Contents

1 Introduction

Gene set enrichment (GSE) testing enables the interpretation of lists of differentially expressed genes (e.g. from RNA-seq), or lists of peaks (e.g. from ChIP-seq), in terms of pathways and other biologically meaningful sets of genes. The chipenrich package was originally designed to perform GSE for ChIP-seq peaks, but it can also be used for genomic regions with different biological meaning. The primary innovation of chipenrich is its accounting for biases that are known to affect the Type I error of such testing. In particular, the length of a gene’s regulatory region affects the probability that a peak will be assigned to it, the number of peaks that will be assigned to it, or the proportion of it covered by peaks.

The chipenrich package includes different enrichment methods for different use cases:

2 Concepts and Usage

library(chipenrich)
## 
## 
## 
## 
## 

2.1 Peaks

A ChIP-seq peak is a genomic region that represents a transcription factor binding event or the presence of a histone complex with a particular histone modification. Typically peaks are called with a peak caller (such as MACS2 or PePr) and represent relative enrichment of reads in a sample where the antibody is present versus input. Typically, peaks are output by a peak caller in BED-like format.

The primary user input for chipenrich(), broadenrich(), or polyenrich() are the peaks called from reads in a ChIP-seq experiment. Lists of genomic regions having other biological meaning can be used, but we shall continue to refer to ‘peaks’. Peaks can be input as either a file path or a data.frame.

If a file path, the following formats are fully supported via their file extensions: .bed, .broadPeak, .narrowPeak, .gff3, .gff2, .gff, and .bedGraph or .bdg. BED3 through BED6 files are supported under the .bed extension (BED specification). Files without these extensions are supported under the conditions that the first 3 columns correspond to chr, start, and end and that there is either no header column, or it is commented out. Files may be compressed with gzip, and so might end in .narrowPeak.gz, for example. For files with extension support, the rtracklayer::import() function is used to read peaks, so adherence to the mentioned file formats is necessary.

If peaks are already in the R environment as a data.frame, the GenomicRanges::makeGRangesFromDataFrame() function is used to convert to a GRanges object. For the acceptable column names needed for correct interpretation, see ?GenomicRanges::makeGRangesFromDataFrame.

For the purpose of the vignette, we’ll load some ChIP-seq peaks from the chipenrich.data companion package:

data(peaks_E2F4, package = 'chipenrich.data')
data(peaks_H3K4me3_GM12878, package = 'chipenrich.data')

head(peaks_E2F4)
##   chrom     start       end
## 1  chr1 156186314 156186469
## 2  chr1  10490456  10490550
## 3  chr1  46713352  46713436
## 4  chr1 226496843 226496924
## 5  chr1 200589825 200589928
## 6  chr1  47779789  47779907
head(peaks_H3K4me3_GM12878)
##   chrom    start      end
## 1 chr22 16846080 16871326
## 2 chr22 17305402 17306803
## 3 chr22 17517008 17517744
## 4 chr22 17518172 17518768
## 5 chr22 17518987 17520014
## 6 chr22 17520113 17520375

2.2 Genomes

Genomes for fly, human, mouse, rat, and zebrafish are supported. Particular supported genome builds are given by:

supported_genomes()
##  [1] "danRer10" "dm3"      "dm6"      "hg19"     "hg38"     "mm10"    
##  [7] "mm9"      "rn4"      "rn5"      "rn6"

2.3 Locus Definitions

A locus definition is a way of defining a gene regulatory region, and enables us to associate peaks with genes. The terms ‘gene’, ‘gene regulatory region’, and ‘gene locus’ are used interchangeably in the vignette. A trivial locus definition might be the gene bodies from the transcription start sites (TSS) to the transcript end sites (TES) for each gene. A locus definition can also express how one expects a transcription factor to regulate genes. For example, a locus definition defined as 1kb upstream and downstream of a TSS (the 1kb definition) would capture TFs binding in proximal-promoter regions.

