Discoveries Log

Raw report: discoveries Feeds into: adversarial-research-quality-assurance, antimicrobial-resistance-fitness-cost, callability-limited-comparative-inference, classifier-database-compatibility-in-taxonomic-quantification, cofitness-network-architecture, community-metabolic-interdependence, composite-functional-annotation, composite-resistance-score-limitations, compositional-robustness-of-differential-abundance-calls, computational-pathway-prediction-validation, condition-dependent-gene-tradeoffs, condition-space-dimensionality, condition-specific-fitness, confirmatory-exploratory-ecological-association-discordance, core-gene-annotation-paradox, core-genome-burden-paradox, costly-dispensable-gene-loss, cross-cohort-microbiome-portability, cross-condition-metabolic-comparability, cross-species-fitness-transferability, cross-tenant-data-bridging, cultivation-collection-bias-in-ecological-genomics, data-landscape-ownership-and-coverage-bias, ecological-cost-of-microbiome-target-depletion, ecotype-clustering-validity, ecotype-environment-gene-content, environment-embedding-geography, environmental-resistome, evidence-triangulation-for-functional-annotation, experimental-prioritization-of-functional-dark-matter, fitness-condition-coverage-prioritization-bias, fitness-module-detection-sensitivity, functional-marker-validation, gene-cooccurrence-ecological-guilds, gene-essentiality, gene-function-acquisition-depth, genetic-perturbation-coverage-bias, genomic-dispersal-functional-coupling, gut-microbiome-ecotypes-as-patient-strata, homology-search-negative-evidence, horizontal-gene-transfer-driven-innovation, lab-field-fitness-concordance, laboratory-fitness-versus-natural-selection, metabolic-capacity-specialization, metabolic-model-gapfilling, metal-cross-resistance, module-level-coinheritance, multi-omics-integration, occurrence-versus-catabolic-activity, ontology-and-category-schema-sensitivity, pangenome-conservation-fitness-decoupling, pangenome-core-boundary-and-clade-size-bias, pangenome-integration, pangenome-openness-determinants, phage-defense-syndromes-and-arms-race, phage-therapy-evidence-translation, phenotype-database-coverage-bias, phylogenetic-confounding-of-pangenome-associations, provenance-aware-resource-discovery, research-attention-inequality, resistance-island-coinheritance, respiratory-capacity-and-nadh-load, sample-size-aware-phenotype-consensus, sampling-depth-and-downsampling-effects, scale-dependent-mobile-element-associations, selection-on-outcome-leakage, shared-stress-versus-stressor-specific-fitness, spatial-sampling-effort-confounding, structural-annotation-gap, study-batch-confounding-of-environmental-associations, subsurface-bacillota-specialization, subsurface-hydrogeological-zonation, taxonomic-nomenclature-reconciliation, taxonomic-resolution-dependent-functional-inference, transposon-callability-bias, two-speed-bacterial-genome, within-species-conservation-between-species-functional-divergence

Overview

This log consolidates cross-project findings from 2026 spanning microbial fitness, pangenome evolution, ecological specialization, metabolic modeling, antimicrobial resistance, subsurface microbiology, microbiome analysis, and research-quality practice. It includes direct measurements, methodological corrections, validation results, contradictions, and reusable analytical rules. ₁

Key Findings

Evolution, gene function, and pangenomes

M22 Sankoff parsimony attribution of 17M gain events on the GTDB-r214 species tree produced function-class acquisition-depth signatures: CRISPR-Cas had a 24.5× recent/ancient ratio, strict housekeeping genes were approximately 1×, and mixed regulatory-plus-metabolic KOs had 2× the Innovator-Exchange rate of pure regulatory or pure metabolic KOs. Leaf consistency independently validated the depth trend, with species-with-KO fractions declining from 0.34 for recent gains to 0.20 for ancient gains. ₁

Within-clade heterogeneity can be hidden by family-level aggregation. For Mycobacteriaceae mycolic-acid KOs, family-rank producer-side Cohen’s d was 0.31, while median leaf consistency was 0.15 — lower than the atlas reference of 0.20; restricting analysis to the mycolic-positive sub-clade increased producer d to +0.394 versus +0.309 at family rank, with 10 of 13 genera having mean leaf consistency ≥ 0.5. The generalizable rule stated in the log is that when a clade-level supported finding has surprisingly low within-clade leaf consistency, the family-rank effect is a population mixture across a positive sub-clade and negative members, so sub-clade-restricted recomputation sharpens the supported claim rather than averaging it away. ₁

A robust clade-level innovation claim should converge across three independent substrates: atlas Cohen’s d, ecology-based clade-by-biome Fisher enrichment, and curated phenotype or model-organism evidence. Each of the three pre-registered atlas findings — mycolic acid, photosystem II, and Bacteroidota polysaccharide utilization loci — produced convergent evidence across these substrates, with BacDive serving as the metabolic-phenotype anchor where coverage existed. The referenced workflow used p<10⁻¹¹ for an ecology enrichment and relied on per-source corroboration coverage whose scope differs by source: 88.8% biosample geographic coverage in kbase_ke_pangenome.ncbi_env, 28.4% biome categorical coverage in kescience_mgnify, 32% phenotype coverage for BacDive through GCF→GCA fallback, and approximately 30 organisms reachable by bidirectional-best-hit (BBH) orthology in kescience_fitnessbrowser — an orthology-access limit, not a statement of unrestricted Fitness Browser coverage. ₁

Across 32 species, 9 phyla and 357,623 genes, novel or singleton genes were enriched in COG (clusters of orthologous groups) L mobile elements by +10.88% at 100% consistency across species — the strongest signal — COG V defense mechanisms by +2.83% at 100% consistency, and COG S unknown function by +1.64% at 69% consistency. The digest lists five further categories under core-gene enrichment while reporting negative enrichment values for them — J translation -4.65% at 97% consistency (labelled the strongest depletion), F nucleotide metabolism -2.09% at 100% consistency, H coenzyme metabolism -2.06% at 97% consistency, E amino acid metabolism -1.81% at 81% consistency, and C energy production -1.75% at 88% consistency — without stating the comparator or sign convention behind those signed values, so the figures are recorded here exactly as reported and no direction of core enrichment or depletion is asserted. All 8 predictions from the initial N. gonorrhoeae analysis were confirmed across the 32 species. Composite assignments such as LV (mobile plus defense) showed +0.34% enrichment and 76% consistency, and composite assignments including EGP (amino acid plus carbohydrate plus inorganic ion) are treated as genuine multi-functional genes rather than annotation artifacts, so multi-function assignments can represent real mobile-defense modules — candidate “mobile defense islands” — rather than annotation noise to be filtered out. ₁

The digest’s reading of this pattern — core genes as an ancient, conserved metabolic engine of translation, energy production and biosynthesis; novel genes as recent acquisitions for ecological adaptation; horizontal gene transfer rather than vertical inheritance as the primary innovation mechanism; the +10.88% COG L enrichment as evidence that most genomic novelty comes from mobile elements; and the universality of the pattern across bacterial phyla as a deep evolutionary constraint — is a broad evolutionary interpretation of enrichment counts rather than a demonstrated mechanism. ₁

Across 27,690 species, 55.9% of dark-gene ortholog groups were kingdom-level — present in thousands of species across multiple phyla — with species counts ranging from 1 to 27,482 and median 135; the top-ranked unknowns COG0468, COG0443 and COG0491 are reported as present in virtually all sequenced bacterial genomes yet carrying zero functional evidence. In a 57,011-gene darkness spectrum, only 4,273 genes (7.5%) were T1 Void with zero evidence, whereas 22,500 (39.5%) were T4 Penumbra with 3–4 converging evidence lines (fitness phenotype, domain annotation, ortholog group, module membership), so the digest reads the bottleneck as targeted experimental validation rather than computational inference. Across 47 testable dark-gene clusters, 29 (61.7%) showed directional concordance between laboratory fitness conditions and carrier-genome environments, with Fisher’s combined p=0.031; NMDC validation confirmed all 4 pre-registered abiotic predictions and all 7 pre-registered trait predictions. ₁ Two qualifications from the originating report must travel with that concordance: a one-sided binomial test of the 29/47 concordance rate against the null of p = 0.5 yields only p = 0.072, consistent with a Wilson score 95% CI of [0.474, 0.742] that includes 0.50, and the NMDC trait correlations likely reflect compositional coupling — genera abundant in a sample contribute to both carrier abundance and community trait scores — rather than independent gene–phenotype validation. ₂

65 dark-gene ortholog groups showed conserved fitness effects across 3+ organisms, and dark-gene concordance levels were statistically indistinguishable from those of annotated genes (Mann–Whitney p=0.17) — a null that refines rather than weakens the concordance result; motility-related dark genes showed the strongest cross-organism concordance, which suggests conserved but incompletely annotated chemotaxis machinery rather than demonstrating it. ₁

Which organisms to characterize first depends on the weighting applied. Shewanella MR-1 contributed 25/100 top fitness-active candidates, 587 scored dark genes in total, and 20.8% of the top-500 in just 3 experiments (stress, nitrogen, carbon), reflecting deep condition coverage (121 conditions) and 142 genes with |fit| >= 4; under conservation-weighted prioritization S. meliloti ranked first instead, with 1,630 ortholog groups and 195 kingdom-level gaps. ₁

The evidence-weighted and conservation-weighted covering sets both selected 42 organisms covering approximately 95% of priority and shared 39 organisms, differing mainly in rank order — MR-1 first on deep condition evidence, S. meliloti first on kingdom-level gaps — so targeted hypothesis testing and novel-function discovery can be served by largely the same organism set. ₁

Adding 25 non-Fitness-Browser organisms (Bacillota, Actinomycetota, Campylobacterota) to the candidate pool selected 50 organisms spanning 6 phyla versus 4 for the Fitness-Browser-only pool, with P. aeruginosa PAO1 ranked #1 (3,713 ortholog groups) and M. tuberculosis reaching #6 as the first Actinomycetota; B. subtilis and S. aureus were not selected, because their kingdom-level ortholog groups were already covered by Pseudomonadota selections, so they add no unique ortholog-group coverage while remaining valuable for studying genes in native Gram-positive context. ₁

Fitness Browser dark genes were substantially reclassified by Bakta v1.12.0: of 39,532 dark genes with pangenome links, 33,105 (83.7%) were reclassified and 6,427 remained truly dark in the strict sense that Fitness Browser and Bakta both call them hypothetical protein. Truly dark genes were shorter (121 versus 194 aa), less conserved (43% versus 73% core), had fewer orthologs (29% versus 64%), and had higher GC deviation (d=0.247); 41% of their neighbors were also hypothetical and 12% were within 2 genes of mobile elements (transposases, integrases, phage proteins), from which concentration in genomic islands is inferred rather than demonstrated. Only 246/6,427 (3.8%) had zero annotation clues, while 84.7% had database cross-references, 43.5% had eggNOG hits (though 55% of COG assignments were “S”/unknown), 29.3% had cross-organism orthologs, and 9.5% sat in ICA fitness modules. ₁

Contrary to expectation, truly dark genes with strong fitness phenotypes were depleted in stress conditions (OR=0.53, p<0.001) and enriched in nutrient, mixed community, and iron conditions. ₁

Bakta v1.12.0 and eggNOG-mapper are complementary across 132.5M gene clusters: eggNOG had higher COG coverage (51% versus 8.2%), KEGG coverage (38.5% versus 17.3%), and Pfam coverage (63% versus 7.7%), while Bakta had higher GO coverage (15% versus 7.4%), product descriptions (71.2% versus 70.4%), and unique UniRef50 links covering 79.2%. Their union increased any-functional-annotation coverage from ~70% to 77.3%, and Bakta rescued 11.2M clusters among 39.2M missed by eggNOG. Only 33.3% of Bakta’s 17.6M distinct UniRef50 IDs existed in the KBase Data Lakehouse UniProt identifier table; among matched IDs the bridge reached RefSeq for 71%, OrthoDB for 46%, STRING for 40%, and KEGG for 36%, while GO, EC, InterPro, and Pfam were missing from the bridge entirely. ₁

