pySAR

Table of Contents

• Introduction
• Requirements
• Installation
• Usage
• Directories
• Tests
• Issues
• Contact
• License
• References

Research Article

The research article that accompanied this software is titled: "Machine Learning Based Predictive Model for the Analysis of Sequence Activity Relationships Using Protein Spectra and Protein Descriptors". This research article is uploaded to the repository as pySAR_research.pdf. The article was published in the Journal of Biomedical Informatics and is available here. There is also a quick jupyter notebook demo of pySAR available here.

How to cite

Mckenna, A., & Dubey, S. (2022). Machine learning based predictive model for the analysis of sequence activity relationships using protein spectra and protein descriptors. Journal of Biomedical Informatics, 128(104016), 104016. https://doi.org/10.1016/j.jbi.2022.104016

Introduction

pySAR is a Python library for analysing Sequence Activity Relationships (SARs) of protein sequences. pySAR offers extensive and verbose functionalities that allow you to numerically encode a dataset of protein sequences using a large abundance of available methodologies and features. The software uses physiochemical and biochemical features from the Amino Acid Index (AAI) database [1] as well as allowing for the calculation of a range of structural, physiochemical and biochemical protein descriptors.

After finding the optimal technique and feature set at which to encode your dataset of sequences, pySAR can then be used to build a predictive regression model with the training data being that of the encoded sequences, and training labels being the experimentally pre-calculated activity values for each protein sequence. This model maps a set of protein sequences to the sought-after activity value, being able to accurately predict the activity/fitness value of new unseen sequences. The use-case for the software is within the field of protein engineering and Directed Evolution, where a user has a set of experimentally determined activity values for a library of mutant protein sequences and wants to computationally predict the sought activity value for a selection of mutated sequences, in the aim of finding the best sequence that minimises/maximises their activity value.

Two additional custom-built softwares were created alongside pySAR - aaindex and protpy. The aaindex software package is used for parsing the amino acid index which is a database of numerical indices representing various physicochemical and biochemical properties of amino acids and pairs of amino acids [1]. protpy is used for calculating a series of protein physiochemical, biochemical and structural protein descriptors. Both of these software packages are integrated into pySAR but can also be used individually for their respective purposes.

Requirements

• Python >= 3.7
• aaindex >= 1.0.4
• protpy >= 1.0.7
• requests >= 2.25.0
• numpy >= 1.24.2
• pandas >= 1.5.3
• scikit-learn >= 1.2.1
• scipy >= 1.10.1
• tqdm >= 4.65.0
• seaborn >= 0.12.2
• biopython >= 1.81
• varname >= 0.11.0

Installation

Install the latest version of pySAR via PyPi using pip:

pip3 install pysar --upgrade

Installation from source:

git clone -b master https://github.com/amckenna41/pySAR.git
python3 setup.py install
cd pySAR

Usage

Confile File

pySAR works through JSON configuration files. There are many different customisable parameters for the functionalities in pySAR including the metaparameters of each of the available protein descriptors, all Digital Signal Processing (DSP) parameters in the pyDSP module, the type of regression model to use and parameters specific to the dataset. These config files offer a more straightforward way of making any changes to the pySAR pipeline. The names of All the parameters as listed in the example config files must remain unchanged, only the value of each parameter should be changed, any parameters not being used can be set to null. An example of the config file used in my research project, with most of the available parameters, can be seen below and in config/thermostability.json.

{
    "dataset": 
      {
        "dataset": "thermostability.txt",
        "sequence_col": "sequence",
        "activity": "T50"
      },
    "model": 
      {
        "algorithm": "plsregression",
        "parameters": "",
        "test_split": 0.2
      },
    "descriptors":
        {
          "descriptors_csv": "descriptors.csv",
          "descriptors": {
            "all_desc": 0,
            "amino_acid_composition": 1,
            "dipeptide_composition": 1,
            ...
        }
      },
    "descriptor_properties":
    {
      "normalized_moreaubroto_autocorrelation":
        {
        "lag":30,
        "properties":["CIDH920105", "BHAR880101", "CHAM820101", "CHAM820102",
          "CHOC760101", "BIGC670101", "CHAM810101", "DAYM780201"]
        },
      ...
      ...
      },
    "pyDSP":
      {
        "use_dsp": 1,
        "spectrum": "power",
        "window": {
          "type": "hamming",
          ...
        },
        "filter": {
          "type": null,
          ...
        }
      }
  }

