Traditional RNA-sequencing, which occurs on homogenized bulk tissue samples, has made it possible to have a global view on the entire transcriptome of a great many tissues and species with relative ease and affordability. Although the technology is mature and its analysis is well-documented, it lacks the ability to capture changes in minor cellular populations. For this purpose, newer technologies (single cell RNA-sequencing) are typically used for analysis at the single cell (rather than whole tissue) level (Kim et al., 2015). Although these enable a powerful new view on transcriptomics, they are technically challenging and artifacts such as so-called “dropout” complicate data analysis (Kim et al., 2015). These and high costs remain limiting factors in the widespread adoption of these methods. To balance between the sensitivity of single cell sequencing and the relative affordability of bulk tissue RNA sequencing (RNA-seq), cell type specific transcriptomes are now also being obtained using a variety of technologies and methods (Kim et al., 2015; Yuan et al., 2017). These methods typically involve the enrichment of a specific cell type of interest by morphological or fluorescence-based approaches, followed by the sequencing of a pool of these cells, usually resulting in one pooled cell type group per sample. They are collectively termed single cell type RNA-seq (sctRNA-seq).

There are a number of advantages of sctRNA-seq compared to bulk tissue RNA-seq or even single cell RNA-seq (scRNA-seq). Relative to bulk RNA-seq, sctRNA-seq (and scRNA-seq) can reveal changes in gene expression in minor or rare cell types, or cell types of a priori interest, which may obscured by changes in more prevalent cellular populations in traditional expression profiling methods (Hwang et al., 2018). Relative to scRNA-seq, sctRNA-seq affords the choice to sequence mRNA from pools of cells to a greater depth, which could greatly simplify the cost for collection, analysis, and interpretation of these data. Another important difference is that with sctRNA-seq (as compared to scRNA-seq), cell type labels are known a priori rather than assigned by cluster associations. For more information on the considerations and differences between these methods, we refer readers to (Hwang et al., 2018; Chen et al., 2019).

Laser capture microdissection coupled with RNA-sequencing (LCM-seq) is one such technology for sctRNA-seq. In LCM-seq, a thin slice of tissue is mounted under a high power microscope with the cells of interest visibly labeled for targeted excision by a laser and collected for sequencing (Bonner et al., 1997; Kummari et al., 2015). There are a number of approaches for labeling cells, including using transgenic animals or through the use of fluorescence in situ hybridization (Progatzky et al., 2013). LCM-seq differs from other methods for sctRNA-seq, including Fluorescence Assisted Cell Sorted (FACS) and immunopurification-based methods such as Translating Ribosome Affinity Purification (TRAP-seq), in that it affords the ability to visually target cells in situ prior to capture. LCM-seq has been widely adopted in the context of cell type specific profiling, including in motor neurons (Bouçanova et al., 2020; Harjuhaahto et al., 2020; Nizzardo et al., 2020), enterocytes (Moor et al., 2018), cortical (Pereira et al., 2017) and hippocampal neurons (Deng et al., 2019), and cortical interneurons (Shukla et al., 2019; Newton et al., 2020).

One potential challenge with LCM-seq is that mRNA from interfering cell bodies and processes may be microdissected along with the cell type of interest (Okaty et al., 2011; Kummari et al., 2015; Rocco et al., 2017). This phenomenon, which we term off-target cell type contamination, may arise in part due to physical overlap or close apposition by surrounding cells or long-range processes (illustrated schematically in Figure 1A). Off-target cellular contamination is especially relevant in brain tissue, in which diverse populations of distinct cell types are densely packed and surrounded by a myriad of other cell types (Kummari et al., 2015; Tremblay et al., 2016). While this potential issue has been noticed previously in the context of microarray data (Okaty et al., 2011) and in other data types, including Patch-seq (Tripathy et al., 2018), to our knowledge there is currently no approach that attempts to systematically identify this issue across sctRNA-seq datasets and assess its potential impact on downstream analyses.

