adder/ms2pipxgb

XGBoost Boosted Trees models for MS2 peak intensity prediction

#ms2pipXGB

###Install

Requirements:

• python numpy
• python pandas
• python multiprocessing
• python XGBoost (only required for training)
• Cython (http://cython.org/)

To compile the Cython .pyx files:

sh compile.sh

###Convert spectral library .msp to PEPREC format

The script

$ python convert_to_mgf.py <file>.msp

converts a spectral library in .msp format into a spectrum .mgf file and a peptide identification file in PEPREC format. This format contains at least three columns:

• spec_id: the id (TITLE) of the spectrum in the .mgf file
• modifications: a string indicating the locations (starting at 1) and the modification types
• peptide: the unmodified amino acid sequence

The files are written as <file>.PEPREC and <file>.PEPREC.mgf.

###Create feature vectors from PEPREC format

The script

usage: peprec2vec.py [-h] [-c INT] <.PEPREC file> <.PEPREC.mgf file>

MS2PIP on XGBoost

positional arguments:
  <.PEPREC file>        file containing peptide identifications
  <.PEPREC.mgf file>    file containing ms2 spectra

optional arguments:
  -h, --help            show this help message and exit
  -c INT, --num_cpu INT
                        number of cores

computes MS2PIP feature vectors from the PEPREC formatted files. The following files are written:

• vectors_b.pkl: all feature vectors for the b-ions
• vectors_y.pkl: all feature vectors for the y-ions
• targets_b_1.pkl: all targets (charge +1) values for the b-ions
• targets_y_1.pkl: all targets (charge +1) values for the y-ions
• psmids.pkl: groups the feature vectors by PSM

Currently (still under optimization) the following features are computed:

• peplen: number of amino acids in peptide
• ionnumber: number of amino acids in ion to predict
• ionnumber_rel: number of amino acids in ion to predict divided by peplen
• pmz: precursor mass
• mean 'chem': mean of 'chem' in peptide
• mz_ion: mass of ion to predict
• mz_ion_other: mass of other ion
• charge: spectrum charge state
• 'chem'_'loc': chem value of amino acid at location 'loc' (nterm, cterm and relative to cleavage pos 'i')
• mean_ion_'chem': mean of 'chem' in ion to predict
• mean_ion_other_'chem': mean of 'chem' in other ion
• min_ion_'chem': min value of 'chem' in ion to predict
• min_ion_other_'chem': min value of 'chem' in other ion
• max_ion_'chem': max value of 'chem' in ion to predict
• max_ion_other_'chem': max value of 'chem' in other ion

The features chem are computed from tables with estimated chemical property values for basisity, hydrophobicity, helicity and pI.

###Optimize and Train XGBoost models

The script

usage: train_xgboost.py [-h] [-p INT]
                        <vectors.pkl> <targets.pkl> <psmids.pkl> <model type>

XGBoost training

positional arguments:
  <vectors.pkl>  feature vector file
  <targets.pkl>  target file
  <psmids.pkl>   PSM groups file
  <model type>   {b,y}

optional arguments:
  -h, --help     show this help message and exit
  -p INT         number of cpu's to use

reads the Pickle files written by peprec2vec.py and trains an XGBoost model. Hyperparameters should still be optimized. You will need to digg into the script for model selection.

This script will also write the XGBoost models as .c files that can be compiled and linked through Cython. Please consult the script for details. To compile the Cython links to the .c models just run the script compile.sh again.

###Run MS2PIP

The script

usage: ms2pipXGB.py [-h] [-c INT] <.PEPREC file> <.PEPREC.mgf file>

MS2PIP on XGBoost

positional arguments:
  <.PEPREC file>        file containing peptide identifications
  <.PEPREC.mgf file>    file containing ms2 spectra

optional arguments:
  -h, --help            show this help message and exit
  -c INT, --num_cpu INT
                        number of cores

will read the compiled .c models and predict the MS2 peak intensities for the <.PEPREC.mgf> file. These will be compared to the observed peak intensities computed from the <.PEPREC> file. The pearson correlation for the b and y-ion models are writted to the file pearson.csv.