Fitness-Module Discovery Depends on Thresholds and Decomposition Constraints

Independent component analysis (ICA), a statistical method that decomposes correlated measurements into components, can identify groups of genes with coordinated fitness patterns across experiments, but the resulting modules depend on membership thresholds, annotation settings, experiment count, experimental diversity, and the allowed number of components. ₁ ICA is therefore a process-level discovery tool: it can associate genes with coordinated biological processes under measured conditions without establishing each gene’s precise molecular function. ₁ This distinction refines cofitness-network-architecture and gene-essentiality, because fitness covariance provides process-level context whereas gene-level assignments require higher-resolution evidence. ₁

Threshold choice changes module quality

The Pan-bacterial Fitness Modules via Independent Component Analysis analyzed RB-TnSeq, a pooled transposon-mutant method for measuring gene fitness across conditions, from 32 organisms. ₁ The initial D’Agostino K-squared membership approach produced modules containing 100-280 genes, with 59% enrichment and 1-17x correlation enrichment. ₁ Replacing that rule with an absolute component-weight threshold of |weight| >= 0.3 and limiting each module to a maximum of 50 genes produced biologically coherent modules, with 94% enrichment and 2.8x correlation enrichment. ₁ This shows that thresholding is not cosmetic: it changes module size and the strength of the recovered cofitness signal. ₁

The cross-project digests support this result and identify the mechanism. ICA weight distributions are intentionally non-Gaussian (heavy-tailed), so the D’Agostino K² normality test always rejected normality and fell back to a permissive z-threshold (2.5) that let 100-280 genes into each module. ₂ The ICA components themselves were fine; the failure lay in deciding which genes belong to each module. ₃ Switching to an absolute cutoff of |Pearson r| ≥ 0.3 with the module profile and a maximum of 50 genes reduced modules to 5-50 genes. ₃ The same switch raised cofitness enrichment from 59% to 94.2% and raised mean within-module |r| across 32 organisms from 0.047 to 0.34. ₃ The project report calls this strict membership threshold critical, and the pitfalls digest calls it the single most impactful parameter in the pipeline. _1,2 The two sources describe the cutoff differently. The project report phrases it as |weight| >= 0.3, whereas the digests phrase it as a correlation with the module profile. The report’s 1-17x correlation figure for the initial approach and the digest’s 0.047 within-module |r| are also different metrics. Neither discrepancy is reconciled here. _1,3

The final analysis identified 1,116 stable modules across 32 organisms, with each organism represented by at least 100 experiments. ₁ Module sizes had a median range of 7-50 genes. ₁ Of these modules, 94.2% showed significantly elevated within-module cofitness by Mann-Whitney U test, a rank-based test of distributional differences, at p < 0.05. ₁ Within-module mean |r| was 0.34 versus 0.12 in the background, corresponding to a 2.8x enrichment, while module genes showed 22.7x genomic-adjacency enrichment. ₁ The adjacency result supports the interpretation that many recovered modules reflect operon co-localization as well as correlated fitness behavior, rather than demonstrating a precise function for every member gene. ₁

Module recovery also depends on the stabilization procedure. Robust ICA combines 30-50 FastICA runs with DBSCAN, a density-based clustering of components, and consistently found 17-52 stable modules per organism across 32 bacteria. ₃ The digest reports the same validation profile: 94.2% of modules showed significantly elevated within-module cofitness (Mann-Whitney U, p < 0.05; mean |r| = 0.34 versus background 0.12), and genomic-adjacency enrichment averaged 22.7× across organisms. It interprets this profile as confirming operon-like co-regulated gene groups. ₃ A purely numerical failure mode sits inside the clustering step. Precomputed cosine distances (1 - |cosine_similarity|) can take tiny negative values from floating-point error, and DBSCAN with a precomputed metric rejects them. Distances must therefore be clipped to the range 0 to 1 before clustering. ₂

