Perturbation-Collection Coverage Bias

Perturbation collections can overrepresent genes that are conserved, annotated, and successfully recoverable as mutants, while underrepresenting less-conserved or poorly annotated genes. ₁ The broader Fitness Browser comparison supports this concern: across approximately 194,000 genes from 43 bacteria, essential genes were 82% core whereas always-neutral genes were 66% core, although fitness importance was only a weak predictor of conservation. ₂

This pattern matters for interpreting gene-essentiality and condition-specific-fitness: apparent functional coverage may reflect which genes enter the perturbation collection rather than the full genomic distribution of gene functions. ₁ Evidence that core genes can themselves be conditionally burdensome refines this interpretation: core status is not equivalent to uniformly low burden or universal essentiality. In the broader Fitness Browser analysis, core genes were more burdensome than non-core genes in several functional categories, and 25,271 genes were classified as condition-dependent trade-offs; these findings are based on laboratory fitness and conservation patterns rather than direct measurement of natural selection. ₃ The new comparison further supports this refinement: core genes showed heavier tails in both negative and positive fitness effects, and 24.4% were ever beneficial versus 19.9% of auxiliary genes. ₂

Evidence from the ADP1 deletion collection

The analysis compared a complete growth matrix of 2,034 genes across 8 carbon sources with TnSeq, or transposon sequencing, essentiality classifications and pangenome annotations. ₁

Among 2,593 TnSeq-dispensable genes, 272 (10.5%) lacked growth data from the deletion collection, whereas 2,321 had growth data. ₁ The genes missing from the deletion collection had a mean length of 813 bp, compared with 981 bp for present dispensable genes. ₁ Missing genes were RAST annotated at 91% versus 100% for present dispensable genes and KO annotated at 49% versus 59%. ₁ The missing genes were also less likely to be pangenome-core: 76.5% were core compared with 93.3% of present dispensable genes, a difference with p = 1.4×10⁻²⁰. ₁

These results support the interpretation that deletion-collection coverage is biased toward more conserved genes, as measured by species-level pangenome core status, and toward genes with stronger functional annotation. ₁ The broader conservation–fitness gradient supports the direction of this association while showing that it is not deterministic: genes affecting 20+ experiments were 79% core, compared with 66% for genes with 0 experiments, with Spearman rho=0.086 and p=8.1e-230. ₂ The core-gene trade-off analysis refines the meaning of this bias by showing that conserved genes can have strong condition-specific fitness effects rather than forming a uniformly inert or universally essential class. ₃ The cross-project discoveries digest restates the ADP1 result: the deletion collection omits 272 of 2,593 TnSeq-dispensable genes, and the omitted genes are shorter, less annotated and less conserved, so the digest describes the collection as biased toward well-characterized, conserved genes. ₄

A respiratory-chain case shows how one uncovered gene can block a planned test. acinetobacter-baylyi-adp1 has 62 respiratory chain genes across 8 subsystems. Of these, 36 have growth data and 26 do not, including NDH-2 and several ATP synthase subunits. ₅ ndh 2 (ACIAD_RS16420, KO K03885) is a one-subunit, non-proton-pumping dehydrogenase. It is TnSeq-dispensable but missing from the deletion collection, so it has no growth data for any tabulated condition. It is a standalone, core-genome gene. The report’s FBA model predicts zero flux through it on all standard carbon sources and routes all NADH through Complex I. FBA is flux balance analysis, a constraint-based metabolic model prediction. ₅ As a result, the report’s central prediction that NDH-2 compensates on glucose cannot be directly tested with this dataset. ₅ This refines the ADP1 coverage association. Missing genes were less often core on average, but a core gene can still be absent, and the absence matters most when it falls on the gene a hypothesis depends on. _1,5

