A predicted biomarker is not a measured biomarker.

One protein can appear in three evidence layers

A protein can enter a research conversation as a laboratory measurement, a model prediction or an association with a health trait. Those statements can sound interchangeable: the protein was measured, the protein was predicted, or the protein was implicated. They are not the same claim.

An assay records a signal from a biological sample under a specific protocol. A genetic prediction model estimates a protein level from variants and relationships learned in a reference population. A downstream association analysis asks whether that predicted component is related to a trait. Each layer can be useful, but each answers a different question and inherits different uncertainty.

The practical teaching rule is simple: name the evidence layer before interpreting the result.

The 2022 proteomics study keeps the layers visible

The peer-reviewed Nature Genetics paper ‘Plasma proteome analyses in individuals of European and African ancestry identify cis-pQTLs and models for proteome-wide association studies’ analyzed 4,657 plasma proteins in 7,213 European American and 1,871 African American participants from the ARIC study, with replication involving 467 African American participants from AASK. The study identified many proteins associated with common variants in nearby cis regions, examined fine-mapping across ancestry groups and built genetic prediction models for proteins.

The paper then applied proteome-wide association analyses to serum urate and gout. That sequence matters. Measured protein data supported pQTL discovery and model development; the models supported prediction; predicted protein components supported downstream association analyses. The paper reports peer-reviewed findings from those analyses. It does not make every predicted protein value equivalent to a direct assay result in a new individual.

Evidence labels are a governance control

This distinction travels beyond genomics. An AI system summarizing biomedical evidence should preserve whether a value was observed, derived, predicted or associated. Collapsing those labels can create false precision: a model estimate may be presented as a direct measurement, or an association may be narrated as a clinical conclusion.

A useful evidence card can carry the layer, biological material, assay or model version, study population, ancestry context, outcome definition, uncertainty, validation setting and source. Provenance should also identify which transformation produced the value. That record does not decide whether a finding is clinically actionable. It helps a reviewer see what kind of evidence is actually present.

Teach the transformation, not only the result

A proposed classroom exercise would give learners short research statements and ask them to classify each as measured, genetically predicted or trait-associated. A second step would require the minimum metadata needed to interpret the claim. Compare accuracy before and after learners see a provenance card, while holding the examples and review time constant.

Measure classification accuracy, unsupported causal language, missing population context, confidence calibration and the time needed to reach the source. Include ambiguous statements that use words such as ‘level,’ ‘signal’ or ‘effect’ without naming the layer.

This exercise and evidence-card design are original teaching and governance proposals, not results from the 2022 study and not clinical guidance. The goal is narrower: when a biomarker claim moves through an AI-assisted workflow, can the evidence keep its identity?