DEWSeq 1.6.0
Wolfgang Huber and Matthias Hentze for mentoring, advice and discussion. Benjamin Lang and Gian Tartaglia for great help with functional analysis and benchmarking, as well as feedback on the vignette. Ina Huppertz for helpful feedback and language improvement on the vignette. Mike Love, Simon Anders, Bernd Klaus and Frederick Ziebell for comments and discussion.
RNA-binding proteins (RBPs) play a key role in the life-time of RNAs. They are involved in RNA synthesis, stability, degradation, transport and translation and add an important layer of regulation in the cell. Over 1,900 murine and over 1,400 human RBPs were detected in different high-throughput detection studies, many of them without known RNA-binding function (Hentze et al. 2018).
It is of great interest to detect an RBP’s binding sites to study the underlying mechanism of its regulatory potential. Individual nucleotide resolution crosslinking and immunoprecipitation (iCLIP) (König et al. 2010) and the further enhanced CLIP (eCLIP) protocol (Van Nostrand et al. 2016) rely on UV crosslinking inducing covalent bonds of RNA and proteins in close proximity. When reverse transcribing the RNA fragment bound to the protein, a majority of the time the reverse transcriptase will terminate at the crosslink site. Although eCLIP introduces updates in chemistry, the use of a size-matched input (SMI) control sample is an essential addition to the protocol which can be also adapted to iCLIP or similar protocols.
In iCLIP and eCLIP, truncation events are extracted as one nucleotide position next to the cDNA fragment (aligned read). In the classical protocols real truncations cannot be distinguished from read-through reads or other reads coming from otherwise truncated reads, which might be caused by RNA modifications or the crosslinking sites of other proteins. This might be different for each individual proteins (and the remaining polypeptide of the digested protein). Other protocols like HITS-CLIP and PAR-CLIP (Hafner et al. 2010) rely exclusively on read-through events (although using other reverse transcriptases). While hybrid approaches exist, the technical difficulty of these protocols requires many optimizations steps, which makes them rather hard to combine.
In summary, iCLIP and eCLIP protocols provide count data for single-nucleotide positions which might be the result of many heuristic events. These are described in the next chapter.
Unlike transcription factors, RNA-binding proteins have many different binding modes (Hentze et al. 2018), some bind in a sequence specific manner, some have preference for structures (like stem-loops), some prefer to bind RNA modifications, others are mostly found at UTRs. A large portion of RBPs do not have a known RNA-binding domain and bind using disordered regions with unknown target preferences.
All these different binding modes can result in different crosslinking patterns, hence different truncation event patterns.
UV crosslinking at 254 nm is used to induce covalent chemical bonds within very close proximity of RNAs to proteins. This provides a “snapshot” of the RNA binding event and ensures that no disassociation of the RNA-protein complexes takes place in further steps. Typically, crosslinking with Stratlinker(R) UV crosslinkers has a limited efficiency of around 1% or less (depending on UV lamp intensity and time of irradiation). Newer crosslinkers with lasers reach up to 5-20% crosslinking efficiency depending on the protein and experimental setup.
Depending on the type of protein and the binding mode, this can result in different crosslinking patterns for each protein. While crosslinking affects all RBPs, the immunoprecipitation (IP) step should only enrich for RNA fragments bound to the RBP of interest. In turn, this means that crosslinked samples show a certain background which is different to normal expression patterns measured by RNA-sequencing.
Immunoprecipitating your (crosslinked) protein and performing sequencing without any RNA digestion is called RIP-seq (without crosslinking) or HITS-CLIP (with crosslinking). These studies nicely show that the enrichment of your protein with antibodies, recognizing protein tags or native epitops will enrich RNA fragments bound to the protein. However, the immunoprecipitated (’IP’ed) samples closely resemble a full transcriptomic background. Depending on the specificity of your antibody, purification method (e.g. type of beads) and many other factors, this degree of resemblance can vary for different proteins. The authors do not know any cases where IP steps exclusively purify targets. If you find any such cases, please contact the authors, for further discussions (we would really like to know).
