--- title: "Input data formats" author: - name: Kellen Cresswell affiliation: - &1 Department of Biostatistics, Virginia Commonwealth University, Richmond, VA - name: Mikhail Dozmorov affiliation: - *1 output: BiocStyle::html_document vignette: > %\VignetteIndexEntry{Input data formats} %\VignetteEncoding{UTF-8} %\VignetteEngine{knitr::rmarkdown} editor_options: chunk_output_type: console bibliography: pack_ref.bib --- ```{r set-options, echo=FALSE, cache=FALSE} options(stringsAsFactors = FALSE, warning = FALSE, message = FALSE) ``` # Introduction TADCompare is an R package for differential analysis of TAD boundaries. It is designed to work on a wide range of formats and resolutions of Hi-C data. TADCompare package contains four functions: `TADCompare`, `TimeCompare`, `ConsensusTADs`, and `DiffPlot`. `TADCompare` function allows for the identification of differential TAD boundaries between [two contact matrices](TADCompare.html#tadcompare). `TimeCompare` function takes a set of contact matrices, one matrix per time point, identifies TAD boundaries, and classifies [how they change over time](TADCompare.html#timecompare). `ConsensusTADs` function takes a list of TADs and identifies a consensus of TAD boundaries across all matrices using our [novel consensus boundary score](TADCompare.html#consensustads). `DiffPlot` allows for [visualization of TAD boundary differences](TADCompare.html#visualization) between two matrices. The required input includes matrices in sparse 3-column format, $n \times n$, or $n \times (n+3)$ formats. This vignette provides a complete overview of input data formats. # Getting Started ## Installation ```{r, eval = FALSE} BiocManager::install("TADCompare") ``` ```{r} library(dplyr) library(SpectralTAD) library(TADCompare) ``` # Working with different types of data ## Working with $n \times n$ matrices $n \times n$ contact matrices are most commonly associated with data coming from the Bing Ren lab (http://chromosome.sdsc.edu/mouse/hi-c/download.html). These contact matrices are square and symmetric with entry $ij$ corresponding to the number of contacts between region $i$ and region $j$. Below is an example of a $5 \times 5$ region of an $n \times n$ contact matrix derived from Rao et al. 2014 data, GM12878 cell line [@Rao:2014aa], chromosome 22, 50kb resolution. Note the symmetry around the diagonal - the typical shape of chromatin interaction matrix. The figure was created using the [pheatmap](https://cran.r-project.org/web/packages/pheatmap/index.html) package. ```{r echo = FALSE} data("rao_chr22_prim") row.names(rao_chr22_prim) <- colnames(rao_chr22_prim) <- format(as.numeric(row.names(rao_chr22_prim)), scientific = FALSE) coords <- 200:275 pheatmap::pheatmap(log10(rao_chr22_prim[coords, coords]), cluster_rows = FALSE, cluster_cols = FALSE) ``` ## Working with $n \times (n+3)$ matrices $n \times (n+3)$ matrices are commonly associated with the `TopDom` TAD caller (http://zhoulab.usc.edu/TopDom/). These matrices consist of an $n \times n$ matrix but with three additional leading columns containing the chromosome, the start of the region and the end of the region. Regions in this case are determined by the resolution of the data. The subset of a typical $n \times (n+3)$ matrix is shown below. ```{r echo = FALSE} coords <- 50:53 sub_mat <- data.frame(chr = "chr22", start = as.numeric(colnames(rao_chr22_prim[coords, coords])), end = as.numeric(colnames(rao_chr22_prim[coords, coords])) + 50000, rao_chr22_prim[coords, coords]) row.names(sub_mat) = NULL sub_mat ``` ## Working with sparse 3-column matrices Sparse 3-column matrices are matrices where the first and second columns refer to region $i$ and region $j$ of the chromosome, and the third column is the number of contacts between them. This style is becoming increasingly popular and is associated with raw data from Lieberman-Aiden lab (e.g., https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE63525), and is the data output produced by the Juicer tool [@Durand:2016aa]. 