triqler 0.5.0

A combined identification and quantification error model of label-free protein quantification

Triqler is a probabilistic graphical model that propagates error information through all steps from MS1 feature to protein level, employing distributions in favor of point estimates, most notably for missing value imputation. The model outputs posterior probabilities for fold changes between treatment groups, highlighting uncertainty rather than hiding it.

The input and output formats are described at the bottom of this page. We also provide converters for output from software packages such as MaxQuant and Quandenser. Instructions on how to run these, as well as other extended functionality, can be found in our wiki.

Method description / Citation

The, M. & Käll, L. (2019). Integrated identification and quantification error probabilities for shotgun proteomics. Molecular & Cellular Proteomics, 18 (3), 561-570.

Requirements

Python 2 or 3 installation

Packages needed:

• numpy 1.12+

• scipy 0.17+

Installation via pip

pip install triqler

Installation from source

git clone https://github.com/statisticalbiotechnology/triqler.git
cd triqler
pip install .

Usage

usage: python -m triqler [-h] [--out_file OUT] [--fold_change_eval F]
                 [--decoy_pattern P] [--min_samples N] [--num_threads N]
                 [--ttest] [--write_spectrum_quants]
                 [--write_protein_posteriors P_OUT]
                 [--write_group_posteriors G_OUT]
                 [--write_fold_change_posteriors F_OUT]
                 IN_FILE

positional arguments:
  IN_FILE               List of PSMs with abundances (not log transformed!)
                        and search engine score. See README for a detailed
                        description of the columns.

optional arguments:
  -h, --help            show this help message and exit
  --out_file OUT        Path to output file (writing in TSV format). N.B. if
                        more than 2 treatment groups are present, suffixes
                        will be added before the file extension. (default:
                        proteins.tsv)
  --fold_change_eval F  log2 fold change evaluation threshold. (default: 1.0)
  --decoy_pattern P     Prefix for decoy proteins. (default: decoy_)
  --min_samples N       Minimum number of samples a peptide needed to be
                        quantified in. (default: 2)
  --num_threads N       Number of threads, by default this is equal to the
                        number of CPU cores available on the device. (default:
                        8)
  --ttest               Use t-test for evaluating differential expression
                        instead of posterior probabilities. (default: False)
  --write_spectrum_quants
                        Write quantifications for consensus spectra. Only
                        works if consensus spectrum index are given in input.
                        (default: False)
  --write_protein_posteriors P_OUT
                        Write raw data of protein posteriors to the specified
                        file in TSV format. (default: )
  --write_group_posteriors G_OUT
                        Write raw data of treatment group posteriors to the
                        specified file in TSV format. (default: )
  --write_fold_change_posteriors F_OUT
                        Write raw data of fold change posteriors to the
                        specified file in TSV format. (default: )

Example

A sample file iPRG2016.tsv is provided in the example folder. You can run Triqler on this file by running the following command:

python -m triqler --fold_change_eval 0.8 example/iPRG2016.tsv

A detailed example of the different levels of Triqler output can be found in Supplementary Note 2 of the Quandenser publication.

Interface

The simplest input format is a tab-separated file consisting of a header line followed by one PSM per line in the following format:

run <tab> condition <tab> charge <tab> searchScore <tab> intensity <tab> peptide     <tab> proteins
r1  <tab> 1         <tab> 2      <tab> 1.345       <tab> 21359.123 <tab> A.PEPTIDE.A <tab> proteinA <tab> proteinB
r2  <tab> 1         <tab> 2      <tab> 1.945       <tab> 24837.398 <tab> A.PEPTIDE.A <tab> proteinA <tab> proteinB
r3  <tab> 2         <tab> 2      <tab> 1.684       <tab> 25498.869 <tab> A.PEPTIDE.A <tab> proteinA <tab> proteinB
...
r1  <tab> 1         <tab> 3      <tab> 0.452       <tab> 13642.232 <tab> A.NTPEPTIDE.- <tab> decoy_proteinA

Alternatively, if you have match-between-run probabilities, a slightly more complicated input format can be used as input:

run <tab> condition <tab> charge <tab> searchScore <tab> spectrumId <tab> linkPEP <tab> featureClusterId <tab> intensity <tab> peptide     <tab> proteins
r1  <tab> 1         <tab> 2      <tab> 1.345       <tab> 3          <tab> 0.0     <tab> 1                <tab> 21359.123 <tab> A.PEPTIDE.A <tab> proteinA <tab> proteinB
r2  <tab> 1         <tab> 2      <tab> 1.345       <tab> 3          <tab> 0.021   <tab> 1                <tab> 24837.398 <tab> A.PEPTIDE.A <tab> proteinA <tab> proteinB
r3  <tab> 2         <tab> 2      <tab> 1.684       <tab> 4          <tab> 0.0     <tab> 1                <tab> 25498.869 <tab> A.PEPTIDE.A <tab> proteinA <tab> proteinB
...
r1  <tab> 1         <tab> 3      <tab> 0.452       <tab> 6568       <tab> 0.15    <tab> 9845             <tab> 13642.232 <tab> A.NTPEPTIDE.- <tab> decoy_proteinA

Some remarks:

• For Triqler to work, it also needs decoy PSMs, preferably resulting from a search engine search with a reversed protein sequence database concatenated to the target database.

• The intensities should not be log transformed, Triqler will do this transformation for you.

• The search engine scores should be such that higher scores indicate a higher confidence in the PSM.

• We recommend usage of well calibrated search engine scores, e.g. the SVM scores from Percolator.

• Do not set –fold_change_eval to 0 or a very low value (<0.2). The fold change posterior distribution always has a certain width, reflecting the uncertainty of our estimation. Even if the fold change is 0, this distribution will necessarily spill over into low fold change values, without there being any ground for differential expression.

• Multiple proteins can be specified at the end of the line, separated by tabs. However, it should be noted that Triqler currently discards shared peptides.

The output format is a tab-separated file consisting of a header line followed by one protein per line in the following format:

q_value <tab> posterior_error_prob <tab> protein <tab> num_peptides <tab> protein_id_PEP <tab> log2_fold_change <tab> diff_exp_prob_<FC> <tab> <condition1>:<run1> <tab> <condition1>:<run2> <tab> ... <tab> <conditionM>:<runN> <tab> peptides

Some remarks:

• The q_value and posterior_error_prob columns represent respectively the FDR and PEP for the hypothesis that the protein was correctly identified and has a fold change larger than the specified –fold_change_eval.

• The protein_id_PEP and diff_exp_prob_<FC> columns are simply the separate probabilities that make up the above hypothesis test, i.e. for correct identification and for fold change respectively.

• The reported fold change is log2 transformed and is the expected value based on the posterior distribution of the fold change.

• If more than 2 treatment groups are present, separate files will be written out for each pairwise comparison with suffixes added before the file extension, e.g. proteins.1vs3.tsv.

• The reported protein expressions per run are the expected value of the protein’s expression in that run. They represent relative values (not log transformed) to the protein’s mean expression across all runs, which itself would correspond to the value 1.0. For example, a value of 1.5 means that the expression in this sample is 50% higher than the mean across all runs. A second example comparing values across samples: if sample1 has a value of 2.0 and sample2 a value of 1.5, it means that the expression in sample1 is 33% higher than in sample2 (2.0/1.5=1.33). We don’t necessarily recommend using these values for downstream analysis, as the idea is that the actual value of interest is the fold change between treatment groups rather than between samples.