--- title: "extraChIPs: Differential Signal Using Sliding Windows" author: - name: Stevie Pederson affiliation: - Black Ochre Data Laboratories, Telethon Kids Institute, Adelaide, Australia - Dame Roma Mitchell Cancer Researc Laboratories, University of Adelaide - John Curtin School of Medical Research, Australian National University email: stephen.pederson.au@gmail.com package: extraChIPs bibliography: '`r system.file("references.bib", package = "extraChIPs")`' output: BiocStyle::html_document vignette: > %\VignetteIndexEntry{Differential Signal Analysis (Sliding Windows)} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} --- ```{r setup} knitr::opts_chunk$set(message = FALSE, crop = NULL) ``` # Introduction The [GRAVI](https://github.com/smped/GRAVI) workflow, for which this package is designed, uses sliding windows for differential signal analysis using strategies defined in the package `csaw` (Lun and Smyth 2016). The GRAVI workflow itself extends to integrating multiple ChIP targets and external data sources, and as such, this package introduces a handful of functions to simplify and enable these analyses. Whilst many existing approaches refer to this type of analysis as Differential Binding analysis, we prefer the term *Differential Signal Analysis* as this more accurately captures the range of ChIP targets which are likely to be investigated, such as H3K27ac or similar. The majority of examples below use heavily reduced datasets to provide general guidance on using the functions. Some results may appear trivial as a result, but will hopefully prove far more useful in a true experimental context. All data, along with this vignette are available [here](https://github.com/smped/extraChIPs_vignette). In order to replicate this vignette, please place all contents of the `data` directory provided in this additional repository into a directory named data in your own working directory. It may also be worth noting that this vignette should run on a conventional laptop with no particularly large performance requirements. However, when following this workflow across an entire genome, memory requirements may exceed those of a standard laptop, and an HPC or high-performance workstation may be required. # Setup ## Installation In order to use the package `extraChIPs` and follow this vignette, we recommend using the package `BiocManager` (hosted on CRAN) to install all packages. Once `BiocManager` is installed, the additional packages required for this vignette can also be installed as shown below. ``` r if (!"BiocManager" %in% rownames(installed.packages())) install.packages("BiocManager") pkg <- c( "tidyverse", "Rsamtools", "csaw", "BiocParallel", "rtracklayer", "edgeR", "patchwork", "extraChIPs", "plyranges", "scales", "glue", "here", "quantro", "cqn", "ggrepel", "BSgenome.Hsapiens.UCSC.hg19" ) BiocManager::install(pkg, update = FALSE) ``` Once these packages are installed, we can load them easily ```{r load-packages, eval = TRUE} library(tidyverse) library(Rsamtools) library(csaw) library(BiocParallel) library(rtracklayer) library(edgeR) library(patchwork) library(extraChIPs) library(plyranges) library(scales) library(glue) library(here) library(magrittr) library(quantro) library(cqn) library(ggrepel) theme_set(theme_bw()) ``` ## Data All data for this vignette is expected to be in a sub-directory of the working directory named “data” (including “data/H3K27ac”), and all paths will be predicated on this. Please ensure you have all data in this location, obtained from [here](https://github.com/smped/extraChIPs_vignette). The data itself is ChIP-Seq data targeting the histone mark H3K27ac, and is taken from the ZR-75-1 cell-line using data from the BioProject , Preprocessing was performed using the [`prepareChIPs`](https://github.com/smped/prepareChIPs) workflow, written in snakemake (Mölder et al. 2021) and all code is available at . H3K27ac signal was obtained under baseline (E2) and DHT-treated (E2DHT) conditions, with alignments being to GRCh37. For this workflow, data has been subset to a region found on chromosome 10 for simplicity. A description of the samples can be found in the file `data/PRJNA509779.tsv`. A pooled Input (IgG) sample was used for the entire dataset and this accession is provided in a separate column. ```{r samples, eval = FALSE} samples <- here("data", "PRJNA509779.tsv") %>% read_tsv() %>% dplyr::filter(target == "H3K27ac") %>% mutate(treatment = factor(treatment, levels = c("E2", "E2DHT"))) accessions <- unique(c(samples$accession, samples$input)) treat_colours <- c("steelblue", "red3", "grey") names(treat_colours) <- c(levels(samples$treatment), "Input") ``` # Differential Signal Analysis ## Sliding Windows The standard approach of tools such as DiffBind (Ross-Innes et al. 2012) is to take a set of peaks, re-centre them, then set all regions to be the same width. From there, data is passed to `edgeR` (Chen, Lun, and Smyth 2016) or `DESeq2` (Love, Huber, and Anders 2014) for analysis. This approach using `extraChIPs` is demonstrated in a separate vignette, however, this workflow instead focusses on a sliding window approach as per the package `csaw` (Lun and Smyth 2016). The resultant *variable width regions* can be particularly advantageous for ChIP targets such as H3K27ac where regions marked by histone-acetylation can vary greatly in size. The starting point for differential signal analyses using `extraChIPs` is to define a set of sliding windows across the genome, then count reads from a set of bam files, defined as a `BamFileList.