Negative homology-search results require sensitivity validation

A negative homology-search result does not by itself demonstrate that a gene or pathway is absent: naming conventions, annotation quality, database coverage, and method-specific limitations can reduce sensitivity. ₁ Claims of absence should therefore be treated as method-dependent until independent evidence supports them. ₁ Negative evidence becomes more credible when searches are benchmarked against known positives and checked with complementary annotations, sequence-based methods, pathway reconstruction, or biological evidence. ₁

The Regulatory and proteomic architecture of Δfur-permitted lipid A loss in *Caulobacter crescentus* provides a direct sensitivity failure case. PaperBLAST returned 0 hits for Caulobacter LpxA versus 11 NCBI hits, 0 for LpxC versus 15, 0 for LpxD versus 15, 1 for LpxB versus 18, and 0 for LpxK versus 18. ₁ The report characterized these results as an approximately 80% false-negative rate for the tested Caulobacter lipid A genes. ₁ Thus, a zero-hit PaperBLAST result can reflect search or representation failure rather than biological absence, and a single-database result should be reported as “not detected by that method” rather than as definitive absence. ₁

The report does not describe these counts consistently, so they cannot be read as a single account. Its detailed NB06b table, which compares the NCBI re-test with PaperBLAST, gives PaperBLAST 1 LpxB hit versus 18 NCBI hits. The table labels this a “PaperBLAST under-count”, not a zero-hit false negative, while LpxA, LpxC, LpxD and LpxK are labelled false negatives. ₁ The report’s scorecard and later prose instead state that LpxA/C/D/B/K “all returned 0” from PaperBLAST but 11-18 each from NCBI. That statement conflicts with the table’s LpxB entry. ₁ This page uses the table’s per-gene counts and treats the all-zero wording as an unresolved internal contradiction rather than correcting either side. ₁

The original NB06 screen used PaperBLAST description-text matching. The report’s adversarial review (I4) flagged an approximately 50% false-negative rate for known Caulobacter essential genes. ₁ NB06b, added after the review, re-ran the comparative panel against NCBI’s protein database via Biopython Entrez, which uses curated genome annotation rather than text matching. ₁ A later limitations passage reports a ~80% false-negative rate for known Caulobacter essential lipid A biosynthesis genes, listing LpxA, LpxC, LpxD and LpxK as 0 hits in PaperBLAST against 11-18 each in NCBI. ₁ The ~50% and ~80% rates refer to differently described gene sets, and the report does not reconcile them. ₁ The report concludes that NB06 should be read as a screen and NB06b as the authoritative comparative measurement for the headline claim. ₁

Benchmark negative results against known positives

The Caulobacter lipid A genes supplied an internal positive-control set because independent NCBI annotation identified genes that PaperBLAST failed to retrieve. ₁ This supports testing recall on genes known to be present in the focal genome before interpreting negative results across comparator genomes. ₁ kescience-paperblast is therefore most informative when its hit behavior is evaluated against an independently annotated reference rather than treated as a complete absence catalog. ₁

Evidence triangulation in the Caulobacter comparison

Despite the PaperBLAST false negatives, NCBI annotation independently supported the cross-species absence claims for the tested comparison. ₁ NCBI annotation found spt and cerR to be Caulobacter-unique across C. crescentus, A. baumannii, N. meningitidis, and M. catarrhalis, with pattern 1000. ₁ ChvG and ChvI also showed pattern 1000, supporting restriction of the ChvG–ChvI regulatory circuit to Caulobacter among the tested species. ₁ The conclusion that comparator species lacked the Caulobacter sphingolipid-substitute pathway and ChvG–ChvI circuit therefore rested on NCBI annotation and literature rather than on the original PaperBLAST screen alone. ₁

