Intensity normalizer¶

Feature list methods → Normalization → Intensity normalizer

The Intensity Normalizer corrects feature abundances (height or area) through a configurable multi-step normalization pipeline. Up to three correction steps are applied in sequence, each optional:

Metadata-based pre-normalization: divides or multiplies each sample's intensities by a numeric metadata value such as sample weight, dilution factor, or injection volume.
Internal standard correction: corrects each feature individually by the nearest or distance-weighted internal standard compound, accounting for extraction efficiency and matrix effects.
Inter-sample correction: removes signal drift within an analytical batch using reference samples (e.g., pooled QCs). Two kinds of methods are available:
- Sample-wide methods (By feature abundances, Spectral TIC) compute one factor per sample and, when multiple batches are present, align batch medians to a global reference median.
- Feature specific QC drift correction (QC-RLSC) fits a separate drift curve for every feature in every batch, so that each feature gets its own correction factor per sample. Batches are aligned in the same step.

Normalized values are written to dedicated Normalized Height and Normalized Area columns and the normalization functions are stored with the feature list so that they can be automatically re-applied to features added later (gap filling, manual integration). Processing steps that change feature intensities or row identities drop or adapt the stored normalization, see When the normalization is dropped.

Recommended citations¶

Info

When using the Feature specific QC drift correction (QC-RLSC), please consider citing:
Dunn W.B., Broadhurst D., Begley P., et al. Procedures for large-scale metabolic profiling of serum and plasma using gas chromatography and liquid chromatography coupled to mass spectrometry. Nature Protocols 6, 1060–1083 (2011), doi:10.1038/nprot.2011.335.

Parameters¶

Feature lists¶

Feature lists to normalize.

Name suffix¶

String appended to the output feature list name (default: norm).

Original feature list¶

Controls whether the original feature list is kept, removed, or modified in-place after normalization.

Feature measurement type¶

Whether feature height or feature area is used as the abundance measure throughout all normalization calculations. The resulting factors are applied to both the normalized height and the normalized area.

Batch correction metadata column (Optional)¶

Metadata column whose values identify the analytical batch a sample belongs to (e.g., Batch). When enabled, the normalization pipeline is run independently per batch:

Steps 2 and 3 are applied within each batch using only that batch's reference samples.
For the sample-wide step 3 methods, a final inter-batch normalization aligns the median reference-sample signal of every batch to the global median, correcting for between-run differences.
For the Feature specific QC drift correction, each feature is fitted within each batch and all batches are rescaled to one common target per feature, which aligns the batches in the same step. No separate inter-batch step is applied (see Feature specific QC drift correction).

Without a batch ID column, all samples are treated as one continuous sequence.

Tip

Use this parameter whenever samples were acquired in multiple separate analytical runs with their own QC injections to avoid over-correcting within a run while still aligning runs to each other.

Metadata normalization column: Step 1 (Optional)¶

Selects a numeric sample metadata column whose value is used to pre-normalize each sample before any other correction. Only a single column, a single value per sample can be used. Typical uses:

Example column	Operation	Effect
Sample weight (mg)	Divide	Normalizes to per-mg intensity
Injection volume (µL)	Divide	Normalizes to per-µL intensity
Dilution factor	Multiply	Corrects for sample dilution

Set a sample's metadata value to 0 to skip normalization for that sample.

Choose divide or multiply depending on whether the metadata column represents a quantity to scale down by (e.g., weight) or scale up by (e.g., dilution factor).

Intra-sample correction: Step 2 (Optional)¶

Applies an internal standard compound correction to each feature individually, accounting for extraction efficiency and matrix effects within a sample. This step runs after step 1 (metadata normalization).

Available options:

No normalization: step 2 is skipped.
Internal standard compounds: loads internal standards from a CSV/TSV file, matches them to feature-list rows by m/z, RT, and optional mobility, then normalizes each feature by the nearest or weighted standard. This workflow is described in more detail here.

Inter-sample correction: Step 3¶

Corrects signal drift across samples using a set of reference injections (typically pooled QCs) or all samples. This step runs after metadata normalization (step 1) and internal standard correction (step 2). For the sample-wide methods, intra-batch correction is applied first; if multiple batches are defined, inter-batch alignment is performed afterward by centering the median batch reference abundance against a global median.

