Summarization and quantitative trait analysis of CNV ranges

In CNVRanger: Summarization and expression/phenotype association of CNV ranges

suppressPackageStartupMessages({ 
    library(CNVRanger)
    library(AnnotationHub)
    library(regioneR)
    library(BSgenome.Btaurus.UCSC.bosTau6.masked)
    library(SummarizedExperiment)
    library(curatedTCGAData)
    library(TCGAutils)
    library(Gviz)
})

Introduction

Copy number variation (CNV) is a frequently observed deviation from the diploid state due to duplication or deletion of genomic regions. CNVs can be experimentally detected based on comparative genomic hybridization, and computationally inferred from SNP-arrays or next-generation sequencing data. These technologies for CNV detection have in common that they report, for each sample under study, genomic regions that are duplicated or deleted with respect to a reference. Such regions are denoted as CNV calls in the following and will be considered the starting point for analysis with the r Biocpkg("CNVRanger") package.

The following figure provides an overview of the analysis capabilities of CNVRanger.

vign.dir <- system.file("vignettes", package = "CNVRanger")
knitr::include_graphics(file.path(vign.dir, "CNVRanger.png"))

(A) The r Biocpkg("CNVRanger") package imports CNV calls from a simple file format into R, and stores them in dedicated Bioconductor data structures, and (B) implements three frequently used approaches for summarizing CNV calls across a population: (i) the CNVRuler procedure that trims region margins based on regional density Kim et al., 2012, (ii) the reciprocal overlap procedure that requires sufficient mutual overlap between calls Conrad et al., 2010, and (iii) the GISTIC procedure that identifies recurrent CNV regions Beroukhim et al., 2007. (C) CNVRanger builds on r Biocpkg("regioneR") for overlap analysis of CNVs with functional genomic regions, (D) implements RNA-seq expression Quantitative Trait Loci (eQTL) analysis for CNVs by interfacing with r Biocpkg("edgeR"), and (E) interfaces with PLINK for traditional genome-wide association studies (GWAS) between CNVs and quantitative phenotypes.

The key parts of the functionality implemented in r Biocpkg("CNVRanger") were developed, described, and applied in several previous studies:

• Genome-wide detection of CNVs and their association with meat tenderness in Nelore cattle da Silva et al., 2016

• Widespread modulation of gene expression by copy number variation in skeletal muscle Geistlinger et al., 2018

• CNVs are associated with genomic architecture in a songbird da Silva et al., 2018

Applicability and Scope

As described in the above publications, r Biocpkg("CNVRanger") has been developed and extensively tested for SNP-based CNV calls as obtained with PennCNV. We also tested CNVRanger for sequencing-based CNV calls as obtained with CNVnator (a read-depth approach) or LUMPY (an approach that combines evidence from split-reads and discordant read-pairs).

In general, r Biocpkg("CNVRanger") can be applied to CNV calls associated with integer copy number states, where we assume the states to be encoded as:

• 0: homozygous deletion (2-copy loss)
• 1: heterozygous deletion (1-copy loss)
• 2: normal diploid state
• 3: 1-copy gain
• 4: amplification (>= 2-copy gain)

Note that for CNV calling software that uses a different encoding or that does not provide integer copy number states, it is often possible to (at least approximately) transform the output to a format that is compatible with the input format of r Biocpkg("CNVRanger"). See Section 4.1 Input data format for details.

r Biocpkg("CNVRanger") is designed to work with CNV calls from one tool at a time. See EnsembleCNV and FusorSV for combining CNV calls from multiple SNP-based callers or multiple sequencing-based callers, respectively.

r Biocpkg("CNVRanger") assumes CNV calls provided as input to be already filtered by quality, using the software that was used for CNV calling, or specific tools for that purpose. r Biocpkg("CNVRanger") provides downstream summarization and association analysis for CNV calls, it does not implement functions for CNV calling or quality control. r Biocpkg("CNVRanger") is applicable for diploid species only.

Key functions

Analysis step | Function
----------------------- | --------------------------------------------- (A) Input | GenomicRanges::makeGRangesListFromDataFrame (B) Summarization | populationRanges
(C) Overlap analysis | regioneR::overlapPermTest (D) Expression analysis | cnvEQTL
(E) Phenotype analysis | cnvGWAS

Note: we use the package::function notation for functions from other packages. For functions from this package and base R functions, we use the function name without preceding package name.

