Processing quantitative metabolomics data with the qmtools package
1 November 2022
1 Introduction
The qmtools package provides basic tools for imputation, normalization, and
dimension-reduction of metabolomics data with the standard
SummarizedExperiment class. It also offers several helper functions to assist
visualization of data. This vignette gives brief descriptions of
these tools with toy examples.
2 Installation
The package can be installed using BiocManager. In R session,
please type BiocManager::install("qmtools").
3 Data preparation
To demonstrate the use of the qmtools functions, we will use the
FAAH knockout LC/MS data,
containing quantified LC/MS peaks from the spinal cords of 6 wild-type and
6 FAAH (fatty acid amide hydrolase) knockout mice.
library(qmtools)
library(SummarizedExperiment)
library(vsn)
library(pls)
library(ggplot2)
library(patchwork)
set.seed(1e8)
data(faahko_se)
## Only keep the first assay for the vignette
assays(faahko_se)[2:4] <- NULL
faahko_se
#> class: SummarizedExperiment
#> dim: 206 12
#> metadata(0):
#> assays(1): raw
#> rownames(206): FT001 FT002 ... FT205 FT206
#> rowData names(10): mzmed mzmin ... WT peakidx
#> colnames(12): ko15.CDF ko16.CDF ... wt21.CDF wt22.CDF
#> colData names(2): sample_name sample_group4 Feature filtering
Metabolomics data often contains a large number of uninformative features that
can hinder downstream analysis. The removeFeatures function attempts to
identify such features and remove them from the data based on missing values,
quality control (QC) replicates, and blank samples with the following methods:
Proportions of missing values: retain features if there is at least one group with a proportion of non-missing values above a cut-off.
Relative standard deviation: remove features if QC replicates show low reproducibility.
Intraclass correlation coefficient (ICC): retain features if a feature has relatively high variability across biological samples compared to QC replicates.
QC/blank ratio: remove features with low abundance that may have non-biological origin.
The FAAH knockout data does not include QC and blank samples. Here, we just illustrate missing value-based filtering.
dim(faahko_se) # 206 features
#> [1] 206 12
table(colData(faahko_se)$sample_group)
#>
#> KO WT
#> 6 6
## Missing value filter based on 2 groups.
dim(removeFeatures(faahko_se, i = "raw",
group = colData(faahko_se)$sample_group,
cut = 0.80)) # nothing removed
#> [1] 206 12
dim(removeFeatures(faahko_se, i = "raw",
group = colData(faahko_se)$sample_group,
cut = 0.85)) # removed 65 features
#> [1] 141 12
## based on "WT" only
dim(removeFeatures(faahko_se, i = "raw",
group = colData(faahko_se)$sample_group,
levels = "WT", cut = 0.85))
#> [1] 104 12In this vignette, we kept all features based on the cut-off: at least one group contains >= 80% of non-missing values.
5 Imputation
Missing values are common in metabolomics data. For example, ions may have
a low abundance that does not reach the limit of detection of the instrument.
Unexpected stochastic fluctuations and technical error may also cause
missing values even though ions present at detectable levels.
We could use the plotMiss function to explore the mechanisms generating
the missing values. The bar plot on the left panel shows the amount of missing
values in each samples and the right panel helps to identify the structure of
missing values with a hierarchically-clustered heatmap.
## Sample group information
g <- factor(colData(faahko_se)$sample_group, levels = c("WT", "KO"))
## Visualization of missing values
plotMiss(faahko_se, i = "raw", group = g)
Overall, the knockout mice have a higher percentage of missing values.
The features on top of the heatmap in general only present at the knockout mice,
suggesting that some of missing values are at least not random
(perhaps due to altered metabolisms by the experimental condition).
In almost all cases, visualization and inspection of missing values are
a time-intensive step, but greatly improve the ability to uncover the nature of
missing values in data and help to choose an appropriate imputation method.