The cross-organism costly–dispensable analysis supports the annotation and conservation concern: among 142,190 genes from 43 bacteria, costly+dispensable genes had SEED annotation rates of 50.8% versus 74.9% for costly+conserved genes and KEGG annotation rates of 20.0% versus 42.7%; 44.5% were orphans with no ortholog group, compared with 13.1% of costly+conserved genes. ₆ It also refines the interpretation by showing that the poorly annotated, narrowly distributed group is enriched for mobile-element signatures rather than being simply a random collection of uncharacterized genes: mobile-element keywords were 7.45x more likely in costly+dispensable than costly+conserved genes (OR=7.45, p=4.6e-71). ₆ This result characterizes the genes most vulnerable to coverage or interpretation bias but does not itself measure mutant-recovery probability. ₆

Annotation bias and uncertain genes

Hypothetical proteins were enriched among genes missing from the deletion collection: 25 were completely unannotated, with q = 2.4×10⁻²⁵, and 48 were annotated as “hypothetical protein,” with q = 3.0×10⁻⁴. ₁ The 313 uncertain-class genes had a mean length of 361 bp, 42% annotation coverage, and 31% pangenome-core status, a profile the report considers consistent with gene fragments or pseudogenes rather than true essential genes. ₁

This annotation pattern connects coverage bias to genomic-under-representation: poorly characterized genes are not merely difficult to interpret after perturbation, but may also be less likely to have usable perturbation phenotypes in the first place. ₁ The Fitness Browser analysis supports retaining this caution: novel singleton genes showed near-zero mean fitness under the tested laboratory conditions, but the report notes that apparent neutrality may reflect poor transposon coverage rather than biological inactivity. ₂ The costly–dispensable analysis refines this caution: 24.2% of costly+dispensable genes were singletons, while no costly+conserved genes were singletons, but within the dispensable category costly genes were only slightly more likely than neutral genes to be singletons (OR=1.09, p=0.02); the report also notes that the first comparison is partly structural because core genes cannot be singletons by definition. ₆

Consequences for inference

The deletion matrix excludes 499 essential genes and 316 genes with incomplete data, so its condition-specific fitness landscape is biased toward dispensable genes with successful deletion mutants. ₁ The ADP1 triple-essentiality analysis supports this structural constraint. All 478 triple-covered genes, meaning genes with TnSeq, FBA and growth data, were TnSeq-dispensable on minimal media, because TnSeq-essential genes have no viable deletion mutants and so cannot have growth-rate measurements. Its FBA–growth comparison can therefore ask only whether FBA predicts which dispensable genes matter for growth. ₇

Two other analyses support this structural exclusion of essential genes. In the module-conservation analysis, 0 essential genes appeared in any ICA module. ICA is independent component analysis, a decomposition that groups genes with shared fitness variation. The report attributes the exclusion to essential genes having no fitness data: ICA modules capture only genes with measurable fitness variation, meaning non-essential genes with transposon insertions. ₈ The functional dark-matter prioritization shows the same exclusion in a candidate ranking. No essential genes appear in its NB05 top 100 candidates, because essential genes have zero rows in genefitness and no fitness scores beyond the essentiality call. Its fitness-importance dimension therefore scores them at the essentiality bonus floor, and its experimental-tractability dimension penalizes them for being non-knockable. ₉

The report therefore supports using perturbation results together with pangenome and annotation data, rather than treating the assayed gene set as an unbiased sample of the genome. ₁ This supports separating collection coverage from biological interpretation of fitness: the core-gene analysis found 28,017 Costly + Conserved genes, indicating that conservation and laboratory burden can coexist, although the result does not directly measure selection in nature. ₃ The broader analysis refines this consequence by finding that strong condition-specific effects were enriched among core genes, with 77.3% of genes tagged with strong specific phenotypes core versus 70.3% without them. ₂ The report raises detectability as a possible explanation, not a demonstrated one: core genes may simply have more detectable conditional effects because they participate in more pathways. If so, the enrichment would partly be a measurement-visibility effect. ₂ Likewise, 14.1% of costly+dispensable genes had condition-specific phenotypes, compared with 16.7% of costly+conserved genes and 2.7% of neutral+dispensable genes, suggesting that some poorly conserved genes can remain phenotypically informative under particular laboratory conditions. ₆

