A small seed of wet-lab data, made to go further.
Our pretrained backbones already encode the rules of protein and RNA structure. A few hundred of your measurements are usually enough to turn that general understanding into an accurate predictor for your specific target.
A comprehensive single-site saturation deep mutational scan on a 500-residue protein typically requires several thousand discrete measurements across synthesis, selection, sequencing, and analysis. The experiment runs for months and is rarely repeated within a project budget.
In a recent held-out benchmark, our protein backbone reached ρ = 0.62 on the remaining 85% of an assay using only 15% of the mutations as training labels, close to the per-protein supervised ceiling reported in the literature.
- 01
Bring a target.
A sequence, optionally a structure or AlphaFold prediction, and a description of the assay or measurement you want predicted.
- 02
We score zero-shot.
The frozen backbone produces a per-position vulnerability map and ranked mutation candidates without any of your labels. This sets a baseline.
- 03
You measure a small seed.
Typically a few hundred mutations, chosen to span the expected fitness range. Smaller than a full DMS by roughly an order of magnitude.
- 04
We fine-tune a head.
A small task-specific adapter trains on your seed in minutes on a single GPU. The backbone stays frozen and untouched.
- 05
You score the rest.
The fine-tuned head predicts the full mutational landscape of your target. The result is yours to use, publish, or fold into your next round of design.
Pilots and academic collaboration
For academic labs, we offer no-cost access to the backbones in exchange for joint publication where appropriate. For industry, we run small per-target pilots before any broader engagement; terms depend on whether the work is purely predictive or extends into model fine-tuning on proprietary data. Either way, the first conversation is short.