0:00
Our next speaker is Beatrice Awasthi, who is right now doing her postdoc in Doctor Will Huang's lab.
0:07
So she has a lot of experience working with single cell spatial technologies.
0:12
And today she's going to be talking about her project in PDAC samples where she has done high plex profiling using CosMx instrument.
0:21
Please welcome Beatrice Awasthi.
0:36
Well, thank you for the introduction.
0:38
My name is Beatrice.
0:39
I'm a postdoc in Will Huang's lab at Master General Hospital.
0:44
And today I'm excited to talk to you about our work on whole transcriptome spatial molecular imaging and match pre and post treatment specimens.
0:54
So our lab studies pancreatic ductal adenocarcinoma or PDAC, which is one of the deadliest and most treatment refractory cancer types.
1:03
PDAC makes up around 3% of cancer diagnosis each year and is the third leading cause of cancer associated deaths across both males and females with an overall five year survival rate of just 13%.
1:15
So it's a pretty urgent clinical need to identify better therapeutics for these patients.
1:22
So there's a number of features in PDAC that make it extremely difficult to treat.
1:26
One of these is that these tumours tend to be very fibrotic, so they contain a large number of fibroblasts that comprise a very dense tumour microenvironment and these fibroblasts support pro malignant signalling, and the density also makes it difficult for therapeutics to diffuse in and out of the tumour.
1:46
Additionally, PDAC tumours as a whole tend to be quite immune evasive, so malignant cells usually have low levels of antigen presentation and the microenvironment is very pro inflammatory and immunosuppressive.
2:01
Malignant cells and PDAC also undergo quite a bit of metabolic reprogramming that gives them a growth and survival advantage.
2:07
And finally, PDAC tumours are highly metastatic, so these tumours typically metastasize early in development and patients are diagnosed at late stages, which really contributes to the low survival rates.
2:25
So a number of studies over the years have described different malignant cell states in PDAC and different studies have kind of described different labels, but the consensus has sort of been that there are two major cell states in PDAC, classical and basal like.
2:41
These cell states have prognostic and predictive value.
2:46
So tumours generally contain a mix of these different cell States and tumours with a higher proportion of classical cells usually have a better prognosis.
2:57
The thing with these bulk drive states is that they were identified and studied and primarily untreated PDAC patients, but usually PDAC patients receive upfront chemotherapy and radiation.
3:09
This is generally needed for them to become surgical candidates in many cases.
3:14
So our question a few years ago was what do cell states actually look like in treated disease and how does this compare to untreated?
3:25
So a few years ago the lab set out to begin to answer this question.
3:31
And this was work that was done before I joined.
3:33
So you'll see the names of the people who did the work at the bottom of the slides.
3:38
But we set out to look at this using single nucleus RNA sequencing, which we used to profile a cohort of treated and untreated PDAC specimens.
3:48
So we were able to annotate a variety of malignant and non-malignant cells within these tumours and further subcluster malignant cells into 7 distinct malignant subtypes.
4:01
So some of these, for example the classical and then also the basal like, squamoid and mesenchymal subtypes that we identified overlapped pretty well with what had been previously found.
4:11
But there were some cell states, most notably the neural like progenitor or the NRP signature that was described here that had not been previously described, showing how taking a more single cell granular look at taxonomy in PDAC can really help to refine the molecular taxonomy.
4:33
Then looking at untreated versus treated samples, we actually found that there was quite a difference in the malignant subtypes that were found in each.
4:42
So for example, the classical signature tended to be enriched in untreated samples but go down relatively in treated samples or is the nerve signature, it was very enriched in the treated samples showing that treatment did seem to have an effect on the malignant subtypes that were found in these PDAC tumours.
5:02
So this point we've identified these malignant subtypes using dissociate single cell approaches.
5:08
But the question is really how are these malignant subtypes actually interacting within tumours as this is really driving tumour phenotype.
5:16
So we turn to spatial profiling technologies.
