Core classification indicates broad pangenome conservation, but it does not guarantee that a gene has substantial representation in characterized sequence space or a well-defined molecular function. The AlphaFold MSA Depth as a Lens on the Bacterial Annotation Gap examines this distinction by relating pangenome gene-cluster classes to AlphaFold multiple-sequence-alignment (MSA) depth, where MSA depth is the number of homologous sequences represented in the alignment. 1 Evidence from Core Gene Paradox — Why Are Core Genes More Burdensome? refines this distinction: core genes can be functionally active and conditionally costly in laboratory assays while still being structurally or functionally uncharacterized. 2 The Gene Conservation, Fitness, and the Architecture of Bacterial Genomes supports this interpretation across 194,216 protein-coding genes from 43 bacteria: essential genes were 82% core, but genes neutral in every experiment were still 66% core, showing that conservation and measured fitness importance are related but non-equivalent. 3 The new Fitness Effects vs Conservation — Quantitative Analysis supports this distinction across approximately 194,000 genes from 43 bacteria: essential genes were 82% core, whereas always-neutral genes were 66% core, so laboratory fitness importance was a statistically robust but weak predictor of conservation. 4 Evidence from Fitness Modules × Pangenome Conservation supports the conservation side of this distinction: co-regulated fitness-module genes were 86.0% core versus 81.5% for all genes, although this enrichment does not establish that the genes are functionally characterized. 5
Core Genes Usually Have Greater Sequence Representation
Among gene clusters bridged to AlphaFold, core clusters had a median MSA depth of 15,308, compared with 5,299 for auxiliary+singleton clusters and 5,527 for auxiliary non-singleton clusters. 1 This corresponds to a 2.89× higher median for core genes than auxiliary+singleton genes and a 2.77× higher median than auxiliary non-singleton genes. 1 The 10th-percentile MSA depth was 334 for core genes versus 25–32 for accessory classes. 1
These distributions are consistent with a broad class-level annotation advantage for core genes, but they do not eliminate poorly represented core proteins. Among bridged clusters, 415,733 core clusters (1.6%) had MSA depth below 10, compared with 245,002 auxiliary non-singleton clusters (4.6%) and 392,959 auxiliary+singleton clusters (5.5%). 1 The fitness analysis supports treating this annotation advantage as distinct from functional importance: core genes were more burdensome in several functional categories, but this does not establish that their molecular functions are characterized. 2 The broader synthesis refines this point: core genes were 1.78x more likely than accessory genes to have strong condition-specific phenotypes, so detectable laboratory activity does not make their molecular roles fully characterized. 3 The new fitness analysis supports this refinement: genes with strong condition-specific effects were 77.3% core, versus 70.3% core among genes without specific phenotypes (OR=1.78, p=1.8e-97), and core genes were more likely than auxiliary genes to show benefits as well as costs, indicating measured activity without implying complete annotation. 4 The module analysis refines this interpretation by showing that conservation also extends to co-regulated units: 577 of 974 mapped modules (59%) were >90% core, while 38 module families had <50% core genes. 5 A separate ICA synthesis identified 1,116 co-regulated fitness modules across 32 organisms, with 86% core genes versus an 81.5% baseline (odds ratio 1.46, p=1.6e-87), supporting coordinated conservation without implying complete functional annotation. 3
A Conserved Low-MSA Core Subset
The analysis identified 415,603 distinct core clusters with MSA depth below 10 across 14,768 species clades. 1 Their mean MSA depth was 4.57 and their median MSA depth was 4.0. 1 Of these clusters, 286,439 (68.9%) were annotated as hypothetical, 137 (0.033%) had an EC annotation, and 346 (0.083%) were mapped to KEGG. 1 EC and KEGG annotation were therefore rare in this subset. This census count of 415,603 distinct clusters is reported separately from the 415,733 (1.6%) very-low-depth core clusters in the bridged-class table, and the two figures should not be treated as interchangeable. 1