Reading and accessing CNV data

vign.dir <- system.file("vignettes", package = "CNVRanger")
knitr::include_graphics(file.path(vign.dir, "Input.png"))

The r Biocpkg("CNVRanger") package uses Bioconductor core data structures implemented in the r Biocpkg("GenomicRanges") and r Biocpkg("RaggedExperiment") packages to efficiently represent, access, and manipulate CNV data.

We start by loading the package.

library(CNVRanger)

Input data format

r Biocpkg("CNVRanger") reads CNV calls from a simple file format, providing at least chromosome, start position, end position, sample ID, and integer copy number for each call.

For demonstration, we consider CNV calls as obtained with PennCNV from SNP-chip data in a Brazilian cattle breed (da Silva et al., 2016).

Here, we use a data subset and only consider CNV calls on chromosome 1 and 2, for which there are roughly 3000 CNV calls as obtained for 711 samples. We use read.csv to read comma-separated values, but we could use a different function if the data were provided with a different delimiter (for example, read.delim for tab-separated values).

data.dir <- system.file("extdata", package="CNVRanger")
call.file <- file.path(data.dir, "Silva16_PONE_CNV_calls.csv")
calls <- read.csv(call.file, as.is=TRUE)
nrow(calls)
head(calls)

length(unique(calls[,"NE_id"]))

We observe that this example dataset satisfies the basic five-column input format required by r Biocpkg("CNVRanger").

The last column contains the integer copy number state for each call, encoded as

• 0: homozygous deletion (2-copy loss)
• 1: heterozygous deletion (1-copy loss)
• 2: normal diploid state
• 3: 1-copy gain
• 4: amplification (>= 2-copy gain)

For CNV detection software that uses a different encoding, it is necessary to convert to the above encoding. For example, GISTIC uses

• -2: homozygous deletion (2-copy loss)
• -1: heterozygous deletion (1-copy loss)
• 0: normal diploid state
• 1: 1-copy gain
• 2: amplification (>= 2-copy gain)

which can be converted by simply adding 2.

In Section 7.2 Application to TCGA data we also describe how to transform segmented log2 copy ratios to integer copy number states.

The basic five-column input format can be augmented with additional columns, providing additional characteristics and metadata for each CNV call. In Section 8 CNV-phenotype association analysis, we demonstrate how to make use of such an extended input format.

Representation as a GRangesList

Once read into an R data.frame, we group the calls by sample ID and convert them to a GRangesList. Each element of the list corresponds to a sample, and contains the genomic coordinates of the CNV calls for this sample (along with the copy number state in the state metadata column).

grl <- GenomicRanges::makeGRangesListFromDataFrame(calls, 
    split.field="NE_id", keep.extra.columns=TRUE)
grl

The advantage of representing the CNV calls as a GRangesList is that it allows to leverage the comprehensive set of operations on genomic regions implemented in the r Biocpkg("GenomicRanges") packages - for instance, sorting of the calls according to their genomic coordinates.

grl <- GenomicRanges::sort(grl)
grl

Representation as a RaggedExperiment

An alternative matrix-like representation of the CNV calls can be obtained with the r Biocpkg("RaggedExperiment") data class. It resembles in many aspects the r Biocpkg("SummarizedExperiment") data class for storing gene expression data as e.g. obtained with RNA-seq.

ra <- RaggedExperiment::RaggedExperiment(grl)
ra

As apparent from the dim slot of the object, it stores the CNV calls in the rows and the samples in the columns. Note that the CN state is now represented as an assay matrix which can be easily accessed and subsetted.

RaggedExperiment::assay(ra[1:5,1:5])

As with r Biocpkg("SummarizedExperiment") objects, additional information for the samples are annotated in the colData slot. For example, we annotate the steer weight and its feed conversion ratio (FCR) using simulated data. Feed conversion ratio is the ratio of dry matter intake to live-weight gain. A typical range of feed conversion ratios is 4.5-7.5 with a lower number being more desirable as it would indicate that a steer required less feed per pound of gain.

weight <- rnorm(ncol(ra), mean=1100, sd=100)
fcr <- rnorm(ncol(ra), mean=6, sd=1.5)
RaggedExperiment::colData(ra)$weight <- round(weight, digits=2)
RaggedExperiment::colData(ra)$fcr <- round(fcr, digits=2)
RaggedExperiment::colData(ra)

Summarizing individual CNV calls across a population

vign.dir <- system.file("vignettes", package = "CNVRanger")
knitr::include_graphics(file.path(vign.dir, "Summarization.png"))

In CNV analysis, it is often of interest to summarize individual calls across the population, (i.e. to define CNV regions), for subsequent association analysis with expression and phenotype data. In the simplest case, this just merges overlapping individual calls into summarized regions. However, this typically inflates CNV region size, and more appropriate approaches have been developed for this purpose.

