The a16z Show - a16z Podcast: When Will Genomics Live Up to the Hype?

Episode Date: February 22, 2017

It's been nearly 15 years since the Human Genome Project was completed. But "are we there yet" in the golden age of genomics? What did we think we'd have by now, what do we actually have, and what do we really still need to make genomics live up to its promise? Well, one thing we now understand is that our DNA isn't static; in fact, it changes at an absolutely crazy rate. We also need to add more context -- about mutations, about somatic tissue, about phenotypes, about each person's unique history -- to make genetic information more complete and accurate. So what does that mean for predictive vs. diagnostic (which are two very different things) genomics? What are the challenges and opportunities for commercialization? The guests in this episode of the a16z Podcast -- Carlos Araya of Jungla,Jeff Kaditz of Q, and Gabe Otte of Freenome -- discuss all this and more with a16z bio fund partner Malinka Walaliyadde in a conversation that took place at our inaugural a16z Summit event. Stay Updated:Find a16z on YouTube: YouTubeFind a16z on XFind a16z on LinkedInListen to the a16z Show on SpotifyListen to the a16z Show on Apple PodcastsFollow our host: https://twitter.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures. Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.

Hi, I'm Hannah and welcome to the A16Z podcast. We've been hearing hype about the possibilities of genomics for decades now since the human genome project first began. So where are we today? In this episode, we look at where genomics actually is right now, what we hoped we would have by now, what we actually do have and don't, and what we still need to learn, given that we now know things like your genome changes at an unbelievable rate, from the phenotypic context needed to make predictions truly accurate to all the challenges and opportunities for commercialization. This episode was recorded at our inaugural summit, moderated by A16Z's Malinka Wallaliate, and joined by Carlos Araya from Jungla, Jeff Kedz from Q, and Gabe Ott from Frinome.

So we're about to get started with a session on genomics. The promise of genomics and where we are today. And we've got some fantastic entrepreneurs joining us. We've got Jeff, the CEO of Q, Carlos, the CEO. of Jongla, and we've got Gabe, the CEO of Freenome. All right. So when we sequence the first human genome as part of the human genome project, the promise was that we would cure all disease.

Now, that was 20 years ago. We certainly haven't lived up to that. And in fact, it's been a little difficult to see exactly where genomics has had a true application today. So I'd actually love to get a sense from the crowd. So I'd love to share a show of hands. How many of you have had your genome sequenced and have had that information meaningfully used to make a change in your health care?

Fairly small fraction, which is exactly what we've been thinking. So there's definitely a gap. There's certainly a gap. And so today we have some fantastic entrepreneurs to help us understand what that gap is and what we need to do to bridge that gap and help fulfill some of the potential here.

So maybe just to get started, let's try and understand what the human genome project was. Why did we do it? Why was it important? And maybe Carlos will start with you because I know you've been in the academic world most recently. Yeah, thank you. The human genome project is really, or was really,

biology's Apollo program. It was the first large-scale biology project that actually, you know, took billions of dollars to complete a coordinated effort across a large number of teams. It set out with the mission of providing basically a first draft sequence that was published in 2001 and later refined to a high-quality reference genome that we can all view. Now, this is a reference x-ray of what a quote-unquote normal genome looks like. And although that functions as a foundation for a lot of basically learning about

our origins, our biology and health and disease, like the Apollo program, a lot of the value from the human genome project isn't just that reference sequence, but it's also the technologies and concepts that were learned and developed therein. Genomics is a fairly abstract concept for many people. How would members of our audience experience genomics today? What are the major use cases? Who are the major players today? Genomics is probably one of the largest big data problems out there. There's three billion bases in a genome. And if that was all, we probably would know a lot more about the genomics and what it can tell us. But unfortunately, that's really just, you know, the beginning of that whole picture, right? Not only the way we call, like right now, each one of those bases

and how we call them is probabilistic, not deterministic in nature, because our technology doesn't necessarily allow us for a deterministic call. So there's sort of a confidence interval around calling mutations, but then there's also this idea that your DNA, your genome is not static throughout your lifetime. And in fact, I'll let Jeff talk about this because he did some math this morning around this, but your genome changes at a ridiculous rate and the fact that we thought taking one person's genome at one snapshot was going to answer every question about diseases and things like that was just ludicrous in retrospect. I agree with that. I think Gabe was just referring to the fact that we were talking a little bit earlier this morning and did some back of the envelope

