speaq 2.0
To illustrate the possibilities of speaq 2.0, the wine dataset is used (available form the University of Copenhagen at models.life.ku.dk). This dataset is also utilized in the paper so with this vignette you will be able to reproduce the results.
Before we start with the example let's first recap what the new method in speaq 2.0 encompass:
- Peak detection:
getWaveletPeaks() - Peak grouping:
PeakGrouper() - Peak filling:
PeakFilling() - Feature matrix construction:
BuildFeatureMatrix()
Plotting the raw wine data
library(speaq)
data(Winedata)
Spectra.wine <- as.matrix(Winedata$spectra )
ppm.wine <- as.numeric(Winedata$ppm)
wine.color <- Winedata$wine.color
wine.origin <- Winedata$origin
# all spectra
speaq::drawSpecPPM(Y.spec = Spectra.wine,
X.ppm = ppm.wine,
title = 'Wine data spectra',
groupFactor = wine.color,
legend.extra.x = 1,
legend.extra.y = 1.1)
The drawSpecPPM() plotting function indicates that there might be a gap in the data (which is correct as the original authors deleted the region between 5.0 and 4.5 ppm). This warning can be of importance for the interpretation of the plot as in this case the original authors deleted the data, not by setting the values to 0, but by effectively removing it from the matrix and ppm vector. This produces a plot that appears continuous but is in fact not.
The drawSpecPPM() function also indicates that the groupFactor is not a factor and (successfully) attempts to do the conversion. The next plot is an excerpt of the wine NMR spectra.
# small excerpt by defining the region of interest
speaq::drawSpecPPM(Y.spec = Spectra.wine,
X.ppm = ppm.wine,
groupFactor = as.factor(wine.color),
title = 'Raw wine data excerpt',
legend.extra.x = 1.1,
legend.extra.y = 1.0,
ROI.ppm = 3.6,
ROI = NULL,
roiWidth.ppm = 0.15,
legendpos = "topright" )
From spectra via peaks to peak groups (features)
Now that we've had a look at the spectra it is time to convert these to peaks by using the getWaveletPeaks() function. This takes about 50 seconds, with nCPU = 4 on a 2.5GHz machine (nCPU is set to 1 for the vignette build but should be changed).
wine.peaks <- speaq::getWaveletPeaks(Y.spec=Spectra.wine,
X.ppm=ppm.wine,
baselineThresh = 10,
SNR.Th = -1,
nCPU = 2,
include_nearbyPeaks = TRUE) # nCPU set to 2 for the vignette build
wine.grouped <- speaq::PeakGrouper(Y.peaks = wine.peaks,
min.samp.grp = 5,
grouping.window.width = 200)There are two notes to be made on some important parameters: 1. The include_nearbyPeaks = TRUE option also detects peaks in the tails of larger other peaks. At first this might seem the preferred option as you want to detect as much peaks as possible but we will see later that often it is better to exclude them as these small peaks can cause problems later on. 2. The grouping.window.width parameter is chosen to be twice the size of the default value: it is advised to choose this larger when working with spectra that exhibit large between sample shifts. The wine dataset is an extreme example of large between sample shifts (caused by the substantial pH differences between wines of different colour)
Now we can plot the detected peaks and the grouped peaks. The dataset after grouping contains both the original ppm values of every peak (in the peakPPM variable) but also the group information (found in the peakIndex variable). By calling the AddPlottingStuff() function the groupPPM variable is added so we also have the ppm value of the groups (the link is: groupPPM <- ppm.wine[peakIndex])
Plotting the peak data
# adding labels to the dat a for plotting and the group ppm values
library(ggplot2)
ROI.ppm <- 1.330
roiWidth.ppm <- 0.025
speaq::ROIplot(Y.spec = Spectra.wine,
X.ppm = ppm.wine,
ungrouped.peaks = wine.peaks,
grouped.peaks = wine.grouped,
ROI.ppm = ROI.ppm,
roiWidth.ppm = roiWidth.ppm,
groupLabels = as.factor(wine.color))
The plot above shows clearly what the basis of speaq 2.0 does, convert spectra accurately to peak data and group these peaks. The quality of the grouping can be checked by plotting the silhouette values. this can be done with the SilhouetR() function.
