Horizontal Gene Transfer as a Driver of Bacterial Gene Novelty

Horizontal gene transfer (HGT), the movement of genetic material between lineages, is proposed here as a major mechanism generating bacterial gene novelty. ₁ The analysis links this proposal to the two-speed-bacterial-genome model, in which conserved core genes coexist with more variable genes associated with mobility, defense, and ecological adaptation. ₁

Evidence from COG functional distributions

The COG Functional Category Analysis compared Clusters of Orthologous Groups (COG) functional categories across 32 species spanning 9 phyla and 357,623 genes. ₁ Novel or singleton genes were enriched in COG L, the mobile-element category, by +10.88%, with 100% consistency across species. ₁ This was the strongest reported signal and supports the interpretation that mobile elements are closely associated with bacterial gene novelty. ₁

Novel or singleton genes were also enriched in COG V, defense mechanisms, by +2.83%, with 100% consistency, and in COG S, unknown function, by +1.64%, with 69% consistency. ₁ The combination of mobile-element, defense, and unknown-function enrichment is consistent with a model in which newly acquired genes contribute to ecological adaptation and lineage-specific functions, although the functional interpretation of unknown-function genes remains limited. ₁

The same analysis lists core genes as consistently enriched in COG J, translation (-4.65%, 97% consistency, labeled the strongest depletion); COG F, nucleotide metabolism (-2.09%, 100% consistency); COG H, coenzyme metabolism (-2.06%, 97% consistency); COG E, amino acid metabolism (-1.81%, 81% consistency); and COG C, energy production (-1.75%, 88% consistency). ₁ The report gives these core-enriched categories negative signs, which appear to be measured relative to the novel or singleton class. Because the sign convention is not stated explicitly, the values are copied as reported and not converted. ₁ These distributions support a two-speed-bacterial-genome interpretation in which conserved genes are concentrated in ancient metabolic and housekeeping functions, while novel genes are more associated with mobility, defense, and niche-specific potential. ₁

Because the observed functional partitioning was reported across the analyzed bacterial phyla, the report treats it as evidence for deep evolutionary constraint rather than a pattern restricted to one lineage. ₁ The report further interprets HGT as the primary innovation mechanism and the +10.88% COG L enrichment as evidence that much genomic novelty may arise through mobile elements. ₁ This is an interpretation of cross-species functional distributions rather than a direct measurement of transfer events, so the causal role of HGT remains a hypothesis requiring transfer-resolved analyses. ₁

The central digest restates these implications as broad evolutionary interpretations. It describes core genes as an ancient, conserved “metabolic engine” for translation, energy, and biosynthesis, and novel genes as recent acquisitions for ecological adaptation. It names HGT, not vertical inheritance, as the primary innovation mechanism, and takes the +10.88% L enrichment to suggest that most genomic novelty comes from mobile elements. _1,2 Category enrichment does not directly measure acquisition timing or adaptive effect. The +10.88% figure is an enrichment, not a measured proportion of all genomic novelty. The claim that mobile elements account for most novelty therefore remains a hypothesis. ₁

The Ecotype Functional Differentiation links the prominence of COG category L in ecotype differentiation to the same two-speed pattern. Category L includes transposases, integrases, and mobile genetic element machinery. The report cites the +10.88% L and +2.83% V novel-versus-core enrichments from the COG project. This cross-project interpretation supports the association but provides no independent evidence of mechanism. ₃

The The 5,526 Costly + Dispensable Genes provides convergent pangenome and fitness evidence: among 142,190 genes from 43 bacteria, 5,526 were both costly in laboratory measurements and dispensable, and this class was 7.45x more likely than costly+conserved genes to contain mobile-element keywords (OR=7.45, p=4.6e-71). ₄ Its 11.7x enrichment of the SEED category “Phages, Prophages, Transposable elements, Plasmids” (FDR=1.3e-17) supports the existing association between mobile elements and gene novelty, while its poor annotation, narrow ortholog breadth, and 24.2% singleton fraction refine that association toward recently acquired or unstable genomic material rather than core metabolism. ₄ This remains indirect support for HGT, because the costly+dispensable classification does not itself reconstruct transfer events. ₄

The costly+dispensable report interprets this class as the “genomic debris of horizontal gene transfer”. In this reading, the class consists of insertion sequences, prophage remnants, transposases, and defense systems that impose metabolic cost on their host and are not conserved across the species pangenome. ₄ The report proposes HGT-mediated genome expansion, rather than degrading metabolic pathways, as the primary source of costly non-conserved genes. For psRCH2, it suggests a recent mobile-element invasion or a genomic expansion specific to that strain. ₄ These are author interpretations rather than direct provenance measurements. They should be weighed against the gene-loss reading discussed in costly-dispensable-gene-loss. ₄

The conservation-fitness synthesis supports this distinction between potentially mobile accessory material and retained core functions: across 194,216 protein-coding genes from 43 bacteria, essential genes were 82% core whereas always-neutral genes were 66% core, and 28,017 genes were both costly in the laboratory and conserved compared with 5,526 that were costly and dispensable. ₅ Thus, the mobile-element enrichment of costly-and-dispensable genes refines the novelty model by separating candidate recent acquisition or gene-loss material from costly genes that may be retained because laboratory burden does not measure natural-environment value. ₅

The new Conservation vs Fitness — Linking FB Genes to Pangenome Clusters supports this core-versus-variable distinction: among 33 organisms, essential genes were 86.1% core versus 81.2% for non-essential genes, with a median odds ratio of 1.56, and 18 of 33 organisms showed significant enrichment after Benjamini-Hochberg false discovery rate correction. ₆ This refines the HGT interpretation by showing that essentiality is associated with retention in the core, but the pangenome comparison does not directly establish transfer for the genes in the auxiliary or unmapped categories. ₆ Essential-auxiliary and essential-unmapped genes were respectively 38.2% and 44.7% hypothetical, making poorly characterized essential functions a distinct possibility alongside recent acquisition or divergence rather than evidence that all novel genes are mobile. ₆

The new T4SS–CAZy analysis supports this indirect functional evidence with transfer-resolved and genomic-context observations. In 30,497 high-quality environmental MAGs, 6,652 (21.8%) carried T4SS or conjugative machinery; GT2 glycosyltransferase neighborhoods showed 77 detected HGT events, including 32 normalized high-confidence cross-phylum events, with the strongest event spanning 8 phyla. ₇ T4SS-proximal CAZy neighborhoods were also enriched in selected environmental biomes, while T4SS-positive genomes had 10× higher mobile genetic element density than other genomes. ₇ These observations refine the COG-based hypothesis by identifying a possible chromosomal or integrative route for dispersing niche-associated functions, rather than treating mobile-element enrichment alone as evidence of transfer. ₇

Acquisition depth, exchange propensity, and donor direction

Tree-based acquisition-depth inference refines the category-based inference; see also gene-function-acquisition-depth. sankoff parsimony is a tree-based method that minimizes the cost of character-state changes. It attributed 17M gain events on the GTDB-r214 species tree to 5 acquisition-depth bins, which produced distinct signatures by control class. ₂ The recent-to-ancient ratio was 24.5× for CRISPR-Cas, interpreted as an HGT-active signature. It was ~1× for strict housekeeping genes (RNAP core, tRNA-synth, ribosomal), interpreted as vertical inheritance. Mixed regulatory-plus-metabolic KEGG Orthology groups (KOs) had 2× the Innovator-Exchange rate of pure regulatory or pure metabolic KOs. ₂ The Gene Function Ecological Agora lists further class-level ratios: CRISPR-Cas 24.5×, tRNA-synth 2.3×, and β-lactamase 9.0×. The report describes these as clean signatures even though the pooled regulatory/metabolic asymmetry is small. ₈ This supports the two-speed reading with gain-depth attribution rather than category enrichment alone. Parsimony on a species tree infers when gains occurred but does not itself resolve transfer provenance or donors. _2,8 The report names López Sánchez et al. 2026 (DOI 10.1177/15578666261426009; PMID 41955011) as a directly parallel Sankoff-Rousseau algorithm for HGT inference that uses KEGG functions on bacterial species. It judges M22 conceptually similar and proposes benchmarking M22 outputs against that algorithm on a shared substrate. ₈

