Skip to contents

Iterate over different ncomp parameters for ICA

ica_evaluate_comp.Rd

This function allows to iterate over a vector of ncomp to identify which ncomp parameter to choose for your data set. The idea is to generate stability profiles over the different ncomps and identify a 'sweet spot' of good stability, low mutual information and good convergence of the identified independent components

Usage

ica_evaluate_comp(
  object,
  ica_type = c("logcosh", "exp"),
  iter_params = params_ica_randomisation(),
  ncomp_params = params_ica_ncomp(),
  ica_params = params_ica_general(),
  random_seed = 42L,
  .verbose = TRUE
)

Arguments

object

The class, see BulkCoExp(). You need to apply ica_processing() before running this function.

ica_type

String, element of c("logcosh", "exp").

iter_params

List. This list controls the randomisation parameters for the ICA runs, see params_ica_randomisation() for estimating stability. Has the following elements:

  • cross_validate - Boolean. Shall the data be split into different chunks on which ICA is run. This will slow down the function substantially, as every chunk needs to whitened again.

  • random_init - Integer. How many random initialisations shall be used for the ICA runs.

  • folds - If cross_validate is set to TRUE how many chunks shall be used. To note, you will run per ncomp random_init * fold ICA runs which can quickly increase.

ncomp_params

List. Parameters for the ncomp to iterate through, see params_ica_ncomp(). In the standard setting, c(2, 3, 4, 5) will be tested and then in steps until max_no_comp will be tested, i.e., c(2, 3, 4, 5, 10, 15, ..., max_no_comp - 5, max_no_comp).

  • max_no_comp - Maximum number of ncomp to test.

  • steps - Integer. In which steps to move from 5 onwards.

  • custom_seq - An integer vector. If you wish to provide a custom version of no_comp to iterate through.

ica_params

List. The ICA parameters, see params_ica_general() wrapper function. This function generates a list containing:

  • maxit - Integer. Maximum number of iterations for ICA.

  • alpha - Float. The alpha parameter for the logcosh version of ICA. Should be between 1 to 2.

  • max_tol - Maximum tolerance of the algorithm.

  • verbose - Controls verbosity of the function.

random_seed

Integer. For reproducibility.

.verbose

Boolean. Controls verbosity.

Value

BulkCoExp with the added information of stability of the components and other data to plot to choose the right ncomp.

Examples

# stability across a small grid of component counts
mat <- t(synthetic_signal_matrix()$mat)
obj <- BulkCoExp(mat, data.table::data.table(sample_id = rownames(mat)))
obj <- preprocess_bulk_coexp(obj, hvg = 0.3, .verbose = FALSE)
obj <- ica_processing(obj, .verbose = FALSE)
obj <- ica_evaluate_comp(
  obj,
  ica_type = "logcosh",
  ncomp_params = params_ica_ncomp(custom_seq = seq(2L, 20L, by = 2L)),
  .verbose = FALSE
)
head(get_ica_stability_res(obj))
#> Key: <no_components>
#>    no_components median_stability UQ_stability LQ_stability converged
#>            <int>            <num>        <num>        <num>     <num>
#> 1:             2        0.3341118    0.3371207    0.3311030      1.00
#> 2:             4        0.5896017    0.6422570    0.5311256      1.00
#> 3:             6        0.6287036    0.6477328    0.5436103      0.34
#> 4:             8        0.4903164    0.5704741    0.4185860      0.00
#> 5:            10        0.6038098    0.7605020    0.4277218      0.58
#> 6:            12        0.4654485    0.5977709    0.3903346      0.30
#>    norm_mutual_information combined_score
#>                      <num>          <num>
#> 1:              0.28908019      0.2375267
#> 2:              0.15172833      0.5001424
#> 3:              0.10588218      0.1911259
#> 4:              0.09540071      0.0000000
#> 5:              0.09111733      0.3182995
#> 6:              0.09419421      0.1264818