The conservation comparison is informative but limited because core/accessory status came from the KBase Data Lakehouse’s species-level Acinetobacter baylyi pangenome and may have less resolution than a population-level analysis. ₁ This limitation refines pangenome-integration: pangenome status can diagnose collection bias, but its interpretation depends on the taxonomic and population scale of the reference pangenome. ₁ The Fitness Browser comparison adds a distinct generalizability limitation: its measurements were biased toward rich media and standard stresses, covered 43 bacteria primarily from Proteobacteria, and used single-gene knockouts that did not capture epistatic interactions. ₂

Coverage gaps also arise in derived data products, not only in mutant collections. The co-fitness and co-inheritance analysis lacked pre-computed co-fitness for Ralstonia and some other organisms. It proposes computing pairwise Pearson correlations directly from raw genefitness data to fill in the missing comparisons. ₁₀

Organism-level coverage of the Fitness Browser panel

Coverage bias also operates at the level of which organisms enter a perturbation panel at all. The essential-genome analysis notes that its 48-organism Fitness Browser panel is biased toward culturable Proteobacteria. As a result, essentiality patterns in uncultured lineages, Actinobacteria and Firmicutes are underrepresented. ₁₁ This extends the generalizability limitation noted above for the 43-bacterium conservation comparison, which drew on a panel described as primarily Proteobacteria. ₂

The functional dark-matter analysis quantifies the skew. It describes the Fitness Browser collection as heavily skewed toward Gammaproteobacteria, with 21/48 organisms and a reported 78% of top-500 dark genes. ₉ The report’s own taxon-coverage table lists 417 Gammaproteobacteria genes among the top 500, which does not match the 78% figure. Both values are reported here as given in the source, without reconciliation. ₉

| Taxon | Organisms in collection | Top-500 dark genes | Gap severity ₉ | |---|---|---|---| | Gammaproteobacteria | 21 organisms | 417 | Saturated | | Betaproteobacteria | 8 organisms | 10 | Moderate | | Alphaproteobacteria | 5 organisms | 71 | Low | | Deltaproteobacteria | 5 organisms | 17 | Low | | Bacteroidetes | 5 organisms | 12 | Low | | Firmicutes | 1 organism | 5 | Critical | | Cyanobacteria | 1 organism | 2 | Moderate | | Archaea | 2 organisms | 0 | Severe | | Actinobacteria | 0 | — | Critical | | Epsilonproteobacteria | 0 | — | High |

Several of these gaps are absent coverage rather than thin coverage. The two Archaea organisms contribute zero top-500 genes. Actinobacteria and Epsilonproteobacteria have zero organisms in the collection, so no top-500 gene count is reported for them. ₉

A later section of the same report frames the bias at phylum level. It reports that 37/48 Fitness Browser organisms are Pseudomonadota (Proteobacteria), with Bacteroidota (4), Cyanobacteriota (1) and Halobacteriota (2) the only other phyla represented. It also reports that Bacillota (Firmicutes), Actinomycetota (Actinobacteria) and Campylobacterota have zero Fitness Browser representation. The report states that kingdom-level ortholog groups, which account for 55.9% of dark-gene ortholog groups, therefore cannot be experimentally addressed in their non-Proteobacterial hosts using Fitness Browser organisms alone. ₉ The report’s figure fig32_organism_taxonomy.png shows the Fitness Browser organism phylum distribution and its sparse taxonomic coverage. ₉

To close these gaps, the report ranks candidate organisms for new RB-TnSeq libraries. RB-TnSeq is random barcode transposon sequencing, a pooled mutant-fitness assay. The ranked candidates are chosen for phylogenetic and ecological breadth. ₉

