Machines & Genes: The Future of AI in Biology with Dr. George Church

00:49:49
https://www.youtube.com/watch?v=8CtR_wxotls

Resumo

TLDRThe video features a conversation with George Church, a leading geneticist who discusses his experience with the Personal Genome Project, ethical concerns in genomics, and the role of AI in advancing biology. Church shares insights from his journey in genetics, his views on the revival of woolly mammoths, and the importance of public access to genomic data. He also touches on the impact of machine learning in protein design and potential changes to healthcare costs. Ethical dilemmas surrounding AI applications in medicine and the future of genomic testing in personal relationships are explored, emphasizing a balance between technological advancement and societal responsibility.

Conclusões

  • 🔬 George Church is a pioneer in genetics and synthetic biology.
  • 📊 The Personal Genome Project emphasizes public access to genomic data.
  • 🤖 AI has transformative potential in protein design and gene therapy.
  • 🧬 Church advocates for understanding genetic risks in relationships.
  • 🌍 Reviving woolly mammoths could aid climate change efforts.
  • 🤝 Ethical concerns about privacy and data sharing are paramount.
  • 💡 Machine learning can significantly reduce healthcare costs.
  • 📈 Genomic testing is becoming more prevalent and accessible.
  • 🎯 Transparency in medical data is essential for ethical research.
  • 🌐 The intersection of technology and ethics remains critical for the future.

Linha do tempo

  • 00:00:00 - 00:05:00

    The speaker discusses their experience with Institutional Review Boards (IRBs) while working on an ethical and technological study, known as the Personal Genome Project. They express concerns about privacy and the inability to share vital health information with patients, advocating for a more open approach to medical records.

  • 00:05:00 - 00:10:00

    The introduction to George Church, a prominent scientist in genetics and synthetic biology. The hosts highlight his groundbreaking work in genomics and his ambitious projects, such as reviving the woolly mammoth and the Personal Genome Project.

  • 00:10:00 - 00:15:00

    George Church shares his early fascination with science and technology, recalling how he began building computers and programming in high school, which led him to pursue a career blending technology and biology.

  • 00:15:00 - 00:20:00

    Church discusses his motivations for wanting to revive the woolly mammoth, focusing on ecological benefits, such as carbon sequestration in the Arctic, rather than merely bringing back an extinct species.

  • 00:20:00 - 00:25:00

    He explains the technical challenges involved in engineering cold-resistant elephants, outlining the process of genome editing necessary to achieve this goal, referencing his previous successful work on pig genome editing for xenotransplantation.

  • 00:25:00 - 00:30:00

    The conversation shifts to Church's personal decision to make his own genomic data publicly available, which stemmed from a desire to explore the implications of genomics on health, privacy, and patient information access.

  • 00:30:00 - 00:35:00

    This public availability has had personal benefits for Church, prompting patient alerts about potential health issues and revealing risk factors that impacted his medical decisions.

  • 00:35:00 - 00:40:00

    An overview of Church's lab’s contributions to using artificial intelligence in biotech, emphasizing their work on protein engineering and enhancing CRISPR technologies by leveraging machine learning's capabilities.

  • 00:40:00 - 00:49:49

    Church reflects on the evolving landscape of artificial intelligence in biology, urging that the integration of machine learning is becoming crucial across all aspects of the field, while emphasizing the ongoing importance of human oversight and engagement.

Mostrar mais

Mapa mental

Vídeo de perguntas e respostas

  • What is the Personal Genome Project?

    The Personal Genome Project is an initiative founded by George Church that aims to make genomic data publicly accessible while emphasizing ethics and participant consent.

  • How does AI impact protein design?

    AI enables massive parallelization in protein design efforts, allowing researchers to achieve higher efficiency in mutagenesis and optimization compared to traditional methods.

  • What are the ethical concerns with genetic data privacy?

    Many researchers, including Church, express skepticism about the feasibility of keeping genetic data private and advocate for informed public sharing of such information.

  • What is George Church’s view on bringing back the woolly mammoth?

    Church argues that reviving the woolly mammoth could help restore Arctic ecosystems and combat climate change by enhancing carbon sequestration.

  • How does Church suggest approaching genomic testing in relationships?

    He advocates for considering genomic testing prior to dating to avoid potential genetic risks in offspring, promoting a more humane approach to genetic matching.

  • Will machine learning reduce healthcare costs?

    Yes, there is potential for machine learning to lower costs in diagnostics and drug development, as evidenced by recent achievements in gene therapies.

  • What role does AI play in the future of genomics?

    AI is expected to contribute significantly to genomics by enhancing data analysis, improving gene therapy results, and enabling personalized medicine.

  • What is one of Church’s most controversial opinions?

    He believes everyone should consider getting their genome sequenced, particularly those of reproductive age, to understand their genetic risks.

  • What companies is George Church involved with?

    Church is involved with several companies including Nabla, Dino Therapeutics, and Manifold Bio, all focused on using machine learning for biotech applications.

  • What keeps George Church up at night regarding AI?

    He concerns himself with the potential for discrimination through biased AI algorithms and the ethical implications of personalized weapons created using AI.