The per-species counts refine that pattern. For spt, PaperBLAST and NCBI agreed at (7,0,0,0) across the four species. ₁ For cerR, the two methods disagreed: PaperBLAST returned (1,1,0,0) and NCBI returned (6,0,0,0). ₁ NCBI counts for ChvG and ChvI were (5,0,0,0) and (6,0,0,0), and the report describes the ChvG–ChvI restriction as consistent with Greenwich 2023. ₁ The bcerS query failed in C. crescentus, but the report states that the other markers confirm the pattern. ₁ The report concludes that absence of the sphingolipid biosynthesis pathway and of ChvG–ChvI in A. baumannii, N. meningitidis and M. catarrhalis is supported by NCBI annotation as well as by PaperBLAST text matching and published literature (Olea-Ozuna 2020/2024; Greenwich 2023). ₁ It argues that the PaperBLAST failures do not threaten this finding, because the finding concerns absence in the comparators, where NCBI confirms. ₁ The cerR disagreement shows that even the positive Caulobacter control was not detected consistently by text matching. ₁

The independent annotation source has its own false negatives. NCBI symbol queries returned C. crescentus counts of 0 for MsbA and LptA, yet MsbA is present as CCNA_00307 (“phospholipid/LPS ABC transporter”), and LptA is present but not annotated with the “lptA” symbol in Caulobacter NA1000. ₁ The report calls these naming-convention false negatives rather than true absences, because Caulobacter is annotated primarily with CCNA_* locus tags rather than canonical Gram-negative gene symbols. ₁ It argues that the sphingolipid and ChvI findings are unaffected because the corresponding C. crescentus genes carry canonical symbols, with NCBI Cc counts of 7 for spt, 5 for ChvG, and 6 for cerR/ChvI. ₁ A symbol-based search can therefore validate an absence only when the focal positive control is itself annotated under the queried symbol. ₁

These independent positive detections refine absence analysis by showing that pathway-level conclusions require searches for both the queried pathway and plausible replacement routes. ₁ The comparison identified species-specific alternatives to lipid A loss: A. baumannii had PBP1A and Ld-transpeptidases, N. meningitidis had 9 capsule-biosynthesis PaperBLAST hits, and late acyltransferase lpxX/lpxL hits numbered 8 in A. baumannii and 3 in N. meningitidis. ₁ The report states that Caulobacter lacks these alternatives and instead uses its sphingolipid-substitution route, but treats the Caulobacter absence conclusions as stronger because they were supported by independent NCBI annotation and literature. ₁ A negative search for one route therefore cannot establish absence of the phenotype or pathway function when unrelated replacement mechanisms remain possible. ₁

Database, annotation, and method coverage

The comparison was vulnerable to naming-convention false negatives, and M. catarrhalis was under-annotated in PaperBLAST with 162 genes total. ₁ The report directs that M. catarrhalis presence/absence rows be treated as advisory only. ₁ This under-annotation further weakened negative results from that comparison and makes database representation and annotation completeness explicit components of the evidential basis for a negative homology result. ₁ This issue connects to environmental-resistome, where inferred biological absence or prevalence can depend on how genes are represented and annotated in the underlying resource. ₁

The report identifies the need for a deeper Pfam HMM search against named RefSeq proteomes to detect unannotated paralogs; an HMM is a profile hidden Markov model built from a protein-family alignment. ₁ This limitation supports structural-annotation-gap and evidence-triangulation-for-functional-annotation: absence claims become stronger when independent annotation systems, sequence-based homology searches, and biological evidence converge. ₁ Pathway absence and gene absence remain distinct claims: the cross-species patterns concerned annotated genes and regulatory components, whereas the functional conclusion concerned availability of alternative routes for lipid A loss. ₁

The report states that the NB06b NCBI re-test confirmed the headline biological claims, but that annotation-based search misses unannotated paralogs. ₁ For full rigor it proposes pfam HMM searches using hmmsearch or pyhmmer. The profiles are PF00155 serine palmitoyltransferase, PF03331 LpxC and PF01475 Fur. The proteomes are GCF_000196795.1 A. baumannii ATCC 17978, GCF_000008805.1 N. meningitidis MC58 and GCF_000092265.1 M. catarrhalis BBH18. ₁ Until that search is run, an unannotated spt paralog in one of the comparators remains an open possibility. ₁

A stronger absence claim should combine database annotation with sensitive sequence-profile searches, independent homology searches, inspection of pathway alternatives, and, where feasible, direct functional or biochemical tests. ₁