The annotation-coverage asymmetry has documented mechanical causes, which the log attributes to Bakta developer Oliver Schwengers across GitHub issues #350, #385, #391, and #393. ₁

• Bakta’s PSC pre-computed annotations are sparse: only ~14% of PSC entries have COG and ~19% have KEGG, because the PSC database is optimized for product descriptions rather than functional categories. ₁
• COG 2024 no longer provides representative sequences, which broke direct mapping and forces a lossy indirect WP accession-to-IPS-to-PSC route. ₁
• Pfam is skipped by design: HMMER runs only on hypotheticals (~7.7%), so the 92% of proteins with PSC matches are never Pfam-searched. ₁
• Search sensitivity is lower in the faster mode: Bakta’s PSC uses DIAMOND fast mode, which is less sensitive than eggNOG’s HMM-based ortholog assignment. ₁
• Integrating eggNOG into PSC representatives was planned (issue #325) and may have partially happened in DB v6.0, but with limited impact; this remains uncertain. ₁

The Arche benchmark paper (Alonso-Reyes & Albarracin 2025, PMID 40811208) independently confirmed Bakta’s conservative annotation transfer relative to other tools. ₁

Pangenome openness had no significant relationship with environment or phylogenetic effects in one analysis (rho=-0.05, p=0.54 and rho=0.03, p=0.73), but variable metabolic pathways strongly predicted openness in another pan-bacterial analysis of 2,810 species with ≥10 genomes each: rho=0.327, p=7.2e-71, and partial rho=0.530, p=2.83e-203 after controlling for genome count, with the signal holding across 18/21 phylogenetic groups tested and amino acid biosynthesis pathways (leu, val, arg, lys, thr) showing the largest accessory dependence gap (~0.14). The digest calls metabolic pathway variability a genuine predictor and much stronger than the null result for environmental and phylogenetic predictors of openness, but those are different predictor sets rather than directly equivalent tests of the same quantity. Across 1,872 species with sufficient AlphaEarth coverage, niche breadth predicted pathway completeness at r=0.392, p=7.1e-70, embedding variance at r=0.412, p=1.8e-77, and geographic range at r=0.360, p=1.8e-58; only 6.8% of species (1,872/27,690) had sufficient coverage. ₁

Metabolic heterogeneity within species tracks the same openness axis, which refines the openness determinants rather than restating them: hierarchical clustering of 225 species (≥50 genomes, ≥3 variable pathways) by GapMind pathway profiles gave a median of 4 metabolic ecotypes per species with a maximum of 8, and more ecotypes correlated with pangenome openness (partial rho=0.322, p=8.0e-07); Alistipes onderdonkii and Barnesiella intestinihominis were the top-ecotyped species, with 8 each. ₁

The joint analysis of 1,872 bacterial species with complete pangenome, GapMind pathway-prediction and AlphaEarth embedding data adds weaker companion correlations that refine rather than replace the openness results above: openness related positively but weakly to pathway completeness (r=0.107, p=3.6e-6) and to pathway variability measured as standard deviation (r=0.066, p=0.004), which the digest reads as accessory genes enabling within-species metabolic heterogeneity rather than demonstrating it, and openness correlated with niche breadth (r=0.324, p=5.6e-47) while core fraction correlated negatively with niche breadth (r=-0.445, p=1.4e-91). Sufficient AlphaEarth coverage meant ≥5 genomes with embeddings, and the 6.8% coverage limit is attributed to missing lat/lon metadata for most NCBI genomes, especially clinical isolates. ₁

In a separate analysis of 172 species with sufficient data, the median partial correlation was 0.0025 for the environment effect and 0.0143 for the phylogeny effect, with phylogeny dominating in 60.5% of species and environment in 39.5%; no significant difference was found between “environmental” bacteria, where latitude/longitude is meaningful, and host-associated bacteria (p=0.66). ₁

AlphaEarth embeddings carry a stronger geographic signal for environmental than for human-associated genomes: environmental samples (Soil, Marine, Freshwater, Extreme, Plant) showed a 3.4x ratio in embedding cosine distance between nearby pairs (<100 km, mean=0.27) and far pairs (>10K km, mean=0.90), whereas human-associated samples (gut, clinical) showed only a 2.0x ratio (0.37 vs 0.75), attributed to hospitals worldwide having similar urban built-environment imagery; the pooled “All samples” curve, also 2.0x, is dominated by the 38% human-associated fraction and substantially underestimates the geographic signal, so embedding-based environment proxies should be stratified by sample type. ₁

Of 83,287 genomes with AlphaEarth embeddings, 38% are human-associated (clinical 20%, gut 16%, other 2%), while soil, marine and freshwater are each 7% — a sampling bias the digest attributes to NCBI’s emphasis on pathogen sequencing, since clinical isolates carry good geographic metadata from epidemiological tracking. ₁

The digest’s own proposal that this clinical bias explains the weak environment effect in the ecotype analysis (median partial correlation 0.003) is contradicted by reanalysis rather than supported: stratifying 213 species by dominant environment type (47% human-associated, 21% environmental) found human-associated species with higher median partial correlation (0.084) than environmental species (0.051), Mann-Whitney p=0.83, so the phylogeny-dominates conclusion holds regardless of sample environment and the embeddings’ geographic signal and the environment–gene-content relationship are distinct phenomena. ₁

Harmonizing isolation_source into 12 categories from 5,774 values classified 106/224 species in the ecotype analysis as majority human-associated (gut plus clinical) and only 47/224 (21%) as majority environmental — a more systematic classification than the original manual species-level categorization, which confirms the clinical sampling bias while showing it does not confound the ecotype results. ₁

Essentiality, fitness, and cofitness

Across 48 bacteria (221,005 genes, 17,222 ortholog families), 15 gene families were essential in every organism: ribosomal proteins (rpsC, rplW, rplK, rplB, rplA, rplF, rps11, rpsJ, rpsI, rpsM), chaperonin groEL, CTP synthase pyrG, translation elongation factor G fusA, valyl-tRNA synthetase valS, and geranyltranstransferase SelGGPS — described as the irreducible essential core of bacterial life. Of 17,222 ortholog families, 859 (5.0%) were universally essential, meaning essential in every organism where the family has members, 4,799 (27.9%) variably essential, and 11,564 (67.1%) never essential, so variable essentiality is the majority case among essential families and genomic context shapes which genes an organism requires for viability. The 7,084 orphan essential genes, which have no orthologs in any other Fitness Browser organism, were 58.7% hypothetical while universally essential genes were 8.2% hypothetical, making the orphans the least characterized yet most functionally important genes in each organism and prime candidates for novel biology. The digest labels this 48-organism cohort bacteria, but the cohort explicitly includes Archaea (Methanococcus), so the bacteria label is the digest’s wording rather than an accurate description of the organism set. ₁

Universally essential genes were 91.7% core versus 80.7% for non-essential genes, and 71% of universally essential families were 100% core within species; orphan essentials were only 49.5% core. Module transfer from non-essential orthologs that participate in ICA fitness modules generated 1,382 function predictions for hypothetical essential genes across 48 organisms, each backed by cross-organism module-family conservation rather than by direct functional measurement, so these remain indirect predictions for genes that fitness-based methods cannot see. ₁

Across 33 bacteria, putative essential genes — those with no transposon insertions in RB-TnSeq (random barcode transposon sequencing) — were 86.1% core versus 81.2% for non-essential genes, with median OR=1.56 and 18/33 organisms significant after BH-FDR (Benjamini–Hochberg false-discovery-rate correction); the enrichment is real but modest, since most genes in well-characterized bacteria are core regardless of essentiality, and the signal is strongest in organisms with larger clades, where more genomes make core classification more reliable. Across 194,216 protein-coding genes in 43 bacteria, core fractions formed a continuous 16-percentage-point gradient from 82% for essential genes to 66% for always-neutral genes, with the same pattern on binning by strongest negative fitness effect (min_fit < -3 gave 78% core versus 66% for min_fit near 0) and by fitness breadth (important in 20+ conditions gave 79% core versus 66% for 0 conditions), which establishes the essentiality–conservation link as quantitative rather than binary. ₁

Essential genes mapping to core clusters were 41.9% enzymes and only 13.0% hypothetical — a well-characterized metabolic backbone of ribosomes, DNA replication, cell wall and cofactor biosynthesis — whereas essential-auxiliary genes, essential but not present in all strains, were 13.4% enzymes and 38.2% hypothetical, and essential-unmapped genes, strain-specific with no pangenome match, were 44.7% hypothetical, making the unmapped set a functional-discovery target. ₁

By SEED functional category, essential-core genes were enriched in Protein Metabolism (+13.7 percentage points versus non-essential), Cofactors/Vitamins (+6.2 pp) and Cell Wall (+3.9 pp), and depleted in Carbohydrates (-7.9 pp), Amino Acids (-5.6 pp) and Membrane Transport (-4.0 pp) — functions the digest reads as conditionally important rather than universally essential. ₁

Fitness modules derived by robust ICA (independent component analysis, 30–50 FastICA runs) with DBSCAN clustering were stable at 17–52 modules per organism across 32 bacteria; 94.2% had significantly elevated within-module cofitness by Mann–Whitney U at p < 0.05, with mean |r|=0.34 versus background 0.12 and mean genomic-adjacency enrichment of 22.7×, consistent with modules capturing operon-like co-regulated gene groups. The ICA components themselves were not the problem: D’Agostino K² normality-based membership thresholding gave biologically meaningless modules of 100–280 genes, and switching to an absolute weight threshold (|Pearson r| ≥ 0.3 with the module profile, maximum 50 genes) reduced modules to 5–50 genes while improving cofitness enrichment from 59% to 94.2% and within-module |r| from 0.047 to 0.34. Cross-organism alignment on bidirectional-best-hit (BBH) ortholog fingerprints found 156 module families spanning at least 2 organisms, including 28 spanning 5+ organisms and one spanning 21 of 32 organisms across Proteobacteria, Bacteroidetes, Firmicutes and Archaea, read as deeply conserved fitness regulons. ₁

This supports a conservation–coregulation link at module level: ICA fitness modules were 86.0% core versus 81.5% baseline across 29 organisms (Fisher OR=1.46, p=1.6e-87; per-organism paired Wilcoxon p=1.0e-03, with 22/29 organisms showing enrichment), and 59% of modules were more than 90% core genes, placing co-regulated fitness-response units preferentially in the conserved genome. ₁

That link does not extend to module breadth: module families spanning more organisms did not have higher core fractions (Spearman rho=-0.01, p=0.914), and because the baseline core rate is approximately 82% the digest reads this null as a ceiling effect rather than as evidence against a conservation–function relationship. ₁

The module census also has a hard blind spot: 0 essential genes appear in any of the 1,116 fitness modules across 32 organisms, because ICA decomposes fitness variation and essential genes — having no transposon insertions and therefore no fitness data — are invisible to it, so modules capture only the non-essential portion of the genome’s functional architecture. ₁

Two per-organism observations bear on how much data module detection needs, without establishing a trend: Caulo, with 198 experiments, showed the weakest pilot signal at 2.9x correlation enrichment versus 31x for DvH, while organisms down to 104 experiments (Ponti) still produced stable modules. Capping components at 40% of experiments was necessary, because higher ratios caused FastICA convergence failures and extreme slowness. ₁

Ortholog scope changed cross-organism results substantially: BBH pairs from 5 organisms (10K pairs, 1,861 ortholog groups) gave 27 module families with no family spanning more than 4 organisms, whereas all 32 organisms (1.15M pairs, 13,402 ortholog groups) gave 156 families, 28 spanning 5+ organisms, and family-backed predictions rising from 31 (4%) to 493 (56%). The recorded lesson is to extract orthologs for every organism in an analysis rather than a pilot subset, because the ortholog graph is not additive — adding organisms creates new transitive connections. ₁

