Predicting m6A sites from miCLIP2 data with m6Aboost
April 26, 2022
Abstract
Precisely identify the m6A sites from miCLIP data is still a challenge in the epigenomics field. Here we present a workflow to determine the m6A sites from the miCLIP2 data set.Package
m6Aboost 1.2.0
1 Introduction
N6-methyladenosine (m6A) is the most abundant internal modification in mRNA. It impacts many different aspects of an mRNA’s life, e.g. nuclear export, translation, stability, etc.
m6A individual-nucleotide resolution UV crosslinking and immunoprecipitation (miCLIP) and the improved miCLIP2 are m6A antibody-based methods that allow the transcriptome-wide mapping of m6A sites at a single-nucleotide resolution (Körtel et al. 2021)(Linder et al. 2015). In brief, UV crosslinking of the m6A antibody to the modified RNA leads to truncation of reverse transcription or C-to-T transitions in the case of readthrough. However, due to the limited specificity and high cross-reactivity of the m6A antibodies, the miCLIP data comprise a high background signal, which hampers the reliable identification of m6A sites from the data.
For accurately detecting m6A sites, we implemented an AdaBoost-based machine learning model (m6Aboost) for classifying the miCLIP2 peaks into m6A sites and background signals (Körtel et al. 2021). The model was trained on high-confidence m6A sites that were obtained by comparing wildtype and Mettl3 knockout mouse embryonic stem cells (mESC) lacking the major methyltransferase Mettl3. For classification, the m6Aboost model uses a series of features, including the experimental miCLIP2 signal (truncation events and C-to-T transitions) as well as the transcript region (5’UTR, CDS, 3’UTR) and the nucleotide sequence in a 21-nt window around the miCLIP2 peak.
The package m6Aboost includes the trained model and the functionalities to prepare the data, extract the required features and predict the m6A sites.
2 Installation
The m6Aboost package is available at
https://bioconductor.org and can be
installed via BiocManager::install:
if (!require("BiocManager"))
install.packages("BiocManager")
BiocManager::install("m6Aboost")A package only needs to be installed once. Load the package into an R session with
library(m6Aboost)3 Pre-requisite
The workflow described herein is based on our published paper (Körtel et al. 2021). Thus we expect the user to preprocess the miCLIP2 data based on the preprocessing pipeline in our article (Körtel et al. 2021). In brief, the preprocessing steps include basic processing of the sequencing reads, such as quality filtering, barcode handling, mapping, generation of the single nucleotide crosslink and the C to T transition bigWig file. After the preprocessing, we expect the user to do the peak calling with the tool PureCLIP (Krakau, Richard, and Marsico 2017).
Overview of the m6Aboost preprocessing workflow.
Note: If you use m6Aboost in published research, please cite:
Körtel, Nadine#, Cornelia Rückle#, You Zhou#, Anke Busch, Peter Hoch-Kraft, FX Reymond Sutandy, Jacob Haase, et al. 2021. “Deep and accurate detection of m6A RNA modifications using miCLIP2 and m6Aboost machine learning.” bioRxiv. https://doi.org/10.1101/2020.12.20.423675.
4 The m6Aboost workflow
Overview of the m6Aboost workflow.
5 Reproducibility filtering
In order to increase the reproducibility of the prediction result, we suggest the user to keep the peaks which present in at least two replicates for the following analysis.
6 Loading the test data set
The package includes a test data set which allows to test all the functions
in the m6Aboost package. The test data set comprises a subset of the
miCLIP2 peak calling result from wildtype mESC cells (Körtel et al. 2021),
including 1,200 peaks (PureCLIP) in three different genes
(ENSMUSG00000026478.14, ENSMUSG00000031960.14, ENSMUSG00000035569.17). These
are encoded in a GRanges object with the following metadata:
- The column
Reads_meancontains the mean value of truncation events at a given peak in three replicate experiments.
- The column
CtoTcontains the mean value of C-to-T transitions that are associcated with a given peak. (Note that the C-to-T transitions at an m6A site do occur on the C flanking the modified A in the DRACH motif.)