` Commonly one or more IP input/control samples is also produced during a ChIP-Seq experiment, and these should be included at this stage of the analysis. The example files provided here contain a small subset of reads from chromosome 10 across two experimental conditions and one input sample, and we will define them all as a `BamFileList`. ```{r bfl, eval = FALSE} bfl <- here("data", "H3K27ac", glue("{accessions}.bam")) %>% BamFileList() %>% setNames(str_remove_all(names(.), ".bam")) file.exists(path(bfl)) ``` ## [1] TRUE TRUE TRUE TRUE TRUE TRUE TRUE `Seqinfo` objects are the foundation of working with GRanges, so defining an object at the start of a workflow is good practice. This is simply enabled with `extraChIPs` by using `defineSeqinfo()` and specifying the appropriate genome. ```{r sq} sq <- defineSeqinfo("GRCh37") ``` **NB:** It should also be noted that counting all reads across a `BamFileList` using sliding windows, **will require \>50Gb of RAM** and will be beyond the capacity of most laptops as of the time of writing. When working with complete datasets, this particular step is best performed on an HPC or a similar interactive server with a large amount of memory. The approach taken below is to first define a set of sliding windows across the genome, using the capabilities of `csaw`. After counting reads across all windows, a set of *pre-defined regions* where H3K27ac signal is confidently found, is then used guide the function `dualFilter()`. This function will discard low-signal windows, retaining only those a) above a minimum signal level and b) with signal notably above that of any input samples. These regions can be obtained from any external resource, or can even be taken from `macs2`-defined peaks from the same samples. Here, the set of regions was defined as those found when merging samples from each treatment, in both treatments, with an FDR \< 0.01. Importantly, our analysis will not be restricted to these regions, but these will instead guide `dualFilter()` as to the best regions to retain for analysis. First we can define our windows and count the alignments using the existing capabilities and functions provided in the `csaw` package. In the following, we’ll use a sliding window of 120bp and a step size of 40bp, meaning each nucleotide is covered by 3 windows. In addition, we’ll exclude blacklisted and grey-listed regions as provided in the dataset. These can be obtained easily by using the `GreyListChIP` package, which is beyond the scope of this vignette, however, code for preparing the GreyList\` is available [here](https://github.com/smped/PRJNA509779/blob/main/workflow/scripts/make_greylist.R). Fragment sizes for each sample were also estimated to be around 200nt. ```{r wincounts, eval = FALSE} greylist <- here("data", "chr10_greylist.bed") %>% import.bed(seqinfo = sq) blacklist <- here("data", "chr10_blacklist.bed") %>% import.bed(seqinfo = sq) rp <- readParam( pe = "none", dedup = TRUE, restrict = "chr10", discard = c(greylist, blacklist) ) wincounts <- windowCounts( bam.files = bfl, spacing = 40, width = 120, ext = 200, filter = length(bfl), param = rp ) seqlevels(wincounts) <- seqlevels(sq) seqinfo(wincounts) <- sq ``` This produces a `RangedSummarizedExperiment` with windows included which passed the minimum threshold of 7 total reads. We can check which windows passed this threshold using `rowRanges()` ```{r rowranges-wincounts, eval = FALSE} rowRanges(wincounts) ``` ## GRanges object with 506538 ranges and 0 metadata columns: ## seqnames ranges strand ## ## [1] chr10 42491801-42491920 * ## [2] chr10 42491921-42492040 * ## [3] chr10 42494601-42494720 * ## [4] chr10 42495641-42495760 * ## [5] chr10 42495681-42495800 * ## ... ... ... ... ## [506534] chr10 99999641-99999760 * ## [506535] chr10 99999681-99999800 * ## [506536] chr10 99999721-99999840 * ## [506537] chr10 99999761-99999880 * ## [506538] chr10 99999801-99999920 * ## ------- ## seqinfo: 25 sequences from GRCh37 genome We can also add some key information to the `colData` element of this object, which will also be propagated to all downstream objects. ```{r coldata-wincounts, eval = FALSE} colData(wincounts) ``` ## DataFrame with 7 rows and 4 columns ## bam.files totals ext rlen ## ## SRR8315186 /home/steviep/github.. 267342 200 75 ## SRR8315187 /home/steviep/github.. 307383 200 75 ## SRR8315188 /home/steviep/github.. 288965 200 75 ## SRR8315189 /home/steviep/github.. 296764 200 75 ## SRR8315190 /home/steviep/github.. 325211 200 75 ## SRR8315191 /home/steviep/github.. 283745 200 75 ## SRR8315192 /home/steviep/github.. 361835 200 75 ```{r update-coldata, eval = FALSE} colData(wincounts) <- colData(wincounts) %>% as_tibble(rownames = "accession") %>% left_join(samples, by = "accession") %>% dplyr::select( accession, all_of(colnames(colData(wincounts))), target, treatment ) %>% mutate( target = str_replace_na(target, "Input"), treatment = str_replace_na(treatment, "Input") %>% as.factor() ) %>% DataFrame(row.names = .$accession) ``` A density plot can be simply drawn of these counts, with the vast majority of windows receiving very low counts, due to the nature of transcription factor binding, where long stretches are unbound. The windows with higher counts tend to be associated with the samples targeting a transcription factor (TF), as seen in the two treatment group samples. ```{r plot-densities, eval = FALSE} plotAssayDensities(wincounts, colour = "treatment", trans = "log1p") + theme_bw() + scale_colour_manual(values = treat_colours) ``` ## Filtering of Sliding Windows After counting all reads in the sliding genomic windows, the next step is to discard windows for which counts are unlikely to represent true signal from our ChIP target. The strategy employed in `extraChIPs` uses a set of pre-defined regions to automatically set thresholds based on 1) counts strongly above the counts from the input sample, and 2) the windows with the overall highest signal. Thresholds are determined such that a proportion (e.g. `q = 0.5`) of the windows which overlap one of the supplied consensus peaks will be returned. Higher values for `q` will return more windows, however many of these will tend to only marginally overlap a peak in one of the tail regions, and these will most likely be covered by neighbouring windows. Experience has shown that values such as `q = 0.5` tend to return a considerable proportion of windows containing true signal from the ChIP target. Higher values will tend to return more sliding windows where the edges overlap the tails of the guide regions. The we can pass these to the function `dualFilter()` which utilises the strategy described above. On large datasets, this can be quite time-consuming, as can the initial counting step. Multiple alternative filtering strategies are also provided by the package `csaw` and these can be accessed using `?csaw::filterWindows` ```{r filter-counts, eval = FALSE} guide_regions <- here("data", "H3K27ac", "H3K27ac_chr10.bed") %>% import.bed(seqinfo = sq) filtcounts <- dualFilter( x = wincounts, bg = "SRR8315192", ref = guide_regions, q = 0.6 ) ``` Thus we have reduced our initial set of 506,538 sliding windows to the 23,417 windows most likely to contain true signal