SilhouetteValues <- speaq::SilhouetR(DataMatrix = wine.grouped$peakPPM,
GroupIndices = wine.grouped$peakIndex)## DataMatrix is not a matrix, attempting conversion with the assumption of only 1 variable (1 column)Silh_plot <- ggplot(SilhouetteValues, aes(SilhouetteValues)) +
geom_freqpoly(binwidth = 0.03) +
theme_bw()
Silh_plot
It is clear that the grouping is very good. To be absolutely sure we can check the mean silhouette value of every group to see if there are groups that should be regrouped. This regrouping, if needed, can be done with the regroupR() function.
groups <- unique(SilhouetteValues$GroupIndices)
Ngroups <- length(groups)
sil_means <- matrix(NA, ncol = 3, nrow = Ngroups)
for(k in 1:Ngroups){
sil_means[k,1] = groups[k]
sil_means[k,2] = mean(SilhouetteValues$SilhouetteValues[SilhouetteValues$GroupIndices==groups[k]])
sil_means[k,3] = mean(wine.grouped$peakSNR[wine.grouped$peakIndex==groups[k]])
}
sil_means <- sil_means[order(sil_means[,2]),]
colnames(sil_means) <- c("groupIndex", "avg_silhouette_val", "avg. SNR")
head(sil_means)## groupIndex avg_silhouette_val avg. SNR
## [1,] 5313 0.2557844 32.261401
## [2,] 5050 0.4549954 11.669748
## [3,] 2330 0.4848325 31.757387
## [4,] 4711 0.4904269 5.754990
## [5,] 5552 0.5011936 6.933832
## [6,] 3287 0.5099114 3.559727As it turns out, there is a group with a low average silhouette value (0.25) but a fairly high average signal-to-noise ratio (if this group had a low SNR ratio it could just be a noise group). This indicates that these are non-noise peaks which are grouped incorrectly, the following plot acknowledges this
faulty.groupIndex <- sil_means[1,1]
ROI.ppm <- ppm.wine[faulty.groupIndex]
roiWidth.ppm <- 0.1
speaq::ROIplot(Y.spec = Spectra.wine,
X.ppm = ppm.wine,
ungrouped.peaks = wine.peaks,
grouped.peaks = wine.grouped,
ROI.ppm = ROI.ppm,
roiWidth.ppm = roiWidth.ppm,
groupLabels = as.factor(wine.color))## Warning: Removed 2 rows containing missing values (geom_point).
Clearly the large peaks and small peaks are grouped together because of the large shifts in the wine dataset making the small and large peaks overlap, this messes up the grouping. There are multiple ways to solve this issue:
By setting
include_nearbyPeaks = FALSEin thegetWaveletPeaks()function these small peaks in the tails of large peaks will not be included, although some other peaks might also be missed. Orby selecting the wrong group and submitting them to the
regroupR()function. This will use both the ppm values and the peak signal-to-noise ratio for the initial between peak distance calculation and group them according to this distance.
We will use option 2 in this case to demonstrate the use of the regroupR() function. Since the wrong grouping of these two overlapping groups also causes a clear misgrouping of the next group we will submit the 3 peak groups (see right half of bottom plot) to the regroupR() function to fix the issues.
wrong.groups <- sort(sil_means[sil_means[,1]>=sil_means[1,1],1])[1:2]
wine.regrouped <- speaq::regroupR(grouped.peaks = wine.grouped,
list.to.regroup = wrong.groups,
min.samp.grp = 5,
max.dupli.prop = 0.1)## 'list.to.regroup' is not a list. Assuming it is a list with only one element.The plot below reveals the problem is clearly fixed now, but it requires more user interaction than wanted.
faulty.groupIndex <- sil_means[1,1]
ROI.ppm <- ppm.wine[faulty.groupIndex]
roiWidth.ppm <- 0.1
speaq::ROIplot(Y.spec = Spectra.wine,
X.ppm = ppm.wine,
ungrouped.peaks = wine.peaks,
grouped.peaks = wine.regrouped,
ROI.ppm = ROI.ppm,
roiWidth.ppm = roiWidth.ppm,
groupLabels = as.factor(wine.color))## Warning: Removed 2 rows containing missing values (geom_point).