If you take a closer look at the first iCLIP analysis protocols (and lot of protocols to date), you will notice that snoRNAs, lincRNAs and other short RNAs are excluded from the analysis, although they make up a majority of the signal. They were either treated as contaminants and dismissed right away, or more recently it was recommended to compare the i/eCLIP data with an orthogonal dataset like an RNA-seq knockout data and only look at the regions of interest that are common in both. As already mentioned, it was shown that size-matched input (SMI) controls can control for different artifacts coming from crosslinking and library preparation (Van Nostrand et al. 2016). After applying an enrichment correction it is possible to detect lincRNA, snoRNA binders.
As we previously discussed, crosslinked samples do have a different background
compared to expressionlevels (as detected with RNA-seq). Therefore, it is of great
importance to compare the IPed sample to a proper input control.
SMI controls can be easily applied to any CLIP protocols.
To narrow down the target RNAs to the binding site region, RNase is used to digest RNA not protected by your protein. Using different RNase concentrations will result in different RNA fragment lengths. In addition, it was shown that the crosslinking pattern will be affected by longer RNA fragments, which occurs as a result of the protein of interest binding closely together, protecting a larger RNA region from digestion (Haberman et al. 2017).
iCLIP and eCLIP depend on the termination of the reverse transcriptase at crosslink sites. In contrast, HITS-CLIP and PAR-CLIP need reverse transcriptases reading through the crosslink sites. Both methods use different reverse transcriptases with different likelihood of truncating or reading through (Hocq et al. 2018). The chance of this event will also be affected by the polypeptide which is left after protein digestion of the protein of interest.
“Read-through” events can truncate at crosslink sites from the same protein, a different protein, an RNA modification, any other unknown events or finish at the end of an RNA fragment. Due to the low efficiency the crosslinking of multiple proteins is usually less likely, however not improbable.
Early truncations can also occur on a long RNA fragment with multiple crosslink sites from different proteins or multiple crosslink sites from the same protein. Again, although low crosslinking efficiency reduces the chance of having multiple crosslink sites at once, those events occur stochastically.
Because of the properties of i/eCLIP data described above, we use a sliding window approach (where different window and step sizes can be chosen in the preprocessing step) and test for significant enrichment in the IP over the control. For this, we extended DESeq2 for a one-sided test: first, we filter windows with negative wald test statistic values (windows with negative log2FoldChange) followed a right-tailed test on the remaining windows. The advantage is that the filtering step increases performance when testing millions of windows.
In contrast to large fragments detected in ChIP-seq, the truncation sites are only one nucleotide. To combine windows in ChIP-seq dataset, csaw defines binding regions which are then subjected to multiple hypothesis correction. In ChIP-seq many windows share coverage from large cDNA fragments and therefore csaw performs multiple hypothesis correction on binding site with the SIMES method (SIMES 1986). SIMES joins coverage blocks and treats them as one entity for multiple hypothesis correction. In large, SIMES punished p-values for long stretches of overlapping windows. Please refer to the csawUserGuide for detailed description.
Single-nucleotide truncation events do not have the same property as cDNA fragments and the transcriptomic background for each gene in eCLIP or iCLIP often results in large coverage blocks. As seen in the picture below, the windows do only share information with the overlapping windows. Therefore, DEWSeq corrects the p-value for each overlapping window with the number of overlaps using Bonferroni correction.
We then apply FDR correction on the p-values corrected for overlaps using Bonferroni correction.
Most importantly, replicates and input controls are required for the significance testing for DEWSeq. Currently, we only allow the use of DESeq2’s wald test and LRT test. Please contact the authors if you would like to use more complex models in your analysis.
