wide compositional differences, rather than plant-specific functional enrichment. Only the 50 OGs surviving phylogenetic control represent candidate plant-interaction genes. Although 48/50 gene families survived phylum and genome-size control, two models were singular, and the earlier phylum-level models had convergence warnings, which the report attributes likely to quasi-separation when some phyla are entirely plant-associated or entirely non-plant. The report states that these warnings mean the reported p-values and odds ratios should be treated as approximate. Finer-grained control proved intractable. Family-level logistic regression (NB08) found insufficient within-family variation for all 10 OGs tested, indicating that plant association is taxonomically clustered at the family level. The planned genus-level fixed effects were not executed in Phase 1 because of computational constraints. The report’s practical reading 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. The causal direction between plant association and enriched functions therefore remains unresolved. ₆₅

The complementarity analysis was limited by 80 GapMind pathways, a potentially over-stringent core-level completeness threshold, and loss of 62/322 genera, or 19.3%, in the NMDC taxonomy bridge, which matched 260 of 322 genera (80.7%) to GTDB. The corrected effect is small, with Cohen’s d = −0.39 for prevalence weighting, but unmatched genera could include complementary partners. The report states that the impact of the 62 lost genera on the H3 complementarity analysis is unknown. If they include key complementary partners, the observed redundancy signal could be artifactual. The adversarial-review issue table records the NMDC taxonomy bridge as 80.7% matched (issue I5), with sensitivity to unmatched genera deferred to future work. The report states that the 80-pathway GapMind level may be too coarse, and that its redundancy result does not rule out a complementarity signal at finer-grained metabolic resolution than the 80 GapMind pathway categories can detect. It adds that reaction-level complementarity or substrate-specific analysis could reveal patterns masked by the current pathway-level aggregation. The report records the max-aggregation inflation (d = −7.54) as partially resolved in NB14 by the prevalence-weighted re-test (d = −0.39, same direction). The core-level completeness threshold, which scored 0% across all compartments, remains a separate open question. ₆₅

HGT evidence is indirect because GeNomad mobile-element annotations were unavailable. The transposase/integrase co-occurrence signal, lower overall marker singleton ratio, and MGnify genus-level mobilome enrichment do not provide a single genome-level estimate of horizontal transfer. MGnify BGC, mobilome, and defense data were available only for the soil biome, so those comparisons do not directly contrast rhizosphere with bulk soil. The current proxies are transposase/integrase singletons and MGnify mobilome aggregates. The report proposes replacing them with direct mobile-element annotations from GeNomad mobile-element predictions, to strengthen the mobilome conclusions and resolve the genus-versus-genome scale-dependent signal. ₆₅

The MGnify and pangenome classifications have only 11.7% Jaccard overlap because they measure different phenomena: isolation metadata versus metagenomic detection. The report notes that this low concordance limits the strength of cross-validation conclusions. The report attributes the low concordance to data-source and sampling differences rather than classification error. It notes that 93 pangenome plant genera were found in MGnify soil but not rhizosphere, which suggests broader environmental associations. It also counts 390 pangenome-only plant genera not found in MGnify rhizosphere, which it suggests likely include opportunistic soil bacteria isolated from plants but not rhizosphere specialists, and 341 MGnify-only rhizosphere genera that may include uncultured lineages absent from GTDB; both explanations are tentative. The subclade analysis is additionally limited by phylogenetic-tree coverage, with only 18/65 candidate species represented. Expanded tree coverage, accessory-gene profiling, gene-content trees, and direct mobile-element annotations are needed to resolve these uncertainties. The current subclade method is core-phylogeny-based. Two analyses remain outstanding: accessory-gene presence/absence profiling as an alternative subclade definition, and gene-content-versus-species-tree comparisons to identify HGT-driven incongruences. Both require larger gene-cluster-by-genome data pulls than the project executed. The report also notes that expanding tree coverage would directly increase the power of any within-species phylogenetic analysis. ₆₅

Prophage-AMR Co-mobilization Atlas

Distances were calculated from ordinal gene positions parsed from gene_id formats rather than base-pair coordinates, so the reported gene distances may differ from true genomic distances. ₆₆

Prophage markers were identified by keyword and Pfam matching in bakta_annotations rather than by dedicated prophage-prediction tools such as PHASTER or geNomad. This approach may include false positives, including phage-defense systems, and may miss divergent prophages. ₆₆

The co-localization analysis sampled 20 genomes per species for the 100-species analysis rather than examining all 293K genomes; exhaustive analysis could strengthen the findings. ₆₆

