Scientist Stories: Mats Nilsson, Sequencing single cells in situ

All right, good morning everyone, and, uh, thanks, Jerry, for the invitation. Um, so I will talk about, uh, um, mapping all these, uh, uh, single cells into their spatial contexts and, and the way we're doing it, um, currently. So I'm Mats Nilsson, I'm a professor of biochemistry at Stockholm University, and I am— my laboratory at at the Science for Life Laboratory in Stockholm. So, again, just a little bit of a background. So single-cell RNA sequencing is the logical way of studying gene expression tissue since it takes cellular heterogeneity into account. So I think that's— quite clear. However, spatial information is missing, so it's hard to reconstruct the architecture of tissue that would, for example, reveal physical cellular interactions. And— but there are new technologies emerging that can generate spatial RNA expression profiles, and I will present our approach on targeted in situ sequencing. So the targeting method we're using is this circularization approach that we call padlock probes. And we're using linear oligonucleotide probes that have target complementary end sequences. And when they interact with the target sequence, they end up head to tail and can be joined by DNA ligase, thus creating a DNA circle. That's the way recognition of an RNA or DNA sequence is done. Turning a linear molecule into a circular one. And there are a number of advantages with that approach. One is that it's really essentially target— absolute target specificity. So if there is no target sequence present in your sample, you won't get this circularization.

We can robustly distinguish single nucleotide variants.

And the last thing here, it's a unimolecular reaction, which means that you can multiplex these very heavily. You can throw in hundreds and thousands of these molecules into one reaction, and they cannot cross-react because it's only the intramolecular ligation reaction that regenerates circles. So the circles we can then make visual, visible, and detectable by rolling circle amplification. So if you have these short DNA circles in your sample, you can add a DNA primer and a good DNA polymerase. We're using Phi29 polymerase.

And initiate a DNA polymerization.

And since this is an endless template for polymerization, this is going on in an endless polymerization reaction. It will eventually stop because of denaturation of the polymerase.

But that happens after hours and thousands of copies of the DNA circle. So as this reaction goes on, it makes more and more copies of the sequence present in the circle, so we get a clonal amplification of the DNA sequence of the circle.

It will be then, of course, that they will be co-localized. They are on the same molecule. So we create this huge molecule in an hour.

We make about 1,000 copies of a circle.

That translates to about a 45-micron-long DNA molecule. This will, in a water solution that we create where they are created, they will coil up spontaneously into submicron-sized blobs of DNA just because of a random collision with water molecules.

And then they just have to sort of collapse into one spot.

It's not a small spot, it's a quite big one. It's a couple of hundred nanometers in diameter. So it's a big molecule. And it's bright. We can easily visualize it by hybridizing shorter oligonucleotide probes.

To this tandem repeated DNA molecule.

We have devised protocols to do this in cells, in tissues, in fixed cells and tissues, and our current main target of analysis is mRNA sequences.

So the protocol we're using, we first transcribe, reverse transcribe all the RNA into cDNA. We degrade the complementary RNA then to make a single-stranded cDNA copy of all RNA in the cell. And then we target a subset of these cDNA molecules with padlock probes, hybridize and ligate. And after the ligation, we initiate the rolling circle amplification. And we're using the cDNA molecule to prime the synthesis. So in the end, our amplification product will be an extension of the cDNA molecules that we're targeting.

To the right here, you see examples how this can look. Is the detection of beta-actin transcript in a human fibroblast cell line.

And the blue dots here are nuclei, and the red dots are mRNA molecules that were detected, or cDNA sequences. And each dot is a rolling circle product, and we can detect and count these rolling circle products, and that's our estimate of expression in that cell. And this, uh, very under these very ideal conditions, we can maybe reach up to 30% detection efficiency. But I think this is extreme.

It's, it's probably more like 5 to 10% detection efficiency.

So if, let's say, there are 1,000 mRNA molecules in a cell while living, after our procedure of fixation, permeabilization, and all these enzymatic steps, we would end up with well, 10—

maybe 10% of that content, so 100 dots. So there are a number of things I could use this for, obviously. I'm here illustrating detection of an oncogenic point mutation in a formalin-fixed, paraffin-embedded lung cancer tissue.

