00:00hi I'm Hannah and welcome to the a 16z
00:02podcast we've been hearing hype about
00:05the possibilities of genomics for
00:07decades now since the human genome
00:09project first began so where are we
00:12in this episode we look at where
00:14genomics actually is right now what we
00:16hoped we would have by now what we
00:18actually do have and don't and what we
00:21still need to learn given that we now
00:23know things like your genome changes at
00:25an unbelievable rate from the phenotypic
00:28context needed to make predictions truly
00:30accurate to all the challenges and
00:32opportunities for commercialization this
00:35episode was recorded at our inaugural
00:37summit moderated by a 16 ZZZ malinka
00:40walala at a and joined by Carlos Araya
00:43from Joomla Jeff cadence from Q and Gabe
00:46OTT from free gnome so we're about to
00:49get started with a session on genomics
00:51the promise of genomics and where we are
00:53today and we've got some fantastic
00:54entrepreneurs joining us we've got Jeff
00:57sieh of cue Carlos the CEO of jungler
01:00and we've got Gabe the CEO of free no
01:03all right so when we sequence the first
01:08human genome as part of the Human Genome
01:10Project the promise was that we would
01:15now that was 20 years ago we certainly
01:19haven't lived up to that and in fact
01:21it's it's been a little difficult to see
01:23exactly where genomics has had a true
01:25application today so it actually love to
01:27get a sense from the crowd so I'd love
01:29to share show of hands how many of you
01:30have had your genome sequenced and have
01:32had and have had that information
01:34meaningfully used to make it change in
01:36your health care fairly small fraction
01:39which is exactly what we've been
01:42thinking so there's there's definitely a
01:43gap there's certainly a gap and so today
01:46we have some fantastic entrepreneurs to
01:48help us understand what that gap is and
01:50what we need to do to bridge that gap
01:52and help fulfill some of the potential
01:55here so maybe just to get started let's
01:58try and understand what the human genome
01:59project was why did we do it why was it
02:03important and maybe corals we'll start
02:04with you because I know you've been in
02:06the academic world most recently yeah
02:08the human genome project is really or
02:10was really biologies Apolo program
02:13it was the first large-scale biology
02:16project that actually you know took
02:18billions of dollars to complete and a
02:20coordinated effort across a large number
02:22of teams it set out with the mission of
02:25providing basically a first draft
02:28sequence that was published in 2001 and
02:30later refined to a high quality
02:32reference genome that we can all view
02:35now this is an a reference x-ray of what
02:37a quote-unquote normal genome looks like
02:40and although that functions as a
02:42foundation for a lot of basically
02:45learning about our origins our biology
02:48and health and disease like the Apollo
02:51program a lot of the value from the
02:54human genome project isn't just that
02:56reference sequence but it's also the
02:58technologies and concepts that were
03:01learned and developed there in genomics
03:04is a fairly abstract concept for many
03:05people how would members of our audience
03:07experience genomics today what are the
03:10major use cases who are the major
03:11players today genomics is probably one
03:14of the largest data problems out there
03:16there's three billion bases in the
03:18genome and if that was all we probably
03:21would know a lot more about the genomics
03:23and what it can tell us but
03:25unfortunately that's really just you
03:27know the beginning of that whole picture
03:29right not only the way we call like
03:33right now each one of those bases and
03:35how we call them is probabilistic not
03:36deterministic in nature because our
03:38technology doesn't necessarily allow us
03:40for a deterministic call so there's sort
03:43of a confidence interval around calling
03:44mutations but then there's also this
03:46idea that your DNA your genome is not
03:49static throughout your lifetime and in
03:51fact I'll let Jeff talk about those
03:53because he did some math this morning
03:54around this but your genome changes at a
03:57ridiculous rate and the fact that we
04:00thought taking one person's genome at
04:02one snapshot was gonna answer every you
04:05know question about diseases and things
04:07like that was just ludicrous in
04:08retrospect I agree that I think Gabe was
04:11just referring to the fact that we were
04:12talking a little early this morning and
04:14did some back of the envelope
04:16calculations and I think the data
04:19transfer rate of somatic DNA in your
04:22body is about 500 terabytes per second
04:25then maybe you can explain what somatic
04:27basically it just means that the rate at
04:29which your DNA is copied in your body
04:31per second is about 500 terabytes
04:34so if you think about the you know error
