lotrs
Pure-Rust bacterial subcellular localization predictor. A compact convolutional network reads a full protein sequence and predicts its compartment — cytoplasm, inner-membrane, periplasm, outer-membrane, extracellular, or cell-wall. Self-contained, database-free, no ML runtime, no Python at inference time.
let m = embedded;
let p = m.predict;
println!; // e.g. "cytoplasm (1.00)"
CLI:
lotrs proteins.faa
#id localization prob
sp|P0A910 outer_membrane 0.991
Model & provenance
A sequence CNN: one-hot(20) → 4 × [Conv1d(k=7, same-pad) → ReLU] → global
average-pool ‖ max-pool → Linear(2C, 6) → softmax. Architecture inspired by DeepLoc
(Almagro Armenteros et al. 2017), but this is an independent clean-room implementation
with its own code and its own weights, trained from scratch on public UniProt/SwissProt
(reviewed bacterial entries with experimental localization, CC-BY 4.0). No DeepLoc source
code or trained weights were used.
Held-out per-class F1 (organism-disjoint): cytoplasm 0.99, inner-membrane 0.97, periplasm
0.82, outer-membrane 0.81, extracellular 0.69 (macro-F1 0.71; well-populated classes are
strongest). Weights are embedded in the crate (weights.json); inference is pure-Rust and
was verified to match the reference PyTorch model to 6 decimal places.
Changelog
0.1.1
- Performance: rewrote the
conv1dkernel (branch-free contiguous AXPY inner loop + flattened weight layout). ~4× faster single-thread; bit-identical output (same floating-point accumulation order — verified byte-for-byte on a full proteome).
0.1.0
- Initial release: distilled/trained pure-Rust model, gzip-embedded weights, PyTorch-parity verified.
License
MIT OR Apache-2.0. Training data derived from UniProt (CC-BY 4.0).