Utilizing Spatial Biology & AI in Late Stage Clinical Development

0:53 
So any further ado, I'd like to hand over to Joe now to begin the presentation. 

 
0:56 
So thank you and over to you, Joe. 

 
1:00 
Thank you. 

 
1:01 
So I’m Joe from Singapore. 

 
1:03 
This is at night here, but good afternoon for most of the people. 

 
1:07 
So I mean, I guess the attendees from UK. 

 
1:10 
So today, my topic is on AI applications on cellular and tissue imaging in drug development. 

 
1:18 
When we talk about drug development, people will be mainly thinking that this is actually quite early stage, but if you think about it, is there in fact drug development is the entire spectrum of most of our biomedical activity. 

 
1:33 
So it's including what we just highlight here, drug screening, clinical trial recruitment and also precision medicine. 

 
1:40 
So in here I'm going to actually share with you AI spatial technology, how to apply across the spectrum. 

 
1:49 
This my disclosure. 

 
1:51 
So after the disclosure is my disclaimer. 

 
1:55 
So I have been actually talking about AI across the world, but I have to declare that I'm not a AI expert that we're not the ones that can do a lot of coding and all these things. 

 
2:05 
And I'm not necessarily the domain expert as well in some of the topic that I'm talking to, but I guess, you know, like we are in the fields long and deep enough to actually serve as a bridge to actually, stitch everything together. 

 
2:22 
So the outline of today is there, you know, talk about, you know, a spatial in the background as a background story. 

 
2:29 
And also how do we think that AI come to pathology and AI is really the end to end solution for entire drug development and hopefully can convince you about this and also share with you why pathologists and also researchers like us that come to the AI world mean like not only just chasing the hot topics, right? 

 
2:52 
There is a reason which I'm going to share with you and then also share with you like two case studies. 

 
2:57 
How do we apply all these things together? 

 
2:59 
And also route to translation. 

 
3:02 
So our group in Singapore, I have in the hospital as well as outside the hospital, we generally do three things right. 

 
3:09 
The histopathology, spatial is the centrepiece because of my training and also the immunology. 

 
3:17 
Our group also does a lot of flow cytometry, that kind of thing and also the AI component to actually stitch everything together. 

 
3:25 
So in internal spatial technology, spatial is a reasonable, it is like a two to three years terminate termination, right? 

 
3:34 
But if you talk about like early things, early technology in the spatial which is like multiplex, IHC and that kind of thing. 

 
3:42 
We has been in the field for more than 10 years, published more than 100 papers and our groups has translated a few assays to the market and also clinical setting, which I believe is something that people are very interested in. 

 
4:01 
And then let me also take the opportunity to share view of fields example. 

 
4:11 
So also like to actually share with you that there's this MxIF-SIG initiated under SITC and also NIH, which we lead and try to write a guideline and recommendation for the spatial technology for translation. 

 
4:26 
So as early as 2019, we actually developed the, I will say maybe the first in the world Multiplex IHC based assay for the clinical setting to actually stain three PD-L1 clones, which is SP142, 22C3 and also SP263 and to precisely quantitate them where she calls them at CK + Epcam CD45 to actually give you CPS, IC and TPS together. 

 
4:55 
So this is something that is really signed and on site in the hospital and then this is marketed out to the even to the region and Hong Kong, Thailand, Philippines, Vietnam, Indonesia. 

 
5:08 
So one test, one slide, three clones of  PD-L1 and I read out and if you want to report the hot and inflamed this is OK as well. 

 
5:19 
So this and in the recent years and as early as 2019 in the recent years. 

 
5:24 
This is actually one of the recent use cases on gastric cancer because gastric cancer needs more than one PD-L1 to make the decision. 

 
5:32 
So we actually have a study and then we reported that the 28-8 is more sensitive than the 22C3, which is the usual assay that people do. 

 
5:43 
And then we our report is actually even earlier than what CheckMate 649 actually later on subsequently validated in their trial and have the same conclusion with us. 

 
5:54 
So this study, you know, even as early as 2019, but in ASCO 2022, which is 2 years ago is come out again as news and also in the gastric cancer setting covered in multiple media. 

 
6:08 
And even because of this, the gastric cancer community in the oncology space, they call for this blueprint study to actually try to harmonise the PD-L1 scoring for gastric cancer. 

 
6:21 
So this is just to give you examples of how this how some of these things can really have a real world impact. 

 
6:30 
So you know, when we talk about Multiplex, I see nowadays in the post pandemic setting. 

 
6:36 
So we usually no longer talk about it's very basic Multiplex, like 6 or 9 cover. 

 
6:41 
We usually talk about, I would say more so we will talk about at least 20 something and then you know, we have been doing at least 40 and then they ask people they claim to do more than 100 as you know, right. 

 
6:52 
So I mean in my presentation, every time you see a QR code, you can take your mobile phone and scan the QR code and usually you can access the image via your mobile phone in the real-life basis. 

 
7:07 
You can toggle it, zoom in, zoom out, change colour. 

 
7:10 
And this is a platform that we set out for community. 

 
7:15 
All of us, all of you are also part of the community for data sharing, data deposition free and also even useful to put it in your publication. 

 
7:24 
If you're interested, reach out. 

 
7:26 
So nowadays when you talk about spatial technology its not only 14 level antibody levels, right, We want to talk about transcriptomics level. 

 
7:35 
For example, our group have this spatial transcriptomics machine for BGI and it actually can do whole, slide whole transcriptomics at single cell levels, 24 samples a day. 

 
7:48 
We have more than one terabyte of data. 

 
7:55 
So this is something that we are very excited for. 

 
7:57 
And also there are already quite some data coming out from there. 

 
8:00 
So this is something that I share with you just now, you know, like, I don't know why the video is not working. 

 
8:07 
I don't know whether it's working at your site, but this is basically the one that show you know, you could take a QR code you go into there and there is a platform not only built for data sharing. 

 
8:19 
And also there's the automation of quantification function there to just handle the big data. 

 
8:26 
So as I say, you know, like we in this part of the world, you know, like Singapore is far from most of you, most of the Western countries, right? 

