This is a proof-of-principle implementation of a method for re-scoring METASPACE annotations to improve the FDR. METASPACE is an engine for metabolite annotations of imaging mass spectrometry data. FDR is the False Discovery Rate and is used in METASPACE to provide results with a given confidence.
Our re-scoring method uses a semi-supervised machine learning approach originally proposed in proteomics. We train a classifier discriminating the most prominent target hits from the most prominent decoy hits and apply this classifier to re-score all annotations.
This repository accompanies the manuscript Silva et al., Data-driven rescoring of metabolite annotations significantly improves sensitivity, Anal. Chem., Just Accepted Manuscript.
IMPORTANT This implementation is a prototype implementation and can be installed and executed by a bioinformatician or computer scientist with experience of installing Python software. This software requires a modified version of the METASPACE engine because we needed additional output (both target and decoy hits opposite to only target hits output by METASPACE, as well as additional spectral and spatial features) which is available here: anasilviacs/sm-engine.
To test this pipeline, we provide an example file: MTBLS415 exported search results.
This file is the result from searching one experiment obtained from the MetaboLights repository (accession number MTBLS415) against HMDB with the modified version of the METASPACE engine found in anasilviacs/sm-engine, and exported into a tab-separated file through use of the export_search_results.py script provided in this repository.
To execute this pipeline the following tools and packages are required:
- Percolator
- Python >= 3
- NumPy >= 1.13.3
- Pandas >= 0.20.2
If you'd like to execute your own search, please refer to anasilviacs/sm-engine for the modified engine's installation and usage instructions.
Clone this repository:
git clone https://github.com/anasilviacs/rescore-metabolites.git
This command creates the rescore-metabolites folder, where all the scripts necessary to run the rescoring pipeline are found.
After running a search with the modified version of the METASPACE engine found in anasilviacs/sm-engine, the results must be exported as a tab-separated, .tsv file. To do so, use the export_search_results.py script as follows:
python export_search_results.py [dataset name] [path to tsv file]
"dataset name" is the name of the dataset of interest in the engine database
"path to tsv file" is where the exported results will be stored
An example of what an exported file looks like can be downloaded from our servers through the following link: MTBLS415 exported search results. This file looks as follows:
formula_db db_ids sf_name sf adduct chaos spatial spectral image_corr_01 image_corr_02 image_corr_03 image_corr_12 image_corr_13 image_corr_23 snr percent_0s peak_int_diff_0 peak_int_diff_1 peak_int_diff_2 peak_int_diff_3 quart_1 quart_2 quart_3 ratio_peak_01 ratio_peak_02 ratio_peak_03 ratio_peak_12 ratio_peak_13 ratio_peak_23 percentile_10 percentile_20 percentile_30 percentile_40 percentile_50 percentile_60 percentile_70 percentile_80 percentile_90 fdr isocalc_sigma isocalc_charge isocalc_pts_per_mz first_peak_mz targets target msm
HMDB {31173} [u'6-Hydroxy-1H-indole-3-acetamide'] C10H10N2O2 +Nd 0.0 0.0 0.0 0.0 0.0 0.0 -0.000410693 0.0 0.0 0.0 0.0 -0.700482 0.955152 -0.0837069 -0.614721 0.0 0.0 0.0 0.0 0.0 0.0 3.4445 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.001238-1 4039 331.982499 [u'-H', u'+Cl'] 0 0.0
HMDB {31173} [u'6-Hydroxy-1H-indole-3-acetamide'] C10H10N2O2 +Ru 0.0 0.0 0.878995 -0.00128352 -0.00047319 -0.00179479 -0.00126883 -0.00481261 -0.00177425 0.0218785 0.999519 -0.888743 0.491929 -0.0492674 0.545825 0.0 0.0 0.0 0.108129 1.01837 0.0624696 9.41809 0.577733 0.0613429 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.001238 -1 4039 285.982374 [u'-H', u'+Cl'] 0 0.0
HMDB {31173} [u'6-Hydroxy-1H-indole-3-acetamide'] C10H10N2O2 +Th 0.0 0.0 0.0 0.0 0.0 -0.000240442 0.0 0.0 0.0 0.0155062 0.99976 -0.285585 -0.00726385 -0.107809 0.6996660.0 0.0 0.0 0.0 0.0 1.00407 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.001238-1 4039 422.112831 [u'-H', u'+Cl'] 0 0.0
HMDB {31173} [u'6-Hydroxy-1H-indole-3-acetamide'] C10H10N2O2 +Rh 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -0.000339876 0.0 0.0 -0.994127 -0.00726385 0.747369 0.512321 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.64987 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.001238-1 4039 292.98
Execute semisupervised.py on the exported tab-separated file::
python semisupervised.py [path to tsv file] [-d] [-k]
"path to tsv file" is where the exported tsv file is
"-d" is a flag which should be used for getting the decoys' q-values
"-k" is a flag which should be used to keep the intermediate files
After execution, a folder called <tsv file name> is created and a comma-separated, results.csv file with the final (median, called combined in the file) q-value for each metabolite ion is written in that folder. It is possible to also save the decoy hits' q-values, although give the nature of the sampling process it is unlikely that all of them will have values, and the intermediate q-values (i.e. the ones obtained in each iteration of the method). The simplest version of the output file looks as follows:
SpecId,combined
C10H10N7O2S-H,0.5886835
C10H10S3-H,0.5578665
C10H11ClFN5O3-H,0.7680545
C10H12ClN5O3-H,0.736791
Please make the most of GitHub issues for any questions, suggestions or requests for support.
This project is licensed under Apache 2.0 license.
This work was funded by the European Horizon2020 project METASPACE (no. 634402).