Validating Functional Markers Before Ecological Inference

Ecological conclusions based on gene markers are only as strong as the correspondence between the marker and the function it is intended to represent. The Subsurface Bacillota_B Specialization — What Distinguishes Deep-Clay Lineages from Soil Congeners? demonstrates this by revisiting a clay-project iron-reduction comparison after finding that its original marker KOs did not represent canonical iron-reduction functions. ₁ The Soil Metal Concentrations Drive Functional Gene Shifts refines this principle at environmental scale: associations between soil metals and functional gene categories can be extensive, but their ecological meaning remains uncertain when co-contamination, spatial matching, project effects, and effect sizes have not been resolved. ₂ The Functional Dark Matter — Experimentally Prioritized Novel Genetic Systems supports treating marker validation as a broader evidence-integration problem: among 57,011 genes classified as dark across 48 Fitness Browser organisms, reannotation, domain evidence, pathway context, synteny, co-fitness, and environmental associations often provide hypotheses but do not by themselves establish function. ₃

Core Principle

A functional marker should be validated at the sequence or domain level, checked for biological specificity, and tested for availability across comparison cohorts before its presence is converted into an ecological inference. ₁ This approach treats marker choice as an empirical validation problem rather than assuming that an annotation or KO label is an adequate ecological proxy. ₁ The soil-metal analysis supports extending this requirement to statistical and spatial validation: its 2,355 significant COG–metal associations across 51,748 samples were observational, and the report cautions that co-varying metals may make apparent single-metal markers reflect multi-metal contamination instead. ₂

The functional-dark-matter analysis refines the sequence-level requirement by showing that annotation status is itself unstable: Bakta v1.12.0 with DB v6.0 reclassified 33,105 of 39,532 pangenome-linked dark genes (83.7%) as not hypothetical, while 6,427 remained hypothetical in both Fitness Browser and Bakta but gained UniRef50 links for 79.4%, UniParc/UniRef100 links for 69.1%, and RefSeq links for 62.4%. ₃ Thus, a marker should be checked against current and independent annotations rather than accepted or rejected solely because one database labels it hypothetical.

The original clay-confined-subsurface analysis used K07811, K17324, and K17323 as iron-reduction markers, but these were identified as TMAO reductase, glycerol ABC ATP-binding, and glycerol ABC permease genes rather than canonical iron-reduction markers. ₁ This mismatch undermined the original interpretation of cohort differences and motivated a corrected detector. ₁ The clay-confined-subsurface report supports this diagnosis and explicitly records that the original iron-reduction analysis was corrected because these KOs were misidentified. ₄

The bacillota report records the thresholds at which the mismatched KOs were called: K07811 (TMAO reductase, threshold 1336.87), K17324 (glycerol ABC ATP-binding, threshold 363.93), and K17323 (glycerol ABC permease, threshold 318.27). ₁ It notes that KEGG has no canonical KO for the Geobacter omcS / Shewanella mtr operon multi-heme outer-surface cytochromes, so the clay project unknowingly substituted unrelated genes. It attributes the former “Mitzscherling rock-attached signature in shallow clay” framing to these mismatched gene IDs combined with the standard eggNOG-mapper threshold, which filters on sequence similarity rather than on KO biological identity. ₁ This is a direct demonstration, not an extrapolation: a passing similarity threshold does not certify that a KO represents the intended function. ₁

The clay report’s own correction notice supports this diagnosis. Its original IR-side comparison of anchor_deep (1/9 IR-positive) against anchor_shallow (15/30 IR-positive) was used to argue for a “Mitzscherling rock-attached IR-rich pattern” in shallow clay, and that comparison was driven by misidentified iron-reduction markers. ₄ The report states that KEGG has no canonical KO for the Geobacter omcS / Shewanella mtr operon multi-heme outer-surface cytochromes, so the original analysis unknowingly substituted unrelated genes. It also records that sister-project NB04 independently flags the K-number problem as systematic: KEGG KO assignment is unreliable for niche genes that lack canonical KOs, such as omcS, mtrCAB and resistance cassettes carried on mobile genetic elements (MGEs). KEGG-based marker mining for such genes can silently substitute unrelated functions. ₄ This extends the clay correction into a general risk for category schemas, discussed in ontology-and-category-schema-sensitivity. ₄

The soil-metal results provide a different but related failure mode: chromium and lead produced the strongest signals among nine metals, while transporters, including ABC and RND systems, and biosynthesis genes dominated the top hits; without effect-size and specificity checks, these categories cannot automatically be treated as metal-resistance markers. ₂ This supports separating marker identity, marker specificity, and ecological association before inferring mechanism. The functional-dark-matter analysis supports the same separation: its GapMind results identified 1,256 organism–pathway pairs co-occurring with strongly fitness-active dark genes, but these were organism-level co-occurrences rather than direct gene-to-enzyme assignments; supplementary domain matching produced 42,239 gene–pathway candidates across 3,186 dark genes, requiring EC matching, structure prediction, enzymology, or other validation. ₃

Corrected Multi-Heme Cytochrome Detector

The corrected workflow combined PFAM PF02085 and PF22678—PFAM is a protein-family domain resource—with a count of at least 4 CXXCH heme-binding motifs in gene-cluster protein sequences. A CXXCH motif is a cysteine-containing sequence pattern associated with c-type heme binding. ₁ A genome was classified as positive when any one of these signals was present. ₁ The new clay report characterizes this as a corrected triple-signal multi-heme cytochrome detector, retaining PF02085, PF22678, and the CXXCH threshold of ≥4. ₄ This refines the workflow by making the signal combination explicit without changing the previously reported positivity rule. ₄

The corrected detector identified multi-heme cytochrome signal in 5/9 anchor_deep genomes, or 55.6%, compared with 1/9, or 11.1%, under the original marker set. ₁ In anchor_shallow, it identified 12/30 genomes, or 40.0%, compared with 15/30, or 50.0%, under the original markers. ₁ In soil_baseline, it identified 61/149 genomes, or 40.9%, compared with 30/149, or 20.1%, under the original markers. ₁ The clay report independently reports the same corrected rates—55.6% for anchor_deep (n = 9), 40.0% for anchor_shallow (n = 30), and 40.9% for soil_baseline (n = 149)—and therefore supports these corrected counts. ₄ The central discoveries digest supports the same correction. It reports anchor_deep 56% (was 11%), anchor_shallow 40% (was 50%) and soil_baseline 41% (was 20%), with none of the corrected cohort comparisons significant (all Fisher p ≥ 0.46). ₅ The digest gives the mismatched KOs’ KOfam thresholds as 1336, 363 and 318, which are less precise than the bacillota report’s values. It also states that the SR-side H3 finding (K11180/K11181/K00394/K00395/K00958, p=4×10⁻¹² versus the Mitzscherling rock-attached null) remains intact. ₅

Pairwise Fisher exact tests—a test of association in contingency tables—found no significant corrected cohort differences: deep versus shallow had odds ratio 1.88 and p=0.46, deep versus baseline had odds ratio 1.80 and p=0.49, and shallow versus baseline had odds ratio 0.96 and p=1.0. ₁ The clay report supports this interpretation: all corrected cohort comparisons had Fisher p ≥ 0.46, and it states that the original shallow-greater-than-deep iron-reduction narrative is withdrawn. ₄ These results therefore contradict the original shallow-enrichment narrative and refine it to a comparison in which corrected multi-heme cytochrome content is statistically similar across the clay cohorts and soil baseline. ₁ Under the corrected detector, shallow (40%) is lower than deep (56%), but the difference is not statistically significant. ₁ The clay report likewise states that the original “shallow >> deep IR” pattern reverses, with deep slightly higher than shallow, and that none of the cohort comparisons is statistically significant after correction. ₄

Marker Availability and Functional Resolution

Within Bacillota_B, PF02085 had 4 hits in 4 clusters, PF00034 had 1 hit in 1 cluster, PF13442 had 1 hit in 1 cluster, PF22678 had 1 hit in 1 cluster, and PF14537 had 0 hits in 0 clusters. ₁ Because the multi-heme cytochrome PFAM signals were sparse, CXXCH motif counting carried most of the corrected iron-reduction signal. ₁ The report describes only ~6 clusters total across all four multi-heme-cytochrome PFAMs within Bacillota_B and confirms that PF14537 is silently absent. It explains that CXXCH motif counting on gene_cluster.faa_sequence is why the corrected IR rates (40–56%) are far higher than what PFAM-only detection would yield. ₁ The absence of PF14537 also shows that a missing domain call is weak negative evidence, a point developed in homology-search-negative-evidence. ₁ This supports combining domain and motif evidence while showing that a composite detector can depend disproportionately on one signal when marker availability is uneven. ₁

The functional-dark-matter analysis supports checking this unevenness across evidence types and cohorts before ecological interpretation. Its conservation-by-ignorance classification assigned 6.0% of dark genes to strong testable hypotheses, 52.5% to weak leads, and 41.5% to true knowledge gaps; 82/100 top candidates had high-confidence hypotheses supported by at least 3 evidence types, indicating that computational agreement can prioritize experiments but does not replace functional validation. ₃

The correction removes statistical support for the original shallow-enrichment interpretation, but it does not establish that all detected multi-heme cytochromes directly mediate iron reduction. ₁ The corrected analysis is therefore best treated as a robust Phase 1 correction of the multi-heme cytochrome comparison rather than a complete functional assignment of iron-reduction activity. ₁ The clay report supports this restraint through the corrected rates of 55.6%, 40.0%, and 40.9% for deep, shallow, and baseline cohorts, respectively, and their non-significant comparisons. ₄

The soil-metal analysis supports the same distinction between statistical detection and functional interpretation. Its copper-specific analysis found 116 COGs at FDR < 0.05—FDR means false discovery rate—among 7,566 samples matched to nearby KBase genomes within 10 km, with positive associations in cell-division and nucleotide-transport categories and negative associations in energy-production categories; the report treats these as observational suggestions of energetic trade-offs, not validated copper mechanisms. ₂ It also found biome-specific metal–COG relationships using PGLS, a comparative regression method that accounts for phylogenetic relationships, rather than a universal resistance programme. ₂

