dna-decode 0.5.1

Deterministic, interpretable genome->phenotype decoder: calls antibiotic/antiviral/antifungal resistance (R/S) from DNA and names the exact genes/mutations driving each call, with its own blind spots + provenance. Spans bacteria, M. tuberculosis, fungi, HIV-1 and SARS-CoV-2. Not a clinical tool.

DNA_SEQUENCE_TO_TRAIT_PREDICTION

dna-decode — a deterministic, interpretable genome→trait decoder. Give it a genome (bacterial, fungal, mycobacterial, or viral); it returns a phenotype call (antibiotic / antiviral / antifungal resistance R/S, or E. coli pathotype) plus the exact genes/mutations that drove the call + its own blind spots + provenance. Mechanism-feature based, not an embedding black-box. Not a clinical tool.

2026-06 — the independent-label breakthrough. The project's binding constraint was a free, independent, measured phenotype label (everything else risks circularity — scoring a rule against another tool's predictions). It is now broken across bacteria (EBI AMR Portal measured AST: E. coli / Salmonella / Klebsiella / Shigella, acc 0.83–0.995), M. tuberculosis (WHO-2023-catalogue rule on N≈2,845 measured-AST isolates: rifampicin acc 0.937, isoniazid 0.914), and HIV-1 (Stanford HIVDB PhenoSense wet-lab fold-change). One legible view across every validation surface — each surface's distinct independence tier preserved, never averaged into a misleading aggregate: wiki/cross_kingdom_validation_summary.md.

What it decodes (v0.5.0)

Tool	Trait	Validation
dna-decode amr (bacterial)	antibiotic R/S — cipro / cef / tet / gent / meropenem across E. coli, K. pneumoniae, P. aeruginosa, S. aureus	6 drugs × 4 organisms, in-cohort + held-out + cross-source (NCBI) + cross-organism; every per-drug rule beats naive AMRFinder. Capstone: wiki/amr_multiorganism_capstone_2026-06-07.md
dna-decode amr (fungal, v0.5.0)	azole / echinocandin R/S — fluconazole / voriconazole / caspofungin / micafungin for Candida auris (BLAST-ERG11/FKS1 target-site engine)	kingdom-jump — same determinant-scan method, validated on a de-confounded C. auris WGS+MIC cohort (Gate G1): sens 1.0 across clades (ERG11 Y132F/F126L), label-limited specificity. wiki/fungal_ep7_g1_closeout_2026-06-08.md
dna-decode pathotype	E. coli pathotype (EPEC/EHEC/ETEC/UPEC/EAEC/…) compatibility + abstention	VirulenceFinder-marker resolver; ExPEC recall 0.917; rest documented scope-limit
dna-decode plasmid (v0.5.0)	plasmid Inc-replicon typing (IncF/IncH/IncI/IncX/IncN/…) — is the resistance plasmid-borne?	deterministic PlasmidFinder-blastn caller (identity 95 / coverage 60); faithful-to-tool (not an independent baseline); offline-safe degrade
dna-decode serotype (new)	E. coli O:H serotype (wzx/wzy/wzm/wzt O-antigen + fliC H-antigen)	deterministic SerotypeFinder-blastn caller (identity 85 / coverage 60); O?/H? when a locus is unresolved; offline-safe
dna-decode resfinder (new)	acquired AMR genes (ResFinder DB) — an independent cross-tool check vs amr	deterministic ResFinder-blastn caller (identity 90 / coverage 60); caller_is_independent_baseline: true (acquired genes only — no point-mutations/efflux); offline-safe
dna-decode pointfinder (new)	chromosomal AMR point mutations (PointFinder; v0 E. coli FQ QRDR gyrA/parC/gyrB/parE)	deterministic blastn + codon-position lookup vs resistens-overview; caller_is_independent_baseline: true (independent of amr's AMRFinder POINT); offline-safe
dna-decode disinfinder (new)	biocide/disinfectant resistance genes (DisinFinder; qac/form quaternary-ammonium + formaldehyde)	deterministic DisinFinder-blastn caller (identity 90 / coverage 60); often plasmid-borne (pair with coloc); offline-safe
dna-decode mlst (new)	MLST sequence type (PubMLST; v0 E. coli Achtman 7-gene) — exact-allele → profile → ST	deterministic blastn 100/100 + PubMLST profile lookup; validated: K-12 MG1655 → ST10; dna-mlst --fetch-db installs the scheme; novel/incomplete → ST not guessed; offline-safe
dna-decode ktype (new)	Klebsiella K-antigen (capsule) type via the wzi allele scheme (BIGSdb Pasteur, Kleborate-bundled) — the serotype sibling	deterministic wzi-blastn caller (identity 90 / coverage 80); self-consistency 15/15 across the DB; faithful-to-tool (wzi→K ~94%, NOT one-to-one); a free measured serological label exists (KlebNET-GSP 731-isolate set) → validatable, full caller-vs-serology run scoped (wiki/ktype_report_card.md); offline-safe

915+ tests green. 9 decoders (shared curated-DB blastn engine dna_decode/typing/blast_caller.py

• codon-mapping dna_decode/typing/codon_map.py) + 3 cross-decoder analyses that compose them:

Analysis	What
dna-decode concordance	AMR cross-tool check — AMRFinder (amr) vs ResFinder (resfinder) acquired-gene calls, gene-family level + Jaccard agreement	the independent second-opinion resfinder was built for
dna-decode profile	run-all — every assembly-FASTA decoder (pathotype+serotype+plasmid+resfinder+pointfinder) + AMR R/S calls with inline trust badges on one genome → one unified honest report	the "tell me everything, honestly" UX; each section degrades independently. AMR R/S needs a cached (--amrfinder-run) or Docker (--run-amrfinder) AMRFinder source; each call carries its validation tier (e.g. INDEPENDENT_MEASURED acc 0.95)
dna-decode coloc	AMR×plasmid co-localization — is this acquired resistance gene on the same contig as a plasmid replicon (likely plasmid-borne)?	turns "both present" into "the gene sits on the plasmid"; same-contig is suggestive, not proof
threshold + AMRFinder-Subclass / QRDR-point / gene-prefix refinement). Engineering principle that held
across every organism: count the drug's specific resistance determinants, not the broad drug-class bag.

Validated resistance cells — by kingdom (each with its own honesty tier)

Beyond the E. coli-family bacterial decoders above, the same determinant-scan method now spans four kingdoms. Each cell is validated on the strongest free label available and carries its honest independence tier — an in-distribution knowledge baseline is never relabelled as independent, and each kingdom keeps a namespace-separate standing report card so the tiers can't be conflated:

Kingdom	Cell	Independent validation	Tier
Bacteria	cipro / cef / tet / gent / meropenem × E. coli · Klebsiella · Salmonella · Shigella	EBI AMR Portal measured AST (free; BioSample/GCA-disjoint), acc 0.83–0.995	provenance-disjoint, measured — non-circular (wiki/amr_portal_independent_report_card.md)
M. tuberculosis	rifampicin (rpoB) + isoniazid (katG/inhA) — WHO-2023 catalogue rule	EBI AMR Portal measured AST, N≈2,845: RIF acc 0.937, INH 0.914	independent, measured (wiki/tb_report_card.md)
Virus — HIV-1	NNRTI / NRTI / PI / INSTI / CAI (RT · protease · integrase · capsid)	Stanford HIVDB PhenoSense wet-lab fold-change (NNRTI EFV AUC 0.962)	free, independent, isolate-level wet-lab label (wiki/hiv_decoder_report_card.md)
Virus — SARS-CoV-2	nirmatrelvir / ensitrelvir (Mpro / 3CLpro)	Stanford CoV-RDB fold-change — in-distribution, underpowered (37R/5S)	knowledge baseline, honestly labelled (wiki/sarscov2_mpro_validation_result_2026-06-23.md)
Fungus — C. auris	fluconazole / voriconazole / caspofungin / micafungin (ERG11 / FKS1)	de-confounded WGS+MIC cohort, sens 1.0 across clades; spec label-limited	kingdom-jump, G1-validated (wiki/fungal_ep7_g1_closeout_2026-06-08.md)

The bacterial 6-drug deployed surface is under a reproducibility freeze (wiki/reproducibility_freeze_2026-06-13.md) — frozen, sha-pinned, one-command reproducible. A learned-embedding alternative to the deterministic rules was tested to a decisive verdict and is a closed 0-for-4 negative (it learned population structure, not mechanism); the deterministic decoder suite is the validated, shippable artifact (wiki/negative_results_map_2026-06-13.md).

Every prediction carries its own trust badge inline. As of the productization pass, each dna-amr call emits a validation: line (and a validation block in the JSON record) reporting that exact cell's honest tier + headline metric + the standing report card it came from — e.g. validation: INDEPENDENT_MEASURED -- acc 0.919 (N=8778) for E. coli ceftriaxone, INDEPENDENT_WETLAB -- AUC 0.962 for HIV efavirenz, IN_DISTRIBUTION for SARS-CoV-2, NO_FREE_PHENOTYPE_SOURCE for the fungal cells, ABSTAINS_BY_DESIGN for the carbapenem abstainers. The honesty discipline is now user-facing at the CLI, not buried in the wiki (dna_decode/data/trust_surface.py; tiers never averaged, metrics never fabricated).

Organism-aware AMR calling (dna-amr --organism)

The per-drug DRUG_RULE is E. coli-tuned. Cross-organism validation (6 organisms × cipro/meropenem, N≈30 each, NCBI AST) found it fails to transfer in three distinct ways — a boundary taxonomy: CONTENT (counts intrinsic genes that don't confer R: Acinetobacter OXA-51, Pseudomonas nalC/oprD → over-call), TUNING (threshold wrong where a single mutation suffices: Campylobacter cipro), and EXPRESSION (regulation/derepression-driven R that gene-presence can't see: Enterobacter AmpC). Map + evidence: wiki/wider_amr_transferability_synthesis_2026-06-08.md.

dna_decode/eval/calibrate_organism.py auto-selects the per-organism config (determinant counter × threshold + intrinsic gene-family exclusions) from a ≥15R/15S labeled cohort by leave-one-out balanced accuracy, and abstains (EXPRESSION_FLOOR) when no presence-based config clears the floor (one-class/ under-powered cohorts → INSUFFICIENT_EVIDENCE). Validated configs ship in the committed registry dna_decode/data/calibrated_amr_rules.json (independent-cohort out-of-sample validated: wiki/calibrated_registry_independent_validation_2026-06-09.md). Pass dna-amr --organism <name> to use a calibrated config when one exists (Campylobacter / Klebsiella / Salmonella cipro); an EXPRESSION_FLOOR organism (Acinetobacter / Pseudomonas carbapenem) prints CALL: ABSTAIN rather than over-calling. The registry is opt-in — the default (no --organism, or an organism with no entry) uses the unchanged DRUG_RULE; calibrated configs are NCBI-AST in-sample-derived and stay opt-in pending a different-lab cohort.