04:36correction codes I mean so my background
04:39it's kind of an honor to be up here
04:41these guys are biologists I'm a physics
04:43and computer science so I look at
04:44biology as an information theory problem
04:46but you know going back to what Gabe
04:49said when you think about alright well
04:51500 terabytes per second what kind of
04:53error correction codes do you need in
04:56order to make sure that information is
04:57copied correctly and when you think
04:59about disease like cancer really those
05:01are information corruption problems and
05:03when we talk about solving cancer it's a
05:05little bit scary also because that
05:06information corruption is what allows us
05:08to evolve so I think you know when I met
05:12Gabe it was very exciting to me because
05:14it was one of the first or the first
05:16person on the biology side that really
05:19thought of Dean is actually this thing
05:20that was much more dynamic than this
05:23thing that could be single you know by
05:25the time the biological sample you take
05:27that is going to get sequence gets to
05:29the lab you genome is basically
05:31different so I think that's an important
05:35realization and then in terms of the use
05:38cases of genomics today so you know
05:42prenatal things like that like what
05:44would you go to the doctor for today
05:45what would the doctor use your genomic
05:47information for today like what's a test
05:49of things that audience members could do
05:51potentially what's multi-tiered right so
05:53you can do things like 23andme which
05:56looks at less than 1% of the entire
05:57genome looking at specific mutations
06:00that you were born with and what that
06:01can tell you about who you're going to
06:02be it's largely predictive and very
06:05probabilistic and all the way to sort of
06:07clinical diagnosis like what you refer
06:09to a non-invasive prenatal testing for
06:12example where they're doing much more
06:15sort of whole genome or certainly whole
06:16chromosome wide sequencing of both the
06:19mother and the fetal DNA to essentially
06:21figure out you know the genomic nature
06:23of the fetus so that's more on the
06:26diagnostic side so we really have
06:27applications all from sort of the term
06:29my mutation detection all the way to
06:31Diagnostics and even prognostic methods
06:33got it so let's get back to trying to
06:36understand that gap that we just talked
06:37about what are the major technical
06:40that are facing the genomics field today
06:42you think okay so I think you know we've
06:44done a pretty good job at being able to
06:47acquire sequence information let's you
06:50know some of the fastest advances in
06:52technology in the history of mankind I'm
06:54told it's actually only beat by one
06:56other technology which is the sort of
06:57the clarity of glass improved at a
06:59faster rate over a period of time but
07:01getting basically you know access to
07:03this information doesn't mean
07:04understanding it and so I think you know
07:07one of our sort of views is that a
07:09critical missing component here is
07:12basically the maps of function for how
07:15we're gonna interpret mutations in here
07:17we're getting large numbers of
07:19individual genomes for people they're
07:21changing because they change over you
07:22know their somatic tissues they develop
07:24you know tumors etc all of that is you
07:27know information that we need to put in
07:30context and we need to be able to
07:31associate mutations that have similar
07:33effects and unfortunately the maps that
07:34we have today are really maps of
07:37function that just say where things that
07:40are things like genes biomolecules where
07:43they are encoded in the genome but it
07:45says really nothing about how they
07:47function and which parts of the genes do
07:50what and that's really what mutations
07:52target so that's I think one fundamental
07:55layer that's really missing for a lot of
07:57the applications that we pursue of
08:00I think applications have also been
08:02extremely limited because one of the
08:04things that you need to understand
08:05genomic data is also phenotypic
08:07information associated with that it's
08:09not what you mean by phenotypic into me
08:10so you know one simple example is when
08:13you get a safe blood sample and I'm
08:15extracting DNA from that root try to
08:17understand the genomics behind whether
08:18this person has cancer or not I need to
08:20know whether that person had cancer or
08:21not I need to know whether that person
08:23was male or female what age some kind of
08:26background information about that person
08:28so that I can properly annotate that the
08:31physical characteristics yeah exactly
08:32it's a physical characteristics and
08:34what's been severely lacking is a deeper
08:37understanding of the phenotypic
08:39information that we can associate back
08:41to genomic information something that
08:43almost every one that's doing research
