Hi there. Thanks for that introduction. So, yeah, I work at the Christopher Shaw Lab, at King’s College London, and, we study neurodegeneration.

And in all spheres so that we use a technique called immunofluorescent staining, whereby we stain for proteins and cells in different fluorescent colors.

Importantly, this allows us to visualize, many proteins in different colors all at the same time so we can see whether they’re, the proteins are interacting with each other, where they’re sort of localized within, the cell and within the brain as well. So it’s quite, important for us. The problem, of course, with using fluorescence is that there’s a lot of background.

Now this background, can be from blood vessels, blood vessels also for us.

I know it could be from other structures as well, and even just the brain tissue itself. Some areas of the brain are more fluorescent than others.

Usually, when you’ve, treated them with formaldehyde in order to preserve it, that increases that fluorescence.

And the problem here is that when you’re trying to quantify your data, if that fluorescence is sort of at a similar level to your antibody, even if you can somewhat distinguish the antibody from the background tissue is it’s very difficult to find a fine line where the target in question is quantified without, the background fluorescence, and indeed, if it’s a blood vessel particularly, that can be entirely, that can’t be avoided.

So it really presents quite an issue when using fluorescent data, trying to quantify it.

And, essentially, what I’m gonna be talking about today, we’ll be using AI to get around this issue so that blood vessels aren’t counted, and just your protein of interest is what’s counted, which, is really sort of a powerful way to do sort of staining in, you know, one brain slice and being able to quantify four different proteins all up all in one go, really quite powerful.

So I’ll be talking about lipofuscin. Now lipofuscin is a hallmark of aging, because essentially, proteins and metals in our body are oxidized or altered phosphorylated in such a way that the body can no longer break them down and, get rid of them, So they start to accumulate, into these aggregates within, within ourselves, and create this autofluorescent structure called lipo fluorescens.

Now this is quite an issue for fluorescent staining usually, because it is highly autofluorescent, and it makes seeing your actual staining very, very difficult, let alone being able to quantify and separate it out.

And, usually, to do it in order to, remove the staining, we have to use a substance called sooten black, but the problem with this is that it increases autofluorescence.

In this case, lipofuscin is shown in green, but soon black will increase the, the red, the red wavelength that we image, at as well. And this essentially prevents multiplexing, not prevents, but makes multiplexing a lot more difficult to do, because then you’ve got a lot more background in your other protein that you’re trying to assess, and that becomes very, very difficult.

Pathologically, lipofuscin is quite interesting because, it builds up, in some cases, the neuronal steroid lipofuscinosis.

It builds up very, very early on. So in normal in a normal adult, for example, the amount that you’re gonna have in a aged normal person, you know, someone who’s eighty or, you in in these diseases, you may see that amount of lipofuscin at a very young age, and, really sort of ramps up the expression of, and build up of lipofuscin very quickly.

And, so if we were quite interested in this pathologically in just general neurodegenerative disorders as well and not just neuron or seroid lipofitinosis and tracking how that changes over time.

So for the purpose of this talk, I’m actually gonna be talking about how we quantified it rather than removed it from our staining.

And I think it’s a good example for other people who have issues with low, with very high background staining in for any protein, because here, we’re not staining for it at all.

So the contrast is really quite difficult to see in some cases, especially sort of here where you can see very tiny bits, but you’ve also got, you know, some background here as well of this sort of green haze, which is gonna make it very difficult for you to get through it.

So how would we go about quantifying this? I mean, this we really spent a long time on this sort of saga. It was about three to six months, really quite, a long period of time where we we took multiple images. We, you know, we first first, we tried to get our imaging accurate, And so we tried a slide scanner, but that was too, too poor quality. We couldn’t really work off that because the background was just too high.

And then we went to sort of confocal, microscopes to really increase the resolution.

And we even, at at certain points, went into different areas of microscopy, such as multispectral imaging where we, tried to analyze the the individual spectra of Lipofuscin, the unique spectra that Lipofuscin has compared to the brain, but, unfortunately, blood vessels are also very similar in that sense, and he amplified the blood vessels as well, so we were not able to use that. So ultimately, we stopped with the confocal images, which is what you see here, and we sort of went about trying, went about quantifying this, and three or four people were involved in, sort of quantifying it separately in order to get sort of the nice accurate results and the best sort of labeling of lipo fusen.

One of the methods was, where we had a set threshold for all the tissue. So having a threshold essentially means having a bar for brightness.

Anything above a certain brightness will be counted as lipofuscin, and anything below that brightness will not be.

The problem with this is that because brains, when they’re dissected, they’re submerged in formaldehyde, and, they’re all slightly different is the, brain autofluorescence varies really considerably between, samples. So when we were getting when we were doing a set threshold for all tissue, we were getting in some, you know, no labeling, at all of lipophilicin even though we can see it, and in other tissues we were getting way too much, than what was what was evident by eye. So we also tried changing the threshold per tissue, which, so so for every brain, we sort of realized that, yes, it will have a different autofluorescent, background, and sort of setting, and we need to adjust for that.

