This vignette describes the basic usage of MOSim package, showing how to perform a simulation, extract and interpret the information from the generated object, touching also some more advanced use cases.
MOSim package simulates multi-omic experiments that
mimic regulatory mechanisms within the cell. Gene expression (RNA-seq
count data) is the central data type and the rest of available omic data
types act as regulators of genes, including DNase-seq (ATAC-seq),
ChIP-seq, miRNA-seq and Methyl-seq. In addition, transcription factor
(TF) regulation can also be modeled.
MOSim algorithm returns the simulated count data
matrices, regulatory connections between genes and omic features and a
detailed description of the simulation settings. Thus, these results can
be used to test new integration methods, to tune or prepare analysis
pipelines, as example data in users’ manuals or for teaching
purposes.
MOSim requires a seed count dataset for each omic to be
simulated. For regulatory omics and TFs, an association table linking
genes to regulators must also be given. For convenience,
MOSim includes the default datasets from STATegra project
(mouse data) for all omics (sampleData). Therefore, the
package can still be used in the absence of custom data.
Due to the potentially great amount of information generated, multiple helper functions are available for both passing the necessary input data and retrieving the generated data.
The outcome of the simulation process depends on two types of input information: the specific parameters of each omic to be simulated and the experimental design.
The experimental design options are flexible. The user can choose the number of experimental groups and the number of replicates. Time series data can also be simulated and, again, the user may decide the number of time points.
The process starts by simulating RNA-seq (gene expression) data. To do that, the program takes a sample from the supplied identifiers (row names of the initial count dataset) and labels them as differentially expressed genes (DEG). The percentage of DEG can be configured by the user, as many other settings that will be described in this vignette. A gene is considered to be differentially expressed if the expression of the gene changes (i) between the reference experimental group (group 1) and at least one of the remaining groups in the experimental design or (ii) across time.
When time course data is to be simulated, MOSim assigns
one of the following profiles to each of the DEGs in each of the
experimental groups:
If a DEG is assigned to a flat profile in all groups, the algorithm will model a change in expression for at least one of the experimental groups (up or down regulation).
For other experimental designs not including time series, all DEGs are labeled as flat and the change in expression is modeled as indicated above, that is, for at least one of the experimental groups with regard to the first group.
Once gene expression settings have been set, regulatory omics are
configured. This is done first by assigning a potential effect
(activation, repression or no-effect) to each regulator. Similarly to
how the percentage of DEG could be specified for RNA-seq, the user can
indicate the percentage of regulators that should be activators,
repressors or with no-effect in each regulatory omic. However, these
initial percentages can be affected, as regulatory relationships must be
coherent with assigned profiles or selection of DEG. For instance, when
a regulator has an effect on more than one DEG and these DEGs have
opposite profiles, the type of effect must be forced to match the
profiles. In fact, and related to this, if a regulator is potentially
associated to more than one DEG with different profiles, the system has
to decide which profile should be assigned to the regulator. For that,
MOSim takes the profile class (combination of profiles for
all groups, for example CI-FL-CR in groups 1, 2 and 3 respectively) with
more DEGs associated to the regulator (majority class) and assigns the
corresponding profile to the regulator: depending on the regulator
effect, the profile will be the same for activation effect (following
the example, CI-FL-CR for groups 1, 2 and 3) or the opposite for
repression (CR-FL-CI). The interactions with genes not included in the
majoritary class are automatically classified as “activator” if the
profiles are the same, “repressor” if the are the totally opposite (CI
vs CR, TI vs TR) or “no-effect” in any other case. Importantly, when
generating user-defined percentages of regulators assigned as
activators, repressors or no-effect, these percentages are subject to
the number of DEG. If the percentage of regulators is higher than what
is possible to generate within the constraint of the simulated DEG,
MOSim generates as many regulator-DEG relationships as
possible, but may generate a lower percentage than requested. In those
cases, inducing a higher percentage of DEG (diffGenes parameter) will
allow for a higher percentage of regulating regulators.
