WEBVTT 1 00:00:01.199 --> 00:00:06.200 Welcome to the Sentient Code, where intelligence is engineered, autonomy 2 00:00:06.280 --> 00:00:10.439 is emerging, and a line between human and machine grows thinner. 3 00:00:10.800 --> 00:00:15.359 Each episode, we decode the algorithms, explore the robotics, and 4 00:00:15.439 --> 00:00:21.920 examine the ideas shaping the future of artificial minds. 5 00:00:23.839 --> 00:00:25.839 I want to start today with a number, and not 6 00:00:25.879 --> 00:00:30.399 just any number. It's a number so incomprehensibly massive that 7 00:00:30.839 --> 00:00:33.759 honestly the human brain just shorts out trying to visualize it. 8 00:00:33.840 --> 00:00:36.039 Oh, I have a sneaking suspicion. I know exactly where 9 00:00:36.039 --> 00:00:38.920 you're going with this. You're looking at the combinatorial math 10 00:00:39.600 --> 00:00:42.840 for biology right now, the protein problem. Absolutely am I 11 00:00:42.880 --> 00:00:44.920 was reading through the research for this deep dive and 12 00:00:44.960 --> 00:00:49.240 I hit this comparison that just stopped me cold. So 13 00:00:49.679 --> 00:00:52.719 for everyone listening, I want you to picture the entire 14 00:00:52.920 --> 00:00:57.920 observable universe right, every star, every planet, every cloud of 15 00:00:57.960 --> 00:01:01.640 interstellar gas made a number of atoms in all of 16 00:01:01.640 --> 00:01:04.599 that is roughly ten to the power of eighty. That 17 00:01:04.719 --> 00:01:06.319 is a one followed by eighty zero. 18 00:01:06.480 --> 00:01:09.280 It's the literal definition of astronomical exactly. 19 00:01:09.280 --> 00:01:12.599 Now, take a single average size protein, just a basic 20 00:01:12.640 --> 00:01:15.519 biological machine made of amino acids. If you wanted to 21 00:01:15.560 --> 00:01:18.599 test every possible combination of amino acids to find the 22 00:01:18.640 --> 00:01:20.879 absolute best version of that protein. 23 00:01:21.079 --> 00:01:23.040 The number of variants you'd have to build is larger 24 00:01:23.079 --> 00:01:24.719 than the number of atoms in the universe. 25 00:01:24.840 --> 00:01:28.560 It is significantly larger, actually, Yeah, and that creates what 26 00:01:28.560 --> 00:01:32.239 we call the fundamental bottleneck of bioengineering. We know that 27 00:01:32.439 --> 00:01:37.079 somewhere in that infinite haystack of combinations, there is a needle, 28 00:01:37.439 --> 00:01:40.439 a perfect needle, right, there is a version of that 29 00:01:40.519 --> 00:01:44.680 protein that cures a specific cancer, or eats plastic waste, 30 00:01:44.840 --> 00:01:49.920 or generates clean energy with perfect efficiency. But we physically 31 00:01:50.000 --> 00:01:51.920 cannot search the whole haystack. 32 00:01:52.040 --> 00:01:52.840 We don't have enough time. 33 00:01:52.840 --> 00:01:55.640 They don't enough time, we don't have enough resources, and frankly, 34 00:01:55.680 --> 00:01:57.920 we don't even have enough atoms on Earth to build 35 00:01:57.959 --> 00:01:59.719 all the test tubes you would need to run that 36 00:01:59.719 --> 00:02:01.319 extras which brings. 37 00:02:01.159 --> 00:02:04.280 Us to today's topic, because if you can't search the haystag, 38 00:02:04.319 --> 00:02:07.400 you're usually stuck just I don't know, picking at the 39 00:02:07.480 --> 00:02:09.400 edges right, hoping you get lucky. 40 00:02:09.479 --> 00:02:10.639 That's historically been the case. 41 00:02:10.719 --> 00:02:14.039 Yeah, But we are looking at a new paper published 42 00:02:14.120 --> 00:02:19.520 just recently in Science February nineteenth, twenty six to be exact, 43 00:02:19.719 --> 00:02:22.599 and it claims to have found a way to cheat 44 00:02:22.639 --> 00:02:23.039 the math. 45 00:02:23.360 --> 00:02:25.439 Well, cheat might be a strong word, but they have 46 00:02:25.599 --> 00:02:29.360 certainly found a massive shortcut. We're talking about a framework 47 00:02:29.400 --> 00:02:33.919 called multi evolve. It's coming out of Patrick's U's Lab 48 00:02:34.039 --> 00:02:37.120 at UC Berkeley and the Ark Institute, and what they've 49 00:02:37.159 --> 00:02:40.280 proposed is the way to use machine learning to navigate 50 00:02:40.360 --> 00:02:44.919 that impossible search space, that universe of options, exactly navigating 51 00:02:44.919 --> 00:02:47.479 it without having to physically test every single one. 52 00:02:47.560 --> 00:02:51.319 It's basically the difference between walking blindfolded through a mountain 53 00:02:51.400 --> 00:02:53.759 range trying to find the highest peak and having a 54 00:02:53.800 --> 00:02:56.080 satellite map that just drops you right at the summit. 