1. bacillus-subtilis 168 (Firmicutes) is ranked first, as the best-studied Gram-positive model with well-established RB-TnSeq protocols. The existing Fitness Browser strain BFirm produced only 5 top-500 genes. ₉
2. Streptomyces coelicolor A3(2) (Actinobacteria) would add an entirely missing phylum. The report cites a large genome (8.7 Mb, ~30% hypothetical) with many biosynthetic gene clusters. ₉
3. Clostridium difficile 630 (Firmicutes, Clostridia) is an anaerobic Firmicute. It would allow anaerobic-phenotype dark genes, which showed a strong NMDC dissolved-oxygen correlation, to be tested in a strict anaerobe. ₉
4. Mycobacterium smegmatis mc²155 (Actinobacteria) is a fast-growing, non-pathogenic, Tn-seq-tractable model for M. tuberculosis. ₉
5. Campylobacter jejuni NCTC 11168 would add the missing Epsilonproteobacteria, and TnSeq has already been applied successfully in this strain. ₉
6. Lactobacillus plantarum WCFS1 (Firmicutes, Bacilli) would complement B. subtilis with a different ecological niche. It would also allow carbon-source dark genes to be tested in a fermentative bacterium. ₉
7. Rhodobacter sphaeroides 2.4.1 (Alphaproteobacteria) would strengthen cross-organism concordance testing for the 71 top-500 Alphaproteobacteria genes. Its metabolic strategy differs from that of S. meliloti. ₉

The report’s projected payoff is a prospective hypothesis, not a tested result. It proposes that adding B. subtilis and S. coelicolor would fill the two largest phylogenetic gaps (Firmicutes depth and Actinobacteria absence). Together they would permit cross-phylum testing of the ~100 “universal” dark-gene families that can currently be studied only in Proteobacteria. ₉

The report also curated an extended tractable-organism list of 73 organisms. These are the 48 Fitness Browser organisms plus 25 literature-curated organisms with published TnSeq or CRISPRi (CRISPR interference knockdown) resources. The additions include six Bacillota with TnSeq and CRISPRi resources listed: B. subtilis 168, S. aureus USA300, S. pneumoniae TIGR4, C. difficile R20291, E. faecium and L. monocytogenes EGD-e. They include two Actinomycetota, also with TnSeq and CRISPRi: M. tuberculosis H37Rv and C. glutamicum ATCC 13032. They also include two Campylobacterota with TnSeq listed: C. jejuni NCTC 11168 and H. pylori 26695. ₉ The same table lists 15 additional Pseudomonadota with TnSeq/CRISPRi resources. It names P. aeruginosa PAO1, Salmonella 14028s, V. cholerae N16961 and A. baumannii 17978, plus 11 others. The report states that the 73-organism list was assembled to address the Proteobacteria-heavy Fitness Browser sampling bias. ₉

An ortholog-group covering analysis refines the case for these additions. Here an OG is an ortholog group, a cluster of genes inferred to share ancestry. A Spark query mapped genus-level OG membership for 25 non-Fitness-Browser organisms. A conservation-weighted covering set over all 73 candidates then produced a 50-organism set covering 98.7% of OGs across 6 phyla, versus 41 organisms, 100%, 4 phyla for Fitness-Browser-only candidates. Coverage at N=5 organisms reached 57.7% versus 38.1% for Fitness-Browser-only candidates (+19.6%). Bacillota organisms (B. subtilis, S. aureus) were not selected, because their OGs are subsets of coverage already provided by Pseudomonadota. The report notes they remain valuable for studying genes in native Gram-positive genomic context. ₉ The 50-organism set includes 16 non-Fitness-Browser organisms. Among them are P. aeruginosa PAO1 (#1, 3,713 OGs), M. tuberculosis (#6, Actinomycetota) and C. jejuni (#48, Campylobacterota). The highest-ranked addition is therefore a Pseudomonadota strain, not a member of a missing phylum. ₉ The report’s limitations section qualifies this gain. It restates that the 48 Fitness Browser organisms are 77% Pseudomonadota (37/48) and that adding 25 non-Fitness-Browser organisms expands phylum coverage from 4 to 6. However, non-Fitness-Browser OG coverage is estimated at the genus level, counting an OG as present if any species in the genus has it, which overestimates individual organism coverage. These organisms also lack Fitness Browser condition profiling, so their dark genes can be assigned only to broad phenotypic screens, not to condition-specific experiments. ₉ The report’s figure fig39_extended_covering_set.png compares the Fitness-Browser-only and 73-organism covering sets, including marginal coverage, phylum diversity and kingdom-level gap coverage. ₉