9 fields of view of a tissue. You can see up here is an image of normal lung. Yeah, we're targeting the EGFR transcript and this particular point mutation. We have detect—

we're detecting the wild-type transcript with the green fluorescence and the mutant transcript with red fluorescence.

And so you see some scars, green dots here in the normal tissue. Here's the border between the normal and the tumor, and now in the tumor, see appearance of red dots. It's about 50/50 green and red dots in this part of the tumor. If you look in this part of the tumor, you basically lost the green dots, and there are abundant red dots, and it's really dominating. So it looks like the wild-type allele has been lost in this in the cancer cells in this part of the tumor.

So we can visualize intratumoral genetic variants. We've recently also used this technology to provide molecular diagnosis of CTCs.

So we used these— wires that are produced by, they're called something.

It's a wire that you put into the vein of the patient and it has EpCAM antibodies on it, so it will enrich, it will capture circulating tumor cells and also a lot of other things.

And then what we did was using this method to target and a transcript variant of the androgen receptor, the B7, in prostate cancer patients, and KRAS mutations in pancreatic cancers. So we have supplied it in situ on these wires, and so we could detect then individual circulating tumor cells, and we could tell whether they were overexpressing this transcript variant or not, and we could tell in these patients whether they are—

whether the CTCs have KRAS mutations or not. So I can go really sensitive down to single-cell genotyping experiments. Then to— in order to multiplex these reactions, we were interested to see whether we could use next-generation sequencing to sequence rolling circle products in these—

in cells. And it's not so far-fetched because complete genomics use— maybe you know or are familiar with this genome sequencing approach where they circularize randomly fragment genomes and then circularize and rolling circle amplify and use those rolling circle amplification products as sequencing substrates, which is distributed on a two-dimensional random array. And then they have used, or at least in the, in this publication, they use sequencing by ligation chemistry. So we just took the protocol from, from the paper and applied it on round of circle products that we had created in, in in situ, and it worked, first experiment.

So it shows that it was actually quite robust protocol that they had published.

So the approach, very same, similar as I just outlined before, but we can do sort of use the sequencing for two purposes. We can use gap-fill padlock probes where we design probes to have a gap between the ends and then provide both the polymerase and the ligase. And then the polymerase can fill this gap and make a copy, copying a short bit of the cDNA sequence and clone it into these small DNA circles. And then we can sequence short bits of cDNA in cells.

Or we just use it to decode heavily multiplexed padlock probe reaction by sequencing sequence barcode that we introduced to the probe molecules to identify the different rolling circuit products after amplification.

So sequencing by ligation works much like Illumina sequencing that you have a base-specific incorporation of fluorescence dye., but using a ligase instead of a polymerase to do that. And, and then you take this through an iteration of biochemistry steps where you incorporate fluorescence in a specific way.

You image that, create a huge number of images of, of your first base incorporation reaction, and then you wash off the fluorescence and then apply as biochemistry to sequence the next base.

And then you take a lot of images and wash off the fluorescence and so on and go on.

And then you compare the different images from the different sequencing cycles and you can read the sequence of base incorporations. So here's the data from our first in situ sequencing experiments. We're sequencing a bit of beta-actin.

In mouse and man in a co-culture of fibroblasts from mouse and human-derived fibroblasts. So here's the incorporation of the first— okay, so in mouse, this sequence we're targeting reads AAAGC. In man, it reads AGGC. So here's the data from the first First base incorporation, you see a lot of blue dots, and that shows incorporation of Cy5, which we read as A, as we can't distinguish these cells. In the second cycle, we still incorporate A dyes in— or Cy5 in this cell, while in these cells we incorporate Cy3. So we read G here and A in these cells.

Cells.

And now we can tell that these are mouse cells and these are human cells.

And as you can see, it's a really clear-cut and clean chemistry, so it's quite easy to interpret. You can even sequence by eye. But that's, of course, not how we're doing it. We have developed base-calling algorithms to call bases in these samples. These huge image stacks that we're creating.

The first multiplexing experiment we did was gene expression profiling. We took the 21 genes that are present in the Oncotype DX for breast tumors.

