Ariadne.ai SPATIAL - Unlocking Neuroscience Applications in Spatial Omics

 If I pronounce that correct, CEO from ariadne.ai and he is the Co-founder and he will speak about unlocking neuroscience applications in spatial omics. 

 
Welcome. 

 
Hi, everyone. 

 
It's my great pleasure to talk today about the work we do at ariadne.ai on unlocking neuroscience applications and spatial omics amongst many different things that we do. 

 
So before I go there though, I would like to just use a couple slides to introduce who we are as a company. 

 
So we are a company that was found in 2014 with a focus on really unlocking biomedical image analysis problems for our clients. 

 
Basically the common thread being that whenever a lab has some degree of automation microscopy and the data analysis in itself becomes the bottleneck, that's where we can come in and build custom software, but also provide access to ready made products for that purpose. 

 
So our team is more than half PhDs and we've been doing this for quite a while and have consistently helped our clients through first class science. 

 
So our clients agree with this assessment. 

 
So they're quite happy with what we've done for them. 

 
And our clients are distributed essentially all over the world with a bit of a geographic focus in the US and Europe, but really all over the world. 

 
So our product line encompasses first of all model engineering. 

 
So what we mean by model engineering is that we're going to be able to build custom tailored machine learning models for a particular application. 

 
And that is an end to end engineering workflow that includes, well, obviously the computer science side of things, but also any amount of data labelling and annotation that might be required to get a state-of-the-art result, which is really sort of the one of the dirty secrets of modern AI is that it doesn't work without that really. 

 
So in addition to that, we've got the software engineering service. 

 
We provide access to models that we've built previously as a service for 3D electron microscopy, confocal microscopy, light sheet microscopy. 

 
LMtrace is the product for microscopy based neuron tracing and then tomoTrace for the reconstruction of sub cellular structures in TM tomography. 

 
But then we're here really because of spatial obviously spatial is a software as a service product. 

 
It's available through a web browser based on a subscription model and it allows you to do end to end analysis of well, mid to high Plex image data, whatever that may be. 

 
So that includes transcriptomics and includes proteomics, but also metabolomics like mass spec based technologies, for example. 

 
So what spatial is really is a device or data type agnostic software and workflow that you access through your web browser. 

 
So because you're accessed through your web browser, you don't need to bring any special hardware or engineering don't know how the pipeline really the heavy lifting computationally runs on our high performance compute infrastructure, the browser. 

 
So the software that runs on your own desktop PC or whatever device you might have is a very lightweight one that's only used for visualising the input and the output. 

 
So today I want to highlight though the unique capabilities that spatial offers specifically in neuroscience. 

 
And I'm going to talk mostly about spatial proteomics-based technologies. 

 
And among the unique capabilities of spatial neuroscience are the detailed delineation of neuronal and glial morphology. 

 
So that gives you really precise morphometrics for all of the cell types in the brain, the astrocytes, the microglia, the neurons, also endothelial cells actually. 

 
So that's really the whole gamut of cell types in the brain. 

 
And it allows you to do a precise classification into subclasses of the various types of pathological structures, typically pathological protein aggregates that are found in the common neuropathology such as Alzheimer's, Parkinson's, frontotemporal dimension, others. 

 
So to be specific, that means that's the aggregation of amyloid beta, of phospho Tau, of alpha synuclein TDP 43 and we are Co-marketing the software for these purposes with a number of spatial proteomics and other device manufacturers. 

 
So to just introduce the general workflow, I think this is not going to be surprising in any way. 

 
That's sort of the baseline workflow that many people in the community use with a couple of details though, that I want to highlight. 

 
So we start with data ingestion. 

 
In our case that means data transfer. 

 
So that data transfer, as long as it's not multiple terabytes, can easily be done through the web browser. 

 
Otherwise there are alternatives. 

 
Now if you have images that are correlated between different modalities for example or between consecutive sections, the first thing you would do would be to apply the registration workflow. 

 
So the registration workflow allows you to do really single cell pixel perfect registration. 

 
So matching between correlated images. 

