Ontology and Category-Schema Dependence of Biological Conclusions

Biological conclusions are partly determined by the ontology, category schema, and matching rules used to represent observations, so analyses should treat schema choice as an explicit source of uncertainty rather than as a neutral preprocessing step. ₁ The findings summarized in discoveries connect this issue to environmental-resistome, composite-functional-annotation, evidence-triangulation-for-functional-annotation, and cross-tenant-data-bridging.

Category definitions can reverse pathway conclusions

In an inflammatory-bowel-disease pathway analysis, a regex-based category scheme identified 3 of 52 CD-up pathways in 7 themes and produced a structurally degenerate result. ₁ A MetaCyc hierarchy assigned 262/409 pathways to at least one of 12 IBD themes and supported iron/heme acquisition with OR=8.1, FDR=7e-6, and 15/52 CD-up pathways. ₁ Here, FDR means false discovery rate, and OR means odds ratio. ₁

This contradicts any interpretation that the biological theme is independent of category construction: the same underlying pathway results yielded materially different counts and conclusions under the two schemas. ₁ The evidence supports using curator-validated ontology hierarchies as the primary analysis and regex categories as a sensitivity analysis. ₁

The project report behind this contrast gives the fuller sequence. Its v1.7 hypothesis H3a was partially supported: 2 of 3 clauses passed, and clause (b) was judged structurally degenerate rather than a fundamental refutation. ₂ Clause (b) was recorded as FAIL (degenerate). Only 44 / 409 background pathways matched the 7 a-priori MetaCyc categories with the v1.7 regex patterns, and only 3 of 52 CD-up passing pathways landed in those categories. The test had ~zero power, because the null also reached 100% top-3 concentration. ₂ Under that schema, unstratified pathway differential abundance did not preferentially load on the classical IBD-themed categories (bile-acid 7α-dehydroxylation, mucin degradation, sulfidogenesis, TMA/TMAO, eut/pdu, polyamine, AA-decarb). The CD signal instead captured broader bacterial-fitness-in-inflamed-gut themes (heme/iron, glyoxylate, fat metabolism, allantoin/purines) that were absent from the prior-literature category set as constructed. ₂

The report offered two non-mutually-exclusive explanations. The first is category-set choice: the categories were drawn from prior pathobiont-mechanism literature, and adding categories such as “alternative-electron-acceptor metabolism”, “iron acquisition” and “fat-utilization-in-inflammation” might recover the signal. The second is that unstratified pathways may be dominated by housekeeping pathways shared across many genera. ₂ NB07b tested the second option. It asked whether species-resolved CD-up pathways, in the HUMAnN3 stratified form PWY-XXX|g__species, concentrate in the 7 categories. ₂ H3a (b) was NOT SUPPORTED for all 6 Tier-A core species. The report calls this verdict structural, because the 7 categories are sparse in per-species pathway repertoires, rather than biological. ₂ Combining NB07a and NB07b, the report concluded that the 7 categories are too narrow to capture HUMAnN3 MetaCyc CD signal at either unstratified or species-resolved level. In its reading, the refutation is real for the v1.7 category set but does not mean that no compositional themes exist; a broader schema, derived programmatically from MetaCyc taxonomy, was needed. ₂ This refines taxonomic-resolution-dependent-functional-inference: moving to species resolution did not rescue a schema that was too narrow at both grains. ₂

Version v1.8 replaced the regex with structured MetaCyc class assignments from the ModelSEEDDatabase MetaCyc_Pathways.tbl file. It expanded the schema to 12 IBD-relevant themes, including iron/heme acquisition, fat metabolism / glyoxylate, anaerobic respiration, purine/pyrimidine recycling, and aromatic AA / chorismate / indole. ₂ The v1.7 regex had silently placed PWY-5920 (superpathway of heme biosynthesis from glycine) in “0_other”, because “heme” and “iron” were not in its pattern set. The ModelSEED class hierarchy assigns PWY-5920 to HEME-SYN, Heme-b-Biosynthesis, Cofactor-Biosynthesis and Tetrapyrrole-Biosynthesis. The report credits this kind of reassignment with reversing the verdict without any change in the underlying data. ₂ The v1.8 cohort test used per-theme Fisher’s exact enrichment (CD-up × in-theme) across 12 themes, with Benjamini–Hochberg FDR correction across themes. A theme counted as supported if FDR < 0.10 AND odds ratio > 1.5. ₂

The regex scheme (44/409 pathways categorized, 7 themes) gave H3a (b) “FAIL — degenerate”. The ModelSEEDDatabase MetaCyc class hierarchy (514/575 pathways categorized, 12 themes; 90 % coverage of 575 HUMAnN3 pathways) gave “SUPPORTED”. Both results came from the same data and the same differential-abundance pipeline. Under the class hierarchy, iron/heme acquisition was the dominant CD-up theme (OR=8.1, FDR 7e-6; 15 of 52 CD-up pathways), alongside H. hathewayi purine/pyrimidine recycling (OR=4.9) and TMA/choline (OR=9.3). ₂ At species level under v1.8, H. hathewayi had two supported themes (16 CD-up pathways): purine/pyrimidine recycling (OR=4.86, FDR=0.048) and TMA/choline (OR=9.33, FDR=0.048). The other Tier-A core species had insufficient species-level CD-up pathway counts for per-theme power: M. gnavus, E. lenta and F. plautii had 0 CD-up pathways, and E. coli and E. bolteae had 1-2. The species-level schema test is therefore power-limited for most species. ₂ The report states that this reverses the v1.7 verdict and is driven entirely by category-schema choice, not by any change in the underlying data: “Same data, opposite verdict”. ₂ It records the episode as a major scientific reversal and encodes the lesson as plan norm N17: prefer ontology or class hierarchy over name-pattern regex for pathway, gene and metabolite categorization wherever feasible, and keep regex as a sensitivity check. Under the same norm, the ModelSEEDDatabase MetaCyc class hierarchy, with 90 % coverage of HUMAnN3 outputs, should be the default for pathway category-enrichment tests. ₂ This supports the digest-level framing above with project-level detail and bears on iron acquisition and pathobiont expansion. The digest and the report state the coverage of the two schemas with different figures (see Tensions). _2,1