Ver mais resumos de vídeos

Obtenha acesso instantâneo a resumos gratuitos de vídeos do YouTube com tecnologia de IA!
Legendas
en
Rolagem automática:
  • 00:00:03
    and so we put in two pages on ethical
  • 00:00:05
    legal social implications and how we
  • 00:00:08
    were going to do the IRB and the process
  • 00:00:10
    this is the first time I've actually
  • 00:00:12
    done an IRB i' collaborated with people
  • 00:00:14
    that had done irbs before this is first
  • 00:00:15
    time I did it and I was just appalled at
  • 00:00:18
    what some of the things that they were
  • 00:00:19
    forcing you to do they they said like
  • 00:00:22
    you have to promise that you'll keep it
  • 00:00:24
    private and I said I don't know how to
  • 00:00:27
    keep my credit card private much less
  • 00:00:29
    somebody's medical records and in fact
  • 00:00:31
    there had been numerous uh hacking
  • 00:00:34
    events and people losing disc drives at
  • 00:00:37
    various hospitals in the Boston area
  • 00:00:38
    that will remain nameless and so I said
  • 00:00:42
    I think the better Assurance would be
  • 00:00:44
    that we get people who are okay with it
  • 00:00:45
    being public that was one thing that I
  • 00:00:48
    was horrified about the other one was
  • 00:00:49
    they said you couldn't return
  • 00:00:51
    information to the patients even if it
  • 00:00:55
    could save their life and I felt that
  • 00:00:57
    was crazy too uh and said said well but
  • 00:01:00
    a consequence of their data being public
  • 00:01:02
    is it also be available to them and so
  • 00:01:05
    they can you know if we give them the
  • 00:01:06
    right software they can draw their own
  • 00:01:08
    conclusions and with proper Consulting
  • 00:01:11
    with their Physicians and Specialists
  • 00:01:13
    they they will make the right decision
  • 00:01:15
    and so we framed it as a ethics study as
  • 00:01:18
    well as a technology study and that
  • 00:01:21
    became the personal Genome
  • 00:01:24
    [Music]
  • 00:01:25
    Project welcome to another episode of
  • 00:01:27
    any Jam AI Grand rounds I'm your host
  • 00:01:30
    Raj Mon and I'm here with my co-host
  • 00:01:33
    Andy beam today we're delighted to bring
  • 00:01:35
    you our conversation with George Church
  • 00:01:37
    of Harvard George is a professor of
  • 00:01:39
    genetics at Harvard Medical School and
  • 00:01:41
    professor of Health Sciences and
  • 00:01:42
    Technology at Harvard and MIT he has
  • 00:01:45
    many other titles and honors Andy I was
  • 00:01:47
    really struck by both the depth and the
  • 00:01:49
    breadth of cuttingedge science that
  • 00:01:51
    George oversees in his laboratory and
  • 00:01:53
    also by how effectively he must contact
  • 00:01:55
    switch he's very well known for his
  • 00:01:57
    pioneering work in genomics and
  • 00:01:58
    synthetic biology his involvement with
  • 00:02:00
    the Human Genome Project and the
  • 00:02:02
    personal Genome Project his work on Gene
  • 00:02:04
    editing and even Revival of the woolly
  • 00:02:06
    mammoth just to name a few areas he also
  • 00:02:09
    shared aspects about his career journey
  • 00:02:11
    and how he approaches both academic and
  • 00:02:13
    entrepreneurial efforts which were
  • 00:02:15
    enlightening you know Raj George has to
  • 00:02:17
    be one of the most fascinating
  • 00:02:18
    individuals I've ever met I think when
  • 00:02:20
    most of us decided that we wanted to be
  • 00:02:22
    scientists we actually probably wanted
  • 00:02:24
    to be someone like George he's always
  • 00:02:26
    working on the most intriguing things
  • 00:02:28
    and I'm constantly Amazed by the range
  • 00:02:30
    of truly sci-fi projects that he has
  • 00:02:31
    going on I really love that he sees his
  • 00:02:34
    mission as making the future happen
  • 00:02:35
    faster you know in that way he really
  • 00:02:37
    reminds me of a scientist from a movie
  • 00:02:38
    almost he's got this big beard this big
  • 00:02:40
    personality and he does crazy things
  • 00:02:42
    like try to bring back the woolly
  • 00:02:44
    mammoth I always enjoy hearing about
  • 00:02:46
    what George is up to and this
  • 00:02:47
    conversation was no
  • 00:02:50
    exception the njm AI Grand rounds
  • 00:02:52
    podcast is sponsored by Microsoft and
  • 00:02:55
    VII we thank them for their support and
  • 00:02:58
    with that we bring you our conversation
  • 00:03:00
    with George
  • 00:03:01
    Church welcome to AI Grand RS George
  • 00:03:04
    we're happy to have you here oh it's
  • 00:03:06
    great to be here so George a question we
  • 00:03:09
    ask all of our guests tell us about the
  • 00:03:11
    training procedure for your own neural
  • 00:03:13
    network how did you get interested in
  • 00:03:15
    science and in AI in particular what
  • 00:03:17
    data and experiences LED you to where
  • 00:03:19
    you are today I've heard there is a
  • 00:03:21
    framable Duke letter as part of this the
  • 00:03:25
    story yeah so I got interested in
  • 00:03:28
    science both both natural natur and
  • 00:03:30
    unnatural my third father was a
  • 00:03:32
    physician and so I was fascinated by the
  • 00:03:34
    technology in his he would do house
  • 00:03:36
    calls back then bit of uh history there
  • 00:03:39
    and then I went to the World's Fair and
  • 00:03:41
    saw computers and robots or animatronics
  • 00:03:45
    and I decided that was definitely cool
  • 00:03:49
    and I came back to Florida where they
  • 00:03:51
    didn't have any science classes to be
  • 00:03:53
    taken that I could find so I just
  • 00:03:55
    started building my own computers
  • 00:03:57
    because I was desperate not very good
  • 00:03:59
    one and so it wasn't until 9th grade
  • 00:04:01
    that I actually got access to a GE 635
  • 00:04:04
    at Dartmouth been programming ever since
  • 00:04:07
    basically that was actually a time
  • 00:04:09
    sharing interactive one this is in 1968
  • 00:04:12
    I'm in ninth grade then when I went to
  • 00:04:15
    the college at Duke was kind of a step
  • 00:04:17
    backwards forwards in terms of science I
  • 00:04:19
    was doing crystallography but backwards
  • 00:04:21
    in ter everything was Punch Cards into
  • 00:04:23
    IBM 360 and then I got two major degrees
  • 00:04:27
    in two years and then proceeded to flunk
  • 00:04:29
    out of graduate school they gave me a
  • 00:04:32
    nice letter hoping that would do well in
  • 00:04:34
    some other field but I stuck to this
  • 00:04:36
    field and my lab has been mostly mixture
  • 00:04:39
    of Technology development and
  • 00:04:41
    computational biology pretty much since
  • 00:04:43
    the beginning in 1986 at AR medical
  • 00:04:46
    school and naturally as the names kept
  • 00:04:49
    changing as to whether it's neural Nets
  • 00:04:51
    or deep learning or machine learning but
  • 00:04:53
    we've been part of that Revolution at
  • 00:04:56
    least as applied to things like Le
  • 00:04:59
    biology and protein design in particular
  • 00:05:02
    thanks that's super interesting and I
  • 00:05:03
    actually didn't know that you had such
  • 00:05:05
    an early interestes in computer science
  • 00:05:06
    and I can see that writing code whether
  • 00:05:08
    it's in bits or in DNA has sort of been
  • 00:05:11
    with you since the beginning that you're
  • 00:05:12
    still programming the software the
  • 00:05:14
    substrate has just changed a little bit
  • 00:05:16
    over time yeah yeah my first computers
  • 00:05:18
    were a analog computer and then a a
  • 00:05:23
    digital mechanical computer so it was
  • 00:05:25
    only the third one that was was a real
  • 00:05:28
    Von noyman digital
  • 00:05:30
    interactive computer we're going to hop
  • 00:05:32
    into your research in just a little bit
  • 00:05:34
    but there's an icebreaker question that
  • 00:05:35
    I think I'm obligated by law to ask you
  • 00:05:38
    and it's about bringing back the woolly
  • 00:05:39
    mammoth and so you've been on record on
  • 00:05:42
    the record is saying that we should
  • 00:05:43
    bring back the woolly mammoth and I'd
  • 00:05:45
    love to hear the motivations for doing
  • 00:05:47
    that and sort of how far are we away
  • 00:05:49
    from being able to do that
  • 00:05:51
    now well this is not going to fit into
  • 00:05:54
    140 or 280 characters but I'll try my
  • 00:05:57
    best so it's really not about bringing
  • 00:06:00
    back an extinct species we're mainly
  • 00:06:02
    interested in endangered species and
  • 00:06:05
    restoring ecosystems and the ecosystem
  • 00:06:08
    that seems like it needs the most help
  • 00:06:10
    is the Arctic where there's a lot of
  • 00:06:12
    carbon sequestered because of the cycles
  • 00:06:14
    of freezing and then summer
  • 00:06:17
    photosynthesis results in sometimes 500
  • 00:06:20
    meters thick a top soil compared to say
  • 00:06:22
    one meter in the lot of the
  • 00:06:25
    rainforests and that's all melting and
  • 00:06:28
    it's melting a lot of it in the form of
