EpiCompare: An online tool to define and explore genomic regions with tissue or cell type-specific epigenomic features

---

The Human Reference Epigenome Map, generated by the Roadmap Epigenomics Consortium, contains thousands of genome-wide epigenomic datasets that describe epigenomes of a variety of different human tissue and cell types. This map has allowed investigators to obtain a much deeper and more comprehensive view of our regulatory genome, for example defining regulatory elements including all promoters and enhancers for a given tissue or cell type. An outstanding task is to combine and compare different epigenomes in order to identify regions with epigenomic features specific to certain type of tissues or cells, for example, lineage-specific regulatory elements. Currently available tools do not directly address this question. This need motivated us to develop EpiCompare that allows investigators to easily identify regions with epigenetic features unique to specific epigenomes that they choose, making detection of common regulatory elements and/or cell type-specific regulatory elements an interactive and dynamic experience. Investigators can design their tests by choosing different combinations of epigenomes, and choosing different classification algorithms provided by our tool. EpiCompare will then identify regions with specified epigenomic features, and provide a quality assessment of the predictions. Investigators can interact with EpiCompare by investigating Roadmap Epigenomics data, or uploading their own data for comparison. Finally, prediction results can be readily visualized and further explored in the WashU Epigenome Browser.

Index

---

• 1. Datasets
• 1.1 Chromatin states
• 1.2 H3K27ac
• 1.3 H3K4me1
• 1.4 H3K4me3
• 1.5 H3K27me3
• 2. Methods
• 2.1 Frequncy cutoff
• 2.2 Fisher's exact test
• 2.3 K-means clustering
• 3. Output
• 3.1 Data output
• 3.2 Validation output
• 3.2.1 Enrichment for H3K27ac peaks
• 3.2.2 Tissue enrichment index on H3K27ac
• 4. How to use the tool
• 4.1 Select a feature
• 4.2 Upload data
• 4.3 Select foreground samples
• 4.4 Select background samples
• 4.5 Select features and samples for visualization
• 4.6 Select methods
• 4.7 Submit
• 5. Reference

1. Datasets

For each feature - the ChromHMM state or epigenomic modification peak below, it is converted into binary presence or absence of the feature in each 200bp window, denoted by 1 or 0. A table is generated for each feature by summarizing the presence or absence of the feature in all samples across windows where at least one sample has the feature.

---

1.1 Chromatin states

Chromatin state data for 15-state model and 18-state model for all tissue/cell types (127 samples have chromatin states from 15-state model and 98 samples from 18-state model) are obtained from Roadmap Epigenomics Project (Roadmap Epigenomics, et al., 2015). Enhancers for 15-state model are defined as state number 6, 7, 12 and enhancers for 18-state model are defined as state number 7, 8, 9, 10, 11, 15. Promoters for 15-state model are defined as state number 1, 2, 10 and promoters for 18-state model are defined as state number 1, 2, 3, 4, 14. Chromatin states are defined on each 200bp window by ChromHMM.

1.2 H3K27ac

H3K27ac peak data for 98 tissue/cell types are obtained from Roadmap Epigenomics Project. The peaks are called by MACS2 (Zhang, et al., 2008). H3K27ac peak data are processed on 200bp window by requiring at least 50bp overlapping with 200bp window in the genome. Only peaks with q-value less than 0.01 are kept.

1.3 H3K4me1

H3K4me1 peak data for 127 tissue/cell types are obtained from Roadmap Epigenomics Project. The peaks are called by MACS2. H3K4me1 peak data are processed on 200bp window by requiring at least 50bp overlapping with 200bp window in the genome. Only peaks with q-value less than 0.01 are kept.

1.4 H3K4me3

H3K4me3 peak data for 127 tissue/cell types are obtained from Roadmap Epigenomics Project. The peaks are called by MACS2. H3K4me3 peak data are processed on 200bp window by requiring at least 50bp overlapping with 200bp window in the genome. Only peaks with q-value less than 0.01 are kept.

1.5 H3K27me3

H3K27me3 peak data for 127 tissue/cell types are obtained from Roadmap Epigenomics Project. The peaks are called by MACS2. H3K27me3 peak data are processed on 200bp window by requiring at least 50bp overlapping with 200bp window in the genome. Only peaks with q-value less than 0.01 are kept.

2. Methods

---

Three methods are used for identifying regions with epigenomic features specific to combinations of tissue or cell types. All methods require the definition of foreground samples and background samples by users. Foreground samples are the group of samples for which we identify specific regions. Background samples are the group of samples against which we compare foreground samples. The principle of all methods is, to define regions with features specific in foreground samples, the features should be enriched in foreground samples but depleted in background samples. Below is visualization of three methods in EpiCompare with a simple example. F represents foreground samples, B represents background samples, R represents regions, C represents clusters.