The pre-registered regulatory-versus-metabolic test (NB11) found regulatory KOs more phylogenetically clumped than metabolic KOs (consumer_z d=−0.211, Mann–Whitney p<10⁻¹⁵, n_reg=1,552 vs n_met=4,964). The report interprets this as less HGT for regulatory KOs. ₈ The “regulatory < metabolic d≥0.3” hypothesis was reframed as a small effect (consumer d=−0.21). In the same comparison, annotation as mobile genetic element (MGE) machinery, meaning transposon, integrase, or phage machinery, was 4.13% for regulatory KOs versus 0.37% for metabolic KOs. The result is direction-consistent with the complexity hypothesis at small effect size, while the strong-form asymmetry is falsified at the d ≥ 0.3 threshold. The project documents both framings. ₈

The report treats this direction as support for Jain 1999’s complexity hypothesis. It states that Burch et al. 2023 (PMID:37232518) independently validated the same direction. As reported, Burch et al. built 74-prokaryote-genome shotgun libraries into E. coli and found that “transferability declines as connectivity increases”, with the strongest effect in translational proteins. ₈ This validation is the report’s characterization of a cited study, not a reanalysis within this corpus. ₈

Exchange propensity also tracked protein architecture. At the atlas level, higher within-KO architectural diversity correlated with elevated cross-clade exchange. ₈ Photosystem II (PSII; photosystem ii), the oxygenic photosynthetic reaction-center complex, had 1 architecture per KO, with clade-restricted flow interpreted as Cyano-internal paralog expansion. The mixed-top-50 subset had a median of 46 architectures per KO and a median dominant fraction of 0.76. The report links this high median diversity to NB11’s 2× Innovator-Exchange rate versus pure regulatory or metabolic categories. ₈ The report presents “a KO’s structural diversity predicts its propensity to flow” as novel at GTDB scale. That novelty claim is the authors’ assessment, and the association does not establish causation. ₈

The report proposes a mechanism: KOs with many possible domain combinations may be more modular and less constrained by host-specific protein-protein interaction networks. This is the complexity hypothesis applied at the architectural level, and it is not directly tested. The report proposed MGE context per gain event (P4-D2) and the Alm r ≈ 0.74 back-test (P4-D3) as formal tests. Its status line now lists both as closed. The Alm 2006 r ≈ 0.74 was NOT REPRODUCED at GTDB scale (P4-D3: r = 0.10–0.29), although methodology generalization held (NB17 architectural concordance r = 0.67). ₈ In the H4 reproduction test, the raw p was 1×10⁻⁴³ and survived Bonferroni correction. The report attributes that significance to large n and judges r=0.10–0.29 too small, so the result is a null for point-estimate reproduction. It frames the project’s intellectual lineage as methodology generalization of Alm 2006, not point-estimate reproduction. ₈ The report also compares this result with Denise 2019, which found that systems encoded in fewer loci were exchanged between taxa more frequently. The two measures are analogous but not identical. ₈

The focal tests separate exchange from donor direction. The pre-registered Cyanobacteria PSII hypothesis tested for high Producer and Participation scores on PSII Pfam architectures. It used a joint Broker-or-Open Innovator-Exchange label and did not distinguish donors. ₈ At the Cyanobacteriia class rank, 21 PSII KOs gave producer d=+1.50 and consumer d=+0.70 (p=2×10⁻⁵ and 2×10⁻⁴). Both criteria passed at α/4 = 0.0125, so H1 was supported; the report’s summary table gives p<10⁻³. ₈ The report also gives 2.77× photic-aquatic enrichment at p<10⁻⁵², a thin BacDive anchor (n=4), and LC=0.88. LC is leaf_consistency, the per-(rank × clade × KO) fraction of species in the clade that carry the KO; the report pairs high LC with low recent gain to mark vertical inheritance. It reports an ancient signal 7× lower than for atlas-wide PSII. _8,2 The report reads the low ancient-gain signal as a donor-origin signature for PSII in or near the Cyanobacteria lineage. That reading draws on cited literature, because the test itself cannot separate Broker from Open donors. ₈ The report cites Benzerara et al. (2026) as direct evidence of recent Cyanobacteria → Gammaproteobacteria HGT of the ccyA gene. It reads that study as independent support for its NB16 framework, in which Cyanobacteria are HGT donors at recent ranks. ₈ Citing Gisriel et al. 2023 (DOI 10.3389/fpls.2023.1289199; PMID 38053766), the report separates within-Cyanobacteria HGT of FaRLiP photosystem I variants from its PSII-core claim. On this reading the study qualifies, but does not contradict, the Cardona 2018 PSII donor-origin framing. Both are characterizations of cited literature, not reanalyses within this corpus. ₈

The Bacteroidota polysaccharide-utilization-locus (PUL) hypothesis failed the pre-registered deep-rank absolute-zero criterion. The report’s headline table pairs the PUL row with the note “Sankoff diagnostic recovers” and d=0.15 (NB08c). However, NB08c reports d=0.15 for normalized Sankoff-score discrimination between positive-HGT control classes (TCS HK, β-lactamase, CAZymes) and housekeeping classes. It is not a Bacteroidota-specific or PUL-specific consumer-z result, and the Phase 1B relative diagnostic found no additional lift for Bacteroidota CAZymes over other phyla’s CAZymes (see below). The attribution of this figure to the PUL hypothesis is therefore the report’s headline framing, not a PUL-specific measurement. Ecological and phenotype anchors were present: 1.40× gut/rumen enrichment (p<10⁻³⁵), a saccharolytic and glycoside-hydrolase-rich phenotype (n=577), and LC=0.41. ₈ One report table labels the result a “qualified pass” and another labels it “falsified” at the absolute-zero criterion. Both labels are preserved here. ₈ The report cites Yu et al. (2025), which found that vertical inheritance dominates the evolution of Bacteroidota alginate utilization loci at the Flagellimonas genus level. The report says this complicates a simple “PUL = HGT” framing. It treats its Phase 1B verdict as consistent with within-Bacteroidota PUL evolution being more vertical than horizontal at fine taxonomic resolution. That verdict is the falsified deep-rank absolute-zero criterion together with what the report describes as a small recovered consumer-z signal, which is subject to the metric caveat above. This is a literature anchor, not a reanalysis. ₈

The deep-rank Producer × Participation categories are Innovator-Isolated, Innovator-Exchange, Sink-Broker-Exchange, and Stable. They are direction-agnostic because donor inference was deferred. The 13.74M-row deep-rank atlas contains 803K Innovator-Isolated, 50K Innovator-Exchange, 742K Sink-Broker-Exchange, 11.83M Stable, and 315K Insufficient-Data classifications. ₈ The family-rank N7 atlas heatmap plots the percentage of (family × KO) tuples that are Innovator-* per top-20 phylum × control class. It shows elevated Innovator-* fractions for AMR and CRISPR-Cas in Cyanobacteriota and Actinomycetota, and low Innovator-* fractions across housekeeping classes. ₈