In addition, the package includes the annotation file test_gff3 which
is a subset of the full annotation in gff3 format downloaded from
GENCODE. The test truncation and C-to-T
transition bigWig files are a subset of the miCLIP2 signal from a wildtype
mESC cells (Körtel et al. 2021).
library(m6Aboost)
## Load the test data
testpath <- system.file("extdata", package = "m6Aboost")
test_gff3 <- file.path(testpath, "test_annotation.gff3")
test <- readRDS(file.path(testpath, "test.rds"))
test## GRanges object with 1200 ranges and 2 metadata columns:
## seqnames ranges strand | CtoTmean WTmean
## <Rle> <IRanges> <Rle> | <numeric> <numeric>
## [1] chr1 153332740 - | 0 17.6667
## [2] chr1 153332699 - | 0 92.3333
## [3] chr1 153332489 - | 0 28.3333
## [4] chr1 153332481 - | 0 20.6667
## [5] chr1 153332480 - | 0 15.0000
## ... ... ... ... . ... ...
## [1196] chr8 111056683 + | 0.00000 9.66667
## [1197] chr8 111057030 + | 0.00000 61.00000
## [1198] chr8 111057335 + | 2.33333 10.33333
## [1199] chr8 111057424 + | 1.66667 11.00000
## [1200] chr8 111057515 + | 0.00000 23.66667
## -------
## seqinfo: 22 sequences from an unspecified genome; no seqlengths7 Read count assignment
After peak calling, the user needs to assign the number of truncation events
and C-to-T transitions that are associated with each peak. In
m6Aboost, we provide two functions, truncationAssignment and
CtoTAssignment, to assign these counts from the imported bigWig files. Note
that the values are first assigned separated for each replicate and then
averaged into a mean count for the subsequent feature extraction by
preparingData.
## truncationAssignment allows to assign the number of truncation events
## The input should be a GRanges object with the peaks and bigWig files
## with the truncation events (separated per strand)
truncationBw_p <- file.path(testpath, "truncation_positive.bw")
truncationBw_n <- file.path(testpath, "truncation_negative.bw")
test <- truncationAssignment(test,
bw_positive=truncationBw_p,
bw_negative=truncationBw_n,
sampleName = "WT1")
## CtoTAssignment allows to assign the number of C-to-T transitions
ctotBw_p <- file.path(testpath, "C2T_positive.bw")
ctotBw_n <- file.path(testpath, "C2T_negative.bw")
test <- CtoTAssignment(test,
bw_positive=ctotBw_p,
bw_negative=ctotBw_n,
sampleName = "CtoT_WT1")Next, the user needs to calculate the mean number of truncation events and C-to-T transition read counts across the replicates.
## E.g. for two replicates, this can be calculated as
peak$WTmean <- (peak$WT1 + peak$WT2)/28 Extract features for the m6Aboost model
For training the m6Aboost model, we used the surrounding nucleotide sequence,
the transcript region harbouring the peak, the number of C-to-T transitions and
the relative signal strength. The function preparingData allows to extract
these features. Please note the function preparingData requires
an annotation file that was downloaded from
GENCODE.
## Extract the features for the m6Aboost prediction
test <- preparingData(test, test_gff3, colname_reads="WTmean",
colname_C2T="CtoTmean")
test## GRanges object with 1200 ranges and 10 metadata columns:
## seqnames ranges strand | CtoTmean WTmean WT1 CtoT_WT1
## <Rle> <IRanges> <Rle> | <numeric> <numeric> <numeric> <numeric>
## [1] chr1 153332740 - | 0 17.6667 12 0
## [2] chr1 153332699 - | 0 92.3333 54 0
## [3] chr1 153332489 - | 0 28.3333 18 0
## [4] chr1 153332481 - | 0 20.6667 10 0
## [5] chr1 153332480 - | 0 15.0000 5 0
## ... ... ... ... . ... ... ... ...
## [1196] chr8 111056683 + | 0.00000 9.66667 6 0
## [1197] chr8 111057030 + | 0.00000 61.00000 39 0
## [1198] chr8 111057335 + | 2.33333 10.33333 8 0
## [1199] chr8 111057424 + | 1.66667 11.00000 8 0
## [1200] chr8 111057515 + | 0.00000 23.66667 10 0
## gene_id RSS CtoT UTR3 UTR5
## <character> <numeric> <numeric> <character> <character>
## [1] ENSMUSG00000026478.14 1.101785 0 NO YES
## [2] ENSMUSG00000026478.14 5.758384 0 NO YES
## [3] ENSMUSG00000026478.14 1.767013 0 NO NO
## [4] ENSMUSG00000026478.14 1.288880 0 NO NO
## [5] ENSMUSG00000026478.14 0.935478 0 NO NO
## ... ... ... ... ... ...
## [1196] ENSMUSG00000031960.14 0.209647 0.00000 YES NO
## [1197] ENSMUSG00000031960.14 1.322945 0.00000 YES NO
## [1198] ENSMUSG00000031960.14 0.224105 2.33333 YES NO
## [1199] ENSMUSG00000031960.14 0.238564 1.66667 YES NO
## [1200] ENSMUSG00000031960.14 0.513274 0.00000 YES NO
## CDS
## <character>
## [1] NO
## [2] NO
## [3] YES
## [4] YES
## [5] YES
## ... ...
## [1196] YES
## [1197] YES
## [1198] NO
## [1199] NO
## [1200] NO
## -------
## seqinfo: 22 sequences from an unspecified genome; no seqlengths9 Prediction of m6A sites
The function m6Aboost performs the prediction, i.e. the classification of the
miCLIP2 peaks into m6A sites and background. As input, the function m6Aboost
uses the output object from the function preparingData. In addition, the user
needs to specify the name of the BSgenome package associated with the species
used for the experiment. Please note that the BSgenome package, which
contains the sequence information, should be downloaded from Bioconductor in
advance.
9.1 Normalization of numerical features
Application of the machine learning model to new data set requires that the
data were generated by the same protocol and thus show an independent and
identical distribution. The m6Aboost model includes two numerical features from
the miCLIP2 data, namely relative signal strength and C-to-T transitions, which
could systematically vary between experiments. Since in the training set, both
features approximated a Poisson distribution. We recommend the user to
normalized the values of each features in the input samples by the ratio of
the mean for this feature between the input dataset and the training set. For
doing this normalization, user just need to set the parameter
normalization = TRUE.
## Note that since the test data set contains only a tiny fraction of the real
## data, and a part of the test data belongs to the training set. Here for
## preventing the unnecessary value change, we set the normalization to FALSE.
out <- m6Aboost(test, "BSgenome.Mmusculus.UCSC.mm10", normalization = FALSE)
out## GRanges object with 272 ranges and 12 metadata columns:
## seqnames ranges strand | CtoTmean WTmean WT1 CtoT_WT1
## <Rle> <IRanges> <Rle> | <numeric> <numeric> <numeric> <numeric>
## [1] chr1 153332308 - | 15 61.33333 24 0
## [2] chr1 153262625 - | 0 35.00000 11 0
## [3] chr1 153255240 - | 0 6.00000 4 0
## [4] chr1 153255239 - | 0 15.00000 8 0
## [5] chr1 153255209 - | 0 9.66667 4 0
## ... ... ... ... . ... ... ... ...
## [268] chr8 111050361 + | 0.00000 14.66667 11 0
## [269] chr8 111054149 + | 0.00000 26.00000 28 0
## [270] chr8 111056349 + | 0.00000 4.66667 5 0
## [271] chr8 111057335 + | 2.33333 10.33333 8 0
## [272] chr8 111057424 + | 1.66667 11.00000 8 0
## gene_id RSS CtoT UTR3 UTR5
## <character> <numeric> <numeric> <character> <character>
## [1] ENSMUSG00000026478.14 3.825064 15 NO NO
## [2] ENSMUSG00000026478.14 2.182781 0 NO NO
## [3] ENSMUSG00000026478.14 0.374191 0 NO NO
## [4] ENSMUSG00000026478.14 0.935478 0 NO NO
## [5] ENSMUSG00000026478.14 0.602863 0 NO NO
## ... ... ... ... ... ...
## [268] ENSMUSG00000031960.14 0.318085 0.00000 NO NO
## [269] ENSMUSG00000031960.14 0.563878 0.00000 NO NO
## [270] ENSMUSG00000031960.14 0.101209 0.00000 YES YES
## [271] ENSMUSG00000031960.14 0.224105 2.33333 YES NO
## [272] ENSMUSG00000031960.14 0.238564 1.66667 YES NO
## CDS class prob
## <character> <character> <numeric>
## [1] YES YES 0.7286185
## [2] YES NO 0.1393625
## [3] YES NO 0.0343775
## [4] YES NO 0.1547813
## [5] YES NO 0.1594595
## ... ... ... ...
## [268] YES NO 0.199353
## [269] YES NO 0.177922
## [270] NO NO 0.058019
## [271] NO YES 0.652307
## [272] NO YES 0.719683
## -------
## seqinfo: 22 sequences from an unspecified genome; no seqlengthsThe function m6Aboost returns a GRanges object with two additional metadata
columns:
- The column class provides the classification result, i.e. whether the
m6Aboost model predicts this site to be an m6A site or not.
- The column prob includes the prediction score for a given site.
All sites with a prediction score > 0.5 are considered as m6A sites, although a more stringent cutoff can be applied if needed.
9.2 Access the machine learning model m6Aboost
The raw model m6Aboost is stored in ExperimentHub. Users can access the raw model with the following code:
## firstly user need to load the ExperimentHub
library(ExperimentHub)
eh <- ExperimentHub()
## "EH6021" is the retrieve record of the m6Aboost
model <- eh[["EH6021"]]
## here shows more information about the stored model
query(eh, "m6Aboost")## ExperimentHub with 1 record
## # snapshotDate(): 2022-04-19
## # names(): EH6021
## # package(): m6Aboost
## # $dataprovider: Zarnack's lab
## # $species: NA
## # $rdataclass: boosting
## # $rdatadateadded: 2021-05-18
## # $title: The m6Aboost machine learning model
## # $description: The machine learning model which use for identify the m6A si...
## # $taxonomyid: NA
## # $genome: NA
## # $sourcetype: RDS
## # $sourceurl: https://github.com/Codezy99/m6Aboost/blob/main/m6Aboost.rds
## # $sourcesize: NA
## # $tags: c("Epigenetics", "ExperimentHubSoftware", "Genetics",
## # "Sequencing")
## # retrieve record with 'object[["EH6021"]]'10 Session info
sessionInfo()## R version 4.2.0 RC (2022-04-19 r82224)
