Hybrid imputation with RF + LCMD

Step 4: Hybrid Imputation (Random Forest + LCMD)

imputed_results <- OmicsProcessing::hybrid_imputation(
  log_transformed_df,
  target_cols = "@",
  method = c("RF-LCMD"),
  oobe_threshold = 0.1
)
imputed_df <- imputed_results$hybrid_rf_lcmd

hybrid_imputation() combines two complementary strategies:

• Random Forest (RF) via missForest::missForest() for features that are well-predicted by other variables (low OOBE).
• Left-Censored Missing Data (LCMD) via imputeLCMD::impute.MAR.MNAR() for features with higher OOBE.

The function fits RF per feature, uses the out-of-bag error (OOBE) to decide whether to keep the RF estimate or switch that feature to LCMD, and returns:

• hybrid_rf_lcmd: the combined result.
• rf / lcmd: per-method outputs.
• oob: OOBE values (helpful for diagnostics).

See the full reference: hybrid_imputation().

Handling QC samples with is_qc

Rows flagged as TRUE by is_qc are excluded from the hybrid imputation calculations. The rationale is that QC samples are repeatedly measured throughout the laboratory workflow and are intended to capture technical variability in the measurement process. Excluding them prevents these rows from influencing the RF and LCMD imputation models fitted to the biological samples.

If QC samples contain missing values that you also want to impute, there are two main options.

A. Impute QC samples together with the rest of the data

To include all rows in the hybrid imputation process, leave is_qc as NULL, or provide a logical vector of the same length as the number of rows in the data, with all values set to FALSE.

imputed_results <- OmicsProcessing::hybrid_imputation(
  log_transformed_df,
  target_cols = "@",
  method = "RF-LCMD",
  oobe_threshold = 0.1,
  is_qc = NULL
)

is_qc_all_false <- rep(FALSE, nrow(log_transformed_df))

imputed_results <- OmicsProcessing::hybrid_imputation(
  log_transformed_df,
  target_cols = "@",
  method = "RF-LCMD",
  oobe_threshold = 0.1,
  is_qc = is_qc_all_false
)

B. Impute QC samples using only QC sample values

If you want to impute QC samples separately, using only information from other QC samples, you can invert the is_qc vector. This excludes all non-QC rows from the imputation calculations and applies the pipeline only to the QC samples.

is_qc <- sample_metadata$is_qc

imputed_qc_results <- OmicsProcessing::hybrid_imputation(
  log_transformed_df,
  target_cols = "@",
  method = "RF-LCMD",
  oobe_threshold = 0.1,
  is_qc = !is_qc
)

imputed_qc_df <- imputed_qc_results$hybrid_rf_lcmd

Customising the RF and LCMD controls

You can tweak both engines via control lists:

my_control_RF <- list(
  parallelize = "no",
  mtry = floor(sqrt(length(target_cols))),
  ntree = 100,
  maxiter = 10,
  variablewise = TRUE,
  verbose = TRUE,
  n_cores = parallel::detectCores()
)

my_control_LCMD <- list(
  method.MAR = "KNN",
  method.MNAR = "QRILC"
)

df_rf_lcmd_hybrid <- OmicsProcessing::hybrid_imputation(
  log_transformed_df,
  target_cols = "@",
  method = c("RF-LCMD"),
  oobe_threshold = 0.1,
  control_LCMD = my_control_LCMD,
  control_RF = my_control_RF
)

Parallelising the RF step (missForest)

missForest can run in parallel when you register a foreach backend and set parallelize:

library(doParallel)

n_cores <- parallel::detectCores(logical = FALSE)
cl <- parallel::makeCluster(n_cores)
doParallel::registerDoParallel(cl)

ctrl_parallel_RF <- list(
  parallelize = "variables", # or "forests"
  mtry = floor(sqrt(length(target_cols))),
  ntree = 200,
  maxiter = 10,
  variablewise = TRUE,
  verbose = TRUE
)

imputed_parallel <- OmicsProcessing::hybrid_imputation(
  log_transformed_df,
  target_cols = "@",
  method = "RF-LCMD",
  oobe_threshold = 0.1,
  control_RF = ctrl_parallel_RF
)

parallel::stopCluster(cl)
doParallel::registerDoSEQ()

Guidance:

• Use parallelize = "variables" for many features; "forests" spreads trees instead.
• Keep ntree reasonable when parallelising to avoid memory pressure.

Tips:

• Keep target_cols explicit when possible for clarity; "@" will use all feature columns resolved via resolve_target_cols().
• Inspect imputed_results$oob to confirm the RF ↔︎ LCMD split aligns with your expectations.
• For very wide matrices, tune ntree, mtry, or the number of worker cores to balance runtime and stability.