55 00:02:56.159 --> 00:02:58.840 That is precisely the shift we are seeing here. It's 56 00:02:58.840 --> 00:03:02.120 a move from discovery by accident to discovery by design. 57 00:03:02.479 --> 00:03:06.960 So we're going to unpack how this works, why previous 58 00:03:07.000 --> 00:03:09.639 methods have failed us, and what this actually means for 59 00:03:09.719 --> 00:03:12.520 things like medicine and green energy. But before we get 60 00:03:12.520 --> 00:03:14.400 into the AI wizardry. 61 00:03:13.919 --> 00:03:15.199 We need to define the hardware. 62 00:03:15.360 --> 00:03:18.439 Yes, what are we actually trying to build here? Because 63 00:03:18.439 --> 00:03:22.000 I'll admit when I hear protein engineering, my mind instantly 64 00:03:22.039 --> 00:03:24.319 goes to bodybuilders drinking shakes at the gym. 65 00:03:24.439 --> 00:03:27.199 Huh, yeah, that is the most common association. But in 66 00:03:27.240 --> 00:03:29.759 this context, when we say protein, we are talking about 67 00:03:29.800 --> 00:03:32.120 the molecular machines that literally run the world. 68 00:03:32.280 --> 00:03:33.840 They do everything, they really do. 69 00:03:34.199 --> 00:03:37.759 Proteins are the hardware of biology. Dn I is just 70 00:03:37.800 --> 00:03:40.719 the code, the blueprint, right, But the protein is the 71 00:03:40.719 --> 00:03:42.719 physical thing that actually does the work. 72 00:03:42.759 --> 00:03:44.439 It catalyzes reactions, it. 73 00:03:44.400 --> 00:03:48.280 Fights off viruses, it transports oxygen, it digests your food. 74 00:03:48.520 --> 00:03:51.199 And we aren't just talking about human biology either. The 75 00:03:51.240 --> 00:03:56.159 research highlights a surprisingly huge range of applications for engineered proteins. 76 00:03:56.240 --> 00:03:59.280 Oh absolutely, and this is crucial to understand if we 77 00:03:59.319 --> 00:04:01.400 want to grasp why the SEE labs work is such 78 00:04:01.400 --> 00:04:05.599 a big deal. The demand for smart, custom built proteins 79 00:04:06.080 --> 00:04:08.840 is exploding across almost every sector of the economy. 80 00:04:08.960 --> 00:04:13.360 Take consumer goods for instance. The paper specifically mentions laundry detergent, right, 81 00:04:13.560 --> 00:04:18.720 which feels incredibly mundane compared to curing disease. But stick 82 00:04:18.720 --> 00:04:21.959 with us here, why on Earth do we need high 83 00:04:22.000 --> 00:04:25.800 tech machine learning design proteins just to wash our socks. 84 00:04:26.000 --> 00:04:29.279 It comes down to energy efficiency and environmental impact. In 85 00:04:29.319 --> 00:04:31.360 the old days, to get a heavy grease stain or 86 00:04:31.399 --> 00:04:34.120 grass stain out of clothes. You needed boiling hot water 87 00:04:34.240 --> 00:04:36.160 and really harsh chemicals. 88 00:04:35.720 --> 00:04:36.759 Which ruins the clothes. 89 00:04:36.759 --> 00:04:39.759 Eventually, it's bad for the fabric, yes, but it's also 90 00:04:40.040 --> 00:04:43.439 terrible for the energy grid. Heating all that water takes 91 00:04:43.480 --> 00:04:48.160 a massive amount of electricity. So manufacturers started adding enzymes 92 00:04:48.240 --> 00:04:51.800 to the detergent in enzymes of proteins exactly. Enzymes are 93 00:04:51.839 --> 00:04:55.360 just proteins that speed up chemical reactions in detergent. They 94 00:04:55.360 --> 00:04:58.040 act like little molecular scissors that cut up the lipids 95 00:04:58.040 --> 00:05:00.600 in grease or the proteins in a grip as stains, 96 00:05:00.639 --> 00:05:02.279 so they just wash away in the water. 97 00:05:02.480 --> 00:05:04.920 But there's a catch, right, because a washing machine isn't 98 00:05:04.959 --> 00:05:06.439 exactly a natural habitat. 99 00:05:06.519 --> 00:05:10.399 It is a totally hostile environment for a biological molecule. 