Metabolic Model Gapfilling and Prediction Reliability

Metabolic-model gapfilling adds reactions or functions needed to complete a model and enable simulated growth, but prediction accuracy can depend on inferred rather than directly supported biology. ₁ The Acinetobacter baylyi ADP1 analysis compares model-derived predictions with complete-gene knockouts, mutant growth, proteomics, TnSeq, pangenome classification, and functional annotations. ₁ The triple-essentiality analysis tests agreement among model predictions, complete-gene knockouts, mutant growth, and proteomics. ₂ The annotation-gap study supports this framing by combining Fitness Browser phenotypes, pangenome annotations, GapMind pathway evidence, and BLAST homology across 14 organisms and 18 carbon sources. ₃ The aromatic-catabolism analysis refines it by showing that increased model-predicted flux can coexist with missed complex-level essentiality. ₄ Formulation design supports condition-specific validation through carbon utilization, growth kinetics, patient activity, and GapMind completeness. ₅ Clay-confined subsurface results refine the self-sufficiency component: cultured deep-clay genomes did not show elevated generalized biosynthetic completeness, so pathway completeness can constrain but cannot establish model sufficiency in a particular environment. ₆ The cross-project synthesis supports treating pathway capacity, condition-specific fitness, and annotation provenance as complementary constraints rather than interchangeable validation of flux balance analysis (FBA), a constraint-based method for predicting feasible flux distributions. ₇

Evidence ceilings: annotation, compounds, and pathway capacity

The ENIGMA Carbon Census refines the framework by showing that the evidence ceiling can precede model construction. Of 83 enrichment compounds, 74 (89%) were organism-dark because utilization determinants were not linkable through queried the KBase Data Lakehouse and curated resources, while only 9 were callable under the project definition. ₈ All 83 compounds resolved to structures with InChIKeys via PubChem, but only 54 linked to KEGG and 9 were callable; 33 KEGG-linked compounds lacked reactions in queried genomes and 29 were fully orphaned. ₈ Thus, a missing reaction may reflect absent biology, missing pathway representation, or incomplete resource linkage rather than demonstrated inability to metabolize a compound. ₈ The 8 callable compounds were concentrated in aromatic, alkaloid, and terpenoid classes, and H1 was not formally supported (χ² = 5.07, p = 0.53, df = 6); xanthine was mis-scored because R02107 represents purine nitrogen acquisition rather than carbon catabolism, making the effective carbon-callable set 8 rather than 9 although committed tables contain xanthine. ₈

The essential-metabolome pilot supports this layered interpretation: GapMind completeness can identify conserved biosynthetic capacity and candidate gaps, but does not establish pathway essentiality, physiological dependency, or reliable flux. ₉ Its 859 universally essential gene families came from 45 organisms, whereas the GapMind comparison used 7; the essential-gene experiments used RB-TnSeq (random-barcode transposon sequencing) in rich media, where nutrient supplementation can make biosynthetic genes appear non-essential. ₉ Only 7 of 45 organisms mapped successfully, the sample was mostly Proteobacteria plus one Deltaproteobacterium, and the 18-pathway result was computational; GapMind can miss non-canonical pathways, divergent enzymes below homology thresholds, or genes absent from annotations. ₉

The functional-dark-matter analysis supports the same layered model while adding a prioritization layer: among 228,709 genes across 48 Fitness Browser organisms, 57,011 (24.9%) were unannotated or poorly characterized, but only 17,344 had strong measured fitness effects or essentiality evidence. ₁₀ Of the 57,011 dark genes, 6.0% were strong testable hypotheses, 52.5% weak leads, and 41.5% true knowledge gaps. ₁₀ Condition-specific fitness can constrain candidate reactions, but does not establish flux direction, stoichiometry, or pathway activity. ₁₀

The FW300-N2E3 cross-database analysis supports layered validation: 17/21 (81%) testable Web of Microbes metabolites were fully concordant, mean concordance was 0.94, and none were fully discordant, but 37/58 metabolites lacked cross-database tests and BacDive agreement was only 3/7. ₁₁ In FW300-N2E3, all 13 Web of Microbes metabolites mapped to GapMind had complete pathways and all 13 showed Fitness Browser growth, while the 94% four-database concordance was structurally driven by Fitness Browser 21/21, GapMind 13/13, and BacDive 3/7. ₇ Across 21 matched metabolites, FW300-N2E3 had significant effects for 601 unique genes and 4,764 total significant gene-condition hits using |fit| > 1 and |t| > 4; 231 genes were significant in at least 3 conditions and the top 18 in all 21. ₁₁ Yet BacDive utilization was only 3/7 and tryptophan was 0+/50− despite production, complete biosynthesis, and significant fitness effects. ₁₁

