winn

winn implements White Noise Normalization for metabolomics LC-MS data. The package applies a decision-guided correction pipeline: each metabolite is first tested for serial dependence, and only metabolites that fail the white-noise gatekeeper are drift-corrected. This keeps stable features untouched while still correcting drift, batch structure, dilution effects, and outliers when they are present.

The package supports both QC-informed tuning and QC-agnostic workflows. If batch labels are unavailable, winn() can infer batch-like segments with an fkPELT-style change-point routine before downstream correction.

Installation

The current release is installed from GitHub. The default ANOVA correction path has no Bioconductor dependency. The optional ComBat correction path uses the Bioconductor package sva.

if (!requireNamespace("remotes", quietly = TRUE)) {
  install.packages("remotes")
}
remotes::install_github("ratschlab/winn")

If you plan to use remove_batch_effects = "combat", install sva separately:

if (!requireNamespace("BiocManager", quietly = TRUE)) {
  install.packages("BiocManager")
}
BiocManager::install("sva")

After a CRAN release is published, install.packages("winn") will also work.

Reproducible Example

The example below creates a small simulated dataset and runs the standard WiNN wrapper with fixed parameters. It is intentionally lightweight and matches the current exported API.

library(winn)

set.seed(1)

n_met <- 20L
n_samples <- 48L
batch <- rep(seq_len(4), each = 12)
run_order <- seq_len(n_samples)
qc_idx <- seq(4, n_samples, by = 8)
study_idx <- setdiff(seq_len(n_samples), qc_idx)

base_log <- matrix(rnorm(n_met * n_samples, mean = 8, sd = 0.15), nrow = n_met)
true_log <- base_log

signal_metabolites <- 1:4
study_signal <- as.numeric(scale(sin(study_idx / 6) + rnorm(length(study_idx), sd = 0.2)))
true_log[signal_metabolites, study_idx] <- true_log[signal_metabolites, study_idx] +
  0.35 * matrix(
    study_signal,
    nrow = length(signal_metabolites),
    ncol = length(study_idx),
    byrow = TRUE
  )

pooled_qc <- rowMeans(true_log[, study_idx, drop = FALSE])
true_log[, qc_idx] <- pooled_qc

dilution <- exp(rnorm(n_samples, sd = 0.05))
drift <- rep(seq(-0.18, 0.18, length.out = 12), times = 4)
batch_shift <- rep(c(-0.12, 0.04, 0.10, -0.06), each = 12)
noise <- matrix(rnorm(n_met * n_samples, sd = 0.08), nrow = n_met)

observed_log <- true_log +
  matrix(log(dilution), nrow = n_met, ncol = n_samples, byrow = TRUE) +
  matrix(drift + batch_shift, nrow = n_met, ncol = n_samples, byrow = TRUE) +
  noise

observed_intensity <- pmax(expm1(observed_log), 0)

corrected_intensity <- winn(
  observed_intensity,
  batch = batch,
  run_order = run_order,
  control_samples = qc_idx,
  parameters = "fixed",
  remove_batch_effects = "anova",
  median_adjustment = "shrink",
  lag = NULL,
  scale_by_batch = FALSE
)

dim(corrected_intensity)

Stepwise Pipeline

Each correction step is also available as a standalone function.

step1 <- adjust_outliers_mad(observed_intensity)

step2 <- autocorrelation_correct(
  step1,
  run_order = run_order,
  batch = batch,
  lag = NULL,
  test = "Ljung-Box",
  ljung_box_fitdf = 0,
  detrend = "mean"
)

step3 <- anova_batch_correction(step2, batch = batch)

step4 <- normalize_by_dilution_factor(
  step3,
  processing = "shrink",
  control_samples = qc_idx
)

step5 <- scale_by_batch(step4, batch = batch)  # optional

When pooled QCs are available and you want the package to search across parameter settings, use parameters = "auto". That mode is slower because it evaluates multiple combinations using the package’s QC quality score. Auto mode requires at least two controls. It searches its internal FDR, spline, and dilution-normalization grids; the corresponding fdr_threshold, spline_method, and median_adjustment arguments apply to fixed mode.