  • 00:06:30
    methane which is 80 times worse than
  • 00:06:32
    carbon dioxide and so we're hoping that
  • 00:06:36
    making cold resistant elephants which
  • 00:06:39
    would help save them from their
  • 00:06:40
    endangered species categorization and
  • 00:06:44
    would also help restore that when we're
  • 00:06:46
    not saying this is the only answer or
  • 00:06:48
    even this is a good answer it's just our
  • 00:06:51
    fot at increasing sequestration and
  • 00:06:54
    decreasing loss in the form of methane
  • 00:06:57
    and we're doing that by genome
  • 00:06:59
    engineering which we've already done in
  • 00:07:02
    pigs for making transplants
  • 00:07:05
    xenotransplants which are making their
  • 00:07:07
    way through now 600 successful days in
  • 00:07:11
    in non-human primates and some and now
  • 00:07:14
    going into human clinical
  • 00:07:16
    trials we can do dozens 42 edits in pigs
  • 00:07:21
    genome dermaline we hope to be able to
  • 00:07:24
    do the same thing in elephants and make
  • 00:07:26
    them cold resistant maybe also resist to
  • 00:07:29
    herpes viruses that are killing them if
  • 00:07:33
    you have to do 42 edits for pigs what is
  • 00:07:36
    the difference in scale to do something
  • 00:07:38
    to like turn the elephant into a cold
  • 00:07:40
    resistant woly
  • 00:07:41
    Mammoth a cold resistant elephant that
  • 00:07:44
    has genes that are that are resurrected
  • 00:07:46
    from woly math we already resurrected
  • 00:07:48
    two genes that seem to have the right
  • 00:07:51
    properties when resurrected we don't
  • 00:07:54
    know the exact number we're doing a
  • 00:07:57
    computational analysis of dozens of of
  • 00:07:59
    elephant and extinct
  • 00:08:02
    relatives and focusing on genes that go
  • 00:08:05
    to fixation that means that they're
  • 00:08:08
    homozygous both their maternal and
  • 00:08:10
    paternal lils are the same and both of
  • 00:08:12
    those are different from the existing
  • 00:08:16
    elephants so basically you've got this
  • 00:08:19
    branching in the philogenetic tree where
  • 00:08:23
    it's gone to fixation where all of the
  • 00:08:26
    mammoth genes are different from the
  • 00:08:29
    existing ones and we can do that for
  • 00:08:30
    multiple mammoths and multiple existing
  • 00:08:33
    genomes both African and Asian elephants
  • 00:08:37
    and we're focusing on those both the
  • 00:08:38
    coding regions and the non-coding
  • 00:08:41
    regions and if we did all of them it
  • 00:08:44
    would be in the hundreds of thousands if
  • 00:08:48
    we just do the ones that we think are
  • 00:08:49
    involved in cold a priori then those
  • 00:08:52
    could be in the dozens so it's somewhere
  • 00:08:54
    in between and we have Technologies
  • 00:08:57
    editing Technologies I think that are up
  • 00:08:59
    up to the high end of the spectrum our
  • 00:09:01
    record so far for repetitive elements is
  • 00:09:03
    editing
  • 00:09:05
    24,000 edits simultaneously in one cell
  • 00:09:09
    that happened to be a human plur poent
  • 00:09:11
    stem cell but it could have been any
  • 00:09:12
    amalian genome as far as we know got it
  • 00:09:17
    thanks so George I want to switch gears
  • 00:09:20
    to your work in genomics you're a
  • 00:09:22
    Pioneer in genome sequencing genetics
  • 00:09:25
    synthetic biology many other areas and
  • 00:09:28
    your work has been influential on many
  • 00:09:30
    scientists myself included uh but I
  • 00:09:32
    actually want to ask you about a
  • 00:09:34
    different perspective which is your
  • 00:09:35
    perspective as a patient you famously
  • 00:09:38
    made your own genome sequence your lab
  • 00:09:41
    values and medical records publicly
  • 00:09:43
    available for anyone to download on the
  • 00:09:45
    internet I want to ask how did you come
  • 00:09:47
    to that decision and what uh you might
  • 00:09:50
    have learned since you've made that
  • 00:09:52
    information
  • 00:09:53
    available yeah that's an interesting
  • 00:09:55
    question so we uh applied for ni Center
  • 00:09:59
    of excellence in genomic science and got
  • 00:10:02
    it and in the process we were proposing
  • 00:10:05
    to develop a new way of doing DNA
  • 00:10:07
    sequencing now this is in 2003 and we
  • 00:10:10
    were partly there anyway so that's why
  • 00:10:12
    we proposed it it's now called nextg
  • 00:10:15
    sequencing but back then it had other
  • 00:10:17
    names and we proposed that in the
  • 00:10:21
    process of the five-year Center of
  • 00:10:23
    Excellence Grant we would get as far as
  • 00:10:26
    doing a 1.7 million based pair bacterial
  • 00:10:30
    genomes this is a very tiny bacterial
  • 00:10:32
    genome not the smallest but very tiny
  • 00:10:34
    with helicobactor pylori the positive
  • 00:10:36
    agent of stomach cancers and
  • 00:10:39
    ulcers and we had already contributed to
  • 00:10:41
    the first time that was sequenced back
  • 00:10:43
    in 199 4ish so in 2003 we proposed to do
  • 00:10:48
    that again after 5 years of work as it
  • 00:10:52
    turned out 8 months into it we had
  • 00:10:54
    finished something three times bigger
  • 00:10:56
    than that and by the end of the project
  • 00:10:59
    we had finished five human genomes at
  • 00:11:01
    six billion base pairs each so diploid
  • 00:11:03
    high quality genomes now that was not
  • 00:11:06
    completely unanticipated we thought we
  • 00:11:08
    might get the human genomes even though
  • 00:11:10
    we were only promising to do a really
  • 00:11:12
    tiny bacterial gen thousand times
  • 00:11:13
    smaller we thought we might get there
  • 00:11:15
    and so we put in two pages on ethical
  • 00:11:18
    legal social implications and how we
  • 00:11:20
    were going to do the IRB and in process
  • 00:11:23
    this is the first time I've actually
  • 00:11:24
    done an IRB i' collaborated with people
  • 00:11:26
    that have done irbs before this is the
  • 00:11:28
    first time I did it and I was just
  • 00:11:29
    appalled at some of the things that they
  • 00:11:31
    were forcing you to do they they said
  • 00:11:34
    like you have to promise that you'll
  • 00:11:35
    keep it private and I said I don't know
  • 00:11:38
    how to keep my credit card private much
  • 00:11:41
    less somebody's medical records and in
  • 00:11:43
    fact there had been numerous uh hacking
  • 00:11:46
    events and people losing disc drives at
  • 00:11:49
    various hospitals in the Boston area
  • 00:11:51
    that will remain nameless and so I said
  • 00:11:54
    I think the better Assurance would be
  • 00:11:56
    that we get people who are okay with it
  • 00:11:58
    being public that was one thing that was
  • 00:12:00
    horrified by the other one was they said
  • 00:12:02
    you couldn't return information to the
  • 00:12:05
    patients even if it could save their
  • 00:12:08
    life and I felt that was crazy too uh
  • 00:12:11
    and I said well but a consequence of
  • 00:12:13
    their data being public is it also be
  • 00:12:15
    available to them and so they can you
  • 00:12:18
    know if we give them the right software
  • 00:12:19
    they can draw their own conclusions and
  • 00:12:21
    with proper Consulting with their
  • 00:12:24
    Physicians and Specialists they they
  • 00:12:26
    will make the right decision and so we
  • 00:12:28
    framed it a ethics study as well as a
  • 00:12:31
    technology study and that became the
  • 00:12:34
    personal Genome
  • 00:12:35
    Project that's fascinating that history
  • 00:12:38
    I'm curious so you also as part of this
  • 00:12:41
    you made your own genome public and I
  • 00:12:44
    think to this day you can go and
  • 00:12:46
    download a copy of your genome as well
  • 00:12:48
    as your medical records and just
  • 00:12:50
    connecting the the dots you were one of
  • 00:12:52
    the participant who was willing to have
  • 00:12:54
    their genome released right you are a
  • 00:12:56
    participant in your own that's correct
  • 00:12:58
    the
  • 00:12:59
    IRB felt that I mean it was an unusual
  • 00:13:03
    proposal to the IRB because I was
  • 00:13:05
    breaking at least two of their rules a
  • 00:13:06
    few others that we don't need to go into
  • 00:13:08
    but but you know transparently and
  • 00:13:10
    politely and it took us about a year to
  • 00:13:13
    negotiate it but that's not that long
  • 00:13:15
    compared to other IRB approvals uh even
  • 00:13:17
    non-controversial ones but part of that
  • 00:13:20
    is they said well you know you should be
  • 00:13:21
    willing to eat your own dog food uh you
  • 00:13:23
    know you we want you to be the sole
  • 00:13:26
    participant and I said well how about 10
  • 00:13:29
    participants let's compromise and I'll
  • 00:13:30
    be the first of the 10 and then we'll
  • 00:13:33
    expand if nobody gets hurt we'll expand
  • 00:13:35
    it from there and John hamka was
  • 00:13:37
    actually number two Esther Dyson was
  • 00:13:39
    number three all their names were known
  • 00:13:42