calculations. And I think the data transfer rate of somatic DNA in your body is about 500 terabytes per second. And maybe you can explain what somatic versus drumline. It basically just means that the rate at which your DNA is copied in your body per second is about 500 terabytes. So if you think about the error correction codes, I mean, so my background, it's kind of an honor to be up here. These guys are biologists. I'm a physics and computer science. So I look at biology is an information theory problem. But going back to what Gabe said, when you think about, all right, well, 500 terabytes per second, what kind of error correction codes do you need in order to make sure that information is copied correctly? And when you think about disease like cancer,

really, those are information corruption problems. And when we talk about solving cancer, it's a little bit scary also because that information corruption is what allows us to evolve. So I think, you know, when I met Gabe, it was very exciting to me because it was one of the first, or the first person on the biology side that really thought of DNA is actually this thing that was much more dynamic than this thing that could be single, you know, by the time the biological sample you take that is going to get sequence gets to the lab, your genome is basically different. So I think that's an important realization. Got it. And then in terms of the use cases of genomics today, so, you know,

prenatal things like that. What would you go to the doctor for today? What would the doctor use your genomic information for today? Like what's a test of things that audience members could do potentially? Well, it's multi-tiered, right? So you can do things like 23MME, which looks at less than 1% of the entire genome, looking at specific mutations that you were born with

and what that can tell you about who you're going to be. That's largely predictive and very probabilistic. And all the way to sort of clinical diagnosis, like what you refer to non-invasive prenatal testing, for example, where they're doing much more sort of whole genome or certainly whole chromosome-wide sequencing of both the mother and the fetal DNA to essentially figure out, you know, the genomic nature of the fetus. So that's more on the diagnostic side. So we really have applications all from sort of germline mutation detection all the way to diagnostics and even prognostic methods. Got it.

So let's get back to try and understand that gap that we just talked about. What are the major technical challenges that are facing the genomics field? today, you think. Okay. So I think, you know, we've done a pretty good job at being able to acquire sequence information that's, you know, some of the fastest advances in technology in the history of mankind. I'm told it's actually only beat by one other technology, which is the sort of the clarity of glass improved at a faster rate over a period of time. But getting basically, you know, access to this information doesn't mean understanding it. And so I think, you know, one of our sort views is that a critical missing component here is basically the maps of function for how we're

going to interpret mutations in here. We're getting large numbers of individual genomes for people they're changing because they change over their somatic tissues, they develop tumors, etc. All of that is information that we need to put in context. We need to be able to associate mutations that have similar effects. And unfortunately, the maps that we have today are really maps of function that just say where things that are, things like genes, biomolecules, where they are encoded in the genome. But it says really nothing about how they function and which parts of the genes do what. And that's really what mutations target. So that's, I think, one fundamental layer that's really missing for a lot of the applications that we pursue of genomics.

I think applications have also been extremely limited because one of the things that you need to understand genomic data is also phenotypic information associated with them. that. Can you describe what you mean by phenotypic information? So, you know, one simple example is when you get a, say, blood sample and I'm extracting DNA from that truth, try to understand the genomics behind whether this person has cancer or not. I need to know whether that person had cancer or not. I need to know whether that person was male or female, what age, some kind of background information about that person so that I can properly annotate that particular data. So for the physical characteristics. Yeah, exactly. It's physical characteristics. And what's been severely lacking is a deeper

understanding of the phenotypic information that we can associate back to genomic information, something that almost everyone that's doing research in genomics would agree it's really hard information to get. And it's really hard to get really clean information around that, even when you get information. So let's switch over to the business side a little bit. What do you think are the major commercial challenges that are facing the genomics industry today? I think one thing we're thinking about is, do we think of the applications of genomics being diagnostic or therapeutic. I mean, we quickly describe what each of those are. Well, sure. Is it a tool that we use to determine if you're sick, or is it a tool that we use to help you heal if you are sick?

