00:00hello and welcome to the a 16z podcast
00:02sequencing the human genome matically
00:04changed how we understand how we as
00:06human beings are coded we're now
00:09entering a phase of building an
00:10applications layer on top of the
00:12sequencing layer so how do we make sense
00:14of and apply all this new information
00:16that genomics gives us and what will
00:18this translate into as it meets the
00:20realities of the healthcare system this
00:22conversation which took place at our
00:23annual summit event in November 2017
00:26includes Carlos Araya co-founder and CEO
00:28of jungle ax and Gabe OTT co-founder and
00:31CEO of freedom and was moderated by a 16
00:34ZJ neural partner Jorge Conde the first
00:37human genome project took 3 billion
00:39dollars over 13 years to generate a
00:41single human genome today we can do that
00:43same amount of work for about a thousand
00:46bucks in a couple of days in 1999 a
00:49writer by the name of Matt Ridley writes
00:51a book called the genome an
00:53autobiography of a species in 23
00:55chapters that book was a fascinating
00:56example of the optimism that we all had
00:59as the first human genome project was
01:02coming to an end but what we've also
01:03learned is that reading the DNA isn't
01:06the same thing as understanding it we've
01:08gone through this period of really
01:10trying to make sense of all of this
01:12information but what's extraordinary is
01:14that now that we can sequence DNA
01:16quickly and reliably and cheaply we've
01:19created this incredible sequencing layer
01:21on top of which we can build
01:22applications so that I think that's one
01:25of the big questions is how do we think
01:26about actionability how do we think
01:29about deriving meaning from our ability
01:31to interpret the both of you are
01:33developing applications in the clinical
01:34space so let's start with June glow
01:36Carlos if memory serves every time a new
01:40human genome sequence is completed there
01:42are on the order of 3 million new
01:44variants identified that's right
01:46so you know people talk about missing
01:48the forest for the trees how do you make
01:51sense of all of that information how do
01:54you figure out which trees matter in
01:55that forest you start by really trying
01:59to form context and so although we
02:02nowadays have access to our genomes
02:05while we go to interpret our genomes
02:08there's really you know only a certain
02:10number of places in the genome that are
02:13for any given condition that we're
02:15considering and putting the information
02:17in context of a condition in context of
02:20the family history in context of other
02:22tests is really imparted and so you
02:24start there and then you then look at
02:26individual genes that are associated
02:28with these typically its genes there are
02:30other types of elements and we look for
02:32the effects of variation in there and
02:35one of the interesting things is that
02:36while yes it costs roughly $1,000 and
02:40going down to the hundreds now per
02:42genome to acquire the data the cost of
02:45interpreting that data is actually
02:47really high although there are three
02:48million variants identified there will
02:50be roughly a hundred variants that are
02:53novel variants in disease associated
02:55genes so these are places of the genome
02:57that we know really matter and
02:58interpreting each one of those under the
03:01current clinical practices cost 50 to
03:03$100 so we're talking about a hundred to
03:06thousand fold increase in the cost of
03:08interpretation relative to the cost of
03:09data acquisition so we build models
03:12computational experimental to provide
03:14variant interpretation teams guidance
03:16that can tell them or help them
03:19understand how these variants relate to
03:22a molecular and cellular effects so they
03:24can interpret these and for the most
03:25part are you looking at variants that
03:27are associated with an increased risk of
03:29disease or are you looking at variants
03:31that are causing a disease most part
03:35it's increasing risk of disease there
03:37are of course causal variants for many
03:40genes there's a spectrum there are
03:42mutations that have very strong effects
03:43and then a gradient of mutations that
03:46have lower effects and understanding
03:48these differences is really you know
03:50it's it's it's an ongoing challenge if
03:52we look across all of the disease
03:53associated genes that we know today
03:55we only have clinical interpretations
03:58for roughly 0.6 percent of the possible
04:00mutations in them so 0.6% 0.6 percent
04:04yeah how do you actually communicate
04:06this information into a position because
04:07the vast majority of doctors out there
04:09are not geneticists agreed and you know
04:11something that interesting that's
04:12happened is that while it is you know
04:14physicians and doctors who order the
04:17tests increasingly this interpretation
04:19is happening in the back end in really
04:23the genetic test provider space where
04:25you have dedicated teams
04:26of interpreters that are doing the work
04:29of classifying these mutations and so
04:32the goal is really I think to give to
04:34physicians really very clear guidance on
04:37what the effects of mutations are and
04:39when we don't know we need to say also
04:40we really don't know what we do is
04:42really we build these models that can
04:45say okay for all these other mutations
04:46you know 9.4% of mutations that we don't
04:49know here is a subset for which we can
04:51make predictions and this is how well
04:52the predictions do and here's the
04:54diagnostic metrics of value they're
04:56prospectively tracked and they can be
