Genomic annotation inference depends not only on genes present in a pangenome but also on whether they bridge to reference databases, structurally characterized sequence space, and published functional knowledge. Here, AlphaFold multiple-sequence-alignment (MSA) depth—the number of homologous sequences represented in an alignment—provides a quantitative proxy for sequence-space representation and annotation richness across bacterial gene clusters. The analysis joined kbase-ke-pangenome, Bakta annotations, interproscan, and kescience-alphafold through 38,051,842 gene cluster–UniProt pairs. 1 The complete report is available at AlphaFold MSA Depth as a Lens on the Bacterial Annotation Gap.
Core evidence
Across 38,051,842 gene cluster–UniProt pairs, MSA depth and domain-hit count were strongly associated (Spearman ρ = 0.7563). 1 In the report’s binned table, whose cluster-count column is labeled core clusters (415,733 in the < 10 bin and 15,993,075 in the ≥ 10,000 bin), mean domain hits increased from 0.59 at MSA depth < 10 to 10.83 at MSA depth ≥ 10,000, while mean distinct InterPro (IPR) families increased from 0.059 to 4.601 across the same bins—an 18× span in mean domain hits. 1 The relationship was monotonic within core, auxiliary non-singleton, and auxiliary+singleton gene classes, with core genes showing slightly higher domain richness than accessory genes at equivalent MSA depth; this within-class pattern is described from binned trends, and no class-specific correlation coefficients were reported. 1 This supports the interpretation that sequence-space representation, rather than pangenome class alone, is a major predictor of available functional annotation. 1
The analysis began with 132,531,501 gene clusters; 38,804,903 (29.3%) had a real UniProt accession and 38,051,842 (28.7%) bridged successfully to AlphaFold MSA depths. 1 The remaining 70.7% lacked UniRef100 identifiers or had UniParc-only identifiers without an AlphaFold entry, so the analysed subset was biased toward better-studied organisms. 1 This suggests the hypothesis that the unrepresented 70.7% contains an even larger annotation gap, although that gap was not directly measured. 1 Discoveries refines this limitation: among 17.6M distinct UniRef50 IDs produced by Bakta, only 33.3% existed in the KBase Data Lakehouse UniProt identifier table, showing that identifier-table incompleteness can also interrupt downstream annotation bridges independently of biological sequence sparsity. 2
Reference coverage is not equivalent to domain-annotation coverage: 111,035,431 gene clusters, or 83.8% of the total, had at least one InterProScan domain annotation, with mean hits of 7.5 and mean distinct IPR families of 3.3, whereas the report’s dataset-coverage section gives 29.3% with a real UniProt accession and 28.7% bridging successfully to AlphaFold MSA depths; elsewhere the report uses 29.3% for the bridged fraction, and this inconsistency is retained as reported rather than treating the two percentages as interchangeable. 1 This contrast indicates that different resources expose different layers of the annotation landscape: a gene can have an InterProScan assignment without a successful AlphaFold bridge, while MSA depth measures how broadly its sequence space is represented. 1 A corpus-wide comparison supports and refines this complementary-resource interpretation: across 132.5M gene clusters, eggNOG had higher COG, KEGG, and Pfam coverage, whereas Bakta had higher GO coverage, product descriptions, and unique UniRef50 links; their union raised any-functional-annotation coverage to 77.3% and Bakta rescued 11.2M clusters among 39.2M missed by eggNOG. 2 Thus, low coverage in one resource should not be treated as equivalent to absence of function.