Composite annotations can carry biological signal

Across 32 species and 9 phyla, novel or singleton genes were enriched in COG L mobile elements by +10.88%, COG V defense mechanisms by +2.83%, and COG S unknown function by +1.64%. ₁ Core genes were depleted in COG J translation by -4.65%, COG F nucleotide metabolism by -2.09%, COG H coenzyme metabolism by -2.06%, COG E amino acid metabolism by -1.81%, and COG C energy production by -1.75%. ₁

Composite assignments such as LV showed +0.34% enrichment and 76% consistency, indicating that multi-function assignments can represent real mobile-defense modules rather than annotation noise. ₁ This supports composite-functional-annotation and cautions against forcing genes with multiple functional labels into a single mutually exclusive category. ₁

Annotation systems define different measurable spaces

Bakta and eggNOG produced complementary functional coverage across 132.5M gene clusters. ₁ eggNOG had higher COG coverage, at 51% versus 8.2%; KEGG coverage, at 38.5% versus 17.3%; and Pfam coverage, at 63% versus 7.7%. ₁ Bakta had higher GO coverage, at 15% versus 7.4%; product descriptions, at 71.2% versus 70.4%; and unique UniRef50 links covering 79.2%. ₁

Their union increased any-functional-annotation coverage to 77.3%, and Bakta rescued 11.2M clusters among 39.2M missed by eggNOG. ₁ Only 33.3% of Bakta’s 17.6M distinct UniRef50 IDs existed in the KBase Data Lakehouse UniProt identifier table. ₁ These results refine evidence-triangulation-for-functional-annotation: apparent functional absence can reflect annotation-system coverage or identifier-table coverage rather than biological absence. ₁

Marker and ontology proxies require validation

Annotation proxies can produce substantially different biological counts depending on the selected preferred name, KO, or product definition. ₁ eggNOG Preferred_name='lanM' produced 505 additional hits with zero overlap with 62 Bakta-validated Lanmodulin genomes. ₁ eggNOG KO K02030 produced 46,369 nonspecific hits, and only 418 of 5,092 genomes with any xoxF marker hit both eggNOG K00114 and Bakta lanthanide-dependent methanol-dehydrogenase products. ₁

The findings support environmental-resistome and homology-search-negative-evidence by showing that generic KOs and stale preferred names should not be treated as definitive markers. ₁ Analyses should report marker definitions explicitly and validate them against an independent annotation or curated reference set. ₁

Schema choice interacts with integration and transfer

The category-schema problem extends beyond pathway enrichment because joins, classifiers, and cross-dataset comparisons depend on compatible representations of the same biological object. ₁ Relative-abundance spaces were cross-cohort-portable because they were unitless and compositional, whereas absolute-intensity metabolomics required explicit batch correction such as ComBat, SVA, RUV, or quantile normalization. ₁ In a pooled HMP2 and Franzosa metabolomics analysis, PCA plus K-means with K=4 separated 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. ₁

A CLR-plus-PCA GMM projection assigned all 26 Kuehl samples to E3 at confidence greater than 0.97 because Kuehl detected only 54% of training species, whereas LDA projection across MetaPhlAn3 and Kaiju namespaces produced plausible Kuehl proportions of 27/42/31% across ecotypes. ₁ These results refine cross-cohort-microbiome-portability: representation and namespace compatibility can determine whether a transfer appears biologically plausible or fails through projection artifacts. ₁

Representation choice also bounds how far one feature space can recover categories built in another. Pathway-based K=4 LDA agreed only partly with the taxon-based consensus_ecotype across 3,145 CMD_IBD samples (ARI 0.113). E1 was most recoverable (65.3 % agreement) and E3 weakest (30.7 %). ₂ This refines ecotype-clustering-validity: an ecotype label is partly a property of the feature space that defines it. ₂

Annotation vocabulary and grain change module and dark-gene conclusions

Annotation vocabulary choice changed how much of a fitness-module set could be interpreted. Adding PFam domains and lowering the enrichment overlap threshold from 3 to 2 increased the module annotation rate from 8% to 80% (92 to 890 modules), yielding 7.6x more function predictions. ₃ The same report found near-zero strict KO precision for Module-ICA and cofitness modules because KEGG KO (KEGG Orthology) groups are gene-level assignments (~1.2 genes per unique KO), so a module with 20 annotated members typically has 20 different KOs. It concludes that PFam provides the broadest annotation coverage, while KEGG KOs are too gene-specific for module-level enrichment. ₃

