Annotation nulls across precomputed substrates: absence signal or method-dependent gap?

Several projects in the corpus report missing gene or domain calls. Whether a marker appears absent changes with the annotation substrate and the marker set used. On a precomputed Pfam (protein-family domain database) substrate, null rates differ between audits. One Pfam family was confirmed silently absent in one lineage. Two annotation pipelines also disagree on whether a cofactor pathway is present. _1,2,3 This matters because downstream claims of gene loss or pathway absence rest on these nulls. Yet the projects describe the disagreements as method-level, and none of them establishes biological absence. _1,2,3 The broader framing is in homology-search-negative-evidence.

Evidence Sides

Side 1 — Pfam null rates vary with marker set and substrate On the same precomputed Pfam substrate, one audit reported 12/22 marker Pfams missing. Another reported 7 of 33 markers with zero clusters. By contrast, interproscan_domains (domain calls from InterProScan, a tool that scans proteins against multiple domain databases) had 1 zero-coverage marker. ₁ The null rate is therefore a property of the substrate and the marker selection, not a stable count of absent functions. ₁

Side 2 — A specific Pfam is silently absent in Bacillota_B PF14537 was confirmed silently absent in Bacillota_B. ₂ The project does not establish this as a biological absence. _1,2,3

Side 3 — eggNOG and bakta disagree on PQQ biosynthesis The lanthanide atlas compares two pipelines on PQQ (pyrroloquinoline quinone, a cofactor) biosynthesis. eggNOG assigns orthologous-group functional annotations, and bakta is a genome annotation pipeline. The two disagree on whether PQQ biosynthesis is present. ₃ The same report says bakta over-calls pqqA-E. It also gives two unreconciled denominators, 2,320 versus 2,185, for genomes carrying xoxF (a lanthanide-dependent methanol dehydrogenase gene) but lacking eggNOG pqq. ₃ One pipeline therefore reports absence where the other reports presence, and the over-calling caveat stops either from being taken as ground truth. ₃

Possible Reconciliations

• Hypothesis: the gap between Pfam audits comes from differences in marker-set composition, not substrate behavior.
• Hypothesis: the precomputed Pfam substrate has coverage gaps that interproscan_domains does not share, which would explain both the marker-level nulls and the PF14537 silent absence.
• Hypothesis: the eggNOG–bakta PQQ disagreement mixes eggNOG under-calls with bakta over-calls. Neither pipeline alone would then give a reliable absence call.
• Hypothesis: the two xoxF denominators come from different pipeline stages or filters, not from an error in either count.

Resolving Work

• Rerun the 22-marker and 33-marker sets against both the precomputed Pfam substrate and interproscan_domains on one shared genome set. Question: how much of the null-rate difference comes from the substrate rather than the marker selection?
• Run a sensitive profile-HMM (hidden Markov model) search for PF14537 directly on Bacillota_B proteomes and nucleotide assemblies. Question: is the silent absence a substrate gap or a true lack of the domain?
• Benchmark eggNOG and bakta pqqA-E calls against curated genomes with known PQQ status. Question: what are each pipeline’s false-positive and false-negative rates?
• Trace the provenance of the 2,320 and 2,185 xoxF genome sets through the lanthanide atlas notebooks. Question: which filter separates them, and which denominator should downstream rates use?