Annotation substrates can silently lack whole domain families

The Caulobacter failure is not isolated. Other projects show that a precomputed domain table can return zero for a family that is present, which supports the general rule that absence must be validated against the annotation substrate. _2,3 In gene_function_ecological_agora, a prior audit found 12/22 marker Pfams silently missing from bakta_pfam_domains (bakta annotation). ₂ The project’s NB13 audit found that 7 of 33 marker Pfams had zero clusters in that table, including all four critical photosystem ii Pfams: PsbA/PF00124, PsbB/PF02530, PsbC/PF02533 and PsbD/PF00421. ₂ By contrast, interproscan_domains (interproscan) had only 1 zero-coverage marker (PF13415, a minor HATPase variant). The median bakta/IPS coverage ratio was 0.102, which the report describes as ~10× higher IPS coverage on average. ₂ The project therefore chose interproscan_domains for all Phase 3 architectural work instead of the planned HMMER recompute fallback. ₂ It notes that the Cyanobacteria PSII hypothesis test would have been structurally impossible under the bakta substrate, which had zero PSII Pfam hits. This supports callability-limited-comparative-inference: a test can be uncallable rather than negative. ₂

Some zero-hit domain results are query-format failures rather than coverage gaps. In plant_microbiome_ecotypes, the Phase 1 NB02 query WHERE pfam_id IN ('PF00771', …) returned zero rows because bakta_pfam_domains stores versioned IDs (PF00771.22; the limitations passage gives PF13629.12 as its example). ₄ The initial Pfam marker-hit output (data/pfam_marker_hits.csv) held 0 rows. Classification instead relied on gene name, KEGG KO and product keyword matches, and NB02 itself fell back to gene-name keyword matching. ₄ The report attributes that initial zero to the query format, not to a domain-coverage gap. It lists the problem as resolved in NB13 Cell 3 via a LIKE 'PF00771%' pattern, while noting that the marker set remains literature-curated. ₄

Prefix matching refines that null into a partial recovery rather than a complete one. Re-running with LIKE 'PF00771%' patterns on the 10 Phase 2b marker Pfams recovered 19,364 domain hits across 7,962 species. ₄ Five of those 10 Pfam IDs produced hits: nitrogenase NifH PF00142 (6,632 hits, 4,040 species, 32.4% core), cellulase GH5 PF00150 (6,898 / 3,234 / 25.9%), pectate lyase 3 PF12708 (4,149 / 2,283 / 24.6%), pectate lyase PF00544 (1,115 / 546 / 20.8%), and T6SS VgrG PF05943 (570 / 343 / 10.9%). The remaining five returned 0 hits. ₄ Fixing the query format therefore does not by itself turn every null into a detection; the residual zeros fall under the coverage audit described next. ₄

The central pitfalls digest records an earlier, unresolved form of the same symptom. Querying bakta_pfam_domains with Pfam IDs like 'PF00771' returned 0 hits across all 11 domains tested. The digest suggests that the table may use a different ID format (domain names, or versioned accessions like PF00771.1) or may need different join logic. ₅ That entry states that the pitfall needs further investigation to determine the correct query format. ₅ The plant microbiome report’s versioned-ID diagnosis refines this open pitfall. The two passages are kept separate here because the digest entry does not itself report the fix. _5,4

Beyond query format, the plant microbiome audit found a genuine coverage gap. It reports that 12 of the 22 marker Pfams are completely absent from bakta_pfam_domains despite being abundant in interproscan_domains, describes them as systematically missing, and states that InterProScan has them all. ₄ The silent-gap set is dominated by secretion-system components: T3SS (PF00771 18.6K IPS hits, PF01313 13.6K, PF01514 15.0K, PF02579 16.8K, PF04610 15.6K, PF07916 8.4K, PF09599 0.3K, PF01312 22.0K), T4SS (PF03135 11.0K, PF03743 17.5K), and T6SS (PF05936 7.5K Hcp tube, PF04183 7.5K). ₄ The 10 Pfams that are in bakta still return only 10–35% of the IPS hit count, so even those are systematically undercounted but non-zero. As the likely cause, the report proposes that bakta uses a reduced “core” Pfam HMM profile set while InterProScan uses the full Pfam-A release. ₄ The report keeps this coverage gap distinct from NB02’s zero-hit query-format failure. It says the project was unaffected because NB10 uses InterProScan. ₄

