Correlating Histone Modification Patterns with Gene Expression Data During Hematopoiesis

Abstract

Hematopoietic stem cells (HSC) in mammals are an ideal system to study differentiation. While transcription factors (TFs) control the differentiation of HSCs to distinctive terminal blood cells, accumulating evidence suggests that chromatin structure and modifications constitute another critical layer of gene regulation. Recent genome-wide studies based on next-generation sequencing reveal that histone modifications are linked to gene expression and contribute to hematopoiesis. Here, we briefl y review the bioinformatics aspects for ChIP-Seq and RNA-Seq data analysis with applications to the epigenetic studies of hematopoiesis and provide a practical guide to several basic data analysis methods.

1 Introduction

Hematopoietic stem cells give rise to all blood cell types while maintaining a capacity of self-renewal [1]. It is known that a core set of transcription factors form a tightly regulated network that controls the spatiotemporal regulation of lineage-specific genes during hematopoiesis [2]. However, the precise molecular mechanisms governing HSC self-renewal and differentiation remain unclear.

There is an increasing awareness of epigenetic mechanisms in controlling the developmental hierarchy of hematopoietic system [3]. Genomic DNA within the eukaryotic nucleus is packaged with histones into a compact form called chromatin. The N-terminal tails of histones are subjected to a variety of posttranslational modifications including acetylation and methylation. Our previous works revealed that histone modifications correlate with gene activities, contribute to T-cell specificity/plasticity, and set stages for hematopoiesis [4–8].

Our knowledge about epigenetic regulation has been greatly advanced by recent development of genome-wide techniques such as ChIP-Seq and RNA-Seq. While ChIP-Seq charts genomic landscapes of transcription factor binding and histone modifications, RNA-Seq quantifies gene expressions. A combinational use of ChIP-Seq and RNA-Seq has been widely applied to the epigenetic studies of hematopoiesis [3]. Here, we have briefl y reviewed the bioinformatical steps commonly used to address epigenetic questions in hematopoiesis by using ChIP-Seq and RNA-Seq.

2 Materials

2.1 Genome Annotation

Genome annotations were downloaded from the online UCSC genome browser [9]:

1. Go to http://genome.ucsc.edu/cgi-bin/hgGateway, choose genome and assembly version.

2. Click “tools” at the top of the browser, and then choose “Table Browser.”

3. Specify the source of genome annotation. The “Table Browser” by default provides download of annotation for the UCSC known genes. One may choose other sources of annotation such as RefSeq and Ensembl from the “track” down-drop list.

4. Type a file name in the “output file” text filed, click the “get output” button, and save the annotation to a local drive.

2.2 Public ChIP-Seq/RNA-Seq Data

Raw fastq sequence files and/or processed BED6 files included in this review were downloaded from Gene Expression Omnibus (GEO: http://www.ncbi.nlm.nih.gov/geo/) [10].

2.3 Software

1. In-house C++ programs (executable files and source code are available upon request):

   1. Sam2Bed6_Bowtie2: Convert a sam file from Bowtie2 to a BED6 file.

   2. RemoveRedundantRead: Remove redundant reads from a BED6 file.

   3. GenerateRPBMBasedSummary: Convert a BED6 file to a BEDGraph file.

   4. AverageDensityAcrossGenes: Generate average read density across gene features.

   5. RPKMCalculator: Calculate RPKM at gene level.

   6. Cat_expr_file: Concatenate gene expression files from different samples.

   7. SortGeneAnnoByExpr: Sort genome annotations by gene expression level.

   8. DensityCalculatorPromoters: Calculate normalized read density for promoters.

2. Microsoft Excel: a spreadsheet editing software.

3. Microsoft Access: a database management system.

4. MeV (Multiple Experiment Viewer): a JAVA application for statistic analysis, clustering, and visualization of gene expression data [11].

3 Methods

We first reviewed several common steps for ChIP-Seq and RNA-Seq data analysis, including (1) inspection of data quality, (2) sequence alignment, (3) data visualization, (4) identification of read-enriched regions, and (5) quantification of gene expression. We then introduced a combinational usage of in-house C++ programs, Excel (Microsoft), Access (Microsoft), and MeV (Dana-Farber Cancer Institute) to correlate histone modification with gene expression. We showed examples by using public ChIP-Seq and RNA-Seq data sets generated for epigenetic studies on hematopoiesis.

3.1 Initial Data Quality Inspection

The first step in processing next-generation data is to check the sequence quality. FastQC (Babraham Institute) is a standalone Java application that outputs a summary statistics for fastq files (see Note 1). It issues warnings for bases with low quality, for bases with abnormal sequence contents, and for primer/adapter contaminations.