2.1 Frequency cutoff

For each region (in this case each 200bp genomic window), the percentages of samples having the feature in foreground samples and background samples are calculated. If the percentage of samples having the feature in the foreground samples is greater than or equal to the defined minimal foreground cutoff (default is 80%) and the percentage of samples having the feature in the background samples is less than or equal to the defined maximal background cutoff (default is 20%), then the region is defined as a positive region. These positive regions are further ranked by the difference between the percentage of samples having the feature in foreground and background samples so users can prioritize top-ranked regions.

2.2 Fisher's exact test

For each 200bp window, a contingency table composed of the number of samples with or without the feature in foreground samples and background samples is calculated. Fisher’s exact test is used to examine whether the percentage of features in foreground samples is significantly greater than in background samples. The p-value is corrected by multiple hypothesis testing using the Benjamini-Hochberg procedure, and regions with q-value less than a cutoff (default is 0.01) are identified and ranked by their q-values. The statistical power of the test depends on the number of foreground and background samples and having more samples can provide more statistical power to identify more significant q-values. Therefore, when the number of foreground samples is small, investigators can use q-value as a ranking measure and obtain the top candidates by setting a higher q-value threshold.

2.3 K-means clustering

First, k-means clustering based on a Jaccard-index distance is performed on the binary data table for each feature, similar to the clustering method used in HoneyBadger2 (Roadmap Epigenomics, et al., 2015). R package flexclust is used for clustering (Leisch, 2006). We chose the optimal cluster number by the elbow method and the silhouette method (Kodinariya and Makwana, 2013). The optimal cluster number for all features is close and around 140, so we fixed the optimal cluster number to be 140. Besides providing the clustering result using the optimal cluster number, we also provided clustering results with three other cluster numbers (90, 200, 250) depending on the users’ need. Next, the percentage of regions having the feature is calculated for each cluster and defined as a feature density table (number of clusters times number of samples). Finally, a cluster specific for a tissue/cell type should have higher feature density in that tissue/cell type than in the background samples. Specifically, to identify clusters specific for foreground samples, we select clusters satisfying the following two conditions: first, the median of feature densities of foreground samples in a cluster is greater than or equal to a threshold (default is 0.4); second, it should also be greater than or equal to the highest feature density in the background samples of that same cluster (this threshold can be set to any percentile of feature densities in the background samples).

3. Output

---

3.1 Data output

Identified regions are displayed on a table whose columns are chr, start, end, and link to WashU Epigenome Browser so users can visualize and compare the regions in different tissue/cell types. For Fisher’s exact test and frequency cutoff method, a summary is provided about the distribution of ranks for identified regions. Links are provided to download unmerged 200bp regions or merged regions which merge neighboring 200bp regions. A link to GREAT analysis for merged regions is also provided.

3.2 Validation output

3.2.1 Enrichment for H3K27ac peaks

Enrichment for H3K27ac peaks in foreground samples and background samples are calculated for identified regions. Enrichment is defined as below: enrichment=((#bp in overlapped regions)⁄(#bp in H3K27ac sites ))/(#bp (identified regions)⁄(#bp in hg19 genome)) . When the number of foreground samples or background samples is bigger than 10, 10 random samples from foreground samples or background samples are chosen to calculate the enrichment.

3.2.2 Tissue enrichment index on H3K27ac

a tissue enrichment index for identified regions is calculated using H3K27ac RPKM (Reads Per Kilobase of transcript per Million mapped reads). Identified regions are filtered using combined H3K27ac peaks from 98 samples and the tissue enrichment index is calculated for filtered regions. Tissue enrichment index has been routinely used to identify tissue-specific genes (Chang, et al., 2011; Yanai, et al., 2005). Generally, a high tissue enrichment index represents tissue-specific regions. The tissue enrichment index we use is a contribution measure (CTM) (Pan, et al., 2013).

4. How to use the tool

---

The tool can be run in Chrome, Internet Explorer, and Firefox web browser. It can be run in Safari if used in a window without other web pages. Opening other pages in the same window in Safari with the tool will make the tool disconnected from the server.

Because the free shiny server can only allow one user at one time, to allow multiple users to use the tool simultaneously, we created 3 copies of the tool named EpiCompare1, EpiCompare2 and EpiCompare3 and used a load balancer named EpiCompare (epigenome.wustl.edu/EpiCompare/) to assign users to the 3 copies. This allows 3 users to use the tool at one time.

Below lists each step of using the tool and specific requirements for each step.

4.1 Select a feature

Select the feature for which you plan to identify the specificity. It can be enhancers or promoters defined from 15 state or 18 state ChromHMM model, histone mark H3K27ac, H3K4me1, H3K4me1, or H3K4me3. Only one feature can be chosen. For example, use the default enhancer state from 18-state ChromHMM model here.