100 00:05:10.920 --> 00:05:14.800 It's soapy, the pH is highly alkaline, it's spinning violently, 101 00:05:14.839 --> 00:05:18.199 and even if we use cold water, it's not the stable, 102 00:05:18.360 --> 00:05:20.199 cozy temperature of a living cell. 103 00:05:20.519 --> 00:05:20.759 Right. 104 00:05:21.480 --> 00:05:24.079 Nature never designed a protein to live inside. 105 00:05:23.720 --> 00:05:25.639 A tide cod so we have to design one. 106 00:05:25.680 --> 00:05:28.519 We have to engineer a protein that is tough enough 107 00:05:28.560 --> 00:05:31.360 to survive the chemical assault of the wash cycle, but 108 00:05:31.519 --> 00:05:34.600 remains active enough to actually eat the stain. That is 109 00:05:34.639 --> 00:05:37.959 a massive, multi variable engineering challenge. 110 00:05:37.600 --> 00:05:40.560 And that's just consumer goods. The paper also goes deep 111 00:05:40.600 --> 00:05:43.040 into the energy sector, specifically biofuels. 112 00:05:43.600 --> 00:05:45.920 This is a huge one for the climate crisis. We 113 00:05:46.000 --> 00:05:49.759 desperately want to be able to take biomass things like cornstalks, 114 00:05:49.759 --> 00:05:53.439 wood chips, switch grass and turn it into clean liquid fuel. 115 00:05:53.600 --> 00:05:55.319 But plants are stubborn. 116 00:05:55.319 --> 00:05:58.519 Very They are made of cellulose, and cellulose is designed 117 00:05:58.519 --> 00:06:01.680 by nature to be incredibly tough. It's literally the structural 118 00:06:01.680 --> 00:06:05.720 support that makes wood hard. It actively resists breaking down. 119 00:06:05.839 --> 00:06:07.759 So we need a protein that can chew through wood. 120 00:06:07.959 --> 00:06:11.639 We need engineered enzymes called celluluses. If we can build 121 00:06:11.639 --> 00:06:14.839 a cellulus that is say, ten times more efficient at 122 00:06:14.879 --> 00:06:18.439 breaking down plant fiber into simple sugars, the production cost 123 00:06:18.480 --> 00:06:20.720 of biofuel drops through the floor. 124 00:06:20.560 --> 00:06:23.720 And suddenly it can actually compete with fossil fuels precisely. 125 00:06:23.879 --> 00:06:27.920 But again, natural celluluses aren't optimized for giant, hot, acidic 126 00:06:27.920 --> 00:06:31.800 industrial vats. They're optimized for a fungus slowly rotting a 127 00:06:31.839 --> 00:06:34.680 log on a forest floor over five years. We need 128 00:06:34.720 --> 00:06:36.360 to upgrade them for industrial speed. 129 00:06:36.439 --> 00:06:40.519 And then, of course there's the big one, therapeutics medicine. 130 00:06:40.639 --> 00:06:43.959 This is where the precision requirement goes entirely off the charts. 131 00:06:44.240 --> 00:06:47.319 The researchers focus heavily on antibodies. 132 00:06:46.800 --> 00:06:49.480 And an antibody is essentially a protein that acts like 133 00:06:49.480 --> 00:06:50.240 a honing beacon. 134 00:06:50.560 --> 00:06:54.600 Exactly. It binds to a specific target, like a viral 135 00:06:54.720 --> 00:06:59.000 spike protein, or a cancer cell, or a signaling molecule 136 00:06:59.040 --> 00:07:01.759 in your body. That's an autoimmune flare up. 137 00:07:02.160 --> 00:07:04.040 But it has to be impossibly specific. 138 00:07:04.439 --> 00:07:07.480 It has to be perfect. If you engineer an antibody 139 00:07:07.519 --> 00:07:09.800 to kill a cancer cell, but it has even a 140 00:07:09.839 --> 00:07:11.920 slight affinity for your healthy. 141 00:07:11.600 --> 00:07:14.480 Liver cells, that's not a drug anymore. That's a poison. Right. 142 00:07:14.680 --> 00:07:17.959 You need absolute high affinity for the target and zero 143 00:07:18.040 --> 00:07:21.199 affinity for everything else. Plus it has to be structurally 144 00:07:21.240 --> 00:07:24.439 stable so it can survive in the bloodstream without falling apart. 145 00:07:24.600 --> 00:07:26.720 It can't clump up in the vial while sitting on 146 00:07:26.759 --> 00:07:27.680 a pharmacy shelf. 147 00:07:27.720 --> 00:07:29.639 It can't trigger a separate allergic reaction. 