Encoding using all 566 AAIndex indices

Encoding protein sequences in dataset using all 566 indices in the AAI database. Each sequence encoded via an index in the AAI can be passed through an additional step where its protein spectra can be generated following an FFT. pySAR supports generation of the power, imaginary, real or absolute spectra as well as other DSP functionalities including windowing, convolution and filter functions. In the example below, the encoded sequences will be used to generate a imaginary protein spectra with a blackman window function applied. This will then be used as feature data to build a predictive model that can be used for accurate prediction of the sought activity value of unseen protein sequences. The encoding class also takes only the JSON config file as input which will have all the required parameter values. The output results will show the calculated metric values for each index in the AAI when measuring predicted vs observed activity values for the unseen test sequences.

from pySAR.encoding import *

'''test_config.json
{
  "dataset": 
    {
    "dataset": "test_dataset1.txt",
    "activity": "sought_activity"
    ...
    }
  "model": 
    {
    "algorithm": "randomforest",
    ...
    }
  "pyDSP": 
    {
    "use_dsp": 1,
    "spectrum": "imaginary",
    "window": "blackman"
    }
}
'''
#create instance of Encoding class, using RF algorithm with its default params
encoding = Encoding(config_file='test_config.json')

#encode sequences using all indices in the AAI if input parameter "aai_indices" is empty/None
aai_encoding = encoding.aai_encoding()

Output results showing AAI index and its category as well as all the associated metric values for each predictive model:

Encoding using list of 4 AAI indices, with no DSP functionalities

Same procedure as prior, except 4 indices from the AAI are being specifically input into the function, with the encoded sequence output being concatenated together and used as feature data to build the predictive PlsRegression model with its default parameters. The config parameter use_dsp tells the function to not generate the protein spectra or apply any additional DSP processing to the sequences.

from pySAR.encoding import *

'''test_config2.json
{
  "dataset": 
    {
    "dataset": "test_dataset2.txt",
    "activity": "sought_activity"
    ...
    }
  "model": 
    {
    "algorithm": "plsreg",
    "parameters": null
    }
  "pyDSP": 
    {
    "use_dsp": 0,
    ...
    }
}
'''
#create instance of Encoding class, using PLS algorithm with its default params
encoding = Encoding(config_file='test_config2.json')

#encode sequences using 4 indices specified by user, use_dsp = False
aai_encoding = encoding.aai_encoding(aai_list=["PONP800102","RICJ880102","ROBB760107","KARS160113"])

Output DataFrame showing the 4 predictive models built using the PLS algorithm, with the 4 indices from the AAI:

Encoding protein sequences using their calculated protein descriptors

Calculate the protein descriptor values for a dataset of protein sequences from the 15 available descriptors in the descriptors module. Use each descriptor as a feature set in the building of the predictive models used to predict the activity value of unseen sequences. By default, the function will look for a csv file pointed to by the "descriptors_csv" parameter in the config file that contains the pre-calculated descriptor values for a dataset. If file is not found then all descriptor values will be calculated for the dataset using the descriptors_ module. If a descriptor in the config file is to be used in the feature data, its parameter is set to true/1. If all_desc is set to true/1 then all available descriptors are calculated using their respective functions.

from pySAR.encoding import *

'''test_config3.json
{
  "dataset": 
    {
    "dataset": "test_dataset3.txt",
    "activity": "sought_activity"
    ...
    }
  "model": 
    {
    "algorithm": "adaboost",
    "parameters": [{
      "estimators": 100,
      "learning_rate": 1.5
      ...
    },
  "descriptors": 
  {
    "descriptors_csv": "precalculated_descriptors.csv",
    "all_desc": 0,
    "moreaubroto_autocorrelation": {
      "lag": 30,
      "properties": ["CIDH920105", "BHAR880101", "CHAM820101", "CHAM820102",
        "CHOC760101", "BIGC670101", "CHAM810101", "DAYM780201"],
      "normalize": 1
    },
    ...
  }
}
'''
#create instance of Encoding class using AdaBoost algorithm, using 100 estimators & a learning rate of 1.5
encoding = Encoding(config_file='test_config3.json')

#building predictive models using all available descriptors
#   calculating evaluation metrics values for models and storing into desc_results_df DataFrame
desc_results_df = encoding.descriptor_encoding()

Output results showing the protein descriptor and its group as well as all the associated metric values for each predictive model:

Encoding using AAI + protein descriptors

Encoding protein sequences in dataset using all 566 indices in the AAI database combined with protein descriptors. All 566 indices can be used in concatenation with 1, 2 or 3 descriptors. E.g: at each iteration the encoded sequences using the indices from the AAI will be used to generate a protein spectra using the power spectrum with no window function applied, this will then be combined with the feature set generated from the dataset's descriptor values and used to build a predictive model that can be used for accurate prediction of the sought activity value of unseen protein sequences. The output results will show the calculated metric values when measuring predicted vs observed activity values for the test sequences.

from pySAR.encoding import *

'''test_config4.json
{
  "dataset": 
  {
    "dataset": "test_dataset4.txt",
    "activity": "sought_activity"
    ...
  }
  "model": 
  {
    "algorithm": "randomforest",
    "parameters": 
      {
      "estimators": 100,
      "learning_rate": 1.5,
      ...
      }
  },
  "descriptors": 
  {
    "descriptors_csv": "precalculated_descriptors.csv",
    "all_desc": 0,
    "moreaubroto_autocorrelation": {
      "lag": 30,
      "properties": ["CIDH920105", "BHAR880101", "CHAM820101", "CHAM820102",
        "CHOC760101", "BIGC670101", "CHAM810101", "DAYM780201"],
      "normalize": 1
    },
    ...
  },
  "pyDSP": 
  {
    "use_dsp": 1,
    "spectrum": "power",
    "window": ""
    ...
  }
}
'''
#create instance of Encoding class using RF algorithm, using 100 estimators with a learning rate of 1.5
encoding = Encoding('test_config4.json')

#building predictive models using all available aa_indices + combination of 2 descriptors,
#   calculating evaluation metric values for models and storing into aai_desc_results_df DataFrame
aai_desc_results_df = encoding.aai_descriptor_encoding(desc_combo=2)

Output results showing AAI index and its category, the protein descriptor and its group as well as the R2 and RMSE values for each predictive model:

Building predictive model from AAI and protein descriptors:

e.g: the below code will build a PlsRegression model using the AAI index CIDH920105 and the 'amino acid composition' descriptor. The index is passed through a DSP pipeline and is transformed into its informational protein spectra using the power spectra, with a hamming window function applied to the output of the FFT. The concatenated features from the AAI index and the descriptor will be used as the feature data in building the PLS model.

import pySAR as pysar   #import pySAR package

'''test_config5.json
{
  "dataset": 
  {
    "dataset": "test_dataset5.txt",
    "activity": "sought_activity"
    ...
  },
  "model": 
  {
    "algorithm": "plsregression",
    "parameters": "",
    ...
  },
  "descriptors": 
  {
    "descriptors_csv": "precalculated_descriptors.csv",
    "all_desc": 0,
    "moreaubroto_autocorrelation": {
      "lag": 30,
      "properties": ["CIDH920105", "BHAR880101", "CHAM820101", "CHAM820102",
        "CHOC760101", "BIGC670101", "CHAM810101", "DAYM780201"],
      "normalize": 1
    },
    ...
  },
  "pyDSP": 
  {
    "use_dsp": 1,
    "spectrum": "power",
    "window": "hamming",
    ...
  }
}
'''
#create instance of PySAR class
pySAR = pysar.PySAR(config_file="test_config5.json")
"""
PySAR parameters:

:config_file : str
    full path to config file containing all required pySAR parameters.

"""
#encode protein sequences using both the CIDH920105 index + aa_composition descriptor.
results_df = pySAR.encode_aai_desc(indices="CIDH920105", descriptors="amino_acid_composition")

Generate all protein descriptors

Prior to evaluating the various available properties and features at which to encode a set of protein sequences, it is reccomened that you pre-calculate all the available descriptors in one go, saving them to a csv for later that pySAR will then import from. Output values are stored in csv set by descriptors_csv config parameter. Output will be of the shape N x 9920, using the default parameters, where N is the number of protein sequences in the dataset, but the size of the 2nd dimension (total number of features calculated from all 15 descriptors) may vary depending on some descriptor-specific metaparameters. Setting all_desc parameter to True means all descriptors will be calculated, by default this is False.

