Skip to contents

SERRF batch correction

Source: R/correct-batch.R
correct_batch_serrf.Rd

[Experimental]

This is an independent re-implementation of SERRF in mrmhub, adapted from the reference code in the malbacR package (not the original authors' package); see Details. Validate results against the reference SERRF for your data.

Normalises systematic error with SERRF (Systematic Error Removal using Random Forest; Fan et al. 2019). For each feature and batch a random forest is trained on the reference QC samples, using the batch's most-correlated features as predictors, and the learned systematic error is removed from all samples. Unlike ComBat, SERRF captures non-linear drift and batch effects jointly and is anchored on the QC samples, matching the QC-based design of the package; it is best suited to larger panels with dense QC coverage.

SERRF operates on the raw abundance scale (no log transform). Features with missing or non-positive values, and batches with fewer than two reference QCs, are left uncorrected. Batch correction is performed after normalization and drift correction in the recommended pipeline.

Usage

correct_batch_serrf(
  data = NULL,
  variable,
  ref_qc_types,
  n_correlated = 10,
  num_trees = 500,
  seed = 1L,
  num_threads = 1L,
  show_progress = TRUE,
  replace_previous = TRUE,
  feature_list = NULL,
  replace_exisiting_trendcurves = FALSE
)

Arguments

data

A MRMhubExperiment object.

variable

The variable to correct: one of "intensity", "norm_intensity", or "conc".

ref_qc_types

Character vector of QC types used as the SERRF training (reference) samples per batch.

n_correlated

Number of most-correlated features used as random-forest predictors per feature. Default 10.

num_trees

Number of trees per random forest. Default 500.

seed

Random seed for ranger::ranger(), for reproducibility. Default 1.

num_threads

Threads per random forest passed to ranger::ranger(). Default 1; kept low because batches are already corrected in parallel and each forest trains on a small QC set.

show_progress

Show a progress bar over batches. Default TRUE.

replace_previous

Replace a previous batch correction (TRUE, default) or apply on top of it.

feature_list

Optional feature selection (character vector or a single regular expression); NULL (default) selects all features.

replace_exisiting_trendcurves

Reseed the plotting trend curves. Default FALSE.

Value

A MRMhubExperiment with corrected data.

Details

The implementation adapts the reference SERRF code in the malbacR package, keeping its normalization arithmetic but selecting each feature's predictors from the per-batch QC-sample Spearman correlation (a simplification of malbacR's QC/sample correlation intersection). Random forests are fit with ranger::ranger() and a fixed seed, so results are reproducible. Users are encouraged to validate results against the reference SERRF implementation for their data.

Batches are corrected independently and in parallel via mirai (through purrr::in_parallel()); set up workers with mirai::daemons() beforehand to use them, otherwise the batches are processed sequentially.

References

Fan S, Kind T, Cajka T, et al. (2019). Systematic Error Removal Using Random Forest for Normalizing Large-Scale Untargeted Lipidomics Data. Analytical Chemistry, 91(5), 3590-3596. doi:10.1021/acs.analchem.8b05592

Implementation adapted from the malbacR package (https://github.com/pmartR/malbacR); random forests via ranger (Wright MN, Ziegler A, 2017, Journal of Statistical Software, 77(1), doi:10.18637/jss.v077.i01 ).

See also

correct_batch_centering(), correct_batch_combat(), correct_drift_loess() and plot_runscatter() for visualisation. The drift and batch correction manual.

Examples

# mexp <- correct_batch_serrf(mexp, variable = "conc", ref_qc_types = "BQC")