Report: Co-fitness and Co-inheritance in Bacterial Pangenomes

Key Findings

Pairwise Co-fitness Weakly Predicts Co-occurrence

Across 9 organisms with co-fitness data (2.25M cofit pairs vs 22.5M prevalence-matched random pairs), co-fit gene pairs show a weak but consistent positive co-occurrence signal. The mean delta phi (cofit - random) is +0.011 across organisms, with 7 of 9 organisms showing a positive effect and 8 of 9 significant at p<0.05 (Mann-Whitney two-sided). However, the overall effect size is small (aggregate delta = +0.003, Mann-Whitney p=1.66e-29) and the Wilcoxon signed-rank test across organisms is not significant (W=9, p=0.13), reflecting high inter-organism variance.

The signal is strongest in Ddia6719 (delta=+0.093), which despite being near-clonal (ANI 99.47%) has sufficient accessory gene variation to detect co-inheritance. Korea shows a negative delta (-0.042) because 95.2% of its cofit pairs have NaN phi (both genes at 100% prevalence across 72 genomes, producing zero-variance vectors). Only ~8,000 of 166,601 pairs are computable, and the negative delta reflects statistical noise in this small effective sample rather than a biological signal.

(Notebook: 02_cooccurrence.ipynb, 04_cross_organism.ipynb)

Operons Are Not a Confound

Only 0.7% of cofit pairs are genomically adjacent (within 5 genes). Excluding them does not change the result pattern.

(Notebook: 02_cooccurrence.ipynb)

ICA Modules Show Co-inheritance, Especially Accessory Modules

At the module level, 195 ICA modules across 6 organisms show within-module co-occurrence (mean phi=0.229) significantly exceeding prevalence-matched null expectations (mean phi=0.177, 1000 permutations), with delta=+0.053 and 51/195 modules (26%) significant at p<0.05. After Benjamini-Hochberg FDR correction, 21/195 modules (11%) remain significant at q<0.05.

Accessory modules show the strongest co-inheritance signal:

The accessory vs core difference trends toward significance (Mann-Whitney p=0.051).

(Notebook: 03_module_coinheritance.ipynb)

Co-fitness Strength Weakly Anti-correlates with Co-occurrence

Stronger co-fitness scores weakly predict lower phi (Spearman rho=-0.109, p<1e-300 across 1.04M pairs). This likely reflects that the strongest co-fitness pairs are core genes with near-universal prevalence, leaving little variance for co-occurrence detection — a prevalence ceiling effect.

(Notebook: 02_cooccurrence.ipynb)

Phylogenetic Distance Stratification

Cofit pair phi is higher among near genomes (mean=0.102) than medium genomes (mean=0.067), as expected from shared ancestry. Most species lack genomes in the “far” stratum (>0.05 branch distance), limiting the ability to fully disentangle functional coupling from phylogenetic signal.

(Notebook: 02_cooccurrence.ipynb)

Results

Data Extraction

11 target species analyzed. Presence matrices extracted via Spark from gene_genecluster_junction joined with gene (two billion-row tables). BROADCAST hints on small filter tables reduced query time to ~210s per organism.

Ralstonia UW163 and GMI1000 have zero co-fitness data in the Fitness Browser despite being in the organism table, excluding them from the primary analysis.

(Notebook: 01_data_extraction.ipynb)

Pairwise Analysis

For each organism, cofit pairs were mapped to pangenome cluster pairs via fb_pangenome_link.tsv, deduplicated, and phi coefficients computed from binary genome x cluster presence vectors. 10 prevalence-matched random pairs were generated per cofit pair, matching each cluster’s prevalence independently (tolerance +/-5%). Adjacency was determined from FB gene coordinates (within 5 loci on the same scaffold).

Aggregate statistics:

• 9 organisms analyzed
• 2,253,491 cofit pairs, 22,534,910 random pairs
• Overall mean phi: cofit=0.092, random=0.089
• Overall delta: +0.003 (Mann-Whitney two-sided p=1.66e-29)
• Organisms with positive delta: 7/9
• Organisms with p<0.05: 8/9
• Wilcoxon signed-rank (delta > 0): W=9, p=0.13

Module Analysis

For 6 organisms with ICA module data (Koxy, Btheta, Putida, Korea, Phaeo, pseudo3_N2E3), mean pairwise phi was computed within each module and compared to 1000 prevalence-matched random gene sets. P-values were corrected for multiple testing using Benjamini-Hochberg FDR.

Per-organism module results:

Btheta shows the strongest module-level signal (61% significant), while Korea (despite being the best pairwise organism) shows 0% significant modules — all Korea modules are >90% core with prevalence near 1.0.

Functional Enrichment

Among high-phi AND high-cofit pairs (top quartile of both), the most common functional categories (SEED annotations) are:

(Notebook: 04_cross_organism.ipynb)

Interpretation

Pairwise co-fitness weakly but consistently predicts co-inheritance: 7 of 9 organisms show a positive delta phi, and the aggregate signal is highly significant (p=1.66e-29). However, the effect size is small (delta=+0.003 overall, +0.011 mean across organisms), indicating that lab-measured functional coupling explains only a tiny fraction of co-inheritance patterns.