Module size itself carries biological signal that interacts with the membership ceiling. The discoveries digest reports that AMR (antimicrobial resistance) genes sit in larger-than-average ICA modules (median 46 vs 27, p=1.7e-8). ₃ Membership is capped at max 50 genes, and a median this close to the cap suggests the hypothesis that the cap truncates AMR-containing modules more than typical modules. The digest quote does not state how many AMR genes were assigned to modules, so this evidence does not show how well modules cover AMR genes. _3,1

The conservation analysis supports this threshold-sensitivity interpretation: its upstream membership rule used |Pearson r| >= 0.3 and a maximum of 50 genes per module, so the resulting conservation profile is conditional on membership constraints rather than an invariant property of the underlying fitness data. ₄ Among 974 modules with at least 3 mapped genes, 577 (59%) were >90% core, 349 (36%) were 50-90% core, and 48 (5%) were <50% core; these classifications use convenient cutoffs that can also shape apparent module architecture. ₄ The new synthesis supports interpreting this enrichment as organized but modest: across 1,116 ICA-derived modules, module genes were 86% core versus an 81.5% baseline, with an odds ratio of 1.46 and p=1.6e-87, while 59% of modules were more than 90% core. ₅ Because these estimates inherit membership and core/accessory thresholds, they strengthen the case for coordinated conservation without making it threshold-independent. ₅

The genome-wide fitness-conservation analysis refines this interpretation by showing that fitness importance is positively but only weakly associated with conservation: across approximately 194,000 genes from 43 diverse bacteria, essential genes were 82% core whereas always-neutral genes were 66% core. ₆ The synthesis reports the same gradient across 194,216 protein-coding genes from 43 bacteria, and therefore supports treating module-level core enrichment as a quantitative tendency rather than a deterministic fitness-to-conservation rule. ₅ Thus, module-level conservation enrichment should not be treated as a direct measure of module fitness importance, and the magnitude of the association remains limited even when fitness is measured across many conditions. ₆

The synthesis also refines the interpretation of conservation by showing that core genes can be burdensome in laboratory conditions: 24.4% had positive fitness when deleted, compared with 19.9% of accessory genes; core genes were 1.78x more likely to have strong condition-specific phenotypes and 1.29x more likely to be important in some conditions but burdensome in others. ₅ Consequently, a conserved module should not be assumed to be an inert housekeeping unit or uniformly beneficial under every assay condition. ₅

Experimental depth and decomposition constrain reliability

The experiment-depth evidence refines the threshold result: organisms represented by fewer experiments produced weaker module structure, whereas broader experiment coverage supported more stable cofitness inference. ₁ The approximately 100-experiment inclusion threshold supports reproducible structure in the reported collection, but it is not a universal reliability cutoff. ₁ Caulo, despite having 198 experiments, showed only 2.9x correlation enrichment, demonstrating that meeting a nominal experiment-count threshold does not guarantee uniformly strong module recovery. ₁ Experiment number alone is therefore insufficient; organism-specific biology, condition composition, and experimental diversity may also affect module quality. ₁

The digest refines the depth result and exposes an inconsistency in the project’s framing. Caulo, with 198 experiments, showed the weakest pilot signal at 2.9x correlation enrichment versus 31x for DvH. Yet organisms down to 104 experiments (Ponti) still produced stable modules. ₃ The project report cites Caulo as an example that organisms with fewer than ~100 experiments produce weaker modules. Caulo’s 198 experiments do not fall below that stated threshold, so the report’s own example does not demonstrate the depth claim. ₁ The digest identifies capping components at 40% of experiments as the key, because higher ratios caused FastICA convergence failures and extreme slowness. ₃

The analysis imposed a 40% component cap, meaning that the number of components could not exceed 40% of the number of experiments, to avoid FastICA convergence failures. ₁ This constraint improves computational stability but may cause some modules to be missed in organisms with few experiments. ₁ Limited depth can consequently weaken cofitness estimates while also restricting the number of components that can be stably extracted. ₁ Differences in depth and the component cap affect both sensitivity and comparability across organisms, so a weak or absent module should not automatically be interpreted as evidence of absent biological conservation. ₁