Typically, CLIP-seq sa mples show a background which comes from a mix of various transcripts with different expression levels. Additionally, UV crosslinking does not affect RNAs in a linear fashion, therefore an appropriate input control is needed for the analysis of eCLIP/iCLIP data. Unfortunately, total RNA-seq does not reflect the UV crosslinked background. Also neither IgG-, empty beads, nor similar controls seem to be appropriate input controls. Mostly, IgG and empty bead controls suggest that the IP input is very clean, which is not what is seen in the data. IP sample often have a broad background signal and bias towards sno and lincRNAs, as well as other RNA types. Often, a majority of the reads in the IP come from such background, although the IgG and empty bead control samples cannot correct for that.
We recommend size-matched input controls (SMI) controls as performed in the eCLIP protocol (Van Nostrand et al. 2016). This type of input control is not protocol specific, therefore can be easily adapted to iCLIP or other CLIP studies. It proves useful to exclude many false positives and can control for truncations caused by interferences.
Empty beads controls are usually crosslinked and control for unspecific background caused through beads, IgG controls are usually crosslinked and control additionally for background caused by the invariable region of the antibody. No crosslink (noCL) controls for background caused by the protein and the functional antibody. SMI controls are crosslinked but not enriched, it controls among others for the background caused by crosslinked RNA.
DEWSeq uses SMI controls to calculate enrichment in the IPed samples. We recommend to use IgG or empty beads controls to flag (expression dependent) regions for optional removal.
Biological replicates in high-throughput studies are needed in the significance testing process to ensure reproducibility of the results. Please refer to the csaw package for discussion on this topic.
In general, we recommend to ask yourself how many replicates would you use when performing RNA-seq. Usually, the answer is, at least three for the sample and three for the control (depending on many experimental factors which also will apply for your CLIP study). As a rule of thumb, your number of replicates should not be less if you are interested in an transcriptome wide analysis of your RNA-binding protein.
DEWSeq needs a countmatrix and preprocessed annotation. The preprocessing is done in two general steps.
From fastq to bam files
eCLIP or iCLIP data preprocessing can be time-consuming. Generally, the pre-processing steps are:
Eventually, you will end up with sorted .bam files.
From bam and gff3 files to count matrix and annotation
htseq-clip is the preprocessing pipeline developed for the use for DEWSeq, but other tools may be used. htseq-clip performs the following steps:
In addition, htseq-clip can also:
htseq-clip can be installed via Python package installer as:
Additional details on htseq-clip functions and required input files are described here
All you need is
An earlier, Python 2.7 compatibile version of htseq-clip is provided as a singularity container image and comes with a Snakemake workflow for processing on local computers, servers, or computer clusters.
Tip: You can use Sailfish, Kallisto, RSEM, Salmon or any (pseudo)aligner to get an estimation of the different expression levels of transcripts. Use this to filter the annotation before flattening it.
DEWSeq uses a sliding window approach with DESeq2 for enriched binding sites.
DEWSeq can be installed from Bioconductor as follows:
First step is to load the required libraries with this command
This example uses ENCODE SLBP data set, contains two IP and one input control (SMI) samples which is not ideal, however it is small enough as test data.
countFile <- file.path(system.file('extdata',package='DEWSeq'),'SLBP_K562_w50s20_counts.txt.gz')
annotationFile <- file.path(system.file('extdata',package='DEWSeq'),'SLBP_K562_w50s20_annotation.txt.gz')
We need to read in the
which can be prepared with htseq-clip (see above)
For this, we recommend data.table
or tidyr
packages to read in the data swiftly.
In our experience, data.table is significantly faster than tidyr.