    that was approved by the IRB as well and
  • 00:13:44
    John was at the time just brand new CIO
  • 00:13:47
    at Best Israel deaconness and also had
  • 00:13:50
    crafted the I think it was Norway's
  • 00:13:52
    government's medical informatics
  • 00:13:54
    policies and so he was really perfect
  • 00:13:57
    second if in fact there's a funny story
  • 00:13:59
    about my posting so I was a patient at
  • 00:14:02
    Beth Israel Deacon this as it turned out
  • 00:14:04
    and before we had the IRB approval I had
  • 00:14:06
    just posted my own medical records as a
  • 00:14:08
    test and some patient had been looking I
  • 00:14:12
    get browsing the internet and had run
  • 00:14:14
    across my medical records and she
  • 00:14:16
    freaked out she thought this is standard
  • 00:14:18
    policy at pical deaconness was to
  • 00:14:20
    release patient medical records on the
  • 00:14:22
    internet and the her complaint worked it
  • 00:14:25
    all the way up to the president and then
  • 00:14:27
    back down to John halamka who said oh I
  • 00:14:30
    know what this is this is George church
  • 00:14:32
    and he contacted me and I said oh yeah
  • 00:14:34
    yeah oh I didn't even think that some
  • 00:14:36
    patient would find this I I put a big
  • 00:14:38
    disclaimer on it I said this is not the
  • 00:14:41
    Beth Israel policy this is just an
  • 00:14:42
    experiment and then they thought that
  • 00:14:44
    was okay and we went
  • 00:14:46
    forward and so have you learned anything
  • 00:14:49
    about your own genome since it's been uh
  • 00:14:52
    publicly available or your own medical
  • 00:14:53
    records since you've made these widely
  • 00:14:56
    available uh oh yeah sure there been
  • 00:14:58
    various advantages there was one case
  • 00:15:02
    where I was in giving a lecture in
  • 00:15:05
    Seattle and somebody in the front row
  • 00:15:09
    said while we were waiting for everybody
  • 00:15:11
    to settle in you know there was kind of
  • 00:15:13
    a
  • 00:15:14
    pause and he said you know you should
  • 00:15:16
    get your blood work checked out because
  • 00:15:19
    according to your public medical records
  • 00:15:20
    you're on Statin and there's no evidence
  • 00:15:23
    that you've checked on your cholesterol
  • 00:15:25
    or on the possible negative consequences
  • 00:15:28
    for muscle damage and I said you know
  • 00:15:31
    you're right I haven't been checked and
  • 00:15:33
    it has been a long time and so I went
  • 00:15:35
    off and I checked and actually there was
  • 00:15:38
    no lowering of cholesterol and there was
  • 00:15:39
    muscle damage so or potentially the
  • 00:15:42
    biomarkers for that and so we changed
  • 00:15:45
    the Statin and the dose and and
  • 00:15:47
    monitored
  • 00:15:48
    the that probably you know gave me extra
  • 00:15:51
    10 years of life for not using drugs
  • 00:15:54
    right and but you know I learned that I
  • 00:15:57
    had risk factors for Peno 1 for Alpha 1
  • 00:16:00
    antirion which put me at risk for a
  • 00:16:02
    whole variety of respiratory problems
  • 00:16:05
    and so I've been cautious about living
  • 00:16:08
    in downtown Beijing and Los Angeles and
  • 00:16:11
    things like that or or hanging around
  • 00:16:14
    with Co too
  • 00:16:15
    much great so we're gonna switch gears
  • 00:16:18
    just a little bit again and talk about
  • 00:16:21
    some of the work that your lab has been
  • 00:16:22
    doing in artificial intelligence and
  • 00:16:24
    deep learning for biotech applications
  • 00:16:26
    so I've you know I tried to pick one
  • 00:16:29
    paper to focus in on but you've done so
  • 00:16:31
    much work on things like protein
  • 00:16:33
    engineering designing aavs for gene
  • 00:16:35
    therapy improving the efficiency of
  • 00:16:37
    crisper so I was hoping you could pick
  • 00:16:39
    one from the list of the many papers
  • 00:16:42
    that you've written in this area and
  • 00:16:44
    help us understand what machine learning
  • 00:16:47
    is bringing to the table so I think
  • 00:16:49
    there's often a lot of confusion about
  • 00:16:52
    when and how and why to use machine
  • 00:16:54
    learning but I was hoping you could
  • 00:16:55
    really help us zero in on the types of
  • 00:16:58
    new question questions that machine
  • 00:16:59
    learning has let your lab ask and
  • 00:17:02
    answer yeah so the one that I would pick
  • 00:17:05
    or the two the first one was the one
  • 00:17:08
    that I would pick is is a nature biotech
  • 00:17:10
    paper we published in year 2000 which is
  • 00:17:13
    Bryant at all Eric kelik uh was uh
  • 00:17:17
    postto on my lab who was a senior author
  • 00:17:18
    on that paper and it was titled
  • 00:17:20
    massively parallel deep diversification
  • 00:17:23
    of aav CD proteins by Machine
  • 00:17:26
    learning and we have been working on
  • 00:17:28
    machine learning for protein design for
  • 00:17:31
    quite a while at that point but that was
  • 00:17:34
    different in it it Illustrated your
  • 00:17:36
    quest your point was when is it time to
  • 00:17:38
    do machine learning when is it not
  • 00:17:40
    typically there's one prerequisite for
  • 00:17:43
    machine learning you have a lot of data
  • 00:17:45
    we had just published some paper that we
  • 00:17:48
    refer to as low-end machine learning but
  • 00:17:51
    it isn't entirely lowend and there's
  • 00:17:53
    some additional background information
  • 00:17:55
    about proteins in general but lowend for
  • 00:17:57
    your particular experiment but but
  • 00:18:00
    anyway that's prerequisite you need to
  • 00:18:01
    have a lot of data typically then
  • 00:18:03
    there's the question of when is it
  • 00:18:04
    better so in this particular case I
  • 00:18:06
    think we showed that it's a lot better
  • 00:18:09
    because we did a comparison of naive or
  • 00:18:13
    random semi- random uh models for
  • 00:18:16
    mutagenizing a stretch uh key stretch of
  • 00:18:20
    28 amino acids that's important to the
  • 00:18:23
    aav capsid for gene therapy delivery and
  • 00:18:27
    we asked you know how many amino acids
  • 00:18:29
    can be changed simultaneously and the
  • 00:18:30
    reason you might want to change a lot of
  • 00:18:32
    them simultaneously is that's where the
  • 00:18:35
    immune system interacts with it and if
  • 00:18:37
    you want to use the gene therapy more
  • 00:18:38
    than once and you want it to be or even
  • 00:18:41
    once you want to make sure the immune
  • 00:18:43
    system doesn't attack your precious
  • 00:18:45
    therapeutic also if you want to Target a
  • 00:18:47
    new tissue you might have to radically
  • 00:18:49
    change the surface of the virus so it
  • 00:18:51
    will Target tissue a and not B so those
  • 00:18:54
    are our
  • 00:18:55
    objectives and we in the naive Eve model
  • 00:18:59
    it was very hard to get more than four
  • 00:19:01
    changes out of 28 and this was very
  • 00:19:04
    consistent with all the work we've done
  • 00:19:05
    in many other proteins before four out
  • 00:19:08
    of 20 was actually a pretty good day in
  • 00:19:10
    one round of diversification and
  • 00:19:13
    selection but using logistic regression
  • 00:19:16
    or various concurrent neural Nets or
  • 00:19:19
    various neural Nets uh we could get up
  • 00:19:22
    to we could get about 90% at 25 out of
  • 00:19:26
    28 about just doing this from memory 70
  • 00:19:29
    80% at 26 out and even 20% of this big
  • 00:19:34
    Library focused on 28 out of 28 so it
  • 00:19:37
    wasn't like one solution it was lots of
  • 00:19:38
    solutions while at four out of 28 with
  • 00:19:42
    the naive model you were getting close
  • 00:19:43
    to zero per. so that to me is a dramatic
  • 00:19:47
    validation that that this is an
  • 00:19:49
    improvement over the the naive models
  • 00:19:52
    yeah so so this all started with uh an
  • 00:19:54
    undergraduate in my lab Harvard
  • 00:19:55
    undergrad Ethan alley and surge biswash
  • 00:19:59
    and others Pierce Ogden who had
  • 00:20:01
    developed unir which was a language
  • 00:20:04
    model so there kind of two broad
  • 00:20:07
    categories in protein structure and
  • 00:20:10
    design one is focusing on the structure
  • 00:20:12
    and that got a lot of attention with
  • 00:20:14
    Alpha fold from Deep Mind Google but it
  • 00:20:17
    just predicts the three-dimensional
  • 00:20:19
    structure of the protein just I should
  • 00:20:21
    say that's a 50-year Holy Grail but from
  • 00:20:23
    a protein design standpoint knowing the
  • 00:20:25
    threedimensional structure is just
  • 00:20:28
    starting point and not necessarily the
  • 00:20:30
    right direction so for example I'm just
  • 00:20:32
    going to end on this example that the
  • 00:20:34
    the just contrasting the language model
  • 00:20:36
    which is essentially linear with the 3D
  • 00:20:39
    structure is if you if you have a
  • 00:20:41
    searing protease which is named because
  • 00:20:44
    sering at the Active side is critical