Core and accessory labels depend on species-level pangenome calling with motupan, so the same gene can have different conservation status in different species. ₆₆

The fitness analysis was limited by the overlap between the 48 fitness-browser model organisms and GTDB pangenome species, preventing an H3 comparison. ₆₆

The H2 association between prophage density and AMR breadth is correlational and does not prove phage-mediated AMR transfer; species with open pangenomes may independently accumulate both prophages and AMR genes. ₆₆

Prophage Gene Modules and Terminase-Defined Lineages Across Bacterial Phylogeny and Environmental Gradients

The analysis identifies prophage-associated genes through eggNOG functional annotations rather than dedicated prophage tools such as geNomad or VIBRANT. Consequently, the reported near-universal prevalence may include domesticated remnants and bacterial homologs of phage genes, including bacterial integrases, and the false-positive rate is uncharacterized. ₆₇

The finding that all 27,702 species carry prophage annotations far exceeds Touchon et al.’s (2016) estimate that ~46% of bacterial genomes harbor prophages. The report attributes this discrepancy to its broader definition: it counts individual prophage-associated gene clusters, including domesticated remnants, rather than intact prophages. The universal-prevalence result is therefore not an intact-prophage prevalence estimate. ₆₇

Genome size was the dominant predictor of prophage burden, with rho=0.717, and residual confounding cannot be excluded despite genome-size stratification and partial correlations because larger genomes contain more genes of all types. ₆₇

NCBI isolation_source metadata were sparse and inconsistently labeled. The 10 environmental categories collapse substantial within-category variation, and the other_unknown category contained 9,659 species (35%), limiting statistical power. ₆₇

NMDC prophage-burden inference was indirect. Genus-level prophage burden scores derived from the pangenome were applied to the 6,365 NMDC metagenomic samples via two-tier taxonomy mapping (a GTDB metadata bridge plus taxonomy_dim genus matching), so burden was not directly measured in the metagenomes. The inference assumed genus-level conservation of prophage content, an assumption that may fail for recently acquired or lost prophages. The approach had been validated for the PHB granule ecology project but had not been independently validated for prophage genes. ₆₇

Module co-occurrence was tested in only 15 species, one per phylum, providing broad phylogenetic coverage but limited within-phylum replication. Head morphogenesis was testable in only 6 species because of sparsity, and the 200-permutation null model was considered adequate but not exhaustive for estimating Z-scores. ₆₇

Only 28% of genomes had AlphaEarth environmental embeddings, producing a biased subsample toward clinically and environmentally well-sampled lineages. ₆₇

The report had not formally quantified the unique variance explained by environment after controlling for phylogeny and genome size simultaneously. Each predictor was tested independently, and the report proposes sequential or partial PERMANOVA as future work. ₆₇

The reported environmental correlations are cross-sectional. Whether prophage burden changes over time within the same site remains untested. The report proposes using NMDC longitudinal samples, if available, to provide causal evidence beyond cross-sectional correlations. ₆₇

Carbon Source Utilization Predicts Ecology and Lifestyle in Pseudomonas

The report identifies sampling bias as a limitation: P. aeruginosa comprised 53% of all genomes (6,760/12,732) because of clinical importance, while many environmental species had fewer than 10 sequenced genomes. This imbalance affects the power of species-level comparisons. ₆₈

Isolation-source classification was based on free-text keywords and introduced approximately 6.7% “unknown” and approximately 29.1% “other” classifications; misclassification could attenuate the ecological signal. Species-level majority-vote assignment also obscures genuinely generalist species that inhabit multiple environments. ₆₈

GapMind’s 62 carbon pathways cover common carbon sources but omit genus-specific capabilities, particularly aromatic degradation pathways such as toluene, naphthalene, and benzoate degradation that are central to P. putida ecology. The report proposes extending the analysis with aromatic-pathway modules, including KEGG modules, to test whether they provide stronger environmental signal. ₆₈

Phylogenetic confounding is substantial: the dominant PCA signal separates subgenera rather than lifestyles, and the analyses do not explicitly control for phylogenetic non-independence among species. The report proposes phylogenetic generalized least squares (PGLS), a regression method that accounts for phylogenetic relatedness, or phylogenetic logistic regression using the GTDB species tree. ₆₈