08:45and genomics would agree it's really
08:47hard information to get and it's really
08:49hard to get really clean information
08:51around that even when
08:52information so let's switch over to the
08:54business side a little bit what do you
08:55think are the major commercial
08:57challenges that are facing the genomics
08:59industry today I think one thing we're
09:01thinking about is do we think of the
09:04application of genomics being diagnostic
09:06or therapeutic mmm-hmm I mean we do
09:08typically describe what what each of
09:09those are well sure is this is it a tool
09:11that we use to determine if you're sick
09:13or is it a tool that we use to help you
09:16heal if you are sick and I think that
09:20you know if you look at just a single
09:23shot whole genome sequencing I think
09:24there's another question to ask if you
09:27want it to use it diagnostically which
09:29is at what point is a prediction a
09:31diagnostic and I think that's a little
09:33bit of a question like saying if I keep
09:35taking a grain of sand off of a pile of
09:37sand at what point is it no longer a
09:38pile of sand because if you think about
09:40the expectations in healthcare a doctor
09:43really most of the time is expected to
09:44give a binary decision of are you sick
09:47or are you not looking at your whole
09:49genome check them for the like vast
09:50majority of cases it's just gonna give
09:52you a statistical likelihood of you know
09:55diseases you may be more predisposed to
09:58than another person but it will be
10:00almost entirely environmental factors
10:02that determines whether that's expressed
10:04I think that the most immediate are
10:07obvious places and I think Russ Altman
10:09at Stanford is doing real interesting
10:10things is you know using genetics to
10:13determine which drugs you're most likely
10:15to respond to I think that that is like
10:17the lowest hanging fruit I think in
10:20order for genomics to be used in
10:24Diagnostics or predictive models or are
10:27you going to get sick I think it has to
10:29be combined with actual time series
10:32biomarker data which is a longer you
10:35know discussion part of the challenges I
10:38think Diagnostics field genomics school
10:39has had is our healthcare system which
10:42is you need to convince a pair an
10:44insurance company to reimburse you then
10:46you need to also convinced and a doctor
10:48to prescribe that test and only once you
10:50get both those parties on board can you
10:52actually go to market that's a lot of
10:54people to convince and really the only
10:57person you should be convincing is the
10:58patient who is nowhere in that in that
11:00equation at all so are there other ways
11:02to get to market that are compelling
11:05that that could break past those
11:07barriers you do point out at a really
11:08interesting point which is when I talk
11:10to clinicians when I talk to pairs it's
11:13often an argument about whether
11:15detecting cancer is a good thing or not
11:17whether that will lead to what they care
11:20about which is sort of savings in the
11:22medical system I think the really
11:24interesting thing is we've been used to
11:27doing things a certain way like in the
11:29field of cancer screening cancer
11:31diagnostics we're used to like really
11:33really bad tests it so like PSA for
11:35prostate cancer detection mammography
11:37for breast cancer detection these things
11:39have false positive rates of anywhere
11:40from 50 to 75 percent all right like
11:43you're literally better off flipping a
11:44coin than taking one of these tests from
11:46false positive perspective and so yes of
11:48course if our tests are that inaccurate
11:50it's going to lead to all sorts of
11:52downstream you know unnecessary
11:54procedures that adds burden but so many
11:57people have that mentality where they
11:59basically say we're not gonna reimburse
12:01this unless you your tests save me money
12:03now right I'm not not ten years from now
12:06not five years from now when this person
12:08is dying of cancer but like is it gonna
12:10save any money now yeah so I think you
12:12know from a business model perspective
12:14there's a lot of opportunities to really
12:17work in I guess what's broadly known as
12:19a wellness space and as our technology
12:21improves and we can start detecting
12:23diseases so early that we can affect
12:26even lifestyle changes to potentially
12:28avoid certain types of diseases I think
12:30that's really the future that we need to
12:31head towards because consumers are
12:33getting screwed over by this sort of old
12:37style mentality of is this really going
12:40to save money for for the payers or not
12:42I think most people in this room can
12:44agree that you know detecting cancer
12:45earlier rather than later is probably a
12:47good thing so I don't think there's an
12:49argument from a consumer perspective
12:50it's really the payers and some of the
12:52clinicians that are being the inhibitors
