Species-level utilization claims should be based on per-strain consensus and reported with the number of strains supporting each result. Per-strain consensus applies a majority vote among duplicate records for each strain before calculating the species-level utilization rate, preventing duplicate records from being treated as independent strains. 1
Why sample size changes interpretation
A utilization percentage without its strain count can obscure whether a result is broadly supported, strain-variable, or based on too little evidence. In the FW300-N2E3 cross-database analysis, BacDive utilization counts ranged from 1 to 51 strains, and raw records could be higher because of duplicate entries per strain. 1 The report therefore treats sample size as part of the evidence rather than as supplementary metadata. 1
The AMR strain atlas supports and broadens this principle: across 180,025 genomes from 1,305 species, 51.3% of 37,444 AMR gene-species records were rare, present in <=5% of strains, while 41.3% were variable and 7.5% were fixed. The median pairwise Jaccard distance was 0.435, showing substantial within-species repertoire variation. 2 Thus, even a species-level database with many genomes can contain strongly strain-dependent traits; a large nominal database is not evidence that a particular phenotype is uniformly distributed. The AMR analysis also found that only 974 species had at least 15 genomes suitable for ecotype clustering, and strict environment-ecotype testing was possible for only 2 species. 2 This refines the denominator rule by making minimum sample size and usable metadata explicit prerequisites for interpreting population structure, not merely reporting conventions.
The BacDive metal-tolerance bridge supports this principle: although it began with 97,334 BacDive strains, only 37,368 (38.4%) matched a pangenome species and metal score, and only 3,994 species had at least 5 phenotype features. Feature-level coverage varied from 5,254 strains for H₂S production to 1,980 for acetate utilization, with species carrying metal scores ranging from 4,531 to 422 across features. 3 Thus, a large database does not guarantee a large or balanced denominator for any particular phenotype. The related BacDive metal-validation bridge further supports denominator-aware interpretation: 42,227 of 97,334 strains (43.4%) matched 6,426 GTDB species, but only 25,089 matched strains had isolation metadata, and 55,107 (56.6%) were unmatched. 4 Its direct metal-utilization comparison contained only 24 matched records, with 8 positive and 16 negative results; the exploratory contrast was inconclusive (Mann–Whitney p=0.14; Cohen’s d=-0.57). 4
The AMR study further supports this warning about usable denominators: 52.7% of genomes had no classifiable isolation_source, and only 70% had parseable collection dates, limiting ecological and temporal inference despite the overall scale. 2 Its host-association result was directionally consistent across two approximate classifiers, but the NCBI keyword classifier assigned environments to 1,190/1,307 species whereas the BacDive approximation classified 459/1,307. 2 This refines interpretation of utilization rates: the denominator should be accompanied by metadata coverage and grouping structure, because nominal strain counts can conceal missing or unbalanced context.
This principle is also supported by the ENIGMA genotype-to-phenotype analysis: a model trained on only 7 strains shifted from genome-scale proxies to condition-specific predictors when expanded to 46,389 genome × condition pairs from 727 genomes, and the report concludes that mechanistic gene-specific prediction requires hundreds of genomes per condition rather than a small number of anchor strains. 5 Although this result concerns prediction rather than BacDive utilization rates, it refines the interpretation rule: both species-level phenotype summaries and genotype-based predictions should expose the strain denominator and avoid treating sparse anchors as population-wide evidence. 5
This approach is especially important when bacdive species-level phenotypes are used to interpret metabolite production by one isolate. BacDive aggregates P. fluorescens strains and, under GTDB reclassification, spans a broad clade identified as Pseudomonas_E fluorescens_E; consequently, a species-level rate may not represent the phenotype of FW300-N2E3 specifically. 1 This limitation connects species-level phenotype inference to environmental-resistome and taxonomic-resolution-dependent-functional-inference.
