The atena package provides methods to quantify the expression of transposable elements within R and Bioconductor.
atena 1.0.1
Transposable elements (TEs) are autonomous mobile genetic elements. They are DNA sequences that have, or once had, the ability to mobilize within the genome either directly or through an RNA intermediate (Payer and Burns 2019). TEs can be categorized into two classes based on the intermediate substrate propagating insertions (RNA or DNA). Class I TEs, also called retrotransposons, first transcribe an RNA copy that is then reverse transcribed to cDNA before inserting in the genome. In turn, these can be divided into long terminal repeat (LTR) retrotransposons, which refer to endogenous retroviruses (ERVs), and non-LTR retrotransposons, which include long interspersed element class 1 (LINE-1 or L1) and short interspersed elements (SINEs). Class II TEs, also known as DNA transposons, directly excise themselves from one location before reinsertion. TEs are further split into families and subfamilies depending on various structural features (Goerner-Potvin and Bourque 2018; Guffanti et al. 2018).
Most TEs have lost the capacity for generating new insertions over their evolutionary history and are now fixed in the human population. Their insertions have resulted in a complex distribution of interspersed repeats comprising almost half (50%) of the human genome (Payer and Burns 2019).
TE expression has been observed in association with physiological processes in a wide range of species, including humans where it has been described to be important in early embryonic pluripotency and development. Moreover, aberrant TE expression has been associated with diseases such as cancer, neurodegenerative disorders, and infertility (Payer and Burns 2019).
The study of TE expression faces one main challenge: given their repetitive nature, the majority of TE-derived reads map to multiple regions of the genome and these multi-mapping reads are consequently discarded in standard RNA-seq data processing pipelines. For this reason, specific software packages for the quantification of TE expression have been developed (Goerner-Potvin and Bourque 2018), such as TEtranscripts (Jin et al. 2015), ERVmap (Tokuyama et al. 2018) and Telescope (Bendall et al. 2019). The main differences between these three methods are the following:
TEtranscripts reassigns multi-mapping reads to TEs proportionally to their relative abundance, which is estimated using an expectation-maximization (EM) algorithm.
ERVmap (Tokuyama et al. 2018) is based on selective filtering of multi-mapping reads. It applies filters that consist in discarding reads when the ratio of sum of hard and soft clipping to the length of the read (base pair) is greater than or equal to 0.02, the ratio of the edit distance to the sequence read length (base pair) is greater or equal to 0.02 and/or the difference between the alignment score from BWA (field AS) and the suboptimal alignment score from BWA (field XS) is less than 5.
Telescope reassigns multi-mapping reads to TEs using their relative abundance, which like in TEtranscripts, is also estimated using an EM algorithm. The main differences with respect to TEtranscripts are: (1) Telescope works with an additional parameter for each TE that estimates the proportion of multi-mapping reads that need to be reassigned to that TE; (2) that reassignment parameter is optimized during the EM algorithm jointly with the TE relative abundances, using a Bayesian maximum a posteriori (MAP) estimate that allows one to use prior values on these two parameters; and (3) using the final estimates on these two parameters, multi-mapping reads can be flexibly reassigned to TEs using different strategies, where the default one is to assign a multi-mapping read to the TE with largest estimated abundance and discard those multi-mapping reads with ties on those largest abundances.
Because these tools were only available outside R and Bioconductor, the atena package provides a complete re-implementation in R of these three methods to facilitate the integration of TE expression quantification into Bioconductor workflows for the analysis of RNA-seq data.
Another challenge in TE expression quantification is the lack of complete TE annotations due to the difficulty to correctly place TEs in genome assemblies (Goerner-Potvin and Bourque 2018). The gold standard for TE annotations are RepeatMasker annotations, available through the RepeatMasker tracks in genome browsers. Other repositories also offer annotations of TEs (see Table 1 in (Goerner-Potvin and Bourque 2018)).
Quantification of TE expression with atena consists in the following two steps:
Building of a parameter object for one of the available quantification methods.
Calling the TE expression quantification method qtex() using the previously
built parameter object.
