Peptide toolkit

This is a toolkit for bioinformatical calculations with peptides (short proteins/amino acid sequences) on Apache Spark.

Features

• Estimation of binding affinities (IC50 value) between peptides and human major histocompatibility complex (MHC) genes
• Generation of all possible peptides with given length
• Inter-peptide distance calculation using PAM/BLOSUM substitution matrices
• Clustering of peptides using k-medoids algorithm
• Clustering of (generated) peptide set around given cluster centers set (of real-world peptides)

Dependencies

• Apache Spark 2.2.0
  Available on https://spark.apache.org/downloads.html
• JUnit 4.12
  Available on https://github.com/junit-team/junit4/
• Apache Maven buildsystem
  Available on http://maven.apache.org/
• spark-kmedoids 0.1.2 : Apache Spark extension implementing k-medoids algorithm.
  Available on https://github.com/tdebatty/spark-kmedoids

3rd-party code included or used

• PSSMHCpan-1.0 : a toolkit for estimation of peptide binding affinities.
  It includes an amount of pre-calculated weight matrices (one for each HLA allele to peptide length pair) and a Perl script for binding affinity estimation.
  Original Perl code from this toolkit was rewritten in Java for Spark.
  One sample weight matrix (for HLA-A0201 allele and 9-meer peptides) is included into this package, the rest should be copied into the working tree from the original PSSMHCpan package as needed.
  Available on https://github.com/BGI2016/PSSMHCpan
• NW-align : Java implementation of Needleman-Wunsch global alignment, included (slightly modified) for benchmarking and comparison purposes only.
  Available on http://zhanglab.ccmb.med.umich.edu/NW-align/
• BLOSUM and PAM amino acid substitution matrices : reference matrices from NCBI are hardcoded.
  Available on ftp://ftp.ncbi.nih.gov/blast/matrices/

Algorithms

In clustering applications, similarity Sab between peptides A and B, using substitution matrix M, is calculated as following :

Sab = 2*SCab/(SCaa + SCbb), where  
SCxy = sum(M(x[i],y[i])), where  
M(k,n) is a substitution matrix value for amino acids k and n, where  
x[i] is the amino acid in peptide x in position i  

Environment

• Netbeans 8.2 IDE
• Java 1.8.0
• executed under Windows 7 x64 and various Linux distributions (Ubuntu 14, CentOs 6 and AltLinux 7.0.5)

Running

• Start Apache Spark master and worker(s) processes
• Submit Spark task :

spark-submit --class <org.package.MainClassName> --master spark://master_hostname:7077 <package_filename.jar> [Config.xml]

• For peptide generation, class name is org.PSSMHC.PSSMHCSpark and package filename is PSSMHCpan-1.0.jar
• For clustering around given centers, class name is org.PeptideClustering.AssignBindersToClusters and package filename is peptide-clustering-1.0.jar
• For k-medoids clustering, class name is org.PeptideClustering.PeptideClusteringMain and package filename is peptide-clustering-1.0.jar

Default config filename is PeptideCfg.xml in current directory.
Configuration is explained in xml comments to the sample config file \PSSMHCpan\src\main\resources\PeptideCfg.xml

So, example command line is :

spark-submit --class org.PSSMHC.PSSMHCSpark --master spark://master_hostname:7077 PSSMHCpan-1.0.jar

Output :

Running each of Spark applications will produce

• some logging in stdout

• depending on xml configuration, some datasets (Spark RDDs saved as text) in output-* folders in Spark work directory.
  In Spark standalone mode, stdout is written the command line windows where you've executed spark-submit, and Spark working directory is your current directory.

  • Peptide generation produces the folder output-pssmhc containing a gzip-ed set of binder peptides, formatted as :
    <peptide>, <binding affinity>
  • Clustering around given centers produces folders
  • output-centers containing a set of binder cluster centers (format described above)
  • output-clusters containing the cluster elements, formatted as :
    (<center peptide>,[<element peptide 1>,<binding affinity 1>, ..., <element peptide N>,<binding affinity N>])
  • K-medoids clustering only logs the solution into stdout in following format :

Medoid : <center peptide> totalSim <sum of similarities between center and elements> avgSim <average similarity between center and elements>  
       Element : <element peptide 1> <binding affinity 1>  
       ...  
       Element : <element peptide N> <binding affinity N>