 
For example, one application that we see this often done for is when our clients want to correlate Visium with Xenium data on consecutive sections. 

 
But also if you have some IF staining protocol that proceeds over multiple cycles and that those stripping steps that occur between the cycles introduce light nonlinear deformations between the consecutive cycles. 

 
Those can be corrected with this registration workflow as well. 

 
So the registration workflow I want to highlight is fully elastic at one. 

 
It's a computationally very expensive algorithm actually. 

 
But because it runs in the cloud, you can really do it quite easily through the web-based environment. 

 
And subsequent to that, like if you had to do that, it's optional, you would run a number of segmentation models for tissue segmentation. 

 
So that gives you tissue subtypes of neuroscience specifically that would potentially be different cortical layers or different hippocampal regions. 

 
You can do cell segmentation. 

 
That's what I mentioned just in the previous slide where you get the precise delineation of morphology and masking also. 

 
So a very common issue really is that you want to exclude certain regions from analysis. 

 
And to this end, you can apply the masking model which will suppress analysis in certain regions that are affected for example by I don't know what one of the common artefacts tissue folds, bubbles. 

 
Subsequent to this, you perform market mapping at which point it sort of it transforms into something that's more similar to a single cell analysis method where you have just feature vectors for each cell. 

 
But of course with the spatial context retained on that. 

 
You do dimensionality reduction for automatic clustering or just old school gating based on your biological priors. 

 
You define cell types and do downstream spatial analysis. 

 
So again, this the broad workflow, this workflow, broadly speaking, I think it's fairly standard. 

 
But what really what's quite unusual about this is that we provide access to a number of pre trained models that we will also fine tune for our clients as part of the subscription package. 

 
The models that are available are for cells and nuclei from all of the common markers in the brain, for tissue regions, for suppression of artefacts, for scoring of marker positivity, and for image registration scoring of marker positivity. 

 
What that refers to is basically that instead of gating to define cell types or just clustering to define cell types, you would actually define the positivity of a given cell from its segmentation based on the raw image for each marker. 

 
And similar to how a person would do it to say OK, this is this cell is clearly positive or negative for this marker. 

 
This would be done by a pre trained model and that unlocks a number of applications such as the analysis of complex cellular morphologies common in the brain but also in other tissues such as skin and muscle for example. 

 
But then also multiomic registration where you can correct the deformations that are introduced by aggressive chemistry or by sectioning and correlate Multiplex IF with H&E with transcripts, for example. 

 
So that also enables you to do serial sections and move to 3D analysis, for example. 

 
So I just want to show a couple of videos to highlight those things that I mentioned sometimes in passing a bit more visually. 

 
So this is sort of what we can do in terms of visualisation business. 

 
Yeah, you can really zoom in from the very fine detail to very large whole slide images that would be multiple centimetres on each side. 

 
That works in 2D and in 3D. 

 
Also we can do the segmentation here based on the markers GFAP, Eva, ONE and UN to get the microglia, astrocytes, microglia and neurons respectively. 

 
And that is a workflow that works. 

 
You just sort of request it through the browser-based interface and then it runs in the cloud or on our compute infrastructure. 

 
That's another visualisation of that, showing how we're capturing the fine arborizations emanating from the Astrocytes. 

 
Based on the GFAP marker. 

 
We can do segmentation on diverse stains. 

 
This is actually haematoxylin, slightly unusual, not even H&E haematoxylin, but that works as well. 

 
This is lymph node tissue actually. 

 
And so if you want to inspect those segmentations up close, feel free to drop by our booth. 

 
We have live demos and can show you. 

 
It's a little hard to judge from videos like that. 

 
Of course that is showing the elastic registration workflow here used not for cross cycle or cross section registration, but rather for whole slide montaging from a series of partially overlapping tiles. 

 
That's a mosaic where there's some lens distortion at the seam. 

 
There's like you can see this cross sort of coming. 

 
That's the tile seam. 

 
And as you can tell, the deformation is really very non linear, like it sort of moves diagonally at the corners of each tile. 

 
And the elastic registration workflow picks up on that and fixes that locally. 