04:57audited speaking of Diagnostics Gabe
04:59you're working on a diagnostic
05:01application as well using DNA so walk us
05:04through what you're doing I want to
05:06understand how what we're doing now with
05:07this kind of technology is different
05:09than how we've all sort of historically
05:11viewed Diagnostics sure it's good to
05:13take a step back and see or talk about
05:15like what do we mean by a gene or what
05:19do we mean by the DNA right now right so
05:21a bunch of people got 23andme done the
05:24truth of the matter is is 23andme
05:25sequences is trying to capture an
05:27information about you from less than 1%
05:29and that's just the static DNA what we
05:31think of as like DNA that we're born
05:33with but DNA is not static DNA is
05:37actually incredibly dynamic and it
05:39changes in all sorts of ways this is the
05:41reason why twins with the same DNA have
05:43very different outcomes think of it
05:46within your own bodies as well right in
05:48your bodies turn neurons that are a
05:51right single-cell meter long and then
05:54there are white blood cells that are
05:56literally turning over every day but
05:57they have the same DNA how does that
05:59happen how do we get these radically
06:01different phenotypes the truth of the
06:02matter is is what is in your DNA is
06:05probably far less consequential than how
06:08your DNA is being used and when these
06:12genes are being turned on and off
06:14because the truth of the matter is is
06:15less than 1% of your DNA is being used
06:17by any particular cell and so it really
06:20matters what that 1% is that ultimately
06:23makes your cells what they are makes you
06:26who you are so when you're looking at it
06:28from sort of a non-static perspective
06:31the DNA that's for example floating in
06:33your blood that is turning over every 20
06:35minutes can give you an insight of
06:38what's happening in your body
06:40why are the cells in your body dying at
06:43that moment and what is the composition
06:45of the cells in your body these are all
06:47relevant dynamic information that can be
06:50read out from your dynamic DNA that you
06:54can't get from your static DNA and the
06:56majority of the focus in the past 10 20
07:00years on our DNA has been really focused
07:03on that static DNA and not to mention a
07:05very small sub fraction of that static
07:07DNA what we're looking at at free gnome
07:10is that dynamic DNA as assayed from our
07:13blood sample that allows us to get an
07:15instantaneous snapshot of your molecular
07:18health which will then allow us to know
07:21whether you have a particular disease
07:22like cancer so this is an important
07:25distinction because I think one of the
07:27things that was so incredibly exciting
07:29and promising regarding this ability to
07:31have a sequencing layer on top of which
07:32we could build diagnostic applications
07:34was this idea that you sequence once and
07:37every time you want to test it's just a
07:40software query so the idea of every time
07:42you want to do a new test you basically
07:43have to redo the biology that goes away
07:46if you're looking at inherited risk but
07:48what you're describing is that since DNA
07:50is dynamic you would actually sequence
07:53in time periods much like we get you
07:55know dental x-rays every year yeah if
07:57you're really talking about being able
07:59to get a sense of how your body is
08:03you'd have to sequence multiple times
08:05because the DNA that you're born with
08:07that static DNA is not deterministic
08:10enough where you can predict everything
08:12from that single point of sequencing and
08:15so we're talking about how your body how
08:17DNA is so dynamic in your body does that
08:19mean that the applications for what you
08:21could diagnose using this approach are
08:23very broad and there's it just cancer
08:25well I think it really depends on what
08:27type of DNA you're looking for right if
08:29you're looking at DNA fragments that are
08:31in your bloodstream that are coming from
08:32the cancer cells and that's all you're
08:34focusing on then presumably you can only
08:37detect cancer what we're detecting is
08:39DNA fragments that are actually coming
08:42from the immune cells that are turning
08:44over in your body and if you can capture
08:47how your immune system is changing at
08:49different times it's sort of the common
08:52denominator to all disease condition
08:54if there's something wrong with you
08:55chances are your immune system is
08:57changing in some way and yes that signal
09:00is extremely convoluted and it's very
09:02hard to figure out what type of change
09:04is specific to a particular disease
09:06state but that's where things like
09:09machine learning artificial intelligence
09:10comes in to help us figure out the
09:12specific signal so that we can turn that
09:14into a specific diagnostic for a disease
09:16but the underlying biology should
09:19theoretically enable us to detect any
09:21diseases where there is an immune change
09:24so you guys have laid out the promise of
09:27this approach let's talk a little bit
09:29about the potential peril Carlos you
09:31were talking about what to do when you
09:33find a variant that has an unknown
09:35clinical significance and there was a
09:37lawsuit in Oregon where a woman had a
09:40hysterectomy because her physician
09:42misinterpreted the genetic test yeah
09:44that's right and it's really unfortunate
09:45because this is a very common situation
09:47where she had a family history she got a
09:51genetic test the genetic test actually
09:53came back negative but they included in
09:55the test information that said that
09:57there was a variant that was found that
09:58was of unknown significance and so the
