Introduction

In this worked example you will replicate a PCA on a published dataset.

The example is split into 2 Parts:

In this Data Preparation phase, you will do the following things:

  1. Load the SNP genotypes in .vcf format (vcfR::read.vcfR())
  2. Extract the genotypes into an R-compatible format (vcfR::extract.gt())
  3. Rotate the data into the standard R analysis format (t())
  4. Remove individuals (rows) from the data set that have >50% NAs (using a function I wrote)
  5. Remove SNPs (columns) that are fixed
  6. Impute remaining NAs (using a for() loop)
  7. Save the prepared data as a .csv file for the next step (write.csv())

Biological background

This worked example is based on a paper in the journal Molecular Ecology from 2017 by Jennifer Walsh titled Subspecies delineation amid phenotypic, geographic and genetic discordance in a songbird.

The study investigated variation between two bird species in the genus Ammodramus: A. nenlsoni and A. caudacutus.

The species A. nenlsoni has been divided into 3 sub-species: A. n. nenlsoni, A.n. alterus, and A n. subvirgatus. The other species, A. caudacutus, has been divided into two subspecies, A.c. caudacutus and A.c. diversus.

The purpose of this study was to investigate to what extent these five subspecies recognized by taxonomists are supported by genetic data. The author’s collected DNA from 75 birds (15 per subspecies) and genotyped 1929 SNPs. They then analyzed the data with Principal Components Analysis (PCA), among other genetic analyzes.

This tutorial will work through all of the steps necessary to re-analyze Walsh et al.s data

Tasks

In the code below all code is provided. Your tasks will be to do 2 things:

  1. Give a meaningful title to all sections marked “TODO: TITLE”
  2. Write 1 to 2 sentences describing what is being done and why in all sections marked “TODO: EXPLAIN”

Preliminaries

Load the vcfR and other packages with library().

library(vcfR)    
## Warning: package 'vcfR' was built under R version 4.2.2
## 
##    *****       ***   vcfR   ***       *****
##    This is vcfR 1.13.0 
##      browseVignettes('vcfR') # Documentation
##      citation('vcfR') # Citation
##    *****       *****      *****       *****
library(vegan)
## Loading required package: permute
## Loading required package: lattice
## This is vegan 2.6-2
library(ggplot2)
library(ggpubr)

Make sure that your working directory is set to the location of the file all_loci.vcf.

getwd()
## [1] "C:/Users/adamf/OneDrive/Documents/Pitt/Year 4/Fall2022/BIOSC1540"
list.files()
##  [1] "07-mean_imputation.docx"                        
##  [2] "07-mean_imputation.html"                        
##  [3] "07-mean_imputation.Rmd"                         
##  [4] "08-PCA_worked.html"                             
##  [5] "08-PCA_worked.Rmd"                              
##  [6] "09-PCA_worked_example-SNPs-part1.html"          
##  [7] "09-PCA_worked_example-SNPs-part1.Rmd"           
##  [8] "10-PCA_worked_example-SNPs-part2.Rmd"           
##  [9] "10252022_InClassTest.docx"                      
## [10] "10252022_InClassTest.Rmd"                       
## [11] "all_loci-1.vcf"                                 
## [12] "all_loci.vcf"                                   
## [13] "BIOSC1540FinalProject"                          
## [14] "bird_snps_remove_NAs.html"                      
## [15] "bird_snps_remove_NAs.Rmd"                       
## [16] "center_function.R"                              
## [17] "feature_engineering.Rmd"                        
## [18] "feature_engineering_intro_2_functions-part2.Rmd"
## [19] "my_snps"                                        
## [20] "removing_fixed_alleles.html"                    
## [21] "removing_fixed_alleles.Rmd"                     
## [22] "rsconnect"                                      
## [23] "SNPs_cleaned.csv"                               
## [24] "transpose_VCF_data.html"                        
## [25] "transpose_VCF_data.Rmd"                         
## [26] "vegan_PCA_amino_acids-STUDENT.html"             
## [27] "vegan_PCA_amino_acids-STUDENT.Rmd"              
## [28] "vegan_pca_with_msleep-STUDENT.html"             
## [29] "vegan_pca_with_msleep-STUDENT.Rmd"              
## [30] "walsh2017morphology.csv"
list.files(pattern = "vcf")
## [1] "all_loci-1.vcf" "all_loci.vcf"