The pitfalls digest turns the cap into an operational rule: keep n_components ≤ 40% of n_experiments. Higher ratios cause frequent convergence failures, and each failed run hits max_iter. For an organism with 150 experiments, this means at most 60 components. ₂ This supports reading the cap as a computational-stability constraint rather than a biologically derived number of components. _2,1

The conservation-linked analysis further supports this qualification: pangenome mappings covered 29/32 organisms because Cola, Kang, and SB2B lacked sufficient GTDB genomes for pangenome construction, making its conservation estimates incomplete across the full module collection. ₄ It also found no family-breadth/conservation association (Spearman rho=-0.01, p=0.914), redirecting interpretation away from module-family scope and toward gene-level conservation under a high approximately 81.5% all-gene core baseline. ₄

The broader fitness-conservation analysis supports treating missingness and callability as important qualifications: singleton and novel genes were near-neutral in tested laboratory assays, but that apparent neutrality may reflect poor transposon coverage rather than true lack of fitness effects. ₆ It also found that essential genes lack viable mutants in the module-conservation analysis, so fitness-module results are not representative of the full genome-wide fitness spectrum. ₄ The new synthesis supports this caveat by distinguishing costly-and-conserved genes, which are evidence for purifying selection rather than direct measurements of natural fitness, from costly-and-dispensable genes that may represent gene loss, mobile elements, or recently acquired functions. ₅

Annotation thresholds do not equal discovery thresholds

Lowering the enrichment-overlap threshold from 3 to 2 and adding PFam domains increased the module annotation rate from 8% to 80%. ₁ The number of annotated modules increased from 92 to 890, and the resulting changes unlocked 7.6x more function predictions. ₁ PFam provided the broadest annotation coverage, whereas KEGG KOs were too gene-specific for module-level enrichment. ₁ Because PFam operates at the domain level, its broader coverage may overcount functional associations. ₁

The digests support this result and explain why: the min_annotated overlap parameter must match the granularity of the annotation database. ₃ KEGG KOs average ~1.2 genes per KO, so almost no KEGG term has 3+ genes in a single 5-50 gene module. With min_annotated=3, only 8% of modules got any enrichment. ₂ Lowering the parameter to 2 and adding PFam domains (814 terms with 2+ genes) alongside TIGRFam (88 terms) raised module annotation from 8% to 80%. ₂ The digest argues that Fisher’s exact test handles small counts correctly. In its view, statistical validity comes from false-discovery-rate (FDR) correction across all tests, not from large per-test overlaps. This qualifies but does not remove the PFam overcounting caveat above. _3,1

These annotation choices differ from the component-membership threshold: the |weight| >= 0.3 rule determines which genes enter a module, whereas the enrichment-overlap threshold determines how module content is associated with functional annotations. ₁ Thus, both module construction and downstream annotation settings can alter apparent biological yield without changing the underlying fitness data. ₁ The conservation analysis supports keeping these layers distinct: its core/accessory classifications describe mapped gene conservation, not functional annotation accuracy, and its accessory-module pattern is only a hypothesis about horizontally transferred or niche-specific units rather than direct evidence of either process. ₄

The fitness-conservation results refine the accessory-module interpretation: condition-specific effects were enriched among core genes, with genes carrying strong condition-specific phenotypes being 77.3% core versus 70.3% core for genes without specific phenotypes, and 4,450 genes fit an ephemeral-niche pattern of overall neutrality but criticality in one condition. ₆ This contradicts the expectation that condition-specific or niche-relevant effects are predominantly accessory, while still not establishing that every condition-specific module is core or that accessory genes lack ecological effects. ₆ The synthesis supports this contradiction at the broader functional level: motility and chemotaxis, RNA metabolism, and protein metabolism showed core-gene burden excesses of +7.8pp, +12.9pp, and +6.2pp respectively, whereas non-core cell-wall genes were more burdensome. ₅

