---

ReDeeM: Repository for Deep Learning Models for Mass Spectrometry

ReDeeM is a Rust workspace for mass spectrometry proteomics, providing deep learning models for peptide property prediction and machine learning classifiers for PSM rescoring. It is designed to be used as a library in other tools (e.g. Sage).

Crates

Crate	Description	Status	Docs
redeem-cli	Command-line interface for ReDeeM	Releases	README
redeem-classifiers	Semi-supervised PSM rescoring (GBDT, XGBoost, SVM)	Not published	README
redeem-openms-ffi	Prebuilt C ABI bridge for OpenMS integration	Releases	README
redeem-properties	Peptide property prediction (RT, CCS, MS2) using candle		README
redeem-properties-py	Python bindings for redeem-properties via PyO3		README · Docs

Installation

Rust

redeem-properties is available on crates.io. Other crates are still in development.

[dependencies]
redeem-properties = "0.1"
redeem-classifiers = { git = "https://github.com/singjc/redeem.git", branch = "develop" }

Python

pip install redeem_properties

For DataFrame output support:

pip install "redeem_properties[pandas]"   # or [polars]

OpenMS FFI bundles

Prebuilt redeem-openms-ffi static-library bundles are published as GitHub release assets for supported OpenMS target platforms. OpenMS consumes those artifacts automatically by platform when WITH_REDEEM=ON:

cmake -S /path/to/OpenMS -B build \
  -DWITH_REDEEM=ON

You can pin a specific release with -DREDEEM_FFI_VERSION=v0.1.0, or override the download entirely with -DREDEEM_FFI_LIBRARY=/path/to/libredeem_openms_ffi.a. See redeem-openms-ffi/README.md for bundle contents, workflow-dispatch testing, and manual packaging instructions.

Quick Example

Python

import redeem_properties as rp

# Create a unified predictor (loads pretrained models by default)
model = rp.PropertyPrediction()

peptides = [
    "SKEEET[+79.9663]SIDVAGKP",
    "LPILVPSAKKAIYM",
    "RTPKIQVYSRHPAE",
]

df = model.predict_df(
    peptides,
    charges=[2, 3],
    nces=20,
    instruments="timsTOF",
    annotate_mz=True,
    annotate_mobility=True
)

>>> df.head()
                    peptide  charge  nce instrument         rt         ccs  ion_mobility  precursor_mz ion_type  fragment_charge  ordinal    intensity          mz
0  SKEEET[+79.9663]SIDVAGKP       2   20    timsTOF  26.884516  535.355408      1.324961     785.35581        b                1        2   790.534363  216.134268
1  SKEEET[+79.9663]SIDVAGKP       2   20    timsTOF  26.884516  535.355408      1.324961     785.35581        b                1        3   822.035767  345.176861
2  SKEEET[+79.9663]SIDVAGKP       2   20    timsTOF  26.884516  535.355408      1.324961     785.35581        b                1        4  1272.754517  474.219454
3  SKEEET[+79.9663]SIDVAGKP       2   20    timsTOF  26.884516  535.355408      1.324961     785.35581        b                1        5  1806.533691  603.262047
4  SKEEET[+79.9663]SIDVAGKP       2   20    timsTOF  26.884516  535.355408      1.324961     785.35581        y                1        9   218.158798  967.449573

See the Python bindings README for full usage, including MS2 prediction, DataFrame output, and the unified PropertyPrediction helper.

Rust

use redeem_properties::pretrained::{locate_pretrained_model, PretrainedModel};
use redeem_properties::models::rt_model::RTModelWrapper;
use candle_core::Device;
use std::sync::Arc;

let model_path = locate_pretrained_model(PretrainedModel::RedeemRtCnnTf)?;
let model = RTModelWrapper::new(&model_path, None::<&str>, "rt_cnn_tf", Device::Cpu)?;

let sequences = vec![Arc::from(b"PEPTIDEK".as_slice())];
let mods = vec![Arc::from(b"".as_slice())];
let mod_sites = vec![Arc::from(b"".as_slice())];
let result = model.predict(&sequences, &mods, &mod_sites)?;

See each crate's README for detailed API documentation and examples.