Data preparation

TODO: Load VCF File

We need to read our VCF file into the R markdown so we have data and SNPs to work with. 1929 Variants should be processed.

snps <- vcfR::read.vcfR("all_loci.vcf", convertNA  = TRUE)
## Scanning file to determine attributes.
## File attributes:
##   meta lines: 8
##   header_line: 9
##   variant count: 1929
##   column count: 81
## 
Meta line 8 read in.
## All meta lines processed.
## gt matrix initialized.
## Character matrix gt created.
##   Character matrix gt rows: 1929
##   Character matrix gt cols: 81
##   skip: 0
##   nrows: 1929
##   row_num: 0
## 
Processed variant 1000
Processed variant: 1929
## All variants processed

TODO: Encode SNPS

TODO: We are encoding our SNPS into the form of 0, 1 or 2. This is a form of dimension reduction and makes the data easier to work with so we can do PCA and other analyses.

snps_num <- vcfR::extract.gt(snps, 
           element = "GT",
           IDtoRowNames  = F,
           as.numeric = T,
           convertNA = T,
           return.alleles = F)

TODO: Transpose Matrix

TODO: The data is loaded with SNPS in rows and people in columns so we transpose the matrix to flip the rows and columns so we can do PCA on the SNPs.

snps_num_t <- t(snps_num)

TODO: Convert the transposed matrix to a data frame so we can work with it easier.

snps_num_df <- data.frame(snps_num_t)

TODO: Locate NAs in Data Frame

TODO: Create a Function to locate which indexes are NA in a vector so we can identify how many NAs are in each column easily.

find_NAs <- function(x){
  NAs_TF <- is.na(x)
  i_NA <- which(NAs_TF == TRUE)
  N_NA <- length(i_NA)
  
  cat("Results:",N_NA, "NAs present\n.")
  return(i_NA)
}

TODO: We need to use a for loop to go through each row (or person) of our data frame and identify where the NAs occur for each person and save the indices of these so we can process which individuals have missing values.

# N_rows
# number of rows (individuals)
N_rows <- nrow(snps_num_t)

# N_NA
# vector to hold output (number of NAs)
N_NA   <- rep(x = 0, times = N_rows)

# N_SNPs
# total number of columns (SNPs)
N_SNPs <- ncol(snps_num_t)

# the for() loop
for(i in 1:N_rows){
  
  # for each row, find the location of
  ## NAs with snps_num_t()
  i_NA <- find_NAs(snps_num_t[i,]) 
  
  # then determine how many NAs
  ## with length()
  N_NA_i <- length(i_NA)
  
  # then save the output to 
  ## our storage vector
  N_NA[i] <- N_NA_i
}
## Results: 28 NAs present
## .Results: 20 NAs present
## .Results: 28 NAs present
## .Results: 24 NAs present
## .Results: 23 NAs present
## .Results: 63 NAs present
## .Results: 51 NAs present
## .Results: 38 NAs present
## .Results: 34 NAs present
## .Results: 24 NAs present
## .Results: 48 NAs present
## .Results: 21 NAs present
## .Results: 42 NAs present
## .Results: 78 NAs present
## .Results: 45 NAs present
## .Results: 21 NAs present
## .Results: 42 NAs present
## .Results: 34 NAs present
## .Results: 66 NAs present
## .Results: 54 NAs present
## .Results: 59 NAs present
## .Results: 52 NAs present
## .Results: 47 NAs present
## .Results: 31 NAs present
## .Results: 63 NAs present
## .Results: 40 NAs present
## .Results: 40 NAs present
## .Results: 22 NAs present
## .Results: 60 NAs present
## .Results: 48 NAs present
## .Results: 961 NAs present
## .Results: 478 NAs present
## .Results: 59 NAs present
## .Results: 26 NAs present
## .Results: 285 NAs present
## .Results: 409 NAs present
## .Results: 1140 NAs present
## .Results: 600 NAs present
## .Results: 1905 NAs present
## .Results: 25 NAs present
## .Results: 1247 NAs present
## .Results: 23 NAs present
## .Results: 750 NAs present
## .Results: 179 NAs present
## .Results: 433 NAs present
## .Results: 123 NAs present
## .Results: 65 NAs present
## .Results: 49 NAs present
## .Results: 192 NAs present
## .Results: 433 NAs present
## .Results: 66 NAs present
## .Results: 597 NAs present
## .Results: 1891 NAs present
## .Results: 207 NAs present
## .Results: 41 NAs present
## .Results: 268 NAs present
## .Results: 43 NAs present
## .Results: 110 NAs present
## .Results: 130 NAs present
## .Results: 90 NAs present
## .Results: 271 NAs present
## .Results: 92 NAs present
## .Results: 103 NAs present
## .Results: 175 NAs present
## .Results: 31 NAs present
## .Results: 66 NAs present
## .Results: 64 NAs present
## .Results: 400 NAs present
## .Results: 192 NAs present
## .Results: 251 NAs present
## .Results: 69 NAs present
## .Results: 58 NAs present
## .