The AlphaFold MSA report also cites outside literature that supports this layering even inside well-characterised references: Schaeffer et al. (2026) classified AFDB Swiss-Prot entries and found > 100,000 domains with no Pfam mapping. 1 This is cited literature, not a measurement made by the project. It indicates that structure-based classification can expose annotation gaps even where reference coverage is high, but it does not quantify that gap in the KBase pangenome. 1
Pangenome structure and research-attention bias
Core gene 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 The core median was therefore 2.89× higher than auxiliary+singleton and 2.77× higher than auxiliary non-singleton clusters; the 10th-percentile depth was 334 for core genes versus 25–32 for accessory genes. 1 Hypothetical-protein rates were 3.8% in core clusters, 11.6% in auxiliary non-singleton clusters, and 13.8% in auxiliary+singleton clusters. 1 Chi-square tests were overwhelmingly significant (χ² > 500,000; p ≈ 0), with odds ratios of 0.25 for core versus auxiliary+singleton and 0.31 for core versus auxiliary non-singleton. 1 These results support pangenome-integration by showing that pangenome class is associated with both sequence-space depth and annotation completeness. 1
The cross-species COG analysis supports and refines this pangenome interpretation: across 32 species spanning 9 phyla and 357,623 genes, core functions were concentrated in conserved metabolic and housekeeping categories, whereas novel or singleton genes were enriched in mobile-element, defense, and unknown-function categories. 3 Novel or singleton genes showed a +10.88% enrichment in COG L (mobile elements), a +2.83% enrichment in COG V (defense mechanisms), and a +1.64% enrichment in COG S (unknown function); the COG L and V patterns were 100% consistent across species. 3 The report therefore provides cross-species functional evidence adjacent to, but not a replacement for, the MSA-depth result: accessory and singleton genes have both lower measured sequence-space representation and a functional distribution enriched for categories plausibly associated with mobility, defense, and niche-specific novelty. 1,3 Because the COG analysis is category-level and the MSA analysis is an annotation-coverage analysis, these results do not establish that mobile transfer caused low MSA depth or hypothetical status. 3
The COG report lists translation, nucleotide, coenzyme, amino-acid, and energy-production categories under the heading that core genes are consistently enriched in them. However, it gives negative enrichment values for these categories and labels COG J (translation), at -4.65%, the strongest depletion. Because the sign convention is ambiguous in the source, this page does not say whether core genes are enriched or depleted in these categories relative to the novel or singleton class. 3 This refines the description of core genes as a conserved functional engine: broad COG distributions identify the functions represented by core and novel classes, whereas MSA depth measures how richly those functions are connected to characterized sequence space. 3 Composite COG assignments were counted once per gene and retained as biologically meaningful multifunctional categories; the LV composite, combining mobile and defense functions, was enriched by +0.34% with 76% consistency. 3 This cautions against treating functional-category complexity as annotation noise when interpreting links between pangenome class and annotation richness. 3
Species-level variation in hypothetical-gene fractions points the same way: in the functional dark matter project, some species have >35% hypothetical genes while others have <15%, and the report attributes this spread to differences in annotation depth rather than true functional content. 4 This supports the reading that hypothetical status partly reflects reference and annotation coverage, but the attribution is the report’s interpretation rather than a test against MSA depth or reference representation. 4
PaperBLAST adds a distinct literature-coverage layer rather than a direct MSA measurement. Homo sapiens accounts for 46.7% of gene-paper records, the top five organisms account for 72.8%, and the top 1,000 of 20,723 organisms capture 94.2% of the literature. 5 Among 841K genes with any text-mined paper link, 551K (65.6%) had exactly one paper, and only 6% of genes (50K) accounted for 57.7% of all gene-paper links. 5 In bacterial literature, the top 100 organisms capture 44.3%, with leading examples including Mycobacterium tuberculosis H37Rv, Escherichia coli K-12, and Pseudomonas aeruginosa PAO1. 5 These findings support the interpretation that successful database bridges can preferentially expose already studied biology, while refining it: literature coverage is not a neutral proxy for biological importance or complete sequence-space representation. 5
