For a detailed explanation of our findings or citation of this work, please see our preprint on BioRxiv: https://doi.org/10.1101/2024.07.14.603469. Please don't hesitate to ask any questions or let us know if you're getting an unexpected result. We have done our best to make this method robust, but data can be weird and noisy, so we're happy to offer our feedback and look into improvements!
Low MALAT1 expression is associated with a lack of a nucleus in single-cell RNA-sequencing data. Cells without nuclei are likely either empty droplets filled with ambient RNA, cell fragments, or mature erythrocytes. Our function define_malat1_threshold takes a vector of normalized MALAT1 expression, and outputs a minimum threshold value that can be used to filter your scRNA-seq object.
We hope to develop a package to allow a user to easily access this function. In the meantime, you can use this the function by either pasting the code directly into your R script, or cloning the GitHub repo, moving the malat1_function.R script into your analysis directory, and adding source("malat1_function.R") to the top of your script to access the function.
We generally recommend to use this function early in a QC pipeline, after reading in and normalizing your data. After filtering for minimum MALAT1 content, you can check for UMI and mitochondrial distribution to see if further filters are necessary, but you may find that this filter is sufficient. We speculate that cells with high MALAT1 but also high mitochondrial content may simply be metabolically active. Doublet filtering is unrelated to this pipeline and can be performed afterwards. Similarly, this function does not correct ambient RNA expression, so correction with e.g. SoupX may be performed after filtering for your final cell matrix, if desired.
This function can also be used to perform additional filtering on a processed dataset. You should get slightly different results (but a consistent broad pattern) if you run this function on individual samples or integrated datasets. Looking at individual samples is likely the best approach, as batch effect may impact the overall MALAT1 distribution of a sample, and these distributions may not integrate perfectly. We do not recommend running this function on individual cell types, as we have found that outliers occur at a sample level rather than a cell-type level. Poor-quality cells tend to cluster together and may therefore appear as a single cell type which will not be filtered accurately. We note that the function tends to work best on larger sample sizes (so it would struggle to find patterns in tiny samples).
To use this function, isolate the normalized MALAT1 expression values from your scRNA-seq object. In a Seurat object, this may look like:
norm_counts <- sobj@assays$RNA@data["MALAT1",]
This can be fed into the MALAT1 threshold function which will return the minimum MALAT1 value that each cell should contain:
threshold <- define_malat1_threshold(norm_counts)
This threshold value can be used to flag or filter cells from your single-cell object. The code below flags cells that don't pass the threshold by using TRUE values to represent good cells, and FALSE to represent cells that don't pass the filter:
malat1_threshold <- norm_counts > threshold
sobj$malat1_threshold <- malat1_threshold
sobj$malat1_threshold <- factor(sobj$malat1_threshold, levels = c("TRUE","FALSE"))
DimPlot(sobj, group.by = "malat1_threshold")
From this, you can use the result to remove cells from your object:
good_cells <- WhichCells(sobj, expression = malat1_threshold == TRUE)
good_sobj <- subset(sobj, cells = good_cells)
We can demonstrate using this function with the Tabula Muris Senis Pancreas dataset which can be downloaded as a Seurat object from cellxgene. The data are also described in this paper:
The Tabula Muris Consortium. A single-cell transcriptomic atlas characterizes ageing tissues in the mouse. Nature 583, 590–595 (2020). https://doi.org/10.1038/s41586-020-2496-1
Here is a brief look at the cells in the dataset:
Here is MALAT1 projected onto the UMAP, and a histogram of the normalized MALAT1 values for that dataset. You can see that certain cells in the pancreatic acinar cell cluster have especially low MALAT1 values and may be suffering from some quality issues. Cells that may actually be empty droplets would be those in the lower MALAT1 expression peak in the histogram, in addition to those with a peak at zero:
This function fits a density function to the histogram, and models a quadratic to the highest MALAT1 expression peak above the normalized expression value of two. It finds this peak by analysing local minima and maxima that appear on the density function. The lower x-intercept of this quadratic is used to define the minimum MALAT1 threshold.
The function outputs the following plots: (1) The density plot with local minima annotated. (2) The density plot with local maxima annotated. (3) The points of the density function in black, with points highlighted in blue covering the range of the data that the quadratic is fit to, with the quadratic fit overtop of the points in red (below). (4) The histogram of normalized MALAT1 counts with the red line indicating the minimum threshold value (below).
Using the code above, we can see which cells passed the filter. Most of the cells that failed the filter are, in fact, pancreatic acinar cells (highlighted below as "FALSE" for having not passed the filter):
This analysis relies on the assumption that there is a MALAT1 peak above the normalized value of one. If such a peak (i.e. local maximum above one) does not exist, the function may call an error. This is probably a good sign to take a closer look at your data anyway, but you can also lower this value by adjusting the parameter chosen_min.
Some histograms are wonky and can have a lot of little peaks, especially if you are working with integrated samples (which may just have different MALAT1 peaks due to batch effect) or samples with very few cells in them (which may have lots of little peaks due to data sparsity). To make the function more robust to these scenarios, there is a smoothing parameter, smooth, which is set to a high value of 1 as default, but can be lowered closer to zero for a tighter fit to the histogram. Further, if the function has trouble finding appropriate minima and maxima, abs_min and rough_max are set to 0.3 and 2 respectively to guide extreme minimum and likely maximum values to help the function work properly and not throw an error. Unless something really weird happens that I haven't predicted yet, you should always end up with a MALAT1 threshold of 0.3 or higher. If you are worried about throwing away too many cells with this lower boundary, you can always change abs_min to zero, but it may not help you keep anything good.
Other parameters that can be modified are bw, lwd, and breaks. Increasing or decreasing bw to say 0.5 or 0.01 respectively will change the plotting of the density function, with higher values creating a function with fewer inflection points (i.e. a "less curvy" function). Modifying lwd changes the thickness of the line on the final plotted histogram, and breaks is the number of buckets used in the histogram.
Worst case scenario, if the function doesn't work for some weird, confusing reason, you can always eyeball your MALAT1 values to try and figure out if there is something fishy going on with your data. You can manually choose your own threshold by looking at the histogram, or just pick out clusters of concerning cells by projecting MALAT1 onto your UMAP.