Independent component analysis of fitness modules

What this entity is

Canonical name: Independent component analysis of fitness modules. ₁

Known aliases: ICA; ICA fitness modules. ₁

Stable external identifier: Not reported in the source documents. ₁

Independent component analysis (ICA) decomposes bacterial fitness matrices into gene modules representing shared fitness patterns across experimental conditions. ₁ The annotation-gap study proposes ICA as complementary to per-gene analysis for detecting functional modules that individual-gene tests may miss. ₂ The counter-ion analysis supports this role by proposing module-level ICA to distinguish shared stress responses from metal-specific fitness signals. ₃

The essential-genome analysis extends ICA from module discovery to indirect function prediction: fitness-module context from non-essential orthologs generated family-backed predictions for hypothetical essential genes, whose lethal disruption prevents direct fitness characterization. ₄ The fitness-modules benchmark refines this interpretation by showing that ICA is substantially stronger for process-level module context than precise gene-level function assignment. ₅ The functional-dark-matter analysis supports ICA as a prioritization layer: 6,142 dark genes belonged to ICA fitness modules, although module membership remains an inference rather than direct functional validation. ₆ The truly-dark-gene analysis extends this use to genes that remain hypothetical after modern annotation: 594 such genes belonged to ICA fitness modules, and module context linked them to candidate functions including phage integrases, metal transporters, chemotaxis systems, and iron regulation. ₇

Evidence and architecture

The AMR cofitness analysis included 28 organisms with antimicrobial-resistance genes, fitness matrices, and ICA modules. ₁ Of 801 AMR genes with fitness data, 192 (24%) were assigned to modules. ₁ AMR-containing modules had a median of 46 genes versus 27 in non-AMR modules (Mann–Whitney U p=1.7×10⁻⁸); efflux and enzymatic AMR modules both had a median size of 48 (p=0.91). ₁ There were 136 unique AMR-containing module families, and 208/209 (99%) AMR gene–module assignments belonged to cross-organism conserved families. ₁ ICA was interpreted as capturing condition-specific co-regulation rather than only shared mean fitness, but module membership does not establish direct transcriptional control. ₁

The broader discoveries synthesis reports the same AMR-module size contrast and InterProScan enrichments for flagellar motility, flagellum assembly, histidine-biosynthesis, and tryptophan-biosynthesis, whereas legacy Fitness Browser SEED annotations detected 0/280 significant enrichments. ₈ The functional-dark-matter analysis extends ICA beyond AMR: among 57,011 dark genes identified among 228,709 genes across 48 organisms, 6,142 belonged to ICA modules; metals, oxidative stress, and osmotic stress dominated strong dark-gene phenotypes. ₆

The cofitness–coinheritance study supports coordinated multi-gene structure: across 195 ICA modules in 6 organisms, within-module co-occurrence exceeded a prevalence-matched null by delta phi=+0.053; 51/195 (26%) were significant at p<0.05 and 21/195 (11%) remained significant at q<0.05 after Benjamini–Hochberg FDR (false discovery rate) correction. ₉ Accessory modules had mean delta phi +0.108 versus +0.059 for core modules, with the difference trending toward significance (Mann–Whitney p=0.051). ₉ This refines the AMR result because module-level structure was stronger than pairwise cofitness, whose aggregate delta was only +0.003. ₉

The fitness-modules analysis used an absolute membership threshold of |weight| >= 0.3 and a maximum of 50 genes, recovering 1,116 stable modules across 32 organisms, each supported by at least 100 experiments. ₅ Module sizes had a median range of 7-50 genes; 94.2% showed significantly elevated within-module cofitness at p < 0.05, with within-module mean |r|=0.34 versus 0.12 in the background, a 2.8x enrichment, and 22.7x genomic-adjacency enrichment. ₅ The initial D’Agostino K-squared membership approach instead produced 100-280 genes per module with 59% enrichment and 1-17x correlation; absolute weight thresholding produced the reported 94% enrichment and 2.8x correlation enrichment. ₅ This refines the discoveries synthesis’s robust-ICA/DBSCAN account, which reported stability at 17–52 modules per organism, 94.2% elevated cofitness, mean |r|=0.34 versus background 0.12, and 22.7× mean genomic-adjacency enrichment. ₈

