BSL (version 3.0.0)

cell: Cell biology example

Description

This example estimates the probabilities of cell motility and cell proliferation for a discrete-time stochastic model of cell spreading. We provide the data and tuning parameters required to reproduce the results in An et al. (2019).

Usage

cell_sim(theta, Yinit, rows, cols, sim_iters, num_obs)
cell_sum(Y, Yinit)
cell_prior(theta)
cell_logPrior(theta)

Arguments

theta

A vector of proposed model parameters, \(Pm\) and \(Pp\).

Yinit

The initial matrix of cell presences of size rows by cols.

rows

The number of rows in the lattice (rows in the cell location matrix).

cols

The number of columns in the lattice (columns in the cell location matrix).

sim_iters

The number of discretisation steps to get to when an observation is actually taken. For example, if observations are taken every 5 minutes but the discretisation level is 2.5 minutes, then sim_iters would be 2. Larger values of sim_iters lead to more ``accurate'' simulations from the model, but they also increase the simulation time.

num_obs

The total number of images taken after initialisation.

Y

A rows by cols by num_obs array of the cell presences at times 1:num_obs (not time 0).

A simulated dataset

An example ``observed'' dataset and the tuning parameters relevant to that example can be obtained using data(cell). This ``observed'' data is a simulated dataset with \(Pm = 0.35\) and \(Pp = 0.001\). The lattice has 27 rows and 36 cols and there are num_obs = 144 observations after time 0 (to mimic images being taken every 5 minutes for 12 hours). The simulation is based on there initially being 110 cells in the assay.

Further information about the specific choices of tuning parameters used in BSL and BSLasso can be found in An et al. (2019).

• data: The rows \(x\) cols \(x\) num_obs array of the cell presences at times 1:144.

• sim_options: Values of sim_options relevant to this example.

• sum_options: Values of sum_options relevant to this example, i.e. just the value of Yinit.

• start: A vector of suitable initial values of the parameters for MCMC.

• cov: The covariance matrix of a multivariate normal random walk proposal distribution used in the MCMC, in the form of a 2 by 2 matrix

Details

Cell motility (movement) and proliferation (reproduction) cause tumors to spread and wounds to heal. If we can measure cell proliferation and cell motility under different situations, then we may be able to use this information to determine the efficacy of different medical treatments.

A common method for measuring in vitro cell movement and proliferation is the scratch assay. Cells form a layer on an assay and, once they are completely covering the assay, a scratch is made to separate the cells. Images of the cells are taken until the scratch has closed up and the cells are in contact again. Each image can be converted to a binary matrix by forming a lattice and recording the binary matrix (of size rows \(x\) cols) of cell presences.

The model that we consider is a random walk model with parameters for the probability of cell movement (\(Pm\)) and the probability of cell proliferation (Pp) and it has no tractable likelihood function. We use the uninformative priors \(Pm ~ U(0,1)\) and \(Pp ~ U(0,1)\).

We have a total of 145 summary statistics, which are made up of the Hamming distances between the binary matrices for each time point and the total number of cells at the final time.

Details about the types of cells that this model is suitable for and other information can be found in Price et al. (2018) and An et al. (2019). Johnston et al. (2014) use a different ABC method and different summary statistics for a similar example.

References

An, Z., South, L. F., Nott, D. J. & Drovandi, C. C. (2019). Accelerating Bayesian synthetic likelihood with the graphical lasso. Journal of Computational and Graphical Statistics. https://doi.org/10.1080/10618600.2018.1537928

Johnston, S., Simpson, M. J., McElwain, D. L. S., Binder, B. J. & Ross, J. V. (2014). Interpreting Scratch Assays Using Pair Density Dynamic and Approximate Bayesian Computation. Open Biology, 4, 1-11.

Price, L. F., Drovandi, C. C., Lee, A., & Nott, D. J. (2018). Bayesian synthetic likelihood. Journal of Computational and Graphical Statistics. https://doi.org/10.1080/10618600.2017.1302882

Examples

# NOT RUN {
require(doParallel) # You can use a different package to set up the parallel backend

# Loading the data for this example
data(cell)
model <- BSLModel(fnSim = cell_sim, fnSum = cell_sum, simArgs = cell$sim_options, 
                  sumArgs = cell$sum_options, theta0 = cell$start, fnLogPrior = cell_logPrior,
                  thetaNames = expression(P[m], P[p]))
true_cell <- c(0.35, 0.001)