Switching substrate recovered the signal. NB10’s refined cohort pipeline already used interproscan_domains (signature_acc field) for T3SS, T4SS, T6SS and CWDE (cell-wall-degrading enzyme) marker detection. It recovered 280,193 marker domain hits across 24,595 species, including 2,804 additional species with CWDE domains (cellulase/pectinase Pfams PF00150, PF00544) not detected in Phase 1’s bakta-only search. The report states that the refined cohort assignments are unaffected. ₄ It recommends treating bakta_pfam_domains as a supplement to, not a substitute for, interproscan_domains. ₄ This supports the gene_function_ecological_agora decision to use interproscan_domains for its Phase 3 architectural work. _4,2

The phage defense arsenal project reports the same gap for defense markers. bakta_pfam_domains misses classical defense markers; for example, CRISPR Cas1 PF01867 returns 0 there but ~25K in InterProScan. ₆ A zero Cas1 count from the bakta table is therefore not evidence that CRISPR systems are absent; see crispr-cas. ₆

The Bacillota_B subsurface project independently reproduces the pattern. PF14537 Cytochrom_NNT had 0 hits and 0 clusters in its PFAM availability check, and the report calls it “confirmed silently absent”, citing a pitfall previously documented in another project. ₃ Multi-heme cytochrome PFAMs were sparse within Bacillota_B, with only ~6 clusters total across all four PFAMs. ₃ Sequence-level CXXCH motif counting on gene_cluster.faa_sequence carried most of the iron-reduction signal, and the report explains that this is why its corrected IR rates (40–56%) are far higher than PFAM-only detection would yield. ₃ Motif-level search is thus an orthogonal check that can overturn a domain-table null; see multi-heme-cytochrome-detection and functional-marker-validation. ₃

The central discoveries digest traces part of this gap to how bakta runs. Its Pfam search is skipped by design, because HMMER runs only on hypotheticals (~7.7%) and the 92% of proteins with PSC matches are never Pfam-searched. ₇ The digest also flags that bakta’s PSC uses diamond fast mode, whereas eggnog uses HMM-based ortholog assignment, and summarizes the difference as “DIAMOND fast < HMM sensitivity”. ₇ This offers a mechanistic explanation, from the digest rather than a direct benchmark, for why bakta Pfam tables under-report families that profile-based substrates detect. ₇

“Dark” status inherits the same limits. Among 5,870 unique truly dark gene clusters, Pfam gave 362 hits for 235 clusters (4.0%). TPR repeats dominated the non-DUF hits, and 56 clusters had DUF-only domains. ₈ The report cautions that some bakta “hypothetical protein” calls may be bakta false negatives: genes with known functions that did not match bakta’s PSC database at the version tested (v6.0). ₈ This supports the digest’s point that PSC-based assignment limits downstream annotation. _8,7 The proposed structural-annotation follow-up targets sequences with no detectable sequence homology. It treats a homology null as the starting point for an orthogonal method, not as evidence that the gene has no function; see structural-annotation-gap. ₈

Different annotation sources can also disagree on whole pathways. In the lanthanide methylotrophy atlas, a pilot exploration found 2,320 xoxf-bearing genomes lacking any eggNOG PQQ-biosynthesis annotation, even though PQQ is an obligate XoxF cofactor. ₉ A later passage of the same report instead counts 2,185 genomes with no eggNOG pqq. The report does not reconcile this with the pilot’s 2,320, so which denominator applies remains unresolved. ₉ Within the xoxF set, complete eggNOG pqqA-E annotation occurred in 33 genomes (0.9 %), and partial eggNOG pqq annotation (1–4 of A-E) in 1,472 genomes (39.9 %). ₉ Bakta-only strong evidence (≥3 PQQ products) occurred in 899 genomes (24.4 %). ₉ Bakta product fields therefore add pqq-biosynthesis annotation where eggNOG has none. However, the same report warns that bakta over-calls pqqA-E and treats eggNOG as primary. These bakta-only calls are annotation evidence that still needs sequence-level validation, not established pathway recovery. ₉ Even so, the disagreement means that an eggNOG-only null for an obligate cofactor pathway should not be read as evidence of absence. ₉

