xcell2 returns scores greater than 1

[Rohit-Satyam]

Just to confirm if I am correct about this, does the corrected scores returned by xCell2Analysis also ranges from 0 to 1 or it's just the raw scores (because I see Max. :1.337908 for SRR6055342)? Besides, since I am using useSpillover=FALSE, should I set rawScores=T and spillover=F?

subudhi.xCell2Ref <- xCell2Train(
  ref = subudhi2020,
  refType = "rnaseq",
  BPPARAM = parallel_param,useSpillover=FALSE,
  useOntology=FALSE
)
table(rownames(toenhake2018) %in% rownames(subudhi2020))

xcell2_results <- xCell2Analysis(
  mix = assay(toenhake2018),
  xcell2object = subudhi.xCell2Ref
)
summary(xcell2_results)
   SRR6055337       SRR6055338        SRR6055339        SRR6055340        SRR6055341        SRR6055342      
 Min.   :0.0000   Min.   :0.00000   Min.   :0.00000   Min.   :0.00000   Min.   :0.00000   Min.   :0.000000  
 1st Qu.:0.0000   1st Qu.:0.00000   1st Qu.:0.00000   1st Qu.:0.00000   1st Qu.:0.00000   1st Qu.:0.000000  
 Median :0.0000   Median :0.00000   Median :0.00000   Median :0.00000   Median :0.00000   Median :0.000000  
 Mean   :0.1201   Mean   :0.09939   Mean   :0.06838   Mean   :0.04287   Mean   :0.04196   Mean   :0.069027  
 3rd Qu.:0.2756   3rd Qu.:0.06468   3rd Qu.:0.04865   3rd Qu.:0.06950   3rd Qu.:0.02740   3rd Qu.:0.004636  
 Max.   :0.6844   Max.   :0.78182   Max.   :0.50924   Max.   :0.23478   Max.   :0.43142   Max.   :1.337908  
   SRR6055343          SRR6055344      
 Min.   :0.0000000   Min.   :0.000000  
 1st Qu.:0.0000000   1st Qu.:0.000000  
 Median :0.0000000   Median :0.000000  
 Mean   :0.0412126   Mean   :0.052271  
 3rd Qu.:0.0008971   3rd Qu.:0.001249  
 Max.   :0.4437633   Max.   :0.389725  

[Rohit-Satyam]

To give some background, I have a published Bulk Query dataset (one 1 sample per time point, no biological replicate) of plasmodium where samples were harvested at 8-hour overlapping intervals to cover the entire 48-hour life cycle of plasmodium in 2018. The reference dataset is another time-series dataset, but higher resolution, where cells were harvested at 2-hour non-overlapping intervals (3 or more replicates). So I wanted to use xcell2 to see if I can capture the heterogeneity of the Query timeseries dataset, which has a much larger scale than the reference.

When using raw scores and plotting them, it appears to match the actual time span provided by the authors of the Query dataset (See the plot below, the author-provided time labels have been mentioned with the SRR number on the x-axis).

When rawScores=F, the score matrix becomes somewhat sparse. I would like to pick your brain on this, whether to use raw scores or not. Note that here we don't have rare cell types like situation, we just wish to use xcell2 to see how heterogeneous or synchronized the bulk RNAseq is, since cell synchronization is usually inspected manually, and achieving 100% cell synchronization is difficult. So we wish to use it more for QC purposes.

[AlmogAngel]

Hi @Rohit-Satyam ,

It's great to see the package being applied to novel use cases like analyzing the heterogeneity of plasmodium.

Understanding the Different xCell2 Scores

xCell2 can produce scores in three "flavors":

1. Raw Scores (rawScores = TRUE): These are the direct output of the underlying enrichment analysis (using the singscore package). They represent the mean enrichment of all signatures for a given "cell type" (or, in your case, time point).
2. Linearly Transformed Scores (rawScores = FALSE, spillover = FALSE): In the original xCell paper, we demonstrated that the raw enrichment scores have a non-linear relationship with the actual proportion of a cell type in a sample. The linear transformation step addresses this by converting the scores to a linear scale. This step is mainly used to prepare the scores for the spillover correction step, as it requires linearity. Some users also find this transformation beneficial on its own as it places the scores on a scale that resembles real proportions. However, it's important to remember, as you noted, that even though these scores are on a linear scale, they are not true proportions and can exceed 1, especially in cases of very high enrichment.
3. Transformed and Spillover-Corrected Scores (rawScores = FALSE, spillover = TRUE): This setting is designed to reduce the inflated enrichment scores that can occur between closely related cell types (e.g., CD4+ and CD8+ T-cells) due to shared gene expression patterns.

Applying xCell2 to Your Time-Series Data

If I understand correctly, you are essentially treating different time points as distinct "cell types" and using the enrichment analysis to track the progression of the organism's state over a broader timeframe.

I think you are correct to disable spillover correction. Since your "cell types" are sequential time points and the spillover correction is not necessary and could potentially introduce unwanted adjustments.

The difference between your two heatmaps might be a matter of scale. The raw enrichment scores are non-linear, which means your first heatmap highlights differences between the highly enriched (high-scoring) time points, while compressing the visual scale for the less-enriched time points. The adjusted scores, because they have been linearly transformed, give more power to resolving differences between the lower-scoring time points. This is visualized in Figure 1C of the first xCell paper, which shows the non-linear "Before" scores being converted to the linear "After" scores.

I hope this explanation is helpful.
Please feel free to reach out if you have any more questions.
Almog

[Rohit-Satyam]

Hi @AlmogAngel, I see that the Vignette has been heavily trimmed down, and a lot of information, which was important, like the intuition about the scores, is missing. Is it possible to get the old vignette?