 
8:34 
So we have, you know, like set this out, all these things for like 10 years anniversary last year. 

 
8:40 
We actually run this I believe is the first in Asia Pacific spatial events in May. 

 
8:46 
And most of the people so that you know, in the region and even some of this from like Carlo or this from flight from US and UK as well. 

 
8:54 
They join this event. 

 
8:55 
And we have bought more than 250 registrants across 12 countries. 

 
8:59 
And most of them are regional pathologists actually really come into our pathology lab to actually learn how to analyse the spatial data. 

 
9:12 
So we are running it again this year. 

 
9:14 
So you know, basically, if you are interested to come Asia Pacific to join some of this event, you know, for example, you know, which I told you our hospital’s immunopathology workshop. 

 
9:25 
And also there is our WIC in Shanghai. 

 
9:29 
And the end of the year, if you want to spend your Christmas in Australia, please come to this multiomics, in which I believe Jasmine is also one of our speakers. 

 
9:38 
So yeah, and talk about conference, we and finally we touching a little bit on the AI, right. 

 
9:44 
So two years ago we also brought this MICCAI, which is like largest medical imaging conference in the world that to Singapore. 

 
9:52 
And one of the highlights of the day is that we actually bring 2000s AIS like this to the night Zoo in Singapore, which I believe the hyenas were quite excited that night. 

 
10:05 
And this year this meeting is going to go to Morocco and then hopefully see you there as well. 

 
10:10 
So enough of this, you know, like a background with how we do things. 

 
10:15 
And you know, let me also come back to the story that I want to share with you, why we believe the AI and also spatial can be something that apply across the entire spectrum of biomedical science. 

 
10:28 
So you know, like as most of the pathologies we are excited that nowadays only certainly AI come to pathology and Paige received approval now. 

 
10:41 
But what really actually excites me is that AI comes to pathology, meaning that AI in healthcare, I finally come to something that is explainable. 

 
10:52 
It's come to something that we call biology, right? 

 
10:55 
Because you know, if the traditionally in the AI healthcare, it's like radiology, surgery, endoscopy, they are like just better recognition. 

 
11:04 
But in the AI pathology, we try to understand the morphology of these particular cells and the distribution of particular cell type and the cell interaction of tumour and immune cells, that kind of thing. 

 
11:16 
This is biology, right? 

 
11:17 
So as long as it is biology and it should be explainable, you should be able to design experiments to be right. 

 
11:27 
So it's no longer a black box. 

 
11:30 
That is the thing that really excites me. 

 
11:32 
And we know the entire drug development is very costly and high cost. 

 
11:43 
So if you look at the entire activity like the drug screening in the very beginning and all the diagnostic and precision at the end of the spectrum and the images that we generated, which potentially can fit for AI is similar, right? 

 
12:02 
If you show these images, any of these images to AI scientists, they're probably working the same and they are going to design a similar AI pipeline to crunch this data, right? 

 
12:13 
So I will say this is really something that we can apply across and forget in the middle of this spectrum is all animal study and animal studies also having a lot of imaging in the models, right? 

 
12:27 
So meaning that if you have AI algorithm to particularly recognise 1 molecular biology in the tissue level, it is based on imaging and it should be able to cut across, apply this algorithm from drug screening and design, animal testing, clinical trial and precision medicine diagnostics, right? 

 
12:52 
This is very real. 

 
12:55 
And there are already companies and pharma working on this. 

 
12:59 
And like, we are part of this project as well. 

 
13:02 
And as I said, you know, why do we actually come in, what role do we really play, right? 

 
13:11 
And what I mean like why we suddenly said that we can do all this AI. The key thing is that the AI is talking about ground truth. 

 
13:21 
How do you give the AI a true answer so they can benchmark this true answer and then predict using something else, right? 

 
13:30 
And hopefully it's even better than whatever label they originally give, right? 

 
13:36 
So that is something that as a scientist or pathologist, we actually can try to generate the data. 

 
13:42 
And we believe that we can generate this data very accurately and also fairly reproducible. 

 
13:47 
So like, why is that? 

 
13:48 
Let me explain to you. 

 
13:50 
So let's use a T cell assay as an example, like you know, in the single cell setting as far as in the tissue level setting, in a single cell setting, we can try to feed AI and single cell data for cytometry data. 

 
14:05 
And for your information, if you are not in the field nowadays for single cell, you can generate single cell images as well. 

 
14:13 
They are pretty high resolution. 

 
14:16 
So this is one of the options that you can give and then you know if you actually can do a single cell imaging basically like this, right? 

 
14:26 
Actually, if that's following my pointer, you know, you have the raw images of the single cell and you label the single cell with antibody base or other things, right? 

 
14:38 
And you know this particular cell type, the phenotype of it, right? 

 
14:42 
And then you have easily 100s of them, millions of them as a crowd true to feed the AI to build a very convincing algorithm. 

 
14:53 
So based on training, we can easily identify the TNF alpha for positive cells. 

 
15:02 
And then just based on these raw images face contrast and then you know, it looks like to you as naked eye, you don't see a difference, right? 

 
15:13 
But the AI algorithm can tell you this is the positive cells for this double positive or this is negative cells. 

 
15:18 
And in the AI for nowadays is explainable, right? 

 
15:22 
They even give you this heat map to show you why the algorithm thinks that this particular region is the what the signal that they actually capture to call this a positive or negative cells and then same things. 

 
15:36 
If you can understand from a single cell point of view in a tissue level, this is a spot you have whatever spatial technology is the high end as it is, and we nowadays can easily generate the H&E images on the same tissue, same section, right. 

 
15:55 
No longer the serial section is actually the exact same section that you do, you have done your spatial and you generate a H&E right? 

 
16:04 
So we have started this work as early as 2019 at that time every time we present, nobody believe it, right? 

 
16:12 
This is something that's easy, everybody actually doing it, right. 

 
16:16 
So if you can do that, meaning that based on your H&E images, which is a good standard for diagnostics, right? 

 
16:25 
You have everything in there. 

 
16:27 
If you do a spatial transcriptomics, you actually have 18,000 genes labelled on that particular substrate. 

