Deep learning model (Temporal conv net, using dilated conv layers) detects post-translationally modified peptides based on their ms/ms spectrum.
This repository is an implementation of our model as described in:
AHLF: ad hoc learning of peptide fragmentation from mass spectra enables an interpretable detection of phosphorylated and cross-linked peptides
Tom Altenburg, Sven Giese, Shengbo Wang, Thilo Muth, Bernhard Y. Renard
bioRxiv 2020.05.19.101345; doi: https://doi.org/10.1101/2020.05.19.101345
System requirements:
- Linux (tested on Ubuntu 18.04),
- Conda installed (e.g. follow documentation at https://conda.io/projects/conda/en/latest/user-guide/install/index.html),
- runs on both CPU and GPU machines. (However, for the optional interpretation code snippet having a GPU is essential).
Note: installation of required packages takes up to several minutes and any individual step (on the provided examples) takes less than one minute.
Start by cloning this repository:
git clone https://gitlab.com/dacs-hpi/AHLF.git
cd AHLF
To install the required packages create a new conda environment (required packages are automatically installed when using ahlf_env.yml):
conda env create -f ahlf_env.yml
conda activate ahlf_env
AHLF accepts MGF-files (.mgf). The provided example/example.mgf contains ms/ms spectra from: 10 phosphorylated peptides and 10 unphosphorylated peptides (in this particular order).
The example looks like this (initial ten lines):
BEGIN IONS
TITLE=controllerType=0 controllerNumber=1 scan=10823
PEPMASS=762.798400878906
RTINSECONDS=2914.18014
CHARGE=2+
SCANS=10823
123.1500854 1432.9295654297
148.0610199 3782.2087402344
168.7863617 1880.0770263672
241.082016 5990.2944335938
You can use the model on the provided ./example/example.mgf by using inference.py:
Select a trained model from the model-dir (e.g. alpha_model_weights.hdf5)
python inference.py --tsv ./model/alpha_model_weights.hdf5 ./example/example.mgf results.tsv
the tsv-file contains the prediction score (column='score') and the predicted binary-labels (column='pred'):
title scan score pred
0 controllerType=0 controllerNumber=1 scan=10823 10823 0.98990786 1.0
1 controllerType=0 controllerNumber=1 scan=10943 10943 0.9966457 1.0
2 controllerType=0 controllerNumber=1 scan=12197 12197 0.9999311 1.0
3 controllerType=0 controllerNumber=1 scan=13176 13176 0.99998116 1.0
4 controllerType=0 controllerNumber=1 scan=13283 13283 1.0 1.0
5 controllerType=0 controllerNumber=1 scan=13395 13395 1.0 1.0
6 controllerType=0 controllerNumber=1 scan=13547 13547 0.844296 1.0
7 controllerType=0 controllerNumber=1 scan=13657 13657 0.94686997 1.0
8 controllerType=0 controllerNumber=1 scan=13778 13778 0.8619728 1.0
9 controllerType=0 controllerNumber=1 scan=14343 14343 0.9951794 1.0
10 controllerType=0 controllerNumber=1 scan=30370 30370 0.25745904 0.0
11 controllerType=0 controllerNumber=1 scan=30531 30531 0.024190009 0.0
12 controllerType=0 controllerNumber=1 scan=39105 39105 0.03308204 0.0
13 controllerType=0 controllerNumber=1 scan=52207 52207 0.083298594 0.0
14 controllerType=0 controllerNumber=1 scan=52282 52282 0.017626554 0.0
15 controllerType=0 controllerNumber=1 scan=52369 52369 0.057228595 0.0
16 controllerType=0 controllerNumber=1 scan=55208 55208 0.44605583 0.0
17 controllerType=0 controllerNumber=1 scan=56041 56041 0.0009965003 0.0
18 controllerType=0 controllerNumber=1 scan=77784 77784 0.27133223 0.0
19 controllerType=0 controllerNumber=1 scan=82582 82582 0.2898352 0.0
- phosphorylated = 1.0
- unphosphorylated = 0.0
AHLFx can be used, simply by specifying the dedicated model weights (./AHLFx/alpha_model_x.hdf5). Here, we apply AHLFx to example data (./AHLFx/cross_linking_example.mgf) containing one spectrum of a cross-linked peptide and a spectrum of linear peptide (in this particular order):
python inference.py --tsv ./AHLFx/alpha_model_x.hdf5 ./AHLFx/cross_linking_example.mgf xlink_results.tsv
title scan score pred
0 example_csm 7787 0.5244542 1.0
1 example_psm 5040 0.04545386 0.0
- cross-linked peptide spectrum match (CSM) = 1.0
- linear peptide spectrum match (PSM) = 0.0
Check out the help-page:
python inference.py -h
usage: inference.py [-h] [--tsv] model_weights mgf_file out_file
Read spectra from an mgf-file and output a prediction score.
positional arguments:
model_weights trained model weights
mgf_file input filename: mgf-file containing ms/ms spectra.
out_file output filename
optional arguments:
-h, --help show this help message and exit
--tsv write a tsv-file
In order to train the model, e.g. on user-specific data, you can use training.py:
(the labels are taken from the filenames: ./training/dummy.phos.mgf and ./training/dummy.other.mgf):
Hyperparameters can be changed in the training.py script directly.
python training.py
In order to fine-tune the model, e.g. to a specific instrument type, you can use finetuning.py:
(the labels are taken from the filenames: ./training/dummy.phos.mgf and ./training/dummy.other.mgf):
Hyperparameters can be changed in the finetuning.py script directly.
python finetuning.py model/alpha_model_weights.hdf5 model/finetuned_alpha_model_weights.hdf5 ./training
AHLF can be applied to RAW data (e.g. .RAW-files) by converting them into MGF-files with third-party packages (e.g. https://github.com/compomics/ThermoRawFileParser).
IMPORTANT: this step requires a GPU with latest drivers installed
For interpretation additional packages need to be installed (specified in the `ahlf_interpretation_env.yml'), it is recommended to simply create a new dedicated conda environment:
conda env create -f ahlf_interpretation_env.yml
conda activate ahlf_interpretation_env
If the above mentioned requirments are met then the following command creates a mirrorplot similar to the one shown below:
python interpretation.py
Tom Altenburg (tzom)