What each normalization method does to your data, what it adjusts, what it leaves untouched, and which to choose for transcriptomics and metabolomics. Every section has a live lab: change the data and watch the method work, or fail.
Every omics measurement mixes biology with the way the sample was handled and measured. Normalization removes the handling, so that what is left is the biology.
Two RNA samples, 400 genes. Of these, 40 truly differ (orange) and 360 do not. Sample B is sequenced more deeply than sample A. Raise the depth ratio and watch what happens to the genes that did not change.
Sequencing depth, how much RNA went into the library, how dilute a urine sample was, how much material was loaded on the column.
Instrument sensitivity changing across a run, reagent lots, operators, days. Often shared by many features at once.
Differences between groups, treatments and time points. Normalization must leave these alone, which is why every method makes an assumption about what stays the same.
If batch and group are the same thing, no normalization can separate them. Randomize and balance first, as in Episode 11.
| Level | What it acts on | Question it answers | Examples in this guide |
|---|---|---|---|
| Sample level | One factor per sample (a column) | Is this sample simply bigger, deeper or more dilute? | CPM, TPM, TMM, median of ratios, PQN, total sum, median, internal standards |
| Distribution level | The whole shape of each sample | Do all samples have the same distribution of values? | Quantile normalization, cyclic loess, median centring |
| Feature level | One transformation per feature (a row) or for all values | Does the spread of a feature depend on its size? Do features have comparable weight? | log2, arcsinh, VST, centring, autoscaling, Pareto |
| Batch and drift | Position in the run, batch label | Did the measurement change over time or between batches? | QC-based LOESS, ComBat, removeBatchEffect, RUV |
Raw RNA-seq counts depend on depth and on gene length. CPM, RPKM, FPKM and TPM each remove a different part of that, and each leaves something behind.
Eight genes of different lengths in two samples. Sample 2 was sequenced 1.6 times deeper. Choose a unit and compare the two samples gene by gene, and compare the totals.
| Unit | Formula for gene g in a sample | Removes | Leaves behind |
|---|---|---|---|
| CPM | count_g / total counts x 10^6 | Sequencing depth | Gene length, composition |
| RPKM, FPKM | count_g / (length in kb x total counts in millions) | Depth and gene length | Sums differ between samples; composition |
| TPM | (count_g / length in kb) / sum over genes of (count / length in kb) x 10^6 | Depth and gene length | Composition. Every sample sums to one million |
FPKM is RPKM for paired-end data, where a fragment is counted once instead of twice. For testing differences between groups, use raw counts in a count-based model (DESeq2, edgeR, limma-voom) and let the model handle depth. CPM, RPKM and TPM are for plots, filtering and descriptions.
Dividing by the total assumes the total amount of RNA is the same in every sample. When a few genes take a large share of the reads, every other gene looks lower, although nothing happened to it.
Two samples with the same number of reads. In sample B, six genes (orange) take a growing share of the reads. The other 494 genes are identical in both samples. Pick a method and see which one still centres them on zero.
| Method | How the factor is found | Assumption | Where it is used |
|---|---|---|---|
| Total count | Library size | The total RNA is the same in every sample | CPM, TPM, plots |
| Upper quartile | 75th percentile of the non-zero counts in each sample | The upper quartile is not driven by changing genes | edgeR (calcNormFactors), older pipelines |
| TMM | Weighted mean of the log ratios against a reference sample, after trimming 30 percent of the ratios and 5 percent of the average expression at each end | Most genes are not differentially expressed | edgeR, limma-voom |
| Median of ratios | Divide each gene by its geometric mean over samples, then take each sample's median ratio | Most genes are not differentially expressed; needs genes with no zeros | DESeq2 size factors |
All four factors are used as scaling factors or offsets in the model, so the counts themselves stay integers. TMM and median of ratios are the usual choices for differential expression of bulk RNA-seq.
Quantile normalization forces every sample to have exactly the same distribution. It is powerful when most features do not change, and wrong when whole distributions truly differ.
Four samples of 300 log2 intensities, each shifted and stretched by the measurement. Apply median centring or quantile normalization, then switch on a real biological shift in sample D to see what the method cannot tell apart.
It was built for microarrays and is common in proteomics. It assumes most features do not change. If one condition really has more of everything, for example a transcription factor that raises most genes, quantile normalization removes the very effect you wanted to measure.
In metabolomics the biggest sample-level problem is often how concentrated the sample was. Each method estimates a dilution factor in a different way.