In order to gain a deeper understanding about how off-target cellular contamination might affect data analysis in sctRNA-seq datasets, we applied a computational method to estimate the extent of off-target cellular contamination in sctRNA-seq datasets. Our approach relies upon the availability of high-quality scRNA-seq data as a reference for comparisons. The output of the analysis is a set of coefficients that estimate, on a per-sample basis, the degree of off-target cellular contamination per cell type in the surrounding tissue. We present two primary use-cases for these cellular contamination estimates in the context of brain sctRNA-seq analysis. The first is in simply assessing the quality and purity of collected sctRNA-seq derived transcriptomes. The second is in using these coefficients in the context of condition-specific differential expression analyses to help correct for the confounds of sample-to-sample and across-condition variances in sample quality.

In this work, we applied an approach for characterizing off-target cellular contamination in RNA-seq-based datasets of pooled cell types (sctRNA-seq datasets). By comparing these sctRNA-seq samples to analogous high-purity scRNA-seq datasets, we were able to derive a series of contamination coefficients that yield an estimated amount of contamination for each off-target cell type. We note that using markers derived from single cell RNA-seq data and applying them to RNA-seq datasets to derive cell type-specific insights is commonly done in the context of cellular deconvolution applications for bulk and spatial transcriptomics datasets (Wang et al., 2019; Elosua et al., 2021), including methods developed by our group previously (Mancarci et al., 2017; Toker et al., 2018). Here, in the context of sctRNA-seq datasets, such coefficients help contextualize the cellular sources and relative amounts of off-target cellular contamination. They can be used to help correct for sample-to-sample differences in purity in downstream analyses. Further, by applying our approach to multiple sctRNA-seq datasets collected using different methodologies for cell type-specific enrichment (LCM, TRAP, and manual purification), we were able to qualitatively compare how these different methodologies compare in terms of sample purity. Though our applications here were focused on sctRNA-seq datasets collected from the mouse neocortex and hippocampus, this method can in principle be applied to any species or tissue.

We found that pyramidal cells and non-neuronal cells, namely oligodendrocytes and astrocytes, were among the greatest contributors to off-target contamination in each dataset and cell purification methodology we re-analyzed. Across methodologies, we observed that manual sorting appears to show the least apparent off-target contamination, followed by LCM-seq and then by TRAP-seq, consistent with previous reports based on microarray datasets (Okaty et al., 2011). We reason that to a large extent, this cellular contamination is likely to reflect underlying biology. For example, apparent pyramidal cell contamination in interneuron-specific samples, as we observed here for LCM-seq and we and others observed previously for Patch-seq (Tripathy et al., 2018; Gouwens et al., 2020; Scala et al., 2020), is likely due to pyramidal cells' large size and extensive projections throughout the neocortex. As such, pyramidal cells are more likely to have their mRNA captured inadvertently when sampling non-pyramidal cell types such as interneurons. Interestingly, we found anecdotal evidence that sources of off-target contamination tended to be similar in different datasets sampled using the same methodology despite being collected by different experimenters such as the LCM-seq mouse Aging (Shukla et al., 2019) and Stress (Newton et al., 2020) datasets reported here. Together, these findings indicate that off-target contamination is widely present, but of overall modest size (relative to expectations based on single cell reference data), and likely an unavoidable consequence of various methodologies for sctRNA-seq.