So we tried that, but the problem with this is therefore it becomes but it can become biased is because one person may change the threshold, and the other person may not agree on it, or they may say, no. I think the threshold should be this. And, it becomes very, very difficult, in replicating it going forward, which, you know, the ability to reproduce data is really paramount in silence.

And then we have the final sort of method, which, we didn’t really look at for too long, because it was so time consuming, was changing the threshold per subregion.

So I recognize that this area of the brain will have a different background to this area of the brain, and even within sort of the same structure. So this is the thalamus of the brain, which is the sort of the relay center of the brain.

This has many sort of nuclei in it that are involved with different, different other areas of the brain or perhaps even going down to the spinal cord.

And these also have different fluorescent properties, so it really becomes quite a logistical nightmare when you want to separate the tissue.

And, the the but the the benefit of doing that is because when when we did a set threshold for all tissue and we applied it, we got a lot of additional signal. Now this area of the brain specifically doesn’t actually really express lipo vousin at all, but as you can see, the program has said, well, I’m recognizing that there’s a lot of lipofuscin here, and biologically, we we just know, and also looking at the image that that is not true. It’s just simply because this area of the brain is very autofluorescent. So they had a lot of issues in this regard.

And this is when we contacted Visiopharm.

We asked them, well, how how would you go about sort of, doing this? You know, how can you help us, in trying to quantify this?

And they went away, and all props to David Mason who, essentially created a artificial intelligence app to recognize life of Fussen.

So I’ll put it to him in this regard.

And having looked at all of the slices and all of the, the image the analysis that he’s done and subsequently produced, it gave us a really, really good and accurate way to identify lipopussin.

And there are some key points that I want to sort of highlight is because you can see that these images are really not that great. There’s they’re all over the place in terms of intensity. This area is very dark. This area is quite bright. This area is also quite bright, and this area is also very dark. And, that presents a real challenge when you’re doing thresholding because, of course, even if you can see a very sort of dark lipoprotein granule that’s brighter relative to this sort of low intensity, it’s not going to be the same threshold as, others that are in higher background environments.

And the the the areas that I want to pick out in terms of this AI app that they developed is that this area of low intensity, you can still see that there’s this pink dot, in the middle and, showing that the app is still recognizing it despite it being very sort of, much lower intensity than the others. Same as in areas where there’s high, high background and high intensity, in the green channel.

There’s not an increase in the labeling such that it’s still only labeling the lab of foods that are not reacting to the intensity, which is really quite good.

And as well in the hippocampus where we previously had issues before where there was a real, over labeling in the previous slide, there isn’t that anymore, and you can also see that there are some blood vessels here, that it’s not picking up, which is, really, good news because we don’t want to be counting blood vessels. We don’t want to be counting, labafucin.

So we’re really, really impressed with that and, really did a great job in helping us moving forward.

In the meantime, while they’re actually setting up this app, however, we also found an antibody that sort of stay in it’s not a direct antibody for lipofuscin, but one of the proteins that accumulates in lipofuscin.

We found an antibody for that and a really, really impressive antibody.

And, so we started staining for that, and the amplification of the antibody was enough that we could just threshold it, and that was it. Because we had some amplification, the difference between the, the now the aggregate and the the background is much higher. So if we compare here, you can see these clear sort of, aggregates and, against the background, the slight haze is a much bigger difference than say these smaller aggregates here in compared to their background, as well as quite visually apparent.

So that was really sort of impressive.

And, and so we actually just, went through the the thresholding, for that, and, that gave us an answer. But, of course, don’t worry. I haven’t come through all of this talk just to say, you know, get a good antibody, and you can threshold, you know, that it did help us in terms of that. But now what we found is that this antibody has different kinds of staining patterns, and this is where we’re sort of looking to now.

As you can see, you have these sorts of punctate, aggregates of this antibody, which is called CMAS.

And, and you also have larger sort of agglomerations that aren’t puncta, even when you reduce sort of the intensity of this signal, they do not separate out into puncture, as well as also having these out of focus aggregates, that create these sort of corona like effects.

So now, yes, we thresholded to sort of look at the the relative expression of the protein, but we can also separate them, into individual, types of its staining.

And, unfortunately, we cannot do that with thresholding, so if we try to do this with thresholding, it would treat everything as the same. So this larger aggregate and these smaller punctate aggregates, and these, particularly these corona and out of focus targets will all be above the threshold part that we set, and they’re all going to be treated the same.

Whereas with AI, what we can do is we can say, look.

I don’t want you. You’re out of focus.

If if you were to be in focus, you’re probably going to be more intense, sort of something similar to this, and probably smaller, than what you’re representing. So we can remove that and say, well, this is not accurate, and it may affect the readouts that we want to produce. For example, if you want to say what’s the intensity of the aggregate, these will affect that if they’re included as well as, the area of the aggregate.

These will also affect that, and they’re not correct reflections of the aggregate when it’s in focus.

So we developed, we went back to Visiopharm, and we said, like, look. We want to try this with AI, and they gave us a sort of a training program, and, I was leading that. And I only had about three to four hours sort of, active work because I was very busy at the time, so I couldn’t dedicate a huge amount of time to it.

