자막 (390)
0:00- As humans, we all want the same thing,
0:02a life that's full of good experiences,
0:05more time with family, with friends,
0:08but sometimes genetic
illness can cut that short
0:12or really, for all of us at some point,
0:14our body breaks down.
0:15And our bodies are genetic machines.
0:18For many diseases, the
cause of the disease
0:21is a mutation in the genome.
0:24Gene therapy is a vision that
many have had for decades,
0:30The power of genetic technology is that
0:32once you get inside of
cells with a DNA molecule,
0:36that molecule can stay there
for the lifetime of the cell.
0:39So it's the potential
for a one-time treatment
0:42for a disease where you wouldn't otherwise
0:45be able to reach the cells
0:46and solve for the root
cause of the disease.
0:49Today though, for the most part,
0:51the genome you're born with
0:52is the genome you die with.
0:54Access to this molecular level
0:56is out of reach for almost all of us.
1:00We've tried many different things,
1:01but have really struggled
to be able to get
1:04enough of the genetic
payload into the cells
1:07where they're gonna be
effective as a therapeutic.
1:09And it's getting inside of the cells
1:12that has really been a
challenge for many, many years.
1:16I'm Eric Kelsic, CEO and
co-founder at Dyno Therapeutics.
1:19For the past 10 years, I've been working
1:21to solve the grand
challenge of gene delivery.
1:31How are we gonna make gene therapy
1:33a mainstream kind of medicine?
1:35We need to solve these grand
challenges like delivery,
1:38being able to deliver
a therapeutic payload
1:40to every organ or every cell
1:43where there might be some
benefit to patient health.
1:46To do that, we're
engineering protein shells
1:48derived from viruses.
1:50Capsids are the protein shells
of adeno-associated virus.
1:55AAVs, adeno-associated virus,
1:57is a parasite of other viruses.
2:00AAV naturally isn't known
to cause any disease.
2:04The reason why AAV gets a lot of attention
2:06is because it's one of
the smallest viruses,
2:09and that enables it to get into
2:11many places all across the body
2:13where we need to deliver
a therapeutic DNA.
2:15We still don't know a lot about
how it functions naturally.
2:18That said, we don't need to understand
2:20everything about how the virus works
2:21in order to adapt it as
a therapeutic technology,
2:24and that's our focus at Dyno,
2:26engineering the capsid sequence
2:28to make capsids a better delivery vehicle
2:32What's amazing about capsids
is they're evolved in nature
2:36to do so many different things.
2:37So they can go through your
body, through the blood,
2:40find a cell, enter the cell,
2:42and then be released
into the cell cytoplasm,
2:44get into the nucleus
through the nuclear pore,
2:46break open the capsid
and release the genome,
2:49and that's where it expresses.
2:51So for a gene therapy, going
from the blood, into cells,
2:55into the cytoplasm, into the nucleus,
2:57and then expressing the genetic payload,
2:59that's entirely the goal.
3:01And when therapeutic genes
are expressed in the nucleus,
3:03they can be treating those cells
3:05for a patient's entire lifetime.
3:07As a one-time treatment, it
can be an effective cure.
3:10However, natural capsids,
they're not efficient enough
3:13for most therapeutic purposes.
3:15So for the past 28 years,
3:18protein engineers have been
working to modify the capsid
3:21to make it better as
a therapeutic protein,
3:23applying a technique
called directed evolution.
3:26Directed evolution is
evolution like occurs in nature
3:29but for a goal that we choose,
3:32and the most common
approach there had been to
3:33randomly change the capsid sequence
3:36to make very large libraries,
3:37millions or even billions of different
3:39capsid sequence molecules.
3:41With a very low quality
library, but a very large one,
3:44you have a chance of getting a good hit,
3:46but it's like a needle in a haystack.
3:49And the reason is because the capsid has
3:51many different functions,
3:52and if you break even one of them,
3:55then, as a therapeutic,
it's essentially useless.
