Contents

1 Introduction

Droplet Digital PCR (ddPCR) is a system from Bio-Rad for estimating the number of genomic fragments in samples. ddPCR attaches fluorochromes to targets of interest, for example, mutant and wild type KRAS. Each sample is then divided into 20,000 droplets and qPCR is run on each droplet. The brightness of these droplets is measured in two channels, each corresponding to our targets. The amplitudes of the droplets can be plotted and the results analysed to see whether droplets can be called as:

There are variations in the brightnesses of the droplets; this can be particularly evident in and around the PP cluster, where there may be some crosstalk due the presence of both fluorochromes. The classification of droplets is therefore not necessarily as simple as deciding on brightness thresholds for each channel, above which a positive reading is called in that channel.

This vignette demonstrates how the twoddpcr package may be used to load data, classified and how to quickly create summaries.

2 Installing the twoddpcr package

The package can be installed from Bioconductor using:

source("https://bioconductor.org/biocLite.R")
biocLite("twoddpcr")

Alternatively, it can be installed from GitHub using:

library(devtools)
install_github("CRUKMI-ComputationalBiology/twoddpcr")

Another alternative is to install the package from source:

install.packages("</path/to/twoddpcr/>", repos=NULL, type="source")

3 Loading the twoddpcr package

Once the package has been installed, it can be loaded in the usual way:

library(twoddpcr)

4 Using the in-built dataset

Our example uses the KRASdata dataset, which comes as part of the package. This dataset was created as a triplicate four-fold serial dilution with 5% A549 mutant KRAS cell lines and 95% H1048 wild type KRAS cells as starting material.

To follow along with this dataset, create a ddpcrPlate object using:

plate <- ddpcrPlate(wells=KRASdata)

To follow along with your own dataset, see the Appendix.

5 Basic plots

All of the droplets can be plotted to see how they tend cluster:

dropletPlot(plate)

This might not be particularly informative; for example, a density plot may be more appropriate:

heatPlot(plate)

It can be seen here that most of the droplets are concentrated in the bottom-left and bottom-right clusters.

To take a different view, all of the wells could be plotted side-by-side:

facetPlot(plate)

6 Plotting the droplets and existing classifications

Since the droplet amplitudes were extracted from Bio-Rad’s QuantaSoft, there may already be some kind of classification. This can be checked with the commonClassificationMethod method, which retrieves the classification methods that exist for all of the wells in the plate:

commonClassificationMethod(plate)
## [1] "None"    "Cluster"

The Cluster classification can be plotted:

dropletPlot(plate, cMethod="Cluster")

Again, these are all of the wells in the plate superimposed onto the same plot. This gives a good overall picture, but the detection of rare alleles in individual wells is where ddPCR is particularly useful. The wells that used in the plate are:

names(plate)
##  [1] "E03" "F03" "G03" "H03" "A04" "B04" "C04" "D04" "E04" "F04" "G04" "H04"

Individual wells can be selected using the [[...]] syntax (as with lists in R). These can be plotted in the same way using dropletPlot. For example:

dropletPlot(plate[["F03"]], cMethod="Cluster")

dropletPlot(plate[["E04"]], cMethod="Cluster")

7 Independent linear gating on the channels (thresholdClassify)

This section illustrates how the Cluster classification was obtained, although the original classification was found using QuantaSoft. The classification here involves setting linear gates (thresholds) for the two channels Ch1.Amplitude and Ch2.Amplitude, above each of which we will call a positive reading for that channel.

plate <- thresholdClassify(plate, ch1Threshold=6789, ch2Threshold=3000)

The commonClassificationMethod method shows that there is now a new classification method:

commonClassificationMethod(plate)
## [1] "None"       "Cluster"    "thresholds"

The thresholds classification can be plotted using dropletPlot but changing the cMethod parameter:

dropletPlot(plate, cMethod="thresholds")