In brief, MOSim usage can be summed up in 3 main steps
that will be described in the next sections:
The MOSim package may be installed as:
if (!require("BiocManager", quietly = TRUE))
install.packages("BiocManager")
BiocManager::install("MOSim")The current development version may be installed as follows:
For the experimental design of the simulation, there are 3 parameters to be set: the number of experimental groups or conditions, the number of time points (1, if time series are not to be considered) and the number of replicates per condition. The only requirements, so that the algorithm can simulate differential expression, are at least 2 groups with no time points, or 1 group with at least 2 time points.
The list of omics to be simulated must be also provided. At this momentMOSim supports the following data types:
The simulation needs gene expression to be present, so RNA-seq will be included always even if it is not specified by the user. The simulation of transcription factors (TF) is also supported as a subset of RNA-seq simulated data.
Optionally, user-defined seed samples to start the simulation can
also be given. They are used for extracting the feature identifiers from
the row names, and as the initial count distribution to generate the
rest of simulated samples. The algorithm includes mouse default samples
from STATegra project (sampleData) but users may provide
seed samples from any other organism or experiment. For each regulatory
omic, the association list linking regulator IDs to genes that they
potentially regulate is also required. If provided, extra care must be
taken to ensure that the identifiers between RNA-seq and the association
lists are correctly matched. Again, for STATegra data
(sampleData), these association files are included in the
package. If TF-target gene associations are provided, TF regulation will
be also simulated.
The structure of the custom data and how to correctly pass it to
MOSim will be described in the following sections.
MOSim simulations are stored in a custom class S4
object, which means that the information is contained in slots and can
be accessed using the standard way (with the operator @). However, this
is not recommended and the preferred way to access the information are
the accessors or utility function provided by the package.
The first of these functions is mosim. This helper
method takes all the options, performs the simulation, and returns the
simulation object.
Internally, this helper function creates a series of S4 objects according to the options passed by the user and calls the required methods to simulate the data, gather the results and return them. The user only needs to set the simulation options, as indicated in the following example:
The main arguments accepted by mosim function are:
- [omics] Character vector containing the names of the omics to simulate, which can be "RNA-seq", "miRNA-seq", "DNase-seq", "ChIP-seq" or "Methyl-seq" (e.g. c("RNA-seq", "miRNA-seq")). It can also be a list with the omic names as names and their options as values, but we recommend to use the argument omicsOptions to provide the options to simulate each omic.
- [omicsOptions] List containing the options to simulate for each omic. We recommend to apply the helper method \Rfunction{omicSim} to create this list in a friendly way, and the function \Rfunction{omicData} to provide custom data (see the related sections for more information). Each omic may have different configuration parameters, but the common ones are:
- [simuData/idToGene] Seed sample and association tables for regulatory omics. The helper function \Rfunction{omicData} should be used to provide this information (see the following section).
- [regulatorEffect] For regulatory omics. List containing the percentage of effect types (repressor, activator or no effect) over the total number of regulators. For example \Robject{list('repressor' = 0.05, 'NE' = 0.95)}. Remember that these numbers might be modified by the algorithm as explained in section ~\@ref(algorithm).
- [totalFeatures] Number of features to simulate. By default, the total number of features in the seed dataset.
- [depth] Sequencing depth in millions of reads. If not provided, it takes the global parameter passed to \Rfunction{mosim} function.
- [replicateParams] List with parameters $a$ and $b$ for adjusting the variability in the generation of replicates using the negative binomial. See section \@ref(advanced) for more information.
- [diffGenes] Number of differentially expressed genes to simulate, given in percentage (0 - 1) or in absolute number (> 1).
- [numberGroups] Number of experimental groups or conditions to simulate.
- [numberReps] Number of replicates per experimetal condition.
- [times] Vector of time points to consider in the experimental design.
- [depth] Sequencing depth in millions of reads.
- [minMaxFC] Vector of minimum and maximum fold-change allowed for differentially expressed features.
- [TFtoGene] A logical value indicating if default transcription factors data should be used (TRUE) or not (NULL), or a 3 column data frame containing custom associations as explained in \@ref(sec:TF). For transcription factors, the count matrix is not simulated like in the other omics but extracted from RNA-seq simulated data. Default value = NULL.The most basic example of mosim function usage is
calling it with only the list of omics to simulate. In this case, the
default values will be used, including the default data samples.