The clay analysis also shows why marker validation should be separated from enrichment discovery. It identified 547 significantly enriched eggNOG orthologous groups, but manual inspection found that keyword categorization had placed relevant anaerobic-respiration and electron-transfer functions in an “other” category, including COG1977 molybdopterin metabolism, a DsrE/DsrF/DsrH-like family, and a 4Fe–4S dicluster associated with 2-oxoglutarate:ferredoxin oxidoreductase. ₁ Manual reclassification suggested that the true anaerobic-respiration-related total was closer to 80–100 of the 547 enriched OGs, showing that false marker assignment and incomplete functional categorization can distort ecological interpretation. ₁ This supports linking marker validation to evidence-triangulation-for-functional-annotation and ec-less-reaction-annotation: sequence, domain, motif, cluster-context, and ecological-statistical evidence should be evaluated together rather than treated as interchangeable. ₁

Even a validated marker set has a resolution limit set by how its annotations are propagated. The clay report notes that eggNOG cluster-level annotations propagate within ≥90% AAI (average amino acid identity) clusters. Strain-level marker variants, such as a single non-functional dsrA in an otherwise-complete operon, may therefore be missed. ₄

Annotation-Source Disagreement for Named Markers

The central discoveries digest supports the page’s core principle with a case in which two annotation sources for the same named marker do not overlap at all. An NB01 calibration on 134K KBase Data Lakehouse pangenome genomes found that the two lanmodulin annotation sources disagree completely. ₅ bakta product='Lanmodulin' flags 62 genomes, all in α-Proteobacterial methylotroph clades (Beijerinckiaceae Methylobacterium, Acetobacteraceae BOG-930, Hyphomicrobiaceae). ₅ eggnog Preferred_name='lanM' flags 505 additional genomes, none of which are also bakta-validated. These are concentrated in unrelated gut Bacillota (Streptococcus pneumoniae 10, Blautia_A wexlerae 9, Enterococcus faecalis 8, Ruminococcus_B gnavus 8). ₅ The digest suggests, without establishing it, that the eggNOG Preferred_name for lanmodulin is a stale or seed-ortholog mislabeling rather than a true homology call. It recommends using bakta product='Lanmodulin' exclusively for these pangenome analyses and treating eggNOG Preferred_name='lanM' as zero-information. ₅ Further detail is on lanmodulin and lanmodulin-narrow-distribution. ₅ The lanthanide methylotrophy atlas supports this from its own marker table. Lanmodulin had 0 calls in both sources, 505 eggNOG-only and 62 bakta-only, and the atlas designates it “bakta only” because eggNOG Preferred_name='lanM' is noise. ₆

The digest extends this case to generic KO proxies. K02030 is the KEGG ortholog “amino acid ABC transporter substrate-binding protein,” a generic family. Filtering eggNOG by KEGG_ko LIKE '%K02030%' returns 46,369 hits across nearly every phylum, almost none of which are xoxJ. ₅ By contrast, bakta’s specific product call (“rare earth element methanol dehydrogenase accessory protein XoxJ”) matches just 61 genomes. ₅ The digest’s guidance is to use bakta product or eggNOG Preferred_name='xoxJ', which is more specific despite the generic KO, and never to use K02030 as a xoxJ proxy. ₅ This parallels the clay KO substitution: in both cases a KO label stood in for a functional marker without representing the intended gene (see xoxj). _5,4 The lanthanide methylotrophy atlas’s marker table gives 41 shared, 46,369 eggNOG-only and 20 bakta-only xoxJ calls. It designates xoxJ “bakta only” because KO K02030 is non-specific. ₆ This supports the digest’s rejection of K02030 as a proxy. The atlas’s bakta-only rule is, however, stricter than the digest’s guidance, which also allows eggNOG Preferred_name='xoxJ'. _6,5

For xoxF the disagreement is partial rather than complete. Of 5,092 genomes carrying any xoxF marker, only 418 (8%) hit both eggNOG K00114 and bakta Lanthanide-dependent methanol dehydrogenase-family products. ₅ The eggNOG-only set (3,272 genomes) appears to include bakta entries labeled Putative dehydrogenase XoxF that the report’s regex did not match, plus other generic dehydrogenases that hit K00114. The bakta-only set (1,402) includes lanthanide-dependent ethanol dehydrogenase variants that the KEGG KO filter excludes. ₅ The digest therefore recommends reporting xoxF results both ways, with the eggNOG-K00114 set as primary and the either-source set as inclusive. It also recommends disclosing the gap and never claiming a single xoxF count without specifying its source. ₅ This refines evidence-triangulation-for-functional-annotation: disagreement between annotation sources is itself a reportable quantity, not noise to be silently resolved (see xoxf). ₅

The lanthanide methylotrophy atlas’s own calibration table extends this source-by-source reporting to three further markers. For xoxF it records 418 shared, 3,272 eggNOG-only and 1,402 bakta-only calls, designating eggNOG K00114 primary with bakta as union. ₆ For mxaf it records 4 shared, 191 eggNOG-only and 8 bakta-only calls and favours eggNOG K14028. ₆ For the pqqA–E markers — the biosynthesis genes for PQQ, pyrroloquinoline quinone, the cofactor of methanol dehydrogenases — it records 50–90K bakta-only calls, characterises bakta as over-calling and treats eggNOG as primary. ₆ The report’s PQQ-supply interpretation nevertheless recommends that pangenome work on PQQ use the union of eggNOG and bakta sources, so that products missed by either annotation are retained — a different recommendation, in a different context, from the calibration table’s eggNOG-primary caution about bakta over-calling pqqA–E. ₆

Across these markers the atlas concludes that for specialty markers like Lanmodulin and XoxJ, bakta’s curated product field outperforms eggNOG’s Preferred_name. ₆ It states that KO K02030 is non-specific for xoxJ and that bakta product is the trustworthy lanmodulin source, making bakta product annotations its calibrated source for lanmodulin calls. ₆ Because eggNOG and bakta calls disagree substantially for several markers (xoxJ, lanmodulin, PQQ biosynthesis), its headline statistics use the calibrated source-of-truth per marker, with secondary union-of-sources analyses reported alongside. ₆ This refines the page’s rule: the response to annotation-source disagreement is a documented per-marker choice plus the alternative, not a single unattributed count. ₆

The atlas’s lanmodulin numbers sharpen the digest’s case. Its 505 eggNOG Preferred_name='lanM' calls are judged false positives concentrated in unrelated gut Bacillota — Streptococcus pneumoniae (10), Blautia_A wexlerae (9), Enterococcus faecalis (8), Ruminococcus_B gnavus (8), Streptococcus pyogenes (7) and similar — none of them lanthanide users, while bakta product='Lanmodulin' matches 62 genomes, all in the canonical α-Proteobacterial methylotroph clades; the atlas therefore directs pangenome lanmodulin work to bakta product exclusively. ₆ The 505 false positives are concentrated in a handful of Streptococcus / Blautia / Enterococcus species, so this annotation failure is lineage-clustered rather than diffuse. ₆

Figure lanM_preferred_name_false_positives.png shows the top 15 families by eggNOG-only lanM calls (false positives) versus bakta-validated calls. ₆

A calibrated marker source still leaves a sequence-level gap. The atlas states that sequence-level evidence is out of scope: xoxF and PQQ presence/absence is at the gene-call level, and it does not screen for pseudogenes, truncated ORFs, or assembly fragmentation. ₆ A marker call of this kind therefore rests on the gene call alone, with the sequence-integrity checks that could overturn it declared out of scope rather than performed; the related problem of unchecked absence is treated in homology-search-negative-evidence. ₆

Resistance-gene catalogues show the same specificity problem. In the AMR fitness-cost analysis, 86% of AMR genes were Tier 2, meaning keyword-matched from bakta annotations, and may include some non-AMR genes such as general efflux transporters. The report says a Tier 1 sensitivity analysis gave consistent results, which mitigates this concern. ₇ The AMR pangenome atlas cautions that AMRFinderPlus scope includes stress-response genes (mercury, arsenic) alongside classical antibiotic resistance, so its 83K hits are not all antibiotic resistance per se. It also cautions that some singleton AMR clusters may reflect annotation artifacts rather than true species-specific resistance genes. ₈ Together these support separating marker identity from functional category before resistance-gene counts are interpreted ecologically. _7,8

Plant-association marker analyses illustrate the gap between marker presence and phenotype. The digest reports a classification of 25,660 species into beneficial, pathogenic, dual-nature and neutral cohorts. Most plant-associated bacterial species carried both plant growth-promoting (PGP) and pathogenic marker genes simultaneously, and marker co-occurrence alone does not establish a beneficial or pathogenic phenotype. ₅ ACC deaminase (acds) shows the strongest compartment-specific enrichment of any single marker gene, with an odds ratio of 69.3 in root-associated species. This estimate is not described as phylogenetically controlled. ₅ After phylogenetic control, 50 novel eggNOG ortholog groups distinguish plant-associated from non-plant species, with cog3569 as the top hit (OR=6.01 phylo-controlled). Plant-interaction functions for these groups remain candidate interpretations. ₅ Single-marker enrichment that has not been phylogenetically controlled should therefore be checked against lineage structure before it is read as a functional marker, as discussed in phylogenetic-confounding-of-pangenome-associations. ₅

Genomic proxies for phenotypes need the same validation as gene markers. The CF formulation report notes that the genome of Pseudomonas aeruginosa (PA) (6.58 Mb, 6,177 CDS) is 2.5–3× larger than those of its commensal species (2.1–2.7 Mb, 2,200–3,000 CDS). Measured growth curves nevertheless show PA14 outgrowing most commensals on its preferred amino acid substrates. ₉ Codon usage bias (CUB) of ribosomal proteins varied across species, but cross-species CUB comparisons were confounded by GC content variation (31–73%), ranging from 31% for G. sanguinis to 73% for M. luteus. Organisms with extreme GC composition have inflated CUB scores regardless of growth optimization. ₉ The report concludes that CUB cannot reliably distinguish growth-rate differences among its species and treats the laboratory growth data as the definitive measure, or ground truth. ₉ This supports the page’s rule that a computational proxy should yield to direct measurement when its confounders are unresolved; see capability-versus-kinetic-predictability. ₉

Broad Domain Anchors Versus Specific Markers

Whether a domain marker is specific enough to stand alone is itself testable. The phage-defense arsenal project tested 19 markers across 7 systems against both interproscan_domains and bakta_pfam_domains: broad-Pfam markers (Retron RVT_1, DISARM PLD/SNF2, Gabija UvrD) required anchor-plus-context filtering at species level, while specific markers (CRISPR Cas1, CBASS SAVED, BREX PglZ, Gabija OLD_TOPRIM_C) can be used as single-marker calls. ₁₀ This refines the page’s validation sequence by making marker specificity a pre-test with two possible outcomes rather than an assumption. ₁₀

