The Science

How Fermvyne generates enzymes that actually work.

Most protein language models are trained on sequences. Ours is trained on reactions — substrate SMILES, product SMILES, cofactor class, and EC hierarchy encoded at every training step. This page explains what that means, what the model can do, and where it doesn't hold up.

Training Data

Sequence is not enough

Sequence-only protein language models learn fold. Fermvyne's model is trained to learn catalysis — substrate SMILES, product SMILES, cofactor class, and EC hierarchy are encoded at every step. The generative decoder conditions on reaction chemistry, not just homology to known sequences.

47M enzyme-reaction pairs, curated from BRENDA, KEGG, MetaCyc, and Swiss-Prot. 3D structure data from AlphaFold2 and PDB contributed to thermostability head training. Substrate and product SMILES encoded via Morgan fingerprints.

BRENDA enzyme database
KEGG reaction repository
MetaCyc metabolic pathways
Swiss-Prot curated entries
AlphaFold2 + PDB structures
47M enzyme-reaction pairs in training set
6,200+ EC numbers represented in training data
340 holdout variants with wet-lab validation data
3 expression host systems (E. coli, Pichia, Bacillus)

Architecture

Encoder–decoder with parallel prediction heads

The reaction encoder processes SMILES and EC context through a stack of transformer blocks. A cross-attention bridge connects encoder outputs to the autoregressive sequence decoder. Three parallel prediction heads — one per biophysical property — run on the decoder's final representations before output is returned.

Reaction Encoder SMILES + EC Cross-Attention Bridge Seq Decoder Autoregressive Expr Head Solubility Tm Head Parallel prediction

Benchmarks

Holdout set: 340 variants with wet-lab validation

Internal validation on a holdout set of 340 enzyme variants not seen during training. Each variant was expressed in the specified host and measured for the target property. These are internal results and have not been submitted to an independent benchmark server.

Expression accuracy ~91% Fermvyne ~50% Seq-only Solubility (binary) 88% Fermvyne ~47% Seq-only Tm correlation R² 0.81 Fermvyne

Seq-only baseline: same transformer architecture, reaction-context encoding removed. All metrics on 340-variant wet-lab holdout, expression host E. coli BL21(DE3). Tm correlation measured against DSF melting curves on the same variant set.

Scope

What Fermvyne doesn't do yet

Generative models that don't fail loudly are dangerous in a wet lab context — you'll order synthesis for sequences the model wasn't built to handle. Fermvyne returns a confidence_flag for out-of-scope requests and documents those boundaries here so you know before you submit.

Multi-domain proteins >500 residues

Prediction confidence drops for proteins with more than 500 residues or multiple independent folding domains. The model was primarily trained on single-domain enzymes.

Allosteric enzyme design

Designing allosteric regulation is an active area of development. We can generate sequences with good catalytic properties but do not yet model allosteric communication paths reliably.

Non-standard expression hosts

Biophysical predictions are validated for E. coli BL21(DE3), Pichia pastoris, and Bacillus subtilis. Expression in other organisms (CHO, yeast strains outside Pichia, plant systems) may have lower accuracy.

Substrate specificity for very novel chemistry

Reactions involving substrate SMILES that fall far outside training distribution (novel heterocycles, unusual functional groups) receive a confidence flag. We show the flag rather than hiding uncertainty.

Try it on your own enzyme targets.

Free tier available. Or read the technical docs first.

Start Free Quickstart guide