corrected_auto <- winn(
  observed_intensity,
  batch = batch,
  run_order = run_order,
  control_samples = qc_idx,
  parameters = "auto",
  lag = NULL,
  scale_by_batch = FALSE
)

If batch labels are missing, set batch = NULL and provide run_order so WiNN can segment the run automatically.

corrected_qc_agnostic <- winn(
  observed_intensity,
  batch = NULL,
  run_order = run_order,
  lag = NULL,
  pelt_penalty = "mbic"
)

Practical Notes

• Use lag = NULL unless you have a study-specific reason to force a fixed Ljung-Box lag.
• The Ljung-Box default is ljung_box_fitdf = 0, because WiNN does not fit an ARMA model before testing. ljung_box_fitdf = 1 is retained only for legacy reproduction and sensitivity analysis.
• With supplied run_order, WiNN sorts samples within each batch for both serial-dependence testing and drift fitting, then restores the input column order. If run_order = NULL, matrix-column order is treated as acquisition order. Run-order values must be unique within each supplied batch.
• Use remove_batch_effects = "anova" for selective, mean-only correction: only features passing its FDR gate are shifted. Use remove_batch_effects = "combat" for global empirical Bayes location/scale adjustment of every feature when sva is available.
• Keep scale_by_batch = FALSE when absolute abundance differences matter for downstream interpretation.
• WiNN’s log-based pipeline requires finite, non-negative quantitative intensities. Numeric zeros are valid observations and are never recoded. Convert explicit non-detection sentinels to missing values and apply a documented imputation policy before calling winn(). Features with a zero PQN reference median are excluded from dilution-factor estimation but remain in the returned matrix; the estimated sample factors are still applied to every feature. Samples whose dilution factor is zero are rejected with an informative error.
• The vignette provides a larger end-to-end example: vignette("winn_tutorial", package = "winn").

Diagnostics and Ablations

The correction operators can return their decision records without changing their usual matrix return values. Set return_diagnostics = TRUE for the drift, batch, or dilution step when an analysis needs to retain test results, selected features, fallback smoothers, correction magnitudes, or sample-level scaling factors.

winn_ablation() provides a fixed-parameter interface for controlled workflow ablations. It uses the same package operators as winn() and can switch the drift and batch gates between selective, forced-all, and disabled modes. The function returns the intermediate matrices, diagnostics, runtimes, and complete configuration used for the run. It is intended for method evaluation and does not replace the standard winn() entry point.

The main wrapper retains its matrix return by default. Set return_details = TRUE to receive a structured result containing $data, the selected fixed or automatic parameters, batch assignments, run order, control indices, and per-stage decisions.

detailed <- winn(
  observed_intensity,
  batch = batch,
  run_order = run_order,
  control_samples = qc_idx,
  parameters = "fixed",
  return_details = TRUE
)

detailed$selected_parameters
head(detailed$stage_decisions$drift)

ablation <- winn_ablation(
  observed_intensity,
  batch = batch,
  run_order = run_order,
  control_samples = qc_idx,
  drift_gate = "selective",
  batch_gate = "all",
  pqn_mode = "shrink"
)

head(ablation$diagnostics$drift)

The MAD preparation step computes each feature’s median, MAD, thresholds, and both outlier masks once from the original finite values. Upper and lower extremes are then shrunk independently in a single non-iterative pass. The thresholds are never recomputed after either tail is adjusted.

Citation

If you use winn, cite the WiNN method paper:

Demler O, Giulianini F, Liu Y, Londschien M, Sjöström A, Tanna T, Luttmann-Gibson H, and Jeanrenaud A (2024). “WiNNbeta: Batch and drift correction method by white noise normalization for metabolomic studies.” arXiv:2404.07906. https://doi.org/10.48550/arXiv.2404.07906.

This is the citation for the underlying method paper. The winn R package is authored and maintained by Tanmay Tanna.