2.3.1 Built-in locus definitions

A number of locus definitions representing different regulatory paradigms are included in the package:

  • nearest_tss: The locus is the region spanning the midpoints between the TSSs of adjacent genes.
  • nearest_gene: The locus is the region spanning the midpoints between the boundaries of each gene, where a gene is defined as the region between the furthest upstream TSS and furthest downstream TES for that gene. If gene loci overlap, the midpoint of the overlap is used as a border. If a gene locus is nested in another, the larger locus is split in two.
  • exon: Each gene has multiple loci corresponding to its exons. Overlaps between different genes are allowed.
  • intron: Each gene has multiple loci corresponding to its introns. Overlaps between different genes are allowed.
  • 1kb: The locus is the region within 1kb of any of the TSSs belonging to a gene. If TSSs from two adjacent genes are within 2 kb of each other, we use the midpoint between the two TSSs as the boundary for the locus for each gene.
  • 1kb_outside_upstream: The locus is the region more than 1kb upstream from a TSS to the midpoint between the adjacent TSS.
  • 1kb_outside: The locus is the region more than 1kb upstream or downstream from a TSS to the midpoint between the adjacent TSS.
  • 5kb: The locus is the region within 5kb of any of the TSSs belonging to a gene. If TSSs from two adjacent genes are within 10 kb of each other, we use the midpoint between the two TSSs as the boundary for the locus for each gene.
  • 5kb_outside_upstream: The locus is the region more than 5kb upstream from a TSS to the midpoint between the adjacent TSS.
  • 5kb_outside: The locus is the region more than 5kb upstream or downstream from a TSS to the midpoint between the adjacent TSS.
  • 10kb: The locus is the region within 10kb of any of the TSSs belonging to a gene. If TSSs from two adjacent genes are within 20 kb of each other, we use the midpoint between the two TSSs as the boundary for the locus for each gene.
  • 10kb_outside_upstream: The locus is the region more than 10kb upstream from a TSS to the midpoint between the adjacent TSS.
  • 10kb_outside: The locus is the region more than 10kb upstream or downstream from a TSS to the midpoint between the adjacent TSS.

The complete listing of genome build and locus definition pairs can be listed with supported_locusdefs():

# Take head because it's long
head(supported_locusdefs())
##     genome              locusdef
## 1 danRer10                  10kb
## 2 danRer10          10kb_outside
## 3 danRer10 10kb_outside_upstream
## 4 danRer10                   1kb
## 5 danRer10           1kb_outside
## 6 danRer10  1kb_outside_upstream

2.3.2 Custom locus definitions

Users can create custom locus definitions for any of the supported_genomes(), and pass the file path as the value of the locusdef parameter in broadenrich(), chipenrich(), or polyenrich(). Custom locus definitions should be defined in a tab-delimited text file with column names chr, start, end, and gene_id. For example:

chr start   end geneid
chr1    839460  839610  148398
chr1    840040  840190  148398
chr1    840040  840190  57801
chr1    840800  840950  148398
chr1    841160  841310  148398

2.3.3 Selecting a locus definition

For a transcription factor ChIP-seq experiment, selecting a particular locus definition for use in enrichment testing implies how the TF is assumed to regulate genes. For example, selecting the 1kb locus definition will imply that the biological processes found enriched are a result of TF regulation near the promoter. In contrast, selecting the 5kb_outside locus definition will imply that the biological processes found enriched are a result of TF regulation distal from the promoter.

Selecting a locus definition can also help reduce the noise in the enrichment tests. For example, if a TF is known to primarily regulate genes by binding around the promoter, then selecting the 1kb locus definition can help to reduce the noise from TSS-distal peaks in the enrichment testing.

The plot_dist_to_tss() QC plot displays where peak midpoints fall relative to TSSs genome-wide, and can help inform the choice of locus definition. For example, if many peaks fall far from the TSS, the nearest_tss locus definition may be a good choice because it will capture all peaks, whereas the 1kb locus definition may not capture many of the peaks and adversely affect the enrichment testing.

2.4 Gene Sets

Gene sets are sets of genes that represent a particular biological function.

