Ortholog-transferred fitness phenotypes are constrained by organism-specific network architecture

Ortholog-transferred fitness phenotypes can suggest conserved functional dependencies, but their interpretation is limited when organisms differ in respiratory-chain composition, pathway wiring, or metabolic capacity. ₁ This concept refines condition-specific-fitness by distinguishing a phenotype caused by a shared substrate-driven demand from one caused by architecture-specific compensation or constraint. ₁

Evidence from Complex I fitness transfer

The kescience-fitnessbrowser ortholog-transferred dataset contains 12,241 entries covering 2,005 genes and 13 conditions. ₁ In that dataset, Complex I orthologs have significantly worse fitness on aromatic conditions than on comparison conditions, with mean fitness values of -1.35 versus -0.77 and Mann-Whitney p < 0.0001. ₁

This result supports the hypothesis that ortholog-level fitness transfer can recover a substrate-associated respiratory demand, but it does not establish that aromatic catabolism itself causes the dependency in every recipient organism. ₁ The largest Complex I defects relative to background occurred on acetate (-1.55) and succinate (-1.39), which are non-aromatic substrates that also generate high NADH flux through the TCA cycle. ₁ Thus, the transferred phenotype refines the interpretation from an aromatic-specific effect toward a dependence associated with NADH-generating substrates. ₁ The cross-project digest records the same ordering — Complex I defects largest on acetate and succinate rather than aromatics — and likewise treats the NDH-2 compensation explanation as tentative, describing the “aromatic support network” as really a high-NADH-flux bottleneck network. ₂

The metabolic capability–dependency analysis supports this caution: among 1,695 complete pathway-organism pairs from 48 organisms, 267 (15.8%) were classified as latent capabilities, meaning complete pathways with fitness-neutral measurements under the study thresholds, while 881 (51.9%) were active and 547 (32.3%) intermediate. ₃ Pathway category strongly predicted dependency class (χ²=163.6, df=4, p=2.5×10⁻³⁴), and carbon-source pathways were more often latent than amino-acid biosynthesis pathways (24.3% versus 6.5%). ₃ This supports interpreting transferred phenotypes as condition-dependent evidence of pressure rather than permanent pathway essentiality. The related pathway-capability report’s finding that, across 161 organism-pathway pairs, complete pathway capability and observed fitness dependency were distinct classifications, with 66 complete but non-important “Latent Capability” pairs, remains relevant; all 66 became fitness-important under at least one tested condition type, although the median-based reclassification threshold is partly circular. ₄

Why architecture limits transferability

Bacterial respiratory chains can contain both proton-pumping Complex I, also called NDH-1, and non-pumping NDH-2, an alternative NADH dehydrogenase. ₁ In the proposed model, NDH-2 can compensate under lower NADH flux but may not match Complex I capacity under high NADH flux. ₁ A fitness phenotype transferred between orthologs therefore depends not only on the ortholog’s molecular role but also on whether the recipient organism has an alternative route with sufficient capacity. ₁

The new ADP1 respiratory-chain analysis supports the architecture-aware interpretation: it found qualitatively different carbon-source wiring, with Complex I required on quinate but dispensable on glucose, and attributed the contrast to substrate-dependent NADH flux and capacity constraints rather than total reducing-equivalent yield. ₅ Proteomics further found similar standard-condition protein levels for Complex I, NDH-2, and ACIAD3522—27.6, 27.0, and 26.2, respectively—supporting simultaneous availability rather than a transcriptional switch. ₅ However, the interpretation remains partly model-based: the proposed quinate NADH burst and capacity limit use theoretical pathway stoichiometry rather than measured flux distributions. ₅

The transferred evidence is not definitive for Acinetobacter baylyi ADP1 because the dataset mixes organisms with different respiratory-chain architectures, and direct Complex I fitness measurements on aromatic substrates in a single organism would be more definitive than the mixed-organism comparison the report could run. ₁ Complex I was reported as dispensable on glucose and lactate, where NDH-2 is held to suffice, but essential on substrates that generate high TCA cycle flux (aromatics, acetate, succinate); this interpretation remains a hypothesis rather than a demonstrated universal rule, and its NDH-2 explanation is inferential. ₁ The ADP1 analysis refines that hypothesis: NDH-2 had no growth data, and its FBA-predicted flux was zero on all standard carbon sources because growth optimization routed NADH through Complex I. ₅