As mentioned in the Introduction, we emphasize the need for quality control of CNV calls and appropriate accounting for sources of technical bias before applying these summarization functions (or in general downstream analysis with CNVRanger).

For instance, protocols for read-depth CNV calling typically exclude calls overlapping defined repetitive and low-complexity regions including the UCSC list of segmental duplications Trost et al., 2018, Zhou et al., 2018. We also note that CNVnator, a very popular read-depth CNV caller, implements the $q0$-filter to explicitely flag and, if desired, exclude calls that are likely to stem from such regions.
If systematically over-represented in the input CNV calls, summarization procedures such as GISTIC will identify these regions as recurrent independent of whether there are biological or technical reasons for that.

Trimming low-density areas

Here, we use the approach from CNVRuler to summarize CNV calls to CNV regions (see Figure 1 in Kim et al., 2012 for an illustration of the approach). This trims low-density areas as defined by the density argument, which is set here to <10\% of the number of calls within a summarized region.

cnvrs <- populationRanges(grl, density=0.1)
cnvrs

Note that CNV frequency (number of samples overlapping each region) and CNV type (gain, loss, or both) have also been annotated in the columns freq and type, respectively.

Reciprocal overlap

We also provide an implementation of the Reciprocal Overlap (RO) procedure that requires calls to sufficiently overlap with one another as e.g. used by Conrad et al., 2010. This merges calls with an RO above a threshold as given by the ro.thresh argument. For example, an RO of 0.51 between two genomic regions A and B requires that B overlaps at least 51\% of A, and that A also overlaps at least 51\% of B.

ro.cnvrs <- populationRanges(grl[1:100], mode="RO", ro.thresh=0.51)
ro.cnvrs

Identifying recurrent regions

In particular in cancer, it is important to distinguish driver from passenger mutations, i.e. to distinguish meaningful events from random background aberrations. The GISTIC method identifies those regions of the genome that are aberrant more often than would be expected by chance, with greater weight given to high amplitude events (high-level copy-number gains or homozygous deletions) that are less likely to represent random aberrations (Beroukhim et al., 2007).

By setting est.recur=TRUE, we deploy a GISTIC-like significance estimation

cnvrs <- populationRanges(grl, density=0.1, est.recur=TRUE)
cnvrs

and filter for recurrent CNVs that exceed a significance threshold of 0.05.

subset(cnvrs, pvalue < 0.05)

We can illustrate the landscape of recurrent CNV regions using the function plotRecurrentRegions. We therefore provide the summarized CNV regions, a valid UCSC genome assembly, and a chromosome of interest.

plotRecurrentRegions(cnvrs, genome="bosTau6", chr="chr1")

The function plots (from top to bottom): (i) an ideogram of the chromosome (note that staining bands are not available for bosTau6), (ii) a genome axis indicating the chromosomal position, (iii) a bar plot showing for each CNV region the number of samples with a CNV call in that region, and (iv) an annotation track that indicates whether this is a recurrent region according to a significance threshold (an argument to the function, default: 0.05).

Overlap analysis of CNVs with functional genomic regions

vign.dir <- system.file("vignettes", package = "CNVRanger")
knitr::include_graphics(file.path(vign.dir, "Overlap.png"))

Once individual CNV calls have been summarized across the population, it is typically of interest whether the resulting CNV regions overlap with functional genomic regions such as genes, promoters, or enhancers.

