Gen dna qc

gen-dna-qc — a legible design-for-manufacture gate for generated DNA

Genomic sequence generators (Evo2, and other DNA language models) emit DNA that scores well on the model's own confidence yet can be unmanufacturable — GC outside the synthesis window, a homopolymer run that slips the synthesizer, a hairpin that won't anneal, a cloning-site collision, or (across a generated batch) two sequences that hybridize to each other instead of their targets.

gen-dna-qc is the programmatic version of that check: a deterministic design-rule check (DRC) that emits a pass/fail verdict with a human-readable reason for every flag — the "unroutable net" report for a generative model's output. It is the complement to a generator, not a competitor — it qualifies the output, it does not generate sequence.

It owns two ownership levels (a per-sequence fact and a design-level invariant no single sequence owns):

contract	tier	catches
GC_OUT_OF_BAND	fails	GC fraction outside the synthesis envelope (default 25–65%)
HOMOPOLYMER_RUN	fails	a run of one base longer than the cap (synthesis slippage)
LENGTH_OUT_OF_RANGE	fails	below the oligo floor / above the single-fragment gene window
STRONG_HAIRPIN	fails	a long self-complementary stem — folds on itself, won't synthesize/anneal
SEVERE_CROSS_HYBRIDIZATION	fails	two sequences in a batch anneal to each other (not their targets)
POLY_G_RUN	discloses	GGGG+ — a G-quadruplex risk
RESTRICTION_SITE	discloses	recognition-site collisions — will be cut if cloned with those enzymes
CROSS_HYBRIDIZATION	discloses	a moderate complementary stretch between two batch sequences

The verdict separates disclosure from condemnation: every hazard is reported, but only the synthesis-breaking ones fail the structure — a restriction site or a poly-G run informs without condemning, because they're cloning/risk notes, not synthesis failures. Thresholds are commercial-synthesis constants / DnaChisel conventions — zero parameters fitted to accuracy.

Install

pip install karyon          # no extras needed — pure stdlib

Usage

--modality dna is required (a .fasta/sequence could equally be a σ70 promoter — see promoter-design):

# A single generated sequence (inline):
karyon qualify ACGTACGT... --modality dna

# A batch (multi-record FASTA, e.g. Evo2 output) — adds the cross-hybridization set check:
karyon qualify evo2_designs.fasta --modality dna

# JSON verdict for piping into an agent / pipeline:
karyon qualify evo2_designs.fasta --modality dna --json

Output is a PASS / FAIL verdict plus, per sequence, one line per fired contract — a · for a disclosed hazard, an ✗ for a condemning one (e.g. "strong hairpin: a 24 bp self-complementary stem (≥12) across a 6 nt loop — folds on itself, won't synthesize/anneal cleanly"). Exit code is non-zero on FAIL so it gates a pipeline directly. --json emits the stable spine schema — {modality, ok, items:[...], batch} — where the set-level cross-hybridization verdict (multi-record only) rides in batch; a design passes iff score == 0.

From Python:

from karyon import qualify
r = qualify("evo2_designs.fasta", modality="dna")     # or qualify("ACGT...", modality="dna")
for name, v in r.items:                               # per-sequence verdicts
    if v.score > 0:
        print(f"REJECT {name} — {v.messages}")
if r.batch and r.batch.score > 0:                     # cross-hybridizing pairs across the batch
    print("batch:", r.batch.messages)

Composition with NVIDIA BioNeMo

Install alongside evo2-nim (DNA generation): the model generates the sequence; this skill gates the batch, so the agent only carries forward synthesizable candidates and can explain every rejection. Generate with the BioNeMo Evo2 skill, then pass the output to gen-dna-qc for a deterministic manufacturability verdict — the packaging mirrors BioNeMo's SKILL.md convention so the two compose cleanly.

Validation

The gate is validated as a real instrument and faithful to two gold-standard packages, with three pre-registered predictions (all PASS):

prediction	result
PI-1 instrument — clean sequences pass, planted decoys flagged	AUROC 1.000, flag-decoy 100%, pass-clean 100%, real E. coli CDS pass 100%
PI-2 faithful — owned verdict vs DnaChisel (EnforceGCContent + AvoidPattern + AvoidHairpins)	per-sequence agreement 100% (GC / homopolymer / hairpin all 100%); hairpin signal vs ViennaRNA ΔG AUROC 0.88
PI-3 effect — synthesis/cloning hazard rates per generator (descriptive)	honest weak-condemn (random DNA is mostly synthesizable under lenient vendor rules) but high disclose rates (restriction sites, poly-G) — the hazards a generator is blind to

Thresholds are commercial-synthesis constants, calibrated to the DnaChisel reference and fixed before the runs — the agreement is not fitted. Qualification, not accuracy: the gate reports what won't synthesize or clone, it does not make the generator better.

Scope (honest)

A fast, legible manufacturability/usability gate (the cheap, certain checks over string geometry: GC, runs, hairpin, cross-hyb, restriction sites), not a folding or expression-strength predictor. Pair it with the karyon expression predictors for the soft, quantitative axis.