Contents

1 Introduction

An ever-growing variety of short read RNA sequencing methods is available to study the various aspects of transcript biosynthesis, processing and degradation (Wang, Gerstein, and Snyder 2010). Some methods, such as mRNA sequencing (mRNA-seq), measure signals derived from processed mRNA and are thus ideally suitable for steady-state mature transcript level measurements and for the investigation of splice variants (Morozova, Hirst, and Marra 2009). Other techniques, such as Global Run-On sequencing (GRO-seq), nuclear RNA-seq (nucRNA-seq) and chromatin-associated RNA-seq (chrRNA-seq) provide information on primary transcription and give a more comprehensive picture of transcriptional activity, also for non-polyadenylated RNA (Mitchell et al. 2012; Werner and Ruthenburg 2015; Hah et al. 2013; Gaidatzis et al. 2015). The differences in the RNA types being sequenced have an impact on the resulting sequencing profiles. mRNA-seq data is enriched with reads derived from exons, while GRO-, nucRNA- and chrRNA-seq demonstrate a substantial broader coverage of both exonic and intronic regions (Zaghlool et al. 2013). The presence of intronic reads in GRO-seq type of data makes it possible to use it to computationally identify and quantify all de novo continuous regions of transcription distributed across the genome. This type of data, however, is more challenging to interpret and less common practice compared to mRNA-seq. One of the challenges for primary transcript detection concerns the simultaneous transcription of closely spaced genes, which needs to be properly divided into individually transcribed units. The R package transcriptR combines RNA-seq data with ChIP-seq data of histone modifications that mark active Transcription Start Sites (TSSs), such as, H3K4me3 or H3K9/14Ac to overcome this challenge. The advantage of this approach over the use of, for example, gene annotations is that this approach is data driven and therefore able to deal also with novel and case specific events. Furthermore, the integration of ChIP- and RNA-seq data allows the identification all known and novel active transcription start sites within a given sample.

transcriptR is an R package with a pipeline that is made up out of two main parts; an RNA-seq and a ChIP-seq part, of which the outputs are integrated to ultimately yield a comprehensive database of de novo identified primary transcripts and their abundance (Figure 1). In the first (RNA-seq) part, strand-specific GRO-seq, nucRNA-seq or chrRNA-seq short-reads in Binary Sequence Alignment Map (BAM) format are converted to coverage profiles. Background noise levels are estimated from regions with the low reads coverage using a Poisson-based approach. Genomic regions with read densities above background levels are considered to be expressed and small gaps in otherwise continuously transcribed regions are bridged when the gaps sizes are below a certain threshold which is extrapolated from the sequencing data and reference gene annotations. The second part, operates on ChIP-seq data and requires two input files: 1) a BAM file with the sequencing reads and 2) a peak file - output of a peak calling algorithm (for example MACS2) (Zhang et al. 2008). As a first step, a classification model, based on the logistic regression, is used to predict and discriminate gene associated peaks from background peaks, using estimated characteristics of the peaks. Next, transcription initiation within a peak region is investigated by comparing RNA-seq read densities upstream and downstream of empirically determined transcription start sites. Putative transcription of both forward and reverse genomic strands is tested and the results are stored with each ChIP-seq peak.

At the end of the pipeline, both parts are combined and, where applicable, closely spaced transcripts are divided into individually transcribed units using the detected active transcription start sites. Additionally, the read count and FPKM value is calculated for each transcript in the dataset to facilitate further quantitative analysis.

The advantage of the two-part approach presented here is that the transcript detection and quantification can still be performed even in the absence of ChIP-seq data, bearing in mind that some adjacent transcripts may be detected as one transcribed unit.

transcriptR workflow

Figure 1: transcriptR workflow

In addition to the features mentioned above transcriptR provides functions for generating BigWig, BedGraph and BED files for the visualization of the coverage profiles, identified transcripts and peaks predictions in the UCSC genome browser.

2 Getting started

The transcriptR package is available in the Bioconductor and can be downloaded as follows:

The following packages are required for the transcriptR workflow.

3 Transcripts detection

3.1 TranscriptionDataSet object construction

The workflow is initiated by creating a TranscriptionDataSet object, which is a container for holding processed sequencing data and the results of all downstream analyses. A function constructTDS initializes the construction of the TranscriptionDataSet object, by providing the paths to the input files and information relevant to the library preparation procedure. Optionally, the extracted reads can be limited to a specific genomic region, by using the region option. This would decrease the run time for the downstream analyses and might be useful for the testing purposes.

