How to calculate the group mean for each group in a very large matrix, after subsetting the matrix according to the rows/column names?

Question

So I have a distance matrix of pairwise distances between DNA sequences where the name of each sequence is something identical to ID0001|Species name. The code I used to generate it is the following:

library(ape)
myseq <- read.dna("sequences.fasta", format = "fasta")
mydist <- dist.dna(myseq)

The resulting matrix looks like this:

What I am trying to do is to calculate the average distance within each genus. The way I have found to do this is by converting the matrix to a data frame, then separating the sequence ID and the species name, and finally extracting the genus (i.e. the first word of the species name):

library(reshape2)
library(stringr)
library(tidyr)

df_dist<-setNames(reshape2::melt(as.matrix(mydist)), c('rows', 'vars', 'values'))
names(df_dist)<-c("record1","record2","dist")
df_dist2<-separate(df_dist,record1,into = c("ID1","species1"),sep = "\\|",remove = FALSE,extra = "merge")
df_dist3<-separate(df_dist2,record2,into = c("ID2","species2"),sep = "\\|",remove = FALSE,extra = "merge")
df_dist3$genus1<- str_extract(df_dist3$species1, '[A-Za-z]+')
df_dist3$genus2<- str_extract(df_dist3$species2, '[A-Za-z]+')

Then what I do is I retain only the distances between the same genus, and after that I remove the instances where the distance was calculated between sequences belonging to the same species within a genus (not useful for my analysis):

congeneric<-df_dist3[df_dist3$genus1==df_dist3$genus2,]
congeneric2<-congeneric[congeneric$species1 != congeneric$species2,]

Then I calculate the mean distance of each genus and save them as a data frame:

congeneric_distances<-congeneric2 %>%
  group_by(genus1) %>%
  summarise_at(vars(dist), list(name = mean))

This code works for some matrices, but for very large ones, I am not being able to run the code, I get returned a message such as this:

Error: cannot allocate vector of size X Gb

I am wondering if there is a way to smooth out the code and make it less memory intensive. One of the steps where the code mostly "freezes" is when using the separate function, but sometimes it fails in other places. I have also been told that I shouldn't convert it to a data frame, as it is more memory intensive than the matrix itself.

I would wrap your code to analyze each genus into a function. Then I would split `myseq` into different objects for each genus, and run each genus separately. Right now, you are calculating lots of things between species that you don't want to compare. — Axeman, May 09 '23 at 16:51
I agree with Axeman. I think you should do the extraction of genus names on the distance matrix row names. Then you can loop through the genus names, creating a submatrix of the large distance matrix each time by subsetting both the rows and columns of the large distance matrix. Then calculate the average of that subset each time. — qdread, May 09 '23 at 17:14
Even the distance matrix may be costly on large datasets. `rownames(myseq)` gives the sample names, and I think you can subset like `myseq[samples_you_want, ]`. — Axeman, May 09 '23 at 17:19

Gerald T · Accepted Answer · 2023-05-24T11:21:17.363

It is a bit hard to answer your question without a reproducible example or some information about your computer. Below is my suggestion.

Some dummy data resembling your dataset:

# Create a dummy example
set.seed(1)
dummy_mat <- matrix(rnorm(n=9*9), ncol = 9, nrow = 9)
diag(dummy_mat) <- 0
dummy_mat[lower.tri(dummy_mat)] <- dummy_mat[upper.tri(dummy_mat)]

# Create col and row names
x <- paste0("Genus", LETTERS[1:3])
y <- paste0(" species", 1:3)

set.seed(1)
ids <- paste0(sample(LETTERS, 9, replace = T), sample(1000:9999, 9))
nms <- paste0(ids, "|", do.call(paste0, expand.grid(x, y)))

colnames(dummy_mat) <- nms
rownames(dummy_mat) <- nms

This is the output:

> dummy_mat
                      Y9788|GenusA species1 D2300|GenusB species1 G9521|GenusC species1 A2798|GenusA species2 B9228|GenusB species2 W2128|GenusC species2 K9920|GenusA species3 N1877|GenusB species3 R9676|GenusC species3
Y9788|GenusA species1             0.0000000            -0.3053884             0.8212212           -1.47075238            -0.3942900            -0.7074952           1.433023702            0.02800216           0.610726353
D2300|GenusB species1            -0.3053884             0.0000000             0.5939013           -0.47815006            -0.0593134             0.3645820           1.980399899           -0.74327321          -0.934097632
G9521|GenusC species1             0.8212212             1.1000254             0.0000000            0.41794156             1.1000254             0.7685329          -0.367221476            0.18879230          -1.253633400
A2798|GenusA species2             0.5939013             0.7631757             1.4330237            0.00000000             0.7631757            -0.1123462          -1.044134626           -1.80495863           0.291446236
B9228|GenusB species2            -1.4707524            -0.7074952             1.9803999            0.02800216             0.0000000             0.8811077           0.569719627            1.46555486          -0.443291873
W2128|GenusC species2            -0.4781501             0.3645820            -0.3672215           -0.74327321             0.1532533             0.0000000          -0.135054604            0.15325334           0.001105352
K9920|GenusA species3             0.4179416             0.7685329            -1.0441346            0.18879230             2.1726117            -1.2536334           0.000000000            2.17261167           0.074341324
N1877|GenusB species3            -0.3942900            -0.1123462             0.5697196           -1.80495863             0.6107264             0.2914462           0.001105352            0.00000000          -0.589520946
R9676|GenusC species3            -0.0593134             0.8811077            -0.1350546            1.46555486            -0.9340976            -0.4432919           0.074341324           -0.58952095           0.000000000