The metabolic capability–dependency analysis supports the need for caution when generalizing across organisms: only 7 of 48 Fitness Browser organisms had matching GapMind genome data, and only 48 organisms were fitness-tested among 293,000 genomes with pathway predictions. ₃ Its clade-level result also linked latent-capability rate to pangenome openness (Spearman ρ = 0.69, p = 0.0004, n = 22 clades), suggesting that genome dynamics and community context can shape whether encoded capabilities become dependencies. ₃ This refines the architecture argument by placing respiratory alternatives within a broader organism-specific network and pangenome context, rather than treating transferred effects as properties of orthologs alone. The earlier comparison remains: matched Tier 1 organisms had near-complete, well-annotated core genomes, with mean core-gene completeness of 0.986 for Active Dependencies and 0.975 for Latent Capabilities. ₄

Relation to direct ADP1 evidence

In Acinetobacter baylyi ADP1, the aromatic-catabolism study identified a 51-gene support network surrounding the beta-ketoadipate-pathway, including 21 Complex I genes. ₁ FBA, or flux-balance analysis, captured 1.76× higher Complex I flux on aromatic substrates than on the comparison condition, with fluxes of 0.55 versus 0.31, but predicted 0% essentiality for Complex I. ₁ Independently, 10/13 Complex I operon subunits produced quinate-specific growth defects. ₁

The direct ADP1 measurements therefore show that a model or transferred phenotype can understate a complex-level dependency when alternative flux routes are available in the representation. ₁ They also provide a species-specific anchor for evaluating whether the cross-species Complex I signal reflects aromatic catabolism, high NADH flux, or differences in respiratory-chain architecture. ₁ The respiratory-chain analysis supports this reading by identifying FBA’s growth optimization and ATP-favorable Complex I routing as reasons it misses alternative, capacity-limited respiratory requirements. ₅

The pathway-capability results refine this comparison by showing that an apparently complete pathway can remain fitness-silent under standard conditions yet become important under stress, nitrogen limitation, or carbon limitation. ₄ Conversely, an Incomplete but Important classification can reflect annotation gaps or salvage routes rather than absence of function. ₄ These categories reinforce that pathway completeness, transferred fitness, and direct organism-level dependency are related but non-equivalent evidence layers. The metabolic analysis further found that the latent fraction varied from 0–31.6% across organisms and from 4.7% to 21.1% across tested threshold combinations, supporting the conclusion that the inferred capability–dependency boundary is condition- and threshold-sensitive. ₃

Ortholog transfer as a function-prediction method

Ortholog transfer is not only a caution in this corpus; it is also used as a method. Because essential genes are invisible to fitness-based module detection while their non-essential orthologs in other organisms do have fitness data and can participate in modules, transferring a module’s functional context across organisms bridges the gap between essentiality, which prevents fitness measurement, and function prediction, which requires fitness data. ₆ On that basis, module transfer via non-essential orthologs in ICA (independent component analysis, a blind source-separation method applied here to fitness profiles) modules generated 1,382 function predictions for hypothetical essential genes, covering 35.3% of predictable targets, all family-backed, with top predicted functions TIGR00254 (signal transduction), PF00460 (flagellar basal body rod), and PF00356 (lactoylglutathione lyase). ₆

The method’s reach is bounded by orthology itself: of the hypothetical essential genes, 3,912 have orthologs and are predictable by this method, while 4,385 are orphans that the method cannot predict. ₆ The predictions are also indirect, because they come from non-essential orthologs in other organisms and the essential gene’s function may have diverged. ₆ This refines the page’s argument rather than contradicting it: the same recipient-specific divergence that weakens transferred fitness magnitudes also weakens transferred functional labels, even where transfer is the only available evidence. ₆

A held-out benchmark supports transfer at the gene level while bounding what module structure adds. Across 32 organisms, ortholog transfer dominated KO-level (KEGG Orthology, a gene-level function identifier) prediction with 95.8% precision and 91.2% coverage, whereas Module-ICA had under 1% precision at the KO level despite validated process-level signal — 94.2% cofitness enrichment and 22.7× adjacency enrichment. ₂ The digest reads this as expected and informative rather than a failure: modules capture process-level co-regulation, not specific molecular function, so module-based predictions should be framed as “involved in [biological process]” rather than “has function [specific KO]”. ₂

