1
00:00:15,356 --> 00:00:15,796
Speaker 1: Pushkin.

2
00:00:20,516 --> 00:00:23,876
Speaker 2: So you know Moore's law, right, the idea that computers

3
00:00:23,916 --> 00:00:28,316
Speaker 2: are specifically chips, get better and cheaper at an exponential rate.

4
00:00:28,716 --> 00:00:32,596
Speaker 2: People who work on researching and developing new drugs, they

5
00:00:32,636 --> 00:00:36,436
Speaker 2: talk about e Room's law. E room is more spelled backwards.

6
00:00:36,716 --> 00:00:38,956
Speaker 2: And this is sort of a half joke, half not

7
00:00:39,036 --> 00:00:41,876
Speaker 2: a joke, way of pointing out that developing new drugs

8
00:00:41,876 --> 00:00:45,436
Speaker 2: has gotten slower and more expensive over the past several decades.

9
00:00:45,836 --> 00:00:50,796
Speaker 2: It has gone in the opposite direction of developing new microchips. Today,

10
00:00:51,156 --> 00:00:54,596
Speaker 2: after all the money and technology and hard work and

11
00:00:54,676 --> 00:00:59,516
Speaker 2: intelligence poured into drug development, something like ninety percent of

12
00:00:59,596 --> 00:01:04,516
Speaker 2: drugs that go into clinical trials fail, only ten percent succeed.

13
00:01:05,476 --> 00:01:06,956
Speaker 1: That means that if we got to.

14
00:01:06,916 --> 00:01:10,716
Speaker 2: A place where even just twenty percent succeeded, still an

15
00:01:10,756 --> 00:01:20,876
Speaker 2: eighty percent failure rate, we would double the rate of progress.

16
00:01:21,156 --> 00:01:23,756
Speaker 2: I'm Jacob Goldstein, and this is what's your problem. My

17
00:01:23,836 --> 00:01:27,836
Speaker 2: guest today is Patrick Sue. Patrick got a PhD from

18
00:01:27,836 --> 00:01:31,596
Speaker 2: Harvard in biochemistry when he was twenty one years old,

19
00:01:31,916 --> 00:01:35,196
Speaker 2: and then four years ago, before he turned thirty, he

20
00:01:35,276 --> 00:01:39,196
Speaker 2: founded the ARC Institute. It's a nonprofit research center in

21
00:01:39,236 --> 00:01:42,476
Speaker 2: the Bay Area that hosts scientists from Stanford, u SE,

22
00:01:42,556 --> 00:01:45,556
Speaker 2: San Francisco, and UC Berkeley, where.

23
00:01:45,316 --> 00:01:46,556
Speaker 1: Patrick is on the faculty.

24
00:01:47,316 --> 00:01:51,276
Speaker 2: Patrick's problem is this, how can you use AI to

25
00:01:51,356 --> 00:01:55,916
Speaker 2: make biological research more efficient, to guide scientists more quickly

26
00:01:56,236 --> 00:01:57,636
Speaker 2: to discoveries that'll lead.

27
00:01:57,516 --> 00:01:59,316
Speaker 1: To new and better truemils.

28
00:01:59,836 --> 00:02:03,116
Speaker 2: As you'll hear later in the conversation, Patrick is particularly

29
00:02:03,116 --> 00:02:07,236
Speaker 2: focused on Alzheimer's disease. To start, he told me about

30
00:02:07,236 --> 00:02:10,556
Speaker 2: why basic biological research is still so slow.

31
00:02:11,076 --> 00:02:15,076
Speaker 3: So if you look at a biology research lab in

32
00:02:15,156 --> 00:02:17,956
Speaker 3: the eighties or the nineties, the early two thousands, of

33
00:02:17,956 --> 00:02:20,836
Speaker 3: the early twenty tens, or today in the twenty twenties,

34
00:02:21,436 --> 00:02:23,316
Speaker 3: they look basically the same.

35
00:02:23,716 --> 00:02:23,956
Speaker 1: Right.

36
00:02:24,116 --> 00:02:26,596
Speaker 3: You have these long benches, you have these two or

37
00:02:26,636 --> 00:02:31,436
Speaker 3: three rows of shelves, these micropipe pets, and various machines

38
00:02:31,476 --> 00:02:36,036
Speaker 3: that look like home kitchen equipment. Right, and so graduate

39
00:02:36,076 --> 00:02:40,276
Speaker 3: students and postdocs or bench scientists generally are like the

40
00:02:40,316 --> 00:02:45,036
Speaker 3: line chefs right inside of you know, a Michelin star kitchen.

41
00:02:45,276 --> 00:02:51,036
Speaker 3: You're prepping the vegetables. You know, you're making the sort

42
00:02:51,076 --> 00:02:54,316
Speaker 3: of initial sauces. Right, So you might be taking tissues

43
00:02:54,356 --> 00:02:59,076
Speaker 3: down or cells, processing them, staining them with antibodies. I

44
00:02:59,116 --> 00:03:04,716
Speaker 3: think the point is that doing these experiments is extremely slow,

45
00:03:05,476 --> 00:03:08,796
Speaker 3: very manual, and requires a huge amount of soft knowledge

46
00:03:08,836 --> 00:03:11,676
Speaker 3: and of how and is really variable lab to lab.

47
00:03:11,956 --> 00:03:14,996
Speaker 3: So we have this multi stack problem where we have

48
00:03:15,836 --> 00:03:21,196
Speaker 3: multi step search. Right, the experiments take months to years

49
00:03:21,556 --> 00:03:26,636
Speaker 3: to run out, and we basically search across the entire

50
00:03:26,716 --> 00:03:30,316
Speaker 3: space in an extremely manual way. And so in the

51
00:03:30,356 --> 00:03:33,116
Speaker 3: modern era of AI, where we believe we can predict

52
00:03:33,116 --> 00:03:36,236
Speaker 3: things across many different fields, the question is can we

53
00:03:36,396 --> 00:03:40,316
Speaker 3: speed up biology by being able to have models that

54
00:03:40,396 --> 00:03:44,036
Speaker 3: actually have predictive value and power. And our goal is

55
00:03:44,076 --> 00:03:49,476
Speaker 3: to be able to create biological intelligence that starts to

56
00:03:49,676 --> 00:03:53,476
Speaker 3: crack that threshold of a cell. Biologist will use the

57
00:03:53,596 --> 00:03:56,916
Speaker 3: model to rank the top twelve things that they'll do

58
00:03:56,996 --> 00:03:59,916
Speaker 3: in the lab, rather than just going into lab and

59
00:03:59,956 --> 00:04:03,156
Speaker 3: trying a bunch of things based on hypothsis driven science.

60
00:04:03,356 --> 00:04:07,756
Speaker 3: But you actually do model first prediction and then a

61
00:04:07,836 --> 00:04:09,076
Speaker 3: lab in the loop experiment.

62
00:04:10,156 --> 00:04:12,396
Speaker 2: And I mean we'll get to the sort of things

63
00:04:12,436 --> 00:04:15,356
Speaker 2: that have to happen for that to happen. But if

64
00:04:15,396 --> 00:04:17,796
Speaker 2: we could get it to work, right, if you could

65
00:04:17,796 --> 00:04:20,836
Speaker 2: get it to work, what's the payoff beyond the level.

66
00:04:20,556 --> 00:04:23,436
Speaker 3: I think the first thing is the practice of how

67
00:04:23,436 --> 00:04:27,036
Speaker 3: you design and do experiments will completely change. Right. And

68
00:04:27,076 --> 00:04:29,996
Speaker 3: so just like you know, graduate students today, whenever they

69
00:04:29,996 --> 00:04:32,516
Speaker 3: want to research a new area, they just use chatch

70
00:04:32,556 --> 00:04:35,716
Speaker 3: to BT, right, or Claude or Gemini or whatever your

71
00:04:35,716 --> 00:04:38,796
Speaker 3: favorite model is. Right, that has just become a fundamental

72
00:04:38,836 --> 00:04:41,316
Speaker 3: part of the workflow. In fact, I don't think most

73
00:04:41,316 --> 00:04:45,556
Speaker 3: people read papers the old fashioned way anymore, right. And

74
00:04:45,596 --> 00:04:47,956
Speaker 3: it used to be, oh, I read them online in

75
00:04:47,956 --> 00:04:50,756
Speaker 3: my browser instead of in a magazine like Nature and Science.