TODO: Identify how many SNPs need to be recorded for an individual to possess at least 50% collected data. We are identifying where that 50% threshold falls on the histogram and we see only a few individuals are missing >50% of the data.

# 50% of N_SNPs
cutoff50 <- N_SNPs*0.5

hist(N_NA)            
abline(v = cutoff50, 
       col = 2, 
       lwd = 2, 
       lty = 2)

TODO: Determine which individuals have >50% NAs and remove them from the population we are considering as valid data. This is likely because this is not data that is representative of the population due to sub-standard data collections for these individuals.

percent_NA <- N_NA/N_SNPs*100

# Call which() on percent_NA
i_NA_50percent <- which(percent_NA > 50) 

snps_num_t02 <- snps_num_t[-i_NA_50percent, ]

TODO: Labeling our Data

TODO: Assign Row Names to identify each SNP and use the regular expression gsub to easily identify population groups.

row_names <- row.names(snps_num_t02) # Key

row_names02 <- gsub("sample_","",row_names)

sample_id <- gsub("^([ATCG]*)(_)(.*)",
                  "\\3",
                  row_names02)
pop_id <- gsub("[01-9]*",    
               "",
               sample_id)

table(pop_id)  
## pop_id
## Alt Cau Div Nel Sub 
##  15  12  15  15  11

TODO: Omit Invariate Data

TODO: Create a function to omit invariant data (SNPs with no minor alleles thus having a standard deviation of 0) and use this function on our current data frame that has omitted persons with >50% missing data. If we don’t omit invariant data, we will be unable to scale our data for PCA later.

invar_omit <- function(x){
  cat("Dataframe of dim",dim(x), "processed...\n")
  sds <- apply(x, 2, sd, na.rm = TRUE)
  i_var0 <- which(sds == 0)
 
  
  cat(length(i_var0),"columns removed\n")
  
  if(length(i_var0) > 0){
     x <- x[, -i_var0]
  }
  
  ## add return()  with x in it
  return(x)                      
}


snps_no_invar <- invar_omit(snps_num_t02) 
## Dataframe of dim 68 1929 processed...
## 591 columns removed

TODO: Mean Imputation

TODO: Generate a function to impute remaining NAs with the mean value for the specific SNP because we need to handle NAs in some way and omitting them would leave us with minimal data.

snps_noNAs <- snps_no_invar

N_col <- ncol(snps_no_invar)
for(i in 1:N_col){
  
  # get the current column
  column_i <- snps_noNAs[, i]
  
  # get the mean of the current column
  mean_i <- mean(column_i, na.rm = TRUE)
  
  # get the NAs in the current column
  NAs_i <- which(is.na(column_i))
  
  # record the number of NAs
  N_NAs <- length(NAs_i)

  # replace the NAs in the current column
  column_i[NAs_i] <- mean_i
  
  # replace the original column with the
  ## updated columns
  snps_noNAs[, i] <- column_i
  
}

Save the data

Save the data as a .csv file which can be loaded again later.

write.csv(snps_noNAs, file = "SNPs_cleaned.csv",
          row.names = F)

Check for the presence of the file with list.files()

list.files(pattern = ".csv")
## [1] "SNPs_cleaned.csv"        "walsh2017morphology.csv"

Next steps:

In Part 2, we will re-load the SNPs_cleaned.csv file and carry an an analysis with PCA.