2.4.1 Built-in gene sets

Gene sets for fly, human, mouse, rat, and zebrafish are built in to chipenrich. Some organisms have gene sets that others do not, so check with:

# Take head because it's long
head(supported_genesets())
##     geneset organism
## 1      GOBP      dme
## 6      GOCC      dme
## 11     GOMF      dme
## 44 reactome      dme
## 2      GOBP      dre
## 7      GOCC      dre

Descriptions of our built-in gene sets:

  • GOBP: Gene Ontology-Biological Processes, Bioconductor ver. 3.4.2. (geneontology.org)

  • GOCC: Gene Ontology-Cell Component, Bioconductor ver. 3.4.2. (geneontology.org)

  • GOMF: Gene Ontology-Molecular Function, Bioconductor ver 3.4.2. (geneontology.org)

  • biocarta_pathway: BioCarta Pathway, Ver 6.0. (cgap.nci.nih.gov/Pathways/BioCarta_Pathways)

  • ctd: Comparative Toxicogenomics Database, Last updated June 06, 2017. Groups of genes that interact with specific chemicals to help understand enviromental exposures that affect human health. (ctdbase.org)

  • cytoband: Cytobands (NCBI). Groups of genes that reside in the same area of a chromosome.

  • drug_bank: Sets of gene that are targeted by a specific drug. Ver 5.0.7. (www.drugbank.ca)

  • hallmark: Hallmark gene sets (MSigDB). Ver 6.0. Specific biological states or processes that display coherent expression. (software.broadinstitute.org/gsea/msigdb/collections.jsp)

  • immunologic: Immunologic signatures (MSigDB). Ver 6.0. Gene sets that represent cell states within the immune system. (software.broadinstitute.org/gsea/msigdb/collections.jsp)

  • kegg_pathway: Kyoto Encyclopedia of Genes and Genomes. Ver 3.2.3. (genome.jp/kegg)

  • mesh: Gene Annotation with MeSH, the National Library of Medicine’s controlled vocabulary for biology and medicine. Useful for testing hypotheses related to diseases, processes, other genes, confounders such as populations and experimental techniques, etc. based on knowledge from the literature that may not yet be formally described in any other gene sets. Last updated ~2013. (gene2mesh.ncibi.org)

  • metabolite: Metabolite concepts, defined from Edinburgh Human Metabolic Network database (Ma, et al., 2007) Contains gene sets coding for metabolic enzymes.

  • microrna: microRNA targets (MSigDB). Ver 6.0. Gene sets containing genes with putative target sites of human mature miRNA. (software.broadinstitute.org/gsea/msigdb/collections.jsp)

  • oncogenic: Oncogenic signatures (MSigDB). Ver 6.0. Gene sets that represent signatures of pathways often disregulated in cancer. (software.broadinstitute.org/gsea/msigdb/collections.jsp)

  • panther_pathway: PANTHER Pathway. Ver 3.5. Contains primarily signaling pathways with subfamilies. (pantherdb.org/pathway)

  • pfam: Pfam. A large collection of protein families. (pfam.xfam.org)

  • protein_interaction_biogrid: Protein Interaction from Biological General Repository for Interaction Datasets. Ver 3.4.151. (thebiogrid.org)

  • reactome: Reactome Pathway Database. Ver 61. (reactome.org)

  • transcription_factors: Transcription Factors (MSigDB). Ver 6.0. Gene sets that share upstream cis-regulatory motifs which can function as potential transcription factor binding sites. (software.broadinstitute.org/gsea/msigdb/collections.jsp)

2.4.2 Custom gene sets

Users can perform GSE on custom gene sets for any supported organism by passing the file path as the value of genesets parameter in broadenrich(), chipenrich(), polyenrich(), or hybridenrich(). Custom gene set definitions should be defined in a tab-delimited text file with a header. The first column should be the geneset ID or name, and the second column should be the Entrez IDs belonging to the geneset. For example:

gs_id   gene_id
GO:0006631  30
GO:0006631  31
GO:0006631  32
GO:0006631  33
GO:0006631  34
GO:0006631  35
GO:0006631  36
GO:0006631  37
GO:0006631  51
GO:0006631  131
GO:0006631  183
GO:0006631  207
GO:0006631  208
GO:0006631  215
GO:0006631  225

