Exposome Data Analysis"

knitr::opts_chunk$set(echo = TRUE, warnings = FALSE, tidy = TRUE)


rexposome is an R package designed for the analysis of exposome data. The exposome can be defined as the measure of all the exposures of an individual in a lifetime and how those exposures relate to health. Hence, the aim or rexposome is to offer a set of functions to incorporate exposome data to R framework. Also to provide a series of tools to analyse exposome data using standard methods from Biocondcutor.


rexposome is currently in development and not available from CRAN nor Bioconductor. Anyway, the package can be installed by using devtools R package and taking the source from Bioinformatic Research Group in Epidemiology's GitHub repository.

This can be done by opening an R session and typing the following code:


User must take into account that this sentence do not install the packages dependencies.


The following pictures illustrates the rexposome's pipeline:

Pipeline for exposome analysis

The first step is to load exposome data on R. rexposome provides to functions for this aim: one to load three TXT files and another to use three data.frames. Then the quantification of missing data and values under limit of detection (LOD) is done, and it helps to consider imputation processes. The exposome characterization is useful to understand the nature of the exposome and the relations between exposures. The clustering processes on individual exposure data is done to find exposure-signatures which association with health outcomes can be tested in the next step. From both exposures and exposure-signatures levels, the association with health outcomes is tested using Exposome-Wide Association Studies (ExWAS).

Data Format

Three table format

rexposome defines the exposome data as a three different data-sets:

1) Description Data 2) Exposure Data 3) Phenotype Data

The description data is a file describing the exposome. This means that has a row for each exposure and, at last, defined the families of exposures. Usually, this file incorporates a description of the exposures, the matrix where it was obtained and the units of measurement among others.

The following is an example of a description data file:

exposure  family  matrix         description
bde100    PBDEs   colostrum       BDE 100 - log10
bde138    PBDEs   colostrum       BDE 138 - log10
bde209    PBDEs   colostrum       BDE 209 - log10
PFOA      PFAS    cord blood      PFOA - log10
PFNA      PFAS    cord blood      PFNA - log10
PFOA      PFAS    maternal serum  PFOA - log10
PFNA      PFAS    maternal serum  PFNA - log10
hg        Metals  cord blood      hg - log 10
Co        Metals  urine           Co (creatinine) - log10
Zn        Metals  urine           Zn (creatinine) - log10
Pb        Metals  urine           Pb (creatinine) - log10
THM       Water   ---             Average total THM uptake - log10
CHCL3     Water   ---             Average Chloroform uptake - log10
BROM      Water   ---             Average Brominated THM uptake - log10
NO2       Air     ---             NO2 levels whole pregnancy- log10
Ben       Air     ---             Benzene levels whole pregnancy- log10

The exposures data file is the one containing the measures of each exposures for all the individuals included in the analysis. It is a matrix-like file having a row per individual and a column per exposures. It must includes a column with the subject's identifier.

The following is an example of a exposures data file:

id    bde100  bde138  bde209  PFOA    ...
sub01  2.4665  0.7702  1.6866  2.0075 ...
sub02  0.7799  1.4147  1.2907  1.0153 ...  
sub03 -1.6583 -0.9851 -0.8902 -0.0806 ... 
sub04 -1.0812 -0.6639 -0.2988 -0.4268 ... 
sub05 -0.2842 -0.1518 -1.5291 -0.7365 ... 
...   ...     ...     ...     ...

The last of the data-sets is the phenotype data files. This file contains the covariates to be included in the analysis as well as the health outcomes of interest. It contains a row per individual included in the analysis and a column for each covariate and outcome. Moreover, it must include a column with the individual's identifier.

The following is an example of a phenotype data file:

id    asthma   BMI      sex  age  ...
sub01 control  23.2539  boy  4    ...
sub02 asthma   24.4498  girl 5    ...
sub03 asthma   15.2356  boy  4    ...
sub04 control  25.1387  girl 4    ...
sub05 control  22.0477  boy  5    ...
...   ...      ...      ...  ...

To properly coordinate the exposome data, the information included in the three data-sets must follow some rules:

This rules are easy seen in the following figure:

Links Between Files

In summary: All the exposures, rows, in the description data file are columns in the exposures data file (plus the column for identifying subjects). All the subjects in the exposures data files are, also, in the phenotype data file.

Single table format

To ease the life of researchers that have their datasets as one big table (exposures and phenotypes combined in a single table), we offer the option of using it as a input to rexposome. Please look into the documentation of the loadExposome_plain() function for further details on how to load this type of data. A few remarks on that methodology:

An example of single table is the following:

id    bde100  bde138  bde209    asthma   BMI      ...
sub01  2.4665  0.7702  1.6866   control  23.2539  ...
sub02  0.7799  1.4147  1.2907   asthma   24.4498  ...  
sub03 -1.6583 -0.9851 -0.8902   asthma   15.2356  ... 
sub04 -1.0812 -0.6639 -0.2988   control  25.1387  ... 
sub05 -0.2842 -0.1518 -1.5291   control  22.0477  ...
...   ...     ...      ...      ...      ...

And a visual representation of this single tables is the following:


rexposome R package is loaded using the standard library command:


rexposome provides two functions to load the exposome data: readExposome and loadexposome. The function readExposome will load the exposome data from txt files and loadExposome will do the same from standard R data.frames. Both functions will create an ExposomeSet object. The ExposomeSet is a standard S4 class that will encapsulate the exposome data.

Loading Exposome Data

From TXT files

The function readExposome will create an ExposomeSet from the three txt files. The following lines are used to locate these three files, that were included in the package for demonstration purposes.

path <- file.path(path.package("rexposome"), "extdata")
description <- file.path(path, "description.csv")
phenotype <- file.path(path, "phenotypes.csv")
exposures <- file.path(path, "exposures.csv")

These files follows the rules described in Data Format section. They are csv files, meaning each values is split from the others by a comma (,). Function readExposome allows to load most any type of files containing exposome data:


readExposome expects, by default, csv files. Changing the content of the argument sep will allow to load other files types. The missing values are set using the argument na.strings. This means that the character assigned to this argument will be interpreted as a missing value. By default, those characters are "NA", "-", "?", " " and "". Then, the columns with the exposures' names and the individual's names need to be indicated. Arguments exposures.samCol and phenotype.samCol indicates the column with the individuals' names at exposures file and phenotypes file. The arguments description.expCol and description.famCol indicates the column containing the exposures' names and the exposures' family in the description file.

exp <- readExposome(
    exposures = exposures, 
    description = description, 
    phenotype = phenotype,
    exposures.samCol = "idnum", 
    description.expCol = "Exposure", 
    description.famCol = "Family", 
    phenotype.samCol = "idnum"

The result is an object of class ExposomeSet, that can show all the information of the loaded exposome:


Under the section exposures description the number of continuous (84) and categorical (4) exposures are shown. The assayData, phenoData and featureData shows the content of the files we loaded with readExposome.

From data.frame

The function loadExposome allows to create an ExposomeSet through three data.frames: one as description data, one as exposures data and one as phenotypes data. The arguments are similar to the ones from readExposome:


In order to illustrate how to use loadExposome, we are loading the previous csv files as data.frames:

dd <- read.csv(description, header=TRUE)
ee <- read.csv(exposures, header=TRUE)
pp <- read.csv(phenotype, header=TRUE)

Then we rearrange the data.frames to fulfil with the requirements of the exposome data. The data.frame corresponding to description data needs to have the exposure's names as rownames.

rownames(dd) <- dd[, 2]
dd <- dd[ , -2]

The data.frame corresponding to exposures data needs to have the individual's identifiers as rownames:

rownames(ee) <- ee[ , 1]
ee <- ee[ , -1]

The data.frame corresponding to phenotypes data needs to have the individual's identifiers as a rownames, as the previous data.frame:

rownames(pp) <- pp[ , 1]
pp <- pp[ , -1]

Then, the ExposomeSet is creating by giving the three data.frames to loadExposome:

exp <- loadExposome(
    exposures = ee, 
    description = dd, 
    phenotype = pp,
    description.famCol = "Family"

Accessing to Exposome Data

The class ExposomeSet has several accessors to get the data stored in it. There are four basic methods that returns the names of the individuals (sampleNames), the name of the exposures (exposureNames), the name of the families of exposures (familyNames) and the name of the phenotypes (phenotypeNames).


fData will return the description of the exposures (including internal information to manage them).

head(fData(exp), n = 3)

pData will return the phenotypes information.

head(pData(exp), n = 3)

Finally, the method expos allows to obtain the matrix of exposures as a data.frame:

expos(exp)[1:10, c("Cotinine", "PM10V", "PM25", "X5cxMEPP")]

Exposome Pre-process

Missing Data in Exposures and Phenotypes

The number of missing data on each exposure and on each phenotype can be found by using the function tableMissings. This function returns a vector with the amount of missing data in each exposure or phenotype. The argument set indicates if the number of missing values is counted on exposures of phenotypes. The argument output indicates if it is shown as counts (output="n") or as percentage (output="p").

The current exposome data has no missing in the exposures nor in the phenotypes:

tableMissings(exp, set = "exposures", output = "n")
tableMissings(exp, set = "phenotypes", output = "n")

Alternatively to tableMissings, the function plotMissings draw a bar plot with the percentage of missing data in each exposure of phenotype.

plotMissings(exp, set = "exposures")

Exposures Normality

Most of the test done in exposome analysis requires that the exposures must follow a normal distribution. The function normalityTest performs a test on each exposure for normality behaviour. The result is a data.frame with the exposures' names, a flag TRUE/FALSE for normality and the p-value obtained from the Shapiro-Wilk Normality Test (if the p-value is under the threshold, then the exposure is not normal).

nm <- normalityTest(exp)

So, the exposures that do not follow a normal distribution are:


Some of these exposures are categorical so they must not follow a normal distribution. This is the case, for example, of G_pesticides. If we plot the histogram of the values of the exposures it will make clear:

plotHistogram(exp, select = "G_pesticides") +
    ggtitle("Garden Pesticides")

Some others exposures are continuous variables that do not overpass the normality test. A visual inspection is required in this case.

plotHistogram(exp, select = "BDE209") + 
    ggtitle("BDE209 - Histogram")

If the exposures were following an anon normal distribution, the method plotHistogram has an argument show.trans that set to TRUE draws the histogram of the exposure plus three typical transformations:

plotHistogram(exp, select = "BDE209", show.trans = TRUE)

Exposures Imputation

The imputation process is out of rexposome scope. Nevertheless, rexposome incorporates a wrapper to run the imputation tools from the R packages and Hmisc. The imputation of the exposures in the ExposomeSet is done by using this code:

exp <- imputation(exp)

To use mice package instead of hmisc, see the vignette entitles Dealing with Multiple Imputations.

Exposures Characterization (i.e. data visualization)

We can get a snapshot of the behaviour of the full exposome using the method plotFamily or its wrapper plot. This function allows drawing a plot of a given family of exposures or a mosaic with all the exposures.

plotFamily(exp, family = "all")

This plotting method allows to group the exposure by a given phenotype using the argument group:

plotFamily(exp, family = "Phthalates", group = "sex")

The same method allows to include a second group using the argument group2:

plotFamily(exp, family = "Phthalates", group = "rhinitis", group2 = "rhinitis")

Exposures PCA

To properly perform a PCA analysis the exposures needs to be standardised. The standardisation is done using function standardize that allows using a normal and a robust approaches or use the interquartile range. The normal aproache scales the exposures using the mean as a centre and the standard variation used as dispersion. In robust aproach the median and the median absolute deviation are used. This transformation are only applied to continuous exposures. When interquartile range is used, the median is used as a center and the coeficient between the interquartile range of the exposure and the normal range between the percentile 75 and 25 as variance.

exp_std <- standardize(exp, method = "normal")

Once the exposures are standardized we can run a PCA on the ExposomeSet using the method pca. Typically, exposome datasets contain both numerical and categorical variables, for that reason, a Factor Analysis of Mixed Data (FAMD) is performed by default rather than a PCA (which only uses numerical variables). To perform a regular PCA, provide the argument pca = TRUE to the pca function.

exp_pca <- pca(exp_std)

exp_pca <- pca(exp_std, pca = TRUE)

The method pca returns an object of class ExposomePCA. This object encapsulates all the information generated by the principal component analysis. The method plotPCA can be used in several ways. The first way is setting the argument set to "all" to create a mosaic of plots.

plotPCA(exp_pca, set = "all")

The plots in the first row correspond to the exposures and samples space. The first plot shows all the exposures on the axis for the first and the second principal components. The second plot shows all the individuals on the axis for the first and second principal components.