To obtain the location of protein-coding genes, we query available Bos taurus annotation from Ensembl

library(AnnotationHub)
ah <- AnnotationHub::AnnotationHub()
ahDb <- AnnotationHub::query(ah, pattern = c("Bos taurus", "EnsDb"))
ahDb

and retrieve gene coordinates in the UMD3.1 assembly (Ensembl 92).

ahEdb <- ahDb[["AH60948"]]
bt.genes <- ensembldb::genes(ahEdb)
GenomeInfoDb::seqlevelsStyle(bt.genes) <- "UCSC"
bt.genes

To speed up the example, we restrict analysis to chromosomes 1 and 2.

sel.genes <- subset(bt.genes, seqnames %in% paste0("chr", 1:2))
sel.genes <- subset(sel.genes, gene_biotype == "protein_coding")
sel.cnvrs <- subset(cnvrs, seqnames %in% paste0("chr", 1:2))

Finding and illustrating overlaps

The findOverlaps function from the r Biocpkg("GenomicRanges") package is a general function for finding overlaps between two sets of genomic regions. Here, we use the function to find protein-coding genes (our query region set) overlapping the summarized CNV regions (our subject region set).

Resulting overlaps are represented as a Hits object, from which overlapping query and subject regions can be obtained with dedicated accessor functions (named queryHits and subjectHits, respectively). Here, we use these functions to also annotate the CNV type (gain/loss) for genes overlapping with CNVs.

olaps <- GenomicRanges::findOverlaps(sel.genes, sel.cnvrs, ignore.strand=TRUE)
qh <- S4Vectors::queryHits(olaps)
sh <- S4Vectors::subjectHits(olaps)
cgenes <- sel.genes[qh]
cgenes$type <- sel.cnvrs$type[sh]
subset(cgenes, select = "type")

It might also be of interest to illustrate the original CNV calls on overlapping genomic features (here: protein-coding genes). For this purpose, an oncoPrint plot provides a useful summary in a rectangular fashion (genes in the rows, samples in the columns). Stacked barplots on the top and the right of the plot display the number of altered genes per sample and the number of altered samples per gene, respectively.

cnvOncoPrint(grl, cgenes)

Overlap permutation test

As a certain amount of overlap can be expected just by chance, an assessment of statistical significance is needed to decide whether the observed overlap is greater (enrichment) or less (depletion) than expected by chance.

The r Biocpkg("regioneR") package implements a general framework for testing overlaps of genomic regions based on permutation sampling. This allows to repeatedly sample random regions from the genome, matching size and chromosomal distribution of the region set under study (here: the CNV regions). By recomputing the overlap with the functional features in each permutation, statistical significance of the observed overlap can be assessed.

We demonstrate in the following how this strategy can be used to assess the overlap between the detected CNV regions and protein-coding regions in the cattle genome. We expect to find a depletion as protein-coding regions are highly conserved and rarely subject to long-range structural variation such as CNV. Hence, is the overlap between CNVs and protein-coding genes less than expected by chance?

To answer this question, we apply an overlap permutation test with 100 permutations (ntimes=100), while maintaining chromosomal distribution of the CNV region set (per.chromosome=TRUE). Furthermore, we use the option count.once=TRUE to count an overlapping CNV region only once, even if it overlaps with 2 or more genes. We also allow random regions to be sampled from the entire genome (mask=NA), although in certain scenarios masking certain regions such as telomeres and centromeres is advisable. Also note that we use 100 permutations for demonstration only. To draw robust conclusions a minimum of 1000 permutations should be carried out.

library(regioneR)
library(BSgenome.Btaurus.UCSC.bosTau6.masked)
res <- regioneR::overlapPermTest(A=sel.cnvrs, B=sel.genes, ntimes=100, 
    genome="bosTau6", mask=NA, per.chromosome=TRUE, count.once=TRUE)
res

summary(res[[1]]$permuted)

The resulting permutation p-value indicates a significant depletion. Out of the r length(sel.cnvrs) CNV regions, r res[[1]]$observed overlap with at least one gene. In contrast, when repeatedly drawing random regions matching the CNV regions in size and chromosomal distribution, the mean number of overlapping regions across permutations was r round(mean(res[[1]]$permuted), digits=1) $\pm$ r round(sd(res[[1]]$permuted), digits=1).

This finding is consistent with our observations across the whole genome (da Silva et al., 2016) and findings from the 1000 Genomes Poject (Sudmant et al., 2015).

plot(res)

Note: the function regioneR::permTest allows to incorporate user-defined functions for randomizing regions and evaluating additional measures of overlap such as total genomic size in bp.