The dataset that will be used to quantify TE expression is a published RNA-seq dataset of Drosophila melanogaster available at the National Center for Biotechnology Information (NCBI) Gene Expression Omnibus (accession no. GSE47006, https://www.ncbi.nlm.nih.gov). The two selected samples are: piwi knockdown and piwi control (GSM1142845 and GSM1142844). These files have been subsampled. The piwi-associated silencing complex (piRISC) silences TEs in the Drosophila ovary, thus, the knockdown of piwi causes the de-repression of TEs. The TE annotations used in following examples consist of 28 and 50 highly expressed TEs and genes, respectively, from Drosophila melanogaster.
To use the ERVmap method in atena we should first build an object of the class ERVmapParam using the function ERVmapParam(). The singleEnd parameter is set to TRUE since the example BAM files are single-end. The ignoreStrand parameter works analogously to the same parameter in the function summarizeOverlaps() from package GenomicAlignments and should be set to TRUE whenever the RNA library preparation protocol was stranded.
One of the filters applied by the ERVmap method compares the alignment score of a given primary alignment, stored in the AS tag of a SAM record, to the largest alignment score among every other secondary alignment, known as the suboptimal alignment score. The original ERVmap software assumes that input BAM files are generated using the Burrows-Wheeler Aligner (BWA) software (Li and Durbin 2009), which stores suboptimal alignment scores in the XS tag. Although AS is an optional tag, most short-read aligners provide this tag with alignment scores in BAM files. However, the suboptimal alignment score, stored in the XS tag by BWA, is either stored in a different tag or not stored at all by other short-read aligner software, such as STAR (Dobin et al. 2013).
To enable using ERVmap on BAM files produced by short-read aligner software other than BWA, atena allows the user to set the argument suboptimalAlignmentTag to one of the following three possible values:
The name of a tag different to XS that stores the suboptimal alignment score.
The value “none”, which will trigger the calculation of the suboptimal alignment score by searching for the largest value stored in the AS tag among all available secondary alignments.
The value “auto” (default), by which atena will first extract the name of the short-read aligner software from the BAM file and if that software is BWA, then suboptimal alignment scores will be obtained from the XS tag. Otherwise, it will trigger the calculation previously explained for suboptimalAlignemntTag="none".
Finally, this filter is applied by comparing the difference between alignment and suboptimal alignment scores to a cutoff value, which by default is 5 but can be modified using the parameter suboptimalAlignmentCutoff. The default value 5 is the one employed in the original ERVmap software that assumes the BAM file was generated with BWA and for which lower values are interpreted as “equivalent to second best match has one or more mismatches than the best match” (Tokuyama et al. 2018, pg. 12571). From a different perspective, in BWA the mismatch penalty has a value of 4 and therefore, a suboptimalAlignmentCutoff value of 5 only retains those reads where the suboptimal alignment has at least 1 mismatch more than the best match. Therefore, the suboptimalAlignmentCutoff value is specific to the short-read mapper software and we recommend to set this value according to the mismatch penalty of that software. Another option is to set suboptimalAlignmentCutoff=NA, which prevents the filtering of reads based on this criteria, as set in the following example.
library(atena)
library(GenomicRanges)bamfiles <- list.files(system.file("extdata", package="atena"),
pattern="*.bam", full.names=TRUE)
TE_annot <- readRDS(file = system.file("extdata", "Top28TEs.rds",
package="atena"))
empar <- ERVmapParam(bamfiles, teFeatures = TE_annot, singleEnd = TRUE,
ignoreStrand = TRUE, suboptimalAlignmentCutoff=NA)
empar## ERVmapParam object
## # BAM files (2): control_KD.bam, piwi_KD.bam
## # features (28): ROO_LTR_174, ..., ROO_LTR_10731
## # single-end, unstrandedIn the case of paired-end BAM files (singleEnd=FALSE), two additional arguments can be specified, strandMode and fragments:
strandMode defines the behavior of the strand getter when internally reading the BAM files with the GAlignmentPairs() function. See the help page of strandMode in the GenomicAlignments package for further details.
fragments controls how read filtering and counting criteria are applied to the read mates in a paired-end read. To use the original ERVmap algorithm (Tokuyama et al. 2018) one should set fragments=TRUE (default when singleEnd=FALSE), which filters and counts each mate of a paired-end read independently (i.e., two read mates overlapping the same feature count twice on that feature, treating paired-end reads as if they were single-end). On the other hand, when fragments=FALSE, if the two read mates pass the filtering criteria and overlap the same feature, they count once on that feature. If either read mate fails to pass the filtering criteria, then both read mates are discarded.