 
16:33 
If you have protein levels, right, you have 40 marker, 100 markers, enough to categorise the character of the very precise subset of the cells, right. 

 
16:47 
Then you can again, you know, run through hundreds of thousands of them, millions of them. 

 
16:50 
You can build algorithms, right. 

 
16:52 
So look at just as easy as CD3 like a plain T cell marker, right. 

 
16:57 
You just to show you this case. 

 
17:00 
How do you register this right? 

 
17:02 
You have H&E, you have the multi place you register them and they are perfectly align and you train the system and how many you put this algorithm, which actually can be as high as 90% accuracy. 

 
17:15 
And you use a H&E to predict them and just give you the label. 

 
17:21 
And this is just to talk about the technique expand the availability of the AI, which I'll skip it now. 

 
17:32 
So this is something we call it virtual staining. 

 
17:34 
So this particular image you see, so if you can see the green colour cells, they are not the actual staining cells. 

 
17:42 
They are the AI generated stain on these particular unseen H&E images. 

 
17:48 
Then to show you this is the T cells. 

 
17:53 
So an example very quickly, this is talking about tumour specific T cells, which I won't go into details about the biology, a lot of this is already published. 

 
18:04 
And then but let me give you why do we look at this particular tumour specific T cells, right? 

 
18:10 
The entire main research question is that. 

 
18:12 
So we all know that in the cancer immunotherapy setting, the PD-L1, TMBs are the two clinical biomarkers that are really routine setting, right? 

 
18:26 
So example in NSCLC, the melanoma setting, but many biomarkers that already was, you know for the past 10 years, right? 

 
18:36 
A lot of this is validated across multiple studies, multiple sites, huge cohorts, right? 

 
18:42 
But very little of this actually translated to the clinics and even easy cues and our CD3, CD8, IHC staining like immunoscore is having a huge difficulty to really reach the routine practise right. 

 
19:07 
And then CD8, CD3, IHC staining is not costly as well. 

 
19:11 
Why are we not doing it? 

 
19:12 
So if we are not doing all these things I think it will be difficult to translate anything above this in a spatial setting. 

 
19:27 
How do we solve this? 

 
19:31 
I believe AI based pathologies is really the solution and that will translate the traditional way. 

 
19:40 
So you have a H&E and some of this site you choose to stain particular marker, for example, CD3, CD8, PD-L1, you make a diagnostic or this is a precision medicine setting, right? 

 
19:50 
And in something called the Next generation of Digital pathology of H&E 2.0 in a 2.0 setting you have a H&E images and then your AI algorithm. 

 
20:06 
Let's say have 100 or even 1000 algorithm there that predict CD3, CD8, PD-L1, PD-1, these kind of things right one by one and give you all this information. 

POOR AUDIO  

 
20:35 
We are not replacing the IHC right. 

 
20:38 
So again, you know, like using this CD39 as one of the example, again, you see this as example of a marker of tumour specific T cell, which is basically the basis of entire immunotherapy like in cancer setting, right? 

 
20:54 
I won't go into the biological details. 

 
20:56 
Multiple papers have published. 

 
20:57 
In fact we have also published a few of them as well and lung cancer, other people publish lung cancer, breast cancer, you know, Melanoma and really this is a huge data set that already published out there and how do we actually translate? 

 
21:12 
You know, some of these papers are huge papers, right? 

 
21:14 
And then we really want to make a real impact. 

 
21:18 
So you know, like designing a Multiplex based assay. 

 
21:23 
It's not difficult, which I already going to do the PD-L1 right? 

 
21:27 
And then it is just as easy to stain 2 markers, right? 

 
21:30 
And then you find double positive cells. 

 
21:32 
And then obviously if you in this area, you can do a 40 marker, 20 marker and give a very, you know, comprehensive report, right? 

 
21:42 
But I’ll take a step back, right? 

 
21:44 
Take a step back and do we really need all these things, right? 

 
21:48 
Do you really need the multiplexing, hyperplexing and spatial technology to be really the one to be translated, right? 

 
21:55 
Can we use AI to actually replace this? 

 
21:57 
But we train from the spatial data. 

 
22:00 
The spatial data is just to train the system, the new algorithm. 

 
22:04 
And ultimately when you translate, it's not using the spatial transpatial data because can be costly and slower, right? 

 
22:13 
So for example, in this case, right, the H&E images these three cells, I mean like if you look at them, they are just cells, right. 

 
22:20 
They look like the lymphocytes, but not necessarily, very typical and classical. 

 
22:26 
But we have the staining prove them. 

 
22:28 
These are the tumour specific T cells and then you train them, you know, hundreds and thousands of them, you build models, then you can actually convince yourself and using an unseen H&E you can predict this and quantity that this precisely like. 

 
22:42 
And again, this is an expendable AI which and then we even build something we call user interface, right? 

 
22:48 
So unseen H&E sitting in front of a pathologist, right? 

 
22:53 
You apply this software, right? 

 
22:56 
And then is at the moment we don't we internally, we call it slider, but we will call a better name in the future. 

 
23:03 
So basically it's just a slide inside out and show you difference markers of the virtuals and in this stainings are virtual, we didn't stain the exact markers on this H&E right. 

 
23:17 
And the key thing is that this is not only technically expandable, this is clinically expandable because if you show to a trained eye, they can actually scrutinise by themselves, like for example PD-L1 positive cells, they should be looks like either tumour cell or macrophages. 

 
23:37 
There shouldn't be CD3 double positive, right? 

 
23:39 
There shouldn't be T cells, right? 

 
23:40 
And then your CD8 T cells should be actually a subset of your CD3 T cells and your tumour specific T cell is actually a subset of your CD8 T cells, right? 

 
23:48 
And then you do a quantification based on this and obviously they see all the checks, they convince themselves and they will apply this in the routine setting, right? 

 
23:58 
And now we have a few of these algorithms. 

 
24:00 
But if you keep building it, you can have hundreds of them and then you can really choose the actual one to actually stay in the comfort of the three IHC, right? 

 
24:10 
So the one that we show you is actually a user interface, but we also build a website which scan in the QR code to actually do this in more of a public domain setting. 