## Warning: replacing previous import 'dplyr::count' by 'matrixStats::count' when
## loading 'MOSim'## Simulation settings of class MOSimulation:
## - Default depth: 74
## - Total genes: 38293
## - Dif. expressed genes: 5744
## - Replicates: 3
## - Factor levels (groups): 2
## - Time vector length: 5## Generating simulation settings for RNA-seq.## Creating settings to change count values on 21 DE genes with the same flat profile on all groups.## Finishing generation of configuration settings.## Configuration generated.## Starting simulation of RNA-seq.## - Simulating count values for group 1.## - Making replicates for group 1 on time 0.## - Making replicates for group 1 on time 2.## - Making replicates for group 1 on time 4.## - Making replicates for group 1 on time 12.## - Making replicates for group 1 on time 24.## - Simulating count values for group 2.## - Making replicates for group 2 on time 0.## - Making replicates for group 2 on time 2.## - Making replicates for group 2 on time 4.## - Making replicates for group 2 on time 12.## - Making replicates for group 2 on time 24.## Rounding RNA-seq count values.The rnaseq_simulation object is a
Simulation class object containing the simulated data, that
can be easily accessed with the helper functions
omicResults and omicSettings, as we will see
in the corresponding section.
Following with that basic example above, we can modify the experimental design to simulate 2 groups, 1 time point and 4 replicates per group:
## Simulation settings of class MOSimulation:
## - Default depth: 74
## - Total genes: 38293
## - Dif. expressed genes: 5744
## - Replicates: 4
## - Factor levels (groups): 2
## - Time vector length: 1## Generating simulation settings for RNA-seq.## Creating settings to change count values on 5744 DE genes with the same flat profile on all groups.## Finishing generation of configuration settings.## Configuration generated.## Starting simulation of RNA-seq.## - Simulating count values for group 1.## - Making replicates for group 1 on time 0.## - Simulating count values for group 2.## - Making replicates for group 2 on time 0.## Rounding RNA-seq count values.Invalid combinations of experimental design arguments will make the algorithm stop with an error message, like this:
To obtain more than one omic data type, the omics list must be modified. RNA-seq is mandatory, and will be automatically included in the simulation, no matter if it is listed or not. Therefore these two simulations would be equivalent:
multi_simulation <- mosim(omics = c("RNA-seq", "DNase-seq"),
times = c(0, 5, 10),
numberGroups = 2,
numberReps = 4,
diffGenes = .3)
dnase_simulation <- mosim(omics = c("DNase-seq"),
times = c(0, 5, 10),
numberGroups = 2,
numberReps = 4,
diffGenes = .3)As it can be seen, mosim function accepts global
simulation parameters, but specific settings for a particular omic
should be provided through two specially designed functions:
omicData and omicSim.
The STATegra dataset Gomez-Cabrero
et al. 2019 is provided in the MOSim package
under sampleData. The dataset contains sample count
matrices for RNA-seq, DNase-seq, miRNA-seq, ChIP-seq and Methyl-seq, as
well as regulator association lists.
This provided dataset may be called by the user:
data("sampleData")
omicDataThe function omicData was designed to help users to
provide their own seed data sets, as follows:
omicData(omic, data, associationList = NULL)
This helper function accepts 3 parameters:
For illustration purposes, consider as our custom data a subset of the default gene expression dataset. To use it as our seed RNA-seq dataset, we can use this code:
# Take a subset of the included dataset for illustration
# purposes. We could also load it from a csv file or RData,
# as long as we transform it to have 1 column named "Counts"
# and the identifiers as row names.
data("sampleData")
custom_rnaseq <- head(sampleData$SimRNAseq$data, 100)
# In this case, 'custom_rnaseq' is a data frame with
# the structure:
head(custom_rnaseq)
## Counts
## ENSMUSG00000000001 6572
## ENSMUSG00000000003 0
## ENSMUSG00000000028 4644
## ENSMUSG00000000031 8
## ENSMUSG00000000037 0
## ENSMUSG00000000049 0
# The helper 'omicData' returns an object with our custom data.
rnaseq_customdata <- omicData("RNA-seq", data = custom_rnaseq)
# We use the associative list of 'omics' parameter to pass
# the RNA-seq object.
rnaseq_simulation <- mosim(omics = list("RNA-seq" = rnaseq_customdata))RNA-seq is a special case of omic data type because it does not require an association list to work. For any other omic, such as DNase-seq, we need two different data frames: the seed sample with the structure already mentioned, and the associations between regulator IDs and genes.
# Select a subset of the available data as a custom dataset
data("sampleData")
custom_dnaseseq <- head(sampleData$SimDNaseseq$data, 100)
# Retrieve a subset of the default association list.
dnase_genes <- sampleData$SimDNaseseq$idToGene
dnase_genes <- dnase_genes[dnase_genes$ID %in%
rownames(custom_dnaseseq), ]
# In this case, 'custom_dnaseseq' is a data frame with
# the structure:
head(custom_dnaseseq)
## Counts
## 1_63176480_63177113 513
## 1_125435495_125436168 1058
## 1_128319376_128319506 37
## 1_139067124_139067654 235
## 1_152305595_152305752 105
## 1_172490322_172490824 290
# The association list 'dnase_genes' is another data frame
# with the structure:
head(dnase_genes)
## ID Gene
## 29195 1_3670777_3670902 ENSMUSG00000051951
## 29196 1_3873195_3873351 ENSMUSG00000089420
## 29197 1_4332428_4332928 ENSMUSG00000025900
## 29198 1_4346315_4346445 ENSMUSG00000025900
## 29199 1_4416827_4416973 ENSMUSG00000025900
## 29200 1_4516660_4516798 ENSMUSG00000096126
dnaseseq_customdata <- omicData("DNase-seq",
data = custom_dnaseseq,
associationList = dnase_genes)
multi_simulation <- mosim(omics = list(
"RNA-seq" = rnaseq_customdata,
"DNase-seq" = dnaseseq_customdata)
)The two exceptions in this section are transcription factors and methylation.
The simulated transcription factor data is extracted from the
generated RNA-seq data but a data frame with 3 columns needs to be
provided: TF column, with the transcription factor
identifiers (any type of identifier can be used); TFgene
column, with transcription factor identifiers that must coincide with
the type of identifier used in RNA-seq; and LinkedGene
column, with the identifier of the target gene (again, the same used in
RNA-seq data). Instead of applying omicData function for
this purpose, the TFtoGene argument in mosim
function must be used:
# Select a subset of the available data as a custom dataset
data("sampleData")
custom_tf <- head(sampleData$SimRNAseq$TFtoGene, 100)
# TF TFgene LinkedGene
# 1 Aebp2 ENSMUSG00000030232 ENSMUSG00000000711
# 2 Aebp2 ENSMUSG00000030232 ENSMUSG00000001157
# 3 Aebp2 ENSMUSG00000030232 ENSMUSG00000001211
# 4 Aebp2 ENSMUSG00000030232 ENSMUSG00000001227
# 5 Aebp2 ENSMUSG00000030232 ENSMUSG00000001305
# 6 Aebp2 ENSMUSG00000030232 ENSMUSG00000001794
multi_simulation <- mosim(omics = list(
"RNA-seq" = rnaseq_customdata,
"DNase-seq" = dnaseseq_customdata),
# The option is passed directly to mosim function instead of
# being an element inside "omics" parameter.
TFtoGene = custom_tf
)For methylation, a seed count sample does not need to be provided
because it will be generated automatically. Methylation simulation only
needs the association list containing the CpG sites to be simulated and
the associated genes. The chromosomal positions for the CpG sites must
be given in the format
<chr>_<start>_<end>, that is, the
chromosome number, start and end positions separated by the char
_.