Detection method alone can move a prevalence estimate by tens of percentage points. eggNOG description-based CRISPR detection dramatically over-counts: 96% species prevalence via eggNOG description matches versus ~55% using the specific Cas1 Pfam (PF01867) ₁₀, an over-count of about 40 percentage points that the report records as a methodological caveat for future pangenome CRISPR comparisons. ₁₀ The 96% figure is therefore an upper bound, and the project’s syndrome analysis uses the combined presence set (eggNOG OR Pfam), which favours recall, so CRISPR-Cas syndromes may be slightly underestimated in specificity but not in significance. ₁₀

disarm shows how a nonspecific anchor propagates into a pangenome structural claim. DISARM is the exception among defense systems in being near-baseline in core fraction, which the report calls almost certainly a detection artefact: the DrmB SNF2 helicase Pfam (PF00176) used as the DISARM anchor is a widespread housekeeping-helicase domain, and its inclusion inflates the DISARM cluster set with non-mobile SNF2 helicases. ₁₀ The report therefore treats the DISARM accessory-enrichment result as unreliable and biased toward core, and proposes a DISARM-specific refinement using PADLOC’s MacSyFinder-style rules, which combine hidden Markov models (HMMs — probabilistic profiles of a sequence family) with genomic context. ₁₀ It argues that the arms-race and syndrome results for DISARM survive this caveat because they rely on species-level presence/absence, not per-cluster core/aux calls. ₁₀ This refines the page’s principle: a specificity failure can invalidate one analysis layer while leaving a coarser layer intact, so the affected layer should be named rather than the whole result withdrawn. ₁₀

retron detection in the same project uses only RVT_1 (PF00078), a broad reverse-transcriptase Pfam, and its stringent variant departs from the planned Millman-2020-effector-Pfam design by requiring the species to carry ≥1 other narrow defense system — a defense-context filter, not a retron-specificity filter. ₁₀ Because narrow defense systems are near-universal across species that carry RVT_1, the candidate → stringent filter is essentially a no-op: 15,109 → 15,098, with 11 species dropped. ₁₀ Results should be interpreted as reverse-transcriptase candidates in defense-syndrome context, not characterized retron systems. ₁₀

A deliberately broad classifier can saturate its own metric. The eggNOG-description prophage classifier picks up integrase, holin, endolysin, CI-like repressor and tail proteins as candidates, terms that also match many non-phage bacterial genes such as host integrases and cell-wall lysins, so n_prophage_modules saturates at 7 modules for 35% of species and is not a specific prophage count. ₁₀ For the arms-race regression the project therefore uses n_prophage_clusters, an unbounded continuous count of matched clusters, as the primary predictor, with n_prophage_modules a coarse categorical secondary. ₁₀ This supports choosing the less saturating marker-derived variable once a panel’s breadth is known. ₁₀

The central pitfalls digest records the mirror-image failure, in which canonical gene names under-detect rather than over-detect. In GenomeDepot-based databases (protect_genomedepot, phagefoundry_paeruginosa_genome_browser), pslA-pslO genes are rarely annotated with canonical gene names — only PAO1’s reference genome has them — and the pangenome bakta_annotations has better coverage but pslA/pslG/pslK are still sparse, so the digest recommends KEGG orthologs (K20997–K21005) via eggnog_mapper_annotations for reliable psl detection at ~95% prevalence versus <1% by gene name alone. ₁₁ Its marker checklist also requires that Pfam accessions be verified against InterPro/UniProt and not assumed from a domain name ₁₁, and that the eggNOG PFAMs column be searched by domain NAME, not accession. ₁₁

Pathway markers inherit the metabolic promiscuity of their genes. The PHB (polyhydroxybutyrate) granule analysis uses phaC — PHA (polyhydroxyalkanoate) synthase, the committed step for PHB biosynthesis — as its pathway marker. ₁₂ It classifies species by pathway completeness: complete (phaC + phaA/phaB, 6,005 species), synthase-only (phaC without precursor enzymes, 62), precursors-only (phaA/phaB without phaC, 12,869), accessory-only (phaP/phaZ/phaR without synthase, 555) and absent (8,199), and notes that the 5/6 detection rate for PHB KOs, with phaR absent, suggests K18080 may be underrepresented in eggNOG annotations. ₁₂ Because phaA and phaB participate in general metabolism beyond PHB, the report judges that the precursors_only category (46.5% of species) likely overestimates partial PHB capability. ₁₂ This supports assessing marker specificity separately from marker presence; see phb-granule-ecology. ₁₂

Marker presence is also not demonstrated function. The plant growth-promoting pangenome analysis states that a gene cluster annotated as pqqC is not necessarily functional, because truncations, frameshifts and pseudogenization are not filtered. ₁₃ This supports the sequence-level gap recorded above for gene-call-level marker counts. _13,6

Pipeline-Level Validation of Annotation Calls

An annotation pipeline’s internal calibration is part of marker validity. The MicrobeAtlas metal-ecology analysis annotates metal AMR genes using amrfinderplus HMM-based detection against the NCBI Bacterial AMR Reference Gene Database, with curated models built from experimentally validated reference sequences such as merA/merB for mercury, arsABCDR for arsenate and copABCD for copper; HMM gathering thresholds are set during model calibration to balance sensitivity and specificity, and no additional identity threshold is applied because HMM scores are not identity values. ₁₄

Its stated residual risk is cross-reactivity: cross-reactive metalloproteases or ABC transporters with structural similarity to metal-resistance proteins could be misclassified, which AMRFinderPlus’s taxon-specific cutoffs reduce rather than eliminate. ₁₄ No manual validation of a subset was performed in that study; spot checks of highly prevalent gene families (arsD, merP) show biologically expected distributions — high in contaminated-environment genera, lower in strict aerobes — which the report offers only as informal support. ₁₄

The project names the missing test instead of claiming it: random sampling of 100 annotated metal AMR gene clusters for manual BLAST verification against reference sequences is proposed to quantify the HMM-based false-positive rate for each metal type category, and is not reported as performed. ₁₄ Its own robustness question is answered only in part — all 7 single-metal exclusions remained positive, so no one metal is the driver, but formal AMRFinderPlus validation was not performed and only spot checks exist. ₁₄ This extends the AMRFinderPlus scope and taxonomic-composition caveats recorded above: an unquantified false-positive rate bounds how far a metal-resistance marker count can be read mechanistically. _14,8,15

Panel Breadth and Phenotype Labels

Panel breadth sets what a marker-based label can mean. The plant-microbiome ecotype analysis reports that the majority of plant-associated species, 65–85% per compartment, carry both plant growth-promoting (PGP) and pathogenic marker genes simultaneously. ₁₆ This supports the digest’s conclusion that marker co-occurrence alone does not establish a beneficial or pathogenic phenotype. _16,5

Two classification schemes over the same markers give different dual-nature rates. Under binary marker presence (NB02), any species carrying ≥1 PGP and ≥1 pathogenic marker is dual-nature, giving 15,474/25,660 species (60.3%), which the report attributes to the broad prevalence of general bacterial functions such as flagella, chemotaxis and T6SS (type VI secretion system) that span both categories. ₁₆ A weighted composite scoring framework (NB07: 40% PGP markers, 20% core fraction, 15% complementarity, 10% metabolic breadth, 15% pathogen penalty), with adaptive thresholds set at the median of non-zero scores, gives neutral 38.0%, pathogenic 29.5%, dual-nature 25.2% and beneficial 7.3%; the passage gives no NB07 denominator. ₁₆

The report reads the 35 percentage point difference between schemes as the distinction between carrying both marker types, nearly universal among bacteria given ubiquitous functions like flagella and T6SS, and scoring highly on both axes when weighted by genomic architecture and metabolic context; the binary classification is more inclusive, the composite classification more discriminating but also more assumption-laden. ₁₆ This refines ontology-and-category-schema-sensitivity: one marker table supports different cohort labels depending on the scoring rule, so the rule must be reported with the rate. ₁₆

The project’s first panel was broad enough to label non-plant enteric bacteria as dual-nature. Phase 1’s 91-marker panel included ubiquitous bacterial functions (flagella, chemotaxis, T6SS, biofilm, quorum sensing, T2SS) that classified even Escherichia and Salmonella as 100% dual-nature; Phase 2 refined this to 17 plant-specific markers (9 PGP + 8 pathogenic) and applied KEGG module completeness gating for multi-gene systems. ₁₆ Requiring ≥3 module genes for T3SS (type III secretion system, M00332) reduced T3SS-positive species from 10,300 to 1,661; the report also describes this as −8,855 false positives and an 86% reduction, and its stated counts and subtraction do not agree, so these figures should be read as reported rather than recomputed. ₁₆ Requiring ≥2 module genes for nitrogen fixation (M00175) reduced nif-positive species by 346, a 14% reduction ₁₆, while the T4SS gate (M00333) had minimal impact because most hits were already multi-gene. ₁₆

Under the stricter panel the dual-nature rate among plant-associated species is 78.7% (878/1,115 species), which the report attributes to genuinely co-occurring markers — phosphate solubilization (57.9% prevalence), CWDE (cell-wall-degrading enzyme) cellulase (63%), phenazine (63%), T4SS (45.8%) — and contrasts with the earlier 60.3%. ₁₆ That contrast does not establish an increase caused by panel refinement, because the 60.3% figure is 15,474/25,660 species overall while the 78.7% figure is restricted to 1,115 plant-associated species; the populations differ. ₁₆

A curated ground-truth panel then showed that the categorical label carries no species-level information. A canonical 18-species panel of model organisms with experimentally confirmed phenotypes — 7 known beneficial (P. simiae WCS417, P. protegens CHA0, B. velezensis FZB42, A. brasilense Sp245, R. leguminosarum, P. phytofirmans PsJN, S. meliloti), 7 known pathogenic (P. syringae, R. solanacearum, X. campestris, A. tumefaciens, E. amylovora, P. carotovorum, C. michiganensis) and 4 neutral non-plant controls — was tested in place of an earlier 92.7% genus-level agreement figure, which was tautological because cohorts are defined from marker presence. ₁₆ All 14 of the beneficial plus pathogenic ground-truth species were assigned to the dual-nature class, so the categorical cohort label cannot discriminate beneficial from pathogenic ground truth, and all 4 neutral controls were also misclassified (E. coli K-12 as dual-nature, M. tuberculosis as dual-nature, D. radiodurans as beneficial, T. thermophilus as pathogenic), which the report attributes to the generality of the marker set rather than to a failure to detect plant adaptation, since flagella, secretion, chemotaxis and biofilm overlap beneficial and pathogenic markers in all bacteria. ₁₆ The discriminative signal comes instead from a continuous metric: the pathogenicity ratio n_pathogen / (n_pgp + n_pathogen) is lower in known beneficial species (median 0.50) than in known pathogenic species (median 0.60), Mann-Whitney U = 9, p = 0.027 on N = 7 vs 7 — a small effect at a small sample size, but statistically consistent. ₁₆