  • 00:20:46
    and you change that sering to an alanine
  • 00:20:48
    so now it's completely dead but it has
  • 00:20:50
    exactly the same threedimensional
  • 00:20:52
    structure now there's the version of
  • 00:20:55
    this Alpha fold which is Alpha fold
  • 00:20:57
    multimer where you can ask whether two
  • 00:20:59
    proteins can stick together that might
  • 00:21:02
    save you in certain circumstances but
  • 00:21:04
    basically point is you can kill a
  • 00:21:05
    protein without changing its
  • 00:21:06
    threedimensional structure but not in
  • 00:21:08
    the language model it's two it it
  • 00:21:10
    catches us stuff right right so that I
  • 00:21:14
    think dovet Tales nicely into a
  • 00:21:15
    follow-up question that I'd like to ask
  • 00:21:17
    is you've seen many Revolutions in
  • 00:21:20
    biology I think it's fair to say um
  • 00:21:22
    genomic sequencing synthetic biology
  • 00:21:25
    could you help us understand um how AI
  • 00:21:28
    is going to make an impact in biology
  • 00:21:30
    over the next 10 years and the sort of
  • 00:21:32
    flip side of that is are there areas of
  • 00:21:35
    biology are there corners of biology
  • 00:21:37
    that you think are immune from sort of
  • 00:21:39
    the AI um incursion that we're seeing
  • 00:21:44
    today yeah so uh I kind of have a policy
  • 00:21:48
    not to say anything's impossible because
  • 00:21:51
    it's one way you can show how foolish
  • 00:21:53
    you are I mean I'm not questioning
  • 00:21:55
    whether I'm foolish it's hard to prove a
  • 00:21:56
    negative yeah and and it's easy to get
  • 00:21:59
    embarrassed two years later when
  • 00:22:01
    somebody shows that it does work so and
  • 00:22:03
    in fact almost everything that I've
  • 00:22:05
    worked on every project I've worked on
  • 00:22:07
    in my 68 years has been 50 years as a
  • 00:22:10
    scientist one person or another has said
  • 00:22:13
    it's impossible sometimes a couple years
  • 00:22:15
    after I published peer- reviewed paper
  • 00:22:17
    on it they still say it's impossible but
  • 00:22:20
    anyway so I I can't think of anything
  • 00:22:22
    where you know the laws of physics or
  • 00:22:24
    Computing would prevent it from making a
  • 00:22:27
    contribution ify may I qualify an area
  • 00:22:30
    of biology that you think has low
  • 00:22:32
    probability of being impacted or or is
  • 00:22:35
    just too hard for AI over the next 10
  • 00:22:38
    years so I've relaxed the condition a
  • 00:22:39
    little bit yeah I still think it's like
  • 00:22:44
    I think they're all sufficiently high
  • 00:22:46
    probability nonzero uh that they that
  • 00:22:50
    they're worth considering I think things
  • 00:22:52
    where it's really hard to get data but
  • 00:22:55
    on the other hand it's not clear any
  • 00:22:57
    kind of intell elligence artificial or
  • 00:22:59
    otherwise is going to solve some of
  • 00:23:01
    those problems so yeah I think there's a
  • 00:23:04
    a good room for it it doesn't mean that
  • 00:23:07
    human beings aren't going to play a
  • 00:23:09
    gigantic role in nudging it programming
  • 00:23:13
    it interpreting it interfacing with
  • 00:23:16
    people who don't need to know all the
  • 00:23:20
    the Gory uh computational details so
  • 00:23:24
    yeah I I think it's going to affect
  • 00:23:25
    everything really and I and I think the
  • 00:23:27
    same thing's true for synthetic biology
  • 00:23:28
    synthetic biology is going to affect
  • 00:23:31
    everything including making computers
  • 00:23:34
    that might be better than current
  • 00:23:36
    computers they'll probably be hybrid
  • 00:23:38
    computers of various sorts uh in in a
  • 00:23:40
    certain sense machine learning is based
  • 00:23:43
    on is inspired at least by um natural
  • 00:23:47
    neuronal computers and I think the
  • 00:23:50
    hybrids will have a good shot at it if
  • 00:23:53
    it was anyone else I would say that was
  • 00:23:54
    a hedge but when George Church says
  • 00:23:56
    Never Say Never I believe it
  • 00:24:00
    100% thank
  • 00:24:02
    you so George I want to ask you about a
  • 00:24:05
    field where you think it won't
  • 00:24:07
    work no I I I've just heard illuminary
  • 00:24:10
    of the field uh advise me against uh ask
  • 00:24:14
    you first
  • 00:24:15
    yeah anyway so George I want to ask you
  • 00:24:19
    about commercialization so you've been
  • 00:24:21
    involved in many companies I think a few
  • 00:24:23
    that are relevant to protein engineering
  • 00:24:26
    and AI including manifold bio nabla Dino
  • 00:24:30
    Therapeutics could you tell us about
  • 00:24:32
    what these companies do and maybe how
  • 00:24:34
    their missions relate to one
  • 00:24:36
    another right so the three that you
  • 00:24:39
    mentioned nabla Dino manifold and two
  • 00:24:43
    more patch and shape Therapeutics all
  • 00:24:45
    are using machine learning for protein
  • 00:24:48
    or or nucleic acid design they're wildly
  • 00:24:52
    different so nabl is focusing on
  • 00:24:54
    antibodies which are one of the key
  • 00:24:56
    therapeutic categor and have diagnostic
  • 00:24:59
    uses as well Dino is on delivery we were
  • 00:25:02
    just talking about aav manifold is very
  • 00:25:06
    interesting and they're developing ways
  • 00:25:09
    to make protein binding pairs but that's
  • 00:25:12
    not the endgame then those can be used
  • 00:25:15
    for testing multiple protein
  • 00:25:17
    Therapeutics simultaneously in one let's
  • 00:25:20
    say expensive test animal for
  • 00:25:24
    preclinical trials let's say a non-human
  • 00:25:26
    primate so you can do thousands of
  • 00:25:29
    simultaneous measurements of pharmac
  • 00:25:32
    kinetics and Dynamics and tissue
  • 00:25:34
    targeting and so forth all at once with
  • 00:25:36
    protein Therapeutics which don't
  • 00:25:38
    normally have a nucleic acid barcode
  • 00:25:41
    that you can follow so this is a protein
  • 00:25:43
    barcode and then shape is working on RNA
  • 00:25:47
    therapies tras and uh ADR and also
  • 00:25:51
    delivery and then patches on CIS
  • 00:25:53
    regulatory elements DNA and RNA and
  • 00:25:56
    there could probably be about 10 more of
  • 00:25:57
    them that would not overlap one another
  • 00:26:00
    it's it's a very important subset of
  • 00:26:03
    things that you can do with machine
  • 00:26:05
    learning so you know as you think about
  • 00:26:07
    those companies that are making use of
  • 00:26:09
    machine learning and artificial
  • 00:26:11
    intelligence and the others that you've
  • 00:26:13
    been involved in I'm curious have you
  • 00:26:16
    identified maybe a set of let's say one
  • 00:26:19
    or two key questions that allow you to
  • 00:26:22
    decide whether an idea is right as a
  • 00:26:24
    commercial entity or whether it maybe
  • 00:26:26
    belongs in
  • 00:26:28
    in Academia for more development before
  • 00:26:30
    it it moves into into a commercial
  • 00:26:33
    entity yeah this is a really tough call
  • 00:26:36
    that that every postto that wants to
  • 00:26:40
    start a company and the pi that maybe
  • 00:26:44
    wants to start it with them or wants to
  • 00:26:46
    jump from Academia into industry and I
  • 00:26:49
    say it's not so bad to be number two to
  • 00:26:52
    be the second one in the field even if
  • 00:26:54
    you thought of it first and it's kind of
  • 00:26:56
    irritates You That Somebody body jumped
  • 00:26:58
    in there with your idea you just want to
  • 00:27:00
    make sure you got enough intellectual
  • 00:27:02
    property that you have freedom to
  • 00:27:03
    operate and it doesn't really matter who
  • 00:27:05
    gets funded first
  • 00:27:07
    necessarily and very often the second
  • 00:27:10
    one does a better job of it either from
  • 00:27:13
    a business side or science engineering
  • 00:27:16
    side so that's part of the
  • 00:27:18
    decision the part where you know that
  • 00:27:20
    it's ready is either you're getting a
  • 00:27:22
    lot of feedback from your peers they
  • 00:27:24
    like it they want it one case you know
  • 00:27:27
    when we were developing DNA synthesis
  • 00:27:31
    the very high throughput DNA synthesis
  • 00:27:32
    like 10,000 times previous throughput
  • 00:27:36
    and suddenly we had a lot of friends
  • 00:27:38
    that wanted to Buddy and and collaborate
  • 00:27:41
    on making big DNA constructs cheaply and
  • 00:27:44
    we said oh this is going to be a
  • 00:27:45
    tremendous academic distraction we
  • 00:27:48
    really have to spin this out just for
  • 00:27:49
    our own sanity so we did that same thing
  • 00:27:52
    with with chrisper that was clearly
  • 00:27:54
    going to be so popular I mean we we
  • 00:27:56
    announced it in January 2013 and by
  • 00:27:59
    March there were like 10,000 users who
  • 00:28:02
    distributed it through adene which is
  • 00:28:04
    nearly free and that would have been a
  • 00:28:07
    real hassle to have done that licking