An additional functionality with respect to the original ERVmap software is the integration of gene and TE expression quantification. The original ERVmap software doesn’t quantify TE and gene expression coordinately and this can potentially lead to counting twice reads that simultaneously overlap a gene and a TE. In atena, gene expression is quantified based on the approach used in the TEtranscripts software (Jin et al. 2015): unique reads are preferably assigned to genes, whereas multi-mapping reads are preferably assigned to TEs.
In case that a unique read does not overlap a gene or a multi-mapping read does not overlap a TE, atena searches for overlaps with TEs or genes, respectively. Given the different treatment of unique and multi-mapping reads, atena requires the information regarding the unique or multi-mapping status of a read. This information is obtained from the presence of secondary alignments in the BAM file or, alternatively, from the NH tag in the BAM file (number of reported alignments that contain the query in the current SAM record). Therefore, either secondary alignments or the NH tag need to be present for gene expression quantification.
The original ERVmap approach does not discard any read overlapping gene annotations. However, this can be changed using the parameter geneCountMode, which by default geneCountMode="all" and follows the behavior in the original ERVmap method. On the contrary, by setting geneCountMode="ervmap", atena also applies the filtering criteria employed to quantify TE expression to the reads overlapping gene annotations.
Finally, atena also allows one to aggregate TE expression quantifications. By default, the names of the input GRanges or GRangesList object given in the teFeatures parameter are used to aggregate quantifications. However, the aggregateby parameter can be used to specify other column names in the feature annotations to be used to aggregate TE counts, for example at the sub-family level.
To use the Telescope method for TE expression quantification, the TelescopeParam() function is used to build a parameter object of the class TelescopeParam.
As in the case of ERVmapParam(), the aggregateby argument, which should be a character vector of column names in the annotation, determines the columns to be used to aggregate TE expression quantifications. This way, atena provides not only quantifications at the subfamily level, but also allows to quantify TEs at the desired level (family, class, etc.), including locus based quantifications. For such a use case, the object with the TE annotations should include a column with unique identifiers for each TE locus and the aggregateby argument should specify the name of that column. When aggregateby is not specified, the names() of the object containing TE annotations are used to aggregate quantifications.
Here, the Telescope annotations will be used and TE quantifications will be aggregated according to the names() of the TE_annot object.
bamfiles <- list.files(system.file("extdata", package="atena"),
pattern="*.bam", full.names=TRUE)
TE_annot <- readRDS(file = system.file("extdata", "Top28TEs.rds",
package="atena"))
gene_annot <- readRDS(file = system.file("extdata", "Top50genes.rds",
package="atena"))
tspar <- TelescopeParam(bfl=bamfiles, teFeatures=TE_annot,
geneFeatures = gene_annot,
singleEnd = TRUE, ignoreStrand=TRUE)
tspar## TelescopeParam object
## # BAM files (2): control_KD.bam, piwi_KD.bam
## # features (GRanges length 78): ROO_LTR_174, ..., FBgn0259229
## # aggregated by: GRanges names
## # single-end; unstrandedIn case of paired-end data (singleEnd=FALSE), the argument usage is similar to that of ERVmapParam(). In relation to the BAM file, Telescope follows the same approach as the ERVmap method: when fragments=FALSE, only mated read pairs from opposite strands are considered, while when fragments=TRUE, same-strand pairs, singletons, reads with unmapped pairs and other fragments are also considered by the algorithm. However, there is one important difference with respect to the counting approach followed by ERVmap: when fragments=TRUE mated read pairs mapping to the same element are counted once, whereas in the ERVmap method they are counted twice.
As in the ERVmap method from atena, the gene expression quantification method in Telescope is based on the approach of the TEtranscripts software (Jin et al. 2015). This way, atena provides the possibility to integrate TE expression quantification by Telescope with gene expression quantification. As in the case of the ERVmap method from atena, either secondary alignments or the NH tag are required for gene expression quantification.
Finally, the third method available is TEtranscripts. First, the TEtranscriptsParam() function is called to build a parameter object of the class TEtranscriptsParam. The usage of the aggregateby argument is the same as in TelescopeParam() and ERVmapParam(). Locus based quantifications in the TEtranscripts method from atena is possible because the TEtranscripts algorithm actually computes TE quantifications at the locus level and then sums up all instances of each TE subfamily to provide expression at the subfamily level. By avoiding this last step, atena can provide TE expression quantification at the locus level using the TEtranscripts method. For such a use case, the object with the TE annotations should include a column with unique identifiers for each TE and the aggregateby argument should specify the name of that column.