8 Classifying using the k-means algorithm (kmeansClassify)

Visually, it appears that the classification in the previous section does not accurately classify a region between the main NP and PP clusters. There are a number of algorithms that could be used to better classify the clusters; one such example is the k-means clustering algorithm. The k-means algorithm is relatively fast but requires that we know how many clusters there are. With this in mind, it helps to classify all of the wells together so that human intervention is not required to judge whether some clusters in individual wells are empty. To run the algorithm on ddpcrPlate objects, the kmeansClassify method is used:

plate <- kmeansClassify(plate)
commonClassificationMethod(plate)
## [1] "None"       "Cluster"    "thresholds" "kmeans"
dropletPlot(plate, cMethod="kmeans")

Notice how the PP cluster incorporates more of the droplets when compared to the thresholds case. Visually, it appears that k-means captures the clustering behaviour of the droplets more accurately.

Using the same wells chosen before, it is interesting to see how the individual wells classify:

dropletPlot(plate[["F03"]], cMethod="kmeans")

dropletPlot(plate[["E04"]], cMethod="kmeans")

9 Adding “rain”

There are regions between clusters where the classification is ambiguous, e.g. above and to the left of the NP cluster. These regions can be labelled as “Rain” and removed from the droplet counts in each of the clusters. To achieve this, the mahalanobisRain method can be used.

plate <- mahalanobisRain(plate, cMethod="kmeans", maxDistances=3)

The classification methods are now:

commonClassificationMethod(plate)
## [1] "None"          "Cluster"       "thresholds"    "kmeans"       
## [5] "kmeansMahRain"

Whenever droplets are relabelled as Rain using the mahalanobisRain method, the character string “MahRain” is appended to the classification name to distinguish it from the original. This classification is plotted as:

dropletPlot(plate, cMethod="kmeansMahRain")

This does not look particularly good; a lot of droplets that should be classified have been labelled as “Rain” instead. To remedy this, the maxDistances parameter can be adjusted to control the maximum (Mahalanobis) distance that droplets can be from the cluster centres. Some fine-tuning of this parameter gives:

plate <- mahalanobisRain(plate, cMethod="kmeans",
                         maxDistances=list(NN=35, NP=35, PN=35, PP=35))
commonClassificationMethod(plate)
## [1] "None"          "Cluster"       "thresholds"    "kmeans"       
## [5] "kmeansMahRain"

The plot now looks slightly different:

dropletPlot(plate, cMethod="kmeansMahRain")

10 Creating a summary

Using the number of droplets in each classification, the Poisson distribution can be used to estimate the number of fragments/molecules in the starting sample. For the k-means classification with rain, this gives the summary:

kmeansMahRainSummary <- plateSummary(plate, cMethod="kmeansMahRain")
head(kmeansMahRainSummary)
##      PP  PN   NP    NN AcceptedDroplets MtPositives MtNegatives WtPositives
## E03 292 273 5775  6229            12569         565       12004        6067
## F03 305 256 5840  5946            12347         561       11786        6145
## G03 236 222 4877  4860            10195         458        9737        5113
## H03  24  95 1630  9931            11680         119       11561        1654
## A04  22 101 1844 10840            12807         123       12684        1866
## B04  19 112 1924 10998            13053         131       12922        1943
##     WtNegatives MtConcentration WtConcentration MtCopiesPer20uLWell
## E03        6502          54.110         775.440            1082.201
## F03        6202          54.707         810.049            1094.135
## G03        5082          54.076         819.050            1081.514
## H03       10026          12.048         179.643             240.956
## A04       10941          11.354         185.264             227.072
## B04       11110          11.867         189.615             237.334
##     WtCopiesPer20uLWell TotalCopiesPer20uLWell  Ratio FracAbun
## E03           15508.792              16590.992 0.0698    6.523
## F03           16200.972              17295.107 0.0675    6.326
## G03           16381.000              17462.514 0.0660    6.193
## H03            3592.853               3833.809 0.0671    6.285
## A04            3705.287               3932.358 0.0613    5.774
## B04            3792.292               4029.626 0.0626    5.890