The Web of Microbes (WoM) explorer supports cross-collection validation but refines what metabolite links establish. In a 2018 snapshot, action codes distinguished amplification (I, present and subsequently higher) from de novo emergence (E, initially absent and newly detected); among 10,744 organism observations, 1,338 were I, 1,155 E, and 7,509 N, with no consumption or decrease action. ₁₂ WoM linked two ENIGMA Pseudomonas strains directly to Fitness Browser, and curated matching found 19 FW300-N2E3-produced metabolites also tested as Fitness Browser carbon or nitrogen sources, including five E products and 14 I metabolites. ₁₂ This supports prioritizing metabolite-to-gene analyses, but production does not establish consumption, uptake bounds, dependency, or essentiality. ₁₂ Of 257 identified compounds, 69 (26.8%) had definitive ModelSEED links, 107 (41.6%) formula-only candidate links, and 81 (31.5%) remained unmatched; the formula-only set expanded to 900 ModelSEED molecules and therefore represents candidate sets rather than definitive reaction identities. ₁₂ Exact name matching linked 69 compounds, whereas formula-only matching added 107 low-confidence sets requiring manual curation; pathway matching was blocked because WoM used internal pathway identifiers rather than simple metabolite names. ₁₂

Reaction conservation, gapfilling burden, and pathway dependence

The ADP1 analysis identified 1,330 unique metabolic reactions across 14 genomes: 1,248 (94%) were shared by all 14 and classified as core, 62 were variable (present in 2-13 genomes), and 20 genome-unique. ₁ Gapfilling accounted for 7.7% of reactions on average and 243 missing functions were cataloged. ₁ This conserved backbone supports joint evaluation of pangenome and reaction conservation, but conserved reaction content does not establish experimental validation. ₁ Pangenome status did not explain FBA/knockout discordance: genes were 93–100% core across discordance classes, with no significant enrichment (OR = 0.89, p = 0.80). ₂ Across 132.5M gene clusters, Bakta and eggNOG had complementary functional coverage and their union reached 77.3% any-functional-annotation coverage, but only 33.3% of Bakta’s 17.6M distinct UniRef50 IDs existed in the KBase Data Lakehouse UniProt identifier table. ₇

Formulation design refines conservation as a prioritization criterion. Across 499 genomes, Micrococcus luteus had 18/18 amino-acid and 39/39 carbon pathways conserved at >95%; Streptococcus salivarius had 18/18 and 32/35; Rothia dentocariosa 14/18 and 39/41; Neisseria mucosa 16/16 and 27/27; and Gemella sanguinis 7/18 and 37/39. ₅ These results support prioritizing conserved functions, but predicted use of myoinositol, xylitol, xylose, and arabinose still requires experiment. ₅

The clay-confined comparison supports distinguishing pathway completeness from environmental sufficiency. Across 18 GapMind pathways, unfiltered mean completeness was 16.22/18 for 9 deep-clay genomes versus 16.66/18 for baseline (Mann–Whitney p = 0.153, Cohen’s d = −0.17); after a CheckM filter of ≥80% completeness and ≤5% contamination, means were 15.50/18 versus 17.14/18 (p = 0.009, d = −0.84). ₆ Within Bacillota_B, means were 16.50 versus 16.79 (p = 0.073, d = −0.13), while shallow clay exceeded baseline before filtering (17.87/18 versus 16.66/18, p = 0.006, d = +0.52) and after filtering (17.87/18 versus 17.14/18, p = 0.029, d = +0.43). ₆

Two capability/dependency analyses refine the assumption that completeness implies dependency. Across 161 organism–pathway pairs from 7 model bacteria and 23 GapMind pathways, 57 (35.4%) were Active Dependency, 66 (41.0%) Latent Capability, 24 (14.9%) Incomplete but Important, and 14 (8.7%) Missing. ₁₃ Active Dependency required a complete pathway containing fitness-important genes; Latent Capability required completeness without significant standard-condition defects; Incomplete but Important combined an incomplete call with important mapped genes; Missing meant neither complete nor important. ₁₃ A separate analysis of 1,695 complete pathway-organism pairs from 48 organisms reported 267 (15.8%) latent capabilities, 547 (32.3%) intermediate cases, and 881 (51.9%) active dependencies; pathway category predicted class (χ²=163.6, df=4, p=2.5×10⁻³⁴). ₁₄ These differing latent fractions are not resolved because organism sets, pathways, mappings, thresholds, and procedures differ. ₁₄

The discoveries synthesis reports the 161-pair distribution as 35.4% Active Dependencies, 41.0% Latent Capabilities, 14.9% Incomplete but Important, and 8.7% Missing, and reports that all Latent Capabilities became fitness-important under condition-specific analyses. ₇ The pathway-capability study used a composite score weighting essentiality 40%, fitness breadth 30%, and fitness magnitude 30%; its 7 Tier 1 organisms were Desulfovibrio vulgaris Hildenborough, Shewanella oneidensis MR-1, Pseudomonas putida KT2440, Pseudomonas stutzeri RCH2, Caulobacter crescentus, Sinorhizobium meliloti, and Azospirillum brasilense, and only 7 of 48 Fitness Browser organisms had matching GapMind data. ₁₃

