---
title: "Compounding (grouping) of LC-MS features"
package: xcms
output:
  BiocStyle::html_document:
    toc_float: true
vignette: >
  %\VignetteIndexEntry{LC-MS feature grouping}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
  %\VignetteDepends{xcms,msdata,BiocStyle,faahKO,pheatmap,MsFeatures}
  %\VignettePackage{xcms}
  %\VignetteKeywords{mass spectrometry, metabolomics}
---

```{r biocstyle, echo = FALSE, results = "asis"}
BiocStyle::markdown()
knitr::opts_chunk$set(echo = TRUE, warning = FALSE, message = FALSE)
```

**Package**: `r Biocpkg("xcms")`<br />
**Authors**: Johannes Rainer<br />
**Modified**: `r file.info("LC-MS-feature-grouping.Rmd")$mtime`<br />
**Compiled**: `r date()`

```{r init, results = "hide", echo = FALSE}
## Silently loading all packages
library(BiocStyle)
library(xcms)
library(MsFeatures)
register(SerialParam())

```

# Introduction

In a typical LC-MS-based metabolomics experiment compounds eluting from the
chromatography are first ionized before being measured by mass spectrometry
(MS). During the ionization different (multiple) ions can be generated from the
same compound which all will be measured by MS. In general, the resulting data
is then pre-processed to identify chromatographic peaks in the data and to group
these across samples in the correspondence analysis. The result are distinct
LC-MS features, characterized by their specific m/z and retention time
range. Different ions generated during ionization will be detected as different
features. *Compounding* aims now at grouping such features presumably
representing signal from the same originating compound to reduce data set
complexity (and to aid in subsequent annotation steps). General MS feature
grouping functionality if defined by the `r Biocpkg("MsFeatures")` package with
additional functionality being implemented in the `r Biocpkg("xcms")` package to
enable the compounding of LC-MS data.

This document provides a simple compounding workflow using
`r Biocpkg("xcms")`. Note that the present functionality does not (yet)
*aggregate* or combine the actual features per values, but does only define
the feature groups (one per compound).


# Compounding of LC-MS data

We demonstrate the compounding (feature grouping) functionality on the simple
toy data set used also in the `r Biocpkg("xcms")` package and provided through
the *faahKO* package. This data set consists of samples from 4 mice with
knock-out of the fatty acid amide hydrolase (FAAH) and 4 wild type
mice. Pre-processing of this data set is described in detail in the *main*
vignette of the *xcms* package. Below we load all required packages and the
result from this pre-processing updating also the location of the respective raw
data files on the current machine.

```{r load-data}
library(MSnbase)
library(xcms)
library(faahKO)
library(MsFeatures)

xmse <- loadXcmsData("xmse")
```

Before performing the feature grouping we inspect the result object. With
`featureDefinitions` we can extract the results from the correspondence
analysis.

```{r fdev}
featureDefinitions(xmse) |> head()
```

Each row in this data frame represents the definition of one feature, with its
average and range of m/z and retention time. Column `"peakidx"` provides the
index of each chromatographic peak which is assigned to the feature in the
`chromPeaks` matrix of the result object. The `featureValues` function allows to
extract *feature values*, i.e. a matrix with feature abundances, one row per
feature and columns representing the samples of the present data set.

Below we extract the feature values with and without *filled-in* peak
data. Without the gap-filled data only abundances from **detected**
chromatographic peaks are reported. In the gap-filled data, for samples in which
no chromatographic peak for a feature was detected, all signal from the m/z -
retention time range defined based on the detected chromatographic peaks was
integrated.

```{r filled-not-filled}
head(featureValues(xmse, filled = FALSE))
head(featureValues(xmse, filled = TRUE))
```

In total `r nrow(featureDefinitions(xmse))` features have been defined in the
present data set, many of which most likely represent signal from different ions
(adducts or isotopes) of the same compound. The aim of the grouping functions of
are now to define which features most likely come from the same original
compound. The feature grouping functions base on the following
assumptions/properties of LC-MS data:

- Features (ions) of the same compound should have similar retention time.
- The abundance of features (ions) of the same compound should have a similar
  pattern across samples, i.e. if a compound is highly concentrated in one
  sample and low in another, all ions from it should follow the same pattern.
- The peak shape of extracted ion chromatograms (EIC) of features of the same
  compound should be similar as it should follow the elution pattern of the
  original compound from the LC.

