Tumor slides can reveal biology that guides personalized care

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 0:41

   Yeah, so Holland S1 is a new cancer biology AI model that we've trained on pathology tissue slides. So the idea here is that, you know, the challenge with cancer specifically, obviously, is that why is it so hard to cure cancer, right? The reason it's so hard to cure cancer is because of the fact that cancer is a very diverse disease, right? Every person has a slightly different kind of disease in terms of, even though they may have the same exact diagnosis, a person may have triple negative breast cancer, but the underlying biology can be quite different.

2. 1:15

   The genes that are being expressed, the interactions of the cells with each other, all that can be quite different. And so the question is, how can we actually map that and measure that? And that can enable us to then more personalize the treatments for these patients. And there are various different tools for doing this molecular test and so on. But a lot of these tools can be quite expensive and also quite time-consuming.

3. 1:38

   And so the idea that we wanted to explore here is, can we extract a lot of those similar insights directly from a tissue slide? So a tissue slide, what we mean by this is that basically whenever a cancer patient has a solid tumor, the typical standard of care or the typical gold standard for getting that cancer diagnosed in the first place is you take a biopsy, right? You take a sample of the tissue and then you put it under the microscope. You put it on a slide and you put it under the microscope and a doctor is analyzing that. And so in this case, what we're saying is we can actually take that same exact slide and actually with the power of AI, we can extract a lot more information about like what's happening at the molecular and biological level and what sorts of pathways are being expressed, what sorts of genes are being expressed, and what makes that tumor unique for

4. 2:24

   that patient. Wow, that is so fascinating. I realize as you're saying this, I actually don't totally know how complex of a problem it is. You hear often, all right, AI is actually going to solve cancer, but we don't talk about how it's going to solve cancer. I know part of your solution, of course, with Holland S1 is having a specific AI model designed for cancer biology. What does that word designed mean? Yeah. I mean, I think the idea is like, we want to be able to, you know, train this model on, yeah, like on the, like the sort of pan cancer, understanding all these different cancer types. I think like, you know, the thing is these sorts of, in my opinion, a lot of these sorts of classifications of cancer types are a bit arbitrary too. Again, like what really matters is what's happening at the molecular and biological level, right? And so we want to train

5. 3:15

   this model on a variety of different cancer types to understand exactly what's going on at this level and being able to connect this across different modalities, including modalities that are more easily accessible. And so that's the real power of these models is to extract information with modalities that, you know, a regular patient may be able to have access to. So specifically, does this mean now that we'll learn more about rare cancers that we've never been able to learn before? Yeah, potentially. I mean, you know, I think a lot of this, of course, needs to further be validated. Of course, this is still a model that, you know, there's a lot more validation needs to be done with this model. But yeah, I mean, you know, we've seen some really interesting results with the model that we currently have. For example, we actually looked at

6. 3:59

   patient with a rare type of bone cancer called osteosarcoma. And our model wasn't actually trained on osteosarcoma, but it was able to extract a lot of biological insights that this patient's medical team had also observed, but that medical team had utilized very expensive molecular tests. And so we were able to identify, for example, this one gene, which was actually then targeted by the medical team for treatment for this patient and led to a successful treatment for that patient. But we're able to do this directly from the whole slide, from the pathology slide image. And so I think like, yeah, there is definitely potential for a lot of this stuff to be able to explore rare cancers. But again, it needs to be further validated. But we're really excited by the potential of this model for sure. Yeah, no, this is a bit of lore, but I actually

7. 4:46

   had a close family friend. He was about 13 years old when he passed away from osteosarcoma. And And it's a really rare cancer. It also affects kids. It's popular amongst kids. It's just really sad. And I know we did a lot when it came to actually raising funding. But one of the hard pieces is that there's not enough information and research, especially when it comes to some of these rare cancers. And so I guess I know people hear often, all right, AI is going to make this more possible.

8. 5:15

   But people still don't get tangibly how that happens. If you're just trying to explain now more to the mainstream audience, how do you actually change maybe some of that problem? Yeah, I mean, I think to take a step back, right, like everyone's talking about, yeah, AI is going to cure cancer. And I think, like, what does it actually mean to cure cancer? In my opinion, curing cancer isn't going to be like finding this magic drug.

9. 5:40

   To me, it's more of a framework around how do we basically design personalized care for a patient? and this sort of framework that is widely successful and generalizable. This is what I think a cure for cancer will actually look like. And so a big part of that will therefore be, you know, how do we measure the tumor's biology? How do we map that? And then based on that, how do we then design appropriate treatments for that specific tumor?

10. 6:07

    And so in this case, what HAL and S1 is doing is more focused on that first part. How do we appropriately map the tumor biology? And then once we are able to map that biology, then we can utilize those insights to then design a treatment for patients. So that's kind of like the idea of what we're trying to do with a model like Holland S1. But yeah, I mean, there's, of course, lots of great work using AI for designing treatments as well.

11. 6:32

    And, you know, of course, lots of different paradigms for this as well. So that's also very exciting. But yeah, I think like often underappreciated is what it takes to actually truly understand that person's biology.