5:20
You just heard a really nice overview on the different technologies, so I won't belabour this, but just broadly we consider spatial profiling.
5:28
There's really 2 main branches, sequencing based spatial omics and imaging based and they both have their strengths and trade-offs.
5:35
So sequencing based approaches tend to allow for unbiased analytic coverage at the cost of resolution.
5:43
It's harder to achieve sub cellular resolution, whereas the imaging based approaches do allow you to achieve cellular resolution.
5:51
However, the number of RNA probes that can be included in a panel for a given experiment could be a bit more restricted.
5:59
And this figure was taken from a review that a few of the graph students in my lab published around a year ago.
6:07
So for the initial exploratory analysis of the spatial environment of PDAC tumours, we elected to go with an unbiased whole transcriptome approach using genomics.
6:21
So with this workflow, what we do is we take H&E sections and use that to guide the selection of regions of interest or Rois in consecutive tumour sections that are then profiled by the genomics.
6:36
So within the genomics, protein stains are used to guide the selection of specific cell compartments, so for example, epithelial versus stromal.
6:46
And then within those cell compartments, transcripts are quantified within specific regions of the tumours.
6:54
We then map the malignant signatures from our single nucleus RNA sequencing on to the slide to identify subtypes in their distribution.
7:03
So using this approach, we're able to identify 3 distinct communities of interacting cells.
7:09
Treatment enriched, squamoid, basaloid and classical, and each community featured a specific malignant subtype or combination thereof, interacting with a very specific combination of immune cells and fibroblasts, which led us to hypothesise that potentially different malignant subtypes are participating in different cellular interactions within these tumours.
7:33
So we really wanted to actually know what these intracellular interactions were, given how important malignant cell microenvironment signalling is for PDAC progression.
7:45
So to be able to study this at single cell resolution we turn into spatial molecular imaging using CosMx and the workflow here is similar in that we use H&E stains to guide the selection of regions of interest in the tumours.
8:01
And then for the actual spatial molecular imaging, labelled probes bind to the slide surface and iterative cycles of imaging are used to identify transcripts and localise them to the slide.
8:17
We then use protein stains to further guide the identification of specific subtypes, including membrane stains to help delineate single cells.
8:26
And putting the protein and transcripts together, we can then assign specific transcripts to certain cell types across the slide.
8:36
So for this initial spatial molecular imaging experiment, we used 1000 Plex RNA probe panel that contain probes against all those cell types we were interested in including malignant subtypes and we applied this panel to again a cohort of treated and untreated PDAC patients.
8:57
So we're able to clearly distinguish malignant and non-malignant cells and then further subtype these into specific subtypes of malignant immune and fibroblast cells.
9:11
So for the intercellular interaction analysis, we developed an algorithm called spatially constrained optimal transport interaction analysis and this was spearheaded by Jingyi and Martin Hemberg's lab.
9:25
So what this algorithm does, we termed it SCOTIA for short and is it takes adjacent source and target cluster pairs.
9:34
So these are known ligand receptor pairs just from previously published databases and it identifies them on the slide and then applies a cost function that takes into account spatial distance and then expression levels within cells that are proximal to each other of both the ligand and the receptor.
9:53
And then based on this cost function, it identifies likely interacting cell types for certain ligand receptor pairs.
10:03
So applying this algorithm to treated and untreated samples, we're able to identify a number of significantly depleted and enriched ligand receptor pairs in both the treated and the untreated samples.
10:18
So on the untreated side, we noted enrichment and WNT signalling and also collagen deposition type of pairs.
10:24
And on the treated side, there was enrichment of chemokine, cytokine signalling and also cell migration.
10:31
Well, one thing that really stood out to us was enrichment of ligand receptor pairs involved in IL6 family signalling in the treated pairs.
10:38
And this was notable because IL6 has long been reported to be implicated in therapeutic resistance in a variety of cancers.
