Building Marker Panels for Cell-Type Annotation

Why Build a Marker Panel?

A list of genes associated with a cell type is not automatically a useful annotation panel. A long list may contain several genes that convey nearly the same information, while omitting a type that is important in the intended analysis. A marker that distinguishes one type in one tissue may also be shared with a close relative or be weakly supported in another tissue. Panel construction is the deliberate selection of a small, interpretable set of genes from established cell-type signatures.

The panel is built for a specified set of labels and an intended tissue scope. The chosen label granularity matters: a gene may distinguish two broad lineages but not two subtypes within one lineage. The earlier material on cell-type annotation as a statistical decision and choosing label granularity with cell ontologies establishes why those choices should be made before marker selection. Here, the task is to choose genes that represent and distinguish the already chosen labels—not to infer a label for an individual cell.

Definition: A marker panel is a selected set of genes, \(S\), intended to provide interpretable evidence about a specified collection of cell types in a stated tissue scope. Coverage describes how well the selected genes represent the chosen types. Specificity describes how well a gene’s support for one type distinguishes it from the relevant competing types. Both are relative to the labels, tissues, and evidence source used to construct the panel.

Set the Scope Before Selecting Genes

Begin with the cell types the panel must represent and the tissues in which it is intended to be used. For each type, identify its closest relevant alternatives in the chosen ontology. These are the comparisons that matter most: a marker shared by several distant cell types may still be informative for a broad label, but a marker shared by two sibling types will not by itself resolve those siblings.

The candidate list should come from established signatures whose evidence and context can be checked. Such signatures may come from prior marker analyses or curated references. The earlier material on marker inference explains why post-clustering evidence needs careful treatment; panel construction should not turn a selected marker list into a claim of independent validation. Record the source, tissue, label definition, and evidence for each candidate. Treat the list as evidence to assess, not as a universal catalogue of markers.

A simple, transparent specificity check uses a gene’s detection fraction: the fraction of cells in a type with detectable expression. For a candidate gene \(g\), target type \(k\), and close alternatives \(j\), define the descriptive contrast

D_{gk}=p_{gk}-\max_{j\in\mathcal{N}(k)}p_{gj},

where \(p_{gk}\) is the detection fraction in type \(k\), and \(\mathcal{N}(k)\) is the set of relevant competing types. A larger positive value indicates greater separation in detection from the most similar alternative in the evidence being considered. This is a screening summary, not a universal test or a guarantee that the gene will distinguish every cell. If detection fractions are unavailable, use documented signature support and comparisons from the source rather than pretending that a quantitative contrast can be calculated.

This comparison assumes that the cell-type labels and tissue context are meaningfully comparable, that detection was measured consistently, and that the candidate evidence is adequate for the intended use. A difference in detection fractions can reflect technical as well as biological variation. It does not establish causation, and a gene with low detection in a reference may still be useful in another context.

Worked Example: Comparing a Candidate With Its Closest Alternatives

Suppose an established, hypothetical tissue-specific signature reports detection fractions for a candidate gene in three related cell types: type A, its close relative B, and type C. The fractions are 0.72, 0.31, and 0.18, respectively. For annotating A, suppose B and C are the relevant alternatives.

The largest competing fraction is \(\max(0.31,0.18)=0.31\). The descriptive contrast is therefore

D_{gA}=0.72-0.31=0.41.

In this hypothetical evidence, the gene is detected in 41 percentage points more of type A than of the most frequently detected alternative, type B. That supports considering it as a candidate for A. It does not mean the gene is absent from B, nor does it show that the gene alone can reliably assign labels. The conclusion is limited to this signature source and tissue.

Choose a Compact Panel by Marginal Coverage

A useful selection rule makes the goal explicit. Let \(K\) be the chosen set of cell types, and let \(w_k\) be the importance assigned to type \(k\), with \(w_k\geq 0\). For a simple panel, set \(a_{gk}=1\) when gene \(g\) has sufficiently supported, relevant marker evidence for type \(k\), and \(a_{gk}=0\) otherwise. This indicator should be based on the evidence and comparisons specified in advance; it is not a claim that expression is biologically impossible when \(a_{gk}=0\).

