Compounds that share a chromophore have similar UV spectra, so the spectra in a peak table can group peaks by compound class. This article shows two ways to do it: correlation with the spectrum of a chosen peak, and hierarchical clustering of all the spectra.
library(chromatographR)
data("Sa_warp")
pks <- get_peaks(Sa_warp, lambdas = 210)
pktab <- get_peaktable(pks)
pktab <- attach_ref_spectra(pktab, ref = "max.cor")After assigning reference spectra to our peak table, we can use
Pearson correlations to locate peaks within a chosen similarity
threshold of a reference spectrum. To do this, we first construct a
correlation matrix using the cor function.
Analysis of spectral correlations
cor_mat <- cor(pktab$ref_spectra)We can then plot spectra falling within a desired similarity
threshold (t) to a reference peak (in this case
V18).
ref <- "V18"
wl <- as.numeric(rownames(pktab$ref_spectra))
par(mfrow=c(2,2))
idx_90 <- which(cor_mat[,ref] > .9)
matplot(wl, pktab$ref_spectra[,idx_90], type='l',
xlab="Wavelength (nm)", ylab="Relative absorbance")
legend("top", "t = 0.90", bty = "n")
idx_95 <- which(cor_mat[,ref] > .95)
matplot(wl, pktab$ref_spectra[,idx_95], type='l',
xlab="Wavelength (nm)", ylab="Relative absorbance")
legend("top", "t = 0.95", bty = "n")
idx_97 <- which(cor_mat[,ref] > .97)
matplot(wl, pktab$ref_spectra[,idx_97], type='l',
xlab="Wavelength (nm)", ylab="Relative absorbance")
legend("top", "t = 0.97", bty = "n")
idx_99 <- which(cor_mat[,ref] > .99)
matplot(wl, pktab$ref_spectra[,idx_99], type='l',
xlab="Wavelength (nm)", ylab="Relative absorbance")
legend("top", "t = 0.99", bty = "n")
We can also extract retention times for spectra above the desired spectral similarity threshold with the reference spectra.
rts_99 = pktab$pk_meta["rt", idx_99]
rts_99
#> V4 V6 V14 V18 V19 V23
#> rt 11.51706 12.19639 14.41222 15.5188 15.75304 16.7874Below we plot traces of our four chromatograms at 210 nm. The 6 peaks that match the reference spectrum at the 99% level are marked with solid black lines and the 7 peaks that match the reference spectrum at the 97% level are marked with dotted gray lines.
par(mfrow = c(1, 1))
plot_chroms(Sa_warp, lambdas=210)
abline(v = rts_99)
abline(v = pktab$pk_meta["rt", idx_97], lty=2, col = "darkgray")
If we are interested in the combined area of these peaks, we could also use these indices to sum them together.
rowSums(pktab$tab[, idx_99, drop = FALSE])
#> 119 121 122 458
#> 175.8716 121.1417 155.2266 207.1897Hierarchical clustering of spectra
chromatographR also includes a function
(cluster_spectra) to group spectra using hierarchical
clustering. We can set the seed with the iseed argument to
ensure repeatability among instantiations of this vignette. Clusters of
20 or more peaks span most of the table and say little about any one
chromophore, so we drop them with max_size.
par(mfrow=c(2,2))
cl <- cluster_spectra(pktab, alpha = 0.05, max_size = 20, iseed = 1)
The output is a dendrogram of peaks, annotated with bootstrap values
(cluster support estimated by resampling) at each node. It also plots
the spectra of clusters whose bootstrap p-value falls below the
significance level specified by alpha. In this case, we
recovered 3 clusters. We can find the cluster containing our reference
peak and access its peaks as follows:
ref_cl <- Filter(function(k) ref %in% k@peaks, cl$clusters)[[1]]
ref_cl@peaks
#> [1] "V4" "V6" "V8" "V14" "V18" "V19" "V23"This cluster holds the same peaks that correlate with V18 above 0.95:
Choosing an approach
A correlation threshold needs a reference peak and a cutoff, and
returns the peaks whose spectra resemble that one. Clustering needs no
reference peak: it groups the spectra of the whole table and estimates
the support for each group by bootstrapping, though you still choose
alpha and max_size. Here both found the same 7
peaks. Use a threshold when you know which compound you are looking for,
and clustering to see which groups of spectra the table contains.