This gain is a coverage gain, not a demonstrated precision gain: PFam annotations are domain-level and may overcount functional associations. ₃ The result refines fitness-module-detection-sensitivity and pathway-versus-reaction-evidence-resolution. Whether a module looks functionally coherent depends partly on whether the vocabulary is coarser or finer than the module itself, and this rests on a single project’s comparison. ₃

A category can also be too coarse to discriminate. In the functional-dark-matter prioritization, breadth classes derived from eggNOG ortholog-group (OG) hierarchies placed 30,721 of 30,756 dark gene clusters (99.9%) in the “universal” class, so the classification did not discriminate among candidates. A species-count metric gave finer resolution, ranging from 1 to 33 species per OG (median=1, mean=2.2). ₄ Adding has_bakta_annotation as a 7th evidence flag changed the darkness tier of 18,019 genes, with the largest shifts from T4 Penumbra to T5 Dawn as genes gained their final missing evidence line. ₄ This supports the Bakta/eggNOG coverage finding above and refines comparative-conservation-metric-calibration: tier and breadth labels depend on which evidence flags and hierarchy levels the schema includes. _4,1

Manual condition and environment categories embed judgment

Keyword-based environment categories leave a large residual. In the environment-embedding analysis, an additional 13,944 genomes (17%) were classified as “Other”. These included site-specific labels such as “Aspo HRL” and “Olkiluoto” (underground research labs), generic terms such as “water” and “bodily fluid”, and clinical sites not captured by current keywords. ₅ In the field-versus-lab fitness comparison, condition classification relied on manual mapping of condition_1 labels. Edge cases such as zinc sulfate (metal versus sulfate source) required subjective judgment. ₆

Upstream field structure can force categories to collapse. The functional-dark-matter project planned 7 condition-environment mappings but implemented 6: stress (consolidating metal, osmotic and oxidative), carbon source, nitrogen source, pH, motility and anaerobic. The consolidation was needed because the Fitness Browser expGroup field uses a broad “stress” class. Motility and anaerobic were added because they emerged as prominent condition classes among dark genes with strong fitness effects. ₄ These caveats refine lab-field-fitness-concordance and environment-embedding-geography: stressor-specific or habitat-specific conclusions are bounded by how finely the source metadata and the analyst’s mapping resolve conditions. _4,6,5

Free-text isolation sources show the same residual problem. In the Pseudomonas carbon-ecology project, keyword-based classification introduced ~6.7% “unknown” and ~29.1% “other” classifications, and the report notes that misclassification could attenuate true ecological signal. ₇ Pathway catalogs impose their own boundary. The 62 GapMind carbon pathways cover common carbon sources but miss genus-specific catabolic capabilities, particularly aromatic degradation pathways (toluene, naphthalene, benzoate) that are central to P. putida ecology. ₇ Both limits refine environment-embedding-geography and cross-condition-metabolic-comparability: habitat and capability calls are bounded by the residual bins and catalog scope. ₇

Sequence-cluster, KO and domain-architecture grains give different evolutionary answers

The gene-function ecological-agora project tested whether the two-component-system histidine kinase (TCS HK) paralog expansion reported by Alm, Huang & Arkin (2006) reproduced across annotation grains. At UniRef50 sequence-cluster resolution the result was negative. The report treats this as a negative result for reproduction at that resolution rather than a methodology failure. Alm 2006 measured expansion at the HK family level (number of distinct HK genes per genome), whereas each UniRef50 typically corresponds to a single HK protein variant and cannot aggregate to the family. ₈ Reproduction had been pre-registered at Phase 2 (KO-level aggregation; KEGG ko02020) and Phase 3 (Pfam multidomain architecture census). The UniRef50 null rules out the simpler interpretation that any sequence-cluster-level methodology would reproduce Alm 2006. ₈

Control-set membership also depended on annotation source. A mid-pilot audit found that kbase_ke_pangenome.interproscan_domains (146 M Pfam hits across 132.5 M cluster representatives; 83.8 % cluster coverage) had not been used in v1. Taking the union of eggNOG name matches with InterProScan accessions increased the detected controls, as listed below. ₈

• ribosomal proteins rose from 9,005 to 19,389 UniRef50s (2.2×);
• tRNA synthetases rose from 5,640 to 8,514 (1.5×);
• TCS HKs rose from 38,488 to 43,217 (1.1×). ₈

This supports evidence-triangulation-for-functional-annotation: a name-match schema alone under-counts even canonical housekeeping families. ₈

At phylum rank, natural-expansion UniRefs had 39.5% more paralogs than the cohort baseline (1.77 vs 1.27 paralogs per phylum). Negative controls (ribosomal, tRNA-synth, RNAP) sat ~12–15% below baseline. TCS HK at UniRef50 sat 18% below cohort, the opposite direction from the Alm 2006 family-level expansion claim. ₈

By the report’s own analysis, this null is not a power artefact. Cohen’s d is a standardized mean difference. A one-sided α=0.05 power analysis gave the minimum detectable d at 80% power and the observed z at each rank, listed below. ₈

• genus (n=73): minimum detectable d 0.29, observed z −0.10;
• family (n=124): 0.23, observed z −0.16;
• order (n=157): 0.20, observed z −0.19;
• class (n=174): 0.19, observed z −0.23;
• phylum (n=186): 0.18, observed z −0.23. ₈

The report states that an effect of d ≥ 0.3 would have been detected with 81–99% probability at all ranks. The consistently negative z indicates slight under-expansion, so underpowering is ruled out: at UniRef50 the TCS HK paralog signal is genuinely absent, and the original effect lives at family resolution. ₈ This refines null-results-under-limited-statistical-resolution: a well-powered null can still be specific to one grain. ₈