# Select a subset of the available data as a custom dataset
data("sampleData")
custom_cpgs <- head(sampleData$SimMethylseq$idToGene, 100)
# The ID column will be splitted using the "_" char
# assuming <chr>_<start>_<end>.
#
# These positions will be considered as CpG sites
# and used to generate CpG islands and other elements.
#
# Please refer to MOSim paper for more information.
#
# ID Gene
# 1 11_3101154_3101154 ENSMUSG00000082286
# 2 11_3101170_3101170 ENSMUSG00000082286
# 3 11_3101229_3101229 ENSMUSG00000082286
# 4 11_3101287_3101287 ENSMUSG00000082286
# 5 11_3101329_3101329 ENSMUSG00000082286
# 6 11_3101404_3101404 ENSMUSG00000082286omicSimAs explained before when describing mosim function,
there are two ways of passing omic configuration options: by giving a
list in the omics parameter or by giving omics
as a character vector including the omics to simulate and also
specifying the simulation options in the parameter
omicsOptions.
The omicsOptions parameter accepts a list, but the
MOSim helper function, omicSim, allows to do
it in a more straightforward way.
Back to the first basic example of RNA-seq simulation using the default dataset, the code to use if we wish to restrict the number of features is:
omic_list <- c("RNA-seq")
rnaseq_options <- omicSim("RNA-seq", totalFeatures = 2500)
# The return value is an associative list compatible with
# 'omicsOptions'
rnaseq_simulation <- mosim(omics = omic_list,
omicsOptions = rnaseq_options)When having multiple omics, we concatenate the information like follows:
omics_list <- c("RNA-seq", "DNase-seq")
# In R concatenaning two lists creates another one merging
# its elements, we use that for 'omicsOptions' parameter.
omics_options <- c(omicSim("RNA-seq", totalFeatures = 2500, depth = 74),
omicSim("DNase-seq",
# Limit the number of features to simulate
totalFeatures = 1500,
# Modify the percentage of regulators with effects.
regulatorEffect = list(
'activator' = 0.68,
'repressor' = 0.3,
'NE' = 0.02
)))
set.seed(12345)
multi_simulation <- mosim(omics = omics_list,
omicsOptions = omics_options)The objects generated by omicData and
omicSim are different and special attention must be paid to
combine them. The following code shows an example on how to do it:
The information contained in a simulation object can be classified in two categories: the simulation settings used to perform the simulation, and the count data matrices generated by the process.
To access this information, MOSim provides two helper
functions: omicSettings to retrieve the simulation settings
and omicResults for accessing the count matrices.
simulation objectAfter running mosim, an S4 object is generated
containing all the information from the simulation. This object may be
parsed using omicResults and omicSettings
functions. The simulation object contains the information
about the global simulation process, including association
of genes with gene classes, regulatory programs and other settings.
For user understanding, the simulation object has the
following structure:
omicSettingsThe helper function omicSettings is used to extract the
settings used in the simulation:
omicSettings(simulation, omics = NULL, association = FALSE, reverse = FALSE, only.linked = FALSE, include.lagged = TRUE)
Users can choose to recover the setting for all the simulated omics or just for some of them:
# This will be a data frame with RNA-seq settings (DE flag, profiles)
rnaseq_settings <- omicSettings(multi_simulation, "RNA-seq")
# This will be a list containing all the simulated omics (RNA-seq
# and DNase-seq in this case)
all_settings <- omicSettings(multi_simulation)For RNA-seq, the settings table has an structure similar to this:
% latex table generated in R 3.4.4 by xtable 1.8-2 package % Tue Jun 19 14:46:02 2018Each column provides different information about the settings used to carry on the simulation:
For regulatory omics, the structure will slightly differ from
RNA-seq, adding additional columns. Note that in this example, for
reading purposes, the last 4 columns containing the
Tmax.GroupX and Lagged.GroupX have been
omitted:
Each row describes a regulator-gene interaction, with the following columns:
To retrieve the original association lists, the parameter
association must be set to TRUE:
# This will be a list with 3 keys: settings, association and regulators
dnase_settings <- omicSettings(multi_simulation, "DNase-seq",
association = TRUE)When setting the association parameter to
TRUE, the output object will be a list of lists with the
following key names:
omicResultsThe last helper function is omicResults:
omicResults(simulation, omics = NULL)
This function can accept 2 parameters: the simulation output object and, optionally, the omics we want to retrieve. As in the previous helper function, retrieving one omic will result in a data frame object, and for more than one a list of data frames will be provided.
