A general substrate for proteins and RNA.
We build foundation models that natively represent biological molecules as structured objects, with their geometry and chemistry intact. The result is one general-purpose backbone per molecule class, and many cheap task heads on top.
Our protein backbone is a deep, structure-native neural network with 105M parameters, trained from scratch on the entire usable Protein Data Bank, roughly 238,000 structures. No weights or representations are inherited from sequence-only language models; the backbone learns the rules of protein structure directly from physical data.
Each residue enters the model carrying its identity, local chemistry, and geometric context. The network reasons over the molecule's structure as a whole, producing a per-residue representation that downstream task heads consume.
The backbone is frozen after pretraining. Every downstream capability, whether mutation effect, stability, or design, lives in a small head trained on top.
We evaluated how much wet-lab data the model needs to predict the mutational landscape of a previously unseen protein. The protocol: freeze the pretrained backbone, randomly select 15% of the protein's deep mutational scan as a support set, train a small task head on that support, and measure Spearman correlation on the remaining 85%.
The finding most relevant to our work with wet labs is that a freshly-initialised head, trained from scratch on a few hundred measurements, reached near-SOTA per-protein accuracy. The frozen backbone already encodes enough general structural understanding that a minimal adapter, on a small seed of wet-lab data, suffices.
The practical implication: a comprehensive deep mutational scan normally requires several thousand measurements to train a competitive per-protein predictor. Our setup reaches comparable accuracy with roughly an order of magnitude less data, a meaningful reduction in wet-lab cost for the same downstream insight.
The RNA backbone is a sibling architecture, motif-aware and multi-scale, designed in collaboration with academic structural-biology groups. It is built to recognise the recurrent structural motifs that the RNA architectonics tradition treats as the load-bearing elements of global structure: pairing, stacking, junctions, and the long-range tertiary interactions like tetraloop receptors and kissing loops.
Initial pretraining draws on a large secondary-structure corpus and the usable RNA chains from the PDB. Parameter count is sized to the data available, smaller than the protein model by a deliberate factor.
- Evolution-aware backbone
- Next-generation protein backbone incorporating evolutionary signal at training time.
- Stability head
- Per-mutation ΔΔG predictor, aimed at drug-target druggability and antibody developability.
- Generation head
- Conditional sampling of novel sequences under structural and functional constraints.
- RNA dual-conformation tool
- Lab tool for collaborators: extract the feature-space difference between two known conformations of the same RNA sequence.
Preprints describing the architecture and benchmarks are in preparation. To request access to the model demo, see Request access. To discuss a collaboration, see Contact.