Horizontal gene transfer (HGT) signal was detectable at UniRef50. CAZymes, β-lactamases and class-I CRISPR-Cas all showed +0.6 to +1.2 σ less clumping than housekeeping at family rank with p << 10⁻⁷. ₈ The pre-registered hypothesis that Bacteroidota act as Innovator-Exchange on PUL CAZymes nonetheless failed at all 4 deep ranks under the absolute-zero criterion. A relative-threshold diagnostic showed that CAZymes do discriminate from housekeeping (+0.78 σ at family rank, p = 1×10⁻⁴³). The effect was small, however, at Cohen’s d ≈ 0.07–0.09, which the report judges below biologically meaningful effect sizes. The hypothesis is therefore not supported at UniRef50 resolution. ₈

The report distinguishes this UniRef50 null from reviewer-cited literature describing rapid CAZyme HGT at PUL or family aggregation level. It attributes the difference to substrate resolution, not necessarily to biological disagreement. ₈ This refines horizontal-gene-transfer-driven-innovation: transfer claims need to name the unit (sequence cluster, KO, family) at which they are tested. ₈

KO-level amplification was set up as a falsification gate rather than assumed. The M18 hard amplification gate required the Phase 2 KO atlas to reach Cohen’s d ≥ 0.3 on Sankoff parsimony for at least one positive HGT control versus negative housekeeping. In the strict retest, a pair passed only at d ≥ 0.3 with a 95% CI lower bound > 0. The plan text gives the UniRef50 baseline as d = 0.15, and the Phase 2 description gives it as d = 0.146 (NB08c). ₈ Failure would have falsified the substrate-hierarchy claim and triggered an M11 redesign. One passage describes that redesign as a switch from Sankoff to gene-tree-vs-species-tree reconciliation; another describes it as a switch to direct phyletic-incongruence on KO presence/absence. Larger KO-level effects were stated as an expectation to test, not a validated prediction. ₈

The first KO-level pass exposed a category-definition failure in the housekeeping controls. The class medians of Sankoff/n_present are listed below; the leaf counts in parentheses are each class’s median n_present_leaves, meaning the median number of tree leaves containing a KO, not class-wide totals. ₈

• neg_rnap_core (19 KOs): 14.50;
• neg_ribosomal (249 KOs; median n_present_leaves ~29): 11.35;
• pos_betalac (110 KOs; median n_present_leaves 390): 2.64;
• pos_crispr_cas (6 KOs; median n_present_leaves 1499): 2.04;
• pos_tcs_hk (310 KOs; median n_present_leaves 918): 1.47;
• neg_trna_synth (54 KOs): 0.30. ₈

The report calls the ribosomal class’s median n_present ≈ 29 leaves out of 18,989 the smoking gun. The description-match rule LOWER(ipr_desc) LIKE '%ribosomal protein%' had captured clade-restricted accessory r-proteins: RimM, RbfA, processing factors, silencing factors, and mitochondrial ribosomal-protein false matches. ₈

The retest redefined housekeeping by the explicit KO ranges listed below. The positive controls were left unchanged because Pfam-accession matching is more specific than description matching. ₈

• ribosomal K02860..K02899 ∪ K02950..K02998: 89 KOs total, 83 in panel;
• tRNA synthetase K01866..K01890: 25 KOs, 20 in panel;
• RNAP core {K03040, K03043, K03046}: 3 KOs, all in panel;
• positive controls: unchanged (Pfam-accession matching). ₈

tRNA-synth and RNAP core behaved as clean universal housekeeping. The strict ribosomal range remained bimodal (q25=0.11, median=18.83 on Sankoff/n_present), with about half its 83 KOs clade-restricted. M21 therefore binds canonical clean housekeeping to tRNA-synth plus RNAP core and reports ribosomal results as informational only. ₈ The report records a generalizable lesson: description matching that was mostly clean at UniRef50 did not survive KO aggregation, because a 50% majority-vote threshold for assigning a KO to a class admits accessory members. ₈ This supports functional-marker-validation and the marker-proxy caveats above. _8,1

The strictness of the housekeeping definition also changed the inheritance signature. Loose housekeeping classes, which include paralogs and weakly-housekeeping members, sat in the middle, while strict housekeeping separated cleanly toward vertical inheritance. ₈ This refines two-speed-bacterial-genome: the apparent split between vertically inherited and exchanged genes depends on how tightly the housekeeping category is drawn. ₈

With the cleanest housekeeping controls (tRNA-synth and RNAP core), M18 passed: 6/9 strict pairs passed at d ≥ 0.3 with 95% CI lower bound > 0. Cohen’s d rose from 0.146 (UniRef50, NB08c, positives versus all housekeeping) to 0.665–3.558 (KO, NB09c, best pairs), which the report describes as a 4–25× amplification depending on positive class. ₈ A reviewer critique held that effect sizes (d=0.146-0.665) were below biological significance. The report judges this only partially correct. The UniRef50 baseline d=0.146 is small, but by Cohen 1988 conventions (d=0.2 small, 0.5 medium, 0.8 large) the Phase 2 KO values of 0.665, 0.753, 0.792, 1.905, 2.504 and 3.558 are not, and the two scales should not be conflated. ₈ The mechanism remains open. A close reading reframed the amplification as possibly metric correction rather than aggregation per se, because Alm 2006 worked at a single-domain level comparable to UniRef50. The report notes that a clearer theoretical framing is still needed. ₈ The pass was declared only after the control category was redefined. This result therefore supports the page’s central claim that category construction is part of the inference, and refines gene-function-acquisition-depth. ₈