The imputation of missing values can be done with the imputeIntensity
function. Several imputation methods are available such as k-Nearest Neighbor
(kNN), Random Forest (RF), Bayesian PCA, and other methods available in
MsCoreUtils. By default, the kNN is used to impute missing values
using the Gower distance. The kNN is a distance-based
algorithm that typically requires to scale the data to avoid variance-based
weighing. Since the Gower distance used, the imputation can be performed
with the original scales, which may be helpful to non-technical users.
se <- imputeIntensity(faahko_se, i = "raw", name = "knn", method = "knn")
se # The result was stored in assays slot: "knn"
#> class: SummarizedExperiment
#> dim: 206 12
#> metadata(0):
#> assays(2): raw knn
#> rownames(206): FT001 FT002 ... FT205 FT206
#> rowData names(10): mzmed mzmin ... WT peakidx
#> colnames(12): ko15.CDF ko16.CDF ... wt21.CDF wt22.CDF
#> colData names(2): sample_name sample_group
## Standardization of input does not influence the result
m <- assay(faahko_se, "raw")
knn_scaled <- as.data.frame(
imputeIntensity(scale(m), method = "knn") # Can accept matrix as an input
)
knn_unscaled <- as.data.frame(assay(se, "knn"))
idx <- which(is.na(m[, 1]) | is.na(m[, 2])) # indices for missing values
p1 <- ggplot(knn_unscaled[idx, ], aes(x = ko15.CDF, y = ko16.CDF)) +
geom_point() + theme_bw()
p2 <- ggplot(knn_scaled[idx, ], aes(x = ko15.CDF, y = ko16.CDF)) +
geom_point() + theme_bw()
p1 + p2 + plot_annotation(title = "Imputed values: unscaled vs scaled")
6 Normalization
In metabolomics, normalization is an important part of data processing to reduce
unwanted non-biological variations
(e.g., variation due to sample preparation and handling).
The normalizeIntensity function provides several data-driven normalization
methods such as Probabilistic Quotient Normalization (PQN),
Variance-Stabilizing Normalization (VSN), Cyclic LOESS normalization, and other
methods available in MsCoreUtils.
Here, we will apply the VSN to the imputed intensities. Note that the VSN
produces glog-transformed (generalized log transform) feature intensities.
The consequence of normalization can be visualized with the plotBox function.
se <- normalizeIntensity(se, i = "knn", name = "knn_vsn", method = "vsn")
se # The result was stored in assays slot: "knn_vsn"
#> class: SummarizedExperiment
#> dim: 206 12
#> metadata(0):
#> assays(3): raw knn knn_vsn
#> rownames(206): FT001 FT002 ... FT205 FT206
#> rowData names(10): mzmed mzmin ... WT peakidx
#> colnames(12): ko15.CDF ko16.CDF ... wt21.CDF wt22.CDF
#> colData names(2): sample_name sample_group
p1 <- plotBox(se, i = "knn", group = g, log2 = TRUE) # before normalization
p2 <- plotBox(se, i = "knn_vsn", group = g) # after normalization
p1 + p2 + plot_annotation(title = "Before vs After normalization")
7 Dimension-reduction
The metabolomics data generally consist of a large number of features, and
dimension-reduction techniques are often used for modeling and visualization to
uncover latent structure underlying many features. The reduceFeatures can be
used to perform dimension-reduction of the data. Currently, Principal Component
Analysis (PCA), Partial Least Square-Discriminant Analysis (PLS-DA) and
t-distributed stochastic neighbor (t-SNE) are supported. The function returns
a matrix containing dimension-reduced data with several attributes that can be
summarized with the summary function.
## PCA
m_pca <- reduceFeatures(se, i = "knn_vsn", method = "pca", ncomp = 2)
summary(m_pca)
#> Reduction method: PCA (SVD)
#> Input centered before PCA: TRUE
#> Input scaled before PCA: FALSE
#> Number of PCs calculated: 2
#> Importance of PC(s):
#> PC1 PC2
#> Proportion of Variance 0.2265 0.1599
#> Cumulative Proportion 0.2265 0.3865## PLS-DA (requires information about each sample's group)
m_plsda <- reduceFeatures(se, i = "knn_vsn", method = "plsda", ncomp = 2, y = g)
summary(m_plsda)
#> Reduction method: PLS-DA (kernelpls)
#> Y responses: WT KO
#> Input centered before PLS-DA: TRUE
#> Input scaled before PLS-DA: FALSE
#> Number of components considered: 2
#> Amount of X variance explained by each component:
#> Comp 1 Comp 2
#> Explained % 21.64 14.02
#> Cumulative % 21.64 35.65The dimension-reduction results can be plotted with the plotReduced function.