 
24:22 
So you can access this, you can see, but not necessarily a user interface. 

24:27 
And also, if you want to develop your algorithm, your images, that is also possible. 

 
24:33 
So if you upload your images and then they can show you this particular area that this is a virtual standing of the CD3 positive T cells. 

 
24:44 
And this is the usual images. 

 
24:48 
And in the future we can all do all these things on antibody based testing, right? 

 
24:53 
It can be transcriptomic, can be metabolomics, can be genomics and all these things can be trained in the H&E staining at $5 H&E images give a hundred million dollar answers right. 

 
25:04 
And before you actually do the development and you can really choose which is the best method to perform the very costly developments. 

 
25:15 
So basically let me recap what we have shown you and how do we potentially bring to a real world impacts, right? 

 
25:25 
So number one, right. 

 
25:27 
So how do you do this in a diagnostic precision setting, which we internally call it like a more routine setting, right. 

 
25:34 
So I think this one, we have talked a lot, but this, I think this can take some time to really get measured. 

 
25:42 
And also I'm more than happy to answer some question because sometimes I after I present this some people may not get the point, but I think the number 2 setting, it will be the easier to get right. 

 
25:53 
So the first is the routine practise number 2 is a clinical trial recruitment. 

 
25:58 
So imagine, you know, just as easy as PD-L1. 

 
26:01 
So the pharma wants to recruit a cohort of patient to actually start a clinical trial in Singapore, let's say, right? 

 
26:11 
So the only criteria here is either PD-L1 positive or is it can be TPS 50 but 50% in the lung setting or is MSI or TMB high, right? 

 
26:22 
And we all know that all this marker is at least 20, right? 

 
26:25 
Around 20% positivity meaning that if I want to recruit 100 patients OK, let me change number, if I want to I recruit 20 patients I need to perform at least 100 tests to get these 20 patients right. 

 
26:41 
The idea here is that if you have AI triaging effort, like I used to show you as a H&E 2.0 prebuild the PD-L1 or MSI or TMV and you actually can just, you know, increase your pool as well. 

 
26:55 
You can even screen through 1000 patient, right, because there's no cost and then in one goal very fast, right? 

 
27:02 
And then you only choose the top 20 to perform the PD-L1 staining or MSI or TMB right, And the win it has to the 20 patient. 

 
27:15 
So your entire clinical recruitment is very fast, very efficient and very low cost number 3 scenario is a global health setting. 

 
27:25 
It's similar to what we just talked about, right? 

 
27:28 
It's just that you're when we triage, this is not only triaging in the local setting, we triaging in the regional setting. 

 
27:36 
So for example, we do all these things in the rural or remote area. 

 
27:43 
So they can do it here, but they can also have a very small scale. 

 
27:49 
But we can easily build us with their computers, right? 

 
27:52 
So we only select the very likely high positive cases to send to the centralised place to actually perform a very auditable high standard IHC to confirm the case. 

 
28:06 
So if you actually do this, even a rural remote area can actually access the best healthcare. 

 
28:13 
So basically this is the idea, right? 

 
28:15 
You know, I just put it in a better diagram setting, right? 

 
28:18 
You have a biomarker 20% positive and you only need to do 20 tests to actually hit the 20 originally you need to do at least 100. 

 
28:29 
So it's a five times more time, money and clinicians and manpower and machine. 

 
28:36 
So that is basically what I want to, you know, like put a comma that on the AI on tissue. 

 
28:42 
But let's go back a little bit. 

 
28:44 
Just I think that this is my last few slides. 

 
28:47 
Go back a little bit on non-tissue setting, right, the single cell setting, right as I showed to you, a lot of single cell setting, which I do not have time to talk, but a single cell setting is more like the early days. 

 
29:00 
So we got clinical recruitment, talk about global health. 

 
29:04 
How do we for remote areas assess the best health care? 

 
29:09 
But you know, talk of remember my topic is the entire spectrum, how one thing that I show to you can be also applied to a very early stage for drug discovery, drug screening, the kind of things, right? 

 
29:21 
So single cell is important here. 

 
29:24 
So you mean like in a single cell setting, you can also do something like what to you can predict them accurately, right? 

 
29:30 
So now this is something I really want to show you. 

 
29:34 
So use H&E as a better setting, but again, you can do this as a single images. 

 
29:41 
Also, I'm sure all of us actually, you know, try the mobile app that you know, you put your own photo and it'll be the either app and it'll tell that I want to see my 70 years old looking or I want to see me in the baby face setting, right? 

 
29:55 
Then the app actually generates either a video or a photo for that, right? 

 
30:03 
So this is something that we built, right? 

 
30:05 
We built training of 100 thousands of cells. 

 
30:09 
We build AI generated video generated through and if you put a single cell in the they will show you the trajectory cells to become a specific T cells or a non specific T cell and give you a likelihood down there. 

 
30:27 
So if you think about it, if you want to build another algorithm particularly for the drugs, the drug screening outcome that you want, right, meaning that your screening readout can be totally no cost, right? 

 
30:42 
So you only need to build the algorithm one time. 

 
30:44 
You can apply in multiple different [unclear] in the AI trial, you use percent of the resource to hit 100% of the outcome. 

 
30:58 
This you only use a few percent of your resources, you hit 100% of your drug development. 

 
31:05 
So basically I think that is my last few slides. 

 
31:09 
I think I can skip the longitudinal and all the other immune things in a clinical trial setting. 

 
31:17 
In SICT we just publish guideline recommendation. 

 
31:26 
I mean I think they are already two in a spatial setting. 

 
31:30 
They are coming out another one which is the one that I call, I show it to you MiX SIG. 

 
31:35 
It should be able to be out by this year somewhere maybe in Q3. 

 
31:40 
So, but just before that I also want to show you this one. 

 
31:43 
This is another series of recommendation that we put a lot of effort to is to show you how to have how to study so-called essential biomarkers in the clinical trial setting, especially for IO, right? 

 
31:57 
And what are the essential biomarker that you should start? 

 
32:00 
You should look at and why and also how that is really a must read. 

