In LieberInstitute/megadepth: megadepth: BigWig and BAM related utilities

library(magrittr)

knitr::opts_chunk$set(echo = TRUE, message = FALSE, warning = TRUE)

Aim: Check the speed of different tools for loading coverage.

Background

There are multiple tools that are able to obtain coverage from BigWig/BAM files. Specifically, here we test the tools megadepth, rtracklayer and pyBigWig. Some operations requiring coverage can be very computationally intensive - for example, the dasper R package requires loading coverage across 3 regions per junction per sample. Here, we aim to find the fastest tool for loading coverage from BigWig files by comparing the speed of the megadepth vs rtracklayer vs pyBigwig.

Input data

First off, we obtain paths to the annotation ranges and BigWig files to load coverage from. For these, we have used the examples hosted on https://github.com/ChristopherWilks/megadepth.

Annotation co-ordinates

• For this we'll download use a set of 5000 ranges included in the dasper package. These are example junctions that

# extract only the required ranges 
annotation <- dasper::junctions_example %>% 
  SummarizedExperiment::rowRanges() 

# keep only ranges on chr 21
GenomeInfoDb::seqlevels(annotation, pruning.mode = "coarse") <- "21"

# add chr to ranges to match bw
GenomeInfoDb::seqlevels(annotation) <- paste0("chr", GenomeInfoDb::seqlevels(annotation))

# take the first 1000 ranges for testing
annotation <- annotation[1:5000]

annotation_path <- file.path(tempdir(), "annotation_example.txt")

rtracklayer::export(annotation, 
                    con = annotation_path, 
                    format = "BED")

BigWig files

• We will obtain the paths to the an example GTEx BigWig using recount2

url <- recount::download_study(project = "SRP012682", 
                               type = "samples", 
                               download = FALSE)

set.seed(32)

# select 1 random sample
bw_remote <- sample(url, 1)

bw_local <- file.path(tempdir(), "example_gtex_bw.bw")
download.file(bw_remote, bw_local)

Are coverage values consistent between methods?

• First, I will check whether the coverage values obtained through rtracklayer match those obtained through megadepth.

# get base-wise coverage as numeric list for summarisation after
rt_cov <- rtracklayer::import.bw(con = bw_local, 
                                 which = annotation, 
                                 as = "NumericList")

md_cov <- megadepth::get_coverage(bigwig_file = bw_local, 
                                  annotation = annotation_path, 
                                  op = "mean")

# round here to match 2 digits returned by megadepth
try(testthat::expect_equivalent(rt_cov %>% BiocGenerics::mean() %>% round(2), 
                                md_cov$cov))

Measure runtime for loading coverage

• Here, we will test the speed of rtracklayer, megadepth and pyBigwig when obtaining coverage from a BigWig file. To make sure the results are comparable, we will use the identical set of ranges and BigWigs initialised above.

rtracklayer & megadepth

• rtracklayer and megadepth can be easily called in R therefore for testing the runtime of these tools I will use the R package microbenchmark.

n_iters <- 10

benchmark_rt_md <- function(bw, times = 10){

  runtime <- microbenchmark::microbenchmark(
    "rt"= { 
      rtracklayer::import.bw(con = bw, 
                             which = annotation, 
                             as = "NumericList")
      }, 
    "md" = {
      megadepth::get_coverage(bigwig_file = bw, 
                              annotation = annotation_path, 
                              op = "mean")
    }, 
    times = times)

  runtime <- runtime %>% 
    dplyr::mutate(time = time / 1e9) %>% # convert time to secs
    dplyr::rename(tool = expr, run_time = time)

  return(runtime)

}

runtime_rt_md <- rbind(benchmark_rt_md(bw = bw_local, times = n_iters) %>% 
                         dplyr::mutate(local_remote = "local"), 
                       benchmark_rt_md(bw = bw_remote, times = n_iters) %>% 
                         dplyr::mutate(local_remote = "remote"))

pyBigWig {.tabset}

• As pyBigWig is a python package, to ensure the paths to the BigWig and annotation ranges remain the same, I will call the script found r here::here("analysis", "test_pyBigWig.py"). This creates a function that calls pyBigWig for a set number of iterations and saves the runtime for each.
• To run this, you must have python3.X available on your OS with the modules pyBigWig, pandas and numpy installed.

benchmark_pybw <- function(bw, times = 10){

  out_path <- paste0(tempdir(), "/pyBigWig_run_time.txt")

  system2("python3", 
          args = c(here::here("analysis", "test_pyBigWig.py"),
                   bw, 
                   annotation_path, 
                   out_path, 
                   times)
  )

  runtime <- readr::read_delim(out_path, delim = "\t")

  return(runtime)

}

runtime_pybw <- rbind(benchmark_pybw(bw = bw_local, times = n_iters) %>% 
                         dplyr::mutate(local_remote = "local"), 
                      benchmark_pybw(bw = bw_remote, times = n_iters) %>% 
                         dplyr::mutate(local_remote = "remote"))

Compare speeds

comp_time <- runtime_rt_md %>% 
  as.data.frame() %>% 
  rbind(runtime_pybw) %>% 
  dplyr::mutate(tool = tool %>% factor(c("md", "rt", "pyBigwig")))

comp_time_plot <- 
  ggplot2::ggplot(comp_time, ggplot2::aes(x = tool, 
                                          y = run_time, 
                                          colour = local_remote)) + 
  ggplot2::geom_boxplot() +  
  ggplot2::geom_jitter() + 
  ggplot2::scale_x_discrete(name = "Tool", 
                            labels = c("megadepth", 
                                       "rtracklayer", 
                                       "pyBigWig")) + 
  ggplot2::scale_y_continuous(name = paste0("Run time for ", length(annotation), " regions (seconds)")) + 
  ggplot2::scale_colour_manual(values = ggpubr::get_palette("npg", 2), 
                               guide = "none") + 
  ggplot2::facet_wrap(~ local_remote, 
                      scales = "free") + 
  ggplot2::theme_bw(base_size = 20) 

comp_time_plot %>% 
  ggplot2::ggsave(filename = "md_rt_pybw_runtime.png", 
                  plot = ., 
                  path = here::here("analysis"), 
                  dpi = 300, 
                  width = 10, height = 5)

Reproducibility

## Session info
library("sessioninfo")
options(width = 120)
session_info()

LieberInstitute/megadepth documentation built on Dec. 14, 2024, 3:30 p.m.