Cofitness Network Architecture, Organism Specificity, and Shared Dispensability

AMR Co-Fitness Support Networks examines antimicrobial-resistance (AMR) gene support networks across 28 organisms using Fitness Browser fitness matrices, independent component analysis (ICA) fitness modules, AMR catalogs, and interproscan annotations. Within-Species AMR Strain Variation tests lineage, co-inherited resistance islands, and environment; Aromatic Catabolism Support Network in ADP1 extends the question to a condition-specific support network around beta-ketoadipate-pathway; Rational Design of Protective Microbiome Formulations for Competitive Exclusion of *Pseudomonas aeruginosa* in Cystic Fibrosis Airways tests organism- and condition-specific interactions in a multispecies formulation against pseudomonas-aeruginosa; and Co-fitness and Co-inheritance in Bacterial Pangenomes tests whether laboratory cofitness predicts gene co-occurrence across genomes. Gene Conservation, Fitness, and the Architecture of Bacterial Genomes integrates RB-TnSeq (random barcode transposon sequencing) fitness data for 194,216 protein-coding genes across 43 bacteria with pangenome conservation and ICA modules. ₁ Counter Ion Effects on Metal Fitness Measurements tests whether shared NaCl responses in metal-fitness profiles reflect counter-ion exposure or broader shared-stress biology. The Pan-Bacterial Essential Genome links essentiality to orthology, pangenome conservation, and ICA module transfer across 221,005 genes from 48 bacteria. ₂ discoveries consolidates evidence on module stability, condition-specific trade-offs, annotation effects, and shared-stress controls. Pan-bacterial Fitness Modules via Independent Component Analysis evaluates ICA for pan-bacterial fitness modules across 32 organisms. Functional Dark Matter — Experimentally Prioritized Novel Genetic Systems and Truly Dark Genes — What Remains Unknown After Modern Annotation? extend the architecture analysis to experimentally actionable genes of unknown function. _3,4 Gene Function Ecological Agora extends it to clade × function producer–participation classes, Sankoff-parsimony acquisition depth, ecological metadata, phenotype records, and gene neighborhoods. Gene-Resolution Metal Cross-Resistance Across Diverse Bacteria adds gene-resolution evidence from 452 metal experiments across 37 organisms and 14 metals. ₅ Fitness Modules × Pangenome Conservation tests whether ICA module genes and cross-organism module families occupy the bacterial core genome. Pan-Bacterial Anti-Phage Defense Arsenal adds pan-bacterial evidence on coordinated anti-phage defense architecture, accessory acquisition, and prophage-associated arms-race signals.

Cofitness measures similarity in genes’ fitness profiles across experimental conditions. Direct co-regulation would require evidence that genes share regulatory machinery; therefore, a cofitness edge is evidence of a shared phenotypic response, not proof of transcriptional control. ₆ The integrated evidence supports structured, condition-dependent organization of fitness phenotypes, while leaving direct co-regulation versus shared dispensability unresolved.

Network Architecture

Among 801 AMR genes with fitness data, 192 (24%) were assigned to ICA fitness modules. ₆ AMR-containing modules had a median of 46 genes versus 27 in non-AMR modules (Mann–Whitney U p = 1.7×10⁻⁸); 136 unique module families contained AMR genes, and 208/209 (99%) AMR gene-module assignments belonged to cross-organism conserved families. ₆ Efflux and enzymatic AMR modules had the same median size of 48 (p = 0.91). ₆

The pairwise analysis found that 769/801 AMR genes (96%) had at least one extra-operon cofitness partner at |r| > 0.3. ₆ The dataset contained 180,370 total cofitness partners, of which 179,375 were extra-operon; only 0.6% were excluded as near-operon pairs. ₆ Mean support-network sizes were 233 genes at |r| > 0.3, 110 at |r| > 0.4, and 71 at |r| > 0.5. ₆ The report’s figure cofitness_threshold_sensitivity.png plots network size against cofitness threshold. ₆ These associations support a structured architecture but do not establish direct regulation: genes can share condition-responsive fitness patterns because both are dispensable in one condition and important in another without sharing a transcription factor or signaling pathway. ₆ Two caveats qualify these counts. First, the report notes that the 233-gene mean network at |r| > 0.3 includes many weak associations and calls for confirmation at |r| > 0.4. Second, the 5-ORF operon exclusion zone used matrix row index as a proxy for genomic proximity, which may not reliably reflect chromosomal gene order; coordinate-based exclusion using the loaded begin, end, and strand columns remains future work, and the current heuristic excluded only 0.6% of pairs. ₆

These results support an architecture in which AMR genes participate in large, conserved, condition-associated modules rather than only mechanism-specific neighborhoods. ₆ Robust ICA and DBSCAN (density-based spatial clustering) across 32 bacteria produced 17–52 modules per organism; 94.2% had significantly elevated within-module cofitness, with mean |r|=0.34 versus background 0.12 and mean genomic-adjacency enrichment of 22.7×. ₇ Absolute ICA-membership thresholding increased enrichment from 59% to 94.2% and within-module |r| from 0.047 to 0.34. ₇ Using |weight| >= 0.3 and a maximum of 50 genes, the analysis identified 1,116 stable modules; 94.2% showed elevated cofitness by Mann-Whitney U at p < 0.05, and within-module mean |r| was 0.34 versus 0.12 in background, a 2.8x enrichment. ₈ The 22.7x adjacency enrichment reinforces operon co-localization while showing that cofitness is not reducible to adjacency. ₈ A single-operon example from SNIPE Defense System in the KBase Data Lakehouse Pangenome supports operon-level coherence: manxyz co-fitness correlations were manX↔manZ r=0.851, manX↔manY r=0.705, and manY↔manZ r=0.725. The report interprets these correlations as confirming that the three genes function as a single operon. ₉ Cross-organism alignment found 156 families spanning at least 2 organisms, 28 spanning 5+ organisms, and one spanning 21 of 32 organisms. ₇

The threshold analysis refines module-stability interpretation: an initial D’Agostino K-squared membership approach produced 100-280 genes per module with 59% enrichment and 1-17x correlation, whereas absolute thresholds produced biologically coherent modules with 94% enrichment and 2.8x correlation enrichment. ₈ A 40% component cap was necessary to avoid FastICA convergence failures but may miss modules in organisms with few experiments. ₈

Module-conservation analysis supports and extends this architecture from module detection to genomic conservation. Module genes were 86.0% core versus 81.5% for all genes, with OR=1.46 and p=1.6e-87; the absolute enrichment was +4.5 percentage points. ₁₀ Among 974 modules with >=3 mapped genes, 577 (59%) were core modules (>90% core genes), 349 (36%) mixed (50-90%), and 48 (5%) accessory (<50%); the median module was 93.4% core. ₁₀ This supports the inference that conserved fitness architecture is not limited to individual genes and refines it by showing that most mapped co-regulated units are embedded in the conserved genome rather than broadly accessory. Family breadth did not predict conservation (Spearman rho=-0.01, p=0.914), so breadth across 2, 5+, or 21 organisms does not imply a higher core fraction. ₁₀ Thirty-eight families had <50% core genes; their possible horizontal-transfer or niche-specific status remains a hypothesis because conservation patterns alone do not measure transfer or regulation. ₁₀

The metal study supports and generalizes conserved condition-dependent architecture beyond AMR. Across 317 organism–metal pair observations involving 28 organisms and 85 unique metal pairs, 98.1% of gene-level fitness correlations were positive (311/317), 99.1% significant (p < 0.05), and all 15 pairs tested in at least 5 organisms showed >90% sign consistency; no pair showed systematically negative cross-resistance. ₅ Strongest associations included Fe-Zn (mean r = 0.61, n = 6), Co-Ni (r = 0.56, n = 28), Co-Zn (r = 0.52, n = 18), Ni-Zn (r = 0.51, n = 18), Cu-Zn (r = 0.48, n = 16), and Co-Cu (r = 0.43, n = 24), while Al was more independent (mean r = 0.34; Al-Co r = 0.30). ₅ Leave-one-out consensus r = 0.41 and Mantel mean r = 0.23 indicate moderate chemistry-specific conservation. ₅

The metal study also supports but qualifies a three-tier architecture: among 8,162 metal-important genes, 1,484 (18.2%) were general stress, 2,306 (28.3%) metal-shared, and 4,372 (53.6%) metal-specific; mean pangenome core fractions were 92.0%, 91.0%, and 89.8%, with >=95%-core percentages of 57.2%, 50.4%, and 45.7%. ₅ This 11.5-percentage-point gradient supports a model from general stress defense to shared metal defense to more specialized resistance, although all classes remain highly core-enriched. ₅

The Gene Function Ecological Agora supports and broadens the conserved-architecture interpretation through producer–participation classes. Across 18,989 GTDB-r214 species representatives, it classified 11,829,746 Stable, 803,196 Innovator-Isolated, 741,587 Sink/Broker-Exchange, 50,026 Innovator-Exchange, and 314,607 insufficient-data clade × function tuples. ₁₁ Median architectures per KO were 1 for PSII, 5 for mycolic-acid KOs, 15 for TCS histidine kinases, and 46 for mixed-category KOs; this supports an association between architectural diversity and cross-clade exchange but does not establish causation. ₁₁

The functional-dark-matter analysis supports and extends this architecture beyond annotated genes. Across 48 organisms, 57,011 of 228,709 genes (24.9%) were hypothetical, DUF, or otherwise uncharacterized; 6,142 dark genes shared ICA fitness modules with annotated genes, which the report describes as co-regulation. This association rests on module membership and does not demonstrate direct regulation. ₃ The co-fitness validation tested 32,075 non-essential operon pairs against the Fitness Browser cofit table of top-20 partners (13.6M rows). It used tiered scores: mutual top-5 cofit scored 1.0, one-directional top-5 scored 0.75, mutual top-20 with cofit > 0.5 scored 0.50, and one-directional top-20 scored 0.25. Of the tested pairs, 2,899 showed co-fitness evidence (rank ≤ 20), including 1,129 mutual top-5 cofit pairs. A further 998 pairs were “double-validated” because they met both the conserved-synteny criterion (>30%) and the strong co-fitness criterion (score ≥ 0.5). ₃ The report states that 85/100 top candidates had module-based function predictions. It treats the 6,691 module-based predictions from fitness_modules as guilt-by-association inferences, so a “high confidence” priority label reflects converging evidence, not experimental proof. Figure fig23_cofit_validation.png shows the tier distribution, a synteny-versus-co-fitness scatter, and the top 20 combined candidates. ₃ These pairs are therefore association-based predictions, not direct functional validation. ₃ The new truly-dark analysis refines this result by separating persistent unknowns from annotation lag: of 57,011 dark genes, 39,532 had pangenome links, 33,105 (83.7%) were reclassified by Bakta, and 6,427 remained hypothetical in both pipelines; 17,479 lacked pangenome links. ₄ Truly dark genes were shorter, less conserved, more taxonomically restricted, and lower in GC content than annotation-lag genes, with median length 121 versus 194 amino acids, core fraction 43.1% versus 72.7%, essential fraction 18.0% versus 13.4%, and ortholog presence 29.3% versus 63.7%. ₄ Thus, module and neighborhood evidence is useful for prioritizing unknown genes, but remains a process-level association rather than proof of direct regulation or molecular function. _3,4

Among 4,394 truly dark genes in ICA organisms, 41% of neighbors were also hypothetical, 12.0% were within 2 genes of a mobile element, and 25.9% showed operon-like cofitness with adjacent genes using |r| ≥ 0.3; 594 belonged to ICA fitness modules associated by guilt-by-association with candidate phage-integrase, metal-transporter, chemotaxis, and iron-regulation functions. ₄ This supports neighborhood-based functional prioritization while refining the interpretation: 246 genes (3.8%) had no annotation clues, whereas 2,314 had partial functional or phenotype signals, and the top 100 candidates were selected across 19 organisms using fitness importance, clues, ortholog breadth, context, and tractability. ₄

The cofitness–co-inheritance analysis refines module interpretation: across 9 organisms, pairwise cofit pairs had aggregate delta phi +0.003 relative to prevalence-matched random pairs (Mann–Whitney p=1.66e-29), but the across-organism Wilcoxon signed-rank test was not significant (W=9, p=0.13). ₁₂ Across 195 ICA modules in 6 organisms, within-module co-occurrence exceeded the null by delta phi=+0.053; 51/195 (26%) were significant at p<0.05 and 21/195 (11%) at FDR q<0.05. ₁₂ Accessory modules had mean delta phi +0.108, versus +0.059 for core and +0.031 for mixed modules; the accessory-versus-core difference trended toward significance (p=0.051). ₁₂ This supports a stronger role for coordinated multi-gene units, especially accessory units, but not a universal rule; only 48/974 mapped modules were accessory. ₁₀

The conservation synthesis supports module-level coordination: ICA genes were 86% core versus an 81.5% baseline (odds ratio 1.46, p=1.6e-87), and 59% of modules were >90% core. ₁ Stable modules ranged from 17–52 per organism; ortholog scope changed results from 27 to 156 families, from 0 to 28 families spanning 5+ organisms, and from 31 to 493 family-backed predictions. ₇ The independent module-conservation analysis covered 29/32 organisms because Cola, Kang, and SB2B lacked sufficient GTDB genomes for pangenome construction. ₁₀

The anti-phage analysis supports and broadens network architecture beyond fitness-derived modules. Across 27,690 species-level pangenomes, markers from seven defense families occurred in 27,626 (99.8%), producing 930,573 marker hits across 825,476 unique gene clusters; the quality-filtered arms-race analysis contained 7,323 species. ₁₃ Twenty-seven of 28 defense-system pairs showed significant positive co-occurrence after 1,000 phylum-stratified permutations and BH-FDR (Benjamini–Hochberg false-discovery-rate correction); restriction-modification Type II × Gabija had 2,429 observed co-occurrences versus a null mean of 1,555, z = 46.1, and OR = 24.0, whereas CRISPR-Cas × CBASS was nonsignificant (z = 0.21, p_emp = 0.98). ₁₃ This supports coordinated units spanning mechanisms, but genomic co-occurrence is not direct evidence of cofitness or regulation.

Six of seven defense systems were enriched in auxiliary and singleton pangenome fractions, with singleton fractions of 61.1–68.8% against a 37.9% background; DISARM was near baseline and unreliable because its PF00176 anchor also identifies widespread housekeeping helicases. ₁₃ Defense count correlated with prophage-cluster burden (Spearman ρ = 0.609, p ≈ 0, n = 7,323), and partial correlation after genome-size and phylum adjustment remained ρ = 0.301 (p = 1.6 × 10⁻¹⁵³). ₁₃ This refines the core-versus-accessory picture: most fitness modules are core-enriched, whereas anti-phage syndromes can be highly accessory. Truly dark genes show a related but weaker novelty signature: 9.2% had strong GC deviation versus 4.0% of annotation-lag genes, and 12.0% were near mobile elements; these patterns support, but do not establish, recent acquisition or HGT (horizontal gene transfer). ₄

Essential genes are absent from ICA modules in the new analysis: zero essential genes appeared in any module. ₁₀ This is consistent with the functional-dark-matter observation that co-fitness cannot be computed for essential genes, because without viable mutants there are no fitness profiles. ₃ This refines rather than contradicts broader essential-genome results because insertion-based ICA cannot observe essential genes lacking viable insertions. Across 48 bacteria, 859/17,222 ortholog families (5.0%) were universally essential, 4,799 (27.9%) variably essential, and 11,564 (67.1%) never essential; only 15 were essential in all 48 organisms. ₂ Truly dark genes nevertheless had an essential fraction of 18.0% versus 13.4% for annotation-lag genes, and 34 of the top 100 unknown-gene candidates were essential. ₄

Pfam domains increased module annotation from 92/1,116 (8.2%) to 890/1,116 (79.7%), predictions from 878 to 6,691, and annotated families from 32 to 145 of 156. ₇ Ortholog transfer achieved 95.8% precision and 91.2% coverage for gene-level KO prediction, whereas Module-ICA had <1% KO-level precision but captured process-level co-regulation. ₇ In the held-out benchmark, ortholog transfer achieved 95.8% strict precision, 91.2% coverage, and 0.934 F1; domain-based prediction achieved 29.1% precision, 66.6% coverage, and 0.401 F1; Module-ICA achieved <1% strict precision and 23.3% coverage; and cofitness voting achieved <1% strict precision and 73.0% coverage. ₈ The new report supports this caution: only 4.0% of truly dark genes had Pfam hits and 4.6% had KEGG KOs, although 43.5% had partial eggNOG signal and 79.4% had UniRef50 links. ₄

The within-species study supports another organization layer: 1,517 resistance islands across 705 species had mean phi coefficient 0.827, and 88% contained genes from multiple resistance mechanisms. ₁₄ This indicates tight genomic co-inheritance, not physical-linkage-driven cofitness or synergy. Only 0.7% of cofit pairs were adjacent within 5 genes, and excluding adjacent pairs did not change the result pattern. ₁₂

The ADP1 study refines conserved-module interpretation: its 51-gene quinate-specific network assigned 44/51 (86%) genes to aromatic degradation, complex-i, iron acquisition, and pqq-biosynthesis, with dispersed subsystems. ₁₅ Cofitness can therefore reveal metabolic coupling among dispersed genes, whereas resistance islands reflect genomic and lineage-linked organization. Each of 8 carbon sources required a distinct respiratory configuration; quinate required only Complex I whereas glucose had no specifically required component. ₇ Within this single-organism network, co-fitness supports subsystem assignment. Of 23 genes initially categorized as “Other” or “Unknown”, 16 were assigned to support subsystems with medium or high confidence, based on growth-profile correlations with known subsystem members. Seven quinate-specific genes remained unassigned because their co-fitness with all subsystems was low. ₁₅ Within-category correlation was much higher than between-category correlation, with Complex I mean r = 0.992 and aromatic pathway r = 0.961. ₁₅ The method recovered pcac (4-carboxymuconolactone decarboxylase), which keyword matching had miscategorized, and nominated two DUF proteins as probable Complex I accessory factors. However, the 11 Complex I-associated co-fitness assignments beyond the core nuo operon rest on phenotypic correlation only. Some may reflect indirect connections rather than physical association with Complex I. ₁₅

The formulation study supports and extends condition-specific architecture to organism interactions: across 142 isolates, metabolic overlap with PA14 predicted planktonic inhibition (r = 0.384, p = 2.3×10⁻⁶), but cross-validated metabolic-model fit was only CV R² = 0.145 ± 0.142. ₁₆ Adding genus increased training fit from R² = 0.274 to R² = 0.360, leaving approximately 73% of inhibition variance unexplained by metabolism alone. ₁₆

The counter-ion analysis supports and narrows condition-specific interpretation. Across 19 organisms and 14 metals, 4,304/10,821 metal-important records (39.8%) were also NaCl-important; excluding the Synechococcus elongatus outlier left 36.7% (3,739/10,183). ₁₇ Zinc sulfate had 44.6% overlap despite 0 mM chloride, versus 41.3% cobalt, 41.0% copper, and 39.3% nickel; chloride-delivered metals averaged 41.6% versus 37.8% for non-chloride metals. ₁₇ This contradicts a chloride-dose explanation and supports shared cellular stress biology, while NaCl remains a mixed sodium-and-osmotic-stress treatment.