In this tutorial we supply an already constructed TranscriptionDataSet, containing nuclear RNA-seq data for human chromosome 15, profiled in prostate cancer LNCaP cells.

Some of the supplied information can already be seen in the constructed TranscriptionDataSet object, whereas other slots are still empty and will be filled once the analysis is conducted.

3.2 Expression background estimation

Gene expression is a stochastic process, which often results in substantial expression noise. To obtain a putative set of transcribed regions, it is necessary to identify those regions that are expressed significantly above the background level. Using a Poisson-based approach for estimating the noise distribution, estimateBackground function returns a coverage cutoff value for a specific False Discovery Rate (FDR). The estimated value is stored in the coverageCutoff slot of the TranscriptionDataSet and will be used in the downstream analysis.

3.3 Coverage profile visualization

RNA-seq coverage profiles for both forward and reverse DNA strand can be visualized separately in the UCSC genome browser using exportCoverage. This function can generate tracks in BigWig and bedGraph formats, which can be uploaded to the genome browser as custom tracks. There is an option to filter coverage profiles by a coverage cutoff value, either estimated for a specific FDR via estimateBackground or a user specified value. By default, the coverage cutoff value stored in the TranscriptionDataSet object is used. In order to make an informed decision about a proper FDR level, it is useful to explore the output at different FDR levels and determine the optimal cutoff value. Additionally, RPM (reads per million mapped reads) normalization is available.

3.4 Parametres tuning

The ultimate goal of transcriptR is to identify continuous regions of transcription. However, in some areas of the genome it is not possible to detect transcription, because of the presence of the low mappability regions and (high copy number) repeats. Sequencing reads can not be uniquely mapped to these positions, leading to the formation of gaps in otherwise continuous coverage profiles and segmentation of transcribed regions into multiple smaller fragments. The gap distance describes the maximum allowed distance between adjacent fragments to be merged into one transcript. To choose the optimal value for the gap distance, the detected transcripts should largely be in agreement with available reference annotations. To accomplish this, the function is build on the methodology proposed by Hah et al. (Hah et al. 2011). In brief, this method uses two types of errors:

  • ‘dissected’ error - the ratio of annotations that is segmented into two or more fragments.

  • ‘merged’ error - the ratio of non-overlapping annotations that merged by mistake in the experimental data.

There is an interdependence between two types of errors. Increasing the gap distance decreases the ‘dissected’ error, by detecting fewer, but longer transcripts, while the ‘merged’ error will increase as more detected transcripts will span multiple annotations. The gap distance with the lowest sum of two error types is chosen as the optimal value.

The function estimateGapDistance uses increasing gap distances (based on the supplied vector) and calculates the associated error rates to determine the optimal gap distance.

Additionally, all intermediate calculations can be accessed by getTestedGapDistances and the output can be presented in a graphical way by plotErrorRate function call. Here, the tested gap distances are plotted on the x-axis and corresponding error rates on the y-axis. Three curved lines depict the two error types calculated by estimateGapDistance and the sum of both errors. The vertical dashed line depicts the gap distance with the smallest sum of two errors.

Gap distance error rate

Figure 2: Gap distance error rate

3.5 Transcript calling

Transcript detection is performed by detectTranscripts. This function uses the two previously determined parameters to identify transcribed regions: coverage.cutoff and gap.dist as calculated by the estimateBackground and estimateGapDistance, respectively and stored in the TranscriptionDataSet object. Alternatively, the user may specify his/her own values to be passed to the function. By increasing the gap.dist, fewer transcripts of longer size will be identified, and an increase in the coverage.cutoff will result in fewer transcripts of shorter size.

If desired, the identified transcripts can be associated with the available reference annotations based on the genomic overlap. To achieve this, the minimal proportion of the overlap between transcript and annotation is controlled by the min.overlap argument.

The detected transcripts can be retrieved by the getTranscripts. This function uses a number of arguments that control the resulting list of reported transcripts. In case, ChIP-seq data is not available, the following method will generates the final set of transcripts.

3.6 Detected transcripts visualization

A convenient graphical way to explore the identified transcripts is to visualize them in the UCSC genome browser. The transcriptsToBed function returns a file in BED format, which can be directly uploaded to the genome browser. To improve the visual perception, transcripts are color-coded by DNA strand orientation.