3:58Roughly 80% of the single changes
4:01that you could make to the capsid
4:02break the most essential function,
4:05which is the assembly and
packaging of the genome.
4:08What that means is that, if by
chance you make any mutation,
4:11four out of five times, it's
gonna break the function.
4:14And that's a problem
for engineering because,
4:16to get improved function,
4:17we're gonna need to make multiple changes,
4:19maybe even hundreds of changes.
4:21So if every time you make a change,
4:22the viability drops down,
4:24it's really hard to have
4:25a library of changes that are going to
4:27do a chance of finding an improved capsid.
4:31And that's just the basics.
4:33You need to be able to
4:34produce and purify that capsid at scale.
4:37It needs to be stable at
low temperature or frozen,
4:40but even when it's in your body,
4:42which is a relatively high temperature,
4:44it also needs to get into the right cells.
4:47For example, there's a lot of unmet need
4:50for gene therapy in the brain
4:51because it's very difficult
to get therapeutic proteins
4:54or other molecules across
the blood-brain barrier.
4:57At a high dose, you might
be able to get into .1%
5:01or maybe a little bit more
of the neurons in the brain.
5:04That's not enough to treat many diseases.
5:07And in addition to that,
5:08most of the capsid delivers
its payload to the liver.
5:12And at a high enough dose,
that can also become toxic.
5:16We need to improve the
efficiency of delivery
5:20Over decades of trying this approach,
5:22we just didn't get
enough improved variants
5:25or variants that were optimized
5:27for all the different
functions that were needed
5:28to make them effective as gene therapies.
5:31I had seen that there was a
new wave of technologies coming
5:34with the potential to change
5:35the way that we engineer
proteins completely.
5:38It starts with this DNA
multiplexing technology.
5:42So we have an idea of an
experiment we want to run,
5:45testing many different capsids.
5:47They might be designed to
bind to a certain receptor
5:50or they might be designed
5:51in a neighborhood of sequence space
5:52that before we found is promising.
5:55And we came up with a way of
5:57building very large libraries of
5:58capsids in which the
sequence was programmed,
6:01meaning we had designed it on a computer,
6:03synthesized that DNA,
6:05and then cloned it into the capsid,
6:07so this could be injected in a few mLs.
6:10The best way we have to make a prediction
6:13about what's gonna be safe
and effective for humans
6:15is to do an animal experiment.
6:17We do most of our screening
in non-human primates,
6:20especially in cynomolgus monkeys.
6:22That's one reason why we
developed this technology
6:24because their lives
are also very precious.
6:26We wanna get as much information as we can
6:29from even one experiment.
6:31In this case, we're measuring
6:32maybe a hundred or two hundred thousand,
6:34sometimes even a million
different capsid sequences
6:38We'll get all these tissues back
6:39from our animal experiments,
6:41and then we wanna learn as much as we can
6:43from that experiment,
meaning look at every organ.
6:46Where did the capsids go
or where did they not go?
6:48Extract the nucleic acids,
6:51purify the DNA, purify the RNA.
6:53You can then work all the way back to
6:56what was the capsid sequence
6:57that this molecule corresponds to?
6:59Is there more or less
of that in the library?
7:01And if there's more, that might mean that
7:03it was functionally improved for delivery.
7:05If there's less, it might
mean that there was a problem
7:08We do this across all of our library.
7:11At Dyno today we've got petabytes of data
7:13from the DNA sequencing.
7:16I had always thought that proteins,
7:18they're too complex for us
to understand as humans,
7:21certainly too complex
for me to understand.
7:23When you look at a string of 735 letters,
7:26it's really hard to notice
all the differences.
7:29But with all that data, what
I could see, even myself, was
7:32there's a lot of patterns in that data,
7:34patterns about which amino
acids work at each position.
7:39My thought was that if I
could recognize those patterns
7:42and the data set is so vast,
7:44there's probably a lot more
information in them as well.