Each point (label) represents a sample. Data ellipses can be visualized.
p_pca <- plotReduced(m_pca, group = g)
p_plsda <- plotReduced(m_plsda, label = TRUE, ellipse = TRUE)
p_pca + p_plsda + plot_annotation(title = "PCA and PLS-DA")
8 Feature clustering
For soft ionization methods such as LC/ESI-MS, a bulk of ions can be generated
from an individual compound upon ionization. Because we typically interested in
compounds rather than different ion species, identifying features from the same
compound is necessary. The clusterFeatures function attempts to cluster
with the following steps:
Clusters features according to their retention times
Based on the initial grouping, clusters features according to the intensity correlations
After the clustering procedures, the function adds the rtime_group and
feature_group columns to the rowData of SummarizedExperiment input.
se <- clusterFeatures(se, i = "knn_vsn", rtime_var = "rtmed",
rt_cut = 10, cor_cut = 0.7)
rowData(se)[, c("rtmed", "rtime_group", "feature_group")]
#> DataFrame with 206 rows and 3 columns
#> rtmed rtime_group feature_group
#> <numeric> <factor> <character>
#> FT001 2789.04 FG.01 FG.01.01
#> FT002 2788.31 FG.01 FG.01.02
#> FT003 2718.79 FG.02 FG.02.01
#> FT004 3024.33 FG.03 FG.03.01
#> FT005 3684.80 FG.04 FG.04.01
#> ... ... ... ...
#> FT202 3001.79 FG.14 FG.14.03
#> FT203 3714.93 FG.24 FG.24.07
#> FT204 3403.03 FG.50 FG.50.04
#> FT205 3614.09 FG.49 FG.49.05
#> FT206 3817.51 FG.41 FG.41.02By default, the retention time-based grouping is performed with a hierarchical clustering based on the Manhattan distance (i.e., differences in retention times). The equivalent steps are
rts <- rowData(se)$rtmed
rt_cut <- 10
fit <- hclust(dist(rts, "manhattan"))
plot(as.dendrogram(fit), leaflab = "none")
rect.hclust(fit, h = rt_cut)
The retention-time based grouping can also be conducted with the algorithms
(groupClosest and groupConsecutive) available in the
MsFeatures package.
Upon the initial grouping, each retention-based time group is further clustered
according to the intensity correlations since features may be originated from
different co-eluting compounds, not from a single entity. By default, the
function creates a graph where correlations serve as edge weights
while low correlations defined by a user-specified cut-off ignored.
cor_grouping = "connected" simply assigns connected features into the same
feature group whereas cor_grouping = louvain further applies the Louvain
algorithm to the graph to identify densely connected features.
The groupSimiarityMatrix approach from the MsFeatures
package is also supported.
The feature clustering results can be visualized with the plotRTgroup
function. A group of features in the same feature group will be displayed with
the same color. Each vertex represents a feature and each weight represent a
correlation between features.
se_connected <- clusterFeatures(se, i = "knn_vsn", rtime_var = "rtmed",
rt_cut = 10, cor_cut = 0.7,
cor_grouping = "connected")
plotRTgroup(se_connected, i = "knn_vsn", group = "FG.22")
se_louvain <- clusterFeatures(se, i = "knn_vsn", rtime_var = "rtmed",
rt_cut = 10, cor_cut = 0.7,
cor_grouping = "louvain")
plotRTgroup(se_louvain, i = "knn_vsn", group = "FG.22")
More details could be plotted by specifying type = "pairs".
plotRTgroup(se_louvain, i = "knn_vsn", group = "FG.22", type = "pairs")
The clustering results can be used to deal with the redundancy of the data with other packages such as QFeatures (aggregation of intensities) and InterpretMSSpectrum (adduct annotation).