148 00:07:29.959 --> 00:07:32.680 It sounds like we're asking for a miracle molecule. We 149 00:07:32.759 --> 00:07:37.439 wanted to do five different, extremely difficult things perfectly, all 150 00:07:37.480 --> 00:07:38.920 at the exact same time. 151 00:07:38.879 --> 00:07:41.920 And that right there is exactly why this is so hard. 152 00:07:42.160 --> 00:07:44.879 That is the high dimensional search problem that Patrick Sue 153 00:07:44.959 --> 00:07:45.720 talks about. 154 00:07:45.439 --> 00:07:48.120 In the paper high dimensional meeting. Lots of variables. 155 00:07:48.600 --> 00:07:52.000 Yes, to get a protein to do all those things, 156 00:07:52.319 --> 00:07:56.480 be stable, be active, be specific. Usually can't just change 157 00:07:56.560 --> 00:07:59.000 one single thing about it. You have to change multiple 158 00:07:59.000 --> 00:08:00.800 parts of the proteins simultaneously. 159 00:08:00.920 --> 00:08:03.879 Okay, so let's get into the how. Let's talk about 160 00:08:03.879 --> 00:08:07.439 the mechanics of changing approaching. A protein is basically a 161 00:08:07.480 --> 00:08:10.360 long spring of amino acids folded up right correct, and 162 00:08:10.639 --> 00:08:12.959 there are twenty different amino acids to choose from. It's 163 00:08:12.959 --> 00:08:14.879 like an alphabet. So if I want to make a 164 00:08:14.879 --> 00:08:16.680 better protein, I just swap out some letters. 165 00:08:16.879 --> 00:08:20.360 In theory, yes, you perform mutations. You swap an alanine 166 00:08:20.360 --> 00:08:23.279 for a tripborn or a glycine for a lysine. 167 00:08:23.360 --> 00:08:26.199 But you mentioned earlier that we can't test all the combinations. 168 00:08:26.759 --> 00:08:31.199 So historically, before this new AI framework came along, if 169 00:08:31.240 --> 00:08:34.120 I wanted to make that better laundry enzyme, what was 170 00:08:34.200 --> 00:08:36.159 the actual process in the lab. 171 00:08:36.519 --> 00:08:38.559 For the last couple of decades, the gold standard has 172 00:08:38.559 --> 00:08:41.080 been a process called directed evolution. 173 00:08:40.840 --> 00:08:42.679 Which won Francis Arnold the Nobel Price. 174 00:08:42.840 --> 00:08:46.360 It did in twenty eighteen. And it's a brilliant technique, 175 00:08:46.480 --> 00:08:50.240 but it fundamentally relies on iteration. It's an educated guess 176 00:08:50.240 --> 00:08:51.039 and check method. 177 00:08:51.320 --> 00:08:53.840 Walk me through that cycle. How does directed evolution work? 178 00:08:54.159 --> 00:08:58.240 You start with your natural baseline protein. You introduce a 179 00:08:58.279 --> 00:09:01.600 bunch of random mutations into the d that codes for it, 180 00:09:01.759 --> 00:09:04.320 usually just changing one amino acid at a time across 181 00:09:04.360 --> 00:09:04.919 a whole batch. 182 00:09:04.960 --> 00:09:08.200 So you make a library of slight variation exactly. 183 00:09:08.799 --> 00:09:12.799 Then you screen those thousands of slightly tweaked variants in 184 00:09:12.840 --> 00:09:16.200 the lab. Now, most of them will be completely broken 185 00:09:16.279 --> 00:09:17.399 by the mutation. 186 00:09:17.240 --> 00:09:19.759 Because random mutations usually break things right. 187 00:09:20.200 --> 00:09:22.399 Some will act exactly the same as the original, but 188 00:09:22.559 --> 00:09:24.960 maybe one or two out of ten thousand are slightly 189 00:09:25.000 --> 00:09:26.840 better at surviving the hot, soapy water. 190 00:09:27.000 --> 00:09:30.399 Okay, so I find a winner. It's let's say five 191 00:09:30.440 --> 00:09:31.000 percent better. 192 00:09:31.080 --> 00:09:33.639 You take that specific winner, and that becomes the parent 193 00:09:33.679 --> 00:09:36.399 for the next generation. You mutate that one, You screen 194 00:09:36.440 --> 00:09:38.519 the new batch, find the best of those, and you 195 00:09:38.639 --> 00:09:40.919 keep climbing that ladder step by step. 196 00:09:41.120 --> 00:09:44.399 That sounds incredibly logical. I mean, it's essentially how natural 197 00:09:44.440 --> 00:09:47.720 evolution works. Survival of the fittest round after round, selecting 198 00:09:47.759 --> 00:09:48.399 for the trait you want. 199 00:09:48.519 --> 00:09:51.600 It is entirely logical, but it has a fatal flaw 200 00:09:51.639 --> 00:09:55.200 when it comes to advanced engineering. It gets stuck stuck out. Okay. 201 00:09:55.240 --> 00:09:57.919 Imagine you're climbing a mountain, but you're in a thick 202 00:09:58.120 --> 00:10:00.919 pea soup fog. You can only I see exactly one 203 00:10:00.919 --> 00:10:03.279 step ahead of you. Okay, And the strict rule of 204 00:10:03.320 --> 00:10:06.120 this climb is you must always take a step up. 205 00:10:06.519 --> 00:10:09.080 If a step goes down in elevation, you are not 206 00:10:09.240 --> 00:10:10.279 allowed to take it. 207 00:10:10.279 --> 00:10:12.799 Because taking a step down means the protein got worse. 