10:47
And indeed, when we looked in vitro so tested the response of pancreatic cancer cells to chemotherapy with and without IL6, we found that the cells that were treated with IL6 were more resistant to the chemotherapy.
11:01
So this was exciting because it showed that our spatial data could actually be used to find therapeutically relevant ligand receptor interactions.
11:10
However, with the 1K data, we're only able to actually profile a fraction of known ligand receptor pairs.
11:20
So it was under 30% of pairs that were actually just included in the play.
11:24
So we wondered what we'd be able to see if we could actually profile these at the whole transcriptome level.
11:31
So recently you've been very excited to be working with the whole transcriptome panel that Bruker has recently published and will be releasing this summer.
11:41
So the whole transcriptome panel contains nearly 38,000 imaging barcodes to cover the entire human protein transcriptome and it is able to cover the whole transcriptome by largely increasing the number of imaging cycles by quite a bit.
12:00
It can cover a variety of transcript lengths, including very short transcripts and well covers every chromosome in the cell.
12:11
So we've been applying this panel to a clinical trial DF/HCC 18-469 in which we collect tissue from patients during their pretreatment diagnostic biopsy and then also during the post treatment surgery.
12:25
So this is surgery that's occurring after the conventional neoadjuvant chemoradiotherapy that patients typically receive.
12:33
And then we're able to profile the match pre and post treatment samples using the spatial molecular imaging.
12:40
And this has been particularly exciting because most PDAC studies so far have been on unmatched samples.
12:47
So this is the first matched spatial analysis of PDAC, which is really allowing us to with a more causative lens, look at pre and post treatment differences.
13:00
So for the actual imaging, we're taking the three pronged approach where we're again using H&E standing to guide the selection of regions of interest in our tumours.
13:09
And then we're performing the spatial molecular imaging at both the thousand Plex and the whole transcriptome levels.
13:16
And we've chosen to do it this way because one thing that we found is that the whole transcriptome data is so information rich that it's very high dimensional and it's quite sparse.
13:27
And we found that there are certain cell types that have been very difficult to annotate in the data.
13:32
So we're using, we're performing the spatial molecular imaging with both panel sizes on consecutive sections.
13:39
And the 1K data is much more easily annotated.
13:42
So by annotating that first, we can then guide the annotation of the whole transcriptome.
13:48
And it's been honestly very cool to see because of the whole transcript on level, we're just able to detect so many genes and transcripts.
13:56
So we've been able to detect a mean of around 1700 transcripts per cell and over 1100 genes, which rivals what we can get with dissociate single nucleus sequencing of these archival samples.
14:13
So we've had to test a number of annotation strategies to really be able to get the most out of this data.
14:19
And this slide is kind of detailing all the different things we've tried so far.
14:24
When you think about conventional annotation of spatial imaging data, frequently people use marker gene based annotation where they cluster the data and then use known signatures and map those onto clusters to assign cell type labels to clusters.
14:40
This worked relatively well.
14:42
However the two caveats are that it can first be a bit difficult, sometimes it can be arbitrary to assign a specific signature to a cluster, and the second is that this is just a very labour intensive approach.
14:54
So what we really wanted to do was be able to find a more automated pipeline for cell typing.
15:00
We then move on to trying InSituType, which is an algorithm that's been specifically developed by NanoString for cell typing of spatial data.
15:11
And this has worked really well for smaller panel slices, but we found that at the whole transcriptome level it was difficult to find a reference that could really help guide annotation of all cell types accurately.
15:24
So then we move on to our third approach, label transfer, which has actually worked quite well.
15:32
So in this approach we take a two-step annotation method where we first use the InSituType algorithm to annotate the 1000 Plex data using our single nucleus RNA sequencing data as a reference.
15:47
We then integrate the 1000 Plex with the whole transcriptome using an algorithm Harmony.
15:53
And because these are consecutive sections, the integration works very well.
15:59
And then we're able to just cluster the data and transfer easily the labels from the 1000 Plex, the whole transcriptome.