Install

uv sync          # or: pip install -e .   -- DETERMINISTIC decoder core (no torch/transformers)
uv build && pip install dist/dna_decode-*.whl     # a built WHEEL also works -- it ships the trust cards
uv sync --extra ml   # ONLY for the foundation-model embedding track (a closed research negative)
# AMR genome mode also needs Docker + an AMRFinderPlus DB (see Gotchas); cached-run + observed-substitution
# modes are pure-Python. The default install no longer pulls the multi-GB torch/transformers/triton stack.

The built wheel ships the standing validation report cards as package data (dna_decode/report_cards/ via the pyproject force-include), so an installed dna-amr / dna-decode serves correct trust badges from the artifact — not just an editable checkout. Verify: uv run python scripts/verify_wheel_ships_cards.py --fresh-env.

New here? See QUICKSTART.md — pure-Python commands (no Docker, no [ml], no network) that run end-to-end, each verified by scripts/verify_quickstart.py.

Quickstart (verified output)

$ uv run dna-decode list
dna-decode 0.5.0 - deterministic genotype->phenotype decoders
  amr        antibiotic resistance R/S - bacterial (cipro/cef/tet/gent/meropenem) + FUNGAL azole/echinocandin (C. auris) + ANTIMALARIAL artemisinin/K13 + chloroquine/pfcrt-K76T (P. falciparum)
  pathotype  E. coli pathotype (EPEC/EHEC/ETEC/UPEC/EAEC/...) compatibility call + abstention

$ uv run dna-decode amr --drug ceftriaxone --amrfinder-run data/amrfinder_runs/GCA_008727135.1
sample: GCA_008727135.1  drug: ceftriaxone
CALL: R  [MODERATE | 1 determinant(s)]
  driven by: blaCMY-2  (CEPHALOSPORIN, 100.00% id)

$ uv run dna-amr --drug fluconazole --observed ERG11:Y132F --sample-id isolate1   # fungal, pure (no BLAST)
sample: isolate1  drug: fluconazole  organism: Candida_auris
CALL: R  [high | 1 determinant(s)]
  driven by: ERG11:Y132F

# Plasmid replicon typing on a genome assembly (blastn + PlasmidFinder DB; composes with amr):
uv run dna-decode plasmid path/to/assembly.fna --sample-id MY_STRAIN
# (downloads the DB once: curl -sSL https://bitbucket.org/genomicepidemiology/plasmidfinder_db/raw/HEAD/enterobacteriales.fsa -o data/plasmidfinder_db/enterobacteriales.fsa)

# Pathotype on a genome assembly (pure-stdlib, no Docker):
uv run dna-decode pathotype path/to/assembly.fna --sample-id MY_STRAIN

# AMR on a novel genome (genome mode — runs AMRFinder via Docker; --organism selects the AMRFinder -O):
uv run dna-decode amr --drug ciprofloxacin --genome-fasta X.fna --organism Klebsiella_pneumoniae

Full capability table + validation provenance: Shipped decoders below. The rest of this README is project history (how the tool was arrived at).

---

Project history — Phase 1 → v0.4.0 (how we got here)

The sections below are the chronological research record (embedding-thesis exploration, Evidence Packets, the deterministic pivot). For using the tool, the section above is all you need.

Status: Phase 1 — CLOSED 2026-05-17 (infrastructure + cross-drug architectural finding)

Phase 1 evidence collection closed 2026-05-17. Cross-drug architectural finding synthesis at wiki/ep1_ep2_cross_drug_architectural_finding_2026-05-17.md:

At 12-strain smoke fidelity, frozen-NT-whole-genome-pooling PASSES on concentrated-signal AMR mechanisms (cipro QRDR point mutations: AUROC 0.750; cef plasmid acquired-gene β-lactamases: AUROC 0.833) AND FAILS on distributed mobile-element mechanisms (tet tet-family efflux + ribosomal protection: AUROC 0.400, anti-predictive). The architecture's failure mode appears mechanism-class-bounded, largely independent of drug identity at smoke fidelity.

EP1 cipro closed internally (wiki/cipro_ep1_closeout_2026-05-17.md) with a 4-tier adversarial audit infrastructure (mechanism × MIC × opacity merge with structurally-enforced SUSPEND gate). EP2 cef + tet smoke fired (cef PASS, tet FAIL, H17 falsified). No Databricks burst spent. External publication deferred per PC1=internal_closeout.