 
32:06 
See if you're running a clinical trial in the IO space. 

 
32:10 
And this is my summary. 

 
32:12 
I will read them and repeat them and if there are any more questions and thank you very much and look forward to seeing you. 

 
32:21 
Thank you. 

 
32:25 
Brilliant. 

 
32:25 
So thanks again Joe for your presentation there. 

 
32:27 
If anyone does have any questions, please can you type this for Q&A or chat box and we can get round to these. 

 
32:34 
Otherwise I would now have to formally sort of introduce the panel discussion element of the session. 

 
32:40 
So as quite a few of you who registered beforehand, you submitted some questions. 

 
32:43 
So I’ll pass these on to Joe and Jasmine. 

 
32:45 
So what they can do is sort of get around to these and obviously any new questions that come through, we can also answer these as well. 

 
32:51 
Joe and Jasmine, I think I did put one that came through immediately as well at the chat box. 

 
32:55 
I don't know if you can see this one. 

 
32:57 
Otherwise you may just want to start from the ones I put in the chat before we start the session and that'd be great. 

 
33:30 
Thanks. 

 
33:31 
So yeah, I think let me see. 

 
33:33 
So clinical show for sure, right? 

 
33:35 
There are so many, right. 

 
33:37 
So I guess there's no doubt that there are so many already apply in the clinical trial space, you know like especially the pharma, the CRO that they are doing it, but they may you may not know. 

 
33:51 
So this for sure, but what we probably need to think more I mean, like again, you know, clinical trial to me is already a clinical space, right? 

 
34:01 
So especially like us, we are training, we are running the clinical trial and location you need in our hospital as well, meaning that I would consider that is a clinical space, right? 

 
34:14 
Because you really do something and then you know that this particular patient should go to which trial and then and that is how it's a decision making and that is something that again, you know, already happening. 

 
34:28 
But what we, I think most of people really discussing and talking about is the if you look at my scenario, it's really the number scenario, right? 

 
34:37 
The routine practise, the one that really you want to do it almost for most of your patients, just like, you know, your stain PD-L1 for everybody in the [UNCLEAR] setting, right? 

 
34:50 
I think that could be technique that needs to take some time and also usually something need to be done in a reflex setting. 

 
34:59 
It's not something that you, a pathology department or the pathologist like to do that. 

 
35:04 
And usually people have a lot of high standard and very cautious to actually have something reflex. 

 
35:10 
Yeah. 

 
35:12 
So that is really a high bar. 

 
35:15 
Yeah, I would say also in terms of AI just in general versus the spatial aspect is that we are already there. 

 
35:22 
So just to kind of elaborate on what Joe said, we currently make clinical decisions based on whether or not for immunotherapy I think is the best example because that crosses all cancer types for PD1 expression and I think Joe showed that we do that already kind of in a pathology lab that's part of diagnostics and decision makings of what treatment you go on. 

 
35:46 
So of course, and it's obviously applicable to all clinical trials. 

 
35:50 
So do you go on immunotherapies and the new immunotherapy targets based on that? 

 
35:55 
But I think the other levels of data now that become important to that is that reflex aspect that Joe just spoke about is that how do we use all these additional markers to kind of inform PD1 expression? 

 
36:07 
Because PD1 expression probably isn't going to be enough because we know a lot in the field that there's a lot of non-responders, even though they have PD1 expression. 

 
36:16 
So that could be based on expression type, it could be on other aspects of the tissue. 

 
36:21 
And we don't know that yet until we do more data, right? 

 
36:24 
So I think what AI will really help us do is fine tune that the more we have data, the more we'll be able to inform those decision trees. 

 
36:33 
So then I guess the next question is what are the expectations of authorities regarding analysis approaches? 

 
36:46 
Yeah, this is tricky, right? 

 
36:48 
So I mean, I guess, you know, if you look at the Q&A, so I mean, just to everybody. 

 
36:53 
So Jasmine is reading the pre submitted questions, right. 

 
36:59 
But then there is also a Q&A. 

 
37:01 
So if you think about the Q&A, the this particular question is similar to what Keith asking, hi Keith. 

 
37:08 
So I guess the key thing is there is also something that we put a lot of thoughts on our works, right? 

 
37:20 
So we have been talking about this AI triage effort and we also we categorise this as assisted too, right? 

 
37:30 
Just like, you know, digital pathology, right? 

 
37:32 
So in digital pathology, yes, we also talk about FDA approval and all these things. 

 
37:39 
But I guess in the world, for example, FDA only approved like 2 scanner, which is like Phillips system and also like I believe so. 

 
37:49 
But if you look at the entire world who actually use digital pathology, there probably is still a minority that really use these two machines, right? 

 
37:59 
So a lot of this is like they will call it assisted tool because you not really using this tool solely making your decisions, right? 

 
38:10 
It's more like giving you more information, just like our triaging efforts. 

 
38:15 
I give you more information and then you make your own decision. 

 
38:19 
And how do you make this entire system more efficient and all these things. 

 
38:24 
I guess if you are thinking about that line, the regulation pathway is actually fairly simple because you do not change the practise, the entire diagnosis and all these things are still similar, right? 

 
38:42 
It's just that one more thing and one more information which is totally independent. 

 
38:47 
So I totally agree with that. 

 
38:48 
I think that in all the fancy spatial transcriptomics, I think that's the discovery place. 

 
38:55 
But I think, I mean, I do work globally as well as Joe. 

 
38:58 
If you really want AI to inform and make an H&E better eventually all this spatial transcriptomics and all the really expensive things are not many things that we're going to federally regulate or make an a CLIA. 

 
39:13 
And I'm an American, a CLIA certified lab. 

 
39:16 
It's going to take existing pathology labs and find the right markers. 

 
39:21 
So that's the H&E 2.0 I think Joe spoke about. 

 
39:24 
I think the very same way is how do we get this plethora of like large data sets of every transcript match RNA into protein and reduce those proteins from like, you know, but make better cell calls of not just an ER positive cell, but maybe you might need three or four additional markers. 

 
39:44 
But I think that's quite capable of doing given most pathology labs can do H&E with additional IF markers. 

