In caravagnalab/mobster: MOBSTER - Model-based tumour subclonal deconvolution

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

library(mobster)
library(tidyr)
library(dplyr)

Bootstrapping a model

This vignette describes how to compute the bootstrap confidence of a MOBSTER model.

Both parametric and nonparametric bootstrap options are available: the former samples data from the model, the latter re-samples the data (with repetitions). Statistics are bootstrap estimates (averages) of the bootstrap fits. In both cases a model bootstrap probability can be computed, as well as the probability of clustering together any two mutations.

We show this with a small synthetic dataset .to speed up the computation.

# Data generation
dataset = random_dataset(
  N = 400, 
  seed = 123, 
  Beta_variance_scaling = 100
  )

# Fit model -- FAST option to speed up the vignette
fit = mobster_fit(dataset$data, auto_setup = 'FAST')

# Composition with cowplot
cowplot::plot_grid(
  dataset$plot, 
  plot(fit$best), 
  ncol = 2, 
  align = 'h') %>%
  print

Now we can compute n.resamples nonparametric bootstraps using function mobster_bootstrap, passing parameters to the calls of mobster_fit. This function by defaults runs the fits in parallel (using a default percentage of the available cores); parallel computing capabilities are achieved using package easypar.

# The returned object contains also the list of bootstrap resamples, and the fits.
bootstrap_results = mobster_bootstrap(
  fit$best,
  bootstrap = 'nonparametric',
  cores.ratio = 0, # can be increased
  n.resamples = 25,
  auto_setup = 'FAST' # forwarded to mobster_fit
  )

The output object includes the bootstrap resamples, the fits and possible error returned by the runs.

# Resamples are available for inspection as list of lists, 
# with a mapping to record the mutation id of the resample data.
# Ids are row numbers.
print(bootstrap_results$resamples[[1]][[1]] %>% as_tibble())

# Fits are available inside the $fits list
print(bootstrap_results$fits[[1]])
plot(bootstrap_results$fits[[1]])

Errors of each run are available, if any.

print(bootstrap_results$errors)

Bootstrap statistics

Bootstrap statistics can be computed with bootstrapped_statistics.

With nonparametric bootstrap the data co-clustering probability is also computed (the probability of any pair of mutations in the data to be clustered together). Note that this probability depends on the joint resample probability of each pair of mutations (each bootstrapped with probability $1/n$, for $n$ mutations).

bootstrap_statistics shows to screen several statistics.

bootstrap_statistics = bootstrapped_statistics(
  fit$best, 
  bootstrap_results = bootstrap_results
  )

Visualising bootstrap results

Object bootstrap_statistics contains tibbles that can be plot with specific mobster functions.

# All bootstrapped values
print(bootstrap_statistics$bootstrap_values)

# The model probability
print(bootstrap_statistics$bootstrap_model)

# The parameter stastics
print(bootstrap_statistics$bootstrap_statistics)

Bootstrapping, one can plot the model frequency across re-samples. A model is identified by its mixture components (e.g., 2 Betas plus one tail).

plot_bootstrap_model_frequency(
  bootstrap_results, 
  bootstrap_statistics
  )

The bootstrap estimates of the parameters can be visualised.

# Plot the mixing proportions
mplot = plot_bootstrap_mixing_proportions(
  fit$best, 
  bootstrap_results = bootstrap_results, 
  bootstrap_statistics = bootstrap_statistics
  )

# Plot the tail parameters
tplot = plot_bootstrap_tail(
  fit$best, 
  bootstrap_results = bootstrap_results, 
  bootstrap_statistics = bootstrap_statistics
  )

# Plot the Beta parameters
bplot = plot_bootstrap_Beta(
  fit$best, 
  bootstrap_results = bootstrap_results, 
  bootstrap_statistics = bootstrap_statistics
  )

# Figure
figure = ggpubr::ggarrange(
  mplot,
  tplot,
  bplot,
  ncol = 3, nrow = 1,
  widths = c(.7, 1, 1)
)

print(figure)

For a nonparametric bootstrap we can plot also the co-clustering probability of the data.

plot_bootstrap_coclustering(
  fit$best, 
  bootstrap_results = bootstrap_results, 
  bootstrap_statistics = bootstrap_statistics
  )

caravagnalab/mobster documentation built on March 25, 2023, 3:40 p.m.