Fitness-panel limits in downstream analyses

The 48-organism Fitness Browser panel also limits analyses that join fitness data to wider genomic collections. The metabolic capability–dependency analysis notes that fitness data for 48 organisms cover only a small fraction of the 293,000 genomes with pathway predictions, so its results may not generalize to all taxonomic groups. ₁₂ The prophage–AMR co-mobilization analysis studied whether prophages co-mobilize AMR (antimicrobial resistance) genes. It found that the 48 model organisms with RB-TnSeq data do not overlap well with the gtdb pangenome species it analyzed. Its hypothesis H3 therefore could not be tested and remains an open question. ₁₃ These are limits on generalizability and an untestable comparison. They are not negative biological findings, and they do not establish any statistical null result. _12,13

Panel composition can also omit ecologically dominant taxa. The lab–field ecology report identifies rhodanobacter as dominating contaminated Oak Ridge wells, citing Carlson et al. (2019), and notes that this genus is not in the Fitness Browser. ₁₄ The dominant genus of these oak-ridge-field-research-center wells therefore has no panel fitness data. ₁₄

Gene-level searches inside the panel are bounded by which organisms it contains. Among the 48 Fitness Browser organisms, manxyz by gene name is present only in E. coli K-12 (Keio collection), which has fitness data for 168 conditions. No Klebsiella strain in the Fitness Browser carries ManXYZ by gene name. ₁₅ In the same panel, 7 genes carry PF13455, the snipe-defense-system nuclease domain, spread across 6 organisms. Only methanococcus-maripaludis has the full two-domain SNIPE architecture (PF13250 + PF13455). ₁₅ phagefoundry has ManYZ annotations in Klebsiella (28 ManZ proteins, 4,619 PTS_EIIC Pfam hits). However, it holds genome annotations only, with no fitness or TnSeq tables, so Klebsiella-specific fitness remains unmeasured there. The report states that this would require new mutant libraries in natural SNIPE-carrying strains such as K. pneumoniae NCTC9140 or HS11286. ₁₅

The metal-fitness-atlas shows the same gap between scoring and validation. Its experiment coverage is broader than its gene-fitness data: 31 organisms have metal data, but only 24 have fitness matrices. ₁₆ The atlas scores 27,702 species but cannot validate predictions for organisms without RB-TnSeq libraries. ₁₆ Its proposed libraries are prospective, and none is reported as completed. Acidithiobacillus ferrooxidans, described as the #1 bioleaching organism, is at the 77th percentile after normalization. The report states that no TnSeq data exist for any Acidithiobacillus species and proposes RB-TnSeq under Fe, Cu, Co, Ni and Zn. Leptospirillum ferrooxidans, at the 91st percentile after normalization, is the highest-scoring bioleaching genus after normalization; RB-TnSeq under Fe(II) oxidation and acid conditions is proposed for it. Geobacter sulfurreducens, a model metal-reducing bacterium at the 53rd percentile after normalization, would be screened under Fe(III), Mn(IV) and U(VI) reduction conditions. ₁₆

Within-panel data joins add a technical layer of attrition. The metal-specificity analysis could not process 7 of 31 metal-tested organisms (ANA3, Dino, Keio, MR1, Miya, PV4, SB2B). Their metal-important gene locusIds did not match the fitness-matrix index format. ₁₇ The 24 included organisms account for 7,609 of 12,838 metal-important gene records (59.3%). The report describes the exclusion of 7 organisms and 5,229 gene records as 40.7% gene attrition. ₁₇ The report states that the excluded organisms are taxonomically diverse and that their absence is not expected to introduce systematic bias, but it does not test this empirically. It also notes that the excluded organisms include Keio (E. coli), MR1 (Shewanella) and ANA3, all important model organisms, and that the excluded genes may have different specificity profiles. ₁₇

