Gene co-occurrence is useful for identifying ecological guilds: recurring combinations of genes can indicate shared environmental selection or complementary ecological roles. However, co-occurrence across genomes does not establish that genes are physically linked in the same operon, genomic island, or locus. 1
Evidence from plant-growth-promoting genes
The PGP Gene Distribution Across Environments & Pangenomes analyzed 13 plant-growth-promoting gene markers across 27,702 species, including 32,736 PGP gene clusters. 1 Across 11,272 species with at least one PGP gene, 8 of 10 focal-gene pairs were significantly associated after Benjamini–Hochberg false-discovery-rate (BH-FDR) correction, with five positive and three negative associations. 1 The report’s co-occurrence heatmap plots log2 odds ratios and marks pairs with FDR q < 0.05. 1
The strongest positive association was between pqqc and acds, with odds ratio (OR) = 7.24, n = 286 co-occurring species, and q = 1.2e-83. The report links these markers to phosphate solubilization (PQQ cofactor biosynthesis) and ethylene reduction (ACC deaminase). 1 The report also says the two genes are nearly always found together when either is present. This page does not adopt that wording. An odds ratio measures relative enrichment, not near-universal overlap, and the report’s own carriage counts do not establish that either gene usually accompanies the other. 1 pqqc also co-occurred significantly with hcnc (OR = 1.91) and ipdc (OR = 1.55), forming a putative non-diazotrophic rhizosphere-effectiveness module. 1
This pattern supports the use of co-occurrence to identify an ecological guild, because the genes recur together more often than expected under the tested association model. 1 It does not prove that pqqc, acds, hcnc, and ipdc occupy the same genomic neighborhood or are transferred as one unit. 1
The negative associations involving nifh further refine the guild interpretation. nifH was negatively associated with hcnC (OR = 0.23, q = 5.8e-29) and pqqC (OR = 0.57, q = 2.9e-19), while its association with ipdC was not significant (OR = 1.13, q = 0.54). 1 These results support ecological separation between diazotrophs and pqqC/acdS-bearing rhizobacteria in this dataset, rather than proving mutually exclusive genomic architectures. 1 The report leaves one question unresolved. It has not established how marine/aquatic diazotrophs (cyanobacteria, Azotobacter), legume-symbiotic rhizobia and free-living soil diazotrophs differ in genome structure and in PGP trait co-occurrence. 1
Only 157 species (1.4%) carried at least three focal traits, indicating that broad multi-trait combinations were uncommon in the analyzed species set. 1 In its trait-combination figure and text, the report calls pqqC + acdS (n = 153) the most common multi-trait genotype. It places nifH + pqqC (n = 225) second. That ranking is internally awkward because the second-ranked count is the larger one. The report attributes the nifH + pqqC co-occurrence, despite the negative overall association, to some generalist lineages. 1 The same report gives pqqC–acdS co-occurrence as n = 286 in its association section but n = 153 in its trait-combination section, and it does not explain the difference. 1 This illustrates why pairwise association statistics and genotype counts should be interpreted together rather than treated as interchangeable evidence of a single guild. 1
The central discoveries digest summarizes the same project. In its account, pqqC and acdS co-occur at OR = 7.24 across 27K GTDB species, and both genes are soil/rhizosphere enriched (acdS OR = 7.0, pqqC OR = 2.9). All 13 PGP genes are predominantly core, at 29.7% mean accessory versus 53.2% genome-wide, and nifH forms a separate ecological guild. 2 However, the digest misstates the trp→ipdC coupling by assigning OR = 2.81 to non-soil species. The project report gives OR = 2.81 overall, OR = 3.56 in non-soil species (n = 10,233, p = 7.7e-13), and OR = 0.30 in soil species (p = 0.02). 2,1 Only 214/11,272 species (1.9%) carry ipdC, so the report treats the soil reversal as hypothesis-generating rather than conclusive. The digest explains the reversal by soil PGPB acquiring tryptophan exogenously from root exudates, and that explanation is untested. 2,1
Environmental support and qualification
The co-occurrence pattern was accompanied by environmental differentiation. Among 1,039 soil/rhizosphere species and 10,233 species from other environments, acdS prevalence was 15.8% versus 2.6% (OR = 7.02, q = 5.1e-62), pqqC prevalence was 43.8% versus 21.2% (OR = 2.90, q = 2.8e-53), and hcnC prevalence was 11.3% versus 6.4% (OR = 1.85, q = 6.1e-08). 1 nifH was depleted in soil-classified species, with prevalence of 12.9% versus 19.7% (OR = 0.60, q = 5.5e-08). 1 The report also highlights hcnC as significantly soil-enriched (OR = 1.85) and as co-occurring with pqqC (OR = 1.91). On this basis it proposes the hypothesis that the hcnA–hcnC operon belongs to a rhizosphere niche toolkit. That role is an inference that needs focused study of hcnA-C phylogeny and of co-occurrence with biocontrol phenotypes. 1
The acdS enrichment remained strong after phylum-level fixed effects in logistic regression (OR = 6.98, p = 4.8e-61) and in a strict rhizosphere-only sensitivity analysis using genomes whose isolation source contained “rhizosphere” or “root nodule” (OR = 10.6, q = 7.6e-38). 1 The agreement between gene co-occurrence and environment-stratified prevalence supports an ecological-guild interpretation more strongly than co-occurrence alone. 1
The Ecotype Correlation Analysis refines this interpretation by showing that environmental similarity was generally weak for whole-genome gene-content similarity: across 172 species, phylogeny generally dominated, with a median partial correlation of 0.0025 for environment versus 0.0143 for phylogeny, and no significant environmental effect in 156 species (90.7%). 3 This does not contradict the gene-specific environmental enrichments above; instead, it supports the hypothesis that ecological selection may be concentrated in particular gene subsets rather than expressed as a strong whole-genome signal. 3
The Pangenome Openness Analysis supports this qualification: among species with both pangenome statistics and ecotype-analysis results, pangenome openness was not significantly related to either environment effect (Spearman rho = -0.05, p-value = 0.54) or phylogeny effect (Spearman rho = 0.03, p-value = 0.73). 4 Thus, an open or closed pangenome did not predict which broad driver dominated gene-content variation, reinforcing that gene-specific ecological signals should not be inferred from a single whole-pangenome summary metric. 4 The report further treats the possibility that core/accessory labels miss functionally adaptive genes, or that horizontal gene transfer is opportunistic rather than environmentally patterned, as hypotheses rather than demonstrations. 4
The Plant Microbiome Ecotypes — Compartment-Specific Functional Guilds and Their Genetic Architecture refines the complementary-function interpretation: corrected analysis of plant-associated co-occurring genus pairs found them slightly less complementary than random pairs, with Cohen’s d ≈ −0.4 and permutation p < 0.001; prevalence-weighted aggregation gave d = −0.39. 5 In absolute terms, the 1,048 co-occurring pairs had mean complementarity of 10.72. A richness-preserving permutation null of 1,000 iterations gave 12.74±0.27. 5 Thus, gene co-occurrence can still mark shared selection or ecological guild structure, but co-occurrence should not be treated as evidence that partners provide complementary metabolic functions. 5
BacDive metabolite-utilization data agreed with GapMind predictions at 83.1% consistency, supporting the reliability of the pathway completeness scores, although the GapMind core-level completeness score was 0% across compartments and may be too stringent for broad taxonomic comparisons. 5 The refined C-score test was structurally underpowered: among 69 NMDC co-occurring genera, there were 0 PGP-dominant, 3 pathogen-dominant, and 66 dual-or-mixed genera. The report records this refined-cohort result as the same pattern seen in Phase 1. 5 This supports retaining prevalence- and power-aware null models when translating co-occurrence into functional complementarity claims. 5
Evidence from Co-fitness and Co-inheritance in Bacterial Pangenomes further refines this gene-specific interpretation: across 9 organisms, pairwise laboratory co-fitness produced only a weak positive co-occurrence signal, with mean delta phi (cofit minus prevalence-matched random pairs) of +0.011 across organisms and aggregate delta = +0.003, despite aggregate Mann-Whitney p=1.66e-29; the across-organism Wilcoxon signed-rank test was not significant (W=9, p=0.13). 6 Thus, co-occurrence can support a guild hypothesis, but pairwise functional similarity is not a uniformly strong predictor of genomic co-occurrence across taxa. 6
The cofitness report also tabulated SEED functional categories for pairs in the top quartile of both co-occurrence and co-fitness. Co-occurrence here was measured by the phi coefficient, a binary-association measure. The leading categories, given as gene counts rather than pair counts, were:
- Metabolism (15,936)
- Transport (11,940)
- Regulation (6,339)
- Motility (2,929)
- Mobile elements (2,716)
- DNA metabolism (2,490) 6
These values are gene counts, not pair counts. The report’s Figure 6 labels them as functional categories enriched among high-phi, high-cofit pairs. 6 The assigned evidence does not include the background distribution behind that enrichment label. The table therefore shows which functions populate these pairs, and should not be read as a demonstrated enrichment. 6
The same report places its module-level result in literature context. Hall et al. (2021) found that escherichia-coli accessory genes co-occur by function and via mobile genetic elements (PMID: 34499026). Choudhury et al. (2025) showed that pseudomonas-aeruginosa phylogroups have distinct co-occurrence networks, which suggests that accessory genome structure is lineage-specific (PMID: 40304385). 6 These are cited prior findings, not measurements made by this project. The second supports the lineage-structure caution under Interpretation limits below. 6
This pattern connects to ecotype-environment-gene-content, because the gene combinations and their negative or positive associations corresponded to different environmental distributions. 1 It also complements module-level-coinheritance, while leaving open whether the observed module is physically coinherited or instead assembled repeatedly by ecological selection. 1
Odds ratios versus co-occurrence share
The ENIGMA Carbon Census — Tiered Knowledge Census of 83 Enrichment Compounds shows that the choice of co-occurrence statistic changes what a guild claim means. Its first analysis used the phi coefficient. φ compresses toward zero when a capacity is rare, and most capacities were present in tens to hundreds of 3109 genomes. A reanalysis therefore recomputed two measures from saved 2×2 counts: a Haldane-corrected odds ratio (an odds ratio with a small-count correction) with 95% confidence interval, and the Jaccard index. 7 The genuine modularity test compared cross-block pairs, meaning pairs with distinct lower pathways. These had median OR 2.28, with 4/15 confidence intervals entirely above 1, but median Jaccard ≈ 0. Within-aromatic pairs had median OR 1.12, with 3/10 intervals entirely > 1 and median Jaccard 0.026. The within-other median OR of 7.43 (1/3) was driven by Haldane correction on near-zero counts. 7
The odds ratio surfaced one cross-block signal that φ had hidden: phenylethylamine co-occurring with aromatic-funnel acids.
- 4 hydroxybenzaldehyde + phenylethylamine: OR 2.84 (95% CI 1.49–5.40, q=0.035, n11=11).
- phthalic acid + phenylethylamine: OR 3.37 (95% CI 1.46–7.75, q=0.071, n11=6). 7
The report reads this overlap as mechanistically coherent. Phenylethylamine is itself aromatic-derived: it is deaminated to phenylacetate and enters the paa/phenylacetyl-CoA route. 7 Even for these enriched pairs the Jaccard stayed ≈ 0.04. Only 25 of 3109 genomes (0.8%) carried both an aromatic and a non-aromatic capacity. The report summarizes the phenylethylamine×aromatic association as OR ≈ 2.8–3.4. It concludes that there is no broad cross-module modularity, so versatility is a clade trait rather than a portable module set. 7 This refines the PGP odds-ratio evidence above. A large multiplicative enrichment can coexist with a tiny share of genomes carrying both traits. Odds ratios and share-based measures should therefore be reported together before co-occurrence is read as a co-assembled guild. 7,1
Taxon-level guild co-occurrence
The SSO Subsurface Community Ecology — Spatial Structure, Functional Gradients, and Hydrogeological Drivers applies co-occurrence at the community scale. It assigned 65 annotated genera to 11 metabolic guilds and analyzed guild co-occurrence across 9 wells. These correlate inferred guild abundances rather than genes within genomes, and they do not demonstrate interactions. 8 It reported three positive correlations:
- Nitrifier × iron oxidizer (ρ = +0.95): both concentrated at U3, the plume entry. The report interprets both as chemolithotrophs exploiting reduced compounds (NH₄⁺, Fe²⁺) arriving in the plume.
- Syntroph × fermenter (ρ = +0.55): interpreted as anaerobic food-web coupling.
- Fermenter × predator (Bdellovibrio) (ρ = +0.85): prey tracking is inferred. 8
The report also found two negative correlations. Denitrifier × syntroph (ρ = −0.67) is consistent with separation between nitrate-available and strictly anaerobic conditions. Sulfate reducer × aerobic heterotroph (ρ = −0.75) is consistent with the two guilds occupying opposite ends of the electron-acceptor hierarchy. 8 The report presents the nitrifier–iron-oxidizer coupling (ρ = +0.95) and the denitrifier–syntroph exclusion (ρ = −0.67) as a functional network topology overlaying the spatial redox gradient. 8 This supports the guild half of this page’s thesis at a different scale. Here, recurring co-occurrence tracks shared or exclusive niches, much as the PGP nifH separation does among genomes. It also connects to subsurface-hydrogeological-zonation. 8,1
These correlations do not demonstrate interactions. The report interprets syntroph–fermenter coupling (ρ = +0.55) through prior syntrophic theory (McInerney et al. 2009). It interprets denitrifier–syntroph exclusion (ρ = −0.67) through canonical redox zonation of contaminated aquifers (Chapelle 2001). The correlations alone do not establish obligate partnerships or causal electron-acceptor separation. 8 The Genotype × Condition to Phenotype Prediction from ENIGMA Growth Curves makes the same caution explicit. Spearman co-occurrence suggests shared or exclusive niches but does not prove direct interaction. That report suggests that compositionally aware SparCC analysis on the full 100WS ASV matrix would strengthen such findings. 9
Co-occurrence is not physical linkage
The report explicitly notes that co-occurrence does not establish physical linkage. 1 The pqqC–acdS association could reflect a shared habitat, complementary functions, correlated lineage structure, repeated independent acquisition, or physical linkage, and the reported analysis does not distinguish among these mechanisms. 1 The report lists one test as a future direction. That test would determine whether pqqC and acdS are physically co-located on the chromosome, in the same genomic island or operon, or co-occur between separate loci. This would distinguish functional operon linkage from independent co-selection. 1
The cofitness analysis supports this caution: only 0.7% of its cofit pairs were genomically adjacent within 5 genes, and excluding adjacent pairs did not change the result pattern. 6 Its stronger signal arose instead at the coordinated multi-gene level: across 195 independent component analysis (ICA) modules in 6 organisms, within-module co-occurrence exceeded a prevalence-matched null by delta phi = +0.053; 51/195 modules were significant at p<0.05 and 21/195 remained significant at q<0.05 after BH-FDR correction. 6 This supports interpreting recurring gene sets as possible selective or ecological units without equating statistical co-inheritance with a physically linked cassette. 6
The core/accessory results also caution against assuming that a co-occurring pair is a horizontally transferred cassette. 1 All 13 PGP genes had higher core fractions than the 46.8% genome-wide baseline, and pqqC was 81.5% core, 7.8% auxiliary, and 10.7% singleton, whereas acdS was 70.4% core, 13.4% auxiliary, and 16.2% singleton. 1 Thus, the pqqC–acdS guild signal is compatible with predominantly conserved gene distributions rather than requiring a single mobile element. 1 The null relationship between openness and broad environment or phylogeny effects supports treating these core/accessory summaries as insufficient on their own to identify the adaptive or mobile basis of a guild. 4 The project report and the discoveries digest agree here. Both reject the hypothesis that PGP genes are predominantly accessory (the report’s H3) and conclude that vertical inheritance dominates, because the genes are predominantly core. 1,2 This page adds a methodological caution of its own, which is not a disagreement between the sources. A predominantly core distribution does not settle whether pqqC and acdS are physically linked, and the report itself leaves their genomic co-location open. 1
The Plant Microbiome Ecotypes — Compartment-Specific Functional Guilds and Their Genetic Architecture refines this conclusion by showing that HGT indicators are scale-dependent: species carrying singleton plant-interaction marker clusters were 16 times more likely to carry transposase or integrase singletons (Fisher OR = 15.95, p = 8.8e-20), yet marker singletons had an overall enrichment ratio of 0.78 relative to the genomic average. 5 At genus scale, plant-associated genera had a median 3.7 mobile elements per genome versus 2.8 in non-plant genera (Mann–Whitney p = 1.49×10⁻⁵). 5 These mixed scales support testing mobility directly rather than inferring a transferred cassette from co-occurrence.