3.2 Sequence Alignment

There are dozens of algorithms to map short reads to a reference genome. Low-quality bases may be clipped off to increase mapping rate (the % of reads mapped to the reference; see Note 2). To minimize ambiguity, reads that are mapped to multiple positions (called multireads) are frequently discarded. Consequently, read enrichments within repetitive regions are underestimated; repetitive sequences are found in constitutive heterochromatin and segmental duplications, both with functions implicated in hematopoiesis [12]. The exclusion of multireads will also underestimate the expression of genes with multiples copies. To address this challenge, several sophisticated probabilistic methods with user-friendly tools have been proposed, for instance, as described in [13].

3.3 Data Visualization

A visualization of the ChIP-Seq and RNA-Seq data in a genome browser such as the UCSC genome browser [9] helps to further inspect the data quality. A local mirror of the UCSC genome browser may be installed. But its maintenance usually requires substantial computational resources. Thus, for a small number of samples, the “custom track” feature from the online UCSC genome browser is recommended (see Note 3). An example is illustrated below about how to upload a ChIP-Seq data set (for transcription factor GATA1) and an RNA-Seq data set (for human CD36+ erythrocyte precursor cells 14) to the online UCSC genome browser:

1. Download the two data sets (GSM651547 and GSM651555) from GEO. The files are in sra format (short read archive). A sub-module called “fastq-dump” from the SRA Toolkit (NCBI) extracts fastq files from sra files (see Note 4).

2. Map the sequences to human genome (hg18) by using Bowtie2, which reports the alignments in sam format [15]. Bowtie2 by default reports the best hit for multireads.

3. Convert sam to BED6 file by using the in-house C++ program “Sam2Bed6_Bowtie2” (see Note 5).

4. Remove redundant reads with “RemoveRedundantRead” (see Note 6). Since the probability that two reads are mapped to the same genomic position is small for ChIP-Seq data, only one read is retained for each genomic position to minimize biases from amplification (see Note 7).

5. Generate genomic distribution of reads with “Generate RPBMBasedSummary.” The program outputs a BEDGraph file, with the first three columns denoting chromosome, starting position and ending position, and the last column denoting the number of reads mapped to the genomic region (see Note 8).

6. Customize the BEDGraph file for the UCSC genome browser. The BEDGraph file acceptable by the online UCSC genome browser reserves the first line for parameters of the track (http://genome.ucsc.edu/goldenPath/help/bedgraph.html). One needs to edit the BEDGraph file from step 5 to accommodate this requirement (see Note 9).

7. Upload the BEDGraph files to the online UCSC genome browser: (1) Go to http://genome.ucsc.edu/cgi-bin/hgGate-way, choose genome (Human) and assembly version (hg18), and click “add custom tracks”; (2) click “Browse” and choose the BEDGraph file and click “Submit” (see Note 10); (3) after uploading a file, the user will be redirected to a page called “Manage Custom Tracks”; from there one may choose to upload more files or go to the genome browser.

8. Save your session (see Note 11). Click the “My Data” option on the top of the genome browser and then choose “sessions.” One needs to register, if not have done, to save the session.

Figure 1 shows a screenshot from the online UCSC genome browser for the two data sets. GATA1 not only occupies the promoters of genes HBB, HBD, and HBBP1 but also binds to the enhancer sites downstream to gene HBE1 (top track). Consistent with their important functions in hematopoiesis, both HBB and HBD are highly expressed (second track). One great advantage of using the online UCSC genome browser is its ability to integrate NGS data sets preexisting in the browser, including those from the ENCODE project. For instance, from the “Encode Regulation …” section, the enhancer regions in the beta-globin domain in erythrocyte precursor cells are marked by H3K4me1 (an enhancer signature) in the distinctive human mammary epithelial cells (third track). However, the enhancers are not likely active, because the nearby globin genes are all silent (forth track).

Fig. 1.

A screenshot of the uploaded ChIP-Seq and RNA-Seq from the online UCSC genome browser. First track: distribution ChIP-Seq reads of GATA1 at the beta-globin domain for human CD36+ erythrocyte precursor cells. Second track: distribution of RNA-Seq reads from the same cells as in the first track. The Y-axis for the two tracks represents the number of tags normalized by total library size and window size. Third track: genomic distribution of H3K4me1 for human mammary epithelial cell line. Enrichment of H3K4me1 in intergenic regions is a marker of enhancer. Forth track: distribution of RNA-Seq reads for the same cells as in the third track. No read was detected in this region. The last two tracks are preexisting in the online UCSC genome browser. Enhancer regions are marked by rectangle

3.4 Identification of Read Enriched Regions

A common step in analyzing ChIP-Seq data is to identify the genomic regions enriched with mapped reads. The general idea is to test whether the number of tags with a genomic region is significantly more than those generated from a background model. An initial check of the read distribution from the genome browser helps to tell whether the read-enriched regions are broad or narrow. While different methods have been developed to address each situation, a combinational usage of the methods is not uncommon in literatures [16].

