This concept examines how phylogenetic depth of gene-function acquisition, clade-specific distribution, ecological association, strain content, pathway capability, realized dependency, defense-system architecture, horizontal gene transfer, annotation quality, and uneven literature coverage jointly explain whether functions are widespread, specialized, exchanged, or stably inherited. 1
The primary evidence comes from Gene Function Ecological Agora, which analyzed a GTDB-r214 atlas of 18,989 species representatives, 13.74M rank × clade × KO producer/participation scores, and 17.07M Sankoff-parsimony gain events. 1 PaperBLAST Data Explorer — Literature Coverage Bias in Protein Sequence Space refines this framework by showing that studied-function distributions are shaped by extreme literature concentration: Homo sapiens accounts for 46.7% of PaperBLAST gene-paper records, the top five organisms account for 72.8%, and the organism-level Gini coefficient is 0.967. 2 Among 15,312 bacterial organisms, the top 100 captured 44.3% of bacterial literature, led by Mycobacterium tuberculosis H37Rv with 9,079 papers, Escherichia coli K-12 with 8,860, and Pseudomonas aeruginosa PAO1 with 5,928. 2
The new Truly Dark Genes — What Remains Unknown After Modern Annotation? refines this acquisition-depth framework by separating annotation lag from persistent functional darkness. Of 57,011 Fitness Browser dark genes, 33,105 (83.7%) were reclassified by Bakta v1.12.0, while 6,427 remained hypothetical in both pipelines and 17,479 lacked pangenome links. 3 Among linked genes, truly dark genes were shorter, less conserved, more taxonomically restricted, and more accessory than annotation-lag genes: median length was 121 versus 194 amino acids, core-genome fraction 43.1% versus 72.7%, ortholog presence 29.3% versus 63.7%, and median ortholog breadth 1 versus 4 organisms. 3 These differences support interpreting acquisition depth and ecological distribution jointly with annotation reliability, but the report does not establish that every truly dark gene was horizontally transferred or that its function is novel. 3
A complementary genotype-to-phenotype analysis integrated 727 genomes and 363 conditions. It refines this framework by showing that KO presence can support prediction of binary capability while failing to predict continuous growth kinetics. 4 The Metabolic Capability vs Dependency supports and refines this distinction at pathway scale: among 161 organism–pathway pairs spanning 7 model bacteria and 23 GapMind pathways, 57 (35.4%) were Active Dependency, 66 (41.0%) Latent Capability, 24 (14.9%) Incomplete but Important, and 14 (8.7%) Missing. 5 Thus, genomic pathway completeness is not equivalent to experimentally observed dependency.
The earlier Metabolic Capability vs Metabolic Dependency found 1,695 complete pathway-organism pairs from 48 organisms, classifying 267 (15.8%) as latent capabilities, 547 (32.3%) as intermediate, and 881 (51.9%) as active dependencies; pathway category strongly predicted class (χ²=163.6, df=4, p=2.5×10⁻³⁴), with carbon-source utilization more often latent than amino-acid biosynthesis (24.3% versus 6.5%). 6 The new analysis refines rather than replaces this result: its 7-organism, 161-pair design used a different mapping and classification scheme, and found that all 66 aggregate Latent Capability pairs became fitness-important under at least one condition type, especially nitrogen limitation, stress, or carbon limitation. 5 Because condition-specific median thresholds can cause reclassification by construction, this is evidence for condition dependence but not independent confirmation of every pathway’s biological dependency. 5
The Metagenome-Prioritized Phage Cocktails for Crohn’s Disease and IBD refines the same distinction by separating species-abundance-mediated from strain-content-mediated associations: Flavonifractor plautii had 0 FDR-passing strain-adaptation genes among 3,245 tested genes, while IBD-biased genes across 59 species were enriched for adaptation with OR = 1.38 and p = 2.4e-6. 7
The Lanthanide Methylotrophy Atlas — Summary adds a large marker-distribution test: across 293,059 genomes, xoxF occurred in 3,690 genomes versus 195 genomes with mxaF, for a global xoxF:mxaF ratio of 18.92 : 1. 8 This supports interpreting broad distribution jointly with phylogeny, environment, and annotation reliability: xoxF rates were high in Acidobacteriota and Gemmatimonadota, but Bakta-validated lanmodulin was restricted to three α-Proteobacterial families and marker calls differed substantially between eggNOG and Bakta. 8
Acquisition depth as a function signature
Sankoff parsimony, a method for inferring the minimum number and placement of character-state changes on a phylogenetic tree, assigned 17,073,194 gain events to recipient-rank depth bins. 1 Recent-to-ancient gain ratios differed by function class: crispr-cas had 58.7% recent and 2.4% ancient gains, a 24.5× ratio; two-component-system histidine kinases had 45.1% recent and 4.4% ancient gains, a 10.3× ratio; beta lactamases had 44.2% recent and 4.9% ancient gains, a 9.0× ratio; and clean tRNA-synthetase controls had 24.7% recent and 10.7% ancient gains, a 2.3× ratio. 1 These ratios describe phylogenetically attributed gains rather than direct donor transfers, since the per-class acquisition-depth distribution is computed over 17M gain events tagged by recipient-rank least-common-ancestor (LCA). 1 The attribution method is Sankoff parsimony with M22 recipient-rank gain attribution, run on 17M gain events over the 18,989-leaf GTDB tree, and it replaced a parent-rank dispersion permutation null that an NB08c diagnostic showed to be the wrong metric. 1
The full per-class distribution behind those ratios is reported as percentages of each class’s gain-event total across five bins — recent (genus-level gain), older_recent, mid, older, and ancient (phylum-or-above) — each row ending with that total: pos_crispr_cas 58.7%, 27.6%, 7.0%, 4.3% and 2.4% of 20,898 gains; pos_tcs_hk 45.1%, 32.9%, 10.6%, 7.0% and 4.4% of 518,106; pos_betalac 44.2%, 32.7%, 10.7%, 7.5% and 4.9% of 139,000; neg_ribosomal_strict 38.9%, 33.3%, 12.4%, 9.1% and 6.3% of 79,168; neg_rnap_core_strict 38.5%, 34.2%, 12.2%, 9.0% and 6.0% of 4,086; and neg_trna_synth_strict 24.7%, 31.8%, 17.5%, 15.2% and 10.7% of 15,060. 1
Two further control ratios from that same table belong beside the four quoted above, and they complicate the housekeeping baseline rather than confirming it. The strict ribosomal class gives 38.9 / 6.3 = 6.2, which the report itself labels mixed, as expected for the contaminated strict-K range, and the strict RNA-polymerase-core class gives 38.5 / 6.0 = 6.4. 1
The report summarizes the set by asserting that the recent/ancient ratio for HGT-active classes is 4–10× higher than for clean housekeeping, but its own clean housekeeping controls do not share one ratio: strict tRNA synthetases sit at 2.3 and strict RNA-polymerase core at 6.4. 1 The asserted 4–10× range therefore does not hold uniformly against the RNA-polymerase-core control, and it is not carried here as an established comparison. 1
Each class signature is read against published transfer work rather than against any internal transfer measurement. The beta lactamases profile, +44.2% recent and 4.9% ancient over n=139,000 gain events, is presented as the GTDB-scale instantiation of a documented pattern, aligned with Smillie et al. 2011 on within-phylum HGT in the human gut microbiome, Forsberg et al. 2012 on the cross-phylum AMR resistome with β-lactamase families spreading in soil and clinical lineages, and Bonomo 2017 on the specific TEM, CTX-M, NDM and OXA-48 families with cross-phylum spread. 1
The crispr-cas recent skew, 58.7% recent at a 24.5× recent-to-ancient ratio, is aligned with Metcalf et al. 2014’s documentation of cross-tree-of-life HGT for class-I CRISPR-Cas systems, and the report reads the strong recent skew as consistent with these systems being dispersed primarily via mobile-element-borne transfers in the recent past. 1 That alignment is a comparison with literature rather than an identification of donors in this substrate. 1
The two component system histidine kinases profile, 45.1% recent at a 10.3× ratio, is called consistent with the Alm 2006 finding that TCS HK abundance correlates with recent-LSE fraction, r = 0.74 in their 207-genome substrate — while the same passage states that the explicit reproduction at r ≈ 0.74 is deferred to Phase 4 (P4-D3) and that this milestone delivers only the underlying methodology that will feed it. 1
The tRNA-synthetase class is the most ancient-skewed of the six reported control classes, at 24.7% recent and 10.7% ancient for a 2.3× ratio, which the report reads as consistent with universal vertical inheritance and presents as extending Phillips et al. 2012’s dosage-balance result — bacterial ribosomal proteins showing low paralogy at a geometric mean of 1.02 paralogs against 1.63 for other universal genes — from low paralogy under dosage constraint to low recent acquisition under dosage constraint, which it calls two manifestations of the same selective regime. 1
The choice of parsimony over reconciliation is argued from precedent as well as tractability: the report cites the Mirkin et al. 2003 and Csurös 2010 precedent for parsimony as an HGT proxy at scales where full duplication-transfer-loss (DTL) reconciliation — AleRax, ALE, Notung — is computationally infeasible, and describes M22 as extracting gain locations from the parsimony reconstruction and assigning each to the rank-LCA of its descendant leaves, which it characterizes as the same level of phylogenetic-origin claim Alm 2006 made via lineage-specific-expansion analysis. 1
The limit of that attribution is stated plainly in the source and qualifies every depth ratio on this page: M22 is recipient-and-depth attribution, not donor-recipient flow. 1 Donor inference at deep ranks is constrained out by codon-amelioration timescales, post-amelioration composition signal degrading for events older than roughly 10–100 Myr, and by computational scale, per-family DTL reconciliation being infeasible at GTDB scale; M22 reports the rank-LCA of leaves under each gain event — the topological location where the gain landed, not who donated it. 1
The same depth binning yields control-class signatures. The cross-project digest records that M22 Sankoff attribution of 17M gain events on the GTDB-r214 species tree (gtdb), classified into five acquisition-depth bins (recent, older_recent, mid, older, ancient), produces clean signatures by control class: crispr-cas at a 24.5× recent/ancient ratio, read as an HGT-active signature; strict housekeeping KOs (RNA polymerase core, tRNA synthetases, ribosomal proteins) at an approximately 1× ratio, read as vertical inheritance; and KOs participating in both regulatory and metabolic pathways at 2× the Innovator-Exchange rate of pure regulatory or pure metabolic KOs. 9 The project report states the same contrast as the atlas centerpiece: CRISPR-Cas at a 24.5× ratio (HGT-active) against housekeeping-strict at approximately 1× (vertical inheritance). 1 This approximately 1× value sits beside the 2.3× ratio the same report gives for tRNA synthetases; the discrepancy is left unresolved here and is recorded under Tensions. 9,1 The synthesis figure for this contrast — Hero 3, the acquisition-depth function spectrum — plots per-control-class recent versus ancient percentages annotated with the recent-to-ancient ratio, and the report presents it as recovering the atlas centerpiece claim that the recent-to-ancient ratio is itself a function-class signature, CRISPR-Cas 24.5× against housekeeping ~1×. 1
Effect sizes in this project live on two scales that must not be conflated: the UniRef50 baseline effect was small at d=0.146, whereas the Phase 2 KO-level values were d=0.665, 0.753, 0.792, 1.905, 2.504 and 3.558, which by Cohen 1988 conventions (d=0.2 small, 0.5 medium, 0.8 large) run from medium-large to very large. 1 The report also separates roles, treating Sankoff as the primary atlas metric — a methodology element — and the KO-level d=0.665–3.558 range as the biological signal. 1 The d=0.146 → 0.665+ change is recorded as empirically observed, and an M14 close-reading reframed it as a metric correction rather than aggregation per se — Alm 2006 having worked at the single-domain, UniRef50-comparable level — while the report concedes that a clearer theoretical framing would help interpretation, so the mechanism behind the amplification is not settled. 1 On the cleanest housekeeping controls — strict tRNA synthetases plus RNAP core — the report records the M18 gate as a PASS and describes the change from d=0.146 at UniRef50 (NB08c, positives against all housekeeping) to d=0.665–3.558 at KO level (NB09c, best pairs) as a 4–25× amplification depending on positive class, with the substrate-hierarchy claim surviving at KO resolution. 1
The per-class signatures are treated as clean and informative independently of the pooled regulatory-versus-metabolic result: per the NB10b M22 attribution, CRISPR-Cas shows a 24.5× recent-to-ancient ratio, tRNA-synth 2.3×, and β-lactamase 9.0×, and the report states these class-level signatures hold even though the regulatory/metabolic pool-level asymmetry is small. 1
