Foundation models for proteins and RNA.
We pretrain structure-native foundation models that read biological molecules as higher-dimensional objects, not as text. The result is a small set of general-purpose backbones, and many cheap task heads that ride on top of them.
Protein world model
A 105M-parameter structure-native backbone pretrained from scratch on every usable structure in the PDB. One frozen substrate; many small, fast-to-train heads for downstream tasks like mutation effect, stability, and design.
Architecture →RNA world model
A motif-aware, structure-native encoder for RNA, built in collaboration with academic structural-biology groups. Targets the long-range tertiary interactions that sequence-only models miss, like tetraloop receptors, kissing loops, and pseudoknots.
Architecture →For wet labs
Our backbones make a small seed of wet-lab data go further. In a recent benchmark, a fraction of one protein's deep mutational scan was enough to recover most of the predictive ceiling.
For labs →ρ = 0.42 → ρ = 0.62
Few-shot mutational-effect prediction on SRC kinase, held out from training. The protein backbone was frozen; only a small head was fine-tuned on a fraction of the assay.
Near-SOTA per-protein accuracy from roughly an order of magnitude less wet-lab data than a conventional fully supervised pipeline would require.
- Assay
- SRC_HUMAN_Ahler_2019
- Support set
- 15% of mutations (505 labels)
- Held-out test
- 85% of mutations (2,867 labels)
- Backbone
- Frozen, 105M parameters
- Head trained
- ~3 min, single GPU
- Implied data saving
- ≈ 85% reduction in DMS labels
For academic groups
We collaborate with structural-biology and protein-engineering labs. If you have an assay, a structure, or a folding problem we might be useful for, write to us. We share access to the backbones in exchange for a chance to publish jointly.
Get in touch →For industry
We pilot with a small number of teams in pharma, antibody engineering, and enzyme design. Reach out about a per-target pilot or evaluation access to the demo.
Request access →