MOGAMUN is a Multi Objective Genetic Algorithm to find active modules
(i.e., highly connected subnetworks with an overall deregulation) in MUltiplex
biological Networks. For a detailed description of MOGAMUN check out the preprint
https://www.biorxiv.org/content/10.1101/2020.05.25.114215v1. All the expression
datasets and networks that we used to obtain the results reported in our
preprint are available in the GitHub repository
https://github.com/elvanov/MOGAMUN-data.
IMPORTANT. Please note that there was a bug in the non-domination sorting process.
All the runs executed between October 15th, 2020 and February 12th, 2021 must be
re-executed. We apologize for the inconveniences.
## Workflow
The workflow of MOGAMUN is composed of 3 main steps: initialization of
parameters, providing input data and running the algorithm.
### Initialization of parameters
We set here the values for the evolution parameters and other general
parameters, such as the minimum and maximum sizes of the subnetworks, via
`mogamun_init`. Please note that we strongly recommend to run the algorithm
with the default values. The only exception is for `MinSize` and `MaxSize`,
which you can adapt to get bigger or smaller subnetworks, knowing that MOGAMUN
tends to give as result subnetworks of sizes near or equal to `MinSize`.
In total, there are 11 customizable parameters:
* `PopSize` is an integer that determines the size of the population
(*default = 100*).
* `MinSize` and `MaxSize` are integers that determine the minimun and maximum
number of nodes that are allowed in each subnetwork (i.e. individual),
respectively (*default min. = 15, max. = 50*).
* `Generations` is an integer that determines the total number of generations
of the evolution process (*default = 500*).
* `TournamentSize` is an integer that determines the size of the tournament,
i.e. the number of individuals that will participate in the tournament for the
selection of parents (*default = 2*).
* `CrossoverRate` is a real value in the [0-1] range that determines the
crossover rate, where 0 means that crossover will never be performed and 1
means that crossover will always be performed (*default = 0.8*).
* `MutationRate` is a real value in the [0-1] range that determines the
mutation rate, where 0 means that mutation will never be performed and 1 means
that mutation will always be performed (*default = 0.1*).
* `JaccardSimilarityThreshold` is an integer that determines the maximum
Jaccard similarity coefficient to consider two subnetworks as different
(*default = 30*). Subnetworks with a Jaccard similarity coefficient higher than
this value are considered as duplicates. Only one of them (the best one,
according to the Pareto dominance criteria) will remain in the population, and
the rest will be replaced by random subnetworks.
* `MaxNumberOfAttempts` is an integer that determines the maximum number of
attempts allowed to try to find compatible parents for the crossover
(*default = 3*).
* `Measure` is a string that can take any of two values: **FDR** or **PValue**.
It determines which statistical value will be taken into account to obtain the
list of significantly differentially expressed genes (*default = FDR*).
* `ThresholdDEG` is a real value that determines the maximum value of the
`Measure` to consider a gene as significantly differentially expressed
(*default = 0.05*).
If MOGAMUN is to be run with the default values, execute
`EvolutionParameters <- mogamun_init()`. Otherwise, specify the parameters to
change, separated by commas. Please be aware than although we recommend to let
the evolution run for 500 generations, this process can be long. For instance,
using a multiplex networks with the three layers that we provide in
https://github.com/elvanov/MOGAMUN-data and the default values for all the
parameters, the process took approximately 12 hours in a desktop computer with
Intel processor i7 at 3.60GHz and 32GB of RAM.
```{r init, echo = TRUE, message = FALSE}
parameters <- mogamun_init(Generations = 1, PopSize = 10)
```
### Providing input data
MOGAMUN uses two sources of information: one or more biological networks, and
the statistical values resulting from a differential expression analysis or any
other test that gives as result *p*-values or False Discovery Rates (FDR)
associated to genes. The second step of the workflow is to provide the input
data using the `mogamun_load_data` function, which has 5 parameters:
* `EvolutionParameters` is the list returned by `mogamun_init`.
* `DifferentialExpressionPath` is a string with the full path to the
**comma separated** file containing the results of the differential expression
analysis. Please note that such file must be composed of at least two columns:
"gene" and `Measure`. The column "gene" must contain **gene names** and the
column `Measure` must be called either "FDR" or "PValue" (depending on which of
the two was seleced in `mogamun_init`) and it should contain the results of the
statistical test. We strongly recommend to have a third columm ("logFC") with
the $log_2(fold\_change)$, i.e. the ratio of difference of the expression
between the two groups under study (e.g. patients vs control). There is a
sample file in *MOGAMUN/extdata/DE/Sample_DE.csv*.
* `NodesScoresPath` is a string with the full path to the **comma separated**
file containing the nodes scores for every node in the network. Please note
that if this file does not exist, MOGAMUN will create it.
* `NetworkLayersDir` is a string with the path to the directory (i.e. folder)
that contains the networks that will compose the multiplex network. Please note
that each biological network must be in a separate file with 2-column
**tab separated** format. The name of each file must start with a different
character, which will be used in the `Layers` parameter. There are two example
files in *MOGAMUN/extdata/LayersMultiplex*. Please note that these files do not
contain full networks and they are only provided to test the algorithm.
* `Layers` is a string composed of a chain of characters that determine which
files from the `NetworkLayersDir` are to be used to build the multiplex network
(e.g. "123" to use the networks which names start with "1", "2", and "3").
