Why Validate a Tissue Niche Across Assays?
A tissue niche is a localized microenvironment characterized by its cells, molecular signals, and spatial relationships. Spatial transcriptomics can identify localized RNA patterns, while spatial proteomics measures proteins in tissue context. Agreement between the assays can strengthen a niche interpretation, but it is not automatic: the technologies may measure different molecules, have different sensitivities, and resolve tissue at different spatial scales.
The practical question is therefore not whether RNA and protein values are numerically identical. It is whether measurements that refer to corresponding tissue locations support a coherent interpretation. A candidate neighborhood might be marked by a transcript associated with a cell type and a protein found in or near that cell type. The evidence is stronger when the neighborhood assignment and the marker pattern agree in matched tissue, and when that agreement is robust to reasonable changes in alignment and analysis choices.
The spatial unit must be stated explicitly. A transcriptomic spot may contain several cells, whereas a protein assay may report measurements for individual segmented cells. Treating these as equivalent observations without an alignment rule can produce misleading comparisons. If the resolutions differ, a defensible approach is to map cells or spots to a common region or aggregate measurements to a shared neighborhood scale. The mapping should preserve the tissue context needed for the scientific question.
A Practical Cross-Assay Validation Workflow
Let \(i\) index matched spatial units, \(g\) an RNA feature, and \(k\) a protein feature. Write \(R_{ig}\) and \(P_{ik}\) for appropriately processed measurements, and let \(Z_i\) indicate whether unit \(i\) belongs to a candidate niche. The values should be normalized within their respective assays as appropriate; RNA counts and protein intensities generally have different scales and should not be compared as if they shared units.
Register tissue images or define a mapping from each assay's units to a common spatial region. Record uncertainty in the registration and avoid claiming cell-level matching when only region-level correspondence is available.
Define which units constitute the candidate neighborhood and identify the RNA and protein features expected to support it. If the same data were used to select the niche and markers, seek validation in held-out markers, sections, or specimens.
Compare corresponding signals or niche assignments using measures suited to the question, such as correlation, mean contrasts, or set overlap. Report the spatial unit and the number of independent specimens.
Inspect registered images and spatial maps, examine alternative aggregation choices, and use a null comparison that preserves relevant spatial structure when making an inferential claim.
For paired continuous measurements, a common descriptive summary is the Pearson correlation:
This measures linear association across matched units, not equality of the measurements. It is meaningful only when the matched pairs are valid and the values vary enough to calculate the denominator. A high correlation can coexist with a systematic offset, and a low correlation can arise because the assays measure related but non-identical biology.
Worked Examples
Worked Example: Correlation Across Matched Regions
Suppose a hypothetical tissue section is summarized into six matched regions. For an RNA marker, the normalized regional values are \(2, 3, 4, 5, 6, 7\); for a related protein, the normalized values are \(1, 2, 2, 4, 5, 6\). Assume the regions were registered consistently and that each value summarizes the same region in both assays.
The means are \(\bar{R}_g=4.5\) and \(\bar{P}_k=20/6=3.333\). The sum of cross-products of deviations is \(18\); the sums of squared deviations are \(17.5\) for RNA and \(19.333\) for protein. Thus,
The normalized RNA and protein values show a strong positive linear association across these six regions. This supports spatially concordant variation in this example, but it does not establish that the assays are interchangeable, that the protein is produced by the same cells as the RNA, or that the association generalizes to other specimens. With only six regions, the value is primarily descriptive.
Worked Example: Overlap Between Two Niche Assignments
In a hypothetical set of 20 common spatial units, an RNA-based analysis assigns 12 units to a candidate niche, and a protein-based analysis assigns 10 units to a corresponding neighborhood. Eight units are in both sets. The Jaccard index is the intersection size divided by the union size:
Thus, 8 of the 14 units assigned by at least one assay are shared. The index gives a concise summary of label overlap, but it does not correct for chance overlap, spatial dependence, or the fact that one assay may systematically define larger neighborhoods. It should be reported alongside the set sizes and a spatial map, not as a stand-alone proof of validation.
Worked Example: Testing Overlap Against a Spatial Null
Continue with a hypothetical analysis in which the observed overlap is eight units. To ask whether this overlap exceeds what might occur under spatially structured chance alignment, the analyst permutes one assay's niche labels within pre-specified tissue blocks, preserving the number of niche labels in each block. Suppose 999 such valid permutations produce overlap counts, and 24 of them are at least eight.