Organism Specificity

Support-network composition was more similar across different AMR mechanisms within the same organism than across the same mechanism in different organisms: mean Jaccard similarity of GO term sets was 0.375 versus 0.207 (p = 4.3×10⁻¹³). ₆ This supports organism-level regulatory, metabolic, and signaling architecture as a stronger determinant than resistance-mechanism class, but does not establish direct co-regulation: organism-specific media responses, metabolic dependencies, conditionally dispensable functions, and regulatory architecture could all contribute. ₆ The conserved core included transmembrane transport in 87–100% of organisms, signal transduction in 87–100%, transcription regulation in 96–100%, and phosphorelay signaling in 91–100%; flagellar motility occurred in 53–61% and amino acid biosynthesis in 30–73%. ₆ Histidine biosynthesis showed the only hint of mechanism specificity: efflux occurred in 68% versus 30% for metal resistance, p = 0.013 uncorrected and q = 0.18 after FDR correction. ₆ Within mechanism-defined networks, enrichment was common: 212 of 9,244 GO tests were significant at FDR < 0.05, and the top signal was efflux-associated histidine biosynthesis in 6 organisms. Efflux AMR genes also showed tryptophan-biosynthesis enrichment in 5 organisms, and metal-resistance genes showed chemotaxis enrichment in 4. However, no GO term differed significantly between mechanisms after FDR correction. ₆ This refines the organism-specificity result: each mechanism carries enrichment, but the mechanisms do not carry detectably distinct enrichment. The generality of the organism-level result is limited because the 28 organisms are lab-adapted, phylogenetically biased toward Pseudomonas, and ecologically limited. ₆ The report lists the H4 reanalysis with InterProScan GO as confirming organism-specificity. ₆ The central digest supports the same contrast, cross-mechanism Jaccard 0.375 versus within-mechanism 0.207 (MWU p=4.3e-13), and describes support networks as organism-specific. ₇

Improved annotation increased within-mechanism Jaccard similarity from 0.069 with old KEGG annotations to 0.207 with InterProScan GO, and cross-mechanism similarity from 0.249 to 0.375; the comparison changed from p = 1.0 to p = 4.3×10⁻¹³. ₆ InterProScan provided 68% gene coverage, reported as 3.6× better than the old SEED annotations. ₆ This refines pangenome-integration by showing that annotation coverage determines detectable architecture, not its regulatory mechanism. Across 132.5M gene clusters, eggNOG had higher COG, KEGG, and Pfam coverage, Bakta had higher GO, product-description, and unique UniRef50 coverage, and their union reached 77.3% any-functional-annotation coverage. ₇ Persistent truly dark genes provide the limiting case: despite 84.7% database cross-references, only 4.0% had Pfam hits and 4.6% KEGG KOs. ₄

The 156 module families, each spanning 2+ organisms within the 32-organism dataset, support organism-spanning architecture. They include 28 families spanning 5+ organisms, 7 families spanning 10+ organisms, and one spanning 21; 145 families (93%) had consensus labels. ₈ Organisms with fewer than approximately 100 experiments produced weaker modules; Caulo, with 198 experiments, showed only 2.9x correlation enrichment. ₈ Experimental coverage is therefore a condition of detectability. Family breadth itself was not associated with core fraction (rho=-0.01, p=0.914), and the pangenome-linked analysis covered only 29/32 organisms. ₁₀

The anti-phage analysis supports but qualifies organism-spanning architecture. CRISPR-Cas prevalence was 96.1%, BREX 80.1%, R-M Type I 76.8%, DISARM 58.4% under the strict call, Retron 54.7%, R-M Type II 38.9%, Gabija 22.8%, and CBASS 7.2%; prevalence varied by phylum, with R-M Type I exceeding 80% in Pseudomonadota and Bacillota and Gabija reaching 30–40% in Bacillota versus below 10% in Actinomycetota. ₁₃ The analysis covered 7,323 species with at least five genomes, or 26% of 27,690 pangenome species, so extrapolation to the full 293K-genome tree remains untested. ₁₃

The metal study supports but qualifies cross-organism signal: cross-resistance direction was consistent across Proteobacteria, Bacteroidetes, Firmicutes, and Actinobacteria; Co-Ni had median r approximately 0.58 across 28 organisms, whereas Al-Co and Al-Ni had median r approximately 0.30. ₅ Pairwise rankings were mostly positive across 10 common metal pairs and 12 organisms with at least 5 metals, but Mantel tests across 351 organism pairs had mean r = 0.23 with 62% positive values, leave-one-out consensus had mean r = 0.41 with only 2/28 organisms individually significant, and metal-label permutation was nonsignificant (p = 0.42). ₅

The Agora supports and qualifies breadth: Cyanobacteriia × PSII was supported at class rank with producer d = +1.50, consumer d = +0.70, producer p = 2×10⁻⁵, and consumer p = 2×10⁻⁴, but genus, family, and order tests were STABLE; the class sample contained n = 21 PSII KOs. ₁₁ Bacteroidota × PUL CAZymes falsified the original absolute-zero Innovator-Exchange criterion at UniRef50 resolution, although Sankoff diagnostics recovered a small relative HGT signal of Cohen’s d = 0.15. ₁₁

Functional-dark-matter evidence supports and qualifies breadth: a conservation-weighted covering set selected 42 organisms covering 95.6% of importance-weighted priority across 28,584 high-priority dark genes, but the Fitness Browser was uneven, with 37/48 organisms in Pseudomonadota and 0 top-500 candidates from Archaea, Actinobacteria, or Epsilonproteobacteria. ₃ An extended 73-organism pool produced a 50-organism set covering 98.7% of ortholog groups across 6 phyla, but non-Fitness-Browser organisms lacked condition profiles. ₃ The truly dark candidate set further emphasizes this limitation: the top 100 span 19 organisms, with Methanococcus_S2 contributing 29, DvH 13, Methanococcus_JJ 13, and MR-1 8. ₄ The report also states that archaea contribute 55% of truly dark genes and cites high Methanococcus cofitness signals (r > 0.97). It proposes, as future work rather than a completed analysis, using these signals to map operonic structure and predict functions for archaeal-specific hypothetical proteins. ₄

The strain-variation results support organism and lineage structure: 701/1,261 species (55.6%) showed significant ANI-distance/AMR-Jaccard correlation at FDR < 0.05, median Mantel r = 0.247, with positive correlation in 87.8%; non-core AMR genes had median Mantel r = 0.222 versus 0.117 for core genes. ₁₄ Host-associated species carried more AMR genes, but metadata sparsity prevents a general environment-defined AMR ecotype claim. ₁₄

The pangenome analysis supports but qualifies specificity: cofit-pair phi was higher among near genomes (mean=0.102) than medium-distance genomes (mean=0.067), while most species lacked a far stratum; organism-level deltas ranged from Ddia6719 +0.093 to Korea -0.042, with Putida -0.000 and p=0.41. ₁₂ Shared ancestry therefore contributes to co-occurrence, and current stratification does not fully separate inheritance from coupling.

