In i3bionet/CoRegFlux: CoRegFlux

knitr::opts_chunk$set(
    warning = FALSE,
    collapse = TRUE,
    comment = "#"
)

library(CoRegFlux)
library(CoRegNet)
library(sybil)
library(latex2exp)

data("SC_GRN_1")
data("SC_EXP_DATA")
data("SC_experiment_influence")
data("SC_Test_data")
data("aliases_SC")
data("iMM904")
metabolites<-data.frame("names" = c("D-Glucose","Ethanol"),
                        "concentrations" = c(16.6,0))

metabolites_rates<- data.frame("name"=c("D-Glucose"),
                               "concentrations"=c(16.6),
                               "rates"=c(-2.81))

model_uptake_constraints <- adjust_constraints_to_observed_rates(model = iMM904,
                            metabolites_with_rates = metabolites_rates)

Testing_influence_matrix <- CoRegNet::regulatorInfluence(SC_GRN_1,SC_Test_data)
experiment_influence<- Testing_influence_matrix[,1]

PredictedGeneState <- predict_linear_model_influence(network = SC_GRN_1,
                    experiment_influence = experiment_influence,
                    train_expression = SC_EXP_DATA,
                    min_Target = 4,
                    model = iMM904,
                    aliases = aliases_SC)

Simulation1<-Simulation(model = iMM904,
                        time = seq(1,20,by = 1),
                        metabolites = metabolites,
                        initial_biomass = 0.45,
                        aliases = aliases_SC)

This Vignette accompanies the CoRegFlux package. It can be used either to get some additional information about the methods or to get examples of the use of the functions. Feel free to ask any question to the package maintainer (coregflux at gmail dot com).

\tableofcontents

Installation

Install CoRegFlux:

``` {r,warning = FALSE, eval = FALSE} if(!requireNamespace("BiocManager", quietly = TRUE)) install.packages("BiocManager") BiocManager::install("CoRegFlux")

Load CoRegFlux:

```r
library(CoRegFlux)

Introduction

The CoRegFlux package aims at providing tools to integrate reverse engineered gene regulatory networks and gene-expression into metabolic models to improve prediction of phenotypes, both for metabolic engineering, through transcription factor or gene (TF) knock-out or overexpression in various conditions as well as to improve our understanding of the interactions and cell inner-working. The following figure summarize the different functions available in the package to constraint the genome-scale metabolic model with gene expression for wild-type or mutants or to run a complete simulation using those constraints.

Data requirement

To use all CoRegFlux features you will need:

1. A genome-scale metabolic model for your model organism as a modelOrg object (see sybilSBML for import) containing gene-association rules for reactions
2. Condition-specific gene expression data / gene states for which the transcription factor influence, a statistical value estimating the TF activity in each sample, will be calculated
3. A gene regulatory network (GRN) as a coregnet objet (see the CoRegNet package for more information about network inference)
4. A large gene expression matrix to train the model

Examples data are provided with the package, including genome-scale metabolic model iMM904 for Saccharomyces Cerevisiae (SC), a gene regulatory network build for SC from the m3D database as described in Trejo Banos et al., "Integrating transcriptional activity in genome-scale models of metabolism", BMC Systems Biology 2017, as well as a subset of the m3D dataset to train the model and provide gene expression data. To study others organims or another regulatory network, you will need to build a GRN using the CoRegNet package and/or choose relevant datasets.

User guide

Computing Influence using CoRegNet package function

data("SC_GRN_1")
data("SC_EXP_DATA")
data("SC_Test_data")

Testing_influence_matrix <- CoRegNet::regulatorInfluence(SC_GRN_1,SC_Test_data)
experiment_influence<- Testing_influence_matrix[,1]

Here are the main functionalities of CoRegFlux

Predict gene state/gene expression level from a condition specific experiment using a linear model

data("aliases_SC")
data("iMM904")
PredictedGeneState <- predict_linear_model_influence(network = SC_GRN_1,
                    experiment_influence = experiment_influence,
                    train_expression = SC_EXP_DATA,
                    min_Target = 4,
                    model = iMM904,
                    aliases = aliases_SC)

GeneState<-data.frame("Name" = names(PredictedGeneState),
                     "State" = unname(PredictedGeneState))