The low-MSA core subset therefore had a 68.9% hypothetical-protein rate, whereas all bridged core genes had a 3.8% hypothetical-protein rate. 1 This contrast supports the interpretation that pangenome conservation and functional characterization measure different properties: a gene can be retained broadly while remaining distant from experimentally or computationally characterized sequence space. 1 The trade-off analysis supports the complementary claim that conservation can reflect active, condition-dependent function rather than annotation completeness: 25,271 genes were classified as true trade-off genes, and trade-off genes were 1.29 times more likely to be core than non-core genes (odds ratio 1.29, p=1.2e-44). 2 The synthesis further identifies 28,017 genes that were both costly in the laboratory and conserved, but treats this category as evidence for possible purifying selection rather than direct measurement of natural selection. 3
The report treats these proteins as candidates for experimental structural characterization, rather than as experimentally validated examples of novel function. 1 The prioritization hypothesis is strongest for proteins that are conserved by pangenome classification but have very shallow alignments and little EC, KEGG, or domain-level annotation. 1 The fitness analysis refines this prioritization logic: novel singleton genes showed near-zero mean fitness in tested laboratory conditions, but that apparent neutrality may reflect poor transposon coverage rather than absence of biological function. 4
The top-ranked low-MSA core proteins at msa_depth = 1 came primarily from poorly characterised marine and soil bacteria (Oceanicoccus, Dwaynesavagella, CAILRJ01). Non-hypothetical entries at msa_depth = 1 included an RNA polymerase ω-subunit family protein and an FXSXX-COOH domain protein. The report describes both as conserved structural components with no solved structure for these specific lineages. These are illustrative named examples, not a validated functional set. 1
Functional Annotation Tracks MSA Depth
Across 38,051,842 gene cluster–UniProt pairs, MSA depth and domain-hit count had a Spearman correlation of ρ = 0.7563. 1 Mean domain hits increased from 0.59 for clusters with MSA depth below 10 to 10.83 for clusters with MSA depth of at least 10,000. 1 Mean distinct InterPro families increased from 0.059 to 4.601 across the same MSA-depth range. 1 The monotone relationship held within core, auxiliary non-singleton, and auxiliary+singleton classes, with core genes showing slightly higher domain richness per MSA bin than accessory genes at equivalent depth. 1
This relationship supports MSA depth as a useful prioritization signal for annotation scarcity, but it does not establish that sequence abundance causes functional knowledge or that shallow MSA depth proves biological novelty. 1 Of all 132,531,501 gene clusters, 111,035,431 (83.8%) had at least one InterProScan domain annotation, whereas only 38,804,903 (29.3%) had a real UniProt accession and 38,051,842 (28.7%) bridged successfully to AlphaFold MSA depths. 1
The fitness results refine interpretation of the shallow-MSA subset by showing that laboratory burden is also condition- and function-dependent: core genes were more burdensome in Protein Metabolism, Motility, and RNA Metabolism, whereas Cell Wall genes showed the reverse pattern, with non-core genes more burdensome. 2 The synthesis quantifies related category-specific contrasts: Motility and chemotaxis had a +7.8pp core-burden excess, RNA metabolism a +12.9pp excess, and protein metabolism a +6.2pp excess, while cell-wall genes again showed the reverse pattern. 3 These measurements indicate functional activity or trade-offs, not verified molecular annotation or direct evidence of natural selection. 2
Two Annotation-Gap Layers
The evidence distinguishes two related annotation-gap layers. 1 First, an MSA-depth-driven gap affects genes in every pangenome class, because domain richness decreases sharply in shallow-MSA bins. 1 Second, a pangenome-class gap is visible in the higher hypothetical-protein rates of accessory classes: 13.8% among 7,095,643 auxiliary+singleton clusters (979,300 hypothetical), 11.6% among 5,384,900 auxiliary non-singleton clusters (622,748 hypothetical), and 3.8% among 25,571,299 core clusters (979,912 hypothetical). 1 The same bridged-class table reports median MSA depths of 15,308 for core, 5,527 for auxiliary non-singleton and 5,299 for auxiliary+singleton clusters. It reports 10th-percentile depths of 334, 32 and 25 and 90th-percentile depths of 19,500, 19,192 and 19,203. Its very-low-depth (below 10) counts are 415,733 (1.6%), 245,002 (4.6%) and 392,959 (5.5%). The 415,733 table count is distinct from the 415,603 count in the low-MSA core census. 1 Chi-square tests were reported as overwhelmingly significant (χ² > 500,000; p ≈ 0), with odds ratios of 0.25 for core versus auxiliary+singleton clusters and 0.31 for core versus auxiliary non-singleton clusters. 1 The report says the MSA-depth-driven gap explains the 18× span in domain hits. It describes the pangenome-class gap as independent of MSA depth and attributes it likely to horizontal transfer, rapid evolution and taxonomically narrow distribution. These are offered explanations rather than directly tested mechanisms. 1