# Performing BSL (reduce the number of iterations M if desired)
# Opening up the parallel pools using doParallel
cl <- makeCluster(detectCores() - 1)
registerDoParallel(cl)
resultCellBSL <- bsl(cell$data, n = 5000, M = 10000, model = model, covRandWalk = cell$cov,
                     parallel = TRUE, verbose = TRUE)
stopCluster(cl)
registerDoSEQ()
show(resultCellBSL)
summary(resultCellBSL)
plot(resultCellBSL, thetaTrue = true_cell, thin = 20)

# Performing uBSL (reduce the number of iterations M if desired)
# Opening up the parallel pools using doParallel
cl <- makeCluster(detectCores() - 1)
registerDoParallel(cl)
resultCelluBSL <- bsl(cell$data, n = 5000, M = 10000, model = model, covRandWalk = cell$cov,
                      method = 'uBSL', parallel = TRUE, verbose = TRUE)
stopCluster(cl)
registerDoSEQ()
show(resultCelluBSL)
summary(resultCelluBSL)
plot(resultCelluBSL, thetaTrue = true_cell, thin = 20)

# Performing tuning for BSLasso
ssy <- cell_sum(cell$data, cell$sum_options$Yinit)
lambda_all <- list(exp(seq(0.5,2.5,length.out=20)), exp(seq(0,2,length.out=20)),
                   exp(seq(-1,1,length.out=20)), exp(seq(-1,1,length.out=20)))
# Opening up the parallel pools using doParallel
cl <- makeCluster(detectCores() - 1)
registerDoParallel(cl)
set.seed(100)
sp_cell <- selectPenalty(ssy, n = c(500, 1000, 1500, 2000), lambda_all, theta = true_cell,
                         M = 100, sigma = 1.5, model = model, method = 'BSL', shrinkage = 'glasso', 
                         parallelSim = TRUE, parallelMain = FALSE)
stopCluster(cl)
registerDoSEQ()
sp_cell
plot(sp_cell)

# Performing BSLasso with a fixed penalty (reduce the number of iterations M if desired)
# Opening up the parallel pools using doParallel
cl <- makeCluster(detectCores() - 1)
registerDoParallel(cl)
resultCellBSLasso <- bsl(cell$data, n = 1500, M = 10000, theta0 = cell$start,
                         covRandWalk = cell$cov, fnSim = cell_sim, fnSum = cell_sum,
                         shrinkage = 'glasso', penalty = 1.3, fnPrior = cell_prior,
                         simArgs = cell$sim_options, sumArgs = cell$sum_options,
                         parallel = TRUE, parallelArgs = list(.packages = 'BSL'),
                         thetaNames = expression(P[m], P[p]), verbose = TRUE)
stopCluster(cl)
registerDoSEQ()
show(resultCellBSLasso)
summary(resultCellBSLasso)
plot(resultCellBSLasso, thetaTrue = true_cell, thin = 20)

# Performing semiBSL (reduce the number of iterations M if desired)
# Opening up the parallel pools using doParallel
cl <- makeCluster(detectCores() - 1)
registerDoParallel(cl)
resultCellSemiBSL <- bsl(cell$data, n = 5000, M = 10000, theta0 = cell$start,
                         covRandWalk = cell$cov, fnSim = cell_sim, fnSum = cell_sum,
                         method = 'semiBSL', fnPrior = cell_prior,
                         simArgs = cell$sim_options, sumArgs = cell$sum_options,
                         parallel = TRUE, parallelArgs = list(.packages = 'BSL'),
                         thetaNames = expression(P[m], P[p]), verbose = TRUE)
stopCluster(cl)
registerDoSEQ()
show(resultCellSemiBSL)
summary(resultCellSemiBSL)
plot(resultCellSemiBSL, thetaTrue = true_cell, thin = 20)

# Plotting the results together for comparison
# plot using the R default plot function
par(mar = c(5, 4, 1, 2), oma = c(0, 1, 2, 0))
combinePlotsBSL(list(resultCellBSL, resultCelluBSL, resultCellBSLasso, resultCellSemiBSL),
    which = 1, thetaTrue = true_cell, thin = 20, label = c('bsl', 'ubsl', 'bslasso', 'semiBSL'),
    col = 1:4, lty = 1:4, lwd = 1)
mtext('Approximate Univariate Posteriors', outer = TRUE, cex = 1.5)
# }
# NOT RUN {
# }