The full eggNOG/bakta partition refines the PQQ null further. Bakta alone detected 1–2 PQQ products (“Bakta-only partial”) in 389 genomes (10.5 %), and 897 genomes (24.3 %) had no PQQ detected by either source. ₉ Of the 2,185 genomes with no eggNOG pqq, 1,288 (59 %) have ≥1 bakta PQQ product. The report calls these pure annotation gaps, where bakta detects PQQ that eggNOG misses. The remaining 897 genomes (24 % of all xoxF carriers) lack PQQ evidence from either source. The report treats them as candidates for assembly incompleteness, pseudogenization, or genuine reliance on community-acquired PQQ. It states that disambiguating these three explanations requires sequence-level inspection, which was out of scope. ₉ The report’s summary restates this as 59 % of xoxF-without-PQQ cases being eggNOG annotation gaps that bakta detects, with the remaining ~24 % of xoxF carriers genuinely lacking PQQ evidence. ₉ The “pure annotation gap” reading sits alongside the report’s own warning, noted above, that bakta over-calls pqqA-E. The rescued genomes therefore rest on annotation evidence that itself awaits sequence-level validation. ₉

The doubly-null 897 genomes are therefore not established absences, and the method sets the limit. xoxF and PQQ presence/absence is called at the gene-call level, with no screen for pseudogenes, truncated ORFs, or assembly fragmentation. Genuinely novel REE-handling (rare-earth-element) enzymes that lack KEGG/RefSeq homologs would not be detected. ₉ A future-directions item covers the ~897 “no-PQQ-detected” xoxF genomes. It calls for per-genome ORF integrity plus completeness scoring (CheckM2) to separate assembly fragmentation, pseudogenization and genuine community-PQQ acquisition; see community-metabolic-interdependence. ₉

Low-frequency annotation calls need the same qualification. In the cystic fibrosis formulation project, ladS was detected in only 0.5% of P. aeruginosa (PA) genomes by bakta annotation, so the rate depends on that annotation. ₁₀ The report interprets PA14’s ladS frameshift mutation as an extreme outlier that locks the strain into an acute-virulence regulatory state, unrepresentative of natural infection dynamics. That characterization is the report’s interpretation and rests on an annotation-qualified detection rate, not on a validated prevalence estimate. ₁₀

Name, text and keyword matching

Gene-name searches fail silently on annotation vocabulary. The PGP pangenome ecology project detected plant-growth-promoting (PGP) genes only through bakta gene annotations matching exact names (nifH, acdS, pqqC, etc.). Genes annotated by product description only, or with variant names, are therefore missed, and the report notes that this particularly affects less-characterized PGP genes. ₁₁ This supports the Caulobacter finding that symbol-based searches are only as sensitive as the annotation’s naming conventions. _11,1

Text matching errs in both directions. In respiratory chain wiring, the ndh 2 gene search in Fitness Browser organisms may miss some orthologs because annotation varies. ₁₂ The same project’s NB04 NDH-2 identification used text matching on gene descriptions. The report says this likely produces false positives for organisms with incompletely annotated Complex I subunits, for example pseudo3_N2E3 with 8 “NDH-2” hits that are probably a Complex I operon. It suggests that KO-based identification (K03885) would be more reliable. ₁₂ This supports the lanmodulin point below that name-based presence calls need the same validation as absence calls. ₁₂

Ortholog-database nulls for single genes carry the same ambiguity. phaR (K18080) was absent from all 27K species in the eggNOG annotations. The PHB granule report reads this as suggesting either poor annotation coverage or misassignment of this regulatory gene, not biological absence. ₁₃ The same report notes that its large “precursors-only” category (46.5%) reflects the pleiotropic nature of phaA (K00626) and phaB (K00023). Both participate in general fatty acid and thiolase metabolism beyond PHB synthesis, so presence calls on those genes are not PHB-specific either; see phb-granule-ecology. ₁₃

Resistance-gene catalogues and identity thresholds

Resistance-gene absence inherits the scope of the catalogue used. amrfinderplus focuses on known resistance genes and may underestimate novel resistance in environmental bacteria. The environmental resistome report notes that this would bias results toward finding more AMR in clinical species, whose resistance genes are well characterized. ₁₄ The strain-variation report similarly states that novel resistance mechanisms not in the AMRFinderPlus database are missed. ₁₅ Two projects independently flag this. Low environmental AMR counts may therefore be undercounts biased in a known direction, and should not be read as evidence that resistance is absent. Neither report quantifies how large the undercount is; see environmental-resistome. _14,15