Scope of the ortholog graph sets measured conservation

How much conservation a transfer analysis finds depends on how the ortholog graph was built. Cross-organism alignment used 1.15M BBH (bidirectional best hit, a reciprocal-best-match orthology criterion) pairs across 32 organisms to form 13,402 ortholog groups. ₇ That alignment yielded 156 module families spanning 2+ organisms, of which 28 spanned 5+, 7 spanned 10+, and 1 spanned 21. ₇ Of those families, 145 carried consensus functional labels, reported as 93%. ₇ The largest aligned family spanned 21 organisms, described as a pan-bacterial fitness module. ₇

The cross-project digest supports the same result from the alignment side: BBH ortholog fingerprints revealed 156 module families spanning 2+ organisms, with the largest spanning 21 of 32 organisms across Proteobacteria, Bacteroidetes, Firmicutes, and Archaea, and 28 families spanning 5+ organisms — read there as evidence of deeply conserved fitness regulons. ₂

The scope dependence is quantitative and large. Using BBH pairs from 5 organisms (10K pairs, 1,861 ortholog groups) versus all 32 organisms (1.15M pairs, 13,402 ortholog groups) changed module families from 27 to 156 (6x), families spanning 5+ organisms from 0 to 28, and family-backed predictions from 31 (4%) to 493 (56%). ₂ The initial run using only 5-organism orthologs found just 27 families with no family spanning more than 4 organisms. ₂

The stated lesson is that the ortholog graph is not additive — adding organisms creates new transitive connections — so orthologs must be extracted for all organisms in an analysis, not just a pilot subset. ₂ The pitfalls digest records the same failure as a methodological rule: using 5-organism orthologs for a 32-organism analysis produced 6x fewer module families and missed all families spanning 5+ organisms, so partial extraction severely underestimates cross-organism conservation. ₈ This refines the architecture argument with a measurement-side counterpart: an apparent ceiling on transferability can come from the coverage of the ortholog graph rather than from recipient biology, and the two causes must be separated before architecture is blamed; see also fitness-module-detection-sensitivity. _2,8

Conserved fitness phenotypes among unannotated genes

Transfer is not confined to annotated genes. Among dark-gene ortholog groups present in 3+ kescience-fitnessbrowser organisms, 65 show measurable fitness concordance — orthologs of the same unknown gene producing fitness effects under the same condition classes across different bacterial species — spanning carbon utilization, stress response, and motility. ₉ These 65 ortholog groups carry conserved phenotypes that could not be identified by studying any single organism. ₉ The digest records the same 65 dark gene ortholog groups with conserved fitness effects across 3+ organisms, adds that dark gene concordance levels are statistically indistinguishable from annotated genes (Mann-Whitney p=0.17), and notes that motility-related dark genes show the strongest cross-organism concordance, suggesting conserved but incompletely annotated chemotaxis machinery. ₂

The functional readings of these groups are hypotheses, not established functions: the report interprets the strongest concordance, in carbon source genes spanning 3–8 organisms and motility genes spanning 3 organisms each, as suggesting conserved but unannotated components of carbohydrate metabolism and chemotaxis machinery respectively, and reads the stress-concordant OG05812 Peptidase_M50 domain (site-2 protease family) as hinting at a conserved regulatory protease under stress. ₉ Its hypothesis-level summary calls cross-organism concordance supported for the 65 testable ortholog groups, with motility-related dark genes showing the strongest cross-organism concordance, consistent with conserved but incompletely annotated chemotaxis machinery. ₉

A matched null control bounds what the 65 groups demonstrate. In the dark-versus-annotated comparison of 65 dark against 490 annotated ortholog groups, dark genes achieve cross-organism concordance levels indistinguishable from annotated genes (Mann-Whitney p=0.17, a rank-based two-sample test). ₉ The matched comparison reports Mann-Whitney p=0.17 and KS p=1.0 (Kolmogorov–Smirnov, a distribution-shape test), with dark median=1.0, annotated median=1.0, dark mean=0.976 and annotated mean=0.985. ₉ The NB06 concordance null control shows dark genes achieving concordance levels indistinguishable from annotated genes. ₉ This refines the reading of the 65 groups: condition-class fitness signal transfers across species for dark genes about as well as for annotated genes, so the result supports general transferability of condition-class signal rather than any special conservation of dark genes; see fitness-matched-null-models. ₉