For a one-sided empirical test, use the observed count plus one in both numerator and denominator:
Under the specified permutation scheme, the observed overlap is unusual relative to the null alignments, with an empirical p-value of 0.025. This is evidence against that particular null of no alignment conditional on the preserved block structure. It does not show that the niche causes a molecular change. The conclusion also depends on whether the within-block permutations are defensible and whether the candidate niche was selected independently of the tested overlap.
Justifying Aggregation and Choosing Comparisons
Aggregation can make cross-assay comparisons more stable when individual spatial units are noisy. Suppose measurements in a region are \(Y_j=\theta+\epsilon_j\), where \(\theta\) is the region-level signal and the errors \(\epsilon_j\) have mean zero, variance \(\sigma^2\), and are independent. The regional average has variance:
Averaging \(m\) independent measurements therefore reduces the noise variance. But spatial measurements are often correlated. If every pair of errors has covariance \(c\), the variance becomes \(\sigma^2/m+(m-1)c/m\), so aggregation does not eliminate shared spatial noise. Moreover, aggregation can hide fine-scale differences. The scale should match the biological claim, and conclusions should be checked at plausible alternative scales.
The main summaries answer different questions. Correlation describes whether paired continuous signals vary together; it does not assess agreement in absolute magnitude. A mean contrast compares signal levels between niche and non-niche regions, but it does not establish spatial overlap of the assignments. The Jaccard index summarizes overlap of two sets, while a spatially constrained permutation asks whether that overlap is greater than expected under a specified null. None is universally best: choose the summary that corresponds to the claim being evaluated.
Assumptions, Diagnostics, and Failure Modes
The analysis depends first on valid spatial correspondence. Registration errors can pair neighboring but biologically distinct regions, lowering apparent agreement or creating false overlap. Review overlays and boundary cases, and repeat the analysis after modest changes to alignment or region definitions. If the assays were performed on serial sections rather than the same section, describe the resulting correspondence as region-level, not cell-level.
A second issue is dependence. Nearby spots or cells are not generally independent replicates. A test that treats thousands of spatial units as independent may report very small standard errors even when the study contains only a few specimens. Permutations should preserve the relevant spatial structure, and biological specimens—not individual spots—remain the basis for generalization across tissues. With few specimens, report specimen-specific patterns and avoid overstating population-level evidence.
Inspect spatial maps and distributions as well as summary statistics. A high correlation can be driven by a broad tissue gradient rather than a niche-specific relationship. A few extreme regions can dominate Pearson correlation; scatterplots and robust or rank-based summaries can help diagnose this, although they do not repair bad matching. Check that each signal varies, that labels are not dominated by one block or specimen, and that the result is not dependent on one threshold or one chosen marker.
A further failure mode is circular validation: identifying a niche from a marker and then using that same marker to claim independent support for the niche. Separate discovery from validation where possible, for example by testing a pre-specified protein marker or an independent tissue section. If many candidate markers are screened, account for the multiplicity or clearly label the findings exploratory. Finally, RNA and protein can genuinely disagree because of translation, degradation, transport, secretion, or differences in assay sensitivity. Discordance is a result to investigate, not automatically evidence that one assay is wrong.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Why is a high RNA–protein correlation not evidence that the two assays measure equal values?
- Two niche assignments contain 9 and 11 units, with 6 units in common. Calculate their Jaccard index.
- Why might unrestricted permutation of niche labels be unsuitable when the labels occur in spatially clustered regions?
- Give one reason to aggregate spatial units and one reason to check results at a finer scale.
- An RNA marker was used to define a niche and is then used as the sole evidence that the niche is validated. Identify the problem and one remedy.
Solutions. (1) Correlation measures linear co-variation, not equality; a constant offset or different scale can leave correlation high. (2) The union contains \(9+11-6=14\) units, so \(J=6/14=0.4286\). (3) Unrestricted permutation destroys spatial clustering and may create an unrealistic null distribution; permuting within defensible spatial blocks can preserve some structure. (4) Aggregation can reduce independent measurement noise, but can hide fine-scale heterogeneity or boundaries. (5) This is circular validation because the same evidence selected and confirmed the niche; test a pre-specified independent marker, held-out section, or independent specimen instead.