Multi-gene modules (ICA-derived co-regulated groups) show a substantially stronger co-inheritance signal (delta=+0.053, 26% significant at p<0.05, 11% after FDR correction), particularly accessory modules (73% significant, 36% after FDR). This suggests that coordinated regulation across multiple genes — not just pairwise functional similarity — is what most strongly constrains co-inheritance. The 48 accessory modules identified in the module_conservation project represent functionally coherent units that travel together through the pangenome.

The pairwise signal is attenuated by a prevalence ceiling: most FB genes map to core clusters (>95% prevalence), where phi approaches 0 for both cofit and random pairs because there is almost no variance to correlate. Species with more gene content diversity (Ddia6719, pseudo3_N2E3) show the largest positive deltas.

Literature Context

• Hall et al. (2021) found that E. coli accessory genes co-occur by function and via mobile genetic elements, consistent with our module-level result showing functional coherence predicts co-inheritance (PMID: 34499026).
• Whelan et al. (2020) developed Coinfinder to detect gene associations in pangenomes while controlling for phylogeny. Our analysis extends this by asking whether fitness-measured functional coupling predicts the associations Coinfinder would detect (PMID: 32100706).
• Choudhury et al. (2025) showed P. aeruginosa phylogroups have distinct co-occurrence networks, reinforcing that accessory genome structure is lineage-specific (PMID: 40304385).
• Price et al. (2018) established the Fitness Browser co-fitness data used here. Co-fitness identifies genes in the same pathway or complex; our results show this lab-measured coupling weakly but significantly predicts co-inheritance in natural populations, though the effect size is small (PMID: 29769716).

Limitations

• Prevalence ceiling: Most FB genes map to core clusters (>95% prevalence). At high prevalence, phi approaches 0 for both cofit and random pairs because there is almost no variance to correlate. The analysis is most informative for species with substantial auxiliary gene content.
• Ralstonia excluded: The two Ralstonia organisms (most phylogenetically diverse, lowest ANI) have zero co-fitness data in the Fitness Browser, excluding the species that would likely be most informative.
• Phylogenetic control limited to two strata: Phylogenetic distance stratification was computed for 7 of 9 organisms, but most species lack genomes in the “far” stratum (>0.05 branch distance). Cofit pair phi is higher among near genomes (mean=0.102) than medium genomes (mean=0.067), as expected from shared ancestry, but the lack of a far stratum limits the ability to fully disentangle functional coupling from phylogenetic signal.
• Pairwise vs multi-gene: Co-fitness captures pairwise gene relationships. ICA modules capture multi-gene coordinated regulation, which may better reflect the selective units that constrain co-inheritance.
• Near-clonal species behave differently: Ddia6719 (ANI 99.47%) shows the strongest pairwise signal (delta=+0.093), while pseudo3_N2E3 (ANI 99.66%) also shows positive delta (+0.026). Despite low overall genome diversity, these species have sufficient accessory gene variation to detect co-inheritance. However, the high baseline phi in these species means the absolute phi values are less interpretable.

Future Directions

1. Restrict to auxiliary-only pairs: Re-analyze using only gene pairs where both clusters are auxiliary (<95% prevalence) to maximize presence/absence variance.
2. Compute co-fitness directly: For Ralstonia and other organisms lacking pre-computed co-fitness, calculate pairwise Pearson correlations from raw genefitness data.
3. Fix phylogenetic control: Resolve the reference genome mapping to enable phylogenetic distance stratification.
4. Module co-transfer networks: Build networks of which modules co-occur across species and test whether co-fitness predicts cross-module co-inheritance.
5. Expand to more diverse species: Target additional organisms with >30% auxiliary genes and existing co-fitness data.

Data

Sources

Generated Data

References

• Price MN et al. (2018). “Mutant phenotypes for thousands of bacterial genes of unknown function.” Nature 557:503-509. PMID: 29769716
• Parks DH et al. (2022). “GTDB: an ongoing census of bacterial and archaeal diversity through a phylogenetically consistent, rank normalized and complete genome-based taxonomy.” Nucleic Acids Res 50:D199-D207. PMID: 34520557
• Hall JPJ et al. (2021). “Gene-gene relationships in an Escherichia coli accessory genome are linked to function and mobility.” Microb Genom 7:000650. PMID: 34499026
• Whelan FJ et al. (2020). “Coinfinder: detecting significant associations and dissociations in pangenomes.” Microb Genom 6:e000338. PMID: 32100706
• Choudhury M et al. (2025). “Phylogroup-specific co-occurrence networks in Pseudomonas aeruginosa.” PMID: 40304385

Supporting Evidence

Figures

Notebooks

Revision History

• v1 (2026-02): Initial report
• v2 (2026-02): Added inline figures, notebook provenance, Data section, References section, renamed Visualizations to Figures