Ortholog coverage also shapes which dark genes are worth testing: 101 ortholog groups in the top 500 span 2+ organisms, which the report treats as amplifying the return on each experiment — the prioritization argument developed in experimental-prioritization-of-functional-dark-matter. ₉ The report further proposes that running stress experiments in MR-1 and P. putida N2C3 alone would provide cross-organism concordance data for OG01997 (stress) and, combined with nitrogen experiments, would test OG02907 and OG01383 across multiple genetic backgrounds; this is an experimental projection rather than a completed test, and it includes OG01383 even though the preceding table assigns that group to pH rather than stress or nitrogen, without explaining how those screens would test its pH association. ₉

Direction transfers better than magnitude

Metal-stress data separate the two things a transferred phenotype can claim. The report states that the direction of cross-resistance (all positive) is universal, while the quantitative magnitude is only moderately conserved across organisms, with leave-one-out consensus prediction r = 0.41 and Mantel mean r = 0.23. ₁₀ That absolute wording is not exact within the same report, which also records that across 317 organism-metal pair observations (28 organisms, 85 unique metal pairs), 98.1% of gene-level fitness correlations are positive (311/317) and 99.1% are statistically significant (p < 0.05); the transferable direction is therefore near-universal rather than strictly universal, and this page keeps the weaker wording. ₁₀ Mantel tests — correlations between two distance matrices — across 351 organism pairs show mean r = 0.23 with 62% positive, described as moderate but consistent conservation of cross-resistance architecture. ₁₀ The report’s metal-label permutation test is non-significant (p = 0.42) because all metal pairs are positive, so shuffling labels does not change the mean; that test therefore cannot be used to claim conserved metal-pair-specific architecture. ₁₀ This supports the page’s distinction between a conserved direction of functional pressure and a non-conserved magnitude of fitness cost, now in a metal-stress rather than respiratory setting; see metal-cross-resistance. ₁₀

The consensus map is a useful but partial predictor of any one organism: leave-one-out consensus prediction had mean r = 0.41, yet only 2/28 organisms reached individual significance, indicating that the consensus captures the average trend while individual organisms have meaningful deviations. ₁₀

Organism context can outweigh mechanism identity

In antimicrobial-resistance cofitness networks, host organism explains more of a gene’s functional neighbourhood than the resistance mechanism does. Mean Gene Ontology (GO, a controlled vocabulary of biological functions) term Jaccard similarity — the overlap of two term sets divided by their union — across resistance mechanisms within the same organism was 0.375, i.e. different AMR genes in the same bug share 38% of cofitness GO terms. ₁₁ For the same mechanism compared across organisms, mean GO-term Jaccard similarity was 0.207, so the same mechanism in different bugs shares only 21%. ₁₁ The within-organism versus across-organism comparison was highly significant, with MWU (Mann-Whitney U) p = 4.3×10⁻¹³. ₁₁ This supports the page’s central claim with an independent data type: a resistance gene’s cofitness context tracks its host network more closely than its mechanism class, so mechanism-level transfer across organisms carries less information than within-organism context; see cofitness-network-architecture. ₁₁

Transfer across wider phylogenetic and mechanistic distance

Some transfer claims in this corpus are explicitly untested predictions. A “structural unavailability” model predicts that introducing the caulobacter-crescentus sphingolipid biosynthesis pathway into acinetobacter baumannii, which lacks it, should be neutral or modestly protective under colistin selection, because the proposed PBP1A/Ldt route in A. baumannii works independently of sphingolipids; a sphingolipid-engineered A. baumannii would then have two alternative routes to colistin resistance, which predicts increased rates of colistin-resistant escapers. ₁₂ The prediction has not been performed, and it states the page’s principle in reverse: whether a transferred pathway produces a phenotype depends on which alternative routes the recipient already carries. ₁₂