Tensions

The observed association between perturbation coverage, annotation, and pangenome-core status could reflect biological conservation, technical difficulty in constructing or measuring mutants, misclassified gene fragments, or some combination of these factors. ₁ The report does not establish which mechanism causes the missing-data pattern, so the association should not be interpreted as proof that evolutionary conservation directly determines deletion-mutant recovery. ₁

The ADP1 collection indicates that successful coverage is enriched among core genes, whereas the broader trade-off analysis shows that core genes can carry substantial laboratory burden and condition-specific costs. This is not a direct contradiction because the studies measure different outcomes, but it creates a boundary on interpretation: core enrichment in a collection cannot by itself establish either low burden or universal essentiality. _1,3 The Fitness Browser result that singleton genes were near-neutral under tested conditions creates a related interpretive tension: neutrality may represent genuine lack of measured effect or inadequate transposon coverage, and the analysis does not distinguish these explanations. ₂

The costly–dispensable analysis adds a related boundary rather than resolving this tension: its classification of burden used max_fit > 1 in any experiment, so a single experiment can classify a gene as burdensome and the result is sensitive to fitness-data noise; its 90% identity DIAMOND threshold may also miss recently acquired genes with low sequence similarity. ₆ Thus, mobile-element enrichment and poor annotation narrow the biological interpretation of costly non-conserved genes toward horizontally acquired DNA, but do not establish that such genes are systematically absent from perturbation collections. ₆

The functional dark-matter report’s organism-coverage figures are not internally consistent, and the report does not reconcile them. First, the 78% Gammaproteobacteria share of top-500 dark genes does not match the 417 genes in its own table. Second, the taxon table lists Firmicutes with 1 organism (the BFirm strain) and Bacteroidetes with 5 organisms. The phylum-level section instead states that Bacillota have zero Fitness Browser representation and lists Bacteroidota (4). Third, B. subtilis is ranked first for a new library on phylogenetic grounds, yet the OG covering set did not select it because its OGs were already covered by Pseudomonadota. This reflects different optimization goals (native-host and phylum depth versus OG breadth) rather than a contradiction in data. ₉

The functional dark-matter report is also inconsistent about whether its covering set already includes non-Fitness-Browser organisms. An earlier NB11b-stage passage presents incorporating them into the Route B covering set as prospective. That passage says incorporation requires a Spark query to map each non-Fitness-Browser genome to dark-gene root_ogs via the pangenome. ₉ The future-directions section likewise states that the current covering sets draw only from 48 Fitness Browser organisms, which are 77% Pseudomonadota. It proposes adding 25 organisms from Bacillota (6), Actinomycetota (2), Campylobacterota (2) and additional Pseudomonadota (15), and notes that a pangenome mapping query is required. ₉ Yet the NB11c results report a 73-candidate covering set that was already run, and the artifact inventory lists an extended covering-set figure. The NB11c and limitations passages also describe the 25 additions as coming from Bacillota, Actinomycetota and Campylobacterota, without mentioning the 15 Pseudomonadota. The report does not reconcile these passages, and this page does not resolve which one reflects its final state. ₉

The same report states that every gene in its catalog has measured fitness phenotypes from controlled lab experiments. This is in tension with its later statement that essential genes have zero genefitness rows and no fitness scores beyond the essentiality call. The first claim holds only if an essentiality call counts as a measured fitness phenotype, and the report does not say that it does. ₉

Source

The underlying analysis is summarized in ADP1 Deletion Collection Phenotype Analysis. ₁

The integrated conservation–fitness evidence is summarized in Core Gene Paradox — Why Are Core Genes More Burdensome?. ₃

The broader Fitness Browser comparison is summarized in Fitness Effects vs Conservation — Quantitative Analysis. ₂

The costly–dispensable gene analysis is summarized in The 5,526 Costly + Dispensable Genes. ₆

