In mathornton01/Genomic-DFT-Yin-R: Fourier Analysis for Genomic Signals

knitr::opts_chunk$set(echo = TRUE)

Genomic Distances

This is an R-Markdown document intended to apply several classification and clustering procedures that are discussed in the "Examining Uses of the DFT-Distance Metric in SARS-CoV-2 Genomes" paper by me (Micah Thornton) and Dr. Monnie McGee. Firstly, there is an R-Package hosted on a git repository that contains a few key features along with the primary data set from the GISAID Initiative. The block immediately to follow need not be run every time, but could be without damaging any dependencies.

Required Packages

This script will examine a few classical distance measures as well as the Fourier Transform distance, and the k-mer frequency vector distances. The Fourier Transform distance produces a power spectra for each of the signals and the k-mer method produces a vector of k-mer counts or frequencies. Distances are then computed (in both cases) by inspecting the Euclidean distance between the vectors or spectra. Other measures such as the classical Jukes-Cantor distance (1969) require that the data be aligned. Hence we can call the k-mer and the Fourier Transform distance 'alignment-free' algorithms, whereas the Jukes-Cantor distance is an example of an algorithm that requires sequence alignment. As such, in order to construct Jukes-Cantor distances the data must be aligned. In the case of the dataset 'sarscvmay' the sequences are not pre-aligned, and therefore would require full multiple sequence alignment prior to the computation of the distances. There are 1,322 total SARS-CoV-2 genomes in the 'sarscvmay' dataset, meaning that a full multiple sequence alignment would take substantial time. As such, sequence alignment is considered for 100 subsamples of 10 sequences, and the distance metrics for these ten are compared by inspecting the correlation prodcued. For the k-mer and Fourier Transform distance measures where an intermediary summary vector is produced, several classification and clustering approaches are examined to determine the utility of these intermediary vectors as numerical summaries for the sequences from which characteristics of interest (particularly geographic region of submission in this case) may be learned.

# Install the Yin Genomic Package if not currently installed. 
YinInstalled <- "YinGenomicDFTDistances" %in% installed.packages(); 
apeInstalled <- "ape" %in% installed.packages(); 
devtoolsInstalled <- "devtools" %in% installed.packages(); 

if (!YinInstalled){
  if (!devtoolsInstalled){
    install.packages("devtools");
  }
  library(devtools); 
  install_github("mathornton01/Genomic-DFT-Yin-R");
}
if (!apeInstalled){
  install.packages("ape");
}
library(YinGenomicDFTDistances);
library(ape);

Timing Analysis

As previously mentioned, the actual time to compute the full multiple sequence alignment for all 1,322 of the SARS-CoV-2 genomes contained within the 'sarscvmay' dataset in the Yin-DFT R-Package would take a considerable amount of time, because the operations generally increase exponentially with the number of sequences being aligned. Therefore, in order to estimate the speed up provided by these alignment free methods over the alignment method, multiple different subsamples of 8 sequences from the original set will be aligned using the multiple-alignment tool clustalw2. In order to use the wrapper function for the aligner in 'ape' you will need to download the executable file for clustalw2 and place this in your system path. You may download the clustalw2 binary files for the Windows OS and Mac OSX here: http://www.clustal.org/download/current/.

M <- 100; # Number of times to re-run the sampling procedure 
n <- 8; # Number of sequences to sample 

alignTimes <- numeric(M);
PScompTimes <- matrix(0,nrow=M,ncol=3); 
merCountTimes <- numeric(M); 

for (i in 1:M){
  sam <- sample(1:nrow(sarscvmay),n)
  dnabinseqs <- as.DNAbin(strsplit(sarscvmay[sam,]$sequences,'')) # Convert to 'Ape' special format DNAbin 
  tsalign <- Sys.time();
  alignsam <- clustal(dnabinseqs);
  tealign <- Sys.time();
  alignt <- tealign-tsalign; 
  alignTimes[i] <- alignt; 
  tsPScompEnc <- Sys.time(); 
  enGenomes <- encodeGenomes(sarscvmay[sam,]$sequences);
  tePScompEnc <- Sys.time();
  PScompEncTime <- tePScompEnc - tsPScompEnc; 
  tsPScompPS <- Sys.time(); 
  PS <- getPowerSpectraEnsemble(enGenomes); 
  tePScompPS <- Sys.time();
  PScompPSTime <- tePScompPS-tsPScompPS; 
  tsPScompES <- Sys.time(); 
  esPS <- evenlyScaleEnsemble(PS);
  tePScompES <- Sys.time(); 
  PScompESTime <- tePScompES-tsPScompES; 


}

Classification with Alignment-Free Intermediary Vectors

For this section of the analysis, the intermediary vectors for the Fourier Distance and the k-mer frequencies are examined to determine whether similar classification accuracy on the general geographic region of interest for the data can be achieved using either of the two methods.

mathornton01/Genomic-DFT-Yin-R documentation built on Dec. 21, 2021, 2:52 p.m.