Available options:

No normalization: step 3 is skipped.
By feature abundances: one factor per sample from a summary of all feature abundances. See sub-parameters below.
Spectral TIC: one factor per sample from the total ion current of the mass lists. See sub-parameters below.
Feature specific QC drift correction (QC-RLSC): one factor per feature and sample from a QC drift curve fitted per feature. See sub-parameters below.

By feature abundances sub-parameters¶

Reference samples¶

Sample types, as defined in the sample metadata in column mzmine_sample_type, used as the normalization reference (default: QC). Non-reference samples are normalized by interpolating correction factors between the nearest reference samples in acquisition order.

Normalize by feature abundances¶

Statistic used to summarize the feature intensity distribution of each sample:

Mode	Description
Median (default)	Median of all feature abundances: robust to outliers
Sum (TIC)	Sum of all feature abundances
Average	Mean of all feature abundances
Max	Maximum feature abundance

The correction factor for each sample is: median across all references / specific reference metric.

Spectral TIC sub-parameters¶

Reference samples¶

Sample types, as defined in the sample metadata in column mzmine_sample_type, used as the normalization reference (default: QC). Non-reference samples are normalized by interpolating correction factors between the nearest reference samples in acquisition order. The distance is defined by the run date in the sample metadata or as read from raw data file metadata.

The correction factor is: median TIC across all references / specific sample TIC, where TIC is computed from mass lists (noise-filtered spectra).

Warning

Spectral TIC is sensitive to noise and baseline signal. Prefer By feature abundances when the feature list has already been filtered, as it is more robust.

Feature specific QC drift correction (QC-RLSC) sub-parameters¶

Fits the drift of every feature in every batch on its pooled QC injections with robust LOESS smoothing and corrects all samples of that batch by the fitted curve (QC-RLSC). The algorithm is described in Feature specific QC drift correction.

Prerequisites

Every raw data file needs an acquisition timestamp (run date) in the sample metadata or in the raw data file metadata. Two samples of the same batch must not share the same timestamp.
Each batch needs at least five usable QC injections per feature (see Minimum usable QCs).
The Feature measurement type must be raw Height or Area.

Reference samples¶

Sample types, as defined in the sample metadata in column mzmine_sample_type, used to fit the drift of each feature (default: QC). Select comparable pooled QC samples only.

Conditioning QCs

Dunn et al. exclude the first QC injections of each batch (three for GC-MS, eight for UPLC-MS) that condition the system. Assign these injections a different sample type than the reference samples, or remove them from the feature list, so that they are not used for fitting.

Smoothing¶

Selects how the drift curve of each feature and batch is smoothed:

QC-RLSC (default): selects the LOESS bandwidth (neighborhood size) and the polynomial degree (linear or quadratic) separately for each feature and batch by leave-one-out cross-validation on the QC injections, as described by Dunn et al. No parameters. Uses two robustness iterations.
Manual LOESS: uses the same bandwidth, polynomial degree, and robustness iterations for all features and batches. No cross-validation.

Manual LOESS parameters:

LOESS bandwidth (default 0.75): fraction of the usable QCs of a batch that form each local fit, rounded up to a whole number of neighbors. At least polynomial degree + 2 neighbors are needed, otherwise the fit of this feature in this batch fails. Larger values give smoother curves.
Polynomial degree (default 1): 1 for local linear, 2 for local quadratic fitting.
Robustness iterations (default 2): number of robust reweighting iterations of the LOESS fit. Each iteration down-weights QC injections with large residuals (bisquare weights), which limits the influence of single outlier QCs. 0 disables the robust reweighting.

Minimum usable QCs¶

Default: 5, minimum 5. Minimum number of usable QC observations (positive, finite abundance) of a feature in a batch. Batches with fewer usable QCs are not corrected for this feature.

Maximum correction factor¶

Default: 10. Largest allowed change of a feature's value, up or down: with 10, values may be multiplied by up to 10 or divided by up to 10 (factors between 0.1 and 10). If any sample of a batch would need a larger correction, the feature gets no normalized values in that batch. The factor is never capped. The factors include both the drift within the batch and the offset of the batch from the common target.

Very large factors usually indicate an unreliable fit, e.g., a feature close to the noise level in the QCs or a curve that drops close to zero between QCs. This limit is an mzmine safeguard and not part of QC-RLSC by Dunn et al., who instead remove features with a high QC RSD after correction.

Global missing QC fraction¶

Default: 0.5. A feature is not corrected in any batch if the fraction of QC injections (across all batches) in which it has no usable abundance exceeds this value. Independently of this setting, a feature is not corrected in a batch where more than 50% of the QCs are missing.