An exploratory genus-rank, tree-based donor inference (M26) combines the M22 gain attribution, a species tree, and a presence matrix. It assigns Open, Broker, Sink, or Closed labels with explicit ambiguity across 4.06M candidate-set rows. ₈ M26 is distinct from the deferred composition-based M25 method. For each recent-rank gain event into family F for KO K, the candidate donor genera are the genera in F with K present, minus the recipient genus. Each genus × KO pair receives a donor_recipient_ratio and is classified as Open-Innovator if ratio≥2, Broker if 0.5≤ratio<2, or Sink if ratio<0.5. Pairs are Closed-Stable if both counts are close to zero and Insufficient if total events <5. ₈ These assignments are descriptive classifications, not null-hypothesis tests, so no multiple-testing correction applies. ₈ The 4.06M-row output reports Open-Innovator 2.38M, Broker 153K, Sink 182K, and Insufficient 1.34M labels. Confidence was high for 1.36M, medium for 646K, and low for 711K. The passage gives no Closed-Stable count. ₈ Results per genus × candidate-KO tuple were as follows. For Mycobacteriaceae genera (28,389): 75.3% (21,365) Open-Innovator, 13.0% (3,701) Broker, 9.2% (2,619) Sink, and 2.5% (704) Insufficient. For Cyanobacteriia genera (54,582): 47.9% (26,158) Open-Innovator, 3.7% (1,999) Broker, 4.4% (2,395) Sink, and 44.0% (24,030) Insufficient. For Bacteroidota genera (439,968): 56.5% (248,581) Open-Innovator, 4.5% (19,663) Broker, 5.4% (23,612) Sink, and 33.7% (148,112) Insufficient. ₈ The report prints a methodological caveat: the algebraic formula counts every genus with the KO present as a potential donor for the gains of all its family-mates, which biases results toward Open-Innovator dominance. Composition-based confirmation would be needed to separate empirically donor-like genera from merely parsimoniously donor-compatible ones, so the tree proxy is reported as exploratory. The high Open-Innovator fractions therefore should not be read as evidence that these clades are net donors. ₈ Full DTL reconciliation cross-validation remains future work; DTL reconciliation maps gene trees onto species trees through duplication, transfer, and loss events. ₈ Cross-validation against the cited DTL methods (Liu 2021 DTLOR / Bansal 2013) was out of scope, and the report acknowledges this as a reportable limit. ₈ Citing Williams et al. 2024 (DOI 10.1093/ismejo/wrae129; PMID 39001714), the report qualifies its “DTL doesn’t scale to GTDB” framing: principled DTL methods are scaling, but not yet to the full 18,989-leaf tree. ₈ M26 was also not validated against documented HGT cases, namely Cyanobacteria → plastid PSII transfer, AMR plasmid transfers, and ICE-mediated PUL transfers per Sonnenburg 2010. The report calls such validation a fair future-work suggestion that is not in scope for the current synthesis pass. ₈

Composition-based donor inference (M25) was deferred. The per-CDS sequence data needed for codon-usage and k-mer analyses are not available in queryable KBase Data Lakehouse schemas. Phase 3 therefore reports the Innovator-Exchange label without distinguishing broker from open donors. ₈ All Phase 2 tests use atlas-derived scores, and tests against independent experimental HGT datasets were deferred to Phase 4. The report now lists Phase 4 as complete, but its Alm back-test did not reproduce the reference point estimate. The Phase 2 results are therefore graded as internally consistent atlas inference, not externally validated transfer evidence. ₈

Rank-resolved exchange signal and positive-control calibration

The Gene Function Ecological Agora pilot formulated, calibrated, and validated at pilot scale a multi-rank Producer × Participation null-model framework. The producer null is clade-matched to neutral families, and the consumer null is a parent-rank dispersion permutation. ₈ At class rank, UniRef50 dispersion approached the random-permutation null, while vertical inheritance dominated at genus through order ranks. The report calls this the first interpretable biological pattern in the atlas and reads it as consistent with cross-class HGT becoming detectable above noise. That reading is consistent with cross-class transfer but does not directly demonstrate it. ₈ The report places the pattern alongside literature holding that most prokaryotic HGT happens within phylum boundaries (Smillie et al 2011; Soucy et al 2015 review), and that cross-phylum HGT is rarer and biased toward specific function classes (Hooper et al 2007). This alignment is literature context, not an additional pilot measurement. ₈

Under the Phase 1A parent-phylum anchor, AMR showed strongly negative consumer z (−4.4 to −4.8) across genus, family, and order ranks. The report argues that this is not absence of HGT. AMR is a documented intra-phylum HGT phenomenon (Forsberg et al 2012 for the soil resistome; Smillie et al 2011 for the gut microbiome), and a null that measures cross-phylum dispersion makes intra-phylum HGT look clumped. ₈ Forsberg 2012 is cited as context for the intra-phylum AMR HGT pattern that motivated M1. Phase 1A drew principally on Alm 2006, a back-test target deferred to Phase 2/3, and on Smillie 2011 as literature context for the cross-rank consumer-z trend. ₈ M1 stratified the consumer-null parent rank per child rank rather than always using phylum. This exposed a rank gradient in informative consumer z means that the parent-phylum anchor had masked: genus (parent family) −10.48, family (parent order) −6.40, order (parent class) −4.18, and class (parent phylum) −1.84. ₈ This refines the acquisition-depth picture in gene-function-acquisition-depth by showing that clumping weakens toward deeper parent ranks. The gradient remains a dispersion statistic rather than a set of reconstructed transfer events. ₈

The Phase 1B pre-registered hypothesis named Bacteroidota as Innovator-Exchange for PUL CAZyme UniRef50s. Innovator-Exchange meant high producer plus high participation, operationalized as both producer and consumer 95 % CI lower bounds > 0. It was a tested hypothesis, not an established finding. ₈ By that absolute criterion it was falsified at all 4 deep ranks (0/4 supported). ₈ Bacteroidota PUL CAZymes showed producer z ∈ [−0.07, −0.10] across deep ranks, with 95 % CIs entirely below zero, and consumer z ∈ [−9.3, −2.8] at parent ranks, so both legs of Innovator-Exchange failed. By the relative criterion, CAZymes still discriminated from housekeeping, but Bacteroidota CAZymes showed no additional lift relative to the CAZymes of other phyla. ₈ This null supports the failed-criterion account of the PUL test given above, and it is why the d=0.15 Sankoff figure should not be read as a Bacteroidota-specific PUL signal. ₈

Phase 1B added known-HGT positive controls to the consumer-null design because the parent-phylum anchor could mask intra-phylum HGT. The controls were β-lactamase families with documented cross-phylum spread (CARD bla group) and class-I CRISPR-Cas systems (Pfam family), chosen per the cross-tree-of-life HGT documented by Metcalf et al 2014. ₈ These controls all sat at strongly negative absolute consumer z. A relative-threshold diagnostic at family rank nonetheless separated several of them from ribosomal controls (results as median consumer z, Δ vs neg_ribosomal, and one-sided Mann-Whitney U p): pos_amr −7.37, −0.37, p 0.94; pos_tcs_hk −6.57, +0.43, p 5×10⁻⁴; pos_betalac −6.37, +0.62, p 5×10⁻⁸; and pos_crispr_cas −5.81, +1.18, p 1×10⁻⁸⁰. ₈ At the family→order comparison, CAZymes (hyp_cazyme) showed +0.78 σ less clumping than ribosomal proteins (median consumer z = −6.21 vs −6.99; Mann-Whitney U one-sided p = 1 × 10⁻⁴³). ₈ The report interprets β-lactamase and class-I CRISPR-Cas as discriminating from housekeeping by 0.6–1.2 σ. The bakta_amr set did not discriminate in the expected direction. The report attributes this to the set’s inherent narrowness, since it is dominated by clade-specific resistance variants, rather than to methodology failure. ₈ It summarized that at UniRef50 the HGT signal is detectable: CAZymes, β-lactamases, and class-I CRISPR-Cas all showed +0.6 to +1.2 σ less clumping than housekeeping at family rank, with p << 10⁻⁷. ₈

