This document introduces ToxPi and describes how to use the
toxpiR
package to easily import, recombine, analyze, and
visualize high dimensional data. The toxpiR
package is an R
implementation of ToxPi that offers new features over what was
previously available for data handling, recombination, and
customization; provides formally packaged, open-source code for ToxPi;
extends the application domain by supporting rapid analysis of massive
datasets; and bridges with the stand-alone, Graphical User Interface (GUI) Java
application and ArcGIS
Toolkit.
The Toxicological Prioritization Index (ToxPi) is a statistical modeling framework that allows transparent integration and visualization of data across disparate sources (i.e. multimodal or multiscale information). This framework aggregates related features into scored ‘slices’, rescales individual slice scores to range 0–1, and then develops an overall score for each sample using a weighted model. The slice weights can be interpreted as the “importance” of categorized features when predicting sample scores and associated ranks. The resulting scores for each sample are visualized as ToxPi profiles (see image below). Slice weights are represented as the arc-width, slice scores are represented as the radius, and the overall sample scores are represented by the combination of slice weights and radii. These visualizations allow for quick comparison of overall feature importance, feature impacts driving a specific sample, relative impact ranking of common features between samples, and overall ranking between samples. As a flexible method capable of integrating data from any source, it has been applied to problems in several different fields. Additional information is linked from toxpi.org.
There are three main steps in using this package:
Loading data: Users can load comma-separated value (csv) formatted data where data features that will make up the slices are stored as columns and attributes are rows. If you are more famililar with the GUI data format, the package also has a function to load data that is formatted in any one of the GUI layouts where slice information is presented in rows beginning with a hash tag(#). However, new users should use a simple .csv file with the row and column layout.
Creating ToxPi Model: ToxPi model can be created by specifying information about each slice and its corresponding weights. Each slice can be made up of one or more features from the input data. You should also name the slices and provide their weights in this step.
Computing ToxPi Scores: Scores can be computed per attribute as well as for each slice. The S4 object resulting from this step will combine and store the ToxPi model and scores all in one place. This final object will contain all the information needed for plotting.
Data contained within different slots of the S4 objects created using this package can be obtained using accessor functions. More details on how this can be done is present later in the document.
Any new features will be first added to the GitHub version of the
package. However, it may be less stable than the release on CRAN. Before
attempting a direct installation from GitHub, please make sure you have
devtools
package on your system. Note, Windows users will
need Rtools to install directly from GitHub.
This flowchart details the basic steps necessary to analyze data
using toxpiR
. A separate vignette describes how to load and
work with the GUI data file using txpImportGui() function. Usage for all
other functions is described below using an example data set. To view a
list of all functions available in this package, please run:
lsf.str(“package:toxpiR”)
Data can either be entered as a regular csv file without the header lines specifying slice information, or in one of the GUI input formats. If you are using the GUI format, please refer to the Import ToxPi GUI files vignette.
The standard data input is a standard R dataframe with rows as your datapoints and columns as features that may be turned into ToxPi slices. A simple example data set called txp_example_input is available within the package. We will use this 10 x 8 toy data to walk through a simple analysis and explain the functions available in this package. These metrics that we have specified can be in completely different scales and data type, but first we will show a simple example of the package usage.
data(txp_example_input, package = "toxpiR")
head(txp_example_input)
#> name metric1 metric2 metric3 metric4 metric5 metric6 metric7 metric8
#> 1 chem01 74 77 25 74 77 97 25 77
#> 2 chem02 28 20 72 28 20 68 72 20
#> 3 chem03 61 3 73 61 3 24 73 3
#> 4 chem04 NA 40 20 NA 40 22 20 40
#> 5 chem05 29 53 44 29 53 4 44 53
#> 6 chem06 12 43 83 12 43 85 83 43
The first step is to specify information about each slice before creating a Toxpi model. This can be done using the TxpSliceList() and TxpSlice() functions. These slices can be made up of one or more features from the input dataset and are specified by listing the features that go into each slice in the TxpSlice() function. We can also specify how we want to transform the slices with the TxpTransFuncList() function. This means we can do a different transformation per slice instead of having to transform the entire input dataset.
