Question: What is the impact of FOV-associated sensitivity variations?
Answer: A full Xenium dataset is composed of many microscope fields-of-view (FOVs), stitched together.
Similar to many optical systems, Xenium has small changes in optical performance across the microscope FOV. This results in slightly lower optical resolution at the edge of each FOV. This can lead to modest reductions in the transcript sensitivity, especially for “dense” gene panels that tag a high density of transcripts in the tissue. We term this effect “roll-off.”
This article explores the strength of the FOV roll-off technical effect and puts it into the context of other sources of technical variability commonly observed in single-cell and spatial transcriptomics assays. We’ll show that the magnitude of the technical variability introduced by roll-off is small compared to other important sources of variability and compared to variability present in single-cell data.
Analysis
Figure 1 shows 1 day old fresh-frozen mouse pup, analyzed with the Xenium Mouse Multi-tissue Panel using XOA v1.5. This sample type yields high transcript densities - roughly 200 decoded transcripts per 100 µm². The two images are the transcript density maps for sequential sections run on two different instruments - one that fails the image quality specifications for commercial Xenium instruments, termed “High Roll-off” and one that passes our specifications, termed “Low Roll-off”. In addition to the organ morphology, the grid pattern of the FOVs is visible, showing the roll-off effect.
Figure 1
| High Roll-off Instrument (failing spec) | Low Roll-off Instrument (passing spec) |
 |
 |
To measure the relative sensitivity across the FOV, we can stack the data for each FOV and take the mean expression level in small spatial bins. In Figure 2, we stratify the genes into 3 groups “Low”, “Med”, and “High” based on expression level, to test whether highly expressed genes show a stronger roll-off pattern. The relative sensitivity metric is the ratio of mean transcript counts per cell of the FOV position compared to the central region of the FOV. We can see the roll-off pattern is similar between the instruments (rows), but the high roll-off instrument drops to 0.6 relative sensitivity at the worst corner of the FOV, while the low roll-off instrument drops to 0.7 relative sensitivity. We can also see that the roll-off pattern is similar for highly and lowly expressed genes.
Figure 2
We can also make a histogram of the relative sensitivities of the FOV bins and gene expression levels. In the low roll-off instrument ~95% of the FOV has a relative sensitivity of >0.8 across all the expression levels (Figure 3).
Figure 3
We can measure the coefficient of variation (CV) of these distributions, as measure of the strength of the variation introduced roll-off. CV is defined as the standard deviation of distribution divided by its mean. The CV of the sensitivity in the low roll-off instrument is 0.07, which can be interpreted as injecting ~7% technical noise into the results.
| Data | Across FOV CV |
| Low Roll-off | 0.07 |
| High Roll-off | 0.103 |
We can also assess whether the FOV position of a cell impacts its cluster assignment. We take the K-means clustering results for each dataset and show the number of cells in each cluster, stratified by 3 groups of “Radius Bin”, or distance from the center of the FOV. We can see that no major shifts in the cluster proportions are observed in different radius bins.
Figure 4
Other sources of technical variability: Z-Position
We can estimate the Z-position of each cell as the mean Z position of all the detected RNAs in the cell. Cells close to the top or bottom of the slice have likely had some of their volume cut away in sectioning, so will show a reduced sensitivity on average. Here we compute the Z-position of each cell and break the cells into equal sized groups according to their Z position, then plot the relative sensitivity of each group. There’s a broad effect of Z-position on sensitivity.
Figure 5
We can measure the CV of these relative sensitivities by Z-position:
| Data | Z-position CV |
| Low Roll-off | 0.113 |
| High Roll-off | 0.126 |
So the technical variation introduced by the Z-position of a cell within the tissue slice is stronger than the technical effect caused by FOV roll-off.
Comparison to Single Cell Variability
We take some high quality single-cell data from 2 replicate PBMC samples. Here we show the histogram of the transcript count within each cell, separated by the unsupervised clustering performed separately for each sample. Different cell types show up with different typical transcript count levels.
Figure 6
We can compute the CV of the total transcript count within each cluster, and we see CVs ranging from 0.11 - 0.18. This variation is a combination of the true biological variability and any technical variability inherent in the single-cell sequencing process. If we ignore the cluster labels and compute a CV over all cells, we see a CV of 0.31. Secondary analysis methods designed for single cell data routinely deal with this level of heterogeneity in the normalization, dimensionality reduction and clustering process.
Figure 7
Summary
The FOV roll-effect is a real source of technical variability in Xenium data that should be considered. However, the magnitude of the technical variability introduced by roll-off is small compared to other important sources of variability, and compared to the total variability present in commonly well-established single-cell data. Thus, popular normalization techniques to single-cell data deal with substantially larger total variation in transcript counts.
Product: Xenium In Situ Gene Expression, Xenium Prime In Situ Gene Expression