Annotation granularity governed how much of the module set could be labelled at all. An initial KEGG-plus-SEED-plus-TIGRFam enrichment pipeline with min_annotated=3 annotated only 92/1,116 modules (8.2%), because KEGG KO (KEGG Orthology) groups average ~1.2 genes per term so modules of 5–50 members almost never carry 3+ genes sharing the same KO; adding Pfam domains, which have 814 terms with 2+ genes versus TIGRFam’s 88, and lowering the overlap threshold to 2 raised annotation to 890/1,116 (79.7%), with Pfam the dominant source because its domain-level labels match module granularity. Downstream, predictions rose from 878 to 6,691 (7.6×) and now covered all 32 organisms, annotated families rose from 32 to 145 of 156 (93%), and three organisms (Caulo, Methanococcus_S2, Korea) went from 0 enrichments to 24–29 enriched modules. The generalizable rule is that min_annotated must be calibrated to the annotation database’s granularity — 3 eliminates nearly all tests for gene-specific KOs, 2 works for domain-level Pfam — since Fisher’s exact test handles small counts correctly and statistical validity comes from FDR correction across all tests rather than from requiring large per-test overlaps. ₁

Held-out KO prediction across 32 organisms showed ortholog transfer achieving 95.8% precision and 91.2% coverage for gene-level function prediction, whereas Module-ICA had <1% KO-level precision; the digest reads this as expected and informative rather than a failure, because modules capture process-level co-regulation (validated by the 94.2% cofitness and 22.7× adjacency enrichment) and an ABC transporter module correctly groups binding, permease and ATPase subunits that each carry a different KO, so module-based predictions should be framed as “involved in [biological process]” rather than “has function [specific KO]”. ₁

Across 25,271 true trade-off genes, 17.8% of genes were true trade-offs, important (fit < -1) in some conditions and burdensome (fit > 1) in others; trade-offs were 1.29× more likely to be core (OR=1.29, p=1.2e-44). Core genes were ever beneficial after deletion in 24.4% of cases versus 19.9% for auxiliary genes (OR=0.77, p=5.5e-48), and condition-specific genes were 77.3% core versus 70.3% without specific phenotypes (OR=1.78, p=1.8e-97). There were 28,017 costly-plus-conserved genes, 5,526 costly-plus-dispensable genes read as candidates for ongoing gene loss, and 21,886 neutral-plus-dispensable genes read as niche-specific. The digest treats genes that are both burdensome in the laboratory and core in the pangenome as the strongest evidence for purifying selection in natural environments — costly to maintain yet kept by every strain, implying nature requires them in conditions the laboratory does not capture — which is an extrapolation from laboratory fitness rather than a field measurement. ₁

The core-burden pattern is not a uniform property of the core genome: core genes were more burdensome in RNA Metabolism (+12.9pp), Motility/Chemotaxis (+7.8pp) and Protein Metabolism (+6.2pp), while Cell Wall reversed, with non-core cell wall genes more burdensome (-14.1pp), so the paradox refines into the trade-off architecture of specific functional systems rather than a genome-wide property. ₁

Condition-specific metabolism and model limitations

In ADP1, each of 8 carbon sources required a distinct set of respiratory chain components — qualitatively different wiring per substrate class rather than a quantitative gradient. Quinate required only Complex I, with all other components dispensable; acetate, the most demanding, required Complex I plus cytochrome bo3 plus ACIAD3522 and more; and glucose required no specific component, showing full redundancy. Quinate produced 0.57 NADH per carbon versus 1.50 for glucose, yet Complex I was more essential on quinate; the digest’s proposed resolution is that β-ketoadipate pathway products (succinyl-CoA plus acetyl-CoA) enter the TCA cycle simultaneously and create a concentrated NADH burst exceeding NDH-2’s reoxidation capacity, a mechanism proposed rather than measured. ₁

Proteomics showed ADP1 expressing its three NADH dehydrogenases at similar levels under standard conditions: Complex I mean 27.6 (66th percentile), NDH-2 mean 27.0 (59th), and ACIAD3522 mean 26.2 (48th), all within 1.4 units of each other around a genome median of 26.4; the digest’s interpretation is a passive, flux-based wiring system in which all three run simultaneously and whichever becomes limiting depends on the NADH flux rate from the carbon source, rather than an active regulatory switch. The three enzymes are assigned distinct condition roles: Complex I (13 subunits, proton-pumping) is quinate-essential and glucose-dispensable; NDH-2 (1 subunit, non-pumping) has no growth data and is only a predicted backup for glucose; and ACIAD3522 (NADH-FMN oxidoreductase) is acetate-lethal at a growth ratio of 0.013 versus 1.39 on both quinate and glucose, making it the most condition-specific gene in the 2,034-gene matrix, while its biological role is unknown and FBA (flux balance analysis) predicts zero flux through it on all conditions. The cross-species NDH-2 compensation hypothesis was not supported after annotation correction, which filtered text-matched candidates to organisms with ≤2 hits per genome: Complex I aromatic deficits were -0.297 with validated NDH-2 versus -0.156 without, p=0.52, and the initial apparent support — 10/14 organisms with NDH-2 — was driven by misannotated Complex I subunits leaking through the text filter in 5 organisms, an annotation artifact rather than a biological signal. ₁

The ADP1 aromatic-catabolism support network contained 8 pathway genes but 51 quinate-specific genes — the digest reports this as a 7:1 support-to-pathway ratio, a figure whose numerator is inconsistent because the originating report counts the 8 pathway genes inside the 51-gene total, so the ratio is recorded as reported and not recomputed here — organized into 21 Complex I genes for NADH reoxidation, 7 iron-acquisition genes supplying Fe²⁺ for the ring-cleavage dioxygenase, and 2 PQQ-biosynthesis genes for the quinate dehydrogenase cofactor. FBA lacked reaction mappings for 30/51 genes (59%) — PQQ biosynthesis, iron acquisition, transcriptional regulators and respiratory chain components are invisible to the metabolic model — and predicted 0% Complex I essentiality despite predicting 1.76× higher aromatic flux, because its linear programming cannot represent the threshold effect of losing a single subunit from a multi-subunit complex. Cross-species ortholog-transferred fitness data showed larger Complex I defects on acetate (-1.55 versus background) and succinate (-1.39) than on aromatics, supporting a high-NADH-flux bottleneck interpretation in which the network is less an aromatic support network than a high-NADH-flux one, with alternative-dehydrogenase compensation on lower-NADH-flux substrates remaining a tentative explanation. ACIAD3137 (UPF0234/YitK) and ACIAD2176 (DUF2280) correlated with Complex I genes at r > 0.98 across 8 growth conditions while lacking FBA reaction mappings and any assigned metabolic function, which suggests physical or regulatory association with Complex I rather than demonstrating it. ₁

Of ADP1’s dispensable genes, 31% had condition-specificity scores ≥ 1.0 and the project headline reports 625 condition-specific genes; that 31% is computed on the 2,034-gene growth matrix rather than on all 2,593 TnSeq-dispensable genes counted in the separate deletion-collection coverage analysis. The top condition-specific gene for each carbon source was the enzyme biochemically predicted to be required — urease subunits for urea, protocatechuate 3,4-dioxygenase for quinate, Entner-Doudoroff enzymes for glucose, the glyoxylate shunt for acetate, and a lactate-responsive regulator for lactate — which the digest reads as validating the growth matrix as a high-quality functional-genomics dataset. ₁

Hierarchical clustering of the 2,034 genes by their 8-condition growth profiles produced an optimal K=3 with silhouette=0.24, that is, no discrete functional modules: the phenotype landscape is a gradient, with the single exception of 24 genes forming a tight quinate-specific module for the aromatic degradation pathway. ₁

The ADP1 deletion collection is itself biased: 272 of 2,593 TnSeq-dispensable genes (10.5%) lack growth data, and those genes are systematically different — shorter (813 versus 981 bp), less annotated (91% versus 100% RAST), less conserved in the pangenome (76.5% core versus 93.3%, p=1.4e-20), and massively enriched for hypothetical proteins (q=2.4e-25) — so the collection favours well-characterized, conserved genes. ₁

Across 7 Fitness Browser organisms and 23 GapMind pathways, 35.4% of pathway-organism pairs were Active Dependencies (complete and fitness-important), 41.0% Latent Capabilities (complete but fitness-neutral in rich media), 14.9% Incomplete but Important, and 8.7% Missing. All Latent Capabilities became fitness-important under condition-specific analyses of nitrogen limitation, stress or carbon-source shifts; the digest supplies no pair count for that claim and does not state how the importance threshold was set, so it is recorded as reported. In Pseudomonas fluorescens FW300-N2E3, all 13 Web of Microbes metabolites that could be mapped to GapMind had complete pathway predictions and all 13 also showed growth in Fitness Browser experiments — agreement limited to those mapped, commonly characterized amino acids and organic acids — and the overall 94% four-database concordance was structurally driven: Fitness Browser 21/21 (100%), GapMind 13/13 (100%), and BacDive only 3/7 (43%), with 37/41 comparisons concordant (90.2%) and BacDive utilization of Web of Microbes-produced metabolites not significantly above the 27.5% species baseline (binomial p=0.40). ₁

A four-line convergence in the same organism qualifies that pathway-completeness agreement: FW300-N2E3 produces tryptophan (Web of Microbes action: Increased), has 231 genes with significant fitness when grown on tryptophan, and has a complete tryptophan biosynthesis pathway by GapMind, yet 0/52 P. fluorescens strains in BacDive can utilize tryptophan as a carbon source. The digest reads this as tryptophan overflow metabolism serving a cross-feeding or signaling rather than catabolic function — an inference from the convergence of four lines rather than a measurement. ₁

Condition breadth separates housekeeping from substrate-specific signal in that dataset: the top 18 genes significant across all 21 Fitness Browser metabolite conditions are amino acid biosynthesis enzymes — homoserine O-acetyltransferase (methionine), ATP phosphoribosyltransferase (histidine), dihydroxy-acid dehydratase (branched-chain), isopropylmalate dehydrogenase (leucine), shikimate dehydrogenase (aromatic) and imidazoleglycerol-phosphate dehydratase (histidine) — essential for growth on any minimal medium, whereas the approximately 370 genes significant in only 1-2 conditions are the truly substrate-specific ones, a distinction the digest treats as critical for interpreting Web of Microbes-to-Fitness-Browser integration. ₁

The Web of Microbes encoding bounds what such integration can claim. Its 4-action system marks organism observations ‘E’ (Emerged — absent from the control medium and now detected, genuine de novo biosynthesis) or ‘I’ (Increased — present in the control and raised), and the two are mutually exclusive, with 0 overlap across 10,744 observations; for the control (“The Environment”) ‘D’ means detected in medium and ‘N’ not detected. Critically, the 2018 snapshot carries no “Decreased” (consumption) action for any organism — all 742 ‘D’ observations belong to the control — so the database records only production, and consumption cannot be inferred from it. ₁

Two Web of Microbes organisms are exact strain matches to Fitness Browser organisms — Pseudomonas sp. FW300-N2E3 = pseudo3_N2E3 (5,854 genes, 211 experiments) and Pseudomonas sp. GW456-L13 = pseudo13_GW456_L13 (5,243 genes, 106 experiments) — and E. coli BW25113 = Keio is the same strain as well, enabling direct metabolite-to-gene-fitness linking; the Fitness Browser carbon-source conditions for these organisms (L-alanine, L-proline, sodium D-lactate and others) test metabolites Web of Microbes detects as produced, with 109 metabolite-condition name overlaps available for prospective cross-dataset analysis. ₁