Process-level modules are not gene-level function calls

In a held-out benchmark withholding 20% of KEGG-annotated genes, ortholog transfer achieved 95.8% strict precision, 91.2% coverage, and 0.934 F1, where F1 is the harmonic mean of precision and recall. ₁ Domain-based prediction achieved 29.1% precision, 66.6% coverage, and 0.401 F1. ₁ Module-ICA achieved <1% strict precision and 23.3% coverage, while cofitness voting achieved <1% strict precision and 73.0% coverage. ₁ These results support using ortholog transfer for specific molecular-function prediction and ICA modules for biological-process context. ₁

Module-ICA and cofitness had near-zero strict KEGG KO precision because KEGG KO groups are gene-level assignments, averaging approximately 1.2 genes per unique KO, whereas a module with 20 annotated members typically contained 20 different KOs. ₁ Module membership can therefore indicate participation in a shared process or co-regulated program, but it cannot establish that a protein has a particular KO-defined molecular function. ₁ The 6,691 function predictions generated for hypothetical proteins should consequently be interpreted as process-level evidence rather than proof of a specific KO-defined function. ₁ Of those predictions, 2,455 were family-backed, representing 37% and carrying cross-organism conservation support, while 4,236 were module-only predictions. ₁ The predictions used enrichment from KEGG, SEED, TIGRFam, and PFam, so their evidential basis ranged from cross-organism conservation to module-level or domain-level association rather than uniformly demonstrating a precise molecular function. ₁

This evidence supports a boundary between process-level and gene-level inference: improved annotation coverage increases the amount of process-level interpretation available from modules, but does not convert module associations or domain-level enrichment into validated gene-level assignments. ₁ The conservation result refines that boundary: zero essential genes appeared in the ICA modules because essential genes lack usable transposon-insertion fitness data, so the observed conservation profile describes the non-essential, insertion-callable portion of the genome rather than essential-gene modules. ₄ The discoveries digest reports the same null directly: 0 essential genes appear in any of the 1,116 fitness modules across 32 organisms. Genes with no transposon insertions have no fitness data for ICA to decompose, so fitness modules capture only the non-essential portion of the genome’s functional architecture. ₃

The genome-wide comparison further supports this caution: core genes showed stronger fitness effects in both negative and positive directions, while novel singleton genes were largely neutral under tested laboratory conditions, with poor transposon coverage a possible explanation. ₆ Therefore, a module’s apparent functional quietness or enrichment cannot by itself distinguish biological neutrality from assay and insertion-callability limits. ₆

Cross-organism stability is alignment- and depth-dependent

Cross-organism alignment produced 1.15M bidirectional-best-hit (BBH) pairs, where BBH denotes reciprocal best sequence matches, across 32 organisms and 13,402 ortholog groups. ₁ It identified 156 module families spanning 2+ organisms, including 28 spanning 5+, 7 spanning 10+, and 1 spanning 21 organisms. ₁ A total of 145 families had consensus functional labels, representing 93% of the families. ₁ These conserved module families provide evidence for recurring fitness-associated co-regulation across diverse bacterial phyla, but BBH-based alignment represents conserved co-regulation patterns rather than proof that every family member has an identical molecular function. ₁

The discoveries digest reports the same alignment result: BBH ortholog fingerprints revealed 156 module families spanning 2+ organisms, 28 families spanning 5+ organisms, and a largest family spanning 21 of 32 organisms across Proteobacteria, Bacteroidetes, Firmicutes, and Archaea. It calls the families evidence of deeply conserved fitness regulons, which is a stronger interpretation than the process-level reading adopted here. ₃ Family counts are themselves sensitive to orthology scope. Using ortholog pairs from a 5-organism pilot for the 32-organism analysis produced 6x fewer module families and missed all families spanning 5+ organisms, because the ortholog graph gains transitive connections from each new organism. ₂ BBH pairs therefore must be extracted for all organisms in the analysis; partial extraction severely underestimates cross-organism conservation. ₂