4 ChIP-seq peaks characterization and classification

When associated ChIP-seq data (e.g. H3K4me3 or H3K9ac) is available, this information can be used to identify the transcript starts. This is particularly useful in situations where genes are densely packed in the genome and the identification of individual transcripts from RNA-seq data is challenging. The inclusion of ChIP-seq peak information has to be performed carefully as not all identified peaks show evidence of active transcription in RNA-seq data. In order to discriminate between peaks with active transcription and background peaks, we use a two step approach. First, we collect the characteristics of ChIP-seq peaks that overlap transcription start sites of annotated genes. These characteristics are used to identify all putatively gene-associated ChIP-seq peaks. A schematic overview of the procedure is shown in Figure 3.

Classification of the peaks on gene associated and background

Figure 3: Classification of the peaks on gene associated and background

Secondly, we incorporate RNA-seq data to find direct evidence of active transcription from every putatively gene-associated peak. In order to do this, we determine the “strandedness” of the ChIP-seq peaks, using strand specific RNA-seq data. The following assumptions are made in order to retrieve the peak “strandedness”:

Average RNA-seq signal across ChIP peak region

Figure 4: Average RNA-seq signal across ChIP peak region

Each peak in the data set is tested for association with transcription initiation on both strands of DNA. A detailed overview of the procedure is depicted in the Figure 5. Steps 1-5 are performed for both forward and reverse DNA strand separately and step 6 combines the data from both strands. If the peak is identified as associated with the transcription on both strands, than it is considered to be a bidirectional.

ChIP peak “strandedness” prediction steps:

  1. Identify a location within the ChIP-seq peak near the transcription start site. This is accomplished by calculating the cumulative distribution of RNA-seq fragments within a peak region. The position is determined where 90% of RNA-seq fragments are located downstream. This approach performs well on both gene-poor and gene-dense regions where transcripts may overlap.

  2. Two equally sized regions are defined (q1 and q2), flanking the position identified in (1) on both sides. RNA-seq fragments are counted in each region.

  3. ChIP peaks with an RNA-seq fragment coverage below an estimated threshold are discarded from the analysis.

  4. The probability is calculated for RNA-seq fragments to be sampled from either q1 or q2. Based on the assumptions we stated above, a ChIP peak that is associated with transcription initiation should have more reads in q2 (downstream of the transcription start position) compared to q1, and subsequently, the probability of a fragment being sampled from q2 would be higher.

  5. ChIP-seq peaks are divided into gene associated and background based on the prediction.

  6. Iteratively, the optimal P(q2) threshold is identified, which balances out the False Discovery Rate (FDR) and False Negative Rate (FNR). Peaks with the P(q2) exceeding the estimated threshold are considered to be associated with the transcription initiation event.

ChIP peak "strandedness" prediction

Figure 5: ChIP peak “strandedness” prediction

4.1 ChipDataSet object construction

We start ChIP-seq analysis by creating a ChipDataSet object, which is a container for storing ChIP-seq peaks information and all the results that will be generated by applying specific functions. The function constructCDS initializes the ChipDataSet object, by providing the paths to the input files and information relevant to the ChIP-seq library preparation procedure. During the object construction the following steps are executed:

  • The peak information is converted into the object of GRanges class.

  • The genomic distribution of the peaks is evaluated (exonic, intronic, intergenic, TSSs).

  • Each peak in the data set is functionally characterized (peak length, total number of reads, pileup, etc.). The estimated features are used to predict which of the peaks are gene associated in the analysis downstream.

As many peak-calling algorithms tend to divide broader peaks into the several narrower closely spaced peaks, it is advised to merge these end-to-end peaks to decrease the number of false positives and prevent unnecessary truncation of transcripts in the downstream analysis.

In this tutorial we supply an already constructed ChipDataSet, containing H3K4me3 active histone mark ChIP-seq peaks from the chromosome 15, profiled in the prostate cancer LNCaP cells.

The peaks stored in ChipDataSet can be retrieved by a simple getPeaks function call.

A simple quality check of the supplied ChIP-seq peaks can be performed by investigating their genomic distribution. Ideally, these peaks should demonstrate substantial enrichment at TSS regions. Enrichment of the peaks at a given genomic feature (e.g. TSS) is defined as the ratio between the observed and expected number of peaks. The expected number of peaks is calculated from the proportion of the genome covered by the given genomic feature.