First we read in counts
or alternatively with tidyr (slower):
Then we read in the annotation data frame
or alternatively with tidyr (slower):
Let’s take a look at the data files
## # A tibble: 6 x 4
## unique_ID IP1 IP2 SMI
## <chr> <dbl> <dbl> <dbl>
## 1 ENSG00000225630.1:exon0001W00014 5 1 4
## 2 ENSG00000225630.1:exon0001W00015 4 0 4
## 3 ENSG00000248527.1:exon0001W00013 2 1 5
## 4 ENSG00000248527.1:exon0001W00014 12 1 54
## 5 ENSG00000248527.1:exon0001W00015 12 1 79
## 6 ENSG00000248527.1:exon0001W00016 9 0 59
## [1] 28569 4
## chromosome begin end width strand unique_id
## 1: chr1 629900 629949 50 + ENSG00000225630.1:exon0001W00014
## 2: chr1 629920 629969 50 + ENSG00000225630.1:exon0001W00015
## 3: chr1 633936 633985 50 + ENSG00000248527.1:exon0001W00013
## 4: chr1 633956 634005 50 + ENSG00000248527.1:exon0001W00014
## 5: chr1 633976 634025 50 + ENSG00000248527.1:exon0001W00015
## 6: chr1 633996 634045 50 + ENSG00000248527.1:exon0001W00016
## gene_id gene_name gene_type gene_region Nr_of_region
## 1: ENSG00000225630.1 MTND2P28 unprocessed_pseudogene exon 1
## 2: ENSG00000225630.1 MTND2P28 unprocessed_pseudogene exon 1
## 3: ENSG00000248527.1 MTATP6P1 unprocessed_pseudogene exon 1
## 4: ENSG00000248527.1 MTATP6P1 unprocessed_pseudogene exon 1
## 5: ENSG00000248527.1 MTATP6P1 unprocessed_pseudogene exon 1
## 6: ENSG00000248527.1 MTATP6P1 unprocessed_pseudogene exon 1
## Total_nr_of_region window_number
## 1: 1 14
## 2: 1 15
## 3: 1 13
## 4: 1 14
## 5: 1 15
## 6: 1 16
## [1] 28569 13
Finally we have to create a sample description
colData <- data.frame(
row.names = colnames(countData)[-1], # since the first column is unique_id
type = factor(
c(rep("IP", 2), ## change this accordingly
rep("SMI", 1)), ##
levels = c("IP", "SMI"))
)
This function will parse the annotation file and create a DESeq object
ddw <- DESeqDataSetFromSlidingWindows(countData = countData,
colData = colData,
annotObj = annotationData,
tidy = TRUE,
design = ~type)
## Warning in DESeqDataSetFromSlidingWindows(countData = countData, colData
## = colData, : countData is a data.table or tibble object, converting it to
## data.frame. First column will be used as rownames
## class: DESeqDataSet
## dim: 28569 3
## metadata(1): version
## assays(1): counts
## rownames(28569): ENSG00000225630.1:exon0001W00014
## ENSG00000225630.1:exon0001W00015 ... ENSG00000210196.2:exon0001W00001
## ENSG00000210196.2:exon0001W00002
## rowData names(8): unique_id gene_id ... Total_nr_of_region
## window_number
## colnames(3): IP1 IP2 SMI
## colData names(1): type
This will return a DESeq object with the coordinates and annotation stores as rowAnnotation.
Prefiltering becomes even more important with large numbers of sliding windows. Please find more details here: DESeq2 vignette on pre-filtering
For library normalisation, rather than modifying the raw counts, DESeq2 estimates size factors which are incorportated in the anlaysis. For more information, please refer to the DESeq2 vignette.
The standard procedure to estimate size factors is
## IP1 IP2 SMI
## 0.8298265 0.4566711 2.7734451
CLIP data often shows high expression levels of small RNAs like snoRNAs, lincRNAs etc, therefore you might want to normalise based on protein coding genes only.
ddw_mRNAs <- ddw[ rowData(ddw)[,"gene_type"] == "protein_coding", ]
ddw_mRNAs <- estimateSizeFactors(ddw_mRNAs)
sizeFactors(ddw) <- sizeFactors(ddw_mRNAs)
sizeFactors(ddw)
## IP1 IP2 SMI
## 0.8298265 0.4807499 2.7734451
In general, when there is asymmetry in the data, like enrichment in IP over control, it is a good pratice to call differentially expressed features with DESeq2 and then exclude them for normalisation. This might be adjusted for your proteins.