Domain-architecture resolution

The Phase 3 Pfam architecture census was focused rather than atlas-wide, covering 376 KOs across 4 subsets. An 833M-row IPS scan with the full 10,750-KO scope was infeasible via Spark Connect within practical timescales. The subsets differed sharply in architectural diversity, as listed below. ₈

• PSII (K02703-K02727): 24 KOs, a median of 1 architecture per KO, dominant fraction 1.00, top architecture PF00124 (PsbA);
• TCS HK: 292 KOs, median 15, dominant fraction 0.66, top PF00072_PF00486 (Response_reg + Trans_reg_C);
• curated mycolic-acid set: 11 KOs, median 5, dominant fraction 0.93, top PF13561 (Enoyl-ACP reductase);
• Mixed-top-50 set (highest Innovator-Exchange in mixed-category): 48 KOs, median 46, dominant fraction 0.76, top PF00005 (ABC transporter). ₈

From this contrast, the report frames architectural promiscuity as correlating with HGT propensity: mixed-category KOs have a median of 46 architectures/KO versus PSII’s 1. It pairs this with an NB17 architectural-versus-KO concordance of r=0.67 on the consumer side, and presents the result as an analog of Denise 2019 (“modular systems exchange more”). ₈ Because the contrast rests on a focused census of selected subsets, it suggests a hypothesis rather than an atlas-wide finding: a single KO label can hide many domain architectures with different transfer behavior. ₈

For TCS HK, 292 candidate KOs decomposed into ~4,000 distinct Pfam architectures across 3.2M gene_clusters. ₈ The NB17 TCS HK architectural backtest compared per-(family × architecture) and per-(family × KO) categories at family rank by Spearman correlation across 8,498 overlapping tuples. Producer correlation was r = 0.093 (exploratory, below the 0.6 threshold) and consumer correlation was r = 0.673 (confirmatory, above it). The backtest script and figure report these as consumer concordance r=0.67 (confirmatory) and producer r=0.09 (exploratory), and Phase 3 was recorded as PASS_MIXED on the same values. ₈ The report attributes the split to measurement grain. Paralog count is intrinsically a per-KO quantity, and decomposing it across architectures dilutes the producer signal. Consumer behavior (cross-clade dispersion), by contrast, is a property of where an architecture lives in the tree. ₈ This refines confirmatory-exploratory-ecological-association-discordance: changing the unit of analysis preserved one inferred role and dissolved the other. ₈

Architecture-level evidence supports Alm 2006 only qualitatively. NB17 found the canonical PF00512 + PF02518 TCS HK architecture concentrated at recent ranks, with consumer-side architectural/KO concordance r = 0.67. However, the explicit r ≈ 0.74 quantitative anchor did not survive scaling from 207 genomes to 18,989. The report calls its “Alm 2006-inspired” framing honest because the methodology generalizes while the point estimate does not. ₈ The two analyses also define lineage-specific expansion differently. Alm 2006 used per-genome paralog counts at the HK family level, capturing expansion at any phylogenetic depth. M22 instead uses Sankoff-parsimony recent-rank (leaf-adjacent) gain attribution on the species tree, which would miss expansions occurring entirely within a single recent species. The two definitions converge for very-recent expansions but diverge for older ones. ₈

Pathway-defined KO sets inherit the looseness of the pathway map. The mycolic-acid candidate set comprised 715 KOs, drawn from KEGG pathways ko00061, ko00071, ko01040 and ko00540 or from a curated list of 11 mycolic-acid-specific KOs, all present in the GTDB substrate. The ko00540 pathway is actually LPS biosynthesis and was included for plan-faithfulness. ₈ The report calls the test somewhat conservative because the pool includes general fatty-acid biosynthesis. Its suggestion that a stricter definition might amplify the signal, at the cost of sample size, is untested. ₈

Category-level contrasts in acquisition depth were small. The most ancient-skewed category was an off-target “other” bin (ko04xxx organismal systems, ko05xxx human diseases). The report suggests these may be mostly evolutionarily older or eukaryotic-derived annotations that pass through PROKKA → eggNOG mapping accidentally. ₈ This refines gene-function-acquisition-depth: a category’s depth profile can reflect annotation-pipeline leakage rather than biology. ₈

Theme backgrounds dilute or restrict enrichment

In Crohn’s disease (CD) metabolomics, theme-level tests were diluted by heterogeneous backgrounds. Only 2 of 22 acyl-carnitines were CD-up: C16 carnitine (+0.41, FDR=0.08) and C18:1 carnitine (+0.39, 0.09). Acyl-carnitine theme enrichment was OR=1.06 and not significant, because the 22-carnitine background is dominated by short/medium-chain forms that do not shift in CD. ₂ Lipid classes (CE / SM / TAG / DAG / ceramide) had 19 CD-up metabolites, but theme-level OR=1.64 (FDR=0.76, not significant). The signal was dominated by cholesteryl esters of long-chain PUFAs (12 CEs, 4 SMs, 1 ceramide, 2 TAGs). The report attributes the null to a 163-metabolite background that is too dilute. ₂

