0:01
Hello everyone.
0:01
Welcome back to the session.
0:03
Welcome back to the best room of in this conference.
0:07
We have a few more talks lined up for us before we break for lunch.
0:12
And the first to speak will be Fabian Svara.
0:14
He's the CEO at Ariadne.AI and he's been building software to scale image analysis since 2006 and Co-founded Ariadne.AI with the same goal in 2014.
0:27
And there are a few items that he would like us to think about before he comes up is why should we go for a cloud based solution and why and how do we deal with batch effects?
0:38
So I hope you have some answers for them.
0:41
Welcome to the stage.
0:48
Thanks so much for the kind introduction.
0:52
Yeah.
0:53
It's a great pleasure to introduce you to Ariadne.AI, the work we do there and specifically our recent advances on unlocking neuroscience applications in mid to high Plex proteomics data, which brings with it some specific challenges that the software and the services that we provide are quite uniquely geared to address.
1:14
So before I talk about that though, I would want to start with a few slides just on the history of Ariadne, who we are as a company of where we come from.
1:24
So we've been around since 2014.
1:27
We're a team that provides software and software related services to help our clients process microscopy data that's in academia and an industry I would say almost completely agnostic towards microscopy methods.
1:42
We originally back in 2014 started building software for applications and three-dimensional electron microscopy.
1:49
We're really, there wasn't any software that was able to handle the deluge of data that was generated or started being generated back then.
1:58
And we've since really branched out to almost every other microscopy technique where usually the common thread is where large amounts of data are being generated and where it's not really that easy anymore for a single person to process or to handle that data actually can I see the time somewhere?
2:25
All right, thank you.
2:27
So one statistic that I just want to briefly highlight is that at least for those clients where we do custom projects and not just the off the shelf software, 25% of those clients publications are in high impact journals enabling beautiful covers as you're seeing here.
2:46
So our clients are really distributed all over the world with a clear geographical focus though in Western Europe and Northern America and include some of the, you know, largest industry players and academic institutions.
3:01
So we offer a number of different products or services.
3:05
First of all, the model engineering service, which is as the name sort of suggests, a service where we're going to build a machine learning model or AI based model for the segmentation analysis of one particular task built to the specifications and built according to the requirements of the client for a particular batch or project, basically of images acquired with a specific microscopy technique.
3:31
Then we've got the software engineering service, which is essentially a bioinformatics engineering service.
3:36
We've got the model as a service, which is essentially models that have been built at some point in the past, now available online as a product.
3:44
That's 3dEMtrace for electron microscopy, LMtrace for neuron tracing in light microscopy data, tomoTrace for electron tomography.
3:52
But then what's of course, much more relevant in the context of this meeting is our spatial software.
3:58
Spatial is a browser based tool for the analysis of multiplexed images.
4:03
So it's a scalable, it's easy to use, it's browser based, it's it runs in the clouds so it doesn't require any particular hardware on the clients.
4:10
And importantly, it's really device agnostic, meaning as long as the data is broadly in the format of an image and that includes things such as single cell resolution, transcriptome data as well, like Xenium or CosMx.
4:26
As long as that's sort of the type of resolution that you're working with, it's going to be possible to just upload the data there and the tool will convert it in a format that makes it possible for you to use it.
4:37
So that's kind of one of our promises that we will support any data format that you throw at it as long as it sort of fulfils these basic requirements of being single cell resolution, doesn't require any coding or special hardware.
4:50
I do want to highlight a few points that really are I think important and interesting to think about as when you when it comes to the question like of why does it make sense to have software in the browser?
5:05
Like why not build some, I don't know, high performance standalone tool that will run on the desktop.
5:13
There are really a couple of points that I think when it comes to large amounts of image data, as many of you here are working with, set that apart and make it in fact quite important for software to be browser based.
5:24
For one, there's a scalability issue.
5:26
So you're not going to process hundreds, maybe thousands of Multiplex images on a single workstation.
5:32
It's just not possible, at least not with state-of-the-art algorithms.
5:35
So you have to scale out.
5:36
What that means is you're going to have to probably work with a cluster of high performance compute environment and all of this is quite technical work that very few groups are going to want to do.
5:48
So that is something that a browser based is to directly gives you access to in its back end.
5:53
And then the collaboration features are something that I want to highlight in particular, meaning that similarly to how you would use, for example, to like Google Maps, you can just, you know, send a link to a location to a friend, a collaborator, a colleague, exactly the same thing as possible with a browser based image analysis tool.