Of 257 identified (non-unknown) Web of Microbes compounds, 69 match ModelSEED molecules by exact name and 107 by formula, totaling 176 (68.5%), which enables reaction-level annotation for the majority of those metabolites. ₁

For gene-to-pathway linking, the Fitness Browser’s own KEGG tables (besthitkegg plus keggmember plus kgroupec) provide direct gene-to-KEGG-KO-to-EC-to-pathway mapping for all 48 Fitness Browser organisms along the chain besthitkegg → keggmember → kgroupec → keggconf → GapMind pathway, which the digest reports as simpler than a DIAMOND blastp link-table pipeline and as covering more organisms. ₁

Across 433 Pseudomonas species (12,732 genomes), GapMind carbon-pathway predictions showed the P. aeruginosa group with near-complete loss of plant-derived sugar catabolism relative to the P. fluorescens/putida group: 43 of 62 pathways differed significantly (FDR < 0.05), with the largest effects in xylose (+74 pp), ribose (+64 pp), arabinose (+63 pp), galacturonate (+60 pp) and myo-inositol (+59 pp), while amino-acid catabolism remained near-universal (>99.5%) in both groups — consistent with literature cited by the digest that cystic-fibrosis sputum is amino-acid-dominated. ₁

Among the free-living species of that Pseudomonas cohort, carbon profiles are significantly associated with isolation environment by PERMANOVA (permutational multivariate analysis of variance; p = 0.006), but predictive accuracy is modest — random-forest balanced accuracy 0.41 versus 0.25 chance — with D-serine, arabinose, rhamnose and fucose the most discriminating pathways; the test establishes an association only, and the modest accuracy leaves any environmental causation of catabolic repertoire untested. ₁

Community-scale NMDC metabolomics across the 131-sample H1 cohort supported the Black Queen Hypothesis (BQH, which predicts loss of costly biosynthetic capacity when the product is available from neighbours): 11/13 amino-acid pathways trended in the predicted direction, with a binomial sign-test p=0.011; leucine had r=-0.390, q=0.022, n=62, and arginine had r=-0.297, q=0.049, n=80, both among the most energetically expensive to synthesize (leucine 37 ATP equivalents; arginine 26 ATP equivalents). Methionine had r=-0.496 but q=0.117 with n=18, the largest effect yet underpowered and not FDR-significant, while tyrosine was an anti-BQH outlier at r=+0.42 and non-significant, for which an alternative tyrosine source via phenylalanine hydroxylation is offered as a likely explanation rather than a demonstrated one. The digest describes all 131 H1 samples as soil ₁, whereas the originating project report re-ran H1 on the 125 Soil-ecosystem samples only, excluding the 6 Unknown-ecosystem samples in the H1 dataset ₃; both descriptions are recorded here rather than reconciled. ₁

The leucine signal was reported robust to soil-only stratification (q = 0.022 unchanged) and to study-level blocking, although the dataset is 95% single-study. The digest adds that cross-study LC-MS protocol heterogeneity is effectively absent as a confounder — contrary to the naive expectation for NMDC metabolomics — and that the leucine Black Queen signal is internally consistent within one analytical protocol; that reassurance carries limited force, since the cohort nonetheless includes 6 samples from a second study. ₁

Carbon utilization, rather than amino-acid biosynthesis, dominated metabolic differentiation in PCA of a 220-sample × 80-pathway GapMind completeness matrix: PC1 explained 49.4% of variance and separated Soil from Freshwater at p<0.0001 (median PC1 Soil = +3.86, Freshwater = −6.28), with loadings dominated by carbon utilization pathways (glucuronate, fumarate, succinate, cellobiose, galactose) rather than amino-acid pathways, while amino-acid pathways loaded more strongly on PC2, which explained 16.6% and separated sub-clusters within Soil. ₁

17/18 amino-acid pathways were significantly differentiated across ecosystem types (q < 0.05, Kruskal–Wallis), but the digest identifies carbon substrate availability rather than biosynthetic capacity as the primary driver, so carbon substrate range will confound amino-acid pathway analyses whenever community metabolic potential is used as a feature space unless it is controlled for by PC1-residualized or ecosystem-stratified analyses. ₁

Antimicrobial resistance and defense

Across 14,723 species, the digest reports that efflux pumps constituted 21% of AMR in human-gut species but 1% in aquatic species (η²=0.127), while metal resistance constituted 45% in soil/aquatic species but 6% in human-gut species (η²=0.107); the denominators behind these mechanism fractions are not documented in the digest. ₁ The originating project report’s own mechanism table disagrees on efflux, giving human gut at 7.0%, aquatic at 1.1% and an η² of 0.055 ₄; both estimates are recorded here as attributed figures rather than reconciled or averaged. ₁ Clinical species had 68% accessory AMR versus 43% in soil. Within 823 multi-environment species, clinical strain fraction predicted total AMR at rho=0.465, p=2.2e-45, a correlation that does not establish causation. Klebsiella pneumoniae had 1,115 AMR clusters, only 7 of them core, across 13,637 genomes. ₁

Across 801 AMR genes in 28 organisms, cofitness neighborhoods were enriched for flagellar motility in 5 organisms, flagellum assembly in 5, histidine biosynthesis in 3, and tryptophan biosynthesis in 3 at FDR<0.05. InterProScan GO annotations, at 68% coverage, detected these signals whereas legacy Fitness Browser SEED annotations detected 0/280 significant enrichments; enrichment alone does not demonstrate co-regulation. Within the same networks, cross-mechanism Jaccard similarity exceeded within-mechanism similarity, 0.375 versus 0.207 (MWU p=4.3e-13), consistent with organism-specific rather than mechanism-specific support networks. AMR genes occurred in larger ICA modules than background genes, with median 46 versus 27 (p=1.7e-8); the log does not state how many of the 801 AMR genes were assigned to modules at all, so this module-size contrast rests on an unstated denominator. ₁

Random-effects meta-analysis of 801 AMR genes across 25 organisms found a pooled knockout fitness shift of +0.086 [+0.074, +0.098], with all 25/25 organisms positive. No difference in this cost was detected by mechanism (KW p=0.89), by core-versus-accessory conservation (p=0.33), or by annotation tier (p=0.26); these are non-significant tests and do not establish independence. The log does not state the comparator set behind the pooled shift, so the figure should not be read as an absolute benefit of deleting an AMR gene. Although no mechanism difference in fitness cost was detected, mechanism strongly predicted conservation status: metal resistance was 44% accessory versus 13% for efflux (chi-square=69.3, p=1.4e-13). Efflux genes showed a stronger antibiotic-dependent fitness flip than enzymatic inactivation genes, +0.094 versus -0.001, MWU p=0.007. ₁

Acquired AMR genes had stronger phylogenetic signal than core AMR genes across 1,261 species: median Mantel r=0.222 versus 0.117, paired t-test p=7.0e-16; core genes have near-zero Jaccard variance by definition (>=95% prevalence), which partly suppresses their distance-based signal, so the contrast is not purely biological and stable vertical inheritance remains an interpretation. Across 180,025 genomes and 1,305 species, 51.3% of within-species AMR occurrences were rare (≤5% prevalence), 41.3% variable (5-95% prevalence), and 7.5% fixed (>=95% prevalence); median strain pairwise Jaccard distance was 0.435. Resistance islands occurred in 705/1,305 species (54%), with 1,517 islands, mean phi=0.827, 88% containing multiple resistance mechanisms, and a maximum of 43 genes. ₁

Within species, 77% of Atlas Core genes were fixed and 79% of singletons were rare. Pangenome openness related weakly and negatively to AMR variability (Spearman rho=-0.193, p=2.2e-12), the opposite of the expected direction; the variability index measures only the 5-95% prevalence zone, so the log’s proposed explanation is that open-pangenome species accumulate more rare or singleton AMR genes below the 5% threshold, deflating the index rather than genuinely diversifying less. ₁

A DUF4041 (PF13250) survey of the 293K-genome KBase Data Lakehouse pangenome found the SNIPE defense system in 4,572 gene clusters across 1,696 species and 33 phyla — far exceeding the >500 homologues reported in Saxton et al. 2026 (Nature) — with 86.7% accessory or singleton genes, consistent with but not direct proof of mobile defense-island carriage, and a patchy distribution across many phyla and families that the digest reads as supporting horizontal gene transfer. Its nuclease domain belongs to Pfam family PF13455 (Mug113), not PF01541 (canonical GIY-YIG restriction enzymes and DNA repair); both sit in the GIY-YIG clan CL0418 but are distinct families, which is why zero gene clusters show DUF4041 plus PF01541 co-occurrence — biological truth rather than an annotation artifact — as verified on the E. coli SNIPE protein A0A0A1A5Z2 (PF13250 at positions 232–333, PF13455 at positions 443–520). ₁

The PhageFoundry phagefoundry_strain_modelling database, separate from the GenomeDepot annotation browsers, contained 17,672 binary infection outcomes from 188 E. coli strains and 96 phages, plus an ML model with AUC=0.883 and SHAP (a feature-attribution method) importances for 1,582 gene cluster features; the reported strain and phage counts imply more possible strain–phage pairs than the 17,672 recorded outcomes, a discrepancy the digest does not resolve. Lambdavirus 411_P1 (phage lambda) infected 1/188 strains (0.5%) versus 43.4% for Myoviridae, which the digest attributes to widespread ManYZ receptor loss — an interpretation rather than a measurement. ₁

Fitness Browser evidence (168 experiments, E. coli K-12 Keio collection) showed that ManXYZ is a mannose/glucosamine transporter rather than a fructose transporter, contradicting ManX’s UniProt annotation “fructose import across plasma membrane” (GO:0005354): the digest reports ManXYZ knockout fitness of -3.93 on D-mannose and -2.75 on D-glucosamine, but +0.18 to +0.66 on D-fructose. ₁ Those substrate assignments contradict the originating report, which gives per-gene values of manX -2.75, manY -3.00 and manZ -2.74 on D-mannose; manX -3.79, manY -2.79 and manZ -3.63 on D-glucosamine; and manX -0.22, manY +0.15 and manZ -0.07 on D-fructose, and which identifies -3.93 as manX’s worst fitness across its 168 experiments rather than as its D-mannose value; both readings are recorded here rather than reconciled. ₅

A full two-domain SNIPE protein was found in Methanococcus maripaludis JJ at locus MMJJ_RS01635, annotated with both PF13250 (DUF4041) and PF13455 (Mug113/GIY-YIG nuclease) plus PF10544 (DUF2525), and carrying 129 experiments of fitness data — the first SNIPE homologue with genome-wide knockout phenotypes available for analysis. The PhageFoundry Klebsiella genome browser carried 1 DUF4041 annotation across 3 proteins at 558 aa, the same length as the E. coli SNIPE protein, plus 4,619 PTS_EIIC (PF02378/ManY family) annotations, making Klebsiella the only PhageFoundry species with both SNIPE and the ManX-family PTS domain (PF00358); a consequent effect on phage-therapy efficacy in this pathogen is speculative. ₁

Microbiome and multi-cohort analysis

Pooling HMP2 and Franzosa metabolomics on 122 m/z-bridge metabolites caused PCA plus K-means K=4 clusters to separate completely by cohort rather than diagnosis; PC1 explained 79% of variance and cross-cohort LOSO ARI was 0.000 versus 0.113 for taxonomic ecotypes. Relative-abundance spaces were cross-cohort-portable because they are unitless and compositional, whereas absolute-intensity metabolomics requires explicit batch correction such as ComBat, SVA, RUV, or quantile normalization. Because batch dominated, within-pooled bootstrap stability metrics measured batch reproducibility rather than biological reproducibility; the failure here is specifically the cross-cohort-portability question, and within-cohort metabolite-feature ecotypes may still be useful inside a single laboratory’s analytical pipeline. ₁