  • 00:28:10
    stamps in our
  • 00:28:11
    office so those kind of things that tell
  • 00:28:14
    you that you're ready don't have to rush
  • 00:28:17
    you can the the more mature it is the
  • 00:28:20
    the slower you will get deluded out
  • 00:28:22
    where if you go there too early you may
  • 00:28:25
    think it's magic money it's easy than a
  • 00:28:27
    grant but pretty soon it'll be taken
  • 00:28:29
    away from you and and it may not go in
  • 00:28:31
    the direction you want it to go once you
  • 00:28:33
    lose control over it yeah I I like the
  • 00:28:36
    emphasis on it's not so bad to be second
  • 00:28:38
    uh there's a saying that I'm reminded of
  • 00:28:40
    that the early bird gets the worm but
  • 00:28:42
    the second mouse gets the cheese and
  • 00:28:44
    that sometimes being second is
  • 00:28:45
    strategically good all right that's good
  • 00:28:48
    second mouse is a cheese yeah yeah and I
  • 00:28:50
    think there a little bit the second worm
  • 00:28:52
    gets to live yeah right yeah yeah right
  • 00:28:56
    um so so George just as a follow on to
  • 00:28:58
    that having overseen many companies from
  • 00:29:02
    genomics and from sequencing genomic
  • 00:29:04
    Technologies and now increasingly with
  • 00:29:07
    AI is there a difference you see in the
  • 00:29:10
    decision- making of when to start a new
  • 00:29:12
    company when genomics in the context of
  • 00:29:14
    your involvement in companies there
  • 00:29:16
    versus now with the more AI focused
  • 00:29:18
    companies or the the principles more or
  • 00:29:20
    less the
  • 00:29:21
    same well there's some very significant
  • 00:29:23
    differences well first of all when I
  • 00:29:25
    started back in the'80s
  • 00:29:28
    I was basically just the way I was
  • 00:29:30
    deciding was if a investor or a
  • 00:29:33
    colleague came forward and said hey we'd
  • 00:29:35
    like to collaborate with you on a
  • 00:29:36
    company and I'd say okay it was fairly
  • 00:29:39
    reactive but recently it's been mostly
  • 00:29:42
    postdoc driven and during that time we
  • 00:29:44
    transitioned from Mostly analytic to
  • 00:29:47
    mostly synthetic and the problem with a
  • 00:29:49
    lot of the DNA sequencing scenarios was
  • 00:29:53
    you had to convince people to get
  • 00:29:56
    sequenced well with the synthetic
  • 00:29:58
    biology which is basic you know in this
  • 00:30:01
    case is basically pharmacology people
  • 00:30:03
    are already consuming drugs and you're
  • 00:30:05
    just making better and better drugs for
  • 00:30:07
    diseases for which there were no drugs
  • 00:30:09
    and it's just a it's an easier pipeline
  • 00:30:11
    while I would say that our first
  • 00:30:13
    sequencing Innovation was in 1984 and
  • 00:30:16
    here we are in almost
  • 00:30:18
    2024 and we still don't have you know 40
  • 00:30:21
    years later we still don't really have
  • 00:30:23
    consensus enough that healthc care
  • 00:30:26
    providers think that it's a good thing
  • 00:30:28
    to give everybody their whole genome
  • 00:30:30
    sequence or act on it in any way and
  • 00:30:33
    most people don't do it on their own so
  • 00:30:35
    there's a disconnect I think it's the 1%
  • 00:30:39
    dilemma it's the seat belts smoking
  • 00:30:43
    global warming and getting your genome
  • 00:30:45
    sequenced they're all like hey I got a
  • 00:30:48
    99% chance of doing okay if I went to
  • 00:30:51
    Las Vegas with those odds I'd be fine
  • 00:30:54
    but it's different for public health and
  • 00:30:56
    it takes special effort to get people to
  • 00:30:58
    stop smoking and to wear seat belts and
  • 00:31:00
    so forth so I think that's what's going
  • 00:31:03
    on here and and no no government agency
  • 00:31:05
    has stepped forward to do what they did
  • 00:31:08
    for seat belts and smoking which was a
  • 00:31:10
    whole whole series of experiments like
  • 00:31:12
    just passing a law to buckle your seat
  • 00:31:14
    belt wasn't enough getting them present
  • 00:31:16
    in every car was not enough they had to
  • 00:31:18
    actually mandate a circuit that would
  • 00:31:21
    close once you buckled it on top of your
  • 00:31:24
    belly Not underneath your you know that
  • 00:31:27
    hasn't happened yet for for even though
  • 00:31:30
    the the carrier status alone plus adult
  • 00:31:33
    onset diseases could save us a trillion
  • 00:31:35
    dollars a year and a lot of pain and
  • 00:31:38
    suffering it just hasn't happened yet
  • 00:31:40
    but with synthetic biology totally
  • 00:31:42
    different thing there's the orphan Drug
  • 00:31:44
    Act that makes it very profitable to go
  • 00:31:47
    after rare things and then there's lots
  • 00:31:49
    of common diseases as well that can be
  • 00:31:51
    treated and my favorite treatment is
  • 00:31:53
    gene therapy for reasons that we could
  • 00:31:55
    go into if you if you want to R shall we
  • 00:31:58
    go to the lightning
  • 00:31:59
    round sure so George we'd like to do a
  • 00:32:02
    quick lightning round if that's all
  • 00:32:04
    right with
  • 00:32:05
    you we're going to ask a series of
  • 00:32:07
    questions and the rules of the game are
  • 00:32:09
    one to two sentence responses uh to each
  • 00:32:12
    question yes or no is is also great does
  • 00:32:15
    that sound all right okay I'm ready and
  • 00:32:18
    some of these will be highly entropic
  • 00:32:20
    questions in the the goal is to to learn
  • 00:32:22
    more about George church and how he
  • 00:32:23
    thinks about the world but you have to
  • 00:32:25
    be brief um so uh the first question is
  • 00:32:29
    kind of a Turing test for biology is the
  • 00:32:31
    way that I think about it and an
  • 00:32:33
    appropriate response to this could be
  • 00:32:35
    that's a dumb question but will AI
  • 00:32:38
    understand biology in any meaningful
  • 00:32:40
    sense where understand is in air quotes
  • 00:32:43
    here I think in a way it already does
  • 00:32:47
    highly Advanced biotechnologists talk to
  • 00:32:50
    each other essentially in biotechnology
  • 00:32:52
    which is not really natural
  • 00:32:55
    language and computers
  • 00:32:57
    I gave an example for protein
  • 00:33:00
    design all right George what's your
  • 00:33:03
    favorite piece of
  • 00:33:05
    music favorite piece of
  • 00:33:09
    music gee you know I I kind of like
  • 00:33:12
    Talking Heads oh nice you know this is
  • 00:33:16
    not my beautiful house I don't know
  • 00:33:19
    given a little more time I could with a
  • 00:33:21
    few others yeah that's the point of this
  • 00:33:23
    though is to inject some entropy so um
  • 00:33:27
    uh you've had a storied career in
  • 00:33:29
    science um what one thing has changed
  • 00:33:31
    the most either from a technology
  • 00:33:33
    standpoint from a society standpoint
  • 00:33:35
    from a political standpoint in science
  • 00:33:37
    over the course of your
  • 00:33:39
    career well I would say all of the above
  • 00:33:43
    in the Genome Project because it the NIH
  • 00:33:46
    was entirely hypothesis driven now
  • 00:33:49
    you've got one Institute that is
  • 00:33:51
    Discovery driven and another one that's
  • 00:33:53
    engineering driven and that coincided
  • 00:33:56
    roughly with a two-fold increase in the
  • 00:33:58
    NIH budget while we were starting the
  • 00:34:00
    Genome Project I can't say it was cause
  • 00:34:02
    and effect but it was it was a nice
  • 00:34:05
    coincidence okay do biologists need to
  • 00:34:08
    understand machine learning to
  • 00:34:10
    contribute to machine learning
  • 00:34:13
    projects do citizens have to understand
  • 00:34:16
    GPS and atomic clocks in order to find
  • 00:34:19
    directions on Google Maps all right
  • 00:34:22
    we'll accept it one sentence we accept
  • 00:34:25
    it as we know the the price of things in
  • 00:34:28
    healthcare tends to be sticky so the
  • 00:34:30
    question is will machine learning
  • 00:34:32
    ultimately reduce costs for Diagnostics
  • 00:34:34
    and
  • 00:34:36
    drugs yes is the short answer uh but I
  • 00:34:40
    don't and I give an example where it's
  • 00:34:41
    not sticky defending on how you define
  • 00:34:43
    things so gene therapy was $2.8 million
  • 00:34:47
    a dose until we got to coid 19 and the
  • 00:34:52
    top five vaccines were all formulated as
  • 00:34:54
    gene therapy some as low as $2 a day do
  • 00:34:57
    so 2.8 to two so give me you know that's
  • 00:35:01
    not very sticky okay our pre-prints and
  • 00:35:05
    net scientific
  • 00:35:06
    good
  • 00:35:09
    preprints I think net is the key word
  • 00:35:11
    yeah know they're good and bad pretty
  • 00:35:13
    high levels of good and bad but yeah I
  • 00:35:16
    think they're that positive okay final
  • 00:35:18
    question of the lightning round if you
  • 00:35:20
    could have dinner with one person Dead
  • 00:35:22
    or Alive who would George Church have
  • 00:35:24
    dinner
  • 00:35:25
    with uh oh jeez I think probably Netty