Tip

Run the Feature specific QC drift correction after gap filling. The QC drift curves are calculated once from the abundances present at that time. Features added later by gap filling or manual integration get the stored factor of their row and sample, but the curves are not refitted.

Algorithm¶

Normalization pipeline¶

The three correction steps are always applied in the following order. Each step is independent and optional; any combination can be enabled:

Step	Parameter	Scope
1	Metadata normalization column	Per sample (constant factor)
2	Intra-sample correction	Per feature, per sample
3	Inter-sample correction: sample-wide methods	Per sample (interpolated)
3	Inter-sample correction: Feature specific QC drift correction	Per feature row, per sample (QC-RLSC)

When a batch ID column is provided, steps 2 and 3 are executed independently within each batch. For the sample-wide step 3 methods, this is followed by an inter-batch alignment step to center the reference abundances across all batches.

Normalization factors¶

Each normalization step contributes a factor per (sample, feature) pair. The factors from all enabled steps are multiplied together into a single composite function per sample. The final normalized abundance is:

\[ \text{NormalizedAbundance} = \text{OriginalAbundance} \times \prod_{i} f_i \]

where each \(f_i\) is the factor from normalization step \(i\). Steps 1 and the sample-wide step 3 methods depend only on the sample, step 2 on the sample and the feature's m/z and RT, and the feature specific QC drift correction on the sample and the feature list row.

If any step has no factor for a feature, the feature receives no normalized values (empty Normalized Height and Normalized Area). There is no fallback to the factors of the remaining steps.

Interpolation to non-reference samples¶

For step 2 and the sample-wide step 3 methods, correction factors may be computed from reference samples (typically pooled QCs). Non-reference samples receive a factor linearly interpolated between the two nearest reference samples in acquisition order (based on acquisition timestamp from metadata):

\[ f_{\text{sample}} = f_{\text{prev}} \cdot w_{\text{prev}} + f_{\text{next}} \cdot w_{\text{next}} \quad \text{where } w_{\text{prev}} + w_{\text{next}} = 1 \]

Samples before the first or after the last reference injection receive the factor of the nearest reference (no extrapolation).

The Feature specific QC drift correction does not use this linear interpolation; it evaluates a smooth curve fitted to the QCs, see below.

Batch-aware inter-batch normalization¶

When a batch ID column is enabled, a sample-wide step 3 method is selected, and multiple batches are present:

Steps 2 and 3 are run independently within each batch.
The median reference-sample metric is recorded for each batch.
A global median is calculated across all batches.
Each batch receives an additional scaling factor: globalMedian / batchMedian.

This aligns the overall signal level of all batches without disturbing the intra-batch corrections. This step is not applied for the Feature specific QC drift correction, which aligns the batches per feature through its common correction target.

Feature specific QC drift correction (QC-RLSC)¶

The Feature specific QC drift correction follows the QC robust LOESS signal correction (QC-RLSC) described by Dunn et al. (2011). It is calculated separately for each feature list row and each batch:

Input abundances: the raw height or area of each feature, multiplied by the factors of steps 1 and 2. Missing features (including features with status unknown), non-positive or non-finite values are not used. Gap-filled features are used.
Acquisition order: within each batch, samples are sorted by their acquisition timestamp and numbered 1, 2, 3, … The drift is modeled against this ordinal position, not against the elapsed time.
Availability checks: the feature is skipped
- in all batches if its fraction of missing QCs over all batches exceeds the Global missing QC fraction,
- in a batch if it has fewer than Minimum usable QCs usable QCs or more than 50% missing QCs in that batch.
Robust LOESS fit: the QC abundances are scaled by their median and smoothed with locally weighted linear or quadratic regression (tricube distance weights), followed by robust bisquare residual reweighting (two iterations for QC-RLSC, Robustness iterations for manual LOESS).
- QC-RLSC: all combinations of degree 1 and 2 and neighborhood sizes from degree + 2 up to the number of usable QCs are scored by the leave-one-out mean squared error over the QCs. The best candidate is selected; ties prefer the larger neighborhood (smoother curve), then the lower degree.
- Manual LOESS: the neighborhood size is ceil(bandwidth × usable QCs) with the selected degree.
Curve: the smoothed QC values are connected by a natural cubic spline. Samples acquired before the first or after the last usable QC of the batch receive the curve value of the nearest QC (no extrapolation).
Correction factor: for every sample \(s\) of the batch, including samples in which the feature was not detected,