A K=4 consensus of LDA (latent Dirichlet allocation) and GMM (Gaussian mixture model) clustering on 8,489 curatedMetagenomicData (CMD) MetaPhlAn3 samples — 5,333 HC (healthy controls) plus 3,156 IBD/other — produced E0 diverse commensal (n=3,604), E1 Bacteroides2 transitional (n=2,601), E2 Prevotella copri enterotype (n=920), and E3 severe Bacteroides-expanded (n=1,364). Healthy samples concentrated in E0 plus E2 at 84% combined, whereas CD and UC distributed across E1 and E3 rather than concentrating in any single ecotype, which is the patient-stratification signal the downstream within-ecotype analyses were built to exploit; data/ecotype_assignments.tsv is the consumable artifact for downstream notebooks. ₁

• E0 diverse commensal, n=3,604: F. prausnitzii / R. bromii / B. uniformis / Bifidobacterium balanced; 66.8% of HC. ₁
• E1 Bacteroides2 transitional, n=2,601: P. vulgatus / B. uniformis / B. dorei; reported as 48% CD, 58% UC, 100% T1D, 97% T2D, 67% nonIBD. ₁
• E2 Prevotella copri enterotype, n=920: P. copri 28%; 16.9% of HC and ~0% disease (non-Western healthy). ₁
• E3 severe Bacteroides-expanded, n=1,364: P. vulgatus 14% plus B. fragilis 3.6%; reported as 50% CD, 40% UC, 67% IBD acute flare, 38% CDI. ₁

Model-selection criteria did not identify K on their own: both LDA training perplexity (sklearn LatentDirichletAllocation) and GMM BIC on CLR (centered log-ratio) plus PCA-20 were monotone with K on this data, always preferring the largest K available, and held-out 5-fold perplexity gave the same monotone signal at this sample size (~8.5K samples). The discriminating signal was the cross-method adjusted Rand index (ARI, a chance-corrected measure of agreement between two clusterings) between LDA and GMM at each K, which has structure because the two model families agree only on K values reflecting real data partitions: ARI peaked at K=7 with 0.140 and at K=4 with 0.131, and a parsimony rule taking the smallest K within 0.02 of the peak selected K=4. ₁

Classifier and projection validation exposed several portability failures. A classifier trained on pooled HC plus IBD samples to predict K=4 ecotype from {is_ibd, sex, age} achieved macro OvR (one-vs-rest) AUC=0.80 in 5-fold cross-validation, exceeding its 0.70 “passes” threshold, yet reached only 41% agreement with the metagenomic projection on UC Davis patients, who are 100% IBD so that is_ibd=1 was constant; 19/22 patients were predicted E1 because the dominant learned rule was “IBD → E1 most of the time”, the marginal mode among IBD training samples. ₁

The AUC and the clinical task were not the same problem: the AUC counted HC-versus-IBD separability toward its score, whereas the UC Davis test requires the harder IBD-internal separability of E1 transitional versus E3 severe. The generalizable fix recorded here is to apply a classifier to an independent cohort with the cohort-axis variable held fixed before quoting an AUC as a measure of clinical utility, and to report the agreement rate on that cohort as the translation-honest metric, which applies to any project building clinical classifiers on pooled disease-plus-control data. ₁

Projection across taxonomic-classifier namespaces was asymmetric: the held-out Kuehl cohort was processed with Kaiju NCBI-NR read classification while the reference embedding was trained on MetaPhlAn3 marker-gene relative abundance. LDA on pseudo-counts was robust, placing UC Davis Kuehl samples in E0/E1/E3 at plausible proportions of 27/42/31%, whereas GMM on CLR plus PCA was fragile: the source states both that Kuehl detected only 54% of training species and that 70% of training species had no Kuehl detection, two percentages it reports without distinguishing their denominators and which are preserved here as a disclosed source inconsistency rather than reconciled; the CLR transform imputed the undetected species’ zeros with a small pseudocount, the resulting sparse vectors landed through PCA in the feature-space region of one dominant Gaussian regardless of actual biology, and all 26 Kuehl samples projected to E3 at GMM confidence >0.97 as an artifact. ₁

The recorded implication is to use pseudo-count or sparse-data-robust methods such as LDA or MMvec as primary when projecting across classifier namespaces, and to treat CLR-based methods such as ANCOM-BC projection or GMM on CLR as advisory at best; the asymmetry generalizes beyond IBD to any project that trains on one microbiome classifier and projects another, including MetaPhlAn↔Kraken, MetaPhlAn↔Kaiju, and 16S↔WGS. ₁

The v2 preliminary IBD report (2026-03-28) called C. scindens CD-enriched at log₂FC=+2.67 from pooled Mann-Whitney differential abundance (DA, testing whether a feature’s abundance differs between groups), and treated that call as suspect because C. scindens is described as a well-documented protective species: a secondary bile-acid producer via 7α-dehydroxylation, a TGR5 activator, present at ~79% prevalence in healthy individuals, and an inhibitor of C. difficile. ₁

That preliminary reading proposed three likely mechanisms for the enrichment, none of them demonstrated at the time — a compositional/relative-abundance artifact in which Ruminococcaceae loss in severe patients inflates the relative abundance of surviving C. scindens strains; strain heterogeneity, since not all C. scindens strains carry the bai operon so species-level presence is not functional presence; and ecotype mixing, since pooled analysis averages across patient subgroups with different microbiomes. Compositional-aware differential abundance (ANCOM-BC / MaAsLin2 / LinDA) run within patient ecotypes was the proposed fix at that point, not a reported completed validation. ₁

The later confound-free analysis contradicts that proposed artifact verdict rather than the pooled measurement: after leakage repair, C. scindens was genuinely CD-increased under within-substudy analysis, with pooled CLR-Δ +1.18, FDR=1e-8, and 4/4 sign concordance, and a leave-one-out refit showed C. scindens CD-increased in both E1 and E3 once the ecotype definition excluded the test species — so the earlier within-ecotype non-significant call, not the enrichment, was the leakage self-selection artifact. ₁

The quantified cost of feature leakage plus confounder non-adjustment was large. The within-ecotype notebook reported a 33-species Tier-A list (18 E1, 15 E3) with the C. scindens paradox marked “RESOLVED by stratification”, and two rigor-repair notebooks then applied three evidence filters to every candidate. ₁

The three filters were a bootstrap confidence-interval lower bound > 0.3 on within-ecotype CLR-Δ, repairing decision logic that read non-significance as resolution; LinDA CD-increase with FDR < 0.10 in the same ecotype, repairing reliance on a single DA method; and a within-substudy CD-versus-nonIBD meta-analysis across 4 IBD sub-studies requiring CD-increase and ≥ 66% sign concordance as a confound-free independent check. Only that third stream is independent of the ecotype definition that generated the candidate list. ₁

In E1, 0 of 18 candidates passed all three filters, and all 14 candidates that passed the first two had negative effects under the confound-free contrast, making them ecotype markers rather than CD drivers; the within-ecotype DA had lifted them to CD-increased through compounded feature leakage and compositional bias. ₁

In E3, 3 of 15 candidates passed all three filters: Mediterraneibacter gnavus (within-substudy CD-increase +5.13, LinDA E3 +1.64), Flavonifractor plautii (within-substudy +1.89, LinDA +2.91), and Blautia wexlerae (within-substudy +0.91, LinDA +2.00). These are the trustworthy Tier-A set. ₁

The ecotype-divergence hypothesis survived strongly once tested against a proper null: permutation-null p=0.000, with observed Jaccard 0.104 against a null mean of 0.785 ± 0.054. Stratification divergence is a real effect, and the earlier effect-size framing of that quantity was the wrong statistic rather than a refutation. ₁

The scope consequence is that per-ecotype Tier-A reduces from 33 candidates across E1 and E3 to 3 candidates in E3 only, and the current within-ecotype DA is not a usable E1 target list. The reformulations recorded are rebuilding ecotypes on a functional (pathway/KEGG) feature matrix to break taxonomic-feature leakage; running within-IBD-substudy CD-versus-nonIBD stratified by ecotype as the primary analysis, which requires substudy × ecotype × diagnosis cells with ≥ 10 samples each and still needs checking; or broadening to UC-versus-HC plus CD-versus-HC combined as IBD-versus-control for power. ₁

The invalidation is scoped rather than total: the ecotype framework itself remains usable for patient stratification, and only the within-ecotype DA was invalidated for target-selection use in the downstream Tier-A scoring and cocktail-draft deliverables. ₁

Pathway category schemas can reverse conclusions. Regex on pathway descriptive names matched 44 of 409 prevalence-filtered pathways to 7 a-priori IBD categories and placed only 3 of the 52 Crohn’s-disease-up (CD-up) pathways in that set; the pre-specified criterion — ”≤ 3 of 7 categories” — was structurally degenerate because the null also sat at 100% top-3, so the recorded fail verdict was uninformative rather than a refutation. The MetaCyc class hierarchy taken from ModelSEEDDatabase’s MetaCyc_Pathways.tbl and overlaid with 12 IBD themes had class data for 357/409 (87%) of pathways, assigned 262/409 (64%) to at least one theme, and supported iron/heme acquisition as the dominant CD-up theme at OR=8.1, FDR=7e-6, with 15 of 52 CD-up pathways. ₁

The regex schema missed iron biology mechanically: the name ‘PWY-5920: superpathway of heme biosynthesis from glycine’ contains no match for ‘iron’, while MetaCyc’s curator-validated hierarchy files that pathway under HEME-SYN, Heme-b-Biosynthesis, Cofactor-Biosynthesis, and Tetrapyrrole-Biosynthesis, so the regex ‘0_other’ bucket was hiding 15+ heme/iron pathways. The earlier ‘fail — degenerate’ verdict was technically correct for its own schema and should have been read as a flag to revisit the schema rather than as a true refutation of the biology. ModelSEEDDatabase’s shipped MetaCyc class hierarchy gives 90%+ coverage of HUMAnN3 outputs, and KEGG BRITE plus the GO (Gene Ontology) biological-process subontology cover analogous categorization needs for KEGG or EC-level data; the log therefore recommends curator-validated ontology hierarchies as primary and regex as sensitivity analysis. ₁

One project produced two independently developed six-line cross-corroboration narratives from a single dataset, each demonstrating the same biological claim across multiple analytical granularities. The iron-acquisition line ran per-actionable MIBiG lookup, sample-level pathway × species attribution, cohort-level pathway-class enrichment, sample-level species × pathway co-variation, genomic BGC (biosynthetic gene cluster) content, and AIEC (adherent-invasive E. coli) literature; the bile-acid 7α-dehydroxylation line ran within-carrier pathway differential abundance (DA), subject-level metabolite DA, a paired sample-level direct substrate–product signature, a strain-level informative null, and mechanism literature. The iron/heme result drove a four-way convergence with E. coli Yersiniabactin/Enterobactin MIBiG matches, heme-biosynthesis attribution to E. coli, and AIEC iron-acquisition literature. ₁

Pathway-level and metabolite-level differential abundance can diverge in direction without contradicting each other: the polyamine/urea pathway was CD-down at pathway level (OR=0.42) while polyamines as a metabolite class were CD-up at OR=14.6, the largest theme-level metabolite effect in that project. Both readings can be correct, because pathway abundance indexes encoded production capacity while a measured metabolite pool reflects the difference between production and consumption rates plus dietary and host inputs; neither quantity is a flux measurement, pool and flux measurements are not interchangeable, and both pool and pathway evidence belong in an analysis when both are available. ₁

For intervention design, each pathobiont target needs three per-target annotations alongside its headline Tier-A score. Whether CD specialization is iron/AIEC-mediated or driven by another mechanism informs whether to target a species broadly or only its AIEC subset. Bile-acid coupling cost matters because F. plautii, E. lenta, and E. bolteae actively 7α-dehydroxylate whereas M. gnavus and E. coli do not, so depleting active 7α-dehydroxylators shifts the bile-acid pool toward inflammatory primary tauro-conjugated forms. Third, the signal must be classed as species-abundance-mediated or strain-content-mediated: the null result of zero strain-adaptation gene signal in Kumbhari suggests the hypothesis that the effect is species-abundance-mediated, under which phage targeting would be expected to deplete activity predictably with no within-species strain-content escape route — an expectation that this single null does not itself demonstrate. ₁

