Functional enrichment in cofitness neighborhoods can arise from genuine shared regulation or from genes having similar fitness behavior under the assay conditions. A fitness-matched null model tests these explanations by comparing AMR-associated neighborhoods with random non-AMR genes whose mean-fitness distribution is matched, rather than matching only conservation class. 1
This problem connects condition-specific-fitness with cofitness-network-architecture: cofitness measures similarity in fitness phenotypes, not direct transcriptional control, so enrichment requires a null model that reproduces the relevant fitness structure. 1 Evidence from a 43-bacterium comparison supports this concern: fitness importance was positively but weakly associated with core status, so conservation class does not fully capture fitness behavior. 2 The same evidence refines the proposed control: fitness breadth and condition-specific effects may also need consideration alongside mean fitness. 2 A broader synthesis further supports this limitation: essential genes were 82% core whereas always-neutral genes were 66% core, while core genes were 1.78x more likely to have strong condition-specific phenotypes. 3 The AMR report accepts the principle, which it attributes to Nichols et al. (2011, Cell), that condition-dependent fitness profiles reveal gene function in E. coli. It adds a caveat: functional enrichment in cofitness neighborhoods can reflect shared experimental context (laboratory conditions) rather than biological co-regulation. 1
Why conservation-matched nulls are insufficient
The AMR cofitness analysis used a permutation that matched conservation class but not mean fitness level. 1 The report identifies this mismatch as the key unresolved limitation. It states that if random genes had been drawn with the same slightly-positive fitness distribution as AMR genes, the flagellar enrichment might disappear entirely. 1 Fitness Browser experiments use shaken liquid culture, where flagella are useless, and often use supplemented media, where biosynthesis is redundant. Under these conditions AMR genes, flagellar genes, and biosynthesis genes are all metabolic burdens. The enrichment may therefore reflect shared dispensability under laboratory conditions rather than genuine co-regulation. 1
The observed enrichment included flagellum-dependent cell motility in 5 organisms with mean odds ratio (OR) 4.7, flagellum assembly in 5 organisms with mean OR 5.3, histidine biosynthesis in 3 organisms with mean OR 5.3, and tryptophan biosynthesis in 3 organisms with mean OR 5.3. 1 These signals therefore do not by themselves distinguish shared regulation from shared dispensability. 1
Evidence consistent with a dispensability-driven signal includes the absence of energy-metabolism enrichment in 0/25 organisms, with a permutation-test fold of 0.91. 1 The report also notes that Pearson correlation removes each gene’s mean fitness before correlating profiles, but genes with similar condition-responsive dispensability can still have correlated fitness patterns without direct co-regulation. 1
The broader fitness-conservation analysis supports treating conservation matching as incomplete: essential genes were 82% core whereas always-neutral genes were 66% core, but the association between fitness breadth and core status was weak (Spearman rho=0.086, p=8.1e-230). 2 It also found that strong condition-specific effects were enriched among core genes (77.3% core versus 70.3% without specific phenotypes), so conservation class cannot be treated as a proxy for either dispensability or condition dependence. 2 The synthesis refines this point by showing that core genes were also more likely to be burdensome in laboratory conditions: 24.4% had positive fitness when deleted versus 19.9% of accessory genes, and core genes were 1.29x more likely to be important in some conditions but burdensome in others. 3
Proposed null model
The primary proposed test is to draw random non-AMR genes while matching the AMR genes’ mean-fitness distribution, including the reported −0.05 to +0.05 range. 1 The report calls this fitness-matched permutation critical and the single most important follow-up analysis. The test has two outcomes. If random non-AMR genes with fitness between −0.05 and +0.05 also show flagellar or biosynthesis enrichment in their cofitness neighborhoods, the enrichment is a shared-dispensability artifact. If AMR genes uniquely enrich for flagella even among fitness-matched genes, the co-regulation interpretation is supported. 1 The resulting randomized neighborhoods should be analyzed with the same cofitness thresholds, module assignments, and enrichment procedure as the observed AMR neighborhoods so that differences reflect AMR association rather than changes in analysis scale. 1
