Why Marker and Reference Evidence Can Disagree
A cell can receive a transferred label from a reference atlas while expressing a marker pattern that appears inconsistent with that label. The disagreement is a diagnostic signal, not proof that either source is wrong. A marker may be shared across cell types, weakly detected, or measured in a state where its expression changes. A reference may not represent the target tissue, condition, or subtype well. Or the two sources may be describing different levels of detail: one may support a broad lineage while the other proposes a fine subtype.
The task here is to locate which of these explanations is plausible before interpreting the label. The checks are complementary: marker specificity asks how discriminating the observed genes are; reference coverage asks whether the reference represents the target cells; and granularity asks whether the evidence supports the label at the level assigned. None of these checks turns a marker list or a transfer score into certainty.
The checks rely on several conditions. Marker comparisons should use appropriately normalized data and be interpreted with awareness of dropout and technical effects; an undetected transcript is not necessarily biologically absent. The reference labels must be sufficiently reliable for the comparison being made. Coverage distances or scores must be calculated in a representation and metric that are meaningful for the data, with any cutoff calibrated using reference validation data rather than chosen after inspecting the disputed cells. Finally, the question of granularity requires a label hierarchy: a subtype label should imply membership in its broader parent class.
Specificity Matters More Than a Marker Name
A gene can be a useful marker because it is commonly expressed in a cell type, yet still be poor evidence for distinguishing that type from its nearest alternatives. Sensitivity measures how often a marker is detected among cells in the proposed class. Specificity measures how often it is not detected among cells outside that class. Both are conditional quantities; neither by itself gives the probability that a cell belongs to the class given that the marker is detected.
Here \(M^+\) denotes detection of a marker, \(M^-\) its non-detection, and \(C\) membership in the proposed class. The probability of class membership given a positive marker, \(P(C\mid M^+)\), also depends on how common the class is before considering the marker. This distinction explains why a marker that is sensitive but shared with other classes can provide only modest support.
Worked Example: A Sensitive Marker That Is Not Highly Specific
Suppose, hypothetically, that a reference set has 30 cells labelled as a candidate class and 70 cells labelled as other classes. A gene is detected in 27 of the candidate cells and 18 of the other cells. Calculate its sensitivity, specificity, positive likelihood ratio, and the fraction of marker-positive cells that belong to the candidate class in this reference set.
Sensitivity. The marker is detected in 27 of 30 candidate cells: \(27/30=0.900\), or 90.0%.
Specificity. It is absent in \(70-18=52\) of the 70 other cells: \(52/70=0.7429\), or 74.3%.
Positive likelihood ratio. A positive result is \(0.900/ (1-0.7429)=0.900/0.2571=3.50\) times as frequent in candidate cells as in other cells. This is evidence for the candidate class, but it is not an overwhelming separation.
Positive predictive value in this set. There are \(27+18=45\) marker-positive cells, of which 27 are candidate cells, so \(27/45=0.600\), or 60.0%. In this hypothetical reference set, a positive marker alone therefore does not distinguish the candidate class from alternatives with high certainty. The fraction depends on the class prevalence in this set and should not automatically be transferred to a differently composed target sample.
The prevalence dependence can be made explicit with Bayes’ rule. If the marker’s sensitivity and specificity apply to the population under consideration, then the posterior odds after observing a positive marker equal the prior odds multiplied by the positive likelihood ratio:
For the example, the prior probability of the candidate class is \(30/100=0.30\), so its prior odds are \(0.30/0.70=0.4286\). Multiplying by the likelihood ratio \(3.50\) gives posterior odds of \(1.50\), which corresponds to a probability of \(1.50/(1+1.50)=0.600\). This agrees with the direct count. The calculation depends on the assumed class prevalence and on the sensitivity and specificity being applicable; it is not a universal probability for a marker-positive target cell.