208 00:10:13.080 --> 00:10:16.080 Exactly, you only keep the mutations that improve the function, 209 00:10:16.559 --> 00:10:19.919 So you keep stepping up and up until eventually every 210 00:10:19.919 --> 00:10:23.000 step around you goes down. You are standing on a peak. 211 00:10:23.320 --> 00:10:24.200 I've reached the top. 212 00:10:24.519 --> 00:10:27.120 You think you've reached the top, but because of the fog, 213 00:10:27.360 --> 00:10:29.279 you don't realize that you are just standing on a 214 00:10:29.320 --> 00:10:33.039 small foothill. The real mountain, the one that's ten times higher, 215 00:10:33.080 --> 00:10:36.440 the truly optimal protein, is a mile away. But to 216 00:10:36.480 --> 00:10:38.120 get there you would have had to walk down into 217 00:10:38.159 --> 00:10:41.159 a valley first, and directed evolution doesn't let you go. 218 00:10:41.200 --> 00:10:44.480 Down, so you're trapped on what they call a local optimum. 219 00:10:44.759 --> 00:10:48.080 You have a decent protein, but mathematically you are nowhere 220 00:10:48.159 --> 00:10:50.120 near the best possible one. 221 00:10:49.919 --> 00:10:53.000 Exactly, And the biological reason you get trapped in that 222 00:10:53.080 --> 00:10:55.480 valley is because of a concept called epistosis. 223 00:10:55.799 --> 00:10:59.519 Okay, yes, I love this word. It sounds incredibly intimidating, 224 00:11:00.080 --> 00:11:02.600 but it is actually the key to the entire puzzle 225 00:11:02.639 --> 00:11:07.480 of why mlt I evolve was created. Let's unpack episticis. 226 00:11:07.399 --> 00:11:10.320 It really is the concept that breaks the old step 227 00:11:10.360 --> 00:11:13.720 wise way of doing things. Epistosis simply means that the 228 00:11:13.720 --> 00:11:16.720 effect of one mutation heavily depends on the context of 229 00:11:16.759 --> 00:11:18.120 the other mutations around it. 230 00:11:18.440 --> 00:11:22.240 So in biology, one plus one does not always equal 231 00:11:22.240 --> 00:11:23.360 to far from it. 232 00:11:23.720 --> 00:11:25.159 Give me a food analogy. 233 00:11:25.240 --> 00:11:28.840 Oh, okay, if you add strawberries to a bowl, that's good. 234 00:11:29.120 --> 00:11:30.879 You add garlic to a pan, that's good. 235 00:11:31.080 --> 00:11:32.679 But if you combine strawberries and. 236 00:11:32.799 --> 00:11:34.360 Garlict an absolute disaster. 237 00:11:34.559 --> 00:11:34.720 Right. 238 00:11:34.840 --> 00:11:38.159 That is negative episticis the combination of the two elements 239 00:11:38.200 --> 00:11:40.639 is vastly worse than the sum of their individual parts. 240 00:11:41.039 --> 00:11:44.279 But epistosis works the other way too. Positive episticis like 241 00:11:44.320 --> 00:11:46.440 a lock and key. Yes, think of a physical key. 242 00:11:46.519 --> 00:11:48.840 If you change one single groove on a key, it 243 00:11:48.879 --> 00:11:51.360 doesn't open the door anymore. It's useless. Right, If you 244 00:11:51.480 --> 00:11:54.159 change a different single groove, it's also useless. But if 245 00:11:54.200 --> 00:11:56.759 you change both grooves at the exact same time to 246 00:11:56.840 --> 00:11:59.759 match a brand new lock, suddenly it works perfectly. 247 00:12:00.159 --> 00:12:01.960 I see. So if you were doing this step by 248 00:12:02.000 --> 00:12:06.000 step directed evolution method, you would have tested the first 249 00:12:06.080 --> 00:12:09.159 groove change. So I didn't open the door and thrown 250 00:12:09.200 --> 00:12:09.519 it in. 251 00:12:09.399 --> 00:12:12.360 The track exactly, you would have said, this mutation is broken. 252 00:12:12.639 --> 00:12:13.879 It's a step down the mountain. 