6:10
You just send a link to what you're looking at in that moment and that link basically becomes a unique reference to exactly that location in your data.
6:19
Also, it enables continuous model fine tuning like where we don't have to ship a model and then hope it's going to work for future data sets as well.
6:28
We can constantly improve models as the platform becomes, as a platform sees more images, so to speak.
6:36
So what exactly is the workflow that spatial enable?
6:39
So we start from the left here with data ingestion and go to registration.
6:44
So we can register elastically between layers and within layers.
6:49
So that's to correct for image deformation introduced by all sorts of different factors, for example, heat induced epitope retrieval or stripping protocols or maybe by just consecutive sectioning.
7:04
So these are all errors that can be corrected to near pixel level precision with this registration pipeline.
7:10
Again, it's an elastic registration pipeline.
7:12
Then once you have that, it's optional.
7:14
Of course, not all data requires it, but once you have that, you can do the masking, cell segmentation and tissue segmentation models, free trade models that the software comes with.
7:23
You can perform marker mapping subsequent to that to get the single cell level vector so to speak.
7:31
That describes the marker expression for a single cell.
7:34
But not just the marker expression, also the morphology of a single cell.
7:39
Once you have that, you can either do old school gating really as if it was a FACS experiment.
7:45
You just define in single cell space what cells you're interested in, then do subsequent analysis on these groups of cells that you've defined.
7:53
Or you do dimensionality reduction clustering to do it automatically.
7:57
Once you have that you can do spatial analysis and I'm going to give a few more concrete examples of what that actually means later on.
8:05
So the software really gives you access to a number of different pre trained models for cells and nuclei, for tissue regions, for artefact suppression, for market positivity.
8:15
So that's a model that decides that takes the binary decision whether a cell is positive for a marker or not to suppress, for example, bleed through between neighbouring cells, but also to give you a more clear signal for markers that are sparsely expressed or that are localised in a small compartment within the cell and for registration.
8:35
And then the applications, these are, you know, that's not an exhaustive list, but some of the key applications are where you have complex cellular morphologies and segmentation with, you know, other techniques.
8:47
Maybe it doesn't lead to satisfactory results anymore.
8:51
And you run into that problem in a number of different tissues and many tissues really, but particularly prominent.
8:57
The problem is particularly prominent in neuro applications, immune applications, skin, muscle.
9:03
You can do multi omic registration thanks to the elastic registration pipeline between Multiplex IF, H&E and also transcript.
9:10
If they're on a different section, you can do a serial section registration.
9:15
So if you have actual consecutive sections and want a three-dimensional reconstruction, you can use that as well to do that.
9:20
And you can do the tissue annotation for brains for brain tissue, for neuroanatomy.
9:28
All right, so let me just run through a number of the key features that I want to highlight.
9:33
First is really the scalability.
9:35
So because it this runs in your browser, all of the data is streamed over the Internet.
9:40
You don't need to use a lot of storage on your workstation.
9:42
You just kind of browse it and the data is loaded as you go along.
9:47
The storage system in the back end is very large, so you can upload hundreds of data sets easily.
9:53
Finally, the pixel level segmentation of cytoplasmic and membrane bound markers highlighted here for EBA1, NeuN, and GFAP in the human cortex.
10:04
So that will allow you to segment cells in the brain.
10:08
Most cells in the brain really all glia neurons essentially at the level where you by I would also be able to do it and where it's really, you know, the image quality, the section thickness and so on that becomes limiting.
10:25
Here's another video showing the same thing.
10:27
This is GFAP, so astrocytes and showing how you can really follow into the neuropil out from the soma using the segmentation pipeline.
10:38
That's the segmentation tool that's part of this package.
10:42
This also works for completely other types of tissue.
10:45
This is actually a lymph node sample processed with haematoxylin.
10:49
So that's a very different kind of staining Right field with the segmentation is very high quality there as well.
10:57
And if you want to learn more about this particular application, I'm not going to talk about it anymore because I want to focus on the neuroscience, but we have an application note and a paper at our booth, if that is relevant to you.
11:09
And OK, last thing I want to just highlight with this video is the segmentation.
11:16
Yeah, I'm not sure if that's playing anyway that should have highlighted the registration.
11:24
And then the 3D support you can just, you know, it's just one more dimension that you can run all of the workflows on.
11:32
All right, So let me get a little bit more concrete now and talk about neuroscience applications as promised that this is data that Bahareh Ajami at OHSU generated.
11:42
This is a Akoya PhenoCycler data from postmortem brain, human brain from a brain bank.