A useful implementation would preserve conservation class while additionally matching mean fitness, because the existing analysis already matched conservation class and the unresolved confound is the lack of mean-fitness matching. 1 Repeating the enrichment tests across many such permutations would provide a null distribution for odds ratios, enriched-term counts, and the number of organisms sharing each term. 1
The broader synthesis reported 86% of fitness-module genes as core versus an 81.5% baseline, with OR=1.46 and p=1.6e-87. 3 The module-level analyses support and refine this design: independent component analysis (ICA), a decomposition method for coordinated fitness modules, identified 1,116 modules across 32 organisms, with 94.2% showing significantly elevated within-module cofitness and module genes showing 22.7x genomic-adjacency enrichment. 4 Module genes were 86.0% core versus 81.5% for all genes (OR=1.46, p=1.6e-87). 5 Thus, conservation matching alone may leave structured module composition in the null, while module-level matching is constrained because essential genes are absent from these insertion-based modules. 5 The new module evidence refines rather than replaces gene-level matching: ICA modules represent process-level co-regulation, not reliable gene-level molecular-function assignments; in held-out benchmarking, Module-ICA had <1% strict KEGG KO precision and 23.3% coverage, whereas ortholog transfer achieved 95.8% strict precision and 91.2% coverage. 4
The new comparison also supports testing module structure as a sensitivity factor rather than treating it as a direct functional label: module sizes had a median range of 7-50 genes and within-module mean |r| was 0.34 versus 0.12 in the background, a 2.8x enrichment. 4 The core genes showed heavier tails in both negative and positive fitness effects, while singleton genes had near-zero mean fitness that may reflect poor transposon coverage rather than true neutrality. 2 Where data permit, the null should therefore assess fitness-effect breadth or coverage in addition to mean fitness, or explicitly test whether those features alter the enrichment result. 2
The null should be evaluated at more than one cofitness threshold because the mean support network contained 233 genes at |r| > 0.3, 110 at |r| > 0.4, and 71 at |r| > 0.5. 1 The report specifically recommends confirmation at |r| > 0.4 because the |r| > 0.3 networks include many weak associations. 1
Interpreting possible outcomes
If flagellar motility and amino acid-biosynthesis enrichment remains stronger than the fitness-matched null, the result would support—but would not by itself prove—shared condition-dependent regulation or another structured biological association. 1 If the enrichment disappears after matching mean fitness, the evidence would favor shared dispensability under the laboratory conditions as the explanation for the original signal. 1
The distinction matters because the original InterProScan analysis found 35/3,193 significant tests at FDR (false discovery rate) < 0.05, whereas the old SEED/KEGG analysis found 0/280 significant tests at FDR < 0.05. 1 A separate old-SEED permutation analysis is reported with 23 significant results among 250 tests, and membrane/cell wall as the top signal (24%, fold 1.12). This permutation result is distinct from the 0/280 enrichment result, and the two should not be conflated. 1 Improved annotation can expose real functional structure, but it can also make it more important to use a null model that controls the fitness properties of the compared genes. The ICA benchmark supports retaining this caution: its module-level predictions were weak for exact KEGG assignments despite identifying coherent fitness modules, so enrichment surviving a module-aware null should still not be interpreted as proof of a specific molecular function. 4
The fitness-matched test should not be used to reinterpret every result in the analysis: AMR-containing ICA modules had median size 46 genes versus 27 genes for non-AMR modules, with Mann–Whitney U (MWU) p = 1.7×10⁻⁸, and the report treats this module-size result as robust to the shared-dispensability concern. 1 Likewise, support-network size was not correlated with AMR fitness cost (Spearman rho = −0.006, p = 0.87, N = 769), although limited variation in fitness cost may reduce the ability to detect such a relationship. 1
Relation to organism-specific network structure
A fitness-matched null is mainly needed to interpret functional enrichment, not to erase the observed organism-specific organization of support networks. 1 Different AMR mechanisms within the same organism had mean Jaccard similarity 0.375, compared with 0.207 for the same mechanism across organisms, with MWU p = 4.3×10⁻¹³. 1 The report describes this relative comparison as robust to the dispensability confound because it tests how support networks are organized within and across organisms rather than relying only on the presence of individual enriched categories. 1