Here, the Telescope annotations will be used and TE quantifications will be aggregated at the repeat name level. This way, the aggregateby argument will be set to aggregateby = "repName".
bamfiles <- list.files(system.file("extdata", package="atena"),
pattern="*.bam", full.names=TRUE)
TE_annot <- readRDS(file = system.file("extdata", "Top28TEs.rds",
package="atena"))
ttpar <- TEtranscriptsParam(bamfiles, teFeatures = TE_annot, singleEnd = TRUE,
ignoreStrand=TRUE, aggregateby = c("repName"))
ttpar## TEtranscriptsParam object
## # BAM files (2): control_KD.bam, piwi_KD.bam
## # features (GRanges length 28): ROO_LTR_174, ..., ROO_LTR_10731
## # aggregated by: repName
## # single-end; unstrandedFor paired-end data (singleEnd=FALSE), the usage of the fragments argument is the same as in TelescopeParam().
Regarding gene expression quantification, atena has implemented the approach of the original TEtranscripts software (Jin et al. 2015). As in the case of the ERVmap and Telescope methods from atena, either secondary alignments or the NH tag are required.
Following the gene annotation processing present in the TEtranscripts algorithm, in case that geneFeatures contains a metadata column named “type”, only the elements with “type” = “exon” are considered for the analysis. Then, exon counts are summarized to the gene level in a GRangesList object. This also applies to the ERVmap and Telescope methods for atena when gene feature are present. Let’s see an example of this processing:
# Creating an example of gene annotations
annot_gen <- GRanges(seqnames = rep("2L",10),
ranges = IRanges(start = c(1,20,45,80,110,130,150,170,200,220),
width = c(10,20,35,10,5,15,10,25,5,20)),
strand = "*",
type = rep("exon",10))
# Setting gene ids
names(annot_gen) <- paste0("gene",c(rep(1,3),rep(2,4),rep(3,3)))
annot_gen## GRanges object with 10 ranges and 1 metadata column:
## seqnames ranges strand | type
## <Rle> <IRanges> <Rle> | <character>
## gene1 2L 1-10 * | exon
## gene1 2L 20-39 * | exon
## gene1 2L 45-79 * | exon
## gene2 2L 80-89 * | exon
## gene2 2L 110-114 * | exon
## gene2 2L 130-144 * | exon
## gene2 2L 150-159 * | exon
## gene3 2L 170-194 * | exon
## gene3 2L 200-204 * | exon
## gene3 2L 220-239 * | exon
## -------
## seqinfo: 1 sequence from an unspecified genome; no seqlengthsttpar_gen <- TEtranscriptsParam(bamfiles, teFeatures = TE_annot,
geneFeatures = annot_gen, singleEnd = TRUE,
ignoreStrand=TRUE)
ttpar_gen## TEtranscriptsParam object
## # BAM files (2): control_KD.bam, piwi_KD.bam
## # features (CompressedGRangesList length 31): ROO_LTR_10494, ..., gene3
## # aggregated by: CompressedGRangesList names
## # single-end; unstrandedLet’s see the result of the gene annotation processing:
features(ttpar_gen)[!attributes(features(ttpar_gen))$isTE$isTE]## GRangesList object of length 3:
## $gene1
## GRanges object with 3 ranges and 10 metadata columns:
## seqnames ranges strand | source type score phase
## <Rle> <IRanges> <Rle> | <factor> <factor> <numeric> <integer>
## gene1 2L 1-10 * | NA exon NA <NA>
## gene1 2L 20-39 * | NA exon NA <NA>
## gene1 2L 45-79 * | NA exon NA <NA>
## repName repClass repFamily name ID isTE
## <character> <character> <character> <character> <character> <Rle>
## gene1 <NA> <NA> <NA> <NA> <NA> FALSE
## gene1 <NA> <NA> <NA> <NA> <NA> FALSE
## gene1 <NA> <NA> <NA> <NA> <NA> FALSE
## -------
## seqinfo: 1539 sequences from an unspecified genome; no seqlengths
##
## $gene2
## GRanges object with 