Methodology
How Proteinvue navigates sequence space.
The model, the training data, and why we use generative search instead of directed evolution simulation.
The model: what it learns and how it searches
Masked language modeling on protein sequences
The core of Proteinvue is a transformer trained with a masked language modeling objective on 200M+ sequences from UniProt, spanning protein families across thermophiles, mesophiles, and mammalian secreted proteins. During training, residues are randomly masked and the model learns to predict them from context. This teaches the model the statistical grammar of protein sequences: which residue identities are compatible with which neighbors, and which combinations appear in functionally viable proteins.
The result is a model that has internalized the co-variation structure of protein families at scale. When you provide a target sequence and fitness objective, the model uses that learned grammar to score candidate mutations and to sample sequences that fit the learned distribution for your protein family.
Fitness landscape sampling via iterative refinement
Generative search works by iteratively proposing and scoring candidate sequences. Starting from your reference, the model samples residue substitutions that improve the predicted fitness score, subject to maintaining the co-variation coherence of the protein family. This is not random mutagenesis. The model explores regions of sequence space that evolutionary pressure has previously visited and found viable, focusing search effort on combinations that are statistically likely to fold and function.
Each iteration refines the candidate pool by selecting for higher predicted fitness and for sequence diversity across the shortlist. The output is a set of candidates that differ meaningfully from each other, so your assay plates are not redundant.
Why sequence co-variation, not structure prediction
Structure prediction tools like AlphaFold2 answer: what is the most likely fold for this sequence? That is valuable, but it is an imprecise proxy for engineering fitness objectives. Binding affinity in a CDR loop depends on fine sequence-level details that the predicted backbone may not resolve. Enzyme active site catalytic efficiency depends on electrostatic environment and conformational dynamics that a static structure does not capture. Thermostability-activity trade-offs involve epistatic interactions across multiple positions that a per-sequence fold prediction does not model.
Proteinvue scores fitness directly from learned sequence-function co-variation, not from a structural intermediate. This is not a claim that structure is uninformative. It is a narrower claim: for the design objectives we support, co-variation in the multiple sequence alignment of evolutionary relatives is a more direct signal than folding geometry alone. We use structure when it is available as an optional input, not as the primary scoring mechanism.
Illustrative residue co-variation signal. Not actual model output. Green = co-varying pairs, amber = active site region.
What the model learned from.
The base model is trained on public sequence databases: UniProt/Swiss-Prot and TrEMBL, RefSeq protein, and curated structure-function datasets. No proprietary customer data is included in base model training. Your sequence data from active programs is used only for your program's retrospective validation and generative search runs, not to retrain the shared model.
The model is retrained periodically as new sequences are deposited. Coverage is strongest for protein families with broad evolutionary representation in public databases: kinases, oxidoreductases, immunoglobulins, proteases, and membrane transporters. Coverage is thinner for recently characterized families and orphan enzymes with few known homologs.
Performance depends on training coverage for your protein family. We discuss coverage before you commit. If retrospective validation does not show meaningful rank correlation on your held-out data, we tell you before you pay for a full program run.
How we assess model quality.
We do not publish a single accuracy benchmark because performance is landscape-dependent. The same model will have very different rank discrimination on a kinase with thousands of characterized variants versus a recently discovered oxidoreductase with a shallow multiple sequence alignment. Publishing a headline number would obscure this difference.
Before onboarding, we run a retrospective validation: we provide ranked candidates for a held-out subset of your known sequence-activity data to verify model discrimination on your specific target. The report shows Spearman rank correlation between predicted fitness and measured activity. If the retrospective does not show meaningful correlation, we tell you.
If the model does not work for your target, we say so. There is no value in selling you a program run on a protein family where training coverage is too thin to produce useful signal. The retrospective validation exists precisely for this reason.Request a retrospective validation
Have questions about the approach?
Elena answers methodology questions directly. If you want to go deeper on MSA depth, co-variation signal, or model limitations for your specific target family, that conversation happens before any commitment.