```{r load, echo = TRUE}
dePath <- system.file("extdata/DE/Sample_DE.csv", package = "MOGAMUN")
scoresPath <-
system.file("extdata/DE/Sample_NodesScore.csv", package = "MOGAMUN")
layersPath <-
paste0(system.file("extdata/LayersMultiplex", package = "MOGAMUN"), "/")
loadedData <-
mogamun_load_data(
EvolutionParameters = parameters,
DifferentialExpressionPath = dePath,
NodesScoresPath = scoresPath,
NetworkLayersDir = layersPath,
Layers = "23"
)
```
### Running the algorithm
Once we have defined all the parameters and provided the input data, we are
ready to run MOGAMUN using the `mogamun_run` function, which has 4 parameters:
* `LoadedData` is the list returned by `mogamun_load_data`
* `Cores` is an integer that determines the number of cores that will be used
to run MOGAMUN. This number must be in line with the number of physical
processor cores (*default = 1*). One core is needed per run. Please note that
the execution in parallel is NOT allowed in Windows.
* `NumberOfRunsToExecute` is an integer that determines the number of runs to
be performed (*default = 1*). If you have enough resources, we strongly
recommend to run MOGAMUN 30 times. Please note that if you wish to run MOGAMUN
a higher number of times than cores you have, the pending runs will be executed
sequentially.
* `ResultsDir` is a string with the full path to the directory where the
results must be stored (*default = '.'*, i.e. current working directory).
```{r run, echo = TRUE}
mogamun_run(LoadedData = loadedData, ResultsDir = '.')
```
In the results directory (`ResultsDir`) you will find a subfolder which name
contains the date when you executed the experiment. Inside, there will be two
files per run (**MOGAMUN_Results_StatisticsPerGeneration_RunN.csv** and
**MOGAMUN_Results__Run_N.txt**). The file
**MOGAMUN_Results_StatisticsPerGeneration_RunN.csv** contains the best values
for the two objectives (average nodes score and density) per generation, which
you can use to check the convergence, for instance. The file
**MOGAMUN_Results__Run_N.txt** contains the complete final population of size
`PopSize` (i.e. all the subnetworks from the last generation), one per row.
The number of elements in every row is variable because the size of each
subnetwork can vary between `MinSize` and
`MaxSize`. If **X_n** is the number of elements in the
**n**-th row: the nodes of the subnetwork are the first **X_n**-4 elements.
The last four elements correspond to the average nodes score, density, rank,
and crowding distance, respectively. The best (non-dominated) subnetworks have
are those with rank = 1.
## Postprocessing of the results
In our preprint (https://www.biorxiv.org/content/10.1101/2020.05.25.114215v1),
we ran MOGAMUN 30 times for each experiment. This increases the chances to find
the global maxima. Given that the result of every run is the set of subnetworks
with rank = 1, if you execute MOGAMUN multiple times the final result is the
union of the results of all the individual runs. But considering that in such
set there might be subnetworks that are better than others (according to the
Pareto dominance), to obtain the final result we calculate the accumulated
Pareto front. To this goal, we re-rank the set composed by the union of all the
results, and leave only those subnetworks with rank = 1. In addition, we
propose to merge the subnetworks that are very similar, in order to avoid
having two different networks if they only differ for one node, for instance
(see `JaccardSimilarityThreshold`).
Depending on the number of runs you execute, there are other plots that might
be generated during the postprocessing, such as scatter plots (always) and
boxplots (only if `NumberOfRunsToExecute` in `mogamun_run` > 1).
Finally, it is possible to visualize the subnetworks from the accumulated
Pareto front in Cytoscape. Please note that the network you used to build the
multiplex will be filtered to leave only those interactions among the genes
that are included in the result. The nodes will be colored according to their
*logFC* value, where green stands for downregulated, red means upregulated, and
white means no deregulation. Nodes corresponding to significantly
differentially expressed genes will have black border, and the color of the
edges will be different for each layer of the multiplex network. To check the
layer an edge corresponds to, click on the edge and check the value of the
`TypeOfInteraction`, in the edge table.
The postprocessing is done with the function `mogamun_postprocess`, which has
4 parameters:
* `ExperimentDir` is a string with the full path to the directory where the
results are stored. It must coincide with the `ResultsDir` folder from
`mogamun_run`.
* `LoadedData` is the list returned by `mogamun_load_data`.
* `JaccardSimilarityThreshold` is an integer that determines the minimum
Jaccard similarity coefficient that a pair of subnetworks must have to be
merged in a single subnetwork (default = 70).
* `VisualizeInCytoscape` is a boolean value (TRUE/FALSE) that determines
whether the accumulated Pareto front is to be displayed in Cytoscape or not.
NOTE. When `VisualizeInCytoscape` is TRUE, Cytoscape needs to be up and running
in the computer (default = TRUE).
Please note that you can make use of `mogamun_postprocess` with any number of
runs. If the experiment to be postprocessed contains a single run, the result
will be the set of subnetworks in the first Pareto front. Otherwise, the result
will be the accumulated Pareto front, as we already explained.
```{r end, echo = TRUE}
mogamun_postprocess(
LoadedData = loadedData,
ExperimentDir = '.',
VisualizeInCytoscape = FALSE
)
```
