Promoter design

Promoter Design DRC

A deterministic design-rule check over a σ70 promoter sequence. A model (or a human) proposes a promoter; this gate qualifies whether it carries the architecture that actually drives transcription — and names every defect — before anything is ordered. Pure stdlib, no model, no network.

Two tiers of contract (each returns a human-readable reason):

• C1–C4 — hard, mechanism-grounded. A −35 box (TTGACA) and −10 box (TATAAT) located by minimum Hamming over the spacer window, the inter-box spacer (15–19 nt, 17 optimal), and a GC band. These are measured-validated: promoters that flag as weak-box express significantly lower on the real Urtecho set (AUROC 0.66, box-OK vs weak), so a flag predicts lower function — not just "looks wrong."
• C5–C6 — calibrated, dormant-by-correctness. Homopolymer-run and rare-forbidden-motif limits read from a reference pool of buildable promoters, so natural tracts / scaffold sites stay silent and only an out-of-distribution run or a genuinely-introduced rare site fires.

Install

pip install karyon          # no extras needed

Usage

--modality promoter is required (a .fasta/sequence could equally be generic synthesizable DNA — see gen-dna-qc):

karyon qualify GCATCG...TTGACA...TATAAT... --modality promoter         # one promoter (inline)
karyon qualify designed_promoters.fasta --modality promoter --json     # a batch

A promoter passes iff score == 0; --json emits the stable spine schema ({modality, ok, items:[{name, ok, score, reasons:[{contract, message, weight}]}], batch}).

From Python — the uncalibrated path (C1–C4 hard rules; C5–C6 fall back to safe defaults):

from karyon import qualify
r = qualify("ATGC...", modality="promoter")
v = r.items[0][1]
if v.score > 0:
    print(v.fired, v.messages)
    # e.g. (['C1 −35 box'], ["weak −35 box: best 'TTGCCA' is 2/6 mismatches from TTGACA"])

Calibrated to a pool of known-buildable promoters (C5/C6 become substrate-relative):

from karyon import promoter_contracts as pc

ctx = pc.calibrate_design(reference_promoters)   # list[str] of deposited, buildable sequences
for seq in generated_promoters:                  # e.g. Evo2 output
    v = pc.validate(seq, ctx)                     # == pc.DESIGN.evaluate(seq, ctx)
    if not v.ok:
        print(f"REJECT — {v.fired}: {v.messages}")

Composition with NVIDIA BioNeMo

Install alongside evo2-nim (genomic / regulatory sequence generation): Evo2 proposes promoters, this skill qualifies which ones carry a real −35/−10 architecture and a buildable sequence before synthesis — the "unroutable net" report for a generated regulatory element. For non-promoter synthesizability (sgRNA GC / homopolymer / Pol-III terminator), pair with the sequence-dfm skill.

Scope (honest)

The box model is a legible best-arrangement scan (no PWM training), so the operating regime is the real σ70 pool and mutated-from-real designs, where the boxes are genuine; on fully-random sequences, chance boxes limit discrimination. This is a design-rule gate (does the regulatory architecture exist, is it buildable), not an expression-strength predictor — pair it with the karyon expression predictors for the quantitative axis.