The first few columns PP, PN, NP and NN are the numbers of droplets in each class, whereas AcceptedDroplets is the sum of these. MtPositives is the number of droplets where a mutant has been called and conversely MtNegatives is the number of droplets with no mutants called. The MtConcentration is the Poisson estimate of how many mutant fragments there are per 1uL, while the MtCopiesPer20uLWell is the same figure multiplied by 20. There are Wt (wild type) analogues of all of these Mt figures. Finally, Ratio is the figure MtConcentration/WtConcentration and FracAbun is the fractional abundance of mutants in the sample, i.e. 100 * MtConcentration/(MtConcentration + WtConcentration).

The summaries for other classifications can still be produced by changing the cMethod parameter to one of those that exist in commonClassificationMethod(plate).

This concludes the main walkthrough of this vignette.

11 Analysis of the data

11.1 Comparison of classification methods

As mentioned above, the KRASdata dataset was created as a triplicate four-fold serial dilution with 5% mutant and 95% wild type starting material. A data frame can be created to reflect this along with the mutant concentration values of each well.

inputNg <- c(rep(64, 3), rep(16, 3), rep(4, 3), rep(1, 3))
mtConcentrations <-
  data.frame(
    x=inputNg,
    Cluster=plateSummary(plate, cMethod="Cluster")$MtConcentration, 
    kmeans=plateSummary(plate, cMethod="kmeans")$MtConcentration, 
    kmeansMahRain=kmeansMahRainSummary$MtConcentration)
knitr::kable(mtConcentrations)
x Cluster kmeans kmeansMahRain
64 51.028 54.152 54.110
64 48.540 54.689 54.707
64 49.161 54.204 54.076
16 12.036 12.036 12.048
16 11.244 11.337 11.354
16 11.848 11.848 11.867
4 3.340 3.435 3.436
4 3.358 3.271 3.272
4 3.042 3.042 3.043
1 0.547 0.547 0.547
1 0.637 0.637 0.637
1 0.470 0.470 0.470

The mutant concentration values can be plotted and the various classification methods compared against each other:

library(ggplot2)
library(reshape2)
mtConcentrationsLong <- melt(mtConcentrations, id.vars=c("x"))
ggplot(mtConcentrationsLong, aes_string("x", "value")) +
  geom_point() + geom_smooth(method="lm") + facet_wrap(~variable)

Numerically, the regression lines have coefficients of determination (R2 values):

bioradLmSummary <- summary(lm(x ~ Cluster, data=mtConcentrations))
kmLmSummary <- summary(lm(x ~ kmeans, data=mtConcentrations))
kmMahRainLmSummary <- summary(lm(x ~ kmeansMahRain, data=mtConcentrations))
knitr::kable(c("Cluster"=bioradLmSummary$r.squared,
               "kmeans"=kmLmSummary$r.squared,
               "kmeansMahRain"=kmMahRainLmSummary$r.squared))
x
Cluster 0.9989393
kmeans 0.9988022
kmeansMahRain 0.9988212

11.2 Discussion

As shown above, the regression lines fit the data from all of the classification methods very well. Moreover, the R2 values are all similar and very close to 1. Therefore all of the approaches are very good and nothing can be said about which of the methods is better or worse. The density plot created by heatPlot above shows that the number of droplets in the PP cluster is relatively small compared to the other clusters, particularly at the bottom of the PP cluster. This explains why the regression lines are very similar.

An advantage of the twoddpcr package’s k-means based approach is that setting thresholds manually can be subjective. In addition, the k-means clustering algorithm is more appropriate for finding clusters when the PN cluster ‘leans’ and the NP cluster ‘lifts’.