One possible coverage objective is the total importance of types represented by at least one selected marker:

C(S)=\sum_{k\in K} w_k\,\mathbf{1}\!\left\{\sum_{g\in S}a_{gk}>0\right\},

where \(S\) is the selected panel and \(\mathbf{1}\{\cdot\}\) equals 1 when its condition holds and 0 otherwise. The compact-panel problem is to choose \(S\) with no more than \(B\) genes while making \(C(S)\) large, subject to requiring that selected candidates pass the chosen evidence and specificity checks. The budget \(B\) reflects practical limits such as assay capacity or interpretability.

The marginal gain from adding a candidate gene \(g\) to the current panel \(S\) is

\Delta(g\mid S)=C(S\cup\{g\})-C(S) =\sum_{k\in K}w_k\,\mathbf{1}\!\left\{a_{gk}=1,\ \sum_{h\in S}a_{hk}=0\right\}.

This follows because adding \(g\) changes the indicator for type \(k\) only if \(g\) supports that type and no already selected gene does. Thus a gene that only repeats support for types already covered has zero gain under this particular objective. Repeated markers may still be desirable for robustness, but that is a different objective and should be stated rather than smuggled into a simple coverage score.

A practical greedy procedure repeatedly selects the eligible gene with the largest marginal gain, recalculating gains after each selection, until the budget is reached or no gene adds coverage. It is easy to audit and prevents a broad signature from dominating solely because it contains many correlated candidates. It does not automatically optimize every desirable property: if robustness, assay cost, or tissue portability matters, include explicit constraints or use a different stated criterion.

Worked Example: Selecting Three Genes for Type Coverage

Consider four hypothetical types with equal importance and a budget of three genes. After applying predeclared evidence and specificity checks, suppose the candidates have these support patterns: gene \(g_1\) supports types A and B; \(g_2\) supports A; \(g_3\) supports C; \(g_4\) supports C and D; and \(g_5\) supports D. These are hypothetical support indicators, not claims about real genes.

Initially, each gene’s marginal gain equals the number of newly covered types: \(g_1\) and \(g_4\) each gain 2, while \(g_2\), \(g_3\), and \(g_5\) each gain 1. Choose \(g_1\) first, breaking the tie with \(g_4\) by a predeclared rule such as stronger evidence in the intended tissue. Types A and B are now covered.

Recalculate. The gain for \(g_2\) is 0 because A is already covered. The gains for \(g_3\), \(g_4\), and \(g_5\) are 1, 2, and 1, respectively. Choose \(g_4\), which adds C and D. All four types are covered with two genes. The third slot is not needed to improve the stated coverage objective. If the panel must contain three genes for a separately stated robustness goal, the choice should be made using evidence for that goal; adding \(g_2\) merely because it is available does not increase coverage.

This example also shows why the budget is a maximum, not a requirement to fill every slot. Under a coverage-only objective, stopping early is sensible once every required type is covered.

Account for Tissue Context

A panel assembled from signatures in one tissue should not silently be treated as universal. Marker expression can vary across tissues, conditions, donors, and measurement platforms. Some differences reflect biology; others may reflect sampling or technical properties of the data. The relevant question is not whether a gene is a marker in the abstract, but whether the evidence supports its intended role in the target tissue scope.

When a panel must cover several tissues, index support by tissue, \(a_{gkt}\), and state which type–tissue combinations are required. One option is to define coverage over those combinations and assign weights \(w_{kt}\). Another is to require each selected marker to have evidence in every intended tissue. These choices answer different questions: weighted coverage allows a strong contribution in one tissue to offset weaker representation elsewhere, whereas a per-tissue requirement does not. Neither should be chosen without reference to the intended use.

Worked Example: Comparing Tissue-Aware Panel Options

Suppose types A and B must be represented in tissues T1 and T2, with equal importance for all four type–tissue combinations. A hypothetical candidate \(g_X\) has established support for A in T1 and B in T1, but no documented support in T2. Candidate \(g_Y\) supports A in both tissues, and \(g_Z\) supports B in both tissues.