12:53to this progress yeah I agree with
12:55everything gave your said but I think
12:57there's also kind of a societal and
12:59ethical question of what rights do we
13:02have as patients to have information
13:05about our bodies because we you know
13:08there's their regulatory bodies that say
13:10you know if you look what happened to
13:1123andme effectively the argument for
13:14shutting them down was well if you say I
13:16have an increased risk of breast cancer
13:18and I go home and cut off my own breasts
13:20and die then 23andme is liable now that
13:26seems a little strange to me considering
13:29we live in a world where there's like a
13:30surgeon general's warning on alcohol and
13:32cigarettes and we know that's just bad
13:34for you so we have to ask ourselves is
13:36is it reasonable that we have access to
13:38this information by the way and can we
13:39be responsible as patients for having
13:42actually federal information rather than
13:43someone telling us it's dangerous for
13:45you to have access to that information
13:46and I think that's a big question that
13:48you know I don't think patients or
13:50clinicians are on the same side of right
13:52now I want to chime in on this on the
13:55challenges for commercialization because
13:56they are really important to all of us
13:58and I completely agree there's
13:59challenges in the regulation side there
14:01are challenges in the reimbursement side
14:04which are coupled to that and there's
14:05challenges also in showing value to
14:08customers showing clear understandable
14:10value of the products that they're
14:12buying in genomics and I think that kind
14:14of couples to what Gabe was saying
14:16earlier that we need more phenotypic
14:18data more clinical data for example to
14:21support the value of decisions or
14:23guidance that we can get out of genetic
14:25information now it stands to reason that
14:27you know we didn't have hundreds of
14:29thousands of genomes a few years ago and
14:31so we didn't have hundreds of thousands
14:32of genomes coupled to you know EHR
14:35systems but that's definitely the way
14:36that a lot of this is moving and being
14:39able to access that information is going
14:41to be able to test what you know how
14:43well do different models distinguish
14:46between different outcomes on the basis
14:48of genetic information and being able
14:50then to evaluate you know okay what is
14:52the value then of guiding decisions on
14:54this basis so it's still early but I
14:57think there's a path forward that's
14:58being set up quickly actually I would
15:00love to hear more about what your
15:01companies are specifically doing we just
15:03talked about some of the challenges both
15:04technical and commercial what your
15:06companies are doing to get over those
15:08obstacles and why now is the time that
15:10that actually works for your companies
15:12sure one thing I always like to point
15:15out is 80% of all the money that we
15:18spend on treating and dealing with
15:21cancer in the healthcare system in the
15:23something about between 75 and 100
15:26billion dollars a year is to help people
15:29die of cancer that's what we spend 80%
15:32on right now and that's you know that's
15:36really unacceptable and the fact that
15:38it's such a high percentage really in
15:40some ways makes my argument is if we
15:43create an accurate enough cancer test
15:45that detects the disease early enough
15:47when it's actually treatable we saved
15:50that eighty percent of the money and so
15:52when dealing with the payers it's really
15:54about listening to them and for them you
15:56know what kind of evidence do they need
15:57to really you know show that we can save
16:01that eighty percent for them that
16:03they're reimbursing and then of course
16:05we already talked about the wellness
16:06angle is if and when I do get fed up
16:08with the payers which may be some time
16:10soon there are other opportunities to
16:12explore especially because FDA has
16:14released guidelines around a wellness
16:16space and they basically said the
16:18patients have the right to choose their
16:20own lifestyle choices their diet and
16:22exercise and how that can potentially
16:24affect their wellness and so one of the
16:27things that we're really working on is
16:28how can we empower the patient with the
16:30right essentially genomic thermometer if
16:32you will to give them a sense of what
16:35kind of things can need and how much
16:37should they exercise for that particular
16:38individual to maximize their wellness
16:40and avoid chances of getting these kinds
16:42I like that model because we don't have
16:44to talk to payers we don't know yep deal
16:46with a lot of people all we have to do
16:48is make sure that that test gets to a
16:50price point that's affordable for the
16:51vast majority of people yeah I think
16:53fundamentally I think the the point that
16:55you know give bringing up is that if you
16:58think about what the payers are in
