---
title: "Processing quantitative metabolomics data with the qmtools package"
author:
  - name: Jaehyun Joo
    affiliation: University of Pennsylvania
    email: jaehyunjoo@outlook.com
output:
    BiocStyle::html_document:
        toc_float: true
vignette: >
  %\VignetteIndexEntry{Quantitative metabolomics data processing}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

```{r, include = FALSE}
knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  message = FALSE,
  fig.height = 10,
  fig.width = 10
)
```

# 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.

# Installation

The package can be installed using `r CRANpkg("BiocManager")`. In R session,
please type `BiocManager::install("qmtools")`.

# Data preparation

To demonstrate the use of the `qmtools` functions, we will use the
[FAAH knockout LC/MS](https://pubs.acs.org/doi/10.1021/bi0480335) data,
containing quantified LC/MS peaks from the spinal cords of 6 wild-type and
6 FAAH (fatty acid amide hydrolase) knockout mice.

```{r setup}
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
```

# 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.

```{r filtering}
dim(faahko_se) # 206 features
table(colData(faahko_se)$sample_group)

## Missing value filter based on 2 groups.
dim(removeFeatures(faahko_se, i = "raw",
                   group = colData(faahko_se)$sample_group,
                   cut = 0.80)) # nothing removed

dim(removeFeatures(faahko_se, i = "raw",
                   group = colData(faahko_se)$sample_group,
                   cut = 0.85)) # removed 65 features

## based on "WT" only
dim(removeFeatures(faahko_se, i = "raw",
                   group = colData(faahko_se)$sample_group,
                   levels = "WT", cut = 0.85))

```

In this vignette, we kept all features based on the cut-off: at least one group
contains >= 80% of non-missing values.

# 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.

```{r plotMiss, fig.wide = TRUE, fig.height = 5}
## 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
`r Biocpkg("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.

```{r knn, fig.wide = TRUE, fig.height = 5}
se <- imputeIntensity(faahko_se, i = "raw", name = "knn", method = "knn")
se # The result was stored in assays slot: "knn"

## 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")
```

# 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 `r Biocpkg("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.

```{r vsn, fig.wide = TRUE, fig.height = 5}
se <- normalizeIntensity(se, i = "knn", name = "knn_vsn", method = "vsn")
se # The result was stored in assays slot: "knn_vsn"

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")
```

# 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.

```{r PCA}
## PCA
m_pca <- reduceFeatures(se, i = "knn_vsn", method = "pca", ncomp = 2)
summary(m_pca)
```
```{r PLSDA}
## 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)
```

The dimension-reduction results can be plotted with the `plotReduced` function.
Each point (label) represents a sample. Data ellipses can be visualized.

```{r plotReduced, fig.wide = TRUE, fig.height = 5}
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")
```

# 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
metabolic features with the following steps:

1. Clusters features according to their retention times

2. 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.

```{r clusterFeatures}
se <- clusterFeatures(se, i = "knn_vsn", rtime_var = "rtmed",
                      rt_cut = 10, cor_cut = 0.7)
rowData(se)[, c("rtmed", "rtime_group", "feature_group")]
```

By 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

```{r rtime hclust, fig.wide = TRUE, fig.height = 5}
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
`r Biocpkg("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 `r Biocpkg("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.

```{r connected, fig.small = TRUE}
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")
```

```{r louvain, fig.small = TRUE}
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"`.

```{r pairs, fig.wide = FALSE}
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 `r Biocpkg("QFeatures")` (aggregation of intensities) and
`r CRANpkg("InterpretMSSpectrum")` (adduct annotation).

# Sample comparison

To test which metabolic features are different between two sets of samples, the
`compareSamples` function provides a convenient way to compute empirical Bayes
statistics using the `r Biocpkg("limma")` package interface. Note that this
function expects log-transformed feature intensities.

```{r compareSamples}
## Compute statisticis for the contrast: KO - WT
fit <- compareSamples(se, i = "knn_vsn", group = "sample_group",
                      class1 = "WT", class2 = "KO")

## List top 5 features
head(fit, 5)
```

Multiple covariates can be included to incorporate important sample and experiment
information.

```{r compareSamples covariates}
## Include covariates
colData(se)$covar <- c(rep(c("A", "B"), 6))
compareSamples(se, i = "knn_vsn", group = "sample_group",
               covariates = "covar", class1 = "WT", class2 = "KO",
               number = 5)
```

For more flexible model specifications (e.g., interaction model, multi-level
model), please use a standard workflow outlined in the `r Biocpkg("limma")`
package user's guide.

# References

Colin A. Smith (2021). faahKO: Saghatelian et al. (2004) FAAH knockout
LC/MS data. http://dx.doi.org/10.1021/bi0480335

Laurent Gatto, Johannes Rainer and Sebastian Gibb (2021). MsCoreUtils: Core
Utils for Mass Spectrometry Data.
https://github.com/RforMassSpectrometry/MsCoreUtils

Johannes Rainer (2022). MsFeatures: Functionality for Mass Spectrometry
Features.
https://github.com/RforMassSpectrometry/MsFeatures

Laurent Gatto and Christophe Vanderaa (2021). QFeatures: Quantitative features
for mass spectrometry data.
https://github.com/RforMassSpectrometry/QFeatures

Jan Lisec (2018). InterpretMSSpectrum: Interpreting High Resolution Mass
Spectra.
https://CRAN.R-project.org/package=InterpretMSSpectrum

Ritchie, M.E., Phipson, B., Wu, D., Hu, Y., Law, C.W., Shi, W., and Smyth, G.K.
(2015). limma powers differential expression analyses for RNA-sequencing and
microarray studies. Nucleic Acids Research 43(7), e47.
https://bioconductor.org/packages/limma

# Session info {-}

```{r session info}
sessionInfo()
```