\[ f_s = \frac{\tilde{y}_{\text{QC}}}{\hat{y}(s)} \]

where \(\hat{y}(s)\) is the fitted curve value at the sample's position in its batch and \(\tilde{y}_{\text{QC}}\) is the median of the usable QC abundances of this feature across all batches. Dividing by the fitted curve removes the drift and the offset of each batch, as in Dunn et al.; multiplying by the common median keeps the abundance scale. After correction, the QCs of all batches are centered on the same level. 7. Rejection: the whole feature/batch combination is rejected if the fit fails, a curve value is not positive or not finite, or any factor lies outside the Maximum correction factor bounds. The remaining batches of the same feature are still corrected.

Tip

The batch column defines which samples share one drift curve. Dunn et al. fit one curve per analytical block (e.g., two batches acquired within one week). Group the samples accordingly if the drift is continuous across several batches.

Unavailable features¶

Features without a factor keep their raw abundances but receive no normalized values for the affected samples (see Normalization factors). The reasons are recorded per row (and batch):

Reason	Scope	Cause
`NO_QC_REFERENCES`	Feature	No sample matches the reference sample types
`GLOBAL_MISSING_QC_THRESHOLD`	Feature	Missing QC fraction over all batches exceeds the global threshold
`BATCH_INSUFFICIENT_USABLE_QCS`	Feature/batch	Fewer usable QCs than Minimum usable QCs
`BATCH_MISSING_QC_THRESHOLD`	Feature/batch	More than 50% of the QCs in this batch are missing
`FIT_FAILED`	Feature/batch	No valid LOESS fit, e.g., too few neighbors for manual LOESS
`INVALID_PREDICTION`	Feature/batch	Fitted curve value not positive or not finite for a sample
`FACTOR_OUT_OF_BOUNDS`	Feature/batch	A factor exceeds the Maximum correction factor bounds

QC diagnostics¶

The task log and the applied method list a summary: the number of fully, partially, and not corrected rows, globally skipped rows, and the number of corrected and unavailable feature/sample pairs.

To inspect the Feature specific QC drift correction per row, open Feature lists tab → right-click a feature list → Show feature list summary, select the intensity normalization in the list of applied methods, and click QC diagnostics. For each row, the dialog shows the availability (fully, partially, or not corrected, or the global skip reason), the usable and total QC count, and the number of batches. Selecting a row shows the per-batch details: skip reason, usable QCs and their position range, number of samples outside the QC range, QC RSD before and after correction, leave-one-out error (QC-RLSC only), factor range, selected degree, span and neighborhood size, correction target (QC median across all batches), and the observed and fitted QC values along the curve.

Persistence and re-application¶

The normalization functions are stored in the settings of the feature list and saved with the project. Diagnostics (messages and Feature specific QC drift correction diagnostics) are stored with the applied method. When features are added later (e.g., by gap filling or manual integration), mzmine retrieves the stored functions and applies the same normalization to the new features.

The Feature specific QC drift correction stores its factors by feature list row ID. Gap-filled or manually integrated features receive the stored factor of their row and sample; the QC curves are not refitted.

When the normalization is dropped, adapted, or blocked¶

Some processing steps after normalization invalidate the stored normalization. mzmine then removes the normalization functions from the resulting feature list and clears the Normalized Height and Normalized Area values. Run the Intensity normalizer again afterward.

Processing step	Effect on the normalization
Feature resolving, smoothing, IMS expander	Dropped: the new features have new intensities
Feature list blank subtraction	Dropped in both result lists: removing features changes the basis of all normalization factors, including sample-wide factors such as the feature median
Split aligned feature list	Dropped in each split list: these lists are meant to be reprocessed per sample group
Row filters or copies that keep no rows	Dropped
Row filters and other steps that copy, filter, or renumber rows	Feature specific QC drift correction factors are remapped to the new row IDs; factors of removed rows are discarded
Rows removed from a feature list in place	Feature specific QC drift correction factors of the removed rows are discarded, so that a later row with a reused ID does not inherit them
Gap filling, manual integration	Normalization is re-applied to the new features with the stored functions
Merging or aligning multiple feature lists	Blocked with an error if any input list has a Feature specific QC drift correction normalization, because row IDs of different feature lists collide. Normalize again after alignment instead

The other normalization methods do not depend on row IDs and are kept when rows are copied, filtered, or removed.

Robin Schmid, Steffen Heuckeroth, Steffen Heuckeroth