Helicase 1
A foundation genome model.
DNA and text in. Edited DNA and text out.
Give Helicase a DNA sequence and a request in words. It returns the edited sequence and a written answer. Helicase 1 is in development and has no published results yet.
Research updates
Introducing the Helicase 1 research programme
Helicase reads DNA in context, interprets an instruction, makes a precise edit, and checks the result — then revises the design from that feedback. Sequence understanding and generation are parts of one workflow, not separate products.
Helicase 1 is in development. We publish results with their sample size, date and failure modes, or we do not publish a number.
Research direction
Our working research target is a native DNA editor: the model receives sequence context and a written instruction, then produces an edited sequence that can be checked against the request.
This is a research programme, not a claim that Helicase 1 can yet perform reliable autonomous editing.
Data RL flywheel
Training data should come from the work the model is expected to do. Each design becomes a measured learning signal instead of an unverified success story.
- 01
Design
Pair DNA with a precise instruction.
- 02
Generate
Produce an edited sequence through the DNA writer.
- 03
Check
Score exact edits, preserved regions, and task constraints from the output bytes.
- 04
Learn
Use accepted and rejected examples to improve the next training round.
Availability and pricing
- Helicase 1
- Public model access, API availability and pricing have not been announced. Join the waitlist for updates.
- GeneLoop
- Coming soon: a workspace for sequence design and analysis, using existing AI models and sequence tools. Join the waitlist. Its product plans are separate from future Helicase access.
What Helicase is for
A workflow that connects what a researcher asks with what changes in their sequence.
Understand sequence context
Read a sequence alongside its context and answer questions relevant to a design, grounded in the supplied DNA and the researcher’s request.
Edit DNA through instructions
Translate a natural-language request into a precise sequence change, preserving the regions and constraints that should stay unchanged.
Check, inspect, and revise
Check the actual edited sequence, inspect the result, and use feedback to revise the design. A convincing explanation alone is not evidence that the requested edit is correct.
How it will be judged
A coding model for DNA is judged on the sequence it produces: instruction following, exact edits, and preservation of design constraints. Three different questions need three different kinds of evidence.
Sequence correctness
Is the edit the one that was asked for, and nothing else? Checked from the bytes of the resulting file.
Model performance
How often, on which tasks, with what sample size and failure modes. Helicase 1 has no published benchmark score yet.
Biological function
Whether a construct works in an experiment. A correct edit does not, by itself, establish this.
Frequently asked questions
What is Helicase 1?
Helicase 1 is GeneLoop’s foundation genome model. You give it DNA and text, such as a sequence and a request in words, and it returns edited DNA and text. It is in development and not yet publicly available.
How is Helicase different from GeneLoop?
GeneLoop is the platform, coming soon: existing AI models driving sequence tools. Helicase is the model we are building underneath it, one that reads and writes DNA directly instead of operating tools from the outside.
Can I use Helicase 1 today?
Helicase 1 is not publicly available. Join the waitlist for availability updates.
How much will Helicase 1 cost?
Helicase 1 pricing has not been announced. GeneLoop’s product plans are separate from future Helicase model access.
How will Helicase’s capabilities be evaluated?
On the sequence it writes: does it follow the instruction, make the requested change exactly, and preserve the constraints it was given — each checked from the bytes of the resulting file, never from the model’s account of what it did. Correct edits, predictive evidence and experimental function are different outcomes and each needs its own evidence.