The GromovMatcher method is an optimal transport based algorithm for matching pairs of LC-MS untargeted metabolomic datasets.
This algorithm is part of our paper: "Optimal transport for automatic alignment of untargeted metabolomic data".
To install GromovMatcher, clone the github repo
git clone git@github.com:sgstepaniants/GromovMatcher.git
GromovMatcher was validated on randomly generated splits of the EXPOsOMICs LC-MS dataset of cord blood samples from 500 newborns (see "Validation and comparison on ground-truth data" in paper). The dataset can be downloaded at
https://www.ebi.ac.uk/metabolights/MTBLS1684/files
at the bottom of the "Files" section under the name `metabolomics_normalized_data.xlsx'.
Our method was experimentally test on LC-MS data from three studies which were part of the European Prospective Investigation into Cancer and Nutrition (EPIC) cohort (see "Application to EPIC data" in paper). This data is not publicly available, but access requests can be submitted to the Steering Committee at
https://epic.iarc.fr/access/submit_appl_access.php
GromovMatcher uses the Unbalanced Gromov-Wasserstein (UGW) algorithm to perform its matchings. Our implementation of the UGW algorithm in src/ugw_sinkhorn_solver.py is an extension of the original code at https://github.com/thibsej/unbalanced_gromov_wasserstein.
On validation and experimental EPIC data, we compare GromovMatcher to two state-of-the-art alignment methods.
The code for the latest M2S toolbox is located in experiments/M2S_experiments/M2S taken from the github repository https://github.com/rjdossan/M2S.
The code for the latest metabCombiner toolbox is located in experiments/mC_experiments/??? taken from the github repository https://github.com/hhabra/metabCombiner.
The GromovMatcher algorithm consists of two main steps:
- Matching metabolic features with Unbalanced Gromov-Wasserstein in
src/ugw_sinkhorn_solver.py - Filtering outlier matches through retention time drift estimation in
src/filtering.py
For details of both these steps see the "Results" and "Methods" sections of our paper.
The entire GromovMatcher algorithm combining both these steps is defined in src/GromovMatcher.py under the GM function.
The application of GromovMatcher is detailed in the notebook tutorial/GM_Example.ipynb. Two toy datasets are provided as an example.