On the gene side, 95 % of FDR-passing genes were unclassified by the gene-symbol regex, because the curated 13-category classifier captures only annotated functional roles. The reported functional enrichment therefore applies to the category-classified subset, not the full gene set. ₂ Together with the regex-versus-MetaCyc pathway contrast above, this supports treating theme definitions and background sets as model choices. It also refines metabolite-class-signatures-of-gut-inflammation: a nonsignificant class-level test can coexist with a coherent sub-class shift. _2,1

Threshold and membership rules define gene categories

Gene categories defined by fitness thresholds are themselves schema choices. The metal-fitness atlas defines metal-important genes as fit < -1 OR n_sick ≥ 1. This captures ~3.3% of genes, most of them general stress genes rather than metal-specific ones, and it differs from an earlier fit < -1, |t| > 4 wording in the same report. ₉ The atlas uses this breadth to explain a disagreement with the prior DvH (Desulfovibrio vulgaris Hildenborough) field-versus-lab analysis. That analysis found heavy-metal resistance genes to be the least conserved condition class (71.2% core), because it isolated condition-specific genes important only for metals. The atlas instead captures all genes with any metal fitness defect, which are dominated by core stress-response functions. ₉ This refines pangenome-conservation-fitness-decoupling: whether metal-relevant genes look core or accessory depends on whether the category is condition-specific or inclusive. ₉

Threshold choice also changes the measured overlap between stress categories. The counter_ion_effects project found 39.8% overlap between metal-important and NaCl-stress genes, whereas the metal-specificity analysis found 14.7% of metal-important genes sick under osmotic stress, a 2.7x discrepancy. ₁₀ The metal-specificity report attributes the difference to method. It used the stricter threshold fit < -1 AND |t| > 4, versus fit < -1 only in counter_ion_effects, and the organism sets differ partially. ₁₀ This refines shared-stress-versus-stressor-specific-fitness: how much of a metal response looks shared with osmotic stress depends on the significance rule that defines category membership. _10,9

Conservation categories carry similar arbitrariness. The module-conservation project’s 90% and 50% cutoffs for core/mixed/accessory module classification are convenient but, by the report’s own assessment, not biologically motivated. ₁₁ In the pangenome table, singleton clusters are a subset of auxiliary clusters (no_singleton_gene_clusters ⊂ no_aux_genome), so the two counts are nested rather than mutually exclusive categories. ₁₂ Both points refine pangenome-core-boundary-and-clade-size-bias. _11,12

Marker-based phenotype classes shift with the classification rule. In the plant-microbiome ecotype project, NB02’s binary rule classed any species carrying ≥1 PGP and ≥1 pathogenic marker as dual-nature, giving 15,474/25,660 species (60.3%). The report says this reflects broadly prevalent functions such as flagella, chemotaxis and T6SS. NB07’s weighted composite score (40% PGP markers, 20% core fraction, 15% complementarity, 10% metabolic breadth, 15% pathogen penalty) used adaptive thresholds set at the median of non-zero scores. It classified 25.2% as dual-nature, alongside neutral 38.0%, pathogenic 29.5% and beneficial 7.3%; the passage gives no NB07 denominator. ₁₃ The report describes the 35 percentage point gap as the difference between carrying both marker types and scoring highly on both axes. The NB02 classification is more inclusive, and the NB07 classification is more discriminating but also more assumption-laden. ₁₃ The report flags T3SS classification as context-dependent. Reclassifying T3SS, T6SS and T2SS from “pathogenic” to “colonization” markers moved 16.4% of dual-nature species to PGP-only, while plant-associated species remained 86.0% dual-nature. The report reads this as robustness of the dual-nature finding. ₁₃ This supports functional-marker-validation and refines dual-nature-plant-associated-bacteria: the overall dual-nature rate depends on the marker schema as well as on the genomes. ₁₃

Coarse habitat categories and leave-one-out sensitivity

MicrobeAtlas uses 13 broad environment categories that lump heterogeneous habitats. The most problematic is “aquatic”, which combines marine, freshwater, estuarine, hydrothermal vent and pond environments that differ enormously in salinity, temperature and chemistry. ₁₄ A taxon detected in both a marine coastal sample and a freshwater lake sample therefore counts as broad-niched. This could reflect closely related but ecologically distinct ecotypes rather than a single generalist population, so the category can inflate inferred niche breadth. ₁₄

The metal-ecology project tested category dependence with a leave-one-out sensitivity analysis (S2). For each of the 13 categories, it excluded that category, row-normalised prevalences over the remaining 12 environments, and recomputed standardized niche breadth (B_std) with J=12. It then re-ran PGLS (phylogenetic generalized least squares, a regression that accounts for shared ancestry). ₁₄ All 13 leave-one-out models were significant (p < 0.05). The weakest result came from excluding “aquatic”, the most heterogeneous and most over-represented category (40,353 OTUs; 40.8% of the dataset). There, β dropped from +0.021 (main) to +0.0085 (−59%) yet remained significant (p = 0.031). ₁₄ Excluding the 3 least-sampled environments gave β = +0.012, +0.013 and +0.010, all p < 0.008: flower (898 OTUs, 0.9%), leaf (1,298 OTUs, 1.3%) and plant (2,647 OTUs, 2.7%). Excluding any environment other than aquatic gave p < 0.01. ₁₄