The main method to perform the feature grouping is called `groupFeatures` which
takes an *xcms* result object (i.e., an `XcmsExperiment` or `XCMSnExp`) as
input as well as a parameter object to chose the grouping algorithm and specify
its settings. *xcms* provides and supports the following grouping approaches:

- `SimilarRtimeParam`: perform an initial grouping based on similar retention
  time.
- `AbundanceSimilarityParam`: perform a feature grouping based on correlation
  of feature abundances (values) across samples.
- `EicSimilarityParam`: perform a feature grouping based on correlation of
  EICs.

Calling `groupFeatures` on an *xcms* result object will perform a feature
grouping assigning each feature in the data set to a *feature group*. These
feature groups are stored as an additional column called `"feature_group"` in
the `featureDefinition` data frame of the result object and can be accessed with
the `featureGroups` function. Any subsequent `groupFeature` call will
*sub-group* (refine) the identified feature groups further. It is thus possible
to use a single grouping approach, or to combine multiple of them to generate
the desired feature grouping in an incremental fashion. While the individual
feature grouping algorithms can be called in any order, it is advisable to use
the `EicSimilarityParam` as last refinement step, because it is computationally
very expensive, especially if applied to a result object without pre-defined
feature groups or if the feature groups are very large.


## Grouping of features by similar retention time

The most intuitive and simple way to group features is based on their retention
time. Before we perform this initial grouping we evaluate retention times and
m/z of all features in the present data set.

```{r feature-rt-mz-plot, fig.width = 8, fig.height = 6, fig.cap = "Plot of retention times and m/z for all features in the data set."}
plot(featureDefinitions(xmse)$rtmed, featureDefinitions(xmse)$mzmed,
     xlab = "retention time", ylab = "m/z", main = "features",
     col = "#00000080", pch = 21, bg = "#00000040")
grid()
```

Several features with about the same retention time (but different m/z) can be
spotted, especially at the beginning of the LC. We thus below group features
within a retention time window of 10 seconds into *feature groups*.

```{r}
xmse <- groupFeatures(xmse, param = SimilarRtimeParam(10))
```

The results from the feature grouping can be accessed with the `featureGroups`
function. Below we determine the size of each of these feature groups (i.e. how
many features are grouped together).

```{r}
table(featureGroups(xmse))
```

In addition we visualize these feature groups with the `plotFeatureGroups`
function which shows all features in the m/z - retention time space with grouped
features being connected with a line.

```{r feature-groups-rtime-plot, fig.width = 8, fig.height = 6, fig.cap = "Feature groups defined with a rt window of 10 seconds"}
plotFeatureGroups(xmse, pch = 21, lwd = 2, col = "#00000040",
                  bg = "#00000020")
grid()
```

Let's assume we don't agree with this feature grouping, also knowing that there
were quite large shifts in retention times between runs. We thus first remove
any defined feature groups assigning them a value of `NULL` and then re-perform
the feature grouping using a larger rt window.

```{r repeat}
## Remove previous feature grouping results to repeat the rtime-based
## feature grouping with different setting
featureDefinitions(xmse)$feature_group <- NULL

## Repeat the grouping
xmse <- groupFeatures(xmse, SimilarRtimeParam(20))
table(featureGroups(xmse))
```

```{r feature-groups-rtime-plot2, fig.width = 8, fig.height = 6, fig.cap = "Feature groups defined with a rt window of 20 seconds"}
plotFeatureGroups(xmse, pch = 21, lwd = 2, col = "#00000040", bg = "#00000020")
grid()
```

Grouping by similar retention time grouped the in total
`r nrow(featureDefinitions(xmse))` features into
`r length(unique(featureGroups(xmse)))` feature groups.


## Grouping of features by abundance correlation across samples

Assuming we are OK with the *crude* initial feature grouping from the previous
section, we can next *refine* the feature groups considering also the feature
abundances across samples. We can use the `groupFeatures` method with an
`AbundanceSimilarityParam` object. This approach performs a pairwise correlation
between all (non-missing) feature values (abundances; across samples) between
all features of a predefined feature group (such as defined in the previous
section). Features that have a correlation `>= threshold` are grouped
together. Feature grouping based on this approach works best for features with a
higher variability in their concentration across samples. Parameter `subset`
allows to restrict the analysis to a subset of samples and allows thus to
e.g. exclude QC sample pools from this correlation as these could artificially
increase the correlation. Other parameters are passed directly to the internal
`featureValues` call that extracts the feature values on which the correlation
should be performed.