Across 194,216 genes, essential genes were 82% core and always-neutral genes 66% core, while module-family breadth did not predict conservation (rho=-0.01, p=0.91). ₁ Universally essential genes were 91.7% core versus 80.7% for non-essential genes, while orphan essentials were 49.5% core. ₂ Variable essentiality was widespread: 41,059/221,005 genes (18.6%) were essential, with rates from 12.2% in Pedo557 to 29.7% in Magneto, and 4,799 families were variably essential. ₂ Truly dark genes refine this organism-specific pattern with a higher essential fraction than annotation-lag genes, but their narrow ortholog breadth and 4–96% organism-level range of dark genes resisting Bakta show that coverage and taxon composition also shape apparent specificity. ₄

The formulation results support organism-specific composition: positive inhibition residuals identified streptococcus salivarius, gemella sanguinis, and neisseria mucosa as candidate dual-mechanism inhibitors, while genus increased explained inhibition variance. ₁₆ The single-cohort planktonic result does not establish a general rule. The ADP1 results support organism-specific respiratory architecture: complex-i comprised 21/51 quinate-network genes, while FBA (flux-balance analysis, a growth-optimizing constraint-based model) predicted 0% essentiality despite 10/13 Complex I subunits producing quinate-specific defects. ₁₅ This suggests that generic models miss some organism-specific dependencies.

The Agora supports environmental embedding as contextual, not causal: Cyanobacteriia showed 2.77× photic-aquatic enrichment (p < 10⁻⁵²), Mycobacteriaceae 7.88× host-pathogen enrichment (p < 10⁻⁴⁵), and Bacteroidota 1.40× gut/rumen enrichment (p < 10⁻³⁵); Mycobacteriaceae was not soil-enriched (0.88×, p = 0.87). ₁₁ AlphaEarth k = 10 placed 35.6% of 90 Cyanobacteriia species in a marine+sponge-dominant cluster, 34.0% of 50 Mycobacteriaceae in a gut+sludge cluster, and 20.0% in another gut cluster, while 729 Bacteroidota species spanned marine and gut-associated clusters. ₁₁

Shared Dispensability and Limits on Co-regulation

InterProScan GO annotations detected enrichment for flagellum-dependent cell motility in 5 organisms with mean OR 4.7, flagellum assembly in 5 with mean OR 5.3, bacterial-type flagellum in 4 with mean OR 4.9, flagellum-dependent swarming in 4 with mean OR 5.0, histidine biosynthesis in 3 with mean OR 5.3, and tryptophan biosynthesis in 3 with mean OR 5.3. ₆ Old SEED/KEGG analysis found 0/280 significant tests at FDR (false discovery rate) < 0.05, whereas InterProScan GO found 35/3,193. ₆ The central digest reports the same pattern for 801 AMR genes across 28 organisms: enrichment for flagellar motility (GO:0071973, 5 organisms FDR<0.05), flagellum assembly (GO:0044780, 5 orgs), histidine biosynthesis (GO:0000105, 3 orgs), and tryptophan biosynthesis (GO:0000162, 3 orgs). The digest interprets this more strongly than the project report: it describes co-regulation with motility and biosynthesis and suggests that AMR costs reflect competition for the proton motive force. ₇ This stronger reading is in tension with the AMR report, which states that enrichment alone does not establish co-regulation and keeps shared dispensability as a competing explanation. The proton-motive-force account should therefore be treated as a hypothesis. ₆ This supports possible connections to flagellar-motility, histidine biosynthesis, and tryptophan-biosynthesis, but not direct co-regulation. Shared dispensability remains a competing explanation because flagella and chemotaxis may be unnecessary in shaken liquid culture, amino-acid biosynthesis may be redundant in rich or supplemented media, and AMR genes may be dispensable without antibiotics. ₆ To distinguish true co-regulation from shared dispensability, the report calls for a fitness-matched permutation, which it describes as the critical follow-up. The permutation would draw random non-AMR genes matched on mean fitness level, not just conservation class. If random genes with fitness between −0.05 and +0.05 also show flagellar/biosynthesis enrichment in their cofitness neighborhoods, the enrichment is a shared-dispensability artifact; if AMR genes uniquely enrich even among fitness-matched genes, co-regulation is supported. ₆ The earlier H1/H3 support-network analysis with old SEED annotations was a null result. The enrichment therefore became detectable only with the InterProScan GO annotation layer. ₆

Evidence favoring shared dispensability includes enrichment for categories expected to be dispensable in shaken-flask culture and the absence of energy-metabolism enrichment in 0/25 organisms in a permutation test with fold 0.91. ₆ That permutation matched conservation class but not mean fitness level, so it does not fully distinguish shared dispensability from shared regulation. ₆ The ICA benchmark supports and sharpens this caution: Module-ICA and cofitness voting each had <1% strict precision for exact KEGG KO assignments despite coverage of 23.3% and 73.0%, respectively. ₈ High coherence therefore indicates process-level architecture, not shared molecular function or direct regulation. The AMR report proposes three further discriminating tests. First, compute cofitness separately for antibiotic conditions and standard growth; AMR–flagella cofitness elevated specifically under antibiotic stress would support a regulatory connection. Second, check whether flagellar knockouts show positive mean fitness across Fitness Browser experiments. Positive values would make shared dispensability plausible, whereas near-zero or negative values would strengthen co-regulation. Third, test whether phage-defense, secondary-metabolite, or other condition-specific gene classes show the same cofitness patterns. If they do, the signal would be a general feature of dispensable genes under lab conditions rather than AMR-specific. These are proposed analyses, not completed ones. ₆

The truly-dark analysis supports using process-level evidence for prioritization: 25.9% of analyzed truly dark genes had operon-like cofitness with adjacent genes, but the report notes that short genes have fewer insertion sites and that strong phenotypes may reflect polar effects. ₄ The reported Pearson correlation removes each gene’s mean fitness before correlating fitness profiles, so uniformly slightly positive fitness values alone would produce zero correlation; nevertheless, genes with similar condition-dependent changes in dispensability can still have high cofitness without direct co-regulation. ₆ Missing fitness values were treated as zero in z-score space through np.nan_to_num, which approximates but does not equal pairwise-complete Pearson correlation; the dense matrices were considered unlikely to substantially alter conclusions, but this remains a limitation for individual edges. ₆ The report attributes to Sagawa et al. (2017) a validation that cofitness recovers transcriptional regulatory relationships in 24 bacteria. This qualifies the dispensability caution: cofitness can carry regulatory signal in general. Even so, the report leaves open whether its own AMR–motility signal reflects regulation or shared dispensability. ₆

Environmental context is relevant to AMR organization: efflux pumps constituted 21% of AMR in human-gut species but 1% in aquatic species, metal resistance 45% in soil/aquatic species but 6% in human-gut species, and clinical species had 68% accessory AMR versus 43% in soil; across 823 multi-environment species, clinical strain fraction predicted total AMR at rho=0.465, p=2.2e-45. ₇

The counter-ion findings strengthen shared-stress explanation: shared NaCl–metal signals involved 4,304/10,821 records (39.8%), and shared-stress genes were important for a mean of 4.1 metals versus 2.5 for metal-specific genes. ₁₇ In DvH, whole-genome correlations with NaCl ranged from zinc r=0.715 and manganese r=0.545 to iron r=0.086; the hierarchy did not follow chloride concentration because zinc sulfate delivered zero chloride yet ranked first. ₁₇ Shared cell-envelope, DNA-repair, and ion-homeostasis stress biology is plausible but remains a hypothesis.