The depth and component-cap evidence qualifies this cross-organism result: differences in experimental depth and extraction limits can influence which modules are detected in each organism, even when shared biology is present. ₁ Stable module families can reveal conserved process-level structure, but their membership, annotation, and cross-organism comparability depend on the analysis regime. ₁ The conservation analysis supports this qualification: module genes were 86.0% core versus 81.5% for all genes (OR=1.46, p=1.6e-87), a statistically strong but modest +4.5 percentage-point enrichment under the high baseline; family breadth itself was not predictive of conservation. ₄

The genome-wide fitness-conservation result supports interpreting this enrichment as modest rather than as a deterministic fitness-to-conservation rule: the broader analysis found a 16-percentage-point core difference between essential and always-neutral genes, but described fitness importance as only a weak predictor of conservation. ₆ It also refines the interpretation of module biology by showing that core genes can be both more strongly costly and more strongly beneficial across conditions, consistent with trade-offs rather than a simple core-versus-accessory activity ranking. ₆ The synthesis supports this trade-off interpretation: 28,017 genes were both costly in the laboratory and conserved in the pangenome, whereas 5,526 were costly and dispensable; the former are evidence for, not direct measurements of, purifying selection in natural environments. ₅

This finding supports cofitness-network-architecture by quantifying how module construction and experiment depth affect recovered network structure, refines condition-specific-fitness by extending condition-dependent fitness analysis from individual genes to independently regulated modules, and supports pangenome-integration by connecting module families to cross-organism ortholog groups. ₁ The conservation result likewise supports pangenome-integration while preserving the distinction between conservation of process context and identity of gene function. ₁

The full project reports are available at Pan-bacterial Fitness Modules via Independent Component Analysis, Fitness Modules × Pangenome Conservation, Fitness Effects vs Conservation — Quantitative Analysis, and Gene Conservation, Fitness, and the Architecture of Bacterial Genomes. _{1 4 6 5}

Discrete modules depend on method and condition space

A single-organism deletion screen in acinetobacter-baylyi-adp1 qualifies the assumption that fitness data naturally partition into discrete modules. Hierarchical clustering of 2,034 genes by their 8-condition growth profiles gave an optimal K = 3. Its silhouette score, a measure of how well separated the clusters are, was low at 0.24 and stayed low across all tested k values (2–25). _7,3 The two large clusters, of 1,160 and 850 genes, were broad ‘generally sensitive’ and ‘generally tolerant’ groups. No specific functional enrichment survived FDR correction, and the module-enrichment output file has 0 rows. ₇ The report interprets this as gene essentiality varying continuously across conditions, consistent with Guzman et al. (2018). _7,3

The one exception was a 24-gene aromatic-degradation module. It showed extreme quinate-specific defects (mean z-score = -7.28 on quinate, near-zero on other conditions) and was the only discrete phenotypic module in that dataset. ₇ The report contrasts this with Nichols et al. (2011), whose E. coli chemical-genetic profiles defined phenotypic modules extending beyond operon boundaries. It proposes, without establishing, that the difference reflects single-gene deletions versus chemical perturbations, or a more interconnected ADP1 metabolic architecture. ₇ This is a single-organism, single-method result. It does not by itself contradict the ICA modules above, which were extracted from organisms with at least 100 experiments rather than 8 conditions. _7,1

Low condition counts also limit cofitness-based assignment. An aromatic-catabolism analysis notes that its co-fitness correlations rely on only 8 conditions (8-dimensional growth vectors), limiting the resolution of gene-gene correlations. ₈ An annotation-gap project proposes applying ICA-based approaches (Borchert et al. 2024) to fitness data to find functional gene modules that per-gene analysis misses. That benefit is a proposal, not a result. ₉