Store individual slice objects.
Arguments
txpValueNames : character vector indicating which data to include in the slice. Can just be the name of a single column if that is all the slice contains.
txpTransFuncs : List of transformation functions to use with one function per entry in txpValueNames. Can be created using TxpTransFuncList() method. Defaults to NULL.
Method to merge all individuals slices together and assign slice names.
Basic Usage
## Goal - Create two slices with transformation functions
# Slice 1: Metric 1, No transformation
# Slice 2: Metric 2 (square) and Metric 3 (no transformation)
slice2.trans <- TxpTransFuncList(func1 = function(x) x^2, func2 = NULL)
f.slices <- TxpSliceList(Slice1 = TxpSlice("metric1"),
Slice2 = TxpSlice(c("metric2", "metric3"),
txpTransFuncs = slice2.trans ))
The ToxPi model object is core for specifying how the created slices should behave when the scores are calculated. In this step, we need to specify what the slices we created above are and assign those slices weights. We can also assign transformation functions at this step. However, it is not necessary if you have already done so when creating the slices. The model is a core component of the analysis because it is going to specify how our calculations are performed.
Arguments
txpSlices : TxpSliceList() object.
txpWeights : numeric vector specifying weight of each slice. By default, all slices are weighted equally (1).
txpTransFuncs : TxpTransFuncList object. Defaults to NULL, if a list is provided, its length must be equal to the number of slices present.
adjusted : Scalar logical to adjust returned weights such that they sum to 1.
Usage
## Goal - Create ToxPi model.
# Slice 1, weight = 2
# Slice 2, weight = 1, apply log transform to final value.
# Object storing list of transformation functions.
final.trans <- TxpTransFuncList(f1 = NULL, f2 = function(x) log10(x))
f.model <- TxpModel(txpSlices = f.slices,
txpWeights = c(2,1),
txpTransFuncs = final.trans)
Calculate ToxPi Scores for the given ToxPi model and input data. This input data is what we read in previously from either a standard .csv file or a GUI compatible one. This step uses the model we have created to specify how the ToxPi scores should be calculated. The results object that is created also stores the model used, allowing for easy reference when you are dealing with many models and results.
Arguments
model : ToxPi model object created using TxpModel() function.
input : dataframe containing input data for ToxPi model.
id.var : Character scalar, column in ‘input’ to store in. Deafults to NULL.
rank.ties.method : character string specifying how ties are treated. Available options “average”, “first”, “last”, “random”, “max” and “min”. Pased to base::rank function.
negative.value.handling : character string specifying how negative values are treated. Available options are “keep” and “missing”. Defaults to keep.
Usage
f.model <- txp_example_model #Load a more complex 4 slice model
f.results <- txpCalculateScores(model = f.model,
input = txp_example_input,
id.var = 'name' )
txpSliceScores(f.results) #ToxPi scores
#> s1 s2 s3 s4
#> [1,] 0.25000000 0.009920635 0.37500000 0.22839506
#> [2,] 0.03893443 0.103174603 0.22026699 0.05246914
#> [3,] 0.07172131 0.105158730 0.17111650 0.00000000
#> [4,] 0.02254098 0.000000000 0.02730583 0.11419753
#> [5,] 0.10860656 0.047619048 0.11468447 0.15432099
#> [6,] 0.05327869 0.125000000 0.28398058 0.12345679
#> [7,] 0.00000000 0.000000000 0.00000000 0.00000000
#> [8,] 0.19467213 0.099206349 0.23118932 0.19444444
#> [9,] 0.21721311 0.075396825 0.22936893 0.25000000
#> [10,] 0.14549180 0.023809524 0.18750000 0.11419753
txpWeights(f.results) #Print weights
#> [1] 2 1 3 2
txpMissing(f.results) #Proportion of missing data within each slice
#> s1 s2 s3 s4
#> 0.100 0.100 0.125 0.100
Two paths exist for visualization, ggplot methods
and grid methods. The ggplot method uses the
ggplot2
package and allows for highly customizable profiles
allowing for several visualization experiences. The grid method uses
advanced grid
graphics and can be implemented using pieGrob
objects that are compatible with the grid system. Additional plots can
be made such as a rank plot by using the data stored in the results
object. The results object also allows for the flexibility to use other
R plotting packages if preferred.