The report corrected the gate-decision framing “HIGH 1 known-HGT controls fail” as overstated. The controls fail the absolute Innovator-Exchange threshold but pass the relative-discrimination diagnostic. The Phase 2 framing therefore became that HGT signal is real at UniRef50 but small in magnitude, with KO aggregation expected, not yet shown, to amplify it. ₈ A later caveat refines that correction. The “+0.6 to +1.2 σ less clumped than housekeeping” framing in REPORT.md v1.2 was expressed in parent-rank z-units rather than Cohen’s d (a standardized mean difference), and it overstated effect magnitude. On Sankoff, Cohen’s d = 0.15, well below the d ≥ 0.3 threshold the project pre-registered for atlas-grade discrimination. ₈

At order rank (parent = class), the diagnostic surfaced an inversion. Positive HGT controls had median consumer z = −4.51 versus −3.69 for negative controls, so they were more clumped, opposite to expectation, with one-sided Mann-Whitney U p = 1.0 (no discrimination in the expected direction). ₈ The report proposed two possible explanations without demonstrating either. First, order rank has only 689 clades and 245 parent classes, which may inflate null variance in the consumer-null permutation. Second, intra-phylum HGT-active genes may cluster within a few orders of a single phylum, making them more clumped at parent = class than housekeeping genes, which spread to many classes. ₈ Diagnostic panel A found median order-rank K = 1 for both positive HGT and negative housekeeping controls. The distributions still differed statistically (p < 10⁻¹⁹), but K bias did not explain the anomaly. ₈ Panel B found lower within-Pseudomonadota dispersion for positive controls (pos median 0.25, neg median 0.33). The report judged this metric biased by class coverage and dropped it, because housekeeping genes occur in every Pseudomonadota family while HGT-active classes are clade-specific within the phylum. ₈ Diagnostic NB08c applied Sankoff parsimony on the GTDB-r214 tree and attributed the inversion to a metric artifact of parent-rank dispersion rather than a methodology failure. It recovered the expected direction, positive HGT > negative housekeeping, at p = 2.1×10⁻⁵. ₈

In the headline panel C, normalized Sankoff scores (parsimony score / n_present_leaves per UniRef) placed positive HGT classes (TCS HK, β-lactamase, CAZymes) above housekeeping classes (ribosomal, tRNA-synth, RNAP). The direction was recovered with Cohen’s d = 0.15, a small effect (p = 2.1×10⁻⁵). ₈ Median Sankoff/n_present values were as follows: random atlas baseline 24.5; pos_tcs_hk 19.75; pos_betalac 19.0; hyp_cazyme 19.0; neg_rnap_core 17.5; pos_crispr_cas 17.0; neg_trna_synth 14.0; neg_ribosomal 12.8, the lowest housekeeping value; pos_amr 12.0; and natural_expansion 4.5. The natural_expansion value is the lowest in the table, which the report attributes to broad conservation by selection. ₈ RNAP core ranking above CRISPR-Cas is a small inversion. The anomalously low pos_amr value is attributed to Pseudomonadota detection bias. AMR was therefore excluded from the Phase 2 positive-control panel, leaving β-lactamase, CRISPR-Cas, and TCS HK. ₈ The report called the substrate-hierarchy claim a partial confirmation. Sankoff restored the expected direction, but the UniRef50 effect size was small (Cohen’s d = 0.146 on Sankoff). Phase 2 KO aggregation had to amplify it to d ≥ 0.3; otherwise an M11 reconciliation-based redesign (gene-tree vs species-tree, AleRax sub-sampling) would be triggered. ₈ These per-UniRef parsimony medians are a different metric from the recent-to-ancient gain ratios reported above, so they neither confirm nor contradict those class-level ratios. ₈

Phase 2 amplification gate and class-level acquisition-depth profiles

The Phase 1B closure refines the calibration above. The report states that the detectable UniRef50 signal is small, well below the absolute Innovator-Exchange threshold that the Phase 1B hypothesis pre-registered. ₈ The report also judged the within-phylum dispersion metric of Diagnostic B wrong, because it was biased by class coverage rather than HGT activity. It named per-phylum Sankoff parsimony as the right within-phylum analog. ₈ Decision M16 therefore made Sankoff parsimony on the GTDB-r214 tree topology the primary atlas metric for Phase 2, and parent-rank dispersion dropped to a diagnostic only. ₈

The closure also states a verdict on the Bacteroidota PUL hypothesis, which supports the failed-criterion reading given above. Bacteroidota → Innovator-Exchange on PUL CAZymes failed at all 4 deep ranks under the pre-registered absolute-zero criterion. ₈ The NB08b relative-threshold diagnostic showed that CAZymes do discriminate from housekeeping (+0.78 σ at family rank, p = 1×10⁻⁴³). The report puts that discrimination at Cohen’s d ≈ 0.07–0.09, below biologically meaningful effect sizes. It states that the hypothesis is genuinely not supported at UniRef50 resolution and that this is not a methodology save. ₈ The closure states that the PUL hypothesis remains falsified by either criterion. On its account, the relative-threshold framing changes only how control-class behaviour is described, not the hypothesis-test outcome. Its M12 relative-threshold framing applies only to control-class methodology validation, meaning whether the method distinguishes HGT-active classes from housekeeping baselines, and not to hypothesis adjudication. ₈ This closure verdict disagrees with the report’s Project Headlines table, which labels the PUL hypothesis itself a “Qualified pass” (deep-rank absolute-zero falsified; Sankoff diagnostic recovers). The report does not reconcile the two verdicts, so both are preserved here. ₈ The report separates this null from reviewer-cited literature on rapid CAZyme HGT. That literature describes transfer at the PUL or family aggregation level, whereas the Phase 1B falsification was at the UniRef50 sequence-cluster level. On the report’s reading, family-level CAZyme HGT exists but operates on aggregated functional units. The difference is therefore one of substrate resolution and need not be a biological disagreement. ₈

Decision M18 turned the expected KO amplification into a hard gate. The Phase 2 KO atlas had to show Cohen’s d ≥ 0.3 on Sankoff parsimony for at least one positive HGT control class versus negative housekeeping. If it did not, the substrate-hierarchy claim would be falsified and the M11 redesign would trigger, switching from Sankoff to gene-tree-versus-species-tree reconciliation per Phase 3’s original sub-sample reservation. ₈ The M18 passage gives the UniRef50 baseline as d = 0.15, and the Phase 2 milestone gives it as d = 0.146 (NB08c). The report labels the gate a falsification gate, not a validated prediction. ₈

The first pass (NB09b) did not clear the gate. Its best comparison, pos_betalac versus neg_trna_synth, gave d = +0.2118 (95% CI [−0.138, +0.531]; one-sided Mann–Whitney p = 1.1×10⁻⁵), and 0/9 pairs passed the threshold of d ≥ 0.3 plus a CI lower bound > 0. ₈ The signed direction reversed against expectation for the ribosomal and RNAP core comparisons (d = −0.66 to −2.97). The report traced this to a description-matched “neg_ribosomal” class dominated by clade-restricted accessory r-proteins rather than universal ones. ₈

NB09c reused NB09b’s per-KO Sankoff scores directly, with no new tree work, and recomputed pairwise Cohen’s d against strict KEGG-KO housekeeping classes. Under that recomputation, 6/9 strict pairs passed at d ≥ 0.3 with a 95% CI lower bound > 0. ₈ The passing contrasts are listed below. Each entry gives the number of KOs in the positive class, the number of KOs in the negative class, d, the 95% CI, and the one-sided Mann–Whitney p.

