uno — MS1-only search engine (Rust)

A Rust MS1-only protein identification engine, in the spirit of DirectMS1 / ms1searchpy. uno accepts a single-run biosaur2 or KOTH feature TSV (--features), finds features with embedded KOTH (--raw), or searches a KOTH aligned intensity matrix across several runs (--matrix). It assigns peptide/protein identities and estimates FDR with target/decoy competition.

See matrix input for multi-run searching and the experimental confidence-aware matrix schema. Matrix input retains cross-run intensities but omits some quality descriptors from individual feature tables.

Digestion follows Sage's database builder (ported into uno-core); retention-time prediction is a pure-Rust learned model, with an on-data calibration step per run.

Experimental quantitative options

uno includes uno quant (details), a Rust DirectMS1Quant-style contrast analyzer in uno-core, and opt-in --out-quant-evidence exports from independent uno searches. The existing KOTH-matrix route remains the second quantitative option. Native-reference parity is tested; biological accuracy and error calibration are pending.

Workspace layout

Install

uno is published on crates.io as uno-ms (the uno name there belongs to an unrelated crate); the installed command is still uno:

cargo install uno-ms

This needs Rust 1.97 or newer. Bruker .d and Thermo .raw input (through koth-ms, pure Rust, no .NET runtime) are built in; cargo install uno-ms --no-default-features drops .raw. Platform: cargo install uno-ms currently supports Linux x86_64 only. LightGBM is linked from a vendored prebuilt lib_lightgbm.so.

Build from source

git clone https://github.com/pgarrett-scripps/uno
cd uno
cargo build --release   # binary at target/release/uno

All dependencies are fetched by Cargo; no sibling checkouts are needed.

Citation

If you use uno, please cite it using the metadata in CITATION.cff.

Run

cargo run --release -p uno-ms -- \
  --features path/to/run.features.tsv \
  --fasta path/to/proteins.fasta \
  --config example_config.toml \
  --out-proteins proteins.tsv

See example_config.toml for all tunables (digest, mass tolerance, top_k) and their documented defaults. Search uses the embedded pretrained CNN by default; --rt-model custom.urtc selects a custom model. Training uses PyTorch; search, model inspection, and evaluation run entirely in Rust.

See Custom RT models for uno rt train, uno rt inspect, uno rt evaluate, the data format, and the Python training API. uno --help lists every flag, including the external predictor RT-library workflow (--dump-candidate-peptides → --rt-library).

Experimental sequence-aware IM

IM models use the same sequence CNN and training workflow as RT, with charge added as an input. uno im train supports training from scratch and full-network adaptation; inspection, prediction, and independent evaluation use the Rust search predictor.

The bundled IM CNN contributes a soft classifier feature by default after the mass/RT shortlist. RT remains the main separator; there is no IM cutoff. Use --no-im to disable it or --im-model adapted.uimc to use custom weights. Missing IM leaves the search unchanged. See the model guide for supported chemistry, units, and limitations.

Search reproducibility

--seed N controls decoy generation and the default RT, cross-validation, and classifier random streams. Reuse the same seed for repeat searches; test multiple seeds to measure whether the result is stable. Keep the inputs, model, config, software version, and experimental environment settings fixed.

Optional scoring.rt_seed, scoring.cv_seed, and scoring.classifier_seed settings override their individual streams, including when --seed is supplied. Leave them absent for shared-seed behavior; see example_config.toml for examples. Changing decoys also changes downstream peptide indexing, so an individual seed intervention is not a pure variance-component estimate.

The current search includes bounded RT recovery: when the usual search and its evidence-retention retry find no target proteins at 1% FDR, UNO attempts one additional broad RT calibration. Successful results are returned unchanged. Validation showed fewer empty low-input runs with 0–1% normal-path runtime overhead; variation between successful seeds remains.

Repository contents

• crates/ — the Rust workspace.
• models/ — notes on the bundled RT and IM CNNs (the weights ship in crates/uno-core/models/).
• docs/ — model, matrix-input and quantification notes.
• tools/quant-reference/ — the DirectMS1Quant reference used for parity checks.
• tests/fixtures/ — software regression-test configurations.

Build artifacts (target/), the Python virtualenv (.venv/), RT training data (data/), and scratch outputs (experiments/) are intentionally not tracked.

Contributing

See CONTRIBUTING.md and the Code of Conduct. Report security issues privately as described in SECURITY.md.

License

MIT © Patrick Garrett