Functional annotation and taxonomy inference from sequence homology.
-
Read docs/README.md and the relevant tool guides before running anything.
-
Normalize tool outputs and generate the complete comparison bundle with the schema-backed driver:
uv run --script skills/bio-annotation/scripts/build_annotation_artifacts.py \
raw_annotations.tsv --genomes genomes.tsv --markers marker_catalog.tsv \
--out results/bio-annotation
The driver refuses a non-empty destination, enforces globally unique protein identifiers, writes normalized Parquet tables, adds explicit absent-marker rows, and computes query-specific/missing/expanded/contracted families against the reference median. The artifact contract is in schemas/artifacts.schema.json.
-
When a nucleotide assembly, MAG, genome, or contig FASTA is available, run /tracking-taxonomy-updates first for the BBTools-container QuickClade percontig domain screen. Use that routing table to choose the right taxonomy/QC path before interpreting protein annotations.
-
For InterProScan, read docs/interproscan-usage.md and validate the exact CLI with --help or --version. Current stable is v5.77-108.0; InterProScan 6 (Nextflow-based) is a forward-looking migration target.
-
Run InterProScan for domain/family annotation.
-
Run eggNOG-mapper v2.1.13+ for orthology-based annotation.
-
Run sequence-vs-database search and resolve taxonomy with TaxonKit v0.20.0+ (required for the March 2025 NCBI rank update that replaces "superkingdom" with "domain" and adds "realm" for viruses).
- Default CPU path: DIAMOND v2.1.20+. For any search against NCBI nr, prefer a clustered nr database (e.g., a
clusterednr build under $BIO_DB_ROOT) — it is dramatically faster than full nr at comparable sensitivity for most annotation tasks. Check whether a clusterednr build is available under the reference root; if not, build one with diamond makedb from a clustered FASTA (MMseqs2/CD-HIT-reduced nr) or fall back to full nr and record the choice in the run log.
- GPU node available (CUDA Turing or newer): MMseqs2-GPU as an alternative to DIAMOND. Published benchmark: 20× faster and ~71× cheaper per query versus 128-core CPU MMseqs2, and 177–199× faster than JackHMMER iterative search for profile-equivalent workflows (Kallenborn et al., Nature Methods 2025, DOI: 10.1038/s41592-025-02819-8). Use
mmseqs easy-search or easy-taxonomy with the --gpu flag.
-
For domain-specific taxonomy after QuickClade:
- Bacteria/Archaea -> run GTDB-Tk when genome/MAG-level sequence is available and cross-check NCBI/DIAMOND lineage assignments.
- Viral/phage -> route to
/bio-viromics; use PHROG/NCVOG markers and vConTACT3 only for phage/prokaryotic-virus contexts.
- Giant-virus/Nucleocytoviricota -> route to
/bio-viromics with GVClass and NCLDV marker-gene phylogeny.
- Eukaryota -> use EukCC for MAG/genome QC and lineage context; avoid CheckM/GTDB-Tk assumptions.
-
For group-appropriate marker families, run HMM searches against the relevant profile libraries (Pfam, TIGRFAM, COG/arCOG, PHROG/NCVOG for viruses, eukaryotic ribosomal/structural HMMs when applicable). Use pyhmmer (Python bindings around HMMER 3.4 with native SIMD and batch-friendly APIs) by default; fall back to the HMMER CLI (hmmsearch / hmmscan) when an upstream tool requires it. The choice of profile libraries is derived from the literature-derived playbook for the inferred group.
-
Build an annotation-wide feature inventory by genome/contig and by gene family/domain/pathway.
-
Marker-gene census — from the literature-derived playbook, list the diagnostic marker / machinery categories for the inferred group (e.g., replication, transcription, translation-related such as ribosomal proteins and translation factors, packaging, capsid/structural, chromatin/SMC/topoisomerase, host-interaction). For EACH query genome and each comparison-set genome supplied, record presence and copy number per category. Save as marker_census.tsv (columns: genome, category, family_id, family_name, copy_number, evidence_source, e_value, notes). Expected-but-absent markers are first-class rows, not silent omissions.
-
Per-family copy-number matrix — build a Pfam/InterPro/HMM-family × genome integer matrix covering queries AND the supplied relatives. Persist as family_copy_number_matrix.parquet. Compute per-family fold change vs the relative median; flag query-specific families, missing-expected families, expansions, and contractions in family_expansion_candidates.tsv.
-
For exploratory work, read the literature-derived analysis playbook for the inferred organism or virus group before deciding what to flag.
-
Mine the inventory for discovery candidates relative to that playbook: expected features, missing expected features, rare or expanded families, unusual combinations, annotation/taxonomy conflicts, and high-value unknowns.
-
For specialized inputs such as viruses, organelles, symbionts, pathogens, or poorly characterized lineages, use the feature classes and outlier dimensions reported in the relevant literature rather than a fixed global checklist.
-
Rank discovery candidates by evidence strength, novelty relative to the comparison baseline, confidence, and follow-up value.