The plots on the second row are a summary of the variability explained by each component. The first plot is a bar plot with the variability explained by each component highlighting the components that are being drawn in the two first plots. The second plot is a line plot indicating the cumulative variability explained until each principal component. The vertical dashed line indicates the last principal component that is drawn in the first two plots. The horizontal dashed line indicates the amount of explained variability.

A second way of using plotPCA is changing the content of the argument set to "samples" to see the samples' space. When the set argument is filled with samples, the argument phenotype can be used to colour each sample with its phenotype value.

plotPCA(exp_pca, set = "samples", phenotype = "sex")

This plot shows the sample space of the first and the second principal component. Each dot is a sample and it is coloured depending on its value in sex. We can see that no cluster is done in terms of sex.

This view be recreated in a 3D space using the function plot3PCA:

plot3PCA(exp_pca, cmpX=1, cmpY=2, cmpZ=3, phenotype = "sex")

Exposures Correlation

The correlation between exposures, in terms of intra-family and inter-family exposures, is interesting to take into account. The correlation of the exposome can be computed using correlation.

exp_cr <- correlation(exp, use = "pairwise.complete.obs", method.cor = "pearson")

The values of the correlation can be obtained using the method extract. This returns a data.frame.

extract(exp_cr)[1:4, 1:4]

The best option to see the inter-family correlations is the circos of correlations while the matrix of correlations is a better way for studying the intra-family correlations. Both of them are drawn using the method plotCorrelation.

plotCorrelation(exp_cr, type = "circos")
plotCorrelation(exp_cr, type = "matrix")

Individuals Clustering

Clustering analysis on exposures data results in exposure profile. The method clustering allows applying most of any clustering method to an ExposomeSet method.

The argument of the method clustering are:


The argument method is filled with the clustering function. This clustering function needs to accept an argument called data, that will be filled with the exposures-matrix. The object obtained from the clustering function needs to have an accessor called classification. Otherwise the argument cmethod needs to be filled with a function that takes the results of the clustering function and returns a vector with the classification of each individual.

In this analysis we apply the clustering method hclust. Hence we create a function to accept an argument called data.

hclust_data <- function(data, ...) {
    hclust(d = dist(x = data), ...)

The argument ... allows passing arguments from recposome's clustering method to hclust.

Then, a function to obtain the classification of each sample is also required. This function will use the cutree function to obtain the labels.

hclust_k3 <- function(result) {
    cutree(result, k = 3)

The new function hclust_k3 is a function that takes the results of hclust_data and applies it the cutree function, requesting 3 groups of individuals.

Having both clustering function (hclust_data) and the classification function (hclust_k3) we use them in the clustering method:

exp_c <- clustering(exp, method = hclust_data, cmethod = hclust_k3)

The profile for each group of individuals can be plotted with plotClassification method.


The classification of each individual can be obtained using the method classification. We can get a table with the number of samples per group with:


As seen, the groups are given as numbers and the plotClassification transforms it to names (Group 1, Group 2 and Group 3).

Exposure associations

Once preprocessed the exposome its association with health outcomes can be tested through three different approaches:

  1. Using the results of the PCA
  2. Testing the association of each exposure to a given trait (single exposure association)
  3. Testing the association of the exposome to a given trait (multiple exposure associations)

Understanding Principal Component Analysis

From the results of the PCA on the exposome data, two measures can be obtained: the correlation of the exposures with the principal components and the association of the phenotypes with the principal components.

The method plotEXP draws a heat map with the correlation of each exposure to the principal components.

plotEXP(exp_pca) + theme(axis.text.y = element_text(size = 6.5)) + ylab("")

From the plot, some conclusions can be obtained:

These conclusions are useful to give a meaning to the Principal Components in terms of exposures.

The method plotPHE test the association between the phenotypes and the principal components and draws a heat map with the score of the association.


The conclusions that can be taken from the heat map are:

Exposure/Enviroment/Exposome Wide Association Studies (ExWAS)

Method exwas performs univariate test of the association between exposures and health outcomes. This method requests a formula to test and the family of the distribution of the health outcome (dependent variable). The models fitted are:

phenotype ~ exposure_1 + covar1 + ... + covarN
phenotype ~ exposure_2 + covar1 + ... + covarN
phenotype ~ exposure_3 + covar1 + ... + covarN
phenotype ~ exposure_M + covar1 + ... + covarN

The following line performs an ExWAS on flu and wheezing adjusted by sex and age. Since the content of flu and others in the ExposomeSet is dichotomous, the family is set to binomial (for more information see ?glm).

fl_ew <- exwas(exp, formula = blood_pre~sex+age, family = "gaussian")
we_ew <- exwas(exp, formula = wheezing~sex+age, family = "binomial")

The method exwas calculates the effective number of tests in base of the correlation between the exposures. This is transformed into a threshold for the p-values of the association. This threshold can be obtained using the method tef.