Organism-level panel coverage is summarized in Functional Dark Matter — Experimentally Prioritized Novel Genetic Systems and The Pan-Bacterial Essential Genome. _9,11

Panel-limit evidence from downstream analyses is summarized in Condition-Specific Respiratory Chain Wiring in ADP1, Fitness Modules × Pangenome Conservation, Pan-Bacterial Metal Fitness Atlas, Metal-Specific vs General Stress Genes, SNIPE Defense System in the KBase Data Lakehouse Pangenome, Lab Fitness Predicts Field Ecology at Oak Ridge, Metabolic Capability vs Metabolic Dependency and Prophage-AMR Co-mobilization Atlas. _{5,8,16,17,15,14,12,13}

Open Directions

• Re-sequence and re-annotate the 272 TnSeq-dispensable genes lacking deletion-collection growth data, then test whether annotation updates explain their missing perturbation phenotypes. ₁
• Compare deletion-mutant construction success and phenotype completeness against gene length, annotation status, pangenome frequency, and pseudogene indicators using a multivariable logistic model to separate technical and evolutionary predictors of coverage. ₁
• Replace the species-level core/accessory labels with population-level pangenome frequencies and test whether the coverage association remains after controlling for gene fragments and uncertain classifications. ₁
• Add the 316 genes with incomplete data and the 499 excluded essential genes where technically possible, then repeat condition-specific fitness analyses to quantify how collection composition changes the inferred essentiality landscape. ₁
• Stratify coverage analyses by functional category and condition, then test whether the observed core-gene burden pattern persists after accounting for which genes are recoverable in the deletion collection. ₃
• Reassess singleton neutrality with coverage-aware models and perturbation assays spanning non-rich-media conditions, testing whether missing transposon coverage or unmeasured ecological conditions explain the apparent lack of fitness effects. ₂
• Test whether the costly+dispensable mobile-element and annotation profile predicts failed mutant construction or incomplete phenotype coverage after controlling for gene length, orthology breadth, and organism. ₆
• Re-assign all 48 Fitness Browser organisms to phyla under one taxonomy, then recompute the top-500 dark-gene shares per taxon. This would settle the report’s inconsistent Gammaproteobacteria, Firmicutes/Bacillota and Bacteroidetes/Bacteroidota counts. ₉
• Build RB-TnSeq or CRISPRi libraries in an Actinomycetota host such as M. tuberculosis H37Rv or S. coelicolor A3(2). Then test whether kingdom-level dark-gene OGs show fitness phenotypes concordant with their Proteobacterial orthologs, closing the native-host gap that the Fitness Browser panel cannot address. ₉
• Compute co-fitness from raw genefitness data for Ralstonia and other organisms lacking pre-computed co-fitness. Then test whether adding them changes co-fitness/co-inheritance conclusions drawn from the pre-computed subset. ₁₀
• Construct or recover an NDH-2 (ACIAD_RS16420) deletion mutant and measure its growth on glucose and the other standard carbon sources. This would directly test the compensation prediction that the current growth data cannot address. ₅
• Repair the locusId mapping for ANA3, Dino, Keio, MR1, Miya, PV4 and SB2B, then rerun the metal-specificity analysis. This would test empirically the report’s expectation that their exclusion introduces no systematic bias. ₁₇
• Replace genus-level OG membership with strain-level genome mapping for the 25 non-Fitness-Browser organisms, then rerun the covering set. This would measure how much genus-level estimation inflates coverage and whether P. aeruginosa PAO1 keeps its top rank. ₉
• As exploratory work, add a Rhodanobacter strain to the RB-TnSeq panel. The genus dominates contaminated Oak Ridge wells but is absent from the Fitness Browser. Screen it under field-relevant conditions to see which genes, if any, show condition-specific fitness. ₁₄
• Build an RB-TnSeq library in Acidithiobacillus ferrooxidans under Fe, Cu, Co, Ni and Zn, as the metal atlas proposes. Then test whether the atlas’s predictions for this species agree with measured laboratory fitness. ₁₆