76
00:04:50,796 --> 00:04:53,796
Speaker 3: And now it's just I throw the pdf into chatch

77
00:04:53,876 --> 00:04:56,556
Speaker 3: BT and I read the summary. Right. That's just a

78
00:04:56,556 --> 00:04:59,036
Speaker 3: completely different workflow than just even a few years ago.

79
00:04:59,156 --> 00:05:02,436
Speaker 3: And I think instead of designing experiments by hand, we're

80
00:05:02,436 --> 00:05:05,596
Speaker 3: gonna have AI models design the experiments for us and

81
00:05:05,716 --> 00:05:07,996
Speaker 3: help us troubleshoot and decide what to do next.

82
00:05:08,436 --> 00:05:11,276
Speaker 2: I've heard you describe the way biology works now as

83
00:05:11,476 --> 00:05:12,676
Speaker 2: guess and check.

84
00:05:14,516 --> 00:05:15,116
Speaker 1: Which I liked.

85
00:05:15,116 --> 00:05:17,236
Speaker 2: Which I guess on one level is kind of the

86
00:05:17,276 --> 00:05:20,916
Speaker 2: scientific method, right, It's kind of like hypothesis and experiment.

87
00:05:21,156 --> 00:05:25,996
Speaker 2: But the guess really leans into the not quite blind,

88
00:05:26,036 --> 00:05:29,276
Speaker 2: but the uncertainty, all the blind alleys that I guess

89
00:05:29,436 --> 00:05:30,716
Speaker 2: you go down exactly.

90
00:05:31,356 --> 00:05:33,276
Speaker 3: You know, it's a lot like that that a childhood

91
00:05:33,316 --> 00:05:37,436
Speaker 3: game battleship where you're trying to sink someone's battleship. You

92
00:05:37,436 --> 00:05:39,436
Speaker 3: don't quite know where it is, and so you're searching

93
00:05:39,516 --> 00:05:44,316
Speaker 3: these different quadrants or like mind sweeper, Right, it's just

94
00:05:44,596 --> 00:05:47,796
Speaker 3: very hard to know that you're searching in the right place.

95
00:05:48,116 --> 00:05:49,836
Speaker 3: And so the first thing is can these A models

96
00:05:49,876 --> 00:05:52,916
Speaker 3: help us search in the right place so that when

97
00:05:52,916 --> 00:05:56,716
Speaker 3: we're peppering things with these individual manual wet lab experiments,

98
00:05:56,796 --> 00:06:00,436
Speaker 3: are hit rate can just significantly go up, whether that's

99
00:06:00,436 --> 00:06:04,156
Speaker 3: for designing a drud or you know, predicting how cell

100
00:06:04,316 --> 00:06:07,316
Speaker 3: respond to you know, some perturbation.

101
00:06:08,116 --> 00:06:11,356
Speaker 2: Yes, and then that's kind of the first order thing.

102
00:06:11,396 --> 00:06:15,716
Speaker 2: So it's basically making basic research more efficient speed. Yeah,

103
00:06:15,756 --> 00:06:19,156
Speaker 2: but yeah, one are the second and third order things

104
00:06:19,156 --> 00:06:22,916
Speaker 2: like how does that translate to clinical outcomes?

105
00:06:23,236 --> 00:06:24,956
Speaker 3: I think they're directly linked, right.

106
00:06:25,156 --> 00:06:25,676
Speaker 1: Yeah.

107
00:06:26,396 --> 00:06:30,196
Speaker 3: In drug discovery, we do a thing called target identification.

108
00:06:30,436 --> 00:06:30,596
Speaker 1: Right.

109
00:06:30,836 --> 00:06:33,516
Speaker 3: You need to be able to find the right drug target,

110
00:06:33,756 --> 00:06:35,716
Speaker 3: and you need to be able to perturb it in

111
00:06:35,716 --> 00:06:38,356
Speaker 3: the right way. You need to find the thing that's

112
00:06:38,396 --> 00:06:41,996
Speaker 3: going wrong, the toxic protein that's causing a disease need

113
00:06:42,036 --> 00:06:44,756
Speaker 3: try to turn it off. Right, So the goals really

114
00:06:44,796 --> 00:06:47,916
Speaker 3: been to just finding right the right drug target and

115
00:06:47,916 --> 00:06:50,236
Speaker 3: then drugging in the right way. But the problem is

116
00:06:50,476 --> 00:06:53,236
Speaker 3: we don't seem to find the right drug target and

117
00:06:53,356 --> 00:06:57,836
Speaker 3: even know what it is for most complex diseases, Alzheimer's disease,

118
00:06:57,916 --> 00:07:01,636
Speaker 3: many different types of cancer, it's aging, autoimmunity, a lot

119
00:07:01,636 --> 00:07:05,836
Speaker 3: of the major killers, it's still the same. Right, we

120
00:07:05,876 --> 00:07:08,636
Speaker 3: don't know what to actually target. We think these models

121
00:07:08,636 --> 00:07:09,516
Speaker 3: could help us do that.

122
00:07:09,996 --> 00:07:12,276
Speaker 2: Let's talk about the ARC Institute, right, which is where

123
00:07:12,916 --> 00:07:14,876
Speaker 2: a lot of this work is happening. A thing you've

124
00:07:14,916 --> 00:07:17,436
Speaker 2: started and you run, like, tell me about the ARK

125
00:07:17,516 --> 00:07:20,316
Speaker 2: Institute and like specifically why you started it.

126
00:07:20,676 --> 00:07:24,676
Speaker 3: ARC is about four years old today, we're about three

127
00:07:24,756 --> 00:07:28,556
Speaker 3: hundred and fifty people, and we're a full stack AI

128
00:07:28,636 --> 00:07:31,956
Speaker 3: and biology research organization. And so by that, I mean

129
00:07:32,316 --> 00:07:38,036
Speaker 3: we try to natively combine experimental science and computational science

130
00:07:38,556 --> 00:07:41,196
Speaker 3: under a single physical roof so that we can do

131
00:07:41,436 --> 00:07:45,756
Speaker 3: native iterative lab in the loop between AI models, both

132
00:07:46,236 --> 00:07:51,836
Speaker 3: training and running them and then actually experimentally validating and

133
00:07:51,956 --> 00:07:55,236
Speaker 3: verifying things in the lab in order to create new

134
00:07:55,276 --> 00:08:00,716
Speaker 3: mechanistic insights, new drug compositions, and identify new therapeutic targets.

135
00:08:01,036 --> 00:08:05,596
Speaker 3: We really have two major goals. The first is to

136
00:08:05,796 --> 00:08:09,076
Speaker 3: create these virtual cell models that can simulate human biology.

137
00:08:09,196 --> 00:08:12,756
Speaker 3: You have foundation models. The second is secure Alzhemer's disease.

138
00:08:13,276 --> 00:08:17,076
Speaker 3: We're interested in the view of ARC as a sort

139
00:08:17,116 --> 00:08:20,996
Speaker 3: of Edison Shop right where you're inventing a whole bunch

140
00:08:20,996 --> 00:08:24,396
Speaker 3: of different things that we can actually productize. And so

141
00:08:24,636 --> 00:08:27,916
Speaker 3: in a way, our ambition from ARC has been increasing

142
00:08:28,156 --> 00:08:32,316
Speaker 3: beyond how do we do breakthrough academic and basic science

143
00:08:32,556 --> 00:08:35,516
Speaker 3: to how do we actually productize this science in a

144
00:08:35,556 --> 00:08:38,316
Speaker 3: way that we can impact human health not just with

145
00:08:38,756 --> 00:08:41,836
Speaker 3: research papers, but with things that people can actually take,

146
00:08:41,916 --> 00:08:42,996
Speaker 3: see and feel and use.