Additionally, the genomic distribution of the peaks can be visualized in two ways, either by observing the total number of peaks overlapping given genomic feature (Figure 6) or by looking at the enrichment levels (Figure 7).

Genomic distribution of the peaks

Figure 6: Genomic distribution of the peaks

Enrichment of peaks at distinct genomic features

Figure 7: Enrichment of peaks at distinct genomic features

4.2 Prediction of the gene associated peaks

In order to discriminate between functional or gene associated peaks and non-functional or background peaks, each peak in the data set is characterized by several features. Among them:

  • length - the length of a peak (in base pairs).

  • fragments - total number of fragments overlapping a peak region.

  • density - number of fragments per base pair of the peak length.

  • pileup - highest fragment pileup in each peak region.

  • tssOverlap - overlap (binary, yes/no) of the peak with the annotated TSS region.

Moreover, the user may wish to supply her/his own list of features with the addFeature. See the manual page for an example of the use of addFeature.

Prior to fitting the logistic model, the relations between predictors and response variable (tssOverlap) can be explored with plotFeatures (Figure 8). Based on the plots, poor predictors can be excluded from the analysis to improve the model fit. In general, the pileup (peak height) is a good predictor on its own, as it can segregate gene associated and background peaks very well.

Estimated features

Figure 8: Estimated features

The logistic regression model fit is accomplished by predictTssOverlap. In order to improve the accuracy of the model the data is internally partitioned into a training and testing data sets. The percent of the data that will be allocated to the training set should be specified by parameter p. A repeated 10-Fold Cross-Validation is used to calculate performance measures on the training data set and to prevent over-fitting. It is possible to specify a subset of features (predictors) to be used in the model fit via the feature argument. By default, all the features will be used.

The result of the logistic regression model fit, can be viewed by getPeaks. Two new columns have been added to the output: 1) predicted.tssOverlap.prob - estimated probability of a peak being gene associated and 2) predicted.tssOverlap - predicted class (binary, yes/no).

The detailed overview of the classification model fit to the data and the significance of each predictor can be accessed by getConfusionMatrix and getPredictorSignificance respectively. Features with a high p-value can be excluded and the analysis should be rerurn with the significant predictors only.

Additionally, the performance of the model can be visualized by a Receiver operating characteristic (ROC) plot (Figure 9). The curve is created by plotting the true positive rate (sensitivity) against the false positive rate (1 - specificity) at various threshold settings. The closer the curve follows the left-hand border and then the top border of the ROC space, the more accurate the test. The area under the curve (AUC) is a measure of accuracy.

Performance of the classification model (gene associated peaks prediction)

Figure 9: Performance of the classification model (gene associated peaks prediction)

4.3 Prediction of the peak strandedness

The Prediction of the peak “strandedness” is accomplished by predictStrand. This function requires information stored in both ChipDataSet and TranscriptionDataSet objects.

The predicted.strand can take one of the four possible variants: 1) “+” plus strand, 2) “-” minus strand, 3) “bi” bidirectional and 4) “.” not associated with a transcription initiation event. All internal calculations performed by predictStrand can be viewed by getQuadProb. It returns a data frame, where each row corresponds to a peak and each column keeps one of the intermediate calculations:

  • max.cov - maximum coverage of the RNA-seq fragments inside the peak region.

  • pass.cov.treshold - whether the max.cov exceeds the coverage.cutoff, either user defined or estimated from RNA-seq data by estimateBackground function call and stored in TranscriptionDataSet object.

  • q1q2.sepline.coord - genomic coordinate corresponding to the transcription start position inside the peak region and which is used to separate q1 and q2.

  • q1.coord - genomic coordinates of q1.

  • q2.coord - genomic coordinates of q2.

  • q1.count - total number of fragments in q1.

  • q2.count - total number of fragments in q2.

  • q1.prob - probability of a fragment being sampled from the q1.

  • q2.prob - probability of a fragment being sampled from the q2 .

The predicted P(q2) threshold, used to select peaks with a putative transcription initiation event, is accessed by getProbTreshold.

4.4 Peaks visualization

A convenient way to explore output of the predictions made on the ChIP peaks is to visualize them in the UCSC genome browser. The peaksToBed function returns a file in BED format, which can be uploaded directly to the genome browser. To improve the visual perception, peaks are color-coded by the predicted strand.

5 Transcript boundaries demarcation

As a last step, both parts of the workflow are combined and, where applicable, closely spaced transcripts are divided into individually transcribed units using the detected active transcription start sites. There is a single function breakTranscriptsByPeaks, which will generate the final set of transcripts.

