CNORfeeder

CNORfeeder is an add-on to CellNOptR that permits to extend a network derived from literature with links derived in a strictly data-driven way and supported by protein-protein interactions as described in:

F. Eduati, J. De Las Rivas, B. Di Camillo, G. Toffolo, J. Saez-Rodriguez. Integrating literature-constrained and data-driven inference of signalling networks. Bioinformatics, 2012 , 28 (18)

E. Gjerga, P. Trairatphisan, A. Gabor, H. Koch, C. Chevalier, F. Ceccarelli, A. Dugourd, A. Mitsos, J. Saez0Rodriguez. Converting networks to predictive logic models from perturbation signalling data with CellNOpt. biRxiv, 2020 , 28 (18)

Requirements

All network modelling steps were performed in R v3.5.1 and visualised using Cytoscape v3.4.

Other prerequisites include downloading and installing the following R package dependencies:

Bioconductor

• CellNOptR
• MEIGOR
• CNORode

CRAN

• doParallel
• readr
• infotheo
• igraph

Installation

CellNOpt-Feeder is currently available for the installation as an R-package from our GitHub page

# Install CARNIVAL from Github using devtools
# install.packages('devtools') # in case devtools hasn't been installed
library(devtools)
install_github('saezlab/CNORfeeder', build_vignettes = TRUE)
# or download the source file from GitHub and install from source
install.packages('path_to_extracted_CNORfeeder-main_directory', repos = NULL, type="source")

Toy Example

Here we show how Dynamic-Feeder can be applied on a single execution script over a simple toy example (ToyMMB).

Sourcing all the necessary packages

library(CellNOptR)
library(MEIGOR)
library(CNORode)
library(doParallel)
library(readr)
library(infotheo)
library(igraph)
library(CNORfeeder)

Loading the toy example

# loading the model
data(ToyModel_Gene, package="CNORfeeder")
# loading the data
data(CNOlistToy_Gene, package="CNORfeeder")
# plotting the model and the data
plotModel(model = model, CNOlist = cnolist)
plotCNOlist(CNOlist = cnolist)

## Loading database
data(database, package="CNORfeeder")

Setting the initial parameters (here parameters ‘k’ and ‘tau’ are optimised and ‘n’ fixed to 3) and optimization settings for the initial training of the model.

ode_parameters=createLBodeContPars(model, LB_n = 1, LB_k = 0,
                                   LB_tau = 0, UB_n = 3, UB_k = 1,
                                   UB_tau = 1, default_n = 3,
                                   default_k = 0.5, default_tau = 0.01, 
                                   opt_n = FALSE, opt_k = TRUE,
                                   opt_tau = TRUE, random = TRUE)

## Parameter Optimization
# essm
paramsSSm=defaultParametersSSm()
paramsSSm$local_solver = "DHC"
paramsSSm$maxtime = 60;
paramsSSm$maxeval = Inf;
paramsSSm$atol=1e-6;
paramsSSm$reltol=1e-6;
paramsSSm$nan_fac=1000;
paramsSSm$dim_refset=30;
paramsSSm$n_diverse=1000;
paramsSSm$maxStepSize=Inf;
paramsSSm$maxNumSteps=10000;
paramsSSm$transfer_function = 4;

paramsSSm$lambda_tau=0.1
paramsSSm$lambda_k=0.01
paramsSSm$bootstrap=F
paramsSSm$SSpenalty_fac=0
paramsSSm$SScontrolPenalty_fac=0

Initial training of the model

opt_pars=parEstimationLBode(cnolist, model, method="essm", 
                            ode_parameters=ode_parameters, paramsSSm=paramsSSm)
simData = plotLBodeFitness(cnolist = cnolist, model = model,
                            ode_parameters = opt_pars, transfer_function = 4)

Identifying the mis-fits (measurements with mse worse than 0.05) and interactions from the database which we want to integrate (on this case only through data-driven method)

## Identifying which measured nodes at which experimental condition have a worse
## fit in terms of mse value compared to the specified threshold mseThresh value
indices = identifyMisfitIndices(cnolist = cnolist, model = model, 
                                    simData = simData, mseThresh = 0.05)
                                    
## Alternatively, users can bypass the initial training of the model and add
## new links through the FEED or database search regardless how well some 
## measurements were already fitted simply by setting simData=NULL
# indices = identifyMisfitIndices(cnolist = cnolist, model = model, simData = NULL)

## Obtaining the object indicating which measurements at which condition are
## poorly fitted
feederObject = buildFeederObjectDynamic(model = model, cnolist = cnolist, 
                                indices = indices, database = NULL, DDN = TRUE)

## Integration of new links
integratedModel = integrateLinks(feederObject = feederObject, 
                                                  cnolist = cnolist)

Plotting the integrated model by highlighting in purple the new added links to the PKN

plotModel(model = integratedModel$model, CNOlist = cnolist, 
                                    indexIntegr = integratedModel$integLinksIdx)

Setting the initial ODE parameters to optimize for the integrated model

ode_parameters=createLBodeContPars(integratedModel$model, LB_n = 1, LB_k = 0,
                                   LB_tau = 0, UB_n = 3, UB_k = 1, UB_tau = 1, 
                                   default_n = 3, default_k = 0.5, 
                                   default_tau = 0.01, opt_n = FALSE, 
                                   opt_k = TRUE, opt_tau = TRUE, random = TRUE)

Optimizing the integrated model with low penalty factor (lambda = 2) of the integrated links - here we observe the effects of the new links on the improvements in the fitting quality

res1 = runDynamicFeeder(cnolist = cnolist, integratedModel = integratedModel, 
        ode_parameters = ode_parameters, paramsSSm = paramsSSm, penFactor_k = 2)

plotLBodeFitness(cnolist = res1$CNOList, model = res1
Integrated-Model`$model, 
                        ode_parameters = res1$Parameters, transfer_function = 4)

Optimizing the integrated model with high penalty factor (lambda_k = 10000) of the integrated links - here we do not observe any effects of the new links on the improvements in the fitting quality

res2 = runDynamicFeeder(cnolist = cnolist, integratedModel = integratedModel, ode_parameters = ode_parameters, penFactor_k = 10000, paramsSSm = paramsSSm)

plotLBodeFitness(cnolist = res2$CNOList, model = res2
Integrated-Model`$model, ode_parameters = res2$Parameters, transfer_function = 4)