The aquatic exclusion is nominally significant but not FDR-significant. The S2 table lists excl_aquatic with n=606, β=+0.0085, SE=0.0039, p=0.031 and BH-FDR q=0.061, marked “no”. ₁₄ The exclusions of every other environment, by contrast, remain significant after the report’s FDR correction, so the aquatic exclusion is the only one that loses FDR significance. ₁₄ The report’s text nonetheless states that the result is not contingent on the heterogeneous “aquatic” category. ₁₄ Leave-one-out λ values were also consistently higher (0.87–0.91 vs 0.71 in the main analysis). The report attributes this to rescaling B_std over fewer categories, which increases the relative weight of shared environments across genera and amplifies phylogenetic covariation. ₁₄ This refines confirmatory-exploratory-ecological-association-discordance: the association survives category removal nominally, but its strength and multiple-testing status depend on the largest and most heterogeneous category. ₁₄

The report’s interpretation of the coefficient change is internally inconsistent. It calls the over-representation of aquatic environments (40.8% of OTUs) a potential source of downward bias when the category is included. Yet excluding it produces a weaker β (+0.0085 vs +0.021), which the report reads as suggesting that aquatic OTUs contribute disproportionately to the association. That reading is a suggestion, not an established finding. ₁₄ The sub-category concern cannot be tested with current data, because MicrobeAtlas does not provide finer environment classification. Splitting aquatic into marine, freshwater and saline sub-categories requires re-analysis of sample metadata not available in the extraction used. The report proposes re-running the niche-breadth computation with a finer taxonomy from MicrobeAtlas sample metadata, to test whether category heterogeneity inflates B_std for certain genera. It scores the question of whether the categories are ecologically meaningful as only partially addressed: leave-one-out was done, but sub-category splitting needs new metadata. ₁₄

The same project’s metal categories have a residual bin too. The “other” metal category (17% of annotated clusters) captures metal-associated genes not classified into the 7 primary types and may include weaker hits. The passage calls its exclusion a “6-metal analysis” yet lists seven metals (Hg, As, Cu, Zn, Cd, Cr, Ni); this inconsistency is kept as written. ₁₄ This refines composite-resistance-score-limitations and environmental-resistome: metal-type profiles depend on how the residual category is handled. ₁₄

Identifier and label vocabularies silently empty categories

Category assignment can fail outright when the code expects one identifier vocabulary and the table stores another. In the PHB granule project, PHA synthase class analysis classified all 11,792 phaC clusters as “other_pfam”. The eggNOG PFAMs column uses domain names (Abhydrolase_1, PhaC_N), whereas the classification code expected Pfam accession IDs (PF00561, PF07167). The class-distribution output covers 6,067 phaC-carrying species, all assigned “other_pfam”. ₁₅ The report lists this as a remaining limitation. It proposes mapping PF00561 -> Abhydrolase_1 and PF07167 -> PhaC_N to enable proper Class I-IV classification. The null is therefore a schema failure, not evidence about synthase-class distribution, and it bears on phb-granule-ecology. ₁₅

Stored label values can differ from expected ones. Filtering metabolic_category == 'amino_acid' returns 0 rows, because the stored values are 'aa' (amino acid pathways) and 'carbon' (carbon source pathways). ₁₂ The Fitness Browser seedclass table (columns orgId, locusId, type, num) stores EC numbers, not SEED subsystem categories. Functional descriptions come from seedannotation (orgId, locusId, seed_desc). ₁₂ When COG categories are mapped to descriptions through a dictionary, composite categories such as “LV” and “EGP” return NaN if absent, and string operations on those NaN values fail. This refines composite-functional-annotation: composite codes need explicit handling in code as well as biological interpretation. ₁₂

Keyword classifiers miss enzymes with non-obvious names. A RAST-description keyword scheme classified ACIAD1710 (4-carboxymuconolactone decarboxylase, EC 4.1.1.44), a core pca pathway enzyme, as “Other”. The keyword list checked for “muconate” but not “muconolactone”. ₁₂ Gene cluster IDs are meaningful only within a species. Cross-species gene-content comparisons should therefore use COG categories, KEGG orthologs or PFAM domains instead of cluster IDs. ₁₂ This supports evidence-triangulation-for-functional-annotation and pangenome-integration. ₁₂

Label inconsistencies also appear inside reports. The snipe defense-system report’s query footnote labels PF00358 as PTS_EIIC and PF02378 as PTS_EIID, which conflicts with its table labels PF00358/PTS_EIIA_1 and PF02378/PTS_EIIC. ₁₆ In the NMDC community metabolic-ecology project, the top soil-directed PC1 loadings were glucuronate (0.146), fumarate (0.144), succinate (0.143), lysine (0.140) and cellobiose (0.139). The report describes all five as carbon-utilization pathways, although lysine is identified elsewhere in the same report as an amino-acid pathway. It reads the loadings as indicating substantially higher predicted carbon metabolic diversity in Soil than in Freshwater communities. ₁₇

Analytical implications

Ontology selection, category construction, marker definition, and identifier mapping should be recorded as model choices that can alter effect sizes, pathway counts, and inferred biological themes. ₁ A robust analysis should compare at least one curator-validated hierarchy with a simpler sensitivity schema, preserve multi-label assignments where biologically justified, and validate positive and negative marker calls against independent evidence. ₁

This approach supports ontology-and-category-schema-sensitivity as a methodological complement to multi-omics-integration and cross-tenant-data-bridging: a cross-modal conclusion is only as stable as the mappings and category definitions used to connect its modalities. ₁