3-column matrices are handled internally in the package by converting them to $n \times n$ matrices using the [HiCcompare](https://bioconductor.org/packages/release/bioc/html/HiCcompare.html) package's `sparse2full()` function. The first 5 rows of a typical sparse 3-column matrix are shown below. ```{r echo = FALSE} data("rao_chr22_prim") head(HiCcompare::full2sparse(rao_chr22_prim)) ``` ## Working with other data types ## Working with .hic files .hic files are a common form of files generally associated with the lab of Erez Lieberman-Aiden (http://aidenlab.org/data.html). To use .hic files you must use the following steps. 1. Download `straw` from https://github.com/aidenlab/straw/ and follow instalation instructions. 2. Download .hic data files. Here, we use data from Rao 2014 and download them using the following commands: `wget ftp://ftp.ncbi.nlm.nih.gov/geo/series/GSE63nnn/GSE63525/suppl/GSE63525_GM12878_insitu_primary_30.hic` `wget ftp://ftp.ncbi.nlm.nih.gov/geo/series/GSE63nnn/GSE63525/suppl/GSE63525_GM12878_insitu_replicate.hic` 3. Extract chromosome 22 at 50kb resolution with no normalization: `./straw NONE GSE63525_GM12878_insitu_primary_30.hic 22 22 BP 50000 > primary.chr22.50kb.txt` `./straw NONE GSE63525_GM12878_insitu_replicate_30.hic 22 22 BP 50000 > replicate.chr22.50kb.txt` 4. Analyze normally: ```{r, eval=FALSE} #Read in data primary = read.table('primary.chr22.50kb.txt', header = FALSE) replicate = read.table('replicate.chr22.50kb.txt', header = FALSE) #Run TADCompare tad_diff=TADCompare(primary, replicate, resolution=50000) ``` ## Working with .cool files Users can also find TADs from data output by `cooler` (http://cooler.readthedocs.io/en/latest/index.html) and HiC-Pro (https://github.com/nservant/HiC-Pro) with minor pre-processing using the [HiCcompare](https://bioconductor.org/packages/release/bioc/html/HiCcompare.html) package. The cooler software can be downloaded from https://mirnylab.github.io/cooler/. A catalog of popular HiC datasets can be found at ftp://cooler.csail.mit.edu/coolers. We can extract chromatin interaction data from .cool files using the following steps: 1. Follow instructions to install the cooler software, https://mirnylab.github.io/cooler/ 2. Download the first contact matrix wget ftp://cooler.csail.mit.edu/coolers/hg19/Zuin2014-HEK293CtcfControl-HindIII-allreps-filtered.50kb.cool 3. Convert the first matrix to a text file using `cooler dump --join Zuin2014-HEK293CtcfControl-HindIII-allreps-filtered.50kb.cool > Zuin.HEK293.50kb.Control.txt` 4. Download the second contact matrix wget ftp://cooler.csail.mit.edu/coolers/hg19/Zuin2014-HEK293CtcfDepleted-HindIII-allreps-filtered.50kb.cool 5. Convert the matrix to a text file using `cooler dump --join Zuin2014-HEK293CtcfDepleted-HindIII-allreps-filtered.50kb.cool > Zuin.HEK293.50kb.Depleted.txt` 6. Run the code below ```{r eval = FALSE} # Read in data cool_mat1 <- read.table("Zuin.HEK293.50kb.Control.txt") cool_mat2 <- read.table("Zuin.HEK293.50kb.Depleted.txt") # Convert to sparse 3-column matrix using cooler2sparse from HiCcompare sparse_mat1 <- HiCcompare::cooler2sparse(cool_mat1) sparse_mat2 <- HiCcompare::cooler2sparse(cool_mat2) # Run TADCompare diff_tads = lapply(names(sparse_mat1), function(x) { TADCompare(sparse_mat1[[x]], sparse_mat2[[x]], resolution = 50000) }) ``` ## Working with HiC-Pro files HiC-Pro data is represented as two files, the `.matrix` file and the `.bed` file. The `.bed` file contains four columns (chromosome, start, end, ID). The `.matrix` file is a three-column matrix where the 1^st^ and 2^nd^ columns contain region IDs that map back to the