ddn Resolved · high auto-curated

Annotation: from legacy to revised

Curation note: 2026-08-26 (feedback from the group's own BPaL/BPaLM resistance-emergence manuscript, mtbc/bpal_resistance_emergence project; submission in preparation for the Journal of Antimicrobial Chemotherapy after a desk-reject by Antimicrobial Agents and Chemotherapy, not yet peer-reviewed). Phylogenetic convergence across a 132,609-strain MTBC pangenome flags position 111 as a phenotype-free candidate priming variant, structurally proximal to the F420 cofactor pocket, uncatalogued and resting on structural/convergence evidence rather than DST (pretomanid has neither a catalogue entry nor public DST data). The manuscript's emergence-profiling test was itself validated on this gene: catalogued drug-era determinants L49P and W139* behave as expected (recent, drug-era), while the catalogued-but-clonal variant W88* is correctly flagged as a single-clone artefact rather than standing variation -- a negative control for the method, not a new finding. Full manuscript and code: https://github.com/cguyeux/deciphering-tuberculosis-with-ai/tree/main/bpal_resistance_emergence.

Auto-curated: this verdict and function were generated by rules from PGAP + Pfam + Foldseek and have not been hand-reviewed.

In the literature (TB corpus sweep) 72 publications

72 TB publications mention this gene. 72 publications discuss this gene. Queried separately, 72 records carry a M. tuberculosis context and 7 a non-TB mycobacterial context (species named in the 7-record verified sample: M. smegmatis (5), M. abscessus (2), M. leprae (1)). The three counts come from three distinct PubMed queries (each with its own context filter), so they neither partition the total nor need to add up to it.

This layer CITES the literature and adds context; it does not change the verdict or the function stated elsewhere in this fiche. This distinguishes a gene that is dark because nobody has looked from one that is dark despite having been studied. Source: PubMed (whole): H37Rv locus tag + GENE NAME + ortholog identifiers (Mb…, MMAR_…, MSMEG_…, ML…, MAB_…), under a mycobacterial context filter; hits verified against the abstract text. Species-context counts distinguish M. tuberculosis literature from literature on other mycobacteria. phase76/phase77, 2026-07-13.

CRISPRi vulnerability

Vulnerability index 1.53 (95% CI -0.55 to 4.96). A more negative index = more vulnerable to knockdown (better drug-target quality); indicative threshold VI ≤ -6 = highly vulnerable.

Quantitative CRISPRi knockdown, graded (finer than binary Tn-seq essentiality). Source: CRISPRi vulnerability index (Bosch 2021, supplementary Table S2 'mmc3.xlsx', bulk import via Europe PMC -- supersedes the per-gene pebble.rockefeller.edu scrape of phase46_crispri_vulnerability.py).

Caveat on this estimate. Bosch et al. flag this gene's vulnerability estimate as NOT certain, and here that call has a mechanical cause visible in the library itself: every guide targeting this gene is >=21 nt (str_span=0.079 on a 0-1 scale, near the genome floor), and guide repression strength saturates from 21 nt on -- so the library never sampled a partial knockdown for this gene, only the near-maximal repression plateau. The dose-response fit (and this VI) is therefore essentially unconstrained by data, not merely uncertain -- a limitation of the CRISPRi library design for the non-essential class (Bosch et al. 2021), not a defect in this specific measurement. n_guides=8 is descriptive only and does not predict this. Do not cite this VI as a standalone vulnerability claim.

Legacy record & comparison (Mycobrowser)

The legacy Mycobrowser record is shown for verification. Mycobrowser is no longer maintained; its EC numbers predate recent nomenclature revisions, so a class change usually reflects re-numbering, not a conflict.

Orthologues (reciprocal best hits across mycobacteria)

Reciprocal-best-hit orthologues (DIAMOND) against the Mycobrowser reference proteomes. A missing species is informative: e.g. a gene absent from M. leprae was likely lost in its reductive genome evolution. Locus tags link to Mycobrowser.

Curated reference (UniProt)

Functional vocabulary (eggNOG-mapper, orthology transfer)

Orthology-based transfer (eggNOG 5.0.2, diamond). EC/KO/GO/CAZy are computed annotations, not manual curation; cross-check against the primary literature before treating a specific reaction as established.

Conservation & selection (intra-MTBC, 142 601 strains, snapshot of 2026-09-29)

Read the disruption percentage against its ceiling, not against 100%. The strain collection is deeply imbalanced by clade, so a loss fixed in an entire lineage still shows a small overall fraction: a gene disrupted in every L6 strain caps at 0.96% of the collection, in every M. caprae strain at 0.20%, in every M. microti strain at 0.02%. Of the 705 clades in this collection, 689 have a ceiling below 1%. A figure of "0.15% of strains" is therefore uninterpretable on its own — it can be noise or a complete, lineage-defining loss. Use the clade attribution row above, which gives the sub-clade sizes, and never read the percentage alone.

pN/pS from segregating SNPs (singletons removed) normalised by possible sites. Low pN/pS = purifying selection (a strong signal that a "hypothetical" is a real, constrained gene). A high pN/pS is ambiguous: relaxed constraint or positive selection (drug resistance, antigenic variation) inflate it; e.g. rpoB/katG/pncA score high here for resistance, not loss of function. A clonal disruption (one allele over a clade) suggests lineage pseudogenisation; a convergent one (many independent alleles) is typical of resistance loss-of-function.