16:06
And this has led to nice annotation of almost the whole transcriptome data.
16:10
And you can see the nice separation in these UMAPs over here.
16:16
We then use protein and H&E stains to further refine our annotations.
16:21
So at the H&E level, pathologists annotate these.
16:25
We're able to identify broad features such as tumour, endocrine cells and nerves.
16:31
And then at the protein stain level, we're able to stain specific cell types of interest.
16:37
So we can set the specific intensity thresholds and ensure that all the cells were cell typing with the transcriptome actually meet these at the protein level.
16:48
And putting it all together, we end up with a nicely annotated whole transcriptome where you can really see the malignant glands, the nerves, where the nerves are expected, endocrine cells where endocrine cells are expected, et cetera.
17:03
So your software applied the whole transcriptome panel to four samples that we've collected and are actively in the process of collecting more.
17:12
So for the samples we've collected so far, we've just done preliminary analysis on cell composition shifts given we observed in our previous studies.
17:21
And he found that as expected in the resection specimens, epithelial proportions went down relative to the biopsy, while five or less went up.
17:31
This is what we'd expect as a treatment effect.
17:35
We also noted a decrease in the classical signature and a relative increase in NRP in the resections compared to the biopsies, which again matched what we saw previously.
17:48
We also, as I said, have been extremely interested in understanding intercellular interactions at the whole transcriptome level.
17:55
So we've again applied the SCOTIA algorithm to our whole transcriptome data.
18:00
Unfortunately, with only four samples, we haven't been powered yet to do statistics.
18:05
So what we've done instead is performed SCOTIA on pretreatment biopsies and resections and then aggregated the average likelihood scores for each.
18:17
So a higher average likelihood score is a higher ligand receptor interaction.
18:23
And you can see that we do have some overlap that we've previously seen.
18:29
For example, in the post treatment resection, we noted that a lot of these inflammatory cytokine signalling pathways weren't had high scores, which kind of matches what we've seen previously.
18:39
But was really exciting with this was that we noted that some of the ligand receptor pairs with the highest scores.
18:48
So these are the strongest interactions we've observed.
18:51
We're actually concluding genes that were not contained in the previous 1K panel.
18:56
So these are shown in red here.
18:58
It's a lot of these laminins, transference signalling.
19:02
So this has been exciting because it shows how using the whole transcriptome can actually illuminate things that you just can't see with smaller panel sizes and that you might not predict.
19:13
And it's just a nice way to be able to, in an unbiased way, discover new intercellular interactions.
19:20
So we're looking forward to collecting more samples and hopefully being powered for statistics soon.
19:26
So just to wrap up, when you think about the future, so you know, it's been such an exciting growth of the spatial technologies over the years with the whole transcriptome, and we've really pushed the boundaries of Plex size and sensitivity.
19:43
So just as we continue to move forward with our studies, we're hoping to continue to be able to increase the number of genes and transcripts we can detect in all of our samples and improve our annotation strategies and be able to really get more out of our samples.
19:59
Another thing that's on the horizon for us is hopefully in the future moving into higher dimensionality in terms of 3D and 4D.
20:06
So looking at tumour architecture actually in 3D and how it changes over time.
20:12
Also eventually including other omics modalities, for example, proteomics is something we're already exploring and then in the future, hopefully genomics and epigenomics as well to really get all we can from these scan patients samples and then finally getting as translatable as we can.
20:30
So we're using matched samples already, and we're hoping to be able to identify biomarkers and prognostic signatures in addition to potentially targetable interactions within our samples.
20:43
And with that, I just like to acknowledge everyone who's been involved with this work, the whole Huang lab, Nicole has been coordinating the whole clinical side of it with Mari and Nick, pathologists on the team, Martin Hemberg's lab and Jingyi have really spearheaded a lot of the computational side.
21:01
And then all our collaborators at Bruker, Ashley, Max, Prajan, everyone else, and thank you for listening.