Setting rain using standard deviation is a commonly used approach in ddPCR analysis to remove false positives. It involves setting Ch1 and Ch2 thresholds for each cluster and removes droplets that are too far from the cluster centres. This method was introduced in (Jones et al. 2014). However, it is not possible to set such thresholds for the NP cluster above because, for example, setting a low Ch1 threshold would exclude too much from the top-right of the cluster, whereas setting a high Ch1 threshold would exclude nothing at all. The twoddpcr package’s Mahalanobis rain method allows this kind of approach to be used, while still respecting the shapes of the clusters.

12 Other classification tools

This section explains how other classification methods can be used. The methods already described above should suffice, but there are some droplet patterns that prevent them from working as well as we would like. The following methods are alternative techniques that can be used.

12.1 Classifying using the k-NN algorithm (knnClassify)

Another classification algorithm is the k-nearest neighbour (k-NN) algorithm. The algorithm is very simple: For each droplet in the plate, look at the classifications of its nearest \(k\)-neighbours in a training set. Assign the majority classification to the droplet.

The challenge now is to find a good dataset that is not too large, since this would slow the algorithm considerably for marginal gains. A training set should also have minimal noise.

To start, two wells E03 and A04 are chosen that reflect the clustering pattern of the plate without too much noise. We create a new (virtual) plate with the amplitudes from these wells.

trainWells <- plate[c("E03", "A04")]
trainPlate <- ddpcrPlate(wells=trainWells)

To create the training classification, the k-means algorithm is useful:

trainPlate <- kmeansClassify(trainPlate)
dropletPlot(trainPlate, cMethod="kmeans")

We see that k-means has worked quite well here, since the training set is relatively noise-free. However, it is clear that there is a PN droplet that is much closer to other PP droplets than the rest of the PN cluster. Noise can be removed by adding rain:

trainPlate <- mahalanobisRain(trainPlate, cMethod="kmeans", maxDistances=3)
dropletPlot(trainPlate, cMethod="kmeansMahRain")

This is a much less noisy classification to use. The training data needs to be a data frame and should also ignore the droplets classified as Rain; the removeDropletClasses method removes these droplets:

trainSet <- removeDropletClasses(trainPlate, cMethod="kmeansMahRain")
trainSet <- do.call(rbind, trainSet)
colnames(trainSet)
## [1] "Ch1.Amplitude" "Ch2.Amplitude" "kmeansMahRain"

We can check that the Rain droplets have been removed:

table(trainSet$kmeansMahRain)
## 
##    NN    NP    PN    PP  Rain   N/A 
## 14866  6418   329   246     0     0

Next, we use this classification as the training set for the k-NN algorithm:

trainAmplitudes <- trainSet[, c("Ch1.Amplitude", "Ch2.Amplitude")]
trainCl <- trainSet$kmeansMahRain
plate <- knnClassify(plate, trainData=trainAmplitudes, cl=trainCl, k=3)

Again, it can be checked that there is a new classification method:

commonClassificationMethod(plate)
## [1] "None"          "Cluster"       "thresholds"    "kmeans"       
## [5] "kmeansMahRain" "knn"

This classification can be plotted in the same way as before:

dropletPlot(plate, cMethod="knn")

12.2 Classifying the four ‘corners’ of a plot (gridClassify)

There may be some datasets where the above classification techniques do not work satisfactorily. As long as the main clusters have good separation from each other, the gridClassify method may be used. This method defines four ‘corner’ regions with linear cut-offs in each channel; the remaining droplets are labelled as “Rain”. To see how this works, consider the following (crude) example:

plate <- gridClassify(plate,
                      ch1NNThreshold=6500, ch2NNThreshold=2110,
                      ch1NPThreshold=5765, ch2NPThreshold=5150,
                      ch1PNThreshold=8550, ch2PNThreshold=2450,
                      ch1PPThreshold=6700, ch2PPThreshold=3870)
dropletPlot(plate, cMethod="grid")

This is not a particularly great classification, but this option exists should it be required. It is tedious to set the parameters above, so it may be helpful to use the Shiny app to aid in this process.

