Expand description
signalox — pure-Rust bacterial signal-peptide / cleavage / transmembrane predictor.
A compact convolutional model with a 2-state linear-chain CRF cleavage decoder and a per-residue transmembrane head. The architecture is inspired by DeepSig (Savojardo, Martelli, Fariselli & Casadio, Bioinformatics 2018) — three cascaded convolutional stages that read the protein N-terminus, a classification head that discriminates signal peptides from N-terminal transmembrane segments, and a CRF that locates the cleavage site. It is not DeepSig: this is an independent clean-room implementation with our own weights, trained from scratch on public UniProt/SwissProt data (CC-BY 4.0). No DeepSig source or model weights are used.
Inference is self-contained and database-free: the trained weights are embedded in
the binary ([weights.json]) and every operation (one-hot encode, conv1d, ReLU,
global-average pool, linear heads, CRF Viterbi) is hand-rolled with no ML runtime.
let m = signalox::Model::embedded();
let p = m.predict(b"MKKTAIAIAVALAGFATVAQAAPKDNTWYTGAKLGWSQYH");
println!("{:?} cleavage={:?}", p.sp_type, p.cleavage);Structs§
- Model
- A loaded signalox model.
- Prediction
- A prediction for one protein.
Enums§
- SpType
- Signal-peptide class predicted for a protein’s N-terminus.