The ENIGMA analysis further supports caution about strain identity: a short-identifier collision caused 12 of 32 genus-level mismatches, including one that matched Rhodanobacter glycinis to Streptococcus pneumoniae and introduced 1,751 spurious clinical genomes into environmental profiles; GTDB-Tk assignments and genus-consistency checks eliminated all false matches. 5 Thus, a strain count is informative only after duplicate records and cross-database identities have been validated. The BacDive metal-tolerance bridge refines this warning because species-name matching recovered 5,647 of 27,702 GTDB species (38.4%), while GCA accession matching was not implemented and could recover additional links. 3 The newer validation likewise used species-name matching, including GTDB suffix removal, and reports that 42,227 strains matched while 56.6% remained unmatched; it therefore supports pursuing accession-based linkage rather than treating the observed denominator as complete. 4
Erratum. The report gives 38.4% for two different denominators. 37,368 of 97,334 matched strains is 38.4%; 5,647 of 27,702 GTDB species, the figure this sentence uses, is about one fifth, roughly half what the percentage suggests. 3
Defining the numerator before the denominator
A strain denominator is only interpretable once the numerator is defined. The pitfalls digest records that BacDive’s metabolite_utilization.utilization field stores four distinct values — + (can utilize), - (cannot utilize), produced (organism produces it), and +/- (variable/ambiguous) — so treating it as a binary +/- field inflates counts: for pseudomonas-fluorescens, the organism used in the worked example, indole has 60 produced entries and only 1 actual utilization test, and computing pct_positive from raw +/- counts without excluding produced and +/- entries gives incorrect species-level utilization rates. 6
The remedy refines the consensus rule rather than replacing it: filter to explicit +/- tests before computing utilization percentages, taking n_tested as the count of + plus the count of - and pct_positive as the count of + divided by n_tested, excluding produced and +/-, and track all four categories separately (n_positive, n_negative, n_produced, n_ambiguous) for full transparency. 6 A reported strain count therefore has to state which record categories it includes, because the same raw table yields different utilization rates depending on that choice. 6
Evidence from FW300-N2E3
The strongest example was tryptophan. FW300-N2E3 increased tryptophan in the WoM exometabolome, had 231 genes with significant fitness effects when grown on tryptophan in the Fitness Browser, and had a complete tryptophan biosynthesis pathway prediction from GapMind. 1 In contrast, 0 out of 50 P. fluorescens strains in BacDive could utilize tryptophan as a carbon source, with high confidence based on n=50 and pct_positive=0.0 using per-strain consensus. 1 This is the most robust production-versus-utilization discordance in the report, but it does not establish that FW300-N2E3 itself cannot grow on tryptophan. 1
The central discoveries digest states the same convergence with the larger denominator: FW300-N2E3 produces tryptophan (WoM: Increased), has 231 genes with significant fitness when grown on tryptophan, and has a complete tryptophan biosynthesis pathway, yet 0/52 P. fluorescens strains in bacdive can utilize tryptophan as a carbon source. 7 The digest reads this convergence of four independent lines of evidence as a case for tryptophan overflow metabolism serving a cross-feeding or signaling function rather than a catabolic one, consistent with Fritts et al. (2021) and Ramoneda et al. (2023) on amino acid cross-feeding and tryptophan auxotrophy in soil communities; that functional role is an inferred explanation rather than a measurement, and the digest’s denominator is the 0/52 variant of the discrepancy recorded below, not the 0+/50- primary result. 7
The evidentiary strength of other BacDive non-utilization results was lower or more heterogeneous. Trehalose was Increased in the WoM exometabolome yet had 1 positive and 5 negative strains, or 1/6, classified with moderate confidence at n=6, with the report reading the compound as a likely osmoprotectant rather than a catabolic substrate — an interpretation, not a measured role. 1 Lysine Emerged in the WoM exometabolome and had 0 positive and 3 negative strains, or 0/3, also with moderate confidence (n=3), a consistent negative on a small sample. 1 Glycine was Increased in the WoM exometabolome and had 0 positive and 1 negative strain, or 0/1, with low confidence (n=1), a single measurement the report calls unreliable. 1 These results support treating trehalose as potentially strain-variable, lysine as consistently negative within a small sample, and glycine as insufficient for a conclusion rather than treating all three as equivalent species-level findings. 1 All BacDive counts in these comparisons use per-strain consensus — a majority vote among duplicate records per strain — and on that basis only tryptophan reaches high confidence for a genuine production-versus-utilization discordance, trehalose is strain-variable at 1/6 positive, and glycine’s single-strain measurement is insufficient to draw conclusions. 1
Thin denominators inflate apparent agreement as well as disagreement: valine Emerged in the WoM exometabolome, had growth with 160 genes, a complete pathway, and BacDive utilization recorded as utilized (1/1), so this four-way concordance rests on a single tested strain. 1 Positive species-level calls therefore require their strain counts to be exposed on the same terms as negative ones. 1
The AMR study supports this distinction between fixed, variable, and rare strain-level outcomes. In that atlas, 77.3% of Core AMR genes were fixed, 78.7% of Singletons were rare, and Auxiliary genes were 57.3% variable and 42.7% rare. 2 These categories are not utilization measurements, but they provide a directly measured cross-species example of why a species label should not be treated as a uniform phenotype without prevalence and strain counts.