First we identify significant windows in IP or the controls.
ddw_tmp <- ddw
ddw_tmp <- estimateDispersions(ddw_tmp, fitType = "local", quiet = TRUE)
ddw_tmp <- nbinomWaldTest(ddw_tmp)
tmp_significant_windows <-
results(ddw_tmp,
contrast = c("type", "IP", "SMI"),
tidy = TRUE,
filterFun = ihw) %>%
dplyr::filter(padj < 0.05) %>%
.[["row"]]
rm("ddw_tmp")
Then we exclude those for the estimation of the size factors
ddw_mRNAs <- ddw_mRNAs[ !rownames(ddw_mRNAs) %in% tmp_significant_windows, ]
ddw_mRNAs <- estimateSizeFactors(ddw_mRNAs)
sizeFactors(ddw) <- sizeFactors(ddw_mRNAs)
rm( list = c("tmp_significant_windows", "ddw_mRNAs"))
sizeFactors(ddw)
## IP1 IP2 SMI
## 0.7663094 0.4391752 3.0000000
We fit parametric and local fit, and decide the best fit following this Bioconductor post
First of all, define a variable called decide_fit
and based on whether the value assigned
to it is TRUE
or FALSE
dispersion fit can be decided
parametric_ddw <- estimateDispersions(ddw, fitType="parametric")
if(decide_fit){
local_ddw <- estimateDispersions(ddw, fitType="local")
}
This is the dispersion estimate for parametric fit
This is the dispersion estimate for local fit, given automated decision fitting is enabled:
This will get the residuals for either fit, only for automated decision fitting
parametricResid <- na.omit(with(mcols(parametric_ddw),abs(log(dispGeneEst)-log(dispFit))))
if(decide_fit){
localResid <- na.omit(with(mcols(local_ddw),abs(log(dispGeneEst)-log(dispFit))))
residDf <- data.frame(residuals=c(parametricResid,localResid),
fitType=c(rep("parametric",length(parametricResid)),
rep("local",length(localResid))))
summary(residDf)
}
## residuals fitType
## Min. : 0.000398 Length:31222
## 1st Qu.: 1.083571 Class :character
## Median :17.621970 Mode :character
## Mean :11.950943
## 3rd Qu.:18.696496
## Max. :19.898749
and we plot histograms of the fits
if(decide_fit){
ggplot(residDf, aes(x = residuals, fill = fitType)) +
scale_fill_manual(values = c("darkred", "darkblue")) +
geom_histogram(alpha = 0.5, position='identity', bins = 100) + theme_bw()
}
Now, we will decide for the better fit based on median
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0.001334 1.338225 18.186270 12.408586 18.948091 19.898749
if(decide_fit){
summary(localResid)
if (median(localResid) <= median(parametricResid)){
cat("chosen fitType: local")
ddw <- local_ddw
}else{
cat("chosen fitType: parametric")
ddw <- parametric_ddw
}
rm(local_ddw,parametric_ddw,residDf,parametricResid,localResid)
}else{
ddw <- parametric_ddw
rm(parametric_ddw)
}
## chosen fitType: local
The next step is to estimate the dispersions and perform the wald test with these two functions.
Please read up on fitType in the DESeq2 package and adjust according to your data set.
You can check the dispersion estimates with the following function:
In the next step we perform a one-sided signficance test, looking for enrichment in the IP samples vs the SMI controls. Also it will perform a correction for p-values of overlapping windows: The function will determine how many overlapping windows for each window there are (this can vary at the end of features, e.g. a gene) and then perform Bonferroni for each window.
These family-wise corrected windows will corrected for multiple testing with Benjamini-Hochberg.