At 50% sequence identity, 9.2% of protein families (31,653) had zero papers across all members and 46.1% (159,046) had exactly one paper, while only 4.3% (14,904) had 20 or more papers. 5 PaperBLAST therefore supports the inference that detectable annotation is coupled to reference and literature coverage, but refines the MSA interpretation by separating two non-equivalent gaps: a sequence family may have sparse homologous representation, sparse published coverage, or both. 5 The analysis identified 5,218 multi-member 50%-identity families, representing 14,534 sequences, with no literature across their members: 95.4% of multi-member 50%-identity clusters had at least one member with a paper, and the remaining 4.6% did not. 5 The report describes these literature-dark families as dominated by REBASE methyltransferases and biolip structural entries, which are highly specialized proteins without published functional studies. They therefore mark a literature gap and do not show sparse sequence-space representation. 5
The family boundaries behind these counts depend on the clustering threshold. mmseqs2 clustering of 815,571 PaperBLAST protein sequences gave 628K clusters at 90% identity, 345K at 50% (protein-family level), and 215K at 30% (superfamily level). Singletons made up 83.3%, 67.5%, and 62.8% of clusters at those thresholds, and the largest 90%-identity cluster was actin (212 members). 5 The largest 50%-identity families were housekeeping proteins: HSP70/BiP (650 members), GAPDH (609), enolase (552), and GroEL (443). Protein family size was positively associated with literature coverage, but the report gives no effect size for that relationship. 5 This refines the MSA-depth account by suggesting the hypothesis that broadly represented families also attract more literature, so sequence-space depth and research attention may be partly confounded rather than independent annotation layers. 5
The core-gene annotation paradox
The class-level pattern does not eliminate a substantial annotation gap within the core genome. The analysis identified 415,603 distinct core clusters with MSA depth < 10 across 14,768 species clades. 1 Their mean and median MSA depths were 4.57 and 4.0, respectively; 286,439 (68.9%) were hypothetical, 137 (0.033%) were EC-annotated, and 346 (0.083%) were KEGG-mapped. 1 This low-coverage core subset had a 68.9% hypothetical rate versus 3.8% for all core genes, while EC and KEGG annotations were below 0.1%. 1 The result supports core-gene-annotation-paradox: conservation-based classification can identify a gene as core while reference coverage remains too sparse for confident functional inference. 1
This subset refines genomic-under-representation by identifying conserved core proteins distant from characterized sequence space. 1 It also supports the hypothesis that functional counts, pathway prevalence, and candidate-gene prioritization can favor taxa and gene classes with stronger reference representation, although downstream ecological or phenotypic effects were not measured. 1 Top-ranked proteins with MSA depth = 1 came primarily from poorly characterized marine and soil bacteria, including Oceanicoccus, Dwaynesavagella, and CAILRJ01. 1 These proteins are candidates for experimental structural characterization, but this is a prioritization hypothesis rather than direct experimental validation. 1
Discoveries refines the boundary between low annotation and irreducible darkness. Bakta reclassified 33,105 (83.7%) of 39,532 Fitness Browser dark genes with pangenome links, leaving 6,427 truly dark genes; only 246/6,427 (3.8%) had zero annotation clues. 2 The remaining truly dark genes were shorter (121 versus 194 aa), less conserved (43% versus 73% core), had fewer orthologs (29% versus 64%), higher GC deviation (d=0.247), and frequently had hypothetical neighbors or nearby mobile elements. 2 This supports treating failed or sparse annotation as a graded evidence problem rather than equating it automatically with functional darkness, while leaving open whether the truly dark subset also has systematically lower MSA depth.