The metal study supports and qualifies this explanation: general stress genes were more conserved than metal-shared and metal-specific genes, but all three tiers remained highly core-enriched, with mean core fractions of 92.0%, 91.0%, and 89.8%. ₅ Its absence of negative metal controls prevents distinguishing universal cross-resistance from a general response to stress. ₅

The anti-phage study supports but does not resolve coordinated defense versus shared ecological exposure. Defense count rose with prophage burden after phylum and genome-size adjustment, but presence/absence data did not establish co-regulation or synergy. ₁₃ CRISPR-Cas prevalence was 96.1% by permissive EggNOG matching but approximately 55% with specific Cas1 PF01867, so the combined call may reduce syndrome specificity. ₁₃

The conservation synthesis supports the competing explanation: core genes were more likely than accessory genes to be burdensome in laboratory conditions, with 24.4% positive fitness when deleted versus 19.9% of accessory genes; core genes were 1.78x more likely to have strong condition-specific phenotypes and 1.29x more likely to show important-and-burdensome trade-offs. ₁ Of 25,271 trade-off genes, 17.8% were important in some conditions and burdensome in others; trade-offs were 1.29× more likely to be core (OR=1.29, p=1.2e-44), and condition-specific genes were 77.3% core versus 70.3% without specific phenotypes (OR=1.78, p=1.8e-97). ₇ This contradicts a simple streamlining expectation that accessory genes are systematically more costly, but not shared dispensability in particular neighborhoods.

The essential-genome analysis supports and qualifies context dependence: 859 families were universally essential, 4,799 variably essential, and 11,564 never essential; variably essential families had median penetrance 33%, with 813 more than 50% essential and 704 less than 10%. ₂ RB-TnSeq essentiality is measured under specific library-construction conditions, typically rich media, and may miss stress-specific requirements.

The formulation study supports condition-specific fitness: PA14 outgrew the average commensal on all 22 substrates, commensals began growing earlier in 43.1% of comparisons, and exceeded PA14’s maximum growth rate in only 13.8%. ₁₆ Metabolic models explained R² = 0.274 before genus, R² = 0.360 after genus, and R² = 0.311 with kinetics in the 29-isolate complete cohort. ₁₆

The ADP1 evidence refines shared-dispensability interpretation. Complex I defects were strongest on acetate (-1.55) and succinate (-1.39), but Complex I was dispensable on glucose and lactate; this suggests dependence on NADH-generating respiratory demand rather than aromatic chemistry alone. ₁₅ PQQ-biosynthesis genes were glucose-specific in a separate deletion analysis. ₁₅ Quinate produced 0.57 NADH per carbon versus 1.50 for glucose, yet Complex I was more essential because β-ketoadipate products created a concentrated NADH burst. ₇

Functional-dark-matter evidence supports and qualifies environmental interpretation. Within-species carrier-versus-non-carrier tests found 10 of 137 environmental-category tests significant at FDR < 0.05 and 1 of 67 AlphaEarth embedding tests significant; a pre-registered lab–field mapping found 29/47 clusters concordant (61.7%). ₃ The one-sided binomial test was p = 0.072, with Wilson 95% CI [0.474, 0.742], while Fisher’s combined probability was p = 0.031. ₃ NMDC validation confirmed 4/4 pre-registered abiotic directions, including nitrogen-source carriers with total nitrogen (ρ = +0.109, n = 1,231, FDR = 2.3e-4) and ammonium nitrogen (ρ = +0.231, n = 1,230, FDR = 8.0e-16), pH carriers with pH (ρ = +0.157, n = 4,366, FDR = 7.4e-25), and anaerobic carriers negatively with dissolved oxygen (ρ = -0.298, n = 272, FDR = 1.5e-6). ₃ However, 441/449 exploratory trait tests were significant, plausibly because of compositional coupling, so these results do not establish gene-specific environmental causation.

The Agora supports and qualifies acquisition-depth differences: M22 Sankoff parsimony assigned 17,073,194 gain events to recipient-rank depth bins; recent-to-ancient ratios were 24.5× for CRISPR-Cas, 10.3× for TCS histidine kinases, 9.0× for β-lactamases, and 2.3× for clean tRNA-synthetase controls. ₁₁ Mycobacteriaceae mycolic-acid gains were 79.87% recent and 0.00% ancient among 53,916 gains, whereas all mycolic-acid gains were 48.79% recent and 3.68% ancient; Cyanobacteria PSII gains had 2.05% ancient events versus 14.90% atlas-wide. ₁₁ Truly dark genes’ narrow ortholog breadth and mobile-element proximity support a related acquisition hypothesis, but GC deviation is an imperfect HGT proxy and neither analysis establishes donor identity. _4,11

Tensions

The central tension remains whether flagellar motility, chemotaxis, and amino-acid biosynthesis enrichment reflects genuine co-regulation or shared dispensability. ₆ The null matched conservation class rather than mean fitness and does not resolve this. ₆

The metal data add a related tension: all tested metal pairs were positive, but no negative controls were included, so universal cross-resistance cannot be distinguished from a general-stress response. ₅ The counter-ion study supports shared cellular stress biology and reports 4,304/10,821 shared NaCl–metal records, but zinc sulfate had 44.6% overlap despite 0 mM chloride. ₁₇

The anti-phage evidence introduces detection and phylogeny tension. Twenty-seven of 28 defense pairings were significant and R-M Type II × Gabija was strong, but CRISPR-Cas prevalence changes from 96% with EggNOG description matching to approximately 55% with Cas1 PF01867, and the analysis controls for phylum rather than using phylogenetically corrected regression. ₁₃

The module-conservation results introduce a measurement-boundary tension rather than a biological contradiction: zero essential genes appeared in ICA modules, whereas essential-genome analyses identify universally and variably essential families. _10,2 Truly dark genes show the opposite comparison at the gene-set level—18.0% essential versus 13.4% for annotation-lag genes—but short genes and insertion bias complicate interpretation. ₄

A second tension concerns process inference versus gene-level identity. Module-ICA and cofitness voting had <1% strict KEGG KO precision, whereas ortholog transfer achieved 95.8% precision, 91.2% coverage, and 0.934 F1. ₈ Functional-dark-matter GapMind and domain results, and truly-dark eggNOG or phenotype clues, identify hypotheses rather than direct gene-to-step assignments. _3,4 The Agora’s M26 tree-based donor inference is exploratory rather than donor identification. ₁₁

Essentiality creates a further tension: 15 families were essential in all 48 organisms, while 4,799 were variably essential and 7,084 essential genes had no detectable ortholog in another Fitness Browser organism. ₂ Orphan essentials were 49.5% core versus 91.7% for universally essential genes, and 58.7% of orphan essentials were hypothetical versus 8.2%. ₂ Conservative BBH orthology can miss divergent homologs, so orphan status is not proof of novelty.