from pySAR.descriptors_ import *

'''test_config6.json
{
  "dataset": 
  {
    "dataset": "test_dataset5.txt",
    "activity": "sought_activity"
    ...
  },
  "model": 
  {
    ...
  }
  "descriptors": 
  {
    "descriptors_csv": "precalculated_descriptors",
    "all_desc": 1,
    "moreaubroto_autocorrelation": {
      "lag": 30,
      "properties": ["CIDH920105", "BHAR880101", "CHAM820101", "CHAM820102",
        "CHOC760101", "BIGC670101", "CHAM810101", "DAYM780201"],
      "normalize": 1
    },
    ...
  },
  "pyDSP": 
  {
    ...
  }
}
'''
#calculating all descriptor values and storing in file named by parameter descriptors_csv
desc = Descriptors("test_config6")

Get record from AAIndex database

The AAIndex class offers diverse functionalities for obtaining any element from any record in the database. Each record is stored in json format in a class attribute called aaindex_json. You can search for a particular record by its index code, description or reference. You can also get the index category, and importantly its associated amino acid values.

from aaindex import aaindex1 

record = aaindex1['CHOP780206'] #get full record
description = aaindex1['CHOP780206'].description #get record's description
refs = aaindex1['CHOP780206'].references     #get record's references 
category = aaindex1['CHOP780206'].category #get record's category
notes = aaindex1['CHOP780206'].notes #get record's notes
correllation_coefficient = aaindex1['CHOP780206'].correllation_coefficient #get record's correllation_coefficient
pmid = aaindex1['CHOP780206'].pmid #get record's pmid
values = aaindex1['CHOP780206'].values  #get amino acid values from record

num_record = aaindex1.num_records()  #get total number of records
record_names = aaindex1.record_names() #get list of all record names
amino_acids = aaindex1.amino_acids() #get list of all canonical amino acids

Directories

• /config - configuration files for the example datasets that pySAR has been tested with, as well as the thermostability.json config file that was used in the research. These config files should be used as a template for future datasets used with pySAR.
• /docs - documentation for pySAR (pending).
• /example_datasets - example datasets used for the building and testing of pySAR, including the thermostability dataset used in the research. The format of these datasets shoould be used as a template for future datasets used with pySAR.
• /images - all images used throughout the repo.
• /pySAR - source code for pySAR software.
• /tests - unit and integration tests for pySAR.

Issues

Any issues, errors or bugs can be raised via the Issues tab in the repository.

Tests

To run all tests, from the main pySAR repo folder run:

python3 -m unittest discover tests

To run tests for specific module, from the main pySAR repo folder run:

python -m unittest tests.MODULE_NAME -v

Contact

If you have any questions or comments, please contact amckenna41@qub.ac.uk or raise an issue on the Issues tab.

License

Distributed under the MIT License. See LICENSE for more details.

References

[1]: Kawashima, S. and Kanehisa, M., 2000. AAindex: amino acid index database. Nucleic acids research, 28(1), pp.374-374. DOI: 10.1093/nar/27.1.368

[2]: Fontaine NT, Cadet XF, Vetrivel I. Novel Descriptors and Digital Signal Processing- Based Method for Protein Sequence Activity Relationship Study. Int J Mol Sci. 2019 Nov 11;20(22):5640. doi: 10.3390/ijms20225640. PMID: 31718061; PMCID: PMC6888668.

[3]: Cadet, F., Fontaine, N., Li, G. et al. A machine learning approach for reliable prediction of amino acid interactions and its application in the directed evolution of enantioselective enzymes. Sci Rep 8, 16757 (2018).

[4]: Lutz S. Beyond directed evolution--semi-rational protein engineering and design. Curr Opin Biotechnol. 2010 Dec;21(6):734-43. doi: 10.1016/j.copbio.2010.08.011. Epub 2010 Sep 24. PMID: 20869867; PMCID: PMC2982887.

[5]: Yang, K.K., Wu, Z. & Arnold, F.H. Machine-learning-guided directed evolution for protein engineering. Nat Methods 16, 687–694 (2019). https://doi.org/10.1038/s41592-019-0496-6

[6]: Yuting Xu, Deeptak Verma, Robert P. Sheridan, Andy Liaw, Junshui Ma, Nicholas M. Marshall, John McIntosh, Edward C. Sherer, Vladimir Svetnik, and Jennifer M. Johnston Journal of Chemical Information and Modeling 2020 60 (6), 2773-2790 DOI: 10.1021/acs.jcim.0c00073

[7]: Medina-Ortiz, D., Contreras, S., Amado-Hinojosa, J., Torres-Almonacid, J., Asenjo, J. A., Navarrete, M., & Olivera-Nappa, Á. (2020). Combination of digital signal processing and assembled predictive models facilitates the rational design of proteins. ArXiv [Cs.CE].

Back to top