resultWindows <- resultsDEWSeq(ddw,
contrast = c("type", "IP", "SMI"),
tidy = TRUE) %>% as_tibble
resultWindows
## # A tibble: 6,428 x 20
## chromosome begin end width strand unique_id gene_id gene_name gene_type
## <chr> <dbl> <int> <int> <chr> <chr> <chr> <chr> <chr>
## 1 chr1 6.30e5 6.30e5 49 + ENSG0000… ENSG00… MTND2P28 unproces…
## 2 chr1 1.06e6 1.06e6 49 + ENSG0000… ENSG00… AGRN protein_…
## 3 chr1 1.06e6 1.06e6 49 + ENSG0000… ENSG00… AL645608… sense_in…
## 4 chr1 1.06e6 1.06e6 49 + ENSG0000… ENSG00… AGRN protein_…
## 5 chr1 1.06e6 1.06e6 49 + ENSG0000… ENSG00… AL645608… sense_in…
## 6 chr1 1.06e6 1.06e6 49 + ENSG0000… ENSG00… AGRN protein_…
## 7 chr1 7.96e6 7.96e6 49 + ENSG0000… ENSG00… PARK7 protein_…
## 8 chr1 1.10e7 1.10e7 49 + ENSG0000… ENSG00… TARDBP protein_…
## 9 chr1 1.10e7 1.10e7 49 + ENSG0000… ENSG00… AL109811… protein_…
## 10 chr1 1.10e7 1.10e7 49 + ENSG0000… ENSG00… TARDBP protein_…
## # … with 6,418 more rows, and 11 more variables: gene_region <chr>,
## # Nr_of_region <int>, Total_nr_of_region <int>, window_number <int>,
## # baseMean <dbl>, log2FoldChange <dbl>, lfcSE <dbl>, stat <dbl>,
## # pvalue <dbl>, pSlidingWindows <dbl>, pSlidingWindows.adj <dbl>
You might be interested to correct for multiple hypothesis testing with IHW.
resultWindows[,"p_adj_IHW"] <- adj_pvalues(ihw(pSlidingWindows ~ baseMean,
data = resultWindows,
alpha = 0.05,
nfolds = 10))
Here some basic stats about the differentially expressed windows:
resultWindows <- resultWindows %>%
mutate(significant = resultWindows$p_adj_IHW < 0.01)
sum(resultWindows$significant)
## [1] 1043
1043 windows are significant.
Here are the top 20 from 105 genes:
resultWindows %>%
filter(significant) %>%
arrange(desc(log2FoldChange)) %>%
.[["gene_name"]] %>%
unique %>%
head(20)
## [1] "HIST1H4J" "HIST1H2BF" "HIST1H3J" "HIST1H4H" "HIST1H2BD" "HIST1H2BL"
## [7] "HIST1H4B" "HIST1H2BJ" "HIST1H2BG" "HIST1H2AG" "HIST1H4C" "HIST1H3D"
## [13] "HIST1H3H" "HIST1H4I" "HIST1H2BN" "HIST1H1E" "H2AFX" "HIST1H4E"
## [19] "HIST2H2BE" "HIST1H2BC"
SLBP is an histone-RNA binding protein, so we are quite happy to see that the majority of the hits are coming from histone genes.
resultWindows
contains information about the differential expression of
windows. Now we would like to combine overlapping windows to a binding region.
resultRegions <- extractRegions(windowRes = resultWindows,
padjCol = "p_adj_IHW",
padjThresh = 0.01,
log2FoldChangeThresh = 0.5) %>% as_tibble
## Warning in extractRegions(windowRes = resultWindows, padjCol = "p_adj_IHW", :
## windowRes is a data.table or tibble object, converting it to data.frame
Since we already corrected for the family wise error, we use the
extractRegions
function to combine the overlapping significant windows and
provide metrics for best and worst p-adjusted value, as well as best and
worst log2 fold change.
## # A tibble: 171 x 21
## chromosome region_begin region_end strand windows_in_region region_length
## <chr> <dbl> <int> <fct> <dbl> <int>
## 1 chr1 21727931 21728020 - 2 89
## 2 chr1 28648640 28648730 + 4 90
## 3 chr1 28648640 28648733 + 4 93
## 4 chr1 35189326 35189375 - 1 49
## 5 chr1 44721786 44721835 - 1 49
## 6 chr1 44731055 44731104 - 1 49
## 7 chr1 149813271 149813460 - 7 189
## 8 chr1 149840678 149840767 - 3 89
## 9 chr1 149842208 149842357 - 6 149
## 10 chr1 149843441 149843533 + 4 92
## # … with 161 more rows, and 15 more variables: padj_min <dbl>, padj_mean <dbl>,
## # padj_max <dbl>, log2FoldChange_min <dbl>, log2FoldChange_mean <dbl>,
## # log2FoldChange_max <dbl>, regionStartId <chr>, gene_id <fct>,
## # gene_name <chr>, gene_type <chr>, gene_region <chr>, Nr_of_region <dbl>,
## # Total_nr_of_region <dbl>, window_number <dbl>, unique_ids <chr>
171 binding regions were found which can exported as BED file so you can browse the regions in your genome browser of choice or use it for further analyses.