from our ChIP target. The returned object will by default contain `counts` and `logCPM` assays, with the complete library sizes used for the calculation of `logCPM` values. Similarly, *the input sample is no longer included* in the data object, although additional columns can easily be added to the returned object using any number of strategies. ``` r dim(wincounts) ``` ## [1] 506538 7 ``` r dim(filtcounts) ``` ## [1] 23417 6 ``` r assays(filtcounts) ``` ## List of length 2 ## names(2): counts logCPM The `rowData` element of the returned object will contain a logical column indicating where each specific retained window overlapped one of the supplied consensus peaks. ``` r rowRanges(filtcounts) ``` ## GRanges object with 23417 ranges and 1 metadata column: ## seqnames ranges strand | overlaps_ref ## | ## [1] chr10 42862921-42863040 * | TRUE ## [2] chr10 42862961-42863080 * | TRUE ## [3] chr10 42863001-42863120 * | TRUE ## [4] chr10 42863041-42863160 * | TRUE ## [5] chr10 42863081-42863200 * | TRUE ## ... ... ... ... . ... ## [23413] chr10 99895121-99895240 * | TRUE ## [23414] chr10 99895161-99895280 * | TRUE ## [23415] chr10 99895201-99895320 * | TRUE ## [23416] chr10 99895241-99895360 * | TRUE ## [23417] chr10 99895281-99895400 * | TRUE ## ------- ## seqinfo: 25 sequences from GRCh37 genome ## Initial Visualisation We can once again check our signal distributions, this time on the logCPM values. We can also check the samples using a PCA plot, again colouring the points by treatment group and adding labels, which will repel by default if the points are shown. ```{r plot-density-pca, eval = FALSE} A <- plotAssayDensities(filtcounts, assay = "logCPM", colour = "treatment") + scale_colour_manual(values = treat_colours) + theme_bw() B <- plotAssayPCA(filtcounts, "logCPM", colour = "treatment", label = "accession") + scale_colour_manual(values = treat_colours) + theme_bw() A + B + plot_layout(guides = "collect") + plot_annotation(tag_levels = "A") ``` ```{r add-dens-pca, echo = FALSE, eval = TRUE, fig.cap = "*A) Densities for logCPM values across all samples after discarding windows less likely to contain H3K27ac signal using `dualFilter()` B) PCA plot based on the logCPM assay*"} knitr::include_graphics("differential_signal_sliding_files/figure-gfm/plot-density-pca-1.png") ``` Another common QC step is Relative Log-Expression (RLE) (Gandolfo and Speed 2018). In the following, we first inspect the RLE across the entire dataset, followed by RLE grouping *within treatments*. This can be particularly useful when distributions vary significantly between treatment groups, such as may occur with a cytoplasmic to nuclear shift by a given ChIP target. Here, however, there is minimal difference between the two approaches as H3K27ac signal tends to be broadly consistent between these treatment groups. ```{r plot-assay-rle, eval = FALSE} a <- plotAssayRle(filtcounts, assay = "logCPM", fill = "treat") + geom_hline(yintercept = 0, linetype = 2, colour = "grey") + scale_fill_manual(values = treat_colours) + ggtitle("RLE: Across All Samples") + theme_bw() b <- plotAssayRle( filtcounts, assay = "logCPM", fill = "treat", rle_group = "treat" ) + geom_hline(yintercept = 0, linetype = 2, colour = "grey") + scale_fill_manual(values = treat_colours) + ggtitle("RLE: Within Treatment Groups") + theme_bw() a + b + plot_layout(guides = "collect") + plot_annotation(tag_levels = "A") ``` ```{r add-rle-plot, eval = TRUE, echo = FALSE, fig.cap = "*RLE plots (A) across all samples and (B) with values calculated within treatment groups*"} knitr::include_graphics("differential_signal_sliding_files/figure-gfm/plot-assay-rle-1.png") ``` ## Statistical Testing Multiple methods are enabled in the package `extraChIPs` via the function `fitAssayDiff()`, with the possibility of incorporating any additional normalisation strategies from external packages. The two basic strategies are 1) Quasi-Likelihood Fits (Lund et al. 2012) and 2) `limma-trend` (Law et al. 2014). The first approach (method = “qlf”) uses counts along with any of the normalisation strategies from `edgeR::calcNormFactors()`, and setting the normalisation method to “none” is the equivalent of library-size normalisation, which replicates the default normalisation strategy from DiffBind (Ross-Innes et al. 2012). If choosing to normalise *within treatment groups*, a factor can be provided via the groups argument, essentially adding this as an option for all methods provided in `edgeR::calcNormFactors()`. The second method (method = “lt”) is specifically for logCPM values and these can be provided as output by `dualFilter()` or may be normalised using any number of additional methods. In addition to the above methods, a range-based $H_0$ (McCarthy and Smyth 2009) can be specified by providing a value to the `fc` or `lfc` arguments, which removes any need for *post-hoc* filtering and correctly controls the FDR, unlike *post-hoc* filtering based on a fold-change threshold. Here, we’ll initially fit our data using Quasi-Likelihood Fits, library-size normalisation and setting a change in signal beyond the range of $\pm$ 20% as being of interest. By default, the returned object, would be a `SummarizedExperiment` object containing the results from model fitting in the `rowData()` element. However, setting `asRanges = TRUE` will simply return the set of GRanges along with the testing results as a stand-alone object. ```{r fit-gr, eval = FALSE} X <- model.matrix(~treatment, data = colData(filtcounts)) fit_gr <- fitAssayDiff(filtcounts, design = X, fc = 1.2, asRanges = TRUE) ``` ## Merging Windows After an analysis has been performed, values contained in the output will be estimated signal (`logCPM`), estimated change (`logFC`) with both raw and adjusted p-values for all sliding windows. Given the dependency of neighbouring windows, any adjusted p-values will not be appropriate and a merging of overlapping and/or neighbouring windows should be performed. Multiple `csaw` methods are wrapped using `mergeByCol()`, `mergeBySig()` with minor changes to the returned object, such as the inclusion of the representative range in the column `keyval_range`. For this vignette, we’ll merge using the asymptotically exact harmonic mean p-value, which can also be used for merging dependent p-values (Wilson 2019). This approach tests H0: no $p$-value in the set of p-values is