BEDMatrix

The BEDMatrix package allows access to genotypes from a BED file without loading the entire matrix in memory. This is very convenient for large datasets, as long as only small portions of the data are pulled at once. However, there are some disadvantages:

• The BEDMatrix return object is not a regular R matrix, which confuses some users.
• The BEDMatrix package provides no way to access the full annotation tables (BIM and FAM files).
  • From BIM, only the locus ID and reference allele are returned, merged into a vector of strings and placed as column names.
  • From FAM, only the family and individual IDs are returned, merged into a vector of strings and placed as row names.
• BEDMatrix does not have any write functions.

Here is an example of that usage and how the data compares. Note in particular that the BEDMatrix representation of the genotype matrix is transposed compared to the genio matrix:

library(BEDMatrix)
# Time it too.
# Although the BIM and FAM tables are not returned,
# they are partially parsed and kept in memory,
# which can take time for extremely large files
time_read_bedmatrix_1 <- system.time(
    X_BEDMatrix <- BEDMatrix(file_plink)
)

# Inspect the first 10 loci and individuals as usual.
# Note it is transposed compared to our X.
# Also note the column and row names are different from genio's.
X_BEDMatrix[1:10, 1:10]

Therefore, for very large datasets, if it suffices to access the genotype data in small portions and the user is willing to deal with the limitations of the object, and no detailed annotation table information is required, then BEDMatrix is a better solution than the read_bed function provided in the genio package.

To compare the two genotype reading functions on an equal footing, let us assume that the entire genotype matrix is required in memory and stored in a regular R matrix.

# This turns it into a regular R matrix.
# Since most of the reading is actually happening now,
# we time this step now.
time_read_bedmatrix_2 <- system.time(
    X_BEDMatrix_Rmat <- as.matrix(X_BEDMatrix)
)
time_read_bedmatrix_2
# Now we can test that the BEDMatrix agrees with the original matrix we simulated.
# Need to transpose first.
stopifnot( all( X == t(X_BEDMatrix_Rmat), na.rm = TRUE) )

snpStats

The snpStats package is the most fully-featured alternative to genio. One of its advantages is its memory efficiency, encoding the entire data in less memory than a native R matrix. However, this memory efficiency comes at the cost of making the data harder to access, especially for inexperienced R users. Like genio, and in contrast to the other packages, snpStats also reads the BIM and FAM tables fully, and provides a BED writer, but it is considerably harder to use.

First we illustrate data parsing. The annotation tables are similar but column names are different, and certain missing values (zero in text files) are converted to NA instead.

library(snpStats)

# Read data, time it.
time_read_snpStats_1 <- system.time(
    data_snpStats <- read.plink(file_plink)
)
time_read_snpStats_1

# Inspect the data

# Genotypes.
# Note data is not visualized this way.
# This matrix is also transposed compared to the genio matrix.
data_snpStats$genotypes

# Locus annotations
head( data_snpStats$map )

# Individual annotations
head (data_snpStats$fam )

As for BEDMatrix, assuming we ultimately desire to convert the entire data into a regular R matrix, an extra step is required:

# Transpose, then convert to a regular R matrix.
# Let's time this step too.
time_read_snpStats_2 <- system.time(
    X_snpStats <- as( t(data_snpStats$genotypes), 'numeric')
)
time_read_snpStats_2

# Now we can visualize the matrix.
# First 10 loci and individuals, as before.
# Note that, compared to (genio, BEDMatrix), alleles are encoded in reverse,
# so 0s and 2s are flipped in this matrix.
X_snpStats[1:10, 1:10]

# Again verify that the matrices are identical.
# (Here 2-X flips back 0s and 2s)
stopifnot( all( X == 2 - X_snpStats, na.rm = TRUE) )

snpStats is the only package other than genio to provide a BED writer. Here we demonstrate how hard it is to use it to write our data.

# Let's write this to another file
file_plink_copy <- tempfile('vignette-random-data-copy')

# Copy objects to not change originals
X_snpStats <- X
bim_snpStats <- as.data.frame(bim) # to use rownames
fam_snpStats <- as.data.frame(fam) # ditto

# All data requires matching row and/or column names.
# These first two were already done above.
#rownames(X_snpStats) <- bim$id
#colnames(X_snpStats) <- fam$id
# Row names here are redundant but required.
rownames(bim_snpStats) <- bim$id
rownames(fam_snpStats) <- fam$id

# We shall time several required steps in order to write genotypes in a standard R matrix,
# and the related annotation tables, to BED.
time_write_snpStats <- system.time({
    # Transpose and convert our genotypes to SnpMatrix object.
    # We flip 0s and 2s before converting
    X_snpStats <- as(2 - t(X_snpStats), 'SnpMatrix')
    