 
39:52 
So we have to kind of spend our time right now in generating data so that we can inform how to reduce those sets because it's going to be tissue specific, it's going to be disease specific, but over time we shouldn't reinvent the wheel. 

 
40:04 
So then those guidelines don't have to be modified too much, right? 

 
40:08 
They have existing microscopes, they have existing ways to kind of stain and really fancy facilities. 

 
40:14 
They may be Leica BOND. 

 
40:15 
And then you and I think we have the capacity more now because to train digital pathologists. 

 
40:21 
So this isn't a question, but I think it comes to all parts of governance. 

 
40:25 
It's not just data governance of how to do it or where the data lies, but also the training. 

 
40:31 
So one of the biggest problems right now in diagnostics is, well, I mean many, but we don't have enough pathologists. 

 
40:38 
And so you can imagine that if we really start to take this approach that both Joe and I are talking about where we just make existing pathology labs better informed from all these data sets, we could train them. 

 
40:50 
I mean, we have we'll have AI software to really allow them to give the right decision calls. 

 
40:56 
I think it will still be a pathologist who gives that decision call, but now we can kind of expand them from the ability to read 10 or 20 or 200 slides a day to thousands. 

 
41:08 
And I think that will really be transformative in the field within the current scope of how a lab and regulations across the world are currently working. 

 
41:19 
Thank you, Jasmine. 

 
41:19 
I think that's fantastic. 

 
41:21 
Which and in fact, if you look at the Q&A, Keith asked another few question. 

 
41:26 
I really would like to invite Keith, if possible to actually speak out, but I don't know. 

 
41:31 
But again, you know, I guess, you know the gist of his question is that, you know, nowadays there are so many different company and different technology in the spatial world and how do we harmonise and reconsult this, right. 

 
41:46 
So this is exactly what I feel ultimately is unlikely. 

 
41:52 
I mean, I don't know, I think maybe they are a spatal technology company do not want to hear my hear me speaking this, but I guess ultimately is unlikely that the particular machine gets translated. 

 
42:03 
I feel so it is more like the biomarker that discover like what Jasmine said, some of this is this machine and get translated by using an intermediate technology, for example, as simple as IHC , as simple as AI, as simple as H&E and that is the one that ultimately gets translated. 

 
42:27 
So I think if you think from that line, I guess we do not need to think too much about how do we harmonise the spatial data as well, which is I think it's impossible. 

 
42:37 
I don't know, what do you think Jamsine? 

 
42:39 
I think I was trained as a geneticist. 

 
42:42 
I feel that it has kind of moved and shown us the way, which is you can still use sequencing in research. 

 
42:54 
It's not going away. 

 
42:55 
So there's no competition across. 

 
42:57 
I mean, there's competition across companies that can let be lively. 

 
43:00 
You can have your favourite sequencer, but if you really don't have all the abilities to do that, we can still do PCR for BRCA1, right? 

 
43:11 
I mean, so there are genetic mutations that we don't need huge sequencers for, but we need huge sequencers to publish and genetics to discover kind of the new mutations, right? 

 
43:21 
Then we can panel size it down in other countries which can't run sequencers. 

 
43:32 
I think that's half of what is already being established. 

 
43:37 
I mean, the amazing part about spatial biology right now, it's both ends, right? 

 
43:41 
It's that research can really start to push the edge, but being able to tell what those new biomarkers could be. 

 
43:48 
But we already have pathology that has some biomarkers that we routinely stay in for diagnostics, right? 

 
43:55 
And so unlike genetics where we had to discover all the mutations, we have some markers already, right? 

 
44:01 
And so all we have to do is really learn from that and better inform that. 

 
44:05 
That doesn't mean the research environment ever goes away. 

 
44:08 
It will always continue to evolve, right? 

 
44:10 
With new drug targets, there'll be new ways to learn from biomarkers. 

 
44:15 
And so then that competition of the variability, which I think Keith was asking about or the variants, I mean, that's where AI will be crucial because we'll learn from all the variable data sets of every different machine of every different kind of tissue type. 

 
44:31 
And over time, that will still, I think go down the trajectory that both Joe and I believe maybe we're biassed is to give the right biomarkers by protein and existing labs. 

 
44:44 
And I guess you know, like I totally agree with what you said, right? 

 
44:47 
Especially, you know, I don't really be very molecular with the genetic settings, right, Because base basically the last rounds of revolutions, right? 

 
44:59 
So I think one of the key things that I believe some of our friends in the fields they are working on, it is also trying to see whether we can harmonise a few similar to the fast Q file in the DNA setting, right? 

 
45:16 
So can we have a metal data or foul type of format that is easily translate and harmonizable? 

 
45:24 
I think that is totally another conversation, another talk. 

 
45:30 
I think yeah, I do know that there are people working on it, but I guess that is quite challenging to me. 

 
45:37 
I don't know what just now we're actually my group with others are working on it right now. 

 
45:42 
The issue becomes I like being proper metrics in the field required, right. 

 
45:48 
So if you recall and see thing, we didn't even keep a fast, we kept BCL2. 

 
45:56 
So for people, sequencer is still an imager and so you have an A,G,T,BCL2s were actually the imaging, the images, they became too big to take over time, but we needed to learn them. 

 
46:09 
And we really know that an A was an A, a C was a C, a T was a T, a G was a G. 

 
46:13 
We now know that. 

 
46:15 
So now we've reduced that into fast Q automatically. 

 
46:18 
Nobody keeps those images. 

 
46:20 
I think to answer that question, Joe, I think over time we'll learn it. 

 
46:24 
I think we're all trying to learn it now so that we can know whether experiments were OK. 

 
46:29 
Can we normalise the data? 

 
46:31 
I think over time, like we're still in the infancy of adopting that spatial transcriptomics. 

 
46:38 
I don't think we're necessarily in the infancy of the IF stuff maybe. 

 
46:44 
I think that probably has been learnt really well. 

 
46:47 
I think H&E for sure. 

 
46:49 
I mean, digital pathology is just moving so fast in so many ways, which makes me really excited that we probably, hopefully not in my lifetime, but hopefully you know what I mean. 