Sequence-identity cut-offs for linking genes across resources create a parallel sensitivity floor. The AMR pangenome atlas used a 100% DIAMOND identity threshold to link Fitness Browser genes to pangenome clusters. The report calls this conservative and notes that closely related allelic variants may be missed, potentially undercounting AMR gene fitness effects. ₁₆ The costly-dispensable-genes project used a 90% identity DIAMOND threshold for the same linking and notes that it may miss recently acquired genes with low sequence similarity. ₁₇ Neither threshold is benchmarked against known positives in these passages, so an unlinked gene is not evidence of an absent ortholog. _16,17

Orthology inference and reference-set limits

bidirectional-best-hit-orthology (BBH, pairs of genes that are each other’s best BLAST hit) is conservative in a direction that inflates apparent absence. The essential-genome report states that BBH misses paralogs, gene fusions and distant homologs, so some “orphan essentials” may have undetected orthologs with diverged sequences. ₁₈ The same report notes the opposite error for connected-component grouping of the BBH graph: transitive connections can merge unrelated genes into one ortholog group, particularly for multi-domain proteins. ₁₈ The costly-dispensable-genes project assigned ortholog groups by BBH across 48 organisms. It notes that genes with orthologs outside this set would be misclassified as orphans, so orphan status depends on the reference set. ₁₇

Pathway-prediction gaps carry the same ambiguity. The essential-metabolome report offers three explanations for the predicted gapmind serine gap in DvH. DvH may use a non-canonical serine pathway not in GapMind’s reference database, the pathway may use divergent enzymes below GapMind’s homology threshold, or the genes may be present but unannotated. ₁₉ The report also notes that E. coli is absent from the dataset, that coverage is unknown for the remaining 38 FB organisms, and that organisms with divergent metabolic genes may be missed. ₁₉ A pathway gap is therefore a hypothesis of auxotrophy, not a measured one; see biosynthetic-prototrophy-and-auxotrophy. ₁₉

Not every null is an artifact, and not every hit is real

Sensitivity validation can also confirm a null. In the SNIPE analysis, the absence of gene clusters co-encoding DUF4041 and PF01541 was explained by family assignment: the SNIPE nuclease domain belongs to Pfam family pf13455 mug113 (Mug113), not PF01541 (canonical GIY-YIG). Both families sit in the GIY-YIG clan (CL0418) but are distinct. ₇ The digest concludes that the zero co-occurrence is “biological truth, not an annotation artifact”. It verified this via InterPro/UniProt analysis of the E. coli SNIPE protein A0A0A1A5Z2, which carries PF13250 at positions 232–333 and PF13455 at positions 443–520. ₇ This refines the page’s thesis: a null becomes credible when an independent reference explains why the queried family should not appear, not merely when a search returns nothing. ₇

Annotation-field format can also manufacture a null. The pitfalls digest records that the eggnog_mapper_annotations.PFAMs column stores Pfam domain names (e.g. DUF4041, GIY-YIG), not Pfam accession IDs (e.g. PF13250, PF01541), so searching by accession returns 0 results. ₅ In its example, WHERE PFAMs LIKE '%PF13250%' returns 0 rows while WHERE PFAMs LIKE '%DUF4041%' returns 2,962. ₅ This parallels the versioned-ID failure in bakta_pfam_domains: in both tables the zero reflects how identifiers are stored, not what the genomes contain. _5,4

Name-to-accession mapping is a second failure point. The digest warns that Pfam domain names can map to wrong accessions if looked up by name alone. PF13250 and PF13291 are numerically close but completely unrelated families, and PF01541 (canonical GIY-YIG) and PF13455 (Mug113) are in the same clan but are distinct families. ₅ The digest gives the symptom as zero co-occurrence hits for domains that should co-occur (e.g. DUF4041 + GIY-YIG in SNIPE proteins), or unexpected functional annotations for domain hits. Its fix is to always verify Pfam accessions against a known protein in InterPro or UniProt. ₅ For SNIPE it gives DUF4041 = PF13250 (IPR025280), not PF13291 (ACT_4), and SNIPE nuclease = PF13455 (Mug113, GIY-YIG clan CL0418), not PF01541 (canonical GIY-YIG); see pf13250 duf4041. ₅