• pos_crispr_cas vs neg_trna_synth_strict: 6, 20, +3.558, [+2.93, +21.53], 4×10⁻⁶.
• pos_crispr_cas vs neg_rnap_core_strict: 6, 3, +1.905, [+1.47, +11.88], 0.012.
• pos_betalac vs neg_trna_synth_strict: 110, 20, +0.812, [+0.72, +0.95], <10⁻¹⁵.
• pos_betalac vs neg_rnap_core_strict: 110, 3, +0.753, [+0.67, +0.86], 0.0017.
• pos_tcs_hk vs neg_trna_synth_strict: 310, 20, +0.690, [+0.60, +0.88], <10⁻¹⁵.
• pos_tcs_hk vs neg_rnap_core_strict: 310, 3, +0.665, [+0.58, +0.85], 0.0015. ₈ These passes come from redefining the control classes, not from new scores. The limitations differ by contrast. Both CRISPR-Cas contrasts rest on a positive class of 6 KOs and have wide intervals, and every RNAP-core contrast rests on a strict class of 3 KOs. By contrast, the β-lactamase and TCS-HK comparisons against tRNA synthetases have 110 or 310 positive KOs against 20 negative KOs, narrow intervals, and p <10⁻¹⁵. They are the best-supported passes. ₈

All three contrasts against the strict ribosomal class still failed with negative d. Each entry gives the KO counts, d, the 95% CI, and the one-sided Mann–Whitney p:

• pos_betalac vs neg_ribosomal_strict (110 vs 83 KOs): −0.836, [−1.17, −0.52], p 0.87.
• pos_crispr_cas vs neg_ribosomal_strict (6 vs 83 KOs): −0.919, [−1.15, −0.66], p 0.71.
• pos_tcs_hk vs neg_ribosomal_strict (310 vs 83 KOs): −1.705, [−2.15, −1.33], p 0.96. ₈ The report attributes these failures to the control definition rather than to universal ribosomal proteins. The strict ribosomal KO range is a pre-2000 KEGG-numbering approximation that still mixes universal r-proteins with accessory variants. The class is bimodal, with about half of its 83 KOs universal and half clade-restricted. ₈ The report’s final verdict is “M18 PASS on the cleanest housekeeping (tRNA-synth + RNAP core)”. Cohen’s d amplified from 0.146 (UniRef50 NB08c, pos vs all housekeeping) to 0.665–3.558 (KO NB09c, best pairs), which the report describes as a 4–25× amplification depending on positive class. The report concludes that the substrate-hierarchy claim survives at KO resolution. ₈ This refines the d = 0.15 baseline account above. The pass is conditional on a post hoc strict-class redefinition, and the ribosomal comparison stays unresolved because that control class is contaminated rather than clean. ₈

Class-level acquisition-depth profiles support the contrast between HGT-active and vertically inherited classes with full bin distributions. The bins are recent (genus-level gain), older_recent, mid, older, and ancient (phylum-or-above), and total gains are given in parentheses:

• pos_crispr_cas: 58.7, 27.6, 7.0, 4.3, 2.4% (20,898).
• pos_tcs_hk: 45.1, 32.9, 10.6, 7.0, 4.4% (518,106).
• pos_betalac: 44.2, 32.7, 10.7, 7.5, 4.9% (139,000).
• neg_ribosomal_strict: 38.9, 33.3, 12.4, 9.1, 6.3% (79,168).
• neg_rnap_core_strict: 38.5, 34.2, 12.2, 9.0, 6.0% (4,086).
• neg_trna_synth_strict: 24.7, 31.8, 17.5, 15.2, 10.7% (15,060). ₈ The reported recent-to-ancient ratios were:
• CRISPR-Cas: 24.5 (58.7 / 2.4), described as highly recent-skewed and consistent with documented HGT activity.
• Two-component-system histidine kinases (TCS HK): 10.3 (45.1 / 4.4).
• β-lactamases: 9.0 (44.2 / 4.9).
• Strict tRNA synthetases: 2.3 (24.7 / 10.7), which the report calls the most ancient-skewed class and consistent with universal vertical inheritance. ₈ The strict ribosomal and RNAP-core classes fall between the positive controls and tRNA synthetases in recent share, and the passages give no ratio for them. ₈

The report aligns these recent-skewed gains with the literature. It relates them to Smillie et al. 2011 (within-phylum HGT in the human gut microbiome) and to Forsberg et al. 2012 (cross-phylum AMR resistome, with β-lactamase families spread in soil and clinical lineages). It cites Bonomo (2017) for specific β-lactamase families with cross-phylum spread (TEM, CTX-M, NDM, OXA-48). It calls the +44.2% recent / 4.9% ancient signature for pos_betalac (n=139,000 gain events) the GTDB-scale instantiation of this pattern. ₈ Elsewhere the report cites Forsberg 2012 for intra-phylum AMR HGT (see above); both characterizations are preserved as written. ₈ The report aligns the CRISPR-Cas recent skew (58.7% recent; 24.5× recent-to-ancient) with Metcalf et al. 2014’s documentation of cross-tree-of-life HGT for class-I CRISPR-Cas systems. It reads that skew as consistent with dispersal primarily via mobile-element-borne transfers in the recent past. ₈ These alignments are literature context and interpretation; the depth profiles do not identify donors or the vector of transfer. ₈

Both depth results are bounded by M22’s scope. Sankoff parsimony at GTDB scale follows the Mirkin et al. 2003 and Csurös 2010 precedent of using parsimony as an HGT proxy where full DTL reconciliation (AleRax, ALE, Notung) is computationally infeasible. M22 assigns each inferred gain to the rank-LCA, the lowest-common-ancestor rank, of its descendant leaves. The report equates this with the level of phylogenetic-origin claim that Alm 2006 made. ₈ The report states that M22 is recipient-and-depth attribution, not donor-recipient flow: it reports where a gain landed, not who donated it. Two constraints rule out donor inference at deep ranks. Codon amelioration degrades the post-amelioration composition signal for events older than ~10–100 Myr, and per-family DTL reconciliation is infeasible at GTDB scale. ₈

Composite functional categories and transferred modules

Composite COG assignments containing multiple functional letters were retained as biologically meaningful categories rather than treated as annotation artifacts. ₁ The LV composite, representing mobile and defense functions, showed +0.34% enrichment with 76% consistency. ₁ The report interprets this pattern as compatible with multifunctional modules such as mobile defense islands, linking HGT-driven novelty to the module-level-coinheritance of functionally related genes. ₁

Composite categories were counted once per gene rather than split across their component letters. ₁ This choice preserves the possibility that a single gene or module participates in coupled mobile and defense functions, but it also means that the reported composite enrichment should not be interpreted as independent enrichment for each component letter. ₁

The T4SS–CAZy result supports this module-level interpretation: GT2 neighborhoods repeatedly co-localized with T4SS markers, and GH23 occurred 106 times in the parsed GT2 neighborhoods, suggesting association with cell-wall-remodeling functions. ₇ However, the association remains observational and the pending synteny-threshold validation means that it does not yet establish a transferred multifunctional module. ₇

The conservation-fitness synthesis refines this interpretation by showing that coordinated fitness modules are not necessarily accessory or transferred: independent component analysis (ICA), a decomposition method for identifying coordinated fitness modules, identified 1,116 modules across 32 organisms, with 86% core genes versus an 81.5% baseline (odds ratio 1.46, p=1.6e-87), and 59% of modules were more than 90% core genes. ₅ Mobile-defense and transferred-module hypotheses therefore concern a functionally distinctive subset of modules, not module-level coordination in general. ₅