Before performing the grouping we could also evaluate the correlation of
features based on their (log2 transformed) abundances across samples with a
heatmap.

```{r abundance-correlation-heatmap, fig.cap = "Correlation of features based on feature abundances.", fig.width = 6, fig.height = 16}
library(pheatmap)
fvals <- log2(featureValues(xmse, filled = TRUE))

cormat <- cor(t(fvals), use = "pairwise.complete.obs")
ann <- data.frame(fgroup = featureGroups(xmse))
rownames(ann) <- rownames(cormat)

res <- pheatmap(cormat, annotation_row = ann, cluster_rows = TRUE,
                cluster_cols = TRUE)
```

Some large correlations can be observed for several groups of features, but many
of them are not within the same *feature group* defined in the previous section
(i.e. are not eluting at the same time). These might thus represent correlated
metabolites, but not ions or adducts from the same metabolite.

Below we use the `groupFeatures` with the `AbundanceSimilarityParam` to group
features with a correlation coefficient higher than 0.7 including both detected
and filled-in signal. Whether filled-in or only detected signal should be used
in the correlation analysis should be evaluated from data set to data set. By
specifying `transform = log2` we tell the function to log2 transform the
abundance prior to the correlation analysis. See the help page for
`groupFeatures` with `AbundanceSimilarityParam` in the `xcms` package for
details and options.

```{r abundance-correlation}
xmse <- groupFeatures(
    xmse, AbundanceSimilarityParam(threshold = 0.7, transform = log2),
    filled = TRUE)
table(featureGroups(xmse))
```

Many of the larger retention time-based feature groups have been splitted into
two or more sub-groups based on the correlation of their feature abundances
(e.g. the feature group *FG.004* has been further split into feature groups
*FG.004.001*, *FG.004.002* and *FG.004.003*). We evaluate this further
refinement for feature group `"FG.040"` by plotting their pairwise
correlation. To better visualize the feature grouping we in addition define a
custom *panel plot* for the `pairs` function that plots data points in blue for
features with a correlation coefficient above the selected threshold (`0.7`) and
red otherwise.

```{r abundance-correlation-fg040, fig.width = 8, fig.height = 8, fig.cap = "Pairwise correlation plot for all features initially grouped into the feature group FG.040."}
cor_plot <- function(x, y) {
    C <- cor(x, y, use = "pairwise.complete.obs")
    col <- ifelse(C >= 0.7, yes = "#0000ff80", no = "#ff000080")
    points(x, y, pch = 16, col = col)
    grid()
}
fts <- grep("FG.040", featureGroups(xmse))
pairs(t(fvals[fts, ]), gap = 0.1, main = "FG.040", panel = cor_plot)
```

Indeed, correlation seems only to be present between some of the features in
this retention time feature group, e.g. clearly between *FT273* and *FT274* and
also between *FT143* and *FT273*. Note however that this abundance correlation
suffers from relatively few samples (8 in total), and a relatively small
variance in abundances across these samples.

After feature grouping by abundance correlation, the
`r nrow(featureDefinitions(xmse))` features have been grouped into
`r length(unique(featureGroups(xmse)))` feature groups.


## Grouping of features by similarity of their EICs

The chromatographic peak shape of an ion of a compound should be highly similar
to the elution pattern of this compound. Thus, features from the same compound
are assumed to have similar peak shapes of their EICs **within the same
sample**. A grouping of features based on similarity of their EICs can be
performed with the `groupFeatures` and the `EicSimilarityParam` object. It is
advisable to perform the peak shape correlation only on a subset of samples
(because peak shape correlation is computationally intense and because
chromatographic peaks of low intensity features are notoriously noisy). The
`EicSimilarityParam` approach allows to select the number of *top n* samples
(ordered by total intensity of feature abundances per feature group) on which
the correlation should be performed with parameter `n`. With `n =3`, the 3
samples with the highest signal for all features in a respective feature group
will be first identified and then a pairwise similarity calculation will be
performed on the EICs between each of these samples. The resulting similarity
score from these 3 samples will then be aggregated into a single score by taking
the 75% quantile across the 3 samples. This value is then subsequently compared
with the cut-off for similarity (parameter `threshold`) and features with a
score `>= threshold` are grouped into the same feature group.

Below we group the features based on similarity of their EICs in the two samples
with the highest total signal for the respective feature groups. By default, a
Pearson correlation coefficient is used as similarity score but any
similarity/distance metric function could be used instead (parameter `FUN` of
the `EicSimilarityParam` - see the respective help page `?EicSimilarityParam`
for details and options). We define as a threshold a correlation coefficient of
0.7.

```{r correlate-eic, message = FALSE, warning = FALSE}
xmse <- groupFeatures(xmse, EicSimilarityParam(threshold = 0.7, n = 2))
```

This is the computationally most intense approach since it involves also loading
the raw MS data to extract the ion chromatograms for each feature. The results
of the grouping are shown below.

```{r correlate-eic-result}
table(featureGroups(xmse))
```