The KO-level amplification gate was designed as a falsification test and passed. The report records M18 passing at d=0.665-3.558 and reproducing at full atlas scale on the NB10 M21 sanity rail, with 6/6 strict pairs passing at 95% confidence-interval lower bounds above zero. 1 Sample size qualifies which pairs carry that weight: the CRISPR-Cas versus strict RNA-polymerase-core pair had n_pos=6 against n_neg=3 with a wide bootstrap 95% CI of [1.55, 11.58], so a pre-registered minimum of n_neg ≥ 20 (M23) made pairs against the n=3 RNAP-core set and the excluded strict ribosomal set informational only, leaving strict tRNA-synthetases at n=20 as the load-bearing housekeeping control. 1 The three passing pairs against that control, at d=0.65, 0.79 and 2.50, survive the constraint with stable estimates, which is why the housekeeping comparisons on this page rest on tRNA synthetases rather than on the smaller control sets. 1 The status of that minimum is described two ways inside the source and must be read with both: the primary-hypothesis-test passage calls n_neg ≥ 20 a pre-registered minimum sample size, while the methodology-change audit lists M23 among three post-hoc metric corrections made on already-collected data, so whether the constraint was fixed in advance or imposed afterwards is not settled here. 1
Each pair in that gate carries its own sample sizes and interval, and all six cleared Cohen’s d ≥ 0.3 with a 95% bootstrap CI lower bound above zero: CRISPR-Cas against strict tRNA synthetases at n_pos=6, n_neg=21, d=2.504, CI [1.31, 16.16]; CRISPR-Cas against strict RNAP core at 6 and 3, d=1.905, CI [1.55, 11.58]; β-lactamase against strict tRNA synthetases at 110 and 21, d=0.792, CI [0.68, 0.90]; β-lactamase against strict RNAP core at 110 and 3, d=0.753, CI [0.66, 0.86]; TCS histidine kinases against strict tRNA synthetases at 310 and 21, d=0.647, CI [0.54, 0.80]; and TCS histidine kinases against strict RNAP core at 310 and 3, d=0.665, CI [0.58, 0.82]. 1 This full-atlas run on the NB10 M21 sanity rail is what reproduces the M18 verdict first obtained on the NB09c 2.69M-row panel subset. 1
What reproduces is the verdict rather than the effect sizes, and the report’s own wording overstates the agreement. It claims the d values match the NB09c panel-only test almost exactly, attributing a small drift to slightly different sample composition, and elsewhere claims the two runs match within Cohen’s d ≤ 0.04 across all comparable pairs; but the NB09c best value it prints is d = 3.558 at n_pos = 6, n_neg = 20, against the corresponding full-atlas CRISPR-Cas versus strict tRNA-synthetase pair at d = 2.504 with n_pos = 6, n_neg = 21, and NB09c passed 6/9 strict pairs at best d = 3.56 where the full atlas passed 6/6 at best d = 2.50. 1 The pass verdict replicated; the near-exact effect-size claim is in tension with the figures the report itself prints, and that tension is recorded here rather than resolved by preferring either number. 1
These depth ratios lacked an independent quantitative anchor when they were reported. The report accepts that the CRISPR-Cas 24.5× and tRNA-synth 2.3× ratios align only qualitatively with the horizontal-gene-transfer literature and are not yet cross-validated against an independent dataset, and that all Phase 2 tests use atlas-derived scores; grounding in phenotype and ecology data — kescience-fitnessbrowser, bacdive, web-of-microbes, nmdc and mgnify — together with the explicit published-correlation reproduction was deferred to Phase 4, as were tests against independent experimental horizontal-transfer datasets. 1 The report’s status line records Phase 4 as complete with those deliverables closed, so the phenotype and ecology grounding was subsequently carried out; the Phase 2 statement that validation against independent experimental transfer measurements was deferred remains the record of the gap that grounding does not fill, so depth signatures here are atlas statistics with ecological grounding rather than independently measured transfer rates. 1 At the review stage that recorded it, the missing-validation-anchor problem was logged against M22 itself and deferred to the Phase 4 deliverables P4-D1 and P4-D3, which qualifies every interpretation resting on M22 gain attribution. 1
Multiple-testing control was hierarchical rather than atlas-wide: Bonferroni α/4 = 0.0125 for the four focal tests at the regulatory-versus-metabolic test and again at the mycolic-acid test, per-pair bootstrap confidence intervals at the amplification gate, and no family-wise correction across the 13.7M atlas scores, which the project frames as exploratory and descriptive by design. 1
The truly dark-gene analysis supports and qualifies this interpretation through independent genomic proxies. Truly dark genes had mean absolute GC deviation of 0.047 versus 0.038 for annotation-lag genes; strong deviation affected 9.2% versus 4.0%, and 12.0% were within 2 genes of a mobile genetic element. 3 Their ortholog presence was 4.2 times less likely than in annotation-lag genes (OR = 0.236), with only 3 of 65 dark-gene ortholog groups containing truly dark genes in cross-organism concordance data. 3 These observations support a hypothesis of recent acquisition or rapid evolution, but GC deviation is an imperfect HGT proxy and mobile-element proximity does not establish transfer direction or mechanism. 3 The cross-project digest qualifies the taxonomic-restriction half of this picture: 55.9% of dark gene ortholog groups are kingdom-level, present across multiple phyla. 9
The PHB analysis supports this interpretation with an independent, more conservative proxy: 311 potential phaC acquisition events occurred in families with less than 20% phaC prevalence and 278 potential losses in families with more than 80% prevalence. 10 Because these events were inferred from phylogenetic discordance and core/accessory status rather than a reconstructed phaC gene tree, they support a hypothesis of horizontal acquisition but do not establish donor identity or transfer direction. 10
The T4SS–CAZy study strengthens the hypothesis that recent or exchange-associated distributions can be organized into cross-phylum transfer events. Of 77 GT2 events, 65 spanned 2 phyla and 12 spanned ≥3 phyla (15.6%); Node_4915 contained 35 genes, 82.9% syntenic, and spanned WOR-3, Desulfobacterota, Patescibacteria, Bacteroidota, Firmicutes_A, Methanobacteriota, Bdellovibrionota, and Acidobacteriota. 11 The negative association between divergence and synteny supports a relationship between phylogenetic distance and neighborhood erosion, but the transfer interpretation remains observational and pending BLAST, null-baseline, and threshold validation. 11
The plant study refines the relationship between acquisition signals and retention depth: beneficial PGP clusters were more core-encoded than pathogenic clusters, but singleton marker clusters were strongly associated with transposase/integrase singletons and MGnify mobilome enrichment was higher at genus scale. 12 The defense analysis supports the interpretation that recent or mobile-associated functions can form coordinated acquisition units. Among 7,323 quality-filtered species, defense-system count correlated with prophage-cluster burden at Spearman ρ = 0.609 (p ≈ 0); after residualizing on log10 median genome size and phylum, partial ρ remained 0.301 (p = 1.6 × 10⁻¹⁵³). 13
Across 46,389 genome × condition pairs, KO × condition interactions increased mean binary-growth AUC from 0.620 to 0.653, with improvement in 80 of 106 held-out genera, but KO presence/absence and bulk genomic features produced negative R² for continuous µmax, lag, and max_A under genus-blocked holdout. AUC is area under the receiver-operating-characteristic curve. 4 The earlier capability analysis supports this binary-capability distinction: latent capability required mean absolute fitness t-score below 1.0 and fewer than 5% essential genes, whereas active dependency required mean absolute t-score above 2.0 or more than 20% essential genes. 6
The pathway analysis refines this result using a four-way classification combining GapMind completeness with gene-level essentiality, fitness breadth, and magnitude. Its composite importance score weighted essentiality at 40%, fitness breadth at 30%, and fitness magnitude at 30%; Active Dependencies had mean core gene completeness of 0.986 versus 0.975 for Latent Capabilities. 5
Truly dark genes provide a further phenotype-side qualification. Their essential fraction was 18.0% versus 13.4% for annotation-lag genes, and 34 of the top 100 prioritized candidates were essential. 3 However, short genes are harder both to annotate and to measure by transposon insertion, and strong fitness phenotypes can reflect polar effects on downstream genes; therefore essentiality and fitness signal narrow experimental hypotheses but do not by themselves establish gene function. 3
Null-model calibration and control-class behavior
Underneath the atlas scores sits a null-model calibration layer built and checked at pilot scale before any hypothesis test. The project formulated, calibrated and validated a multi-rank Producer × Participation null-model framework for clade-level innovation atlases on UniRef50-scale substrates, pairing a producer null drawn from clade-matched neutral families with a consumer null built by parent-rank dispersion permutation. 1
The producer null was checked against a positive control of 200 UniRef50 clusters with documented within-species paralog count ≥ 3 and cross-species presence in ≥ 5 pilot species — the natural_expansion class — which showed positive producer z-scores at all five ranks with effect size growing monotonically with rank: genus +0.13 σ, 95% CI [0.08, 0.18], n=975; family +0.19 σ, [0.13, 0.26], n=802; order +0.31 σ, [0.22, 0.39], n=654; class +0.50 σ, [0.38, 0.61], n=482; and phylum +0.55 σ, [0.43, 0.67], n=457. 1 The report reads this as validating that the clade-matched neutral-family null detects real paralog expansion above cohort baseline, without which all subsequent Phase 1A scoring would have rested on a null that cannot discriminate signal from noise. 1
The housekeeping negative controls ran in the opposite direction, and the report corrected its own criterion rather than its data: ribosomal proteins, tRNA synthetases and RNAP core subunits all showed negative producer z, −0.15 to −0.24 σ across ranks with 95% confidence intervals entirely below zero, which the report calls biologically correct because dosage-constrained genes carry fewer paralogs than typical genes at matched prevalence; it therefore records the pre-registered “near zero” criterion as wrong and the correct criterion as ”≤ 0 with CI not strongly positive”. 1 This is a post-hoc criterion change on an already-specified control, and it qualifies the housekeeping comparisons that the rest of this page rests on. 1
The consumer side carries the rank gradient. Pilot consumer z means over informative tuples were −3.77 at genus (n=423) and −3.95 at family (n=329), read as strong vertical inheritance; −4.10 at order (n=267), the peak of vertical clumping; and −1.36 at class (n=172), where the report reads clumping as weakening and a cross-class horizontal-transfer signal as emerging. 1 The M1 fix — stratifying the consumer-null parent rank per child rank instead of always anchoring at phylum — turned that pattern monotone: genus against a family parent −10.48, family against order −6.40, order against class −4.18, and class against phylum −1.84. 1 The earlier parent-phylum-only anchor had produced a flat-then-jump profile of roughly −4 σ at deep ranks and −1.36 at class, which the report reads as having masked a gradient that is in fact monotone. 1
The two-component-system histidine kinases (two component system histidine kinases) failed as a positive control at UniRef50 resolution, and the report treats that failure as informative rather than as missing power. Their UniRef50 clusters showed mean producer z ≈ −0.2 across all ranks, with only 4–11% of TCS HK UniRefs above zero: at genus 73 clusters scored, 8 (11%) positive, 2 above 2σ, mean producer z −0.10; at family 124, 8 (6%), 2, −0.16; at order 157, 7 (4%), 3, −0.19; at class 174, 9 (5%), 2, −0.23; and at phylum 186, 9 (5%), 2, −0.23. 1 A power calculation gives 81–99% detection probability at all ranks for an effect of d ≥ 0.3, so the report rules out underpowering as the explanation: at UniRef50 the TCS HK paralog signal is genuinely absent and slightly negative. 1 At that point the report explained the absence by substrate hierarchy — that the original Alm 2006 effect lives at family resolution, where a single histidine-kinase family aggregates many UniRef50 clusters that UniRef50-level scoring cannot recover — but a later Alm 2006 close-reading retracts that explanation, recording that the project had been misreading Alm 2006 and that Alm worked at single-domain level, approximately UniRef50 rather than “family”, so the substrate-hierarchy claim needed reformulation from “UniRef50 is too narrow, aggregate to KO” to “we used the wrong metric on the right substrate”. 1 The family-aggregation reading is therefore carried here as the earlier interpretation, superseded inside the source rather than endorsed; the diagnostic that settled it, NB08c Sankoff parsimony on the GTDB-r214 tree, found the order-rank anomaly to be a metric artifact, with Sankoff recovering the expected direction of positive HGT above negative housekeeping at p = 2.1×10⁻⁵. 1
The full Phase 1A census is reported per rank × control class as observed paralog mean, cohort mean, raw difference, percentage relative to cohort, and producer z. At phylum rank: natural_expansion n=457, 1.77 against a cohort mean of 1.27, +0.50 raw, +39.5%, z +0.55; pos_tcs_hk n=186, 1.07 against 1.30, −0.24, −18.4%, z −0.23; neg_ribosomal n=315, 1.04 against 1.18, −0.13, −11.4%, z −0.15; neg_rnap_core n=284, 1.01 against 1.20, −0.18, −15.2%, z −0.24; neg_trna_synth n=248, 1.02 against 1.20, −0.18, −15.0%, z −0.19; and pos_amr n=272, 1.17 against 1.34, −0.17, −12.6%, z −0.16. 1 For natural_expansion at the finer ranks: class n=482, 1.74 against 1.28, +0.46, +36.0%, z +0.50; order n=654, 1.56 against 1.28, +0.28, +22.3%, z +0.31; family n=802, 1.45 against 1.27, +0.18, +14.3%, z +0.19; and genus n=975, 1.31 against 1.20, +0.11, +9.3%, z +0.13. 1 The report’s reading is that at phylum rank natural_expansion UniRefs carry 39.5% more paralogs than the cohort baseline, 1.77 versus 1.27 paralogs per phylum; that the ribosomal, tRNA-synthetase and RNAP negative controls sit about 12–15% below cohort baseline, the dosage-constrained signature; and that TCS HK at UniRef50 sits 18% below cohort — the opposite direction from the Alm 2006 family-level expansion claim. 1