Identification of read-enriched regions is also justified for RNA-Seq data under certain circumstances. It is known that reads from the 3′-end of an RNA molecule are more likely sampled than those from the 5′-end, especially for single cell RNA-Seq [17]. In this situation, normalizing the read count within a gene by the size of read-enriched regions rather than simply by the gene length would improve the quantification of gene expression. The 3′-end- biased sequencing data provide valuable information on the exact ending positions of transcripts, of which the boundaries can be defined by read-enriched islands.

3.5 Gene Expression Quantification

The abundance of mRNA of a gene is quantified by RPKM (the number of reads per kilobases of exon model per million reads) for RNA-Seq, which normalizes the length of RNA species and sequencing depth [18]. The expression level can be measured at both gene and isoform levels and the choice is project specific. The differentially expressed (DE) genes are identified by examining whether or not the difference in read counts between two conditions is significantly higher than expectation. Different probabilistic distributions are proposed to model read count from RNA-Seq data, including Poisson and negative binomial, with representative tools such as edgeR [19].

3.6 Average Density Profile from ChIP-Seq Data

We previously generated a large number of ChIP-Seq data sets for histone methylations and acetylations in human hematopoietic stem cells, erythrocyte precursors, CD4⁺ T cells, and B cells [5–8]. Analysis of these data sets revealed that different histone modifications show distinctive preferences in genomic localizations. A plot for the average density of reads for a histone modification surrounding and across genic regions helps to reveal its localization preference. Below is an example for how to obtain the average distribution of H3K4me3 across a genic region from human hematopoietic stem cells:

1. Download the BED file for the H3K4me3 ChIP-Seq data from GEO (GSM317587) [6]. Note that the BED file is based on hg18.

2. Download genome annotation from the online UCSC genome browser following instructions in Subheading 2.1 (choose Human and assembly version hg18).

3. Calculate the average density of H3K4me3 across genes by usinganin-houseC++program“AverageDensityAcrossGenes”: It divides the promoter region (TSS ± 2 Kbps) into 20 equal size bins, separates gene body region into 10 fractions, and extends to 2 Kbps after TES (10 equal size bins) (see Note 12). It outputs the density for each bin/fraction in a fl at text file.

4. Visualize the average density with any spreadsheet software such as Excel (Fig. 2a).

Fig. 2.

H3K4me3 read density at promoters positively correlates with gene expression level. (a) Average H3K4me3 read density across promoter region, gene body region, and region 2 Kbps downstream to transcription ending site (TES). The promoter region (TSS ± 2 Kbps) is divided into 20 equal size bins. Gene body region, excluding the first 2 Kbps, is separated into ten fractions. Region downstream of TES is divided into 10 equal size bins. (b) Similar to panel a, except that the average density is plotted independently for four groups of genes, sorted based gene expression levels. (c) The average density is visualized as a heatmap. Genes were sorted into 200 equal size groups by gene expression level. Each row corresponds to a group of genes. Each column corresponds to a bin/fraction of genomic region as defined in a

3.7 Correlating Histone Modifications with Gene Expression Level

After sorting genes based on their expression levels into equal size groups, two strategies are introduced to visualize the correlation between histone modifications and gene expression: (1) Plot the average density profile of a histone modification (see Note 13), and (2) visualize the read density by using heatmap.

Below is a step-by-step guide about how to visualize H3K4me3 densities across groups of genes sorted by gene expression level (both from human hematopoietic stem cells):

1. DownloadtheChIP-SeqBEDfileforH3K4me3(GSM317587) [6] from GEO and genome annotation (hg18) from the online UCSC genome browser.

2. Download the RNA-Seq BED file from GEO (GSM651554) [14].

3. An in-house C++ program “RPKMCalculator” calculates expression at gene level by taking a BED file and a genome annotation file (see Note 14).

4. Sort gene annotations with the expression file from step 3 with “SortGeneAnnoByExpr.”

5. Set the desired number of groups as the last input parameter of the in-house C++ program: “AverageDensityAcrossGenes.” It will output a file containing a matrix of average read density with each row corresponding to a bin or fraction and each column corresponding to a gene group. The matrix can be visualized by any spreadsheet software such as Excel (Fig. 2b). If the number of groups is large, the density matrix can be imported into and visualized as a heatmap by MeV (Fig. 2c).