With this peaks regrouped we can continue the analysis. first peak filling is applied and secondly the grouped peak data is converted to a so called feature matrix with samples for rows and features (peak groups) for columns whereby 1 matrix element indicates the peakvalue for a specific group and sample combo.
wine.filled <- speaq::PeakFilling(Y.grouped = wine.regrouped,
Y.spec = Spectra.wine,
max.index.shift = 50,
nCPU = 1) # nCPU set to 1 for the vignette build
wine.Features <- speaq::BuildFeatureMatrix(wine.filled)Now that we have these peak data they can either be incorporated in larger metabolomic experiments, where NMR spectroscopy data is often combined with LC-MS data (which is processed always in peak data format), or it can be analysed on its own, as we will demonstrate here.
Intermezzo: PCA
Now that we have the feature matrix we can quickly perform a PCA (principal component analysis) as a way of visualizing potential trends and groups in the data. Before any PCA it is advised to scale and center the data, here we will use the pareto scaling but other are available (see the SCANT() help file)
wine.Features.scaled <- speaq::SCANT(data.matrix = wine.Features,
type = c("pareto", "center")) common.pca <- prcomp(wine.Features.scaled)
loadings <- common.pca$rotation
scores <- common.pca$x
varExplained <- common.pca$sdev^2
barplot(varExplained/sum(varExplained),
main="Scree Plot",ylab="Proportion of variance explained",
xlab = "Principal comonent",
names.arg = as.character(seq(1,length(varExplained))))
plot.marks <- as.numeric(wine.color)
plot.marks[plot.marks == 1] <- 8
plot.marks[plot.marks == 2] <- 15
plot.marks[plot.marks == 3] <- 1
cp1 <- 1
cp2 <- 2
plot(scores[,cp1]/max(scores[,cp1]), scores[,cp2]/max(scores[,cp2]),
main=paste("score plot, PC",cp1," vs. PC",cp2,sep=""),
xlab=paste("PC",cp1,round(varExplained[cp1]/sum(varExplained),digits=2),""),
ylab=paste("PC",cp2,round(varExplained[cp2]/sum(varExplained),digits=2),""),
pch = plot.marks)
text(scores[,cp1]/max(scores[,cp1]),scores[,cp2]/max(scores[,cp2]), wine.color, cex=0.5, pos=4, col="red")
lines(x = c(-100,100), y = c(0,0))
lines(x = c(0,0), y = c(-100,100))
legend("topleft",
legend = c("red ","rosé ","white "),
pch = c(8,15,1),
y.intersp = 1.9)
From the PCA analysis it is already clear that there is a large difference between on the one side red wine and on the other side white and rosé wine.
Next we can use the relevant.features() function in speaq which uses linear models to produce a p-value corresponding to the null hypothesis that feature x is not related to the outcome vector. After a p-value correction for multiple testing (as with increasing numbers of even randomly generated features some will eventually be significantly related to the output vector) we can identify the relevant features. As noted in the paper this method is perfectly suited for outcome vectors containing either 2 classes or more classes with a numeric relationship. This is unfortunately not the case here as red, white and rosé are 3 categorical variables. Since there are only 2 rosé wines we will perform the analysis leaving these ones out. This does require however to rescale the raw data without the rosé wines, as otherwise they still have an influence on the dataset ### removing the rose wine
red.white.scaled <- speaq::SCANT(wine.Features[wine.color!="rose",],
type = c("pareto", "center"))
red.white.colors <- as.factor(as.character(wine.color)[wine.color!="rose"])A differential analysis based on linear models
p.all_bonf <- speaq::relevant.features.p(datamatrix = red.white.scaled,
responsevector = red.white.colors,
p.adj = "bonferroni")## Warning in speaq::relevant.features.p(datamatrix = red.white.scaled,
## responsevector = red.white.colors, : 'responsevector' is deprecated, please use
## the more general 'response' variable.## [1] "response was transformed to numeric"significant.features <- p.all_bonf[p.all_bonf$p.values<=0.05, ]
# order from most significant
significant.features <- significant.features[order(significant.features[,2]),]
head(significant.features)## index p.values
## 126 5170 2.880545e-18
## 125 5146 1.006329e-12
## 201 7867 3.355267e-07
## 93 4201 8.326499e-07
## 83 3852 2.797353e-06
## 124 5121 7.234151e-06It is clear that there are features well below the 0.05 p-value threshold. This was to be expected as the difference between the two classes was already apparent in the PCA analysis. Often when the PCA analysis does not reveal obvious results the differential analysis bases on linear models will nonetheless find significant features. We can now easily find which features correspond to the low p-values and consequently plot them.