Laboratory burden conflicts with natural-selection interpretation: 28,017 genes were costly in the laboratory and conserved in the pangenome versus 5,526 costly and dispensable genes. ₁ These are evidence for, not direct measurements of, purifying selection in natural environments.

Pairwise and module-level generality diverge. Pairwise cofit pairs had mean delta phi +0.011 across organisms but aggregate delta +0.003 and Wilcoxon p=0.13, whereas ICA modules had delta phi=+0.053 overall and accessory modules +0.108. ₁₂ Modules may therefore be stronger units, but Korea had no significant modules because all were >90% core with prevalence near 1.0, and accessory-versus-core differences were only near significant (p=0.051). ₁₂ Only 5% of mapped modules were accessory, so accessory-module advantages cannot be generalized. ₁₀

Functional coupling remains confounded by ancestry: near genomes had mean phi=0.102 versus 0.067 for medium-distance genomes, and most organisms lacked a far stratum. ₁₂ The metal study likewise requires PGLS (phylogenetic generalized least squares) or independent contrasts; its organisms were not phylogenetically independent, matrices ranged from 3 to 112 metal experiments, and metal concentrations differed. ₅

The metal–environment validation is unresolved. Multi-metal tolerance scores did not correlate with BacDive isolation from metal environments at the Fitness Browser species scale (Spearman rho approximately -0.02, p > 0.8); after matching and collapsing strains, only 20 independent species remained, although the limitations report n = 26 at organism scale. ₅ This contradicts treating cross-resistance as already validated by isolation environment, but not the prior Metal Fitness Atlas pangenome-scale result of Cohen’s d = +1.0 across 42K strains. ₅

The functional-dark-matter analysis adds coverage tension: 57,011 dark genes were found, but the Fitness Browser contained 37/48 Pseudomonadota and no Archaea, Actinobacteria, or Epsilonproteobacteria among the top 500 candidates. ₃ The truly-dark analysis adds a 17,479-gene pangenome-linkage gap and estimates approximately 2,841 additional truly dark genes among them at the 16.3% linked rate; this is an estimate, not an observed count. ₄ The functional-dark-matter report also contains an internal scope conflict. One section describes its 32,075 co-fitness-tested pairs as providing the strongest functional inference for essential genes that lack direct fitness data. Its Finding 12 instead specifies 32,075 non-essential operon pairs and states that co-fitness is unavailable for essential genes. The tested-pair scope and the claimed applicability therefore conflict. ₃

The Agora adds rank-resolution tension. Cyanobacteriia × PSII was STABLE at genus, family, and order despite class-level producer d = +1.50 and consumer d = +0.70, while Mycobacteriaceae mycolic-acid signal varied with aggregation rank. ₁₁ These findings contradict treating a clade-level architecture label as rank-invariant.

Network size does not explain AMR fitness cost: Spearman rho = −0.006, p = 0.87, N = 769; within mechanisms, rho = −0.049 for efflux, +0.038 for enzymatic resistance, and −0.031 for metal resistance, all p > 0.4. ₆ Uniform resistance cost was +0.086 and was not explained by neighborhood size. ₆ Across 801 AMR genes and 25 organisms, pooled knockout fitness shift was +0.086 [+0.074, +0.098]; cost was mechanism-independent (KW p=0.89), conservation-independent (p=0.33), and tier-independent (p=0.26), although efflux genes showed a stronger antibiotic-dependent flip than enzymatic inactivation genes, +0.094 versus -0.001, MWU p=0.007. ₇

The ADP1 model-versus-phenotype tension remains: FBA predicted 0% Complex I essentiality while 10/13 subunits produced quinate-specific defects, and 30/51 network genes lacked FBA reaction mappings. ₁₅ The corrected cross-species test did not support the NDH-2 compensation hypothesis: aromatic deficits were -0.297 with validated NDH-2 versus -0.156 without, p=0.52. ₇

Resistance-island co-inheritance creates caution rather than resolution: only 190/974 species (19.5%) formed at least two AMR ecotypes, and environmental association testing was limited by 52.7% of genomes lacking a classifiable isolation_source. ₁₄

The Agora’s quantitative Alm 2006 comparison remains unresolved: four framings produced Pearson r values of 0.288, 0.157, 0.100, and 0.105 rather than the reported r ≈ 0.74; consumer-side KO-to-architecture concordance remained r = 0.673, while producer concordance was exploratory at r = 0.093. ₁₁

