In msenosain/denoisingCTF: CyTOF Data Noise Removal R Package

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

knitr::opts_knit$set(root.dir = "../inst/doc/model_eval")

load('beads_TrainTest.RData') 
load('beads_RFmodel.RData')

Beads model evaluation

The diagram below shows the strategy to build the training and test sets for the beads classification model. A total of 170 000 and 60 000 events (i.e. cells, rows) were used for training and test sets, respectively.

{width=100%}

As previously explained, we used the Beads_TrainTest function to obtain training and test data sets.

denoisingCTF::Beads_TrainTest(sample_size = 40, method = 'k_means', 
    bsample = 5000, class_col = 'BeadsSmp_ID', ...)

We used the TrainModel function to train a Random Forest classification model:

denoisingCTF::TrainModel(train_set, test_set, alg = 'RF', class_col = 'BeadsSmp_ID', 
    seed = 40, name_0 = 'cells', name_1 = 'beads', label = 'beads',
    allowParallel = T, free_cores = 2)

To tune the training hyperparameters we used repeated 10-fold CV (x3). The metrics of the model are shown below:

model_rf

plot(model_rf)

Feature importance plot:

plot(ftimp_rf)

Assessing model accuracy in test set:

conf_rf

Debris model evaluation

load('debris_TrainTest.RData') 
load('debris_RFmodel.RData')

The diagram below shows the strategy to build the training and test sets for the debris classification model. A total of 220 000 and 80 000 events (i.e. cells, rows) were used for training and test sets, respectively.

{width=100%}

As previously explained, we used the pre_gate function to perform row indexing which is needed when comparing pre-gated and post-gated files (the latter are manually gated using user's preferred platform (e.g. Cytobank, FlowJo)) to successfully label debris and obtain training and test data sets with the post_gate function.

# Removal of zeros, beads and addition of row ID column
denoisingCTF::pre_gate(sample_size=30, model_beads=model_beads, alg_bd = 'RF') 

# Manually gate debris and dead cells using Gaussian Parameters and a live/dead cell marker.

# Comparison of pre-gated and post-gated files, noise labeling and training/testing data sets generation.
denoisingCTF::post_gate(bsample = 5000, path_pregated = '../') 

We used the TrainModel function to train a Random Forest classification model:

denoisingCTF::TrainModel(train_set, test_set, alg = 'RF', class_col = 'GP_Noise', 
    seed = 40, name_0 = 'cells', name_1 = 'debris', label = 'debris', 
    allowParallel = T, free_cores = 2)

To tune the training hyperparameters we used repeated 10-fold CV (x3). The metrics of the model are shown below:

model_rf

plot(model_rf)

Feature importance plot:

plot(ftimp_rf)

Assessing model accuracy in test set:

conf_rf

msenosain/denoisingCTF documentation built on Jan. 28, 2021, 2:23 a.m.