The module-conservation analysis supports and qualifies this architecture: among 974 modules with >=3 mapped genes, 577 (59%) were >90% core genes, 349 (36%) were mixed (50-90% core), and 48 (5%) were <50% core; the median module was 93.4% core. ₁₀ 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 because the baseline was already high. ₁₀ This extends pangenome conservation from individual genes to co-regulated functional units. ₁₀ Thirty-eight families had <50% core genes; their possible horizontal-transfer or niche-specific-operon interpretation is a hypothesis, because the analysis measured conservation patterns rather than directly demonstrating transfer or niche-specific regulation. ₁₀ Family breadth did not predict conservation (Spearman rho=-0.01, p=0.914), redirecting interpretation toward gene-level conservation rather than cross-organism module scope. ₁₀

The truly-dark-gene analysis supports the use of ICA for annotation-resistant genes but also narrows the inference: among 4,394 truly dark genes in organisms with integrated condition analysis data, 41% of neighboring genes were also hypothetical, 25.9% showed operon-like cofitness with adjacent genes using |r| ≥ 0.3, and 594 genes belonged to ICA fitness modules. ₇ The resulting guilt-by-association links to phage integrases, metal transporters, chemotaxis systems, and iron regulation are candidate functional hypotheses, not validated assignments. ₇

Cross-organism alignment produced 1.15M bidirectional-best-hit (BBH) pairs across 32 organisms and 13,402 ortholog groups. ₅ It identified 156 module families spanning 2+ organisms, including 28 spanning 5+, 7 spanning 10+, and 1 spanning 21 organisms; 145 families had consensus functional labels, representing 93% of families. ₅ This supports the discoveries result that alignment found 156 families, including 28 spanning 5+ organisms and one spanning 21 of 32 organisms, while indicating conserved fitness architecture rather than identical molecular functions for every member. _8,5

Ortholog scope increased module families from 27 to 156, families spanning 5+ organisms from 0 to 28, and family-backed predictions from 31 to 493. ₈ Pfam-domain evidence 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. ₈ Adding Pfam domains and lowering the enrichment-overlap threshold from 3 to 2 increased module annotation from 8% to 80%, expanded annotated modules from 92 to 890, and unlocked 7.6x more function predictions; Pfam provided the broadest coverage, whereas KEGG KOs were too gene-specific for module-level enrichment. ₅

The analysis generated 6,691 function predictions for hypothetical proteins: 2,455 were family-backed (37%) and 4,236 were module-only predictions, using enrichment from KEGG, SEED, TIGRFam, and Pfam. ₅ Ortholog transfer achieved 95.8% strict precision, 91.2% coverage, and 0.934 F1 in held-out benchmarking; 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. ₅ Thus, the results refine ICA’s role toward process-level module discovery rather than precise gene-level assignment. _8,5 Near-zero strict KEGG KO precision is expected because KEGG KO groups are gene-level assignments averaging approximately 1.2 genes per unique KO, so a module with 20 annotated members typically contained 20 different KOs. ₅

The essential-genome analysis supports and qualifies this role: module transfer produced 1,382 function predictions for hypothetical essential genes, equal to 35.3% of predictable targets, across all 48 organisms, including TIGR00254 signal transduction, PF00460 flagellar basal body rod, and PF00356 lactoylglutathione lyase among top predicted functions. ₄ The functional-dark-matter accounting reports 6,691 module predictions while 6,142 dark genes belonged to modules; these prediction and module-associated-gene counts should not be treated as identical. ₆