significant. When merging windows using the harmonic mean p-values, instead of values from a representative window, weighted averages for the expression and logFC estimates are returned using the weights $w_i = \frac{1}{p_i}$. A representative window, corresponding to the original window with the lowest p-value is returned. ```{r results-gr, eval = FALSE} results_gr <- mergeByHMP(fit_gr, inc_cols = "overlaps_ref", merge_within = 120) %>% addDiffStatus(drop = TRUE) arrange(results_gr, hmp)[1:5] ``` ## GRanges object with 5 ranges and 10 metadata columns: ## seqnames ranges strand | n_windows n_up n_down ## | ## [1] chr10 79257441-79258720 * | 30 3 0 ## [2] chr10 43689241-43690160 * | 21 2 0 ## [3] chr10 74008161-74008760 * | 13 2 0 ## [4] chr10 63858841-63859040 * | 3 1 0 ## [5] chr10 63859401-63859760 * | 7 1 0 ## overlaps_ref keyval_range logCPM logFC hmp ## ## [1] TRUE chr10:79257761-79257880 7.01563 1.82950 2.08382e-12 ## [2] TRUE chr10:43689681-43689800 6.49125 1.78743 8.37664e-09 ## [3] TRUE chr10:74008441-74008560 6.48379 1.81692 9.27730e-09 ## [4] TRUE chr10:63858881-63859000 5.98361 1.75311 6.38262e-07 ## [5] TRUE chr10:63859601-63859720 5.98163 1.82260 1.02640e-06 ## hmp_fdr status ## ## [1] 1.49827e-09 Increased ## [2] 2.22346e-06 Increased ## [3] 2.22346e-06 Increased ## [4] 1.14728e-04 Increased ## [5] 1.47597e-04 Increased ## ------- ## seqinfo: 24 sequences from GRCh37 genome In the above, we returned 19 ranges which we might consider using the significance threshold $\alpha$ = 0.05. As is common, we can assess our results using an MA plot. However, given that testing is performed using sliding windows and merging windows using the harmonic mean will introduce a bias, we can check using the complete set of sliding windows and the final set of merged windows. Any bias present in our data will be visible when using the sliding windows, whilst our final results can be inspected using the merged windows. ```{r plot-ma-libsize, eval = FALSE} A <- fit_gr %>% as_tibble() %>% ggplot(aes(logCPM, logFC)) + geom_point(alpha = 0.6) + geom_smooth(se = FALSE, method = "loess") + geom_label( aes(label = label), data = . %>% summarise( n = dplyr::n(), logCPM = max(logCPM) - 0.15 * diff(range(logCPM)), logFC = max(logFC) - 0.05 * diff(range(logFC)), label = glue("{comma(n)}\nSliding Windows") ) ) + ylim(range(fit_gr$logFC)) + ggtitle("MA Plot: All Retained Sliding Windows") B <- results_gr %>% as_tibble() %>% mutate(`FDR < 0.05` = hmp_fdr < 0.05) %>% ggplot(aes(logCPM, logFC)) + geom_point(aes(colour = `FDR < 0.05`), alpha = 0.6) + geom_smooth(se = FALSE, method = "loess") + geom_label_repel( aes(label = range, colour = `FDR < 0.05`), data = . %>% dplyr::filter(hmp_fdr == min(hmp_fdr)), show.legend = FALSE ) + geom_label( aes(label = label), data = . %>% summarise( n = dplyr::n(), logCPM = max(logCPM) - 0.15 * diff(range(logCPM)), logFC = max(fit_gr$logFC) - 0.05 * diff(range(fit_gr$logFC)), label = glue("{comma(n)}\nMerged Windows") ) ) + scale_colour_manual(values = c("black", "red")) + ylim(range(fit_gr$logFC)) + ggtitle("MA Plot: Merged Windows") A + B + plot_annotation(tag_levels = "A") ``` ```{r add-ma-plot, echo=FALSE, eval=TRUE, fig.cap="*MA Plots using A) all sliding windows (after filtering before merging) and B) the final set of merged windows. The most highly ranked merged window is also labelled in the set of merged windows. Blue lines represent a loess curve through the data.*"} knitr::include_graphics("differential_signal_sliding_files/figure-gfm/plot-ma-libsize-1.png") ``` A particularly beneficial feature of this approach is that the final ranges will be of highly variable width, with this select region of chromosome 10 producing merged windows ranging from 120bp to 18.88kb, as may be expected for H3K27ac signal. ## Alternative Normalisation Approaches Using Library Size normalisation, as above, is a conservative approach and other methods such as RLE (Love, Huber, and Anders 2014) or TMM (Robinson and Oshlack 2010) may be appropriate. However, these methods assume that signal distributions are drawn from the same distributions across all samples, which unlike most RNA-Seq, is not always able to be safely assumed for ChIP-Seq data. An example may be when an transcription factor is primarily cytoplasmic in one treatment group, before being translocated to the nucleus in another treatment. The Androgen Receptor (*AR*) is a common example of this behaviour. Whilst H3K27ac signal should remain approximately constant between these two treatments, we can easily check for differences in data distributions as well as any GC-bias. The package `quantro` (Hicks and Irizarry 2015) provides a useful test for any difference in medians as well as differences in the underlying distributions. ```{r qtest, eval = FALSE} set.seed(100) qtest <- assay(filtcounts, "counts") %>% quantro(groupFactor = filtcounts$treatment, B = 1e3) qtest ``` ## quantro: Test for global differences in distributions ## nGroups: 2 ## nTotSamples: 6 ## nSamplesinGroups: 3 3 ## anovaPval: 0.49809 ## quantroStat: 0.19728 ## quantroPvalPerm: 0.68 Here, no differences were evident across median values (ANOVA p = 0.498) or between distributions (p = 0.68) and as such, TMM/RLE normalisation across all samples may be appropriate. ### TMM Normalisation To perform an analysis using TMM-normalisation, we can simply specify this using the argument `norm = "TMM"` when calling `fitAssayDiff()` ```{r fit-tmm, eval = FALSE} tmm_gr <- fitAssayDiff( filtcounts, design = X, fc = 1.2, norm = "TMM", asRanges = TRUE ) tmm_results <- mergeByHMP(tmm_gr, inc_cols = "overlaps_ref", merge_within = 120) %>% addDiffStatus(drop = TRUE) table(tmm_results$status) ``` ## ## Unchanged Increased ## 698 21 As might be expected, the results are highly concordant, with TMM-normalisation providing a moderate improvement in statistical power, returning 21 windows with evidence of differential signal, instead of the initial 19. Any of the normalisation methods taken by `edgeR::calcNormFactors()` can be used here. If a difference in distributions is found between groups, the normalisation step can be set to only occur within treatment groups, by passing the grouping column name to the `groups =` argument. However, this may exaggerate differences between groups and may introduce false positives, so should only be used when compelling evidence exists and the