Gene-expression signatures

In one sentence

A gene-expression signature combines selected RNA measurements using defined rules to describe a biological pattern or predict a specified outcome.

The intuition

A weather index combines several readings into one useful number. You need to know its formula and the question it answers. The same applies to a gene-expression signature. The analogy has limits: RNA patterns depend on specimen mixture and assay preparation, and a biological label does not make a score a direct measurement of that process.

How it works

Start with ribonucleic acid (RNA) measurements for a chosen gene set. The algorithm specifies how those measurements are prepared, weighted, compared and combined. The reference data and software version are part of the definition. A gene list alone is not a reproducible test.

Different signatures answer different questions:

Kind	Question	Boundary
Pattern or module score	Do chosen messages show a specified pattern?	A summary of messages does not directly measure pathway flux
Subtype classifier	Which reference pattern is most similar?	A category depends on the classifier and reference
Outcome predictor	How does the model predict a defined endpoint?	Performance belongs to a population, endpoint and time horizon

For example, the original PAM50 work used a defined 50-gene method to assign intrinsic breast-cancer subtypes and develop particular risk models. It does not validate every later score carrying a similar label. Parker 2009.

Scoring methods also differ. One module-score implementation subtracts expression of matched control genes from the selected genes' average expression. Gene set enrichment analysis (GSEA) instead tests whether a gene set accumulates toward one end of a ranked list in its defined comparison. These outputs cannot share a cutoff just because they use the same genes. Official module-score documentation, Subramanian 2005.

The rules and reference are part of the test; a score's usefulness needs separate validation.

For a clinical predictor, the complete procedure should be locked: specimen requirements, measurement, preprocessing, algorithm and decision threshold are fixed before validation. Changing them creates a new version requiring evidence. Analytical repeatability asks whether the test measures consistently. Clinical validity asks whether its result relates to the intended outcome. Clinical utility asks whether using it improves a clinical decision or outcome. McShane 2013.

Why it matters in cancer

A bulk inflammatory signature can rise because there are more immune cells, because those cells change state, or both. Single-cell profiles can help examine origin, with their own sampling limits.

A stress score can support a hypothesis about a program. It cannot establish protein activation, metabolic rate or dependence on a drug target. Outcome prediction also needs calibration in the intended setting; the original treatment era and endpoint matter.

Assay card

Field	What to retain
Measures and how	A defined composite: obtain RNA data, apply fixed preprocessing, compute the specified algorithm and interpret against its reference
Input and tissue cost	The underlying RNA assay consumes material according to its protocol; rescoring an existing data table consumes no additional tissue, but may be incompatible with the model
Output and units	Named subtype, unitless score, rank or predicted probability; retain exact scale, model version, endpoint and horizon
Thresholds	The validated cutoff for that complete test and intended population; a percentile or another model's cutoff is not interchangeable
Failure modes	Missing genes, changed normalization, batch effects, specimen mixture and population shift
Limits	No automatic cell attribution, protein-function measurement or universal therapy-selection rule
Validation tier	Exploratory scores, research predictors and clinically validated tests are different; regulatory/laboratory status must be checked for the exact test and intended use

Common confusions

Gene list versus test: computation and reference define the output.
Association versus treatment prediction: a prognostic pattern need not identify who benefits from a therapy.
Higher score versus stronger dependence: the score summarizes measured messages under chosen rules.

Try it

A fictional laboratory replaces a signature's normalization method. It retains the old “high” cutoff and recommends a treatment. Is that justified?

Answer: No. First establish whether the new implementation reproduces the validated test. Treatment selection then requires evidence for that test, disease and setting. The old cutoff cannot supply the missing validation.

Explain it back

Complete: “To interpret this signature, I need the genes, _____ and _____.”

One possible answer: The complete algorithm and its validated reference/intended question.

Takeaway

Interpret a signature as a defined test answering a defined question, with evidence tied to that version and setting.

Sources and scope

Source check: October 10, 2026. Examples are fictional. Expert and learner review remain pending. The cited methods illustrate different algorithms; they are not an inventory of currently approved clinical tests.