The report further notes that the moderate classifier accuracy may reflect small sample sizes per class, overlap between related environments such as soil and rhizosphere, and the coarse resolution of the 62 GapMind pathways. Several species, such as P. fluorescens and P. putida, are isolated from multiple environments, so the report proposes a within-species analysis of pathway variation to test whether it reveals metabolic ecotypes, meaning subpopulations adapted to different niches. Environment prediction in this report was conducted at species level. The report proposes moving to genome-level prediction using the full 789K genome-pathway matrix with genome-specific isolation sources, to increase statistical power. It also proposes cross-referencing carbon pathway predictions with RB-TnSeq (random barcode transposon sequencing) fitness data from the kescience_fitnessbrowser (Fitness Browser) collection. That comparison would experimentally test which pathways are functionally important on specific carbon sources. None of these analyses is reported as performed. ₆₈

Condition-Specific Respiratory Chain Wiring in ADP1

• NDH-2 has no growth data, so the central prediction that NDH-2 compensates on glucose cannot be directly tested in the current dataset. ₆₉
• The stoichiometry analysis uses theoretical pathway biochemistry rather than measured flux distributions. ₆₉
• The cross-species comparison has only 4 organisms without NDH-2 and is insufficient for statistical significance. ₆₉
• The NDH-2 gene search may miss orthologs because annotation varies among Fitness Browser organisms. ₆₉
• ACIAD3522 may not be a true respiratory NADH dehydrogenase; its NADH-FMN oxidoreductase annotation could represent another metabolic function. ₆₉
• Text matching on gene descriptions likely produced false-positive NDH-2 calls, especially in organisms with incompletely annotated Complex I subunits; for example, pseudo3_N2E3 has 8 “NDH-2” hits that are probably a Complex I operon. KO-based identification using K03885 would be more reliable. ₆₉
• The planned pangenome KO co-occurrence analysis was not performed, leaving NDH-2/Complex I co-occurrence across Acinetobacter species untested. ₆₉
• The proposed wiring model can be tested by constructing an NDH-2 deletion mutant, measuring NADH/NAD⁺ ratios on each carbon source, expanding the cross-species K03885 and K00330–K00343 analysis across 27K species, characterizing ACIAD3522, and reanalyzing quinate-versus-succinate proteomics for respiratory-chain proteins. ₆₉
• The central prediction is that NDH-2 deletion makes Complex I essential on glucose. Only an experiment can test it, and an ADP1 NDH-2 knockout would be the definitive test. ₆₉
• The flux-rate hypothesis predicts higher NADH/NAD⁺ ratios in quinate-grown cells than in glucose-grown cells. Metabolomics could test this directly, but the ratios have not been measured. ₆₉
• It remains unresolved whether ACIAD3522 is a true respiratory NADH dehydrogenase or a metabolic enzyme. Its acetate lethality and FMN cofactor suggest a specific role in acetate catabolism that needs further investigation. ₆₉
• The report has not established whether ADP1 upregulates Complex I on aromatic substrates or NDH-2 on glucose. It proposes reanalyzing the quinate-versus-succinate proteomics of Stuani et al. (2014) for respiratory-chain proteins to test this. ₆₉

SNIPE Defense System in the KBase Data Lakehouse Pangenome

The report states that eggNOG Pfam annotations may miss divergent SNIPE homologues, and that only 54/4,572 DUF4041 clusters showed Mug113 co-annotation, suggesting that many SNIPE nuclease domains may go undetected. DUF4041 is strongly associated with SNIPE but may also occur in non-SNIPE proteins. ₇₀

The AlphaEarth analysis covered 28.4% of genomes and was skewed toward environmental isolates with geographic metadata, limiting generalization to the full pangenome. The planned comparison of COG V defense-gene density could not be completed through the REST API because it required Spark Connect for a 93M × 132M row join. ₇₀

NMDC metagenome ecosystem analysis was limited because only 2 samples matched the API query format; the report identifies interactive MCP-tool queries as a way to improve coverage. PhageFoundry provides genome annotations in GenomeDepot format but no fitness/TnSeq tables for Klebsiella, despite its ManYZ annotations there, so Klebsiella-specific SNIPE or ManYZ fitness remains unmeasured in that resource; strain-specific mutant libraries in natural SNIPE-carrying strains such as K. pneumoniae NCTC9140 or HS11286, or curated external K. pneumoniae Tn-Seq datasets, would be needed to test SNIPE and ManYZ fitness directly. ₇₀

The report’s H1 assessment is supported across prevalence, accessory status, environmental association, and the proposed ManYZ trade-off, while H0 is rejected for taxonomic distribution and environmental niche association. However, the environmental result establishes statistical association rather than a specific causal habitat mechanism, and the archaeal Fitness Browser example does not establish the Enterobacterales phage-lambda mechanism in archaea. ₇₀