The Oncotype DX is a gene expression predict— treatment predictive gene expression profile. That basically reads the—

it's qPCR-based, and you run a lot of PCRs and calculate from the CT values of the individual genes, you calculate the recurrence score, which is a very obviously empirical score. This is how it looks. It contains genes that are involved in proliferation. There are some hormonal and other marker genes here. There are some genes that are expressed by the stroma, actually, and other breast cancer-relevant marker genes. And also CD68 macrophage markers.

It's sort of a mixed pot of many different things that this recurrence score is measuring. In the bulk extracts of RNA. So we were curious to see how that looks in a spatial context.

So what you see here is a 0.6 square millimeter-sized bit of breast cancer tissue. It's an HER2-positive and ER-negative one, and it has about 2,000 cells.

So we applied these 21 genes and a couple of others, so we had 39 genes. And what I'm showing here is the top 30 of these genes. Genes. So here's now—

so we're plotting the in situ sequencing data just as symbols.

So it's really an interpretation of many images. So for every pixel, we have sequenced about— we have about 20 images.

And then we plot symbols, and you can see that the combination of colors and symbols are really different in different areas of this tissue. Tissue, and there is a prominent yellow-pinkish area here and then a green-bluish area here, which translates to the cancer cell compartment and the surrounding stroma. This is an H&E stained—

well, so after the sequencing, we stained the tissue with the H&E, so it's the very same tissue that we have now the morphology and the molecular content.

So I can look then at all the distributions of these 30 genes, and so you can look here at the— on the HER2 background. So this is the distribution of the HER2 sequencing reads, and it really outlines very nicely the cancer cells in the tumor. And then on top of this, we plot plot in blue the vimentin here onto the HER2, and if its colocalization is white, an ER here on top of HER2, it's an ER-negative, so we don't see much ER in this tumor, and all the other genes. And I can see that at the proliferation, for example, it looks like there is a little bit more activity on the right part of this tiny bit of the tumor.

Remember, it's only 0.6 square millimeters, and still one can see a little bit of heterogeneity within this cancer cell population.

I mentioned this stromalize F3 gene.

It's expressed by fibroblasts, and it produces a protein that degrades extracellular matrix, and it has—

is implied as a bad prognostic marker for invasion. And this is really exactly the location of these— even if it's just a handful of sequencing reads, they all end up in the periphery of the tumor and outside the tumor.

So it looks like we have fibroblasts here that are trying to release cancer cells from the tumor. To get a little bit of quality check, we also extracted RNA from a consecutive section and ran a standard standard RNA sequencing on the bulk of that section.

And this is the correlation between the— this must be the— okay, here's RNA sequencing and in situ sequencing. And it's a striking correlation, actually. So it looks like it's a pretty quantitative approach. If we look at the HER2— no, this is the vimentin-positive cells in our data. We've got the best match with the normal breast tissue that we found in the databases. And for the HER2-positive cell population, we've got the best match with a breast cancer-derived cell line.

So a number of things one can use this kind of data for. So of course I can try to make recurrent scores and we're working on that, but it's actually not completely super simple to convert this bulk scoring approach to a spatially resolved approach. So I think we have to maybe rethink that a little bit. Okay, so but another thing I can start to look for just using unsupervised clustering, try to define to find different cell populations among the cancer cells.

So this—

the approach we're taking here is just based on the DAPI stains, expanded DAPI, the nuclear stain, expand that until it hits another expanding cell and then define that as a cell. That's, of course, a very crude guess of the extent of a cell, but it gives at least some approximation of a cell. And then we measure the in situ RNA content and cluster. And here I think it's over-clustering, but we can have a look at least at these two cell populations, the population 9 and 14.

14 is located mainly here to the left in this tumor. And the population 9 to the right. And if one then pick 9 random nuclei, so the cells now are in the middle of this field of view, and this is how they look. They look messy as tumor cells of— tumor nuclei often do. But the interesting is the population 14 looks even messier. They have huge vacuoles in the nucleus.

So, we found we can use these molecular data to find morphologically distinct subpopulation of cancer cells. So, that's reassuring that we're actually finding something biologically relevant.