Conservation did not distinguish active from latent pathways: mean conservation was 0.829 for 755 active dependencies, 0.907 for 508 intermediate pathways, and 0.869 for 248 latent capabilities; active-versus-latent testing was non-significant (Mann–Whitney U p = 0.94; rank-biserial r = 0.052). ₁₄ Latent-capability rate correlated with pangenome openness (Spearman ρ = 0.69, p = 0.0004, n = 22 clades). ₁₄ Variable pathway count correlated with openness at raw rho=0.327, p=7.2e-71 and partial rho=0.530, p=2.83e-203 after controlling for genome count; the signal was positive in 13/18 genera and significant in 5/18. ₁₃ Thus pangenome dynamics provide context, not a rule that conserved completeness implies FBA-relevant dependency. _14,13

Core-only versus all-gene GapMind completeness supports an accessory-genome contribution to modelable capacity: leucine and valine were 0.614 all-gene, 0.468 core-only, gap 0.146; arginine 0.613, 0.472, 0.141; lysine 0.804, 0.664, 0.140; and threonine 0.803, 0.663, 0.140. ₁₃ These differences suggest, but do not demonstrate, community sharing under Black Queen dynamics. ₁₃

Gapfilling and growth-prediction reliability

ADP1 evaluated 121,519 growth phenotype predictions across 14 genomes; 105,376 (87%) required at least one gapfilled reaction. ₁ This supports the conclusion that prediction accuracy is tightly coupled to gapfilling quality. False negatives had higher mean gap counts than correct predictions, suggesting—but not proving—that models requiring more inferred reactions are more vulnerable to missing or incorrectly selected functions. ₁

The annotation-gap study refines this conclusion by separating model completion from gene assignment. Across 574 organism–carbon-source combinations, baseline FBA had 42.5% accuracy, 86.5% recall (244 of 282 growth-positive conditions correctly predicted), and 42.5% precision (244 of 574 growth predictions were correct), with 330 false positives; because the reported accuracy and precision carry the same value, they should not be read as independently verified metrics. ₃ The report attributes the low baseline to overly permissive draft models built from automated RAST annotations, which it describes as containing systematic errors and predicting growth on carbon sources the organisms cannot use. ₃ Conditional gapfilling for 38 false negatives added 219 reactions: 201 enzymatic, 14 transport, and 12 exchange, averaging 5.8 reactions per case; the listed category counts do not add up to the stated total of 219, and the source leaves that numerical conflict unresolved. ₃ Gapfilling corrected some false negatives within models that already overpredicted growth, making reaction selection and environmental constraints jointly important. ₃ Candidate genes were assigned to 96/201 enzymatic reaction-organism pairs (47.8%), including 44 high-confidence (21.9%), 19 medium-confidence (9.5%), and 33 low-confidence (16.4%); 105 (52.2%) remained unresolved. BLAST alone resolved 70 pairs (34.8%), while the full pipeline resolved 96 (47.8%), a 13-percentage-point gain. ₃

Pathway-level prediction partially corroborates gapfilling rather than confirming it. Of 104 GapMind–gapfill pathway pairings, GapMind pathway predictions partially corroborated the gapfilling results, and GapMind called pathways incomplete (not_present or steps_missing) for many of the same carbon sources where ModelSEED required gapfilling, with “many” left unquantified in the source. ₃ Multiple valid gapfill solutions may exist for each false-negative case, so the default minimum-reaction ModelSEED solution is not necessarily the biologically correct one; the report proposes gapseq in place of RAST/ModelSEED for initial reconstruction as a hypothetical route to fewer false-positive FBA predictions and a more targeted gapfill set. ₃

Of 201 gapfilled reactions, 50 (24.9%) were EC-less dark reactions; only 8/50 (16%) were resolved versus 88/151 (58.3%) with known EC numbers. ₃ The 243 missing functions are therefore evidence gaps affecting interpretation, not merely formatting omissions. ₁

The capability/dependency analysis supports provenance and condition matching. Fitness Browser genes were mapped through the native KEGG route besthitkegg-to-keggmember-to-EC-to-KEGG-map-to-GapMind pathway. ₁₃ Two GapMind pathways lacked matching SEED role descriptions, alanine was excluded because fewer than 3 SEED-annotated genes met the coverage threshold, and direct GapMind per-step assignments were identified as a needed improvement. ₁₄ An apparent pathway gap can therefore be a database or membership-matching gap rather than biological absence, while a complete pathway can remain latent under uninformative conditions. ₁₄