 
47:01 
We won't. 

 
47:01 
I mean, you showed the slide for that, right? 

 
47:03 
We'll just take H&E because we'll have all these data sets have trained the H&E to tell us the cell type we never knew before. 

 
47:10 
But my faith is an AI because I'm like Joe, we're experts talking about AI because we're doing AI. 

 
47:19 
Spatial omics has allowed us to have a lot of training data, but I don't think it is an AI. 

 
47:27 
I'm all paper out last year and I was like, this is really not my cup of tea. 

 
47:31 
But the one reason I think we're both in it is I give this to everybody that's thinking about it is we know ChatGPT is an amazing for language models, right? 

 
47:42 
But spatial, but these AI models are amazing for spatial data in a different way. 

 
47:49 
So they're not language processing models. 

 
47:50 
You can open your phone, right? 

 
47:52 
Everybody has an iPhone or Samsung literally open your phone based on your face, which tells you AI is going to get us through it for spatial. 

 
48:01 
Because if you train your phone enough with enough faces, it's going to be able to identify my face and Joe's face and Keith's face independently, right? 

 
48:14 
Because there are enough features. 

 
48:16 
And that's the key I think to establish the metrics of what a Q file is. 

 
48:21 
We have to, and we're throwing it out because the data is so big. 

 
48:25 
So my worry in the field is we're using a lot of commercialised machines. 

 
48:30 
We've compressed everything. 

 
48:32 
So really we already did it into a cell, the Pixel. 

 
48:40 
POOR AUDIO 

 
49:10 
Hi both sorry my office Wi-Fi keeps crashing. 

 
49:14 
I mean you guys still have around 10 minutes, so feel free to continue. 

 
49:18 
OK, thanks. 

 
49:29 
So I guess you know, can we recognise all the cells in the H&E setting, which is I'm not sure especially how do we define cells, right? 

 
49:44 
If you are benchmarking your single cell, which to me is strong long because yeah, if you benchmark your single cell, and I don't think you can, if you benchmark like category code, which is like traditional where that pathologist has been using their eye to tell that this is fibroblast, this is macrophages, this is T cell, this looks like B cells, and this has they can tell, AI can tell, right? 

 
50:09 
So I guess and a lot of this in the pathology diagnostic setting, they're actually just seeing one particular feature or even seeing particular cells that they make the diagnosis. 

 
50:20 
The famous one will be the Reed-Sternberg cells in the lymphoma setting, the Hodgkin lymphoma, right? 

 
50:27 
I would think that in the moving fields moving in, I mean like we are mainly talking about the AI but AI and the spatial here. 

 
50:34 
But the relevant topic will be the AI pathology and the diagnostic pathology. 

 
50:41 
My point is that in our era, there are many things will be discovered similar to Reed Sternberg cells. 

 
50:47 
So how they discovered in the 50 years ago, that kind of thing is very difficult, right? 

 
50:52 
Somebody see this show to another one. 

 
50:55 
And this could be in London and then after 10 years and somebody else pick up outside London and then this. 

 
51:01 
But in this world we pick up something we can show everybody. 

 
51:05 
We can call for a meeting like this. 

 
51:06 
We can show this in ACR, and a lot of people will see it and they can validate them using their own data and their images and we can discover multiple different Reed-Sternberg cells in the future. 

 
51:21 
I would think with AI, so if you want to go deep into the modelling, when you miscall it's called a hallucination. 

 
51:36 
But then mathematically and computationally, we can keep training so those hallucinations start to get minimised, right. 

 
51:44 
So unlike kind of my pathology call versus someone else's pathology call versus someone else's pathology call, it's still manually fed in the sense that it's it is still a manual annotation from training. 

 
51:59 
We don't know why, we don't know why that could be altered, but computationally we could figure it out, right? 

 
52:06 
The more metrics we have so those models can get retrained and you know, iPhones couldn't open on your face alone. 

 
52:13 
And I love this. 

 
52:14 
There's an AI generated models of like cats early in the day where one eye was here and one eye was there, right? 

 
52:22 
And like the mouse was in another place and the ears were over here. 

 
52:26 
But over time those models get better and those computational models with the hallucinations get trained based on more and more data. 

 
52:35 
So the idea that's hard to think about right now, Keith, because we're not there is that H&E is getting, we see neighbourhoods or cells based on what we already previously know in that epithelial cell. 

 
52:47 
So we're just calling it an epithelial cell relative to what it's around. 

 
52:53 
But the more information, and I think that was Joe's slide of H&E 2.0. 

 
52:57 
There's going to be multi omics data underneath here that will be reduced into an H&E over time, right? 

 
53:05 
That the computer will pick up that type of H&E when it has these types of neighbourhoods around it is epithelial cell type 8, right versus, but right now we're just calling them all epithelial cells if that makes sense. 

 
53:21 
But we're not there yet. 

 
53:22 
But do I think we'll get there? 

 
53:23 
I mean, I don't know what the trajectory is, but I think that we could get there spatially more than most other kind of multi omic data sets. 

 
53:34 
Totally agree. 

 
53:37 
OK, let's see where else we're at, what kind of resolution for H&E is necessary. 

 
53:46 
So John asked, maybe this is known already, but do you think H&E images will be able to capture represent all kinds of multi omic markers and what kind of resolution of H&E will be necessary? 

 
53:57 
So we usually do for the X, right this in the clinical setting. 

 
54:01 
So I guess this is yeah, so for the X is the key, right. 

 
54:05 
So again, I think, you know, I think nobody knows whether we can capture all the feature, but I guess this should be a lot, especially I do think that we, I mean, again, our group mainly work on immune cells. 

 
54:21 
I mean, like we only do some tumour markers because some company pay us to do that. 

 
54:26 
But I feel I mean, I might be wrong because I think my AI scientist told me that they think totally oppositely, but I feel that immune cells can be more recognisable. 

 
54:41 
But I mean, I think we need more training though. 

 
54:44 
So like depend on technologies, right? 

 
54:47 
I think that they're, you know, in single cell, they got captured really well in spatial because of that cell segmentation. If we learn off pixels, I think that'll be true. 