 
And yeah, we support 3D as well. 

 
It's a modality that's at this point somewhat rare still, but we as a company originally come actually from large scale 3D analysis and microscopy. 

 
So this is something that we've built in from the very start. 

 
So for those still somewhat rare data set where that's relevant, we can do that. 

 
Now I want to switch gears a little and move towards a more specific application. 

 
So we did this in collaboration with Ajami lab at OHSU. 

 
That's Oregon Health and Science. 

 
She has a phenocycler data that's [unclear], sometimes also referred to as Codex, where she stayed for about 45 or so proteins and the cortex. 

 
Those are human postmortem samples from a brain bank and she's staying for amongst other things, those markers shown on the upper left. 

 
So those the last four, DAPI, GFAP, IBA1 and MAP2, those are used for cell segmentation. 

 
That gives you well the nuclear obviously, but then the astrocytes, microglia and neurons respectively, but then also amyloid beta, that's the aggregating protein Alzheimer's disease. 

 
And we can use those markers to detect the blood vessels in her case, she actually used that for masking because those caused some false positive stain. 

 
But you might also do spatial analysis based on blood vessel proximity, for example. 

 
The cells, so they're shown in rainbow colours here because you know, we're emitting instances based on those markers. 

 
And then also the amyloid beta stain can be used for first of all detection and instantiation of the plaques, but then sub classification also. 

 
So pink versus blueish here is dense quart plaques versus diffuse plaques, which is quite helpful to know because it's actually, as some of you may have heard, probably the core plaques that are actually toxic or that cause a lot of the disruption and Alzheimer's. 

 
And finally, the software allowed her to do this classification of layers in the cortex. 

 
So that's layers 1 outside towards and then 2,3,4,5 towards the inside of the brain. 

 
So what has she used that for? 

 
She used that to get, first of all, as I mentioned, a precise segmentation including the morphology of all of the cells in her samples. 

 
That's neurons, astrocytes, microglia and then oligodendrocytes respectively in the upper right panel. 

 
And she uses that then to detect, so to classify them or to cluster them based on morphology and then to look at proximity based on that classification also. 

 
So that's what we're showing here or what she shows here really. 

 
This paper is currently in review in Nature Neuroscience, so fingers crossed you'll be able to actually go into the details yourself very soon. 

 
There's this GFAP positive cell, I believe. 

 
No, sorry, that's a microglia IBA1 positive. 

 
We extract all of those morphometrics and do clustering based on the morphometrics and that reveals a number of subtypes within the microglia class. 

 
And we can do that or they did that on the proteins as well. 

 
So with this combined classification based on morphometrics and proteins, they have been asked whether there are systematic differences in how these subclasses occur in the Alzheimer's versus the control samples. 

 
And it turns out that indeed there's a number of classes in particularly, in particular the one they term BLM here that seems to be much more common in Alzheimer's samples. 

 
And also that associates with, but the Alzheimer's pathology the next they then use the classification of amyloid beta aggregates into dense and diffuse to ask whether those subtypes also differentially associate with dense versus diffuse plaques in the Alzheimer's cases. 

 
And that is indeed the case of the BLMs do very much prefer being in close proximity to the dense amyloid beta plaques versus the diffuse amyloid beta plaques. 

 
So they are very interested in pursuing this further. 

 
So far it's sort of like a correlation and observation. 

 
But since there's sort of this working hypothesis in the field or particularly pursued by Bahari and her team, that it's sort of the immune response, microglia responding to the presence of dense port amyloid beta plaques that really contribute to the massive disruption in the brain that is then sort of causing the behavioural deficits or the behavioural issues with Alzheimer's. 

 
And that is really the microbial response. 

 
And that this might then, and they're now testing the hypothesis more closely that this, the cell type is part of the pathway really of microbial activation that leads to these downstream disruptions. 

 
So I'm basically done. 

 
So I would love to show you more drop by the booth for more hands on or more close demos. 

 
I want to thank the R.E.M AI Spatial team in Heidelberg and in Lucerne and I would like to thank you for your attention.