Feature IDs, Annotation, and Cross-Layer Mapping

Why Feature Labels Need Interpretation

An omics feature table can contain rows labelled with gene symbols, stable accessions, probe IDs, protein accessions, or compound names. These labels are not interchangeable. A label may identify a biological entity, a particular assay feature, or only a candidate interpretation of a measured signal. Before comparing or combining features across assays, establish what each label denotes and how its annotation was produced.

The earlier material on feature granularity distinguishes measurement targets such as variants, transcripts, proteins, and metabolite signals. Annotation adds a layer of correspondence: it proposes which biological entity or entities a feature refers to. That correspondence can be unique, ambiguous, many-to-one, or absent. It also depends on the feature definition and the reference resources used.

Identifiers, Symbols, and Labels

A stable identifier, such as an accession from a genome or protein reference system, is intended to identify an entity more consistently than a descriptive name. “Stable” does not mean unchanging under every circumstance: records can be merged, retired, split, or revised as references are updated. An accession should therefore be interpreted together with its namespace and, when relevant, the reference release, genome build, or database version.

Gene symbols are compact and readable, but they are names rather than guaranteed permanent keys. Symbols can have aliases, can change over time, and may be interpreted differently across resources. A symbol is useful for communication, but a reproducible mapping generally needs the identifier namespace and reference version as well as the symbol. Similar-looking names should not be assumed to identify the same entity.

A probe ID identifies an assay feature on a particular platform. It is not itself a gene identifier. A probe can match more than one genomic region or gene, match a region that has been reannotated, or fail to map under a newer reference. Conversely, several probes may target different regions of the same gene. Transcript identifiers add another level of distinction: multiple transcripts can arise from one gene, so a transcript-level feature cannot automatically be treated as a gene-level feature.

Compound labels require similar care. A name may be a synonym, a broad chemical class, or a candidate identification rather than a unique chemical identity. A measured mass-spectrometry feature can reflect an ion, adduct, isotope, or unresolved signal, while a compound database record refers to a chemical entity. One entity may give rise to multiple features, and one feature may be consistent with several candidate compounds.

Representing a Mapping and Checking Its Direction

Let \(A\) be a set of features in one assay and \(B\) a set of entities or features in a reference system. Represent the proposed mapping as a relation \(R\subseteq A\times B\). A pair \((a,b)\in R\) means that feature \(a\) is linked to entity \(b\) under the stated annotation procedure. The mapping is directional in use: asking “which genes do these probes map to?” is different from asking “which probes map to this gene?”

For a feature \(a\in A\), define its mapping degree as the number of distinct targets linked to it:

Thus \(d(a)=0\) indicates no recorded target, \(d(a)=1\) indicates a unique target, and \(d(a)>1\) indicates an ambiguous one-to-many mapping in this direction. Many-to-one correspondence is detected by reversing the question and counting how many source features link to the same target. A mapping is bijective only if every feature on both sides has exactly one partner and no feature is left unmatched. In practice, many annotation tables do not meet those conditions.

Two useful summaries are the mapping coverage, the fraction of source features with at least one target, and the unique-mapping fraction, the fraction with exactly one target. With \(N=|A|\), these are:

These summaries answer different questions. A feature with three candidate targets contributes to coverage but not to the unique-mapping fraction. Report the denominator and mapping direction, since a percentage without them can be misleading.

Ambiguity, Incomplete Annotation, and Aggregation

Ambiguity and missing links have different meanings. An ambiguous mapping records more than one candidate target for a feature. An unmapped feature has no target recorded in the annotation being used. It may be unmapped because the feature is novel, the reference is incomplete, the assay feature is low-confidence, or the identifier namespace or version does not match. “Unmapped” is therefore a statement about the available annotation, not proof that the feature has no biological meaning.

Many-to-one mappings also need careful interpretation. Several transcript features can belong to one gene; several probes can target one gene; and multiple measured signals can be associated with one compound. These relationships do not automatically justify collapsing rows. Aggregation changes the unit represented by the table and can lose feature-specific variation. Before aggregation, one needs a scientific reason for combining the features and a rule appropriate to the measurement scale. For example, summing raw counts is not interchangeable with averaging normalized intensities, and neither operation resolves an ambiguous assignment by itself.

A cross-layer link can be even less direct. A gene may encode a protein, but a gene-level RNA measurement and a protein abundance measurement are different quantities with different assay behavior. A metabolite may be linked to a pathway or enzyme without being a direct counterpart of a gene or protein feature. Annotation can support a biological interpretation or define candidate correspondences; it does not establish equivalence, causal direction, or matched measurement units.

A Practical Annotation Audit

Before using mapped features in an analysis, inspect the annotation as data rather than treating it as a finished lookup. A concise audit makes the assumptions visible and helps identify errors that could otherwise be mistaken for biological results.

This audit assumes that the annotation file is available in a form that identifies its reference context and that the analysis preserves the original feature IDs. If either assumption fails, mapping rates can reflect inconsistent identifier handling rather than true annotation gaps. Keep the original IDs and annotation fields so a mapping decision can be revisited.

Alternatives, Diagnostics, and Common Pitfalls

There is no single mapping policy appropriate for every analysis. Retaining all candidate links preserves uncertainty but can make a feature contribute to multiple targets. Restricting an analysis to uniquely mapped features is straightforward to describe, but can discard substantial data and may disproportionately exclude particular feature types. Selecting one target per feature can simplify a table, but is defensible only when a prespecified evidence-based rule supports the selection. Aggregating multiple features to a shared target changes the measurement unit and requires a justified aggregation method.

Warning signs include unexpectedly low coverage, a large share of features with multiple targets, many duplicate target labels, abrupt changes in mapping rates between annotation releases, and identifiers that fail to match the stated namespace. Compare mapping summaries before and after any cleaning step. A surprising result may indicate a version mismatch or parsing error rather than an unusual biological pattern. These diagnostics concern annotation; they are distinct from diagnosing missing measured values in the feature table.

Common mistake: “Both rows have the same gene symbol, so they are the same feature and can be collapsed.” This reasoning confuses a shared annotation target with feature identity. Two distinct probes or transcripts can map to one gene while measuring different regions or molecular quantities. Keep their feature IDs distinct unless a scientifically justified procedure defines a target-level summary.