Generate
PF-1 proposes active-site geometries around the target transition state, then builds a backbone that holds them. Roughly 105 candidate sequences per campaign, none of them derived from a natural homologue.
Proteaform — Computational enzyme design
Enzymes designed, not discovered.
Most industrial reactions have no natural enzyme to borrow. We generate catalysts for them from scratch — so steps that need palladium, chlorinated solvents and cryogenic reactors can run in water, at room temperature.
PF-1 · designed backbone
78 residues · 3 helices · 2 strands
01 — The bottleneck
Chemistry built for refineries, used to make medicine.
A modern small-molecule drug takes eight to twelve synthetic steps. Most of them still run on chemistry inherited from petroleum processing: precious-metal catalysts, halogenated solvents, sub-zero reactors, and a waste stream twenty to a hundred times the mass of the product.
Enzymes do the same transformations at ambient temperature, in water, with the stereochemistry right the first time — when one happens to exist. For most reactions a process chemist actually needs, none does. The industry's answer has been to screen natural diversity for a distant relative and then evolve it: eighteen to thirty months of mutagenesis to reach a catalyst that survives a manufacturing plant, and no guarantee of a starting point at all.
We think that is a design problem wearing a search problem's clothes. Given the transition state you want stabilised, the active site that stabilises it can be generated — and the protein that holds it in place can be generated around it.
02 — Platform
PF-1, and the loop that keeps making it better.
PF-1 is a sequence–structure diffusion model conditioned on reaction geometry rather than on a homologous protein family. It is trained on public structural data and on our own paired library of variants and measured kinetics — the part that cannot be downloaded. Every campaign runs the same four stages, and every campaign ends by feeding the model.
Filter in silico
Folding, docking and stability models score every candidate for expressibility, melting temperature and predicted transition-state binding. Around a thousand survive to the bench — the step that makes the wet lab affordable.
Express
Genes are synthesised and expressed in E. coli, arrayed in 384-well plates. Cell-free expression runs alongside it for the designs we want an answer on this week rather than next month.
Measure
Kinetics by LC–MS on every variant, on the real substrate, under conditions a plant would use. Failures are measured with the same care as hits, because they carry as much signal.
↺ Every measurement returns to training. Each campaign shortens the next.
Numbers from the first eighteen months.
14
Weeks, median, from target reaction to a catalyst that holds up at scale. Directed evolution takes 18–30 months.
2.1M
Variants in our internal training set with kinetics measured in-house, not inferred.
41×
Median gain in kcat/KM over the best natural starting point we could find for the same reaction.
68%
Lower process mass intensity than the incumbent chemical route on our lead program.
Five programs, one platform.
Reaction classes chosen for how often they appear in routes that process chemists would rather not run.
| Program | Reaction class | Sponsor | Stage |
|---|---|---|---|
| PF-201 | Stereoselective C–N coupling | Undisclosed pharma partner | Scale-up · 100 L |
| PF-118 | Ketone → chiral amine (transaminase) | Internal | Lead optimisation |
| PF-095 | Late-stage C–H hydroxylation | Internal | Lead optimisation |
| PF-233 | Nitrile hydrolysis, non-natural substrate | Internal | Discovery |
| PF-260 | Macrocyclisation | Internal | Discovery |
05 — Compute
The wet lab is the slow part. Nothing else should be.
A campaign is worth running only if the design and ranking stages finish faster than the plates come back. That makes infrastructure a scientific constraint, not an IT one.
- TrainingPF-1 trained on multi-node NVIDIA H100 clusters; checkpoints and datasets versioned in object storage so any result can be traced to the exact model that produced it.
- InferenceStructure prediction and docking fan out across interruptible GPU capacity — tens of thousands of candidates scored per campaign, at a cost per design that lets us stay generous with what we test.
- Assay dataInstrument output lands in a lakehouse within minutes of the plate finishing, keyed to the design that generated it. No spreadsheet ever becomes the source of truth.
- ReproducibilityEvery campaign is a pinned pipeline: same seeds, same containers, same model weights, re-runnable a year later.
06 — Contact
If you run a route you wish you didn't, tell us about it.
We take on a small number of partnered programs a year, alongside our internal pipeline. We are also hiring scientists and engineers who want to work where the model meets the plate.