Relationship to bacterial pangenomes

The findings refine pangenome-integration by assigning a functional signature to the distinction between conserved core genes and novel or singleton genes. ₁ They also support genomic-dispersal-functional-coupling, because mobile-element enrichment provides a functional route by which genes can be dispersed among bacterial lineages. ₁ The costly+dispensable analysis supports this pangenome interpretation: costly+dispensable genes had 44.5% orphan genes with no ortholog group, compared with 13.1% among costly+conserved genes, and a median ortholog breadth of 15 organisms versus 31. ₄ The evidence does not establish that every novel gene was horizontally transferred, because the analysis classified genes by novelty and COG category rather than directly reconstructing their evolutionary histories. ₁

The conservation-vs-fitness integration refines this interpretation with explicit linkage limits: 44 of 48 Fitness Browser organisms mapped to pangenome species clades, but only 33 organisms entered the downstream analysis, and pangenome clades with only 2 genomes can have trivially high core fractions. ₆ Thus, the association between essentiality and core conservation strengthens the retained-core side of the model without converting core/accessory status into evidence of HGT. ₆

The T4SS–CAZy analysis supports the proposed coupling between accessory gene content and dispersal by linking GT2 neighborhoods to 32 high-confidence cross-phylum HGT events and to integrative or chromosomal transfer contexts; CAZy genes were not detected on plasmids by ICEfinder, while 12 integrative mobilizable elements occurred among the top 100 accumulators. ₇ This refines the pangenome interpretation toward multiple mobility routes rather than assuming plasmid mobilization, but the report explicitly treats the mechanism as a hypothesis requiring experimental validation. ₇

Co-occurrence and neighborhood evidence supports a mobile-element association without proving transfer; see scale-dependent-mobile-element-associations. Species carrying singleton plant-interaction marker gene clusters were 16x more likely to also carry transposase/integrase singletons (OR=15.95, p=8.8e-20). ₂ Among the neighbors of truly dark genes, 41% were also hypothetical; the digest also reports a 12% figure for proximity within 2 genes of mobile elements. ₂ The project report assigns that proximity figure to the focal genes, not their neighbors: 12.0% of truly dark genes are within 2 genes of a mobile genetic element. ₉ The co-occurrence result suggests horizontal transfer, and the proximity result is used to infer concentration in genomic islands, but neither directly shows transfer. ₂

The Pan-Bacterial AMR Gene Landscape contrasts intrinsic resistance genes with acquired resistance genes such as blaTEM, tet cassettes, and aminoglycoside-modifying enzymes like ant(2”)-Ia. The acquired genes are fully accessory, often 0% core, and appear as singletons or in a minority of genomes. ₁₀ The report calls these horizontally transferred elements that spread under antibiotic selection pressure, which is interpretive. It supports the link between accessory status and acquisition for resistance genes specifically (environmental-resistome). ₁₀

Direct tests of MGE context refine the link between gene gain and mobile elements. MGE-context products occurred in 248,908 of 18,204,007 KO-bearing gene clusters, an atlas-wide baseline of 1.37%. ₈ Regulatory KOs showed a ~10× higher MGE-machinery rate (4.13%) than metabolic (0.37%) or mixed (0.27%) KOs. The report attributes this to regulatory products that include transposon-bound and phage-bound regulators. ₈ The three focal recent-gain sets had near-zero MGE-machinery fractions: Mycobacteriaceae × mycolic acid (6,683) 0.57%, Cyanobacteriia × PSII (236) 0.00%, and Bacteroidota × PUL (1,023) 0.00%. The report judges all three focal KO groups not phage-borne. It bases this on these per-cluster rates plus the PSII gene-neighborhood rate, which was at baseline (10.91% versus a Poisson expectation of 10.6%). ₈

The PSII gene-neighborhood scan covered 27,148 focal features and 218,321 neighbor pairs. The mean MGE-neighbor fraction was 1.65%, and 10.91% of focal features had at least one MGE neighbor within ±5kb. This is approximately the 10.6% Poisson expectation, so no MGE-cargo enrichment was detected. ₈ The corresponding output table has 26,864 rows. The report does not identify that row count as the analysis denominator, and the two counts are not reconciled here. ₈

PUL and mycolic-acid neighborhoods were not scanned; the report judged that such scans would refine the picture but not change the headline. ₈ The report concludes that mycolic-acid, PSII, and PUL genes flow via chromosomal recombination and integrative conjugative elements (ICEs) rather than phage-mediated transduction. This is consistent with cited mechanism literature, but for PUL and mycolic acid it goes further than the direct cargo tests support. For those two gene sets, the route remains a hypothesis (chromosomal-and-integrative-gene-transfer). ₈

Further transfer proxies across projects

The T4SS–CAZy Environmental HGT gene-tree analysis refines the T4SS–CAZy evidence above with event-level detail. Of the 77 HGT events detected in the GT2 gene tree, 32 were normalised high-confidence cross-phylum events; 65 events spanned 2 phyla and 12 spanned ≥3 phyla (15.6%). ₇ The report describes the strongest event, Node_4915, as confirmed. It comprises 35 genes, is 82.9% syntenic, and spans 8 phyla (WOR-3, Desulfobacterota, Patescibacteria, Bacteroidota, Firmicutes_A, Methanobacteriota, Bdellovibrionota, Acidobacteriota), with Max_Divergence = 4.843. ₇ Across events, phylogenetic divergence and syntenic percentage were negatively associated (Spearman ρ = −0.615, p<0.001), which the report reads as sequence divergence after transfer. The phyla most involved across all events were Firmicutes_A (27), Pseudomonadota (22), Bacillota_A (19), and Actinomycetota (10). ₇ These remain gene-tree inferences, and the BLAST and threshold validation listed under Limitations was still pending. ₇

The Truly Dark Genes — What Remains Unknown After Modern Annotation? supplies compositional and positional proxies. GC content deviation from the host-genome mean is significantly higher for truly dark genes (mean |ΔGC| = 0.047 vs 0.038, d = 0.247, p = 1.3e-43). Strong deviation (|z| > 2) affects 9.2% of truly dark genes versus 4.0% of annotation-lag genes. ₉ Among truly dark genes in ICA organisms (4,394 genes; ICA is independent component analysis), several neighborhood patterns appear:

• 41% of neighboring genes are also hypothetical.
• 12.0% are within 2 genes of a mobile element (transposase, integrase, or phage protein).
• 25.9% show operon-like cofitness (|r| ≥ 0.3) with adjacent genes.
• 594 belong to ICA fitness modules. ₉ This scope refines the 12.0% figure cited above. The Key Findings section states it for truly dark genes without qualification, but the Results section restricts it to the 4,394 genes in ICA organisms, so it should not be read as covering every truly dark gene. ₉ The report notes that GC deviation is an imperfect HGT indicator, because it can also reflect gene-specific composition biases (e.g., membrane proteins) or amelioration over time. ₉

The truly-dark report interprets several properties together as a consistent picture of recently acquired, rapidly evolving genes that have outpaced the growth of annotation databases. Those properties are shorter sequences, accessory-genome enrichment, higher GC deviation, proximity to mobile elements, and narrow taxonomic breadth. ₉ The report suggests that the 41% hypothetical-neighbor pattern marks genomic “dark islands”, contiguous regions of unknown function. It proposes, without demonstrating, three likely origins: recently acquired genomic islands, phage-derived regions, or species-specific neighborhoods that arose after divergence from well-studied organisms. It lists characterizing these islands as future work, noting that some may be prophage remnants and others recently acquired metabolic cassettes. ₉ Each line of evidence is indirect, so recent acquisition of truly dark genes is a hypothesis supported by convergent proxies, not a measured transfer history. ₉