An earlier Phase 2b draft’s “77.8% relaxed accuracy” — arithmetically 14/18, counting dual-nature as correct for both beneficial and pathogenic ground truth — is explicitly withdrawn as a discriminative metric, because the rule cannot fail for any species the classifier assigns to dual-nature and therefore contains no discriminative information. ₁₆ The report’s practical conclusion is that the refined cohort labels (beneficial / pathogenic / dual-nature / neutral) are a coarse screen, not a species-level classifier, with the continuous pathogenicity ratio doing the discriminating ₁₆, and that a stronger follow-up validation would include near-pure beneficial organisms (Methylobacterium extorquens) and near-pure pathogens (Xylella fastidiosa) with known minimal PGP/pathogenicity machinery. ₁₆

The named dual-nature genera are pseudomonas e (358 species; PGP: HCN, phenazine, phosphate solubilization; pathogenic: T2SS, T3SS, T6SS, effectors), bradyrhizobium (73 species; PGP: nitrogen fixation, ACC deaminase; pathogenic: T3SS, T4SS, CWDE) and Streptomyces (261 species; PGP: phenazine, phosphate solubilization; pathogenic: CWDE, T3SS). ₁₆

Within-genus label permutation narrows the species-level marker list much further: only 3 of 15 testable markers survive — nitrogen fixation (permutation p < 0.001), ACC deaminase (permutation p < 0.001) and T3SS (permutation p = 0.005). ₁₆ The test has its own resolution limits: n = 200 permutations is under-resolved for p ≤ 0.005, so T3SS is marginal, and no multiple-testing correction was applied across 15 markers, with Bonferroni α = 0.0033 noted as demoting T3SS while nitrogen fixation and ACC deaminase remain significant. ₁₆ The report’s reading is that most marker-level plant-association signals in the project are genus-scale rather than species-scale; this does not invalidate its H1/H2 compartment and architecture findings, which already operate at genus, family and phylum scale, but it does mean the dual-nature rate and the compartment-specific marker enrichments should be interpreted as genus-level genomic backgrounds of plant-associated lineages rather than as species-level functional evidence of plant adaptation. ₁₆ This supports phylogenetic-confounding-of-pangenome-associations, where the acds root-association odds ratio of 69.3 is likewise flagged as not phylogenetically controlled. _16,5

The project’s marker panel is itself reported with an internal inconsistency worth recording. A curated set of 91 marker genes — described as 39 plant growth-promoting (PGP), 44 pathogenic and 9 colonization markers — was searched across bakta annotations, Pfam domains, KEGG KOs and product descriptions, yielding 588,098 marker gene clusters across 25,660 species. ₁₆ The stated category counts do not add up to the stated panel size of 91, so the panel’s composition should be read as reported rather than recomputed. ₁₆

The most prevalent functional categories in that search were T6SS products (80,324 clusters), chemotaxis (69,986), quorum sensing (57,201), T3SS products (52,247) and T4SS (48,145); among PGP functions, phosphate solubilization (23,876), phenazine (12,727), biofilm (7,512) and nitrogen fixation (6,139) were the most common. ₁₆ The panel’s most abundant hits are therefore general bacterial systems rather than plant-specific functions, which supports the page’s requirement that marker specificity be established before counts are converted into ecology. ₁₆

The refinement step is documented with its criterion rather than asserted. The Phase 1 91-marker panel was reduced to 17 plant-specific markers by removing 6 ubiquitous categories (flagella, chemotaxis, T6SS, biofilm, quorum sensing, T2SS) that had only 1.2–2.4× enrichment in plant versus non-plant species, and KEGG module completeness gating was applied to multi-gene systems: M00175 (nitrogen fixation, threshold ≥2 genes), M00332 (T3SS, ≥3 genes) and M00333 (T4SS, ≥3 genes). ₁₆ This refines the panel-refinement record above by supplying the removal criterion and the per-module thresholds. ₁₆

Under the refined panel the control organisms behave closer to expectation: beneficial-plus-dual rates were Rhizobium 97%, Bradyrhizobium 100%, Azospirillum 100% and Mesorhizobium 99%, while PGP-only rates among negative controls improved to Escherichia 25% (down from ubiquitous dual-nature classification in Phase 1), Salmonella 0%, Staphylococcus 18% and Mycobacterium 6%. ₁₆

The broad panel’s negative controls nevertheless show how far marker breadth inflates a phenotype label. Non-plant genera Escherichia, Salmonella and Clostridioides were each classified 100% dual-nature and Staphylococcus 98%, because flagella, chemotaxis, biofilm, quorum sensing and secretion systems are ubiquitous across bacteria rather than plant-specific. ₁₆ Combined with the genome-size confound, the report judges that the NB02 dual-nature rate of 60.3% substantially overestimates genuine plant-adapted duality, and that the NB07 composite rate of 25.2% only partially addresses this because it remains dependent on arbitrary scoring weights. ₁₆

Within the dual-nature class the drivers are again general functions: the most prevalent pathogenic markers were T6SS products (64%), T2SS (55%), chemotaxis (52%) and T3SS products (50%), and the most prevalent PGP markers were quorum sensing (49%), flagella (40%) and phenazine (34%), so general motility functions contribute substantially to the classification. ₁₆

Moving markers between categories changes the label without eliminating the pattern. Reclassifying T3SS, T6SS and T2SS from pathogenic to colonization markers changed 16.4% of dual-nature species to PGP-only, yet plant-associated species remained 86.0% dual-nature under the revised classification. ₁₆ The report separately logs T3SS dual interpretation as documented and context-dependent rather than resolved. ₁₆ This refines ontology-and-category-schema-sensitivity: the category a marker is assigned to is an analytic choice with a measurable effect on cohort rates, so the assignment must be reported with the rate. ₁₆

The underlying biology explains why that assignment is contestable. The report states that classification as beneficial or pathogenic from genomic markers alone is insufficient, because the same T3SS that delivers pathogenic effectors in Pseudomonas syringae facilitates beneficial nodulation signalling in Rhizobium, with host genotype, environmental conditions and community composition determining the outcome. ₁₆

The breadth of co-carriage is reported at several scales. An initial estimate found that 60–85% of plant-associated bacteria carry both PGP and pathogenic markers ₁₆, a figure stated elsewhere in the same project as 65–85% per compartment; the two ranges are preserved as reported and not reconciled here. ₁₆ The 30 genus dossiers likewise showed that plant-associated genera nearly universally carry both marker types. ₁₆ The Phase 2 refined 17-marker panel with KEGG module gating confirmed the pattern persists at 78.7% after removing ubiquitous bacterial functions, which remains a genomic-marker classification rather than a measured phenotype. ₁₆

Co-occurrence of the two marker sets does not establish simultaneous expression. The report states that transcriptomic or experimental validation is needed to determine whether co-carriage represents genuine lifestyle flexibility or annotation artifacts — for example T6SS serving inter-bacterial competition rather than pathogenicity. ₁₆ It also states plainly that the dual-nature classification is based on gene presence. ₁₆

Continuous genus-level summaries carry more information than the cohort label, and their definitions are scoped. The report’s PGP ratio is sum(n_pgp) / sum(n_pgp + n_pathogen) across plant-associated species in a genus, with values below 0.5 indicating a pathogen-marker-rich genus such as Xanthomonas at 0.28 or Chryseobacterium at 0.23. ₁₆ Its cluster-robust markers are those surviving the cluster-robust model at q < 0.05 Benjamini-Hochberg false discovery rate, and the MGnify biosynthetic-gene-cluster counts in the same dossier table are soil-biome counts only, so a dash means the genus is absent from MGnify soil rather than lacking clusters. ₁₆

The project grades its own mechanism hypotheses instead of asserting them: nifHDK nitrogen fixation, siderophore iron acquisition, T3SS effector injection and cell-wall degradation are rated High confidence, while ACC deaminase stress relief, DAPG/phenazine biocontrol antimicrobials, context-dependent T6SS and PGP-plus-secretion-system co-occurrence are rated Medium, and all are presented as proposed mechanisms. ₁₆

A model-organism check supports the continuous axes rather than the labels: B. subtilis, R. leguminosarum, B. japonicum and S. meliloti cluster in the expected quadrants of the PGP-versus-pathogen scatterplot. ₁₆

Genome size confounds any raw marker count. Genome size, measured as gene cluster count, correlates moderately with total marker count (r=0.44, p<1e-300), so larger genomes accumulate more markers by chance. ₁₆ Cross-tabulation showed that genome-size normalization shifts 33% of dual-nature species to neutral, an effect the report calls meaningful but not sufficient to eliminate the dual-nature pattern. ₁₆

Measured directly on refined cohort data, the dual-nature rate among plant-associated species is 54% in the bottom genome-size quartile versus 87% in the top three quartiles. ₁₆ Read together with the normalization result, the report concludes that the headline 78.7% dual-nature rate masks a real genome-size gradient, and that the gradient is observational rather than model-dependent. ₁₆

The project’s novel ortholog groups survive genome-size control, but with two failures that must not be counted as successes. After phylum and log₁₀(genome_size) control, 48 of 50 ortholog groups retained a positive plant coefficient at BH-FDR q<0.05, with top-10 controlled odds ratios of 4.4–6.6, while COG2351 and COG5343 failed with singular design matrices rather than biological rejection. ₁₆ An earlier recap stating that all 50 ortholog groups retained adjusted positive coefficients is superseded by this canonical per-species analysis and should not be read as 50 successful adjusted fits. ₁₆

Phylogenetic control sorts the panel into tiers rather than validating it as a whole. A cluster-robust generalized linear model (GLM — a regression with errors clustered, here by genus) found 8 of 14 markers surviving at q<0.05 with Wald confidence intervals excluding zero, identified iaa_biosynthesis as one earlier L1-bootstrap false positive, and combined with the within-genus shuffle result of 3 of 15 to give a three-tier verdict: 3 species-level, 5 cassette-level and 6 not robust. ₁₆