Phase 1 code: all 18 implementation steps shipped Wave 0-7 (2026-05-11 → 2026-05-12) + 3 hardening waves; cross-drug Evidence Packet evidence collection completed 2026-05-17 per the Evidence Packets framing reset 2026-05-15. Phase 2 entry fired 2026-05-18: BV-BRC strict-MIC 4-drug feasibility census ran (scripts/bvbrc_strict_mic_4drug_census.py + wiki/bvbrc_strict_mic_4drug_census_2026-05-18.{md,json}) — NO drug clears N=150 per-class at either strict-MIC or relaxed-MIC bars; structural bottleneck is assembly_accession. North star clarified: AI DNA decoder tool, not papers. v0 UX + success criteria LOCKED at wiki/decoder_v0_ux_and_success_criterion.md (CLI via pipeline.py predict, LOSO AUROC ≥ 0.70, cipro v0 / cef v0.1, JSON + markdown sidecar). 3 of 5 v0 criteria green via 24 new tests; 2 gated on Databricks N=147 cipro cache landing.

Phase 2 in-flight (2026-05-22 → 2026-05-24): cipro interpretability audit completed on Precision 7780 (RTX 3500 Ada) by parallel Codex CLI session. Bounded-falsifier coordination plan + post-falsifier ship-path technical plan covered all 4 verdict branches × 3 gate states pre-committed. Codex on Precision 7780 ran the falsifier 2026-05-23 — verdict = FAIL (ranking-only rescue did not improve the ELX-family failure case on 12-strain Bucket B). Per the FAIL branch + north star, v0 shipped 2026-05-24 as a cached-strain cipro predictor (scripts/pipeline.py predict --strain-id ...) with a documented scope-limit. v0 spec RELOCKED at wiki/decoder_v0_ux_and_success_criterion.md to match the implemented cached-strain surface (not the original genome-input decoder concept). Leakage-safe retrain on leave_one_accession_out CV yielded AUROC 0.8697. v0 closeout handoff: wiki/dna_decoder_v0_closeout_handoff_2026-05-24.md.

v0.1 cipro genome-input slice 1 LANDED (2026-05-25): Codex shipped pipeline.py predict --genome-fasta X.fna --annotations Y.gff3 end-to-end. Cross-path concordance validated on 4-strain mixed panel + same-strain parity (max prob delta 0.011599). Live embedding in batched chunks (OOM fix). Audit fallback to cohort-level framing when sample missing from per_strain.

v0.1 cef cached-strain LANDED + validated (2026-05-25 → 2026-05-26): Codex shipped a dedicated 67-strain NT cache + trained cef classifier (CV AUROC 0.895 / AUPRC 0.838 on N=49 usable, 25R/24S). Duplicate-accession audit PASS (no LOSO leakage). Full-panel cached-vs-genome-input validation: 49/49 prediction concordance, 47/49 label alignment, max prob delta 0.063. 2 shared model misses (562.28389 FP, 562.7695 FN) at decision-boundary probabilities. Currently debug-mode (no audit sidecar yet).

v0.1 cef audit-aware packet (IN FLIGHT 2026-05-26): closeout slice per plans/Cef_Audit_Aware_Packet_Design.md — 5 artifacts (AMRFinder mechanism audit + cef MIC tier audit + new scripts/drug_mechanism_phenotype_merge.py + 4 canonical predict examples in canonical_audit_aware mode + release packet update with pre-committed verdict-branch wording). ~3.5-4 hr Precision 7780 compute. Closes the last debug-mode gap.

Long-horizon roadmap drafted (2026-05-26): plans/Trait_Decoding_Roadmap.md maps Phase 0 (v0 cipro) → Phase 6 (eukaryotic organisms) with per-phase terminal claims + dataset prerequisites + falsifier triggers. EP-4 first non-AMR phenotype scoping: pathotype prediction (EnteroBase substrate; multiclass EPEC/EHEC/ETEC/UPEC/EAEC/commensal) per plans/EP_4_Non_AMR_Phenotype_Candidates.md.

Test count: 847 green (as of v0.4.0, 2026-06-07).

See plans/EP1_EP2_Cross_Drug_Synthesis_Plan.md for the synthesis plan; plans/Cipro_Decision_Bundle_Plan.md + plans/Cipro_Decision_Bundle_Technical_Plan.md for the EP1 closeout planning chain; plans/EP2_Cef_Tet_Smoke_Design_Plan.md for the EP2 design. See plans/Ecoli_G2P_Phase1_Ship_Path_Plan.md for the original Phase 1 contracted ship-path. See plans/Ecoli_G2P_Platform_Technical_Plan.md for the full Phase 1 plan with Tier 1-5 attribution-success framework. See docs/ARCHITECTURE.md for the module map.

What runs end-to-end today:

Surface	Entry point	Notes
Pilot gate (HARD)	python -m scripts.pilot_gate --ast-tsv <path>	Validates per-drug strain counts before ingestion fires. Exit 0=GO, 1=NO-GO, 2=PilotGateError, 3=no source.
Full pipeline	python -m scripts.pipeline {ingest, train, predict, attribute}	Single CLI with 4 subcommands; shared config-driven path resolution.
Smoke regression	python scripts/smoke_pipeline.py	<60s synthetic-fixture end-to-end via MockFoundationModel; asserts AUROC ≥0.85 + top-1 attribution = seeded gene.
Leaderboard fan-out	python scripts/leaderboard.py --drugs ... --models evo,dnabert2	Loops pipeline.py train per (model × drug); writes data/processed/leaderboard.md.
Quant-fidelity check	python scripts/quantize_fidelity_check.py --full-precision-attributions <manifest.json> --quantized-attributions <manifest.json>	One-time 4-bit vs full-precision ISM concordance check; gates whether Phase 1 attribution numbers are quantization-conditional.
Viz	dna_decode.viz.browser.render_attribution_plot + export_attribution_tsv	matplotlib PNG + TSV export; pygenometracks deferred to Phase 2.
BV-BRC strict-MIC 4-drug feasibility census	python -m scripts.bvbrc_strict_mic_4drug_census	Phase 2 entry (2026-05-18). Per-drug feasibility at strict + relaxed bars for cipro/cef/tet/gent. Writes wiki/bvbrc_strict_mic_4drug_census_<date>.{md,json}. Imports from dna_decode/data/mic_tiers.py (shared per-drug catalogs).
v0 decoder predict	python -m scripts.pipeline predict --strain-id X --model-path M.pkl --cache C.h5 --annotations G.gff3 --audit-merge-json A.json --output Y.json	v0 schema per wiki/decoder_v0_ux_and_success_criterion.md (2026-05-18 LOCKED). Emits JSON + markdown sidecar with prediction + calibrated_probability + confidence_tier + top_k_attribution + audit_verdict (SUSPEND propagation) + provenance.
Provenance-disjoint validation	python scripts/provenance_disjoint_validate.py --drug X --organism Y ...	Scores a deployed decoder on a submitter/lab-disjoint NCBI-PD cohort. Leakage exclusion via the data-driven dna_decode/eval/cohort_manifest.py (EXACT-self identity over all raw+parquet cohorts); FAILS CLOSED on an incomplete manifest (INCOMPLETE_MANIFEST, exit 2) unless --allow-incomplete-manifest. Powering census scripts/ncbi_pd_provenance_census.py self-persists to wiki/provdisjoint_census_results.json.
Decoder-suite validation report card	python scripts/build_validation_report_card.py	Read-only roll-up (exit 0 always — a report, not a gate). Rows = deployed-claim surface (dna_decode/data/shipped_decoder_surface.py) ∪ observed cells; honest per-cell tier, no aggregate headline. Renders a Lineage disclosure (clonality-corrected) table from the lineage sidecar when present. Writes wiki/decoder_validation_report_card.{md,json}.
Lineage-disclosure metrics	python scripts/compute_lineage_metrics.py	Recomputes per-cohort clonality-corrected sens/spec — the report card's raw sens/spec counts one vote per isolate, so an over-sampled clone inflates it. Greedy-representative Mash clustering (dna_decode/eval/clonality.py; chaining-resistant, NOT single-linkage), cluster-weighted confusion + Wilson CI + effective-lineage-N, graded lineage bucket. M4 reconciles raw sens/spec vs the committed artifact before trusting any weighted number. Needs Docker (Mash). Writes wiki/provdisjoint_lineage_metrics.json (schema provdisjoint-lineage-metrics-v1).
External re-validation — Gate-0 preflight	python scripts/external_cohort_preflight.py --project PRJNA604975 --cohort-name oxford [--mic-open|--mic-gated]	Wave-0 go/no-go for re-validating the FROZEN decoder on an INDEPENDENT measured-MIC cohort. Bidirectional Entrez-primary/ENA-fallback BioSample resolver (dna_decode/eval/biosample_resolver.py) closes the accession-string leakage blind spot; emits assembly-availability (FREE vs ASSEMBLY-REQUIRED) + a BioSample-level leakage verdict (FAIL-CLOSED if any tuning overlap, >5% unresolved, or Entrez/ENA disagreement) + MIC-openness. Writes wiki/external_preflight_<cohort>_<date>.json.
External re-validation — scorer	python scripts/external_cohort_revalidate.py --cohort oxford --drug ciprofloxacin --labels-dir D --preflight-json P.json	Mirrors provenance_disjoint_validate (ensure_run → call_resistance → _conf) on the external cohort, into a SEPARATE external-validation-v1 namespace (evidence_tier=external_clinical). Organism triple pinned VERBATIM from the frozen E. coli cells (AMRFinder -O Escherichia, registry Escherichia_coli_Shigella). MIC→R/S via dna_decode/data/external_mic_labels.py (classify_tier strict=HIGH_R/HIGH_S primary, relaxed=+DECISIVE secondary, bucket counts — NOT naive thresholding). Fail-closed unless preflight PASS. Writes wiki/external_validation_<cohort>_<drug>_<date>.json.
External re-validation — roll-up	python scripts/build_external_validation_report.py --run-id R [--allow-degraded]	Run-scoped (refuses glob-all without --run-id/--artifacts/--allow-unscoped-glob); globs the separate external_validation_* namespace → wiki/external_validation_report_card.{md,json}; per cell raw strict/relaxed + cluster-weighted STRICT sens/spec + Wilson CI + effective-lineage-N (reuses clonality.py math inline; Docker Mash, degrades to raw on non-Docker hosts). Skips powering.hard_fail cells; degraded only with --allow-degraded. FROZEN decoder report card + compute_lineage_metrics NOT touched (Fix C).
Oxford ingestion — W0 probe	python scripts/oxford_w0_probe.py --project PRJNA604975 --mic-table T --key-col K --drug-col COL=alias	Pins the MIC-table schema + crosswalk feasibility BEFORE the ingester codes against it: ENA candidate-field cardinality, row/key/dup summary, operator/censoring distribution, MIC-key→BioSample resolution rate → wiki/oxford_w0_probe_<date>.json.
Oxford ingestion — labels + manifest	python scripts/build_oxford_labels.py --project PRJNA604975 --mic-table T --key-col K --drug-col COL=alias --run-id R	Ingest → alias→BioSample crosswalk (ABORTS on hard conflict) → per-drug selected_{strict,relaxed}.tsv + buckets + the single cohort_manifest_external_<run_id>.json (the exact scored-cohort definition).
Oxford ingestion — one-command driver	python scripts/run_oxford_revalidation.py --project PRJNA604975 --mic-table T --key-col K --drug-col COL=alias --drugs ciprofloxacin	Chains W0 probe → labels → exact-set preflight (abort != PASS) → per-drug scorer → roll-up ONLY IF every drug run is acceptable (driver gating). Exit = worst child (3 hard-fail > 1 degraded > 2 gate > 0).