The SNIPE project report supports the digest’s reading of the PF01541 null. It states that searching for PF01541 co-occurrence with DUF4041 yields zero hits, not because of annotation artifacts, but because these are genuinely different Pfam families. ₂₀ It also refines the picture with a sensitivity caveat on the correct family. eggNOG Pfam annotations may miss divergent SNIPE homologues, and only 54/4,572 DUF4041 clusters show Mug113 co-annotation, which the report says suggests many SNIPE nuclease domains go undetected; see snipe-defense-system. ₂₀ The PF01541 null is therefore validated as biological. The low PF13455 co-annotation is a sensitivity-limited count, not a measured frequency of SNIPE loci lacking a nuclease. ₂₀

The converse error, a false positive from annotation, also occurs. The digest suggests that eggNOG’s Preferred_name for lanmodulin appears to be a stale or seed-ortholog mislabeling rather than a true homology call. This is presented as a likely explanation, not an established one. ₇ Annotation-name searches can therefore err in both directions, and presence calls need the same validation as absence calls; see functional-marker-validation. ₇

The same logic extends beyond sequence search to literature search. The ENIGMA carbon census describes its unannotated compounds as resource-darkness, not proof of absence from science. Class-level catabolic literature exists for monoterpenes (Marmulla & Harder, Front. Microbiol. 2014) and for alkaloids such as nicotine (Huang et al., Front. Microbiol. 2020). ₂₁ Its literature-rescue channel (NB02b) returned zero only because it used a shallow PubMed-title-only screen. The report calls this a method floor rather than evidence of absence. ₂₁

The PaperBLAST explorer project names the missing control directly. Without negative controls, it cannot easily distinguish genes that are genuinely unstudied from genes whose literature was missed by text mining. ₂₂ A gene with no linked papers is therefore an unvalidated literature null, which supports the ENIGMA census’s treatment of resource-darkness; see research-attention-inequality. _22,21

Tensions

PaperBLAST produced strong negative evidence for several Caulobacter lipid A genes, while NCBI annotation produced positive evidence for the same genes, creating a method-level conflict rather than a biological contradiction. ₁ The discrepancy is provisionally handled by treating PaperBLAST negatives as sensitivity-limited and retaining cross-species absence claims only where independent annotation and literature support them. ₁

The Caulobacter report is also internally inconsistent about the size of the PaperBLAST failure. One side is its NB06b table, which gives LpxB as 1 PaperBLAST hit versus 18 NCBI hits and labels it an under-count. ₁ The other side is its scorecard and later prose, which state that LpxA/C/D/B/K all returned 0 from PaperBLAST. ₁ Separately, the adversarial review’s ~50% false-negative rate for known Caulobacter essential genes sits alongside the later ~80% rate for lipid A genes, without reconciliation. ₁ Neither discrepancy changes the comparator-absence conclusion as the report frames it. Both mean the false-negative rate should be cited with its gene set and source passage, not as a single settled figure. ₁

Across projects, the same precomputed Pfam substrate yields different null rates depending on the marker set. The reported rates are 12/22 marker Pfams missing in one audit and 7 of 33 with zero clusters in another, against 1 zero-coverage marker in interproscan_domains. ₂ PF14537 was confirmed silently absent in Bacillota_B. ₃ In the lanthanide atlas, eggNOG and bakta disagree on PQQ-biosynthesis presence; the report also says bakta over-calls pqqA-E and gives unreconciled denominators (2,320 versus 2,185) for xoxF genomes lacking eggNOG pqq. ₉ These are method-level disagreements. None of them establishes biological absence. _2,3,9

The sources also give different accounts of why bakta Pfam tables come up empty. The plant microbiome report treats its original zero as a resolved versioned-ID query-format problem. It attributes the separate 12/22 coverage gap to a likely reduced “core” Pfam profile set in bakta. ₄ The discoveries digest instead attributes the under-reporting to bakta running its Pfam search only on hypotheticals. ₇ The pitfalls digest’s entry on 0 hits across all 11 domains tested still describes the correct query format as unestablished. ₅ None of these passages tests the mechanism proposed by another, so the explanations stand unreconciled. _4,7,5