Outgroup conservation (beyond the MTBC) Actinomycetia

Two orthogonal outgroup signals. M. canettii (the immediate outgroup) gives a deep-divergence dN/dS (a low value confirms a constrained, real gene; shown as confident only at ≥8 substitutions, else flagged low-power). Genus-wide presence/absence (tblastn vs assembled non-MTBC genomes) places the gene on the ancient-core ↔ MTBC-specific axis: a gene absent even from the closest MTBAP relatives is a candidate MTBC-specific innovation (possible host-adaptation factor, to confirm by synteny). The phylostratum extends that axis outside the genus (tblastn vs 13 reference genomes spanning Mycobacteriaceae → Corynebacteriales → Actinomycetia → outside the phylum): it is the deepest clade in which a homolog is still detected, i.e. a proxy for gene age. Read it with the null model in mind: a shallow (young) stratum can also reflect homology-detection failure for short or fast-evolving ORFs, so it is a descriptive axis, not a proof of novelty.

Essentiality (transposon mutagenesis)

Genome-wide Himar1 transposon essentiality in H37Rv (DeJesus 2017). An essential call (ES/ESD/GD) is strong, independent evidence that a "hypothetical" locus encodes a functional, selectively required gene — orthogonal to intra-species conservation.

Conditional fitness (RB-TnSeq, 95 conditions) stress

Randomly-barcoded transposon screen across 95 carbon/nitrogen sources, pH, stressors and antibiotics (2 condition-specific phenotype(s) for this gene). A conditional fitness phenotype is a context lead, not a proven function, and never changes the verdict here. Note the blind spot: RB-TnSeq cannot measure essential genes. Source: RB-TnSeq 95-condition barcoded transposon screen, Mtb (PLoS Biol 2026, doi:10.1371/journal.pbio.3003529).

Drug resistance (WHO catalogue) delamanid

This gene carries mutations classed Associated with resistance (WHO grade 1–2) in the consolidated catalogue (WHO 2nd ed. 2023 + tb-profiler). Only the R-associated tier is shown; "uncertain" and empirical-only signals are excluded. Test a specific strain or variant with the resistance tester. Research context, not a clinical diagnostic.

Proteomics (mass spectrometry) detected

Detection by mass spectrometry is direct, experimental evidence that the protein product exists — orthogonal to sequence conservation and to Tn-seq essentiality, and especially decisive for a "hypothetical" locus. Read the dataset count as abundance, not as replication. Measured across this atlas (3,491 detected proteins), the number of datasets a protein appears in tracks its abundance almost monotonically: median 1.2 ppm at 1 dataset of 16, 27 ppm at 10, 392 ppm at 15, 1,325 ppm at 16 — three orders of magnitude. A high count therefore says the protein is abundant more than it says several independent experiments confirmed it. What the count does establish, at any value, is that the product exists.

Predicted localisation (DeepTMHMM + lipobox) no export signal, no TM helix

A verified absence of signals, which is not a compartment. No Sec/Tat signal peptide, no transmembrane helix and no lipobox were found on this sequence. The predictor reports topology and export signals; it does not measure where the protein ends up, so this page draws no cytoplasmic conclusion from it. In M. tuberculosis that reservation is not theoretical: the 12 Esx paralogues of this atlas — EsxA/ESAT-6 and EsxB/CFP-10 among them, the best-established secreted proteins of the species — carry this very call, because substrates of the type VII (ESX) secretion system leave no recognisable signal. « Secreted » and « no signal peptide » are true together here.

Transmembrane topology and signal peptide from DeepTMHMM (deep-learning reference predictor); lipoproteins from a (myco)bacterial lipobox motif. A sequence-based prediction of subcellular context.

Physico-chemical properties (computed, ProtParam)

Computed from the ancestral MTBC0 sequence with the ExPASy ProtParam method (Biopython). Descriptive biophysical context: a positive GRAVY flags a hydrophobic (often membrane) protein, a high instability index (>40) predicts a short in-vitro half-life, an extreme pI hints at compartment or binding partner.

Domains (Pfam, hmmscan --cut_ga)

3D structure (interactive — Mol*/pdbe-molstar)

Experimental structure 3r5l (PDBe), the strongest available structural evidence for this gene. Drag to rotate, scroll to zoom; hetero-groups (ligands, ions, waters hidden by default) shown as sticks.

Experimental structures (Protein Data Bank) 4 solved

Experimentally solved structures mapped from the UniProt accession via PDBe/SIFTS (4 total; up to 8 shown, ranked by sequence coverage then resolution). An experimental structure is direct proof of the folded product and the strongest structural evidence — superseding the predicted ESMFold/AlphaFold models below for any covered region.