A table with the associations between the exposures and flu is obtained with method extract:


A Manhattan-like plot with the p-values of the association between each exposure and asthma, coloured by families of exposures, is draw by method plotExwas.

clr <- rainbow(length(familyNames(exp)))
names(clr) <- familyNames(exp)
plotExwas(fl_ew, we_ew, color = clr) + 
      ggtitle("Exposome Association Study - Univariate Approach")

Then a plot for the effects of a given model can be obtained with plotEffect:


Sensitivity analysis: stratified models

No direct method is implemented to perform a stratified exposome wide analysis, however it may be of interest for some researchers, so a small code sample is provided to perform such studies. On this example a stratified analysis using the sex phenotype as stratifying variable is used, the formula associates the exposures to the phenotype cbmi with no covariates.

strat_variable <- "sex"
formula <- cbmi ~ 1
family <- "gaussian"

strat_ex <- lapply(levels(as.factor(pData(exp)[[strat_variable]])), function(i){
      mask <- pData(exp)[[strat_variable]]==i
      exwas_i <- rexposome::exwas(exp[,mask], formula = formula,
                                  family = family, tef = FALSE)
      exwas_i@formula <- update.formula(exwas_i@formula, 
                                        as.formula(paste0("~ . + strata(", strat_variable, 
                                                          "_", gsub("[[:space:]]|-|+|(|)", "", i), ")")))

We have created a list of ExWAS objects that we can plot together using the following:

do.call(plotExwas, strat_ex)

Inverse ExWAS

Method invExWAS performs a similar association test between exposures and health outcomes. The method asks for a formula to indicate the health outcome and covariables of interest for the association test. The difference to the regular ExWAS is on the model fitted, which on the inverse ExWAS is:

exposure_1 ~ phenotype + covar1 + ... + covarN
exposure_2 ~ phenotype + covar1 + ... + covarN
exposure_3 ~ phenotype + covar1 + ... + covarN
exposure_M ~ phenotype + covar1 + ... + covarN

Since not all exposures on a dataset have to be of the same family, linear models (lm) are fitted for numerical exposures, while multinomial log-linear models (nnet::multinom) are fitted for categorical variables.

The following examples perforns an inverse ExWAS analysis on flu adjusted by sex, note that no left hand side term needs to be supplied, since the left term will be all the exposures on the dataset, if a left han side term is supplied it will be ignored.

inv_ex <- invExWAS(exp, formula = ~ flu + sex)

Since the object returned by invExWAS is of class ExWAS, all the previous manipulation explained can be used on it. Following the example we can extract the table of results:


And we can also use the visualization methods:

clr <- rainbow(length(familyNames(exp)))
names(clr) <- familyNames(exp)
plotExwas(inv_ex, color = clr) + ggtitle("Inverse Exposome Association Study - Univariate Approach")

Variable selection ExWAS

The last approach is a multivariate analysis in order to find the group of exposures related to the health outcome. This can be done using methods like Elastic Net. The method mexwas applies elastic net to the exposures given a health outcome of interest.

bl_mew <- mexwas(exp_std, phenotype = "blood_pre", family = "gaussian")
we_mew <- mexwas(exp_std, phenotype = "wheezing", family = "binomial")

The coefficient of each exposure is plotted with plotExwas. The method draws a heat map with two columns and the exposures as rows. The heat map is coloured with the coefficient of each exposure in relation with the health outcome, so the ones in white are not interesting. The two columns of the heat map correspond to the minimum lambda (Min) and to the lambda which gives the most regularised model such that error is within one standard error of the minimum (1SE).

plotExwas(bl_mew, we_mew) + ylab("") +
        ggtitle("Exposome Association Study - Multivariate Approach")

Session info {.unnumbered}


Try the rexposome package in your browser

Any scripts or data that you put into this service are public.

rexposome documentation built on March 13, 2021, 2:01 a.m.