Open Directions

• Search named RefSeq proteomes with Pfam HMMs and other profile-based or sequence-based homology methods to determine whether unannotated paralogs explain or undermine the reported absence of LpxA, LpxC, LpxD, LpxB, LpxK, spt, cerR, ChvG, and ChvI in comparator species. ₁
• Run the report’s proposed hmmsearch/pyhmmer screen (PF00155, PF03331, PF01475 against GCF_000196795.1, GCF_000008805.1 and GCF_000092265.1) to close the residual concern about an unannotated spt paralog. ₁
• Build a sensitivity benchmark from the 11 NCBI LpxA hits, 15 NCBI LpxC hits, 15 NCBI LpxD hits, 18 NCBI LpxB hits, and 18 NCBI LpxK hits, then compare recall across PaperBLAST, NCBI annotation, and Pfam HMM searches. ₁
• Re-derive the PaperBLAST LpxA/C/D/B/K counts from the raw NB06 outputs to settle whether LpxB returned 0 or 1, and report the false-negative rate with an explicit gene-set denominator. ₁
• Reanalyze the four comparator proteomes, especially M. catarrhalis, with consistent gene calling, complete quality-controlled proteomes, and annotation pipelines to test whether the reported absence claims persist after replacing its 162-gene PaperBLAST representation. ₁
• Compare the Caulobacter sphingolipid-substitution route with PBP1A, Ld-transpeptidase, capsule-biosynthesis, and lpxX/lpxL alternatives using profile-HMM, synteny, and genome-resolved pathway searches to ask whether replacement pathways are genuinely absent or merely poorly annotated. ₁
• Combine genome-resolved pathway reconstruction with targeted lipidomics to test whether annotated absence of comparator lipid A-loss alternatives corresponds to absence of predicted biochemical products. ₁
• Perform cross-species genetic engineering or complementation to test whether the Caulobacter sphingolipid-substitution route is functionally distinct from the alternative routes identified in A. baumannii and N. meningitidis. ₁
• Recompute Bacillota_B multi-heme cytochrome and PF14537 calls on interproscan_domains, then compare them with CXXCH motif counts to measure how much of the iron-reduction signal each substrate recovers. _3,2
• Run PQQ-biosynthesis profile HMMs on the xoxF-bearing genomes that lack eggNOG pqq annotation, to test whether bakta-only calls are true PQQ operons or bakta over-calls, and to reconcile the report’s 2,320 and 2,185 denominators. ₉
• Re-link Fitness Browser genes to pangenome clusters across a range of DIAMOND identity thresholds, testing whether AMR fitness effects and orphan or acquired-gene calls change between the 100% and 90% settings. _16,17
• Screen environmental genomes with profile-based resistance-family searches alongside AMRFinderPlus, to test whether the clinical-versus-environment AMR contrast survives a catalogue-independent detector. _14,15
• Re-run the pitfalls digest’s 11-domain bakta_pfam_domains query with versioned-ID prefix matching and compare each domain with interproscan_domains, to separate query-format zeros from bakta coverage gaps. _5,4
• Query interproscan_domains for the five Phase 2b marker Pfams that still returned 0 hits after prefix matching, to classify each as a bakta coverage gap or a true absence in the plant-associated genomes. ₄
• Score the ~897 xoxF genomes with no PQQ detected by either source using CheckM2 completeness and ORF-integrity checks, to separate assembly fragmentation and pseudogenization from community-acquired PQQ. ₉
• Re-identify NDH-2 in Fitness Browser organisms by KO K03885 and compare with the NB04 text-matching calls, starting with the 8 putative NDH-2 hits in pseudo3_N2E3. ₁₂
• Run PF13455 profile searches through InterProScan on the DUF4041 clusters lacking Mug113 co-annotation, to measure how many of the 4,572 clusters carry an undetected SNIPE nuclease domain. ₂₀
• Search the 27K species with a phaR (K18080) profile HMM, to test whether its eggNOG absence is an annotation-coverage or misassignment artifact. ₁₃