Stress, conservation, and limitations

The field-versus-lab analysis refines the conservation interpretation in a single-organism test: across 52 ICA modules in Desulfovibrio vulgaris Hildenborough, mean core fraction was 0.886 and median 1.000, and conservation did not significantly correlate with field-condition activity (Spearman rho=0.071, p=0.62). ₁₁ Using 0.886 as the threshold, ecological modules averaged 0.980 core, conserved-quiet modules 0.983, field-variable modules 0.829, and lab modules 0.516; the 21 ecological modules contained 239 genes, including 52 unannotated candidates for environmental-adaptation functions. ₁₁

The counter-ion analysis refines the stress interpretation: across 19 organisms and 14 metals, 4,304 of 10,821 metal-important gene records (39.8%) were also NaCl-important, and it proposes ICA to separate shared NaCl–metal signals from metal-specific effects without reporting a completed test. ₃ Zinc had a DvH whole-genome correlation of r=0.715 with NaCl despite zinc sulfate delivering 0 mM chloride, whereas iron had r=0.086; these are fitness-profile extrapolations rather than direct biochemical tests. ₃

The metal-fitness atlas supports core-enriched module architecture at broader scale: among 183 metal-responsive modules with conservation data, mean core fraction was 0.826 and median 0.929. ₁₂ It identified 600 responsive module records with |z| > 2.0 among 19,453 module × metal-experiment records (3.1%), including 47 DvH modules across 12 metals and 11 psRCH2 modules across six metals. ₁₂ The metal-specificity analysis contradicts a strong expectation of metal-specific ICA structure but not the broader module model: it identified 0 metal-specific modules. ₁₃ Per-module z-normalization yielded maximum absolute z values <2.0 for most metal experiments because those experiments were a small fraction of each organism’s experiments; raw scores also differed from the precomputed z-scored profiles used in the Metal Atlas. ₁₃ The negative result therefore refines rather than disproves module organization and motivates use of precomputed z-scored activities. ₁₃

The conservation analysis found 0 essential genes in any module. ₁₀ Because ICA requires measurable transposon-insertion fitness variation and essential genes lack usable insertions, this describes the non-essential portion of the genome rather than demonstrating that essential genes lack modules. ₁₀ The essential-genome result likewise shows that transfer from non-essential orthologs is indirect and may fail when an essential gene’s function has diverged. ₄

Module size does not establish direct transcriptional co-regulation. _1,9 ICA modules indicate shared fitness phenotypes, and AMR-module size may reflect shared dispensability. _1,9 The fitness-modules predictions should therefore be treated as biological-process hypotheses: Pfam provides broad domain-level coverage and may overcount functional associations, while Module-ICA had <1% strict KEGG KO precision. ₅ The truly-dark-gene results reinforce this caution because module membership and adjacent-gene cofitness supplied candidate functions for genes lacking direct annotation rather than experimental confirmation. ₇

The 40% component cap, imposed because components could not exceed 40% of the number of experiments to avoid FastICA convergence failures, may cause modules to be missed in organisms with few experiments. ₅ Organisms with fewer than approximately 100 experiments produced weaker modules; Caulo, with 198 experiments, showed only 2.9x correlation enrichment. ₅ The strict threshold and annotation results depend on membership and enrichment-overlap choices, and BBH-based families represent aligned conservation patterns rather than proof of identical molecular functions. ₅ The module-conservation analysis covered a 29/32 organism subset because Cola, Kang, and SB2B lacked sufficient GTDB genomes for pangenome construction; its >90% and <50% core cutoffs are convenient rather than biologically motivated boundaries. ₁₀

The field-versus-lab analysis found cross-validated AUC 0.517 for field fitness alone, 0.531 for lab fitness alone, 0.548 for combined field plus lab fitness, and 0.645 after adding gene length; gene length is confounded with fitness-measurement quality and core status. ₁₁ The functional-dark-matter collection contained 37/48 Pseudomonadota, and its top 500 dark-gene candidates contained 417 Gammaproteobacteria, 71 Alphaproteobacteria, 17 Deltaproteobacteria, 12 Bacteroidetes, 10 Betaproteobacteria, 5 Firmicutes, 2 Cyanobacteria, and 0 Archaea, Actinobacteria, or Epsilonproteobacteria, limiting cross-phylum generalization. ₆ Korea had no significant co-inheritance modules because all Korea modules were >90% core with prevalence near 1.0. ₉

