Expand description
nomenox — pure-Rust protein-function classifier for hypothetical proteins.
A compact convolutional network that reads a protein sequence and predicts Pfam family membership over a fixed vocabulary, for CDS that have no homology hit (otherwise-“hypothetical” proteins). It was distilled from an ESM-2 teacher (frozen embeddings + MLP head over the vocabulary) into a small sequence CNN so it can run pure-Rust on CPU; per-family calibrated thresholds mean it only emits a call when confident and otherwise stays silent (leaving the protein hypothetical — it never fabricates a function). Architecture is our own, trained from scratch on public UniProt/SwissProt (CC-BY); no external tool’s code or weights are used.
Inference is self-contained: weights embedded at compile time, every op hand-rolled
(one-hot encode, conv1d + ReLU, global-average pool, linear head, per-family
sigmoid + threshold).
let m = nomenox::Model::embedded();
for call in m.predict(b"MSKIVKIIGREIIDSRGNPTVEAEVHLEGGFVGMAAAPSGASTGSRE") {
println!("{} {} ({:.2})", call.pfam, call.name, call.score);
}