Enumerating transcription factors

Given the importance of the topic, it is not surprising that a number of bioinformatic research groups have published catalogs of transcription factors along with metadata about their features. Standard nomenclature for TFs has yet to be established. Gene symbols, motif sequences, and position-weight matrix catalog entries have all been used as TF identifiers.

In TFutils we have gathered information from four widely used resources: Gene Ontology (GO, Ashburner et al. (2000), in which GO:0003700 is the tag for the molecular function concept “DNA binding transcription factor activity”), CISBP (Weirauch et al. (2014)), HOCOMOCO (Kulakovskiy et al. (2018)), and MSigDb (Subramanian et al. (2005)). Figure @ref(fig:lkupset) depicts the sizes of these catalogs, measured using counts of unique HGNC gene symbols. The enumeration for GO uses Bioconductor’s org.Hs.eg.db package to find direct associations from GO:0003700 to HGNC symbols. The enumeration for MSigDb is heuristic and involves parsing the gene set identifiers used in MSigDb for exact or close matches to HGNC symbols. For CISBP and HOCOMOCO, the associated web servers provide easily parsed tabular catalogs.

Sizes of TF catalogs and of intersections based on HGNC symbols for TFs.

Classification of transcription factors

As noted by Weirauch et al. (2014), interpretation of the “function and evolution of DNA sequences” is dependent on the analysis of sequence-specific DNA binding domains. These domains are dynamic and cell-type specific (Gertz et al. (2013)). Classifying TFs according to features of the binding domain is an ongoing process of increasing intricacy. Figure @ref(fig:TFclass) shows excerpts of hierarchies of terms related to TF type derived from GO (on the left) and TFclass (Wingender et al. (2018)). There is a disagreement between our enumeration of TFs based on GO in Figure @ref(fig:lkupset) and the 1919 shown in AmiGO, as the latter includes a broader collection of receptor activities.

Screenshots of AmiGO and TFClass hierarchy excerpts.

Table @ref(tab:classtab) provides examples of frequently encountered TF classifications in the CISBP and HOCOMOCO catalogs. The numerical components of the HOCOMOCO classes correspond to TFClass subfamilies (Wingender et al. (2018)).

Enumerating TF targets

The Broad Institute MSigDb (Subramanian et al. (2005)) includes a gene set collection devoted to cataloging TF targets. We have used Bioconductor’s GSEABase package to import and serialize the gmt representation of this collection.

TFutils::tftColl

## GeneSetCollection
##   names: AAANWWTGC_UNKNOWN, AAAYRNCTG_UNKNOWN, ..., GCCATNTTG_YY1_Q6 (615 total)
##   unique identifiers: 4208, 481, ..., 56903 (12774 total)
##   types in collection:
##     geneIdType: EntrezIdentifier (1 total)
##     collectionType: NullCollection (1 total)

Names of TFs for which target sets are assembled are encoded in a systematic way, with underscores separating substrings describing motifs, genes, and versions. Some peculiarity in nomenclature in the MSigDb labels can be observed:

grep("NFK", names(TFutils::tftColl), value=TRUE)

## [1] "NFKAPPAB65_01"         "NFKAPPAB_01"           "NFKB_Q6"              
## [4] "NFKB_C"                "NFKB_Q6_01"            "GGGNNTTTCC_NFKB_Q6_01"

Manual curation will be needed to improve the precision with which MSigDb TF target sets can used.

Cataloging TF targets

The MSigDb collection is provided primarily for the purpose of defining gene sets in terms of TF targets. We use the GSEABase package GeneSetCollection class to manage these sets.

Quantitative predictions of TF binding affinities

The FIMO algorithm of the MEME suite (Grant, Bailey, and Noble (2011)) was used to score the human reference genome for TF binding affinity for 689 motif matrices to which genes are associated. Sixteen (16) tabix-indexed BED files are lodged in an AWS S3 bucket for illustration purposes.

library(GenomicFiles)
data(fimo16)
fimo16

## GenomicFiles object with 0 ranges and 16 files: 
## files: M0635_1.02sort.bed.gz, M3433_1.02sort.bed.gz, ..., M6159_1.02sort.bed.gz, M6497_1.02sort.bed.gz 
## detail: use files(), rowRanges(), colData(), ...

head(colData(fimo16))

## DataFrame with 6 rows and 2 columns
##          Mtag        HGNC
##   <character> <character>
## 1     M0635_1      DMRTC2
## 2     M3433_1       HOXA3
## 3     M3467_1        IRF1
## 4     M3675_1      POU2F1
## 5     M3698_1        TP53
## 6     M3966_1       STAT1

We harvest scores in a genomic interval of interest (bound to fimo16 in the rowRanges assignment below) using reduceByFile. This yields a list with one element per file. Each such element holds a list of scanTabix results, one per query range.

library(BiocParallel)
register(SerialParam()) # important for macosx?
rowRanges(fimo16) = GRanges("chr17", IRanges(38.077e6, 38.084e6))
rr = GenomicFiles::reduceByFile(fimo16, MAP=function(r,f)
  scanTabix(f, param=r))

scanTabix produces a list of vectors of text strings, which we parse with data.table::fread. The resulting tables are then reduced to a genomic location and -log10 of the p-value derived from the binding affinity statistic of FIMO in the vicinity of that location.

asdf = function(x) data.table::fread(paste0(x, collapse="\n"), header=FALSE)
gg = lapply(rr, function(x) {
       tmp = asdf(x[[1]][[1]]) 
       data.frame(loc=tmp$V2, score=-log10(tmp$V7))
     })
for (i in 1:length(gg))  gg[[i]]$tf = colData(fimo16)[i,2]

It turns out there are too many distinct TFs to display individually, so let’s also label the scores with the TF families as defined in CISBP.

matchcis = match(colData(fimo16)[,2], cisbpTFcat[,2])
famn = cisbpTFcat[matchcis,]$Family_Name
for (i in 1:length(gg))  gg[[i]]$tffam = famn[i]
nn = do.call(rbind, gg)

A simple display of predicted TF binding affinity in a genomic region is then

library(ggplot2)
ggplot(nn, aes(x=loc,y=score,group=tffam, colour=tffam)) + geom_point()

Summary

We have compared enumerations of human transcription factors by different projects, provided access to two forms of binding domain classification, and illustrated the use of cloud-resident genome-wide binding predictions. In the next section we review selected details of data structures and methods of the TFutils package.

Implementation

data(tftColl)
data(tftCollMap)
data(cisbpTFcat)

TFs_MSIG = TFCatalog(name="MsigDb.TFT", nativeIds=names(tftColl),
 HGNCmap=data.frame(tftCollMap,stringsAsFactors=FALSE))

TFs_CISBP = TFCatalog(name="CISBP.info", nativeIds=cisbpTFcat[,1],
 HGNCmap = cisbpTFcat)

TFs_MSIG

TFs_CISBP