This is a python package for scoring SWATH-MS data using static generalizable machine learning models trained on large curated datasets. Current support is for OpenSwath, but could be expanded to other tools quite easily.
The recommended way to install GPS is into a virtual environment to make sure that all dependencies work correctly.
This can be done using your method of choice. The following demonstration will be using miniconda
conda create -n gpsenv -c conda-forge python=3.10 pip
conda activate gpsenv
With your environment activated, you can then install via pip
pip install gps-ms
GPS is now installed and ready to use!
GPS is very easy to use. To get started scoring a processed file, you simple run the score command:
gps score --input extracted_peakgroups.osw --output extracted_peakgroups.scored.tsv
This command will take in output from OpenSwath, score the extracted peakgroups, and write a tsv file with the q-values, scores, probabilities, etc.
To increase the number of identifications at a particular q-value cutoff, you can enable PIT estimation and correction. This will use a novel denoising algorithm to estimate the false target probability distribution of the target labels, and weight the decoys during q-value calculation.
gps score --input extracted_peakgroups.osw --output extracted_peakgroups.scored.tsv --estimate-pit
You can also make use of multiple cores using the --threads option
gps score --input extracted_peakgroups.osw --output extracted_peakgroups.scored.tsv --threads 10
Once all individual files are scored using GPS, it is very straightforward to build models to control the global levels of peptides and proteins in the analysis.
You can specify the level of the model to build using the --level cli option
gps build --level peptide --input *.scored.tsv --output peptide.model
The above command will take all scored files in at once using wildcard command line options, build a peptide level model, and estimate the global PIT for q-value correction
To build a protein level model, you only need to change the --level option.
gps build --level protein --input *.scored.tsv --output protein.model
Once all files have been scored and the global models have been built, GPS can combined all files into a quantiative matrix for convinient downstream analysis.
gps combine --input-files *.scored.tsv --peptide-model peptide.model --protein-model protein.model --output quantitative_matrix.tsv --max-peakgroup-q-value 0.01
The maximum q-value for the included precursors can be indicated if you would like to be more, or less, lenient on the identifications that you include in your final quantitative_matrix
GPS can easily be used to train your own model using just 2 commands.
First, we need to export data and perform any denoising to remove false target precursors from the dataset.
gps export --input extracted_peakgroups_1.osw --output extracted_peakgroups_1_training_data.npz
This filters the data using the denoising algorithm and then exports it to the numpy .npz format.
Next, we just need to train a model using the exported/filtered training data.
gps train --input *_training_data.npz --model-output trained_model.model --scaler-output trained_scaler.scaler
This trains a model using your input data and outputs the model and associated scaler so that it can be applied to new data effectively.
These models can be easily used to predict and score new data:
gps predict --input extracted_peakgroups.osw --output extracted_peakgroups.predicted.tsv --scoring-model trained_model.model --scaler trained_scaler.scaler
gps score --input extracted_peakgroups.osw --output extracted_peakgroups.scored.tsv --scoring-model trained_model.model --scaler trained_scaler.scaler
The rest of the downstream analysis can be the same.