The planned null model therefore complements cofitness-network-architecture: it asks whether particular functional categories are overrepresented beyond expected fitness similarity, while the network-organization comparison asks whether organism context structures support relationships more strongly than AMR mechanism. 1 The 156 cross-organism ICA module families, including one spanning 21 organisms, support testing whether such conserved process-level structure also contributes to enrichment, while the benchmark refines interpretation by showing that conservation of a module does not establish identical molecular functions for every member. 4
Background-calibrated comparators beyond cofitness
A Caulobacter envelope-stress analysis extends the background-matching argument from cofitness enrichment to fitness-threshold tests. Its envelope-stress subset covered 22 experiments. The genome background phenotype-bearing rate (|fitness t| > 4 in ≥ 2 envelope-stress experiments) was 33.25% (n=3943 Caulobacter genes; K=1311 phenotype-bearing). 6 The pre-registered ≥10% threshold sits below this background rate. It is therefore essentially a presence test rather than an enrichment test, and passing it is by itself uninformative about specific mechanistic importance. The revised report uses hypergeometric enrichment against background as the formal H2 verdict. 6 The adversarial review drew a general methodological lesson for Fitness Browser projects. Pre-registered fitness-phenotype thresholds must be calibrated against the genome-wide background rate rather than set as a fixed percentage, because a ≥10% threshold against a 33% background passes any random gene set. The report recommends that future thresholds use fold-enrichment over background. 6
Under the recalibrated test, Path A (concordant_strong, clean Fur signature) had 17 of 32 genes phenotype-bearing (53.1%; also reported as 53% observed vs 33% background). This is a fold of 1.60× with hypergeometric p = 0.016; the report describes it as marginal enrichment and records Path A as SUPPORTED. 6 Path B (SspB-buffered, cbb3/fix-rich) had 9 of 26 genes phenotype-bearing (34.6%), a fold of 1.04× with hypergeometric p = 0.515. This rate is indistinguishable from the 33.25% genome background, and Path B is recorded as NOT SUPPORTED. 6 The report attributes the earlier result to the pre-registered ≥10% threshold sitting below the 33.25% background rate. It therefore treats the “respiratory ATP required” arm of its dual-release switch model as a working hypothesis rather than an established finding. 6 This single-organism result supports the page’s central caution that passing a criterion is not the same as exceeding an appropriate background. Its background, however, is a genome-wide phenotype rate rather than a draw of genes matched on mean fitness. 6
A null control in the functional-dark-matter project supports using matched comparators for concordance claims. For genes with orthologs in 3+ organisms, dark and annotated genes did not differ significantly in cross-organism concordance (Mann-Whitney p=0.17, KS p=1.0). Both groups had a median of 1.0; the dark mean was 0.976 and the annotated mean was 0.985. The report treats this NB06 null control as supporting H1. 7 However, the same report notes that equivalent null controls were not run for its biogeographic and lab-field concordance analyses, so those concordance results lack a comparable background check. 7
The IBD phage-targeting project supports the same principle from the opposite direction, showing how an unmatched background inflates an effect size. Its NB08a Tier-A comparator was the full biosynthetic gene cluster (BGC) catalog (10,060 BGCs across all species), not a matched comparator. The 44× Fisher OR for iron_siderophore is partly inflated by the catalog-wide rarity of iron MIBiG matches (51 / 9,774 = 0.5 %). 8 The report suggests a more conservative matched-niche pathobiont comparator: gut Proteobacteria plus CD-associated Firmicutes at similar genome-assembly depth. It states that the 44× figure should not be treated as a precise effect size. It nonetheless judges the qualitative finding robust across choice of comparator: E. coli alone among actionable Tier-A carries the iron-siderophore signature. 8 Per-species CB-ORF CD-up rates at FDR<0.10 were likewise compared against a catalog background (2.5 %) rather than a stated matched comparator. 8
Evidence status