toBED(windowRes = resultWindows,
regionRes = resultRegions,
fileName = "enrichedWindowsRegions.bed")
Further analyses which can be done is enrichment of gene types, sequence motif analysis of the regions, secondary structure analysis of binding regions and other functional analyses.
The authors very much welcome any feedback, comments and suggestions.
## R version 4.1.0 (2021-05-18)
## Platform: x86_64-pc-linux-gnu (64-bit)
## Running under: Ubuntu 20.04.2 LTS
##
## Matrix products: default
## BLAS: /home/biocbuild/bbs-3.13-bioc/R/lib/libRblas.so
## LAPACK: /home/biocbuild/bbs-3.13-bioc/R/lib/libRlapack.so
##
## locale:
## [1] LC_CTYPE=en_US.UTF-8 LC_NUMERIC=C
## [3] LC_TIME=en_GB 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] parallel stats4 stats graphics grDevices utils datasets
## [8] methods base
##
## other attached packages:
## [1] data.table_1.14.0 forcats_0.5.1
## [3] stringr_1.4.0 dplyr_1.0.6
## [5] purrr_0.3.4 readr_1.4.0
## [7] tidyr_1.1.3 tibble_3.1.2
## [9] ggplot2_3.3.3 tidyverse_1.3.1
## [11] IHW_1.20.0 DEWSeq_1.6.0
## [13] BiocParallel_1.26.0 DESeq2_1.32.0
## [15] SummarizedExperiment_1.22.0 Biobase_2.52.0
## [17] MatrixGenerics_1.4.0 matrixStats_0.58.0
## [19] GenomicRanges_1.44.0 GenomeInfoDb_1.28.0
## [21] IRanges_2.26.0 S4Vectors_0.30.0
## [23] BiocGenerics_0.38.0 R.utils_2.10.1
## [25] R.oo_1.24.0 R.methodsS3_1.8.1
## [27] BiocStyle_2.20.0
##
## loaded via a namespace (and not attached):
## [1] colorspace_2.0-1 ellipsis_0.3.2 XVector_0.32.0
## [4] fs_1.5.0 rstudioapi_0.13 farver_2.1.0
## [7] bit64_4.0.5 AnnotationDbi_1.54.0 fansi_0.4.2
## [10] lubridate_1.7.10 xml2_1.3.2 splines_4.1.0
## [13] cachem_1.0.5 geneplotter_1.70.0 knitr_1.33
## [16] jsonlite_1.7.2 broom_0.7.6 annotate_1.70.0
## [19] dbplyr_2.1.1 png_0.1-7 BiocManager_1.30.15
## [22] compiler_4.1.0 httr_1.4.2 backports_1.2.1
## [25] assertthat_0.2.1 Matrix_1.3-3 fastmap_1.1.0
## [28] cli_2.5.0 htmltools_0.5.1.1 tools_4.1.0
## [31] gtable_0.3.0 glue_1.4.2 GenomeInfoDbData_1.2.6
## [34] Rcpp_1.0.6 slam_0.1-48 cellranger_1.1.0
## [37] jquerylib_0.1.4 vctrs_0.3.8 Biostrings_2.60.0
## [40] xfun_0.23 ps_1.6.0 rvest_1.0.0
## [43] lifecycle_1.0.0 XML_3.99-0.6 zlibbioc_1.38.0
## [46] scales_1.1.1 hms_1.1.0 RColorBrewer_1.1-2
## [49] yaml_2.2.1 memoise_2.0.0 sass_0.4.0
## [52] stringi_1.6.2 RSQLite_2.2.7 highr_0.9
## [55] genefilter_1.74.0 rlang_0.4.11 pkgconfig_2.0.3
## [58] bitops_1.0-7 evaluate_0.14 lattice_0.20-44
## [61] lpsymphony_1.20.0 labeling_0.4.2 bit_4.0.4
## [64] tidyselect_1.1.1 