TuringDNA · A method · v 0.1 · 2026

Engineer proteins
with AI.

A directed-evolution engine: design with ESM-2, build, edit with CRISPR, and learn from every bench result. Four tools, one loop, in the browser.

Open the engine  See how it fits together

5–50×
hit rate over
random mutagenesis
<60 s
sequence to a
ranked library
1,000+
REBASE enzymes
screened per map
9
published methods,
no black boxes

Free to try · your sequences stay private

Rendering a predicted fold…
Real PDB structure · drag to rotate
A real protein structure, rendered live in your browser right now — the same viewer that sits inside the engine over a predicted fold. A molecule, not a decorative double-helix.

The loop

Design, build, edit, learn.

TuringDNA is one loop, not four separate apps. A variant library comes out of step one — build it, edit it if the project needs CRISPR, then close the loop with what you measure at the bench.

round 2 01 Design DIRECTED EVOLUTION 02 Build PLASMID · PRIMERS 03 Edit CRISPR 04 Learn LOG RESULTS → ROUND 2 one loop — NOT FOUR APPS
Figure 1. A ranked variant library comes out of Design; you build and edit it, measure results at the bench, and Learn feeds round two back into Design — the dashed return — now conditioned on your own data, not just the model’s priors.
11503004506017519011,051T7 promoter · promoter · 7–24 (+)M13 fwd · primer_bind · 25–41 (+)lac operator · regulatory · 48–71 (+)ORF (286 aa) · CDS · 72–932 (+)M13 rev · primer_bind · 939–955 (+)ORF (286 aa)fuGFP1,201 bp · circular
Figure 2. The engine’s own output — a construct mapped in seconds: backbone, annotated features (promoter, CDS, primer sites), and a restriction ruler. Not a mock-up; this is the live map view.

Calibration

A predictor, not an oracle.

Every score is a prediction — verify on the bench. The directed-evolution numbers come from retrospective benchmarks across published single-protein campaigns, and every method traces to a primary source: ESM-2 zero-shot scoring, Doench CFD off-target scoring, SantaLucia nearest-neighbor Tm, REBASE, and AlphaFold. Nothing here is a black box.

Open

Free to use.
Your sequences stay private.

Everything runs without sending your sequence off-server — the only network calls are opt-in, per run. Only anonymous, aggregated signals improve our models. See how.

Open the engine 

Hosted on Hugging Face Spaces · sleeps after 48 h of inactivity · cold start ~30 s