morinlab / dryclean

Irons out wrinkles in noisy coverage data using robust PCA

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Build Status codecov.io

dryclean

dryclean

Robust PCA based method to de-noise genomic coverage data.

Installations

Install devtools from CRAN

install.packages('devtools')

Set this to allow dependencies that throw warnings to be installed.

Sys.setenv(R_REMOTES_NO_ERRORS_FROM_WARNINGS = TRUE)

Install dependent packages and latest Bioconductor (if you haven't already)

source('https://bioconductor.org/biocLite.R')
biocLite('GenomicRanges')

Install mskilab R dependencies (gUtils)

devtools::install_github('mskilab/gUtils')

Install dryclean

devtools::install_github('mskilab/dryclean')

(after installing R package) Add dryclean directory to PATH and test the executable

$ export PATH=${PATH}:$(Rscript -e 'cat(paste0(installed.packages()["dryclean", "LibPath"], "/dryclean/extdata/"))')
$ drcln -h ## to see the help message

Tutorial

dryclean is a robust principal component analysis (rPCA) based method. dryclean uses a panel of normal (PON) samples to learn the landscape of both biological and technical noise in read depth data. dryclean then uses this landscape significantly reduce noise and artifacts in the signal for tumor samples. The input to the algorithm is a GenomicsRanges object containing read depth. You can either use read count from your favorite tools (there are many fast tools out there, for example: megadepth). Using uncorrected read count as input for dryclean works well from our experience, but if you wish, you can use the GC amd mappability corrected read depth data from fragCounter that can be found at: https://github.com/mskilab/fragCounter .

1. Creating Panel of Normal aka detergent

For creating PON the following factors are needed:

  1. Tumor and normal sample fragCounter outputs should be stored in two different directories
  2. A data.table with two columns: a. "sample" column contains the sample name you will use to index the sample b. "normal_cov" is a column with paths to the normal samples to be used

Following is an example of such a table

normal_table_example = readRDS("~/git/dryclean/inst/extdata/normal_table.rds")
normal_table_example
samplenormal_cov
samp1 ~/git/dryclean/inst/extdata/samp1.rds
samp2 ~/git/dryclean/inst/extdata/samp2.rds
samp3 ~/git/dryclean/inst/extdata/samp3.rds

There are three ways to make the PON:

  1. Using all normal samples availabble. PON can be made with all available normal sample. In this case set use.all = TRUE
detergent = prepare_detergent(normal.table.path = "~/git/dryclean/inst/extdata/normal_table.rds", path.to.save = "~/git/dryclean/inst/extdata/", num.cores = 1, use.all = TRUE)
  1. Using random subset of normal samples availabble.
detergent = prepare_detergent(normal.table.path = "~/git/dryclean/inst/extdata/normal_table.rds", path.to.save = "~/git/dryclean/inst/extdata/", num.cores = 1, use.all = FALSE, choose.randomly = TRUE)
  1. Cluster based approach. In order to keep the size of PON as small as possible but maximize the information in the PON. This is acheived by clustering the genomic background of normal samples and selecting normal samples from each cluster. Hierarchical clustering is used on L matrix after decomposing a small genomic region of all normal samples.
detergent = prepare_detergent(normal.table.path = "~/git/dryclean/inst/extdata/normal_table.rds", path.to.save = "~/git/dryclean/inst/extdata/", num.cores = 1, use.all = FALSE, choose.by.clustering = TRUE)
names(detergent)
head(detergent$L)
  1. 'L'
  2. 'S'
  3. 'k'
  4. 'U.hat'
  5. 'V.hat'
  6. 'sigma.hat'
0.11864910.12936250.1181923
0.34289640.37385830.3415761
0.11894280.12968280.1184849
0.12309010.13420450.1226161
0.10944830.11933090.1090269
0.12330530.13443920.1228305

The detergent is a list with the following elements:

  1. L: This is the L low ranked matrix of all the PONs calculated by batch robust PCA mathod
  2. S: This is the S sparse matrix of all the PONs calculated by batch robust PCA mathod
  3. k: This is estimated rank of a matrix where coverage values from each normal sample forms a column
  4. U.hat: svd decompsed left sigular matrix of L required for online implentation of rPCA
  5. V.hat: svd decompsed right sigular matrix of L required for online implentation of rPCA
  6. sigma.hat: svd decompsed first k sigular values of L required for online implentation of rPCA

2. Identifying germline events

Since a PON is used for decomposing the tumor samples, a method is required identify and remove germline events. This is achieved by looking at all normal samples as a population and infer the markers that have a copynumber events at a given frequency, set by user.