Soil Microbial Dark Matter and the Clay Shield Null Result

The negative out-of-sample R² diagnosis is incomplete. The three model families predict worse than the training mean, but the analyses do not distinguish among train/test distributional shift caused by spatial autocorrelation within GroupKFold folds, high-leverage outlier samples dominating test-fold mean squared error, and genuine unpredictability of functional gene counts from the measured predictors at global scale. Only the third explanation would support a strong biological null interpretation; the first two would indicate modelling failure. ₇₁

The GDI is a novel index without published precedent. Because GDI = Richness / (Mean_Completeness + 1), it can equal 902 even when there are zero genomes, because completeness = 0 makes the denominator 1. The index also conflates sampling gap and OTU richness, potentially allowing the richness term to dominate; separate reporting of richness and completeness with a two-dimensional scatterplot may be more interpretable. ₇₁

The Forest GDI = 902.36 versus Cropland GDI = 890.82 difference is 1.3%; without bootstrap confidence intervals, the two values are not meaningfully distinguishable. The report therefore recommends describing forest and cropland as jointly the highest-GDI biomes rather than implying a ranked ordering. ₇₁

The +0.8 pH unit discovery-bias gap requires a reverse-causality check. Alkaline soils may have fewer genomes in databases because fewer samples from those pH ranges were sequenced, rather than because alkaline microbes are harder to assemble or less studied. Controlling for the number of 16S samples in each pH bin before computing GDI would distinguish assembly or annotation gaps from sampling gaps. ₇₁

Pending analyses are: decomposing negative R² into distributional shift, outlier leverage, and true unpredictability using spatial blocking; computing rarefaction-corrected GDI after uniform 16S sequencing-depth correction; estimating bootstrap 95% CIs for biome-level GDI rankings; controlling pH discovery bias for the number of 16S samples per pH bin; and reporting Forest versus Cropland GDI with uncertainty. These NB05 validations require re-running from the BERIL Observatory 16S tables and kbase_ke_pangenome completeness data because no local CSV output is available. ₇₁

Soil Metal Concentrations Drive Functional Gene Shifts

All reported associations are observational. Co-contamination is a major confound because chromium, copper, lead, and zinc co-vary in many industrial soils; it is therefore unresolved whether individual COG–metal associations are metal-specific or reflect a generic response to multi-metal contamination. Partial-correlation models are planned to address this distinction. ₇₂

The reported db-RDA R² = 0.799 is conditional: project accession was removed before fitting metal predictors. If project effects explain most community COG variance, this value describes metals’ explanation of residual variance rather than total variance. The unconditional R² for metals alone was not reported and may be substantially lower; both values should be reported. ₇₂

The 2,355 discoveries among 3,915 implied tests (nine metals × 435 COGs) represent a 60% discovery rate, but the report cautions that metal co-contamination makes tests non-independent. Benjamini–Hochberg FDR correction may therefore be anti-conservative under positive correlation, and the true FDR may be higher than reported. The report names the partial-correlation model (NB05 item 3) as the correct remedy; this is the report’s methodological assessment, not a validated recalculation of FDR. ₇₂

Effect sizes have not been systematically reported across all 2,355 associations. The planned effect-size audit will examine the Spearman rho distribution and flag associations with rho < 0.05, because statistical significance may not imply substantial biological effect. ₇₂

The copper-specific analysis used a 10 km proximity criterion to match soil samples with KBase genomes. Some genomes designated as nearby may not be co-located with the copper measurements, so copper–COG attribution requires sensitivity analyses at 5 km and 20 km. ₇₂

The planned validation includes Moran’s I spatial-autocorrelation testing on residuals, with SEVM if spatial autocorrelation is significant; partial correlation of COG ~ Cr | Cu + Zn + Pb; mechanistic classification of significant COGs into resistance, stress, membrane, energy, and unknown categories; reporting unconditional db-RDA R²; and copper proximity sensitivity at 5 km and 20 km. ₇₂

NB05 items 1–4 require rerunning analyses from the kescience_mgnify and kbase_ke_pangenome Spark tables. No local CSV containing the Spearman rho values or model residuals is available, so these validations cannot be completed without Spark access. ₇₂