CNV-expression association analysis

vign.dir <- system.file("vignettes", package = "CNVRanger")
knitr::include_graphics(file.path(vign.dir, "Expression.png"))

Studies of expression quantitative trait loci (eQTLs) aim at the discovery of genetic variants that explain variation in gene expression levels (Nica and Dermitzakis, 2013). Mainly applied in the context of SNPs, the concept also naturally extends to the analysis of CNVs.

The r Biocpkg("CNVRanger") package implements association testing between CNV regions and RNA-seq read counts using r Biocpkg("edgeR"), which applies generalized linear models based on the negative-binomial distribution while incorporating normalization factors for different library sizes.

In the case of only one CN state deviating from 2n for a CNV region under investigation, this reduces to the classical 2-group comparison. For more than two states (e.g. 0n, 1n, 2n), edgeR’s ANOVA-like test is applied to test all deviating groups for significant expression differences relative to 2n.

Application to individual CNV and RNA-seq assays

We demonstrate the functionality by loading RNA-seq read count data from skeletal muscle samples for 183 Nelore cattle steers, which we analyzed for CNV-expression effects as previously described (Geistlinger et al., 2018).

rseq.file <- file.path(data.dir, "counts_cleaned.txt")
rcounts <- read.delim(rseq.file, row.names=1, stringsAsFactors=FALSE)
rcounts <- as.matrix(rcounts)
dim(rcounts)
rcounts[1:5, 1:5]

For demonstration, we restrict analysis to the 939 genes on chromosome 1 and 2, and store the RNA-seq expression data in a r Biocpkg("SummarizedExperiment").

library(SummarizedExperiment)
rranges <- GenomicRanges::granges(sel.genes)[rownames(rcounts)]
rse <- SummarizedExperiment(assays=list(rcounts=rcounts), rowRanges=rranges)
rse

Assuming distinct modes of action, effects observed in the CNV-expression analysis are typically divided into (i) local effects (cis), where expression changes coincide with CNVs in the respective genes, and (ii) distal effects (trans), where CNVs supposedly affect trans-acting regulators such as transcription factors.

However, due to power considerations and to avoid detection of spurious effects, stringent filtering of (i) not sufficiently expressed genes, and (ii) CNV regions with insufficient sample size in groups deviating from 2n, should be carried out when testing for distal effects. Local effects have a clear spatial indication and the number of genes locating in or close to a CNV region of interest is typically small; testing for differential expression between CN states is thus generally better powered for local effects and less stringent filter criteria can be applied.

In the following, we carry out CNV-expression association analysis by providing the CNV regions to test (cnvrs), the individual CNV calls (grl) to determine per-sample CN state in each CNV region, the RNA-seq read counts (rse), and the size of the genomic window around each CNV region (window). The window argument thereby determines which genes are considered for testing for each CNV region and is set here to 1 Mbp.

Further, use the filter.by.expr and min.samples arguments to exclude from the analysis (i) genes with very low read count across samples, and (ii) CNV regions with fewer than min.samples samples in a group deviating from 2n.

res <- cnvEQTL(cnvrs, grl, rse, window = "1Mbp", verbose = TRUE)
res

The resulting GRangesList contains an entry for each CNV region tested, storing the genes tested in the genomic window around the CNV region, and (i) log2 fold change with respect to the 2n group, (ii) edgeR's DE p-value, and (iii) the (per default) Benjamini-Hochberg adjusted p-value.

Application to TCGA data stored in a MultiAssayExperiment

In the previous section, we individually prepared the CNV and RNA-seq data for CNV-expression association analysis. In the following, we demonstrate how to perform an integrated preparation of the two assays when stored in a r Biocpkg("MultiAssayExperiment"). We therefore consider glioblastoma GBM data from The Cancer Genome Atlas TCGA, which can conveniently be accessed with the r Biocpkg("curatedTCGAData") package.

library(curatedTCGAData)
gbm <- curatedTCGAData::curatedTCGAData("GBM", 
        assays=c("GISTIC_Peaks", "CNVSNP", "RNASeq2GeneNorm"), 
        dry.run=FALSE)
gbm

The returned MultiAssayExperiment contains three assays:

• the SNP-based CNV calls stored in a RaggedExperiment (GBM_CNVSNP),
• the recurrent CNV regions summarized across the population using the GISTIC method (GBM_GISTIC_Peaks), and
• the normalized RNA-seq gene expression values in a SummarizedExperiment (GBM_RNASeq2GeneNorm).