2.5 Mappability

We define base pair mappability as the average read mappability of all possible reads of size K that encompass a specific base pair location, \(b\). Mappability files from UCSC Genome Browser mappability track were used to calculate base pair mappability. The mappability track provides values for theoretical read mappability, or the number of places in the genome that could be mapped by a read that begins with the base pair location \(b\). For example, a value of 1 indicates a Kmer read beginning at \(b\) is mappable to one area in the genome. A value of 0.5 indicates a Kmer read beginning at \(b\) is mappable to two areas in the genome. For our purposes, we are only interested in uniquely mappable reads; therefore, all reads with mappability less than 1 were set to 0 to indicate non-unique mappability. Then, base pair mappability is calculated as:

\[ \begin{equation} M_{i} = (\frac{1}{2K-1}) \sum_{j=i-K+1}^{i+(K-1)} M_{j} \end{equation} \]

where \(M_{i}\) is the mappability of base pair \(i\), and \(M_{j}\) is mappability (from UCSC’s mappability track) of read \(j\) where j is the start position of the K length read.

2.5.1 Built-in mappability

Base pair mappability for reads of lengths 24, 36, 40, 50, 75, and 100 base pairs for hg19 and for reads of lengths 36, 40, 50, 75, and 100 base pairs mm9 a included. See the complete list with:

# Take head because it's long
head(supported_read_lengths())
##   genome locusdef read_length
## 1   hg19     10kb         100
## 2   hg19     10kb          24
## 3   hg19     10kb          36
## 4   hg19     10kb          40
## 5   hg19     10kb          50
## 6   hg19     10kb          75

2.5.2 Custom mappability

Users can use custom mappability with any built-in locus definition (if, for example, the read length needed is not present), or with a custom locus definition. Custom mappability should be defined in a tab-delimited text file with columns named gene_id and mappa. Gene IDs should be Entrez Gene IDs, and mappability should be in [0,1]. A check is performed to verify that the gene IDs in the locus definition and mappability overlap by at least 95%. An example custom mappability file looks like:

mappa   gene_id
0.8 8487
0.1 84
0.6 91
1   1000

2.6 Testing for enrichment

As stated in the introduction, the chipenrich package includes three classes of methods for doing GSE testing. For each method, we describe the intended use case, the model used for enrichment, and an example using the method.

2.6.1 broadenrich()

Broad-Enrich is designed for use with broad peaks that may intersect multiple gene loci, and cumulatively cover greater than 5% of the genome. For example, ChIP-seq experiments for histone modifications.

The Broad-Enrich method uses the cumulative peak coverage of genes in its logistic regression model for enrichment: GO ~ ratio + s(log10_length). Here, GO is a binary vector indicating whether a gene is in the gene set being tested, ratio is a numeric vector indicating the ratio of the gene covered by peaks, and s(log10_length) is a binomial cubic smoothing spline which adjusts for the relationship between gene coverage and locus length.

gs_path = system.file('extdata','vignette_genesets.txt', package='chipenrich')
results = broadenrich(peaks = peaks_H3K4me3_GM12878, genome = 'hg19', genesets = gs_path,
    locusdef = "nearest_tss", qc_plots = FALSE, out_name = NULL, n_cores=1)
results.be = results$results
print(results.be[1:5,1:5])
##    Geneset.Type Geneset.ID Description      P.value          FDR
## 1 user-supplied GO:0002521  GO:0002521 9.600908e-06 8.000756e-05
## 2 user-supplied GO:0031400  GO:0031400 7.564088e-05 4.727555e-04
## 3 user-supplied GO:0022411  GO:0022411 1.775843e-04 8.879214e-04
## 4 user-supplied GO:0071845  GO:0071845 6.040128e-04 1.888373e-03
## 5 user-supplied GO:0022604  GO:0022604 3.868645e-03 9.671612e-03

2.6.2 chipenrich()

ChIP-Enrich is designed for use with 1,000s or 10,000s of narrow peaks which results in fewer gene loci containing a peak overall. For example, ChIP-seq experiments for transcription factors.