The class-level comparison is dominated by the higher average MSA depth of core genes, while the low-MSA core subset reveals a severe annotation gap within the core itself. 1 This distinction refines pangenome-integration by showing that core status should not be treated as a proxy for complete annotation. 1 It also supports interpreting the costly-and-conserved subset as a functional-prioritization signal rather than as evidence that every conserved gene is already understood: the trade-off study reported 28,017 Costly + Conserved genes, but its selection interpretation rests on laboratory fitness and conservation patterns rather than direct natural-selection measurements. 2 The synthesis additionally reports 5,526 genes that were costly and dispensable, treating them as candidates for mobile-element association, recent acquisition, or gene loss rather than as established examples of any one process. 3
The module analysis refines the class-level interpretation by finding no relationship between module-family breadth and conservation (Spearman ρ=-0.01, p=0.914), redirecting attention from the scope of a module family toward gene-level conservation. 5 Because ICA requires measurable fitness variation from transposon-insertion data, zero essential genes appeared in the modules; this supports treating module conservation as complementary evidence rather than a complete account of conserved functional units. 5 The independent synthesis similarly found no module-family breadth gradient (rho=-0.01 and p=0.91), supporting the conclusion that coordinated conservation does not by itself resolve annotation status. 3
Essential Genes Split by Conservation Status
Essentiality in the conservation-versus-fitness analysis is putative. The analysis identified 27,693 putative essential genes (18.6% of 148,826 protein-coding genes across 33 organisms) as genes lacking fitness data. The report calls this definition an upper bound. Genes may lack transposon insertions because they are short, sit in low-complexity regions or lie at scaffold edges, not necessarily because they are essential. RB-TnSeq (random barcode transposon sequencing) essentiality is also defined under the specific library construction conditions, so genes essential only under stress would not be captured. 6 Within this putative-essential set, annotation tracked conservation status. The 22,751 essential-core genes were 41.9% enzymes and 13.0% hypothetical. The 3,683 essential-auxiliary genes were 13.4% enzymes and 38.2% hypothetical. The 1,259 essential-unmapped genes were 18.2% enzymes and 44.7% hypothetical. The 124,744 non-essential genes were 21.5% enzymes and 24.5% hypothetical. That tabulated non-essential count is presented separately from the report’s 148,826-protein-coding-gene denominator, and this page does not reconcile the two. 6 This supports the class-level annotation advantage of core genes described above. It also refines the paradox: putatively essential genes outside the core are the least characterized. 6
Essential-core genes were 41.9% enzymes and 87% known-function. Relative to non-essential genes, they were enriched in Protein Metabolism (+13.7 percentage points), Cofactors/Vitamins (+6.2%), Cell Wall (+3.9%) and Fatty Acid biosynthesis (+3.1%). They were depleted in Carbohydrates (-7.9%), Amino Acids (-5.6%) and Membrane Transport (-4.0%). The report reads the depleted functions as conditionally important rather than universally essential, but that reading is an inference. 6
The central digest supports this gradient. It describes essential-core genes (41.9% enzymes, 13.0% hypothetical) as the well-characterized metabolic backbone. It reports essential-auxiliary genes as 13.4% enzymes and 38.2% hypothetical. It describes essential-unmapped genes, at 44.7% hypothetical, as prime targets for functional discovery. 7
As secondary literature context, the report cites Hutchison et al. (2016), which describes a minimal Mycoplasma genome of 473 genes in which 149 essential genes (31%) had unknown function. The report compares this with its own 44.7% hypothetical essential-unmapped category. 6
The digest further reports 7,084 essential genes with no orthologs in any other Fitness Browser organism. Of these, 58.7% are annotated as hypothetical proteins, versus 8.2% of universally essential genes. 7 This suggests the hypothesis that taxonomically restricted essential genes are among the least characterized. However, absence of orthologs among Fitness Browser organisms reflects that database’s organism sampling and does not by itself establish taxonomic restriction. 7