The report’s Tier 1, which it characterizes as robust at species level under the tests it ran, comprises nitrogen fixation, ACC deaminase and T3SS — large coefficients, Wald confidence intervals not crossing zero, and survival of within-genus permutation — so on that reading specific plant-associated species in a genus carry the function while non-plant sister species reliably do not. ₁₆ The T3SS placement is marginal under the project’s own resolution caveats, a disagreement recorded under Tensions below rather than resolved here. ₁₆ Tier 2, robust under the cluster-robust GLM but failing the within-genus shuffle, comprises phenazine, CWDE cellulase, CWDE pectinase, phosphate solubilization and effector markers; the report reads these as cassette-level signals that tend to be acquired when a genus is plant-associated but are not species-specific within it, ecologically meaningful yet not species-level distinguishers. ₁₆ Tier 3, not significant in any test, comprises DAPG biocontrol, T4SS, hydrogen cyanide, IAA biosynthesis, siderophore and acetoin/butanediol, including signals previously suggested by L1 bootstrap. ₁₆

Stated as a null result, 12 markers — phenazine, IAA biosynthesis, CWDE cellulase, CWDE pectinase, effector, DAPG biocontrol, hydrogen cyanide, siderophore, acetoin/butanediol, phosphate solubilization, T4SS and phytotoxin — fail the within-genus permutation test, so none of them can be read as species-level plant-association markers. ₁₆ Only three — the nif nitrogen-fixation operon, the ACC deaminase enzyme and the hrp/hrc T3SS structural cluster — are carried by specific plant-associated species within genera where non-plant species do not reliably have them. ₁₆

The report therefore bounds what its headline numbers license: the 78.7% dual-nature rate and the compartment-specific marker enrichments should be interpreted as genus-level genomic backgrounds associated with plant-associated lineages, not as species-level functional evidence of plant adaptation. ₁₆ This supports phylogenetic-confounding-of-pangenome-associations and refines the page’s validation sequence by adding an explicit step: name the taxonomic scale at which a validated marker claim holds. ₁₆

A panel’s validity also depends on which annotation table it is queried against and on how that table stores its identifiers, and the project records two distinct failures. The first is a query-format failure: the Phase 1 NB02 query WHERE pfam_id IN ('PF00771', …) returned zero rows because bakta_pfam_domains stores versioned IDs (PF00771.22) rather than bare accessions, and re-running with LIKE 'PF00771%' patterns recovered the hits. ₁₆ The second, separate from that format problem, is an annotation-coverage gap: an audit found that 12 of the project’s 22 marker Pfams are silently missing from bakta_pfam_domains, mostly secretion-system T3SS/T4SS/T6SS domains, while InterProScan carried all 22 and the InterProScan-based pipeline was unaffected. ₁₆ A missing marker call can therefore be a property of the query syntax or of the annotation source rather than of the genome, which supports homology-search-negative-evidence. ₁₆

Wrong-Family and Clan-Level Marker Confusion

A marker can sit in the right superfamily and still be the wrong family. In the snipe-defense-system analysis, the SNIPE nuclease domain is PF13455 (Mug113, “Meiotically up-regulated gene 113”), which belongs to the GIY-YIG clan CL0418 but is not the canonical GIY-YIG family. ₁₇ Canonical GIY-YIG is pfam PF01541 / IPR000305, listed with 68,148 proteins covering restriction enzymes and DNA repair, and is not the SNIPE nuclease. ₁₇

InterPro and UniProt analysis of the Escherichia coli SNIPE protein (A0A0A1A5Z2) assigns the nuclease to pf13455 mug113 rather than PF01541; the originating paper describes the nuclease as “GIY-YIG” at the superfamily level, but at Pfam family level the two are distinct entries. ₁₇ The protein is 558 aa, with DUF4041 / SNIPE (PF13250, DNA/TMP binding) at residues 232–333 and Mug113 (nuclease) at residues 443–520. ₁₇

The consequence is a null result that is a family distinction, not a detection failure. Searching for PF01541 co-occurrence with DUF4041 yields zero hits, which the report attributes to the two being genuinely different Pfam families rather than to annotation artifacts. ₁₇ Zero gene clusters carried both DUF4041 and canonical GIY-YIG PF01541 in their eggnog Pfam annotations. ₁₇ The 74,686 PF01541 clusters the project found are therefore not SNIPE nucleases but restriction enzymes, DNA repair enzymes and other GIY-YIG family members. ₁₇ This refines homology-search-negative-evidence: an absent co-occurrence can be the correct answer once the marker’s family identity is checked. ₁₇

The project’s marker definition follows that diagnosis. SNIPE homologues were identified by the DUF4041 domain (pf13250 duf4041, PF13250 / IPR025280) in eggNOG annotations, with PF01541 queried independently and then rejected as the wrong Pfam family. ₁₇ InterPro and UniProt record 1,612 SNIPE proteins across 1,373 species under IPR025280, saved as an external anchor for the marker. ₁₇

The chosen marker is incomplete in the opposite direction. Among the 4,572 DUF4041 clusters, 54 carry the description “Meiotically up-regulated gene 113”, consistent with full-length SNIPE proteins annotated with both domains. ₁₇ Because only 54 of 4,572 clusters show Mug113 co-annotation, the report states that eggNOG Pfam annotations may miss divergent SNIPE homologues, suggesting that many SNIPE nuclease domains go undetected. ₁₇ The co-annotation rate is consistent with under-detection but does not by itself establish that divergent homologues, rather than some other cause, explain the missing annotations. ₁₇

The report also quantifies its false-positive exposure instead of asserting specificity. The top two descriptions among DUF4041 clusters — T5orf172 (2,109), the gene-family name for DUF4041, and “Domain of unknown function (DUF4041)” (1,566) — account for 3,675 of 4,572 clusters (80.4%) and are treated as bona fide SNIPE-family proteins, while the remaining ~20% (897 clusters) carry primary annotations such as histidine kinase (319) and seryl-tRNA aminoacylation (238), where DUF4041 was likely detected as a secondary or minor domain in multi-domain proteins whose primary function is unrelated to phage defense and which the report reads as annotation noise. ₁₇ DUF4041 (T5orf172) is strongly associated with SNIPE, but some hits may represent non-SNIPE proteins. ₁₇

The cluster annotations are at least consistent with the system’s biology. Across the DUF4041-containing clusters the COG category distribution was D (cell cycle, 1,533), T (signal transduction, 1,118), J (translation, 1,066) and M (cell wall/membrane, 834), with the membrane-associated COG M category aligning with SNIPE’s inner-membrane localization; these counts are reported for the DUF4041 cluster set as a whole, and no COG counts are reported for any description-filtered subset. ₁₇ Separately, a stricter filter requiring a T5orf172 or DUF4041 description would retain ~80% of clusters as a high-confidence SNIPE set, and the report states that the 86.7% accessory rate and the taxonomic distribution patterns are not substantially affected by that subset, since the non-SNIPE annotations are spread across the same phyla. ₁₇ This refines the page’s rule in the same way the DISARM case does: a specificity caveat should be bounded to the analysis layers it actually moves. ₁₇

Within one report, the label attached to an accession need not be stable: its query footnote labels PF00358 as PTS_EIIC and PF02378 as PTS_EIID, while its own table labels PF00358 as PTS_EIIA_1 and PF02378 as PTS_EIIC. ₁₇ The conflict is recorded rather than reconciled, and it supports ontology-and-category-schema-sensitivity together with the central pitfalls requirement that Pfam accessions be verified against InterPro/UniProt rather than assumed from a domain name. _17,11

Free-text description matching produces the same class of error for respiratory markers. The respiratory-chain analysis notes that its NDH-2 (type II NADH dehydrogenase) identification uses text matching on gene descriptions, which likely produces false positives for organisms with incompletely annotated Complex I subunits — the cited case being pseudo3_N2E3 with 8 apparent “NDH-2” hits that are probably a Complex I operon — and states that KO-based identification (K03885) would be more reliable. ₁₈ This supports the page’s preference for accession- or motif-level markers over description matches, and bears on negative calls too, since a text-matched panel can both over- and under-count (homology-search-negative-evidence). ₁₈

Annotation-Proxy Panels for Prophages

Prophage detection in this corpus is annotation-proxy based, and both projects that do it disclose the consequence. The prophage-ecology analysis uses eggnog functional annotations as proxies for prophage genes rather than dedicated prophage detection tools such as genomad or VIBRANT, which likely inflates prophage prevalence by including domesticated remnants and bacterial homologs of phage genes such as bacterial integrases, with an uncharacterized false-positive rate. ₁₉ The co-mobilization analysis makes the same disclosure for keyword and Pfam matching on bakta annotations rather than PHASTER or geNomad, which may include false positives such as phage-defense systems and may miss divergent prophages. ₂₀

A prevalence figure from such a panel is therefore not a prevalence of intact elements. The finding that all 27,702 species carry prophage annotations far exceeds a cited estimate that ~46% of bacterial genomes harbour prophages, and the report attributes the discrepancy to its broader definition: it identifies individual prophage-associated gene clusters, including domesticated remnants, not intact prophages. ₁₉

Within the panel, modules differ in what they license. Modules A, D and F likely include many domesticated prophage remnants — defective prophages under purifying selection that retain packaging, lysis and regulatory genes — while the structurally variable modules B, C and G are more informative for understanding prophage ecology because their presence indicates more complete prophage elements. ₁₉

The panel does carry positive validation: all 7 modules were detected in every one of 8 well-characterized prophage-carrying species. ₁₉ The validation set was E. coli, S. enterica, S. aureus, P. aeruginosa, M. tuberculosis, B. subtilis, V. cholerae and S. pyogenes. ₁₉

An internal coherence test gives only partial support for the module definitions. Across 15 phylogenetically stratified species (one per phylum, at most 300 genomes each, 200 null permutations), 44 of 95 module-species tests (46.3%) showed significantly higher within-module gene co-occurrence than expected by chance at one-tailed z > 1.96, with mean contig co-localization across all tests of 0.769. ₁₉ The report’s own reading is that the module definitions are validated for the core prophage backbone of packaging, lysis and lysogenic regulation (A, D, F) but less so for the accessory integration and anti-defense functions (E, G). ₁₉ This refines the page’s principle that a composite panel must be validated signal by signal rather than as a whole, as the multi-heme cytochrome detector showed for domains versus motifs. _19,1