17:00healthcare right now and you think about
17:02the actuarial models that have been
17:03built they're completely reactive and
17:05backwards looking and all of these new
17:07technologies with genomics
17:09transcriptomics proteomics metabolomics
17:10microbiome acts all this stuff is really
17:13much more powerful as a preventative
17:16tool and so how do you convince an
17:18entire industry that looks as spending a
17:21dollar as a dollar lost right that when
17:25you have these tools that you say can
17:27end over the long run save money but
17:29they're preventative it's not like we're
17:31not we don't want to spend money when
17:33somebody's already sick and that's the
17:35fundamental problem that any of these
17:36new technologies have when you're trying
17:38to find who the payers and so I think
17:40it's not clear to me that the existing
17:43payers will actually ever come around to
17:45it's a fair point I think there's a
17:48really good opportunity as you bring
17:50down the cost of these tests and you
17:51have consumers able to directly pay for
17:53them or sometimes actually go outside
17:55the US there's a lot of countries where
17:57it's either single pair or it's very
17:59much self-pay and consumers are using
18:01used to paying for testing for them by
18:03themselves like in India
18:04don't think about dental care in the
18:06United States yes that's actually a
18:07preventive I would argue that that's the
18:09best one of the health care systems in
18:11the world how does it work well twice a
18:14year you go get the same set of things
18:16basically measured about your body and
18:17we develop hundreds of millions of
18:19longitudinal you know medical dental
18:24records tied to outcomes so there's a
18:26positive feedback loop and the dental
18:28industry where actually if you look at
18:30the cost of dental care over time it's
18:32flat or down in inflation-adjusted
18:34dollars and the quality of the care has
18:36gone up if you look at over the exact
18:38same period of time in health care it's
18:39the exact opposite trend care in a lot
18:42of ways is getting worse costs are
18:43skyrocketing and so I think it's worth
18:46kind of asking what is the difference is
18:48why and in the dental care system you
18:50know there are dentists who give away
18:52checkups mm-hmm to get you as a customer
18:55because they know eventually you're
18:57gonna need a root canal or right and
18:59that's when you pay yep right and so
19:01they take the preventative approach
19:02saying you know it's never we're all
19:04gonna get sick we're all gonna die like
19:06and that's when we should get paid we're
19:08gonna try and keep you healthy until
19:10then yeah yeah it's very fundamentally
19:13different approach to but health care
19:15and the less I think it's important to
19:17remember the patients and the consumers
19:19role in this because it's really not
19:21going to work without the patients and
19:23consumers opting into these kinds of
19:25behavioral changes effectively we're
19:27asking people you know to look at your
19:29health in a different way because you
19:31know a Jeff brought up the dentist
19:32example right we brush our teeth every
19:33day or and hopefully every day twice
19:36twice a day and you know usually have
19:39annual checkups and things like that and
19:40and we don't really expect to go to the
19:43dentist after 20 years of not brushing
19:44our teeth and expecting them to you know
19:46make our teeth perfect right but that's
19:48exactly what we ask of our doctors today
19:50yeah it's like people treat their bodies
19:52terribly and then when they are you know
19:56on the verge of death they go to doctors
19:58perfect then the insurance companies
20:00complain because it's expensive to fix
20:01you wrap so so you know it really needs
20:04sort of patient and consumer buy-in but
20:06that's partly on them partly also on the
20:09technology companies to making it easier
20:11for them or enabling them or certainly
20:13motivating them in certain ways actually
20:15as you're talking about that in order
20:17for us to get to that well we do
20:18definitely need to have better
20:19diagnostic tests there's higher quality
20:22and and we're very long a I in Silicon
20:25Valley we think AI plus any industry it
20:27leads to a giant company in that
20:28industry generally yeah I plus cars AI
20:30plus radiology what is what is AI plus
20:33genomics look like so we at Google have
20:36sort of realized that in this explosion
20:39of you know genomics entering more and
20:43more the clinic what you see is that
20:45there is this really large growth in the
20:48amount of uncertainty around genetic
20:50tests and it's kind of ironic that when
20:51you look at you know even some of the
20:53most abundant why are just volume
20:55genetic tests these are cancer gene
20:58panel tests for basically it you know
21:00hereditary risk of developing cancer
21:03those tests will basically find 95