We can explore which transcripts are broken by peaks, by simply intersecting two sets of transcript outputs (before breaking transcripts by peaks, and after).

6 Session Information

Here is the output of sessionInfo() on the system on which this document was compiled:

## R version 4.0.3 (2020-10-10)
## Platform: x86_64-pc-linux-gnu (64-bit)
## Running under: Ubuntu 18.04.5 LTS
## 
## Matrix products: default
## BLAS:   /home/biocbuild/bbs-3.12-bioc/R/lib/libRblas.so
## LAPACK: /home/biocbuild/bbs-3.12-bioc/R/lib/libRlapack.so
## 
## locale:
##  [1] LC_CTYPE=en_US.UTF-8       LC_NUMERIC=C              
##  [3] LC_TIME=en_US.UTF-8        LC_COLLATE=C              
##  [5] LC_MONETARY=en_US.UTF-8    LC_MESSAGES=en_US.UTF-8   
##  [7] LC_PAPER=en_US.UTF-8       LC_NAME=C                 
##  [9] LC_ADDRESS=C               LC_TELEPHONE=C            
## [11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C       
## 
## attached base packages:
## [1] stats4    parallel  stats     graphics  grDevices utils     datasets 
## [8] methods   base     
## 
## other attached packages:
##  [1] caret_6.0-86                           
##  [2] ggplot2_3.3.2                          
##  [3] lattice_0.20-41                        
##  [4] TxDb.Hsapiens.UCSC.hg19.knownGene_3.2.2
##  [5] GenomicFeatures_1.42.0                 
##  [6] AnnotationDbi_1.52.0                   
##  [7] Biobase_2.50.0                         
##  [8] GenomicRanges_1.42.0                   
##  [9] GenomeInfoDb_1.26.0                    
## [10] IRanges_2.24.0                         
## [11] S4Vectors_0.28.0                       
## [12] BiocGenerics_0.36.0                    
## [13] transcriptR_1.18.0                     
## [14] BiocStyle_2.18.0                       
## 
## loaded via a namespace (and not attached):
##   [1] colorspace_1.4-1            hwriter_1.3.2              
##   [3] ellipsis_0.3.1              class_7.3-17               
##   [5] XVector_0.30.0              farver_2.0.3               
##   [7] bit64_4.0.5                 prodlim_2019.11.13         
##   [9] lubridate_1.7.9             xml2_1.3.2                 
##  [11] codetools_0.2-16            splines_4.0.3              
##  [13] knitr_1.30                  pROC_1.16.2                
##  [15] Rsamtools_2.6.0             dbplyr_1.4.4               
##  [17] png_0.1-7                   BiocManager_1.30.10        
##  [19] compiler_4.0.3              httr_1.4.2                 
##  [21] assertthat_0.2.1            Matrix_1.2-18              
##  [23] htmltools_0.5.0             prettyunits_1.1.1          
##  [25] tools_4.0.3                 gtable_0.3.0               
##  [27] glue_1.4.2                  GenomeInfoDbData_1.2.4     
##  [29] reshape2_1.4.4              dplyr_1.0.2                
##  [31] rappdirs_0.3.1              ShortRead_1.48.0           
##  [33] Rcpp_1.0.5                  vctrs_0.3.4                
##  [35] Biostrings_2.58.0           nlme_3.1-150               
##  [37] rtracklayer_1.50.0          iterators_1.0.13           
##  [39] timeDate_3043.102           gower_0.2.2                
##  [41] xfun_0.18                   stringr_1.4.0              
##  [43] lifecycle_0.2.0             XML_3.99-0.5               
##  [45] zlibbioc_1.36.0             MASS_7.3-53                
##  [47] scales_1.1.1                ipred_0.9-9                
##  [49] hms_0.5.3                   MatrixGenerics_1.2.0       
##  [51] SummarizedExperiment_1.20.0 RColorBrewer_1.1-2         
##  [53] yaml_2.2.1                  curl_4.3                   
##  [55] memoise_1.1.0               biomaRt_2.46.0             
##  [57] rpart_4.1-15                latticeExtra_0.6-29        
##  [59] stringi_1.5.3               RSQLite_2.2.1              
##  [61] highr_0.8                   foreach_1.5.1              
##  [63] e1071_1.7-4                 BiocParallel_1.24.0        
##  [65] lava_1.6.8                  chipseq_1.40.0             
##  [67] rlang_0.4.8                 pkgconfig_2.0.3            
##  [69] matrixStats_0.57.0          bitops_1.0-6               
##  [71] evaluate_0.14               purrr_0.3.4                
##  [73] labeling_0.4.2              GenomicAlignments_1.26.0   
##  [75] recipes_0.1.14              bit_4.0.4                  
##  [77] tidyselect_1.1.0            plyr_1.8.6                 
##  [79] magrittr_1.5                bookdown_0.21              
##  [81] R6_2.4.1                    magick_2.5.0               
##  [83] generics_0.0.2              DelayedArray_0.16.0        
##  [85] DBI_1.1.0                   pillar_1.4.6               
##  [87] withr_2.3.0                 survival_3.2-7             
##  [89] RCurl_1.98-1.2              nnet_7.3-14                
##  [91] tibble_3.0.4                crayon_1.3.4               
##  [93] BiocFileCache_1.14.0        rmarkdown_2.5              
##  [95] jpeg_0.1-8.1                progress_1.2.2             
##  [97] grid_4.0.3                  data.table_1.13.2          
##  [99] blob_1.2.1                  ModelMetrics_1.2.2.2       
## [101] digest_0.6.27               openssl_1.4.3              
## [103] munsell_0.5.0               askpass_1.1