Mis-scored Reactions and Single-Signature Calls

Marker validation also applies to reaction-level signatures. The ENIGMA carbon census found that xanthine was mis-scored as carbon-catabolic. Its allowlisted reaction R02107 (xanthine→urate, xanthine oxidase) belongs to the purine pathway, which is nitrogen acquisition rather than carbon catabolism, so the carbon-callable set is effectively 8, not 9. ₂₁ The NB03 carbon allowlist has since been corrected to exclude R02107. The committed tables predate that fix and were not regenerated, so they still list xanthine (groundwater, Alkaloids, T3_kegg) with 13 strains, 0 high-certainty, 6 genera, 4 field genera and top field prevalence 0.75. ₂₁ The caveat concerns this one reaction and does not by itself rule out other purine carbon-utilization routes. The census assigns phenylethylamine to the phenylacetyl-CoA (paa) ring-cleavage route, whereas xanthine was mis-included from a purine nitrogen route. ₂₁ This supports the clay KO lesson: an allowlisted identifier can pass a pipeline filter while encoding a different biological role. _21,1

Callable counts can also overstate certainty. For terephthalic acid (necromass, Shikimates/Phenylpropanoids, T2_3_reaction), the census table reports 34 strains, 34 high-certainty, 17 genera, 10 field genera and top field prevalence 0.67. ₂₁ The report states that the “34 all high-certainty” result is a single-reaction-signature artifact, not a strength. ₂₁

All callable census evidence is catabolic-direction by construction of the NB03 filter, comprising 846 Tier-2 pathway and 102 Tier-3 signature genome–compound calls. Biosynthetic-only signatures are excluded and land in the dark set. ₂₁ The 6 biosynthesis-known compounds are tyramine, guanidineacetic acid, cinnamic acid, caffeic acid, palmitic acid and farnesol. Each has annotated biosynthetic-direction signatures but no catabolic call, and is flagged for triage against MIBiG and the biosynthetic literature as an external consult. MIBiG was not queried in the KBase Data Lakehouse. ₂₁ Reaction direction, and not only presence, is therefore part of marker identity; see pathway-versus-reaction-evidence-resolution. ₂₁

Control Panels for Marker-Based Metrics

Control panels used to validate marker-based metrics need the same scrutiny. In the gene-function ecological agora, a first-pass methodology test compared control classes using Cohen’s d (cohens d), a standardized mean difference. The values compared were per-KO Sankoff parsimony scores (sankoff parsimony), which count the minimum gains and losses on a tree. The result was 0/9 pairs passing the d ≥ 0.3 plus CI-lower-bound > 0 threshold. The signed direction also reversed against expectation for the ribosomal and RNAP-core comparisons (d = −0.66 to −2.97). ₁₅ The class medians of Sankoff/n_present were:

• neg_rnap_core, 14.50 (19 KOs);
• neg_ribosomal, 11.35 (249 KOs; median n_present leaves ~29);
• pos_betalac, 2.64 (110 KOs; 390 leaves);
• pos_crispr_cas, 2.04 (6 KOs; 1499 leaves);
• pos_tcs_hk, 1.47 (310 KOs; 918 leaves);
• neg_trna_synth, 0.30 (54 KOs). ₁₅ The report calls the neg_ribosomal median of ≈ 29 leaves (out of 18,989) the smoking gun, because genuine universal r-proteins should be in essentially all species. The class was built with the description-match rule LOWER(ipr_desc) LIKE '%ribosomal protein%'. It was dominated by clade-restricted accessory r-proteins (RimM, RbfA, processing factors, silencing factors) and by mitochondrial ribosomal protein false matches. ₁₅ This directly shows that a text-matched category can invert a control, paralleling ontology-and-category-schema-sensitivity. ₁₅

A retest reclassified the housekeeping controls by KO range:

• ribosomal K02860..K02899 ∪ K02950..K02998, 89 KOs total and 83 in panel;
• tRNA synthetase K01866..K01890, 25 KOs and 20 in panel;
• RNAP core {K03040, K03043, K03046}, 3 KOs, all in panel. The Pfam-accession-matched positive controls were left unchanged. ₁₅ Reusing the same per-KO Sankoff scores, 6/9 strict pairs then passed at d ≥ 0.3 with 95% CI lower bound > 0. ₁₅ All three comparisons against neg_ribosomal_strict still failed with negative d, reported with 95% confidence intervals:
• pos_betalac (110 vs 83), d = −0.836, 95% CI [−1.17, −0.52];
• pos_crispr_cas (6 vs 83), d = −0.919, 95% CI [−1.15, −0.66];
• pos_tcs_hk (310 vs 83), d = −1.705, 95% CI [−2.15, −1.33]. Their one-sided Mann–Whitney (MW) p values are 0.87, 0.71 and 0.96 respectively. ₁₅ This refines the first-pass null: much of that failure traced to control definition rather than to the scores. ₁₅

The ribosomal_strict pairs still failed because the K02860-K02899 + K02950-K02998 range is a pre-2000 KEGG-numbering approximation that itself mixes universal r-proteins with accessory variants. ₁₅ neg_ribosomal_strict (83 KOs) had median Sankoff/n_present 18.83 (q25=0.11, q75=23.13, bimodal) and median presence of 13 leaves. By comparison, neg_rnap_core_strict (3 KOs) had 0.073 and 18,758 leaves, and neg_trna_synth_strict (20 KOs) had 0.032 and 18,917 leaves. ₁₅ tRNA synthetase and RNAP core are clean universal housekeeping, whereas about half of the 83 ribosomal_strict KOs are universal and half clade-restricted. ₁₅

Under commitment M21 (plan v2.8), M18-style methodology validation and any d-based metric across Phase 2/3 use only tRNA-synth (K01866..K01890) and RNAP core ({K03040, K03043, K03046}) as canonical clean housekeeping. Both ribosomal classifications — description-match and the strict range — are judged too contaminated at KO level and are reported as informational only. ₁₅ The limitations restate that the strict range is bimodal (q25=0.11, median=18.83 on Sankoff/n_present). ₁₅ The ribosomal strict control’s ratio was 6.2 (38.9 / 6.3), described as mixed, as expected for the contaminated strict-K range. ₁₅ Even a KO-range definition is thus an unvalidated marker set until its presence distribution is checked against the expected universality. ₁₅

Positive controls can carry detection bias. The pre-registered cross-phylum HGT (horizontal gene transfer) positive controls (β-lactamase and class-I CRISPR-Cas) all sat at strongly negative absolute consumer z. At family rank, the relative-threshold diagnostic gave these results against neg_ribosomal (one-sided MW p):

• pos_amr, median −7.37, Δ −0.37, p = 0.94;
• pos_tcs_hk, median −6.57, Δ +0.43, p = 5×10⁻⁴;
• pos_betalac, median −6.37, Δ +0.62, p = 5×10⁻⁸;
• pos_crispr_cas, median −5.81, Δ +1.18, p = 1×10⁻⁸⁰. ₁₅ This diagnostic is run on UniRef50 families (its AMR panel is described in terms of AMR UniRefs), a different resolution from the KO-level ribosomal classes that M21 demotes; M21 states the ribosomal contamination at KO level, so it does not by itself invalidate the UniRef50-level reference used here. ₁₅ Diagnostic panel D showed that only AMR was strongly Pseudomonadota-biased, with a median of 50% and 49% of AMR UniRefs >50% Pseudomonadota; all other classes had ≤ 4% median. The report attributes this confound to bakta_amr being sourced from amrfinderplus’s Pseudomonadota-heavy reference set. ₁₅ pos_amr also had an anomalously low median Sankoff/n_present of 12.0, attributed to Pseudomonadota detection bias. ₁₅

The report therefore excluded AMR from the Phase 2 positive-control panel, keeping β-lactamase, CRISPR-Cas and TCS HK. Under M17, AMR appears only as info_amr (54.8% recent, 4.1% ancient) for transparency and is not load-bearing. ₁₅ This extends the AMRFinderPlus scope caveat above: a reference set’s taxonomic composition can bias where markers are detected at all. _15,8

Calibrating a panel to one lineage creates a similar trap. On the report’s leaf-consistency (LC) score — defined per rank × clade × KO as the fraction of clade species carrying the KO — g__Mycobacterium itself sat in the mycolic-low sub-clade (mean LC=0.36, median 0.08), while the mycolic-positive sub-clade comprises the broader Mycobacteriales/Corynebacteriales lineages that uniformly carry the panel. ₁₅ The project’s 11-KO mycolic panel is calibrated to long-chain Mycobacterium-style mycolate biosynthesis, and only the M. tuberculosis complex carries all panel KOs at high prevalence; many environmental Mycobacterium species do not. ₁₅ A panel’s negative call can therefore reflect calibration rather than absence of the function, which supports homology-search-negative-evidence. ₁₅

Phenotype databases can anchor marker-based inferences only where coverage permits. The NB24 BacDive profiles matched clade biology for two groups. mycobacteriaceae were 89.4% aerobic-leaning, Gram-positive, non-motile, rod-shaped and catalase-positive. Bacteroidota were saccharolytic, glycoside-hydrolase-rich and 1.5× anaerobe enriched. ₁₅ The Bacteroidota saccharolytic profile (maltose 230+ utilized, raffinose 155+, esculin 228+, plus cellobiose, glucose and lactose) was the strongest available bacdive anchor for the Phase 1B PUL (polysaccharide utilization locus) Innovator-Exchange finding, matching the signature predicted by PUL biology. ₁₅ For Cyanobacteriia, BacDive coverage was too thin (n=4 species) for phenotype anchoring. The anchor load instead rests on NB23 biome enrichment (63.8% photic aquatic, p < 10⁻⁵²) and NB25 env-cluster concentration (35.6% in cluster 0, which is marine- and sponge-dominant). ₁₅

Curated Lists and Coarse Proxies

The harvard forest warming analysis tested a curated 62-KO carbon-cycling list (CAZymes, peptidases, TCA, β-oxidation, aromatic catabolism, methane, C1; committed at user_data/c_cycling_kos.tsv). The test asked whether these KOs were enriched among heated-up KOs (q<0.10, log2 FC > 0), using Fisher’s exact test per pool × horizon with an H1 sensitivity check in direct samples. ₂₂

Coarse proxies dilute specificity. The ENIGMA contamination analysis depends on gtdb and eggnog conventions, and at genus-level aggregation broad COG-style signals can dilute pathway-specific metal-response effects. ₂₃ Its COG-fraction proxies are coarse summaries, not curated resistance pathways. ₂₃ This supports the soil-metal caution above that COG hits cannot automatically be read as metal-resistance markers. _23,2

Taxonomy-Inferred Function as a Marker

Assigning function from taxonomy is a weaker marker layer; see taxonomic-resolution-dependent-functional-inference. The ENIGMA sso analysis inferred biogeochemical processes from 16S composition at phylum, class and genus resolution and mapped them across the SSO 3×3 grid. ₂₄ Class-level inference covered 22 classes and 78%, and genus-level inference covered 65 annotated genera and 21%. ₂₄ Genus-based trait inference places the O₂ → NO₃⁻ → Fe(III) → SO₄²⁻ → fermentation sequence and putative hotspots on the well grid; these processes were inferred, not measured. ₂₄

