Read the Full Video Transcript

Felix Feng: So today, I've been asked to talk about whether genomics or [inaudible 00:00:05] AI can resolve the nomenclature question. I want to begin with what I think is a bias here in the States, and Laura Esserman pointed this out earlier, which is that our bias tends to be that more is better. More testing somehow will solve this particular issue of what we should do with the patient. And fundamentally, I think we have to remember the reason why we care about all these questions is that we want to know how to better manage our caseloads. If a particular test doesn't change the intervention, then it's probably of very limited use.

And in terms of functional genomics, you guys probably know I'm a big genomics person, but sadly, I think that the evidence out there currently is that the current genomic tools do not help us in this particular situation. Why is that? Well, Dan [inaudible 00:00:57] and his team showed that [inaudible 00:01:00] test really doesn't add beyond clinical variables in the context of active surveillance in the [inaudible 00:01:07] cohort, and then Matt, Peter, and team just reported an abstract at the Western Sectional looking at Decipher, and you see us in that active surveillance cohort showing us that prognostic.

And then, Polaris, as Dan mentioned, is [inaudible 00:01:22] enough not to [inaudible 00:01:24]. And my joke about active surveillance is that active surveillance is where biomarkers go to die. Literally.

So, I think it's very safe to say, based on the available data at this point in time, unfortunately, the current genomic tools just really haven't validated in the active surveillance study, and therefore, it's really hard to make changes in management based on that.

However, you can ask why are they not working as well, and as [inaudible 00:01:53] mentioned earlier, maybe it's because most of these studies are done without MRI. So, maybe you haven't biopsied the right patient yet. That could definitely be it. And if that's the answer, then probably nothing is going to work, because if [inaudible 00:02:08] what are you going to get back, right?

But I don't think it's going to be completely that, in the sense of, let's say you missed the lesion 5% of the time. You can still derive a signal in 90% where something's the matter.

So, what are other possibilities? Well, heterogeneity, and that's been mentioned quite a bit. Part of the issue is that, at least in Grade 1 cancer, there's not much of it. So, it's not like we have 100% tumor, we have 5% of a single port. And the problem is, when you run an RNA spectrum, if you have 5% tumor and 95% something benign, sometimes that will get washed out. You can imagine the situation where you have 10,000 cancer cells, and 3 of them are really bad; genomics, transcriptomics, may not be able to detect that.

So fundamentally, we have to account for that heterogeneity. We need to go to single-cell approaches to find those 3 bad cells. Single-cell genomics is just never going to work [inaudible 00:03:11]—it's too expensive and too complicated—and that's where AI solutions may play a role. I am bullish on pathology behind [inaudible 00:03:22] because I have seen presentations at ASCO GU and other conferences suggesting that these approaches are prognostic. At the same time, they have never been tested in this setting, so I can't say that they would work better, to be honest with you. But I think that it gives single-cell resolution in a study where you need single-cell resolution, and if pathology AI doesn't work to change management in this situation, then I'm not sure we are going to find anything.