Therapeutic consortium design

Across 142 CF (cystic fibrosis) airway isolates, metabolic overlap with PA14’s amino acid preferences (proline, histidine, ornithine, glutamate, aspartate) predicts planktonic inhibition at r=0.384, p=2.3e-6. The digest groups an R²=27.4% with that correlation, but the originating report attributes R²=0.274 to a four-feature multivariate model rather than to metabolic overlap alone, and reports five-fold cross-validated R²=0.145 ± 0.142 as an overfitting warning ₆; the strongest FDA-safe inhibitors — S. salivarius, N. mucosa and G. sanguinis — show 57-74% more inhibition than metabolic overlap predicts, from which dual-mechanism (metabolic plus direct antagonism) action is inferred rather than established by that residual. ₁

PA14 is a metabolic generalist: it outgrows the average commensal on all 20 tested amino acids plus glucose and lactate, with a selectivity ratio <1.0 for all, and on that basis the digest proposes that competitive exclusion requires community-level niche coverage by 3+ organisms rather than individual superiority. ₁ That comparison is against the average commensal on endpoint growth and does not mean no commensal ever has an advantage: the originating report separates endpoint growth from kinetics, with commensals beating PA14’s maximum growth rate on 13.8% of substrate comparisons and starting growth before PA14 — a lag advantage greater than 0 — on 43.1% of comparisons. ₆

GapMind pathway comparison, as the digest states it, identifies 6 carbon sources where formulation commensals have 100% pathway completeness but PA14 has 0%: xylitol, myoinositol, xylose, arabinose, fucose and rhamnose. ₁ The originating report qualifies both halves of that framing: each of the six pathways is carried by only 1–2 of the 5 core commensal species, so these are species-specific capabilities rather than consortium-wide ones, and PA completeness is 0% only for myoinositol, xylitol, xylose and arabinose, with fucose and rhamnose at 1% (selectivity 0.99 rather than 1.00), so neither categorical absence in PA14 nor consortium-wide commensal coverage holds; the report accordingly pairs prebiotics per species — xylitol for S. salivarius, xylose/arabinose/fucose for N. mucosa and G. sanguinis — and suggests a multi-prebiotic cocktail. ₆ These remain genomic prebiotic predictions rather than tested interventions; the digest notes that the compounds are FDA-GRAS, that xylitol is already used in CF airway products, and that patient metatranscriptomics shows commensal-specific PTS sugar transport expression in vivo as supporting evidence. ₁

In the KBase Data Lakehouse pangenome, 38% of R. dentocariosa genomes and 33% of N. mucosa genomes come from respiratory sources — disproportionately lung-associated relative to other formulation species — which the digest uses to anchor the FDA-safe formulation on inhibition, engraftability and lung tropism. ₁

Environmental, plant, and subsurface ecology

In a 3×3 SSO well grid spanning approximately 6 m, community similarity showed distance decay with Mantel rho=0.323 and p=0.029, driven by the east-west axis rather than the hillslope gradient; a U3-M6-L7 diagonal followed the inferred NE-to-SW plume flow path, and well M5 hosted a Rhodanobacter denitrification hotspot of 7.7% at the plume mixing zone. The redox ladder across the grid was inferred from genus-level composition rather than measured. Groundwater communities were temporally stable, with well R²=49.9% and date R²=0.8% over 9 days. Candidatus Nitrosotalea, an archaeal ammonia oxidizer, and Sideroxydans, an iron oxidizer, co-occurred at rho=+0.95 across the 9 wells, concentrated at U3; a shared niche is an inference from that co-occurrence rather than a measurement. ₁

The SSO campaign registered 221 METALS/ICTOC/ISOTOPES/NH3NO2 sample tubes in ENIGMA CORAL, but zero Assay Geochemistry processes were linked, so the analytical geochemistry values were not ingested. This limits validation of the contamination-plume model. ₁

In the ENIGMA contamination-functional project (108 overlap samples), primary univariate associations were non-significant across defense, stress, metabolism and mobilome outcomes in both mapping modes; the contamination–defense association became significant only after coverage-aware sensitivity modelling (coverage-adjusted ordinary-least-squares contamination p = 3.98e-4 relaxed, 3.54e-3 strict; high-coverage subset Spearman p = 0.0207 relaxed, 0.0098 strict), so the strongest signal is conditional on mapping coverage and model specification rather than broad across outcomes. ₁

After explicit confirmatory-versus-exploratory labelling and global BH-FDR across all reported p-values, the confirmatory defense Spearman tests in genus-level modes remained null (p=0.546/0.483; q=0.862/0.849), while the relaxed coverage-adjusted defense model retained q=0.046 and the strict and other sensitivity signals attenuated above q<0.1 — so the null confirmatory evidence takes priority and the surviving positive result should be framed as exploratory. ₁

Re-testing the confirmatory defense endpoint under four contamination-index constructions — all-metals composite, uranium-only, top-3 variance metals and PCA-PC1 — produced no significant confirmatory association after FDR correction (all q=0.546 across 8 confirmatory-variant tests), indicating the null confirmatory result is not an artifact of one index formulation. ₁

Strict-versus-relaxed mapping sensitivity is analytically meaningful and should not be approximated by copied feature tables: computing both modes independently changed 1,140 of 1,590 genus-feature pairs (approximately 71.7%) and changed site stress scores for all 108 samples. ₁

enigma_coral.ddt_brick0000454 provides taxonomy only through Genus, with no species or strain rows; a species-proxy mode (species_proxy_unique_genus) keeping only unique genus-to-single-GTDB-clade mappings retained 150 genera but collapsed mapped abundance to mean 0.031 versus 0.343 in the strict and relaxed genus modes, and in that low-coverage regime the defense trend was positive (rho=0.169) but non-significant (p=0.081), indicating that true species-level gains likely require higher-resolution ENIGMA taxonomy or metagenomic mapping. ₁

Among 25,660 plant-associated species, 60–85% were classified as dual-nature, carrying both plant growth-promoting (PGP) and pathogenic marker genes simultaneously; marker co-occurrence alone does not establish a beneficial or pathogenic phenotype. Plant compartment (root, leaf, rhizosphere and others) explained 53% of functional-profile variance by PERMANOVA (permutational multivariate analysis of variance; R²=0.53, pseudo-F=235, p=0.001), a single estimate that does not separate compartment-location effects from dispersion or taxonomic sampling; PGP genes were 64.6% core versus 45.2% for pathogenic genes (Mann–Whitney p=3.4e-125), which the log reads as suggesting beneficial functions are under stronger purifying selection — an interpretation rather than a measurement; and ACC deaminase (acdS) showed the strongest compartment-specific enrichment of any single marker gene, with an odds ratio of 69.3 in root-associated species — an estimate not described as phylogenetically controlled. ₁

Co-occurring plant-associated genera showed functional redundancy rather than complementarity, with permutation p=1.0 and Cohen’s d=-7.54; the reported p=1.0 does not itself support a significant permutation result, and no corrected estimate is given. After phylogenetic control, 50 novel eggNOG ortholog groups distinguished plant-associated species from non-plant species, led by COG3569 at OR=6.01; the plant-interaction functions of these groups remain candidate interpretations. Singleton marker-gene clusters co-occurred with transposase/integrase singletons at OR=15.95, p=8.8e-20 (described as 16x more likely), which suggests but does not demonstrate horizontal transfer. ₁

The pqqC–acdS pair co-occurred across 27K GTDB species at OR=7.24 — far stronger than any other PGP gene pair; acdS and pqqC were enriched in soil/rhizosphere environments at OR=7.0 and OR=2.9, respectively. All 13 PGP genes were predominantly core, with 29.7% mean accessory versus 53.2% genome-wide, rejecting the tested HGT hypothesis in favor of vertical inheritance. Nitrogen fixation (nifH) forms a separate ecological guild: negatively associated with pqqC and hcnC, and depleted in soil relative to aquatic environments. ₁ The digest’s trp→ipdC wording misassigns OR = 2.81 to non-soil species; the originating report gives OR = 2.81 overall, OR = 3.56 in non-soil species (n = 10,233, p = 7.7e-13) and OR = 0.30 in soil species (p = 0.02), and because only 214/11,272 species (1.9%) carry ipdC it treats the soil reversal as hypothesis-generating rather than conclusive, with the proposed root-exudate explanation untested. ₇

Within Bacillota_B, deep-clay specialization was associated with expansion rather than streamlining: the 10 deep-clay anchor genomes had CheckM-rescaled size 4.3 Mbp versus 3.2 Mbp in 62 phylum-matched soil genomes (d=+1.37, Mann–Whitney p=0.013), reported as 35% larger, and 2,771 versus 2,233 eggNOG ortholog groups (d=+1.32, p=0.009), with comparable completeness of 94.7 versus 94.3 (p=0.93) ruling out a MAG-quality artefact — so the “small is mighty” subsurface streamlining pattern documented for Patescibacteria/CPR does not generalize to these cultivable subsurface Firmicutes, which show the opposite pattern under the same nominal subsurface oligotrophy. The clay project’s iron-reduction side used K07811, K17324 and K17323 as markers, which are instead TMAO reductase (KOfam threshold 1336) and glycerol ABC transport ATP-binding (363) and permease (318) — none are iron reduction, and KEGG has no canonical KO for the Geobacter omcS / Shewanella mtr multi-heme outer-surface cytochromes. Re-running with a corrected triple-signal detector (PF02085 Cytochrom_CIII plus PF22678 Cytochrom_c_NrfB-like plus ≥4 CXXCH heme-binding motifs) changed deep, shallow, and soil rates to 56% (was 11%), 40% (was 50%), and 41% (was 20%); all cohort comparisons had Fisher p≥0.46, invalidating the original iron-reduction interpretation, while the sulfate-reduction side (K11180/K11181/K00394/K00395/K00958, p=4×10⁻¹² against the rock-attached null) remains intact. A separate and distinct failure mode affects the cohort-level summary scores: unadjusted toolkit-score enrichment of deep 1.89 versus baseline 0.39 looks like a strong deep-clay-versus-soil signal, but Bacillota_B — the phylum containing Desulfosporosinus, BRH-c4a and BRH-c8a, the lineages that dominate cultured clay isolates — carries Wood-Ljungdahl plus group 1 [NiFe]-hydrogenase at high background rates even in soil samples (Bacillota_B baseline mean toolkit 1.65), so this apparent habitat signal is attributed to phylogenetic confounding and biased habitat sampling — a phylum-level signature — rather than to marker-detection error. The generalizable warning for subsurface comparative genomics is that an apparent habitat signal at cohort level may be a phylum-level signature picked up by biased habitat sampling. ₁

The clay cohort was enriched for sulfate-reduction markers relative to the Mitzscherling rock-attached null: 5/9 deep-clay genomes were SR-positive, with binomial p=4×10⁻¹². However, within Bacillota_B, Wood-Ljungdahl and group 1 [NiFe]-hydrogenase were not enriched in deep clay, with 5/5 versus 15/19 and 5/5 versus 14/19, respectively, both p=0.54; only sulfate reduction was enriched within the phylum, at 5/5 versus 4/19 with p_BH=0.04. ₁

Two published paradigms describe qualitatively different microbial communities at the same Mont Terri Opalinus clay site: porewater dominated by sulfate reducers, rock surfaces by Geobacter/Geothrix iron reducers. The log converts this into a cultivation-bias diagnostic — for any cohort of cultured clay-isolated genomes, compute sulfate-reduction marker (dsrAB-aprAB-sat) and iron-reduction marker (omcS / mtrC / mtrA) prevalences and compare them against the rock-attached frequencies of SRB ~0.2% and IRB ~7% by binomial test. On that test the 9 anchor_deep clay genomes read as the porewater fraction rather than the rock-attached fraction, and the framework is proposed as generalizable to any future cultured cohort drawn from comparable pangenome resources. ₁