The T4SS–CAZy Environmental HGT gives a related neighbourhood-level association. It compared MAGs (metagenome-assembled genomes) with and without gt2 glycosyltransferases in type-iv-secretion-system-proximal neighbourhoods. GT2-neighbourhood MAGs (n=376) had mean 0.045 metal resistance types, versus 0.004 for non-GT2 MAGs (n=260,276; Mann-Whitney p=8.6e-27). The report separately describes these genomes as carrying 11× more metal resistance genes. 10 The report links this result to a metal-resistance niche-breadth hypothesis and calls these genomes hubs for GT2 horizontal transfer. That hub description is interpretive. Genome-level co-carriage of a syntenic CAZy–T4SS neighbourhood and metal resistance functions does not by itself show co-transfer. 10
Single-species co-annotation is weaker evidence still. In the SNIPE Defense System in the KBase Data Lakehouse Pangenome, klebsiella was the only phagefoundry species with both the snipe-defense-system and the PF00358 ManX-family PTS domain, matched by exact Pfam accession. With one species, this co-annotation can motivate a functional or physical pairing but cannot test one. 11
The ICA results refine this conclusion by showing stronger co-inheritance for accessory modules: accessory modules had mean delta phi +0.108, compared with +0.059 for core modules and +0.031 for mixed modules, although the accessory-versus-core difference trended toward significance (Mann-Whitney p=0.051). 6 This makes accessory gene sets plausible candidates for ecological guilds or coordinated selective units, but does not establish that their members share physical loci. 6
The analysis also identified pqqD as an exception, with 55.5% core and 27.5% singleton status, suggesting that pqqD can sometimes spread as a standalone gene. 1 This distinction between statistical co-occurrence, shared ecological function, and physical coinheritance is central to interpreting gene-cooccurrence-ecological-guilds. 1
Interpretation limits
The environmental classification was conservative and noisy: 1,637 species (5.9% of species with an environment label) were classified as soil/rhizosphere dominant, while 291,279 genomes had isolation-source metadata and 93.5% were classifiable. 1 NCBI sampling is biased toward clinical and host-associated genomes, so the observed acdS and pqqC enrichment may underestimate true rhizosphere enrichment. 1
The ecotype analysis further refines this limitation: AlphaEarth embeddings covered only 28.4% of genomes, and geographic coordinates were often missing or imprecise; for host-associated organisms they may describe collection sites rather than biologically relevant microenvironments. 3 Its partial-correlation framework also assumed linear relationships between distance matrices, so weak whole-genome environmental effects may reflect both biology and limited environmental representation. 3
The pangenome-openness analysis adds that its sample was limited to species with both pangenome statistics and ecotype-analysis results, and that openness is only a single summary metric; the upstream ecotype analysis also had limited power for some species with few genomes. 4 These caveats support interpreting its null correlations as limits on prediction by this metric, not as proof that pangenome structure is independent of ecological dynamics. 4
The cofitness study adds a related prevalence limitation: its strongest co-fitness pairs were often core genes with near-universal prevalence, producing a prevalence ceiling with little co-occurrence variance; co-fitness strength was weakly anti-correlated with co-occurrence (Spearman rho=-0.109, p<1e-300 across 1.04M pairs). 6 This supports treating accessory enrichment and prevalence-matched nulls as important qualifications when translating pairwise co-occurrence into ecological claims. 6
The plant-microbiome analysis supports an additional caution about marker-defined guilds: 878 of 1,115 plant-associated species (78.7%) were classified as dual-nature under refined marker rules, but categorical cohorts did not usefully discriminate a curated validation panel; all 14 known beneficial or pathogenic species were dual-nature and all four neutral controls were misclassified. 5 This supports interpreting marker co-occurrence as a coarse ecological screen, not as proof of a unified phenotype or co-expressed guild.