The respiratory-chain study supports this optimization warning. FBA predicted zero NDH-2 and ACIAD3522 flux on standard media because growth optimization routes NADH through Complex I, which produces more ATP per NADH, although experimental fitness showed condition-specific requirements and ACIAD3522 had a growth ratio of 0.013 on acetate. ₁₅ That project attributes the failure to predict condition-specific respiratory requirements to missing capacity constraints — the model assumes the optimal pathway is always used, missing constraints that would force cells onto alternative, suboptimal pathways under high flux — and notes that its stoichiometry analysis uses theoretical pathway biochemistry rather than measured flux distributions. ₁₅ Complex I growth ratios were 0.37 on quinate, 1.44 on glucose, and 0.49 on acetate; ACIAD3522 ratios were 1.39 on quinate and glucose but 0.013 on acetate. ₁₅ Proteomics found Complex I mean 27.6 (66th percentile), NDH-2 27.0 (59th), and ACIAD3522 26.2 (48th), versus genome median 26.4, supporting a passive flux-wiring hypothesis rather than a transcriptional switch. ₁₅ NDH-2 lacked growth data and ACIAD3522 may have another function. ₁₅

The aromatic network supports the interpretation that unmapped support functions create blind spots: 30/51 quinate-specific genes had no FBA mappings, including cofactor-supply, iron-acquisition, regulatory, and newly identified respiratory functions. ₄ FBA captured 1.76× higher Complex I flux on aromatic substrates (0.55 versus 0.31) but predicted 0% essentiality, whereas 10/13 Complex I operon subunits independently produced quinate-specific growth defects. ₄ Two structural reasons are given: FBA models optimize for growth rate and can redistribute flux through alternative pathways, so a bottleneck need not appear; and Complex I operates as a single multi-subunit complex in which loss of any subunit eliminates function, a threshold effect that FBA’s linear-programming framework does not capture. ₄ The cross-project synthesis restates both points, recording that 30 of 51 quinate-specific genes (59%) have no FBA reaction mappings and that linear programming cannot represent the threshold effect of losing a single subunit from a multi-subunit complex. ₇ Nine of 11 aromatic-degradation genes were FBA-discordant (OR = 9.70, FDR-adjusted q = 0.012), with FBA under-prediction enriched (OR = 12.0, q = 0.004). ₂ Almost all of that aromatic discordance runs in one direction — FBA under-prediction: the model assigns zero flux to these genes, yet their deletion impairs growth. ₂ Beta-ketoadipate genes predicted blocked but associated with growth defects suggest missing aromatic substrates or mismatched environmental assumptions, a hypothesis requiring measured media and revised constraints. ₂ The discordant aromatic functions include named beta-ketoadipate pathway enzymes — 4-carboxymuconolactone decarboxylase and beta-ketoadipate enol-lactone hydrolase — and the FBA model assumes minimal media without aromatic compounds, while experimental media may contain trace aromatics from yeast extract or agar impurities; absent modeled substrates, trace aromatics, and moonlighting functions remain competing explanations that the project does not resolve. ₂ By contrast, lipid metabolism was depleted among discordant genes (OR = 0.34, q = 0.042), with only 7 of 46 lipid-metabolism genes discordant, which the report reads as lipid metabolism being relatively well captured by FBA. ₂

Plant marker analysis supports provenance caution: 12/22 marker Pfams were absent from Bakta Pfam domains despite abundance in InterProScan, while the ten non-zero Bakta Pfams had only 10–35% of InterProScan hit counts; versioned Pfam identifiers recovered 19,364 hits across 7,962 species for ten queried markers. ₁₆ The Pseudomonas analysis likewise refines environment-linked pathway evidence: among 54 species with at least 5 genomes and at least 60% majority-environment agreement, carbon profiles associated with isolation environment (999-permutation p = 0.006; between-group mean distance 2.054 versus within-group 1.890), but Random Forest balanced accuracy was 0.408 +/- 0.169 against a 0.250 chance baseline. ₁₇ The first five principal components explained 74.9% of variance, with the dominant axis separating Pseudomonas s.s. from Pseudomonas_E; lifestyle categories overlapped within Pseudomonas_E. ₁₇ These observations do not justify direct environment-specific reaction additions. ₁₇

Flux predictions and experimental essentiality

FBA and TnSeq calls could be compared for 866 genes: 639 (73.8%) were concordant and 227 discordant. ₁ The discordant 26% — genes where the model and the experimental data disagree — are candidates for model refinement or may reflect regulatory effects not captured by FBA. ₁ FBA flux classes changed between rich and minimal media for 177/866 genes (20%), supporting condition-specific-fitness while remaining limited to the overlap. ₁ Essentiality also differed by medium: 499 genes were essential on minimal media versus 346 on LB. ₁ This supports treating validation as condition-specific rather than transferring validation from one medium to another. ₁