And I think that, you know, if you look at just a single shot whole genome sequencing, I think there's another question to ask if you want it to use it diagnostically, which is at what point is a prediction a diagnostic? And I think that's a little bit of a question like saying, if I keep taking a grain of sand off of a pile of sand, at what point is it no longer a pile of sand? Because if you think about the expectations in healthcare, a doctor really most of the time is expected to give a binary decision of, are you sick or are you not? Looking at your whole genome-segon for the vast majority of cases, it's just going to give you a statistical likelihood of diseases you may be more predisposed to than another person. But it will be almost entirely environmental factors that determines whether that's expressed. I think that the most immediate or obvious places, and I think Russ Altman at Stanford is doing

really interesting things here is using genetics to determine which drugs you're most likely to respond to. I think that that is like the lowest hanging fruit. I think in order for genomics to be used in diagnostics or predictive models of are you going to get sick, I think it has to be combined with actual time series biomarker data, which is a, a, a longer potential discussion. Part of the challenges, I think, the diagnostics field, genomics field has had, is our healthcare system, which is you need to

convince a payer, an insurance company to reimburse you, then you need to also convince a doctor to prescribe that test, and only once you get both those parties on board, can you actually go to market? That's a lot of people to convince, and really the only person you should be convincing is the patient

who is nowhere in that equation at all. So are there other ways to get to market that are compelling? Maybe could break past those barriers. You do point out at a really interesting point, which is when I talk to clinicians, when I talk to pairs, it's often an argument about whether detecting cancer is a good thing or not, whether that will lead to what they care about, which is sort of savings in the medical system.

I think the really interesting thing is we've been used to doing things a certain way, like in the field of cancer screening, cancer diagnostics, we're used to like really, really bad tests. So like PSA for prostate cancer detection, mammography for breast cancer detection, these things have false positive rates of anywhere from 50 to 75%. Right. Like you're literally better off flipping a coin than taking one of these tests from a false positive perspective. And so, yes, of course, if our test are that inaccurate, it's going to lead to all sorts of downstream, you know, unnecessary procedures that adds burden.

But so many people have that mentality where they basically say, we're not going to reimburse this unless your test saves me money now. Right? Not 10 years from now, not five years from now when this person is dying of cancer, but like, is it going to save me money now? Yeah. So I think, you know, from a business model perspective, there's a lot of opportunities to really work in, I guess, what's broadly known as the wellness space. And as our technology improves, and we can start detecting diseases so early that we can

affect even lifestyle changes to potentially avoid certain types of diseases, I think that's really the future that we need to head towards because consumers are getting screwed over by this sort of old-style mentality of, is this test really going to save money for the payers or not? I think most people in this room can agree that, you know, detecting cancer earlier rather than later is probably a good thing. So I don't think there's an argument from a consumer perspective. It's really the payers and some of the clinicians that are being the inhibitors to this progress. I agree with everything Gabe just said. But I think there's also kind of a societal and ethical question of what rights do we have as patients to have

information about our bodies. Because we, you know, there's, there are regulatory bodies that say, you know, if you look what happened to 23 and me, effectively the argument for shutting them down was, well, if you say I have an increased risk of breast cancer and then I go home and cut off my own breasts and die, then 23 in me is liable. Now, that seems a little strange to me considering we live in a world where there's like a surgeon general's warning on alcohol and cigarettes, and we know that's just bad for you. So we have to ask yourselves, is it reasonable that we have access to this information about it? And can we be responsible as patients for having access to that information rather than something telling us it's dangerous for you to have access

to that information? And I think that's a big question that, you know, I don't think patients or clinicians are on the same side of right now. I want to chime in on this on the challenges for commercialization, because they are really important to all of us. And I completely agree there's challenges in the regulation side. There are challenges in the regulations side. There are challenges in in the reimbursement side, which are coupled to that. And there's challenges also in showing value to customers, showing clear, understandable value of the products that they're buying in genomics. And I think that kind of couples to what Gabe was saying earlier, that we need more phenotypic data, more clinical data, for example, to support the value of decisions or guidance that we can get

out of genetic information. Now, it stands to reason that, you know, we didn't have hundreds of thousands of genomes a few years ago. And so we didn't have hundreds of thousands of genomes coupled to, you know, EHR systems. But that's definitely the way that a lot of this is moving. And being able to access that information is going to be able to test what, you know, how well do different models distinguish between different outcomes on the basis of genetic information and being able then to evaluate, you know, okay, what is the value then of guiding decisions on this basis? So it's still early, but I think there's a path forward that's being set up quickly. Actually, I would love to hear more about what your companies are specifically doing.