# multi_simulation is an object returned by mosim function.
# This will be a data frame with RNA-seq counts
rnaseq_simulated <- omicResults(multi_simulation, "RNA-seq")
# Group1.Time0.Rep1 Group1.Time0.Rep2 Group1.Time0.Rep3 ...
# ENSMUSG00000073155 4539 5374 5808 ...
# ENSMUSG00000026251 0 0 0 ...
# ENSMUSG00000040472 2742 2714 2912 ...
# ENSMUSG00000021598 5256 4640 5130 ...
# ENSMUSG00000032348 421 348 492 ...
# ENSMUSG00000097226 16 14 9 ...
# ENSMUSG00000027857 0 0 0 ...
# ENSMUSG00000032081 1 0 0 ...
# ENSMUSG00000097164 794 822 965 ...
# ENSMUSG00000097871 0 0 0 ...
# This will be a list containing RNA-seq and DNase-seq counts
all_simulated <- omicResults(multi_simulation)The structure of the final count matrix will have the features as row
names, and the conditions as column names following the scheme
<Group>.<Timepoint>.<Replicate>.
Alternatively a ExpressionSet object can be returned by
setting the format argument to
ExpressionSet.
experimentalDesignDownstream analysis of the simulated datasets will require an
experimental design matrix depicting the simulated comparisons between
groups and throughout time. To generate an experiental design matrix,
function experimentalDesign needs a simulaton object.
plotProfileGraphical plots are useful to check a feature profile or to compare
gene & regulator interactions. To generate them, the function
plotProfile needs the simulation object, the omic or two
omics to plot, and one feature for each omic.
# The methods returns a ggplot plot, if called directly
# it will be automatically plotted.
plotProfile(multi_simulation,
omics = c("RNA-seq", "DNase-seq"),
featureIDS = list(
"RNA-seq" = "ENSMUSG00000024691",
"DNase-seq" = "19_12574278_12574408"
))The returned ggplot can be stored in a variable to customize other attributes.
library(ggplot2)
# Store the plot in a variable
profile_plot <- plotProfile(multi_simulation,
omics = c("RNA-seq", "DNase-seq"),
featureIDS = list(
"RNA-seq" = "ENSMUSG00000024691",
"DNase-seq" = "19_12574278_12574408"
))
# Modify the title and print
profile_plot +
ggtitle("My custom title") +
theme_bw() +
theme(legend.position="top")Most of the common users’ needs should be covered by the previously showed examples, but this section describes some advanced settings as the variability of the negative binomial distribution that is used to generate replicates.
In a negative binomial distribution, the variance depends on the
mean. In MOSim, the generated counts for a given condition
(and/or time point) are taken as the mean of the distribution. To model
the dependence between the negative binomial mean and variance for each
omic, we analyzed several data sets and experiments, and observed a
linear relationship between log-transformed count values of means and
variances (R^2 > 0.95 for all models): sigma^2 = 10^a * (mu + 1)^b -
1. To assure a minimum variance we really used the maximum of this value
and 0.03. A regression model was applied to estimate coefficients a and
b and these estimations were used as default values. The default values
for a and b can be changed by users to
increase or decrease the default variability in each omic data type.
In some special cases, it may be necessary to transform a regulator matrix into 0/1 format (e.g. to simulate datasets counting presence/absence of an event). We can generate a dataframe of 1 and 0 by running:
Transcription factors may be simulated using MOSim, but
are a special case, as they represent a subset of the simulated RNA-seq
data. They are the result of including parameter
TFtoGene = TRUE in the simulation using mosim
function. It is important to remember Transcription
Factors are not simulated when included in the input list to
(omic) parameter.
However, similar to other omic simulation results, the simulated
dataframe for TF is extracted by default (if
TFtoGene = TRUE) by omicResults, and the
association list and regulator effect by omicSettings.
MOSimThe MOSim package is currently on biorxiv:
Monzó C., Martínez-Mira C., Arzalluz-Luque A., Conesa A., Tarazona S. (2024) MOSim: bulk and single-cell multi-layer regulatory network simulator DOI: 10.1101/421834