These calibrations held at full scale, with the positive control strengthening. At Phase 1B scale the natural_expansion class showed monotone-growing positive producer z — genus +0.13 σ (+9.3 % raw paralog above cohort), family +0.19 σ (+14.3 %), order +0.31 σ (+22.3 %), class +0.77 σ (+55.2 %) and phylum +0.89 σ (+64.5 %) — stronger than at pilot scale, while the ribosomal, tRNA-synthetase and RNAP-core negative controls sat about 12 % below cohort with z ≈ −0.15, the dosage-constrained signature. 1 The Phase 1B closure states the methodology framework as validated: multi-rank null-model construction works at full GTDB scale, the M1 rank-stratified parent ranks give a clean monotone gradient in clumping signal, and the producer null is responsive, with the natural_expansion class at +64.5 % paralog above cohort at phylum, p ≈ 10⁻⁸⁰. 1
The pilot and full-scale percentages must not be read as a contradiction. A review item comparing natural_expansion “+64.5% above cohort” against a recomputed “+39.4%” was identified as a phase-mixing error: the recomputation used the Phase 1A pilot means, observed 1.77 against cohort 1.27 for a ratio of +39.4%, while the compared statement reports the Phase 1B full-scale figure, observed 2.09 against cohort 1.27 for +64.5%; both are correctly computed on their respective scales, and the report describes the pilot-to-full-scale change, +39.5% → +64.5%, as methodology validation. 1
One further control check sits on the gain side rather than the paralog side: the natural_expansion class had the lowest median Sankoff / n_present value in its class table, 4.5, which the report scores as correctly lowest and attributes to broad conservation by selection. 1
Three diagnostics underwrite the scores rather than test hypotheses. The producer-z null-cohort diagnostic computes per (rank, prevalence_bin) cohort moments — mean and standard deviation of paralog count — from neutral-family UniRef50 clusters in the same cohort, and finds that all bins have ≥ 30 cohort members at all 5 ranks, which the report reads as validating that producer-z scoring against this null is statistically tractable and not pathologically thin in any bin. 1
The consumer-z null diagnostic, a permutation null on UniRef50 prevalence patterns, produced a unimodal null shape per rank with distinct shapes per rank, which the report reads as validating that consumer-z is comparable across ranks. 1 This single-rank version was superseded by the per-rank null introduced at NB02 Phase 1A v1.1 together with a Phase 1B M16 Sankoff diagnostic, and the source separately records it as superseded by the NB08c Sankoff parsimony diagnostic (M14), which uses tree-aware reconciliation; it is kept for provenance. 1
The third diagnostic is a pre-scoring sanity check on control-class paralog distributions: a boxplot of paralog count per UniRef50 within each control class on the 1K species × 1.2K UniRef50 pilot, in which natural_expansion shows the wide distribution expected of paralog-rich families, the AMR and CRISPR-Cas positive controls show modest paralog inflation, and the negative controls cluster tightly at 1 paralog, consistent with single-copy housekeeping. 1
Pre-registered hypothesis adjudication
Of four pre-registered hypotheses, two were confirmed: Mycobacteriaceae × mycolic-acid as Innovator-Isolated (NB12) and Cyanobacteria × PSII as Innovator-Exchange at class rank (NB16). 1 Both surviving hypotheses withstood D2 residualization and were grounded in their expected biomes at p < 10⁻¹¹. 1 The project’s own tally of the four currently-testable pre-registered hypotheses is 2 confirmed, 1 falsified at UniRef50, and 1 reframed at small effect size in the literature-predicted direction. 1
Of the project’s 16 formal hypothesis tests, 14 survived family-wise Bonferroni correction; the two that did not — Mycobacteriaceae × soil and the NB11 regulatory-versus-metabolic test — are reported as correct null results rather than failures. 1
That correction is reported family by family, with family-specific α: across the 16 formal tests grouped into three families, 14 of 16 survive Bonferroni and the 2 non-surviving tests are correctly null results. 1 In the pre-registered family, H1 Bacteroidota PUL Innovator-Exchange survives at raw p = 1×10⁻³ as a qualified pass, H2 Mycobacteriota mycolic Innovator-Isolated at family rank survives at raw p = 1×10⁻⁶, and H3 Cyanobacteria PSII Innovator-Exchange at class rank survives at raw p = 2.1×10⁻⁵. 1 The H2 raw value in this family-wise table, 1×10⁻⁶, does not agree with the per-test family-rank producer p-value of 2×10⁻⁶ recorded elsewhere in the same report and quoted above; this is a further reporting discrepancy inside the source, distinct from the summary-bound discrepancy discussed below, and it is not resolved here. 1 H4, the Alm 2006 r ≈ 0.74 reproduction, survives Bonferroni at raw p = 1×10⁻⁴³ but is recorded as NOT REPRODUCED, its p-value significant only because of large n while r = 0.10–0.29 is too small. 1
The biome-enrichment family is tabulated the same way: Cyanobacteriia × marine survives at raw p = 8.3×10⁻¹⁹ and Cyanobacteriia × photic aquatic at 1.5×10⁻⁵³; Mycobacteriaceae × host-pathogen survives at 3.6×10⁻⁴⁶ and Mycobacteriaceae × (soil OR host-pathogen) at 8.0×10⁻¹²; Bacteroidota × gut/rumen survives at 3.6×10⁻³⁶; and Mycobacteriaceae × soil does not survive, at raw p = 0.87 with 0.88× fold, which the report marks as correctly null and states as “soil NOT enriched”. 1
The third family is the six P4-D5 residualization replication tests at Bonferroni α = 0.00833, of which 5 of 6 survive: the NB12 Mycobacteriaceae × mycolic family-rank producer_z at d = +0.31, p = 2×10⁻⁶, and its family-rank consumer_z at d = −0.16, p = 1.6×10⁻⁵; the NB12 Mycobacteriales × mycolic order-rank producer_z at d = +0.29, p = 5×10⁻⁶; and the NB16 Cyanobacteriia × PSII class-rank producer_z at d = +1.50, p = 2.1×10⁻⁵ with its class-rank consumer_z at d = +0.63, p = 2.3×10⁻⁵. 1 The one that does not survive is the residualized NB11 regulatory-versus-metabolic producer_z, d = +0.06 at p = 0.69, which the report marks as correctly null — the original H1 criterion of d ≥ 0.3 falsified and the finding reframed in the small-effect complexity-hypothesis direction — rather than as a false positive. 1
Three of the methodology changes behind these numbers are post-hoc metric corrections on already-collected data: M14 Sankoff parsimony replacing parent-rank dispersion, M21 strict housekeeping, and the M23 minimum of n_neg ≥ 20. 1 Each was triggered by a specific failed diagnostic and is reported together with the diagnostic that prompted it, which makes the corrected criteria traceable rather than silent, but they remain corrections made after the data were in hand. 1
The multiple-testing burden at the pre-registration layer is correspondingly light, and the report says so: only 4 pre-registered hypotheses are tested — Bacteroidota PUL, Mycobacteriota mycolic, Cyanobacteria PSII, and the Alm 2006 r ≈ 0.74 reproduction — so with Bonferroni at α = 0.05 / 4 = 0.0125 the significant findings it lists there, NB12 mycolic-acid p<10⁻⁶, NB16 PSII class p<10⁻⁵ and the P4-D1 enrichments all p<10⁻¹¹, survive trivially. 1 Those summary bounds do not agree with the per-test values the same report records elsewhere and quoted above — NB12 producer p=2×10⁻⁶ at family rank and p=5×10⁻⁶ at order rank, and the NB16 class-rank producer_z at p=2.1×10⁻⁵ — so the summary bounds are carried here as a reporting discrepancy rather than as the tests’ own p-values, and the evidence does not show which figure the summary intends. 1
The scope of the pre-registered tally also differs between passages of the same report. Its own count of four currently-testable pre-registered hypotheses is 2 confirmed, 1 falsified at UniRef50 and 1 reframed at small effect size, whereas the four named in the multiple-testing passage include the non-reproduced Alm 2006 comparison, and the formal pre-registered-verdict file records five entries across atlas, ecology, phenotype and MGE columns: H1 a qualified pass, H2 supported, H3 supported at class rank, H4 not reproduced, and NB11 reframed. 1 Whether the reframed NB11 result belongs inside the pre-registered four or beside them as a bonus is not settled in the source, so the differing tallies are recorded here rather than merged. 1
H1, Bacteroidota → Innovator-Exchange on polysaccharide utilization loci (PUL), was recorded as a qualified pass: the deep-rank absolute-zero criterion was falsified while the Sankoff diagnostic recovered the direction, with d=0.15, 1.40× gut/rumen enrichment at p<10⁻³⁵, a saccharolytic, GH-rich BacDive anchor at n=577, 0% MGE-machinery, and LC=0.41 (twice the atlas reference, a phylum tendency with within-phylum heterogeneity). 1 The report maintains that this pre-registered hypothesis remains genuinely falsified at UniRef50 under its absolute-zero criterion, with later methodology QC governing quality control only. 1 The hypothesis-status table records the Bacteroidota PUL Innovator-Exchange prediction as falsified at the UniRef50 absolute-zero criterion in NB07, with the KO-level retest deferred rather than run. 1 A later review pass restates the same verdict and confines the M12 scope clarification to methodology quality control, so the falsification stands on its own criterion. 1 The hypothesis-status row for H1 pairs the falsified deep-rank absolute-zero criterion with the atlas-level diagnostic that recovers a small consumer-z signal, the d=0.15 UniRef50 estimate, the 1.40× gut/rumen enrichment at p<10⁻³⁵, the saccharolytic glycoside-hydrolase-rich BacDive anchor at n=577, and 0% MGE-machinery with ICE-mediated transfer cited per Sonnenburg 2010. 1
A review pass set that verdict beside literature pointing the same way at finer resolution: Yu et al. 2025, Microbiology Spectrum 13(4):e00917-25, doi:10.1128/spectrum.00917-25, finds vertical inheritance dominating Bacteroidota alginate-utilization-locus evolution at the Flagellimonas genus level. 1 The report reads this as complicating a simple “PUL = HGT” framing and as consistent with its own Phase 1B verdict, in which the deep-rank absolute-zero criterion was falsified and only a small consumer-z signal was recovered — that is, within-Bacteroidota PUL evolution is more vertical than horizontal at fine taxonomic resolution. 1
H2, Mycobacteriota → Innovator-Isolated on mycolic acid, was supported with producer d=+0.31 and consumer d=−0.19, 7.88× host-pathogen enrichment at p<10⁻⁴⁵, an aerobic-rod, Gram-positive, catalase BacDive anchor at n=318, and 0.57% MGE-machinery; LC=0.15 restricts the claim to sub-clades rather than to the whole family. 1 The hypothesis-status table records the Mycobacteriaceae × mycolic-acid Innovator-Isolated prediction as supported at family and order ranks, with producer d=0.31, consumer d=−0.19 and p<10⁻¹⁵, alongside a depth profile of 79.9% recent and 0% ancient gains; the 7.88× host-pathogen figure above is an ecological-enrichment result and does not stand in for that hypothesis-test outcome. 1 The hypothesis table names the M. tuberculosis and M. leprae complexes as the mycolic-positive species, but the project’s own later sub-clade recomputation does not agree with that identification, so which sub-clades carry the function is treated here as unsettled within the source and is recorded under Tensions. 1
The pre-registered statement of the mycolic-acid hypothesis is rank-explicit: it names GTDB f__Mycobacteriaceae / o__Mycobacteriales and predicts Innovator-Isolated — high Producer plus low Participation — on the mycolic-acid pathway KO set at deep ranks, family and order, at q < 0.0125 (α/4 Bonferroni). 1 At family rank, 582 Mycobacteriaceae clade × KO tuples gave producer d=+0.309 (p=2×10⁻⁶) and consumer d=−0.193 (p<10⁻¹⁵), classified Innovator-Isolated; at order rank, 614 Mycobacteriales tuples gave producer d=+0.288 (p=5×10⁻⁶) and consumer d=−0.285 (p<10⁻¹⁵), also classified Innovator-Isolated. 1
The pre-registered effect-size criterion is met unevenly across those two ranks, and the report says so: family-rank producer d=0.309 just exceeds the d ≥ 0.3 threshold while order-rank d=0.288 is just below it, with the hypothesis judged supported because both ranks carry the expected producer-high, consumer-low directions and Mann–Whitney p-values well below α/4 = 0.0125. 1 Two later review passes repeat the qualification that the strict d ≥ 0.3 criterion is met cleanly only at family rank, making the conservative reading that the Innovator-Isolated signature is strongest for Mycobacteriaceae and weaker but directionally consistent for Mycobacteriales. 1 The hypothesis-status table nonetheless records the finding as supported, with both criteria passing at family and order and d=0.31 with strong p and a striking depth signature. 1