Activity normalization determines which modules look condition-responsive

Once modules exist, whether a module counts as responsive to a condition class depends on the activity scale. The metal-fitness-atlas analysis (NB05) used z-scored module profiles across 31 organisms, standardized per module across all experiments. It found 600 of 19,453 module × metal-experiment records (3.1%) above |z| > 2.0. DvH led with 47 responsive modules across 12 metals, and psRCH2 had 11 modules across 6 metals. ₁₀ The 183 metal-responsive modules with conservation data had a mean core fraction of 0.826 and a median of 0.929. The report reads this as consistent with core enrichment of metal fitness genes. ₁₀ An initial attempt using raw activity scores from module_conditions.csv produced 0 responsive modules, because raw scores (max 0.96) are on a different scale from z-scores. The report states that using per-experiment profiles with z-normalization resolved this scale mismatch. ₁₀

The metal-specificity analysis reached a null at the module level: z-scored activity profiles yielded 0 metal-specific modules. Per-module z-normalization produced max |z| values < 2.0 for most metal experiments, because metal experiments are a small fraction of each organism’s total experiments. ₁₁ The pitfalls digest gives an example, 12/176 metal experiments for MR1, in which the z-score denominator is dominated by the non-metal majority. ₂ The digest attributes the Atlas’s success to its use of precomputed z-scores from metal_modules.csv. It states that recomputing z-scores from raw module_conditions.csv activity does not reproduce the same responsiveness threshold. The specificity report likewise notes that its raw activity scores were on a different scale from the Atlas’s z-scored profiles. _2,11 The 0-module null may therefore partly reflect normalization rather than an absence of metal-related modules. ‘Responsive’ and ‘specific’ are also different criteria. _11,2

Whether metal cross-resistance genes cluster into coherent regulatory units remains untested. The cross-resistance report proposes applying ICA from the fitness_modules project specifically to metal conditions as future work. ₁₂

Tensions

The module-conservation analysis found an 86.0% core fraction for module genes versus 81.5% for all genes, while the genome-wide fitness-conservation analysis found only a weak association between fitness importance and conservation and emphasized that novel singleton genes may appear neutral because of poor transposon coverage. _{4 6} These findings are not numerically identical estimands: the former is conditional on detected, non-essential module members and selected module thresholds, whereas the latter spans broader fitness categories and approximately 194,000 genes. _{4 6}

The new synthesis supports the module-level enrichment but qualifies its interpretation: it reports 86% core module genes against an 81.5% baseline and 1,116 modules, while also finding only a modest conservation gradient and documenting that core genes may be laboratory-burdensome. ₅ Thus, module conservation is compatible with coordinated evolutionary retention, but it does not establish that modules are uniformly fitness-beneficial, environmentally essential, or independent of threshold and callability constraints. ₅

The metal-module analyses disagree about what z-scoring does. The Metal Fitness Atlas reports z-scored profiles that identified 600 responsive records. The metal-specificity analysis reports that per-module z-normalization produced max |z| < 2.0 for most metal experiments and found 0 metal-specific modules. _10,11 The pitfalls digest describes the Atlas z-scores as standardized per module across all experiments and attributes the difference to precomputed versus recomputed z-scores. The Atlas report itself credits per-experiment profiles with z-normalization. The procedural discrepancy is therefore documented but not explained. _2,10

The ADP1 deletion screen found a continuous phenotype gradient with only one discrete 24-gene module (silhouette 0.24). In contrast, robust ICA recovered 1,116 stable modules across 32 organisms, 94.2% of them with elevated within-module cofitness. _7,3,1 The two analyses differ in method (hierarchical clustering versus ICA), perturbation type (single-gene deletions versus pooled transposon mutants) and condition count (8 conditions versus at least 100 experiments). They are therefore not direct contradictions, and which factor drives the difference is unresolved. _7,1