Now we can try to answer your question.

1] First I would suggest to list all unique genera you have in your dataset. I will use the library stringr (but it can also be done in base R) and the fact that all the genera are delimited by "|" and " ".

# Identify all genera you have
nms <- colnames(dummy_mat)

library(stringr)
nms <- stringr::str_extract(nms, "(?<=\\|).*(?= )") # delimiters of the Genus "|" and " " (whitespace)
nms <- unique(nms)

2] Now we can create a data.frame that will store all your average distances (preallocation is always a good thing when dealing with large datasets)

# Create a data frame to pre allocate memory 
df <- data.frame(genus = nms, meanDist = rep(as.numeric(NA), length(nms)))

> df
   genus meanDist
1 GenusA       NA
2 GenusB       NA
3 GenusC       NA

3] Compute the average distance. Within each iteration we only need the upper triangular matrix (diagonal is always 0, lower triangular is equal to upper triangular matrix). Instead of recreating new data.frames at each iteration, we will try to subset the existing matrix.

for(n in seq_along(df$genus)){
  
  pattern <- df$genus[n] 
  ind     <- grep(pattern, colnames(dummy_mat))
  
  m <- dummy_mat[ind, ind]
  
  df$meanDist[n] <- mean(m[upper.tri(m)])
  
}

4] Our results

> df
   genus   meanDist
1 GenusA -0.3606211
2 GenusB  0.2209894
3 GenusC -0.1613317

Let me know if this solve your problem.

UPDATE

Using the profiling tool provided by the library profvis we can see that te loop itself doesn't use up much memory even with a distance matrix 10.000x10.000

library(profvis)

profvis({
    
    # Create a dummy example
    set.seed(1)
    dummy_mat <- matrix(rnorm(n=10000*10000), ncol = 10000, nrow = 10000)
    diag(dummy_mat) <- 0
    dummy_mat[lower.tri(dummy_mat)] <- dummy_mat[upper.tri(dummy_mat)]
    
    # Create col and row names
    x <- paste0("Genus", LETTERS[1:10])
    y <- paste0(" species", 1:1000)
    
    set.seed(1)
    ids <- paste0(sample(LETTERS, 9, replace = T), sample(1000:9999, 9))
    nms <- paste0(ids, "|", do.call(paste0, expand.grid(x, y)))
    
    colnames(dummy_mat) <- nms
    rownames(dummy_mat) <- nms
    
    
    # Identify all genera you have
    nms <- colnames(dummy_mat)
    
    library(stringr)
    nms <- stringr::str_extract(nms, "(?<=\\|).*(?= )") # delimiters of the Genus "|" and " " (whitespace)
    nms <- unique(nms)
    
    # Create a data frame to pre allocate memory 
    df <- data.frame(genus = nms, meanDist = rep(as.numeric(NA), length(nms)))
    
    
    # Compute mean distances
    for(n in seq_along(df$genus)){
      
      pattern <- df$genus[n]
      ind     <- grep(pattern, colnames(dummy_mat))
      
      m <- dummy_mat[ind, ind]
      
      df$meanDist[n] <- mean(m[upper.tri(m)])
      
    }
    
}) #profvis end

score 1 · Answer 2 · answered May 25 '23 at 14:12

Using the sample data from Gerald T, here is an alternative approach:

## save species names, just in case
nms_save = colnames(dummy_mat)

## replace row and column names with genus
nms = sub(".*\\|([^ ]+) .*", "\\1", nms_save)
rownames(dummy_mat) = nms
colnames(dummy_mat) = nms

## compute distance means for each genus
gen_means = by(dummy_mat, nms, function(x) {
    ## keep only columns for current genus
    x_subset = x[,colnames(x) == rownames(x)[1]]
    mean(x_subset[upper.tri(x_subset)])
})

Result:

> data.frame(gen_means)
        gen_means
GenusA -0.3606211
GenusB  0.2209894
GenusC -0.1613317

How to calculate the group mean for each group in a very large matrix, after subsetting the matrix according to the rows/column names?

2 Answers2

UPDATE