Metal tolerance and metabolic ecology

Across 7,609 metal-important gene records in 24 organisms, 4,177 (55%) were genuinely metal-specific — sick under metal stress but with a <5% sick rate across 5,945 non-metal experiments (antibiotics, osmotic, carbon sources and others) — while 38% were “general sick”, important across many conditions, and 7% were metal-plus-stress, shared with other stresses; on that basis the digest concludes the metal-important gene set is not dominated by general stress response. ₁

Metal-specific genes were 84.8% core (pooled) versus 90.2% for general-sick genes, a significant difference by Cochran–Mantel–Haenszel test (p=0.011), with both categories enriched above the 79.8% baseline — which only partially supports a two-tier model of core general stress response versus accessory specialized metal resistance, since both are predominantly core and the accessory genome contributes only modestly to metal-resistance machinery. ₁

UCP030820/OG01015 (67% metal-specific, oxidoreductase, 7 metals), YebC/OG01383 (58%, transcriptional regulator/translation factor, 6 metals, 11 organisms) and DUF1043-YhcB/OG03264 (50%, cell division protein, 5 metals) were the most metal-specific novel candidates, whereas YfdZ (15%) and the Mla/Yrb system (0-25%) are pleiotropic and DUF39 (0%, sick rate 0.64) is a general fitness factor despite spanning 8 metals. ₁

A proposed but untested explanation for YebC’s metal specificity is that YebC, recently shown to be a translation factor for proline-rich proteins, is needed under metal stress because upregulation of proline-rich metal transporters — P-type ATPases CopA/ZntA and CDF transporters, which carry proline-rich cytoplasmic loops — creates a translation bottleneck that YebC resolves, with those transporters not highly expressed and YebC dispensable under non-metal conditions. ₁

An initial 0% metal-specificity for essential metals (Mo, W, Se, Mn) was an artifact of a locusId type mismatch that had silently excluded DvH, whose 608 non-metal experiments otherwise make specificity hard to attain; after the fix, manganese reached 61%, molybdenum 61%, tungsten 57% and selenium 46%, so the experiment-count bias is real but less severe than initially thought. The digest’s summary range of 47-61% does not match its own listed selenium value of 46%, and both are recorded as reported. ₁

Counter-ion analysis found that 39.8% of metal-important genes overlapped NaCl-important genes across 19 organisms and 14 metals, which the digest attributes to shared stress biology (cell envelope, ion homeostasis, DNA repair) rather than chloride counter-ion contamination: zinc sulfate, which contains 0 mM Cl⁻, had 44.6% NaCl overlap and the highest DvH profile correlation, r=0.715 — higher than most chloride-delivered metals. After removing the approximately 40% shared-stress fraction, core enrichment persisted for 12 of 14 metals, and essential metals showed stronger enrichment after correction, with the molybdenum delta moving from +0.132 to +0.145, so the original atlas conclusion (OR=2.08) is not an artifact. Synechococcus elongatus (SynE) was an experiment-count outlier with 12 NaCl dose-response experiments spanning 0.5–250 mM against 1–6 for every other organism, which makes the n_sick >= 1 importance threshold much easier to satisfy and produced an 88.6% shared-stress rate; excluding it reduced the overall overlap from 39.8% to 36.7%, so analyses using NaCl fitness data should account for experiment count when setting importance thresholds. ₁

Whole-genome DvH fitness correlation with NaCl ranks the metals Zn (r=0.72) > Mn (0.55) > Cu (0.53) > Co (0.50) > Hg (0.48) > Ni (0.45) > Al (0.42) > Mo (0.40) > U (0.35) > Se (0.34) > Cr (0.32) > W (0.30) > Fe (0.09), which the digest reads as a general-to-specific toxicity gradient: metals that broadly displace cofactors share more genes with ionic stress, while metals targeting specific pathways (Mo/W for molybdopterin, Fe for iron-sulfur clusters) are distinct from general stress. ₁

BacDive metal-tolerance scores predicted isolation from heavy-metal-contaminated environments at Cohen’s d=+1.00, Mann–Whitney p=0.006, n=10 — a full standard deviation above the environmental baseline, with the signal reportedly holding within Pseudomonadota and Actinomycetota after phylogenetic stratification; the dose pattern was heavy metal +1.00, waste/sludge +0.57, all contamination +0.43, and industrial +0.20. Host-associated bacteria (n=12,086) unexpectedly scored higher than environmental bacteria at d=+0.14, p<0.0001, contrary to the expectation that host niches have lower metal exposure; genome-size and taxonomic-composition confounding — BacDive host-associated strains dominated by large-genome Pseudomonadota pathogens with correspondingly more KEGG-annotated gene clusters, inflating the normalized score — is proposed rather than established. ₁

Analytical rigor and data-quality lessons

Across 9 adversarial review rounds, two citation failures recurred: PMID hijacking, where real PMIDs pointed to unrelated papers, and hallucinated author lists. Load-bearing citations require verification of author, title, journal, PMID, and DOI through PubMed, not merely DOI resolution. ₁

Methodology revisions M1–M26 are not automatically a multiple-testing family. The referenced project had 4 pre-registered hypotheses, for which Bonferroni α=0.0125 was appropriate and all 4 survived; only M14, M21, and M23 were genuine post-hoc metric corrections. Family-wise error correction should be applied to actual hypothesis-test families, not to transparent documentation of methodological changes. ₁

Standard review and adversarial review identified different classes of problems. Standard review called one project exceptionally sophisticated with no critical issues and recommended proceeding with the existing Tier-A candidate list, whereas adversarial review of the same files — framed explicitly to find flaws and not be diplomatic — found 5 critical and 6 important issues: feature leakage in within-cluster DA, which clustered samples on taxon abundances and then tested the same taxa within cluster; hard-coded “RESOLVED” verdict logic that treated non-significant results as positive evidence; zero confounder adjustment despite citing Vujkovic-Cvijin 2020, which calls for it; Jaccard 0.14 reported as supporting the divergence hypothesis without any null distribution; and an ecotype framework with 48.9% cross-method agreement and no external replication. Plan revisions and load-bearing notebooks should receive both review modes before downstream claims are accepted. ₁

Generalizing from that single paired-review example, the digest offers its own assessment of the division of labour between the two modes: it holds that the standard reviewer is strong at surface flaws — missing sections, undocumented dependencies, broken figures, simple statistical errors — and appropriate for routine submissions, but systematically weak at structural and inferential issues such as selection-on-outcome confounding, missing null distributions, post-hoc choice of decision rules, and narrative-versus-evidence mismatch, which the digest argues require an explicit adversarial framing that the default reviewer prompt does not invoke. That characterization is the digest’s judgement rather than a separately demonstrated property of the reviewers. ₁

Five failure classes recurred across review rounds: effect-size thresholds stated without null distributions (Jaccard 0.14 in one notebook, |ρ| > 0.4 in a plan revision); substudy or cohort sample-size claims that did not match the data on inspection; hypothesis labels that did not match the operationalized test (“H2c RESOLVED” treating non-significance as positive, and a “butyrate-producer” anchor set composed mostly of non-butyrate-producers); same-axis feature leakage in stratified analyses (clustering on taxa then testing taxa, and stratifying pathway DA by a pathway-feature ecotype); and falsifiability rules loose enough to pass under random data, as ”≤ 5 categories of 7” passes on random allocation. ₁

The common thread is pre-registration discipline: what the actual sample size is given the actual data, what null distribution any positive claim is being made against, and what each operational threshold means given that null. Standard review treats these as advanced questions; adversarial review treats them as entry-level diligence. ₁

Adversarial review applied at plan scope was equally productive: review of one plan revision found 4 critical and 10 important issues, including feature leakage in a dual-basis design, a ‘four-substudy meta-viable’ count inherited without re-verification that was actually 3+1, hypothesis thresholds set without null distributions, a ‘butyrate-producer ↔ pathobiont’ label whose 4 of 6 named anchors were not butyrate producers, a strain-adaptation hypothesis that collapses to species level by construction, and a cohort of n=67 across 3 sites described as a 130-subject study. Catching such issues at plan scope was cheaper than at notebook scope: the comparable notebook-scope repair took 7 notebooks, whereas the plan-scope repair was a single plan-edit pass. ₁

Large KBase Data Lakehouse catalogs should be discovered through access-aware helpers rather than historical inventories, using get_databases(), get_tables(), and get_table_schema(): historical static inventory files lagged the live catalog and omitted databases that were in fact present. The live catalog included kescience_mgnify, PhageFoundry GenomeDepot databases, kescience_interpro, kescience_pubmed, pangenome_bakta, arkinlab_microbeatlas, protect_integration, and user-owned Klebsiella databases that were absent from older snapshots. ₁

• kescience_mgnify — EBI MGnify, the metagenomics service, relevant for IBD-cohort cross-validation, and the correct spelling behind a mis-recalled ke_science_magnify. ₁
• phagefoundry_ecoliphages_genomedepot, plus a duplicate _genomedepot variant — an E. coli phage genome collection directly relevant for AIEC targeting; PhageFoundry has 7 databases in the tenant, not the 5 documented. ₁
• kescience_interpro — InterPro domain annotations, the primary source for virulence-factor classification. ₁
• kescience_pubmed — literature text mining, sibling to kescience_paperblast. ₁
• pangenome_bakta — a separate Bakta-pangenome slice distinct from kbase_ke_pangenome. ₁
• arkinlab_microbeatlas — 464K global 16S samples, a useful prevalence and biogeography baseline. ₁
• protect_integration — a second Protect-tenant database beyond protect_genomedepot. ₁
• u_kazakov__klebsiella_genomedepot and u_kazakov__strain_modelling — user-owned Klebsiella phage-host work worth checking with the owner. ₁

Local data marts benefit from lineage.yaml, schema_overview.yaml, per-table YAML dictionaries, and sentinel-code tables distinguishing pending data from true NULL values. The UC Davis / Arkin CrohnsPhage data mart is the worked example: lineage.yaml carries ETL provenance, changelog and known gaps, schema_overview.yaml a table inventory by category, per-table *.yaml dictionaries list columns, dtypes, null counts, unique counts and sample values, and a ref_missing_data_codes table holds sentinel codes such as PENDING_DAVE_LAB, PENDING_KUEHL and PENDING_HMP2_RAW that distinguish real NULLs from known-pending data, making large local marts agent-readable without any live queries. Large Spark joins touching the 2.5B-row UniProt identifier table can exceed the 1GB spark.driver.maxResultSize through broadcast exchange and should disable broadcast joins with SET spark.sql.autoBroadcastJoinThreshold = -1 to force sort-merge joins; notebooks executed through nbconvert should cache large Spark outputs as CSV because direct toPandas() calls on large result sets can kill the kernel (DeadKernelError) — the recorded workaround adds cache-check logic to Spark cells (if the CSV path exists, load it; otherwise run Spark), so live Spark outputs are captured interactively on JupyterHub and nbconvert then re-executes the notebook locally from cached CSVs. ₁

Two further query-construction notes are recorded: querying average nucleotide identity for all 13K+ genomes in a single IN clause fails or times out, so the query must iterate over species and request ANI per species; and the pangenome table already contains no_core, no_aux_genome, no_singleton_gene_clusters and no_gene_clusters, so these counts need not be computed from the gene_cluster table. ₁

Several Fitness Browser table schemas differ from the documented schema: keggmember uses keggOrg/keggId rather than orgId/locusId and must be joined through besthitkegg; kgroupec uses ecnum rather than ec; seedclass carries orgId, locusId, type, num rather than a subsystem/category hierarchy; and the fitbyexp_* tables are long format rather than pre-pivoted as documented. ₁

