CODEWITHSUNDEEP
rsapply
Federico Giorgi
Federico Giorgi

Reputation: 10735

Efficient coordinate match between data.frames in R with sapply

I'm trying to obtain a vector telling me which rows in a data.frame (transcriptcoords)

              chr  start  end
NONHSAT000001 chr1 11868 14409
NONHSAT000002 chr1 11871 14412
NONHSAT000003 chr1 11873 14409
NONHSAT000004 chr1 12009 13670
NONHSAT000005 chr1 14777 16668
NONHSAT000006 chr1 15602 29370

have start/end coordinates loosely contained (with a +/- 10 tolerance) in another data.frame (genecoords)

              chr  start  end
NONHSAG000001 chr1 11869 14412
NONHSAG000002 chr1 14778 29370
NONHSAG000003 chr1 29554 31109
NONHSAG000004 chr1 34554 36081
NONHSAG000005 chr1 36273 50281
NONHSAG000006 chr1 62948 63887

To do so, I am looping with sapply over the row indeces of the first data.frame, matching the coordinates with any row in the second data.frame. I have a solution (described below), but it seems to be rather slow (around six seconds with a slice of 2000 rows):

   user  system elapsed 
   6.02    0.00    6.04

I'm trying to understand which parts of the sapply could be optimized. Is it the if/else block? Or the comparison lines (==, <=, >=)? Or more simply, is it an intrinsically slow algorithm?

Thank you! The code I generated is below:

load(url("http://www.giorgilab.org/stuff/data.rda"))

# Pre-vectorize the data frames
g0<-rownames(genecoords)
g1<-genecoords[,1]
g2<-as.integer(genecoords[,2])
g3<-as.integer(genecoords[,3])

t0<-rownames(transcriptcoords)
t1<-transcriptcoords[,1]
t2<-as.integer(transcriptcoords[,2])
t3<-as.integer(transcriptcoords[,3])

system.time(gs<-sapply(1:2000,function(i){
            t<-t0[i]
            chr<-t1[i]
            start<-t2[i]
            end<-t3[i]

            # Find a match (loose boundaries +/- 10)
            right1<-which(g1==chr)
            right2<-which(g2<=start+10)
            right3<-which(g3>=end-10)
            right<-intersect(right3,intersect(right1,right2))
            right<-g0[right]

            if(length(right)==1){
                g<-right
            } else if(length(right)>1){
                # Get the smallest match
                heregenecoords<-genecoords[right,]
                size<-apply(heregenecoords,1,function(x){abs(as.numeric(x[3])-as.numeric(x[2]))})
                g<-names(which.min(size))
            } else {
                g<-t
            }
            return(g)           
        }
))

Upvotes: 4

Views: 546

Answers (3)

Martin Morgan
Martin Morgan

Reputation: 46866

With your data

tx0 <- read.table(textConnection("chr  start  end
NONHSAT000001 chr1 11868 14409
NONHSAT000002 chr1 11871 14412
NONHSAT000003 chr1 11873 14409
NONHSAT000004 chr1 12009 13670
NONHSAT000005 chr1 14777 16668
NONHSAT000006 chr1 15602 29370"))

gene0 <- read.table(textConnection("chr  start  end
NONHSAG000001 chr1 11869 14412
NONHSAG000002 chr1 14778 29370
NONHSAG000003 chr1 29554 31109
NONHSAG000004 chr1 34554 36081
NONHSAG000005 chr1 36273 50281
NONHSAG000006 chr1 62948 63887"))

The GenomicRanges package in Bioconductor does this easily and efficiently (for millions of overlaps).

library(GenomicRanges)
tx <- with(tx0, GRanges(chr, IRanges(start, end)))
gene <- with(gene0, GRanges(chr, IRanges(start, end)))

## increase width by 10 on both sides of the center of the gene range
gene <- resize(gene, width=width(gene) + 20, fix="center")
## find overlaps of 'query' tx and 'subject' gene, where query is within subject
olaps <- findOverlaps(tx, gene, type="within")

Showing, e.g., that 'query' (tx) 1, 2, 3, and 4 are within 'subject' (gene) 1.

> findOverlaps(tx, gene, type="within")
Hits of length 6
queryLength: 6
subjectLength: 6
  queryHits subjectHits 
   <integer>   <integer> 
 1         1           1 
 2         2           1 
 3         3           1 
 4         4           1 
 5         5           2 
 6         6           2

and that gene 1 is overlapped by 4 transcripts, gene 2 by 2 transcripts.