We just talked about some of the challenges about technical and commercial, what your companies are doing to get over those obstacles and why now is the time that actually works for your companies. Sure. One thing I always like to point out is 80% of all the money that we spend on treating and dealing with cancer in the healthcare system in the United States, something about between $75 and $100 billion a year is to help people. die of cancer. That's what we spend 80% of the money on right now. And that's, you know, that's really unacceptable. And the fact that it's such a high percentage really, in some ways, makes my argument is if we create an accurate enough cancer test that detects the disease

early enough when it's actually treatable, we save that 80% of the money. And so when dealing with the payers, it's really about listening to them. And for them, you know, what kind of evidence do they need to really, you know, show that we can save that 80% for them, that they're reimbursing. And then, of course, we already talked about the wellness angle is if and when I do get fed up with the pairs, which may be sometimes soon. There are other opportunities to explore, especially because FDA has released guidelines around a wellness space, and they basically said that patients have the right to choose their own lifestyle choices, their diet and exercise, and how that can potentially affect their wellness. And so one of the things that we're really working on is how can we empower the patient with the right, essentially genomic thermometer, if you will, to give them a sense of what kind of things can they eat and how much should they exercise for that particular individual to maximize their wellness and avoid chances of getting these kinds of diseases. I like that model because we don't have to talk to payers.

We don't have to deal with a lot of people. All we have to do is make sure that the test gets to a price point that's affordable for the best majority of people. Got it. Yeah, I think fundamentally, I think the point. that you know, bring up is that if you think about who the pairs are

in healthcare right now and you think about the actuarial models that have been built, they're completely reactive and backwards looking. And all of these new technologies with genomics,

transcriptomics, proteomics, metabolomics, microbiomics, all this stuff is really much more powerful as a preventative tool. And so how do you convince

an entire industry that looks as spending a dollar is a dollar lost, right? that when you have these tools that you say can in the long run save money but they're preventative. It's not like we're not, we don't want to spend money when somebody's already sick. And that's the fundamental problem that any of these new technologies have when you're trying to find who the payers. And so I think it's not clear to me that the existing payers will actually ever come around to that.

That's a fair point. I think there's a really good opportunity. As you bring down the cost of these tests and you have consumers able to directly pay for them, Or sometimes actually go outside the U.S. There's a lot of countries where it's either single pair or it's very much self-pay and consumers are used to paying for tests by themselves, like in India. Well, think about dental care in the United States. That's actually a preventative.

I would argue that that's one of the best preventative health care systems in the world. How does it work? Well, twice a year you go get the same set of things basically measured about your body and we develop hundreds of millions of longitudinal, you know, medical, dental records tied to outcomes. So there's this positive feedback loop in the dental industry where actually if you look at the cost of dental care over time, it's flat or down in inflation adjusted dollars and the quality of the care has gone up. If you look at over the exact same period of time in health care, it's the exact opposite trend. Care in a lot of ways is getting

worse. Costs are skyrocketing. And so I think it's worth kind of asking what is the difference is why. And in the dental care system, you know, there are dentists who give away checkups. to get you as a customer because they know eventually you're going to need a root canal. Right? And that's when you pay. Yep. Right? And so they take the preventative approach saying, you know, it's inevitable.

We're all going to get sick. We're all going to die. Like, and that's when we should get paid. We're going to try and keep you healthy until then. Yeah. And it's very fundamentally different approach to, but health care nonetheless. I think it's important to remember the patients and the consumer's role in this because it's really not going to work without the patients and consumers.

opting into these kinds of behavioral changes. Effectively, we're asking people to look at your health in a different way. Because Jeff brought up the dentist's example, right? We brush our teeth every day, or hopefully every day, twice a day. And usually have annual checkups and things like that. And we don't really expect to go to the dentist after 20 years of not brushing our teeth and expecting them to make our teeth perfect. But that's exactly what we ask of our doctors today.

