Skip to contents

Batch correction using SERRF

serrf-normalisation.Rmd

Batch correction using Systematical Error Removal using Random Forest (SERRF) 

out_serrf <- OmicsProcessing::normalise_SERRF(
  imputed_df,
  target_cols = "@",
  is_qc = grepl("^sQC", imputed_df$sample_type),
  strata_col = "batch"  # omit to treat all samples as one strata
)

normalise_SERRF() models unwanted technical variation using QC samples within each batch/stratum and applies SERRF normalization to the specified feature columns. Use target_cols to select features by name, tidyselect helper, or regex (resolved via resolve_target_cols()), and supply a logical is_qc flag for each row. See the reference: normalise_SERRF().

Input requirements

  • strata_col: provide the batch/strata column when it exists; if you omit it, a dummy single-level strata is created so all samples are treated as one batch. If supplied, it is coerced to a factor (no NAs).
  • is_qc is a logical vector the same length as nrow(df); ensure every batch has enough QC samples for the SERRF model.
  • All target_cols must be numeric and contain no NA values (impute beforehand).
  • At least three target features are required (SERRF needs predictors to model each feature).

Tuning and tips

  • num_screen_SERRF (default 10) controls how many correlated features are screened per target; increase cautiously if you have many features and stable QC coverage.
  • Inspect before/after QC CVs or PCA to verify batch drift is reduced.
  • Keep batches balanced: very small batches or few QC points can lead to unstable fits; consider merging sparse strata or adding QC if possible.