10:00test very clearly indicated there was no
10:03clinically significant mutation found by
10:05the sort of practicing guidelines it was
10:08a negative test yet for I think reasons
10:11that these to me are unknown the
10:13recommendation for her was that she
10:15would have a bilateral mastectomy and
10:18the hysterectomy and it's I think it was
10:20a particularly sad because the mutation
10:23was actually in a gene that's not
10:24associated strongly at least with breast
10:27cancer it was in mlh1 I believe and you
10:30know there's a strong association with
10:31colorectal cancer and endometrial cancer
10:33but there really was no strong support
10:35for a breast cancer association to me
10:38this poses the challenges of how complex
10:41really this data has gotten to
10:42communicate even to physicians and then
10:45to have that message go clearly to
10:47patients that's incredible story yeah so
10:50look thinking about freedome you know
10:51perhaps this is unfair comparison or an
10:54example but if we think about early
10:55screening we've been using mammograms
10:57for 30 40 years and the data suggests
11:01that while we've actually done a lot of
11:04early detection of breast cancer using
11:07number of late stage cancers breast
11:10cancer sac she hasn't gone down so that
11:12implies that we've actually over
11:13diagnosed things right in some cases
11:15diagnosed the wrong things
11:16is there an analogy here or is that a
11:19worry for you and if so how do you
11:20control it so there's there's a lot of
11:23concerns in terms of launching a
11:24diagnostic I think what you're talking
11:26about is really under technology to
11:28science side of things right and this is
11:30one of the really interesting things for
11:32us because breast cancer and
11:33specifically mammography as a screening
11:35method has a false positive rate of 50%
11:3750% 50% so from a false positive
11:40perspective you're better off flipping a
11:42coin than you know doing a mammography
11:44really and there's a reason for this
11:46which is that no clinical trial that
11:48we've done has ever been large enough
11:50you don't actually compensate for all
11:55the false positive cases that could
11:57potentially happen all the false
11:58negative cases that could potentially
12:00happen in a clinical trial and so once
12:02you launch these Diagnostics into the
12:05market the only direction the
12:07performance goes is down it really
12:09doesn't really go up right because there
12:12are all these edge cases that you didn't
12:13account for them for the first time in
12:16our history were able to compensate for
12:17that problem where because we can now
12:21make a diagnostic that's fundamentally
12:23AI based even after we launched a test
12:25into market we can actually work with
12:28our partners to get results of the tests
12:30that we sell back so that we can teach
12:33the artificial intelligence that it made
12:36mistakes after we've sort of launched
12:38app s and it never makes that mistake
12:40again so for the first time we have an
12:42opportunity to make the direction of the
12:45accuracy of a test after we launch it go
12:47up as opposed to go down and that's
12:50really the only way we're going to
12:51compensate for this because the largest
12:52clinical trial that's ever been
12:53announced but hasn't been performed yet
12:55120,000 people last year 35 million
12:59people should have gotten screened for
13:01colorectal cancer in the United States
13:02alone and didn't 35 million 35 million
13:05right so if you can capture even a
13:07fraction of that market and learn from
13:10that information and make sure that you
13:12don't make mistakes from those tests
13:14ever again then all of a sudden you have
13:16a clinical trial that's an order of
13:17magnitude if not two greater than the
13:20coach all that's ever been announced so
13:22what you're both doing is so potentially
13:25transformative for our healthcare system
13:26who pays for this how does this get
13:28covered especially on the diagnostic
13:31side reimbursement is a particularly
13:32tough issue there's so many stakeholders
13:35in health care that it is not clear who
13:38pays for it if you're looking at the
13:39average statistics generally when you're
13:42launching a diagnostic test only about
13:4520% of the tests that you sell actually
13:47get fully reimbursed so 80% of tests
13:50that you're selling is actually not
13:52being paid for it properly the long
13:54story short is most of the time payers
13:57don't see a clear return on investment
14:01that the diagnostic tests represent so
14:04if I pay five hundred dollars for this
14:06test now am I actually going to make
14:08that money back because I'm detecting
14:10this disease earlier so we don't have to
14:12spend as much money curing this person
14:15and when the disease is progressed
14:16further that's a question that we need
14:19to be able to answer clearly before we
14:23can get these tests paid for by the
14:25pair's I think we're like stuck in this
14:28model where we're relying on payers to
14:30pay for these tests there are new models
14:32that are coming out that's leveraging
14:34life insurance companies that's
14:35leveraging these closed systems where
14:38these hospitals there are own pairs that
14:40are easier to sort of convince of the
14:42value of the tests I think we're going
14:44to see leveraging of these new models
14:46much more but it's still in the early
14:48days excellent Thank You Carlos
14:51Thank You Gabe for being here you're
14:52working in a fascinating space and it
14:55clear you're gonna change how we think
14:57about disease forever and for always so
14:59thank you thank you thank you guys