In acinetobacter-baylyi-adp1, essential genes were more annotation-rich than dispensable genes. COG (Clusters of Orthologous Groups) assignments covered 33% of essential versus 5% of dispensable genes, and KEGG KO assignments covered 92% versus 53%. 8 This single-organism result supports the same direction. The ~8% of essential genes without KO assignments are a research lead for novel essential functions, not an established result. 8
Broadly Distributed Dark Ortholog Groups
The central digest reports that 55.9% of dark-gene ortholog groups (OGs) are kingdom-level, present across multiple phyla. In a 27,690-species pangenome analysis, dark-OG species counts ranged from 1 to 27,482 (median 135). The digest names COG0468, COG0443 and COG0491 as top-ranked unknowns that are present in virtually all sequenced bacterial genomes yet have zero functional evidence. 7 This refines the paradox by extending it beyond species-clade core status: functional darkness also affects pan-bacterially distributed gene families. 7
The 11,774 root OGs were classified by the narrowest taxonomic rank containing all member species: kingdom 55.9%, class 11.0%, family 10.5%, genus 6.9%, mobile 6.5%, phylum 4.8%, order 3.9%, species 0.5%. A gene-level count was also reported, covering 57,011 dark genes including 19,014 FB-only fallback assignments: kingdom 51.4%, species 29.8%, genus 6.2%, class 4.0%, family 3.5%, mobile 1.9%, phylum 1.9%, order 1.3%. 9
A conservation × ignorance importance score ranked all 11,774 dark OGs, with kingdom-level true knowledge gaps scoring highest. Route B prioritizes broadly conserved genes with zero functional evidence. 9 The top three entries were cog0468 (27,427 species, 142 phyla, importance 23.9), cog0443 (27,279 species, 142 phyla, importance 23.9) and cog0491 (27,393 species, 142 phyla, importance 23.9). The last of these was assigned zero evidence across all inference layers. COG0172 (27,431 species) was also cited among the leading OGs, but as a weak lead rather than a true knowledge gap. 9
Figure fig35_top_knowledge_gaps.png shows the top 25 true-knowledge-gap OGs by pangenome species count. 9
The digest names COG0468, COG0443 and COG0491 as the top-ranked unknowns. It describes them as present in virtually all sequenced bacterial genomes yet with zero functional evidence. 7
The report cautions that these broadly conserved, uncharacterized OGs are invisible to its fitness-based Route A. They may lack condition-specific fitness effects, for example because they are essential under all conditions. 9 This refines the proposed fitness joins in Open Directions: an absent condition-specific signal would not show that a conserved dark gene lacks function. 9
Among the most widespread top-500 dark gene clusters (present in 15+ pangenome species), the most common domain families were TauE (sulfonate export, 7 clusters), DUF444 (5 clusters), Cu-oxidase_4 (5 clusters), EamA (4 clusters, including the #1 candidate) and DUF484/DUF971/DUF934 (3+ clusters each). 9 This supports the distinction drawn above between domain-level annotation and characterized function. 9
A plant-microbiome analysis gives a counterpoint. Phase 2 annotation found that all 50 OGs examined were well-characterized proteins, none hypothetical. Despite this, 19/50 (38%) had characterized domains but no clear plant-interaction mechanism and remain candidates for experimental characterization. 10 This refines the paradox: full annotation does not resolve a gene’s context-specific biological role. 10
Scope and Interpretation
Only 38,051,842 of 132,531,501 total gene clusters (28.7%) bridged successfully to AlphaFold MSA depths, while 70.7% lacked UniRef100 IDs or had UniParc-only identifiers without an AlphaFold entry. 1 Because the analysed subset is biased toward better-studied organisms, the annotation gap among the remaining 70.7% is likely larger, although that inference was not directly measured. 1
MSA depth was obtained for each cluster’s representative sequence, so the analysis did not measure sequence-level variation within each cluster. 1 The 293K genomes were not phylogenetically balanced, and common taxa such as Pseudomonas and E. coli were over-represented. 1 The reported Spearman ρ = 0.7563 was calculated on the full 38,051,842-pair dataset without subgroup stratification. 1 These limitations mean that the low-MSA core set is a defensible discovery and prioritization set, not a complete census of conserved functionally dark genes. 1 The trade-off study adds a related limitation: Fitness Browser conditions are biased toward experimentally convenient rather than ecologically representative conditions, so costly core genes cannot by themselves be taken as evidence of selection in nature. 2 The synthesis likewise states that laboratory measurements capture gene-maintenance costs while pangenome conservation captures evolutionary retention, not direct natural-environment fitness. 3 The new fitness analysis supports this caveat by noting that its assays were biased toward rich media and standard stresses, that measurements used single-gene knockouts without epistatic interactions, and that the 43-bacterium coverage was primarily Proteobacteria. 4 The module-conservation estimate is likewise limited to a 29/32 organism subset, because three organisms lacked pangenome links. 5