12.3 Adding rain with sdRain

Since droplets tend to cluster into ellipse-like structures, the mahalanobisRain method should usually suffice for labelling ambiguous droplets as “Rain”. An alternative way is to use the mean and standard deviation of each of the clusters (in both channels). To do this, use the sdRain method:

plate <- sdRain(plate, cMethod="kmeans")
dropletPlot(plate, cMethod="kmeansSdRain")

As is the case with Mahalanobis rain, the rain levels could be tweaked a little:

plate <- sdRain(plate, cMethod="kmeans",
                errorLevel=list(NN=5, NP=5, PN=3, PP=3))
dropletPlot(plate, cMethod="kmeansSdRain")

12.4 Custom classifications

If you wish to use your own classification methods, the droplet information would need to be extracted and can also be added to the ddpcrPlate object. The basic workflow would be:

  1. Retrieve the droplet amplitudes using amplitudes and combine them in a single data frame:

    allDrops <- amplitudes(plate)
    allDrops <- do.call(rbind, amplitudes)
  2. Classify the droplets using your own method:

    allDrops$class <- someClassificationMethod(allDrops)
  3. Add the classification to plate:

    plateClassification(plate, cMethod="nameOfCMethod") <- allDrops$class

The ddpcrPlate class only understands classifications if it is a factor with levels c("NN", "NP", "PN", "PP", "Rain", "N/A"). If the result of your custom classification method returns a vector/factor with four classes (with maybe some “Rain” or “N/A”), then the vector/factor may be relabelled by:

relabelClasses(allDrops, classCol="class")

If there are fewer than four classes, relabelClasses will try to guess which of the classes are present. To help the method correctly label the clusters, set the presentClasses parameter:

relabelClasses(allDrops, classCol="class", presentClasses=c("NN", "NP", "PN"))

13 Appendix

13.1 Shiny-based GUI for non-R users

A Shiny app is included in the package, which provides a GUI that allows interactive use of the package for ddPCR analysis. This can be run from an interactive R session using:

shinyVisApp()

This can also be accessed at http://shiny.cruk.manchester.ac.uk/twoddpcr/.

To run on your own Shiny server, a file called app.R should be created with the following code:

library(shiny)
library(twoddpcr)

# Disable warnings.
options(warn=-1)

shiny::shinyApp(
  ui=shinyVisUI(),
  server=function(input, output, session)
  {
    shinyVisServer(input, output, session)
  }
)

13.2 Exporting droplet amplitudes from QuantaSoft to CSV files

If you have run your own two channel ddPCR experiments that have produced a QuantaSoft Plate (.qlp) file, then the raw droplet amplitudes can be extracted for use with the twoddpcr package. To do this:

  1. Run QuantaSoft.
  2. Load a plate (from a QuantaSoft Plate .qlp file).
  3. Select the samples to use by using the Ctrl and/or Shift key with the mouse.
  4. Click Options in the top-right.
  5. Click Export Amplitude and Cluster Data.
  6. Select a location to export the amplitude files to. This can take a while to complete.

The amplitudes will be exported to a number of CSV files in the chosen location, with one file for each well. Each file is named <PlateName>_<WellNumber>_Amplitude.csv, where <PlateName> is the name of the .qlp file without the extension and <WellNumber> is the position in the plate, e.g. B03. These amplitude files are now ready to be loaded using the twoddpcr package.

13.3 Using other datasets

The example in this vignette can be followed using a different dataset, such as those from your own ddPCR experiments. To load a dataset:

  1. Follow the instructions in the exporting droplet amplitudes section.
  2. The droplets can be imported using:

    plate <- ddpcrPlate(well="data/amplitudes")

    Here, data/amplitudes should be changed to the directory containing the droplet amplitude files.

13.4 Problems reading files

While loading data, the following error message may appear:

Error in read.table(file = file, header = header, sep = sep, quote = quote,  :
  duplicate 'row.names' are not allowed

Possible solution: The number of columns in the header row might differ from the number of columns in the other rows. For example, there may be extra commas/tabs at the end of some lines. In such cases, the removal of ‘empty’ columns should fix the problem.