The candidate KO definition qualifies how specific the result is to mycolic acid. The set comprises 715 KOs with KEGG pathway membership in ko00061 fatty-acid biosynthesis, ko00071 fatty-acid degradation, ko01040 biosynthesis of unsaturated fatty acids and ko00540 — which the report notes is actually LPS biosynthesis, included for plan-faithfulness — or membership in a curated list of 11 mycolic-acid-specific KOs: Pks13/K11778, KasA/K11533, KasB/K11534, InhA/K00208, the Ag85 mycolyltransferases K01205/K11211/K11212, the MmaA methyltransferases K20274/K11782, K00667 and K00507, all 11 of which are present in the GTDB substrate. 1 Because the pool is broader than strict mycolic acid and includes general fatty-acid biosynthesis, the report calls the test somewhat conservative and records a tradeoff rather than a result: a more strict definition might amplify the signal further, but at the cost of sample size, and neither the amplification nor its cost was tested. 1
The substrate was 459 Mycobacteriaceae species × 715 mycolic-acid candidate KOs under sankoff parsimony attribution, which the report describes as quantitative phylogenomic confirmation of a mechanistic prediction drawn from the literature: Marrakchi et al. 2014, Chemistry & Biology 21(1):67–85, doi:10.1016/j.chembiol.2013.11.011, PMID:24374164, documents mycolic acids as major and specific lipid components of the mycobacterial cell envelope, essential for the survival of members of the genus Mycobacterium. 1 “Confirmation” here is the report’s own reading of a correlative phylogenomic signature; no experimental test of mycolic-acid acquisition was performed. 1
H3, Cyanobacteria → Innovator-Exchange on PSII, was supported at class rank with producer d=+1.50 and consumer d=+0.70, 2.77× photic-aquatic enrichment at p<10⁻⁵², and LC=0.88; the hypothesis-status row records class rank as the supported rank because genus rank was too sparse, and its BacDive anchor was thin at n=4, with Cluster 0, marine plus sponge, dominant in its place. 1 As registered the prediction was a joint Broker-or-Open Innovator-Exchange with donor classes left undistinguished per M25, and the status table records it supported at class rank with producer d=1.50, consumer d=0.70 and p<10⁻³, together with a 7× lower ancient fraction than atlas-wide PSII that the report reads as a donor-origin signature. 1
Annotation-density residualization (D2) closed a long-standing pre-registration debt and supports the producer-side results specifically: producer_z was bias-immune at R²=0.000 while consumer_z carried small bias at R²=0.053, which changed no verdict. 1
The replication table is per-test. NB11’s regulatory-versus-metabolic producer_z stood at d=+0.06 before and +0.06 after residualization with p=0.69, a null either way. 1 The same table records NB11’s regulatory-versus-metabolic consumer_z unchanged at d=−0.21 before and after residualization at p<10⁻¹⁰, the NB12 Mycobacteriaceae × mycolic family-rank producer_z unchanged at d=+0.31 at p=2×10⁻⁶, that pair’s family-rank consumer_z attenuating from d=−0.19 to −0.16 at p=1.6×10⁻⁵, and the NB12 Mycobacteriales × mycolic order-rank producer_z unchanged at d=+0.29 at p=5×10⁻⁶. 1 The NB12 Mycobacteriales × mycolic order-rank consumer_z moved from d=−0.29 to −0.28 at p<10⁻¹⁰; the NB16 Cyanobacteriia × PSII class-rank producer_z held at d=+1.50 both before and after, at p=2.1×10⁻⁵; and the same pair’s class-rank consumer_z moved from d=+0.70 to +0.63 at p=2.3×10⁻⁵. 1
Across the 13.74M-row D2-residualized atlas and its 8 replicated hypothesis tests, all hypothesis verdicts survive annotation-density residualization: the largest attenuation is the NB12 family-rank consumer side, d=−0.19 → −0.16, roughly 16% relative attenuation and still significant, while NB11 reframed, NB12 supported and NB16 supported at class rank each preserve direction and significance. 1
Those verdicts postdate an earlier state of the same report, kept here as an audit trail rather than as a live status. At the NB10/NB10b milestone the pre-registered hypotheses had not yet been tested: NB10 and NB10b delivered the atlas methodology and substrate, while NB11 (regulatory-vs-metabolic Tier-1) and NB12 (Mycobacteriota mycolic-acid) were named as the notebooks that would test the actual research-plan hypotheses, with the report stating that the atlas-as-deliverable interpretation is independent of those tests’ outcomes. 1 The NB11 plan specified categorizing KOs by KEGG BRITE B-level as regulatory — 09120 Genetic Information Processing plus 09130 Environmental Information Processing — or metabolic, 09100 Metabolism, and comparing per-class acquisition-depth profiles and Producer × Participation populations between the two categories at FWER<0.05; that BRITE-level definition sits beside the ko-number definition the executed test used, recorded below, and the source does not reconcile the two. 1 The NB12 plan was to filter the atlas to mycolic-acid pathway KOs — KEGG ko00540 plus FAS-II plus mycolyl transferase — and to Mycobacteriota clades or the family f__Mycobacteriaceae, testing Producer × Participation = Innovator-Isolated at q < 0.0125, the Bonferroni threshold for four focal tests. 1
Two uncertainty items were open at that same milestone. Phylogenetic non-independence was only partially addressed: KO aggregation collapses intra-Pfam-family UniRef duplication and M19 cluster-bootstrap confidence intervals handle function-class-level Cohen’s d, but per-(clade × KO) tuple-level confidence intervals would have required additional bootstrap work and were deferred. 1 Annotation-density residualization (D2) was at that point not yet implemented at atlas level — per-genome annotated_fraction was computed and stored but not regressed out as a nuisance covariate per rank × KO score, with P4-D5 named as the deliverable that closes it — which it later did, yielding the producer-side R²=0.000 and consumer-side R²=0.053 results recorded above. 1
Clade-specific depth signatures
The Mycobacteriaceae mycolic-acid gain set contained 53,916 gains: 79.87% recent, 16.72% older-recent, 3.41% mid, 0.00% older, and 0.00% ancient, compared with 48.79%, 31.81%, 9.60%, 6.12%, and 3.68% across all 1.47M mycolic-acid gains atlas-wide; both rows are percentages of gain events, not gain counts. 1 A Mycobacteriaceae sub-clade analysis found producer d = +0.394 for the mycolic-positive sub-clade comprising 10 of 13 genera, versus family-rank d = +0.309 and mycolic-low sub-clade d = +0.211; consumer d was +0.006 for the positive sub-clade, and leaf_consistency was 0.15. 1 The cross-project digest records the same structure at family rank: producer-side Cohen’s d=0.31 for Mycobacteriaceae × mycolic-acid supported H2 Innovator-Isolated, while median leaf consistency for mycolic-acid KOs in the family was LC=0.15, lower than the atlas reference of 0.20. 9 Its sub-clade-restricted recomputation kept the mycolic-positive sub-clade — 10 of 13 genera with mean LC ≥ 0.5 — giving producer d=+0.394 against family-rank d=+0.309. 9 The project frames this as a modest amplification, producer d=0.39 versus family-rank d=0.31. 1 Because of the low leaf consistency, the supported claim narrows from “Mycobacteriaceae innovate mycolic-acid” to sub-clades of Mycobacteriaceae being mycolic-acid Innovator-Isolated, with the family-rank d=0.31 effect treated as a population-mixture average. 1 The report states the refined claim as “the mycolate-producing sub-clade of Mycobacteriaceae (10 of 13 genera) is mycolic-acid Innovator-Isolated, producer d=0.39 at genus rank”, adding that the consumer-side Innovator-Isolated component is weaker than the original family-rank framing implied. 1 It reads the family-rank effect as a population-mixture average across mycolic-positive (Innovator-Isolated) and mycolic-negative members, and as consistent with the 7.88× host-pathogen biome enrichment, which it says is also driven primarily by the mycolic-positive pathogenic sub-clades. 1 The source does not settle which sub-clades those are, and its recomputation’s consumer-side signal does not persist at genus rank, which weakens the Innovator-Isolated reading; both points are recorded under Tensions rather than resolved. 1
The report emphasizes zero older-class and zero ancient gains for the Mycobacteriaceae mycolic-acid acquisitions, placing them at genus rank (recent, 79.9%) or family rank (older_recent, 16.7%), against an atlas-wide mycolic-acid distribution that spans the full depth range with 9.8% in deeper bins (older + ancient combined). 1 Its prose overstates this twice — “essentially all” and then “every single one of 53,916 gain events is recent or older-recent” — which contradicts its own table entry of 3.41% mid gains; the mid fraction is retained here and the stronger phrasing is not. 1
Within the same 582 family-rank tuples, the Producer × Participation quadrants were 67 Innovator-Isolated at 11.51% of the subset against an atlas-wide reference of 5.85% (803K / 13.7M), 497 Stable at 85.40% against 86.10%, 17 Sink/Broker-Exchange at 2.92% against 5.40%, and 1 Innovator-Exchange at 0.17% against 0.36%. 1 The report reads this as 2.0× Innovator-Isolated enrichment over the atlas baseline — substantially more lineage-specific paralog expansion with within-clade containment than a typical clade × KO tuple — with Innovator-Exchange essentially absent at 1 tuple against the ≥ 50K typical of HGT-active classes. 1
A family-rank atlas heatmap, figure N7, extends the same quadrant view across clades: it plots the percentage of family × KO tuples that are Innovator-* per top-20 phylum × control class, and shows elevated Innovator-* fractions for antimicrobial-resistance (AMR) and CRISPR-Cas KOs in Cyanobacteriota and Actinomycetota, with housekeeping classes low across the board. 1
Cyanobacteria PSII gains contained 2.05% ancient events versus 14.90% atlas-wide; the class-level sample contained n = 21 PSII KOs, while genus, family, and order results were STABLE. 1 The PSII verdict is rank-dependent: genus rank n=2,350 gave d=0.08 and a STABLE call, class rank n=21 gave producer d=+1.50 with consumer d=+0.70 and an Innovator-Exchange call, and phylum rank gave d=1.60 and Innovator-Isolated; the report treats class rank as biologically appropriate, the genus null as expected under that framing, and the phylum reversal as a substrate artifact from insufficient reference consumer data. 1 ENIGMA identified eight metabolic guilds across 20 taxonomic orders among 123 strains, with genome sizes from 2.6 to 11.4 Mb. 4
The sub-class STABLE calls are fully specified. At genus rank, 137 pooled Cyanobacteria genera and 2,350 KOs gave producer d=+0.08 and consumer d=−0.53 with both Mann–Whitney p values at 1.00; at family rank, 42 families and 705 KOs gave producer d=+0.20 and consumer d=−1.28 with p=0.97 and 1.00; at order rank, 19 orders and 301 KOs gave producer d=+0.19 and consumer d=−1.42 with p=0.99 and 1.00. 1 These support the rank-dependence reading: the same clade × function pair is scored Stable at three consecutive ranks and Innovator-Exchange at class rank, so an acquisition-depth verdict is a property of a rank-and-clade pairing rather than of the clade alone. 1
The report’s own framing of that pattern is that PSII is a class-level innovation rather than a within-genus one, citing Cardona, Sánchez-Baracaldo, Rutherford & Larkum 2019, Geobiology 17(2):127–150, doi:10.1111/gbi.12322, PMID:30548831, for the statement that “the evolution of PSII predates the diversification of Cyanobacteria”. 1 On that framing the genus, family and order STABLE verdicts are biologically expected: within Cyanobacteria, PSII is a near-universal feature with deep ancestry, so within-genus tests ask whether genera are innovating a function they already share, and the d ≈ 0.08–0.20 producer signals at finer ranks reflect that expected near-uniformity. 1 The report preserves the matching caveat explicitly: the rank-dependence pattern shows PSII Innovator-Exchange to be specifically a class-rank phenomenon rather than a generic Cyanobacteria-wide HGT signature at every taxonomic resolution, so the supported claim is narrower than “Cyanobacteria innovate PSII at all ranks” and is instead that Cyanobacteriia as a class exchange PSII KOs more than reference clades do, anchored to the photic-aquatic biome at 2.77× enrichment, p<10⁻⁵², with a depth signature of 2.05% ancient against 14.9% atlas-wide. 1
The PSII depth profiles behind the ancient-gain contrast are, for the 1,026 Cyanobacteria PSII gains, 32.26% recent, 36.26% older_recent, 12.96% mid, 16.47% older and 2.05% ancient, against 27.37%, 30.58%, 12.85%, 14.31% and 14.90% for all 11K PSII gains atlas-wide. 1 The report’s phrase “7× fewer ancient PSII gains” compares these ancient-gain percentages and not absolute ancient-gain counts, so it is read here strictly as a percentage contrast. 1
The report interprets the depleted ancient fraction as a donor-origin signature — PSII originated in or near the Cyanobacteria lineage, so Cyanobacteria does not itself receive ancient cross-phylum PSII transfers — and reads the atlas-wide 14.9% ancient fraction as PSII gains in non-Cyanobacterial phyla, Chloroflexota and certain Proteobacteria, that are ancient for the recipient. 1 This interpretation is layered on a presence-pattern statistic: the pre-registered test identified no donors, so the direction of flow remains a hypothesis. 1 The external anchor cited for it, Cardona et al. 2018, is phylogenomic and Bayesian relaxed-molecular-clock evidence that a water-oxidizing photosystem appeared in the early Archean about 1 Gyr before the most recent common ancestor of described Cyanobacteria and before the diversification of anoxygenic photosynthetic bacteria. 1 A later review pass adds Ślesak 2024 alongside Cardona 2018 as jointly consistent with the donor-origin reading of NB16’s 7× lower ancient percentage in Cyanobacteria; both remain supporting literature rather than a direct donor-composition test. 1
A further Cyanobacteria × PSII hypothesis was left open: the Broker reading, planned as a Phase 3 architectural deep-dive at genus rank, is recorded as not yet tested, so the corpus carries no result for it. 1