Open Directions

• Recompute AMR and non-AMR permutations matched on mean fitness, including the −0.05 to +0.05 range, and test whether flagellar and amino-acid enrichment persists; recompute truly-dark neighborhood enrichment with insertion-density and mean-fitness controls. _6,4
• Test whether phage-defense, secondary-metabolite, and other condition-specific gene classes reproduce AMR-like flagellar/biosynthesis cofitness neighborhoods, which would mark the signal as a general lab-dispensability feature rather than AMR-specific. ₆
• Use the high Methanococcus cofitness signals (r > 0.97) to map operonic structure and generate functional predictions for archaeal-specific truly dark hypothetical proteins. ₄
• Partition fitness matrices into antibiotic-exposure and standard-growth conditions; compare direct transcription-factor targets and expression responses with cofitness edges; measure mean fitness for flagellar knockouts and other conditionally dispensable genes. ₆
• Recalculate networks at |r| > 0.4 and |r| > 0.5, using coordinate-based operon exclusion rather than matrix row position, and test whether organism-specificity and enrichment remain stable. ₆
• Apply Pfam enrichment, replace the row-index operon heuristic with genomic coordinates, and reassess extra-operon network sizes. ₆
• Apply PADLOC MacSyFinder-style multi-PFam and gene-order rules to Retron, DISARM, and Gabija; compare Cas1-specific, EggNOG, and context-refined CRISPR-Cas calls; test whether defense-pair enrichment survives stricter calls. ₁₃
• Fit the defense–prophage relationship with a GTDB-tree phylogenetic mixed-effects model and refit the negative-binomial model with estimated dispersion. ₁₃
• Experimentally compare R-M Type II-only and R-M Type II-plus-Gabija species in matched phage-challenge assays; expand defense analysis beyond seven systems and cross-reference syndromes with isolation-source metadata or AlphaEarth embeddings. ₁₃
• Reanalyze ICA with |weight| >= 0.3, a 50-gene cap, and a 40% component cap across organisms matched for experiment count. ₈
• Experimentally test module-only, family-backed, and truly-dark top-100 predictions under module-informed conditions without treating them as exact KO assignments; prioritize PV4/5210953 motility, ANA3/7026383 nitrogen limitation, DvH/206658 stress, and Methanococcus_S2/MMP_RS06570 fluoride-efflux-associated tests. _8,4
• Extend pangenome linkage to the 17,479 unlinked dark genes; test whether the estimated approximately 2,841 additional truly dark genes and their 2,208 estimated strong-phenotype subset are enriched in dark islands, mobile elements, or ICA modules. ₄
• Use AlphaFold2 or ESMFold followed by Foldseek for the top 100 candidates, test growth under predicted conditions, apply mobile-CRISPRi, and distinguish polar effects from gene-autonomous phenotypes. ₄
• Restrict cofitness–co-inheritance analyses to auxiliary-only pairs below 95% prevalence, build module co-transfer networks, and expand to species with >30% auxiliary genes. ₁₂
• Recover raw genefitness data for Ralstonia UW163 and Ralstonia GMI1000 and add a well-populated far phylogenetic stratum. ₁₂
• Map resistance islands to plasmids, chromosomes, integron boundaries, and insertion sequences, then stratify by island context and AMR ecotype. ₁₄
• Test the ADP1 NADH-flux hypothesis by deleting NDH-2 across quinate, acetate, succinate, glucose, and lactate, then add benzoate, catechol, vanillate, iron limitation, and respiratory inhibitors. ₁₅
• Validate ACIAD3137 and ACIAD2176 by interaction or co-purification assays and compare Complex I retention across aromatic degraders. ₁₅
• Measure the complete 10-pair formulation matrix, repeat assays with PAO1 and 3–5 mucoid clinical PA isolates, and test structured biofilms. ₁₆
• Classify 5,526 costly-and-dispensable genes as mobile, recently acquired, or undergoing loss, and test whether module membership distinguishes transient accessory architecture. ₁
• Use AlphaEarth data to test whether environmental variability predicts core trade-offs; identify universally essential families; and test whether 48 accessory modules retain their co-inheritance advantage after phylogenetic and prevalence matching. ₁
• Match metal and NaCl profiles on mean fitness and conservation, apply formal COG, KEGG, and Pfam enrichment, and use choline chloride or KCl controls with matched CuCl₂/CuSO₄, ZnCl₂/ZnSO₄, and CoCl₂/CoSO₄ experiments. ₁₇
• Apply ICA to metal–NaCl matrices to test whether shared-stress modules explain condition-specific architecture. ₁₇
• Hold ortholog scope and Pfam coverage constant when testing whether process-level module predictions replicate across 32 organisms. ₇
• Test whether the 1.29× core enrichment and 17.8% trade-off rate persist after matching prevalence, mean fitness, organism, condition, and experimental coverage. ₇
• Characterize 7,084 orphan essentials for mobile-element association, recent acquisition, rapid evolution, and missed divergent orthology; test the 1,382 module-transfer predictions by CRISPRi under module-informed conditions and link variable essentiality to pathway completeness. ₂
• Expand Fitness Browser taxonomic coverage and compare universal essential families with the Database of Essential Genes. ₂
• Test the 57,011 dark-gene spectrum using CRISPRi for essential candidates, RB-TnSeq or targeted deletion for non-essential candidates, and matched conditions across organisms; ask whether module, synteny, and cofitness evidence predicts reproducible phenotypes. ₃
• Reanalyze the 1,256 GapMind gaps with direct EC and structure-based matching, then experimentally test the highest-confidence gene–pathway assignments rather than treating organism-level co-occurrence as assignment. ₃
• Repeat lab–field validation with compositional-control permutations, annotated-accessory controls, genome-level AlphaEarth coverage, and sample-label permutation tests; determine whether the 29/47 concordance and NMDC directions remain gene-specific. ₃
• Extend conservation and covering-set analyses beyond the Pseudomonadota-heavy Fitness Browser, then test whether the 42-organism set remains optimal after adding condition coverage and phylum balance. ₃
• Recompute Producer × Participation classifications across matched taxonomic ranks and report leaf_consistency alongside module conservation to test whether apparently conserved architecture is uniform within clades. ₁₁
• Cross-validate the 17,073,194 Sankoff gain events with DTLOR or another reconciliation method, bootstrap individual acquisition-depth estimates, and test whether recent-to-ancient ratios persist after controlling for annotation density. ₁₁
• Obtain per-CDS sequence data and perform composition-based confirmation and full DTL reconciliation for M26 donor labels; test whether donor inference changes Open-Innovator classifications. ₁₁
• Repeat PSII, mycolic-acid, and PUL neighborhood scans with memory-scalable joins, and test whether the 10.91% PSII MGE-neighbor fraction remains distinct from the 10.6% Poisson baseline. ₁₁
• Expand Cyanobacteriia phenotype coverage beyond n = 4 and test whether AlphaEarth environmental clusters predict function-class architecture after phylogenetic and sampling controls. ₁₁
• Repeat metal cross-resistance with non-metal stress controls, normalize fitness by concentration relative to MIC, and use PGLS or independent contrasts to test whether positive directionality persists after phylogenetic correction. ₅
• Apply ICA to metal-condition matrices and compare universal-stress, metal-shared, and metal-specific modules after matching organisms for experiment count. ₅
• Map cross-resistance gene signatures with KEGG/PFAM across 27K species and test whether the 318 conserved ortholog groups predict metal-associated environments at pangenome scale rather than at the underpowered species scale. ₅
• Use AlphaFold-based structural analysis to test whether metal-shared proteins have convergent binding or stress-response features, and experimentally validate the strongest shared-versus-specific candidates. ₅
• Reanalyze module conservation after matching organisms for pangenome availability and experiment count, and test whether the 86.0% versus 81.5% core difference persists under alternative ICA thresholds and module-size caps. ₁₀
• Test whether the 38 accessory module families retain apparent co-inheritance or niche-specific signatures after controlling for phylogeny, prevalence, genome size, and annotation coverage; distinguish horizontal transfer from other causes of low core fraction. ₁₀
• Use CRISPRi or other perturbations to determine whether essential genes form condition-specific modules invisible to transposon-based ICA. ₁₀