The FBA/TnSeq comparison supports empirical essentiality as a validation layer, while its 227 discordances identify a refinement set — candidates for pathway- or regulatory-function enrichment analysis and model refinement — rather than proving either method universally correct. ₁ FBA covered only 15% of genes, so the 73.8% estimate applies only to the measured subset. ₁ The capability/dependency study refines endpoint interpretation: active pathways required mean absolute fitness t-score >2.0 or >20% essential genes, while latent pathways had mean absolute t-score <1.0 and <5% essential genes. ₁₄ In the larger study, 32.3% of complete pathways were intermediate and latent fractions shifted from 4.7% to 21.1% across thresholds, so pathway-level constraints should retain uncertainty. ₁₄

Direct comparison with complete-gene knockouts refines the concordance picture by medium. In rich media across 724 genes, FBA reached 60.8% recall, 64.0% precision, F1 0.624, and Cohen’s kappa 0.486 (a chance-corrected agreement statistic), described as moderate agreement with knockout essentiality. ₂ In minimal media across 833 genes, FBA reached 65.6% recall, 69.2% precision, F1 0.673, and Cohen’s kappa 0.493, also moderate. ₂ The underlying confusion matrices were TP = 164, FP = 92, FN = 106, TN = 362 with specificity 79.7% in rich media, and TP = 229, FP = 102, FN = 120, TN = 382 with specificity 78.9% in minimal media. ₂

The report reads these numbers as FBA performing better in minimal media (F1 = 0.67) than rich media (F1 = 0.62), likely because metabolic constraints are tighter there and FBA assumptions better match reality. ₂ That explanation is an interpretation rather than something the comparison establishes, and the FBA endpoint is itself a modeled quantity — whether flux through a gene is required for growth under defined-medium assumptions — not a measurement. ₂ On that basis the report recommends FBA only as a first-pass screen at ~60-70% accuracy (κ≈0.49), better in minimal than rich media, with all predictions validated experimentally. ₂

The same project contradicts, within its tested scope, the expectation that FBA class explains residual growth phenotypes. Among 478 triple-covered genes, all TnSeq-dispensable, FBA class was not significantly associated with growth-defect status (chi-squared = 0.93, p = 0.63, 2 df): defect rates were 73.1% (57/78 genes) for FBA-essential, 73.5% (150/204 genes) for FBA-variable, and 69.4% (136/196 genes) for FBA-blocked. ₂ Mean growth rates likewise did not differ by FBA class (Kruskal-Wallis H = 1.67, p = 0.43, a rank-based test of location across groups). ₂ Because the defect rate was uniform across classes, growth measurements did not differentially validate FBA-essential versus FBA-blocked predictions and did not resolve the FBA–TnSeq disagreement in this selected sample. ₂

Within the FBA-blocked set (zero predicted flux, n = 196), 136/196 (69%) showed growth defects — FBA under-predicting — while 60/196 (31%) grew normally, consistent with FBA. ₂ The null result is scoped to those 478 triple-covered, TnSeq-dispensable genes in minimal media across the tested carbon-source conditions and does not extend to dispensable genes generally; the report records 70% of genes showing condition-specific growth defects, weak flux correlations of ρ = -0.26 to +0.25, and no capture by FBA essentiality class (p = 0.63). ₂ Sensitivity analysis across Q10, Q15, Q20, Q25, Q30, Q35, and Q40 defect thresholds reports that the null held from Q10 through Q35, with Q25 the standard choice; only at Q40 did a marginal association appear (p = 0.048), so the null is threshold-dependent at that one setting. ₂

These results refine rather than overturn the original comparison: the original study found that FBA does not predict growth defects among TnSeq-dispensable genes (p = 0.63), while the refined study found moderate concordance with knockout lethality (κ≈0.49). ₂ Both rest on the Durot et al. (2008) ADP1 FBA metabolic model, whose original publication reported ~75% FBA-TnSeq concordance. ₂

Condition-specific FBA flux correlated weakly and inconsistently with growth: glucose ρ = -0.021 (p = 0.677; n = 387), acetate ρ = -0.153 (p = 0.004; n = 352), asparagine ρ = -0.257 (p < 0.001; n = 286), butanediol ρ = -0.145 (p = 0.092; n = 137), glucarate ρ = +0.246 (p = 0.005; n = 127), and lactate ρ = -0.160 (p = 0.065; n = 135). ₂ The positive glucarate association is opposite the expected direction and supports environmental-assumption limits. ₂ Carbon-source mutant fitness also showed mean pairwise correlation 0.44 across eight carbon sources; urea and quinate were nearly uncorrelated (r = 0.11), while butanediol-acetate and butanediol-lactate were strongest (r = 0.58 and r = 0.53). ₁