Module map: dna_decode/data/ (ingestion) + dna_decode/models/ (foundation wrappers + cache + classifiers + classical baselines; cache.verify_complete integrity gate added 2026-05-15) + dna_decode/interp/ (ISM + Tier 1-5 attribution) + dna_decode/eval/ (CV + metrics + batched-call phylogeny + clade-only baseline) + dna_decode/viz/ (browser) + tools/ (Stage 2 bioinformatics-tool runner via Docker Desktop — Mash + AMRFinderPlus + Bakta).

Phase 1 scope

Aspect	Value
Organism	E. coli
Phenotypes	Ciprofloxacin, ceftriaxone, tetracycline binary resistance
Foundation models	Evo (primary), DNABERT-2, Nucleotide Transformer, GENA-LM (leaderboard)
Classical baselines	AMRFinder gene calls, k-mer logreg + XGBoost, gene-presence XGBoost (Step 18)
Baseline ML	Frozen foundation-model embeddings + XGBoost per drug
Attribution	In-silico mutagenesis (gene-level + nucleotide-level saturation)
CV	Leave-one-Mash-clade-out + clade-only baseline + per-clade reporting
Target	AUROC ≥0.80 SLO / ≥0.85 stretch; clade-baseline-gap ≥0.10 on ≥75% of held-out clades; ≥3pp gap vs best classical baseline on ≥2 of 3 drugs
Horizon	3 months Phase 1; 12 months Phase 1+2+3
Compute	Local GTX 860M (4 GiB Maxwell, NT v2 only — verified 2026-05-13) + Databricks burst for larger cohorts. 4-bit Evo unavailable (bitsandbytes requires CC ≥ 7.0). Original target was RTX 4090 + 4-bit Evo; never materialized.

Long-term vision

Multimodal genotype-phenotype platform — start with bacterial AMR (Phase 1), expand toward eukaryotes + image-paired phenotype data in later phases. NOT a direct stepping stone to "DNA → animal image" prediction; that would require a parallel multimodal track.

Setup

# 1. Install uv (if not already on PATH)
#    Windows PowerShell:  irm https://astral.sh/uv/install.ps1 | iex
#    Linux/macOS:         curl -LsSf https://astral.sh/uv/install.sh | sh

# 2. Sync deps. pytest is in default deps (Wave 1.5 hardening fix):
uv sync

# 3. Run the test suite
uv run pytest tests/ -v

# 4. Optional: install dev tooling (ruff + pytest-cov)
uv sync --extra dev

# 5. (Advanced, gated on hardware) install bitsandbytes for 4-bit Evo quantization
#    Requires CC ≥ 7.0 GPU (Ampere / Ada / Hopper). NOT compatible with the project's
#    actual GTX 860M (CC=5.0). Skip unless running on A100+ or similar.
uv sync --extra quantize

Phase 1 quickstart

End-to-end Phase 1 run (assumes uv sync + BV-BRC AST TSV downloaded + Mash CLI installed):

# 1. HARD gate: confirm you have enough labeled strains per drug
uv run python -m scripts.pilot_gate \
  --drugs ciprofloxacin,ceftriaxone,tetracycline \
  --target-per-drug 150 \
  --ast-tsv path/to/bvbrc_ast.tsv

# 2. Smoke test: <60s end-to-end on synthetic fixtures (sanity check before real run)
uv run python scripts/smoke_pipeline.py

# 3. Ingest: build cohort + download cohort genomes
#    For real-data runs, pass --assembly-metadata-csv pointing at the BV-BRC
#    Genomes-tab export (CSV). The adapter at dna_decode/data/bvbrc_genome.py
#    feeds contig_count + N50 + MLST + assembly_accession into
#    candidates_from_bvbrc_ast. Both --assembly-metadata (legacy YAML) and
#    --assembly-metadata-csv are mutually exclusive.
uv run python -m scripts.pipeline ingest \
  --drugs ciprofloxacin,ceftriaxone,tetracycline \
  --ast-tsv path/to/BVBRC_genome_amr.csv \
  --assembly-metadata-csv path/to/BVBRC_genome.csv \
  --download-genomes

# 4. Populate the embedding cache (deferred — see ARCHITECTURE.md for the wiring;
#    embedding cache populate is invoked from a Phase 2 helper script that hasn't
#    shipped yet; Phase 1 callers populate the cache externally via cache.populate()).