underlying biology supports this. Checking MA-plots shows that the very slight negative bias in higher-signal windows when using library size normalisation, as seen above, was lessened after TMM normalisation. ```{r plot-ma-tmm, eval = FALSE} tmm_gr %>% as_tibble() %>% ggplot(aes(logCPM, logFC)) + geom_point(alpha = 0.6) + geom_smooth(se = FALSE, method = "loess") + ggtitle("MA Plot: TMM Normalisation") ``` If choosing more nuanced normalisation strategies, passing any offset matrices can be simply done using the `offset` argument, which will over-ride any other normalisation settings. # Mapping of Windows ## Mapping to Genes Once the changes in signal for our given ChIP target have been determined, a common next step is to assess which genes are likely to be impacted. Whilst no definitive, single methodology exists for this process, the function `mapByFeature()` offers an intuitive approach, taking into account any previously defined regulatory features. These regulatory features may be defined by simple proximity to TSS regions, by histone marks, downloaded from external repositories or any other possibility. Whilst these features can improve the precision of mapping, even without these this function can still enable a useful assignment of target gene to binding event. The process undertaken inside `mapByFeature()` is a sequential checking of each range’s association with regulatory features and the most likely target as a result. These steps are: 1. **Check for any HiC interactions** - All genes which directly overlap an interaction anchor are considered part of the regulatory network for that interaction, and as such, all genes associated with both anchors are assigned to a peak which overlaps a HiC Interaction 2. **Check for any overlaps with a promoter** - All genes regulated by that promoter are assigned as regulatory targets. By default, this is by direct promoter/gene overlap (`prom2gene = 0`) 3. **Check for any overlaps with an enhancer** - Peaks which overlap an enhancer are assigned to *all* genes within the distance specified by `enh2gene` (default = 100kb) 4. **Check for genes with no previous mappings** - Peaks *with no previous mappings* are assigned to all directly overlapping genes, or the *nearest* gene within a specified distance (default `gr2gene` = 100kb) If no promoters, enhancers or long-range interactions are supplied, all genes will be simply mapped to ChIP-Seq regions using step 4. A GTF of annotations can be used, with the following being the genes, transcripts and exons from chr10, taken from [Gencode v43](https://www.gencodegenes.org/human/release_43lift37.html). Whilst this example relies on this format, any additional resource such as an `EnsDb` or `TxDb` object could easily be used. Here, we’ll load the data then split into a `GRangesList` with separate elements for genes, transcripts and exons. ```{r gencode, eval = FALSE} gencode <- here("data/gencode.v43lift37.chr10.annotation.gtf.gz") %>% import.gff() %>% filter_by_overlaps(GRanges("chr10:42354900-100000000")) %>% split(.$type) seqlevels(gencode) <- seqlevels(sq) seqinfo(gencode) <- sq ``` Now we’ve loaded our gene annotations at multiple levels, we can easily define Transcription Start Sites (TSS) using transcript definitions and the function `reduceMC()`. This function is an extension of `GenomicRanges::reduce()` which retains all supplied columns in the `mcols()` element by collapsing these into `CompressedList` columns, and simplifying to vectors where possible. ```{r tss, eval = FALSE} tss <- gencode$transcript %>% resize(width = 1, fix = "start") %>% select(gene_id, ends_with("name")) %>% reduceMC(min.gapwidth = 0) ``` Checking the TSS for *PPIF* shows we have one site which 3 transcripts start at, along with two more unique to specific transcripts. ```{r subset-tss, eval = FALSE} subset(tss, vapply(gene_name, function(x) "PPIF" %in% x, logical(1))) ``` ## GRanges object with 3 ranges and 3 metadata columns: ## seqnames ranges strand | gene_id gene_name ## | ## [1] chr10 81107225 + | ENSG00000108179.14_6 PPIF ## [2] chr10 81107256 + | ENSG00000108179.14_6 PPIF ## [3] chr10 81107414 + | ENSG00000108179.14_6 PPIF ## transcript_name ## ## [1] PPIF-205,PPIF-203,PPIF-201 ## [2] PPIF-204 ## [3] PPIF-202 ## ------- ## seqinfo: 25 sequences from GRCh37 genome We can also add the TSS overlap status to our set of results ```{r overlaps-any, eval = FALSE} tmm_results$tss <- overlapsAny(tmm_results, tss) ``` Next we’ll want to map our regions to genes. To use the methods available in `mapByFeature()` let’s define a set of promoters. First we’ll create a promoter for each transcript, then we can merge any overlapping promoters using `reduceMC()`. (Setting `simplify = FALSE` avoids an issue in this subset of genes where one gene has multiple IDs, and helps retain a 1:1 mapping between IDs and genes.) Importantly, this will give a set of promoters which are unique and non-overlapping but with variable width and with the complete set of transcripts that may be regulated by the promoter listed. As can be seen, 4 potential promoter regions are found for *CRTAC1* ```{r promoters, eval = FALSE} promoters <- gencode$transcript %>% select(gene_id, ends_with("name")) %>% promoters(upstream = 2500, downstream = 500) %>% reduceMC(simplify = FALSE) tail(promoters) ``` ## GRanges object with 6 ranges and 3 metadata columns: ## seqnames ranges strand | ## | ## [1] chr10 99609056-99612055 - | ## [2] chr10 99635155-99638154 - | ## [3] chr10 99643500-99646805 - | ## [4] chr10 99695536-99698535 - | ## [5] chr10 99770595-99773594 - | ## [6] chr10 99789879-99793085 - | ## gene_id gene_name ## ## [1] ENSG00000227356.2_10 GOLGA7B-DT ## [2] ENSG00000265398.1 AL139239.1 ## [3] ENSG00000095713.14_13,ENSG00000095713.14_13 CRTAC1,CRTAC1 ## [4] ENSG00000095713.14_13 CRTAC1 ## [5] ENSG00000095713.14_13 CRTAC1 ## [6] ENSG00000095713.14_13,ENSG00000095713.14_13 CRTAC1,CRTAC1 ## transcript_name ## ## [1] GOLGA7B-DT-201 ## [2] AL139239.1-201 ## [3] CRTAC1-205,CRTAC1-206 ## [4] CRTAC1-204 ## [5] CRTAC1-201 ## [6] CRTAC1-203,CRTAC1-202 ## ------- ## seqinfo: 25 sequences from GRCh37 genome ```{r map-by-feature, eval = FALSE} tmm_results <- mapByFeature( tmm_results, genes = gencode$gene, prom = select(promoters, gene_id, gene_name) ) tmm_results %>% filter(hmp_fdr < 0.05) %>% arrange(hmp) ``` ## GRanges object with 21 ranges and 13 