Open Directions

• Re-run ICA across organisms while varying the |weight| threshold, 50-gene maximum, and 40% component cap; test which settings preserve the 94.2% significant-module rate, 2.8x correlation enrichment, and 22.7x genomic-adjacency enrichment. ₁
• Use matched experiment subsets, bootstrap or split-sample ICA, and convergence diagnostics to quantify recovery below and above approximately 100 experiments, including Caulo’s 198 experiments and 2.9x enrichment. ₁
• Balance condition classes across organisms to test whether experimental diversity, rather than experiment count alone, explains module-reliability differences. ₁
• Compare PFam, KEGG KO, and other annotation layers at fixed memberships to determine whether the increase from 8% to 80% reflects genuine coverage or domain-level overcounting. ₁
• Combine the 6,691 hypothetical-protein predictions with ortholog transfer, domain architectures, and targeted gene-level experiments to test which module-only predictions acquire specific molecular-function support. ₁
• Reanalyze the held-out benchmark with process-level gold standards in addition to KEGG KO labels to ask whether Module-ICA has useful precision when evaluated against biological-process membership rather than gene-level identity. ₁
• Benchmark module-level predictions against independently measured phenotypes or perturbation data to test whether process-level predictions improve interpretation despite <1% strict KEGG KO precision. ₁
• Use larger, condition-diverse RB-TnSeq datasets and compare module stability before and after the 40% component cap to determine which conserved modules are missed in organisms with fewer experiments. ₁
• Reconstruct cross-organism families under alternative orthology definitions and after downsampling every organism to the same experiment count; test whether the 156 families, including the family spanning 21 organisms, remain stable without assuming identical molecular functions. ₁
• Test the 156 cross-organism module families with gene-neighborhood, domain-combination, and experimental perturbation data to determine whether shared module membership reflects conserved regulation, conserved pathway involvement, or recurrent genomic organization. ₁
• Recalculate module conservation under alternative membership and core/accessory thresholds, with essentiality-aware missingness and matched pangenome coverage, to test whether the 86.0% versus 81.5% enrichment and the rho=-0.01 family-breadth result persist. ₄
• Reanalyze module conservation with fitness breadth, strongest fitness effect, condition-specific phenotype tags, and transposon-callability metrics to test whether the apparent core enrichment reflects broad functional importance, conditional effects, or ascertainment. ₆
• Use condition-diverse assays and targeted perturbations for singleton and ephemeral-niche genes to distinguish true laboratory neutrality from poor transposon coverage and to test whether condition-specific module effects are core-enriched outside standard laboratory conditions. ₆
• Partition the 28,017 costly-and-conserved and 5,526 costly-and-dispensable genes by module membership, mobile-element evidence, and environmental breadth; test whether the two classes differ in callability, recent acquisition, or loss signatures. ₅
• Reassess the 86% versus 81.5% module-core enrichment after matching organisms on experiment count, condition diversity, pangenome coverage, and essential-gene missingness to determine how much of the signal is biological versus ascertainment. ₅
• Apply ICA to the ADP1 deletion fitness data, as the annotation-gap project proposes, and compare it with hierarchical clustering on the same 8 conditions. This tests whether the gradient (silhouette 0.24) reflects the clustering method or the small condition space. _7,9
• Score the same 19,453 module × metal-experiment records with precomputed metal_modules.csv z-scores, recomputed per-module z-scores and condition-class-stratified normalization. This tests whether the 600-record versus 0-module discrepancy is a normalization artifact. _10,11,2
• Run ICA restricted to metal conditions to test whether cross-resistance genes form coherent co-regulated modules. ₁₂
• Rebuild module families from BBH pairs extracted for growing organism subsets to quantify how family counts depend on orthology scope, given the 6x shortfall from a 5-organism pilot. ₂
• Re-run module membership with a relaxed 50-gene cap to test whether AMR-containing modules (median 46 vs 27) are truncated by the cap. ₃