Worked CITE-seq Analysis of RNA and Surface Proteins

From Paired Counts to a Joint CITE-seq Neighborhood

CITE-seq measures RNA and antibody-derived tags (ADTs) in the same cells. The pairing is valuable: a cell’s transcript profile and its measured surface proteins can be compared directly. But the two count matrices have different measurement properties. RNA counts reflect sequencing of transcripts; ADT counts reflect antibody binding and can include background signal. A useful joint analysis therefore does not treat the two raw matrices as interchangeable. It normalizes each modality appropriately, constructs representations on which distances are meaningful, and then asks how cells relate across both sources of evidence.

This tutorial follows a hypothetical analysis of 300 immune cells. The goal is to identify a population with a distinctive joint neighborhood structure and examine an apparent mismatch between a marker transcript and its surface protein. The analysis describes patterns in the measured data; a mismatch alone does not establish a regulatory mechanism, a change over time, or a causal relationship.

The basic assumptions for the workflow are that cell barcodes have been correctly matched across modalities; counts have passed appropriate cell- and feature-level quality checks; and both modalities are available for the cells included in the joint graph. The retained features and dimensions should represent interpretable cellular variation rather than being dominated by library size, low-quality cells, batch, or technical background. If cells have missing modalities, the distance below cannot simply be calculated for them as written; the missingness requires an explicit method rather than silently treating absent values as zero.

Normalize RNA and Protein Separately

For RNA, a straightforward library-size normalization divides each cell’s gene counts by its total RNA count and multiplies by a fixed scale factor. A log transformation reduces the influence of very large values. For cell \(i\) and gene \(g\), let \(Y_{ig}\) be the RNA count and \(L_i=\sum_g Y_{ig}\) the total over the selected RNA features. One possible normalized value is

This transformation adjusts for differences in total captured RNA counts; it does not make all genes equally variable, remove batch effects, or guarantee that the total RNA content of cells is biologically identical. The denominator must be defined consistently, and cells with extremely low \(L_i\) should be investigated rather than assumed to be reliable.

ADT counts require a separate treatment. As a simple relative normalization, the centered log-ratio (CLR) transform can be applied within each cell across the chosen antibody panel. With a pseudocount \(\delta>0\), define

where \(X_{ij}\) is the ADT count for cell \(i\) and antibody \(j\), and \(P\) is the number of antibodies in the panel. The result is a relative, centered value: it indicates whether an antibody’s transformed count is high or low compared with the other antibodies in that cell. It is not an absolute protein concentration. Choice of panel, pseudocount, and handling of background can affect the results. Isotype controls, negative controls, and the distributions of individual ADTs can help identify background, but CLR is not itself a background-subtraction model.

Before building a graph, inspect modality-specific distributions and technical covariates. RNA library size and detected genes can reveal low-quality cells; ADT distributions can reveal unusually high background, poor staining, or a small number of dominant features. Compare results across batches and, where available, controls. If a marker’s ADT values are not cleanly separated from background, a high/low label needs cautious interpretation. Do not apply an RNA normalization to ADTs merely because both are count matrices.

Define a Joint Distance Before Finding Neighbors

A common workflow reduces each normalized modality to a lower-dimensional representation, such as retained principal component scores. Let \(z_{ir}\) denote the RNA representation for cell \(i\) and \(u_{ip}\) its protein representation. The exact distance must be specified: using raw concatenated values can let one modality dominate because of its scale or number of features. In the following illustrative approach, the retained dimensions within each modality are scaled comparably, and each modality contributes a chosen weight to squared distance.

For a chosen number \(k\), the \(k\) cells with smallest distances to a focal cell are its nearest neighbors. A graph can then represent those neighbor relations, using, for example, a union of directed neighbor links or mutual links. These choices can yield different graphs, so report them. The distance and graph are analytical constructions, not direct measurements of cell-to-cell biological interaction.

The division by \(D_R\) and \(D_P\) makes each modality’s term an average squared difference across its retained dimensions, rather than a sum that automatically grows with the number of dimensions. It does not remove the need to scale dimensions sensibly or choose weights carefully.

Investigate a Discordant Marker Population

A joint graph can help identify cells that are similar in overall multimodal representation, after which marker-level measurements can be examined. It is important not to define a population solely by the same marker contrast that will later be presented as evidence of discordance: that creates a circular description. Instead, identify a neighborhood or population using a stated procedure, then inspect the marker distributions within it. Define RNA and ADT scales and thresholds separately, using controls or distributional evidence where possible.

Diagnostics, Alternatives, and Limits

A joint graph is useful only if its neighborhoods are stable enough to support the question being asked. Check whether cells cluster mainly by batch, library size, or other technical variables. Review how many neighbors are shared across RNA-only and protein-only graphs, and whether the joint graph connects groups that are clearly separated in one modality for a known technical reason. Repeat the analysis over defensible choices of retained dimensions, \(k\), modality weights, and graph symmetrization. If a small change in settings entirely rearranges the population of interest, report that instability rather than presenting one graph as definitive.

The equal-weight distance above is transparent but not uniquely correct. Separate RNA and protein graphs preserve modality-specific structure and are valuable comparisons, although neither alone provides a joint neighborhood. Simple concatenation of coordinates is another option, but unless scales and contributions are controlled, the larger-scale or higher-dimensional block may dominate. Adaptive approaches, including weighted-neighbor methods, can allow a cell’s neighborhood to reflect which modality is more informative for it. Their weights and resulting neighborhoods still require diagnostics; adaptivity does not automatically correct poor measurements or guarantee biological validity. The earlier material on benchmarking joint multimodal representations motivates comparing these alternatives rather than assuming one representation is best.

Thresholded discordance is also limited: it discards information within the high and low groups, and results can change with threshold choice. Show the underlying distributions and state how thresholds were obtained. A high ADT count is not automatically evidence of high surface abundance unless background and assay behavior support that reading. Likewise, an RNA count of zero can reflect sampling and detection limits, not necessarily absence of transcript. Associations between RNA and protein in a selected neighborhood are descriptive and should not be reported as causal effects.

A Practical Analysis Sequence

The central practical point is that normalization, graph construction, and marker interpretation are linked but distinct decisions. Separate normalization prevents the raw count scales from being conflated; an explicit joint distance makes the neighborhood definition inspectable; and a careful discordance analysis keeps a measured mismatch from being overstated as a biological conclusion.