The genus-inferred hotspots are M5 denitrification 7.7% (Rhodanobacter), U3 iron oxidation 2.8% (Sideroxydans) and U3 nitrification 2.3% (Ca. Nitrosotalea). Further hotspots are U1 iron reduction 2.3% (anaeromyxobacter), M4 sulfur oxidation 1.9% (Arcobacter, Thiobacillus), M4 methanotrophy 1.9% (Ca. Methanoperedens) and L9 fermentation 5.3% (Spirochaeta, Paenisporosarcina). ₂₄ Because even gene-level iron-reduction markers were misassigned in the clay project, the genus-based U1 call suggests the hypothesis of local iron reduction rather than establishing it. _24,1

Genus-level functional annotation covers only 21% of total reads (65 of 1,038 genera), so process abundance estimates are lower bounds, and the unclassified 56% of reads could harbor additional capacity. ₂₄ Species-level classification is ~0%, and genus-level classification covers only 44% of sediment reads. Function is assigned from literature-based traits rather than direct genomic evidence. ₂₄ Phylum- and class-level trait scores are consensus estimates, not measurements for these populations. ₂₄

All environmental inferences come from community composition alone and await geochemical confirmation. ₂₄ The denitrification and iron-oxidation hotspots rest on sediment data, without groundwater validation at these wells. ₂₄ The 18 SSO isolate genomes from M6-C2 are predicted to encode fermentation or sulfate reduction, but that confirmation is not reported. ₂₄ Separate ORR work on 77 sediment and 33 groundwater metagenomes is cited as complementary; those are distinct data. See subsurface-hydrogeological-zonation. ₂₄

Clinical and Serological Proxy Markers

Serological markers raise the same identity-and-resolution questions. In the IBD phage-targeting cohort of 32 Crohn’s disease (CD), 14 ulcerative colitis (UC) and 21 nonIBD subjects, the report’s sanity check found that canonical IBD-serology patterns hold: the ANCA mean was 42.6 in UC versus 17.8 in CD, matching the canonical UC-high pANCA pattern, while CD means exceeded UC for ASCA (0.28 versus 0.0), CBir1 anti-flagellin (29.0 versus 13.7), IgA ASCA (14.2 versus 1.2) and IgG ASCA (15.0 versus 3.6). OmpC was only slightly higher in CD (9.3 versus 7.6) and is recorded as a partial match. ₂₅

Marker specificity nevertheless limits what this serology supports: OmpC and ANCA cross-react with related antigens beyond their named targets, so per-species attributions are correlative rather than exclusive. ₂₅ The report concludes that serology is not yet a quantitative target-prioritization signal at the per-target level for this cohort. The canonical CD-versus-UC stratification axes (ANCA-UC, ASCA-CD, CBir1-CD) are intact, so serology can be used as a CD-versus-UC stratifier but not as a per-target abundance predictor: the |r|≈0.31 ceiling at cohort level means a single subject’s anti-M. gnavus ANCA titer cannot be used to confidently predict that subject’s M. gnavus abundance, with multi-cohort meta-analysis (out of project scope) named as the structural next step. ₂₅ This extends the page’s rule from sequence markers to clinical assays: a marker may be valid for group-level stratification and still be unusable as a per-target quantitative proxy. ₂₅

Proposed clinical proxies in the same report are explicitly provisional. mediterraneibacter gnavus qPCR (quantitative polymerase chain reaction) is offered as a candidate cheap clinical proxy for ecotype-state monitoring, avoiding full metagenomics at every visit, with a 5-fold change in M. gnavus abundance triggering full ecotype re-test; this is a proposal rather than a validated threshold. ₂₅ Likewise, the NB17 patient design categories split active from quiescent at a fecal-calprotectin threshold of 250 μg/g, which is a clinical convention not validated against per-patient response data; some patients with calprotectin 41–248 μg/g, currently in C_quiescent, may benefit from an active cocktail, and the threshold should be re-evaluated against clinical response in a pilot. ₂₅ These proxies belong with proxy-assays-for-community-state-assignment and phage-therapy-evidence-translation rather than with validated measurements. ₂₅

Preserving Supported and Unsupported Ecological Inferences

The corrected analysis does not eliminate every ecological inference from the clay project. Its SR-side finding remained robust: 5/9 deep-cohort genomes were SR-positive versus a Mitzscherling rock-attached null rate of 0.2%, with binomial p=4×10⁻¹². The bacillota report states that the SR markers (K11180, K11181, K00394, K00395, K00958) were correctly identified. ₁ It describes the “Mitzscherling rock-attached vs Bagnoud porewater” framing as half-supported (SR side) and half-unsupported (IR side, after marker correction). ₁ The new clay report likewise describes a dissimilatory sulfate-reduction signal: 5/9 deep genomes were positive versus 0.018 expected among 9 under the Mitzscherling rock-attached null, with binomial p = 4.0×10⁻¹². ₄ Thus, the porewater-bias interpretation remains supported through this retained reduction signal but loses support on the corrected iron-reduction side. _{1 4}

The central digest turns this split into a reusable diagnostic. For any cohort of cultured clay-isolated genomes, it computes SR-marker (dsrAB-aprAB-sat) and IR-marker (omcS / mtrC / mtrA) prevalences and compares them against Mitzscherling rock-attached frequencies (SRB ~0.2%, IRB ~7%) with a binomial test. ₅ The clay correction showed that KEGG has no canonical KO for omcS / mtr outer-surface cytochromes. This suggests the hypothesis that the IR arm of the diagnostic should rely on a validated detector such as multi-heme-cytochrome-detection rather than on KO lookup; the combination is untested. _5,1 The diagnostic is developed further in cultivation-collection-bias-in-ecological-genomics. ₅

This distinction supports evidence triangulation: an ecological pattern should be retained only for the function whose marker set remains biologically defensible and statistically supported. ₁ The soil-metal analysis refines this rule by showing that even statistically strong multivariate results require checks for project conditioning, co-contamination, spatial autocorrelation, and biological effect size before being interpreted as ecological functions. ₂

The functional-dark-matter analysis supports preserving this distinction when computational markers are linked to field observations. In a pre-registered lab–field mapping, 29 of 47 testable clusters were concordant (61.7%), but the one-sided binomial test gave p = 0.072; Fisher’s combined probability was p = 0.031. ₃ NMDC independently confirmed all 4 testable pre-registered abiotic predictions, yet 441 of 449 exploratory tests also reached FDR < 0.05, which the report attributes largely to compositional coupling. ₃ These results refine ecological marker validation by showing that directional replication can be informative while broad correlation screens remain vulnerable to abundance and composition effects.

Tensions

The plant-association project’s three-tier verdict places T3SS among its three species-level-robust markers, while the same report’s resolution caveats weaken that placement: n = 200 permutations is under-resolved for p ≤ 0.005, so T3SS is marginal, and no multiple-testing correction was applied across 15 markers, with Bonferroni α = 0.0033 noted as demoting T3SS while nitrogen fixation and ACC deaminase remain significant. ₁₆ The report also logs T3SS dual interpretation as documented and context-dependent. ₁₆ The tier assignment is therefore carried here as attributed rather than as an unconditional robustness claim, and settling it requires the stricter permutation run with more permutations and multiple-testing correction listed under Open Directions. ₁₆

The original analysis supported a shallow-enrichment narrative for iron reduction, whereas the corrected detector found no significant deep-versus-shallow, deep-versus-baseline, or shallow-versus-baseline difference. ₁ The disagreement is attributable to marker definition rather than to silently reconciling numerical results: K07811, K17324, and K17323 were replaced by PFAM and motif-based signals, and the resulting cohort counts and tests changed. ₁

The correction is considered robust for the multi-heme cytochrome signal but remains a Phase 1 correction that does not fully determine whether the original clay-project comparison with the Bagnoud porewater pattern should be retained. ₁ The report states that the SR side is robust, that the IR-side narrative loses force, and that a revised clay-project report amendment will note this. ₁

The bacillota report labels the retained signal only with the abbreviation “SR”, which has been read as “sulfite reduction”. The clay report describes it as dissimilatory “sulfate reduction”. Both report 5/9 positives and the same null scale. _{1 4} The bacillota report lists its SR markers as K11180, K11181, K00394, K00395 and K00958, and the digest defines the SR-marker set as dsrAB-aprAB-sat. This weakens the terminology tension in favour of the sulfate-reduction reading. The label-to-marker mapping should still be verified against the underlying marker definitions and source tables rather than assumed. _1,5

The soil-metal report presents a related interpretive tension between a conditional db-RDA result and the total environmental effect: conditioning on batch and project effects produced R² = 0.799 and p = 0.005 with 999 permutations, but the unconditional R² for metals alone was not reported and may be substantially lower. ₂ This does not contradict the conditional result; it refines what that result can support, because the reported value describes residual variance after project accession was removed rather than necessarily total community COG variance. ₂ Likewise, the 2,355 discoveries among 3,915 implied tests represent a 60% discovery rate, but co-contamination may make tests non-independent and the true FDR may be higher than reported. ₂

The functional-dark-matter analysis adds a tension between computational prioritization and functional certainty: 33,105 of 39,532 pangenome-linked dark genes received non-hypothetical Bakta descriptions, but the report still treats pathway, domain, module, and environmental links as hypotheses requiring validation. ₃ This refines rather than overturns the page’s principle: broader annotation improves candidate identification, but a database product description is not equivalent to a biologically specific ecological marker.