7:47That's actually the
perfect type of problem
7:48for a machine learning model.
7:51We can use AI to automate
the analysis of all that data
7:56and to find even more nuanced patterns
7:58to maximize the chances of success,
8:00the expected value of
finding an improved variant.
8:03We call that AI-guided design.
8:06But once we have that data
8:07and we've trained models on it,
8:09we can now query those models
8:12billions and billions of times.
8:13So our ability to scale
the computational work
8:17is even higher than the molecular side.
8:19We can't possibly test
everything in an experiment.
8:22But with machine learning,
8:23we can test many different
sequences in silico,
8:25meaning on a computer,
8:28and the models will tell us which ones
8:29they think are better,
8:31or we might try many models,
8:33tens or hundreds of different models
8:34that each have a different insight.
8:36And we compare the opinions
of all those different experts
8:38to choose the ones that
we're very confident
8:41are worth investing in
8:42as we bring them forward
into the next experiment.
8:46So it's this iterative cycle
of making libraries in DNA,
8:50measuring their properties,
8:53to analyze and understand
those properties.
8:55Then querying the models to know
8:57where are the most promising regions
8:59of sequence space that we should go next.
9:00And then going back to
design a new library,
9:02turning that into DNA.
9:05We're using a lot of technology,
9:06but there's always human
judgment at some point
9:09before we do another round of experiments.
9:12I think that, over time, what we wanna do
9:14is put humans at an even
higher point of leverage
9:17so that they're able to use
their exceptional judgment
9:20and shift some of the more
routine tasks to AI agents
9:24or even just to simple scripts
that run on the computer.
9:29Being able to collaborate
with AIs more effectively
9:32is where we'd like to go so that we can,
9:34for example, give instructions to the AI
9:36to automate how we analyze the
library or how we design it
9:40and get back the answers that we expect.
9:43We wanna get the results
as fast as we can.
9:46Patients are waiting for better medicines.
9:48We wanna make sure that,
if there's anything wrong,
9:51and for that, we need a human in the loop.
9:54The power of genetic technology is that
9:57once you get inside of
cells with a DNA molecule,
10:00that molecule can stay there
10:03for the lifetime of the cell.
10:04So for example, in the neurons
where they're not dividing,
10:08getting the right therapeutic
DNA sequence into the neuron
10:11can be effectively a cure for
a patient's entire lifetime.
10:14That's the reason why at Dyno,
10:17and myself personally, and many of us
10:19are so excited about the
potential of gene therapy.
10:23A good example of this would be Zolgensma,
10:25which is now an approved medicine
10:26and was really a breakthrough drug.
10:29SMA, spinal muscular atrophy,
was the leading cause of death
10:33from a genetic disease in
children prior to this treatment.
10:36The problem is that the SMN1 gene
10:39is not functional in patients.
10:42This disease, prior to gene
therapy, was always fatal
10:46at a very young age. Children would die
10:48usually around two or three years old.
10:51With Zolgensma though,
10:52if children are treated very early,
10:55in the first few weeks of life, say,
10:57the gene therapy can restore
the function of that gene.
11:00It can, with a one-time treatment,
11:02completely cure the disease.
11:04And it's an example of the
11:06amazing potential of gene therapy.
11:09What's unfortunate is that there's,
11:10today, just a handful of
FDA-approved gene therapies,
11:13but there's thousands of genetic diseases
11:16that we know about, 7,000 or more.
11:19And for most of them,
11:20we have no good treatment
options available.
11:23What we want to be able to do
11:25with our capsid engineering work,
11:27by solving deliveries,
make it easier to get into
11:29all the cells that will
enable us to then apply
11:32the knowledge we have
from genome sequencing
11:34and from systems biology
11:35to develop therapies that are gonna treat
11:37the underlying cause of those diseases.
11:40Today, most of our attention,
11:42most of the industry's attention
11:44is focused on a smaller
number of diseases,
11:46diseases that could benefit more patients,
11:49and where the markets are large enough
11:50to justify commercial investment.
11:53There's also a long tail of
rare and ultra-rare diseases
11:56where there might only be 10
11:58or even a single patient
in the entire world
12:00who has a certain disease.
12:02Today, gene therapies are very expensive.
12:05A single dose might cost
millions of dollars.
12:08Our goal is to bring the cost
of delivery down to zero,
12:10or very, very close to it.
12:12To do that, we also need
to enable there to be
12:14many more genetic medicines
12:16so that there's good
competition between developers.
12:19We can also look to other
industries where there's been
12:22really dramatic changes
in the cost efficiencies
12:25and the scale economies over time.
12:27One of them is in semiconductors,
12:29as we've been able to dramatically improve
12:31the number of transistors
that you can put on a chip.
12:34Other areas like solar where
12:36we'd been able to bring the cost down
12:38more than 200 fold over five decades
12:40and increase deployment
over 100,000 times.
12:44These phenomena are all
called Wright's Law,
12:47which is basically that with
every doubling of production,
12:49there's a percentage decrease in the cost.
12:52And in gene therapy,
12:53I think there will be something similar.
12:54So we can go from a gene
therapy that might cost
12:57hundreds of thousands of dollars
12:59down to something that costs $10,000
13:00or even $1,000 to develop.
13:03To the point where rare
diseases or ultra-rare diseases,
13:06non-profit efforts could be fully funding
13:08the treatments for patients.
13:10Obviously, there's a lot
of things we need to do
13:12in order to achieve that.
13:13Solving delivery is just one of them.
13:15The therapies are complex
13:16and we don't understand exactly how to
13:18design them in a way they're
gonna work in humans,
13:22but AI may be part of that solution
13:24because if an AI could design
13:26a therapy just for one patient
13:28and customize it to their genome sequence,
13:32customize it to their goals,
13:35that could be done on-demand,
13:37and that AI could even chart out
13:39how to develop the
therapy, how to produce it,
13:41how to test it, how to
ensure that it's safe.
13:44This could be done in a
massively scalable way.
13:46And I think that's the
path that we can use
13:48to solve for the long tail of disease
13:50and to help patients who, today,
13:53we understand their genetics.
13:56We know what the problem is.
13:58We even, in many cases, know how to design
14:01a therapy that could help them,
14:03but we need to be able to bring that
14:04to the patient directly,
14:06and AI is a way that
they can get the benefit
14:08from all this innovation
14:10in a way that's economically affordable.
14:14As there's more gene therapies,
14:16one thing that we may want to do is
14:18to be able to reset or
remove prior gene therapies.
14:21It's far away because it's
not the urgent priority today.
14:24But for a future with genetic agency
14:26where patients are making the best choice
14:28for them to live a healthy life,
14:30they may want to be able to
upgrade their therapy in time.
14:33This ability to reset would
give them that potential.
14:36For example, if there's a new approach
14:38that's even more effective,
14:39a patient wouldn't think
twice about taking that now,
14:42knowing that they could
remove it in the future
14:43and replace it with a better therapy
14:45that might come along in 10 or 20 years.
14:48That makes gene therapy a
much more routine decision.
14:52What that means is that
14:54we can think about genetic technologies
14:56less as really a part of us,
14:58but just something that we choose to use
15:00in the same way that I might
wear one set of clothes a day
15:03or a different set next year,
15:04but that's not really part of who I am.
15:07And I think about that
very differently than
15:09today how I think about my genome,
15:11which has always been a part of me.
15:12And up until very recently,
15:13I thought I would die
15:15with the same genome that I was born with.
15:17Because of the genetic technologies,
15:19I think we're gonna no longer
15:21associate the genetics that
we have with who we are,
15:24and it's more a decision
for who we want to be
15:28or what we want to become,
15:30and you'll be able to have
much more control over that
15:32so you can live the
very best possible life.