  • 00:35:29
    Stevens you've probably never heard of
  • 00:35:31
    her but she was on the little Google
  • 00:35:33
    logo but I knew her before that in
  • 00:35:35
    around 1910 1915 she found the
  • 00:35:37
    chromosome theory of inheritance along
  • 00:35:41
    with Morgan but we're separate from
  • 00:35:43
    Oregon about the same time all right all
  • 00:35:47
    right so we're gonna move to the final
  • 00:35:49
    segment of the episode we're GNA talk
  • 00:35:50
    about some big picture things I think
  • 00:35:52
    it's fair to say we've already touched
  • 00:35:53
    on some big picture topics so far but
  • 00:35:56
    we're going to try and broad the
  • 00:35:57
    aperture just a little bit
  • 00:35:58
    further we've talked a lot about
  • 00:36:01
    biotechnology and your work in the area
  • 00:36:03
    I want to come back a little bit to a
  • 00:36:06
    clinical focus and given sort of what
  • 00:36:08
    you see happening either in Diagnostics
  • 00:36:11
    or gene therapy what Medical Specialties
  • 00:36:14
    do you think are most likely to be
  • 00:36:15
    changed and impacted by
  • 00:36:18
    AI Medical Specialties hopefully
  • 00:36:21
    genetics interpreting The genome is
  • 00:36:23
    increasingly engaging polygenic risk
  • 00:36:26
    score
  • 00:36:27
    and I think that probably could be done
  • 00:36:30
    better um and then that could have
  • 00:36:32
    impact on almost every field of medicine
  • 00:36:35
    the other thing is age related diseases
  • 00:36:38
    I think there there's an opportunity of
  • 00:36:39
    having multiple genes involved in gene
  • 00:36:44
    therapy and possibly even personalized
  • 00:36:47
    or personalized medicine in general but
  • 00:36:49
    especially related to aging because
  • 00:36:52
    aging affects every disease basically
  • 00:36:54
    almost every form of human morbidity
  • 00:36:57
    immortality is impacted so I think those
  • 00:36:59
    are a cluster of three things that
  • 00:37:02
    interact with one another genetics aging
  • 00:37:05
    and machine learning so thinking about
  • 00:37:08
    our listeners who are clinicians and in
  • 00:37:10
    particular early career clinicians Med
  • 00:37:13
    students residents what do you think
  • 00:37:15
    those folks should know about AI to help
  • 00:37:18
    them prepare for a career in
  • 00:37:21
    medicine I think that they have a fairly
  • 00:37:24
    high level view of it it's it's like we
  • 00:37:26
    no longer most of us don't program in
  • 00:37:30
    zeros and ones we program with high
  • 00:37:32
    level languages like python or or maybe
  • 00:37:35
    even HTML or Excel or something so it'll
  • 00:37:38
    be like that hopefully it'll be very
  • 00:37:40
    easy to interface with as is the case
  • 00:37:43
    for most really awesome software but
  • 00:37:45
    they will have to know it and they might
  • 00:37:47
    not have to memorize as much when I was
  • 00:37:50
    a boy you know he had to memorize all
  • 00:37:52
    the biochemical Pathways and all the
  • 00:37:54
    pathologies and hopefully it'll be like
  • 00:37:57
    how do you look for it how do you
  • 00:37:58
    interface with the machine learning and
  • 00:38:01
    the big databases so you can't know it
  • 00:38:05
    all anymore but how do you know where to
  • 00:38:08
    look for the answer there's not going to
  • 00:38:11
    be a KB cycle of AI for for Physicians
  • 00:38:14
    to memorize hopefully not on the
  • 00:38:17
    contrary it's going to mean fewer people
  • 00:38:19
    learning crb cycle excellent and they'll
  • 00:38:23
    learn instead oh uh isocitrate dehydra
  • 00:38:27
    enase is very impactful on certain gomas
  • 00:38:31
    right and so it's one of the most
  • 00:38:33
    treatable of the of the otherwise nasty
  • 00:38:36
    category of cancer that hits the the
  • 00:38:39
    brain yeah I'm married to a clinician
  • 00:38:41
    and I think that the Mandate that I've
  • 00:38:44
    been given is that if I develop anything
  • 00:38:46
    in the AI space that makes her job more
  • 00:38:48
    difficult if I have a new KB cycle that
  • 00:38:50
    she has to memorize then that thing is
  • 00:38:51
    not going to get very far clinically so
  • 00:38:54
    I yeah yeah there's there
  • 00:38:57
    in this day and age there's really no
  • 00:39:00
    excuse for poor user interface in in
  • 00:39:02
    Computing or another checkbox that you
  • 00:39:05
    have to click or something like that
  • 00:39:07
    exactly that's true there there's plenty
  • 00:39:10
    of bat softare nevertheless but yeah but
  • 00:39:12
    there's no excuse for it yeah so I'm
  • 00:39:15
    glad that I get to ask this question
  • 00:39:17
    because I'm sure that we'll get a great
  • 00:39:19
    answer what is your most controversial
  • 00:39:23
    opinion my most controversial opinions
  • 00:39:26
    were
  • 00:39:27
    placed on me not from me okay so like
  • 00:39:30
    things having to do with Advocate I
  • 00:39:32
    don't preced this but I don't Advocate
  • 00:39:35
    but that people felt that I was
  • 00:39:38
    advocating cloning human
  • 00:39:41
    neanderthals so uh a controversial
  • 00:39:44
    opinion that is yours not attributed to
  • 00:39:46
    you one that's mine is probably that
  • 00:39:50
    everybody should seriously consider
  • 00:39:51
    getting their genome sequence in
  • 00:39:53
    particular if they're of reproductive
  • 00:39:56
    age age you know let's say 16 and up
  • 00:40:00
    especially for men that keep going that
  • 00:40:04
    they should know their carrier status uh
  • 00:40:06
    and that could influence who they date
  • 00:40:08
    or various other things I think the idea
  • 00:40:10
    of dating app that is aware of your
  • 00:40:14
    carrier status is the most Humane place
  • 00:40:16
    to do it but the controversy is they
  • 00:40:19
    either think that that's Eugenics which
  • 00:40:21
    it isn't or it's controversial because
  • 00:40:25
    yeah don't want to der
  • 00:40:27
    romanticize something by being so
  • 00:40:29
    technical but anyway I think that's a
  • 00:40:32
    huge missed opportunity it's more Humane
  • 00:40:34
    because I'm glad we're not in the
  • 00:40:36
    lightning round it's more Humane because
  • 00:40:39
    if you do it after you're pregnant then
  • 00:40:43
    you have the tough decision about
  • 00:40:45
    termination of pregnancy which is tough
  • 00:40:49
    for essentially everybody where you're
  • 00:40:51
    pro-choice or pro- lifee and if you do
  • 00:40:53
    it after you're married then you got the
  • 00:40:55
    tough decision are you going to have
  • 00:40:57
    children with this person which means
  • 00:40:59
    you're going to do invitro fertilization
  • 00:41:01
    which is no walk in the park and so
  • 00:41:03
    that's bad news but if you do it before
  • 00:41:05
    you've even met the person then it means
  • 00:41:07
    that out of a thousand people you could
  • 00:41:09
    date you're going to date you know 990
  • 00:41:12
    of them and at most and so you're going
  • 00:41:16
    to eliminate a few and and there's no
  • 00:41:17
    false positive problems at that point
  • 00:41:20
    there's definitely a false positive
  • 00:41:21
    problem if you're doing IVF or or
  • 00:41:25
    termination but or or worse yet you know
  • 00:41:28
    doing a surgery to remove organs that
  • 00:41:31
    might be at risk for cancer but there's
  • 00:41:33
    essentially no false positive problem
  • 00:41:34
    when you're rejecting 3% of the
  • 00:41:38
    potential suitors and at the risk of
  • 00:41:40
    misattributing another quote to you I
  • 00:41:42
    think I've heard you talk about this
  • 00:41:44
    before and it would go something like
  • 00:41:45
    there's a dating app like you said and
  • 00:41:47
    silently behind the scenes you're
  • 00:41:49
    getting screened out from people who
  • 00:41:51
    have the same carrier status as you so
  • 00:41:53
    that you're never matched with someone
  • 00:41:55
    who would you so that two recessive
  • 00:41:57
    geneses would would come together right
  • 00:41:59
    so you You' kind of not even know that
  • 00:42:01
    it was going on behind the scenes you
  • 00:42:03
    would just be matched with people that
  • 00:42:05
    you would not have this problem with
  • 00:42:07
    right that's accurate and and that's
  • 00:42:09
    Humane in addition to all the things I
  • 00:42:11
    mentioned that's Humane because another
  • 00:42:14
    awkward time is you've decided to marry
  • 00:42:16
    somebody and then you get the score and
  • 00:42:18
    then you decide not to uh this is or or
  • 00:42:22
    for that matter you get people find out
  • 00:42:24
    you know you're all ready to get married
  • 00:42:25
    and then marriage is off um and then
  • 00:42:28
    everybody knows that both of you are
  • 00:42:30
    carriers and so in a certain sense maybe
  • 00:42:32
    in less accepting parts of society you
  • 00:42:36
    both get branded as why should anybody
  • 00:42:38
    date them and in fact right everybody
  • 00:42:40