 
54:58 
But I think because the immune cells got overrepresented in single cell data, we don't really know true ground truth now because they not always get counted as well in the spatial omic data. 

 
55:13 
I think as that field evolved. 

 
55:15 
I mean, back to Keith's question, do you think we'll ever do that? 

 
55:18 
I was like, I don't think anyone's going to be out of a job anytime soon, right? 

 
55:21 
Like I don't think AI is replacing any of the pathology calls because you have so many different components, we actually see more effect and kind of the other types of immune cells, not the T cells, but the macrophages, you know, the things that kind of got ignored a little bit and all the other kind of, you know, early site seq data, those kinds of things. 

 
55:50 
I think immune cells and the neighbourhoods by which they infiltrate will really get us there. 

 
55:57 
But I think it's the overall cyto architecture that we don't have good markers for, right? 

 
56:02 
So we call a fibroblast a fibroblast. 

 
56:04 
But kind of as you dig through the spatial specs in the cancer stuff, it's the immune cells relative to kind of these protective layers in the fibroblast for like the non responders. 

 
56:15 
You know, we couldn't see that in the single cell data because you didn't have the context because we were just counting cell numbers, right? 

 
56:22 
Like less T cells, more T cells, but I don't know, I know you can see this a lot in your data. 

 
56:28 
It's like there's these weird kind of protective neighbourhoods, right? 

 
56:32 
And as you evolve through that, like we know CAR T cells can't get in, right? 

 
56:37 
Are those like the mechanisms by which they get in? 

 
56:41 
Are they able to get in? 

 
56:42 
Can they infiltrate? 

 
56:43 
2 separate questions, but I think right? 

 
56:50 
And immune cell, see those kind of things. 

 
56:53 
So I think that will be really teased out with all this spatial data. 

 
56:59 
Yep. 

 
57:00 
Do you agree? 

 
57:01 
Yes. 

 
57:02 
Then we are kind of like less 2 to 3 minutes. 

 
57:05 
So let me also ask you this. 

 
57:07 
I mean, I think I like your analogy about like, since we can use our phone, our face to open our phone, so we should be able to do this. 

 
57:16 
I think the key challenge that I can think of, which I would think that this is a perfect platform to say this and somebody will pick this out and work on it. 

 
57:27 
And especially like a group that from you just been so for example, it's so easy to use your face to open your phone is because we know that the structure of a usual face, right? 

 
57:43 
There's two eyes, there's one nose and there's two ears and one mouth. 

 
57:48 
I think those are the standard structure, right? 

 
57:52 
And in a few years, right, there are a few publications from our group. 

 
57:56 
We call this thing spatial anchor, you know, like anchors, right? 

 
58:01 
So you need to, we need to define the anchor nicely. 

 
58:06 
So that everyone generated the data that in the world that ultimately they can fit into something that is usable, right. 

 
58:19 
So otherwise everybody using different ankles and different rotation difference, you know, things they are not just like you said, they are not two eyes, one nose is like one eye here, one idea. 

 
58:32 
So we can't harmonise this data, so this slows down the entire process. 

 
58:38 
Yeah, I think I lean in on people that do like structures, you know, like lean in on the brain. 

 
58:47 
People who like have absolute structures because, you know, people more like us. 

 
58:51 
I do brain too. 

 
58:52 
But in tumours, you know, you'll have structures for prostate ducts or right, like ductal structures and breasts. 

 
58:59 
But they're not anchored, like you said, because they can change based on the tumour microenvironment, how much epithelia or mood cells are there. 

 
59:06 
So maybe we can lean in, right. 

 
59:08 
It's not just strictly, you know, one type of disease by which we'll learn the spatial. 

 
59:13 
And I think that's the really exciting part is that, you know, we have so much anatomy, right? 

 
59:19 
And if we can learn from the anatomy, maybe it's not. 

 
59:22 
Maybe it's going to be learned on mouse brain. 

 
59:24 
And then over time, those applications will be used as cell typing in heart or maybe we'll learn about epithelial cells better and the ileum, which has this beautiful anatomy that's very typical that has amazing Histology behind it. 

 
59:42 
So I think it's not separable. 

 
59:44 
I think the field is kind of separable about it now. 

 
59:47 
But I think over time, if we took that approach of like merging faces, right, if we start to merge kind of cell types kind of like the human cell Atlas, but of like spatial that like cell types. 

 
1:00:00 
Over time we learn on immune cells and epithelial cells across all the organ types, right? 

 
1:00:06 
And maybe we can start to anchor better based off because then you might get neighbourhoods out for inflammation that may reflect neighbourhood right from Crohn's disease, let's say that aren't exactly a tumour one, but maybe they will start to get some anchors that make sense for a neighbour. 

 
1:00:25 
And I think we're out of time but yeah I totally agree with you. 

 
1:00:28 
So anatomical is the anchor, but that mean also means that only the normal tissue can be easily anchored. 

 
1:00:36 
The tumour can be difficult. 

 
1:00:38 
That is why we actually want emails because I think the T cell in the tumour and in the normal that it should be looks like quite similar, right. 

 
1:00:48 
So that can be an angle, right yeah. 

 
1:00:51 
And also I look at the chat that, you know, Ryan was asking how do we make sure this is transferable to another institute. 

 
1:01:00 
So usually in the just to just since there's this question. 

 
1:01:03 
So I take the chance to address that concern also. 

 
1:01:05 
So nowadays in the AI pathology, usually people talk about repeatability, reproducibility and whether you can do this in across observer and across the performer as well as the performer in here usually is a scanner. 

 
1:01:21 
So if you have an algorithm, can you do this across different rending of the scanner and also even the same branding, on two different machines can you perform the same performance in there? 

 
1:01:36 
So those are the things that is standard in a few people test those. 

 
1:01:41 
Yeah. 

 
1:01:43 
And I think we're done for time. 

 
1:01:45 
Is that right? 

 
1:01:46 
Yep. 

 
1:01:58 
Thank you again, Joe for your presentation and thank you Jasmine as well for being a panellist and thank you for having for their questions.