    # This complicated command is required to write the data.
    # Although X, fam, and bim are passed as for genio's write_plink,
    # here the name of every column must be specified (there are no reasonable defaults).
    # Interestingly, the parameter names of snpStats' write.plink
    # do not agree with read.plink from the same package.
    write.plink(
        file_plink_copy,
        snps = X_snpStats,
        subject.data = fam_snpStats,
        pedigree = fam,
        id = id,
        father = pat,
        mother = mat,
        sex = sex,
        phenotype = pheno,
        snp.data = bim_snpStats,
        chromosome = chr,
        genetic.distance = posg,
        position = pos,
        allele.1 = ref,
        allele.2 = alt
    )
})

# remove the new file, no longer need it
delete_files_plink(file_plink_copy)

Overall time comparisons

Note that reading performance varies on different machines depending on the balance between hard drive access and processor speeds (where the relative bottleneck is). That being said, the genio reader is consistently the fastest, if not a close second (see below for tests on your machine), for several reasons. Genotypes are read very efficiently using Rcpp code, and stored directly into an R matrix, which is most efficient if that is the end goal. The annotation tables are also read most efficiently, using the readr package internally.

The BEDMatrix reader is also written in Rcpp, but its main goal is to provide efficient random access to genotypes stored in a file, which makes obtaining an R matrix more awkward, though surprisingly without paying a time penalty for it. The annotation tables are read with data.table, which is actually the fastest, though the difference in performance is small compared to readr for reasonably-sized files. However, BEDMatrix does not process or return full annotation tables, which gives it an unfair advantage compared to genio, as BEDMatrix takes shortcuts to only read the columns it needs. If the annotation tables are needed, reading them will incur an additional (and in some cases considerable) time penalty.

The snpStats reader is written in C, so it is also very fast for its initial step, but converting the genotypes from the snpStats format into a native R matrix proves too expensive. The annotation tables are read with the base function read.table, which is also the slowest. In terms of completeness of output (full genotypes and annotations tables), only this package matches genio, which makes it the fairest comparison.

# Extract component 3 of each time object,
# which is is total time elapsed.
# Sum the two steps it takes for each of BEDMatrix and snpStats to obtain a native R matrix.
times_read <- c(
    time_read_genio[3],
    time_read_bedmatrix_1[3] + time_read_bedmatrix_2[3],
    time_read_snpStats_1[3] + time_read_snpStats_2[3]
)
names_read <- c(
    'genio',
    'BEDMatrix',
    'snpStats'
)
# Create barplot
barplot(
    times_read,
    names.arg = names_read,
    main = 'BED reader runtimes',
    xlab = 'packages',
    ylab = 'runtime (s)'
)

Now we compare the writers. Only snpStats and genio have a BED writer. Not only was the genio writer much easier to use, it is also considerably faster.

times_write <- c(
    time_write_genio[3],
    time_write_snpStats[3]
)
names_write <- c(
    'genio',
    'snpStats'
)
# Create barplot
barplot(
    times_write,
    names.arg = names_write,
    main = 'BED writer runtimes',
    xlab = 'packages',
    ylab = 'runtime (s)'
)

Memory comparison

High memory consumption is the main sacrifice for the ease of use of native R matrices, and in this respect genio consumes the most memory.

We use the pryr package to compare the memory usage of each native genotype object.

library(lobstr)
# Store directly into a vector
sizes <- c(
    obj_size( X ),
    obj_size( data_genio$X ),
    obj_size( X_BEDMatrix ),
    obj_size( data_snpStats$genotypes )
)
names_sizes <- c(
    'original',
    'genio',
    'BEDMatrix',
    'snpStats'
)
# Create barplot
barplot(
    as.numeric( sizes ),
    names.arg = names_sizes,
    main = 'Native genotype object sizes',
    xlab = 'packages',
    ylab = 'memory (bytes)'
)

The BEDMatrix package is ideal for very large files, since data is loaded into memory only as needed. So the main object does not actually hold any genotypes in memory. It is up to the user to decide how much data to access at once to trade-off speed and memory consumption.

snpStats manages memory better than genio, but strictly by a factor of 4. Each genotype is encoded as an integer in R (in general) and genio (in particular), which uses 4 bytes (this is actually platform dependent). On the other hand, snpStats uses strictly one byte per genotype.

Overall, there is a narrow window in data sizes in which a genotype matrix is too large for a native R matrix but small enough for a snpStats object. That, and the much more complex interface of snpStats, makes it not very worthwhile in my opinion, unless there is need for functions provided only by that package. I prefer to use BEDMatrix for large datasets.