And then, current work is combining the mutation profiling and expression profiling. Here is a breast tumor again, a piece of a breast tumor. That has been deeply sequenced, and so have a mutation profile of— we had the mutation profile of this tumor based on deep DNA sequencing. We designed probes for all these mutations and applied it onto this tumor tissue. And here you can see the location of 3 different mutations, the blue mutations, the green, and the red ones, and you can see that they actually form we can visualize different genetically distinct subpopulations of cells in this, again, pretty small piece of tumor. And here is the mutation profile of these different regions, so region 1 and 2 look like this. They are sort of similar in— well, they are similar mutation profile that you can see from the image. And 5, the green patches here, and 4— sorry, 3 and 5, and 4, the red one.

They all look a little bit different.

And it's the blue one that is present, so this is ductal carcinoma in situ. We also had invasive lesions from the same patient, and in that, this blue population.

Clearly dominated. So on top of that, on the same or on a consecutive section, we applied an expression profile.

So we had 96 genes that we measured, just used to measure expression, and then we can look at the expression profile of these particular subpopulations of cells.

So that's another kind of layer we want to put on the tumor tissues. So we get the genotype-phenotype correlation.

And also look at this immune cell activity in these different genetically defined or distinct cell populations. So here we're looking at— this is another tumor where we've done this. So we looked at the tumor compartment and then tumor compartments in white here and then surrounding stroma in yellow. And derived immune cell profiles. So this is both cell marker and activity markers in this, and what we can see is that the immune cells within the tumor compartments have a completely different profile than the immune cells in the surrounding stroma, which may not be very surprising.

Okay. So that's our work towards mapping and characterizing tumor heterogeneity. Next step is using this technology to make spatial cell maps out of single-cell RNA sequencing data. So we start with single-cell RNA sequencing. We have these clusters that we want to find out where they are located in the tissue.

Right, so this is actually— this is cells that are derived from CA1 in mouse hippocampus, and they form these kind of clusters. We then take that data and and pick genes that will collectively be able to define these different cell types.

So we select marker genes, and then we do this massively parallel in situ expression profiling of the marker genes in mouse brain sections.

And then for every cell in the section, we will— derive an expression profile and then match it to the expected expression profile from the single-cell RNA sequencing. So, it's really trying to map these cells onto a spatial map.

So, it's a probabilistic approach where we try to match the in situ expression profiles with the single-cell-derived expression profiles.

So, and here we're targeting for this expedition of making a complete cellular map of mouse hippocampus, CA1. We designed 99 probes, and we did run it through 5 cycles of sequencing this time to get a little bit more data from the imaging. And we applied it on full coronal sections. So here's the distribution of the 99 genes in a full coronal section of the mouse brain, adult. And then we zoom into the hippocampus. It looks like this. And then in this area of the hippocampus, you see these cells that we're targeting. So we're not targeting all cells here. We're only targeting the neurons. And then down to the single cells. And this is the data we're going to—

to match to this single-cell RNA sequencing data. Some example distributions of marker genes in our data and compared to the Allen Brain Atlas, which has comprehensively mapped all mouse RNAs onto mouse brain tissues gene by gene, so they don't have— a combined profile as we have, but it's just a single gene kind of map. Here are the genes we're targeting. And we have compared these, our distributions with Allen Brain map for all the genes, and they are— when we can compare, they look very comparable.

All right. So next step is then to assign an identity to every individual cell in the tissue. And we do, as I said, we use a probabilistic approach that has been developed by Kenneth Harris at UCL. So it actually uses also the single-cell RNA sequencing definition to guide mapping of the reads to cells. So it really looks for what combination of markers makes sense in the local environment. So here it assigns reads. And then make the comparison with the cell definitions we have and give it the probability. So this one must be a CGA-NGF cell. It's 100% probability. This one is a little bit less likely, so we make these pie charts of probabilities. And this one is this dark purple or this pinkish cell. Line, and most likely this one. If it isn't, it must be this one. So these are the probabilities. And then we map these pie charts onto the DAPI image basically of the brain. And here's a complete map of CA1 of mouse hippocampus, and we can provide identity to the absolute majority of the genes in this tissue. According to these definitions. This is about the 20 cell types that were neurons that we want to map.

