retention time alignment tool for large cohort LC-MS data analysis
Project description
DeepRTAlign v1.1.3
Overview
DeepRTAlign is a deep learning-based retention time alignment tool for large cohort LC-MS data analysis.
Installation
- Install Python3
- Install Pytorch CPU version, please refer to https://pytorch.org/
- Install DeepRTAlign by command
pip install deeprtalign
Getting Started
-
Feature lists and sample list should be prepared before running DeepRTAlign. Feature lists are the output of feature extraction tools (DeepRTAlign supports Dinosaur, OpenMS, MaxQuant and XICFinder). The sample list is an excel file recording the correspondences between feature files and sample names. You can find the test data in the example_files folder. Note that if you use MaxQuant as the feature extraction tool, you should use the allPeptides.txt as the input file or files, and the sample list should correspond to the first column of allPeptides.txt file. DO NOT use "allPeptdes.txt" as the file name in sample list if you use MaxQuant.
-
You can get the help information by command
python -m deeprtalign -h
, the basic arguments are as follows:--method {Dinosaur,XICFinder,OpenMS,MaxQuant}, -m {Dinosaur,XICFinder,OpenMS,MaxQuant} the feature extraction method, support Dinosaur, XICFinder, OpenMS and MaxQuant --file_dir FILE_DIR, -f FILE_DIR the data folder contain feature lists --sample_file SAMPLE_FILE, -s SAMPLE_FILE the sample file
As an example, to handle the Dinosaur test data in example_files folder, you can create a new folder and put the file_dir (containing result files from feature extraction tool ) and sample_file in, switch the working directory to this folder, then use command
python -m deeprtalign -m Dinosaur -f file_dir -s sample_file.xlsx
.optional arguments:
--processing_number PROCESSING_NUMBER, -pn PROCESSING_NUMBER processing number, choose according to the number of CPUs --percent PERCENT, -pt PERCENT skip the bins with sample numbers below the percent of total sample numbers --bin_width BIN_WIDTH, -bw BIN_WIDTH the bin width, choose according to the feature extraction step --bin_precision BIN_PRECISION, -bp BIN_PRECISION the decimal place of bins, choose according to the feature extraction step --dict_size DICT_SIZE, -ds DICT_SIZE the dict size, choose according to the memory size --keep_temp KEEP_TEMP, -kt KEEP_TEMP if keep the temp files, 0 remove, 1 keep --begin_step BEGIN_STEP, -bs BEGIN_STEP begin from any step
processing_number (int, default 1) depends on your hardware, percent (float:0-1, default 0) is a threshold, DeepRTAlign will skip the bins with sample numbers below the percent of total sample numbers. bin_width(float, default 0.03) and bin_precision (int, default 2) depends on your feature extraction parameters. dict_size(int , default 1024) depends on your memory size, default 1024MB. If you want to keep the temp files, set the keep_temp(int, default 0) to 1. You can begin from any begin_step(int, default 1).
-
The results will output to the mass_align_all_information folder. In order to avoid a single file from being too large, a single result file contains at most 1,000 groups. Each group contains the features from different samples aligned by DeepRTAlign.
Note
Do not run the different projects under a same folder, the results will be overwritten.
License
GPLv3 (General Public License version 3.0), details in the LICENSE file.
Contacts
For any questions involving DeepRTAlign, please contact us by email.
Yi Liu, leoicarus@163.com
Cheng Chang, changchengbio@163.com or changcheng@ncpsb.org.cn
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Hashes for deeprtalign-1.1.3-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | db79cd153dc6d1e8878e48339e2e5a0ffa9b94bb8e11532b036aa42a7b828a18 |
|
MD5 | 78f9d435c736f31bb70c736ccb07e707 |
|
BLAKE2b-256 | ad758eb19a8c3063df14e5967823a593d33bd913a94a02a10071460963ef439c |