> table(subjectHits(olaps))

1 2 
4 2

See also this publication. Using the larger data set:

tx <- with(transcriptcoords, GRanges(V1, IRanges(V2, V3, names=rownames(tx0))))
gene <- with(genecoords, GRanges(V1, IRanges(V2, V3, names=rownames(gene0))))

with some timings:

system.time(gene <- resize(gene, width=width(gene) + 20, fix="center"))
##   user  system elapsed 
##  0.056   0.000   0.057 
system.time(findOverlaps(tx, gene, type="within"))
##   user  system elapsed 
##  2.248   0.000   2.250

I think this is approximately the time for the data.table solution from @danas.zuokos with just 1000 transcripts

system.time({
    dt <- genecoords[transcriptcoords, allow.cartesian = TRUE]; 
    res <- dt[start <= start.1 + tol & end >= end.1 - tol, 
         list(gene = gene[which.min(size)]), by = transcript]
})
##    user  system elapsed 
##   2.148   0.244   2.400

Upvotes: 2

danas.zuokas
danas.zuokas

Reputation: 4643

I'm using data.table library.

rm(list = ls())
load(url("http://www.giorgilab.org/stuff/data.rda"))
library(data.table)
tol <- 10 # tolerance
id <- 1:2000 # you can comment this out, but make sure you have big RAM

Convert to data.table format. Additionaly compute size (I'm not sure why you take abs, isn't end always bigger than start?).

genecoords <- data.table(genecoords, keep.rownames = TRUE)
setnames(genecoords, c("gene", "chr", "start", "end"))
genecoords[, size := end - start]
transcriptcoords <- data.table(transcriptcoords, keep.rownames = TRUE)[id] # comment out [id] when running on all trascripts
setnames(transcriptcoords, c("transcript", "chr", "start", "end"))

And this gives the result.

setkeyv(genecoords, "chr")
setkeyv(transcriptcoords, "chr")
dt <- genecoords[transcriptcoords, allow.cartesian = TRUE]
res <- dt[start <= start.1 + tol & end >= end.1 - tol, list(gene = gene[which.min(size)]), by = transcript]

Aware that this will not include transcripts, which do not meet the condition (g<-t in your code).

Upvotes: 1

cbare
cbare

Reputation: 12468

Ha! Martin beat me to it with a better answer. It's practically always better to use working tested code in a well established library rather than roll your own. Definitely use Martin's solution, not this one.

But, just for laughs, here's another way to do it.

First, make up some genes and transcripts:

gs = 1:10*500
genes = data.frame(start=gs, end=gs+400)
rownames(genes) = sprintf('g%05d', 1:nrow(genes))

ts = sample(1:max(genes$end), size=10)
transcripts = data.frame(start=ts, end=ts+60)
rownames(transcripts) = sprintf('t%05d', 1:nrow(transcripts))

We can vectorize the comparison using outer which applies a function to each combination of its two vector arguments.

overlaps = function(genes, transcripts, min_overlap=1) {
  b1 = outer(genes$end, transcripts$start, min_overlap=min_overlap, 
             function(e,s,min_overlap) e-s+1>min_overlap)
  b2 = outer(genes$start, transcripts$end, min_overlap=min_overlap,
             function(s,e,min_overlap) e-s+1>min_overlap)
  result = b1 & b2
  rownames(result) = rownames(genes)
  colnames(result) = rownames(transcripts)
  return(result)
}

For our genes and transcripts, we might get something like:

> genes
       start  end
g00001   500  900
g00002  1000 1400
g00003  1500 1900
g00004  2000 2400
g00005  2500 2900
g00006  3000 3400
g00007  3500 3900
g00008  4000 4400
g00009  4500 4900
g00010  5000 5400

> transcripts
       start  end
t00001   535  595
t00002  2502 2562
t00003  4757 4817
t00004  3570 3630
t00005  3094 3154
t00006  1645 1705
t00007  5202 5262
t00008    13   73
t00009   788  848
t00010  4047 4107

o1 = overlaps(genes, transcripts, 1)

The result is a boolean matrix that tells you whether each transcript overlaps with each gene.

> o1
       t00001 t00002 t00003 t00004 t00005 t00006 t00007 t00008 t00009 t00010
g00001   TRUE  FALSE  FALSE  FALSE  FALSE  FALSE  FALSE  FALSE   TRUE  FALSE
g00002  FALSE  FALSE  FALSE  FALSE  FALSE  FALSE  FALSE  FALSE  FALSE  FALSE
g00003  FALSE  FALSE  FALSE  FALSE  FALSE   TRUE  FALSE  FALSE  FALSE  FALSE
g00004  FALSE  FALSE  FALSE  FALSE  FALSE  FALSE  FALSE  FALSE  FALSE  FALSE
g00005  FALSE   TRUE  FALSE  FALSE  FALSE  FALSE  FALSE  FALSE  FALSE  FALSE
g00006  FALSE  FALSE  FALSE  FALSE   TRUE  FALSE  FALSE  FALSE  FALSE  FALSE
g00007  FALSE  FALSE  FALSE   TRUE  FALSE  FALSE  FALSE  FALSE  FALSE  FALSE
g00008  FALSE  FALSE  FALSE  FALSE  FALSE  FALSE  FALSE  FALSE  FALSE   TRUE
g00009  FALSE  FALSE   TRUE  FALSE  FALSE  FALSE  FALSE  FALSE  FALSE  FALSE
g00010  FALSE  FALSE  FALSE  FALSE  FALSE  FALSE   TRUE  FALSE  FALSE  FALSE

Upvotes: 1

Related Questions