metadata columns: ## seqnames ranges strand | n_windows n_up n_down ## | ## [1] chr10 79257441-79258720 * | 30 3 0 ## [2] chr10 43689241-43690160 * | 21 2 0 ## [3] chr10 74008161-74008760 * | 13 2 0 ## [4] chr10 63858841-63859040 * | 3 1 0 ## [5] chr10 63859401-63859760 * | 7 1 0 ## ... ... ... ... . ... ... ... ## [17] chr10 79307081-79307960 * | 20 2 0 ## [18] chr10 74014401-74015360 * | 22 3 0 ## [19] chr10 90615361-90615640 * | 4 2 0 ## [20] chr10 63696921-63697240 * | 6 3 0 ## [21] chr10 74878001-74878360 * | 7 3 0 ## overlaps_ref keyval_range logCPM logFC hmp ## ## [1] TRUE chr10:79257761-79257880 7.02793 1.85621 6.20766e-13 ## [2] TRUE chr10:43689681-43689800 6.51086 1.81157 4.45187e-09 ## [3] TRUE chr10:74008441-74008560 6.49275 1.84270 5.07996e-09 ## [4] TRUE chr10:63858881-63859000 5.99257 1.77724 4.03575e-07 ## [5] TRUE chr10:63859601-63859720 5.99366 1.84777 6.37775e-07 ## ... ... ... ... ... ... ## [17] TRUE chr10:79307441-79307560 6.45271 1.15247 0.000563495 ## [18] TRUE chr10:74015201-74015320 6.14440 1.26391 0.000592892 ## [19] TRUE chr10:90615361-90615480 5.97947 1.14308 0.000804490 ## [20] TRUE chr10:63697121-63697240 6.00058 1.11473 0.001074487 ## [21] TRUE chr10:74878201-74878320 6.19539 1.05769 0.001437350 ## hmp_fdr status tss gene_id gene_name ## ## [1] 4.46331e-10 Increased FALSE ENSG00000156113.25_17 KCNMA1 ## [2] 1.21750e-06 Increased FALSE ENSG00000198915.12_14 RASGEF1A ## [3] 1.21750e-06 Increased FALSE ENSG00000166295.9_6 ANAPC16 ## [4] 7.25426e-05 Increased FALSE ENSG00000150347.17_12 ARID5B ## [5] 9.17121e-05 Increased FALSE ENSG00000150347.17_12 ARID5B ## ... ... ... ... ... ... ## [17] 0.0236827 Increased FALSE ENSG00000156113.25_17 KCNMA1 ## [18] 0.0236827 Increased FALSE ENSG00000289506.2_2 ENSG00000289506 ## [19] 0.0304436 Increased FALSE ENSG00000152766.6_8 ANKRD22 ## [20] 0.0386278 Increased FALSE ENSG00000150347.17_12 ARID5B ## [21] 0.0492121 Increased FALSE ENSG00000166321.14_7 NUDT13 ## ------- ## seqinfo: 25 sequences from GRCh37 genome So now we have our regions showing differential signal, along with which gene they are most likely to be impacting, and whether they directly overlap the TSS. This can all be summarised into a single MA-plot for presentation of final results. ```{r plot-ma-mapped, eval = FALSE} status_colours <- c(Unchanged = "grey70", Increased = "red3", Decreased = "royalblue") tmm_results %>% as_tibble() %>% ggplot(aes(logCPM, logFC, colour = status, shape = tss)) + geom_point() + geom_label_repel( aes(label = label), data = . %>% arrange(hmp) %>% dplyr::slice(1:20) %>% mutate( label = vapply(gene_name, paste, character(1), collapse = "; ") %>% str_trunc(30) ), fill = "white", alpha = 0.6, show.legend = FALSE ) + scale_colour_manual(values = status_colours) + scale_shape_manual(values = c(19, 21)) ``` ```{r add-ma-mapped-plot, eval = TRUE, echo = FALSE, fig.cap = "*MA-plot with the top 20 regions by significance labelled according to the most likely gene influenced by the detected signal. Regions which directly overlap a TSS are shown as empty circles*"} knitr::include_graphics("differential_signal_sliding_files/figure-gfm/plot-ma-mapped-1.png") ``` ## Mapping to Regions In addition to which gene is likely to be directly impacted by the detected signal, knowing which type of regulatory region signal was observed in is another question commonly asked of ChIP-Seq data. The function `deinfeRegions()` takes, gene-, transcript- and exon-level information from an annotation source and creates unique region annotations across the genome. These are for in hierarchical order as: 1) Promoters, 2) Upstream Promoters, 3) Exons, 4) Introns, 5) Proximal Intergenic and, 6 Distal Intergenic regions. ```{r define-regions, eval = FALSE} regions <- defineRegions( genes = gencode$gene, transcripts = gencode$transcript, exons = gencode$exon ) regions ``` ## GRangesList object of length 6: ## $promoter ## GRanges object with 1491 ranges and 5 metadata columns: ## seqnames ranges strand | region ## | ## [1] chr10 42644498-42647497 * | Promoter (-2500/+500) ## [2] chr10 42678287-42681286 * | Promoter (-2500/+500) ## [3] chr10 42702938-42705937 * | Promoter (-2500/+500) ## [4] chr10 42729924-42732923 * | Promoter (-2500/+500) ## [5] chr10 42735669-42738668 * | Promoter (-2500/+500) ## ... ... ... ... . ... ## [1487] chr10 99695536-99698535 * | Promoter (-2500/+500) ## [1488] chr10 99770595-99773594 * | Promoter (-2500/+500) ## [1489] chr10 99789879-99793085 * | Promoter (-2500/+500) ## [1490] chr10 99891881-99894913 * | Promoter (-2500/+500) ## [1491] chr10 99920140-99923543 * | Promoter (-2500/+500) ## gene_id gene_name ## ## [1] ENSG00000229485.1_5 KSR1P1 ## [2] ENSG00000237592.2_5 IGKV1OR10-1 ## [3] ENSG00000271650.1_7 ENSG00000271650 ## [4] ENSG00000264398.1 AL031601.1 ## [5] ENSG00000290458.1_2 ENSG00000290458 ## ... ... ... ## [1487] ENSG00000095713.14_13 CRTAC1 ## [1488] ENSG00000095713.14_13 CRTAC1 ## [1489] ENSG00000095713.14_13 CRTAC1 ## [1490] ENSG00000166024.14_8 R3HCC1L ## [1491] ENSG00000166024.14_8 R3HCC1L ## transcript_id ## ## [1] ENST00000446298.1_2 ## [2] ENST00000442306.2_2 ## [3] ENST00000605702.1_2 ## [4] ENST00000580460.1 ## [5] ENST00000622823.4_3 ## ... ... ## [1487] ENST00000413387.5_4 ## [1488] ENST00000309155.3_5 ## [1489] ENST00000370597.8_11,ENST00000370591.6_9 ## [1490] ENST00000370584.7_7,ENST00000298999.8_8 ## [1491] ENST00000370586.6_5,ENST00000613938.4_5,ENST00000612478.4_2,... ## transcript_name ## ## [1] KSR1P1-201 ## [2] IGKV1OR10-1-201 ## [3] ENST00000605702 ## [4] AL031601.1-201 ## [5] ENST00000622823 ## ... ... ## [1487] CRTAC1-204 ## [1488] CRTAC1-201 ## [1489] CRTAC1-203,CRTAC1-202 ## [1490] R3HCC1L-203,R3HCC1L-201 ## [1491] R3HCC1L-204,R3HCC1L-206,R3HCC1L-205,... ## ------- ## seqinfo: 25 sequences from GRCh37 genome ## ## ... ## <5 more elements> We can now use the function `bestOverlap()` to assign each window with ChIP-Seq signal to the region which it shows the greatest overlap, setting this as a factor which reflects the hierarchy with which the regions were defined. ```{r bet-overlap, eval = FALSE} tmm_results$bestOverlap <- bestOverlap(tmm_results, regions) %>% factor(levels = names(regions)) filter(tmm_results, hmp_fdr < 0.05, bestOverlap == "promoter") ``` ## GRanges object with 2 ranges and 14 metadata columns: ## seqnames ranges strand | n_windows n_up n_down ## | ## [1] chr10 76779561-76780200 * | 14 4 0 ## [2] chr10 79623321-79635960 * | 314 13 0 ## overlaps_ref keyval_range logCPM logFC hmp ## ## [1] TRUE chr10:76779721-76779840 6.21638 1.244849 0.000194483 ## [2] TRUE chr10:79630321-79630440 7.71543 0.906496 0.000100434 ## hmp_fdr status tss ## ## [1] 0.00932221 Increased FALSE ## [2] 0.00515801 Increased TRUE ## gene_id gene_name ## ## [1] ENSG00000156650.14_17 KAT6B ## [2] ENSG00000204049.1_8,ENSG00000151208.17_11 ENSG00000204049,DLG5 ## bestOverlap ## ## [1] promoter ## [2] promoter ## ------- ## seqinfo: 25 sequences from GRCh37 genome Given the region column contains nicely formateed output, we can also use this ```{r reg-as-factor, eval = FALSE} reg_levels <- vapply(regions, function(x) x$region[1], character(1)) tmm_results$bestOverlap <- tmm_results %>% bestOverlap(unlist(regions), var = "region") %>% factor(levels = reg_levels) ``` # Visualisation of Results ## Pie Charts Now that we’ve assigned each window to a genomic regions, we can use `plotPie()` to view the distribution of ChIP-Seq signal within these regions. First we’ll define a consistent colour palette for these regions in all plots. ```{r reg-colours, eval = FALSE} region_colours <- hcl.colors(length(reg_levels), "Viridis", rev = TRUE) names(region_colours) <- reg_levels ``` Labels produced using `plotPie()` use the `glue()` syntax and are able to access the summarised data using `N` as the overall total number of sites. as well as `.data[[fill]]` to access data in the column selected to fill the pie segments. Totals and percentages within each segment are accessible using `n` and `p` respectively as these are columns within the internal `data.frame` formed when creating the plot. External formatting functions such as `comma()`, `percent()` or `str_to_title()` can also be put to good use here. ```{r plot-pie-labels, eval = FALSE} plotPie( tmm_results, fill = "bestOverlap", min_p = 0.05, cat_glue = "{.data[[fill]]}\n{n}\n({percent(p, 0.1)})" ) + scale_fill_manual(values = region_colours) ``` ```{r add-pit, eval = TRUE, echo = FALSE, fig.cap = "*Pie chart showing customised labels. Here the regions have been modified to use title case in the labels, but without modifying the underlying data.*"} knitr::include_graphics("differential_signal_sliding_files/figure-gfm/plot-pie-labels-1.png") ``` In addition to highly customisable labels, segments can be scaled by width to produce segments representing the overlaps in kb instead of counts, by setting `scale_by = "width"` ## Split Donut Charts Pie charts can be further extended into Split-Donut charts, with data in two rings. The default settings can be used to add the status of each window to the overlapping regions. ```{r plot-donut-default, eval = FALSE} plotSplitDonut( tmm_results, inner = "bestOverlap", outer = "status", inner_palette = region_colours ) ``` ```{r add-donut, eval = TRUE, echo=FALSE, fig.cap = "*Split-Donut chart showing status with overlapping region*"} knitr::include_graphics("differential_signal_sliding_files/figure-gfm/plot-donut-default-1.png") ``` A key advantage of `plotSplitDonut()` is the ability to provide separate colour palettes for the inner & outer rings. Once again these are heavily customisable using the `glue()` syntax for labels and exploding key segments of interest. ```{r plot-donut-customised, eval = FALSE} plotSplitDonut( tmm_results, inner = "bestOverlap", outer = "status", inner_palette = region_colours, outer_palette = status_colours, inner_glue = "{str_replace_all(.data[[inner]], ' ', '\n')}\nn = {comma(n)}\n{percent(p,0.1)}", outer_glue = "{.data[[outer]]}\n{str_replace_all(.data[[inner]], ' ', '\n')}\nn = {n}", explode_outer = "(In|De)creased", explode_r = 0.3, outer_label = "text", outer_min_p = 0, outer_max_p = 0.02 ) ``` ```{r add-donut-customised, eval=TRUE, echo=FALSE, fig.cap = "*Split-Donut chart exploding key regions and customising labels.*"} knitr::include_graphics("differential_signal_sliding_files/figure-gfm/plot-donut-customised-1.png") ``` ## Coverage Plots In order to show our changed signal in context we can show the coverage using a `BigWigFileList` and the function `plotHFGC()`, which stands for plot **H**iC, **F**eatures, **G**enes and **C**overage, with tracks appearing in that order. `plotHFGC()` wraps multiple calls to the package `Gviz` (Hahne and Ivanek 2016) to simplify repetitive plot generation but retain significant flexibility. All tracks are options, with the minimal requirements being cytogenetic bands along with a specified GenomicRange. However, this does restrict plotting to the “Standard Chromosomes”. Coverage is produced using BigWig Files, passed to `plotHFGC()` as either a `BigWigFileList` (for separate tracks) or a list of `BigWigFileList` objects, for multiple tracks with individual samples overlaid. ```{r bwfl, eval = FALSE} bwfl <- here::here( "data", "H3K27ac", glue("{levels(samples$treatment)}_cov_chr10.bw") ) %>% BigWigFileList() %>% setNames(levels(samples$treatment)) ``` We can also include cytogenetic bands for these coverage plots, some of which are provided with `extraChIPs`. ```{r cytobands} data("grch37.cytobands") head(grch37.cytobands) ``` Let’s check the coverage across the most highly-ranked region with default settings. ```{r gr, eval = FALSE} gr <- filter(tmm_results, hmp_fdr < 0.05, str_detect(bestOverlap, '^Prom'))[1] ``` The most important option to consider when showing coverage is whether to display tracks individually or as overlapping tracks. Passing a BigWigFileList will plot each file on a separate track by default, with y-axis limits able to be specified as consistent across all tracks. Alternatively, by passing the coverage as a list of BigWigFileList objects, each list element will be drawn as a single track with coverage overlaid. For one or more overlaid tracks, colours will need to now be provided in a list with matching structure. The code for separate tracks would be: (not shown) ```{r plot-hfgc, eval = FALSE} plotHFGC(gr, cytobands = grch37.cytobands, coverage = bwfl) ``` Overlaying coverage onto a single track can instead be highly informative and enables the visualisation of multiple ChIP targets. (Plot not shown in vignette) ```{r cov-list, eval = FALSE} cov_list <- list(H3K27ac = bwfl) cov_colour <- list(H3K27ac = treat_colours[levels(samples$treatment)]) plotHFGC( gr, cytobands = grch37.cytobands, coverage = cov_list, linecol = cov_colour, zoom = 50, rotation.title = 90 ) ``` By default, the specified range will also be outlined (i.e. highlighted) with a blue rectangle which is able to be turned off and customised as the user sees fit. ### Adding Annotations To Coverage An indication of which regions are associated with increased or decreased ChIP signal can also be a useful annotation to add to plots such as the above. Using the status from our differential signal testing, we can create a `GRangesList` with the various regions annotated as Increased, Unchanged etc, Similar to the features track, where the names of `GRangesList` elements denote the different feature types, able to then assigned a colour, coverage annotation tracks follow these same rules. For each coverage track being annotated, a `GRangesList` object can denote the ranges which can be assigned different colours. Annotation tracks should be passed to `plotHFGC()` using the same structure as the coverage tracks, so here if we pass the annotations as a list with the single element `H3K27ac` these annotations will be added directly above the coverage track for H3K27ac signal. ```{r cov-annot, eval = FALSE} cov_annot <- splitAsList(tmm_results, tmm_results$status) %>% endoapply(granges) cov_annot <- list(H3K27ac = cov_annot) ``` ### Displaying Genes and Features Next we might like to add gene models to provide the regulatory context. We’ll use the exons element of our `gencode` object to create the format used by the defaults of the `GeneRegionTrack()` function, with all exons and transcripts annotated. ```{r gene-models, eval = FALSE} gene_models <- gencode$exon %>% select( type, gene = gene_id, exon = exon_id, transcript = transcript_id, symbol = gene_name ) ``` Another useful track to add might be some key features such as promoters and other annotated regions, as we have formed earlier in the workflow. Features must **always** be a `GRangesList`, with each element defining a different type of feature, as we already have in our `regions` object. For multiple feature tracks, a list of `GRangesList` objects can be passed. In our final plot, we’ll zoom out to the entire region covered by *KAT6B*, showing all transcripts, with out original region highlighted in blue. All merged windows for which a differential signal test was performed will be annotated with colours indicating the result for that window. As can be seen multiple H3K27ac signal regions appear to be responding to DHT treatment for *KAT6B* with a possible intronic enhancer and alternative promoter being most likely impacted, including the neighbouring region to our initial range. In contrast, the primary promoter appears unaffected by DHT treatment. ```{r plot-hfgc-with-annot, eval = FALSE} names(region_colours) <- names(regions) plotHFGC( gr, cytobands = grch37.cytobands, features = regions, featcol = region_colours, featstack = "dense", coverage = cov_list, linecol = cov_colour, genes = gene_models, genecol = "wheat", collapseTranscripts = FALSE, zoom = 350, shift = -1e5, rotation.title = 90, covsize = 10, genesize = 1, featsize = 5, annotation = cov_annot, annotcol = status_colours[c("Unchanged", "Increased")] ) ``` ```{r add-hfgc, eval=TRUE, echo=FALSE, fig.cap = "The addition of an annotation track for the coverage tracks shows which regions were retained during the analysis, as well as those which were considered as showing changed or unchanged signal"} knitr::include_graphics("differential_signal_sliding_files/figure-gfm/plot-hfgc-with-annot-1.png") ``` Plots are able to be tweaked considerably further via multiple parameters, however these basic approaches cover the core functionality of `plotHFGC()` for enabling simple & reproducible plotting across regions for multiple sites within a larger experiment. If long-range interactions are available, these can also be provided as a GenomicInteractions object, completing all available options for the HFGC components. # Session Info ```{r session-info} sessionInfo() ``` # References
Chen, Yunshun, Aaron A T Lun, and Gordon K Smyth. 2016. “From Reads to Genes to Pathways: Differential Expression Analysis of RNA-Seq Experiments Using Rsubread and the edgeR Quasi-Likelihood Pipeline.” *F1000Research* 5: 1438. .
Gandolfo, Luke C, and Terence P Speed. 2018. “RLE Plots: Visualizing Unwanted Variation in High Dimensional Data.” *PLoS One* 13 (2): e0191629.
Hahne, Florian, and Robert Ivanek. 2016. “Statistical Genomics: Methods and Protocols.” In, edited by Ewy Mathé and Sean Davis, 335–51. New York, NY: Springer New York. .
Hicks, Stephanie C, and Rafael A Irizarry. 2015. “Quantro: A Data-Driven Approach to Guide the Choice of an Appropriate Normalization Method.” *Genome Biol.* 16 (1): 117.
Law, Charity W, Yunshun Chen, Wei Shi, and Gordon K Smyth. 2014. “Voom: Precision Weights Unlock Linear Model Analysis Tools for RNA-seq Read Counts.” *Genome Biol.* 15 (2): R29.
Love, Michael I., Wolfgang Huber, and Simon Anders. 2014. “Moderated Estimation of Fold Change and Dispersion for RNA-Seq Data with DESeq2.” *Genome Biology* 15: 550. .
Lun, Aaron T L, and Gordon K Smyth. 2016. “Csaw: A Bioconductor Package for Differential Binding Analysis of ChIP-Seq Data Using Sliding Windows.” *Nucleic Acids Res.* 44 (5): e45.
Lund, Steven P, Dan Nettleton, Davis J McCarthy, and Gordon K Smyth. 2012. “Detecting Differential Expression in RNA-sequence Data Using Quasi-Likelihood with Shrunken Dispersion Estimates.” *Stat. Appl. Genet. Mol. Biol.* 11 (5).
McCarthy, Davis J, and Gordon K Smyth. 2009. “Testing Significance Relative to a Fold-Change Threshold Is a TREAT.” *Bioinformatics* 25 (6): 765–71.
Mölder, Felix, Kim Philipp Jablonski, Brice Letcher, Michael B Hall, Christopher H Tomkins-Tinch, Vanessa Sochat, Jan Forster, et al. 2021. “Sustainable Data Analysis with Snakemake.” *F1000Res.* 10 (January): 33.
Robinson, Mark D, and Alicia Oshlack. 2010. “A Scaling Normalization Method for Differential Expression Analysis of RNA-seq Data.” *Genome Biol.* 11 (3): R25.
Ross-Innes, Caryn S., Rory Stark, Andrew E. Teschendorff, Kelly A. Holmes, H. Raza Ali, Mark J. Dunning, Gordon D. Brown, et al. 2012. “Differential Oestrogen Receptor Binding Is Associated with Clinical Outcome in Breast Cancer.” *Nature* 481: –4. .
Wilson, Daniel J. 2019. “The Harmonic Mean *p*-Value for Combining Dependent Tests.” *Proc. Natl. Acad. Sci. U. S. A.* 116 (4): 1195–1200.