coordinates in the bed file, and the third column contains the number of contacts between the two regions. In this example we analyze two matrix files `sample1_100000.matrix` and `sample2_100000.matrix`and their corresponding bed files `sample1_100000_abs.bed` and `sample2_100000_abs.bed`. We do not include HiC-Pro data in the package, so these serve as placeholders for the traditional files output by HiC-Pro. The steps for analyzing these files is shown below: ```{r eval = FALSE} # Read in both files mat1 <- read.table("sample1_100000.matrix") bed1 <- read.table("sample1_100000_abs.bed") # Matrix 2 mat2 <- read.table("sample2_100000.matrix") bed2 <- read.table("sample2_100000_abs.bed") # Convert to modified bed format sparse_mats1 <- HiCcompare::hicpro2bedpe(mat1,bed1) sparse_mats2 <- HiCcompare::hicpro2bedpe(mat2,bed2) # Remove empty matrices if necessary # sparse_mats$cis = sparse_mats$cis[sapply(sparse_mats, nrow) != 0] # Go through all pairwise chromosomes and run TADCompare sparse_tads = lapply(1:length(sparse_mats1$cis), function(z) { x <- sparse_mats1$cis[[z]] y <- sparse_mats2$cis[[z]] #Pull out chromosome chr <- x[, 1][1] #Subset to make three column matrix x <- x[, c(2, 5, 7)] y <- y[, c(2, 5, 7)] #Run SpectralTAD comp <- TADCompare(x, y, resolution = 100000) return(list(comp, chr)) }) # Pull out differential TAD results diff_res <- lapply(sparse_tads, function(x) x$comp) # Pull out chromosomes chr <- lapply(sparse_tads, function(x) x$chr) # Name list by corresponding chr names(diff_res) <- chr ``` ## Effect of matrix type on runtime The type of matrix input into the algorithm can affect runtimes for the algorithm. $n \times n$ matrices require no conversion and are the fastest. Meanwhile, $n \times (n+3)$ matrices take slightly longer to run due to the need to remove the first 3 columns. Sparse 3-column matrices have the highest runtimes due to the complexity of converting them to an $n \times n$ matrix. The times are summarized below, holding all other parameters constant. ```{r message=FALSE} library(microbenchmark) # Reading in the second matrix data("rao_chr22_rep") # Converting to sparse prim_sparse <- HiCcompare::full2sparse(rao_chr22_prim) rep_sparse <- HiCcompare::full2sparse(rao_chr22_rep) # Converting to nxn+3 # Primary prim_n_n_3 <- data.frame(chr = "chr22", start = as.numeric(colnames(rao_chr22_prim)), end = as.numeric(colnames(rao_chr22_prim))+50000, rao_chr22_prim) # Replicate rep_n_n_3 <- data.frame(chr = "chr22", start = as.numeric(colnames(rao_chr22_rep)), end = as.numeric(colnames(rao_chr22_rep))+50000, rao_chr22_rep) # Defining each function # Sparse sparse <- TADCompare(cont_mat1 = prim_sparse, cont_mat2 = rep_sparse, resolution = 50000) # NxN n_by_n <- TADCompare(cont_mat1 = prim_sparse, cont_mat2 = rep_sparse, resolution = 50000) # Nx(N+3) n_by_n_3 <- TADCompare(cont_mat1 = prim_n_n_3, cont_mat2 = rep_n_n_3, resolution = 50000) # Benchmarking different parameters bench <- microbenchmark( # Sparse sparse <- TADCompare(cont_mat1 = prim_sparse, cont_mat2 = rep_sparse, resolution = 50000), # NxN n_by_n <- TADCompare(cont_mat1 = rao_chr22_prim, cont_mat2 = rao_chr22_rep, resolution = 50000), # Nx(N+3) n_by_n_3 <- TADCompare(cont_mat1 = prim_n_n_3, cont_mat2 = rep_n_n_3, resolution = 50000), times = 5, unit = "s" ) summary_bench <- summary(bench) %>% dplyr::select(mean, median) rownames(summary_bench) <- c("sparse", "n_by_n", "n_by_n_3") summary_bench ``` The table above shows the mean and median of runtimes for different types of contact matrices measured in seconds. As we see, `TADCompare` is extremely fast irrespectively of the parameters. However, sparse matrix inputs will slow down the algorithm. This can become more apparent as the size of the contact matrices increase. # Session Info ```{r} sessionInfo() ``` # References