253 00:12:14.080 --> 00:12:16.559 You never would have found the brilliant combination because you 254 00:12:16.639 --> 00:12:17.559 never tried them together. 255 00:12:17.720 --> 00:12:23.240 Precisely, the step wise approach blindly discards mutations that look 256 00:12:23.320 --> 00:12:27.120 detrimental on their own but are actually essential ingredients for 257 00:12:27.159 --> 00:12:28.759 a future breakthrough. 258 00:12:28.360 --> 00:12:31.080 Because they need a partner mutation to make sense. 259 00:12:30.919 --> 00:12:34.600 Structurally, yes, and because we are trying to optimize for 260 00:12:34.720 --> 00:12:39.039 these highly complex, high dimensional traits like stability and activity 261 00:12:39.039 --> 00:12:43.159 and specificity all at once. The true solution almost always 262 00:12:43.200 --> 00:12:48.000 requires a combination of mutations that have this complex epistatic relationship. 263 00:12:48.320 --> 00:12:50.840 So we fundamentally need a way to see the garlic 264 00:12:50.879 --> 00:12:53.360 and strawberries problem before we taste it. You need to 265 00:12:53.360 --> 00:12:55.320 predict these hidden interactions, and. 266 00:12:55.279 --> 00:12:58.120 That is exactly where a MULTII evolve enters the picture. 267 00:12:58.399 --> 00:13:00.600 The goal of Patrick Sue and his team was to 268 00:13:00.600 --> 00:13:02.720 stop walking step by step in the fog. 269 00:13:02.840 --> 00:13:04.039 They wanted the satellite map. 270 00:13:04.200 --> 00:13:06.559 They wanted to build a machine learning system that could 271 00:13:06.639 --> 00:13:09.000 predict the view from the top of the highest mountain 272 00:13:09.440 --> 00:13:12.840 without having to physically climb every single hill to get there. 273 00:13:13.159 --> 00:13:17.360 So let's break down the actual EMUALTI evolve architecture because 274 00:13:17.440 --> 00:13:20.000 reading the paper, it's very clear this isn't just a 275 00:13:20.039 --> 00:13:22.639 situation where they dumped a bunch of raw data into 276 00:13:22.679 --> 00:13:25.240 a black box AI and asked it to do magic. 277 00:13:25.399 --> 00:13:26.120 No, not at all. 278 00:13:26.159 --> 00:13:28.720 Three specific three step workflow correct. 279 00:13:28.919 --> 00:13:32.720 And the structure of this workflow is incredibly deliberate. It's 280 00:13:32.799 --> 00:13:37.320 designed specifically to capture and learn that epistatic interaction data 281 00:13:37.320 --> 00:13:38.200 we just talked about. 282 00:13:38.279 --> 00:13:40.000 Okay, Step one, walk us through it. 283 00:13:40.360 --> 00:13:43.559 Step one is establishing the baseline prediction. This is where 284 00:13:43.600 --> 00:13:45.720 they look at single mutations. 285 00:13:45.519 --> 00:13:47.480 Just the individual alphabet swaps, right. 286 00:13:47.559 --> 00:13:51.200 They use existing historical data, or they run a quick 287 00:13:51.360 --> 00:13:54.159 initial round of lab experiments to see what happens to 288 00:13:54.240 --> 00:13:57.320 the protein when you change just one single amino acid 289 00:13:57.399 --> 00:13:58.799 at a time across the board. 290 00:13:59.080 --> 00:14:02.000 So this established the floor. It tells the system what 291 00:14:02.120 --> 00:14:04.720 each individual piece does in total isolation. 292 00:14:05.080 --> 00:14:08.799 Exactly. It gives the AI the additive model if there 293 00:14:08.799 --> 00:14:11.639 were no epistosis, if there were no interactions, this is 294 00:14:11.679 --> 00:14:13.879 literally all the data you would need. You'd just pick 295 00:14:13.919 --> 00:14:16.600 the five best single mutations and mash them together. 296 00:14:16.799 --> 00:14:19.519 But we know that doesn't work the garlic and the. 297 00:14:19.440 --> 00:14:23.200 Strawberries, right, So that limitation leads directly to step two, 298 00:14:23.440 --> 00:14:26.919 which is arguably the most innovative part of the entire protocol, 299 00:14:27.120 --> 00:14:28.440 the epistatic step. 300 00:14:28.559 --> 00:14:30.000 And this is where they actually go back into the 301 00:14:30.000 --> 00:14:34.000 wet lab. Right. They don't just simulate this on a supercomputer. 