11:50
And she shows she has something like 40 markers in there, protein markers and uses the GFAP, DAPI, IBA1 and MAP2 for cell segmentation and then amyloid beta to relate the phenotypes of the cells to more to pathology.
12:13
Now she used spatial to 1st segment the blood vessels for artefact suppression, then the cells again, that's from the markers listed up here, the amyloid beta aggregates and then the amyloid beta aggregate model will actually subdivide the aggregates into core that's pink and diffuse in blue.
12:38
And as many of you have probably heard, it's really the core aggregates, the dense core amyloid better aggregates that correlate very much with or much better than the diffuse ones with the symptomatic presentation of Alzheimer's disease.
12:55
And Bahareh and her lab are really quite interested in understanding how plaques and Alzheimer's and in particular core plaques correlate with microglial activation.
13:07
And the whole idea behind this experiment was finding subtypes of microglia that would somehow potentially correlate with amyloid beta core plaques in diseased brains.
13:22
So then she also used the neuroanatomy model, which gives you in this case the cortical layers and then the white matter down here.
13:35
And next she was interested in, well, first of all, obtaining the segmentation.
13:42
I think I already said it as GFAP believe that's EBA1.
13:46
Oh yeah, that's amyloid beta.
13:47
And that would yeah, I'm not sure.
13:54
Oh yeah, sorry.
13:54
That's what I was looking for.
13:57
So we've got the NeuN, GFAP, EBA1, and Olig2 segmentation from those markers respectively.
14:03
And as you can see again, the key point here is that it tracks precisely the morphology of the cells to the extent that is actually possible in those images given the imaging conditions of sample preparation and the sample thickness.
14:17
And that is of course, very important when it comes to cells such as the microglia where, you know, there are, it's thought that their morphology changes a lot with their activation.
14:30
But then also in proteomics is particularly in the brain, it's important to be able to map markers, not just if they're close to the nucleus, right?
14:39
I mean, if you do it nuclear segmentation and then expand the nuclei and use that to map the markers, you're going to miss all of the stain that's more distally detected either out in the neuron or somewhere out here in the sheets emanating from the astrocytes and the microglia.
14:57
And then the other thing is that if you have this type of segmentation, you can actually use that to measure to create morphometrics or actual quantitative measurements or quantitative descriptions of the morphology of a cell and use that for gating as well.
15:17
So use that as another input or for dimension reduction, clustering or manual gating.
15:24
OK, so with all of the things in hand, what Bahareh did was exactly this, basically looking at microglia in this case, extracting morphometrics and doing dimension reduction and clustering.
15:36
So, and all of this, what I'm showing is these are obviously figures from a manuscript, but all of this you can do in spatial in the browser based tool directly.
15:43
So she did clustering based on morphology, finds a number of different subtypes based on morphology, but then also on protein expression.
15:53
And she names these clusters here.
15:55
So this is the heat map showing a selection of the markers and then the protein based clusters, MO, PVM, and BLM and so on.
16:05
That's, you know, just the names of the clusters.
16:09
And then the next step for her was then to actually compare between cohorts in order to create hypothesis really that have hopefully mechanistic relevance and relate that to mechanisms of really microglial activation in the brain.
16:24
So we've got the AD cohort on the left, the control cohort on the right here.
16:28
And then again horizontally you hear those are the protein based clusters from the previous slide.
16:34
And what you find is that there's really this one population here, BLM where they differ quite a bit between the AD cohort and the controller cohort.
16:43
And then the next step obviously was to ask whether those somehow interact with the amyloid beta plaques.
16:52
And to this end, she used the pathology segmentation pretrained model from spatial to get the dense cord versus diffuse segmentation and use that then to relate those protein based clusters to diffuse plaques on the right and dense plaques on the left.
17:13
So what we're looking at here, those are histograms, we're looking out from the plaque so that these are short distances or 0 Micron distances.
17:21
And then it goes further away from the plaque towards the right.
17:24
And what we can see is that there's really one colour here that differs substantially between the dense and diffuse.
17:30
And that's the orange and that's the here, the BLM population, which is something that is basically exactly what Bahareh was looking for.
17:42
So basically hoping that's really what she was hoping to find that there's one particular microglial sub population that associates with dense a better plaque specifically and which might be part of the body's immune system response to those plaques, which hypothetically might contribute to the downstream deleterious effects of those plaques.
18:03
So with that, I just want to briefly say that there's a many of our clients and collaborators have nice things to say about us and we're very happy to show you much more at our booth.
18:17
And thanks for the entire spatial team that built the software and thanks for your attention.
