A protein-language-model engine for directed evolution

Engineer proteins
with AI.

Describe what you want in plain language. Turing fetches the gene, scores every mutation, and hands you a ranked library — ready to order.

Open the engine Meet Turing

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

The conversational core

One conversation.
The whole lab.

Turing drives every tool in plain language — fetch a gene, evolve it, design guides, check primers. It runs the right one, in the right order, and remembers what it found.

Open Turing

TuringDesign
Find the CDS for human MC1R and design a stronger-expressing variant.
TuringPulling the real coding sequence first, then scoring substitutions.
fetch_sequenceDone
resolvedMC1R · ENST…147 · 954 nt
sourceEnsembl
design_variant_libraryDone
rank 1V60L · F76Y
rank 2T95A
positions scored317
TuringThree ranked variants ready — mutated residues in red. Want the DNA to order, or send the top pick to the Plasmid Editor?
Design Build Edit Learn ONE LOOP NOT FOUR APPS

The method

Design, build,
edit, learn.

01
Design
ESM-2 scores every substitution; a ranked variant library falls out.
02
Build
Map the construct, design primers, simulate the digest.
03
Edit
CRISPR guides, scored and specificity-checked.
04
Learn
Log bench results; round two comes back smarter.

Calibration

A predictor,
not an oracle.

Every score is a prediction — verify it on the bench, and the next round comes back sharper. Log what you measured; Turing fits a model on YOUR results, tells you exactly which of its guesses were right, and re-ranks accordingly. Nothing here is a black box; each number traces to a published method.

ESM-2Doench CFDSantaLucia TmREBASEAlphaFold

Your individual sequences stay private — never shared, exposed, or reproduced. Only anonymous, aggregated signals improve the models (how it works). You keep every right to what you design.

The Commons

It gets smarter
with every lab that uses it.

Every measurement you log — de-identified, never your sequence — sharpens the model for everyone who comes after you. A closed system only improves when its own team retrains it. This one compounds on its own.

276
Substitution effects pooled from independent labs
0.52
Spearman ρ vs. real published bench data
3
Public datasets validated, zero fabricated numbers

Checked against Stiffler et al. 2015 Cell, Melamed et al. 2013 RNA, and McLaughlin et al. 2012 Nature — the same public benchmarks the field uses to grade protein language models. Every number here comes from a live run against that data, never hand-entered.

Start with a sequence.

Open the engine 

Free to try · your sequences stay private