Interpreting Cell-Specific Modality Weights

What a Cell-Specific Modality Weight Says

Weighted nearest neighbors (WNN) combine modality-specific neighborhood information for the same cells. The central interpretive feature is that the contribution of a modality can differ from cell to cell. A cell whose local RNA neighborhood is more informative than its protein neighborhood may receive a larger RNA weight; another cell may show the reverse pattern. These weights describe the relative usefulness of the measured modalities for constructing local neighborhoods under the chosen representation and analysis settings.

This is a statement about the analysis, not a direct measurement of biological importance. A high protein weight does not mean that proteins are globally more important than RNA, that the cell contains more protein than RNA, or that protein caused the cell’s identity. It indicates that, for that cell and those processed data, the protein modality contributes relatively strongly to the WNN neighborhood construction.

For two modalities, RNA and protein, a cell with weights 0.8 and 0.2 has an RNA weight four times its protein weight on the normalized weight scale. It is not correct to say that there is an 80% probability that RNA is the true modality, or that RNA explains 80% of the cell’s biology. The normalization makes the weights relative: if one modality’s contribution rises, the other modality’s share can fall even if its own measurement has not changed.

The interpretation also depends on paired measurements. RNA and protein values should refer to the same cells, and each modality should have an appropriate representation for measuring neighborhoods. The representations and their scales should be suitable for the WNN procedure; a weight cannot correct an unsuitable embedding or invalid measurements. Weights are not hypothesis tests, confidence intervals, or estimates with a sampling error attached by default.

From Relative Reliability to a Weight

At a conceptual level, WNN assigns more influence to a modality when its local structure provides more reliable neighborhood information for a cell. The assessment is cell-specific: one modality can be useful for distinguishing one local group and less useful for another. In a common conceptual description, a modality’s local predictive performance is compared with that of the other modalities, and the resulting nonnegative scores are normalized to obtain relative weights. The exact score calculation and implementation details can vary; interpretation should follow the software and settings actually used.

The normalization step has a simple justification. Suppose the procedure produces nonnegative local reliability scores \(a_{im}\) for cell \(i\), with at least one score greater than zero. Define the normalized weight by dividing each score by their sum:

Each weight is nonnegative, and summing over modalities gives \(\sum_m \omega_{im}=\sum_m a_{im}/\sum_r a_{ir}=1\). Thus, normalization makes the weight a share of the combined reliability scores, not an absolute accuracy measure. If every score were zero, this expression would be undefined; a valid implementation must handle that case rather than interpret the resulting weights as meaningful.

In a two-modality setting, if RNA and protein have scores \(a_{iR}\) and \(a_{iP}\), respectively, their normalized weights are \(a_{iR}/(a_{iR}+a_{iP})\) and \(a_{iP}/(a_{iR}+a_{iP})\). The ratio of the two weights equals the ratio of the scores when both are positive. This is useful for describing relative contribution, but does not turn either score or weight into a probability of biological truth.

Interpreting Patterns Across Cell Populations

Weights can reveal that different modalities contribute differently across cell populations. For example, a surface protein may distinguish two closely related immune populations even when their RNA profiles are similar at the measured depth. Conversely, an RNA program may separate cell states that have similar measured antibody-derived tag profiles. Such patterns are plausible interpretations to investigate, not conclusions established by the weights alone.

A useful summary retains the cell-level nature of the quantity. Report distributions or summaries within biologically defined groups, alongside the number of cells and the procedure used to define those groups. A single average across all cells can conceal opposite patterns in distinct populations. If cell labels were assigned using the same combined embedding, the interpretation should also acknowledge that the group definitions and weight patterns are not independent pieces of evidence.

Assumptions, Diagnostics, and Common Pitfalls

A meaningful interpretation requires that the data and analysis support the comparison being made. At minimum, the modalities must be measured on the same cells being combined; the chosen features and representations must contain relevant variation; and normalization, scaling, and dimensional reduction must not make one modality’s distances dominate merely because of arbitrary scale. The paired-measurement structure established in earlier material on CITE-seq and Multiome is important: an apparent cell-specific advantage is not interpretable as a comparison of paired neighborhood information if the modalities belong to different cells.

Weights near zero or one deserve attention, but are not automatically errors. They may reflect genuine local differences in information, or may arise from sparse features, extreme measurement noise, batch effects, a poorly chosen embedding, or a modality that contributes little usable variation in a region of the data. Check whether extreme weights concentrate in a particular sample, batch, depth range, or small group. Compare weight distributions across relevant populations and inspect whether conclusions change under reasonable alternative feature or preprocessing choices. These checks do not substitute for a full diagnosis of multimodal neighborhoods; they help prevent an unsupported biological reading of the weights.

• Do not read the weight as a probability. A value of 0.8 is not an 80% chance that the modality is correct.
• Do not read a relative weight as absolute quality. A modality can receive most of the weight even if both modalities are weak, because one is merely less weak relative to the other.
• Do not equate weight with biological importance. The weight concerns local neighborhood construction under the analysis, not the causal role or overall importance of a molecular layer.
• Do not rely only on a global mean. A pooled summary can hide cell populations in which different modalities dominate.

A frequent mistake is to see a high protein weight in an immune population and conclude, “Protein determines the identity of these cells.” The first part of the reasoning may be supported if the protein weights are high and the analysis is appropriate; the causal or biological claim is not. A more defensible statement is: “In this analysis, protein features contributed more strongly than RNA features to the local WNN neighborhoods of these cells; this pattern is consistent with protein providing useful local discrimination.” The distinction keeps the interpretation tied to what the weights encode.

How Weighting Choices Differ

Cell-specific weighting is not the only way to combine modalities. Equal weighting gives each modality the same contribution for every cell. A global weighting scheme uses one set of relative contributions across all cells. A single-modality analysis uses only one representation. Each answers a different practical question.

These alternatives are not merely technical competitors. They make different assumptions about whether modality usefulness is uniform across cells. When cell-specific weights vary substantially, replacing them with a global average may erase meaningful local structure; when weights are unstable or driven by artifacts, flexibility can instead amplify noise. The comparison should therefore be guided by the scientific question and by the reliability of the measured modalities, not by a preference for the most complex combination.