It's like people treat their bodies terribly. terribly. And then when they are on the verge of death, they go to the doctors and then say, you know, make me perfect. Then the insurance company is complaining because it's expensive to fix you. Right. So, so, you know, it really needs sort of patient and consumer buy-in, but that's partly on them, partly also on the technology companies, making it easier for them or enabling them or certainly motivating them in certain ways. Actually, as you talk about that, in order for us to get to that, well, we do definitely need to have better diagnostic tests, higher quality, lower cost.

And we're very long AI in Silicon Valley. We think AI plus any industry leads to a giant company in that industry generally. AI plus cars, AI plus radiology. What does AI plus genomics look like? So we at Jungler have sort of realized that in this explosion of genomics entering more and more of the clinic, what you see is that there is this really large growth in the amount of uncertainty around genetic tests. And it's kind of ironic that when you look at even some of the most abundant, large volume, genetic tests, these are cancer gene panel tests for basically, you know,

hereditary risk of developing cancer. Those tests will basically find 95 mutations that they have absolutely no clue of what the effects of those mutations are in these important cancer-associated genes per each mutation that is known to cause disease. So when you can only interpret basically one out of 90-19. mutations in this space, it's kind of ironic that we call it precision medicine. We see that as a fantastic place for machine learning. And in fact, the American College of Medical Genetics has guidelines for the use of computational tools in this space. And so that's really, you know,

one of the key places where we're focusing on entering because we know that today we can make those tools 35% better. And that we're, you know, happy to distribute these tools as a horizontal across lots of different genetic test providers and not in containing the singles. To add to that, when you think about, you know, really what, I mean, AI is it just lets us combine a lot of different variables and make predictions. But, you know, we live in a world where Google and Facebook use millions of variables to predict which ad we're going to click on or, you know, whether or not we can repay a loan. But we try and reduce disease to single variables. And it's happening in genetics right now with a single nucleotide varying. We want to say if you're going to get

breast cancer. That's crazy. Honestly, human body is an extremely complicated system to build predictive models based on a single variable, which is exactly what clinical studies do, because that's what they're designed to do, really, is pretty assinine. And I think we have to think about, you know, why it's true in one place and not the other part of society. And I think that applying machine learning is really just applying a tool that lets us look at far more information. I mean, can you imagine somebody in real time trying to figure out which ad you're going to click on better than Google? It's just two things to look at, right? And that's what these tools allow us to do. Look at so many things, way more powerful than the human brain could ever look at.

And I think that's really what it's about. Yeah, and just adding to that, because I completely agree with that. From the patient's perspective and the consumer's perspective, what the learning engine can really provide is a way to make sense of all of that data in a way that pertains to them. These kinds of things, Apple watches, Fitbits and things like that, phenomenal at gathering data, right? Very bad at telling you what that actually does for you or what that, you know, means for you, right?

On average, an average human being makes, you know, 12 cancer cells every minute, right? Sorry to freak out, you know, people in the audience, but, you know, you need to, like, really think about that and realize, you know, there is an optimum for you in terms of what kind of food you eat, how you exercise, you know, how much you sleep, things along those lines that really minimize those kinds of events. Right now, patients have zero idea, right? How what they do in their lives actually correlates to any kind of changes in that space. That is a huge big data problem that's just starting to get deconvaluted. and that's where the machine learning can really step in

and figure up those kinds of interactions and correlations for us so that we can provide a consumer. Here I think there's a really important distinction to make in genetics, which is there are Mendelian disorders where the phenotype is correlated with single genes, and there's estimates are that there's at least 7,000 of them, and we know like 3,500 of them. And then there's complex sort of phenotypes.

Maybe give examples of what each is. Yeah, a standard sort of Mendelian disorder would be cystic fibrosis, where you have basically issues in the lung, and it really is driven by mutations in a chlorine channel. In the case of sort of complex disorders, we're talking about things more... Diabetes, you know, things related to sort of metabolism, where you have lots of genes contributing. Those are definitely controlled by lots of different variables. And genetics plays a role in both of those, and I think it's really important that we remember, you know, which type of disorder we're talking about when we think about how we're applying AI and

We think very carefully like, okay, what is the features and types of data that we're correlating with these? And I, yeah. Great. Well, thank you so much for that session. Let's thank our guests.