The truly dark genes project refines this digest-level profile with its own comparisons against annotation-lag genes: median length 121 versus 194 aa (d = −0.432), core-genome fraction 43.1% versus 72.7% (OR = 0.284), essential fraction 18.0% versus 13.4% (OR = 1.420), mean GC content 0.542 versus 0.584 (d = −0.395), ortholog presence 29.3% versus 63.7% (OR = 0.236), and median ortholog breadth 1 versus 4 organisms (d = −1.072). The p-values were < 1e-100 for length, core fraction, ortholog presence and ortholog breadth, < 1e-10 for essential fraction, and 2e-115 for GC content. 6 The higher essential fraction among truly dark genes suggests that persistent annotation failure does not imply that a gene is biologically dispensable. 6
Truly dark genes are not evenly distributed: Methanococcus strains S2 and JJ account for 55% of all truly dark genes, and across organisms 4–96% of dark genes resist bakta annotation, with archaeal organisms at the high end. 6 The report attributes this concentration to under-representation of archaea in annotation databases, which supports genomic-under-representation as a driver of persistent darkness. Because of this concentration, however, two archaeal genomes weigh heavily on the aggregate profile. 6
Truly dark genes are also nearly invisible to cross-organism analysis: only 3 of 65 dark-gene ortholog groups with cross-organism concordance data contain truly dark genes. 6 The report summarizes the ortholog contrast as truly dark genes being “4.2× less likely” to have cross-organism orthologs. That phrase describes the odds ratio (OR = 0.236) imprecisely: the underlying proportions are 29.3% versus 63.7%, so the result is not a 4.2-fold probability comparison. 6
A database link is not the same as a functional assignment for these genes. The report’s Key Findings state that 79.4% of truly dark genes have UniRef50 links and 84.7% have database cross-references (RefSeq, UniParc), yet only 4.0% have Pfam domain hits and only 4.6% have KEGG KOs. 6 Its Results section, however, places the cross-reference, Pfam and KEGG-KO percentages under 5,870 unique truly dark gene clusters rather than all 6,427 truly dark genes, so these percentages should not be treated as gene-level rates without checking the denominator. 6,2 Among those 5,870 clusters, Pfam yielded 362 hits for 235 clusters (4.0%), with TPR repeats dominating non-DUF hits and 56 clusters carrying DUF-only domains. 6 This refines the MSA-depth interpretation. The AlphaFold analysis links deeper sequence-space representation to richer annotation, but most truly dark genes are already linked to reference databases and still lack assignable function, so database representation alone is not enough for annotation. 1,6
Gene length is a candidate confound across these comparisons. Short genes are harder to annotate because they carry fewer domains and homologs, and harder to measure for fitness because they contain fewer TA sites, so the length difference (d = −0.432) may partly explain other observed differences. 6 This connects to transposon-callability-bias and suggests the hypothesis that part of the association between low MSA depth and sparse annotation is length-mediated; none of these analyses tested this directly. 6
Beyond homology: convergent functional evidence
The annotation-gap discovery study refines this sequence-space account by showing that functional recovery also depends on convergent evidence beyond homology. Across 201 gapfilled enzymatic reaction-organism pairs, its integrated pipeline assigned candidate genes to 96 pairs (47.8%), while BLAST alone resolved 70 (34.8%); no single evidence stream exceeded 35% resolution. 7 Thus, MSA depth supports homology-based annotation potential, while fitness, pangenome conservation, pathway evidence, and metabolic-model evidence can add resolution where sequence evidence alone is insufficient. 7
The same study supports the interpretation that annotation difficulty is function-dependent: 8 of 50 EC-less “dark reactions” (16%) were resolved, compared with 88 of 151 reactions with known EC numbers (58.3%). 7 This provides an independent, reaction-level analogue of the low-depth core gap, but does not establish that low MSA depth caused those unresolved reactions. 7 The complementary Bakta–eggNOG result supports this multimodal interpretation: Bakta rescued 11.2M clusters missed by eggNOG, while the union reached 77.3% any-functional-annotation coverage. 2
Interpretation and limits
The evidence supports a two-layer model. An MSA-depth-driven gap affects core, auxiliary, and singleton classes, while a pangenome-class gap produces higher hypothetical-protein rates among accessory and singleton genes; the latter may be associated with horizontal transfer, rapid evolution, and taxonomically narrow distribution, but those mechanisms were not directly established here. 1 The COG comparison supports the functional contrast between conserved core and novel classes across diverse bacteria, but its enrichment of mobile and defense categories does not by itself prove that horizontal gene transfer is the primary cause of every novel gene or of the annotation gap. 3 The cross-class comparison is dominated by the higher average MSA depth of core genes, whereas the low-depth core subset demonstrates that conservation does not guarantee rich annotation. 1