# 5. Train per-drug classifier + run CV + emit clade-only baseline + validation gate
uv run python -m scripts.pipeline train \
  --drug ciprofloxacin --model evo --include-clade-baseline

# 6. Run ISM attribution + Tier 1-5 classification for one strain
uv run python -m scripts.pipeline attribute \
  --strain-id <bvbrc-strain-id> \
  --drug ciprofloxacin \
  --card-path path/to/card.json \
  --amrfinder-path path/to/amrfinder.tsv \
  --output data/processed/attribution_report.json

# 7. Build leaderboard across foundation models + classical baselines
uv run python scripts/leaderboard.py \
  --drugs ciprofloxacin,ceftriaxone,tetracycline \
  --models evo,dnabert2

# 8. (Optional, gated on CC ≥ 7.0 GPU) Validate that 4-bit Evo attribution matches full-precision
uv run python scripts/quantize_fidelity_check.py \
  --full-precision-attributions full_manifest.json \
  --quantized-attributions quantized_manifest.json \
  --drug ciprofloxacin

Decoder v0 quickstart (Phase 2 in-flight)

The v0 AI DNA decoder operates on cached strains — a strain whose NT embeddings already live in the HDF5 cache (built by pipeline ingest + the Databricks N=147 cipro populate). UX + success criteria locked in wiki/decoder_v0_ux_and_success_criterion.md.

# Predict cipro R/S for a cached strain, with top-K attribution + audit-verdict propagation.
uv run python -m scripts.pipeline predict \
  --model-path data/processed/models/ciprofloxacin_nucleotide_transformer.pkl \
  --strain-id 562.12345 \
  --cache D:/dna_decode_cache/embeddings/nt_n147_cipro.h5 \
  --annotations D:/dna_decode_cache/refseq/GCF_xxx.x/annotations.gff3 \
  --audit-merge-json wiki/cipro_mechanism_phenotype_merge_2026-05-17.json \
  --output result.json

Writes result.json + result.md (markdown sidecar) per the v0 schema:

• prediction (R/S) + calibrated_probability + confidence_tier (HIGH/MEDIUM/LOW)
• top_k_attribution — gene-level ISM hits with resistance-catalog tier labels (Tier 1–5)
• audit_verdict — propagated from the merge gate; explicit suspend_gate_fired flag + verdict explanation when training cohort had SUSPEND_CONDITION_4
• provenance — model, training cohort, LOSO AUROC, trained-on date

Not a clinical decision support tool. Audit verdict + provenance must accompany any downstream interpretation. See wiki/decoder_v0_ux_and_success_criterion.md for full v0 schema + success criteria.

Shipped decoders (v0.4.0) — two interpretable E. coli genome→trait tools

The project's delivered value is two deterministic, interpretable decoders (installable console commands after uv sync / pip install -e .). Both take a genome assembly and emit a call + the exact genes/mutations that drove it + provenance — biologically interpretable, not embedding black-boxes.

Command	Trait	What it reports	Validation
dna-pathotype	E. coli pathotype (EPEC/EHEC/ETEC/UPEC/EAEC/…)	virulence-cluster compatibility call + abstention + canonical-VirulenceFinder diff	compatibility resolver; ExPEC/EPEC/ETEC supported, rest documented scope-limit
dna-amr	antibiotic resistance (R/S) — cipro / cef / tet / gent / meropenem; E. coli + Klebsiella + Pseudomonas (--organism)	R/S call + the curated AMRFinder resistance determinants driving it (e.g. gyrA_S83L, blaCTX-M-15, aac(3)-IIa) + undetectable_mechanisms blind-spots on S calls	cipro E. coli N=147 acc 0.925 (cross-organism QRDR-POINT rule); cef N=60 0.933; gent N=128 0.945; tet N=12 0.833. Cross-source (NCBI, zero BV-BRC overlap): cipro 1.0 / cef 0.864 / gent 1.0 / tet 0.909, beating naive AMRFinder. Cross-ORGANISM: full Klebsiella 5-drug matrix — cipro 1.0 / cef 0.80 / gent 0.867 / meropenem 0.867 (all ✅); tet 0.80 (sens-limited by efflux). 3rd organism Pseudomonas cipro N=30 acc 0.867; 1st Gram-positive S. aureus oxacillin/mecA sens 1.0 (genotype transfers; oxacillin-label-confounded spec — wiki/staphylococcus_aureus_oxacillin_validate_2026-06-07.md). Per-drug rules in amr_rules.py::DRUG_RULE; genome mode takes --organism

# Unified entry (dispatches to the trait decoders):
uv run dna-decode list                                  # what it decodes + per-trait validation status
uv run dna-decode pathotype path/to/assembly.fna --sample-id MY_STRAIN
uv run dna-decode amr --drug ciprofloxacin --amrfinder-run data/amrfinder_runs/GCA_xxx.x

# The per-decoder entries remain independently usable:
uv run dna-pathotype path/to/assembly.fna --sample-id MY_STRAIN
uv run dna-amr --drug ceftriaxone --genome-fasta X.fna   # genome mode needs Docker + AMRFinder DB

Why deterministic, not embeddings: the frozen-genome-embedding (NT-mean-pool) thesis was tested to a decisive verdict and found to have no E. coli AMR niche — on the cleanest substrate (cipro) it lost to the QRDR-POINT knowledge baseline and within-lineage scored at chance (it learned lineage, not mechanism). See plans/AMR_embedding_niche_decision_2026-06-05.md. For AMR, mechanism features win and are interpretable; the embedding architecture's remaining open frontier is non-AMR phenotypes lacking a curated knowledge baseline (gated on a de-confounded labeled substrate; wiki/HANDOFF_session_2026-06-05.md).