In the AlphaFold analysis, core status means presence in the majority of genomes within each of 14,768 species clades, not universality across all bacteria. 1 The report reads core status as implying strong purifying selection and likely important biological function. It reads MSA depth < 10 as a genuinely unprecedented fold. Both readings are interpretations, not direct measurements. 1
Tensions
The report flags that core clusters’ low overall hypothetical rate appears to contradict its hypothesis H2 as literally stated. It argues that the low-MSA census resolves the tension: within core genes, the MSA-depth < 10 subset has a 68.9% hypothetical rate, versus 3.8% for all core genes. 1 This resolution comes from the report itself and has not been independently tested. 1
The report’s interpretation says the bottom 10% of core genes still have MSA depth ≥ 334, while the bottom 10% of accessory genes have depth ≤ 25–32. 1 This contradicts the same report’s table, which lists 415,733 (1.6%) core clusters with depth below 10. A 10th-percentile value of 334 marks the upper edge of the lowest tenth, not a floor beneath it. This page treats the statement as an internal contradiction, not as a finding about the lowest-depth core genes. 1
Open Directions
- Join the 415,603 low-MSA core clusters to kescience-fitnessbrowser measurements and use condition-specific fitness analysis to ask whether structurally isolated core proteins have essential or condition-dependent phenotypes. 1
- Recompute MSA depth and InterPro-family richness after stratifying the 38,051,842 bridged pairs by pangenome class and organismal lineage, using Spearman correlations and balanced taxon sampling to ask whether the ρ = 0.7563 relationship is consistent across taxa. 1
- Compare representative-sequence MSA depths with member-level sequence diversity for the 415,603 low-MSA core clusters, using within-cluster pangenome sequence analysis to ask whether representatives conceal deeper or shallower support among cluster members. 1
- Prioritize the 286,439 hypothetical low-MSA core clusters for structure prediction and experimental characterization, using the 137 EC-annotated and 346 KEGG-mapped clusters as annotated comparators to ask which structural features predict recoverable function. 1
- Extend the bridge analysis to the 70.7% of clusters without AlphaFold MSA depth, using alternative sequence-similarity and domain resources to ask whether the unbridged population contains a larger conserved annotation gap. 1
- Intersect low-MSA core clusters with the 28,017 Costly + Conserved genes and test their fitness across ecologically relevant conditions, asking whether shallow sequence representation predicts condition-dependent burden or essentiality. 2
- Test whether low-MSA core clusters are overrepresented in the 38 accessory module families, and whether module membership or module-family breadth predicts annotation recovery after controlling for pangenome class. 5
- Intersect the low-MSA core set with the 5,526 costly-and-dispensable genes and test mobile-element, acquisition, and gene-loss hypotheses using genomic-context and phylogenetic analyses. 3
- Intersect low-MSA core clusters with the fitness-effect categories from the 43-bacterium dataset and test whether shallow MSA representation predicts condition-specific burden, benefit, or apparent neutrality after controlling for transposon coverage. 4
- Map the 415K paradox clusters onto the gtdb phylogeny, as the report proposes, to identify which bacterial phyla and families carry the highest densities of conserved-yet-novel proteins. The report does not give phylum- or family-level densities. 1
- Test COG0468, COG0443 and COG0491 against essentiality calls rather than condition-specific fitness, asking whether Route A misses them because they are essential across conditions. 9
- Compare annotation status among essential-core, essential-auxiliary and essential-unmapped genes against the low-MSA core set, asking whether the 44.7% hypothetical essential-unmapped category overlaps with shallow-MSA proteins. 6