Check Whether the Reference Covers the Target Cells
A transferred label is a comparison with the available reference, not a guarantee that the reference contains the correct target class. A closed-set classifier must select among its represented labels even if the target cell is unlike all reference cells. A high class score can therefore coexist with poor coverage. The earlier material on detecting cell types missing from the reference and recognizing reference distribution shift motivates treating coverage as a separate question from the assigned label.
One practical diagnostic is to compare target-to-reference distances with distances observed for held-out reference cells under the same representation, metric, and preprocessing. For example, a cutoff might be the 95th percentile of held-out distances for a relevant class. Such a cutoff is an empirical reference for unusual distances, not a universal boundary between covered and uncovered cells. The fraction of target cells beyond it is descriptive evidence of mismatch; it is not, by itself, a formal test that the cell type is absent.
Worked Example: A Confident Label With Weak Coverage
In a hypothetical analysis, a reference classifier assigns 20 target cells to the same broad lineage, with a reported class score of 0.82 for the cluster. The score is not known to be a calibrated probability. On a shared embedding, the nearest-reference distance for each target cell is compared with held-out reference distances for that lineage. The held-out 95th-percentile cutoff is 1.20 distance units. Only 6 of the 20 target cells have distances at or below 1.20; the other 14 exceed it. Assume that target and held-out cells were processed comparably and that the distance metric was selected before examining this cluster.
The fraction within the reference’s empirical distance range is \(6/20=0.300\), or 30.0%; the fraction beyond it is \(14/20=0.700\), or 70.0%. Thus the transfer score indicates that the model prefers this lineage among its available labels, while the distance check indicates that most target cells are poorly represented relative to held-out examples of that lineage.
These results do not establish that the lineage assignment is false. They do show that the score should not be read as evidence of good reference coverage. The next interpretation should consider whether the reference lacks a relevant subtype, state, or tissue context, and whether the distance result could instead reflect technical or preprocessing differences.
Coverage diagnostics can fail when the target and reference were processed differently, when the embedding is distorted by batch effects, or when the distance metric is dominated by irrelevant features. A cutoff calibrated on one lineage may also be inappropriate for another if their within-class variability differs. Examine distances by class and, where possible, use held-out cells processed in a way that resembles the target. Do not infer missing cell types solely from a large distance: technical mismatch and biological novelty can produce similar patterns.
Match the Strength of the Label to Its Granularity
A broad label can be well supported even when its subtypes are not distinguishable. For example, several genes may support a T-cell lineage without separating memory, regulatory, and other T-cell subtypes. Reference scores can likewise favor a parent lineage while splitting among its children. The appropriate response to that pattern is not to force a fine label from whichever child has the largest score. It is to ask what level the evidence can actually distinguish.
This check assumes that the proposed labels form a meaningful hierarchy and that the markers being used are relevant to the distinction in question. A marker associated with a parent lineage may confirm that lineage but provide no evidence between sibling subtypes. Conversely, a subtype marker observed in a few cells may be insufficient to assign the subtype to an entire cluster, especially when detection is sparse or the gene is not specific.
Worked Example: Evidence for a Parent Label but Not a Subtype
Consider 12 hypothetical cells from a cluster. The reference model reports scores of 0.87 for the broad T-cell lineage and, among its fine labels, 0.46 for CD4 memory, 0.41 for regulatory T cells, and 0.08 for CD8 T cells. These are model scores, not calibrated probabilities. Marker detection is as follows: CD3D in 10 of 12 cells, CD4 in 8 of 12, FOXP3 in 3 of 12, and IL2RA in 4 of 12.
CD3D is detected in \(10/12=0.833\), or 83.3%, of cells, providing supporting evidence for the broad T-cell lineage in this example. CD4 is detected in \(8/12=0.667\), or 66.7%, which is compatible with a CD4-associated population but does not by itself establish a specific CD4 subtype. The fine-label scores differ by only \(0.46-0.41=0.05\), and the two genes shown as associated with regulatory T cells are detected in only \(3/12=0.250\) and \(4/12=0.333\) of cells, respectively.