Phylogenetic distance also limits what a single-organism phenotype says about a system elsewhere. The methanococcus-maripaludis snipe-defense-system gene shows mild fitness defects under specific conditions, with minimum fitness = -1.16 on formate/acetate, but is dispensable under most conditions; it is the first organism with both a confirmed SNIPE protein and genome-wide fitness data, though as an archaeon its phage biology context differs significantly from the Enterobacterales phage lambda system. ₁₃ Reading the lambda-based expectation onto this archaeal measurement, or the reverse, is therefore weak evidence about the system’s role in either lineage. ₁₃

Tensions

The aromatic-catabolism analysis proposes that NDH-2 compensation can explain lower Complex I dependence on some substrates, while the cross-species respiratory-chain comparison found no supporting compensation pattern: among organisms retained after filtering, validated NDH-2 organisms had a larger mean Complex I aromatic deficit (-0.297 versus -0.156), with p = 0.52. ₅ The comparison included only 4 organisms lacking NDH-2 and may therefore be underpowered, insufficient for statistical significance; the two results should not be reconciled as a universal rule. ₅ The former remains a substrate- and architecture-based hypothesis ₁, whereas the latter weakens its generalization across species. ₅

The digest records how that reversal arose and repeats the corrected numbers. After correcting NDH-2 false positives by filtering text-matched candidates to organisms with ≤2 hits per genome, the cross-species compensation test reverses: organisms with validated NDH-2 show larger Complex I aromatic deficits (mean -0.297) than those without (-0.156, p=0.52). ₂ The initial analysis — 10/14 organisms with NDH-2, apparent support for compensation — was driven by misannotated Complex I subunits leaking through the text filter in 5 organisms, which the digest presents as a case where a finding that looked supportive was an annotation artifact. ₂ On the project’s own corrected counts, after filtering to organisms with 1-2 NDH-2 hits and excluding those with >2 hits that likely represent misannotated Complex I subunits, 5 of 14 organisms have validated NDH-2; those organisms show a mean Complex I aromatic deficit of -0.297 versus -0.156 for organisms without NDH-2, a difference that is not significant (p = 0.52), so NDH-2 presence does not predict whether Complex I is dispensable on aromatics and the ADP1 respiratory wiring pattern may be species-specific rather than a general rule. ₅ The same report’s limitation list instead states that the cross-species comparison has only 4 organisms without NDH-2; the 5-of-14 and 4-without counts are not reconciled in the source, and both are preserved here as an unresolved discrepancy rather than combined into a single cohort description. ₅ This weakens the compensation hypothesis as a cross-species rule and shows that a transferred-fitness comparison can be dominated by gene-identification error rather than by recipient biology. ₂

The metabolic capability analysis introduces a related limitation rather than resolving this tension: pathway-level conservation did not distinguish latent capabilities from active dependencies, with mean conservation of 0.869 for 248 latent capabilities versus 0.829 for 755 active dependencies and Mann–Whitney U p = 0.94 for active > latent. ₃ Thus, conserved pathway presence alone does not establish conserved fitness dependence, supporting the caution that ortholog transfer cannot substitute for recipient-specific network and condition measurements. ₃

A smaller tension sits inside the dark-gene evidence itself: the report’s interpretation section calls cross-organism concordance supported for the 65 testable ortholog groups, with motility-related dark genes showing the strongest concordance, while its Finding 4 instead describes carbon-source genes as having the strongest concordance, so the leading condition class is inconsistent within the report. ₉ The two statements should not be averaged; which condition class carries the strongest transferable dark-gene signal is unsettled in this corpus. ₉

Interpretation and limitation

Ortholog-transferred fitness is strongest as evidence for a conserved direction of functional pressure when the compared organisms share relevant pathway and respiratory architectures. ₁ It is weaker as evidence for a conserved magnitude of fitness cost or for a universal condition-specific essentiality rule when alternative dehydrogenases, pathway bypasses, or complex capacities differ between organisms. ₁ The 12,241-entry dataset provides a comparative signal, whereas direct Complex I fitness measurements on aromatic substrates in a single organism would provide a stronger test of the causal role of aromatic catabolism. ₁

This limitation connects to cross-species-fitness-transferability and respiratory-capacity-and-nadh-load: transferred phenotypes should be interpreted against recipient-specific network architecture rather than treated as architecture-independent annotations. ₁ The metabolic capability study supports this interpretation while showing that environment-linked metabolic clusters are not universal: all 10 target species formed clusters, but significant environment–cluster associations were reported only for Salmonella enterica and Phenylobacterium sp.; the ecological interpretation was observational and lacked explicit phylogenetic correction. ₃ The full ADP1 evidence and its limitations are summarized in Condition-Specific Respiratory Chain Wiring in ADP1, while the capability–dependency evidence is summarized in Metabolic Capability vs Metabolic Dependency.