PGP detection relied on Bakta annotations matching exact gene names such as nifH, acdS, and pqqC, so product-only annotations and variant gene names could have been missed. 1 Gene clusters were not functionally validated for truncations, frameshifts, or pseudogenization, meaning that an annotated co-occurring cluster was not necessarily functional. 1
Phylogenetic structure and uneven sampling can also contribute to apparent co-occurrence, so the guild interpretation should be tested with lineage-aware models and genome-context analyses. 1 The cofitness analysis likewise found higher mean phi among near genomes than medium-distance genomes (0.102 versus 0.067), while most species lacked genomes in the far stratum (>0.05 branch distance), limiting separation of functional coupling from shared ancestry. 6 These limitations connect the concept to phylogenetic-confounding-of-pangenome-associations, environmental-resistome, and evidence-triangulation-for-functional-annotation.
Tensions
The PGP analysis supports a pqqC–acdS ecological-guild interpretation through strong pairwise association and environmental enrichment, while the plant-microbiome complementarity analysis found co-occurring genus pairs slightly less complementary than random pairs (Cohen’s d ≈ −0.4, permutation p < 0.001). 1 5 These findings are not mutually exclusive: the tension is whether a guild denotes shared environmental selection or complementary metabolic provisioning, and it should not be resolved without pathway- and activity-level tests.
The pangenome-openness analysis found no significant relationship between openness and either environment or phylogeny effects (Spearman rho = -0.05, p-value = 0.54; Spearman rho = 0.03, p-value = 0.73), whereas the PGP analysis found strong environment-associated enrichment for particular genes, including acdS and pqqC. 4 1 This is a scale and representation tension rather than a direct contradiction: a whole-pangenome openness metric may fail to predict broad eco-phylogenetic structure while selected gene subsets retain environmental associations.
Open Directions
- Use complete or well-assembled genomes from the pqqC–acdS-positive set, together with operon and genomic-island analysis, to test whether the two genes are physically linked or occur at separate loci. 1
- Apply phylogeny-aware co-occurrence models to the PGP gene matrix and ask whether the pqqC–acdS association persists after accounting for shared ancestry. 1
- Compare gene neighborhoods, synteny, mobile-element context, and gene-tree reconciliation for pqqC, acdS, hcnC, and ipdC to distinguish vertical conservation from repeated horizontal acquisition. 1
- Reanalyze nifH associations after stratifying by environment, lineage, and diazotrophic ecological context to determine whether the negative nifH–pqqC and nifH–hcnC associations reflect ecological separation or sampling structure. 1
- Functionally validate representative co-occurring clusters with transcript or phenotype data to test whether annotated pqqC, acdS, hcnC, and ipdC genes are active components of the proposed guild. 1
- Reanalyze gene-specific guild associations with alternative environmental distances and direct metadata, then compare them with whole-genome and ecotype-cluster contrasts to test whether environmental signal is concentrated in selected loci. 3
- Stratify pangenome openness analyses by gene function, including L (mobile) and V (defense) categories, and test auxiliary fraction, Heap’s law alpha, and pangenome fluidity to determine whether specific functional subsets recover environment-associated guild signals. 4
- Test openness-by-lifestyle interactions, such as open pathogen versus open environmental species comparisons, rather than treating openness as a universal predictor. 4
- Restrict co-fitness comparisons to auxiliary-only pairs below 95% prevalence and test whether the pairwise guild signal strengthens after removing the prevalence ceiling. 6
- Build module co-transfer networks and compare their accessory-module membership with environmental gene guilds, testing whether multi-gene co-inheritance predicts ecological distribution better than pairwise co-occurrence. 6
- Test the plant-associated co-occurrence and complementarity tension with reaction-level pathway comparisons, transcript measurements, and direct mobile-element or gene-neighborhood data. 5
- Recompute Haldane-corrected odds ratios with 95% confidence intervals and Jaccard indices for the PGP gene pairs, as in the carbon-census reanalysis. This would test whether the pqqC–acdS enrichment reflects a substantial share of carriers or only a small overlap, as with the aromatic cross-block pairs. 7,1
- Reconcile the differing pqqC–acdS counts (n = 286 versus n = 153) and the nifH + pqqC ranking directly from the species-by-gene matrix before using genotype counts as guild evidence. 1
- Test the ipdC soil reversal with expanded ipdC-positive sampling, since only 214/11,272 species carry ipdC, and with direct tryptophan-availability data for the untested root-exudate explanation. 1,2
- Run compositionally aware SparCC on the full 100WS ASV matrix to test whether the well-level guild correlations (for example, nitrifier × iron oxidizer and denitrifier × syntroph) persist under compositional correction. 9,8