147
00:08:43,556 --> 00:08:45,676
Speaker 2: When you talk about the Edison Shop, I mean obviously

148
00:08:45,716 --> 00:08:49,076
Speaker 2: the Edison Shop was a business right, not a philanthropy,

149
00:08:49,596 --> 00:08:53,116
Speaker 2: like are you how are you thinking about it commercially?

150
00:08:53,996 --> 00:08:56,716
Speaker 3: So ARC is a nonprofit right, and so you know,

151
00:08:56,956 --> 00:09:00,116
Speaker 3: I ARC will the ARC Institute will always remain this

152
00:09:00,236 --> 00:09:04,596
Speaker 3: nonprofit discovery oriented mothership, but we also will certainly have

153
00:09:04,676 --> 00:09:08,116
Speaker 3: the capacity and the interest to be able to spin

154
00:09:08,196 --> 00:09:11,676
Speaker 3: out entities that can take on more focused commercial capital

155
00:09:11,916 --> 00:09:13,596
Speaker 3: in order to productize things.

156
00:09:14,076 --> 00:09:17,756
Speaker 2: Let's go back to the machine learning slash AI work

157
00:09:17,796 --> 00:09:18,236
Speaker 2: you're doing.

158
00:09:20,196 --> 00:09:22,476
Speaker 1: I think we should talk about EVO. Tell me about EVO.

159
00:09:23,036 --> 00:09:28,076
Speaker 3: EVO is essentially a chat GPT, but only with DNA, right, okay,

160
00:09:28,436 --> 00:09:32,556
Speaker 3: And so it's DNA in and DNA out right, So

161
00:09:32,636 --> 00:09:35,916
Speaker 3: you can talk to it by typing in nucleotides and

162
00:09:35,956 --> 00:09:41,356
Speaker 3: you'll receive from the model some corresponding set of nucleotides.

163
00:09:41,636 --> 00:09:43,836
Speaker 3: So just like I could say to be or not

164
00:09:43,956 --> 00:09:47,316
Speaker 3: to the model would tell me, you know, you know,

165
00:09:47,356 --> 00:09:49,036
Speaker 3: to be or not to be? That is the question,

166
00:09:49,356 --> 00:09:51,196
Speaker 3: you know, and then it will keep going and the

167
00:09:51,196 --> 00:09:54,356
Speaker 3: rest of hamlets delilokey. Right. If I gave it a

168
00:09:54,436 --> 00:10:00,516
Speaker 3: fragment of let's say, a mitochondrial genome or you know,

169
00:10:01,636 --> 00:10:05,836
Speaker 3: e Coli genome, it will then try to autocomplete the

170
00:10:05,956 --> 00:10:08,996
Speaker 3: rest of that fragment. It's not just memorizing in regard

171
00:10:09,116 --> 00:10:13,076
Speaker 3: diating information from the training data, which is what's really important, right.

172
00:10:13,356 --> 00:10:18,476
Speaker 3: It's doing a semantic diversification in a way of what

173
00:10:18,716 --> 00:10:20,956
Speaker 3: it thinks is in the meaning of what it needs

174
00:10:20,996 --> 00:10:21,356
Speaker 3: to make.

175
00:10:21,556 --> 00:10:21,756
Speaker 2: Right.

176
00:10:21,996 --> 00:10:23,396
Speaker 1: Yeah, that's the wild part.

177
00:10:23,476 --> 00:10:23,596
Speaker 3: Right.

178
00:10:23,636 --> 00:10:26,716
Speaker 2: If it was just regurgitating training data, it wouldn't be AI.

179
00:10:26,836 --> 00:10:29,196
Speaker 1: It would just be sort of a library, right exactly.

180
00:10:29,276 --> 00:10:31,156
Speaker 3: And this is why you can ask chatch bt to

181
00:10:31,716 --> 00:10:33,916
Speaker 3: you know, here's what I want to see in my email,

182
00:10:33,916 --> 00:10:35,956
Speaker 3: write me the email, and if you gave it that

183
00:10:36,036 --> 00:10:39,276
Speaker 3: thing ten different times, it would write you similar but

184
00:10:39,516 --> 00:10:40,476
Speaker 3: different emails.

185
00:10:40,716 --> 00:10:41,236
Speaker 1: Right yeah.

186
00:10:41,276 --> 00:10:44,076
Speaker 3: And just similarly, you can make different crisper systems that

187
00:10:44,156 --> 00:10:48,116
Speaker 3: are similar but not the same, different mitochondrial genomes that

188
00:10:48,156 --> 00:10:51,676
Speaker 3: are similar but not the same. And you might say, okay,

189
00:10:51,716 --> 00:10:53,876
Speaker 3: well that's fun. Now you have a bunch of sequences

190
00:10:53,876 --> 00:10:57,556
Speaker 3: in the computer. What's interesting about that? The point is

191
00:10:57,596 --> 00:11:01,836
Speaker 3: that we can then chemically synthesize those DNA strands in

192
00:11:01,876 --> 00:11:06,196
Speaker 3: the lab and test them and see what their function is. Now,

193
00:11:06,356 --> 00:11:10,076
Speaker 3: we could then take those the numbers that we've measured

194
00:11:10,076 --> 00:11:12,116
Speaker 3: in the lab and feed it back into the model

195
00:11:12,716 --> 00:11:15,836
Speaker 3: and then tell the model for things like make it

196
00:11:15,876 --> 00:11:19,996
Speaker 3: better at this task that I've experimentally verified, and then

197
00:11:20,036 --> 00:11:22,436
Speaker 3: you can This is the closing the lab and the

198
00:11:22,436 --> 00:11:25,316
Speaker 3: loop that we've built ARC around in order to he'll

199
00:11:25,396 --> 00:11:29,156
Speaker 3: climb on. Really any experimental accut that you care about.

200
00:11:29,436 --> 00:11:32,556
Speaker 2: I know there was a project that one of your

201
00:11:32,596 --> 00:11:37,276
Speaker 2: colleagues did last year using evo to to basically to

202
00:11:37,396 --> 00:11:40,396
Speaker 2: design a new version of an existing phage. A phage

203
00:11:40,436 --> 00:11:44,556
Speaker 2: is a virus that infects bacteria. So, like, tell me

204
00:11:44,596 --> 00:11:48,316
Speaker 2: about the project with the fix phage and why it's meaningful.

205
00:11:48,596 --> 00:11:53,796
Speaker 3: So we have these databases where whenever people published papers,

206
00:11:54,636 --> 00:11:58,676
Speaker 3: they deposit the genetic sequences into this database. So of

207
00:11:58,716 --> 00:12:02,596
Speaker 3: all fix and FIX related phages, we have a sense

208
00:12:02,676 --> 00:12:06,476
Speaker 3: of the evolutionary distribution that is out there that we've

209
00:12:06,516 --> 00:12:09,876
Speaker 3: been able to detect so far as scientists, right, Yeah,

210
00:12:09,916 --> 00:12:13,956
Speaker 3: And it turns out evo is able to make new

211
00:12:14,116 --> 00:12:19,116
Speaker 3: versions of fi X that are as evolutionarily distinct as

212
00:12:19,276 --> 00:12:22,676
Speaker 3: everything else that we've seen. So we're not just copying

213
00:12:23,036 --> 00:12:26,516
Speaker 3: and making things that look like a Corolla or another

214
00:12:26,596 --> 00:12:29,436
Speaker 3: a chord That would be very boring, but we can make,

215
00:12:29,636 --> 00:12:33,716
Speaker 3: you know, a fundamentally new type of car. Right, Yeah,

216
00:12:33,876 --> 00:12:37,436
Speaker 3: it's still a car, but it's a completely new type

217
00:12:37,436 --> 00:12:40,356
Speaker 3: of design. Let's say, something like as unique as a

218
00:12:40,356 --> 00:12:42,156
Speaker 3: pet cruiser, but much much cooler.

219
00:12:42,356 --> 00:12:46,636
Speaker 2: Okay, And so so let's talk about the meaning of

220
00:12:46,676 --> 00:12:49,716
Speaker 2: this as a research tool and then the clinical implications, right, Like,

221
00:12:50,196 --> 00:12:52,596
Speaker 2: why is it meaningful as a research tool? Why is

222
00:12:52,636 --> 00:12:56,356
Speaker 2: this a meaningful proof of concept.

223
00:12:56,676 --> 00:13:00,156
Speaker 3: The first thing is that there they work in the lab,

224
00:13:00,476 --> 00:13:04,076
Speaker 3: right that when you actually create these aid designed sequences,

225
00:13:04,316 --> 00:13:08,356
Speaker 3: you can actually package real phage particles and they can

226
00:13:08,396 --> 00:13:13,156
Speaker 3: actually in effect living bacteria in the lab the way

227
00:13:13,196 --> 00:13:16,276
Speaker 3: that a normal one would. Now that's sort of step one.

228
00:13:16,316 --> 00:13:20,036
Speaker 2: Just to be clear, this is a thing, a virus

229
00:13:20,116 --> 00:13:24,756
Speaker 2: that has never existed before exactly that the AI invented exactly,

230
00:13:24,796 --> 00:13:27,476
Speaker 2: just worked. The machine made it up and it worked,

231
00:13:27,476 --> 00:13:30,476
Speaker 2: and it was quasi alive because it's a virus.

232
00:13:30,316 --> 00:13:34,716
Speaker 3: Yeah, exactly, And so that that alone is pretty cool.

233
00:13:35,396 --> 00:13:38,716
Speaker 3: The second thing is that we could steer the generation

234
00:13:39,196 --> 00:13:42,516
Speaker 3: and so you can use it to infect specific strains

235
00:13:43,076 --> 00:13:45,996
Speaker 3: of E. Coli that you cared about and not others

236
00:13:46,036 --> 00:13:49,116
Speaker 3: that you didn't. And so you know, in phage therapy

237
00:13:49,156 --> 00:13:51,836
Speaker 3: for example, this is the therapeutic implication of this. Right,

238
00:13:52,076 --> 00:13:55,516
Speaker 3: if you were to you know, target the gut, you

239
00:13:55,556 --> 00:13:58,916
Speaker 3: would want to target specific microbes for example, that are

240
00:13:58,956 --> 00:14:02,036
Speaker 3: like bad gut bacteria and not others that are good.

241
00:14:02,356 --> 00:14:04,556
Speaker 3: And so you don't want to just do a broad

242
00:14:04,676 --> 00:14:09,316
Speaker 3: spectrum wiping out of you know, the entire complex gut community,

243
00:14:09,436 --> 00:14:12,316
Speaker 3: which is what broad spectrum antibiotics do. But you could do,

244
00:14:12,996 --> 00:14:17,596
Speaker 3: in principle, much more selective deletion of something that you want.

245
00:14:17,756 --> 00:14:17,956
Speaker 2: Right.

246
00:14:18,516 --> 00:14:21,596
Speaker 3: So, from a basic science point of view, the controllability

247
00:14:21,956 --> 00:14:26,596
Speaker 3: of generation from the model can lead to like important

248
00:14:26,756 --> 00:14:30,116
Speaker 3: precise functional outcomes in the lab. And then this has

249
00:14:30,156 --> 00:14:32,556
Speaker 3: I think, you know, as I just explained really interesting

250
00:14:32,796 --> 00:14:37,396
Speaker 3: therapeutic implications for phage therapy or a targeted therapy, a

251
00:14:37,516 --> 00:14:43,796
Speaker 3: targeted therapy, right, and in a way both medicines, Well,

252
00:14:44,716 --> 00:14:47,996
Speaker 3: we care a lot about precision and selectivity, right, yeah,

253
00:14:48,036 --> 00:14:51,756
Speaker 3: and that's really the entire game of on target therapeutic

254
00:14:51,836 --> 00:14:54,876
Speaker 3: efficacy and side effect profiles. Right.

255
00:14:55,156 --> 00:14:57,756
Speaker 2: Right, You wanted to do one thing really well, and

256
00:14:57,796 --> 00:14:59,196
Speaker 2: you wanted to do nothing else.

257
00:14:59,676 --> 00:15:01,716
Speaker 3: You want it to be safe. You want it to

258
00:15:01,756 --> 00:15:05,756
Speaker 3: be safe. Yeah, yeah, and effective, exactly safe and effective.

259
00:15:05,996 --> 00:15:08,356
Speaker 2: Tell me about the work you're doing with EVO and

260
00:15:08,396 --> 00:15:12,036
Speaker 2: the BRACA one gene, This gene that's implicated in some

261
00:15:12,116 --> 00:15:13,276
Speaker 2: cases of breast cancer.

262
00:15:13,676 --> 00:15:16,076
Speaker 3: You know, Broca one is an important gene that can

263
00:15:16,156 --> 00:15:17,876
Speaker 3: cause breast ravarian cancer.

264
00:15:18,476 --> 00:15:22,236
Speaker 1: Yeah. The thing that we.

265
00:15:22,156 --> 00:15:25,076
Speaker 3: Wanted to see with the model is can it understand

266
00:15:25,116 --> 00:15:28,196
Speaker 3: whether or not a genetic mutation will cause disease, and

267
00:15:28,276 --> 00:15:31,836
Speaker 3: so you can feed in, you know, someone if someone

268
00:15:31,876 --> 00:15:35,076
Speaker 3: comes in and gets a you know, gets their genome sequence,

269
00:15:35,116 --> 00:15:36,916
Speaker 3: then they want to know. You know, as a woman,

270
00:15:37,396 --> 00:15:40,036
Speaker 3: does I have a mutation in the broco one gene?

271
00:15:40,756 --> 00:15:43,516
Speaker 3: Do I need like is am I higher risk of

272
00:15:43,556 --> 00:15:46,916
Speaker 3: itating breast cancer? Do I need to get a double mistectomy?

273
00:15:46,956 --> 00:15:49,476
Speaker 3: If I have a causal variant? Many you know patients

274
00:15:49,476 --> 00:15:51,996
Speaker 3: elect to do that or do I just do an

275
00:15:51,996 --> 00:15:55,076
Speaker 3: annual mammogram and I just monitor? Right, And most of

276
00:15:55,116 --> 00:15:58,476
Speaker 3: the time your mutation is different from the ones that

277
00:15:58,516 --> 00:16:02,516
Speaker 3: are recorded in existing clinical databases, so it becomes classified

278
00:16:02,516 --> 00:16:06,116
Speaker 3: as a variant of unknown significance, right, And the model

279
00:16:06,356 --> 00:16:10,596
Speaker 3: has a very good sense of whether not that variant

280
00:16:10,836 --> 00:16:12,836
Speaker 3: that has never been seen before, you don't know what

281
00:16:12,916 --> 00:16:15,716
Speaker 3: it does, whether or not it will cause disease or not. Right,

282
00:16:15,996 --> 00:16:17,956
Speaker 3: And it does that not just for BROCA one, but

283
00:16:17,996 --> 00:16:18,836
Speaker 3: for any gene.

284
00:16:19,276 --> 00:16:21,476
Speaker 2: How do you know that? How do you test that

285
00:16:21,596 --> 00:16:22,956
Speaker 2: sort of out of sample?

286
00:16:23,996 --> 00:16:30,516
Speaker 3: Well, essentially you benchmark it right against there are these

287
00:16:30,556 --> 00:16:34,356
Speaker 3: like ground truth databases created by the National Institutes of

288
00:16:34,356 --> 00:16:37,596
Speaker 3: Health that we've tested against. And then you can also

289
00:16:38,116 --> 00:16:41,596
Speaker 3: again you can run experiments, so you can for example,

290
00:16:41,756 --> 00:16:44,876
Speaker 3: generate those variants of known significance in the lab and

291
00:16:44,956 --> 00:16:47,996
Speaker 3: actually test them in the lab assay and compare them

292
00:16:48,676 --> 00:16:52,316
Speaker 3: against the causal mutations or the benign mutations and essentially

293
00:16:52,396 --> 00:16:55,956
Speaker 3: see where they rank in terms of you know, this

294
00:16:56,236 --> 00:17:01,476
Speaker 3: metric of pathogenicity. So folks have done those experiments and

295
00:17:01,716 --> 00:17:03,916
Speaker 3: you can use those as benchmark data sets.

296
00:17:04,556 --> 00:17:06,916
Speaker 2: So I mean that one seems to have kind of

297
00:17:07,076 --> 00:17:10,396
Speaker 2: immediate clinical relevance, does it? Or am I missing something?

298
00:17:10,716 --> 00:17:14,956
Speaker 3: We certainly don't think that folks should be using EVO

299
00:17:15,076 --> 00:17:18,996
Speaker 3: designer on the ARC website, putting in their genetic sequence

300
00:17:19,196 --> 00:17:22,716
Speaker 3: and then trying to make clinical decisions, right, Sure, but

301
00:17:23,596 --> 00:17:27,116
Speaker 3: we are making newer versions of the model, which we're

302
00:17:27,116 --> 00:17:29,676
Speaker 3: training internally that we think can be state of the

303
00:17:29,796 --> 00:17:32,996
Speaker 3: art at doing this type of genetic diagnostic and so

304
00:17:33,036 --> 00:17:36,356
Speaker 3: we're excited about improving these models over time. I think

305
00:17:36,436 --> 00:17:40,876
Speaker 3: in order for them to be used by a doctor

306
00:17:41,156 --> 00:17:44,556
Speaker 3: in the context of the clinical diagnosis, right, there's a

307
00:17:44,556 --> 00:17:48,836
Speaker 3: lot of testing and validation and regulatory oversight that will

308
00:17:48,836 --> 00:17:52,156
Speaker 3: need to happen to make sure that this can be

309
00:17:52,276 --> 00:17:54,956
Speaker 3: used and certified in the right way. But from a

310
00:17:55,036 --> 00:17:57,676
Speaker 3: research point of view, right, I think this is extremely

311
00:17:57,716 --> 00:18:01,436
Speaker 3: exciting given the performance of the model already, let alone

312
00:18:01,476 --> 00:18:03,396
Speaker 3: the new things that we're doing to it under the hood.

313
00:18:07,116 --> 00:18:19,356
Speaker 2: We'll lead back in just a minute. Earlier in the conversation,

314
00:18:19,596 --> 00:18:22,756
Speaker 2: Patrick was talking about the BRACA one gene with that

315
00:18:22,916 --> 00:18:27,156
Speaker 2: single gene link to breast cancer. But for most diseases

316
00:18:27,156 --> 00:18:30,636
Speaker 2: with genetic components, there is not one single gene. There

317
00:18:30,636 --> 00:18:32,956
Speaker 2: are lots of genes, and in lots of cases we

318
00:18:32,996 --> 00:18:35,716
Speaker 2: still don't really understand what's going on. And so I

319
00:18:35,756 --> 00:18:39,196
Speaker 2: asked him how that universe of diseases fits into his work.

320
00:18:39,476 --> 00:18:43,276
Speaker 3: Yeah, so this is some of the fun part, right obviously.

321
00:18:44,436 --> 00:18:46,556
Speaker 3: You know, so so far we've been talking about individual

322
00:18:46,676 --> 00:18:50,716
Speaker 3: mutations in one gene, right, Yeah, and you know, in

323
00:18:50,756 --> 00:18:53,196
Speaker 3: this very kind of well controlled setting, in a very

324
00:18:53,196 --> 00:18:56,756
Speaker 3: well understed gene for a well reasonably well understood disease.

325
00:18:57,356 --> 00:18:57,556
Speaker 1: Right.

326
00:18:58,276 --> 00:19:01,716
Speaker 3: Most of biology is not that simple, right, And so

327
00:19:01,796 --> 00:19:05,196
Speaker 3: the reason why we have GA studies or polygenic risk

328
00:19:05,196 --> 00:19:08,676
Speaker 3: scores is because we have complex traits, lots of different

329
00:19:08,756 --> 00:19:11,276
Speaker 3: mutations that create lots of different phenotypes.

330
00:19:11,476 --> 00:19:15,316
Speaker 2: Gis genome wide association. Let's look at the whole genome

331
00:19:15,436 --> 00:19:17,276
Speaker 2: and see what correlates exactly.

332
00:19:17,436 --> 00:19:21,436
Speaker 3: It's basically the ultimate human genetics fishing experiment, What the

333
00:19:21,476 --> 00:19:26,116
Speaker 3: hell is the genetic reason why someone has a given trait? Right,

334
00:19:26,276 --> 00:19:28,916
Speaker 3: And that can be literally any trait, And you get

335
00:19:28,956 --> 00:19:31,156
Speaker 3: a bunch of people that have that trade a and

336
00:19:31,196 --> 00:19:34,356
Speaker 3: a bunch of people that seem normal, sequence them and

337
00:19:34,396 --> 00:19:37,636
Speaker 3: look at the statistical association for what genes seem to

338
00:19:37,676 --> 00:19:42,516
Speaker 3: be enriched in the people who have the desired or

339
00:19:42,636 --> 00:19:45,836
Speaker 3: bad trade right, and that could be height, that could

340
00:19:45,836 --> 00:19:51,436
Speaker 3: be hair thickness, that could be likelihood to get Alzheimer's disease.

341
00:19:51,916 --> 00:19:55,316
Speaker 2: Yeah, and it seems like that was one that people

342
00:19:55,316 --> 00:19:57,756
Speaker 2: were very hopeful about, I don't know, ten years ago

343
00:19:57,876 --> 00:20:01,196
Speaker 2: or something in the kind of post human genome project era.

344
00:20:01,596 --> 00:20:04,116
Speaker 2: That turned out to be a lot harder than anybody thought.

345
00:20:03,956 --> 00:20:04,476
Speaker 1: Is that right?

346
00:20:04,996 --> 00:20:08,116
Speaker 3: Yeah? I think our initial goal with sequencing the human

347
00:20:08,156 --> 00:20:09,996
Speaker 3: genome was that we would find a whole bunch of

348
00:20:10,076 --> 00:20:13,436
Speaker 3: drug targets. It turned out we needed the ability to

349
00:20:13,556 --> 00:20:19,196
Speaker 3: sequence many many more genomes, potentially everybody's genomes, and compute

350
00:20:19,756 --> 00:20:23,876
Speaker 3: over all of this inscrutable complex data with AI right

351
00:20:24,036 --> 00:20:27,676
Speaker 3: in order to actually understand how these mutations actually interact

352
00:20:27,716 --> 00:20:31,316
Speaker 3: and work. Right. And then so that's the thing that

353
00:20:31,716 --> 00:20:35,516
Speaker 3: we're very excited about with next generation versions of EVA

354
00:20:35,956 --> 00:20:40,276
Speaker 3: is to try to understand genetic interactions and polygenic traits

355
00:20:40,556 --> 00:20:44,556
Speaker 3: rather than monogenetic traits that are caused by single gene

356
00:20:44,596 --> 00:20:45,516
Speaker 3: and a single mutation.

357
00:20:46,036 --> 00:20:49,716
Speaker 2: Because most diseases, or the diseases that affect most people,

358
00:20:49,796 --> 00:20:54,676
Speaker 2: certainly have lots of genetic inputs and it's exactly one

359
00:20:54,716 --> 00:20:55,676
Speaker 2: messed up gene.

360
00:20:56,036 --> 00:20:58,956
Speaker 3: So in high school you learned that genotype and the

361
00:20:59,076 --> 00:21:02,476
Speaker 3: environment collaborate to create phenotype, which is a fancy way

362
00:21:02,516 --> 00:21:07,756
Speaker 3: of saying there were nature and nurture collectively create behavior

363
00:21:08,276 --> 00:21:12,916
Speaker 3: and all like biology, right, And so this is really

364
00:21:12,916 --> 00:21:16,956
Speaker 3: where our EGO work, in our virtual cell work starts

365
00:21:16,956 --> 00:21:21,716
Speaker 3: to collide a model of genotype and a model of

366
00:21:22,196 --> 00:21:26,476
Speaker 3: cellular response and behavior, and how do we actually connect

367
00:21:26,716 --> 00:21:30,196
Speaker 3: these two modules in order to make more accurate models

368
00:21:30,236 --> 00:21:31,116
Speaker 3: of cell biology.

369
00:21:31,956 --> 00:21:34,436
Speaker 2: So you brought up the virtual cell, which we haven't

370
00:21:34,476 --> 00:21:38,956
Speaker 2: really talked about, which seems like kind of the other

371
00:21:38,996 --> 00:21:41,156
Speaker 2: at least another big project at ARC.

372
00:21:41,276 --> 00:21:43,876
Speaker 1: Right, tell me about the virtual cell.

373
00:21:44,596 --> 00:21:47,876
Speaker 3: So we talked earlier about GOS, right, and GS is

374
00:21:47,996 --> 00:21:53,956
Speaker 3: fundamentaliation exactly geno wide association. It's an association, right, there's

375
00:21:53,996 --> 00:21:58,636
Speaker 3: no causality, right, And so in many ways. The breakthrough

376
00:21:58,676 --> 00:22:03,076
Speaker 3: of Crisper and genome editing was to basically take mutations

377
00:22:03,316 --> 00:22:07,276
Speaker 3: that are associated causally creating them and testing them in

378
00:22:07,316 --> 00:22:09,956
Speaker 3: the lab and living cells and seeing how the cells

379
00:22:09,996 --> 00:22:14,196
Speaker 3: respond and behave when you make this disease associated mutation,

380
00:22:14,636 --> 00:22:17,596
Speaker 3: do you get cancer from a normal self?

381
00:22:17,596 --> 00:22:20,116
Speaker 2: So it's allowing us to say, well, we know there's

382
00:22:20,196 --> 00:22:22,836
Speaker 2: correlation here, but we don't know it's causative. It allows

383
00:22:22,916 --> 00:22:25,996
Speaker 2: us to answer that question at least in some content exactly.

384
00:22:26,156 --> 00:22:29,196
Speaker 3: And these are experiments that you do in the lab.

385
00:22:29,516 --> 00:22:31,796
Speaker 3: Now the question is how much of this can we

386
00:22:31,916 --> 00:22:34,756
Speaker 3: push to an AI model, so instead of doing the

387
00:22:34,836 --> 00:22:37,476
Speaker 3: Chrisper experiment, the model can just tell you.

388
00:22:37,716 --> 00:22:40,436
Speaker 2: And is this going back to the original problem of

389
00:22:40,476 --> 00:22:43,316
Speaker 2: like it's wildly slow to have to do these one

390
00:22:43,356 --> 00:22:44,836
Speaker 2: by one in the lab with Crisper.

391
00:22:45,036 --> 00:22:49,156
Speaker 3: That's exactly right with Crisper, with a small molecule library

392
00:22:49,556 --> 00:22:52,076
Speaker 3: or anything else. Right. The point is we want a

393
00:22:52,156 --> 00:22:55,356
Speaker 3: model that can actually predict, for example, the result of

394
00:22:55,476 --> 00:23:02,116
Speaker 3: Chrisper experiments or a drug perturbations right and guide the

395
00:23:02,116 --> 00:23:03,756
Speaker 3: things that people are going to do in the lab.

396
00:23:04,836 --> 00:23:08,116
Speaker 2: So it's basically saying, if we took a cell and

397
00:23:08,156 --> 00:23:11,036
Speaker 2: we changed the output of this gene. Or if we

398
00:23:11,076 --> 00:23:15,396
Speaker 2: took a cell and we used this drug on it,

399
00:23:15,756 --> 00:23:17,916
Speaker 2: for lack of a better phrase, what would happen?

400
00:23:18,116 --> 00:23:19,276
Speaker 1: Is that the kind of question?

401
00:23:19,556 --> 00:23:22,836
Speaker 3: Yeah, the maybe the mental model for me is, imagine

402
00:23:22,836 --> 00:23:25,436
Speaker 3: you're like a DJ, right, and you have the most

403
00:23:25,556 --> 00:23:29,916
Speaker 3: complex mixer on the planet, right, and you're in the

404
00:23:29,996 --> 00:23:32,076
Speaker 3: middle of this Las Vegas club. You have all these

405
00:23:32,156 --> 00:23:35,476
Speaker 3: knobs and dials, and every time you move them around,

406
00:23:35,516 --> 00:23:40,476
Speaker 3: the music changes, and you're controlling the music of the cell, right,

407
00:23:40,636 --> 00:23:42,396
Speaker 3: And so when you turn this up all the way,

408
00:23:42,716 --> 00:23:45,516
Speaker 3: you know, for some given you know sound, you can

409
00:23:45,556 --> 00:23:50,956
Speaker 3: get this kind of chaotic super loud disease, like like noise.

410
00:23:50,996 --> 00:23:53,796
Speaker 3: It doesn't sound mollifluous and interesting, right, And then you

411
00:23:53,836 --> 00:23:56,076
Speaker 3: want to basically figure out how to turn it off.

412
00:23:56,276 --> 00:24:00,076
Speaker 3: But if it's actually a bunch of different sounds that

413
00:24:00,156 --> 00:24:03,556
Speaker 3: are contributing to the noise, right, you need to figure

414
00:24:03,556 --> 00:24:06,876
Speaker 3: out the fastest way to find out which of these

415
00:24:07,156 --> 00:24:09,876
Speaker 3: you know, you know, tens of thousands of jobs are

416
00:24:09,876 --> 00:24:12,836
Speaker 3: out there should be actually changed in the right way.

417
00:24:13,076 --> 00:24:13,316
Speaker 1: Now.

418
00:24:13,396 --> 00:24:16,596
Speaker 3: Today we try to find this out by trying to

419
00:24:16,636 --> 00:24:20,276
Speaker 3: guess which knobs we need to be dialing and then

420
00:24:20,276 --> 00:24:22,596
Speaker 3: we like just messed around and do it. And this

421
00:24:22,716 --> 00:24:27,716
Speaker 3: is extremely slow, and it's this comminentorial search that's incredibly slow.

422
00:24:28,076 --> 00:24:30,036
Speaker 2: It's not one thing. It's not even that you have

423
00:24:30,076 --> 00:24:32,596
Speaker 2: to guess the one thing. It's like it's many different

424
00:24:32,596 --> 00:24:35,516
Speaker 2: things have to happen, and so the math of that,

425
00:24:35,556 --> 00:24:38,476
Speaker 2: if you're doing trial and error, is just crazy and

426
00:24:38,556 --> 00:24:39,116
Speaker 2: it takes forever.

427
00:24:39,196 --> 00:24:42,516
Speaker 3: Yeah, it's exclusively combinatorial, and the goal of the model

428
00:24:42,636 --> 00:24:44,956
Speaker 3: is to be a copilot where it would be the

429
00:24:44,956 --> 00:24:49,116
Speaker 3: equivalent of your like metaar glasses that tells you it's

430
00:24:49,196 --> 00:24:51,276
Speaker 3: this knob and this knob and this knob, or it's

431
00:24:51,476 --> 00:24:54,796
Speaker 3: you should try these five knobs and these five knobs,

432
00:24:54,876 --> 00:24:56,756
Speaker 3: and then you can, you know, in a much more

433
00:24:56,796 --> 00:25:00,036
Speaker 3: targeted way, you know which things to dial right. And

434
00:25:00,076 --> 00:25:05,436
Speaker 3: that's how we could accelerate the experimental verification of the

435
00:25:05,476 --> 00:25:10,276
Speaker 3: model's predictions in order to try to find the perturbations

436
00:25:10,316 --> 00:25:14,316
Speaker 3: that could treat disease cells for example. But this of

437
00:25:14,476 --> 00:25:17,876
Speaker 3: course is a platform capability that can be useful across

438
00:25:17,956 --> 00:25:22,516
Speaker 3: basic science but also very much for a therapeutic target idea,

439
00:25:22,556 --> 00:25:24,796
Speaker 3: which is something that we deeply care about accelerating.

440
00:25:25,596 --> 00:25:29,076
Speaker 2: Yeah, I mean therapeutic target idea is basically like what

441
00:25:29,116 --> 00:25:31,996
Speaker 2: should we send a drug to go fix?

442
00:25:32,356 --> 00:25:32,796
Speaker 3: Exactly?

443
00:25:32,956 --> 00:25:33,196
Speaker 1: Change?

444
00:25:33,276 --> 00:25:33,396
Speaker 3: Right?

445
00:25:33,436 --> 00:25:33,876
Speaker 1: Exactly?

446
00:25:34,516 --> 00:25:37,476
Speaker 2: I mean you've talked about, I think in this context,

447
00:25:37,476 --> 00:25:40,036
Speaker 2: I've heard you talk about the the failure rate of

448
00:25:40,076 --> 00:25:44,036
Speaker 2: clinical trials, right, like something like ninety percent of drugs

449
00:25:44,076 --> 00:25:47,316
Speaker 2: that go into clinical trials, which already is far along

450
00:25:47,396 --> 00:25:52,676
Speaker 2: in the development process, right, ninety percent of those drugs fail.

451
00:25:52,916 --> 00:25:54,636
Speaker 1: Like how does how could this help?

452
00:25:55,436 --> 00:25:57,796
Speaker 3: So we're not very good at making drugs? Right? The

453
00:25:57,876 --> 00:26:02,836
Speaker 3: statistics are clear, and that really means we're that at

454
00:26:02,996 --> 00:26:07,356
Speaker 3: two different things, finding the right drug target and then

455
00:26:07,436 --> 00:26:10,516
Speaker 3: actually having a drug that actually targets it in the

456
00:26:10,556 --> 00:26:10,996
Speaker 3: right way.

457
00:26:11,316 --> 00:26:11,556
Speaker 1: Right.

458
00:26:11,916 --> 00:26:15,036
Speaker 3: So, so, for example, if you found the right drug target,

459
00:26:15,276 --> 00:26:18,196
Speaker 3: let's say it's like a runaway car and you have

460
00:26:18,276 --> 00:26:22,436
Speaker 3: an inhibitor, but it doesn't stop it, It just slows

461
00:26:22,476 --> 00:26:25,676
Speaker 3: it down, right, Yeah, it wouldn't actually be curative and

462
00:26:25,756 --> 00:26:26,556
Speaker 3: fix the problem.

463
00:26:26,676 --> 00:26:26,916
Speaker 1: Right.

464
00:26:26,996 --> 00:26:30,236
Speaker 3: So you need the right drug composition and you need

465
00:26:30,236 --> 00:26:33,316
Speaker 3: the right drug target, and because we seem to be

466
00:26:33,396 --> 00:26:36,796
Speaker 3: bad at both of these things, we get a ninety

467
00:26:36,796 --> 00:26:37,716
Speaker 3: percent failure rate.

468
00:26:38,996 --> 00:26:41,636
Speaker 1: Why did you decide to focus on Alzheimer's.

469
00:26:41,956 --> 00:26:45,436
Speaker 3: So Alzheimer's is one of the major killers no one

470
00:26:45,476 --> 00:26:48,276
Speaker 3: has agreed on or we certainly don't know what the

471
00:26:48,356 --> 00:26:51,116
Speaker 3: right drug target is for Alzheimer's disease. And we think

472
00:26:51,156 --> 00:26:54,356
Speaker 3: of it as a textbook example of a complex human

473
00:26:54,396 --> 00:26:57,956
Speaker 3: disease that has a bunch of genetic mutations that we

474
00:26:57,996 --> 00:27:00,796
Speaker 3: don't understand. We don't know what genes they control, what

475
00:27:00,836 --> 00:27:03,996
Speaker 3: pathways they control, so we just don't know what the

476
00:27:04,036 --> 00:27:07,276
Speaker 3: hell is happening, right, And it's a you know, you know,

477
00:27:07,316 --> 00:27:11,436
Speaker 3: they're they're also inputs from the environment, infection age obviously,

478
00:27:12,396 --> 00:27:16,956
Speaker 3: you know, potentially diabetes and others that increase your risk.

479
00:27:17,436 --> 00:27:21,796
Speaker 3: So it's this really interesting example of having a risk

480
00:27:21,996 --> 00:27:26,796
Speaker 3: conferring genetic state penetr and then pushed over the edge

481
00:27:26,956 --> 00:27:32,636
Speaker 3: by environmental perturbations in order to create dementia. Right, And

482
00:27:32,676 --> 00:27:35,156
Speaker 3: we think, first of all, it would be very impactful

483
00:27:35,236 --> 00:27:38,196
Speaker 3: to human health if we can solve it, and second,

484
00:27:38,636 --> 00:27:41,676
Speaker 3: we can use it as an example as a blueprint

485
00:27:41,956 --> 00:27:44,356
Speaker 3: for how we can cure other complex diseases if we

486
00:27:44,396 --> 00:27:47,636
Speaker 3: can make progress. And so our goal has been to

487
00:27:47,756 --> 00:27:50,636
Speaker 3: try to figure out ways that we can model this

488
00:27:51,196 --> 00:27:57,156
Speaker 3: both experimentally at scale and computationally with model like AI

489
00:27:57,316 --> 00:28:00,516
Speaker 3: models of the major cell types in the brain in

490
00:28:00,636 --> 00:28:05,156
Speaker 3: order to try to you know, model these different combinatorial

491
00:28:05,276 --> 00:28:11,876
Speaker 3: changes over time much more effectively, basically much faster. Yeah,

492
00:28:11,916 --> 00:28:13,676
Speaker 3: and that's sort of one of the ways that we're

493
00:28:13,676 --> 00:28:17,236
Speaker 3: going to actually internally test the output of our virtual

494
00:28:17,276 --> 00:28:19,996
Speaker 3: cell models. But of course, any other lab will be

495
00:28:19,996 --> 00:28:23,196
Speaker 3: able to take these models and test them for some

496
00:28:23,316 --> 00:28:27,596
Speaker 3: heart disease pathway, let's say, or some inflammation pathway that

497
00:28:27,716 --> 00:28:31,196
Speaker 3: might be involved in glaucoma and eye disease, or you know,

498
00:28:31,316 --> 00:28:34,236
Speaker 3: whatever basic science mechanism that they might care about.

499
00:28:34,436 --> 00:28:34,956
Speaker 1: So that's the.

500
00:28:34,956 --> 00:28:39,756
Speaker 3: Scientific reason why we care about Alzheimer's. On a personal note,

501
00:28:39,756 --> 00:28:43,756
Speaker 3: it's also why I became a scientist. My my grandfather,

502
00:28:43,996 --> 00:28:47,756
Speaker 3: like many other people, you know, got Alzheimer's when I

503
00:28:47,796 --> 00:28:50,356
Speaker 3: was eleven. He was living with us at the time,

504
00:28:50,436 --> 00:28:55,236
Speaker 3: and I watched him go through, you know, from mild

505
00:28:55,276 --> 00:29:00,156
Speaker 3: cognitive impairment and confusion to you know, flipping shopping carts

506
00:29:00,196 --> 00:29:03,476
Speaker 3: in costco to you know, being in a nursing home

507
00:29:03,516 --> 00:29:07,516
Speaker 3: and dying. And you know, I think that gave me

508
00:29:07,756 --> 00:29:10,436
Speaker 3: very clear early miss and I think really in many

509
00:29:10,476 --> 00:29:14,076
Speaker 3: ways redirected my life from probably becoming some you know,

510
00:29:14,196 --> 00:29:17,876
Speaker 3: software founder of some kind too, you know, joining a

511
00:29:17,876 --> 00:29:20,956
Speaker 3: lot in high school and picking up a pipet and

512
00:29:21,156 --> 00:29:23,716
Speaker 3: doing that gusts and check search that we're now trying automate.

513
00:29:27,916 --> 00:29:30,036
Speaker 1: We'll be back in a minute with the lightning round.

514
00:29:31,836 --> 00:29:42,556
Speaker 2: Mm hmm, okay, let's finish with the lightning round. Who

515
00:29:42,636 --> 00:29:44,516
Speaker 2: is the most underrated medieval king?

516
00:29:45,196 --> 00:29:49,796
Speaker 3: The most medieval See? I feel like this is basically

517
00:29:51,716 --> 00:29:55,356
Speaker 3: well there, there's there's there's so many that you could

518
00:29:55,836 --> 00:29:58,956
Speaker 3: that you could pick from. I'm going to pass on

519
00:29:59,036 --> 00:30:02,596
Speaker 3: this question. This is this is very this is very

520
00:30:02,676 --> 00:30:06,156
Speaker 3: unfair of me. But I mean I'll have to ponder.

521
00:30:06,196 --> 00:30:06,916
Speaker 3: I'll have to ponder.

522
00:30:07,396 --> 00:30:09,076
Speaker 1: Yeah, why do you love medieval history?

523
00:30:09,556 --> 00:30:12,596
Speaker 3: I studied it for four years as a as a

524
00:30:12,636 --> 00:30:13,836
Speaker 3: young homeschooled student.

525
00:30:14,236 --> 00:30:17,076
Speaker 1: Yeah, why did you study it for four years?

526
00:30:17,436 --> 00:30:20,996
Speaker 3: There's a reason why we love like swords and sorcerers

527
00:30:21,036 --> 00:30:25,796
Speaker 3: and you know, in Game of Thrones and the Battle

528
00:30:25,836 --> 00:30:29,796
Speaker 3: of Aquitaine and whatever. You know, it's I think it's

529
00:30:30,156 --> 00:30:32,756
Speaker 3: one of the periods of time that we felt like

530
00:30:32,916 --> 00:30:33,996
Speaker 3: true optimism.

531
00:30:34,276 --> 00:30:36,716
Speaker 1: Do you think so? I don't think. I mean, you

532
00:30:36,796 --> 00:30:38,716
Speaker 1: know more about it than I do. But I'm shocked.

533
00:30:39,156 --> 00:30:40,596
Speaker 1: Is that not just a projection.

534
00:30:40,956 --> 00:30:45,316
Speaker 3: It's evolutionary selection playing out in real time. It's very

535
00:30:45,396 --> 00:30:51,156
Speaker 3: like visible and well recorded history of evolutionary struggle. I

536
00:30:51,196 --> 00:30:54,156
Speaker 3: think maybe that's why I found deeply fascinating about it.

537
00:30:54,316 --> 00:30:55,636
Speaker 1: Uh huh.

538
00:30:55,676 --> 00:30:59,756
Speaker 2: That's the biologist view I guess of the medieval period.

539
00:30:59,476 --> 00:31:00,516
Speaker 3: Maybe one way to frame it.

540
00:31:01,396 --> 00:31:03,356
Speaker 1: If you weren't working on biology, what would you be

541
00:31:03,396 --> 00:31:03,876
Speaker 1: working on?

542
00:31:04,356 --> 00:31:06,716
Speaker 3: You know, I'd probably be a music talent scout.

543
00:31:07,036 --> 00:31:10,396
Speaker 2: Yeah, what should I listen to? Who should I sign to?

544
00:31:10,476 --> 00:31:10,996
Speaker 1: Our label?

545
00:31:12,996 --> 00:31:16,916
Speaker 3: My favorite band it is the XX I discovered them

546
00:31:16,956 --> 00:31:21,956
Speaker 3: in two thousand and nine or two thousand and eight, yeah, early, Yeah,

547
00:31:21,996 --> 00:31:24,516
Speaker 3: It's like, you know, if I were a seed investor,

548
00:31:24,596 --> 00:31:26,396
Speaker 3: this would be this would be one of my kind

549
00:31:26,396 --> 00:31:27,996
Speaker 3: of key seed investments.

550
00:31:28,236 --> 00:31:29,916
Speaker 2: I'm going to read you a thing you wrote and

551
00:31:29,916 --> 00:31:32,996
Speaker 2: then ask you a question about it you wrote. In

552
00:31:33,036 --> 00:31:36,876
Speaker 2: my view, researchers also need to stoke two warring urges.

553
00:31:37,036 --> 00:31:40,876
Speaker 2: One an opinionated sense of taste in a relentless search

554
00:31:40,876 --> 00:31:44,876
Speaker 2: for beauty, and two a yeoman's grindset for the unglamorous,

555
00:31:44,876 --> 00:31:48,836
Speaker 2: dirty work required to get things done, which is lovely.

556
00:31:48,876 --> 00:31:54,316
Speaker 2: A lovely sentence, And I'm curious for your thoughts about

557
00:31:54,356 --> 00:31:59,236
Speaker 2: that in the context of AI. Basically, like you think

558
00:31:59,316 --> 00:32:00,236
Speaker 2: AI can do two?

559
00:32:00,676 --> 00:32:02,676
Speaker 1: Number two? Do you think it can do one?

560
00:32:02,716 --> 00:32:02,796
Speaker 3: Like?

561
00:32:02,876 --> 00:32:05,756
Speaker 1: Where where does AI fit in the context of that sentence?

562
00:32:06,196 --> 00:32:08,476
Speaker 3: Early on, you know, maybe a year year and a

563
00:32:08,476 --> 00:32:11,876
Speaker 3: half ago, you know, I think there was lots of discussion.

564
00:32:11,916 --> 00:32:14,036
Speaker 3: You know, I want AI to you know, do my

565
00:32:14,156 --> 00:32:18,396
Speaker 3: laundry and taxes so I can make art and music

566
00:32:18,916 --> 00:32:21,476
Speaker 3: and not the other way around, right, because you know,

567
00:32:21,476 --> 00:32:23,996
Speaker 3: early on, we we had all these models that were

568
00:32:24,236 --> 00:32:27,036
Speaker 3: you know, generating new images and making AI art that

569
00:32:27,116 --> 00:32:29,556
Speaker 3: was really controversial, and making new types of songs. And

570
00:32:29,556 --> 00:32:31,636
Speaker 3: you're like, wellhy can't I do the basic things so

571
00:32:31,676 --> 00:32:35,596
Speaker 3: I can live and create and dream? And I think

572
00:32:35,836 --> 00:32:39,996
Speaker 3: today it's very clear that it is helping us with both, right,

573
00:32:40,076 --> 00:32:43,836
Speaker 3: that these bicycles for the mind are high quality enough

574
00:32:43,876 --> 00:32:47,196
Speaker 3: to allow us to create not just intellectual insights, but

575
00:32:47,636 --> 00:32:51,556
Speaker 3: new like you know, visual design styles and paradigms, new

576
00:32:51,556 --> 00:32:54,476
Speaker 3: types of sound. And we're just very much in the

577
00:32:54,516 --> 00:32:57,436
Speaker 3: early innings of using these in a way that feel

578
00:32:57,476 --> 00:33:02,196
Speaker 3: augmentive to human creative capacity rather than you know, some

579
00:33:02,196 --> 00:33:06,636
Speaker 3: somehow like aggressive or replacing. You know, that's and I'm

580
00:33:06,636 --> 00:33:09,836
Speaker 3: a big AI Bowl. But above all, I'm a big

581
00:33:10,476 --> 00:33:13,116
Speaker 3: believer in how humans can use them to benefit.

582
00:33:19,916 --> 00:33:22,676
Speaker 2: Patrick Sue is the co founder of the Art Institute

583
00:33:22,996 --> 00:33:26,356
Speaker 2: and Assistant Professor of Bioengineering at UC Berkeley.

584
00:33:27,276 --> 00:33:30,596
Speaker 1: Please email us at problem at pushkin dot fm. We

585
00:33:30,676 --> 00:33:33,116
Speaker 1: are always looking for new guests for the show.

586
00:33:33,996 --> 00:33:37,756
Speaker 2: Today's show was produced by Trinamanino and Gabriel Hunter Chang.

587
00:33:38,156 --> 00:33:42,156
Speaker 2: It was edited by Alexander Garretson and engineered by Sarah Bruguer.

588
00:33:42,596 --> 00:33:44,756
Speaker 2: I'm Jacob Goldstein and we'll be back next week with

589
00:33:44,796 --> 00:33:50,676
Speaker 2: another episode of What's Your Pop