But what you can see is that larger granules are being detected. Smaller granules are being detected.

And what I really want to highlight is that the out of the corona, I was even in that small time period was able to say I do not want to include the corona, and these were avoided by the AI, and we’re able to separate them out, which gives us a lot more pure data of what’s actually in focus and saying well, this is the intensity of everything that’s in focus, and this is the area of everything in focus. And that way, we can really sort of narrow it down and keep everything sort of as accurate possible.

Going forward, we haven’t actually returned to this app, since actually buying AI, although we will do it soon. Going forward, we’re probably going to create an app which, gives different labeling for the puncta compared to the larger aggregates. And this may be a way to track better disease pathology and what may correlate more to disease phenotypes that we see. For example, if this was related in locomotion or memory tests, you know, do the punctate do the does the amount of punctate correlate better with disease, or do the large aggregates correlate better with disease, and, so on and so forth?

But it doesn’t stop there. We’ve been looking at other staining and how we can improve that, and, I, I have a colleague at work who, wanted to investigate these puncta, in these blue, nuclei.

And, she’d been using a Bayesian classifier, but had been, struggling because it was picking up sometimes, yes, the puncture, but it was also picking up other areas. And it was she couldn’t get it to a point where it was accurate enough. And so she essentially gave up on the idea and just said, well, I can’t quantify it.

Even though it’s extremely clear and visible to, to us that, yes, it does exist. It is there. It is completely different to this kind of staining, which is the full nuclei, just brown.

And so when we got AI, I said, well, you know, let’s let’s revisit that. Let’s have a look at it. And indeed in so probably about an hour of our work, at work, it’s already really nicely picking up these puncture. It’s not complete. There are some it’s missing some areas.

But, again, this was within about one hour of us training it, and it’s already at this stage. And this picture is actually also from when we run the app, and it’s not picking up anything here, not even these smaller dots that it may sort of think of a puncture. So really doing quite well there, and it’s it’s, you know, sort of reignited her fire. She’s now able to quantify these, and that can go quite a way in terms of progressing her experiments.

We, on the other hand, we’re also looking at inflammatory cells in the brain microglia, and we’d always been told by several people that, you know, inflorescence do not bother with this. Don’t quantify, because the ramifications, these sort of trees that come out of the the cell body, these sort of circles, are very similar intensity to the actual cell body. And so doing thresholding and other classifiers is just extremely difficult.

And they just sort of said, don’t bother with it.

But now with AI, we’ve gone back and said, well, no. Let’s see if we can do it. And yes, indeed, it looks really, really clean, really, really clear.

I’m gonna say clever, but AI is clever.

And it looks really, really impressive.

And we are now able to quantify that. And even we are separating them now based on their expression of another protein called CD sixty eight, which is an a marker for when they’re activated.

And we are able to distinguish between these two, because it’s recognizing that, yes, this green channel is present. Yes. You’re a microglial cell. You’re expressing, the, protein in green, but also underneath, you’re expressing this protein in orange or CD sixty eight, as a marker of activity. Therefore, we’re labelling you red now. You’re active.

And microglia, in general, become active towards sort of insult and injury or inflammation, neurodegeneration. So they’re quite important for us to sort of quantify.

And AI has allowed us to do that, which is not, which before with thresholding and other sort of techniques, it wouldn’t be possible, to identify specifically the cell body, as a as a finite count and then, colocalize that with CD sixty eight. We could do it by area, but then we’re also including ramifications, which then becomes a bit hazy because the ramifications change when they’re activated and so on. So it it would give us very unreliable data.

And finally, we’re looking at neurons, very briefly. This is what we’re we’re working on now. As you can see, it’s quite a very difficult stain because you’ve got overlap of neurons, and we’ve got these really swollen, dying neurons as well.

And getting an app to recognize all of them as the same is really quite difficult. It’s, impossible with sort of thresholding analysis, really, really difficult, especially with these sort of hollowed out structures, which I really struggle with, as well. And the ongo I mean, this is still in its, sort of early stages. There’s a bit of over quantification, and in some cells, it’s splitting it up. However, it’s doing a really great job at identifying these swollen neurons in full as well as being able to separate these overlapping neurons, really quite well. So I just want to thank, the people in my lab, Christopher Shaw, Natalia and Alexandra for, you know, helping me and sort of, facilitating with the lipo flucyn going forward and all of that. Natalia and Alex, certainly kept me sane through all because it was really quite a challenge at points.

And then I want to thank George, at the imaging facility because he did some for some of us, the imaging, taught us about the multi multispectral imaging and, so on and so forth. And I even did some of the quantification as well. So really, big thanks to him. And, of course, Tilly Hawkins, who I mentioned, in regards to helping, my colleague in the lab, for allowing me to present some of her data. And, of course, David Mason, who facilitate, who made some of the apps and Georgia for organizing, the communication between King’s and Visiopharm, and, of course, for inviting me to this conference. So thank you very much.