Data analysis on AlpsNMR can be performed on both nmr_dataset_1D full spectra as well as nmr_dataset_peak_table peak tables.
nmr_data_analysis(
dataset,
y_column,
identity_column,
external_val,
internal_val,
data_analysis_method,
.enable_parallel = TRUE
)An nmr_dataset_family object
A string with the name of the y column (present in the metadata of the dataset)
NULL or a string with the name of the identity column (present in the
metadata of the dataset).
A list with two elements: iterations and test_size.
See random_subsampling for further details
An nmr_data_analysis_method object
Set to FALSE to disable parallellization.
A list with the following elements:
train_test_partitions: A list with the indices used in train and test on each of the cross-validation iterations
inner_cv_results: The output returned by train_evaluate_model on each inner cross-validation
inner_cv_results_digested: The output returned by choose_best_inner.
outer_cv_results: The output returned by train_evaluate_model on each outer cross-validation
outer_cv_results_digested: The output returned by train_evaluate_model_digest_outer.
The workflow consists of a double cross validation strategy using random
subsampling for splitting into train and test sets. The classification model
and the metric to choose the best model can be customized (see
new_nmr_data_analysis_method()), but for now only a PLSDA classification
model with a best area under ROC curve metric is implemented (see
the examples here and plsda_auroc_vip_method)
# Data analysis for a table of integrated peaks
## Generate an artificial nmr_dataset_peak_table:
### Generate artificial metadata:
num_samples <- 32 # use an even number in this example
num_peaks <- 20
metadata <- data.frame(
NMRExperiment = as.character(1:num_samples),
Condition = rep(c("A", "B"), times = num_samples / 2)
)
### The matrix with peaks
peak_means <- runif(n = num_peaks, min = 300, max = 600)
peak_sd <- runif(n = num_peaks, min = 30, max = 60)
peak_matrix <- mapply(function(mu, sd) rnorm(num_samples, mu, sd),
mu = peak_means, sd = peak_sd
)
colnames(peak_matrix) <- paste0("Peak", 1:num_peaks)
## Artificial differences depending on the condition:
peak_matrix[metadata$Condition == "A", "Peak2"] <-
peak_matrix[metadata$Condition == "A", "Peak2"] + 70
peak_matrix[metadata$Condition == "A", "Peak6"] <-
peak_matrix[metadata$Condition == "A", "Peak6"] - 60
### The nmr_dataset_peak_table
peak_table <- new_nmr_dataset_peak_table(
peak_table = peak_matrix,
metadata = list(external = metadata)
)
## We will use a double cross validation, splitting the samples with random
## subsampling both in the external and internal validation.
## The classification model will be a PLSDA, exploring at maximum 3 latent
## variables.
## The best model will be selected based on the area under the ROC curve
methodology <- plsda_auroc_vip_method(ncomp = 3)
model <- nmr_data_analysis(
peak_table,
y_column = "Condition",
identity_column = NULL,
external_val = list(iterations = 3, test_size = 0.25),
internal_val = list(iterations = 3, test_size = 0.25),
data_analysis_method = methodology
)
## Area under ROC for each outer cross-validation iteration:
model$outer_cv_results_digested$auroc
#> # A tibble: 3 × 3
#> cv_outer_iteration ncomp auc
#> <int> <int> <dbl>
#> 1 1 1 0.933
#> 2 2 1 0.875
#> 3 3 2 1
## Rank Product of the Variable Importance in the Projection
## (Lower means more important)
sort(model$outer_cv_results_digested$vip_rankproducts)
#> Peak2 Peak6 Peak17 Peak11 Peak3 Peak14 Peak20 Peak12
#> 1.259921 1.587401 3.556893 5.646216 6.316360 7.651725 8.962809 9.049114
#> Peak9 Peak19 Peak15 Peak1 Peak4 Peak5 Peak7 Peak13
#> 9.165656 9.654894 11.100998 11.686316 12.091887 12.428930 12.632719 13.782348
#> Peak16 Peak18 Peak10 Peak8
#> 13.924767 14.227573 14.986655 17.324782