The pathway analysis supports the importance of strain and clade content: all 10 target species formed metabolic clusters, with silhouette scores from 0.35 for PALSA-747 sp. to 0.89 for Salmonella enterica. Cluster–isolation-environment associations were significant for Salmonella enterica (χ²=1570.2, df=25, p<0.0001) and Phenylobacterium sp. (χ²=12.2, df=1, p=0.0005), but not for Stutzerimonas, Alteromonas, Pelagibacter, or Prochlorococcus. 6 In the 2,810-species analysis, binary completeness profiles across 80 pathways yielded a median of 4 metabolic ecotypes per species and a maximum of 8 in Alistipes onderdonkii and Barnesiella intestinihominis. 5
The truly dark-gene result refines clade-specific interpretations by showing that the residual unknown set is unevenly distributed: Methanococcus strains S2 and JJ account for 55% of truly dark genes, and the fraction resisting Bakta ranges from 4–96% across organisms, with archaeal organisms at the high end. 3 This concentration is compatible with lineage-specific novelty or annotation limitations, but the 17,479-gene pangenome-linkage gap prevents equivalent assessment across all dark genes. 3
Plant communities were dominated by Rhizobium, Mesorhizobium, and Bradyrhizobium in roots, Sphingomonas and Methylobacterium in the phyllosphere, and Pseudomonas_E across all compartments. 12 Only 5 of 17 testable species passed Bonferroni-corrected within-species tests; strongest evidence occurred for Xanthomonas campestris toward Brassica hosts, with 46/47 genomes in one subclade and corrected p = 3.3×10⁻¹¹, and Xanthomonas vasicola toward Zea mays, with 47/52 genomes in one subclade and corrected p = 1.7×10⁻¹⁰. 12
The PGP analysis supports clade- and environment-dependent function profiles: among 1,039 soil/rhizosphere species versus 10,233 other species, acdS, pqqC, and hcnC were enriched, with acdS prevalence 15.8% versus 2.6% (OR = 7.02, q = 5.1e-62), pqqC 43.8% versus 21.2% (OR = 2.90, q = 2.8e-53), and hcnC 11.3% versus 6.4% (OR = 1.85, q = 6.1e-08); nifH was depleted, 12.9% versus 19.7% (OR = 0.60, q = 5.5e-08). 14
Figure: environmental enrichment of PGP genes — prevalence comparison and log2 odds ratios (figures/env_enrichment_barplot.png). 14
Three of those five focal genes were significantly enriched in soil at Benjamini–Hochberg false-discovery-rate (BH-FDR) q < 0.05 in the comparison of 1,039 soil/rhizosphere species against 10,233 species from other environments; ipdC was not, at 1.5% versus 1.9%, OR = 0.79, q = 0.47. 14
The acdS enrichment is the most robust of the set, and it supports treating the association as stable under coarse taxonomic adjustment: the soil/rhizosphere versus other-environment contrast of 15.8% against 2.6%, OR = 7.02, survives phylum-level fixed effects in logistic regression at OR = 6.98, p = 4.8e-61, and holds in a strict rhizosphere-only sensitivity analysis using only genomes with “rhizosphere” or “root nodule” in their isolation source, OR = 10.6, q = 7.6e-38. 14 Two qualifications attach. Phylum-level fixed effects do not exclude finer-scale shared-ancestry confounding, so this is robustness to coarse taxonomic structure rather than a demonstration that the enrichment is ecological rather than phylogenetic. 14 And the report’s accompanying phrasing that ACC deaminase is “7× more prevalent in soil/rhizosphere species than in other environments” sits beside its own prevalence pair of 15.8% versus 2.6% and its odds ratio of 7.02; odds ratios and prevalence ratios are not interchangeable, so that multiplier is carried here as the source’s wording rather than endorsed, and no replacement figure is computed. 14
The phylum-stratified pass qualifies the nifH depletion as non-universal: Bacillota_A is an exception with nifH enriched in soil at OR = ∞, q = 2.5e-4, which the report attributes to soil-specific anaerobic diazotrophs such as clostridia. 14
Prevalence across the PGP set is highly uneven, which bears on how much weight any single marker’s distribution can carry. pqqd, at 55.5% core and present in 11,040 species, is the most widespread PGP gene and the most promiscuous pqq gene, appearing as a single-gene orphan in many species lacking other pqq genes; excluding pqqD, pqqc is the most prevalent focal trait at 3,028 species, 26.8% of species with any PGP gene, and ipdc is the rarest at 214 species, 1.9% — a count the report distinguishes explicitly from its 226 ipdC gene clusters. 14
pqqD is also a core-fraction outlier that qualifies the core-encoded reading of PGP traits: at 55.5% core against 63–81% for the other PGP genes, and with the highest singleton fraction at 27.5%, the report reads it as occasionally spreading horizontally as a standalone gene, while the functional pqq operon taken as a unit — pqqB–pqqC — is predominantly core. 14
The acdS result is framed as agreeing with Nascimento et al. 2014, which used phylogenetic analysis to show that acdS is predominantly vertically inherited despite some HGT events, including documented transfer in some Proteobacteria. 14 The project’s rejection of its own horizontal-spread hypothesis therefore refines rather than removes the transfer channel: predominance of vertical inheritance does not establish that transfer never occurs. 14
Gene-by-environment interaction cuts the other way for ipdC, and contradicts any single global association between pathway completeness and carriage: within soil/rhizosphere species (n = 1,039) the tryptophan-completeness-to-ipdC association reverses to OR = 0.30, p = 0.02, so trp-complete soil species are less likely to carry ipdC, while in non-soil species (n = 10,233) the positive association holds strongly at OR = 3.56, p = 7.7e-13. 14
The report’s proposed explanation for that reversal — that soil-associated plant-growth-promoting bacteria obtain tryptophan exogenously from plant root exudates, relaxing selective pressure to maintain autonomous trp biosynthesis while retaining ipdC for indole-3-acetic-acid production from plant-supplied substrate — is offered as a possibility and is not tested here. 14
Within-clade consistency and ecological distribution
Mean leaf_consistency was 0.34 for recent gains and 0.20 for ancient gains; hypothesis-specific values were 0.88 for Cyanobacteria × PSII, 0.41 for Bacteroidota × PUL, and 0.15 for Mycobacteriaceae × mycolic acid, against an atlas reference of 0.20. 1 These values, together with ENIGMA guilds that tracked but did not reproduce taxonomy, support caution that recent or clade-associated signals can conceal within-clade heterogeneity. 1,4 The cross-project digest treats the leaf-consistency layer as an independent validation of the depth binning, because the per-rank × clade × KO fraction of species carrying the KO decreases monotonically across depth bins, from 0.34 in the recent bin to 0.20 in the ancient bin. 9 Figure panel S7 plots the per-hypothesis leaf-consistency distribution — PSII LC=0.88 as a canonical class-defining innovation, PUL 0.41 as a phylum tendency, and mycolic 0.15 below the atlas reference 0.20, revealing within-Mycobacteriaceae heterogeneity — alongside the H3-B control-class signature plane and the S8 atlas confidence ridge. 1
That layer has an explicit substrate. A 13.75M-row lookup gives, per rank × clade × KO, the number of species carrying the KO over the total clade species. 1 It is joined to the 17.07M M22 gain events together with n_leaves_under, whose median is 10 and 75th percentile 39, yielding leaf_consistency at 97.8% coverage and the monotonic decline from a mean of 0.34 in the recent bin to 0.20 in the ancient bin. 1
The same layer is described elsewhere in the report against different substrate and event counts, and both descriptions are carried here as given rather than reconciled. There, the full 28M-row species-level KO presence table — the substrate the atlas was built from — was read to compute, per rank × clade × KO, species-with-KO over total-species-in-clade, then vectorized-merged to 16.7M gain events at 97.8% coverage, giving per-depth-bin values of recent 0.34, older_recent 0.30, mid 0.26, older 0.23 and ancient 0.20. 1 The same pass gives per-control-class medians that support the housekeeping-versus-HGT-active framework directly: strict RNAP core 0.97, strict tRNA synthetases 0.94 and strict ribosomal proteins 0.92 at the top, against CRISPR-Cas 0.29, TCS histidine kinases 0.13 and AMR 0.11 at the bottom. 1
Leaf consistency also separates control classes directly, and because it was not used in any pre-registered hypothesis test the report calls it its cleanest independent validation of the housekeeping-versus-HGT-active framework. 1 Strict housekeeping classes occupy the low-recent, high-consistency corner — RNA-polymerase core strict at 0.97, strict tRNA synthetases at 0.94, strict ribosomal proteins at 0.92 — with KOs present in nearly all clade members and low recent-gain rates, while the high-recent, low-consistency corner holds patchy distributions with active recent gains: CRISPR-Cas at 0.29, AMR at 0.11, and two-component-system histidine kinases at 0.13. 1 The ancient-patchy corner of the same plane is sparsely populated, holding only paralog-inflated loose-housekeeping classes and little else. 1
The per-hypothesis leaf-consistency rows carry their own sample sizes and ratios. Cyanobacteriia × PSII, focal n=1,005, had median LC 0.88 against the atlas reference of 0.20, 4.4× higher, which the report reads as PSII KOs present in 88% of Cyanobacteriia species and therefore a canonical class-defining innovation on which the class-rank d=1.50 verdict sits as a near-uniform population. 1 Bacteroidota × PUL, focal n=6,070, had median LC 0.41 against 0.20, 2.0× higher, read as PUL KOs in 41% of Bacteroidota species — a phylum tendency with substantial within-phylum heterogeneity. 1 Mycobacteriaceae × mycolic, focal n=33,643, had median LC 0.15 against 0.20, 0.75× — lower than the reference — meaning mycolic gains land in clades where the mycolic KO is present in only 15% of family members, which the report reads as mycolic-acid-positive species forming a sub-clade within Mycobacteriaceae. 1
Figure panel S8, the atlas confidence ridge, plots that decline per depth bin at recent 0.34, older_recent 0.30, mid 0.26, older 0.23 and ancient 0.20; the report reads it as independent validation of the M22 acquisition-depth signal, with deeper gains diversifying and subsequently being lost while recent gains land in clades where the KO is propagated. 1
The ridge form of that panel recovers the same signal as the summary table, a decline from recent 0.34 to ancient 0.20, but as full distributions rather than summary statistics: recent gains have a long right tail toward 1.0, clades where the KO is fully fixed, while ancient gains compress toward 0, the KO retained in a small fraction of phylum members after diversification plus loss. 1 The report presents the panel as the project’s atlas-wide signal-confidence map — attribution is cleanest for recent gains and uncertainty grows toward ancient ones, where the underlying signal is diluted by subsequent within-clade losses — so by the report’s own framing the ancient-bin claims on this page are the least reliable of the depth signatures. 1
Bootstrap-based per-event uncertainty on the gain attribution was not computed, although uncertainty proxies were. It was deferred at roughly 50 hr of compute, with n_leaves_under shipped as the primary proxy. 1 A later status entry repeats the deferral as item C9, again at ~50hr compute, and describes leaf_consistency as one uncertainty proxy over n=16.7M gain events with per-depth-bin median ranging 0.05 ancient to 0.26 recent; those medians stand beside the per-bin means of 0.34 recent to 0.20 ancient quoted above and the source does not reconcile the two sets of values. 1
Latent capability rates ranged from 0–31.6% across organisms, with a mean of 15.0%, and correlated positively with pangenome openness across 22 species clades (Spearman ρ = 0.69, p = 0.0004, n = 22 clades). 6 The pathway analysis supports this association at larger scale: variable pathway count correlated with pangenome openness at raw Spearman rho=0.327, p=7.2e-71, and partial rho=0.530, p=2.83e-203 after controlling for genome count; the signal was positive in 13 of 18 genera and significant in 5: Clostridium, Eubacterium, Mesorhizobium, Pseudomonas, and Streptomyces. 5
The PGP result contrasts with these positive openness associations: PGP richness correlated negatively with openness (Spearman ρ = −0.195, p = 2.0e-97, n = 11,272 species with ≥2 genomes), and the mean PGP accessory fraction was 29.7% versus 53.2% genome-wide. 14 This is not a direct contradiction because the analyses use different units and traits.
The truly dark-gene result adds a related qualification: most persistent hypothetical genes are not functionally empty, because 79.4% have UniRef50 links and 84.7% have database cross-references, yet only 4.0% have Pfam hits and 4.6% have KEGG KOs. 3 Of 6,427 truly dark genes, 246 (3.8%) had zero annotation clues, while 1,603 (24.9%) had phenotype or module signal without functional annotation and 711 (11.1%) had partial Pfam, COG, or eggNOG function. 3 Thus, pangenome absence, narrow orthology, and accessory status can identify candidates for functional discovery, but sequence recognition alone does not establish realized capability.
The plant study found co-occurring plant-associated genus pairs slightly less complementary than random pairs (Cohen’s d ≈ −0.4, permutation p < 0.001; prevalence-weighted d = −0.39). 12 BacDive metabolite-utilization data agreed with GapMind predictions at 83.1% consistency, but 66 of 69 NMDC co-occurring genera were dual-or-mixed and the core-level GapMind completeness score was 0% across compartments. 12 Thus, accessory pathway variation and co-occurrence do not demonstrate metabolic complementarity or community dependency.