Structural search (AlphaFold DB model, Foldseek vs PDB — genome-wide) pLDDT 91.7

Foldseek search of the AlphaFold DB model (mean pLDDT 91.7, gated at 70) against the PDB — a genome-wide extension of the ESMFold dark-gene search that also covers proteins beyond the single-sequence length limit. Confident structural neighbours (E < 0.01) shown.

Genomic context (neighbours & predicted operon)

Neighbours from the H37Rv annotation (+ strand). The operon is predicted by co-directional intergenic distance (same strand, gaps ≤50 bp) — a transcription-unit hypothesis, not a mapped TSS. For a "hypothetical", co-transcription with a characterised operon is a concrete functional lead (complements the STRING neighborhood channel below).

Published operon evidence (two independent sources)

Two independent published operon maps, cited alongside the distance heuristic above: an association corroborating or contradicting the heuristic, never proof of shared function. REMap's RNA-seq coverage continuity is the most direct of the three signals; Roback's logistic regression and the intergenic-distance heuristic are statistical proxies.

Transcriptional regulation (signed TRN: ChIP-seq + TFOE)

Regulatory edges from the ISB signed transcriptional regulatory network (TF ChIP-seq binding, Minch 2015 + TF-overexpression response, Rustad 2014). An edge is regulatory evidence (binding and/or expression change), not necessarily direct. For a "hypothetical", membership in a known regulon (e.g. DosR dormancy, PhoP virulence) is a physiological-context lead, not a function.

Condition: Condition-blind by construction. These edges come from transcription-factor OVEREXPRESSION in a single laboratory condition, so a listed regulator is not a claim that it governs this gene in every circumstance. Measured counter-examples exist within this atlas (see Rv2516c, where the DosR regulon collapses under moxifloxacin while the gene itself rises). Read a regulator as 'capable of shifting this gene', and look for the condition before building a hypothesis on it.

Node colour = verdict, dashed = still uncharacterised; edge colour = evidence (green experimental, orange genomic-context, grey co-expression), width ∝ score. Click a partner to open its page; "Explore full network" to walk the graph.

Closest characterised functional partner: fadA5 (acetyl-CoA acetyltransferase FadA), high confidence (score 778, all STRING channels except text-mining; the partner is chosen on its genomic-context evidence).

Showing the 15 strongest of this gene's 46 partners at medium confidence (17 at high confidence). The table is an extract, not a neighbourhood: across the proteome the real median degree is 34 while the served median is 15, so no hub, degree distribution or centrality may be read on this table. None of the 46 partners rests on an interaction measured in M. tuberculosis itself.

STRING combines evidence channels (neighborhood, fusion, cooccurrence, coexpression, experimental, database, text-mining) into a 0–1000 score. The ctx badge marks edges carried by the genomic-context channels (conserved neighborhood, fusion, phylogenetic co-occurrence), which are independent of orthology and structure and the strongest signal for an unknown gene. The no text-mining column recomputes the score from data alone, so a link that does not depend on the literature is visible. Association is a function hypothesis, not proof: corroborate with the operon context and the primary literature before assigning a function.

Where the experimental evidence comes from, and why it matters here. STRING's experimental channel fuses interactions measured in M. tuberculosis with interactions measured in another organism and transferred to this pair by orthology. The two are not equivalent for this atlas, whose standing rule is that better studied elsewhere is not a function established in M. tuberculosis. Measured on the served network: of the 38 229 partner pairs shown across the atlas, 13 483 partner rows carry a non-zero experimental channel, and only 106 pairs rest on an interaction measured in M. tuberculosis itself — 0.3 % of the network. Those 106 are long-established complexes (ESX-1/ESAT-6, RNA polymerase with CarD and RbpA, the Pup–Mpa proteasome, gyrase, Emb, GarA with PknB and PknG, the Rel modules); no still-uncharacterised gene is among them. The reason is not a gap in this atlas but in the literature: there is almost no experimental interactome of this bacterium, IntAct holding 134 curated protein pairs for a ~4 000-gene proteome. A exp (transferred) badge is therefore a lead, never a measurement, and an absence of badge is not evidence that the pair does not interact.

Evidence

• Legacy H37Rv annotation: deazaflavin-dependent nitroreductase
• MTBC0 PGAP product: deazaflavin-dependent nitroreductase Ddn
• Pfam (hmmscan --cut_ga): F420H2_quin_red PF04075.21 (E=5e-58)
• (auto-curated by rules from PGAP + Pfam + Foldseek; not hand-reviewed)

Ancestral MTBC0 protein sequence

>mtbc0_003764|Rv3547|ddn
MPKSPPRFLNSPLSDFFIKWMSRINTWMYRRNDGEGLGGTFQKIPVALLTTTGRKTGQPRVNPLYFLRDGGRVIVAASKGGAEKNPMWYLNLKANPKVQVQIKKEVLDLTARDATDEERAEYWPQLVTMYPSYQDYQSWTDRTIPIVVCEP