bio-metabolomics-normalization-qc

What this skill does

## Version Compatibility

Reference examples tested with: pmp 1.14+, statTarget 1.30+, imputeLCMD 2.1+, missForest 1.5+, sva 3.50+

Before using code patterns, verify installed versions match. If versions differ:
- R: `packageVersion('<pkg>')` then `?function_name` to verify parameters

Valid drift correction requires QC injections that bracket the samples at both ends and sample the drift curve (~1 QC every 5-10 injections); conditioning injections must be excluded. Valid batch correction requires biological groups randomized across batches; a confounded design cannot be rescued by any algorithm.

If code throws ImportError, AttributeError, or TypeError, introspect the installed
package and adapt the example to match the actual API rather than retrying.

# Metabolomics Normalization and QC

**"Normalize my metabolomics data and correct for batch effects"** -> Filter junk features by QC quality, correct within-batch drift against injection order, normalize per-sample dilution, and impute by missingness mechanism -- each step verified against held-out QCs, not just QC clustering.
- R drift correction: `QCRSC()` (pmp), `shiftCor()` (statTarget)
- R normalization: `pqn_normalisation()` (pmp)
- R imputation: `mv_imputation()` (pmp), `impute.QRILC()` (imputeLCMD), `missForest()` (missForest)

## The Single Most Important Insight -- Normalization Is a Modeling Decision, Not a Cleanup Step

Every preprocessing step imposes an assumption about where the *unwanted* variance lives; if that assumption is wrong the result is not noisier, it is confidently wrong. Three corollaries reorganize the whole skill. (1) QC-based correction assumes the pooled QC's per-feature drift trajectory *is* the samples' trajectory -- false for subgroup-specific features (the pool dilutes them toward absence) and for features at different abundance in samples vs pool (suppression is concentration-dependent), so correcting them extrapolates from noise. (2) The order/batch/biology confound is information-theoretically unwinnable post hoc: if group is collinear with batch or injection order, no estimator can attribute the shared variance to one source -- it only redistributes it, wrongly. Randomization at the bench is the only real fix. (3) Over-correction is invisible to the metric everyone reports: "QC RSD dropped / QCs cluster tighter" is exactly what a too-flexible model games (a cubic spline threading every QC drives QC RSD to ~0% while raising biological-sample RSD). Validate on held-out QCs and dilution-QC linearity, never on the metric the model optimized.

## The Four Orthogonal Operations (Do Not Conflate)

| Operation | Acts on | Removes | Methods |
|---|---|---|---|
| Drift / signal correction | each feature, within a batch, vs injection order | longitudinal intensity decay/rise (column fouling, sensitivity loss) | QC-RLSC (LOESS), QCRSC (spline), QC-RFSC (RF vs order), SERRF (RF across correlated features) |
| Batch correction | each feature, across batches | step-changes between analytical batches | QC-anchored median/reference alignment; ComBat (reserved, dangerous) |
| Sample normalization | each sample (column) | dilution / total-amount differences | PQN, MSTUS, TIC/sum, median, internal standard |
| Transformation / scaling | each feature (row) | mean-variance dependence; range dominance | log/glog; Pareto/auto -> defers to metabolomics/statistical-analysis |

TIC normalization does not handle drift, and -- because of closure -- can spread one feature's change across all others. Keep the axes separate.

## Pipeline Order (and Why Order Matters)

| # | Step | Why here | Tool |
|---|---|---|---|
| 0 | Exclude conditioning injections | Pre-equilibrium signal warps a LOESS edge and corrupts RSD/blank filters | manual (drop first ~8 QC) |
| 1 | Blank filter -> detection-rate filter | Removes background/contaminant and mostly-absent features before any model trains on them | `filter_peaks_by_blank`, `filter_peaks_by_fraction` (pmp) |
| 2 | Within-batch drift correction | Flattens order-dependent trend per feature before cross-sample comparison | `QCRSC` (pmp), `shiftCor` (statTarget) |
| 3 | QC RSD / D-ratio filter | Drift correction *should* improve RSD; filter after so reproducibility reflects corrected data (report both stages) | `filter_peaks_by_rsd` (pmp), `dratio_filter` (structToolbox) |
| 4 | Between-batch alignment | QC-anchored offsets removed after within-batch drift is flat | median-of-QC / batchCorr |
| 5 | Missing-value imputation | Filter aggressively first, then impute only the sparse residual holes by mechanism | `mv_imputation` (pmp), `impute.QRILC`, `missForest` |
| 6 | Sample normalization | Dilution correction on quality features, after junk removed | `pqn_normalisation` (pmp) |
| 7 | Transformation + scaling | Defers to metabolomics/statistical-analysis | `glog_transformation` (pmp) |

Detection-rate filtering must precede imputation: never impute a feature that is 90% missing, which would fabricate 90% of it.

## Decision Tree -- Sample Normalization by Matrix

| Matrix / situation | Use | Why |
|---|---|---|
| Urine / variable-dilution biofluid | PQN or MSTUS (osmolality/SG if measured) | Dilution varies wildly; PQN's median-quotient isolates the common dilution factor; MSTUS excludes drug/diet xenobiotics that corrupt TIC |
| Plasma / serum | PQN or median (TIC only if no dominant peak) | Volume relatively constant; closure risk lower but still present |
| Tissue / cells | Per measured amount (mass, protein, cell count) at the bench | The confounder (input amount) is known -- more honest than any data-driven post-hoc method |
| Targeted / few analytes | Per-class internal standards | One IS cannot represent all chemical classes/RT regions |
| Global profile genuinely differs between groups | Avoid quantile normalization | It forces all samples to one distribution, erasing real distributional biology |
| Creatinine for urine | Avoid as sole method | Fails under renal impairment / muscle-mass differences (Warrack 2009) |

When >50% of features move coherently (potent drug, gross pathology), the PQN median-quotient measures the biology, not dilution, and subtracts it out -- switch to a measured external quantity and check whether the normalization factor correlates with the phenotype.

## Decision Tree -- Drift Correction Method

| Situation | Do | Why |
|---|---|---|
| Smooth monotonic drift, frequent QCs, small/medium study | QCRSC (spline) or QC-RLSC | Per-feature fit vs order; CV-select span to avoid overfit |
| Non-smooth / multi-pattern drift within a batch | QC-RFSC (statTarget) or batchCorr clusters | RF / cluster-based captures non-monotonic trend |
| Large cohort (>~500), complex multi-source error, want lowest RSD | SERRF | Borrows strength across correlated features (~5% RSD on >800-sample cohorts, Fan 2019) |
| Sparse QCs (<5-6 spanning the batch) | Coarse median-of-QC offset or no within-batch correction | LOESS/spline with too few QCs produces gaps/garbage |
| Feature weak/absent in QCs | Exclude from correction | Correcting it extrapolates from noise |
| No detectable drift in a feature | Do not correct it | Correcting a flat QC trajectory only adds the model's wiggle |
| Run order confounded with biology | Do not drift-correct; fix design or caveat | A smooth function of order absorbs and subtracts the biological trend |

Flexible ML methods (SERRF/RF/adversarial) win on large complex cohorts but are *more* prone to learning-and-removing biology that tracks order/batch. Always confirm QC RSD dropped AND biological-sample RSD did not rise.

## Filter Features by QC Quality (RSD and D-ratio)

**Goal:** Keep only reproducible features whose technical variance is small relative to biological variance.

**Approach:** Compute per-feature QC RSD and the robust D-ratio (technical SD / biological SD), then apply a boolean mask. Lead with D-ratio: CV alone is matrix-blind, scoring a precisely-measured-but-flat feature as good and a noisy-but-