The Plant Microbiome Ecotypes — Compartment-Specific Functional Guilds and Their Genetic Architecture evaluated three HGT proxies and got mixed results, which refines the co-occurrence result cited above. Among 1,136 plant-associated species, those carrying singleton marker gene clusters were 16× more likely to also carry transposase/integrase singletons (Fisher’s exact OR=15.95, p=8.8e-20). This proxy drew on 986,464 transposase/integrase singleton clusters (722,674 transposase + 263,790 integrase). ₁₁ However, the overall singleton enrichment ratio for markers was 0.78, so markers are less mobile than the genomic average (Wilcoxon p<1e-300). Cross-species context variation was higher for pathogenic markers (0.534) than beneficial ones (0.450), but not significantly so (Kruskal-Wallis p=0.122). ₁₁ The report suggests that individual HGT events are common while plant-interaction markers as a class are under stronger purifying selection than the genomic average, likely because they are functionally important once acquired. This is an interpretation of mixed proxies, not a demonstration of selection. Without GeNomad mobile-element annotations, HGT was assessed only indirectly, and the report says the mixed signal highlights the limitations of proxy approaches. ₁₁

A per-marker deep dive in the same project found the strongest transposase co-occurrence for plant-growth-promoting (PGP) markers: DAPG biocontrol (OR=8.75), ACC deaminase (OR=6.43), and nitrogen fixation (OR=3.76). Among pathogenic markers, T4SS (OR=3.23), T6SS (OR=2.66), and effectors (OR=2.67) also showed significant enrichment. ₁₁ Contig co-location found 498,677 marker-transposase pairs on shared contigs across 18,569 species. T4SS had the most co-located pairs (84,411) with the closest median distance (238 genes). Fifteen pairs were immediately adjacent (gene-number distance 1); they spanned effectors, T6SS, T3SS, quorum sensing, cellulase, and T4SS, in species including Rhizobium ecuadorense, Phytobacter ursingii, and Burkholderia puraquae. ₁₁ Pangenome position differed by function. Biofilm (83.3% core), IAA biosynthesis (78.1%), nitrogen fixation (72.3%), and phosphate solubilization (70.9%) had the highest core fractions. T4SS (48.5% singleton), coronatine toxin (45.2% singleton), and effectors (39.8% singleton) were strongly accessory or singleton-enriched. ₁₁

The PGP Gene Distribution Across Environments & Pangenomes tested and rejected an HGT-dominated hypothesis for PGP genes. All 13 PGP genes showed significantly higher core fractions than the genome-wide baseline of 46.8% core (chi-square vs baseline; Benjamini–Hochberg false discovery rate, BH-FDR, q < 0.05 for all genes). The mean accessory fraction across PGP genes was 29.7%, compared with 53.2% genome-wide. ₁₂ The report states that its HGT hypothesis (H3) is clearly rejected and that PGP genes are inherited vertically with the core genome. It cites Nascimento et al. (2014), whose phylogenetic analysis showed acdS to be predominantly vertically inherited despite some HGT events. The rejection therefore concerns HGT as the dominant mode, not the occurrence of every transfer. ₁₂ This contradicts reading every ecologically important trait as transfer-driven. The disagreement with the plant-marker transposase co-occurrence is recorded under Tensions. _12,11

The Pangenome Openness Analysis raises, without testing it, the possibility that HGT does not track environmental similarity, because gene acquisition might be opportunistic rather than niche-specific. Its finding that openness is uncorrelated with environment effects is described as consistent with opportunistic rather than ecologically directed HGT. A null association does not establish that mechanism (pangenome-openness-determinants). ₁₃

The Polyhydroxybutyrate Granule Formation Pathways — Distribution Across Clades and Environmental Selection used phylogenetically discordant phaC presence as a transfer proxy (phb-granule-ecology). It identified 311 potential HGT acquisition events, which are species carrying phaC despite being in families with <20% phaC prevalence. It also identified 278 potential HGT loss events, which are species lacking phaC in families with >80% prevalence. ₁₄ Core gene clusters were defined as those present in >90% of genomes within a species, and all others were accessory. The accessory rate in phylogenetically discordant species was 60.1%, versus 32.3% overall. The report presents this as pangenome-scale evidence for ongoing horizontal transfer of phaC, consistent with earlier single-species case studies (Kalia et al. 2007; Catone et al. 2014). ₁₄ The events are potential, not confirmed, and the accessory-rate contrast is indirect evidence. The paired loss count also leaves gene loss as a competing source of discordance. ₁₄

Defense-system distributions give similar indirect support (phage-defense-syndromes-and-arms-race). In the Pan-Bacterial Anti-Phage Defense Arsenal, singleton fractions for accessory-enriched defense systems were 1.5–1.8× above the background 37.9 % baseline. The report reads this as consistent with active horizontal transfer of defense loci through mobile genetic elements, a core prediction of the defense-islands hypothesis of Rocha & Bikard 2022. ₁₅ The SNIPE Defense System in the KBase Data Lakehouse Pangenome describes the patchy distribution of snipe-defense-system across many phyla and families as consistent with horizontal transfer of defense islands. ₁₆ Both are distributional patterns that do not demonstrate transfer, so they support the mobile-defense reading of the COG V and LV enrichments only indirectly. _15,16

Prophage evidence links mobility to repertoire breadth at the species level. In the Prophage-AMR Co-mobilization Atlas, species with higher prophage marker density carried significantly broader AMR gene repertoires (Spearman rho=0.572, p<10^-300, n=4,770 species). A log-log regression found that a 10-fold increase in prophage density predicts a ~6.6-fold increase in AMR breadth (slope=0.823, R²=0.30). The association held after controlling for genome count (partial Spearman rho=0.464, p=1.0×10^-253). ₁₇ The report notes that this is consistent with two non-exclusive mechanisms, and the correlation does not distinguish them. One is direct prophage mobilization of resistance genes via specialized or generalized transduction. The other is a shared high recombination potential for acquiring genes from other mobile elements. The report also states that the effect is consistent across all five major phyla, arguing against a purely phylogenetic explanation. ₁₇ The Prophage Gene Modules and Terminase-Defined Lineages Across Bacterial Phylogeny and Environmental Gradients found module-level environmental effects beyond phylogeny but no lineage-level effects. It proposes modular exchange rather than whole-phage adaptation to environments. ₁₈ Together with the agora null for phage-borne focal innovations, this suggests the hypothesis that phage-associated acquisition matters for resistance breadth even where the PSII, PUL, and mycolic-acid sets showed no MGE-cargo signal. _17,8

Limitations and tensions

COG annotations covered approximately 70% of genes, so unassigned genes may skew the observed functional distributions. ₁ The comparison included 32 species, and a larger sample could reveal phylum-specific patterns that are not visible in the current analysis. ₁ The use of eggnog v6 annotations may produce assignments that differ from original COG assignments. ₁

A further limitation is that mobile-element enrichment is indirect evidence for HGT: mobile functions can facilitate transfer without proving that a particular gene moved between lineages. ₁ The costly+dispensable result is also sensitive to how burden is defined: burden was assigned when max_fit > 1 in any experiment, so a single noisy experiment can classify a gene as costly. ₄ Direct comparison with gene-tree/species-tree discordance, synteny, genomic-context evidence, and transfer networks would therefore be needed to distinguish HGT from other explanations for gene novelty. ₁

The conservation-fitness synthesis adds a related qualification: costly-and-conserved genes are evidence for, rather than a direct measurement of, purifying selection in natural environments, because laboratory fitness and pangenome retention capture different dimensions of selection. ₅ This refines the interpretation of costly-and-dispensable genes as recent acquisitions or genes undergoing loss rather than establishing either history. ₅

The conservation-vs-fitness analysis adds that essentiality was inferred from RB-TnSeq, a random-barcode transposon sequencing approach, under library construction and growth conditions represented in the Fitness Browser; genes without insertions can also be missed because of short length, low-complexity regions, or scaffold edges. ₆ Its slightly shorter essential genes and the exclusion of ten organisms with less than 90% DIAMOND coverage refine the strength of the conservation comparison, while its single-condition measurement leaves stress-specific conservation patterns unresolved. ₆