The report also records a denominator discrepancy that should remain visible: its primary tryptophan result reports 0+/50- with n=50, while its literature-context discussion states that only tryptophan, with n=52, has sufficient data for a confident conclusion. 1 The values should not be silently reconciled or averaged.
The BacDive metal-tolerance study supports the same distinction between nominal effect size and evidentiary strength. Its H₂S association had d = -0.867 but only 8 negative controls, making the estimate unreliable and likely inflated by small-sample bias; acetate utilization was based on only 1,980 strains at the feature level. 3 It also found that apparent Gram-stain, urease, and catalase associations were largely between-lineage or reversed within taxonomic classes, showing that a phenotype summary with a sufficient-looking denominator can still be misleading if its observations are phylogenetically unbalanced. 3 The metal-validation study further supports this caution: its heavy-metal contamination comparison had Cohen’s d=+1.00 but only n=10 matched isolates, and the observed effect barely exceeded an approximately d=0.93 minimum detectable effect at 80% power. 4 Its result is therefore ecological validation of the scoring approach, not a reason to treat a small utilization sample as conclusive.
Relationship to cross-database concordance
BacDive was the variable component of the cross-database comparison. The report describes the 94% concordance as largely structural, because Fitness Browser and GapMind comparisons are invariably concordant with WoM production and the only variable component is BacDive. 1 Among 41 individual metabolite-database comparisons, 37 were concordant, or 90.2%; Fitness Browser comparisons were concordant for 21/21, or 100%, and GapMind comparisons were concordant for 13/13, or 100%, whereas BacDive utilization was positive for only 3/7 matched metabolites, or 43%. 1 The 3/7 rate for WoM-produced metabolites was compared with the overall P. fluorescens baseline of 22/80, or 27.5%, using a binomial test that gave p = 0.40. 1
These results refine multi-omics-integration by showing that apparent cross-database disagreement can reflect the resolution and sample size of the phenotype resource, not only biological inconsistency. The AMR atlas supports this interpretation at species scale: 701/1,261 species (55.6%) showed significant correspondence between ANI and AMR Jaccard distances at FDR < 0.05, but ecotype-environment testing was severely limited by missing isolation_source metadata and insufficient within-species environmental diversity. 2 A large strain collection can therefore reveal structure while still being unable to support every ecological or phenotype-specific comparison.
The ENIGMA analysis supports this interpretation: binary growth prediction varied substantially by condition class, with AUC 0.775 for amino acids, 0.780 for nucleosides, 0.695 for carbon sources, 0.619 for antibiotics, 0.605 for metals, and 0.435 for nitrogen, demonstrating that a pooled performance value can conceal condition-specific heterogeneity. 5 They also support phenotype-database-coverage-bias because only 8/58 WoM metabolites, or 14%, could be matched to BacDive P. fluorescens utilization data. 1
The distinction between production and utilization is also necessary. WoM exometabolomics can measure overflow metabolism, biosynthetic byproduct release, or active secretion, whereas BacDive growth assays measure utilization as a growth capability. 1 Therefore, production by FW300-N2E3 and non-utilization at the aggregated species level are not intrinsically contradictory. 1
Consensus beyond utilization: majority votes and fractions
Majority-vote consensus is applied across this corpus to more than metabolite utilization, and its costs are recorded. The plant-microbiome ecotype project assigned species-level plant-association classifications by majority vote with mean confidence 0.88. 8 The Pseudomonas carbon-ecology project qualifies this practice: assigning species to a single environment category based on majority vote obscures genuinely generalist species that inhabit multiple environments. 9 A majority vote is therefore well suited to collapsing duplicate records within a strain, as in the BacDive per-strain step, but it discards within-species heterogeneity when the vote is taken over strains or genomes instead. 1,9
Reporting the fraction rather than the winner is one alternative. The Pseudomonas carbon-ecology project computed species-level pathway completeness as the fraction of genomes in which each of 62 gapmind carbon pathways was scored “complete” or “likely_complete”. 9 Comparing the 7 Pseudomonas s.s. species with the 189 Pseudomonas_E species, each with >=5 genomes, Mann-Whitney U tests with Benjamini-Hochberg false-discovery-rate correction (control of the expected proportion of false positives among significant calls) identified 43 of 62 pathways as significantly different (q < 0.05). 9 This supports the denominator rule from the other direction: a fraction-based species summary built on an explicit per-species genome minimum sustained pathway-level significance testing across 43 of 62 pathways while retaining the within-species spread that a single majority-vote label would collapse. 9
The aggregation choice also has a mechanical form. In the pitfalls digest’s guidance for kbase.ke_pangenome.gapmind_pathways, where score_simplified contains only 0.0 (pathway incomplete) or 1.0 (pathway complete), the table is genome-level, with one row per genome-pathway pair, and should be aggregated to species level with MAX(score_simplified) GROUP BY clade_name, pathway before joining to species data. 6 That procedure is scoped to this join: MAX records whether a pathway is complete in any genome of the clade, which is a different quantity from the genome-fraction conservation metric above and measures neither strain prevalence nor consensus. 6 A species-level pathway claim that skips the aggregation step is instead counting genome rows, so its denominator is not the species count it appears to be. 6