If the target is T1 only, \(g_X\) covers two type–tissue combinations, while either \(g_Y\) or \(g_Z\) covers one. For a two-tissue target, selecting \(g_Y\) and \(g_Z\) covers all four combinations: A–T1, A–T2, B–T1, and B–T2. Selecting \(g_X\) with \(g_Y\) covers A in both tissues and B in T1, leaving B–T2 without documented support. Thus the panel that looks broader in T1 is not necessarily the better panel for a multi-tissue application.

“Without documented support” is the appropriate conclusion for the missing combination. It is not proof that \(g_X\) is biologically absent from B in T2. If the intended application includes T2, the evidence gap should be resolved or reported, not converted into an unverified negative claim.

Diagnostics, Alternatives, and Limitations

Before finalizing the panel, inspect the selected genes against the intended labels and tissues. A compact coverage table makes omissions visible: list each required type–tissue combination and the selected genes that support it. Review the close alternatives for every claimed type-specific marker. If a candidate’s apparent specificity depends on comparing a target with distant types while omitting its closest sibling, the comparison is not adequate for the annotation task.

Redundancy: several selected genes may support exactly the same types. Under a coverage objective, later additions with no marginal gain are redundant; retain them only if a separate robustness or assay-design rationale is documented.
Uneven coverage: a large number of markers for common types can obscure a type with little evidence. Use stated type weights or minimum coverage requirements that reflect the intended application.
Weak tissue transfer: a gene supported in one tissue may lack evidence in another. Mark unsupported combinations explicitly and avoid assuming portability.
Ontology mismatch: a panel selected for broad labels may not distinguish finer subtypes. Revisit the target granularity rather than treating a broad-lineage marker as subtype-specific.
False certainty from a binary table: coding a candidate as support/no support simplifies selection, but discards strength and uncertainty. Preserve the underlying evidence and use the binary rule only as a transparent selection aid.

The coverage objective is not the only reasonable approach. A manual expert panel can incorporate biological knowledge that a simple indicator table misses, but decisions may be less reproducible unless the criteria are recorded. A ranked list of genes preserves more information about evidence strength, but does not by itself specify how many genes to select or how to balance types. A full optimization procedure can incorporate costs, tissue constraints, and redundancy penalties, but added complexity can make the result harder to explain and more sensitive to uncertain inputs. In each case, the panel should be judged against its stated purpose rather than by size alone.

A common mistake is to reason, “This gene is strongly associated with type A, so it is a good marker for distinguishing A from B.” Association with A does not establish specificity against B. The correction is to compare the gene with the closest relevant alternatives, check the tissue context, and state exactly which types the evidence supports. A panel is a structured summary of evidence, not a guarantee that every selected gene is exclusive to one type or that every cell can be assigned from the panel.

Key takeaway: Build panels for a fixed set of labels and tissues. Select candidates from established evidence, compare each marker with close alternatives, and make coverage, budget, and tissue requirements explicit. Keep unsupported comparisons visible rather than turning them into claims of absence.

Check Your Understanding

Try each problem, then compare your reasoning with the solutions.

A candidate gene has detection fractions 0.64 in target type A and 0.27 and 0.39 in its two closest alternatives. What is its descriptive contrast against the closest alternative, and what does it mean?
Under the stated coverage objective, a type already has one selected supporting gene. What is the marginal coverage gain from adding another gene that supports only that type? Give one reason it might still be retained.
A panel covers every type in tissue T1 but has no documented signature evidence for one type in T2. What conclusion is justified, and what conclusion is not?
Why can a marker that distinguishes two broad lineages be inadequate for distinguishing two sibling subtypes?

Solutions. (1) The largest alternative fraction is 0.39, so \(D=0.64-0.39=0.25\), a 25-percentage-point difference in this evidence source; it does not establish exclusive expression or reliable cell-level assignment. (2) Its gain is 0 because that type is already covered. It could be retained for a separately justified robustness objective, provided its evidence supports that role. (3) The panel has documented coverage in T1 and an evidence gap for that type in T2; one cannot conclude that the marker or type is biologically absent in T2. (4) The sibling types may share the broad-lineage marker, so it may support their common parent while failing to separate them; specificity must be checked against the closest alternatives at the intended label granularity.