T4SS–CAZy Environmental HGT

The observed T4SS co-localization with GT2 glycosyltransferase cassettes and phylogenetic incongruence in the GT2 gene tree provide positive evidence for 32 cross-phylum HGT events. The absence of plasmid-borne CAZy genes in the ICEfinder analysis, together with 10× higher MGE density in T4SS-positive genomes, is consistent with chromosomal or integrative mechanisms involving IMEs or integrative conjugative elements rather than plasmid mobilization. These results support the hypothesis that T4SS machinery mediates environmental dissemination of carbohydrate-active enzyme diversity across phylogenetically distant bacteria, but all associations are observational and require experimental validation for mechanistic confirmation. ₇₃

The report identifies four pending validation tasks: a synteny-threshold permutation test requiring unfiltered Spark data; BLAST validation of Node_4915 against NCBI nr; a housekeeping-gene null baseline; and biome-enrichment factorization using θ = OR(T4SS-CAZy) / [OR(T4SS) × OR(CAZy)]. ₇₃

Truly Dark Genes — What Remains Unknown After Modern Annotation?

The pangenome linkage gap prevents assessment of 17,479 dark genes (31%); the report estimates that approximately 2,841 of them may be truly dark, describing this as a 31% coverage gap in the prioritized list. ₇₄

Bakta may produce false-negative functional calls: some genes labeled hypothetical may have known functions that did not match the Bakta PSC database at the tested version, v6.0. ₇₄

Bidirectional best hit (BBH) ortholog coverage, which treats reciprocal top-scoring matches between genomes as orthologs, includes only 32 of 48 Fitness Browser organisms, so truly dark genes from the remaining 16 organisms lack concordance and ortholog-breadth data. ₇₄

GC deviation is an imperfect HGT indicator because gene-specific composition biases, including those associated with membrane proteins, and amelioration over time can also produce deviation. ₇₄

Short genes are inherently harder to annotate because they contain fewer domains and have fewer homologs, and they are harder to measure for fitness because they provide fewer transposon-insertion sites. Consequently, the length difference of d = −0.432 may partially explain other observed differences. ₇₄

Some strong fitness phenotypes may result from polar effects on downstream genes rather than from the truly dark gene itself. ₇₄

The proposed biological interpretation is therefore strongest for the directly measured differences and weaker for causal explanations: accessory-genome enrichment, GC deviation, mobile-element proximity, and narrow taxonomic breadth support recent acquisition and rapid evolution, but do not by themselves establish HGT or gene function. ₇₄

Web of Microbes Data Explorer

The report identifies the absence of consumption data as the fundamental limitation. The snapshot records what organisms produce but not what they consume, which prevents testing whether consumed metabolites predict gene essentiality. The data are also small, with 37 organisms (20 experimental) across 5 projects from a single laboratory. The report says this is too few for statistical analyses across organism groups. ₇₅

The snapshot is frozen in 2018 and was accessed through the Wayback Machine, so it may not reflect the current state of WoM. In its literature context, the report cites de Raad et al. (2022). That study developed the Northen Lab Defined Medium (NLDM), which supports growth of 108/110 phylogenetically diverse soil bacteria, with all metabolites trackable via LC-MS/MS (liquid chromatography–tandem mass spectrometry). The report states that this dataset “likely contains consumption data” and “may be available” in the current GNPS2 version of WoM. It suggests that the GNPS2-hosted version or Northen laboratory datasets would be substantially richer, but these were not analyzed. ₇₅

The report’s future-directions section makes a stronger claim than its literature context. It states that the NLDM study tested 110 soil bacteria with full consumption and production tracking. It also states that ingesting that dataset would resolve the no-consumption limitation and increase the organism count 3–5×. This projection is conditional: neither ingestion nor the increase occurred in this project. ₇₅

GapMind pathway matching failed because pathway names did not contain simple metabolite names, so a dedicated pathway-to-metabolite mapping table is needed. WoM coverage of E. coli BW25113 is especially limited despite the richness of its Keio Fitness Browser data, with only 12 WoM observations focused on sulfur and cysteine. ₇₅

The report proposes five next steps:

• Obtain the current WoM dataset from GNPS2 or the Northen laboratory.
• Build a GapMind pathway-to-metabolite lookup table.
• For pseudo3_N2E3 and pseudo13_GW456_L13, identify which genes are fitness-important on carbon sources that the same organism produces, to test whether metabolic output predicts gene essentiality.
• Test whether higher E/(E+I) ratios associate with more open pangenomes or more accessory genes.
• Re-ingest WoM when consumption data become available. ₇₅

These are future work, not demonstrated results. ₇₅