9 References
Colin A. Smith (2021). faahKO: Saghatelian et al. (2004) FAAH knockout LC/MS data. R package version 1.32.0. http://dx.doi.org/10.1021/bi0480335
Laurent Gatto, Johannes Rainer and Sebastian Gibb (2021). MsCoreUtils: Core Utils for Mass Spectrometry Data. R package version 1.4.0. https://github.com/RforMassSpectrometry/MsCoreUtils
Johannes Rainer (2022). MsFeatures: Functionality for Mass Spectrometry Features. R package version 1.3.0. https://github.com/RforMassSpectrometry/MsFeatures
Laurent Gatto and Christophe Vanderaa (2021). QFeatures: Quantitative features for mass spectrometry data. R package version 1.2.0. https://github.com/RforMassSpectrometry/QFeatures
Jan Lisec (2018). InterpretMSSpectrum: Interpreting High Resolution Mass Spectra. R package version 1.2. https://CRAN.R-project.org/package=InterpretMSSpectrum
Session info
sessionInfo()
#> R version 4.2.1 (2022-06-23)
#> Platform: x86_64-pc-linux-gnu (64-bit)
#> Running under: Ubuntu 20.04.5 LTS
#>
#> Matrix products: default
#> BLAS: /home/biocbuild/bbs-3.16-bioc/R/lib/libRblas.so
#> LAPACK: /home/biocbuild/bbs-3.16-bioc/R/lib/libRlapack.so
#>
#> locale:
#> [1] LC_CTYPE=en_US.UTF-8 LC_NUMERIC=C
#> [3] LC_TIME=en_GB 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] stats4 stats graphics grDevices utils datasets methods
#> [8] base
#>
#> other attached packages:
#> [1] patchwork_1.1.2 ggplot2_3.3.6
#> [3] pls_2.8-1 vsn_3.66.0
#> [5] qmtools_1.2.0 SummarizedExperiment_1.28.0
#> [7] Biobase_2.58.0 GenomicRanges_1.50.0
#> [9] GenomeInfoDb_1.34.0 IRanges_2.32.0
#> [11] S4Vectors_0.36.0 BiocGenerics_0.44.0
#> [13] MatrixGenerics_1.10.0 matrixStats_0.62.0
#> [15] BiocStyle_2.26.0
#>
#> loaded via a namespace (and not attached):
#> [1] bitops_1.0-7 webshot_0.5.4 RColorBrewer_1.1-3
#> [4] httr_1.4.4 tools_4.2.1 bslib_0.4.0
#> [7] utf8_1.2.2 R6_2.5.1 affyio_1.68.0
#> [10] DBI_1.1.3 lazyeval_0.2.2 colorspace_2.0-3
#> [13] nnet_7.3-18 sp_1.5-0 withr_2.5.0
#> [16] tidyselect_1.2.0 gridExtra_2.3 compiler_4.2.1
#> [19] preprocessCore_1.60.0 cli_3.4.1 TSP_1.2-1
#> [22] DelayedArray_0.24.0 plotly_4.10.0 labeling_0.4.2
#> [25] bookdown_0.29 sass_0.4.2 scales_1.2.1
#> [28] lmtest_0.9-40 DEoptimR_1.0-11 robustbase_0.95-0
#> [31] affy_1.76.0 proxy_0.4-27 stringr_1.4.1
#> [34] digest_0.6.30 rmarkdown_2.17 ca_0.71.1
#> [37] XVector_0.38.0 pkgconfig_2.0.3 htmltools_0.5.3
#> [40] fastmap_1.1.0 limma_3.54.0 highr_0.9
#> [43] htmlwidgets_1.5.4 rlang_1.0.6 jquerylib_0.1.4
#> [46] generics_0.1.3 farver_2.1.1 zoo_1.8-11
#> [49] jsonlite_1.8.3 car_3.1-1 dendextend_1.16.0
#> [52] dplyr_1.0.10 RCurl_1.98-1.9 magrittr_2.0.3
#> [55] GenomeInfoDbData_1.2.9 Matrix_1.5-1 Rcpp_1.0.9
#> [58] munsell_0.5.0 fansi_1.0.3 MsCoreUtils_1.10.0
#> [61] abind_1.4-5 viridis_0.6.2 lifecycle_1.0.3
#> [64] stringi_1.7.8 yaml_2.3.6 carData_3.0-5
#> [67] MASS_7.3-58.1 zlibbioc_1.44.0 plyr_1.8.7
#> [70] grid_4.2.1 lattice_0.20-45 magick_2.7.3
#> [73] knitr_1.40 pillar_1.8.1 igraph_1.3.5
#> [76] ranger_0.14.1 boot_1.3-28 reshape2_1.4.4
#> [79] codetools_0.2-18 glue_1.6.2 evaluate_0.17
#> [82] laeken_0.5.2 data.table_1.14.4 BiocManager_1.30.19
#> [85] vcd_1.4-10 vctrs_0.5.0 foreach_1.5.2
#> [88] VIM_6.2.2 gtable_0.3.1 purrr_0.3.5
#> [91] tidyr_1.2.1 clue_0.3-62 heatmaply_1.4.0
#> [94] assertthat_0.2.1 cachem_1.0.6 xfun_0.34
#> [97] e1071_1.7-12 class_7.3-20 viridisLite_0.4.1
#> [100] seriation_1.4.0 tibble_3.1.8 iterators_1.0.14
#> [103] registry_0.5-1 cluster_2.1.4