3.8 Visualization of Histone Modifications and Gene Expression During Hematopoiesis

The dynamics of histone modifications and gene expression during early stages of T-cell development was extensively characterized by Dr. Ellen Rothenberg’s laboratory [20]. They generated genome- wide ChIP-Seq data for several histone modifications and RNA- Seq data during the differentiation from “early T-cell precursors” (DN1) to CD4 and CD8 double-positive cells [20]. Using these data sets, we show below an example of a combinational usage of in-house programs, Access/Excel and MeV, to visualize the dynamics of H3K4me2 enrichment at promoters during the early stages of T-cell development. For a concise result, we limited the data analysis to genes that are specifically expressed in DN1 cells. Examples of Access/Excel files are available upon request:

1. Obtain gene expression data. (1) Download RNA-Seq raw sequence data from GEO (GSE31235; “sample1”) [20]. Generate BED6 files as described in steps 1–3 of Subheading 3.3. (2) Download UCSC genome annotation (mm9) as described in Subheading 2.1. (3) Apply the in-house C++ program “RPKMCalculator” to calculate RPKM values for each BED6 file. (4) Apply the in-house C++ program “Cat_expr_file” to concatenate the expression files from step 3. It outputs a file containing a matrix of expression values, with each row corresponding to one gene and each column corresponding to one sample.

2. Define DN1 specific genes. (1) Open the expression matrix file from Excel. (2) Create a new column and define the values for the column by using the formula (Fig. 3a).

3. Create a database using Access (see Note 15). Choose “Blank Database” from the templates that appear after running Access to create a blank database, name, and save it to a local drive.

4. Import the expression data into Access. Click the “External Data” panel and choose “Import/Excel File,” which activates the “Import Text Wizard,” to import the excel file generated from step 2. During the process, note that (1) the first row of the density file contains field names and (2) attribute “ID” should be specified as primary key of the table.

5. Obtain read density of H3K4me2 at promoters. (1) Download ChIP-Seq data from GEO (GSE31235) and process the sra files to BED6 files. (2) Apply the in-house C++ program “DensityCalculatorPromoters” to calculate the normalized read density at promoters for all samples (see Note 16).

6. Import the read density into Access. Similar to step 4, click “External data” panel, and chose “Import/Text” to import the text file generated from step 5 into the database.

7. Intersect the expression and density tables. (1) Click the “create” panel and choose “query design” to activate the “show table” dialog. (2) Add the two tables to the query panel from the dialog. (3) Click the “ID” attribute of one table, hold, drag, and release it to the “ID” attribute of the other (Fig. 3b, dashed arrow 1) to create a join link that implements an intersection operation between the two tables through the specified attributes.

8. Extract read density for DN1 specific genes. (1) Click, hold, drag, and release the attributes associated with read densities to the bottom panel (Fig. 3b, dashed arrow 2). (2) Use “Criteria” in the bottom panel to restrict the query results to DN1-specific gene (Fig. 3b, dashed arrow 3). (3) Execute (“Design/Results/Run”) and save the query (“ctrl+s”). (4) Export the results to a fl at text file (“External Data/Export/Text File”). During this process, choose to include Field name and “Tab” as delimiter.

9. Visualize read density by using MeV (Fig. 3c). (1) Import the fl at text file from step 6 into MeV (“File/Load Data”). (2) Normalize the read density to highlight changes of H3K4me3 enrichment across samples (“Adjust Data/Gene/Row Adjustments/Normalize Genes/Rows”). (3) Cluster genes based on their read density across samples (“Analysis/Clustering/HCL”; see Note 17). A “HCL: Hierarchical clustering” dialog will show up before the clustering for users to set parameters. (4) Choose color theme (“Display/Color Scheme”), adjust color scale (“Display/Set Color Scale Limits”), and set size (“Display/Set Element Size”) of the heatmap.

Fig. 3.

A combinational use of Excel/Access and Mev to visualize changes of histone modifications. (a) A new column “DN1-specific” is created and the value is defined by the formula. As shown in the formula, a gene is defined as DN1-specific if the expression is higher than 3 (“B2 > 3,” where “B2” means column “B” row “2”) and is at least twofold higher than any other stages (“B2/(max(C2:F2) + 0.001) > 2,” where 0.001 is a “pseudo count”). (b) Screenshot of the query window. The dashed arrows are explained in main text. (c) Hierarchical clustering and heatmap visualization of H3K4me2 density at promoters of DN1-specific genes during the differentiation from DN1 to DP cells. H3K4me2 is highly enriched at the promoters of most genes at DN1 stage and decreases during differentiation. However, about 1/3 of genes also show high enrichments of H3K4me2 at promoters at later development stages