The engine
The engine behind unnatural structure, natural function.
Generative layers are commoditizing fast. Our advantage sits above them — a function-conditioned objective layer and a closed wet-lab loop whose data no public model has seen.
The loop
A design–build–test–learn loop for molecular machines
The generative layers are increasingly commodity. Our moat is above them: a function-conditioned objective layer that scores catalytic competence, and a closed wet-lab loop whose proprietary assay data retrains the models. Models commoditize; functional data does not.
- 01
Objective
Encode the task as a scorable target: substrate-binding geometry, transition-state preorganization, processivity, foldability and expressibility.
- 02
Backbone generation
Diffusion / flow-matching over backbone space, conditioned on the objective — searching a learned manifold of designable structures, not enumerating sequence space.
- 03
Sequence design
Inverse-folding models assign sequences that fold to the generated backbone and present the catalytic geometry.
- 04
Structure & physics
Structure prediction filters for fold confidence; MD / FEP / QM-MM stress-test the active site. In-silico scores are proxies — never the final word.
- 05
Build & assay
Express the survivors and measure real turnover in the wet lab. The assay is the ground truth the models are missing.
- 06
Learn
Functional data flows back into the objective layer and the generative models. The loop is the asset.
The objective layer
Function, encoded as something a model can search toward.
Before we generate anything, we turn the task into a scorable target. Function-conditioned scoring is the difference between a structure that merely looks right and one that actually catalyzes.
Transition-state geometry
The active site is scored on how precisely it positions catalytic residues around the reaction's transition state — the geometry that lowers the activation barrier, not just a pocket that binds substrate.
Electrostatic preorganization
We reward designs whose electrostatic environment is pre-arranged to stabilize developing charge in the transition state, the property that separates a competent catalyst from an inert binder.
Processivity
For machines that act on polymers, the objective encodes substrate threading and translocation — the ability to stay engaged and turn over repeatedly, not catalyze a single cut.
Foldability & expressibility
A design that does not fold or express is worthless. Fold confidence and soluble expression are first-class terms in the objective, weighed alongside catalysis from the start.
The stack
Where the work is commodity, and where the moat lives.
We build on the best generative and physics layers in the field — and we are clear-eyed that those layers are increasingly available to everyone. The defensible value is above them.
Commodity layers
- Backbone generation (diffusion / flow-matching)
- Inverse folding for sequence design
- Structure prediction & fold confidence
- Physics: MD, FEP, QM-MM
Powerful, improving, and available to the whole field. We adopt the best of each — none of them is a moat.
Where our moat lives
Function-conditioned objective layer
The scoring of catalytic competence — transition-state geometry, electrostatic preorganization, processivity, foldability — is where designs are won or lost. It is not a commodity model; it is our accumulated theory of what makes a catalyst.
Closed-loop assay data flywheel
Every expressed design and measured turnover is proprietary data that retrains the objective and the generative models. The loop compounds: better data sharpens the objective, which yields better designs to assay.
Substrate-targeting for hard substrates
We design against insoluble, crystalline substrates that off-the-shelf pipelines ignore. Targeting the substrate where the economics actually live is a deliberate, defensible choice.
Models commoditize. Functional data does not.
How we hold ourselves honest
Confidence is a measurement, not a claim.
De novo enzymes have historically been slow catalysts; the field's job is climbing rate enhancement by orders of magnitude. We don't claim that's solved. We claim a loop that climbs it — and we report what the loop actually measures.
In-silico scores are proxies
MD, FEP and QM-MM approximate catalysis; they don't equal k_cat. Every computational claim is paired with wet-lab turnover before we stand behind it.
We track the designability gap
Many designs don't fold or express. We measure fold and expression yield as first-class objectives, not afterthoughts.
Benchmarks, not adjectives
Cellulose-to-glucose turnover is our public yardstick. Numbers replace this scaffold as designs clear assay — we won't pre-announce them.