The focal clades were ecologically consistent: Cyanobacteriia showed 2.77× photic-aquatic enrichment with p < 10⁻⁵², Mycobacteriaceae 7.88× host-pathogen enrichment with p < 10⁻⁴⁵, and Bacteroidota 1.40× gut/rumen enrichment with p < 10⁻³⁵; Mycobacteriaceae was not soil-enriched (0.88×, p = 0.87). 1 BacDive profiles found 95% rod-shaped, 99% non-motile, and 89.4% aerobic-leaning records among 318 Mycobacteriaceae species; anaerobic Bacteroidota phenotypes were 33.2% versus 22.0% atlas-wide among 577 species. 1 The summary passage reporting these biome enrichments does not itself state a sample-level denominator, though the underlying enrichment table does, at n=309 Cyanobacteriia and n=459 Mycobacteriaceae. 1 BacDive phenotype anchors confirmed the mycobacterial aerobic-rod and Bacteroidota saccharolytic profiles, and AlphaEarth environment clustering recovered ecological structure with the focal clades concentrated in expected environment clusters. 1 The cross-project digest supports this triangulation: for the mycolic, PSII and Bacteroidota PUL findings, three independent measurement substrates produced convergent evidence — atlas Cohen’s d on Sankoff-derived scores, per-clade × biome Fisher’s enrichment at p<10⁻¹¹, and a BacDive metabolic-phenotype anchor where coverage exists. 9 The project states the grounding as counts: all three pre-registered atlas findings — NB12 mycolic-acid Innovator-Isolated, NB16 PSII Innovator-Exchange and the Phase 1B Bacteroidota PUL result — ground cleanly in expected biomes at p < 10⁻¹¹, with two of three also confirmed by BacDive phenotype anchors and three of three by AlphaEarth environment-cluster concentration. 1
The enrichment table itself is clade-explicit. Cyanobacteriia (n=309) were 63.8% photic-aquatic, marine plus freshwater, against a background of 23.0%, a 2.77× enrichment at p < 10⁻⁵², and on the narrower marine category 35.3% against 15.1%, 2.33× at p < 10⁻¹⁸. 1 Mycobacteriaceae (n=459) were 16.6% host-pathogen against 2.1%, a 7.88× enrichment at p < 10⁻⁴⁵, while the narrower soil category was not enriched at 12.9% against 14.5%, 0.88× at p = 0.87. 1 A combined soil-or-host-pathogen row for the same Mycobacteriaceae set (n=459) gives 29.0% against 16.4%, a 1.76× enrichment at p < 10⁻¹¹, and the Bacteroidota row (n=2,581) gives 36.1% gut/rumen against 25.8%, 1.40× at p < 10⁻³⁵. 1 The report reads the pair of Mycobacteriaceae categories as locating the mycolic-acid Innovator-Isolated signature in host-pathogen rather than soil mycobacteria, treating the 7.88× host-pathogen fold-enrichment as the clean anchor and the soil enrichment as a null. 1
The report also limits what these anchors can show. The enrichment results — Mycobacteriaceae 7.88× host-pathogen, Cyanobacteriia 2.77× photic aquatic, Bacteroidota 1.40× gut/rumen — demonstrate statistical association between focal clades and expected biomes and do not establish causal mechanisms: whether environment drives function-class innovation, function-class innovation drives niche occupancy, or both are driven by a third factor is untested. 1 The same limit applies to the triangulation argument used throughout this page: the three-substrate convergence framework of atlas, ecology and phenotype provides convergent evidence — three independent measurement substrates pointing to the same biology — but does not establish causation. 1
Recent-gain density also varies across the project’s environment clusters, reported as per-species mean recent-gain counts by regulatory, metabolic and mixed KO category: cluster 5 (mouse-gut + marine + acid mine drainage) at 3,178 / 10,891 / 6,418; cluster 8 (marine + human-gut + soil) at 3,091 / 10,063 / 6,008; cluster 1 (gut + sludge) at 2,652 / 8,081 / 5,780; cluster 3 (gut) at 2,627 / 8,173 / 5,446; cluster 6 (hypersaline + Microcystis) at 490 / 1,475 / 1,100; and cluster 9 (soil + grassland) at 1,303 / 5,314 / 3,087. 1 The report reads the mixed-biome clusters 5 and 8 as carrying the highest recent-gain density and the hypersaline/extremophile cluster 6 as carrying the lowest. 1 These are per-cluster mean counts together with each cluster’s biome composition, not measured per-species niche breadth and not transfer rates, so the idea that acquisition rate tracks the breadth of a species’ environments is at most a hypothesis this comparison suggests and is not established by it. 1
MicrobeAtlas refines this ecology–content connection: metal-type diversity remained associated with inferred niche breadth after pangenome coverage, genome size, and sampled-species count adjustment (β = +0.0218, SE = 0.0061, p = 3.5×10⁻⁴ in a 527-genus model), but strict 5% within-environment prevalence yielded β = +0.0166, SE = 0.0099, p = 0.092. 15
Phylogenetic signal of niche and resistance traits
The Metal Resistance Ecology — Phylogenetic Conservation vs. Environmental Selection adds a trait-distribution measure that does not depend on gain attribution: Pagel’s λ, a phylogenetic-signal statistic estimating how closely a trait’s distribution across taxa tracks the tree. λ indexes covariance with phylogeny, not the timing or depth at which a function was acquired, so λ values and the recent-to-ancient gain ratios above are not on a common scale. Across 1,264 bacterial genera with ≥ 3 OTUs in MicrobeAtlas, Levins’ B_std — a standardized niche-breadth index — shows strong phylogenetic signal at λ = 0.787, p = 7.9×10⁻¹⁰² by likelihood-ratio test (LRT), and habitat range, the number of environment categories detected, is even more conserved at λ = 0.909, p = 1.4×10⁻¹⁵⁷. 15
Metal antimicrobial-resistance (AMR) traits carry intermediate but significantly non-zero signal on that same scale. Among 606 bacterial genera on GTDB r214, Pagel’s λ is intermediate and significantly non-zero for all three metrics: total AMR cluster count λ = 0.260, p = 6.1×10⁻²⁸; core AMR fraction λ = 0.441, p = 1.8×10⁻⁸; and metal type diversity λ = 0.335, p = 1.1×10⁻²³. 15 The report reads the ordering — core fraction λ > type diversity λ > cluster count λ — as consistent with a model in which constitutive core metal resistance is mainly vertically inherited while accessory gene accumulation and metal-type expansion are shaped by horizontal transfer and local metal exposure; this is a consistency argument from a signal ordering and establishes neither inheritance mode. 15
The trait-class contrast is the load-bearing comparison: in bacteria, niche breadth and habitat range are more phylogenetically conserved than metal AMR traits, λ_niche ≈ 0.79–0.91 against λ_AMR ≈ 0.26–0.44. 15 This supports, on a substrate independent of Sankoff gain attribution, the page’s separation of tightly tree-tracking traits from exchange-associated ones, while measuring a different quantity than the depth ratios and so not validating them. 15 A metabolic positive control anchors the high-signal end: nitrification carries near-maximal signal, λ = 0.939 in bacteria and λ = 1.000 in archaea — reported as λ = 0.94–1.00 in both domains — which the report attributes to ancient, vertically inherited metabolic entrenchment in lineages such as Nitrospirota and Thaumarchaeota and treats as an internal comparison strengthening confidence in the pipeline. 15
Archaea were fitted separately and give a mixed picture at far smaller n. Archaeal niche breadth has detectable but much weaker signal than its bacterial counterpart, B_std λ = 0.197, p = 1.1×10⁻⁵ at n = 132, while archaeal habitat-range count is strongly conserved, λ = 0.898, p = 4.6×10⁻¹⁴ at the same n = 132; the report reads the low B_std λ as consistent with the phylogenetically sparse and ecologically distinct nature of the archaeal genera sampled by short-read 16S amplicon surveys. 15
One archaeal resistance trait is a null: AMR cluster count shows approximately zero phylogenetic signal across the 48 archaeal genera carrying that measurement, at p = 1.0. 15 This null is specific to cluster count in those 48 genera and does not establish absent signal for archaeal core-AMR fraction or metal-type diversity, for which no archaeal λ is quoted here. 15
Two methodological qualifications bound every λ quoted here. The niche-breadth signal is robust to abundance filtering: against 0.787 across all 1,264 bacterial genera and 0.763 recomputed on the same 576 genera, the ≥ 0.01% relative-abundance filtered subset gives λ = 0.750 at LRT p = 2.8×10⁻⁴⁰, a change of −0.013 (−1.7%). 15 The estimate is nonetheless model-dependent: standalone Pagel’s λ for B_std is 0.787 while the λ estimated within the phylogenetic generalized least-squares (PGLS) model — a regression that models residual covariance from the tree — is 0.708, because the PGLS removes some phylogenetically structured variance by including metal type diversity, itself carrying signal at λ = 0.335, as a predictor. 15 A reported λ is therefore a property of the model it was fitted in rather than of the trait alone. 15
Accessory genes, producer behavior, and exchange
The pathway analysis supports a gene-function acquisition-depth interpretation through accessory content: across approximately 293,000 genomes and 80 GapMind pathways, core-only versus all-gene completeness gaps were 0.146 for leucine and valine biosynthesis, 0.141 for arginine, 0.140 for lysine, and 0.140 for threonine; the corresponding all-gene/core-only values were 0.614/0.468, 0.614/0.468, 0.613/0.472, 0.804/0.664, and 0.803/0.663. 5 These exact gaps are consistent with a Black Queen hypothesis and possible community sharing but do not demonstrate metabolite exchange. 5
The PGP analysis refines this model: all 13 PGP genes had higher core fractions than the 46.8% genome-wide baseline; pqqC was 81.5% core, 7.8% auxiliary, and 10.7% singleton, acdS 70.4% core, 13.4% auxiliary, and 16.2% singleton, nifH 63.8% core, 15.1% auxiliary, and 21.0% singleton, and pqqD 55.5% core, 17.0% auxiliary, and 27.5% singleton. 14
The defense analysis supports a complementary accessory-exchange model. Relative to 46.8% core and 37.9% singleton background, Retron had 8.2% core and 65.8% singleton, Gabija 9.1% and 68.8%, R-M Type II 9.4% and 68.2%, R-M Type I 9.8% and 65.3%, CBASS 11.8% and 61.1%, BREX 20.7% and 55.8%, CRISPR-Cas 27.4% and 47.2%, and DISARM 40.6% and 42.3%. 13
The atlas classified clade × function tuples as Innovator-Isolated, Innovator-Exchange, Sink/Broker-Exchange, or Stable: 11,829,746 Stable at low producer plus low participation, 803,196 Innovator-Isolated at high producer plus low participation, 741,587 Sink/Broker-Exchange at low producer plus high participation, 50,026 Innovator-Exchange at high producer plus high participation — rare, and the v2 plan’s hardest call — and 314,607 insufficient_data tuples, whose n_clades_with was too low for the consumer null. 1 Innovator-Exchange, the methodology’s hardest defensible call without donor inference, constitutes 0.36% of scored tuples. 1 These rank-aware results connect to cofitness-network-architecture and pangenome-integration. The classification is the atlas’s organizing framework: a direction-agnostic per-clade Producer × Participation categorization into Innovator-Isolated, Innovator-Exchange, Sink/Broker-Exchange and Stable, applied at deep ranks where duplication-transfer-loss (DTL) reconciliation is intractable, with a per-rank null model built from prevalence-binned cohorts. 1 This deep-rank Producer × Participation atlas spans 13.74M rows at all deep ranks, with Innovator-Isolated 803K, Innovator-Exchange 50K, Sink-Broker-Exchange 742K, Stable 11.83M and Insufficient-Data 315K. 1 Production rate within that framework is paralog expansion above the per-rank null, carried by the producer_z score from the M5/M6 null-model machinery, and the report presents the per-rank null plus Producer × Participation categorization as its own direction-agnostic construction rather than an inherited scheme, noting that Alm 2006 did not categorize this way. 1 A separate genus-rank tree-based parsimony donor inference reuses the M22 gains, the species tree and the presence matrix to emit Open, Broker, Sink and Closed labels with explicit ambiguity, while the composition-based donor inference was deferred. 1 A review pass positioned the method against literature it had not previously cited: duplication-transfer-loss reconciliation (Bansal 2013) is the principled methodology, and this project’s Sankoff parsimony is its lightweight tractable alternative at GTDB scale, which is a tractability argument rather than a demonstration of equivalence. 1
That tree-based donor inference is a different layer from the deep-rank quadrants above: its Open-Innovator, Broker, Sink and Closed-Stable labels are donor-proxy calls made only at genus rank on the Phase 3 candidate set, not the deep-rank Producer × Participation categories. Its 4.06M rows carry those labels as Open-Innovator 2.38M, Broker 153K, Sink 182K and Insufficient 1.34M, with no Closed-Stable count given in the same listing, split by confidence into high 1.36M, medium 646K and low 711K. 1 The report prints a methodological caveat that qualifies every one of those labels: the algebraic formula counts every genus with the KO present as a potential donor for all family-mates’ gains, biasing toward Open-Innovator dominance, and composition-based confirmation (M25, deferred) would be needed to resolve which Open-Innovator-tagged genera are empirically donor-like rather than merely parsimoniously donor-compatible; the tree proxy is reported as exploratory. 1
Joining those labels back to the genus-rank atlas yields an 8.64M-row concordance table in which 169K comparisons agree, 2.55M conflict and 5.92M are insufficient. 1 The conflict column must not be read as 2.55M substantive classification disagreements: the per-tuple conflict analysis classifies most of them as expected “deep-stable + genus-active” rank relationships rather than real conflict, leaving approximately 22K real rank-dependent conflicts. 1 So the comparison qualifies any attempt to treat the deep-rank quadrant classification and the genus-rank tree-proxy donor labels as interchangeable, chiefly by showing that the two layers speak about different ranks. 1
Per focal clade at genus rank, the tree proxy distributes its labels unevenly. Across 28,389 Mycobacteriaceae genus × candidate-KO tuples it calls 75.3% (21,365) Open-Innovator, 13.0% (3,701) Broker, 9.2% (2,619) Sink and 2.5% (704) Insufficient. 1 Across 54,582 Cyanobacteriia tuples it calls 47.9% (26,158) Open-Innovator, 3.7% (1,999) Broker, 4.4% (2,395) Sink and 44.0% (24,030) Insufficient; across 439,968 Bacteroidota tuples, 56.5% (248,581) Open-Innovator, 4.5% (19,663) Broker, 5.4% (23,612) Sink and 33.7% (148,112) Insufficient. 1 The large Insufficient fractions for Cyanobacteriia and Bacteroidota qualify any donor-side reading at genus rank for those clades, while Mycobacteriaceae is the one focal clade where the proxy returns a label for nearly every tuple; given the proxy’s Open-Innovator bias and exploratory status, that contrast cannot by itself be read as more transfer into Mycobacteriaceae. 1
Truly dark genes extend this network interpretation: among 4,394 truly dark genes in organisms with integrated condition data, 25.9% showed operon-like cofitness with adjacent genes using |r| ≥ 0.3, 41% of neighboring genes were also hypothetical, and 594 genes belonged to integrated-condition-analysis fitness modules. 3 Module associations implicated candidate phage integrases, metal transporters, chemotaxis systems, and iron regulation, while the top-ranked candidates included genes associated with TatC, ABC transport, and fluoride-efflux contexts. 3 These are guilt-by-association hypotheses rather than demonstrated biochemical functions.