Practical Validation Sequence

1. Identify the biochemical function required by the ecological claim rather than beginning with an inherited KO list. ₁
2. Verify each candidate marker against protein-family domains, motifs, gene-cluster context, and known alternative functions. ₁
3. Quantify marker availability separately for every cohort before interpreting prevalence differences. ₁
4. Recompute cohort contrasts with the corrected detector and report exact counts, effect sizes, and significance values. ₁
5. Preserve supported and unsupported functional inferences separately when only part of a multi-function ecological narrative survives correction. ₁
6. For environmental associations, report effect-size distributions, test partial correlations among co-varying exposures, report both conditional and unconditional multivariate fit, and test spatial residuals and proximity thresholds before assigning metal-specific function. ₂
7. Treat pathway co-occurrence, fitness modules, conservation, reannotation, and environmental correlation as evidence-weighted hypotheses; require direct sequence, biochemical, expression, or perturbation validation before assigning a dark gene to an ecological function. ₃
8. When annotation sources disagree on a named marker, report counts from each source and their overlap rather than a single unsourced count, and never use a generic KO as a proxy for a specific gene. ₅

Relation to the Source Reports

The source report presents this workflow as reusable for comparisons in which functional labels were assigned from mismatched or weakly specific genes. ₁ The full cohort construction, detector definitions, marker counts, and statistical results are documented in Subsurface Bacillota_B Specialization — What Distinguishes Deep-Clay Lineages from Soil Congeners?. ₁ The corresponding named detection workflow is represented by multi-heme-cytochrome-detection. ₁ The clay-confined analysis applies the corrected detector to its deep, shallow, and soil cohorts and documents the resulting non-significant iron-reduction comparisons in Self-Sufficiency, Anaerobic Toolkit, and Cultivation Bias in Clay-Confined Cultured Bacterial Genomes. ₄

The bacillota report’s correction outputs are threefold. 06_clay_h3_correction.ipynb applies the triple-signal multi-heme cytochrome detector to the clay-project cohort as a Phase 1 correction. data/clay_h3_ir_orig_vs_corrected.tsv (5 rows) is the headline cohort-level table of original versus corrected IR rates. data/clay_h3_ir_corrected_fisher.tsv (3 rows) holds pairwise Fisher tests on the corrected multi-heme cytochrome rates. ₁ The report proposes a clay-report amendment clarifying that the IR-side H3 finding is artefactual while the SR side stands. ₁ The clay report now carries a correction notice that withdraws its IR-side narrative. ₄

The clay report’s original H3 outputs must be read through that correction notice. data/genome_features.parquet (61 genomes) and data/baseline_features.parquet (150 rows) carry per-genome marker booleans. The second file covers the phylum-stratified soil baseline with the same schema. Both include the original IR field, which must not be treated as validated iron-reduction evidence. ₄ data/h3_vs_mitzscherling.tsv holds 2 binomial tests against the Mitzscherling 2023 rock-attached null. data/h3_cohort_pairwise.tsv holds 6 rows of pairwise cohort Fisher tests on SR and purported IR markers, and its IR-side comparisons are not validated. data/h3_marker_class_table.tsv holds a 4-row SR_only / IR_only / both / neither breakdown, and its IR-dependent classes require re-evaluation with corrected detection. ₄ The 06_h3_porewater_bias.ipynb notebook and the h3_porewater_vs_rock.png figure depict the original SR/IR analysis. Their SR comparison remains supported, but their IR rates and IR-dependent marker classes use misidentified markers. ₄ Under the correction notice, the corrected IR rates are 55.6% (n=9 deep), 40.0% (n=30 shallow), and 40.9% (n=149 baseline), with all corrected cohort-comparison Fisher p ≥ 0.46. ₄

The soil-metal report documents the environmental-scale association analysis and its pending validation requirements in Soil Metal Concentrations Drive Functional Gene Shifts. ₂ Its planned partial-correlation, effect-size, unconditional-db-RDA, spatial-autocorrelation, and 5 km/20 km proximity analyses extend the marker-validation sequence from biological specificity to exposure specificity and spatial attribution. ₂

The functional-dark-matter report documents the complementary genome-scale prioritization workflow in Functional Dark Matter — Experimentally Prioritized Novel Genetic Systems. ₃ Its integrated use of fitness, pangenome, domain, pathway, synteny, co-fitness, reannotation, and environmental evidence supports extending marker validation from correcting named ecological markers to experimentally prioritizing uncertain ones. ₃

Open Directions

• Apply the corrected multi-heme cytochrome detector to the full clay-project branch and test whether the Bagnoud porewater comparison changes when PFAM, motif, and cluster-context evidence are evaluated together. ₁ The reported cohort reanalysis has now been performed; the remaining task is to verify marker definitions and compartment annotations in the underlying tables. ₄
• Use gene-cluster context, protein-domain annotation, and targeted biochemical or expression data to determine which CXXCH-positive clusters are plausibly involved in iron reduction rather than other electron-transfer processes. ₁
• Increase the number of rock-attached and porewater genomes, then repeat the corrected Fisher tests to ask whether substrate association explains multi-heme cytochrome distribution. ₁ Add deep-subsurface MAGs from Mont Terri, Olkiluoto, MX-80 bentonite, and Oak Ridge to test whether the cultured-cohort result generalizes beyond porewater isolates. ₄
• Compare the corrected detector with independently curated iron-reduction markers and metatranscriptomic or proteomic measurements to test whether marker positivity predicts active iron reduction. ₁
• Reclassify the 462 “other_or_unannotated” enriched OGs with an LLM-based or manual functional scan, then test whether the estimated 80–100 anaerobic-respiration-related OGs resolve into reproducible marker sets. ₁
• Decompose corrected marker prevalence by genus and cohort, using the 10-genome anchor and 62-genome baseline design, to test whether apparent ecological signals are lineage markers. ₁
• Benchmark the same marker-validation workflow in other phylum-matched subsurface comparisons to determine whether marker correction repeatedly changes ecological conclusions. ₁
• Resolve the sulfite-versus-sulfate terminology by auditing the retained marker set and linking genome presence to Bagnoud’s metaproteomic evidence. ₄
• Reanalyze the soil-metal associations with partial correlations such as COG ~ Cr | Cu + Zn + Pb, report Spearman rho values and unconditional db-RDA R², and classify significant COGs into resistance, stress, membrane, energy, and unknown categories. ₂
• Test Moran’s I on model residuals and apply SEVM if spatial autocorrelation is significant; repeat copper–COG attribution using 5 km and 20 km genome–soil proximity thresholds. ₂
• Use the available Spark tables from kescience_mgnify and kbase_ke_pangenome to determine whether the 2,355 associations survive effect-size, co-contamination, spatial, and project-conditioning audits. ₂
• Test the highest-priority dark-gene markers with targeted knockdown or deletion, direct biochemical assays, and expression or proteomic measurements to distinguish conserved computational hypotheses from true ecological functions. ₃
• Reanalyze the 29/47 lab–field concordance with compositionality-aware permutation tests and annotated-accessory-gene controls, then ask whether marker-specific associations persist beyond genus abundance. ₃
• Re-run the clay report’s 4-row H3 marker-class table and 6-row pairwise Fisher output with the corrected multi-heme cytochrome detector, so that the IR_only, both, and neither classes rest on validated markers. ₄
• Compare lanmodulin, xoxJ, and xoxF prevalence across environments separately for bakta-only, eggNOG-only, and both-source genome sets, to test whether ecological conclusions depend on annotation source. ₅
• Repeat AMR ecological analyses restricted to Tier 1 genes, with AMRFinderPlus mercury and arsenic stress-response hits separated from antibiotic-resistance genes, to test whether Tier 2 keyword matches or stress genes drive the results. _7,8
• Re-estimate the acdS root-association odds ratio with phylogenetic control, as was done for the 50 eggNOG ortholog groups, to test whether the single-marker enrichment survives lineage correction. ₅
• Replace cross-species codon-usage-bias comparisons with GC-corrected measures and test them against measured growth curves. ₉
• Regenerate the carbon-census tables with the corrected NB03 allowlist and audit other single-reaction signatures such as terephthalic acid for multi-reaction support. ₂₁
• Rebuild the ribosomal control from a curated universal r-protein set and re-run the three failing ribosomal_strict comparisons. ₁₅
• Re-test AMR consumer z with a taxonomically balanced reference set, and recalibrate the 11-KO mycolic panel for environmental Mycobacterium. ₁₅
• Test the SSO genus-inferred hotspots with gene-level markers in the 18 M6-C2 isolate genomes and with groundwater geochemistry. ₂₄
• Re-test the anti-M. gnavus serology-to-abundance relationship in a multi-cohort meta-analysis, and re-evaluate the 250 μg/g calprotectin split against per-patient clinical response in a pilot. ₂₅
• Run the sequence-level screen the lanthanide atlas declares out of scope — pseudogenes, truncated ORFs, assembly fragmentation — on its xoxF and PQQ gene calls, to test whether annotation-level marker counts survive. ₆
• Recount xoxF, mxaF and pqqA–E prevalence under the calibrated per-marker source and under the union of eggNOG and bakta, to test whether the PQQ union recommendation and the eggNOG-primary calibration give different ecological conclusions. ₆
• Perform the proposed random sample of 100 annotated metal AMR gene clusters with manual BLAST verification, to quantify the AMRFinderPlus false-positive rate per metal-type category. ₁₄
• Rebuild DISARM calls with HMM-plus-context rules and retron calls with a Millman-2020 effector-Pfam filter, then re-run the accessory-enrichment and syndrome analyses on the corrected cluster sets. ₁₀
• Report CRISPR-Cas prevalence from eggNOG descriptions and from Cas1 PF01867 on the same species set, with the gap stated, before any cross-study CRISPR comparison. ₁₀
• Recount psl-like under-named operons by KEGG ortholog rather than gene name across GenomeDepot-based and pangenome annotation sources, and verify each panel’s Pfam accessions against InterPro/UniProt. ₁₁
• Re-estimate PHB partial capability after excluding phaA/phaB-only species, and test whether the undetected phaR reflects an eggNOG annotation gap for K18080. ₁₂
• Repeat the plant-association validation with near-pure beneficial and near-pure pathogen reference organisms and report the continuous pathogenicity ratio rather than cohort labels; run the stricter permulation-style test against the full GTDB tree with more permutations and multiple-testing correction, so that the 12 markers failing within-genus permutation are reported at the genus scale the report assigns them. ₁₆
• Re-query the project’s 22 marker Pfams through InterProScan and through bakta_pfam_domains on the same species set, using the versioned-ID query pattern, and report the per-marker coverage gap separately from the format failure, before any secretion-system prevalence is compared across annotation sources. ₁₆
• Re-estimate the dual-nature rate within genome-size strata and with the continuous pathogenicity ratio, so the 54%-versus-87% quartile gradient is reported instead of a single pooled rate. ₁₆
• Re-detect SNIPE homologues with a PF13455 (Mug113) scan alongside DUF4041, then test whether the 54-of-4,572 co-annotation rate reflects divergent nuclease domains or genuinely truncated loci. ₁₇
• Re-identify NDH-2 by KO K03885 across the same organisms and compare the call set with the description-matched set, to quantify how many apparent NDH-2 hits are Complex I subunits. ₁₈
• Re-run prophage detection with geNomad or PHASTER on the same species set and report the prevalence gap against the annotation-proxy panel, so the all-species annotation result can be separated from intact-element prevalence. _19,20