What happens if you combine SORT-seq and microscopy-based spatial annotation? A group of scientists made that high-tech merger happen, generating a single-cell dataset with high transcript counts and tissue-context information. They used it to investigate a source of heterogeneity and metastasis in cancer.
We often want to have the best of both worlds.
With transcriptome-wide spatial profiling, current technologies lack single-cell resolution and/or have relatively low transcript counts compared to single-cell RNA sequencing. As a result, spatial transcriptomic data likely omits lowly expressed genes or genes expressed in only a few cells. Single-cell RNA sequencing technologies like SORT-seq, on the other hand, have high transcript sensitivity and reach the single-cell level. Yet, SORT-seq lacks spatial information about the analyzed tissue.
That is, unless you add a phototagging microscope.
To overcome both technologies’ limitations, a team of Erasmus MC researchers extended SORT-seq with spatial data using a custom-built microscope. For their proof-of-concept study, published in Frontiers in Bioengineering and Biotechnology, they applied it to unravel tumor heterogeneity. We describe their technology in this blog.
Causes of cellular heterogeneity
Therapy-resistant cells arising from cellular heterogeneity within a tumor is often identified as a source of treatment failure and therapy resistance. One of the causing mechanisms of that heterogeneity is arguably a molecular program called epithelial-to-mesenchymal transition (EMT). It stands for a change in tumor or developing cells from epithelial-like traits to mesenchymal traits. Beyond fueling heterogeneity, EMT can cause tumor cells to lose adhesion and gradually gain migratory and invasive characteristics. Hence, EMT is also an important subject of study for scientists working on cancer metastasis.
Location and tissue context seem to be important factors in EMT. For example, a study found that EMT marker genes are more highly expressed in cells at the migrating front of tumors. Another study showed this was also true in a tumor model system.
The Erasmus MC team, led by Dr. Miao-Ping Chien, decided to take an in-depth look at EMT in these outwardly migrating tumor cells to answer the question, what transcriptional changes underlie their EMT? For this study, they cultured a tumor model which heterogeneously showed EMT and applied a custom-made technology combining microscopy-based spatial annotation with SORT-seq.
Select cells by any microscopically observable trait
This technology, dubbed FUNseq, was developed by Chien and colleagues to make microscopy-based cell isolation compatible with SORT-seq. It employs live-cell imaging to identify cells with interesting phenotypes. Then it tags these with a dye using a laser (phototagging). Using FACS and SORT-seq, you can isolate all phototagged cells and perform single-cell sequencing.
In principle, the authors claim, you can select cells with this technology by any trait observable with a microscope. These could be static features such as morphology or size or dynamic features such as movement or cell division rate.
FUNseq helps identify cell subpopulations
In a proof-of-concept experiment for FUNseq, the team focused on the morphological traits of EMT cells. They detected and isolated fast-moving cells and cells with mesenchymal-like morphology, an EMT marker from a heterogenous cell culture. Then they performed single-cell RNA sequencing.
Instead of the usual unsupervised clustering analysis, the team performed cell clustering informed by the FUNseq-provided annotations on movement and morphology. That helped them perform successful clustering, marker gene identification, and cell type annotation. As a result, they could find EMT gene expression patterns in the fast-moving cells and cells with mesenchymal-like morphology.
On the contrary, unsupervised clustering, the most frequently used way of clustering in single-cell data analysis, seemed to fail on those fronts. According to the authors, the gradual change in gene expression during EMT probably hinders it. But this is rescued by incorporating FUNseq-provided annotations in the clustering process.
In summary, by placing functional single-cell selection upstream of SORT-seq, the team successfully identified a subpopulation of tumor cells in vitro undergoing EMT. This subpopulation would have been disregarded by regular unsupervised clustering.
This raises the question, how did they translate that technology to a spatial application?
Adding the dimension of space to SORT-seq
The team experimented on a mammary epithelial tumor model, again focusing on EMT. This model has a specific spatial heterogeneity. Namely, cells in the outermost layer of the epithelial patch (the ‘invasive edge’) seem to be undergoing EMT. To research the transcriptomic changes underlying EMT in these cells, they applied FUNseq.
Instead of labeling the cells based on their morphology or dynamics, the team now labeled them based on their distance from the patch’s center. With their custom-built microscope and FACS, they can tag and isolate a circle of cells with a desired spatial bandwidth. They tagged and isolated cells from the outermost invasive edge and several bands of more central cells at two different bandwidths.
In total, the team RNA sequenced 743 single cells phototagged with the larger bandwidth and 696 with the smaller bandwidth. They then performed supervised clustering based on the FUNseq-provided annotations for both sets separately. Although the larger bandwidth did not give clear clustering of the cells, the smaller bandwidth did.
The results confirmed the team’s expectations. Transcriptional analysis of the clusters showed highly expressed EMT genes in the outermost cells. In addition, gene expression analysis focused on cell-cell interactions could show enrichment of several EMT-inducing signaling molecules and their receptors. These include tumor necrosis factor (TNFA) and genes involved in the epidermal growth factor pathway (CD44, EGFR, EPGN, and HBEGF). That probably means these EMT cells respond to local environmental cues and can stimulate EMT in neighboring cells via these receptors.
In July 2023, the authors published the spatially annotated SORT-seq protocol in STAR protocols.
Conclusions and limitations
It seems that with FUNseq, you can have the best of the spatial and single-cell transcriptomics world, after all. Yet, it’s important to delineate what FUNseq exactly means for other scientists working with spatial and single-cell transcriptomics.
FUNseq compared to other spatial methods
The authors mention that the in-depth transcriptional analysis of the low number of cells was made possible by SORT-seq’s technical specifications. They note it enabled them to profile gene expression using 34,000 transcripts per cell.
That is very high compared to other spatial transcriptomics techniques, which one benchmarking study reports analyzed to be in the 1,000–2,000 transcripts range. The missed transcripts would likely have affected results and prevented the team from finding genes inducing EMT in these cells. However, nobody has directly compared FUNseq and these technologies on the same tissue.
On the other hand, FUNseq as applied in the paper has less spatial information than other spatial transcriptomics methods. Since FUNseq pools cells per bandwidth, it retains information on cells’ distance to the center. Yet, unlike other spatial methods, whether a cell laid north or south of the center is unknown. In other words, FUNseq loses data on cells’ coordinates and cannot place cells on a 2D plane.
FUNseq compared to SORT-seq–only
What would happen if we compared FUNseq results with SORT-seq–only results? It’s possible that gene set enrichment, EMT score, and cell-cell interaction analysis on the EMT cells would have been less fruitful without spatial annotations. In the earlier mentioned proof-of-concept experiment for FUNseq, its spatial annotation enabled clustering analysis impossible with SORT-seq alone. There was no direct comparison in the second paper. However, supervised clustering with extra annotations can help when gene expression changes are gradual or diffuse, as in EMT. In such cases, FUNseq is at least an exciting method to attempt.
To fully explore FUNseq’s potential, the authors write that the next step would be to profile tumor sections. These are more complex and probably more heterogeneous than cultured tumor models. It would be interesting to see if the authors can find the same EMT markers.
We will keep an eye out for that research. In the meantime, you can read the full paper here:
