Latest / What's Your Problem? / Can AI Help Solve Alzheimer’s?

Transcript

Can AI Help Solve Alzheimer’s? · What's Your Problem?

0:15
Speaker 1: Pushkin.
0:20
Speaker 2: So you know Moore's law, right, the idea that computers
0:23
Speaker 2: are specifically chips, get better and cheaper at an exponential rate.
0:28
Speaker 2: People who work on researching and developing new drugs, they
0:32
Speaker 2: talk about e Room's law. E room is more spelled backwards.
0:36
Speaker 2: And this is sort of a half joke, half not
0:39
Speaker 2: a joke, way of pointing out that developing new drugs
0:41
Speaker 2: has gotten slower and more expensive over the past several decades.
0:45
Speaker 2: It has gone in the opposite direction of developing new microchips. Today,
0:51
Speaker 2: after all the money and technology and hard work and
0:54
Speaker 2: intelligence poured into drug development, something like ninety percent of
0:59
Speaker 2: drugs that go into clinical trials fail, only ten percent succeed.
1:05
Speaker 1: That means that if we got to.
1:06
Speaker 2: A place where even just twenty percent succeeded, still an
1:10
Speaker 2: eighty percent failure rate, we would double the rate of progress.
1:21
Speaker 2: I'm Jacob Goldstein, and this is what's your problem. My
1:23
Speaker 2: guest today is Patrick Sue. Patrick got a PhD from
1:27
Speaker 2: Harvard in biochemistry when he was twenty one years old,
1:31
Speaker 2: and then four years ago, before he turned thirty, he
1:35
Speaker 2: founded the ARC Institute. It's a nonprofit research center in
1:39
Speaker 2: the Bay Area that hosts scientists from Stanford, u SE,
1:42
Speaker 2: San Francisco, and UC Berkeley, where.
1:45
Speaker 1: Patrick is on the faculty.
1:47
Speaker 2: Patrick's problem is this, how can you use AI to
1:51
Speaker 2: make biological research more efficient, to guide scientists more quickly
1:56
Speaker 2: to discoveries that'll lead.
1:57
Speaker 1: To new and better truemils.
1:59
Speaker 2: As you'll hear later in the conversation, Patrick is particularly
2:03
Speaker 2: focused on Alzheimer's disease. To start, he told me about
2:07
Speaker 2: why basic biological research is still so slow.
2:11
Speaker 3: So if you look at a biology research lab in
2:15
Speaker 3: the eighties or the nineties, the early two thousands, of
2:17
Speaker 3: the early twenty tens, or today in the twenty twenties,
2:21
Speaker 3: they look basically the same.
2:23
Speaker 1: Right.
2:24
Speaker 3: You have these long benches, you have these two or
2:26
Speaker 3: three rows of shelves, these micropipe pets, and various machines
2:31
Speaker 3: that look like home kitchen equipment. Right, and so graduate
2:36
Speaker 3: students and postdocs or bench scientists generally are like the
2:40
Speaker 3: line chefs right inside of you know, a Michelin star kitchen.
2:45
Speaker 3: You're prepping the vegetables. You know, you're making the sort
2:51
Speaker 3: of initial sauces. Right, So you might be taking tissues
2:54
Speaker 3: down or cells, processing them, staining them with antibodies. I
2:59
Speaker 3: think the point is that doing these experiments is extremely slow,
3:05
Speaker 3: very manual, and requires a huge amount of soft knowledge
3:08
Speaker 3: and of how and is really variable lab to lab.
3:11
Speaker 3: So we have this multi stack problem where we have
3:15
Speaker 3: multi step search. Right, the experiments take months to years
3:21
Speaker 3: to run out, and we basically search across the entire
3:26
Speaker 3: space in an extremely manual way. And so in the
3:30
Speaker 3: modern era of AI, where we believe we can predict
3:33
Speaker 3: things across many different fields, the question is can we
3:36
Speaker 3: speed up biology by being able to have models that
3:40
Speaker 3: actually have predictive value and power. And our goal is
3:44
Speaker 3: to be able to create biological intelligence that starts to
3:49
Speaker 3: crack that threshold of a cell. Biologist will use the
3:53
Speaker 3: model to rank the top twelve things that they'll do
3:56
Speaker 3: in the lab, rather than just going into lab and
3:59
Speaker 3: trying a bunch of things based on hypothsis driven science.
4:03
Speaker 3: But you actually do model first prediction and then a
4:07
Speaker 3: lab in the loop experiment.
4:10
Speaker 2: And I mean we'll get to the sort of things
4:12
Speaker 2: that have to happen for that to happen. But if
4:15
Speaker 2: we could get it to work, right, if you could
4:17
Speaker 2: get it to work, what's the payoff beyond the level.
4:20
Speaker 3: I think the first thing is the practice of how
4:23
Speaker 3: you design and do experiments will completely change. Right. And
4:27
Speaker 3: so just like you know, graduate students today, whenever they
4:29
Speaker 3: want to research a new area, they just use chatch
4:32
Speaker 3: to BT, right, or Claude or Gemini or whatever your
4:35
Speaker 3: favorite model is. Right, that has just become a fundamental
4:38
Speaker 3: part of the workflow. In fact, I don't think most
4:41
Speaker 3: people read papers the old fashioned way anymore, right. And
4:45
Speaker 3: it used to be, oh, I read them online in
4:47
Speaker 3: my browser instead of in a magazine like Nature and Science.
4:50
Speaker 3: And now it's just I throw the pdf into chatch
4:53
Speaker 3: BT and I read the summary. Right. That's just a
4:56
Speaker 3: completely different workflow than just even a few years ago.
4:59
Speaker 3: And I think instead of designing experiments by hand, we're
5:02
Speaker 3: gonna have AI models design the experiments for us and
5:05
Speaker 3: help us troubleshoot and decide what to do next.
5:08
Speaker 2: I've heard you describe the way biology works now as
5:11
Speaker 2: guess and check.
5:14
Speaker 1: Which I liked.
5:15
Speaker 2: Which I guess on one level is kind of the
5:17
Speaker 2: scientific method, right, It's kind of like hypothesis and experiment.
5:21
Speaker 2: But the guess really leans into the not quite blind,
5:26
Speaker 2: but the uncertainty, all the blind alleys that I guess
5:29
Speaker 2: you go down exactly.
5:31
Speaker 3: You know, it's a lot like that that a childhood
5:33
Speaker 3: game battleship where you're trying to sink someone's battleship. You
5:37
Speaker 3: don't quite know where it is, and so you're searching
5:39
Speaker 3: these different quadrants or like mind sweeper, Right, it's just
5:44
Speaker 3: very hard to know that you're searching in the right place.
5:48
Speaker 3: And so the first thing is can these A models
5:49
Speaker 3: help us search in the right place so that when
5:52
Speaker 3: we're peppering things with these individual manual wet lab experiments,
5:56
Speaker 3: are hit rate can just significantly go up, whether that's
6:00
Speaker 3: for designing a drud or you know, predicting how cell
6:04
Speaker 3: respond to you know, some perturbation.
6:08
Speaker 2: Yes, and then that's kind of the first order thing.
6:11
Speaker 2: So it's basically making basic research more efficient speed. Yeah,
6:15
Speaker 2: but yeah, one are the second and third order things
6:19
Speaker 2: like how does that translate to clinical outcomes?
6:23
Speaker 3: I think they're directly linked, right.
6:25
Speaker 1: Yeah.
6:26
Speaker 3: In drug discovery, we do a thing called target identification.
6:30
Speaker 1: Right.
6:30
Speaker 3: You need to be able to find the right drug target,
6:33
Speaker 3: and you need to be able to perturb it in
6:35
Speaker 3: the right way. You need to find the thing that's
6:38
Speaker 3: going wrong, the toxic protein that's causing a disease need
6:42
Speaker 3: try to turn it off. Right, So the goals really
6:44
Speaker 3: been to just finding right the right drug target and
6:47
Speaker 3: then drugging in the right way. But the problem is
6:50
Speaker 3: we don't seem to find the right drug target and
6:53
Speaker 3: even know what it is for most complex diseases, Alzheimer's disease,
6:57
Speaker 3: many different types of cancer, it's aging, autoimmunity, a lot
7:01
Speaker 3: of the major killers, it's still the same. Right, we
7:05
Speaker 3: don't know what to actually target. We think these models
7:08
Speaker 3: could help us do that.
7:09
Speaker 2: Let's talk about the ARC Institute, right, which is where
7:12
Speaker 2: a lot of this work is happening. A thing you've
7:14
Speaker 2: started and you run, like, tell me about the ARK
7:17
Speaker 2: Institute and like specifically why you started it.
7:20
Speaker 3: ARC is about four years old today, we're about three
7:24
Speaker 3: hundred and fifty people, and we're a full stack AI
7:28
Speaker 3: and biology research organization. And so by that, I mean
7:32
Speaker 3: we try to natively combine experimental science and computational science
7:38
Speaker 3: under a single physical roof so that we can do
7:41
Speaker 3: native iterative lab in the loop between AI models, both
7:46
Speaker 3: training and running them and then actually experimentally validating and
7:51
Speaker 3: verifying things in the lab in order to create new
7:55
Speaker 3: mechanistic insights, new drug compositions, and identify new therapeutic targets.
8:01
Speaker 3: We really have two major goals. The first is to
8:05
Speaker 3: create these virtual cell models that can simulate human biology.
8:09
Speaker 3: You have foundation models. The second is secure Alzhemer's disease.
8:13
Speaker 3: We're interested in the view of ARC as a sort
8:17
Speaker 3: of Edison Shop right where you're inventing a whole bunch
8:20
Speaker 3: of different things that we can actually productize. And so
8:24
Speaker 3: in a way, our ambition from ARC has been increasing
8:28
Speaker 3: beyond how do we do breakthrough academic and basic science
8:32
Speaker 3: to how do we actually productize this science in a
8:35
Speaker 3: way that we can impact human health not just with
8:38
Speaker 3: research papers, but with things that people can actually take,
8:41
Speaker 3: see and feel and use.
8:43
Speaker 2: When you talk about the Edison Shop, I mean obviously
8:45
Speaker 2: the Edison Shop was a business right, not a philanthropy,
8:49
Speaker 2: like are you how are you thinking about it commercially?
8:53
Speaker 3: So ARC is a nonprofit right, and so you know,
8:56
Speaker 3: I ARC will the ARC Institute will always remain this
9:00
Speaker 3: nonprofit discovery oriented mothership, but we also will certainly have
9:04
Speaker 3: the capacity and the interest to be able to spin
9:08
Speaker 3: out entities that can take on more focused commercial capital
9:11
Speaker 3: in order to productize things.
9:14
Speaker 2: Let's go back to the machine learning slash AI work
9:17
Speaker 2: you're doing.
9:20
Speaker 1: I think we should talk about EVO. Tell me about EVO.
9:23
Speaker 3: EVO is essentially a chat GPT, but only with DNA, right, okay,
9:28
Speaker 3: And so it's DNA in and DNA out right, So
9:32
Speaker 3: you can talk to it by typing in nucleotides and
9:35
Speaker 3: you'll receive from the model some corresponding set of nucleotides.
9:41
Speaker 3: So just like I could say to be or not
9:43
Speaker 3: to the model would tell me, you know, you know,
9:47
Speaker 3: to be or not to be? That is the question,
9:49
Speaker 3: you know, and then it will keep going and the
9:51
Speaker 3: rest of hamlets delilokey. Right. If I gave it a
9:54
Speaker 3: fragment of let's say, a mitochondrial genome or you know,
10:01
Speaker 3: e Coli genome, it will then try to autocomplete the
10:05
Speaker 3: rest of that fragment. It's not just memorizing in regard
10:09
Speaker 3: diating information from the training data, which is what's really important, right.
10:13
Speaker 3: It's doing a semantic diversification in a way of what
10:18
Speaker 3: it thinks is in the meaning of what it needs
10:20
Speaker 3: to make.
10:21
Speaker 2: Right.
10:21
Speaker 1: Yeah, that's the wild part.
10:23
Speaker 3: Right.
10:23
Speaker 2: If it was just regurgitating training data, it wouldn't be AI.
10:26
Speaker 1: It would just be sort of a library, right exactly.
10:29
Speaker 3: And this is why you can ask chatch bt to
10:31
Speaker 3: you know, here's what I want to see in my email,
10:33
Speaker 3: write me the email, and if you gave it that
10:36
Speaker 3: thing ten different times, it would write you similar but
10:39
Speaker 3: different emails.
10:40
Speaker 1: Right yeah.
10:41
Speaker 3: And just similarly, you can make different crisper systems that
10:44
Speaker 3: are similar but not the same, different mitochondrial genomes that
10:48
Speaker 3: are similar but not the same. And you might say, okay,
10:51
Speaker 3: well that's fun. Now you have a bunch of sequences
10:53
Speaker 3: in the computer. What's interesting about that? The point is
10:57
Speaker 3: that we can then chemically synthesize those DNA strands in
11:01
Speaker 3: the lab and test them and see what their function is. Now,
11:06
Speaker 3: we could then take those the numbers that we've measured
11:10
Speaker 3: in the lab and feed it back into the model
11:12
Speaker 3: and then tell the model for things like make it
11:15
Speaker 3: better at this task that I've experimentally verified, and then
11:20
Speaker 3: you can This is the closing the lab and the
11:22
Speaker 3: loop that we've built ARC around in order to he'll
11:25
Speaker 3: climb on. Really any experimental accut that you care about.
11:29
Speaker 2: I know there was a project that one of your
11:32
Speaker 2: colleagues did last year using evo to to basically to
11:37
Speaker 2: design a new version of an existing phage. A phage
11:40
Speaker 2: is a virus that infects bacteria. So, like, tell me
11:44
Speaker 2: about the project with the fix phage and why it's meaningful.
11:48
Speaker 3: So we have these databases where whenever people published papers,
11:54
Speaker 3: they deposit the genetic sequences into this database. So of
11:58
Speaker 3: all fix and FIX related phages, we have a sense
12:02
Speaker 3: of the evolutionary distribution that is out there that we've
12:06
Speaker 3: been able to detect so far as scientists, right, Yeah,
12:09
Speaker 3: And it turns out evo is able to make new
12:14
Speaker 3: versions of fi X that are as evolutionarily distinct as
12:19
Speaker 3: everything else that we've seen. So we're not just copying
12:23
Speaker 3: and making things that look like a Corolla or another
12:26
Speaker 3: a chord That would be very boring, but we can make,
12:29
Speaker 3: you know, a fundamentally new type of car. Right, Yeah,
12:33
Speaker 3: it's still a car, but it's a completely new type
12:37
Speaker 3: of design. Let's say, something like as unique as a
12:40
Speaker 3: pet cruiser, but much much cooler.
12:42
Speaker 2: Okay, And so so let's talk about the meaning of
12:46
Speaker 2: this as a research tool and then the clinical implications, right, Like,
12:50
Speaker 2: why is it meaningful as a research tool? Why is
12:52
Speaker 2: this a meaningful proof of concept.
12:56
Speaker 3: The first thing is that there they work in the lab,
13:00
Speaker 3: right that when you actually create these aid designed sequences,
13:04
Speaker 3: you can actually package real phage particles and they can
13:08
Speaker 3: actually in effect living bacteria in the lab the way
13:13
Speaker 3: that a normal one would. Now that's sort of step one.
13:16
Speaker 2: Just to be clear, this is a thing, a virus
13:20
Speaker 2: that has never existed before exactly that the AI invented exactly,
13:24
Speaker 2: just worked. The machine made it up and it worked,
13:27
Speaker 2: and it was quasi alive because it's a virus.
13:30
Speaker 3: Yeah, exactly, And so that that alone is pretty cool.
13:35
Speaker 3: The second thing is that we could steer the generation
13:39
Speaker 3: and so you can use it to infect specific strains
13:43
Speaker 3: of E. Coli that you cared about and not others
13:46
Speaker 3: that you didn't. And so you know, in phage therapy
13:49
Speaker 3: for example, this is the therapeutic implication of this. Right,
13:52
Speaker 3: if you were to you know, target the gut, you
13:55
Speaker 3: would want to target specific microbes for example, that are
13:58
Speaker 3: like bad gut bacteria and not others that are good.
14:02
Speaker 3: And so you don't want to just do a broad
14:04
Speaker 3: spectrum wiping out of you know, the entire complex gut community,
14:09
Speaker 3: which is what broad spectrum antibiotics do. But you could do,
14:12
Speaker 3: in principle, much more selective deletion of something that you want.
14:17
Speaker 2: Right.
14:18
Speaker 3: So, from a basic science point of view, the controllability
14:21
Speaker 3: of generation from the model can lead to like important
14:26
Speaker 3: precise functional outcomes in the lab. And then this has
14:30
Speaker 3: I think, you know, as I just explained really interesting
14:32
Speaker 3: therapeutic implications for phage therapy or a targeted therapy, a
14:37
Speaker 3: targeted therapy, right, and in a way both medicines, Well,
14:44
Speaker 3: we care a lot about precision and selectivity, right, yeah,
14:48
Speaker 3: and that's really the entire game of on target therapeutic
14:51
Speaker 3: efficacy and side effect profiles. Right.
14:55
Speaker 2: Right, You wanted to do one thing really well, and
14:57
Speaker 2: you wanted to do nothing else.
14:59
Speaker 3: You want it to be safe. You want it to
15:01
Speaker 3: be safe. Yeah, yeah, and effective, exactly safe and effective.
15:05
Speaker 2: Tell me about the work you're doing with EVO and
15:08
Speaker 2: the BRACA one gene, This gene that's implicated in some
15:12
Speaker 2: cases of breast cancer.
15:13
Speaker 3: You know, Broca one is an important gene that can
15:16
Speaker 3: cause breast ravarian cancer.
15:18
Speaker 1: Yeah. The thing that we.
15:22
Speaker 3: Wanted to see with the model is can it understand
15:25
Speaker 3: whether or not a genetic mutation will cause disease, and
15:28
Speaker 3: so you can feed in, you know, someone if someone
15:31
Speaker 3: comes in and gets a you know, gets their genome sequence,
15:35
Speaker 3: then they want to know. You know, as a woman,
15:37
Speaker 3: does I have a mutation in the broco one gene?
15:40
Speaker 3: Do I need like is am I higher risk of
15:43
Speaker 3: itating breast cancer? Do I need to get a double mistectomy?
15:46
Speaker 3: If I have a causal variant? Many you know patients
15:49
Speaker 3: elect to do that or do I just do an
15:51
Speaker 3: annual mammogram and I just monitor? Right, And most of
15:55
Speaker 3: the time your mutation is different from the ones that
15:58
Speaker 3: are recorded in existing clinical databases, so it becomes classified
16:02
Speaker 3: as a variant of unknown significance, right, And the model
16:06
Speaker 3: has a very good sense of whether not that variant
16:10
Speaker 3: that has never been seen before, you don't know what
16:12
Speaker 3: it does, whether or not it will cause disease or not. Right,
16:15
Speaker 3: And it does that not just for BROCA one, but
16:17
Speaker 3: for any gene.
16:19
Speaker 2: How do you know that? How do you test that
16:21
Speaker 2: sort of out of sample?
16:23
Speaker 3: Well, essentially you benchmark it right against there are these
16:30
Speaker 3: like ground truth databases created by the National Institutes of
16:34
Speaker 3: Health that we've tested against. And then you can also
16:38
Speaker 3: again you can run experiments, so you can for example,
16:41
Speaker 3: generate those variants of known significance in the lab and
16:44
Speaker 3: actually test them in the lab assay and compare them
16:48
Speaker 3: against the causal mutations or the benign mutations and essentially
16:52
Speaker 3: see where they rank in terms of you know, this
16:56
Speaker 3: metric of pathogenicity. So folks have done those experiments and
17:01
Speaker 3: you can use those as benchmark data sets.
17:04
Speaker 2: So I mean that one seems to have kind of
17:07
Speaker 2: immediate clinical relevance, does it? Or am I missing something?
17:10
Speaker 3: We certainly don't think that folks should be using EVO
17:15
Speaker 3: designer on the ARC website, putting in their genetic sequence
17:19
Speaker 3: and then trying to make clinical decisions, right, Sure, but
17:23
Speaker 3: we are making newer versions of the model, which we're
17:27
Speaker 3: training internally that we think can be state of the
17:29
Speaker 3: art at doing this type of genetic diagnostic and so
17:33
Speaker 3: we're excited about improving these models over time. I think
17:36
Speaker 3: in order for them to be used by a doctor
17:41
Speaker 3: in the context of the clinical diagnosis, right, there's a
17:44
Speaker 3: lot of testing and validation and regulatory oversight that will
17:48
Speaker 3: need to happen to make sure that this can be
17:52
Speaker 3: used and certified in the right way. But from a
17:55
Speaker 3: research point of view, right, I think this is extremely
17:57
Speaker 3: exciting given the performance of the model already, let alone
18:01
Speaker 3: the new things that we're doing to it under the hood.
18:07
Speaker 2: We'll lead back in just a minute. Earlier in the conversation,
18:19
Speaker 2: Patrick was talking about the BRACA one gene with that
18:22
Speaker 2: single gene link to breast cancer. But for most diseases
18:27
Speaker 2: with genetic components, there is not one single gene. There
18:30
Speaker 2: are lots of genes, and in lots of cases we
18:32
Speaker 2: still don't really understand what's going on. And so I
18:35
Speaker 2: asked him how that universe of diseases fits into his work.
18:39
Speaker 3: Yeah, so this is some of the fun part, right obviously.
18:44
Speaker 3: You know, so so far we've been talking about individual
18:46
Speaker 3: mutations in one gene, right, Yeah, and you know, in
18:50
Speaker 3: this very kind of well controlled setting, in a very
18:53
Speaker 3: well understed gene for a well reasonably well understood disease.
18:57
Speaker 1: Right.
18:58
Speaker 3: Most of biology is not that simple, right, And so
19:01
Speaker 3: the reason why we have GA studies or polygenic risk
19:05
Speaker 3: scores is because we have complex traits, lots of different
19:08
Speaker 3: mutations that create lots of different phenotypes.
19:11
Speaker 2: Gis genome wide association. Let's look at the whole genome
19:15
Speaker 2: and see what correlates exactly.
19:17
Speaker 3: It's basically the ultimate human genetics fishing experiment, What the
19:21
Speaker 3: hell is the genetic reason why someone has a given trait? Right,
19:26
Speaker 3: And that can be literally any trait, And you get
19:28
Speaker 3: a bunch of people that have that trade a and
19:31
Speaker 3: a bunch of people that seem normal, sequence them and
19:34
Speaker 3: look at the statistical association for what genes seem to
19:37
Speaker 3: be enriched in the people who have the desired or
19:42
Speaker 3: bad trade right, and that could be height, that could
19:45
Speaker 3: be hair thickness, that could be likelihood to get Alzheimer's disease.
19:51
Speaker 2: Yeah, and it seems like that was one that people
19:55
Speaker 2: were very hopeful about, I don't know, ten years ago
19:57
Speaker 2: or something in the kind of post human genome project era.
20:01
Speaker 2: That turned out to be a lot harder than anybody thought.
20:03
Speaker 1: Is that right?
20:04
Speaker 3: Yeah? I think our initial goal with sequencing the human
20:08
Speaker 3: genome was that we would find a whole bunch of
20:10
Speaker 3: drug targets. It turned out we needed the ability to
20:13
Speaker 3: sequence many many more genomes, potentially everybody's genomes, and compute
20:19
Speaker 3: over all of this inscrutable complex data with AI right
20:24
Speaker 3: in order to actually understand how these mutations actually interact
20:27
Speaker 3: and work. Right. And then so that's the thing that
20:31
Speaker 3: we're very excited about with next generation versions of EVA
20:35
Speaker 3: is to try to understand genetic interactions and polygenic traits
20:40
Speaker 3: rather than monogenetic traits that are caused by single gene
20:44
Speaker 3: and a single mutation.
20:46
Speaker 2: Because most diseases, or the diseases that affect most people,
20:49
Speaker 2: certainly have lots of genetic inputs and it's exactly one
20:54
Speaker 2: messed up gene.
20:56
Speaker 3: So in high school you learned that genotype and the
20:59
Speaker 3: environment collaborate to create phenotype, which is a fancy way
21:02
Speaker 3: of saying there were nature and nurture collectively create behavior
21:08
Speaker 3: and all like biology, right, And so this is really
21:12
Speaker 3: where our EGO work, in our virtual cell work starts
21:16
Speaker 3: to collide a model of genotype and a model of
21:22
Speaker 3: cellular response and behavior, and how do we actually connect
21:26
Speaker 3: these two modules in order to make more accurate models
21:30
Speaker 3: of cell biology.
21:31
Speaker 2: So you brought up the virtual cell, which we haven't
21:34
Speaker 2: really talked about, which seems like kind of the other
21:38
Speaker 2: at least another big project at ARC.
21:41
Speaker 1: Right, tell me about the virtual cell.
21:44
Speaker 3: So we talked earlier about GOS, right, and GS is
21:47
Speaker 3: fundamentaliation exactly geno wide association. It's an association, right, there's
21:53
Speaker 3: no causality, right, And so in many ways. The breakthrough
21:58
Speaker 3: of Crisper and genome editing was to basically take mutations
22:03
Speaker 3: that are associated causally creating them and testing them in
22:07
Speaker 3: the lab and living cells and seeing how the cells
22:09
Speaker 3: respond and behave when you make this disease associated mutation,
22:14
Speaker 3: do you get cancer from a normal self?
22:17
Speaker 2: So it's allowing us to say, well, we know there's
22:20
Speaker 2: correlation here, but we don't know it's causative. It allows
22:22
Speaker 2: us to answer that question at least in some content exactly.
22:26
Speaker 3: And these are experiments that you do in the lab.
22:29
Speaker 3: Now the question is how much of this can we
22:31
Speaker 3: push to an AI model, so instead of doing the
22:34
Speaker 3: Chrisper experiment, the model can just tell you.
22:37
Speaker 2: And is this going back to the original problem of
22:40
Speaker 2: like it's wildly slow to have to do these one
22:43
Speaker 2: by one in the lab with Crisper.
22:45
Speaker 3: That's exactly right with Crisper, with a small molecule library
22:49
Speaker 3: or anything else. Right. The point is we want a
22:52
Speaker 3: model that can actually predict, for example, the result of
22:55
Speaker 3: Chrisper experiments or a drug perturbations right and guide the
23:02
Speaker 3: things that people are going to do in the lab.
23:04
Speaker 2: So it's basically saying, if we took a cell and
23:08
Speaker 2: we changed the output of this gene. Or if we
23:11
Speaker 2: took a cell and we used this drug on it,
23:15
Speaker 2: for lack of a better phrase, what would happen?
23:18
Speaker 1: Is that the kind of question?
23:19
Speaker 3: Yeah, the maybe the mental model for me is, imagine
23:22
Speaker 3: you're like a DJ, right, and you have the most
23:25
Speaker 3: complex mixer on the planet, right, and you're in the
23:29
Speaker 3: middle of this Las Vegas club. You have all these
23:32
Speaker 3: knobs and dials, and every time you move them around,
23:35
Speaker 3: the music changes, and you're controlling the music of the cell, right,
23:40
Speaker 3: And so when you turn this up all the way,
23:42
Speaker 3: you know, for some given you know sound, you can
23:45
Speaker 3: get this kind of chaotic super loud disease, like like noise.
23:50
Speaker 3: It doesn't sound mollifluous and interesting, right, And then you
23:53
Speaker 3: want to basically figure out how to turn it off.
23:56
Speaker 3: But if it's actually a bunch of different sounds that
24:00
Speaker 3: are contributing to the noise, right, you need to figure
24:03
Speaker 3: out the fastest way to find out which of these
24:07
Speaker 3: you know, you know, tens of thousands of jobs are
24:09
Speaker 3: out there should be actually changed in the right way.
24:13
Speaker 1: Now.
24:13
Speaker 3: Today we try to find this out by trying to
24:16
Speaker 3: guess which knobs we need to be dialing and then
24:20
Speaker 3: we like just messed around and do it. And this
24:22
Speaker 3: is extremely slow, and it's this comminentorial search that's incredibly slow.
24:28
Speaker 2: It's not one thing. It's not even that you have
24:30
Speaker 2: to guess the one thing. It's like it's many different
24:32
Speaker 2: things have to happen, and so the math of that,
24:35
Speaker 2: if you're doing trial and error, is just crazy and
24:38
Speaker 2: it takes forever.
24:39
Speaker 3: Yeah, it's exclusively combinatorial, and the goal of the model
24:42
Speaker 3: is to be a copilot where it would be the
24:44
Speaker 3: equivalent of your like metaar glasses that tells you it's
24:49
Speaker 3: this knob and this knob and this knob, or it's
24:51
Speaker 3: you should try these five knobs and these five knobs,
24:54
Speaker 3: and then you can, you know, in a much more
24:56
Speaker 3: targeted way, you know which things to dial right. And
25:00
Speaker 3: that's how we could accelerate the experimental verification of the
25:05
Speaker 3: model's predictions in order to try to find the perturbations
25:10
Speaker 3: that could treat disease cells for example. But this of
25:14
Speaker 3: course is a platform capability that can be useful across
25:17
Speaker 3: basic science but also very much for a therapeutic target idea,
25:22
Speaker 3: which is something that we deeply care about accelerating.
25:25
Speaker 2: Yeah, I mean therapeutic target idea is basically like what
25:29
Speaker 2: should we send a drug to go fix?
25:32
Speaker 3: Exactly?
25:32
Speaker 1: Change?
25:33
Speaker 3: Right?
25:33
Speaker 1: Exactly?
25:34
Speaker 2: I mean you've talked about, I think in this context,
25:37
Speaker 2: I've heard you talk about the the failure rate of
25:40
Speaker 2: clinical trials, right, like something like ninety percent of drugs
25:44
Speaker 2: that go into clinical trials, which already is far along
25:47
Speaker 2: in the development process, right, ninety percent of those drugs fail.
25:52
Speaker 1: Like how does how could this help?
25:55
Speaker 3: So we're not very good at making drugs? Right? The
25:57
Speaker 3: statistics are clear, and that really means we're that at
26:02
Speaker 3: two different things, finding the right drug target and then
26:07
Speaker 3: actually having a drug that actually targets it in the
26:10
Speaker 3: right way.
26:11
Speaker 1: Right.
26:11
Speaker 3: So, so, for example, if you found the right drug target,
26:15
Speaker 3: let's say it's like a runaway car and you have
26:18
Speaker 3: an inhibitor, but it doesn't stop it, It just slows
26:22
Speaker 3: it down, right, Yeah, it wouldn't actually be curative and
26:25
Speaker 3: fix the problem.
26:26
Speaker 1: Right.
26:26
Speaker 3: So you need the right drug composition and you need
26:30
Speaker 3: the right drug target, and because we seem to be
26:33
Speaker 3: bad at both of these things, we get a ninety
26:36
Speaker 3: percent failure rate.
26:38
Speaker 1: Why did you decide to focus on Alzheimer's.
26:41
Speaker 3: So Alzheimer's is one of the major killers no one
26:45
Speaker 3: has agreed on or we certainly don't know what the
26:48
Speaker 3: right drug target is for Alzheimer's disease. And we think
26:51
Speaker 3: of it as a textbook example of a complex human
26:54
Speaker 3: disease that has a bunch of genetic mutations that we
26:57
Speaker 3: don't understand. We don't know what genes they control, what
27:00
Speaker 3: pathways they control, so we just don't know what the
27:04
Speaker 3: hell is happening, right, And it's a you know, you know,
27:07
Speaker 3: they're they're also inputs from the environment, infection age obviously,
27:12
Speaker 3: you know, potentially diabetes and others that increase your risk.
27:17
Speaker 3: So it's this really interesting example of having a risk
27:21
Speaker 3: conferring genetic state penetr and then pushed over the edge
27:26
Speaker 3: by environmental perturbations in order to create dementia. Right, And
27:32
Speaker 3: we think, first of all, it would be very impactful
27:35
Speaker 3: to human health if we can solve it, and second,
27:38
Speaker 3: we can use it as an example as a blueprint
27:41
Speaker 3: for how we can cure other complex diseases if we
27:44
Speaker 3: can make progress. And so our goal has been to
27:47
Speaker 3: try to figure out ways that we can model this
27:51
Speaker 3: both experimentally at scale and computationally with model like AI
27:57
Speaker 3: models of the major cell types in the brain in
28:00
Speaker 3: order to try to you know, model these different combinatorial
28:05
Speaker 3: changes over time much more effectively, basically much faster. Yeah,
28:11
Speaker 3: and that's sort of one of the ways that we're
28:13
Speaker 3: going to actually internally test the output of our virtual
28:17
Speaker 3: cell models. But of course, any other lab will be
28:19
Speaker 3: able to take these models and test them for some
28:23
Speaker 3: heart disease pathway, let's say, or some inflammation pathway that
28:27
Speaker 3: might be involved in glaucoma and eye disease, or you know,
28:31
Speaker 3: whatever basic science mechanism that they might care about.
28:34
Speaker 1: So that's the.
28:34
Speaker 3: Scientific reason why we care about Alzheimer's. On a personal note,
28:39
Speaker 3: it's also why I became a scientist. My my grandfather,
28:43
Speaker 3: like many other people, you know, got Alzheimer's when I
28:47
Speaker 3: was eleven. He was living with us at the time,
28:50
Speaker 3: and I watched him go through, you know, from mild
28:55
Speaker 3: cognitive impairment and confusion to you know, flipping shopping carts
29:00
Speaker 3: in costco to you know, being in a nursing home
29:03
Speaker 3: and dying. And you know, I think that gave me
29:07
Speaker 3: very clear early miss and I think really in many
29:10
Speaker 3: ways redirected my life from probably becoming some you know,
29:14
Speaker 3: software founder of some kind too, you know, joining a
29:17
Speaker 3: lot in high school and picking up a pipet and
29:21
Speaker 3: doing that gusts and check search that we're now trying automate.
29:27
Speaker 1: We'll be back in a minute with the lightning round.
29:31
Speaker 2: Mm hmm, okay, let's finish with the lightning round. Who
29:42
Speaker 2: is the most underrated medieval king?
29:45
Speaker 3: The most medieval See? I feel like this is basically
29:51
Speaker 3: well there, there's there's there's so many that you could
29:55
Speaker 3: that you could pick from. I'm going to pass on
29:59
Speaker 3: this question. This is this is very this is very
30:02
Speaker 3: unfair of me. But I mean I'll have to ponder.
30:06
Speaker 3: I'll have to ponder.
30:07
Speaker 1: Yeah, why do you love medieval history?
30:09
Speaker 3: I studied it for four years as a as a
30:12
Speaker 3: young homeschooled student.
30:14
Speaker 1: Yeah, why did you study it for four years?
30:17
Speaker 3: There's a reason why we love like swords and sorcerers
30:21
Speaker 3: and you know, in Game of Thrones and the Battle
30:25
Speaker 3: of Aquitaine and whatever. You know, it's I think it's
30:30
Speaker 3: one of the periods of time that we felt like
30:32
Speaker 3: true optimism.
30:34
Speaker 1: Do you think so? I don't think. I mean, you
30:36
Speaker 1: know more about it than I do. But I'm shocked.
30:39
Speaker 1: Is that not just a projection.
30:40
Speaker 3: It's evolutionary selection playing out in real time. It's very
30:45
Speaker 3: like visible and well recorded history of evolutionary struggle. I
30:51
Speaker 3: think maybe that's why I found deeply fascinating about it.
30:54
Speaker 1: Uh huh.
30:55
Speaker 2: That's the biologist view I guess of the medieval period.
30:59
Speaker 3: Maybe one way to frame it.
31:01
Speaker 1: If you weren't working on biology, what would you be
31:03
Speaker 1: working on?
31:04
Speaker 3: You know, I'd probably be a music talent scout.
31:07
Speaker 2: Yeah, what should I listen to? Who should I sign to?
31:10
Speaker 1: Our label?
31:12
Speaker 3: My favorite band it is the XX I discovered them
31:16
Speaker 3: in two thousand and nine or two thousand and eight, yeah, early, Yeah,
31:21
Speaker 3: It's like, you know, if I were a seed investor,
31:24
Speaker 3: this would be this would be one of my kind
31:26
Speaker 3: of key seed investments.
31:28
Speaker 2: I'm going to read you a thing you wrote and
31:29
Speaker 2: then ask you a question about it you wrote. In
31:33
Speaker 2: my view, researchers also need to stoke two warring urges.
31:37
Speaker 2: One an opinionated sense of taste in a relentless search
31:40
Speaker 2: for beauty, and two a yeoman's grindset for the unglamorous,
31:44
Speaker 2: dirty work required to get things done, which is lovely.
31:48
Speaker 2: A lovely sentence, And I'm curious for your thoughts about
31:54
Speaker 2: that in the context of AI. Basically, like you think
31:59
Speaker 2: AI can do two?
32:00
Speaker 1: Number two? Do you think it can do one?
32:02
Speaker 3: Like?
32:02
Speaker 1: Where where does AI fit in the context of that sentence?
32:06
Speaker 3: Early on, you know, maybe a year year and a
32:08
Speaker 3: half ago, you know, I think there was lots of discussion.
32:11
Speaker 3: You know, I want AI to you know, do my
32:14
Speaker 3: laundry and taxes so I can make art and music
32:18
Speaker 3: and not the other way around, right, because you know,
32:21
Speaker 3: early on, we we had all these models that were
32:24
Speaker 3: you know, generating new images and making AI art that
32:27
Speaker 3: was really controversial, and making new types of songs. And
32:29
Speaker 3: you're like, wellhy can't I do the basic things so
32:31
Speaker 3: I can live and create and dream? And I think
32:35
Speaker 3: today it's very clear that it is helping us with both, right,
32:40
Speaker 3: that these bicycles for the mind are high quality enough
32:43
Speaker 3: to allow us to create not just intellectual insights, but
32:47
Speaker 3: new like you know, visual design styles and paradigms, new
32:51
Speaker 3: types of sound. And we're just very much in the
32:54
Speaker 3: early innings of using these in a way that feel
32:57
Speaker 3: augmentive to human creative capacity rather than you know, some
33:02
Speaker 3: somehow like aggressive or replacing. You know, that's and I'm
33:06
Speaker 3: a big AI Bowl. But above all, I'm a big
33:10
Speaker 3: believer in how humans can use them to benefit.
33:19
Speaker 2: Patrick Sue is the co founder of the Art Institute
33:22
Speaker 2: and Assistant Professor of Bioengineering at UC Berkeley.
33:27
Speaker 1: Please email us at problem at pushkin dot fm. We
33:30
Speaker 1: are always looking for new guests for the show.
33:33
Speaker 2: Today's show was produced by Trinamanino and Gabriel Hunter Chang.
33:38
Speaker 2: It was edited by Alexander Garretson and engineered by Sarah Bruguer.
33:42
Speaker 2: I'm Jacob Goldstein and we'll be back next week with
33:44
Speaker 2: another episode of What's Your Pop