302 00:14:33.519 --> 00:14:36.279 Yet, No, And that is a critical point. AI and 303 00:14:36.360 --> 00:14:40.960 biology is completely useless without high quality physical data. So 304 00:14:41.039 --> 00:14:44.879 in step two they physically synthesize and chest a library 305 00:14:44.879 --> 00:14:47.240 of proteins that have exactly two mutations. 306 00:14:47.320 --> 00:14:49.840 Wait, why just two? If the goal is a protein 307 00:14:49.919 --> 00:14:52.919 with ten mutations, why not generate data on groups of 308 00:14:52.919 --> 00:14:53.559 three or four? 309 00:14:53.759 --> 00:14:56.919 Because pairs are the fundamental building block of interaction. If 310 00:14:56.960 --> 00:15:00.279 you can deeply understand how mutation A affects mutation B, 311 00:15:00.519 --> 00:15:03.720 and how B effects C exactly, you start to build 312 00:15:03.759 --> 00:15:06.279 a web of logic you don't actually need to physically 313 00:15:06.320 --> 00:15:09.519 test every trio or quartet if the computer has a 314 00:15:09.559 --> 00:15:13.600 really robust grasp of the pairwise physics. It's about data efficiency. 315 00:15:13.799 --> 00:15:16.639 Oh, that makes sense. So they are generating this massive, 316 00:15:16.759 --> 00:15:20.879 incredibly clean data set of just how these two specific 317 00:15:20.879 --> 00:15:23.000 things get along in a given protein structure. 318 00:15:23.240 --> 00:15:28.360 Yes, they are essentially teaching the machine learning model the biochemical. 319 00:15:27.720 --> 00:15:30.000 Rules of the road, things like spatial reasoning. 320 00:15:30.159 --> 00:15:35.519 Spatial reasoning charge polarity. The model learns, for example, when 321 00:15:35.559 --> 00:15:39.080 a positively charged amino acid is placed directly next to 322 00:15:39.120 --> 00:15:41.480 a negative one, they stabilize each other. 323 00:15:41.480 --> 00:15:42.159 Right, they attract. 324 00:15:42.240 --> 00:15:45.440 But when two incredibly bulky amino acids are mutated next 325 00:15:45.440 --> 00:15:48.360 to each other, they physically clash and the protein falls apart. 326 00:15:48.720 --> 00:15:53.159 The model learns the underlying grammar of the protein's physical structure. 327 00:15:53.120 --> 00:15:56.360 And once the model has internalized that grammar from the 328 00:15:56.399 --> 00:15:57.919 double mutation data, that. 329 00:15:57.840 --> 00:16:01.320 Brings us to step three. High order. Now that the 330 00:16:01.320 --> 00:16:04.559 model truly understands the interactions, you ask it too. 331 00:16:04.399 --> 00:16:06.799 Extrapoly, You take off the training wheels. 332 00:16:06.440 --> 00:16:09.320 You do you say, okay, computer, you know exactly how 333 00:16:09.360 --> 00:16:14.200 these pieces interact in pairs. Now mathematically predict what happens 334 00:16:14.240 --> 00:16:17.600 if I put five of them together or seven or 335 00:16:17.679 --> 00:16:19.399 then this is the massive leap. 336 00:16:19.799 --> 00:16:21.840 This is what the paper refers to when it talks 337 00:16:21.840 --> 00:16:24.799 about condensing iterative cycles into a single round. 338 00:16:24.960 --> 00:16:29.399 It is instead of doing five years of mutate test, select, repeat, 339 00:16:29.480 --> 00:16:32.600 over and over, you do one single round of baseline 340 00:16:32.639 --> 00:16:35.639 and pairwise data gathering. You train the model on that 341 00:16:35.720 --> 00:16:39.279 specific protein's rules, and then the model just spits out 342 00:16:39.279 --> 00:16:42.960 a list of highly complex, multi mutated designs that it 343 00:16:43.080 --> 00:16:45.000 confidently predicts will be successful. 344 00:16:45.080 --> 00:16:47.399 And then you just go build those specific designs. 345 00:16:47.480 --> 00:16:50.519 Yes, instead of physically testing a million random variants in 346 00:16:50.519 --> 00:16:53.639 the lab to find a marginal improvement, you synthesize maybe 347 00:16:53.639 --> 00:16:56.320 the top twenty or fifty designs. The computer suggests. 348 00:16:56.440 --> 00:16:59.519 It sounds incredibly efficient on paper, but the proof is 349 00:16:59.559 --> 00:17:01.200 in the pudding, or I guess in this case, the 350 00:17:01.240 --> 00:17:03.639 proof is in the protein. Right, did it actually work? 351 00:17:03.720 --> 00:17:07.480 Because computer theory is one thing, but actual physical biology 352 00:17:07.559 --> 00:17:09.799 is notoriously messy and unpredictable. 