osmo_refdb: a DIAMOND+HMM reference database for osmoadaptation gene detection in metagenomic data

Reference DIAMOND (protein sequence search) and HMM (profile hidden Markov model) database for detecting osmoadaptation genes in metagenomic sequencing data, for use with osmotool. Covers 43 gene families spanning compatible-solute transport and biosynthesis (glycine betaine, proline, carnitine, choline, trehalose), Na⁺/K⁺ transport and antiporter complexes, mechanosensitive channels, glutathione redox cycling, cold-shock response, and related stress-response genes — including a dedicated Firmicutes/Bacillota gap-filling panel (Bacillus-type Opu transporters, the Mrp/Mnh Na⁺/H⁺ antiporter complex substituting for nhaA, and the Ktr K⁺-uptake system substituting for Trk) added to close phylum-level detection gaps in the original panel: Ectoine synthesis: ectA, ectB, ectC Trehalose synthesis: otsA, otsB Glutathione redox cycle: gshA, gshB, gshF, gor Choline → glycine betaine oxidation: betA, betB Compatible-solute transporters: betL, proX, proP, opuAA, opuBA, opuCA, opuAB, opuBB, opuCB, opuAC, opuBC, opuCC Mrp/Mnh Na⁺/H⁺ antiporter complex: mrpA, mrpB, mrpC, mrpD, mrpE, mrpF, mrpG Other Na⁺/H⁺ and K⁺ transport: nhaA, kdpA, trkA, trkH, ktrA, ktrB, ktrD Mechanosensitive channels: mscL, mscS Housekeeping / co-occurrence markers: galE, mazG, murB Cold shock: cspA Reference sequences curated from UniProt, with per-family gene-symbol queries checked for common annotation gaps (documented synonyms, organism-specific numbered paralogs). Negative-pool mislabeling fixed for 8 families (murB, mrpB, mrpE, mrpF, mrpG, gshB, cspA, mscS): for a family whose Pfam domain is essentially diagnostic for the gene itself, the standard "same domain, different gene symbol" recipe for building a hard-negative set instead selects real, unlabeled orthologs of the target gene. Six of these families previously deferred to Pfam's own curated HMM and gathering cutoff directly; that cutoff turned out to be miscalibrated for the short, fragment-level reads this database is actually used against and has been replaced with a custom-built, locally calibrated model for all six. Recall improved substantially across the board (e.g. gshB's HMM F1 0.47→0.88, mrpE 0.20→0.72), confirmed in a full 43-family rebuild with real cross-family competition present. Hard-negative pools were also audited for fused-ORF proteins (e.g. mrpA+mrpB, otsA+otsB — bifunctional/complex-forming variants under bare locus tags) sitting in a family's own negative set. The correction was kept only where full-panel benchmarking showed it helped (mrpA, otsA); it measurably hurt DIAMOND calibration for mrpB and otsB and was reverted there. Six families (murB, otsA, mrpC, trkH, ktrB, ktrD) are built and searchable for genome-level co-occurrence checks (osmotool annotate) but excluded from osmotool profile's reported read-level output, due to unresolved negative-pool or close-paralog issues (murB is excluded by original design, as a Culligan et al. 2012 co-occurrence marker). A handful of close-paralog transporter-subunit groups (proX/opuAC/opuBC/opuCC; the opuA/opuB/opuC ATPase and permease subunit trios; mrpA/mrpD) have a similar, unresolved specificity limitation. Documented per-family in this release's qc_scorecard.tsv and in the Github repository's README/CHANGELOG. Per-family score cutoffs calibrated against held-out sequence data and validated against both simulated and real-genome short reads. Panel-wide (read-volume-weighted), this build improves on the pre-fix baseline at DIAMOND F1 0.827→0.832 and HMM F1 0.693→0.722. Contents: osmo_refdb.dmnd — DIAMOND protein database osmo_refdb.train_refs.faa — the underlying reference sequences (train positives + decoy/fused-ORF references) DIAMOND was built from, in plain FASTA hmms/osmo_refdb.hmm (+ .h3f/.h3i/.h3m/.h3p) — pressed HMM database with per-family gathering (GA) cutoffs osmo_refdb.diamond_cutoffs.tsv, osmo_refdb.profile_cascade.tsv — per-family DIAMOND cutoffs and the DIAMOND+HMM cascade configuration used by osmotool profile osmo_refdb.profile_excluded_families.txt, osmo_refdb.annotate_excluded_families.txt — families excluded from reported output in each osmotool mode refs/, alignments/, results/, qc_scorecard.tsv — full build and benchmark provenance (QC'd reference sequences, alignments, per-family precision/recall/F1 benchmark output), included for reproducibility