00:27:05 · Helmy Eltoukhy
[00:27:05] S0: Yeah, so it's frankly a really exciting time. Mean, one of the things we, our sort of tagline or mission statement is conquering cancer with data, which you're part of in terms of when we sort of revamped that. And we don't take that lightly. We have, I now one point three million cancer patients that we've tested. We have 700,000 epigenomic profiles, which I don't think anyone has. And we have all of that married up with a lot of real world evidence and outcomes data and so on. But the thing that we are missing, I would say, is enough bandwidth and ability to sort of do the analytics on that and drive a lot of the insights. And now with a lot of the new AI tools, it's really given us this amazing increase in velocity essentially. Every single one of our R and D individuals is now a manager of multiple agents essentially being able to go in and do all of these analyses and develop these applications on that data. There's a lot of excitement in the greater field about sort of like AI being used to cure disease or discover new drugs and so on, or foundation models for cells, and I would say that 90% of that is probably overhyped in some ways. And the reason for that is that the underlying data that's collected to essentially look at those models is very under sampled. It's looking at a lot of these cells or these processes or RNA transcripts with very low resolution. 60% of a cell is mostly housekeeping genes, and if you under sample it, all you're seeing is essentially the standard stuff, not the stuff that is actually making those cells unique. And so I think we took things the other way where we developed, I would say, very exquisite, like very deep technology to look at like a lot of those things, but we just didn't have the bandwidth to sort of analyze a little bit. And now we do. And so it's really been a, I would say, sort of a renaissance for us in terms of like, you know, the velocity of new applications we're developing. And it's really been a godsend, I would say, in terms of what we're doing. And now we're putting AI on other reports, and we're able to take stuff that I think you, like we all envisioned when you were there, that just wasn't possible technologically at the time.
00:29:40 · Mark Jacobstein
[00:29:40] S1: We had big plans in 2018. As I recall, I mean, there were a lot of issues. I mean, biology's complicated and then like, okay, we've got at the time hundreds of thousands, not millions of metastatic patients. We have a view across their genome. We've got an extra tube of blood in a freezer, but you needed phenotypic information and downstream stuff. It took a while, but play this out three, four years. What do you think? Now it's working. You got modern AI tools. You have over a million metastatic patient samples in the freezer and a lot of information on them already. What's the most optimistic scenario here? What falls out of this?
00:30:23 · Helmy Eltoukhy
[00:30:23] S0: Yeah, so I guess I can maybe go backwards and show you kind of the evolution that's happened. A story I tell is this tenth patient we sequenced, which we tested in 2014. She was a, I think, Bay Area woman, 48 year old woman diagnosed with pancreatic cancer initially. We ran a test, physicians said, this doesn't look like pancreatic cancer, maybe it's ovarian cancer. Ran a few more tests, found out it was ovarian cancer, and then put her on a drug that's been around for thirty years and shared a complete response. And our test was not designed to do that. It just gave a few clues that perhaps she was misdiagnosed. Fast forward now, our test, that same test that was ordered, can tell the physician that it is exactly this type of disease, that it's a subset of that type of ovarian cancer, that type of pancreatic cancer. It can tell the physician essentially likely response rates for certain drugs. The reports can now actually tell you the last thousand patients we tested that looked like this, seventy three percent of them were put on this drug, fifteen percent on this one, and so on, almost build you a sort of GPS of what is likely to sort of come next for that patient. And a lot of that was sort of powered by us moving from genomics to epigenomics, basically going from essentially the genetics, which you can think of as the hardware of every cell. Every cell in your body has the same DNA, yet they look very different. Your liver cells look different than your skin cells and so on, and a lot of that is because of the epigenetics. You can think of epigenetics as a software program that is actually being run on that cell, and so we're able to see that in the blood. And I think what's been exciting is that when we made that transition, not only allowed us to see the sort of phenotype of these cells or disease in much higher resolution, or cancer in much higher resolution, but it allowed us to go beyond cancer as well. And when you think about genomics, there are three areas that we've made big progress because of genomics: cancer, because the genome changes rare diseases, because you're born with mutations that maybe causes diseases and infectious diseases because by definition organisms have different DNA than human DNA. But for most other diseases, diabetes or cardiovascular disease or all this Autoimmune.
00:33:00 · Helmy Eltoukhy, Mark Jacobstein
Autoimmune, they don't change the genome, but they do change the epigenome. And so that's what's been exciting about marrying up, now having this early detection test where we're testing hundreds of thousands and soon millions of people a year that fortunately very few percentage have cancer, but all those individuals have many other diseases. And because we've built this architecture where we can see the claims data, we can see the outcomes data of almost every patient we test in a de identified sort of manner, we can now throw that into all the AI machine learning sort of scripts we've created, and we have labeled data now. So we can say, okay, these ten thousand people are in HUMIRA. They have rheumatoid arthritis. Are there any signals there we can sort of like, and so we're actually doing all of that in parallel now in a massive way, and we're seeing exquisite signals for some diseases that essentially would just be a software update on top of the existing tests that we have. [00:34:06] S1: That's amazing. And what's so interesting is that a lot of those disease areas have never had any kind of diagnostic. It's like they put you on Humira until it doesn't work. And [00:34:17] S0: there's [00:34:17] S1: not a biomarker, there's no understanding really, I think, of the mechanisms of action and why somebody is a likely responder or not, and presumably therapeutics will fall out of that also. I don't know if that's on the roadmap or if you can talk about that.
00:34:17 · Mark Jacobstein, Helmy Eltoukhy
[00:34:17] S1: not a biomarker, there's no understanding really, I think, of the mechanisms of action and why somebody is a likely responder or not, and presumably therapeutics will fall out of that also. I don't know if that's on the roadmap or if you can talk about that. [00:34:31] S0: Yeah, so I think the idea is like, we've talked about going, like the next thing is essentially thinking about non oncology applications, basically using our blood test as a sort of everything test to detect hundreds, if not thousands of diseases. We have about 500 labeled diseases in our database now that we're seeing, so it's really exciting in terms of like, I'm as excited now as I was 14 ago just because we can do some pretty massive things potentially. Then I think what's exciting is working with pharma in non oncology applications because there really is a lack of precision medicine in most areas and if you can essentially build these surrogate endpoints where you can measure disease more quantitatively at a much faster scale, you can decrease the cycles of learning and improve the progress considerably.
00:35:23 · Mark Jacobstein, Steven
[00:35:23] S1: There's an interesting macro lesson here for everybody in the room sort of thinking about AI, is all these things scale with data and the challenge is how do you get your data and Guardant has managed to sort of unlock a way to actually not only get the data, get paid to develop the data, which certainly makes it an awful lot easier. All right, well, thank you. I'm going to open it up now to other questions, but I would be remiss. So for those of you who don't, Steven is a South Park Commons member. He's also a former Guardant employee. I think you wrote the first bioinformatic pipeline that Guardant ever had, but also always had the first question at every all hands meeting at Guardant. So I'm very happy to have Stephen ask the first question. No pressure. [00:36:03] S2: So one of the things that I took away from Guardant Health was that we had the core value of putting patients first, and that's always been important to the company, it's been important to me. One of the things that happens as companies mature is you have the opportunity to democratize access to technology. How are you guys thinking about that?