Formulation design supports condition dependence: PA14 endpoint OD was 0.60 for proline, 0.56 histidine, 0.46 ornithine, 0.40 glutamate, 0.36 aspartate, 0.36 isoleucine, 0.35 arginine, and 0.22 glucose; threonine, methionine, cysteine, serine, and glycine each supported growth at <0.07. ₅ All 66 aggregate Latent Capability pairs became fitness-important under at least one condition type, most often nitrogen limitation, stress, or carbon limitation, but median-based subset thresholds can cause category crossings by construction. ₁₃

Community analyses refine the endpoint distinction. In 13 testable amino-acid pathways, 11 (85%) had negative community-completeness–metabolite-intensity correlations and the sign test was significant (p = 0.011), but the H1 dataset was dominated by one study (125/131 samples, 95%), all 62 leucine samples came from it, and the measure was genomic potential rather than expression. ₁₈ The largest methionine effect was r = −0.496, n = 18, q = 0.117; tyrosine was an anti-BQH outlier (r = +0.419, ns), and isoleucine showed no signal (r = −0.057, n = 18, q = 0.823). ₁₈ In the merged NMDC matrix, 174 samples had mean taxonomy-bridge coverage 94.6%; all 220 samples passed the 30% quality-control threshold and 92% mapped at least 85% of community abundance to GTDB pangenome species. ₁₈ Approximately 1,352 Centrifuge taxa matched multiple GTDB clades within a genus, and alphabetical tiebreaking selected representatives; these ambiguous taxa accounted for approximately 6.5% of mapped abundance. ₁₈

Environmental and chemical matching remain constraints rather than validation. The NMDC H1 matrix had 131 samples and the full matrix 174; pH, temperature, and total organic carbon were all NaN, preventing partial correlations controlling for environmental gradients. ₁₈ Metabolite matching was string-based; an isoleucine/leucine collision was corrected, KEGG compound IDs had only a 2% annotation rate, and cysteine, histidine, and lysine were untestable because corresponding compounds were absent. ₁₈ Shikimic acid and 3-dehydroshikimic acid were only upstream proxies for chorismate, making the chorismate correlation (r = −0.038) especially uncertain. ₁₈

Tensions

The 73.8% FBA/TnSeq concordance for 866 genes ₁ sits alongside moderate FBA/knockout concordance and a null association between FBA class and growth defects within TnSeq-dispensable genes. ₂ These endpoints and gene sets differ; matched-gene comparisons using the same medium, knockout definition, FBA class, and continuous growth endpoint are needed.

The 42.5% baseline FBA accuracy and 330 false positives ₃ contrast with ADP1-centered evidence that FBA often agrees with TnSeq. ₁ The organisms, media mappings, and endpoints differ; harmonized simulations are required. The 73.8% concordance therefore does not erase the broader accuracy result. ₁

WoM production evidence creates a coverage and endpoint tension with model support: 19 production-to-Fitness-Browser matches lacked consumption actions in the 2018 snapshot, while FW300-N2E3 showed GapMind 13/13 and Fitness Browser 21/21 concordance but BacDive only 3/7. _12,7,11 Production, capability, utilization, dependency, and flux remain distinct endpoints.

Respiratory optimization creates a model–phenotype tension: FBA routed NADH through Complex I and predicted zero NDH-2 and ACIAD3522 flux, while Complex I ratios were 0.37 on quinate and 1.44 on glucose, and ACIAD3522 growth was 0.013 on acetate. ₁₅ This does not establish that FBA is incorrect because NDH-2 lacks growth data, ACIAD3522 may have another function, and the NADH-capacity explanation is theoretical. ₁₅

The aromatic analysis likewise found 1.76× higher Complex I flux but 10/13 operon-subunit growth defects. ₄ Transferred defects were strongest on acetate and succinate rather than exclusively aromatic substrates; ACIAD3522 ratios were 0.013 on acetate and 1.39 on quinate and glucose, and the cross-species NDH-2 compensation hypothesis was not supported after annotation correction (Complex I aromatic deficits −0.297 with validated NDH-2 versus −0.156 without, p = 0.52). ₇ The synthesis calls this ACIAD3522 acetate effect — 0.013 on acetate against no defect on quinate (1.39) or glucose (1.39) — the largest condition-specific effect of any single gene in the 2,034-gene growth matrix, with FBA predicting zero flux through ACIAD3522 on all conditions and its biological role unknown. ₇

Conservation and dependency are also in tension: active-versus-latent conservation was non-significant (p = 0.94), although latent rate correlated with openness (ρ = 0.69, p = 0.0004, n = 22), and variable pathway count correlated with openness at partial rho=0.530, p=2.83e-203. _14,13 The 41.0%, 35.4%, and 15.8% latent fractions from different analyses must not be averaged. _7,13,14