The T4SS–CAZy findings refine rather than eliminate this limitation: gene-tree incongruence, synteny, and T4SS proximity provide stronger transfer evidence than functional-category enrichment, but the associations remain observational. ₇ Threshold validation, BLAST validation of the strongest GT2 event, a housekeeping-gene null baseline, and biome-enrichment factorization were still pending. ₇

Tensions

Mobile-element primacy versus focal innovations with no detected MGE-cargo signal. The COG analysis interprets the +10.88% COG L enrichment as suggesting that most genomic novelty comes from mobile elements. The costly+dispensable report proposes HGT-mediated genome expansion as the primary source of costly non-conserved genes. _1,4 By contrast, the agora analysis found MGE-machinery rates of 0%, 0%, and 0.57% for the PSII, PUL, and mycolic-acid focal sets. Its PSII neighborhood was at baseline (10.91% versus 10.6%), so no MGE-cargo enrichment was detected for its best-supported cross-clade innovation. ₈ A null enrichment does not exclude transfer by mobile elements. The report’s proposed route for these genes includes ICEs, which are themselves mobile elements, so the disagreement concerns phage or MGE-cargo signatures, not mobility as such. The two lines of evidence also measure different units: novel or singleton genes versus focal KO gains. The sources do not establish what share of novelty each route carries. Because the tension spans three projects, it is a candidate for a dedicated conflict page. _1,4,8

Regulatory genes: more MGE-associated yet less transferred. Regulatory KOs had a higher MGE-machinery rate than metabolic KOs (4.13% versus 0.37%). Yet they were more phylogenetically clumped (consumer_z d=−0.211), which the report interprets as less HGT. The report attributes the MGE rate to transposon-bound and phage-bound regulators. The passages assigned here do not report a per-gain-event MGE result that reconciles the two observations. ₈

Housekeeping acquisition ratio. The digest reports a ~1× recent-to-ancient ratio for strict housekeeping classes that include tRNA-synth. The project report lists tRNA-synth at 2.3×. The sources do not reconcile the class definitions behind these figures. _2,8 The strict-class depth table refines this tension. The project names neg_trna_synth_strict, at 2.3 (24.7 / 10.7), as its most ancient-skewed class. Its neg_ribosomal_strict and neg_rnap_core_strict classes had 6.3% and 6.0% ancient gains, but the assigned passages give no ratio for them. The digest’s ~1× figure for strict housekeeping therefore remains unexplained. _8,2

Bacteroidota PUL verdict: qualified pass or falsified? The report’s Project Headlines table labels the PUL hypothesis a “Qualified pass”. Its Phase 1B closure states that the hypothesis remains falsified by either criterion and limits the relative-threshold framing to control-class methodology validation. The report does not reconcile the two labels. ₈

Plant-growth-promoting traits: vertical core genes or transposase-linked cargo? The PGP project found all 13 PGP genes more core than the 46.8% baseline and rejected its HGT hypothesis, citing predominantly vertical inheritance of acdS. ₁₂ The plant-microbiome deep dive instead found PGP markers with the strongest transposase co-occurrence, including ACC deaminase (OR=6.43) and nitrogen fixation (OR=3.76), although nitrogen fixation was also 72.3% core there. ₁₁ The two projects measure different things: within-species core fraction versus species-level co-occurrence with transposase singletons. The plant report itself suggests that individual HGT events are common while markers as a class are under purifying selection. Neither side establishes what share of PGP carriage arrived by transfer. _12,11

Open Directions

• Expand the comparison beyond the 32 analyzed species and test with phylum-stratified models whether the +10.88% COG L, +2.83% COG V, and +1.64% COG S enrichments remain consistent across additional taxonomic groups. ₁
• Reanalyze the novel and singleton genes with gene-tree/species-tree reconciliation and genomic-context methods to test whether COG L enrichment corresponds to directly inferred HGT events. ₁
• Examine COG L and COG V genes as linked neighborhoods using synteny and module-level coinheritance analyses to test the hypothesis that mobile defense islands explain the +0.34% LV enrichment. ₁
• Join COG assignments to environmental metadata and use habitat-stratified comparisons to ask whether mobile, defense, and unknown-function enrichment varies by habitat. ₁
• Quantify how genes without COG annotation change the inferred novelty partition using alternative annotation resources and sensitivity analyses. ₁
• Reanalyze costly+dispensable genes using gene presence fractions rather than binary core/accessory labels, and transfer-resolved methods, to test whether mobile-element enrichment tracks recent HGT or subsequent gene loss. ₄
• Complete the T4SS–CAZy synteny permutation test, GT2-event BLAST validation, housekeeping-gene null comparison, and biome-enrichment factorization to test whether the apparent cross-phylum transfer signal exceeds threshold, homology, and background-association artifacts. ₇
• Partition the 28,017 costly-and-conserved genes from the 5,526 costly-and-dispensable genes by mobile-element content, ortholog breadth, and transfer-resolved evidence to test whether retention and novelty reflect distinct evolutionary histories. ₅
• Extend the conservation analysis from binary essentiality to condition-specific fitness, especially fitness < -2 under stress conditions, and test whether conditionally important genes show different core, auxiliary, and transfer-resolved patterns. ₆
• Characterize the 3,683 essential-auxiliary genes with ortholog, synteny, and mobile-element analyses to distinguish strain-specific replacements of core functions from recently acquired or unstable genes. ₆
• Run the PUL and mycolic-acid gene-neighborhood scans that the report chose not to run, to test whether the not-phage-borne conclusion holds beyond PSII. Use the closed per-gain-event MGE context outputs (P4-D2) to test whether the regulatory-KO MGE association reflects transfer cargo. ₈
• Cross-validate the M26 Open/Broker/Sink/Closed labels with full DTL reconciliation. Once per-CDS sequence is accessible, add composition-based donor inference to separate Broker from Open donors for the Cyanobacteriia PSII result. ₈
• Score the M26 genus-rank labels on documented transfers (Cyanobacteria → plastid PSII, AMR plasmid transfers, ICE-mediated PUL transfers) to test whether the Open-Innovator-biased tree proxy recovers known donors. ₈
• Benchmark M22 gain attributions against the López Sánchez et al. 2026 Sankoff-Rousseau algorithm on a shared KEGG substrate to test whether acquisition-depth calls are method-robust. ₈
• Run a Bacteroidota-specific normalized-Sankoff comparison of PUL CAZymes against other phyla’s CAZymes, so the PUL headline no longer rests on the pooled NB08c control-class effect. ₈
• Intersect the M22 acquisition-depth bins with the COG L and COG V novel or singleton genes to test whether category-enriched novelty corresponds to recent gain events. _1,2
• Split the bimodal neg_ribosomal_strict class into universal and clade-restricted KOs and re-run the M18 contrasts. This would test whether the negative d against ribosomal controls (−0.836 to −1.705) disappears once accessory r-protein variants are removed. ₈
• Apply per-phylum Sankoff parsimony, the report’s named within-phylum analog, to test whether intra-phylum transfer of AMR and CAZymes is detectable where parent-rank dispersion failed. ₈
• Resolve the 311 potential phaC acquisitions and 278 potential losses with gene-tree reconciliation and synteny to separate transfer from loss. ₁₄
• Re-run the plant-marker HGT proxies with geNomad mobile-element annotations, and test whether the fifteen immediately adjacent marker-transposase pairs lie inside mobile elements. ₁₁
• Intersect the 4,394 ICA-organism truly dark genes with prophage calls and GC-deviation scores to test whether dark islands are prophage remnants or acquired metabolic cassettes. ₉
• Test whether AMR genes lie inside or adjacent to prophage regions across the 4,770 species, to separate direct transduction from shared acquisition propensity. ₁₇