In order to do so, normal samples are treated as tumors and all copy number changes in the normal samples are extracted using thr PON created. Here is an example:

decomp.1 = start_wash_cycle(cov = sample.1, detergent.pon.path = "~/git/dryclean/inst/extdata/", whole_genome = TRUE, chr = NA, germline.filter = FALSE)

Once all normal samples are decomposed, the data.table is updated to reflect that:

normal_table_example = readRDS("~/git/dryclean/inst/extdata/normal_table.rds")
normal_table_example
samplenormal_covdecomposed_cov
samp1 ~/git/dryclean/inst/extdata/samp1.rds ~/git/dryclean/inst/extdata/decomp1.rd
samp2 ~/git/dryclean/inst/extdata/samp2.rds ~/git/dryclean/inst/extdata/decomp2.rd
samp3 ~/git/dryclean/inst/extdata/samp3.rds ~/git/dryclean/inst/extdata/decomp3.rd

With this table we can run the following:

grm = identify_germline(normal.table.path = "~/git/dryclean/inst/extdata/normal_table.rds", path.to.save = "~/git/dryclean/inst/extdata/", signal.thresh=0.5, pct.thresh=0.98)

Now we are ready for tumor decomposition

3. Running dryclean on tumor sample within R

Following is a dummy example. The data diretory has a dummy coverage gRanges object which requires "reads.corrected" field

coverage_file = readRDS("~/git/dryclean/data/dummy_coverage.rds")
coverage_file
GRanges object with 50 ranges and 1 metadata column:
       seqnames    ranges strand | reads.corrected
          <Rle> <IRanges>  <Rle> |           <numeric>
   [1]       22       1-3      * |    1.64885252481326
   [2]       22       3-5      * |    3.81186937098391
   [3]       22       5-7      * |    2.58672125521116
   [4]       22       7-9      * |   0.606155182467774
   [5]       22      9-11      * |    4.83087254804559
   ...      ...       ...    ... .                 ...
  [46]       22     91-93      * |    3.85029493598267
  [47]       22     93-95      * |    2.96715440694243
  [48]       22     95-97      * |    3.56770512764342
  [49]       22     97-99      * |    4.38074178760871
  [50]       22    99-101      * |     0.1496995636262
  -------
  seqinfo: 1 sequence from an unspecified genome; no seqlengths

In order to run dryclean, simply invoke the following function

cov_out = start_wash_cycle(cov = coverage_file, detergent.pon.path = "~/git/dryclean/inst/extdata", whole_genome = TRUE, chr = NA, germline.filter = TRUE, germline.file = "~/git/dryclean/inst/extdata/germline.markers.rds")
head(cov_out)
GRanges object with 6 ranges and 12 metadata columns:
      seqnames  ranges  strand |      background.log      foreground.log
         <Rle> <IRanges> <Rle> |      <numeric>             <numeric>
  [1]        22 1-3          * |     0.169769767795501 -0.0568088592719299
  [2]        22 3-5          * |    0.0177158854407206  0.0770640446340152
  [3]        22 5-7          * |    0.0664955328192282  0.0347592933585855
  [4]        22 7-9          * |     0.230039681056936  -0.100959892306004
  [5]        22 9-11         * | -0.000366098004722278 -0.0812780098385511
  [6]        22 11-13        * |     0.135769072540305  0.0878847580129722
          reads                gc       map   input.read.counts
      <numeric>         <numeric> <numeric>          <numeric>
  [1]       762 0.612756264236902         1   2.78713582604055
  [2]       324 0.587301587301587         1    1.1277123686483
  [3]       474 0.494505494505495         1   1.60676731785653
  [4]      1131 0.568181818181818         1   3.81814927898723
  [5]       254 0.592592592592593         1  0.890949237364552
  [6]       314 0.697819314641745         1   2.43451610256352
             median.chr blacklisted      foreground        background
              <numeric>   <logical>         <numeric>         <numeric>
  [1] 0.608371537098667       FALSE 0.944774637021039  1.18503198738368
  [2] 0.608371537098667       FALSE  1.08011124950716   1.0178737425542
  [3] 0.608371537098667       FALSE  1.03537045825707  1.06875618973741
  [4] 0.608371537098667       FALSE 0.903969288282395  1.25864995349903
  [5] 0.608371537098667       FALSE 0.921937348159965 0.999633969000975
  [6] 0.608371537098667       FALSE  1.09186228639812  1.14541735473137
               log.reads germline.blk
               <numeric>    <logical>
  [1]   1.02501448288477        FALSE
  [2]  0.120191128226724        FALSE
  [3]  0.474224283401017        FALSE
  [4]   1.33976582327867        FALSE
  [5] -0.115467825789155        FALSE
  [6]  0.889748010939351        FALSE
  -------
  seqinfo: 24 sequences from an unspecified genome

The output has following metadata fields:

  1. background.log: This is the L low ranked vector after decomposition and represent the background noise separated by dryclean in the log space
  2. foreground.log: The S vector with the inferred copy number signal separated by dryclean, that forms foreground, in the log space
  3. reads: Raw read counts
  4. gc: GC score
  5. map: mappability score
  6. input.read.counts: This is the fragCounter input in linear space
  7. median.chr: median chromosome signal
  8. blacklisted: if off target marker list is available and used
  9. foreground: Foreground signal, that forms SCNAs (S vector) in read count/ratio space
  10. background: This is the L low ranked vector after decomposition and represent the background noise separated by dryclean in read count/ratio space
  11. log.reads: log of the fragCounter signal
  12. germline.blk: germline marker based on the inferred germline function

3. Running dryclean on tumor sample from command line

./drcln -i inst/extdata/samp1.rds -p inst/extdata/detergent.50.rds 
Rprofile Loading
Rprofile Finished Loading


▓█████▄   ██▀███  ██   ██▓ ▄████▄   ██▓    ▓█████ ▄▄▄       ███▄    █
 ██▀ ██▌ ▓██   ██  ██  ██  ██▀ ▀█  ▓██▒    ▓█   ▀ ████▄     ██ ▀█   █
░██   █▌ ▓██ ░▄█    ██ ██  ▓█    ▄  ██░    ░███   ██  ▀█▄   ██  ▀█ ██▒
░▓█▄   ▌ ▒██▀▀█▄   ░ ▐██▓ ▒▓▓▄ ▄██▒ ██░    ░▓█  ▄ ██▄▄▄▄█   ██▒  ▐▌██▒
░▒████▓  ░██▓  ██  ░ ██▒    ▓███▀ ░░█████ ▒█████▒ █     █▒ ██░   ▓██░
 ▒ ▓  ▒  ░  ▓ ░▒▓░  ██    ░ ░▒ ▒  ░░ ▒░▓  ░░░ ▒░ ░▒▒   ▓▒█░░ ▒░   ▒ ▒
 ░ ▒  ▒    ░▒ ░  ░  ░░▒░   ░  ▒   ░ ░ ▒  ░ ░ ░  ░ ▒   ▒▒ ░░ ░░   ░ ▒░
 ░ ░  ░    ░░   ░   ░  ░░  ░          ░ ░  ░    ░    ░   ▒      ░   ░ ░
   ░        ░     ░ ░     ░ ░          ░  ░   ░  ░     ░  ░     ░   ░
 ░               ░ ░     ░       ░    ░     ░     ░      ░     ░ 


(Let's dryclean the genomes!)

Loading PON a.k.a detergent from path provided
Let's begin, this is whole exome/genome
Initializing
Using the detergent provided to start washing
lambdas calculated
Here begins the wash cycle
calculating A and B
calculating v and s
Updating subspace
Combining matrices with gRanges
Giddy Up!

All the options and usage is as follows

./drcln -h
Rprofile Loading
Rprofile Finished Loading
Usage: ./drcln [options]


Options:
	-i INPUT, --input=INPUT
		Path to cov.rds file, fragCounter output for sample under consideration

	-p PON, --pon=PON
		Path to Panel Of Normal (PON) on which batch rPCA have been run. Within the file should be L, S matrices, estimated rank for burnin samples and svd decomposition matrices for the same

	-m NAME, --name=NAME
		Sample / Individual name

	-b BLACKLIST, --blacklist=BLACKLIST
		blacklisted makers

	-w WHOLEGENOME, --wholeGenome=WHOLEGENOME
		If TRUE then it will process all chromosomes and parallelize it

	-C CHROMOSOME, --chromosome=CHROMOSOME
		If wholeGenome is FALSE, specify the chromosome to process

	-g GERMLINE.FILTER, --germline.filter=GERMLINE.FILTER
		If PON based germline filter is to be used for removing some common germline events, If set to TRUE, give path to germline annotated file

	-f GERMLINE.FILE, --germline.file=GERMLINE.FILE
		Path to file annotated with germline calls, if germline.filter == TRUE

	-c CORES, --cores=CORES
		How many cores to use

	-o OUTDIR, --outdir=OUTDIR
		output directory

	-k COLLAPSE, --collapse=COLLAPSE
		collapse 200bp fragCounter to 1kb

	-h, --help
		Show this help message and exit

Panel of Normal for 1kb WGS (hg19)

The Panel of Normal samples (PON) of 395 TCGA WGS normal samples was created using hierarchical clustering approach described above and filtered for CNPs.

The file is 17G in size.

WGS 1 kb PON: https://mskilab.s3.amazonaws.com/hg19/WGS/detergent.rds

About

Irons out wrinkles in noisy coverage data using robust PCA


Languages

Language:R 100.0%