To annotate the genomic coordinates of the genes measured in the RNA-seq assay, we use the function symbolsToRanges from the r Biocpkg("TCGAutils") package. For demonstration, we restrict the analysis to chromosome 1 and 2.

library(TCGAutils)
gbm <- TCGAutils::symbolsToRanges(gbm, unmapped=FALSE)
for(i in 1:3) 
{
    rr <- rowRanges(gbm[[i]])
    GenomeInfoDb::genome(rr) <- "NCBI37"
    GenomeInfoDb::seqlevelsStyle(rr) <- "UCSC"
    ind <- as.character(seqnames(rr)) %in% c("chr1", "chr2")
    rowRanges(gbm[[i]]) <- rr
    gbm[[i]] <- gbm[[i]][ind,]
}
gbm

We now restrict the analysis to intersecting patients of the three assays using MultiAssayExperiment's intersectColumns function, and select Primary Solid Tumor samples using the splitAssays function from the r Biocpkg("TCGAutils") package.

gbm <- MultiAssayExperiment::intersectColumns(gbm)
TCGAutils::sampleTables(gbm)
data(sampleTypes, package="TCGAutils")
sampleTypes
gbm <- TCGAutils::splitAssays(gbm, sampleCodes="01")
gbm

The SNP-based CNV calls from TCGA are provided as segmented log2 copy number ratios.

ind <- grep("CNVSNP", names(gbm))
head( mcols(gbm[[ind]]) )
summary( mcols(gbm[[ind]])$Segment_Mean )

It is thus necessary to convert them to integer copy number states for further analysis with r Biocpkg("CNVRanger").

In a diploid genome, a single-copy gain in a perfectly pure, homogeneous sample has a copy ratio of 3/2. On log2 scale, this is log2(3/2) = 0.585, and a single-copy loss is log2(1/2) = -1.0.

We can roughly convert a log ratio lr to an integer copy number by

round( (2^lr) * 2)

Note that this is not the ideal way to calculate absolute integer copy numbers. Especially in cancer, differences in tumor purity, tumor ploidy, and subclonality can substantially interfere with the assumption of a pure homogeneous sample. See ABSOLUTE and the r Biocpkg("PureCN") package for accurately taking such tumor characteristics into account.

However, without additional information we transform the log ratios into integer copy number states using the rough approximation outlined above.

smean <- mcols(gbm[[ind]])$Segment_Mean
state <- round(2^smean * 2)
state[state > 4] <- 4
mcols(gbm[[ind]])$state <- state
gbm[[ind]] <- gbm[[ind]][state != 2,]
mcols(gbm[[ind]]) <- mcols(gbm[[ind]])[,3:1]
table(mcols(gbm[[ind]])$state)

The data is now ready for CNV-expression association analysis, where we find only four CNV regions with sufficient sample size for testing using the default value of 10 for the minSamples argument.

res <- cnvEQTL(cnvrs="01_GBM_GISTIC_Peaks-20160128", 
    calls="01_GBM_CNVSNP-20160128", 
    rcounts="01_GBM_RNASeq2GeneNorm-20160128_ranged", 
    data=gbm, window="1Mbp", verbose=TRUE)
res

We can illustrate differential expression of genes in the neighborhood of a CNV region of interest using the function plotEQTL.

res[2]
(r <- GRanges(names(res)[2]))
plotEQTL(cnvr=r, genes=res[[2]], genome="hg19", cn="CN1")

The plot shows consistent decreased expression (negative log2 fold change) of genes in the neighborhood of the CNV region, when comparing samples with a one copy loss (1$n$) in that region to the 2$n$ reference group.

Note that a significant expression change is not only observed for genes locating within the CNV region (dosage effect, here: PARK7), but also genes locating in close proximity of the CNV region (neighborhood effect, here: CAMTA1 and RERE). This is consistent with previous observations in mouse Cahan et al., 2009 and our observations in cattle Geistlinger et al., 2018.

CNV-phenotype association analysis

vign.dir <- system.file("vignettes", package = "CNVRanger")
knitr::include_graphics(file.path(vign.dir, "Phenotype.png"))

Specifically developed for CNV calls inferred from SNP-chip data, r Biocpkg("CNVRanger") allows to carry out a probe-level genome-wide association study (GWAS) with quantitative phenotypes. CNV calls from other sources such as sequencing data are also supported by using the start and end position of each call as the corresponding probes.