The need for a fitness-matched null is a methodological conclusion supported by a specific unresolved confound, not evidence that the observed enrichment is artifactual. 1 The current data support the hypothesis that flagellar, chemotaxis, and amino acid-biosynthesis enrichment may reflect shared dispensability, while leaving genuine co-regulation as an alternative explanation. 1 The fitness-conservation comparison supports the need to control assay-visible fitness structure, but its weak conservation association and coverage caveat mean that it does not establish that mean-fitness matching alone is sufficient. 2 The synthesis supports retaining this cautious interpretation: its costly-and-conserved category is evidence for, rather than a direct measurement of, natural-environment purifying selection. 3 The module-conservation result supports adding module composition to the sensitivity analysis, but its ICA modules exclude essential genes and therefore cannot establish how essential-gene structure would affect the null. 5 The fitness-modules benchmark refines the evidence standard: module coherence and cross-organism conservation can support process-level structure, but exact gene-function claims require stronger sequence-based evidence. 4
The source reports are summarized at AMR Co-Fitness Support Networks, Fitness Effects vs Conservation — Quantitative Analysis, Gene Conservation, Fitness, and the Architecture of Bacterial Genomes, Fitness Modules × Pangenome Conservation, Pan-bacterial Fitness Modules via Independent Component Analysis, Regulatory and proteomic architecture of Δfur-permitted lipid A loss in *Caulobacter crescentus*, Functional Dark Matter — Experimentally Prioritized Novel Genetic Systems, and Metagenome-Prioritized Phage Cocktails for Crohn’s Disease and IBD.
Open Directions
- Use the existing AMR and non-AMR fitness matrices, conservation classes, and mean-fitness values to run conservation- and fitness-matched permutations; test whether flagellar and amino acid-biosynthesis enrichment remains significant. 1
- Recompute cofitness separately for antibiotic-exposure and standard-growth conditions, then apply the matched null; ask whether enrichment is specific to resistance-relevant conditions or persists under general laboratory growth. 1
- Compare observed odds ratios and enriched-term counts with null distributions at |r| > 0.3, |r| > 0.4, and |r| > 0.5; ask whether the functional signal survives removal of weaker associations. 1
- Measure mean fitness directly for flagellar knockouts and other conditionally dispensable gene classes, then use those distributions in the null; ask whether their observed neighborhoods are predictable from dispensability alone. 1
- Extend the matched-null analysis to phage-defense and secondary-metabolite genes; ask whether enrichment of other conditionally dispensable classes is similarly explained by fitness structure. The report notes that if these classes show the same cofitness patterns as AMR genes, the signal is a general feature of dispensable genes under laboratory conditions rather than an AMR-specific one. 1
- Replace the operon-exclusion row-index heuristic with coordinate-based genomic filtering before permutation testing; ask whether local-gene structure changes the enrichment estimates. 1
- Add fitness breadth, condition-specific-effect status, and transposon-callability measures to the matching or stratification scheme; ask whether the enrichment survives controls motivated by the weak conservation–fitness association and possible singleton coverage bias. 2
- Test whether the 1,116 conservation-enriched fitness modules alter functional-enrichment null distributions when module composition, rather than only gene-level conservation and mean fitness, is matched. 3,4
- Re-run the module-composition sensitivity analysis on the 29/32 organisms with pangenome links, and separately assess whether exclusion of essential genes changes enrichment conclusions. 5
- Compare enrichment against nulls that preserve ICA module membership but randomize within module, then test whether surviving terms remain process-level associations rather than exact KEGG KO predictions. 4
- Re-test the Caulobacter Path B gene set with fold-enrichment over the 33.25% envelope-stress background and with size-matched random gene draws; ask whether any envelope-stress subset separates the respiratory-ATP arm from background. 6
- Run annotated-gene null controls equivalent to NB06 for the biogeographic and lab-field concordance analyses; ask whether dark-gene concordance in those analyses exceeds what matched annotated genes achieve. 7
- Recompute the iron_siderophore Fisher OR against a matched-niche pathobiont comparator at similar genome-assembly depth; ask how far the 44× estimate changes while the qualitative E. coli signal is retained or lost. 8