Pangenome conservation as a strain-denominator test
The most direct test of whether species-level phenotype claims generalize across strains comes from the cystic-fibrosis formulation-design project. Its formulation design is based on the carbon utilization profiles of specific PROTECT isolates and could fail when different strains are used clinically if metabolic capabilities vary significantly between strains of the same species, so it examined gapmind metabolic pathway conservation across 499 pangenome genomes — the pooled gene content of all sequenced genomes of a species — in its 5 core species. 10
- micrococcus luteus: 295 genomes, with 18/18 (100%) amino-acid pathways and 39/39 (100%) carbon pathways above 95% conservation. 10
- streptococcus salivarius: 153 genomes, with 18/18 (100%) amino-acid and 32/35 (91%) carbon pathways meeting the conservation criterion. 10
- rothia dentocariosa: 29 genomes, with 14/18 (78%) amino-acid and 39/41 (95%) carbon pathways meeting the criterion. 10
- neisseria mucosa: 15 genomes, with 16/16 (100%) amino-acid and 27/27 (100%) carbon pathways; the pangenome table’s amino-acid denominator here is 16, whereas the project’s own core-species profile reports 16/18 for the same species, and the two denominators should stay visible rather than be reconciled. 10
- gemella sanguinis: only 7 genomes were evaluated, with 7/18 (39%) amino-acid pathways but 37/39 (95%) carbon pathways meeting the criterion. 10
Before this profile is read as validation of the formulation isolates, the project’s clade-selection caveat has to be carried with it: the pangenome contains two N. mucosa clades, the conservation notebook uses s__Neisseria_mucosa_A (15 genomes), but the PROTECT isolate reference genome (GCA_003028315.1) maps to s__Neisseria_mucosa (8 genomes). 10 A sensitivity check on the 8-genome clade shows stronger conservation — 18/18 AA pathways at >95%, versus 16/18 for the 15-genome clade, and 37/62 carbon pathways versus 27/62, counted against a different denominator than the main table — and the 8-genome clade also has 1 respiratory genome; the project reads this as meaning its conservation claims for the formulation anchor species are, if anything, understated by the primary analysis. 10 The denominator used for a species-level claim is therefore a clade-choice decision as well as a sample-size one. 10
The consequence is visible in the project’s own core-species profile, where gemella sanguinis is retained as a core species with 85% best inhibition on a profile of only 7 pangenome genomes. 10 The project characterizes it as strong inhibition (85%) by a dual mechanism (residual +62%) that adds depth at k=5, while naming the smallest pangenome (7 genomes) as the reason strain selection is most critical for this species. 10
The project concludes that its conservation hypothesis H5 is strongly supported — the metabolic capabilities it measured are species-level traits conserved across hundreds of genomes — that G. sanguinis shows the most pathway variability (small pangenome, 7 genomes) so strain selection matters most there, and that for the other four species any well-characterized strain should provide equivalent metabolic competition. 10 This refines rather than contradicts the consensus rule: a species-level phenotype claim can be well supported when the genome denominator is large and the pathway is conserved across it, and the same accounting names where support is thinnest — the 7-genome species, whose amino-acid pathways are the most variable at 7/18 even though 37/39 (95%) of its carbon pathways meet the conservation criterion, so the reported remedy is careful strain selection rather than discarding the species-level claim. 10 The conserved quantity is predicted pathway presence rather than measured growth, so the result bounds strain variability in genomic capability and does not by itself demonstrate equivalent phenotypes. 10
Interpretation rule
A species-level utilization claim should report at least the per-strain positive and negative counts, the resulting utilization fraction, the confidence or evidentiary category, and whether duplicate records were collapsed by per-strain consensus. 1 Claims based on large samples, such as tryptophan at 0+/50- with n=50, can support stronger conclusions than claims based on 0/1 or 0/3 strains. 1 Even a high-confidence species-level result remains a population-level constraint on interpreting FW300-N2E3, not a direct measurement of that isolate’s phenotype. 1
The AMR results refine this rule: strain consensus should report not only the denominator but also prevalence class, lineage structure, and metadata completeness when those factors can alter interpretation. AMR resistance islands occurred in 705 of 1,305 species, and their mean pairwise co-inheritance phi coefficient was 0.827; however, co-occurrence does not establish co-selection or functional synergy. 2 This illustrates why a consensus result can identify a structured strain-associated pattern without proving its mechanism. The AMR study also found no significant temporal trend in AMR gene count among 513 species after Benjamini-Hochberg FDR correction, while noting sparse and noisy collection-date metadata. 2 Null or positive population-level summaries therefore require the same denominator and metadata qualification as utilization claims.