peak_of_interest <- 5 # change this to the peak you want
interest.groupIndex <- significant.features$index[peak_of_interest]
interest.peakIndex <- as.numeric(rownames(significant.features))[peak_of_interest]
ROI.ppm <- ppm.wine[interest.groupIndex]
roiWidth.ppm <- 0.03
ggplot(p.all_bonf, aes(x=as.numeric(rownames(p.all_bonf)), y= -log10(p.values) )) +
geom_point(data = p.all_bonf[-interest.peakIndex,],
aes(x=as.numeric(rownames(p.all_bonf[-interest.peakIndex,])), y= -log10(p.values) ),
shape = 16) +
geom_point(data = p.all_bonf[interest.peakIndex,], aes(x=interest.peakIndex, y= -log10(p.values) ),
shape = 18,
size = 3,
colour ="#00B0F6" ) +
xlab("feature index") +
ylab("- log10 p-value") +
ggtitle("Bonferroni corrected p-values") +
geom_hline(aes(yintercept= -log10(0.05), color="red"),linetype = 2) +
guides(color=FALSE) +
theme_bw() +
theme(plot.title = element_text(lineheight = 0.8, face="bold", margin = margin(12,0,13,0),hjust = 0.5, size = 15),
text = element_text(size=14))## Warning: `guides(<scale> = FALSE)` is deprecated. Please use `guides(<scale> =
## "none")` instead.
speaq::ROIplot(Y.spec = Spectra.wine,
X.ppm = ppm.wine,
ungrouped.peaks = wine.peaks,
grouped.peaks = wine.regrouped,
ROI.ppm = ROI.ppm,
roiWidth.ppm = roiWidth.ppm,
groupLabels = as.factor(wine.color))
Or just plot the spectra of a single feature
peak_of_interest <- 4# change this number to any of the peaks you want to see
drawSpecPPM(Y.spec = Spectra.wine[wine.color != "rose", ],
X.ppm = ppm.wine,
groupFactor = red.white.colors,
title = paste("significant feature, p-value =",
format(significant.features$p.values[peak_of_interest],
scientific = TRUE,
digits=2),
sep=" "),
legend.extra.x = 1.1,
legend.extra.y = 0.9,
ROI = significant.features$index[peak_of_interest],
roiWidth = 100,
legendpos = "topright" )
Now that we have identified relevant features we can use the original speaq package to plot the spectra more clearly. Specifically we can use the dohCluster() function from speaq to align the raw spectra and consequently plot these with the drawSpecPPM() function as in the plot above.
peakList <- speaq::detectSpecPeaks(as.matrix(Spectra.wine),
nDivRange = 128,
scales = seq(1, 16, 2),
baselineThresh = 100,
SNR.Th = -1,
verbose=FALSE)
resFindRef <- findRef(peakList)
refInd <- resFindRef$refInd
aligned.spectra <- speaq::dohCluster(as.matrix(Spectra.wine),
peakList = peakList,
refInd = refInd,
maxShift = 200,
acceptLostPeak = TRUE, verbose=FALSE)
speaq::drawSpecPPM(Y.spec = aligned.spectra[wine.color != "rose", ],
X.ppm = ppm.wine,
groupFactor = red.white.colors,
title = paste("significant feature, p-value =",
format(significant.features$p.values[peak_of_interest],
scientific = TRUE,
digits=2),
sep=" "),
legend.extra.x = 1.1,
legend.extra.y = 0.9,
ROI = significant.features$index[peak_of_interest],
roiWidth = 100,
legendpos = "topright" ) 