MuMDIA

MuMDIA is a search engine for data-independent acquisition (DIA) proteomics, written in Rust. It reads Orbitrap and Astral mzML and Bruker diaPASEF .d files (natively, with ion mobility), searches them against an in-silico FASTA digest or an imported predicted spectral library, and reports peptides and protein groups under target-decoy FDR control with label-free quantities. Optional Python sidecars add machine-learning components: DeepLC retention time, MS2PIP, AlphaPeptDeep and IM2Deep predictions, and a PyTorch or mokapot rescorer. A desktop application wraps the engine for users who prefer not to use a terminal.

The engine is validated on public DIA data with entrapment experiments. The command line, configuration and artifact formats can still change between minor versions.

What it does

One binary with one subcommand per stage. Each stage also runs standalone on the outputs of the previous one.

FASTA -> digest -> peptidoforms -> predict-frag --+
imported spectral library ------------------------+-> search-seed -> rt-im-train
mzML / diaPASEF .d / vendor file -> convert ------+   (multi-head DeepLC calibration)
                                                                  |
          extract -> (retrace, diaPASEF) -> features -> compete -> rescore -> quant -> report

mumdia run runs the whole chain. Given several files it searches them as one experiment: a pooled rescore, a second pass that matches between runs, and cross-run MaxLFQ. Every artifact is Parquet with a report.json that records its schema version, row count, content hash, resolved parameters and the model that ran, and every run writes a manifest.json.

Development is measured on three separate objectives: identification sensitivity, FDR validity under entrapment, and quantification accuracy on known-ratio data. A higher identification count is not evidence for the other two (docs/20).

Install

• Desktop application. Each release has a Windows installer (.msi) and a Linux .AppImage. It sets up the Python environments, builds libraries and runs searches (desktop/README.md).

• Docker, the reference environment with the sidecars installed:

  docker run --rm -v "$PWD:/data" --user "$(id -u):$(id -g)" ghcr.io/compomics/mumdia \
      run --fasta /data/proteome.fasta --mzml /data/sample.mzML \
          --out-dir /data/results --config /opt/mumdia/config.dia.json

  --user is required when writing into a bind mount (on Windows PowerShell, drop it and use -v "${PWD}:/data"). The image carries two configurations: /opt/mumdia/config.dia.json (FASTA digest, MS2PIP, DeepLC, mokapot) and /opt/mumdia/config.diann-lib.json (imported library, multi-head DeepLC calibration, nn_torch).

• Release archive, one per platform (x86_64-unknown-linux-musl, x86_64-pc-windows-msvc, aarch64-apple-darwin), with the binary, the sidecar scripts/, the conda specifications in env/, example configurations and the docs.

• From source:

  git clone https://github.com/CompOmics/MuMDIA.git
  cd MuMDIA/rust/mumdia && cargo build --release --locked

  The toolchain is pinned in rust-toolchain.toml; a C compiler is needed. On a synced drive (OneDrive, Dropbox), move the Cargo target directory off it first (rust/mumdia/.cargo/config.toml.example).

Outside Docker, the sidecars need a conda environment, then a check:

conda env create -f env/mumdia-deeplc.yml     # DeepLC, MS2PIP and the nn_torch rescorer
conda env create -f env/mumdia-peptdeep.yml   # AlphaPeptDeep, for the diaPASEF recipe
mumdia doctor --config configs/examples/diann-library.json

doctor resolves every interpreter as a run would, reports the versions it found and exits non-zero if the configuration cannot run. DeepLC 4.5.0 or newer is required.

Quickstart

# no Python at all: proves the installation and the input files
mumdia run --fasta proteome.fasta --mzml sample.mzML --out-dir results \
           --config configs/examples/native.json

# imported spectral library with the sidecars (the most sensitive workflow)
mumdia run --lib-precursors lib_precursors.parquet --lib-fragments lib_fragments.parquet \
           --mzml sample.mzML --out-dir results --config configs/examples/diann-library.json

# several runs as one experiment; repeat --mzml
mumdia run --lib-precursors lib_precursors.parquet --lib-fragments lib_fragments.parquet \
           --mzml a.mzML --mzml b.mzML --mzml c.mzML --out-dir exp \
           --config configs/examples/diann-library.json

# Bruker diaPASEF, read natively with ion mobility
mumdia run --fasta proteome.fasta --mzml sample.d --out-dir results \
           --config configs/examples/diapasef.json

Use a fresh output directory: run always recomputes its outputs. Building an imported library (predicted library, import, paired decoys with predicted spectra) is described in docs/13_sidecars.md, and configs/README.md explains each example configuration. docs/19_getting_started.md has complete worked runs.

Output

file	contents
peptides.tsv, proteins.tsv	identified precursors and protein groups, for reading
psms_scored.parquet	every scored PSM with all q columns; the unit for analysis
peptide_quant.parquet, protein_group_quant.parquet	quantities for analysis
lfq_maxlfq.parquet	cross-run protein matrix (experiments)
manifest.json, *.report.json	provenance, hashes, parameters, model identities

Use the Parquet tables for analysis; the TSVs are rounded for reading. Each q column controls its own row unit (run_psm_q per run, precursor_q per precursor, peptide_q_value per peptide, pg_q_value per protein group), so name the column with any count. peptides.tsv lists precursors but is filtered on the peptide-level q. In an experiment the TSVs are experiment-wide, with one quantity column per run. Details in docs/11_compete_rescore_fdr.md and docs/15_data_dictionary.md.

Limits

• Input is mzML or a Bruker diaPASEF .d. Thermo .raw is converted with ThermoRawFileParser; other vendor formats go through ProteoWizard msconvert and are not verified end to end (docs/04_convert.md).
• Wildcard and terminal variable modifications are not supported.
• CI builds and tests the engine on three platforms and runs an end-to-end smoke test, but does not run the Python sidecars on data. The PyTorch rescorer and the DeepLC fine-tune are not bit-reproducible, so counts vary slightly between runs.
• A peak cap at conversion (--top-peaks-ms2) deletes fragment evidence. The default is uncapped; set a cap only from a sweep on your own acquisition.

Documentation

• docs/README.md: the developer guide, one document per subsystem.
• docs/19_getting_started.md and docs/17_troubleshooting.md: worked runs and common failure modes.
• docs/35_diapasef.md: diaPASEF support and its measurements.
• docs/23_cli_reference.md and docs/24_config_reference.md: every subcommand and configuration field, generated from the code.
• bench/README.md: benchmarks and how to reproduce them.
• CHANGELOG.md, CONTRIBUTING.md, SECURITY.md.

Citation

There is no CITATION.cff or DOI yet. Cite the repository URL with the release tag you used.

License and attribution

Apache-2.0, see LICENSE. The notices of the statically linked Rust crates are in THIRD_PARTY_LICENSES.md, and sbom.cdx.json lists them in CycloneDX form.

Developed at CompOmics, VIB-UGent Center for Medical Biotechnology, Ghent University. This is a ground-up reimplementation of the earlier MuMDIA, which is archived on the legacy-python branch (tag legacy-python-v1). MuMDIA keeps a clean-room boundary against closed implementations. It does not ship or invoke DIA-NN; users build imported libraries under their own DIA-NN licence.