4 ranges and 10 metadata columns:
## seqnames ranges strand | source type score phase
## <Rle> <IRanges> <Rle> | <factor> <factor> <numeric> <integer>
## gene2 2L 80-89 * | NA exon NA <NA>
## gene2 2L 110-114 * | NA exon NA <NA>
## gene2 2L 130-144 * | NA exon NA <NA>
## gene2 2L 150-159 * | NA exon NA <NA>
## repName repClass repFamily name ID isTE
## <character> <character> <character> <character> <character> <Rle>
## gene2 <NA> <NA> <NA> <NA> <NA> FALSE
## gene2 <NA> <NA> <NA> <NA> <NA> FALSE
## gene2 <NA> <NA> <NA> <NA> <NA> FALSE
## gene2 <NA> <NA> <NA> <NA> <NA> FALSE
## -------
## seqinfo: 1539 sequences from an unspecified genome; no seqlengths
##
## $gene3
## GRanges object with 3 ranges and 10 metadata columns:
## seqnames ranges strand | source type score phase
## <Rle> <IRanges> <Rle> | <factor> <factor> <numeric> <integer>
## gene3 2L 170-194 * | NA exon NA <NA>
## gene3 2L 200-204 * | NA exon NA <NA>
## gene3 2L 220-239 * | NA exon NA <NA>
## repName repClass repFamily name ID isTE
## <character> <character> <character> <character> <character> <Rle>
## gene3 <NA> <NA> <NA> <NA> <NA> FALSE
## gene3 <NA> <NA> <NA> <NA> <NA> FALSE
## gene3 <NA> <NA> <NA> <NA> <NA> FALSE
## -------
## seqinfo: 1539 sequences from an unspecified genome; no seqlengthsqtex()Finally, to quantify TE expression we call the qtex() method using one of the previously defined parameter objects (ERVmapParam, TEtranscriptsParam or TelescopeParam) according to the quantification method we want to use. The qtex() method returns a SummarizedExperiment object containing the resulting quantification of expression in an assay slot. Additionally, when a data.frame, or DataFrame, object storing phenotypic data is passed to the qtex() function through the phenodata parameter, this will be included as column data in the resulting SummarizedExperiment object and the row names of these phenotypic data will be set as column names in the output SummarizedExperiment object.
In the current example, the call to quantify TE expression using the ERVmap method would be the following:
emq <- qtex(empar)## iteration: 12emq## class: RangedSummarizedExperiment
## dim: 28 2
## metadata(0):
## assays(1): counts
## rownames(28): ROO_LTR_174 ROO_LTR_196 ... ROO_LTR_10666 ROO_LTR_10731
## rowData names(10): source type ... ID isTE
## colnames(2): control_KD piwi_KD
## colData names(0):colSums(assay(emq))## control_KD piwi_KD
## 29 17In the case of the Telescope method, the call would be as follows:
tsq <- qtex(tspar)## iteration: 12tsq## class: RangedSummarizedExperiment
## dim: 78 2
## metadata(0):
## assays(1): counts
## rownames(78): ROO_LTR_174 ROO_LTR_196 ... FBgn0028496 FBgn0259229
## rowData names(15): source type ... transcript_name isTE
## colnames(2): control_KD piwi_KD
## colData names(0):colSums(assay(tsq))## control_KD piwi_KD
## 150 150For the TEtranscripts method, TE expression is quantified by using the following call:
ttq <- qtex(ttpar)## iteration: 12ttq## class: RangedSummarizedExperiment
## dim: 1 2
## metadata(0):
## assays(1): counts
## rownames(1): ROO_LTR
## rowData names(0):
## colnames(2): control_KD piwi_KD
## colData names(0):colSums(assay(ttq))## control_KD piwi_KD
## 149 122As mentioned, TE expression quantification is provided at the repeat name level. All 28 TE features share the same repeat name. Thus, the expression levels of the initial 28 TE features have been a aggregated into 1 element with repeat name ROO_LTR.