The reported ρ = 0.7563 was computed on the full 38,051,842-pair dataset without subgroup stratification, so it may differ among core, auxiliary, and singleton clusters and among organisms with different annotation gaps. 1 The report itself lists separate core, auxiliary, and singleton runs as an unresolved next step, to determine whether the MSA depth → domain-richness gradient is steeper in one class and whether structural novelty has the same functional meaning across the pangenome. The binned within-class monotonicity described above is therefore descriptive rather than a class-specific estimate. 1 The analysis used one representative sequence per gene cluster, so within-cluster sequence diversity was not captured and the representative could have higher or lower MSA depth than typical members. 1 The 293K genomes were not phylogenetically balanced, with common taxa such as Pseudomonas and E. coli over-represented. 1 The report gives 29.3% of gene clusters with a real UniProt accession and 28.7% successfully bridged to AlphaFold MSA depths, while elsewhere using 29.3% for the bridged fraction. In either case, the inference that the remaining 70.7% contains a larger annotation gap is plausible but not directly measured. 1 The analysis also used a static version-6 KBase Data Lakehouse AlphaFold snapshot, and later UniProt deposits may change MSA depths. 1
Platform pitfalls refine what MSA depth can and cannot signal. About 9% of AlphaFold entries have MSA depth < 30 and are flagged as low-confidence predictions, which should be filtered by msa_depth for structure-based analyses. 8 The low-depth core subset (MSA depth < 10) falls under this warning threshold. No structural-confidence (pLDDT) measurement was reported for that subset, so its predicted structures are not shown to be low-confidence; the warning only argues for caution in structure-based use. 8,1 AlphaFold coverage is UniProt-only, so proteins absent from UniProt have no AlphaFold entry, and per-residue pLDDT confidence is stored in the structure files rather than the metadata. SIFTS coverage is likewise limited to SwissProt/TrEMBL, leaving some protein-data-bank chains without any UniProt mapping. 8 Under these coverage rules, a failed AlphaFold or PDB bridge can reflect identifier coverage rather than biology, which is consistent with the 70.7% of clusters that the AlphaFold MSA report describes as lacking a usable link. 8,1
Annotation-density bias can also be measured and removed directly. In the gene-function ecological agora project, annotation-density residualization left producer_z bias-immune (R²=0.000) and consumer_z with a small bias (R²=0.053) that did not change any verdict. All hypothesis verdicts were preserved in the residualized outputs, which have 13.74M and 8 rows. 9 This refines the bias concern raised by the MSA-depth gradient and by gene-function-acquisition-depth: uneven annotation can be tested for, and in that atlas it did not change conclusions. A single analysis does not show that annotation-density effects are negligible elsewhere. 9
The COG analysis covered 32 species across 9 phyla, but its approximately 70% COG-annotation coverage leaves unassigned genes that could skew category distributions; its larger-sample recommendation also leaves open whether phylum-specific patterns are hidden. 3 Composite categories were counted once per gene, and eggnog v6 annotations may differ from original COG assignments. 3
PaperBLAST mines PubMed Central full-text articles, so paywalled papers are missed; 35% of organisms were assigned to an Unknown domain by heuristic classification; 50% sequence identity is a conventional rather than universal family boundary; and only 19% of the estimated 570K reviewed SwissProt entries were included. 5 Consequently, zero or low literature coverage cannot distinguish genuinely unstudied proteins from proteins missed by text mining, just as failed AlphaFold bridging cannot establish functional darkness. 5
Connections to other research