References

Gaidatzis, Dimos, Lukas Burger, Maria Florescu, and Michael B Stadler. 2015. “Analysis of Intronic and Exonic Reads in Rna-Seq Data Characterizes Transcriptional and Post-Transcriptional Regulation.” Nat Biotechnol 33:722–29. https://doi.org/10.1038/nbt.3269.

Hah, Nasun, Charles G. Danko, Leighton Core, Joshua J. Waterfall, Adam Siepel, John T. Lis, and W. Lee Kraus. 2011. “A Rapid, Extensive, and Transient Transcriptional Response to Estrogen Signaling in Breast Cancer Cells.” Cell 13:622–34. https://doi.org/10.1016/j.cell.2011.03.042.

Hah, Nasun, Shino Murakami, Anusha Nagari, Charles G. Danko, and W. Lee Kraus. 2013. “Enhancer Transcripts Mark Active Estrogen Receptor Binding Sites.” Genome Res 23 (8):1210–23. https://doi.org/10.1101/gr.152306.112.

Mitchell, Jennifer A., Leuan Clay, David Umlauf, Chih-yu Chen, Catherine A. Moir, Christopher H. Eskiw, Stefan Schoenfelder, Lyubomira Chakalova, Takashi Nagano, and Peter Fraser. 2012. “Nuclear Rna Sequencing of the Mouse Erythroid Cell Transcriptome.” PLoS One 7. https://doi.org/10.1371/journal.pone.0049274.

Morozova, Olena, Martin Hirst, and Marco A. Marra. 2009. “Applications of New Sequencing Technologies for Transcriptome Analysis.” Annual Review of Genomics and Human Genetics 10:135–51. https://doi.org/10.1146/annurev-genom-082908-145957.

Wang, Zhong, Mark Gerstein, and Michael Snyder. 2010. “RNA-Seq: A Revolutionary Tool for Transcriptomics.” Nat Rev Genet 10 (1):57–63. https://doi.org/10.1038/nrg2484.

Werner, Michael S., and Alexander J. Ruthenburg. 2015. “Nuclear Fractionation Reveals Thousands of Chromatin-Tethered Noncoding Rnas Adjacent to Active Genes.” Cell Rep 12 (7):1089–98. https://doi.org/10.1016/j.celrep.2015.07.033.

Zaghlool, Ammar, Adam Ameur, Linnea Nyberg, Jonatan Halvardson, Manfred Grabherr, Lucia Cavelier, and Lars Feuk. 2013. “Efficient Cellular Fractionation Improves Rna Sequencing Analysis of Mature and Nascent Transcripts from Human Tissues.” BMC Biotechnol 13. https://doi.org/10.1186/1472-6750-13-99.

Zhang, Yong, Tao Liu, Clifford A Meyer, Jérôme Eeckhoute, David S Johnson, Bradley E Bernstein, Chad Nusbaum, et al. 2008. “Model-Based Analysis of Chip-Seq (Macs).” Genome Biol 9. https://doi.org/10.1186/gb-2008-9-9-r137.