nrow(ttq)## [1] 1sessionInfo()## R version 4.1.2 (2021-11-01)
## Platform: x86_64-w64-mingw32/x64 (64-bit)
## Running under: Windows Server 2012 R2 x64 (build 9600)
##
## Matrix products: default
##
## locale:
## [1] LC_COLLATE=C
## [2] LC_CTYPE=English_United States.1252
## [3] LC_MONETARY=English_United States.1252
## [4] LC_NUMERIC=C
## [5] LC_TIME=English_United States.1252
##
## attached base packages:
## [1] stats4 stats graphics grDevices utils datasets methods
## [8] base
##
## other attached packages:
## [1] atena_1.0.1 SummarizedExperiment_1.24.0
## [3] Biobase_2.54.0 GenomicRanges_1.46.1
## [5] GenomeInfoDb_1.30.0 IRanges_2.28.0
## [7] S4Vectors_0.32.2 BiocGenerics_0.40.0
## [9] MatrixGenerics_1.6.0 matrixStats_0.61.0
## [11] BiocStyle_2.22.0
##
## loaded via a namespace (and not attached):
## [1] Rcpp_1.0.7 bslib_0.3.1 compiler_4.1.2
## [4] BiocManager_1.30.16 jquerylib_0.1.4 XVector_0.34.0
## [7] bitops_1.0-7 tools_4.1.2 zlibbioc_1.40.0
## [10] digest_0.6.28 jsonlite_1.7.2 evaluate_0.14
## [13] lattice_0.20-45 rlang_0.4.12 Matrix_1.3-4
## [16] DelayedArray_0.20.0 parallel_4.1.2 yaml_2.2.1
## [19] xfun_0.28 fastmap_1.1.0 GenomeInfoDbData_1.2.7
## [22] stringr_1.4.0 knitr_1.36 Biostrings_2.62.0
## [25] sass_0.4.0 grid_4.1.2 R6_2.5.1
## [28] BiocParallel_1.28.1 rmarkdown_2.11 bookdown_0.24
## [31] magrittr_2.0.1 sparseMatrixStats_1.6.0 Rsamtools_2.10.0
## [34] GenomicAlignments_1.30.0 htmltools_0.5.2 stringi_1.7.5
## [37] RCurl_1.98-1.5 SQUAREM_2021.1 crayon_1.4.2Bendall, Matthew L, Miguel De Mulder, Luis Pedro Iñiguez, Aarón Lecanda-Sánchez, Marcos Pérez-Losada, Mario A Ostrowski, R Brad Jones, et al. 2019. “Telescope: Characterization of the Retrotranscriptome by Accurate Estimation of Transposable Element Expression.” PLoS Computational Biology 15 (9): e1006453.
Dobin, Alexander, Carrie A Davis, Felix Schlesinger, Jorg Drenkow, Chris Zaleski, Sonali Jha, Philippe Batut, Mark Chaisson, and Thomas R Gingeras. 2013. “STAR: Ultrafast Universal Rna-Seq Aligner.” Bioinformatics 29 (1): 15–21.
Goerner-Potvin, Patricia, and Guillaume Bourque. 2018. “Computational Tools to Unmask Transposable Elements.” Nature Reviews Genetics 19 (11): 688–704.
Guffanti, Guia, Andrew Bartlett, Torsten Klengel, Claudia Klengel, Richard Hunter, Gennadi Glinsky, and Fabio Macciardi. 2018. “Novel Bioinformatics Approach Identifies Transcriptional Profiles of Lineage-Specific Transposable Elements at Distinct Loci in the Human Dorsolateral Prefrontal Cortex.” Molecular Biology and Evolution 35 (10): 2435–53.
Jin, Ying, Oliver H Tam, Eric Paniagua, and Molly Hammell. 2015. “TEtranscripts: A Package for Including Transposable Elements in Differential Expression Analysis of Rna-Seq Datasets.” Bioinformatics 31 (22): 3593–9.
Li, Heng, and Richard Durbin. 2009. “Fast and Accurate Short Read Alignment with Burrows–Wheeler Transform.” Bioinformatics 25 (14): 1754–60.
Payer, Lindsay M, and Kathleen H Burns. 2019. “Transposable Elements in Human Genetic Disease.” Nature Reviews Genetics 20 (12): 760–72.
Tokuyama, Maria, Yong Kong, Eric Song, Teshika Jayewickreme, Insoo Kang, and Akiko Iwasaki. 2018. “ERVmap Analysis Reveals Genome-Wide Transcription of Human Endogenous Retroviruses.” Proceedings of the National Academy of Sciences 115 (50): 12565–72.