Skip to contents

Compute imputation quality metrics (NRMSE, SOR, PSS, ACC_OI)

Source: R/Imputation_Metrics.R
im_compute_metrics.Rd

Extracts complete-case rows from the assay, introduces artificial NAs, re-imputes with each method, and computes four quality metrics plus a combined ranking.

Usage

im_compute_metrics(
  se,
  assay_name = NULL,
  methods = .IM_BENCH_METHODS,
  combo_methods = list(),
  condition_col = "Condition",
  na_prop = 0.2,
  seed = 42L,
  pattern = "random",
  method_args = list(),
  with_value = NA_real_,
  verbose = TRUE
)

Arguments

se

SummarizedExperiment (from import_imp_matrices() or pipeline)

assay_name

Character. Name of the assay to use as starting point (typically the normalized assay, pre-imputation, with real NAs).

methods

Character vector of individual imputation methods to benchmark. Default: .IM_BENCH_METHODS (16 methods). Use character(0) or NULL to skip individual methods.

combo_methods

Named list of combo/softHybrid configurations. Each element is a list with: mode ("combo" or "softHybrid", default "combo"), mar_method, mnar_method. The list name is used as the method label in results. Default: empty list (no combo methods).

condition_col

Character. Column in colData with condition labels, required for combo methods. Default "Condition".

na_prop

Numeric (0-1). Proportion of artificial NAs. Default 0.20. Only used when pattern = "random". Ignored when pattern = "from_data" (NA proportion is derived from the data, replicating NAguideR).

seed

Integer. Random seed. Default 42.

pattern

Character. NA introduction pattern: "random" for uniform random NAs, or "from_data" to replicate NAguideR's simulation strategy (row ratio + per-column distribution from the actual data).

method_args

Named list of per-method argument lists.

with_value

Constant value for method "with".

verbose

Logical. Print progress. Default TRUE.

Value

data.frame with columns: Method, NRMSE, SOR, PSS, ACC_OI, NRMSE_Rank, SOR_Rank, PSS_Rank, ACC_OI_Rank, Rank_Mean. Ordered by Rank_Mean (best first).

Examples

data(nadia_dia)
se <- im_prepare_se(nadia_dia, norm_method = "cycloess", verbose = FALSE)

# Ground-truth simulation: complete rows are masked and re-imputed, so the
# error of each method can be measured against the values it did not see
metrics <- im_compute_metrics(se, methods = c("min", "zero", "knn"),
                              verbose = FALSE)
metrics[, c("Method", "NRMSE", "SOR", "Rank_Mean")]
#>   Method     NRMSE  SOR Rank_Mean
#> 1    knn 0.1293948 1349         1
#> 2    min 3.1934374 2698         2
#> 3   zero 6.3558709 4047         3