Despite the modest effect sizes of contamination associated with single cell type captures, we show analytical improvement when correcting for off-target contamination in the context of group-level modeling analyses, such as case/control differential expression analyses. We note that this is just one possible use of such covariates: they may also be used in other ways, including to remove samples with outlying levels of contamination, as covariates in analyses such as limma (Ritchie et al., 2015) or ComBat (Johnson et al., 2007) to normalize gene-level read counts for contamination, etc. Applying our approach initially within the context of a simulation-based framework, we found that incorporating contamination coefficients was effective in improving true calls of differentially expressed genes without a corresponding increase in false calls. This approach was also effective when contamination was confounded with or biased toward a particular group (e.g., greater contamination in the “disease” group). Applying our approach to real-world LCM-seq datasets, we found a greater number of differentially expressed genes per cell type. Importantly, given that the models that included our coefficients provided a statistically improved fit to these data relative to models without them, we further posit that correcting for off-target cellular contamination is likely to improve the quality of differentially expressed gene calls (rather than simply increasing the number thereof). However, we note that without experimental validation, we do not know which genes are truly differentially expressed between condition and control, making it difficult to rigorously evaluate the accuracy and validity of this approach.

Based on our simulation and real-data analyses, we expect that including covariates for sample-specific off-target contamination will tend to improve such group-level analyses. Nonetheless, correcting for off-target contamination when it is mild and consistent across experimental conditions is of minimal benefit. As we have described, the physical source of off-target contamination is necessarily due to the presence of cellular soma and processes surrounding the cell of interest (see Figure 1). Indeed, many psychiatric and neurological disorders are characterized by altered neuronal and astroglial morphology, which pose potential confounds to the interpretation of LCM and TRAP-seq data. Examples include reduced soma size of hippocampal neurons in major depressive disorder (MDD) (Stockmeier et al., 2004), reduced pyramidal cell dendritic branching in natural aging (Luebke et al., 2015) and MDD (Qiao et al., 2016), reduced astrocyte size and density in MDD, bipolar disorder, and schizophrenia (Cotter et al., 2002), and regionally-varying atrophy and hypertrophy of astrocytes in Alzheimer's disease (Verkhratsky et al., 2019). The morphological changes outlined above represent sources of contamination that are confounded with disease and thus may introduce false differentially expressed genes or mask the true differential expression of others. These kinds of conditions are thus likely to benefit the most from the approach we outline.

The major assumption underlying our approach is that high-purity scRNA-seq datasets can provide a useful proxy for sctRNA-seq samples. This assumption, while reasonable, has a number of limitations. First and foremost, this approach requires the availability of scRNA-seq reference datasets, such as those we used here from the mouse neocortex (Tasic et al., 2018). Second, there may be genuine differences in sctRNA-seq expression profiles relative to scRNA-seq that are not the result of contamination; for example, such differences might arise due to technical factors such as differences in library preparation and collection protocol [which may also include transcriptional differences related to cellular dissociation-based cell stress (Adam et al., 2017)]. Similarly, true biological differences can result in large changes in cell-state, including condition- or disorder-specific transcriptional regulation (Lee and Young, 2013), or cell differentiation. As our approach relies on the quality of cell type-specific marker genes: unsuitable markers (e.g., those that are markers of cell state but not cell type) or those with sufficiently non-binary expression patterns may render the derived contamination coefficients unusable. A related limitation is that as the markers derived here are markers based on scRNA-seq samples from dissociated cell bodies, they may not be entirely reflective mRNA in more distal cellular processes (Glock et al., 2017). Each of these caveats are likely to affect each dataset to varying degrees and we advise that users carefully consider these assumptions prior to the use of these methods.

Despite the fact that some amount of contamination seems to be inevitable by current methodologies, cell type specific RNA sequencing offers several advantages compared to traditional bulk or single cell RNA sequencing. Our approach for estimating sample-to-sample differences in purity provides a simple yet flexible approach to account and control for their effect in downstream analyses, allowing for more meaningful interpretation of sctRNA-seq datasets.

The original contributions presented in the study are publicly available. The code needed to perform the analyses can be found as an R script online at: https://github.com/jsicherman/ sct-contamination, built on R version 3.6.0 (Planting of a Tree).

ST and ES designed the study. ST, PP, and ES supervised the study. ST, PP, ES, and DN provided input on data analysis. JS performed data analysis. JS and ST wrote the manuscript with input from all authors. All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.