What TuringDNA does

Everything on this page describes shipped behaviour. Planned work is labelled as such.

Start here

TuringDNA runs five tools behind one interface. Turing is a conversational agent that calls the other four for you; each is also available directly, for hands-on work without the chat. Chatting is free; running a tool requires a free account.

The shortest path to a real result is three steps:

1. Give it a sequence. Paste DNA or protein, or type a gene symbol like TP53 and let the engine resolve it against Ensembl or NCBI.
2. Pick the tool and its two or three settings. Most defaults are already the right answer — SpCas9 with an NGG PAM, knockout mode, 50 guides.
3. Read the ranked table, then export or copy. Every tool produces a ranked result you can sort, filter, download as CSV or Excel, or copy straight into a vendor order form.

Overview

Your individual sequences stay private, never shared, exposed or reproduced, and leave the machine only on the explicit opt-in network steps listed in Local vs. network. Only anonymous, aggregated signals improve our models. See the Privacy Policy.

Turing, the agent

Describe the job (fetch a gene, evolve it, design guides, check primers) and Turing calls the right tool in the right order, in a single reply where the task needs several.

• Tools it can call. fetch_sequence (gene symbol or accession → real DNA, via Ensembl/NCBI), design_variant_library, design_crispr_guides, design_primers. These are the same engines the direct tools use, not a separate model.
• Memory. It remembers what it fetched or designed earlier in the same conversation, so you can refer back to “it” or “that gene” instead of re-pasting.
• Honesty. Where something is out of scope, whether a capability that isn’t wired up or a request needing bench validation it can’t do, it says so rather than fabricating a result.

Because Turing calls the same scoring and design code, every caveat in the sections below applies to its answers too.

CRISPR guide design

Open the CRISPR designer

Two nucleases are supported, with the geometry each one requires:

For every candidate guide the designer reports:

• On-target activity: a Doench-style sequence-feature heuristic, with strand and position.
• Input off-target (CFD): cuts at other PAM-adjacent sites within the sequence you pasted, scored with the CFD matrix of Doench et al. (2016). This is the check that matters when you are editing a tagged construct whose tag may itself harbour a cut site. Genome-wide search is a separate step.
• Knockout efficacy: predicted indel spectrum, frameshift percentage, out-of-frame dominance and a loss-of-function likelihood, from frameshift probability and cut-position bias.
• Cloning oligos: ready-to-order sense/antisense pairs with the overhangs for the standard vectors (BbsI / BsmBI, and Cas12a geometry), copyable straight into a vendor order.

Genome-wide off-target search

Screen a guide against a genome

Choose an organism and your top-ranked guides are searched against its index; leave it unset and only the sequence you pasted is checked. Coverage differs by organism, and the difference matters:

For the four complete-genome organisms there is no coding-only caveat: an off-target anywhere, whether intronic, intergenic or regulatory, is found. Reference sequence comes from NCBI (E. coli) and Ensembl release 112 (the rest).

How the search works. Every 23-mer matching {spacer}{NGG} is extracted on both strands and grouped by its 8-nucleotide PAM-proximal seed (spacer positions 13–20 plus the 3-nt PAM), so a query scores a short candidate list rather than brute-forcing the whole index. For E. coli that is roughly 240,000 sites indexed in about five seconds. Index build cost scales at roughly one second per megabase, which is what bounds the organism list above.

Base editing

Open base-edit mode

Two editor families, each with its own activity window:

• CBE: cytosine base editors, deaminating C→T on the protospacer (non-target) strand.
• ABE: adenine base editors, deaminating A→G in a comparable window.

Where the knockout pipeline asks “where will Cas9 cut, and how cleanly?”, base editing asks “which single-base change does this guide install, and what does that do to the protein?”. Most competing tools only flag that a guide is base-editable. TuringDNA reports the predicted edit or edits, warns about bystanders, scores editability, and, when the reading frame is known, gives the codon and amino-acid consequence, including premature stops for DSB-free knockout (CRISPR-STOP / iSTOP).

Coordinates. Spacer positions are 1-indexed from the 5′, PAM-distal end. For SpCas9 the 20-nt protospacer occupies positions 1–20 with the NGG PAM at 21–23; editing windows are reported in those coordinates.

Primer analysis

Open primer analysis

The usual workflow for a handful of candidate primers is to BLAST each one separately and eyeball the results. This does the deterministic half in one shot:

• Scoring. Nearest-neighbour melting temperature (SantaLucia), GC content, 3′-end stability, GC clamp, and hairpin and self-dimer heuristics for every primer.
• Template binding. 3′-anchored binding sites and uniqueness against the template you are amplifying.
• In-silico PCR. Pairs any forward-oriented priming site with a downstream reverse-oriented site within a sane product size, the same logic as Primer-BLAST’s specificity check, and reports the intended amplicon plus any off-target products and their sizes.
• Amplicon checks. Product size, Tm match between the pair, and 3′ cross-dimer risk.
• Specificity (opt-in). BLASTs candidates against NCBI nt for a chosen organism and separates intended from off-target hits. This step sends the sequence to NCBI; it is off by default.