4.2 Upload data

Upload your own data for comparison analysis besides using default Roadmap data. Skip this step if you don't want to use your own data. You can upload one file or multiple files. Select one to upload one file and select multiple to upload multiple files together. The files must have only three columns (chromosome, start, end) specifying the location of the feature. The coordinates can be merged or not. The tool will map the coordinates to 200bp window by requiring at least 1bp overlapping. After uploading files, the name of each file will be listed on top of Roadmap samples for selection. The name of the file must not have space. Only frequency cutoff and Fisher's exact test can be used to analyze uploaded data while k-means clustering method cannot be used.

1.Here is a test file . Download the file to your local drive.

2.Upload this file locally. The file will be uploaded and processed.

3.After processing, the name of uploaded file will be listed as user-defined samples.

4.3 Select foreground samples

Select the group of samples for which you identify specific regions. They can be chosen from Roadmap samples or uploaded samples. Click finish selection button after finishing selection. Selected sample IDs for foreground samples will be listed. For example, select 5 adult brain samples here.

4.4 Select background samples

Keep selected foreground samples in step 3 unchanged and select background samples, which are the group of samples against which we compare foreground samples. The selected foreground samples in step 3 must not be unselected because the tool subtracts the selection in step 3 from all selections in step 4 to obtain selected background samples . Click finish selection button after finishing selection. Selected sample IDs for background samples will be listed. If background samples are not specified, the tool will identify shared enhancers for foreground samples. Only frequency cutoff and k-means clustering method can be used without background samples, while Fisher's exact test cannot. For example, select all adult blood samples here.

4.5 Select features and samples for visualization

Select features and samples for visualization in WashU Epigenome Browser. Click finish selection button after finishing selection. Selected sample IDs will be listed. If both features and samples are not chosen, the default values for them will be used, which are the feature selected in step 1 and foreground samples selected in step 3. For example, select ChromHMM18 and H3K27ac as features and 2 brain samples from foreground samples and 2 blood samples from background samples as samples here.

4.6 Select a method

Select frequency cutoff, Fisher's exact test or k-means clustering method and the parameters for each method.

1) Frequency cutoff method: a region is defined as a positive region if the percentage of samples having the feature in the foreground samples is greater than or equal to a cutoff (default is 80%, called foreground cutoff) and the percentage of samples having the feature in the background samples is less than or equal to a cutoff (default is 20%, called background cutoff).

2) Fisher's exact test method: A region is defined as a positive region if the q-value from Fisher's exact test is less than a cutoff (default is 0.01, called q-value cutoff). When the number of foreground samples is small, Fisher’s exact test method cannot identify any regions with q-value threshold less than 0.01. In this case, investigators can use q-value as a ranking measure and obtain the top candidates by setting a high q-value cutoff.

3) K-means clustering method: Clusters specific for foreground samples should satisfy the following two conditions. First, the median of feature densities of foreground samples in this cluster is greater than or equal to a cutoff (default is 0.4, called foreground density cutoff); second, the median of feature densities of foreground samples is also greater than or equal to the percentile cutoff of feature densities of the background samples (default is 100%, which is the maximal feature density of the background samples, called percentile cutoff). The feature density in one cluster is the proportion of regions for one sample having the feature in the cluster. A value of 50% means half of the regions have the feature in the cluster.

Fisher's exact test method is not applicable when background samples are not specified. K-means clustering method is not applicable for analyzing uploaded data.

For example, use the default frequency cutoff method here.

4.7 Submit

Click submit button and start analysis. The result will be available in about 3 minutes.

This is the table of identified regions.

This is the validation results: enrichment for H3K27ac peaks and tissue enrichment index on H3K27ac.

This is the visualziation in Washu Epigenome Browser for an identified region.

5. Reference

Chang, C.W., et al. Identification of human housekeeping genes and tissue-selective genes by microarray meta-analysis. PLoS One 2011;6(7):e22859.

Leisch. A Toolbox for K-Centroids Cluster Analysis. Computational Statistics and Data Analysis, 51 (2), 526-544, 2006.

Pan, J.B., et al. PaGenBase: a pattern gene database for the global and dynamic understanding of gene function. PLoS One 2013;8(12):e80747.

Roadmap Epigenomics, C., et al. Integrative analysis of 111 reference human epigenomes. Nature 2015;518(7539):317-330.

Zhang, Y., et al. Model-based analysis of ChIP-Seq (MACS). Genome Biol 2008;9(9):R137.p>

Yanai, I., et al. Genome-wide midrange transcription profiles reveal expression level relationships in human tissue specification. Bioinformatics 2005;21(5):650-659.

Any questions are welcome. Please contact yu.he (at) wustl.edu .

A paper about EpiCompare is pubished in Bioinformatics journal. Please cite the tool as:

Yu He, Ting Wang; EpiCompare: An online tool to define and explore genomic regions with tissue or cell type-specific epigenomic features. Bioinformatics 2017 btx371. doi: 10.1093/bioinformatics/btx371