As previously described da Silva et al., 2016, we construct CNV segments from probes representing common CN polymorphisms (allele frequency >1\%), and carry out a GWAS as implemented in PLINK using a standard linear regression of phenotype on allele dosage.

For CNV segments composed of multiple probes, the segment p-value is chosen from the probe p-values, using either the probe with minimum p-value or the probe with maximum CNV frequency.

For demonstration, we use CNV data of a wild population of songbirds (da Silva et al., 2018).

cnv.loc <- file.path(data.dir, "CNVOut.txt") 
cnv.calls <- read.delim(cnv.loc, as.is=TRUE) 
head(cnv.calls)

Here, we use the extensibility of the basic five-column input format described in Section 4.1. In addition to the required five columns (providing chromosome, start position, end position, sample ID, and integer copy number state), we provided three optional columns storing the number of probes supporting the call, and the Affymetrix ID of the first and last probe contained in the call.

As these columns are optional, it is not ultimately necessary to provide them in order to run a CNV GWAS. However, we recommend to provide this information when available as it allows for a more fine-grained probe-by-probe GWAS.

As described in Section 4.2, we store the CNV calls in a GRangesList for further analysis.

cnv.calls <- GenomicRanges::makeGRangesListFromDataFrame(cnv.calls, 
    split.field="sample.id", keep.extra.columns=TRUE)
sort(cnv.calls)

In the following, we use genomic estimated breeding values (GEBVs) for breeding time (BT) as the quantitative phenotype, and accordingly analyze for each CNV region whether change in copy number is associated with change in the genetic potential for breeding time.

Setting up a CNV-GWAS

For compatibility with PLINK's fam file format, we read phenotype information from a tab-delimited file, containing exactly four columns: sample ID, family ID, sex, and the quantitative phenotype (here: breeding time, BT).

phen.loc <- file.path(data.dir, "Pheno.txt")
phen.df <- read.delim(phen.loc, as.is=TRUE)
head(phen.df)

Although family ID and sex are listed as columns, this info is not considered in the current implementation and can be set to NA.

As described in Section 4.3, we combine the CNV calls with the phenotype information in a RaggedExperiment for coordinated representation and analysis.

re <- RaggedExperiment::RaggedExperiment(cnv.calls, colData=phen.df)
re

If probe information is available and has been annotated to the CNV calls, as we did above, the probe IDs and corresponding genomic positions should be provided in a separate file.
For compatibility with PLINK's map file format, this is expected to be a tab-delimited file containing exactly three columns: probe ID, chromosome, and the position in bp.

map.loc <- file.path(data.dir, "MapPenn.txt")
map.df <- read.delim(map.loc, as.is=TRUE)
head(map.df)

Given a RaggedExperiment storing CNV calls together with phenotype information, and optionally a map file for probe-level coordinates, the setupCnvGWAS function sets up all files needed for the GWAS analysis.

The information required for analysis is stored in the resulting phen.info list:

phen.info <- setupCnvGWAS("example", cnv.out.loc=re, map.loc=map.loc)
phen.info

The last item of the list displays the working directory:

all.paths <- phen.info$all.paths
all.paths

For the GWAS, chromosome names are assumed to be integer (i.e. 1, 2, 3, ...). Non-integer chromosome names can be encoded by providing a data.frame that describes the mapping from character names to corresponding integers.

For the example data, chromosomes 1A, 4A, 25LG1, 25LG2, and LGE22 are correspondingly encoded via

# Define chr correspondence to numeric
chr.code.name <- data.frame(   
                    V1 = c(16, 25, 29:31), 
                    V2 = c("1A", "4A", "25LG1", "25LG2", "LGE22"))

Running a CNV-GWAS

We can then run the actual CNV-GWAS, here without correction for multiple testing which is done for demonstration only. In real analyses, multiple testing correction is recommended to avoid inflation of false positive findings.

segs.pvalue.gr <- cnvGWAS(phen.info, chr.code.name=chr.code.name, method.m.test="none")
segs.pvalue.gr

The CNV-GWAS uses the concept of CNV segments to define more fine-grained CNV loci within CNV regions.

knitr::include_graphics(file.path(vign.dir, "CNVsegments.png"))

This procedure was originally proposed in our previous work in Nelore cattle (da Silva et al., 2016) and defines CNV segments based on CNV genotype similarity of subsequent SNP probes.