Only 48 organisms had fitness tests among 293,000 genomes with pathway predictions, and laboratory conditions may make environmentally important pathways appear latent. 6 Conservation did not distinguish latent from active pathways: mean conservation was 0.869 for 248 latent pathways versus 0.829 for 755 active dependencies, with Mann–Whitney U p = 0.94 for active > latent and rank-biserial r = 0.052. 6 The pathway analysis extends this limitation: only 7 of 48 Fitness Browser organisms had matching GapMind data. 5
Architectural diversity and exchange rate
Within-KO Pfam architectural diversity — the number of distinct Pfam domain architectures carried by genes assigned to one KO — tracks exchange behavior across four focused KO subsets rather than across whole function classes. For the 11 NB12-curated mycolic-acid-specific KOs, the median was 5 architectures per KO with a dominant-architecture fraction of 0.93, alongside the NB12 Innovator-Isolated call at family and order ranks in Mycobacteriaceae. 1 For the 292 KOs of the two-component-system histidine-kinase control class, the median was 15 architectures per KO with a dominant fraction of 0.66, alongside an atlas-wide M21 sanity-rail PASS at d=0.65-0.79. 1
The report states that higher architectural diversity per KO correlates with elevated cross-clade exchange at the atlas level, contrasting photosystem ii’s near-zero diversity of 1 architecture per KO — corresponding to clade-restricted, Cyanobacteria-internal paralog expansion — with the median of 46 architectures per KO in the Mixed-top-50 subset, the 48 mixed-category KOs with the highest Innovator-Exchange scores, and their elevated Innovator-Exchange rate. 1 Because that subset was selected on the very behavior it is compared against, and because the census covers chosen subsets rather than the whole mixed category, this is an association across selected KO sets rather than a demonstrated cause of exchange; it refines the pathway-overlap asymmetry recorded in the next section by suggesting that architectural heterogeneity, not regulatory status, is the property co-varying with exchange in these comparisons. 1 The report frames the generalization as novel — a KO’s structural diversity predicting its propensity to flow, which to its authors’ knowledge no prior atlas-style work at GTDB scale has framed as within-KO architectural promiscuity — and that novelty assessment is the authors’ own rather than an independent survey result. 1
Regulatory versus metabolic function classes
The Phase 2 headline test (NB11) asked whether KOs in regulatory KEGG categories (Genetic Information Processing ko03xxx plus Environmental Information Processing ko02010–ko02065) show different innovation and acquisition signatures than KOs in metabolic categories (ko00xxx–ko019xx), under a pre-registered criterion of Cohen’s d ≥ 0.3 with a 95% confidence interval excluding zero and Mann–Whitney p < α/4 = 0.0125, the Bonferroni threshold for four headline tests. 1
No test achieved the pre-registered d ≥ 0.3 threshold; two of the four showed small but statistically significant signals after Bonferroni correction, and two were null. 1
Test T1, producer z-median, was null with n_reg=1,554 and n_met=4,970, d=+0.059, 95% bootstrap CI [−0.011, +0.124], and two-sided Mann–Whitney p=0.71. 1 Test T2, consumer z-median, gave a small effect in the opposite direction with n_reg=1,552 and n_met=4,964, d=−0.211, 95% bootstrap CI [−0.288, −0.150], and p<10⁻¹⁵. 1 Test T3, recent-acquisition percentage, gave a small effect with n_reg=1,554 and n_met=4,970, d=+0.141, 95% bootstrap CI [+0.092, +0.189], and p=3.0×10⁻⁶. 1 Test T4, number of clades with maximum score, showed no signal with n_reg=1,552 and n_met=4,964, d=+0.020, 95% bootstrap CI [−0.045, +0.082], and p=0.12. 1
The strong-form regulatory-versus-metabolic asymmetry hypothesis is therefore not supported at GTDB scale, and the project fell back to its pre-registered outcome of continuing as a descriptive atlas. 1 The T2 direction contradicts the prior that regulatory genes flow more than metabolic genes: regulatory KOs were more clumped (lower consumer z, implying more vertical inheritance), and although the two-component-system histidine-kinase signal against tRNA synthetases is real at d=0.65, it does not generalize across the regulatory category as a whole. 1
A softer version of the hypothesis survives weakly: regulatory KOs showed slightly more recent acquisitions than metabolic KOs, 47.3% versus 45.9% recent, but at d=0.14 this sits below the project’s biological-significance threshold. 1 The bonus NB11 comparison was accordingly reframed as a small effect, consumer d=−0.21 in the Jain 1999 direction, with regulatory KOs at 4.13% MGE-machinery against metabolic KOs at 0.37% and mixed KOs at 0.27% — a roughly 10× higher regulatory rate, which the report reads as consistent with regulatory products including transposon-bound and phage-bound regulators. 1
The asymmetry that does exist in the atlas is between pathway-overlap and pathway-pure KOs rather than between regulatory and metabolic classes: KOs annotated to both regulatory and metabolic pathways occupy the Innovator-Exchange quadrant at 0.57%, twice the rate of pure regulatory (0.27%) or pure metabolic (0.28%) KOs. 1 Regulatory KOs were also somewhat elevated in Sink/Broker-Exchange at 7.16% against metabolic 5.20%, suggesting more cross-clade exchange participation even where the direction of exchange is unclear. 1
The quadrant breakdown by category runs as follows. Metabolic: 5,038,487 clade × KO tuples, 0.28% Innovator-Exchange, 4.24% Innovator-Isolated, 5.20% Sink/Broker-Exchange, 87.85% Stable. 1 Regulatory: 1,609,066 tuples, 0.27%, 5.72%, 7.16%, 84.50%. 1 Mixed: 4,888,614 tuples, 0.57%, 8.04%, 5.71%, 83.68%. 1 Other: 398,805 tuples, 0.33%, 4.08%, 9.96%, 82.73%. 1 Unannotated: 1,804,190 tuples, 0.13%, 4.88%, 2.52%, 89.98%. 1
Acquisition-depth profiles across the same categories, in the order recent, older_recent, mid, older, ancient, were: metabolic 45.92%, 32.17%, 10.23%, 7.04%, 4.64%. 1 Regulatory 47.27%, 32.07%, 9.68%, 6.64%, 4.34%. 1 Mixed 47.62%, 32.75%, 9.82%, 6.15%, 3.66%. 1 Other 39.99%, 30.17%, 11.38%, 9.71%, 8.75%. 1 Unannotated 46.88%, 31.37%, 9.79%, 6.95%, 5.01%. 1
The report cautions that differences across these KEGG categories are small, and that the most ancient-skewed profile — the “other” category of off-target pathways, ko04xxx organismal systems and ko05xxx human diseases — is consistent with mostly evolutionarily older or eukaryotic-derived annotations passing accidentally through PROKKA → eggNOG mapping. 1 That explanation is offered as consistent with the profile rather than demonstrated, so the “other” row may partly reflect annotation-pipeline provenance and cannot be read as a purely biological acquisition-depth signature. 1
Formally the outcome was a fallback to the null: the project’s hypothesis-status table records the regulatory-versus-metabolic asymmetry test as reframed with an H0 fallback, with signals existing at d=0.14-0.21 but below the d ≥ 0.3 threshold, and the report places that range inside the d=0.0-0.3 band Cohen 1988 calls small, so the statistically significant signals are descriptive rather than evidence for a strong-form asymmetry. 1
The direction of the consumer-side signal is nonetheless the one the Jain 1999 complexity hypothesis expects, and the report records NB11’s T2 finding as supporting that hypothesis at small effect size. 1 The report reads regulatory KOs as more phylogenetically clumped than metabolic KOs — consumer_z d=−0.211, Mann–Whitney p<10⁻¹⁵, n_reg=1,552 versus n_met=4,964 — that is, as showing less horizontal transfer, exactly as the complexity hypothesis predicts. 1 This supports that prior in sign only: it is one of the two signals that were statistically significant after Bonferroni correction, T2 and T3, and neither of those two reaches the pre-registered effect-size threshold. 1 The report keeps both framings side by side and documents both: NB11 is direction-consistent with Jain 1999 at small effect size, while the strong-form asymmetry is falsified at the d ≥ 0.3 threshold. 1 It further records the reframed NB11 direction, consumer d=−0.21, as independently empirically validated by Burch et al. 2023, PMID:37232518, which it reports as recovering the same direction at Lactobacillaceae scale; this is an external direction-only comparison at a different taxonomic scale, not a re-test of the atlas effect size. 1 At the function-class pool level the finding is exactly the one reported above — regulatory KOs more clumped than metabolic, d=−0.21 at p<10⁻¹⁵. 1
The project treats the reframing itself as a finding: it falsifies the strong-form regulatory-versus-metabolic asymmetry prior at GTDB scale and at the pre-registered d ≥ 0.3 threshold, and the novel observation that replaces it is that mixed-category KOs, those participating in both regulatory and metabolic pathways, show 2× the Innovator-Exchange rate of pure regulatory or pure metabolic categories, 0.57% versus 0.27% / 0.28%. 1 This remains hypothesis-generating: no mechanism linking dual pathway membership to exchange was tested. 1
Two scoping statements follow from the pooled null. The atlas’s 50K Innovator-Exchange clade × KO tuples remain the substantive map, and the regulatory-versus-metabolic test confirms they are not concentrated by that split. 1 The null also does not invalidate the four pre-registered weak-prior hypotheses, each of which tests a specific clade × function tuple and survives it; what it removes is the ability to use a regulatory/metabolic asymmetry to predict which tuples will be innovators. 1
Mobile-genetic-element context and tensions
The atlas-wide MGE-machinery baseline was 1.37% of KO-bearing gene clusters, versus 0.57% for Mycobacteriaceae mycolic-acid KOs and 0.00% for Cyanobacteriia PSII and Bacteroidota PUL KOs. 1 PSII neighborhood analysis examined 27,148 focal features and 218,321 neighbor pairs: 10.91% had at least one MGE neighbor within ±5 kb versus a Poisson baseline of 10.6%, with mean MGE-neighbor fraction 1.65%. 1 On this basis the report judges all three pre-registered focal KO groups — PSII, PUL and mycolic acid — to be not phage-borne, citing per-cluster MGE-machinery rates of 0%, 0% and 0.57% together with the PSII gene-neighborhood rate of 10.91% against a Poisson expectation of 10.6%. 1
The truly dark-gene analysis supports retaining this distinction between MGE association and demonstrated transfer: 12.0% of truly dark genes were within 2 genes of a mobile genetic element, but the report explicitly notes that GC deviation, short length, gene-specific composition bias, amelioration, and neighborhood association cannot independently establish HGT. 3
The T4SS–CAZy study supports the same caution. T4SS-positive genomes had 10× higher MGE density, but CAZy genes were not detected on plasmids by ICEfinder; 12 IMEs occurred among the top 100 accumulators, and chromosomal or integrative transfer remains a hypothesis rather than an established mechanism. 11
Tensions and limits
The project did not reproduce the Alm 2006 quantitative correlation at GTDB scale: across 18,989 representatives, four framings yielded Pearson r values 0.288, 0.157, 0.100, and 0.105, and Spearman r values 0.333, 0.222, 0.152, and 0.106 rather than r ≈ 0.74. 1 Differences in the original 207-genome substrate, M22 tree-attributed gains versus per-genome paralog expansion, and “recent” definitions partly explain the tension. The report frames the outcome as methodology generalization holding while point-estimate reproduction does not: r = 0.10–0.29 across n=18,989 rather than the published r ≈ 0.74, with NB17 architectural concordance at r = 0.67. 1 The report’s closure note on the failed anchor states the position squarely: the project’s intellectual lineage is methodology generalization of Alm 2006, not point-estimate reproduction, and it names its own atlas-as-flow-and-acquisition-map framing as the alternative framework. 1 The reproduction joined per-genome histidine-protein-kinase counts, taken as PF00512 hits from InterProScan domains, to each genome’s recent-LSE fraction under M22 recent-rank gain attribution across all 18,989 P1B-qualified species representatives, and tested four framings of the correlation. 1 NB17 retains the qualitative result at architectural resolution — the canonical PF00512 + PF02518 architecture concentrated at recent ranks, with consumer-side concordance between architectural and KO-level signals at r = 0.67 — which the report calls its strongest connection to Alm 2006 even though the quantitative anchor does not survive scaling. 1
The four framings, each at n = 18,989, were: HPK count versus recent two-component-system gains at genus rank, Pearson 0.288, Spearman 0.333, p < 10⁻³⁰⁰; HPK count versus recent-LSE fraction per GC, Pearson 0.157, Spearman 0.222, p < 10⁻¹⁰⁵; HPK count versus recent-LSE per TCS histidine-kinase KO, Pearson 0.100, Spearman 0.152, p < 10⁻⁴³; and the Alm-2006-style HPK fraction versus recent-LSE fraction, Pearson 0.105, Spearman 0.106, p < 10⁻⁴⁷. 1 The verdict recorded on this set is NOT REPRODUCED: all four correlations are statistically significant at n = 18,989 but biologically modest, r = 0.10–0.29 against Alm 2006’s r ≈ 0.74 in 207 genomes, and the strongest framing — HPK count versus recent TCS gains at genus rank — recovers r = 0.29, which the report calls less than half of Alm 2006. 1
The report names three identifiable reasons for the dilution rather than leaving the gap unexplained. Substrate scale: Alm 2006 used 207 genomes, mostly cultivated isolates with high annotation density and broad taxonomic spread, whereas this substrate is 18,989 species representatives spanning the full GTDB tree — including DPANN, CPR and weakly-annotated lineages where HPK detection is genuinely sparse — so aggregating across the full tree dilutes a taxonomically heterogeneous signal. 1 Lineage-specific-expansion detection method: Alm 2006 counted per-genome paralogs at the histidine-kinase family level, while this analysis detects via M22 Sankoff-parsimony recent-rank gain attribution on the species tree, a tree-aware approach that would miss expansions occurring entirely within a single recent species, so the two metrics measure overlapping but non-identical phenomena. 1 Tree-rank granularity: “recent” in M22 is a leaf-adjacent gain on the species tree, capturing within-genus diversification, whereas Alm 2006’s lineage-specific expansion was operationalized as a per-genome paralog count capturing all expansion regardless of phylogenetic depth; the two definitions converge for very-recent expansions but diverge for older ones. 1
That comparison has a chronology inside the report, and the earlier entries are kept here as an audit trail rather than as a live status. The hypothesis-status table and the Phase 3 architectural note record the explicit reproduction of the Alm 2006 two-component-system histidine-kinase result at r ≈ 0.74 as not yet tested and deferred to Phase 4, and the report notes that Alm 2006 measured per-genome histidine-kinase count against lineage-specific-expansion fraction across 207 genomes broadly, not Mycobacteriaceae mycolic acid, so the mycolic-acid Innovator-Isolated finding is at most biologically consistent with it, both being clade-level paralog-expansion signatures. 1 In place of the correlation those earlier sections report an architecture-level profile: the canonical HisKA + HATPase_c architecture, equivalent to Alm’s IPR005467 single-domain detection scaled to multidomain, appears at 25K gain events with a 44.7% recent / 4.5% ancient profile, a recent-skewed signature consistent with lineage-specific expansion, while the full per-genome reproduction stayed deferred at that point. 1 Phase 4 was subsequently closed and the attempted reproduction did not recover the published value, giving the r = 0.10–0.29 range recorded above; the deferral notices are historical status, not an alternative current result. 1
