Worked Example of Tissue-Specific Reference Mismatch

When the Reference Does Not Represent the Query Tissue

A cell-type annotation method can only choose among the labels it has been given. When a query comes from a tissue or cell state that the reference atlas represents incompletely, the method may assign it to the nearest available label even if that label is biologically wrong. A high prediction score does not by itself rule out this kind of reference mismatch: it may indicate only that one available label fits better than the other available labels.

The practical task is to assess whether the prediction is credible at the available level of detail. This tutorial focuses on combining prediction scores with marker evidence to decide whether to retain a broad lineage label, report a candidate identity with uncertainty, or use an unknown label. A mismatch is a possibility to investigate, not a conclusion to draw from one low score or one unusual marker.

What Prediction Scores Can and Cannot Establish

Let \(q_{ik}\) be the prediction score for query cell \(i\) and reference label \(k\), with \(K\) labels in the model. If the scores are probabilities over the reference labels, then \(q_{ik}\geq 0\) and \(\sum_{k=1}^{K}q_{ik}=1\). Even a calibrated model can be calibrated only with respect to the labels and conditions represented in its validation data. Calibration among known classes does not establish that the set of classes is complete.

This limitation follows directly from the closed-set setup. Suppose a model converts compatibility values \(L_k\) into scores proportional to \(L_k\pi_k\), where \(\pi_k\) is the reference prevalence for class \(k\). If a true class \(u\) is omitted, its compatibility value is not included in the normalization. The scores for the represented labels still sum to one:

Thus, the largest score can be substantial even if every represented label fits poorly relative to the missing class. These scores are useful for describing how the model distributes support across the available labels; they are not, on their own, probabilities that the true identity is among those labels. The introductory material on detecting cell types missing from the reference establishes why open-set evidence matters; here the emphasis is on evaluating that possibility in a tissue-specific case.

For a cluster \(C\) of \(m\) cells, a descriptive summary is the average score for label \(k\), \(\bar q_{Ck}=m^{-1}\sum_{i\in C}q_{ik}\). The average can be useful, but it must be read alongside cell-level scores. A cluster average may conceal a mixture—for example, half the cells may strongly support one label and half another. Inspect score distributions, top-label frequencies, and consistency across biological replicates or donors, rather than treating the cluster mean as a prediction for every cell.

A Tissue-Aware Assessment

A reference-mismatch interpretation is more persuasive when multiple kinds of evidence agree. First, check whether the query cluster is technically credible: assess basic quality-control measures, possible doublets, ambient RNA, and whether a batch effect could account for the pattern. Then ask whether the leading reference label is supported by its expected markers and whether the tissue context makes an omitted type plausible. Marker evidence should be interpreted as a coherent pattern, not as a single detected gene. Single-cell measurements are sparse, and marker expression can vary with state, depth, and assay.

A useful distinction is between the level of identity supported by the data and the most specific name one can imagine. If several independent markers support a parent lineage but evidence for a particular subtype is incomplete, retain the parent label and flag the fine identity as uncertain. If even the parent lineage is unsupported, an unknown label may be more honest. A biologically plausible candidate can be noted as a candidate without presenting it as a settled annotation.

Worked Examples

Why the Reasoning Works, and Where It Can Fail

A simple score property explains why maximum score alone cannot establish a match. Since the \(K\) scores sum to one, at least one must be at least \(1/K\); otherwise every score would be less than \(1/K\), and their sum would be less than one. This lower bound says nothing about biological correctness. For example, with four available labels, a maximum score of 0.40 is mathematically compatible with the true type being absent. More generally, renormalizing compatibility across only represented classes forces the model to distribute all its score among those classes.

This reasoning assumes that scores are meaningfully comparable for the model being used and that the query cells and markers have passed appropriate quality checks. It does not assume that a marker list is exhaustive, or that a particular score cutoff is universal. Thresholds for abstention or mismatch should be selected using validation data relevant to the tissue, reference, and assay. A score threshold validated on one tissue should not automatically be transferred to another.

Several alternatives help diagnose a suspected mismatch, but answer different questions. A calibrated classifier estimates reliability within the evaluated label system; it cannot guarantee that an omitted class is represented. An abstention rule can prevent an unsupported specific call, but abstention does not identify the missing type. A missing-class detector can provide additional evidence, but its performance depends on how it was validated and on the shift between reference and query. Marker-based annotation adds biological evidence, but can be misleading when markers are nonspecific, sparse, or affected by contamination. No one of these approaches replaces checking cell-level coherence and tissue context.

Warning signs include a top predicted label whose expected markers are consistently absent, a coherent set of markers not represented by any reference label, markedly different predictions across donors, or a cluster average that hides distinct per-cell groups. These signs have alternatives: low sequencing depth can cause marker dropout; a batch effect can change expression patterns; ambient RNA can create misleading marker detections; and a mixed cluster can combine distinct identities. If the pattern does not replicate across cells or donors, or disappears after quality checks, the evidence for a missing tissue-specific type weakens.

Common Mistake