Open Directions

• Measure Complex I and NDH-2 deletion fitness directly in ADP1 on quinate, glucose, acetate, and succinate to test whether the phenotype tracks aromatic catabolism or NADH-generating substrate load. ₁
• Search the ADP1 genome for NDH-2 and combine deletion experiments with respiratory-capacity measurements to determine whether alternative NADH dehydrogenase activity explains the glucose and lactate results. ₁
• Stratify the kescience-fitnessbrowser ortholog-transferred data by organism and respiratory-chain architecture, then compare effect sizes across architectures to quantify transferability. ₁
• Expand comparative pangenome analysis across aromatic-degrading species to test whether Complex I retention and NDH-2 presence predict transferred fitness phenotypes. ₁
• Rebuild the ADP1 FBA model with PQQ biosynthesis, iron homeostasis, and respiratory-chain capacity constraints, then test whether architecture-aware constraints reduce the gap between predicted 0% essentiality and the observed defects in 10/13 Complex I subunits. ₁
• Reanalyze transferred fitness alongside pathway completeness and condition-specific fitness across organisms with matched GapMind and Fitness Browser data to test whether architecture-aware transfer predicts Active Dependency, Latent Capability, or Incomplete but Important classifications. ₄
• Use KO-based K03885 and K00330–K00343 searches across the planned 27K-species comparison, then test whether NDH-2 presence predicts Complex I fitness only within matched respiratory architectures. ₅
• Re-report the NDH-2 cohort explicitly so that the 5-of-14 validated-NDH-2 count and the stated 4 organisms without NDH-2 refer to the same filtered organism set before any compensation test is repeated. ₅
• Measure NADH/NAD⁺ ratios and respiratory-chain protein abundances across ADP1 carbon sources to distinguish the proposed flux-capacity mechanism from transcriptional regulation. ₅
• Use direct GapMind per-step gene assignments and matched fitness assays across multiple organisms to test whether architecture-aware ortholog transfer predicts latent versus active pathway dependency; this addresses the current SEED-proxy mapping limitation and the 4.7%–21.1% threshold sensitivity range. ₃
• Re-extract BBH orthologs for every organism in any planned cross-organism fitness comparison before estimating conservation, since the 5-organism versus 32-organism contrast moved family-backed predictions from 31 (4%) to 493 (56%); this closes the gap between measured and real transferability. ₂
• Benchmark module-transfer predictions for the 3,912 orthology-covered hypothetical essential genes against direct assays, and design a non-orthology approach for the 4,385 orphans this method cannot reach. ₆
• Run the proposed MR-1 and P. putida N2C3 stress screens, with nitrogen experiments added, to obtain cross-organism concordance data for OG01997 and to test OG02907 and OG01383 across genetic backgrounds; the proposal covers those groups rather than all 65 concordant dark-gene groups, its outcome would be condition-class concordance rather than guaranteed molecular-function assignment, and the unexplained OG01383 pH assignment would need resolving. ₉
• Test whether the per-organism deviations from the metal cross-resistance consensus map — mean r = 0.41 with only 2/28 organisms individually significant — are predicted by transporter and efflux gene content in the same fitness dataset. ₁₀
• Compute cofitness GO-term Jaccard similarity for matched resistance mechanisms in further organism pairs to test whether the 0.375 within-organism versus 0.207 across-organism gap holds beyond the sampled organisms. ₁₁
• Perform the sphingolipid-pathway transfer into A. baumannii under colistin selection to test the predicted neutral-to-modestly-protective effect and the predicted increase in resistant escapers. ₁₂
• Measure SNIPE fitness in an Enterobacterales host carrying genome-wide fitness data to test whether the archaeal minimum fitness of -1.16 on formate/acetate reflects a conserved conditional role. ₁₃