So we're involved in two Human Cell Atlas pilot projects. One is to compare this spatial cell typing approach with other spatial cell typing approaches. So I think we have basically all published variants of spatial expression profiling methods in the consortium. And it's led by Allen Brain Institute. We're going to test these methods on mouse brain tissues and human brain tissues that they distribute to all participants. And then we generate data and send it back to Allen, and they will interpret and draw some conclusions. And the general idea is to see whether those technologies are mature enough, or then we would see some convergence hopefully between the methods. If they all give us different spatial maps, maybe they are not really mature yet. We will see.

Another is a human developmental cell atlas coordinated by Stenlin Årsjön and Karolina Skjaer.

And here we are going to try to make spatial maps of human developmental tissue, complete embryos and parts of embryos, brain, lung, and heart. We'll go deeper into, but we also try to make complete maps of embryos. And here we're using single-cell RNA sequencing, we're using spatial transcriptomics, and in situ sequencing and try to merge these kind of datasets.

And we have done an initial trial just to see whether we can go from spatial transcriptomic data to generate the targets for the in situ sequencing. So this is how spatial transcriptomic data looks like.

What it looks like. It's a method where you take a section of your tissue, put it onto a DNA microarray of spatial barcoded oligonucleotide primers, and blot over the RNA to the slide and perform cDNA harvest and put into a DNA sequencer. And then you get full transcriptome-wide mapping of the transcriptome.

With a certain resolution. And this is the kind of resolution you get. This is the density of the arrays.

And so Joachim Lönnberg and colleagues, they selected a couple of genes that they thought had an interesting spatial distribution based on the spatial transcriptomics. And we converted that into an in situ sequencing assay. And here I'll show you an example of these 4 genes.

So this MYH7 has a prominent expression in this part of the developmental heart, and this is exactly where we see it as well, of course. And the OGN is expressed up in this corner, and this is really where we see it, and so on. So it's really a very nice correlation.

But of course, we get a more detailed view since we have single-cell resolution, so we can look, we can see whether there is co-expression or not, and we can define— get a more fine definition of the distribution of these. So it looks like a very promising approach, not only to go from single-cell RNA sequencing data, but also using the spatial transcriptomics data to inform which markers to use.

So to sum up this, So what I didn't say, but maybe was clear from—

because we can do sort of wide-field imaging, unlike many single-molecule FISH-based methods, which use 100x objectives, so you really have to do imaging of very faint fluorescent spots. We have super bright spots, as we use 20x imaging, which makes our imaging at least 100 times faster than the single-molecule.

Single-molecule FISH imaging. So we can— we usually put 4 brain sections of mouse brain sections on a slide, which gives us about 400,000 cells information per slide. And imaging is done in a couple of hours. And we use a standard epifluorescence microscope. And it's just reagents. It's just enzymes and oligos. It's nothing expensive.

It's essentially a 3D method, but we're not really using that.

But we're going to go there. It works on, on the formalin-fixed paraffin-embedded materials.

We have done now also human brain samples with high autofluorescence, and still our signals are prominently bound, that kind of autofluorescence that you have in human brain tissues. We can map isoforms, and it's really up and running now. We have worked out automation of microscopy and and the biochemistry. So it's— and we're setting up a core facility at SciLife Lab in Stockholm to offer this kind of service of mapping transcripts in tissues. I'll just finish with this one showing that sequencing isn't that hard.

So we thought since we have these very bright fluorescent objects, it should be bright enough to be able to be captured by a mobile phone camera.

So we engaged in this collaboration with Eydogan Özkan at the UCL, who is an expert in mobile phone microscopy.

And he constructed this 3D-printed device that you can put on top of your mobile phone, aligning the lens of the mobile phone with a 2x lens in this device and then equipped it with some very cheap, simple lasers like this one, the green one. And we use that and the white light LED as well, so we can make bright-field images. And on top of that, we get the fluorescence imaging from DNA sequencing chemistry. So we're really actually sequencing DNA using a mobile phone, and we could use that to map mutations also in tumor tissues. So, prominent contributors present and former group members and collaborators and funders.

Thank you so much.