The ChIP-Enrich method uses the presence of a peak in its logistic regression model for enrichment: peak ~ GO + s(log10_length). Here, GO is a binary vector indicating whether a gene is in the gene set being tested, peak is a binary vector indicating the presence of a peak in a gene, and s(log10_length) is a binomial cubic smoothing spline which adjusts for the relationship between the presence of a peak and locus length.

# Without mappability
gs_path = system.file('extdata','vignette_genesets.txt', package='chipenrich')
results = chipenrich(peaks = peaks_E2F4, genome = 'hg19', genesets = gs_path,
    locusdef = "nearest_tss", qc_plots = FALSE, out_name = NULL, n_cores = 1)
results.ce = results$results
print(results.ce[1:5,1:5])
##    Geneset.Type Geneset.ID Description      P.value          FDR
## 1 user-supplied GO:0034660  GO:0034660 5.435777e-05 0.0004529814
## 2 user-supplied GO:0007346  GO:0007346 8.592104e-05 0.0005370065
## 3 user-supplied GO:0031400  GO:0031400 1.164884e-03 0.0058244176
## 4 user-supplied GO:0009314  GO:0009314 2.166075e-02 0.0676898435
## 5 user-supplied GO:0051129  GO:0051129 1.196240e-01 0.2718727018
# With mappability
gs_path = system.file('extdata','vignette_genesets.txt', package='chipenrich')
results = chipenrich(peaks = peaks_E2F4, genome = 'hg19', genesets = gs_path,
    locusdef = "nearest_tss", mappability=24, qc_plots = FALSE,
    out_name = NULL,n_cores=1)
results.cem = results$results
print(results.cem[1:5,1:5])
##    Geneset.Type Geneset.ID Description      P.value          FDR
## 1 user-supplied GO:0034660  GO:0034660 4.552997e-05 0.0003794164
## 2 user-supplied GO:0007346  GO:0007346 7.076743e-05 0.0004422965
## 3 user-supplied GO:0031400  GO:0031400 1.045192e-03 0.0052259579
## 4 user-supplied GO:0009314  GO:0009314 2.358027e-02 0.0736883370
## 5 user-supplied GO:0043623  GO:0043623 1.125058e-01 0.2412736564

2.6.3 polyenrich()

Poly-Enrich is also designed for narrow peaks, but where there are tens of thousands of peaks which results in nearly every gene locus containing a peak. For example, ChIP-seq experiments for transcription factors or repeated elements.

The Poly-Enrich method uses the number of peaks in genes in its negative binomial regression model for enrichment: num_peaks ~ GO + s(log10_length). Here, GO is a binary vector indicating whether a gene is in the gene set being tested, num_peaks is a numeric vector indicating the number of peaks in each gene, and s(log10_length) is a negative binomial cubic smoothing spline which adjusts for the relationship between the number of peaks in a gene and locus length.

gs_path = system.file('extdata','vignette_genesets.txt', package='chipenrich')
results = polyenrich(peaks = peaks_E2F4, genome = 'hg19', genesets = gs_path,  method = 'polyenrich',
    locusdef = "nearest_tss", qc_plots = FALSE, out_name = NULL, n_cores = 1)
results.pe = results$results
print(results.pe[1:5,1:5])
##    Geneset.Type Geneset.ID Description      P.value          FDR
## 1 user-supplied GO:0007346  GO:0007346 7.530343e-05 0.0006275286
## 2 user-supplied GO:0031400  GO:0031400 5.373275e-04 0.0033582970
## 3 user-supplied GO:0009314  GO:0009314 1.529658e-03 0.0063735763
## 4 user-supplied GO:0051129  GO:0051129 1.428948e-01 0.4344936832
## 5 user-supplied GO:0043623  GO:0043623 2.040071e-01 0.4344936832

2.6.4 hybridenrich()

The hybrid method is used when one is unsure of which method, between ChIP-Enrich or Poly-Enrich, is the optimal method. It will be more conservative than either test individually, but also much more powerful than using the wrong test.

The hybrid p-value is given as 2*min(chipenrich_pvalue, polyenrich_pvalue). This test will retain the same Type 1 level and will be a consistent test if one of chipenrich or polyenrich is consistent. This can be extended to any number of tests, but currently we only allow a hybrid test for chipenrich and polyenrich. For more information about chipenrich or polyenrich, see their respective sections.

gs_path = system.file('extdata','vignette_genesets.txt', package='chipenrich')
results = hybridenrich(peaks = peaks_E2F4, genome = 'hg19', genesets = gs_path,
    locusdef = "nearest_tss", qc_plots = F, out_name = NULL, n_cores = 1)
results.hybrid = results$results
print(results.hybrid[1:5,1:5])
##   Geneset.ID  Geneset.Type Description          FDR     Effect
## 1 GO:0001558 user-supplied  GO:0001558 3.910701e-01 -0.1554272
## 2 GO:0002521 user-supplied  GO:0002521 2.869857e-01 -0.1787083
## 3 GO:0003013 user-supplied  GO:0003013 5.871998e-07 -0.6905978
## 4 GO:0006631 user-supplied  GO:0006631 3.910701e-01 -0.1429443
## 5 GO:0007346 user-supplied  GO:0007346 5.370065e-04  0.4998322

2.7 QC Plots

Each enrich function outputs QC plots if qc_plots = TRUE. There are also stand-alone functions to make the QC plots without the need for GSE testing. The QC plots can be used to help determine which locus definition to use, or which enrichment method is more appropriate.

2.7.1 Peak distance to TSS distribution

This plot gives a distribution of the distance of the peak midpoints to the TSSs. It can help in selecting a locus definition. For example, if genes are primarily within 5kb of TSSs, then the 5kb locus definition may be a good choice. In contrast, if most genes fall far from TSSs, the nearest_tss locus definition may be a good choice.

# Output in chipenrich and polyenrich
plot_dist_to_tss(peaks = peaks_E2F4, genome = 'hg19')
E2F4 peak distances to TSS

Figure 1: E2F4 peak distances to TSS

2.7.2 Presence of peak versus locus length

This plot visualizes the relationship between the presence of at least one peak in a gene locus and the locus length (on the log10 scale). For clarity of visualization, each point represents 25 gene loci binned after sorting by locus length. The expected fit under the assumptions of Fisher’s Exact Test (horizontal line), and a binomial-based test (gray curve) are displayed to indicate how the dataset being enriched conforms to the assumption of each. The empirical spline used in the chipenrich method is in orange.

# Output in chipenrich
plot_chipenrich_spline(peaks = peaks_E2F4, locusdef = 'nearest_tss', genome = 'hg19')
E2F4 chipenrich spline without mappability

Figure 2: E2F4 chipenrich spline without mappability

2.7.3 Number of peaks versus locus length

This plot visualizes the relationship between the number of peaks assigned to a gene and the locus length (on the log10 scale). For clarity of visualization, each point represents 25 gene loci binned after sorting by locus length. The empirical spline used in the polyenrich method is in orange.

If many gene loci have multiple peaks assigned to them, polyenrich is likely an appropriate method. If there are a low number of peaks per gene, then chipenrich() may be the appropriate method.

# Output in polyenrich
plot_polyenrich_spline(peaks = peaks_E2F4, locusdef = 'nearest_tss', genome = 'hg19')
E2F4 polyenrich spline without mappability

Figure 3: E2F4 polyenrich spline without mappability

2.7.4 Gene coverage versus locus length

This plot visualizes the relationship between proportion of the gene locus covered by peaks and the locus length (on the log10 scale). For clarity of visualization, each point represents 25 gene loci binned after sorting by locus length.

# Output in broadenrich
plot_gene_coverage(peaks = peaks_H3K4me3_GM12878, locusdef = 'nearest_tss',  genome = 'hg19')