Tensions

Within one project, whether Alm 2006-style TCS HK expansion appears depends on the grain of analysis. At UniRef50 the signal was absent and slightly negative (TCS HK 18% below cohort; negative observed z at every rank tested) despite adequate modelled power. ₈ At architecture level the qualitative recent-rank result held, with consumer concordance r = 0.67, but producer concordance was r = 0.09 and Alm 2006’s r ≈ 0.74 did not reproduce. ₈ The KO-level Cohen’s d range of 0.665–3.558 measures a different outcome: the separation of positive HGT controls from clean housekeeping on Sankoff/n_present. It is not a TCS HK paralog-expansion effect and should not be read as showing Alm 2006-style expansion at KO resolution. The corpus does not settle whether that amplification reflects aggregation or metric correction. ₈

The project records its UniRef50 baseline as d = 0.15 in one passage and d = 0.146 in another; both values are kept as written. ₈ The CAZyme null at UniRef50 also sits against literature evidence for CAZyme transfer at PUL or family level. The report treats this as a difference in substrate resolution rather than a contradiction. ₈

The discoveries digest and the IBD project report give different coverage figures for the two schemas in the same reversal. The digest says the regex identified 3 of 52 CD-up pathways in 7 themes, and the MetaCyc hierarchy assigned 262/409 pathways to at least one of 12 IBD themes. ₁ The project report gives 44/409 background pathways for the regex. For MetaCyc it gives 514/575 pathways categorized, with 90 % coverage of 575 HUMAnN3 pathways. ₂ The MetaCyc-side denominators differ, so the counts may describe different pathway sets. The corpus does not reconcile them, and both are kept as written. _1,2

Within the metal-ecology project, the aquatic leave-one-out result is described as significant (p = 0.031) and as showing that the result is not contingent on the aquatic category. The S2 table, however, gives BH-FDR q=0.061 for the same exclusion and marks it not FDR-significant. The same report calls aquatic inclusion a potential downward bias, although β is larger with aquatic included. ₁₄

Metal–osmotic overlap estimates show a 2.7x discrepancy between counter_ion_effects (39.8%) and metal_specificity (14.7%). The metal-specificity report attributes this to threshold choice and partly different organism sets, not to biology. ₁₀

Open Directions

• Re-run the 52 CD-up pathway analysis with the regex scheme, the curator-validated MetaCyc hierarchy, and a third ontology while holding the statistical model and multiple-testing procedure fixed; test which iron/heme-acquisition conclusion remains stable. ₁
• Build a benchmark of independently validated lanthanide-dependent methanol-dehydrogenase genomes and compare preferred-name, KO, product-description, and profile-based markers; quantify precision, recall, and disagreement across marker definitions. ₁
• Recompute the 32-species COG enrichment analysis with mutually exclusive versus multi-label category assignments; test whether the +0.34% LV enrichment and 76% consistency persist under each representation. ₁
• Reconcile MetaPhlAn3 and Kaiju species namespaces before cross-cohort projection, then compare LDA and CLR-plus-PCA GMM using held-out samples; determine whether the Kuehl E3 assignment is a namespace artifact or a reproducible ecotype signal. ₁
• Audit pathway and metabolite joins using explicit identifier dictionaries and collision tests, including leucine versus isoleucine; measure how many inferred pathway associations change after schema-to-value-space validation. ₁
• Extend the architecture census beyond the 376-KO focused subsets toward the 10,750-KO scope, for example by sampling the 833M-row IPS table; test whether the median 46 versus 1 architectures/KO contrast predicts transfer propensity atlas-wide. ₈
• Re-run the mycolic-acid test using only the 11 curated KOs versus the 715-KO pathway-derived set; test whether the stricter definition amplifies the signal, and at what cost in sample size. ₈
• Recompute acyl-carnitine and lipid-class theme enrichment with chain-length-restricted backgrounds instead of the 22-carnitine and 163-metabolite backgrounds; test whether the long-chain CD-up signal becomes significant. ₂
• Map the 13,944 “Other” genomes to a curated environment ontology and re-run the environment-embedding comparison; measure how many habitat associations change. ₅
• Re-classify fitness conditions with sub-typed stress classes (metal, osmotic, oxidative) and dual labels for edge cases such as zinc sulfate; test whether condition-specific conclusions change. _4,6
• Re-run the MicrobeAtlas niche-breadth PGLS with aquatic split into marine, freshwater and estuarine sub-categories from sample metadata. Test whether category heterogeneity inflates B_std for certain genera and whether the excl_aquatic q=0.061 result changes. ₁₄
• Re-run PHA synthase class assignment after mapping PF00561 -> Abhydrolase_1 and PF07167 -> PhaC_N; test how the 11,792 phaC clusters distribute across Class I-IV. ₁₅
• Recompute metal–osmotic overlap under both fit < -1 and fit < -1 AND |t| > 4 on a shared organism set; determine how much of the 39.8% versus 14.7% gap is threshold-driven. ₁₀
• The v1.8 species-level 12-theme test could evaluate only H. hathewayi, because M. gnavus, E. lenta and F. plautii had 0 CD-up pathways and E. coli and E. bolteae had 1-2. Pool additional IBD cohorts, or relax the per-species differential-abundance threshold as a sensitivity analysis, to raise species-level CD-up counts. Then test whether the H. hathewayi purine/pyrimidine and TMA/choline themes replicate and whether the other Tier-A species carry the cohort-level iron/heme theme. ₂