Pathotype resolver (E. coli) — v0 tool (SHIPPED 2026-06-04, tag pathotype-v0)

A self-contained, pure-stdlib CLI that takes an E. coli genome assembly (FASTA) and emits an auditable pathotype-compatibility call with virulence-cluster provenance + a side-by-side diff against canonical VirulenceFinder. Honest framing: it is a marker-based compatibility resolver with abstention, NOT a clinical predictor. Supported (externally-valid) classes = ExPEC / EPEC / ETEC; EAEC / commensal / clean-EHEC are a documented scope-limit (the resolver reports their modules but flags low external validity).

# After `uv sync` (or `pip install -e .`), the `dna-pathotype` command is available:
uv run dna-pathotype path/to/assembly.fna --sample-id MY_STRAIN --out result.json
# or equivalently:  uv run python -m dna_decode.pathotype path/to/assembly.fna ...

Emits provenance JSON + a human summary:

• derived_call — 11-class honest surface (EHEC/STEC/tEPEC/aEPEC/ETEC/EAEC/UPEC/HYBRID/AMBIGUOUS/ UNCLASSIFIED/COMMENSAL) with confidence_tier + external_validity + abstention rules. --legacy-6class preserves the original 6-class promise.
• cluster_profile + marker_hits — which virulence clusters drove the call (k=15 k-mer-seed coverage over the VirulenceFinder E. coli allele DB; ≥0.80 = confident).
• vf_diff — canonical VirulenceFinder side-by-side via real blastn over the SAME VF DB: per-gene + per-cluster concordance. HONESTY: both callers use the same DB, so caller_is_independent_baseline: false + a same-DB caveat ship in every diff — it is an AUDIT of the fast caller, not independent validation. Degrades to status: unavailable (never dropped) when blastn is absent. Use --no-vf-diff to skip.

BLAST+ for the diff: install NCBI BLAST+ (blastn + makeblastdb) and either put it on PATH or set $BLASTN_BIN. Without it, the resolver still runs fully; only the canonical diff degrades to unavailable.

Marker DB: data/virulencefinder_db/virulence_ecoli.fsa (fetch from the VirulenceFinder Bitbucket DB; the CLI prints the exact curl command if it's missing). DB checksum is pinned in every provenance record.

Phase 1 success criteria

Phase 1 ships when:

• Smoke pipeline passes (scripts/smoke_pipeline.py returns exit 0)
• LOMO-clade-out CV AUROC ≥0.80 SLO / ≥0.85 target per drug
• Embedding model AUROC ≥0.10 above clade-only baseline on ≥75% of held-out clades
• Top-K=20 attribution-tier distribution: cipro ≥40% Tier 1-3 hits; ceftriaxone ≥25%; tet ≥30%; all ≤20% Fail
• Best foundation model beats best classical baseline by ≥3pp AUROC on ≥2 of 3 drugs
• Quantization-fidelity check returns GO (mean Spearman ≥0.7, intersection ≥0.6)

Phase 2 redesign trigger: classical baselines win on ≥2 drugs. The Step 18 classical-baselines control wires this empirically — see plans/Ecoli_G2P_Platform_Technical_Plan.md validation-gate section.

Pilot gate alternate inputs

• BVBRC_AST_TSV=path/to/ast.tsv env var
• bvbrc_ast.local_tsv_path: path/to/ast.tsv in config/datasources.yaml

Optional: route caches to a USB drive

Phase 1 runtime needs ~25GB (foundation models + strain genomes + embeddings). If your C: drive is tight, route caches to external storage:

# Replace E: with your drive letter (Windows) or /mnt/usb (Linux)
export HF_HOME=E:/hf_cache               # HuggingFace tokenizer + model cache
export DNA_DECODE_CACHE_ROOT=E:/dna_decode_cache

Then edit config/datasources.yaml to point cache_dir fields at the USB-backed path.

Project workflow

Built using a personal Claude Code skill ladder for project planning:

• /idea-anchor → /project-init → /brainstorm ×3 → /technical-plan → /probe → /execute-plan
• Project ledger maintained via /project-state skill
• Execution state tracked in .claude/execute-plan-state/Ecoli_G2P_Platform_Technical_Plan.json
• All planning artifacts captured as audit trail

See project_state/dna-decode-2026-05-11.md for full decision history (17 hypotheses, 12+ decisions made, 54+ action-log entries as of 2026-05-17; "Phase 1 / 2 / 3" labels retrospective-only — new work tracked as Evidence Packets per the 2026-05-15 framing reset; Phase 1 evidence collection CLOSED 2026-05-17 with the cross-drug architectural finding synthesis).