Identifier namespaces inside Fitness Browser are also not uniform: the aaseqs file (fit.genomics.lbl.gov/cgi_data/aaseqs) uses RefSeq-style locus tags such as ABZR86_RS* for some organisms while the gene table uses the original annotation locus tags such as N515DRAFT_*, which caused a complete join failure for Dyella79 at a 0% merge rate; only 1 of 34 organisms was affected, but any pipeline joining aaseqs-derived data with gene-table data should verify locus-tag consistency first. ₁

The notebook-utility Spark session helper (setup_spark_session.get_spark_session) works from regular Python scripts on JupyterHub and not only from notebook kernels, so full analysis pipelines can run from the command line without jupyter nbconvert; the auto-import in /configs/ipython_startup/00-notebookutils.py affects notebook kernels only. ₁

Caveats

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. ₁

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. ₁

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. ₁

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. ₁

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. ₁

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. ₁

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. ₁

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. ₁

Slots Into

• gene-essentiality — universal, variable, orphan, and module-transferred essentiality findings provide cross-organism evidence for the structure of bacterial essential genomes. ₁
• pangenome-integration — pangenome conservation, accessory-gene innovation, dark matter, plant-associated genes, and ecological breadth findings extend cross-source genome-content synthesis. ₁
• condition-specific-fitness — ADP1 respiratory wiring, metabolic dependencies, trade-offs, AMR costs, and pathway-capability findings refine condition-specific fitness interpretation. ₁
• cofitness-network-architecture — ICA module stability, cross-organism module families, annotation granularity, and AMR support networks add architecture and validation evidence. ₁
• metabolic-model-gapfilling — FBA blind spots, GapMind-versus-fitness comparisons, and pathway completeness versus dependency directly identify model gaps. ₁
• multi-omics-integration — metabolite pools versus pathway capacity, NMDC community metabolomics, and multi-line microbiome corroboration support cross-modal integration. ₁
• environmental-resistome — environment-dependent AMR composition, phylogenetic retention, resistance islands, and metal-specificity findings expand the environmental resistome synthesis. ₁
• subsurface-bacillota-specialization — clay cohort bias, Bacillota_B genome expansion, sulfate-reduction enrichment, and corrected iron-reduction markers refine subsurface specialization. ₁
• functional-marker-validation — corrected multi-heme cytochrome detection overturns the clay iron-reduction signal and establishes a marker-validation requirement. ₁
• cross-tenant-data-bridging — live catalog discovery, BacDive-to-GTDB matching, PhageFoundry data, and KBase Data Lakehouse data-mart conventions provide cross-tenant integration lessons. ₁
• condition-specific-fitness — microbiome ecotype leakage, classifier mismatch, and cohort-batch effects define limits on transferring condition-specific models across cohorts. ₁
• gene-function-acquisition-depth — Sankoff-attributed gain depths and leaf-consistency validation date function classes by acquisition depth. ₁
• horizontal-gene-transfer-driven-innovation — the CRISPR-Cas recent-versus-ancient contrast against strict housekeeping anchors an HGT-active signature. ₁
• taxonomic-resolution-dependent-functional-inference — family-rank effects as sub-clade mixtures, and strain-content signals that collapse to species level by construction. ₁
• phenotype-database-coverage-bias — per-source corroboration coverage limits which clades can be anchored to curated phenotypes. ₁
• adversarial-research-quality-assurance — citation-verification failures, review-mode contrast, plan-scope adversarial review, and correct scoping of the multiple-testing family. ₁
• selection-on-outcome-leakage — feature leakage in a dual-basis design caught at plan scope. ₁
• cross-cohort-microbiome-portability — metabolomics clusters separating by cohort versus the portability of compositional taxonomic spaces. ₁
• study-batch-confounding-of-environmental-associations — batch-dominated bootstrap stability measures batch reproducibility, not biology. ₁
• ontology-and-category-schema-sensitivity — regex names versus a curated MetaCyc hierarchy reverses the iron/heme conclusion. ₁
• ecological-cost-of-microbiome-target-depletion — bile-acid coupling cost and mediation-mode annotations for pathobiont targets. ₁
• occurrence-versus-catabolic-activity — metabolite pools reflect production minus consumption plus dietary and host inputs, not flux. ₁
• cultivation-collection-bias-in-ecological-genomics — a marker-prevalence diagnostic showing a cultured clay cohort samples the porewater fraction. ₁
• subsurface-hydrogeological-zonation — porewater versus rock-attached community paradigms at the same clay site. ₁
• evidence-triangulation-for-functional-annotation — eggNOG-KO and Bakta-product overlap as a reportable annotation gap rather than a single count. ₁
• homology-search-negative-evidence — preferred-name hits with zero validated overlap treated as zero-information. ₁
• gut-microbiome-ecotypes-as-patient-strata — the K=4 CMD ecotype consensus and its disease distribution define candidate patient strata. ₁
• ecotype-clustering-validity — monotone perplexity and BIC versus cross-method ARI as the K-selection signal. ₁
• classifier-database-compatibility-in-taxonomic-quantification — Kaiju-versus-MetaPhlAn3 namespace mismatch breaking CLR-based projection but not pseudo-count LDA. ₁
• compositional-robustness-of-differential-abundance-calls — Tier-A candidates surviving bootstrap, LinDA and within-substudy gating. ₁
• provenance-aware-resource-discovery — live access-aware catalog discovery versus lagging static inventories. ₁
• data-landscape-ownership-and-coverage-bias — tenant-owned and user-owned databases absent from documented inventories. ₁
• phylogenetic-confounding-of-pangenome-associations — Bacillota_B background rates turning a habitat contrast into a phylum-level signature. ₁
• antimicrobial-resistance-fitness-cost — the pooled knockout fitness shift, the absence of detected mechanism, conservation and tier differences in that cost, and the efflux-versus-enzymatic antibiotic-dependent flip. ₁
• condition-dependent-gene-tradeoffs — efflux genes showing a larger mean antibiotic-dependent fitness flip than enzymatic-inactivation genes, a comparative difference rather than an absence of effect in the latter. ₁
• resistance-island-coinheritance — 1,517 tightly co-occurring resistance islands across 705 of 1,305 species. ₁
• pangenome-openness-determinants — openness versus AMR variability and the null environment/phylogeny relationships with openness. ₁
• pangenome-core-boundary-and-clade-size-bias — prevalence-threshold definitions suppressing core-gene distance-based signal, and core-versus-singleton fixation rates. ₁
• experimental-prioritization-of-functional-dark-matter — the annotation-clue inventory that leaves only 3.8% of truly dark genes clueless. ₁
• fitness-condition-coverage-prioritization-bias — dark genes with strong phenotypes depleted in stress and enriched in nutrient, mixed-community and iron conditions. ₁
• fitness-module-detection-sensitivity — the AMR module-size contrast resting on an unstated module-assignment denominator. ₁
• research-attention-inequality — large-scale reclassification of previously dark genes once updated annotation is applied. ₁
• environment-embedding-geography — phylogeny-versus-environment partial correlations across 172 species and the null host-association contrast. ₁
• two-speed-bacterial-genome — plant growth-promoting genes core versus pathogenic genes accessory. ₁
• community-metabolic-interdependence — functional redundancy rather than complementarity among co-occurring plant-associated genera. ₁
• scale-dependent-mobile-element-associations — singleton marker-gene clusters co-occurring with transposase/integrase singletons. ₁
• respiratory-capacity-and-nadh-load — ADP1’s per-substrate respiratory wiring, similar standard-condition dehydrogenase expression, and the corrected NDH-2 compensation null. ₁
• cross-species-fitness-transferability — ortholog-scope effects on module families and the ortholog-transfer-versus-Module-ICA prediction benchmark. ₁
• genomic-dispersal-functional-coupling — ACIAD3137/ACIAD2176 near-perfect cofitness with Complex I and the quinate-specific support-gene network around the β-ketoadipate pathway. ₁
• composite-functional-annotation — composite COG assignments such as LV and EGP treated as genuine multi-functional genes. ₁
• phage-defense-syndromes-and-arms-race — the mobile-plus-defense LV signal as a candidate mobile defense island. ₁
• genetic-perturbation-coverage-bias — the 272 TnSeq-dispensable genes missing from the ADP1 deletion collection. ₁
• core-gene-annotation-paradox — essential-core genes enzyme-rich while orphan and unmapped essentials are hypothetical-rich. ₁
• pangenome-conservation-fitness-decoupling — the quantitative core-fraction gradient across essentiality, fitness-effect and fitness-breadth bins. ₁
• costly-dispensable-gene-loss — costly-plus-dispensable genes as candidates for ongoing gene loss. ₁
• laboratory-fitness-versus-natural-selection — costly-plus-conserved genes read as purifying selection in conditions the laboratory does not capture. ₁
• sampling-depth-and-downsampling-effects — pilot-subset ortholog graphs and single-genome pangenomes filtered from species metadata. ₁
• core-genome-burden-paradox — core genes more often beneficial when deleted, and the function-specific reversal for cell wall genes. ₁
• module-level-coinheritance — ICA fitness modules enriched in core genes, with no further gradient by module-family organism breadth. ₁
• transposon-callability-bias — zero essential genes in any fitness module, because insertion-free genes carry no fitness data. ₁
• ecotype-environment-gene-content — the reanalysis that keeps phylogeny dominant despite human-associated sampling bias. ₁
• spatial-sampling-effort-confounding — high-density coordinate clusters that include legitimate DOE field sites. ₁
• confirmatory-exploratory-ecological-association-discordance — null confirmatory ENIGMA contamination endpoints against a coverage-adjusted exploratory signal. ₁
• callability-limited-comparative-inference — amino-acid pathways left untestable by absent compound detections. ₁
• lab-field-fitness-concordance — dark-gene laboratory-versus-environment concordance indistinguishable from annotated genes. ₁
• metabolic-capacity-specialization — the P. aeruginosa group’s near-complete loss of plant-derived sugar catabolism. ₁
• within-species-conservation-between-species-functional-divergence — carbon-pathway divergence between Pseudomonas groups alongside near-universal amino-acid catabolism. ₁
• cross-condition-metabolic-comparability — the Web of Microbes production-only action encoding, name-based compound matching, and structurally driven four-database concordance. ₁
• computational-pathway-prediction-validation — GapMind completeness checked against Fitness Browser growth and Web of Microbes production in FW300-N2E3. ₁
• biosynthetic-prototrophy-and-auxotrophy — latent biosynthetic capabilities that become fitness-important only under condition-specific analysis. ₁
• sample-size-aware-phenotype-consensus — the 0/52 BacDive tryptophan-utilization result set against 231 fitness-significant genes. ₁
• condition-space-dimensionality — housekeeping genes significant in all 21 metabolite conditions versus the ~370 genes significant in only 1-2. ₁
• shared-stress-versus-stressor-specific-fitness — metal-specific versus general-sick classification and the NaCl shared-stress correction. ₁
• metal-cross-resistance — zinc sulfate’s chloride-free overlap, the per-metal NaCl correlation gradient, and the metal-specific candidate genes. ₁
• composite-resistance-score-limitations — BacDive tolerance scores confounded by genome size and lineage composition. ₁
• taxonomic-nomenclature-reconciliation — BacDive-to-GTDB species matching leaving 56.6% of strains unmatched. ₁
• phage-therapy-evidence-translation — the PhageFoundry infection matrix and Klebsiella SNIPE carriage as therapy-relevant but indirect evidence. ₁
• competitive-exclusion-consortium-design — PA14’s generalist nutrition, commensal-only prebiotic sugars, and respiratory-source formulation anchors. ₁
• gene-cooccurrence-ecological-guilds — nifH as a guild separate from pqqC/acdS, and the corrected trp→ipdC soil reversal. ₁