14 Citing twoddpcr

If you use the twoddpcr package in your work, please cite the Bioinformatics paper:

citation("twoddpcr")
## 
## To cite twoddpcr in publications, please use:
## 
##   Anthony Chiu, Mahmood Ayub, Caroline Dive, Ged Brady, Crispin J
##   Miller; twoddpcr: An R/Bioconductor package and Shiny app for
##   Droplet Digital PCR analysis. Bioinformatics 2017 btx308. doi:
##   10.1093/bioinformatics/btx308
## 
## A BibTeX entry for LaTeX users is
## 
##   @Article{,
##     author = {Anthony Chiu and Mahmood Ayub and Caroline Dive and Ged Brady and Crispin Miller},
##     title = {twoddpcr: An R/Bioconductor package and Shiny app for Droplet Digital PCR analysis},
##     journal = {Bioinformatics},
##     publisher = {Oxford University Press},
##     year = {2017},
##   }

15 Further reading

(Rödiger et al. 2015) describes how to use R in order to analyse ddPCR data using the dpcR package.

16 Session information

Here is the output of sessionInfo() on the system on which this document was compiled:

## R version 3.4.3 (2017-11-30)
## Platform: x86_64-pc-linux-gnu (64-bit)
## Running under: Ubuntu 16.04.3 LTS
## 
## Matrix products: default
## BLAS: /home/biocbuild/bbs-3.6-bioc/R/lib/libRblas.so
## LAPACK: /home/biocbuild/bbs-3.6-bioc/R/lib/libRlapack.so
## 
## locale:
##  [1] LC_CTYPE=en_US.UTF-8       LC_NUMERIC=C              
##  [3] LC_TIME=en_US.UTF-8        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] stats     graphics  grDevices utils     datasets  methods   base     
## 
## other attached packages:
## [1] reshape2_1.4.3  ggplot2_2.2.1   hexbin_1.27.1   twoddpcr_1.2.1 
## [5] BiocStyle_2.6.1
## 
## loaded via a namespace (and not attached):
##  [1] Rcpp_0.12.14        highr_0.6           compiler_3.4.3     
##  [4] pillar_1.0.1        RColorBrewer_1.1-2  plyr_1.8.4         
##  [7] class_7.3-14        tools_3.4.3         digest_0.6.13      
## [10] evaluate_0.10.1     tibble_1.4.1        gtable_0.2.0       
## [13] lattice_0.20-35     rlang_0.1.6         shiny_1.0.5        
## [16] yaml_2.1.16         parallel_3.4.3      stringr_1.2.0      
## [19] knitr_1.18          S4Vectors_0.16.0    stats4_3.4.3       
## [22] rprojroot_1.3-2     grid_3.4.3          R6_2.2.2           
## [25] rmarkdown_1.8       bookdown_0.5        magrittr_1.5       
## [28] backports_1.1.2     scales_0.5.0        htmltools_0.3.6    
## [31] BiocGenerics_0.24.0 mime_0.5            colorspace_1.3-2   
## [34] xtable_1.8-2        httpuv_1.3.5        labeling_0.3       
## [37] stringi_1.1.6       lazyeval_0.2.1      munsell_0.4.3

References

Jones, M., J. Williams, K. Gartner, R. Phillips, J. Hurst, and J. Frater. 2014. “Low Copy Target Detection by Droplet Digital PCR Through Application of a Novel Open Access Bioinformatic Pipeline, ‘Definetherain’.” J Virol Methods 202 (2). Elsevier: 46–53.

Rödiger, Stefan, Michał Burdukiewicz, K. A. Blagodatskikh, and P. R. Schierack. 2015. “R as an Environment for the Reproducible Analysis of DNA Amplification Experiments.” The R Journal 7 (2). The R Foundation: 127–50.