Everything rolls into a transparent fitness score in [0, 1] and a plain-English recommendation of the pair to use.

Plasmid editor

Open the plasmid editor

• Import. GenBank, FASTA, EMBL, SnapGene-style features, or raw DNA. Annotations are read from the record where present.
• Maps and editing. Annotated circular and linear maps, plus a sequence view with feature colouring, find, translate, reverse-complement and in-place editing.
• Restriction. Scans the full REBASE enzyme set (1,000+ enzymes) via Biopython, separates cutters from non-cutters, and simulates a digest on a virtual agarose gel.
• Cloning. Simulates Gibson-style overlap assembly (also NEBuilder HiFi, SLIC, CPEC), detecting a fragment order whose ends overlap and designing an overlap with homology-tailed primers where none exists; Golden Gate via Type IIS enzymes (BsaI, BbsI), chaining parts whose 4-nt overhangs match into a circle; and classic restriction-ligation. All three report fragment order, junctions and the assembled product.
• Export. GenBank or FASTA, plus a per-user construct library kept for 45 days.

Directed evolution

Open library design

Provide a wild-type sequence (or send a CDS across from the plasmid editor) and receive a library of multi-mutant variants ranked by predicted evolutionary fitness, codon-optimised for your host and ready to order.

• Parse and translate. FASTA, SnapGene, GenBank, EMBL, raw DNA or raw protein. Plasmid files surface a CDS picker; raw DNA with multiple stop codons triggers six-frame ORF discovery.
• Zero-shot scoring. ESM-2 computes ΔLL = log P(mutant | x_WT) − log P(WT | x_WT) for every position × 19 substitutions, using the wild-type marginal scheme of Meier et al. (2021).
• Combinatorial search. Simulated annealing over the top-percentile pool of single-site mutations, with multiple restarts, cumulative ΣΔLL as fitness, and penalties for stop codons and duplicated positions.
• Codon optimisation. Reverse-translation against a host codon-usage table, then synonymous scrubbing of BsaI (GGTCTC), BsmBI (CGTCTC) and NotI (GCGGCCGC) sites, on both strands, since Type IIS enzymes cut either, so the insert survives Golden Gate assembly and vendor prep.

Approximate functional retention as mutation count rises, which is the reason to stay shallow:

Local vs. network

Every step below runs inside the engine unless it is in the opt-in list. Sequences do not leave the machine for local steps.

Genome reference FASTAs are downloaded from NCBI and Ensembl on first use for the organism you pick; that request carries no sequence of yours.

Limits

Collected in one place, because knowing where a tool stops is the difference between using it and misusing it.

• Human and mouse off-target search is coding-sequence only. Intronic, intergenic and regulatory off-targets are outside the index.
• On-target activity is a heuristic in the style of Doench et al., not a re-trained or independently benchmarked model.
• Base-editing windows are literature consensus, not a trained efficiency predictor.
• Cloning, digests and in-silico PCR are predictions. Confirm before committing reagents.
• Directed evolution is not de novo design, and ESM-2 cannot see structure, active-site necessity, or epistasis between chosen mutations.
• Primer analysis ranks primers you supply; it does not generate candidates.
• Research use only. Not for clinical, diagnostic, therapeutic, prophylactic, food, feed or cosmetic use. See Terms §12 and the Biosecurity Policy.

References

The published methods the engine implements or builds on:

• Meier, J. et al. (2021). Language models enable zero-shot prediction of the effects of mutations on protein function. NeurIPS 34.
• Lin, Z. et al. (2023). Evolutionary-scale prediction of atomic-level protein structure with a language model. Science 379:1123–1130.
• Doench, J. G. et al. (2016). Optimized sgRNA design to maximize activity and minimize off-target effects of CRISPR-Cas9. Nature Biotechnology 34:184–191.
• Hsu, P. D. et al. (2013). DNA targeting specificity of RNA-guided Cas9 nucleases. Nature Biotechnology 31:827–832.
• Komor, A. C. et al. (2016) and Gaudelli, N. M. et al. (2017). Programmable base editing of C·G and A·T pairs. Nature 533:420 and 551:464.
• Allawi, H. T. & SantaLucia, J. (1997). Thermodynamics and NMR of internal G·T mismatches in DNA. Biochemistry 36:10581–10594.

Bundled components and licences

• ESM-2 weights: MIT (Meta / FAIR)
• Transformers, Accelerate: Apache 2.0 (Hugging Face)
• PyTorch: BSD-style (Meta)
• Biopython (parsing, restriction, Tm): BSD-derived
• Mol* viewer: MIT (PDBe / RCSB)
• Inter, JetBrains Mono: SIL Open Font License
• Lucide icons: ISC

How to cite

If TuringDNA contributed to published work, cite the tool alongside the underlying method (the CFD matrix, ESM-2 or the thermodynamic parameters) rather than in place of it. The method papers are listed under References.

TuringDNA (2026). TuringDNA: an integrated engine for CRISPR guide design,
base editing, primer analysis and directed-evolution library design.
https://turingdna.com/

Please state which organism index you screened against and whether coverage was whole-genome or coding-sequence only. The distinction in Genome-wide off-target search changes what a reported off-target count means.