## Platform: x86_64-pc-linux-gnu (64-bit)
## Running under: Ubuntu 20.04.4 LTS
##
## Matrix products: default
## BLAS: /home/biocbuild/bbs-3.15-bioc/R/lib/libRblas.so
## LAPACK: /home/biocbuild/bbs-3.15-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] ExperimentHub_2.4.0 AnnotationHub_3.4.0
## [3] BiocFileCache_2.4.0 dbplyr_2.1.1
## [5] BSgenome.Mmusculus.UCSC.mm10_1.4.3 BSgenome_1.64.0
## [7] rtracklayer_1.56.0 Biostrings_2.64.0
## [9] XVector_0.36.0 m6Aboost_1.2.0
## [11] GenomicRanges_1.48.0 GenomeInfoDb_1.32.0
## [13] IRanges_2.30.0 adabag_4.2
## [15] doParallel_1.0.17 iterators_1.0.14
## [17] foreach_1.5.2 caret_6.0-92
## [19] lattice_0.20-45 ggplot2_3.3.5
## [21] rpart_4.1.16 S4Vectors_0.34.0
## [23] BiocGenerics_0.42.0 BiocStyle_2.24.0
##
## loaded via a namespace (and not attached):
## [1] plyr_1.8.7 splines_4.2.0
## [3] BiocParallel_1.30.0 listenv_0.8.0
## [5] digest_0.6.29 htmltools_0.5.2
## [7] fansi_1.0.3 magrittr_2.0.3
## [9] memoise_2.0.1 recipes_0.2.0
## [11] globals_0.14.0 gower_1.0.0
## [13] matrixStats_0.62.0 hardhat_0.2.0
## [15] colorspace_2.0-3 blob_1.2.3
## [17] rappdirs_0.3.3 xfun_0.30
## [19] dplyr_1.0.8 crayon_1.5.1
## [21] RCurl_1.98-1.6 jsonlite_1.8.0
## [23] survival_3.3-1 glue_1.6.2
## [25] gtable_0.3.0 ipred_0.9-12
## [27] zlibbioc_1.42.0 DelayedArray_0.22.0
## [29] future.apply_1.9.0 scales_1.2.0
## [31] DBI_1.1.2 Rcpp_1.0.8.3
## [33] xtable_1.8-4 bit_4.0.4
## [35] lava_1.6.10 prodlim_2019.11.13
## [37] httr_1.4.2 ellipsis_0.3.2
## [39] pkgconfig_2.0.3 XML_3.99-0.9
## [41] nnet_7.3-17 sass_0.4.1
## [43] utf8_1.2.2 tidyselect_1.1.2
## [45] rlang_1.0.2 reshape2_1.4.4
## [47] later_1.3.0 AnnotationDbi_1.58.0
## [49] munsell_0.5.0 BiocVersion_3.15.2
## [51] tools_4.2.0 cachem_1.0.6
## [53] cli_3.3.0 generics_0.1.2
## [55] RSQLite_2.2.12 evaluate_0.15
## [57] stringr_1.4.0 fastmap_1.1.0
## [59] yaml_2.3.5 ModelMetrics_1.2.2.2
## [61] knitr_1.38 bit64_4.0.5
## [63] purrr_0.3.4 KEGGREST_1.36.0
## [65] future_1.25.0 nlme_3.1-157
## [67] mime_0.12 compiler_4.2.0
## [69] filelock_1.0.2 curl_4.3.2
## [71] png_0.1-7 interactiveDisplayBase_1.34.0
## [73] tibble_3.1.6 bslib_0.3.1
## [75] stringi_1.7.6 Matrix_1.4-1
## [77] vctrs_0.4.1 pillar_1.7.0
## [79] lifecycle_1.0.1 BiocManager_1.30.17
## [81] jquerylib_0.1.4 data.table_1.14.2
## [83] bitops_1.0-7 httpuv_1.6.5
## [85] R6_2.5.1 BiocIO_1.6.0
## [87] bookdown_0.26 promises_1.2.0.1
## [89] parallelly_1.31.1 codetools_0.2-18
## [91] MASS_7.3-57 assertthat_0.2.1
## [93] SummarizedExperiment_1.26.0 rjson_0.2.21
## [95] withr_2.5.0 GenomicAlignments_1.32.0
## [97] Rsamtools_2.12.0 GenomeInfoDbData_1.2.8
## [99] grid_4.2.0 timeDate_3043.102
## [101] class_7.3-20 rmarkdown_2.14
## [103] MatrixGenerics_1.8.0 pROC_1.18.0
## [105] Biobase_2.56.0 shiny_1.7.1
## [107] lubridate_1.8.0 restfulr_0.0.13References
Körtel, Nadine#, Cornelia Rückle#, You Zhou#, Anke Busch, Peter Hoch-Kraft, FX Reymond Sutandy, Jacob Haase, et al. 2021. “Deep and accurate detection of m6A RNA modifications using miCLIP2 and m6Aboost machine learning.” Nucleic Acids Research. https://doi.org/10.1093/nar/gkab485.
Krakau, Sabrina, Hugues Richard, and Annalisa Marsico. 2017. “PureCLIP: capturing target-specific protein RNA interaction footprints from single-nucleotide CLIP-seq data.” Genome Biology 18 (1). https://doi.org/10.1186/s13059-017-1364-2.
Linder, Bastian, Anya V Grozhik, Anthony O Olarerin-George, Cem Meydan, Christopher E Mason, and Samie R Jaffrey. 2015. “Single-nucleotide-resolution mapping of m6A and m6Am throughout the transcriptome.” Nat Methods. https://doi.org/10.1038/nmeth.3453.