A reporting discrepancy inside this source is unresolved. The report headline and the cross-project digest both give an approximately 1× recent-to-ancient ratio for strict housekeeping, a set the digest defines as RNA polymerase core, tRNA synthetases and ribosomal proteins, while elsewhere the same report pairs CRISPR-Cas 24.5× with tRNA-synth 2.3×, the figure reported above on this page for the clean tRNA-synthetase control class. 9,1 The available evidence does not show that the two values describe different control definitions, and the discrepancy must not be resolved by averaging them or by preferring either figure. 9,1 The per-class gain table widens that discrepancy rather than closing it, because the three strict housekeeping classes do not share one ratio there either: strict tRNA synthetases 2.3, strict ribosomal proteins 6.2 — which the report flags as mixed, as expected for the contaminated strict-K range — and strict RNA-polymerase core 6.4. 1
A second tension is internal to the mycolic-acid result. The pre-registered hypothesis table identifies the mycolic-positive species as the M. tuberculosis and M. leprae complexes, whereas the project’s v3.5 sub-clade recomputation places Mycobacterium in the mycolic-low sub-clade — the one scored at d = +0.211 — and draws the mycolic-positive sub-clade from the other genera among the 10 of 13 that met the mean LC ≥ 0.5 cut. 1,9 That recomputation also reports that the consumer-side signal does not persist at genus rank, so the Innovator-Isolated interpretation is weakened and rank-dependent rather than a property of the family at every rank. 1
The pangenome-openness cross-validation was null at atlas scale (Spearman r = −0.011 across 894 genera), while targeted Mycobacteriaceae and Cyanobacteriia tests were underpowered (n = 10 and n = 83 genera). 1 The pathway analysis contrasts with this atlas-scale null at a different aggregation level: variable pathway count correlated with openness at partial rho=0.530, p=2.83e-203 across 2,810 species, and PGP richness correlated negatively with openness (ρ = −0.195, p = 2.0e-97). 5,14 These findings must not be combined as one universal openness effect. The project reads its own atlas-scale result (P4-D4) as an informative null in which M22 gain attribution and pangenome openness measure distinct evolutionary phenomena. 1
That null is specified test by test across the same 894 genera. Atlas-wide openness versus log recent gains gave Pearson +0.017 and Spearman −0.011 with a 95% CI of (−0.08, +0.06), scored null. 1 Restricted to regulatory KOs it gave Pearson +0.023 and Spearman −0.013, CI (−0.08, +0.05), also null; restricted to metabolic KOs, Pearson +0.010 and Spearman −0.017, CI (−0.08, +0.05), null again. 1
The reason the report gives is definitional. An M22 recent-rank gain is a between-species KO presence/absence transition on the GTDB species tree inferred via Sankoff parsimony across species representatives — a KO acquired by one species but absent in its sister species — whereas pangenome openness indexes within-species strain diversity, so openness is not a valid cross-substrate validation of M22: P4-D4 was designed on the working assumption that the two would correlate, and the empirical answer is that they do not, because they measure distinct phenomena. 1 The null changes no headline verdict — NB11 reframed, NB12 supported and NB16 supported all stand, and the atlas-wide null does not undermine any pre-registered hypothesis result. 1 Cross-substrate validation of M22 therefore remained an open project commitment after P4-D4, with phenotype and ecology grounding named as the remaining path: whether high recent acquisition in a function class tracks the biomes and phenotypes expected for that class. 1
A review pass recorded a methodological limit of a different kind: the project’s Sankoff parsimony approximation operates without principled benchmarking against modern duplication-transfer-loss reconciliation methods. 1 Two comparators were added to the reference list rather than run against the atlas. Liu, Mawhorter, Liu et al. 2021, BMC Bioinformatics 22(Suppl 10):394, doi:10.1186/s12859-021-04290-6, PMID:34348661, extends DTL to the DTLOR model, handling two events common in microbial evolution — gene origin from outside the sampled species tree, and rearrangement of gene syntenic regions — which the report flags as directly relevant to GTDB-scale HGT analysis. 1 Kundu and Bansal 2018, BMC Bioinformatics 19(Suppl 9):290, doi:10.1186/s12859-018-2269-0, PMID:30367593, documents that a large fraction of gene trees have multiple optimal rootings and quantifies which aspects of DTL reconciliation are conserved across rootings — a rooting-uncertainty quantification this project’s Sankoff approximation lacks. 1
The report’s stated position on that gap is a tractability argument, not a claim of equivalence: Sankoff parsimony is computationally tractable at full GTDB scale, 17M gain events on an 18,989-leaf tree, where exact DTL methods are not; the Producer × Participation framework (M5/M6) supplies a complementary signal that requires no donor inference; and M22 acquisition-depth attribution surfaces recipient-and-when signal at scale even without principled DTL. 1 The corresponding limits are stated plainly. Per-family DTL reconciliation with AleRax or ALE does not run at full GTDB scale, so the four-quadrant donor labels are produced only at genus rank on the Phase 3 candidate set of roughly 10K KOs rather than atlas-wide; this restriction applies to the Open/Broker/Sink/Closed donor-proxy labels and not to the deep-rank Producer × Participation categories, which are reported across the whole 13.74M-row deep-rank atlas. 1 A comprehensive cross-validation against Liu 2021 DTLOR or Bansal 2012/2013 SPR-based DTL on a representative GTDB subset was declared out of scope, requiring external tool integration plus GTDB-scale per-family gene-tree estimation, and was left as a future direction. 1
Uncertainty is also not propagated end to end. The report logs its error-propagation item as open: Phase 1 UniRef50→KO projection errors, Phase 2 Sankoff approximation errors, Phase 3 architectural census coverage gaps and Phase 4 residualization assumptions are none of them carried into a unified atlas confidence interval, Monte Carlo error propagation was estimated at one to two weeks of additional work and deferred, and readers are directed instead to weight findings by convergence across the three independent measurement substrates — atlas effect size, ecology grounding, phenotype anchor — rather than by single-pipeline confidence intervals. 1 External tool integration with PaperBLAST, AlphaFold and eggNOG likewise remained future work, the project closing at the atlas-construction and interpretation level rather than at experimental validation. 1
The project’s own list of genuinely remaining open items names four: possible sample-size effect inflation, especially the PSII class-rank n=21; phylogenetic non-independence, with no phylogenetic-independent-contrasts (PIC) correction implemented — PIC being a method that adjusts comparative statistics for shared ancestry; unquantified cross-phase error propagation; and mixed two-component-system architectural concordance. 1 That PIC item sits in tension with an earlier review pass in the same source, which retires M9 (PIC) on the ground that per-leaf Sankoff parsimony already encodes tree topology so PIC stacked on Sankoff is double-counting, and names UniRef-cluster non-independence — multiple UniRef50s within a Pfam family inheriting related phylogenetic distributions, so that Mann–Whitney across UniRefs treats them as independent and inflates effective N — as the legitimate residual concern, bound instead to M19 cluster-bootstrap at Pfam-family granularity with B = 200 resamples per function class. 1 Both the retirement of PIC and the later listing of PIC as unimplemented are recorded here rather than reconciled, and these limits bound every atlas-derived effect size quoted on this page. 1
The truly dark-gene result adds a tension between functional darkness and apparent sequence recognition: 79.4% of truly dark genes had UniRef50 links and 84.7% had database cross-references, but only 4.0% had Pfam hits, 4.6% had KEGG KOs, and 17.0% had COG categories; 55.4% of COG assignments were category S, meaning function unknown. 3 This refines the concept of “darkness”: sequence-level homology and identifiers can be widespread while experimentally useful function remains sparse.
The genotype-to-phenotype tension remains: binary growth was only moderately predictable (AUC 0.620), while continuous µmax, lag, and max_A were not transferable across genera. 4 Complete pathways were often fitness-neutral, latent fractions were pathway-category-dependent and threshold-sensitive, and conservation did not distinguish latent from active pathways. 6
A separate marker/function tension remains: xoxF was broad and had global fraction 0.9498 among xoxF + mxaF calls, but validated lanmodulin occurred in only 62 genomes across 10 species and 100 % of those genomes were in Beijerinckiaceae, Acetobacteraceae, or Hyphomicrobiaceae. 8 Truly dark genes reinforce this caution: many genes have sequence identifiers without reliable functional assignments, and Bakta may produce false-negative calls when a function is absent from the tested PSC database. 3
Environmental interpretation remains associative: ENIGMA used Spearman correlation and incomplete sample-to-location resolution; IBD ecotype stability was low; methylotrophy H2 was only partially supported (OR = 3.51, p_BH = 0.082); and PHB inference scores had only modest correlations with measured variables, all |rho| < 0.12. 4,7,8,10
Open Directions
- Combine per-CDS sequence data with composition-based donor detection and DTL reconciliation; complete deferred PSII, PUL, and mycolic-acid neighborhood scans; and bootstrap M22 gain events to test transfer histories and recent-to-ancient ratio stability. 1
- Run the deferred composition-based donor confirmation (M25) on the Open-Innovator-tagged genera to separate empirically donor-like genera from merely parsimoniously donor-compatible ones, closing the bias the tree proxy’s potential-donor counting introduces. 1
- Benchmark M22 Sankoff attribution against Liu 2021 DTLOR and Bansal SPR-based DTL reconciliation on a representative GTDB subset, and quantify gene-tree rooting uncertainty per Kundu and Bansal 2018, to close the gap that the atlas’s transfer inference currently has no principled reconciliation baseline. 1
- Propagate per-phase uncertainty — UniRef50→KO projection, Sankoff approximation, architectural census coverage, residualization — into a unified atlas confidence interval, so that depth-signature claims carry error bars instead of requiring readers to weight them by substrate convergence. 1
- Run the deferred bootstrap of per-event M22 uncertainty, estimated at roughly 50 hr of compute, so that depth-bin and clade × function effect sizes carry per-event error bars instead of resting on the
n_leaves_underand leaf_consistency proxies. 1 - Settle which residual non-independence correction the atlas effect sizes need, given that the project lists PIC as unimplemented among its open limits while a review pass retires M9 (PIC) as double-counting topology: run the M19 cluster-bootstrap on UniRefs at Pfam-family granularity, B = 200 resamples per function class, recomputing per-leaf Sankoff Cohen’s d and reporting 95% CI, and ask whether the reported effect sizes — including small samples such as the PSII class-rank n=21 — survive UniRef-cluster resampling. 1
- Examine the approximately 22K tuples the conflict analysis classes as real rank-dependent disagreements between the deep-rank quadrants and the genus-rank tree-proxy labels, asking which layer tracks independent transfer evidence in the focal clades. 1
- Use matched species-clade pangenomes and fitness-tested strains to test whether the ρ = 0.69 openness association, partial rho=0.530 variable-pathway association, and partial ecotype rho=0.322 survive correction for species composition, sampling imbalance, genome count, and phylogeny. 6,5
- Extend pangenome linkage to the 17,479 unlinked dark genes, then test whether the estimated approximately 2,841 additional truly dark genes reproduce the linked set’s length, orthology, GC-deviation, mobile-element, and fitness patterns. 3
- Structure-predict the top 100 truly dark candidates with AlphaFold2 or ESMFold and search Foldseek; test PV4/5210953 by motility assays, ANA3/7026383 under nitrogen limitation, and selected candidates with mobile-CRISPRi. 3
- Test whether truly dark-gene operon-like cofitness and ICA-module associations predict functions by measuring neighboring genes, module-linked conditions, and polar-effect controls in the prioritized organisms. 3
- Reanalyze MicrobeAtlas with finer environment categories, alternative trees, discrete/count phylogenetic models, rarefaction, manual validation of 100 AMRFinderPlus clusters, and metal-stress microcosms. 15
- Use PaperBLAST’s 129,823 VIMSS links to test whether sequence-family darkness predicts missing fitness measurements, and repeat acquisition-depth and ecological analyses after stratifying by literature coverage, organism research intensity, and 50% identity cluster size. 2
- Analyze PGP gene neighborhoods and operonic context, validate pqqC, acdS, hcnC, nifH, and pqqD with gene-integrity and expression tests, and determine whether the pqqC–acdS association reflects physical linkage, repeated ecological selection, or independently inherited loci. 14
- Replace defense-domain presence calls with PADLOC MacSyFinder-style multi-Pfam and gene-order rules for Retron, DISARM, and Gabija; quantify how refined calls alter accessory enrichment and defense-system prevalence. 13
- Reconstruct a phaC gene tree and use phylogenetic logistic regression or phylogenetic independent contrasts to test whether the 311 putative acquisition events, 278 putative losses, and elevated accessory fraction remain after shared ancestry is modeled. 10
- Perform the pending T4SS–CAZy synteny-threshold permutation test using unfiltered Spark data, BLAST-validate Node_4915 against NCBI nr, construct a housekeeping-gene null baseline, and calculate θ = OR(T4SS-CAZy) / [OR(T4SS) × OR(CAZy)] for biome enrichment. 11
- Measure growth across environmental media for pathways classified as latent in Fitness Browser conditions, asking whether latent capability becomes condition-specific dependency and whether continuous performance remains poorly predicted by pathway presence alone. 5
- Recompute Innovator-Exchange rates for mixed, pure regulatory and pure metabolic KOs under alternative KEGG category assignments and the deferred composition-based donor inference, asking whether pathway overlap rather than regulatory status is the real axis of exchange. 1
- Apply the Pfam architectural lens to the 67 Mycobacteriaceae × mycolic Innovator-Isolated tuples to identify which mycolic-acid-related Pfam architectures are most lineage-specific, closing the gap that architecture-specific lineage restriction was not established. 1
- Recompute producer and consumer effect sizes for mycolic-acid KOs per Mycobacteriaceae genus and per depth bin, and re-derive the housekeeping recent-to-ancient ratio from the gain table under each control definition, to settle which sub-clades are Innovator-Isolated and whether the housekeeping ratio is approximately 1× or 2.3×. 1,9
- Fit Pagel’s λ to per-function-class presence patterns on the same GTDB-r214 tree used for the MicrobeAtlas niche and metal-AMR traits, asking whether classes separated by recent-to-ancient gain ratio are also separated by phylogenetic signal; because both quantities would come from the same tree and atlas, this tests internal consistency between two trait summaries and would not supply the independent transfer anchor the atlas still lacks. 1,15
- Test whether the acdS, pqqC and hcnC soil enrichments survive a phylogenetic model rather than phylum-level fixed effects, and whether the soil-specific tryptophan–ipdC reversal reproduces inside the strict rhizosphere/root-nodule subset, to separate repeated ecological selection from finer-scale shared ancestry in the plant-growth-promoting gene set. 14