The NMDC negative completeness–metabolite associations suggest a community-scale signal but do not establish individual flux or causality because all 33 Freshwater samples lacked paired metabolomics and abiotic covariates were unavailable. ₁₈ The Carbon Census similarly found broad occurrence—implicated genera appeared in 83/86 genera across 1,719 NMDC metagenomes—but measured occurrence rather than degradation or activity. ₈

The plant study adds a complementarity tension: corrected genus-pair analysis found weakly negative complementarity (d ≈ −0.4; prevalence-weighted d = −0.39), whereas the original analysis reported d = −7.54 because it divided by the standard deviation of permutation means rather than raw pair-level values. ₁₆ The corrected result does not establish metabolite exchange: 62/322 genera (19.3%) were lost from the NMDC bridge, the C-score test found 0 PGP-dominant, 3 pathogen-dominant, and 66 dual-or-mixed genera, and core-level GapMind completeness was 0% across compartments. ₁₆

Interpretation limits

The 87% dependence of 105,376/121,519 growth predictions on gapfilled reactions limits interpretation as direct consequences of genome-supported metabolism, and the 243 missing functions represent genuine gaps in genomic evidence that affect prediction reliability. ₁ The ADP1 comparison cannot determine whether each discordance reflects an incorrect reaction, unmodeled regulation, or condition differences. ₁ No gene had measurements across all six ADP1 modalities, so integrated validation is necessarily partial. ₁

Default ModelSEED gapfilling minimizes added reactions but does not guarantee biological optimality; manual curation mapped 109 carbon sources, Fitness Browser significance depends on thresholds and experiment counts, and GapMind covers about 80 carbon and amino-acid pathways rather than full metabolism. ₃ The capability studies used SEED-proxy membership in one analysis, only 7 of 48 Fitness Browser organisms had matching GapMind data in the newer study, and the broader analysis covered 80 pathways across approximately 293,000 genomes. _14,13 The broader study included 2,810 GTDB species with at least 10 genomes, but more-sequenced species were more likely to show pathway variation and ecotypes; partial correlations did not remove sampling bias and full phylogenetic independent contrasts were not computed. ₁₃ The related ecotype analysis found phylogeny dominated gene content in 60.5% of species. ₁₃

WoM was a 2018 Wayback snapshot containing 37 organisms across 5 ENIGMA projects and 589 metabolites, of which 332 (56.4%) were unidentified with an Unk_ prefix. ₁₂ Only two direct Fitness Browser strain matches were available; E. coli BW25113 had 12 WoM observations versus 168 Fitness Browser experiments for the matching Keio strain. ₁₂ Formula-only ModelSEED matches, internal GapMind identifiers, and absent species-level pangenome matching limit reaction-level interpretation. ₁₂

The respiratory interpretation is limited by absent NDH-2 growth data, theoretical rather than measured flux, possible ACIAD3522 misclassification, annotation-dependent cross-species calls, and an underpowered comparison with only 4 organisms lacking NDH-2. ₁₅ Text matching produced likely false-positive NDH-2 calls and the planned pangenome KO co-occurrence analysis was not performed. ₁₅ The aromatic analysis used only 8 conditions and 8-dimensional growth vectors, and 11 non-core Complex I assignments relied on correlation rather than direct physical evidence. ₄

The Pseudomonas analysis covered 12,732 genomes across 433 species clades, but P. aeruginosa comprised 53% (6,760/12,732); source metadata were classifiable for 8,171/12,732 (64.2%), approximately 6.7% were unknown and approximately 29.1% other, and many environmental species had fewer than 10 genomes. ₁₇ GapMind omitted genus-specific capabilities including toluene, naphthalene, and benzoate degradation. ₁₇ Plant-associated analysis likewise had only 7,995/293,059 genomes (2.7%) with plant annotations, only 29 endophyte species, and 11.7% Jaccard overlap between pangenome-derived and MGnify rhizosphere genera. ₁₆

Formulation evidence was limited to planktonic PA14 assays, a 22-substrate panel omitting mucins, lipids, iron, polyamines, and identified sugar alcohols and pentoses, 142 isolates representing 62/211 species (29%), kinetics for 32 isolates, and 21 lung genomes across 5 species. ₅ Clay comparison began with 9 deep genomes and retained 6 after filtering, versus 30 shallow and 137 baseline genomes; labels were inferred from isolation-source keywords and the cohort represented cultivable isolates. ₆

The FW300-N2E3 study could test only 21/58 WoM metabolites (36%); manual matching associated 28/58 (48%) with Fitness Browser conditions, 8/58 (14%) with BacDive, and 13/58 (22%) with GapMind. ₁₁ WoM used R2A rich medium, Fitness Browser minimal single-source media, and BacDive aggregated P. fluorescens strains across a broad GTDB-reclassified clade. ₁₁