The evidence is therefore more coherent at the parent level than at the subtype level: the broad lineage score is high relative to the listed child scores, while the fine-label scores are close and the subtype-associated markers are not widespread in this small cluster. A defensible interpretation is that the cells support a T-cell label, while the available evidence does not clearly separate CD4 memory from regulatory T cells. The numbers do not prove that either subtype is absent; sparse detection, mixed cells, or an incomplete marker panel could contribute.
A Focused Diagnostic Sequence
When marker and transferred labels conflict, the following sequence keeps the checks distinct. It is not a universal scoring rule: each result must be interpreted using the assay, reference, and biological question at hand.
Compare marker expression in the proposed class with plausible alternatives. Separate evidence that supports a broad lineage from evidence that separates sibling subtypes.
Ask whether the relevant tissue, condition, state, and label classes are represented. Compare target-to-reference distances with an appropriate held-out reference benchmark where available.
Determine whether the evidence supports the assigned subtype, only its parent class, or neither. Do not treat a fine label as justified just because it is the top-scoring option.
Keep the marker result, transfer result, and coverage result conceptually separate. A disagreement can have more than one cause, and the available evidence may not identify a unique explanation.
Comparing the Main Evidence Strategies
| Strategy | What it can contribute | Main limitation |
|---|---|---|
| Marker-only interpretation | Connects expression patterns to biologically interpretable genes and may support a broad lineage. | Markers may be shared, state-dependent, or sparsely detected; a marker’s presence does not establish a label. |
| Reference-only transfer | Compares cells against a consistent set of reference labels and can provide a ranked assignment. | A model may choose among incomplete classes; a score does not guarantee coverage or calibration. |
| Joint diagnostic interpretation | Uses marker specificity, reference coverage, and label hierarchy to locate the source and level of a disagreement. | Evidence sources can share biases, and their apparent agreement is not independent confirmation. |
A common mistake is to count two correlated marker genes and a reference score as three independent votes for one label. That reasoning is not justified unless their dependence and calibration have been addressed. Genes in the same expression program may be correlated, and a reference model may use those same genes. The correction is to ask what each source contributes: whether the markers distinguish the proposed class, whether the reference covers the target, and whether the evidence resolves the label at the claimed granularity.
These checks have limits. Marker expression can change with activation or disease state; dropout can obscure genuine expression; reference distance depends on representation choices; and a cluster may contain more than one population or a continuum of states. A numerical score cannot replace biological context, and agreement between methods does not guarantee correctness if they rely on the same data or assumptions. The result of this tutorial’s diagnostic process is a more defensible interpretation of the disagreement, not an automatic final label.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A marker is detected in 80% of cells from a candidate class and 50% of cells outside it. What are its sensitivity and specificity? Why is sensitivity alone insufficient?
- A reference assigns a target cluster a high class score, but most cells are farther from that class than held-out reference cells. What does this establish, and what does it not establish?
- Markers support a broad lineage, while two sibling subtype scores are close. What label granularity is most defensible from those facts alone?
- Why can a marker-positive fraction from a reference set not automatically be interpreted as the probability that a target cell belongs to the class?
Solutions. (1) Sensitivity is 80.0%. Specificity is \(1-0.50=0.50\), or 50.0%. Sensitivity describes detection within the candidate class; it does not show how often the marker also occurs among alternatives. (2) The score shows the model’s preference among its available labels, while the distance comparison suggests weak coverage relative to the held-out reference benchmark. Neither result alone proves the label wrong or proves that the cell type is missing from the reference. (3) The broad lineage is more defensible; the close sibling scores do not resolve the subtype, so the evidence given does not justify choosing one fine label. (4) The fraction depends on the class prevalence in the reference set. A target sample with a different class prevalence can have a different \(P(C\mid M^+)\), even if marker sensitivity and specificity are unchanged.