Fourteen urine-like samples (7 and 7), 60 metabolites. Each sample was diluted by a different amount, and group B is on average more dilute. Six metabolites truly rise in B. Choose a method and see whether it recovers the real dilution and what it does to the results.
| Method | Factor for sample i | Strength | Weakness |
|---|---|---|---|
| Total sum | Sum of all feature intensities | Simple, widely used for NMR and MS | One abundant feature that changes drags all others the opposite way (closure) |
| Median | Median intensity of the sample | Not moved by a few abundant features | Needs most features detected in all samples |
| IQR | Interquartile range of the sample | Robust, uses the spread as well as the centre | Poor when many values are missing or the distributions differ in shape |
| PQN | Median of the quotients of the sample against a reference spectrum, after sum normalization | Designed for dilution, robust to large changes in a few signals | Assumes most signals do not change |
| Internal standard | Signal of a compound added at a known amount | Corrects extraction and injection variation | Does not correct biological dilution, because it is added after the sample is collected |
Creatinine is sometimes used to normalize urine. It varies with age, sex, muscle mass and kidney function, so it is not a neutral reference. Check that your design allows it before using it.
In counts and intensities the spread grows with the mean. A transformation makes the spread similar across the range, so low and high features can be compared fairly.
300 simulated features, abundance from 1 to 100,000 counts, 10 samples each. Every point is a feature, with its standard deviation after the transformation. A good transformation makes the cloud flat.
| Transformation | Handles zeros | What it does | Typical use |
|---|---|---|---|
| log2(x + c) | With a pseudocount | Makes fold changes additive; the pseudocount sets how low counts are treated | Intensities, log-CPM, plots |
| Square root | Yes | Stabilizes Poisson noise only; weaker than log | Counts with low overdispersion |
| Generalized log, arcsinh | Yes | Linear near zero, logarithmic for large values | Mass spectrometry, cytometry, the basis of VSN |
| VST and rlog (DESeq2) | Yes | Uses the fitted mean and dispersion relationship of the data | PCA and clustering of RNA-seq |
| voom (limma) | Yes | Log-CPM plus precision weights from the mean and variance trend | Differential expression with linear models |
A transformation changes the scale on which a difference is measured. After a log2 transformation a difference of 1 means two-fold. Report results on the scale you used.
Multivariate methods such as PCA treat large variances as important. Scaling decides which features get a voice.
Forty samples, 12 features. Three abundant features vary together because of the measurement (orange). Seven scarce features carry the real group difference (blue). Two scarce features are only noise (grey). Choose a scaling and watch the PCA.
| Scaling | Formula | Effect | Caution |
|---|---|---|---|
| Centring | x - mean | Moves every feature to zero mean. Keeps the original scale | Large features still dominate |
| Autoscaling (z-score) | (x - mean) / s | Every feature has variance 1 | Inflates noise in features that are barely measured |
| Pareto | (x - mean) / sqrt(s) | Reduces the weight of large features without removing it | A compromise, common in metabolomics |
| Range | (x - mean) / (max - min) | Every feature spans a similar range | Very sensitive to outliers |
| Vast | (x - mean) / s x (mean / s) | Autoscaling, weighted toward stable features (low variation) | Needs positive, reliable means |
| Level | (x - mean) / mean | Shows relative change | Needs positive means |
Scaling changes the weight a feature has in PCA, clustering and machine learning. It does not change a univariate t-test on the same feature, because the group difference and its spread are scaled together.
Instruments drift. Pooled quality control samples, measured at regular intervals, show how much, and let you correct for it.
60 injections, every third a pooled QC (dark dots). Group B truly produces 12 percent more signal and was run later on average. Turn on the signal drift, then correct with a LOESS curve fitted to the QCs. Half of the QCs fit the curve and half are held out to check it.
| Problem | Tool | Needs | Risk |
|---|---|---|---|
| Drift within a run | QC-based LOESS (QC-RSC), SERRF | Pooled QCs spread across the run | Too few QCs, or a span that is too small, fits noise |
| Batches that differ in mean or spread | ComBat, removeBatchEffect, RUV | Batch labels, or control features | If batch and group are the same, the group effect is removed too |
| Batch as part of the analysis | Include batch in the model (Episode 11) | A balanced design | Often the safest choice for testing |
Correct drift with the QCs and test with the samples. Never judge a correction by the QCs that were used to fit it; hold some QCs back, as the lab does.
Every method in one table. Filter by omics, by the level it works at, or by what you need it to adjust.
Pick a kind of data and a problem. The table keeps only the methods that apply, and says what each one changes and what it leaves untouched.
There is no best normalization, only the one whose assumption your data satisfy. Describe your study and get a starting pipeline with the reason for every step.
Choose the data, the goal and anything that applies. The steps follow the usual order: filter, correct, normalize, transform, scale. Episode 5 explains why the order matters.
Every method assumes something stays constant: the total, most features, the distribution, a spiked standard. Check yours.
Plot the distributions, a PCA and the QCs before and after. If normalization added structure, stop.
In prediction, learn scaling factors inside each cross-validation fold, so the test samples never influence them.
Name the method, the software version and the order in your methods section, so someone else can repeat it.
Tick as you go. Your ticks stay in this browser.
This series is built from the questions researchers actually ask. Tell us where you got stuck and what you want next. It takes under a minute.
Turn on JavaScript to send your questions, or write to wael@omvitalis.com.
Primary sources for the methods in this guide. Check each DOI against the publisher page before you cite.