NMDC included 126 Soil, 33 Freshwater, and 61 unrecorded-ecosystem samples; H1 was soil-dominated because all Freshwater samples lacked paired metabolomics, and 125/131 H1 samples (95%) came from nmdc:sty-11-r2h77870. ₁₈ Plant phylogenetic-tree coverage existed for only 18 of 65 plant-associated species with at least 20 genomes; only 5 of 17 testable species passed Bonferroni correction. ₁₆ Genome size correlated with total marker count (r = 0.44, p < 1e-300), and refined dual-nature rates were 54% in the lowest genome-size quartile versus 87% in the top three. ₁₆

Open Directions

• Join the 243 missing functions and 105,376 gap-dependent predictions to pangenome conservation and annotations; test whether core/accessory status predicts false-negative frequency and mean gap count. ₁
• Re-run enrichment on the 227 FBA–TnSeq discordances, stratify 121,519 predictions by gap count, reaction class, and phenotype, and compare rich/minimal flux for the 177 condition-changing genes. ₁
• Build matched-gene comparisons of FBA, knockout lethality, TnSeq fitness, and growth; reconstruct 38 false negatives with curated media and alternative gapfilling and test whether 42.5% accuracy and 330 false positives improve without reducing 86.5% recall. _2,3
• Prioritize the 44 high-confidence assignments and characterize the 50 EC-less reactions. ₃
• Add measured aromatic compounds, re-simulate beta-ketoadipate cases, and test whether 9/11 discordance and FBA under-prediction decline. ₂
• Test whether unconfirmed trace aromatics explain the model–experiment mismatch by adding trace aromatic compounds to the modeled minimal-medium definition and checking whether aromatic-degradation discordance resolves, and regress continuous FBA flux values against growth rate instead of using binary essentiality class, closing the gap left by a class-level null (p = 0.63) that quantitative effects could still occupy. ₂
• Measure ADP1 NDH-2 deletion, NADH/NAD⁺, respiratory proteins, and growth on quinate, glucose, acetate, lactate, and urea; characterize ACIAD3522 and add enzyme-capacity constraints. ₁₅
• Expand K03885 and K00330–K00343 searches across 27K species and perform Acinetobacter pangenome co-occurrence analysis. ₁₅
• Re-run the 23 pathway-organism comparisons with condition-specific fitness and benchmark Bakta, eggNOG, and their union on unresolved gapfilled pairs. ₇
• Check lower-confidence DvH serine predictions with serine-free medium, complementation, and targeted validation. ₉
• For 1,256 functional-dark-matter organism–pathway pairs, compare GapMind with matched Fitness Browser phenotypes and test EC-prefix, Pfam, and keyword evidence. ₁₀
• Map 601 FW300-N2E3 fitness-active genes to GapMind steps and distinguish biosynthetic, catabolic, regulatory, and housekeeping signals. ₁₁
• Re-run capability/dependency analyses with direct GapMind assignments, matched media, explicit clade linkage, independent threshold calibration, and phylogenetic contrasts; test whether the 15.8% latent fraction, ρ = 0.69, rho=0.530, and rho=0.322 persist. _14,13
• Measure growth, uptake, secretion, and flux for latent, intermediate, and active pathways; test whether class predicts FBA reliability. ₁₄
• Obtain current WoM/GNPS2 data, recover consumption actions, build pathway-to-substrate/product mappings, curate 107 formula-only candidate sets, and test whether exact identity improves reaction-level gapfilling. ₁₂
• Add measured metabolite concentrations, uptake bounds, metatranscriptomics, and study-aware mixed models to NMDC; obtain metabolomics for 33 Freshwater samples and test whether 11/13 negative directions persist after abiotic adjustment. ₁₈
• Resolve 61 Unknown-ecosystem samples through NMDC ENVO annotations and improve KEGG mapping for cysteine, histidine, and lysine. ₁₈
• Test accessory-dependent amino-acid completeness experimentally using community composition, extracellular leucine, valine, arginine, lysine, and threonine, uptake, secretion, and expression. ₁₃
• Recalculate plant-associated completeness with relaxed core thresholds, direct GapMind assignments, and matched BacDive strains; test whether 83.1% GapMind–BacDive consistency and 0% core completeness persist. ₁₆
• Pair plant markers with transcriptomics, uptake, secretion, and growth under defined conditions; test whether 78.7% dual-nature classification predicts activity. ₁₆
• Restore the 62/322 lost NMDC genera and reassess plant complementarity using reaction-level profiles and exchange measurements. ₁₆
• Extend GapMind with toluene, naphthalene, benzoate, and other aromatic pathways; apply PGLS or phylogenetic logistic regression using the GTDB tree and balanced sampling. ₁₇
• Use the 789,012-row carbon-pathway matrix to model within-species variation in P. fluorescens and P. putida and cross-reference ecotypes with RB-TnSeq fitness data. ₁₇