21:06mutations that they have absolutely no
21:08clue of what the effects of those
21:10mutations are in these important cancer
21:12associated genes per each mutation that
21:15is known to cause disease so when you
21:18can only interpret basically one out of
21:2096 you know mutations in this space it's
21:22kind of ironic that we call it precision
21:24medicine we see that as a fantastic
21:27place for machine learning and in fact
21:30the American College of medical genetics
21:32has guidelines for the use of
21:35computational tools in this space and so
21:38that's really you know one of the key
21:40places where we're focusing on entering
21:41because we know that today we can make
21:43those tools thirty-five percent better
21:45and that we're you know happy to
21:47distribute these tools as a horizontal
21:50across lots of different genetic test
21:52providers and not in continuous singles
21:55to add to that when you think about you
21:57know really what I mean AI is it just
22:00lets us combine a lot of different
22:02variables and make predictions but you
22:06know we live in a world where Google and
22:08Facebook use millions of variables to
22:10predict which add we're going to click
22:12or you know whether or not we can repay
22:14a loan but we try and reduce disease to
22:17single variables and and that's
22:19happening in genetics right now with a
22:20single nucleotide variant say if you're
22:23gonna get breast cancer that's that's
22:25it's honestly human bodies they
22:27extremely complicated system to build
22:30predictive models based on a single
22:31variable which is exactly what clinical
22:33studies do because that you know it's
22:35what they're designed to do really is is
22:38pretty asinine and and I think we have
22:40to think about you know you know why
22:43it's true in one place and not the other
22:45part of society and I think that
22:46applying machine learning is really just
22:48applying a tool that lets us look at far
22:51more information I mean can you imagine
22:53somebody in real time trying to figure
22:55out what you add you're gonna click on
22:56better than Google no it's just doing
22:58the things to look at right and that's
22:59what these tools allow us to do look at
23:01so many things way more powerful making
23:04more you could ever look at and I think
23:06that's really what it's about
23:08yeah and just just adding to that
23:10because I completely agree with that
23:11from the patient's perspective and the
23:13consumers product perspective what the
23:15learning engine can really provide is a
23:18way to make sense of all of that data in
23:20a way that pertains to them these kinds
23:23of things Apple watches fit bets and
23:25things like that phenomenal at gathering
23:27data right very bad at telling you what
23:30that actually does for you or what that
23:33you know means for you right on average
23:36an average human being makes you know
23:38twelve cancer cells every minute right
23:40sorry to freak out you know people in
23:42the audience but um you know you need to
23:45like really think about that and realize
23:48you know there is an optimum for you in
23:50terms of what kind of food you eat how
23:52you exercise you know how much you sleep
23:54things along those lines that really
23:56minimize those kinds of events right now
23:58patients have zero idea right how what
24:02they do in their lives actually
24:04correlates to any kind of changes in
24:07that space that is a huge big data
24:10problem that's just starting to get D
24:13convoluted and that's where the machine
24:15learning can really step in and figure
24:17out those kinds of interactions and
24:19correlations for us so that we can
24:21provide it to consumer you know I think
24:23there's a really important distinction
24:25which is there are Mendelian disorders
24:27where the phenotype is correlated with
24:29single genes and there's estimates are
24:32that there's at least 7,000 of them and
24:34we know like 3,500 them and then there's
24:36complex sort of maybe give examples of
24:40what each is yeah a standard sort of
24:42Mendelian disorder would be cystic
24:43fibrosis okay where you know you have
24:46basically issues in the lung and it
24:47really is driven by mutations in a
24:49chlorine channel in the case of sort of
24:52complex disorders we're talking about
24:54diabetes yeah diabetes you know things
24:56related to sort of metabolism where you
24:58have lots of genes contributing those
25:00are definitely controlled by lots of
25:02different variables and genetics plays a
25:04role in in both of those and I think
25:06it's really important that we remember
25:07you know which type of disorder we're
25:10talking about when we think about how
25:11we're applying AI and we think you know
25:13very carefully like okay what is you
25:15know the features and types of data that
25:18we're correlating with these and I yeah
25:20great well thank you so much for that
25:23session let's thank our guests