353 00:17:09.880 --> 00:17:13.359 It did work brilliantly, And to prove it wasn't a fluke, 354 00:17:13.519 --> 00:17:17.200 the huslab didn't just pick one easy, well understood target. 355 00:17:17.920 --> 00:17:23.319 They tested the MLTI evolve framework on three completely different 356 00:17:23.359 --> 00:17:26.599 classes of proteins to prove it's a universal solution. 357 00:17:26.720 --> 00:17:28.759 Okay, let's go through these case studies because this is 358 00:17:28.799 --> 00:17:33.440 where rubber meets the road case steady a the autoimmune antibody. 359 00:17:33.680 --> 00:17:37.160 Right, So they took an existing antibody that is clinically 360 00:17:37.240 --> 00:17:41.440 relevant for treating autoimmune diseases. The engineering goal here was 361 00:17:41.559 --> 00:17:43.240 affinity maturation. 362 00:17:43.000 --> 00:17:45.559 Meaning they wanted it to bind tighter to its target. 363 00:17:45.680 --> 00:17:49.039 Exactly. If the antibody binds tighter, it's more effective. And 364 00:17:49.079 --> 00:17:51.599 if it's more effective, you can give the patient a lower. 365 00:17:51.359 --> 00:17:53.440 Dose, which means fewer side effects. 366 00:17:53.200 --> 00:17:57.640 Fewer side effects, lower manufacturing costs, better patient outcomes. So 367 00:17:57.720 --> 00:18:01.839 they ran THEMLTI evolve workflow. They gathered the pairwise data, 368 00:18:01.960 --> 00:18:04.039 trained the model, and the model predicted a set of 369 00:18:04.160 --> 00:18:06.960 variants with multiple simultaneous mutations, and. 370 00:18:07.000 --> 00:18:08.519 When they actually built them in the lab. 371 00:18:08.440 --> 00:18:11.960 They found variants that significantly outperformed the natural version. But 372 00:18:12.160 --> 00:18:15.759 here is the absolute picker. The individual mutations that made 373 00:18:15.839 --> 00:18:16.960 up that winning combination. 374 00:18:18.000 --> 00:18:20.039 Some of them, if you looked at them completely on 375 00:18:20.079 --> 00:18:23.240 their own, were neutral or even slightly detrimental. 376 00:18:23.319 --> 00:18:26.680 Wow. So the old step by step directed evolution method 377 00:18:26.680 --> 00:18:28.279 would have entirely missed them. 378 00:18:28.519 --> 00:18:32.839 Guaranteed, the traditional approach would have screened those single mutations 379 00:18:32.839 --> 00:18:35.759 in round one, seeing that they didn't improve binding, and 380 00:18:35.880 --> 00:18:36.880 thrown them in the trash. 381 00:18:37.119 --> 00:18:39.839 But maltii evolve saw the hidden connection. 382 00:18:40.200 --> 00:18:44.319 It saw that when you combined those specific detrimental mutations together, 383 00:18:44.720 --> 00:18:48.839 they created a completely new structural feature that locked onto 384 00:18:48.880 --> 00:18:53.839 the autoimmune target like a vice. It successfully navigated the epistosis. 385 00:18:54.000 --> 00:18:56.920 That is a massive validation of the whole concept. Okay, 386 00:18:57.079 --> 00:18:59.599 case study B. This is a buzzword pretty much everyone 387 00:18:59.599 --> 00:18:59.920 knows by now. 388 00:19:00.319 --> 00:19:05.319 Crisper, yes the genetic engineering tool. Specifically, the researchers were 389 00:19:05.359 --> 00:19:07.160 looking at the CAS proteins. 390 00:19:07.240 --> 00:19:09.599 These are the actual molecular scissors that cut the DNA 391 00:19:09.640 --> 00:19:10.720 strand right exactly. 392 00:19:11.039 --> 00:19:14.119 Christ is the guidance system and CAST nine or CAST 393 00:19:14.160 --> 00:19:15.240 twelve is the blade. 394 00:19:15.359 --> 00:19:16.839 Why do we need to engineer those? 395 00:19:16.880 --> 00:19:17.160 Though? 396 00:19:17.480 --> 00:19:21.559 I thought Crisper was already this miraculous, perfect tool. It's miraculous, 397 00:19:21.759 --> 00:19:25.359 but it is definitely not perfect. The natural CAST proteins 398 00:19:25.400 --> 00:19:29.079 were evolved by bacteria to fight off viruses. They weren't 399 00:19:29.079 --> 00:19:32.279 evolved to do precision gene therapy in human cells. Oh 400 00:19:32.279 --> 00:19:34.119 that makes sense, so they can be a bit sluggish, 401 00:19:34.480 --> 00:19:37.640