The ENIGMA results refine this rule by showing that sample size must be paired with condition coverage and identity quality: the integrated corpus included 46,389 pairs across 363 conditions, yet approximately 76% of ENIGMA conditions lacked GapMind pathway coverage or Carbon Source Phenotypes training data, and prediction on those conditions fell to approximately AUC 0.63. 5 A large denominator therefore cannot by itself establish broad utilization if the relevant conditions are poorly represented or the linked strains are misassigned.
The BacDive metal-tolerance results further support reporting both denominator and grouping structure. In a direct Fitness Browser–BacDive validation, 12 organisms represented 6 unique species; all Gram-typed organisms were Gram-negative, all urease-typed organisms were urease-negative, and the single anaerobe, methanococcus-maripaludis, had only 1 metal tested versus 4–5 for aerobes. 3 These design imbalances prevented interpretable within-set tests, so strain counts should be accompanied by coverage across taxa, conditions, and phenotype classes rather than reported alone. The new BacDive validation supports the same rule: contamination effects were significant within Pseudomonadota and Actinomycetota but not within Bacillota or Bacteroidota, where contamination groups were small (n=19 and n=5, respectively), so pooled or subgroup claims require their taxonomic denominators to remain visible. 4
The detailed evidence is summarized in BacDive Isolation Environment × Metal Tolerance Prediction and Within-Species AMR Strain Variation.
Open Directions
- Extend the FW300-N2E3 crosswalk to additional isolates, including Pseudomonas stutzeri RCH2, and test whether BacDive utilization rates are reproducible across isolates rather than driven by the current strain composition. 1
- Expand WoM–BacDive matching with InChIKey or CHEBI identifiers, then reassess whether the observed 8/58 BacDive coverage and 3/7 utilization rate change when nomenclature-based missed matches are reduced. 1
- Reanalyze BacDive records with explicit per-strain consensus and sample-size thresholds, asking which utilization claims remain stable after duplicate records and low-n comparisons are separated. 1
- Compare strain-resolved BacDive phenotypes with isolate-specific Fitness Browser and WoM data, asking whether species-level non-utilization predicts or obscures FW300-N2E3-specific growth and secretion. 1
- Apply the ENIGMA identity-validation and condition-coverage checks to the expanded FW300 crosswalk, then test whether utilization consensus remains stable when predictions are evaluated across hundreds of genomes per condition rather than sparse strain anchors. 5
- Extend BacDive matching with GCA accessions and use class-matched or phylogenetically blocked analyses to determine whether low-n phenotype associations persist after coverage and lineage imbalance are controlled. 3
- Expand the BacDive metal-utilization extraction beyond the 24 matched records, including MIC and growth-inhibition data where available, and test whether utilization conclusions remain stable after accession-based matching and explicit per-strain consensus. 4
- Use the AMR atlas presence/absence matrices and per-species ecotype summaries to test whether phenotype consensus changes across lineage-defined strain groups, and whether conclusions remain stable after restricting analyses to species with adequate isolation_source coverage. 2
- Recompute the FW300-N2E3 BacDive rates after filtering to explicit +/- tests and tracking
producedand+/-records separately, asking whether the 3/7 and 22/80 rates move onceproducedentries of the kind seen for indole are excluded. 6,1 - Repeat the 499-genome GapMind conservation test on Pseudomonas_E fluorescens_E genomes, reporting per-pathway genome fractions and the clade actually matched to each reference isolate, and asking whether the utilization-relevant pathways behind the FW300 crosswalk are as variable as the amino-acid pathways of the 7-genome G. sanguinis pangenome. 10
- Replace majority-vote species-level environment and plant-association labels with fraction-based summaries at a stated genome minimum, testing whether species currently collapsed into one category change the ecological comparisons. 9,8