Open directions and related pages

Fitness-matched permutation sampling of random non-AMR genes in the proposed −0.05 to +0.05 mean-fitness range could test whether AMR-module enrichment reflects condition-specific organization or shared dispensability. ₁ Carbon-source fitness data could test whether ICA adds functional structure beyond individual-gene associations. ₂ Precomputed z-scored module activities could determine whether the zero metal-specific-module result reflects scale and sparsity rather than biology. ₁₃ Metal-responsive modules should be tested with concentration-normalized fitness, regulatory or expression data, and direct perturbations; CRISPRi knockdown under module-informed conditions could validate the 1,382 transferred essential-gene functions. _12,4 Validation should use held-out annotations and perturbations rather than treating module membership as a KO assignment, and module-associated dark genes should be tested under stress, carbon-source, and nitrogen-source conditions. _5,6 The truly-dark-gene study further proposes testing its top candidates with structure prediction, module-informed growth assays, mobile-CRISPRi, and targeted analysis of dark genomic islands. ₇

• Fitness Modules × Pangenome Conservation — source summary of ICA-module conservation, core/accessory architecture, family breadth, and essential-gene coverage. ₁₀
• Metal-Specific vs General Stress Genes — source summary of metal-specific versus general-stress gene classification and its ICA module test. ₁₃
• Pan-Bacterial Metal Fitness Atlas — source summary of the pan-bacterial metal fitness atlas and its ICA module analysis. ₁₂
• Functional Dark Matter — Experimentally Prioritized Novel Genetic Systems — source summary of dark-gene scale, ICA-associated candidates, conservation, and experimental prioritization. ₆
• Pan-bacterial Fitness Modules via Independent Component Analysis — source summary of ICA thresholds, cross-organism architecture, annotation coverage, and held-out function-prediction benchmarking. ₅
• AMR Co-Fitness Support Networks — source summary describing the ICA module analysis and its interpretation. ₁
• Annotation-Gap Discovery via Phenotype-Fitness-Pangenome-Gapfilling Integration — source summary proposing ICA analysis of fitness data to identify functional modules missed by per-gene analysis. ₂
• Co-fitness and Co-inheritance in Bacterial Pangenomes — source summary testing ICA-module co-inheritance in bacterial pangenomes. ₉
• Gene Conservation, Fitness, and the Architecture of Bacterial Genomes — synthesis connecting ICA module architecture with gene conservation across organisms. ₁₄
• Counter Ion Effects on Metal Fitness Measurements — analysis proposing ICA to separate shared-stress from metal-specific fitness architecture. ₃
• discoveries — cross-project synthesis of ICA module stability, cross-organism architecture, annotation, and AMR-network validation. ₈
• The Pan-Bacterial Essential Genome — pan-bacterial essentiality analysis using ICA modules for indirect function prediction. ₄
• Field vs Lab Gene Importance in *Desulfovibrio vulgaris* Hildenborough — DvH analysis testing whether field-versus-lab fitness modules predict pangenome conservation. ₁₁
• Truly Dark Genes — What Remains Unknown After Modern Annotation? — analysis using ICA modules and cofitness neighborhoods to prioritize genes that remain hypothetical after modern annotation. ₇
• condition-specific-fitness — cofitness and ICA modules may reflect condition-specific shared dispensability rather than direct co-regulation. ₁
• gene-essentiality — fitness phenotypes and dispensability are central to interpreting AMR-containing modules. ₁
• cofitness-network-architecture — ICA modules show stronger co-inheritance structure than pairwise cofitness and are enriched for core genes. _9,14
• metal-cross-resistance — metal-responsive ICA modules connect shared fitness architecture with cross-metal stress responses. ₁₂
• multi-omics-integration — module-level fitness activity motivates integration with regulatory and expression measurements. ₁₂
• kescience-fitnessbrowser — Fitness Browser matrices supplied the fitness data analyzed with ICA. ₁