Simulations

For each simulation step, the function receives a metabolic model and performs:

• update fluxes by metabolites concentrations

• update fluxes by coregnet and influence value

• update fluxes by gene state from the GRN simulator

The simulation result is a list containing:

• objective_history: time series of objective function value for the linear program

• metabolites: metabolites concentrations over time

• fluxes_history: time series of the fluxes values for all the time series

• metabolites_concentration_history: time series of metabolite concentrations

• metabolites_fluxes_history: time series of the metabolites fluxes during the simulation

• rate_history: time series of the growth rate values for all simulation

• time: vector containing the simulation times

• gene_state_history: list containing the values for the gene state during the simulation

The fluxes for the simulation time are stored in a matrix which row names are the fluxes reaction id.

data("aliases_SC")
data("iMM904")
metabolites<-data.frame("names" = c("D-Glucose","Ethanol"),
                        "concentrations" = c(16.6,0))

Simulation1<-Simulation(model = iMM904,
                        time = seq(1,20,by = 1),
                        metabolites = metabolites,
                        initial_biomass = 0.45,
                        aliases = aliases_SC)

Simulation1$fluxes_history[1:10,1:5]

To have access to the gprRules users can use the sybil package, which returns a vector of size equal to the number of fluxes and the associated genes.

library(sybil)
gpr(iMM904)[1:5]

If you only wish to know which gene affects which reaction; the sybil objects have a slot for obtaining the flux-gene matrix.

rxnGeneMat(iMM904)[1:10,1:10]

Simulate a dFBA over time (here 20h) without constraint

metabolites<-data.frame("names" = c("D-Glucose","Ethanol"),
                        "concentrations" = c(16.6,0))

Simulation1<-Simulation(model = iMM904,
                        time = seq(1,20,by = 1),
                        metabolites = metabolites,
                        initial_biomass = 0.45,
                        aliases = aliases_SC)

Simulation1$biomass_history
Simulation1$met_concentration_history

Constraining the model

When provided with different kind of constraints, CoRegFlux process the given information in the following order:

• Gene expression is first integrated
• TF KO or OV is carried out, starting with the first line in the regulator table and going along the rows.
• Gene KO or OV is carried out. starting with the first line in the gene table and going along the rows. Thus order in the regulator table and gene table might play a role and potentially give different results.

The different functions used by CoRegFlux to constraint the model are individually accessible to allow the combination of CoRegFlux's models with other algorithms and parameters, provided by sybil for instance. A model can be constrain iteratively through the different function. In that case, the recommended order is as follows: uptake constraint, gene expression, TF KO or OV, gene KO or OV.

regulator_table <- data.frame("regulator" = c("MET32","CAT8"),
                              "influence" =  c(-1.20322,-2.4),
                              "expression" = c(0,0),
                              stringsAsFactors = FALSE)

model_TF_KO_OV_constraints <- update_fluxes_constraints_influence(model= iMM904,
                                           coregnet = SC_GRN_1,
                                           regulator_table = regulator_table,
                                           aliases = aliases_SC )

sol<-sybil::optimizeProb(model_TF_KO_OV_constraints) 
#Additional parameters from sybil can then be integrated such as the chosen 
# algorithms

sol

Simulate a dFBA with gene expression as a constraint

Simulation2<-Simulation(model = iMM904,
                        time = seq(1,20,by = 1),
                        metabolites = metabolites,
                        initial_biomass = 0.45,
                        aliases = aliases_SC,
                        gene_state_function = function(a,b){GeneState})

Simulation2$biomass_history
Simulation2$met_concentration_history

Simulate a dFBA with TF knock-out (KO) while constraining the model with gene expression

If the simulated mutant have several TFs KO or OV, CoRegFlux will constrain the model according to the order of the TFs in the regulator table. While this example also constraint the model with gene expression, it is possible to run the simulation without such constraints.

regulator_table <- data.frame("regulator" = "MET32",
                              "influence" =  -1.20322,
                              "expression" = 0,
                              stringsAsFactors = FALSE)

SimulationTFKO<-Simulation(model = iMM904,
                        time = seq(1,20,by = 1),
                        metabolites = metabolites,
                        initial_biomass = 0.45,
                        aliases = aliases_SC,
                        coregnet = SC_GRN_1,
                        regulator_table = regulator_table ,
                        gene_state_function = function(a,b){GeneState})

SimulationTFKO$biomass_history ## This KO is predicted as non-lethal 

Simulate a dFBA with TF over-expression (OV) while constraining the model with gene expression

If the simulated mutant have several TFs KO or OV, CoRegFlux will constrain the model according to the order of the TFs in the regulator table. While this example also constraint the model with gene expression, it is possible to run the simulation without such constraints.

regulator_table <- data.frame("regulator" = "MET32",
                                  "influence" = -1.20322 ,
                                  "expression" = 3,
                                  stringsAsFactors = FALSE)

SimulationTFOV<-Simulation(model = iMM904,
                            time = seq(1,20,by = 1),
                            metabolites = metabolites,
                            initial_biomass = 0.45,
                            aliases = aliases_SC,
                            coregnet = SC_GRN_1,
                            regulator_table = regulator_table,
                            gene_state_function = function(a,b){GeneState})

SimulationTFOV$biomass_history ## This OV is predicted as non-lethal

Simulate a dFBA with gene(s) knock-out or over-expression simulation while constraining the model with gene expression

If the simulated mutant have several gene KO or gene OV, CoRegFlux will constrain the model according to the order of the genes in the gene table. While this example also constraint the model with gene expression, it is possible to run the simulation without such constraints.

gene_table <- data.frame("gene" = c("YJL026W","YIL162W"),
                                  "expression" =c(2,0),
                                  stringsAsFactors = FALSE)

SimulationGeneKO_OV<-Simulation(model = iMM904,
                                time = seq(1,20,by = 1),
                                metabolites = metabolites,
                                initial_biomass = 0.45,
                                aliases = aliases_SC,
                                coregnet = SC_GRN_1,
                                gene_table = gene_table,
                                gene_state_function = function(a,b){GeneState})

SimulationGeneKO_OV$biomass_history ## This OV is predicted as non-lethal

Constraining the model according to gene expression, TF KO or OV, gene KO or OV to run various FBA using sybil

The different functions used by CoRegFlux to constraint the model are individually accessible to allow the combination of CoRegFlux's models with other algorithms and parameters, provided by sybil for instance. A model can be constrain iteratively through the different function. In that case, the recommended order is as follows: uptake constraint, gene expression, TF KO or OV, gene KO or OV.

metabolites_rates <- data.frame("name"=c("D-Glucose"),
                               "concentrations"=c(16.6),
                               "rates"=c(-2.81))

model_uptake_constraints <- adjust_constraints_to_observed_rates(model = iMM904, 
                                    metabolites_with_rates = metabolites_rates)

model_gene_constraints <- coregflux_static(model= iMM904,
                                           predicted_gene_expression = 
                                               PredictedGeneState,
                                           aliases = aliases_SC)$model

model_TF_KO_OV_constraints <- update_fluxes_constraints_influence(model= iMM904,
                                           coregnet = SC_GRN_1,
                                           regulator_table = regulator_table,
                                           aliases = aliases_SC )

model_gene_KO_OV_constraints <- update_fluxes_constraints_geneKOOV(
                                            model= iMM904,
                                            gene_table =  gene_table,
                                            aliases = aliases_SC)

sol <- sybil::optimizeProb(model_TF_KO_OV_constraints)   

sol

From observations to fluxes

Here we will compute the fluxes from the observed growth rates (which can be obtained directly from the growth curves)

Assuming we have an observed growth rate of 0.3

fluxes_obs <- 
  get_fba_fluxes_from_observations(iMM904,0.3)
fluxes_obs[1:10,]

Given that the fba solution is not unique, if you wish to see the intervals of maximum and minimum allowed fluxes for a reaction, flux variability analysis should be used

fluxes_intervals_obs <-
  get_fva_intervals_from_observations(iMM904,0.3) 
fluxes_intervals_obs[1:10,]

It worth noting that none of the two methods guarantee that the observed growth rate will be reached.

fluxes_obs[get_biomass_flux_position(iMM904),]
fluxes_intervals_obs[get_biomass_flux_position(iMM904),]

This could mean that the uptake rates for the limiting substrates (most commonly glucose uptake rate) does not allow for higher growth.