The coverage gradient strengthens pangenome-integration by connecting core/accessory structure to reference representation and hypothetical-protein rates. 1 The COG analysis supports this connection by showing that the core/novel functional partition is reproducible across 32 species and is strongest for mobile-element, defense, and conserved metabolic categories. 3 The sequence-space and literature evidence together support structural-annotation-gap as a measurable bridge between representation and functional annotation, while the PaperBLAST results show that research attention is an additional, uneven layer. 1,5 The annotation-gap discovery results further support multi-omics-integration by demonstrating that homology is most informative when combined with fitness and metabolic-model evidence. 7 Discoveries supports this connection by showing that Bakta and eggNOG recover complementary annotation layers rather than a single universal annotation source. 2 Joining the 415,603 low-depth core clusters to kescience-fitnessbrowser measurements could test whether structurally isolated proteins have condition-specific or growth-essential phenotypes, connecting this problem to gene-essentiality and multi-omics-integration. 1
The AMR pangenome atlas names AlphaFold and protein-data-bank structures as prospective resources for cross-referencing AMR proteins to identify novel resistance folds. This is a proposed analysis, not a reported result, but it would extend the structural-annotation axis from unannotated proteins to a functionally labeled protein class. 10
Tensions
The AlphaFold MSA report gives two counts for its lowest-depth core population that it does not reconcile: its binned table lists 415,733 core clusters in the MSA depth < 10 bin, whereas the low-depth core analysis reports 415,603 distinct core clusters with MSA depth < 10. The difference may reflect distinct versus non-distinct counting, but the source does not say, so both figures are retained as reported. The same report gives 29.3% for clusters with a real UniProt accession and 28.7% for clusters that bridge successfully to AlphaFold, yet elsewhere uses 29.3% for the bridged fraction; this inconsistency is also left unresolved. 1
Open Directions
- Recompute the bridge using all 132,531,501 gene clusters, including the 70.7% without a successful AlphaFold entry, with identifier recovery and stratified missingness analysis to test whether the unrepresented population has a larger annotation gap. 1
- Stratify the 38,051,842 pairs by core, auxiliary non-singleton, and auxiliary+singleton status and by phylogenetically balanced taxon strata; recompute Spearman correlations and domain-richness curves to test whether ρ = 0.7563 and the 2.89× and 2.77× median-depth ratios persist. 1
- Compare representative-sequence MSA depths with member-level depth distributions and InterProScan annotations to quantify whether representatives overestimate or underestimate typical reference coverage. 1
- Join the 415,603 core clusters with kescience-fitnessbrowser measurements and test condition-specific associations between low MSA depth, hypothetical status, essentiality, and fitness effects. 1
- Reprocess the static version-6 AlphaFold snapshot against later UniProt deposits and compare MSA-depth changes, measuring how database growth alters inferred annotation coverage. 1
- Cross-tabulate MSA depth, InterProScan status, PaperBLAST paper counts, and pangenome class for the same sequences, testing whether literature-poor families are also sequence-space-poor or instead reflect PMC/text-mining limitations. 5
- Apply the integrated gapfilling, fitness, pangenome, and homology evidence pipeline to low-depth core clusters to test whether multimodal evidence preferentially resolves candidates that sequence-space depth alone leaves poorly annotated. 7
- Separate biological sparsity from identifier and resource incompleteness by repeating the AlphaFold bridge after recovering the 33.3% of Bakta UniRef50 IDs present in the KBase Data Lakehouse UniProt identifier table, and test whether the missingness gradient changes. 2
- Compare low-MSA-depth core clusters with the 6,427 truly dark genes after Bakta reclassification to test whether sequence-space depth predicts persistent darkness rather than merely initial hypothetical status. 2
- Reanalyze COG L, V, and S enrichments after stratifying by MSA-depth bin, pangenome class, and phylum to test whether mobile, defense, and unknown-function categories independently predict sparse sequence-space representation. 3
- Retain composite COG categories while comparing their annotation richness and MSA-depth distributions with single-category genes, testing whether multifunctional categories expose or obscure the sequence-space annotation gradient. 3
- Compute ESMFold-based novelty, which does not use MSA depth, for the 415,603 low-depth core clusters and compare it with AlphaFold MSA depth to test whether an orthogonal structural novelty signal flags the same proteins; the report proposes this axis but reports no comparison. 1
- Run AlphaFold2 or ESMFold on the truly dark genes project’s top-100 candidates and search Foldseek for remote structural homologs, testing whether structure recovers function where sequence homology fails; this is the report’s proposed next step, not a validation result. 6
- Join truly dark gene clusters to AlphaFold MSA depth and stratify by gene length to test whether low MSA depth predicts persistent darkness independently of the d = −0.432 length difference. 6