    you know 97% of people should date them
  • 00:42:42
    just not the 3% that are mismatched and
  • 00:42:45
    so by avoiding anybody knowing that your
  • 00:42:48
    carrier
  • 00:42:49
    status I think it's the most Humane
  • 00:42:52
    thing so both more Humane and no scarl
  • 00:42:55
    letter no Stigma
  • 00:42:56
    like none of none of those issues yeah
  • 00:42:59
    right so I think that's the time to do
  • 00:43:00
    it or we could destigmatize everything
  • 00:43:03
    but that's I think that's harder it's
  • 00:43:05
    hard to say which is hard but it could
  • 00:43:06
    be hard yeah George are there any
  • 00:43:08
    examples of a I guess I think you called
  • 00:43:10
    a dating app are there any examples of
  • 00:43:12
    that in place or is this is just an idea
  • 00:43:15
    there are not exactly that not an app
  • 00:43:18
    and it's not quite but it's like that
  • 00:43:20
    which is dorya sharim which was started
  • 00:43:23
    by a rabbi I think mid 80s so it's been
  • 00:43:26
    around for a while because he had I
  • 00:43:29
    think four of his children had taacs
  • 00:43:31
    which is a serious disease kills kids
  • 00:43:34
    painfully at age
  • 00:43:36
    4ish and he just decided that there
  • 00:43:39
    should at least be the option among
  • 00:43:42
    his congregation is anybody that that
  • 00:43:45
    could have similar afflictions and it
  • 00:43:48
    scaled up to I think eight or nine genes
  • 00:43:50
    typically that are enriched in the
  • 00:43:52
    ashkanazi population but in a certain
  • 00:43:55
    sense we're all at risk for those eight
  • 00:43:57
    or nine genes and about a thousand more
  • 00:44:00
    and it's not clear why it hasn't spread
  • 00:44:02
    it's been very successful in in the
  • 00:44:04
    populations that use it where it's
  • 00:44:07
    lowered the risk of such births by at
  • 00:44:10
    least a factor of 10 and why it hasn't
  • 00:44:12
    spread to other populations is I don't
  • 00:44:14
    think it's because one population knows
  • 00:44:17
    more or less science than than the other
  • 00:44:19
    ones it's something
  • 00:44:21
    else it's not that one population has
  • 00:44:24
    necessarily more genetic diseases than
  • 00:44:25
    other it is true that some inbred
  • 00:44:28
    populations have slightly higher but
  • 00:44:30
    that's not point is we're all at at
  • 00:44:32
    least a 3% risk
  • 00:44:34
    right so our our final question for you
  • 00:44:37
    George and you can take this in multiple
  • 00:44:39
    possible ways it's up to you what
  • 00:44:42
    applications of AI to biology keep you
  • 00:44:45
    up at
  • 00:44:47
    night oh yeah well first of all I'm
  • 00:44:51
    genetically narcoleptic and so nothing
  • 00:44:53
    keeps me up at night it's about 30
  • 00:44:55
    seconds is the median time to falling
  • 00:44:58
    asleep but what keeps me up during the
  • 00:45:00
    day is anything involving discrimination
  • 00:45:04
    so artificial intelligence could more so
  • 00:45:08
    it's an interesting question when we
  • 00:45:10
    worry about discrimination are we
  • 00:45:11
    worried about it being too inaccurate no
  • 00:45:15
    words we're we're stereotyping an entire
  • 00:45:17
    people or category of people that maybe
  • 00:45:21
    have a a priori a low probability of
  • 00:45:24
    living up to the stereotyp type or are
  • 00:45:26
    we worried it's too accurate is it are
  • 00:45:28
    we worried that it's not accurate enough
  • 00:45:30
    or it's too accurate and I think it's
  • 00:45:32
    Case by case but any case it could have
  • 00:45:35
    enough imp premature of accuracy that it
  • 00:45:38
    would be used but still inaccurate
  • 00:45:41
    enough that it could be abuse so that's
  • 00:45:42
    one scenario the other scenario is they
  • 00:45:45
    could use it you know make personalized
  • 00:45:47
    weapons you know once once the
  • 00:45:48
    Terminators come then are human failures
  • 00:45:53
    could be in a certain sense public to
  • 00:45:55
    the machine
  • 00:45:56
    even though they're not public in other
  • 00:45:58
    words I can't tell it but they could
  • 00:46:01
    figure it
  • 00:46:02
    out I guess a follow-up question to that
  • 00:46:05
    is I know like you're working Gene
  • 00:46:07
    editing you spend a lot of time thinking
  • 00:46:08
    about how when this technology becomes
  • 00:46:11
    democratized and you can buy like a a
  • 00:46:13
    reagent kit for $10 like how do we what
  • 00:46:16
    are a set of ethics and what are a set
  • 00:46:18
    of protocols that we can use in a world
  • 00:46:19
    like that is that at all similar to how
  • 00:46:22
    you think about what's happening with AI
  • 00:46:24
    either generally or a biology because a
  • 00:46:27
    model that cost $10 million to create
  • 00:46:29
    today will you know cost $10 to create
  • 00:46:32
    five years from now so how do we think
  • 00:46:34
    about these powerful technologies that
  • 00:46:37
    are also being democratized at like a
  • 00:46:39
    very very quick Pace how can we sort of
  • 00:46:42
    balance safety and progress in that kind
  • 00:46:45
    of
  • 00:46:47
    world well we've demonstrated that we
  • 00:46:50
    that there's no such thing as a slippery
  • 00:46:52
    slope that is to say there are
  • 00:46:55
    documented cases where we were able to
  • 00:46:58
    keep ourselves off the slippery slope
  • 00:46:59
    and other cases where we are were not at
  • 00:47:01
    least not for the whole population
  • 00:47:04
    there's always some percentage of the
  • 00:47:05
    population that falls into the Trap so
  • 00:47:07
    for example speed limits there is no
  • 00:47:10
    magic point where suddenly it becomes
  • 00:47:13
    unsafe but people tend to stay pretty
  • 00:47:16
    close to the speed limits yeah I guess
  • 00:47:18
    the question was about either
  • 00:47:21
    professional societal Norms so I know
  • 00:47:23
    that there's been a lot of this in gene
  • 00:47:25
    editing where there are groups that meet
  • 00:47:27
    to discuss safety and come up with
  • 00:47:30
    regulation is there any of that lessons
  • 00:47:33
    from that community that transport to AI
  • 00:47:38
    generally or AI in
  • 00:47:41
    biology right so you don't need special
  • 00:47:44
    groups to monitor Gene editing uh you
  • 00:47:48
    have the FDA and the FDA is very
  • 00:47:50
    effective at keeping us bringing things
  • 00:47:52
    out that are safe and effective that
  • 00:47:55
    also applies to medical devices so it is
  • 00:47:57
    possible the AI would fall in that
  • 00:48:00
    category but it's also possible you can
  • 00:48:02
    evade the category by making something
  • 00:48:04
    that's not
  • 00:48:05
    recognizable for example dating apps
  • 00:48:07
    that don't seem to be reg regulated by
  • 00:48:09
    the FDA even though they could have a
  • 00:48:11
    trillion dollar impact on medicine in a
  • 00:48:14
    certain sense they're not medicine but
  • 00:48:16
    nevertheless so even though people try
  • 00:48:20
    to make regulations on top of
  • 00:48:21
    regulations so they wanted to have a
  • 00:48:23
    moratorium on Gene editing on top of the
  • 00:48:26
    FDA which has a moratorium on all new
  • 00:48:28
    drugs I thought that was a little crazy
  • 00:48:31
    redundant and it didn't really
  • 00:48:34
    happen but with AI if it does slip
  • 00:48:37
    between the cracks then there should be
  • 00:48:38
    some kind of safety now our track record
  • 00:48:41
    for that not so great in Computing uh if
  • 00:48:44
    you look at the internet there was very
  • 00:48:46
    little of the foresight that existed
  • 00:48:48
    very little of it made it in in time so
  • 00:48:51
    there's wide open doors for hacking for
  • 00:48:54
    computer viruses for identity theft for
  • 00:48:58
    abuse of children and pornography and so
  • 00:49:02
    forth so we didn't do such a great job
  • 00:49:04
    there and I hope we do a better job with
  • 00:49:06
    AI partly because of fantastic
  • 00:49:10
    educational media by which I mean the
  • 00:49:15
    Terminator all right uh that maybe we'll
  • 00:49:18
    edit it so that we don't end on a dower
  • 00:49:20
    note like that but I think that yeah
  • 00:49:22
    that would be
  • 00:49:23
    better ending on The Terminator yeah but
  • 00:49:26
    I can't remember the last time I've had
  • 00:49:27
    a conversation where we discussed
  • 00:49:29
    resuscitating Willie mammoths and dating
  • 00:49:31
    apps within the same hour so it's it's
  • 00:49:33
    been it's been a really special
  • 00:49:35
    conversation thank you so much George
  • 00:49:37
    for being on AI Grand rounds it's been a
  • 00:49:38
    pleasure thank you it was great I I look
  • 00:49:41
    forward to hearing it yeah great
  • 00:49:45
    [Music]
  • 00:49:47
    thanks
Etiquetas
  • George Church
  • Genetics
  • AI in Biology
  • Personal Genome Project
  • Ethical Implications
  • Genomic Data
  • Woolly Mammoths
  • Machine Learning
  • Protein Design
  • Healthcare Costs