The default is min.sim=0.95, which will continuously add probe positions to a given CNV segment until the pairwise genotype similarity drops below 95%. An example of detailed up-down CNV genotype concordance that is used for the construction of CNV segments is given in S12 Table in da Silva et al., 2016.

As PLINK returns a p-value for each probe, only one of the p-values of the probes contained in a CNV segment is chosen as the segment p-value.

This is similar to a common approach used in differential expression (DE) analysis of microarray gene expression data, where typically the most significant DE probe is chosen in case of multiple probes mapping to the same gene.

Here, the representative probe for the CNV segment can be chosen to be the probe with lowest p-value (assign.probe="min.pvalue", default) or the one with highest CNV frequency (assign.probe="high.freq").

Multiple testing correction based on the number of CNV segments tested is carried out using the FDR approach (default). Results can then be displayed as for regular GWAS via a Manhattan plot (which can optionally be exported to a pdf file).

## Define the chromosome order in the plot
order.chrs <- c(1:24, "25LG1", "25LG2", 27:28, "LGE22", "1A", "4A")

## Chromosome sizes
chr.size.file <- file.path(data.dir, "Parus_major_chr_sizes.txt")
chr.sizes <- scan(chr.size.file)
chr.size.order <- data.frame(chr=order.chrs, sizes=chr.sizes, stringsAsFactors=FALSE)

## Plot a pdf file with a manhatthan within the 'Results' workfolder
plotManhattan(all.paths, segs.pvalue.gr, chr.size.order, plot.pdf=FALSE)

Producing a GDS file in advance

The genomic data structure (GDS) file format supports efficient memory management for genotype analysis. To make use of this efficient data representation, CNV genotypes analyzed with the cnvGWAS function are stored in a CNV.gds file, which is automatically produced and placed in the Inputs folder (i.e. all.paths[1]).

Therefore, running a GWAS implies that any GDS file produced by previous analysis will be overwritten. Use produce.gds=FALSE to avoid overwriting in the GWAS run.

For convenience, a GDS file can be produced before the GWAS analysis with the generateGDS function. This additionally returns a GRanges object containing the genomic position, name and, frequency of each probe used to construct the CNV segments for the GWAS analysis.

Note that probes.cnv.gr object contains the integer chromosome names (as the GDS file on disk). Only the segs.pvalue.gr, which stores the GWAS results, has the character chromosome names.

## Create a GDS file in disk and export the SNP probe positions
probes.cnv.gr <- generateGDS(phen.info, chr.code.name=chr.code.name)
probes.cnv.gr

## Run GWAS with existent GDS file
segs.pvalue.gr <- cnvGWAS(phen.info, chr.code.name=chr.code.name, produce.gds=FALSE)

Using relative signal intensities

SNP-based CNV callers such as PennCNV and Birdsuit infer CNVs from SNP-chip intensities (log R ratios, LRRs) and allele frequencies (B allelel frequencies, BAFs). As an auxiliary analysis, it can be interesting to carry out the GWAS based on the LRR relative signal intensities itself (da Silva et al., 2018).

To perform the GWAS using LRR values, import the LRR/BAF values and set run.lrr=TRUE in the cnvGWAS function:

# List files to import LRR/BAF 
files <- list.files(data.dir, pattern = "\\.cnv.txt.adjusted$")
samples <- sub(".cnv.txt.adjusted$", "", files)
samples <- sub("^GT","", samples)
sample.files <- data.frame(file.names=files, sample.names=samples)

# All missing samples will have LRR = '0' and BAF = '0.5' in all SNPs listed in the GDS file
importLrrBaf(all.paths, data.dir, sample.files, verbose=FALSE)

# Read the GDS to check if the LRR/BAF nodes were added
cnv.gds <- file.path(all.paths[1], "CNV.gds")
(genofile <- SNPRelate::snpgdsOpen(cnv.gds, allow.fork=TRUE, readonly=FALSE))
gdsfmt::closefn.gds(genofile)

# Run the CNV-GWAS with existent GDS
segs.pvalue.gr <- cnvGWAS(phen.info, chr.code.name=chr.code.name, produce.gds=FALSE, run.lrr=TRUE)

Session info

sessionInfo()

CNVRanger documentation built on Dec. 12, 2020, 2 a.m.