magrittr_2.0.1 bookdown_0.22
## [67] R6_2.5.0 magick_2.7.2 generics_0.1.0
## [70] DelayedArray_0.18.0 DBI_1.1.1 pillar_1.6.1
## [73] haven_2.4.1 withr_2.4.2 survival_3.2-11
## [76] KEGGREST_1.32.0 RCurl_1.98-1.3 modelr_0.1.8
## [79] crayon_1.4.1 fdrtool_1.2.16 utf8_1.2.1
## [82] rmarkdown_2.8 locfit_1.5-9.4 grid_4.1.0
## [85] readxl_1.3.1 blob_1.2.1 reprex_2.0.0
## [88] digest_0.6.27 xtable_1.8-4 munsell_0.5.0
## [91] bslib_0.2.5.1
Haberman, Nejc, Ina Huppertz, Jan Attig, Julian König, Zhen Wang, Christian Hauer, Matthias W. Hentze, et al. 2017. “Insights into the design and interpretation of iCLIP experiments.” Genome Biology 18 (1). https://doi.org/10.1186/s13059-016-1130-x.
Hafner, Markus, Markus Landthaler, Lukas Burger, Mohsen Khorshid, Jean Hausser, Philipp Berninger, Andrea Rothballer, et al. 2010. “Transcriptome-wide Identification of RNA-Binding Protein and MicroRNA Target Sites by PAR-CLIP.” Cell 141 (1): 129–41. https://doi.org/10.1016/j.cell.2010.03.009.
Hentze, Matthias W., Alfredo Castello, Thomas Schwarzl, and Thomas Preiss. 2018. “A brave new world of RNA-binding proteins.” https://doi.org/10.1038/nrm.2017.130.
Hocq, Rémi, Janio Paternina, Quentin Alasseur, Auguste Genovesio, and Hervé Le Hir. 2018. “Monitored ECLIP: High accuracy mapping of RNA-protein interactions.” Nucleic Acids Research. https://doi.org/10.1093/nar/gky858.
König, Julian, Kathi Zarnack, Gregor Rot, Toma Curk, Melis Kayikci, Blaž Zupan, Daniel J. Turner, Nicholas M. Luscombe, and Jernej Ule. 2010. “ICLIP reveals the function of hnRNP particles in splicing at individual nucleotide resolution.” Nature Structural and Molecular Biology 17 (7): 909–15. https://doi.org/10.1038/nsmb.1838.
SIMES, R. J. 1986. “An improved Bonferroni procedure for multiple tests of significance.” Biometrika 73 (3): 751–54. https://doi.org/10.1093/biomet/73.3.751.
Van Nostrand, Eric L., Thai B. Nguyen, Chelsea Gelboin-Burkhart, Ruth Wang, Steven M. Blue, Gabriel A. Pratt, Ashley L. Louie, and Gene W. Yeo. 2017. “Robust, cost-effective profiling of RNA binding protein targets with single-end enhanced crosslinking and immunoprecipitation (SeCLIP)” 1648: 177–200. https://doi.org/10.1007/978-1-4939-7204-3_14.
Van Nostrand, Eric L, Gabriel A Pratt, Alexander A Shishkin, Chelsea Gelboin-Burkhart, Mark Y Fang, Balaji Sundararaman, Steven M Blue, et al. 2016. “Robust transcriptome-wide discovery of RNA-binding protein binding sites with enhanced CLIP (eCLIP).” Nature Methods 13 (November 2015): 1–9. https://doi.org/10.1038/nmeth.3810.