To constraint the model using the substrate uptake rate, the user must also provide the metabolites_rates argument

metabolites_rates <- data.frame("name"=c("D-Glucose","Ethanol"),
                               "rates"=c(-10,-1))
fluxes_obs <- 
  get_fba_fluxes_from_observations(
    model = iMM904,
    observed_growth_rate =  0.3,
    metabolites_rates = metabolites_rates) 

fluxes_obs[get_biomass_flux_position(iMM904),]

fluxes_interval_obs <- 
  get_fva_intervals_from_observations(
    model = iMM904,
    observed_growth_rate =0.3,
    metabolites_rates = metabolites_rates) 
fluxes_interval_obs[get_biomass_flux_position(iMM904),]

Adjusting the fluxes bounds based on observed growth rates, and visualized its effects on metabolic genes

During this step, you might get a message from R.cache to choose where the cached files should be saved. Since those files are only temporary files, you can create a dedicated folder in your working directory which you can remove afterward, or pick a location near the installation folder of the R.cache package.

FBA_bounds_from_growthrate<- get_fba_fluxes_from_observations(
    model = iMM904,observed_growth_rate = 0.3,
    metabolites_rates = metabolites_rates)

FVA_bounds_from_growthrate<- get_fva_intervals_from_observations(
    model = iMM904,observed_growth_rate = 0.3,
    metabolites_rates = metabolites_rates)

ODs<-seq.int(0.099,1.8,length.out = 5)
times = seq(0.5,2,by=0.5)

ODcurveToMetCurve<- ODCurveToMetabolicGeneCurves(times = times,
                             ODs = ODs,
                             model = iMM904,
                             aliases = aliases_SC,
                             metabolites_rates = metabolites_rates) 

visMetabolicGeneCurves(ODcurveToMetCurve,genes = "YJR077C")

ODtoflux<-ODCurveToFluxCurves(model = iMM904,
                              ODs = ODs,
                              times = times,
                              metabolites_rates = metabolites_rates)

visFluxCurves(ODtoflux, genes ="ADK3")

Calibration: identifying the softplus parameter using bayesian optimization

To translate the gene expression to fluxes in the GEM, CoRegFlux use the softplus function

library(ggplot2)
library(latex2exp)
eq_title<-latex2exp::TeX('$v_{i}\\leq\\ln\\left(1+ \\exp\\left(\\theta+gpr_{i}\\left(X\\right)\\right)\\right)$')
fun_1 <- function(x)log(1+exp(x))

p <- ggplot2::ggplot(data = data.frame(x = 0), mapping = ggplot2::aes(x = x))
p + ggplot2::stat_function(fun = fun_1,colour="red") + ggplot2::xlim(-5,5)  + 
    ggplot2::geom_vline(xintercept = 0) + ggplot2::geom_hline(yintercept = 0) + ggplot2::ggtitle(eq_title)

where $\theta$ is the softplus parameter applied to all fluxes, $gpr_{i}\left(X\right)$ is the result of evaluating the gene-protein-reaction rules for a set of gene expression levels of the metabolic genes $X$. These rules relate genes to reactions and are logical form. CoRegflux transform these rules as follows

• AND are substitued by MIN()
• OR are substituted by SUM().

Given a known growth rate and predicted gene expressions obtained through the function predict_linear_model_influence, the users have the possibility to adjust the softplus parameter $\theta$ to calibrate the integration of the gene expression in the GEM. This step requires the installation of the package rBayesianOptimization.

library(rBayesianOptimization)
gRates <- 0.1

opF<-function(p){
        CoRegFlux_model<-coregflux_static(model = model_uptake_constraints,
                                          gene_parameter = p,
                                          predicted_gene_expression = 
                                              PredictedGeneState)
        ts<-optimizeProb(CoRegFlux_model$model)
        list(Score=-1*log(abs(lp_obj(ts)-gRates)/gRates),Pred=0)
    }

result<-BayesianOptimization(FUN = opF,
                             bounds = list(p = c(-10,10)),
                             data.frame(p = seq(-10,10,by =  0.5)),
                             n_iter = 10, 
                             verbose = TRUE)

Session Info

sessionInfo()

i3bionet/CoRegFlux documentation built on May 31, 2019, 1:50 a.m.