Summary: The ggPlot graphics path will be the
best option for users wanting to create ToxPi profiles, with many
options for customization. The grid graphics path is for users wanting
to modify very specific, basic elements of ToxPi profiles and/or
recombine profiles with other graphics. The ggplot2
package
(high-level) is built atop grid
graphics (low-level), so
the two systems interact well.
grid
package and the “grob” (grid graphical object)
frameworkThis method uses the ggplot2
package to draw highly
customizable ToxPi profiles with multiple different aesthetics and
information options. The most notable aesthetic addition is the
inclusion of information regarding the amount of missing data each slice
has, which is represented by the a small circle of interior slices as is
shown in profiles generated by the ToxPi GUI. Each interior slice is
colored on a grey scale, with black representing all missing data and
white representing no missing data. Each new customization option, along
with some other notable differences, are described below.
Customization Options (Arguments)
fills : Vector containing slice colors. Defaults to a color scheme matching the ToxPi GUI.
showMissing : Boolean to show missing data information. Defaults to TRUE.
showCenter : Boolean to show inner circle. Defaults to TRUE. FALSE overrides showMissing.
showScore : Boolean to show overall profile scores. Defaults to TRUE.
ncol : Integer specifying number of columns in plot. Default is an automatic calculation by ggplot.
bgColor : Text specifying background color for profiles as either hex codes or R recognized colors. Defaults to “gray80”.
borderColor : Text specifying color of max radius ring as either hex codes or R recognized colors. Defaults to “white”.
sliceBorderColor : Text specifying color of slice borders as either hex codes or R recognized colors. Defaults to “white”.
sliceValueColor : Text specifying color of slice scores as either hex codes or R recognized colors. Defaults to NULL.
sliceLineColor : Text specifying color of slice guidelines as either hex codes or R recognized colors. Defaults to NULL.
Usage
Depending on export method, plot resolution may end up poor. SVG and PDF images tend to produce better results than PNG. Usage and example plot output are shown below.
Sometimes the inner circle of the ToxPi figure can show plotting
artifacts. A possible workaround for this is to create a wrapper
function around ggplot2:::coord_munch
as shown below.
## Optional code to make smoother lines
# Plot before updating coord_munch
plot(f.results["chem01"], package = "gg")
# Save the original version of coord_munch
coord_munch_orig <- ggplot2:::coord_munch
# Make a wrapper function that has a different default for segment_length
if (length(formals(coord_munch_orig)) == 5) {
coord_munch_new <- function(coord, data, range, segment_length = 1/1000,
is_closed = FALSE) {
coord_munch_orig(coord, data, range, segment_length, is_closed)
}
} else {
coord_munch_new <- function(coord, data, range, segment_length = 1/1000) {
coord_munch_orig(coord, data, range, segment_length)
}
}
# The environment may need to be set
#environment(coord_munch_new) <- environment(coord_munch_orig)
# Replace ggplot2:::coord_munch with coord_munch_new
assignInNamespace("coord_munch", coord_munch_new, ns = "ggplot2")
# Plot after updating coord_munch
plot(f.results["chem01"], package = "gg")
# Revert to original coord_munch ater plotting if desired
#assignInNamespace("coord_munch", coord_munch_orig, ns = "ggplot2")
Customization examples