In many cases, pre-defined feature groups (by the retention time similarity and
abundance correlation) were not further subdivided. Below we evaluate some of
the feature groups, starting with *FG.008.001* which was split into two
different feature groups based on EIC correlation. We first extract the EICs for
all features from this initial feature group. With `n = 1` we specify to extract
the EIC only from the sample with the highest intensity.

```{r}
fidx <- grep("FG.013.001.", featureGroups(xmse))
eics <- featureChromatograms(
    xmse, features = rownames(featureDefinitions(xmse))[fidx],
    filled = TRUE, n = 1)
```

Next we plot the EICs using a different color for each of the subgroups. With
`peakType = "none"` we disable the highlighting of the detected chromatographic
peaks.

```{r example-1-eic, fig.width = 8, fig.height = 6, fig.cap = "Feature EICs per sample for features from a feature group defined by rentention time and feature abudances across samples. Features with high correlation of their EICs are shown in the same color."}
cols <- c("#ff000080", "#0000ff80")
names(cols) <- unique(featureGroups(xmse)[fidx])

plotChromatogramsOverlay(eics, col = cols[featureGroups(xmse)[fidx]],
                         lwd = 2, peakType = "none")
```

```{r example-1-eic-norm, fig.width = 8, fig.height = 6, fig.cap = "Feature EICs per sample normalized to an absolute intensity of 1 for features from a feature group defined by rentention time and feature abudances across samples. Features with high correlation of their EICs are shown in the same color."}
plotChromatogramsOverlay(normalize(eics),
                         col = cols[featureGroups(xmse)[fidx]],
                         lwd = 2, peakType = "none")
```


One of the features (highlighted in red in the plots above) within the original
feature group was separated from the other two because of a low similarity of
their EICs. In fact, the feature's EIC is shifted in some samples along the
retention time dimension and can thus not represent the signal from the same
compound.

We evaluate next the sub-grouping in another feature group.

```{r}
fidx <- grep("FG.045.001.", featureGroups(xmse))
eics <- featureChromatograms(
    xmse, features = rownames(featureDefinitions(xmse))[fidx],
    filled = TRUE, n = 1)
```

Next we plot the EICs using a different color for each of the subgroups.

```{r example-2-eic, fig.width = 8, fig.height = 6, fig.cap = "Feature EICs per sample for features from a feature group defined by rentention time and feature abudances across samples. Features with high correlation of their EICs are shown in the same color."}
cols <- c("#ff000080", "#0000ff80")
names(cols) <- unique(featureGroups(xmse)[fidx])

plotChromatogramsOverlay(eics, col = cols[featureGroups(xmse)[fidx]],
                         lwd = 2, peakType = "none")
```

```{r example-2-eic-norm, fig.width = 8, fig.height = 6, fig.cap = "Feature EICs per sample normalized to an absolute intensity of 1 for features from a feature group defined by rentention time and feature abudances across samples. Features with high correlation of their EICs are shown in the same color."}
plotChromatogramsOverlay(normalize(eics),
                         col = cols[featureGroups(xmse)[fidx]],
                         lwd = 2, peakType = "none")
```

Based on the EIC correlation, the initial feature group *FG.045.001* was grouped
into two separate sub-groups.

The grouping based on EIC correlation on the pre-defined feature groups from the
previous sections grouped the `r nrow(featureDefinitions(xmse))` features into
`r length(unique(featureGroups(xmse)))` feature groups.


## Grouping of subsets of features

In the previous sections we were always considering all features from the data
set, but sometimes it could be desirable to just group a pre-defined set of
features, for example features found to be of particular interest in a certain
experiment (e.g. significant features). This can be easily achieved by assigning
the features of interest to a initial feature group, using `NA` as group ID
for all other features.

To illustrate this we *reset* all feature groups by setting them to `NA` and
assign our features of interest (in this example just 30 randomly selected
features) to an initial feature group `"FG"`.

```{r reset-feature-groups}
featureDefinitions(xmse)$feature_group <- NA_character_

set.seed(123)
fts_idx <- sample(1:nrow(featureDefinitions(xmse)), 30)
featureDefinitions(xmse)$feature_group[fts_idx] <- "FG"
```

Any call to `groupFeatures` would now simply sub-group this set of 30
features. Any feature which has an `NA` in the `"feature_group"` column will be
ignored.

```{r rtime-grouping}
xmse <- groupFeatures(xmse, SimilarRtimeParam(diffRt = 20))
xmse <- groupFeatures(xmse, AbundanceSimilarityParam(threshold = 0.7))
table(featureGroups(xmse))
```


# Session information

```{r sessionInfo}
sessionInfo()
```

# References