Pod Insights Lex Fridman Podcast #113 – Manolis Kellis: Human Genome and Evolutionary Dynamics About this episode Manolis Kellis is a professor at MIT and head of the MIT Computational Biology Group. He is interested in understanding the human genome from a computational, evolutionary, biological, and other cross-disciplinary perspectives.
Support this podcast by supporting our sponsors:
– Blinkist: https://blinkist.com/lex
– Eight Sleep: https://eightsleep.com/lex
– MasterClass: https://masterclass.com/lex
If you would like to get more information about this podcast go to https://lexfridman.com/ai or connect with @lexfridman on Twitter https://twitter.com/lexfridman, LinkedIn https://www.linkedin.com/in/lexfridman/, Facebook https://www.facebook.com/lexfridman, Medium https://medium.com/@lexfridman, or YouTube https://www.youtube.com/lexfridman where you can watch the video versions of these conversations. If you enjoy the podcast, please rate it 5 stars on Apple Podcasts https://podcasts.apple.com/us/podcast/artificial-intelligence/id1434243584, follow on Spotify https://open.spotify.com/show/2MAi0BvDc6GTFvKFPXnkCL, or support it on Patreon https://www.patreon.com/lexfridman.
Here’s the outline of the episode. On some podcast players you should be able to click the timestamp to jump to that time.
OUTLINE:
00:00 – Introduction
03:54 – Human genome
17:47 – Sources of knowledge
29:15 – Free will
33:26 – Simulation
35:17 – Biological and computing
50:10 – Genome-wide evolutionary signatures
56:54 – Evolution of COVID-19
1:02:59 – Are viruses intelligent?
1:12:08 – Humans vs viruses
1:19:39 – Engineered pandemics
1:23:23 – Immune system
1:33:22 – Placebo effect
1:35:39 – Human genome source code
1:44:40 – Mutation
1:51:46 – Deep learning
1:58:08 – Neuralink
2:07:07 – Language
2:15:19 – Meaning of life
Transcript 24,472 words · 3 speakers
Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.
0:00 Lex Fridman? The following is a conversation with Manolis Kellis. He's a professor at MIT and head of the MIT Computational Biology Group. He's interested in understanding the human genome from a computational, evolutionary, biological, and other cross-disciplinary perspectives. He has more big impactful papers and awards than I can list. But most importantly, he's a kind, curious, brilliant human being, and just someone I really enjoy talking to. His passion for science and life in general is contagious. The hours honestly flew by, and I'm sure we'll talk again on this podcast soon. Quick summary of the ads. Three sponsors: Blinkist, Aidsleep, and MasterClass.
0:43 Lex Fridman? Please consider supporting this podcast by going to blinkist.com/lex, aidsleep.com/lex, and signing up at masterclass.com/lex. Click the links, buy the stuff, get the discount. It's the best way to support this podcast. If you enjoy this thing, subscribe on YouTube, review it with five stars on Apple Podcasts, support it on Patreon, or connect with me on Twitter at @LexFriedman. As usual, I'll do a few minutes of ads now and never any ads in the middle that can break the flow of the conversation. This episode is supported by Blinkist, my favorite app for learning new things. Get it at blinkist.com/lex for a seven-day free trial and 25% off afterwards.
1:27 Lex Fridman? Blinkist takes the key ideas from thousands of nonfiction books and condenses them down into just 15 minutes that you can read or listen to. I'm a big believer in reading at least an hour every day. As part of that, I use Blinkist every day to try out a book I may otherwise never have a chance to read. And in general, it's a great way to broaden your view of the ideal landscape out there and find books that you may want to read more deeply. With Blinkist, you get unlimited access to read or listen to a massive library of condensed nonfiction books. Go to blinkist.com/lex to try it free for seven days and save 25% off your new subscription. That's blinkist.com/lex. Blinkist, spelled B-L-I-N-K-I-S-T.
2:15 Lex Fridman? This show is also sponsored by Eight Sleep and its Pod Pro mattress that you can check out at eightsleep.com/lex to get $200 off. It controls temperature with an app. It can cool down to as low as 55 degrees on each side of the bed separately. Research shows that temperature has a big impact on the quality of our sleep. Anecdotally, as been true for me, it's truly been a game changer. I love it. The Pod Pro is packed with sensors that track heart rate, heart rate variability, and respiratory rate, showing it all in their app. The app's health metrics are amazing, but the cooling alone is honestly worth the money. Check it out at eightsleep.com/lex to get $200 off. This show is also sponsored
2:58 Lex Fridman? by MasterClass. Sign up at masterclass.com/lex to get a discount and to support this podcast. When I first heard about MasterClass, I thought it was too good to be true. For $180 a year, you get an all-access pass to watch courses from, to list some of my favorites, Chris Hatfield on space exploration, Neil deGrasse Tyson on scientific thinking and communication, Will Wright, one of my favorite game designers, Carlos Santana, one of my favorite guitar players, Garry Kasparov, of course, the greatest chess player of all time, I'm not biased, Daniel Negrono on poker, and many more. Chris Hatfield explaining how rockets work and the experience of being launched into space alone is worth the money. By the way, you
3:40 Lex Fridman? can watch it on basically any device. Once again, sign up at masterclass.com/lex to get a discount and to support this podcast. And now, here's my conversation with Manolis Kellis.
4:11 Lex Fridman? What to you is the most beautiful aspect of the human genome?
4:15 Manolis Kellis? Don't get me started.
4:19 Manolis Kellis? So.
4:19 Lex Fridman? We've got time.
4:21 Manolis Kellis? The first answer is that the beauty of genomes transcends humanity. So it's not just about the human genome. Genomes in general are amazingly beautiful. And again, I'm obviously biased. So in my view, the way that I like to introduce the human genome and the way that I like to introduce genomics to my class is by telling them, you know, we're not the inventors of the first digital computer. We are the descendants of the first digital computer. Basically, life is digital, and that's absolutely beautiful about life. The fact that at every replication step, you don't lose any information because that information is digital. If it was analog, if it was just protein concentrations, you'd lose it after a few generations. It would just dissolve away.
5:03 Manolis Kellis? And that's what the ancients didn't understand about inheritance. The first person to understand digital inheritance was Mendel, of course. And his theory, in fact, stayed in a bookshelf for like 50 years while Darwin was getting famous about natural selection. But the missing component was this digital inheritance, the mechanism of evolution that Mendel had discovered. So that aspect, in my view, is the most beautiful aspect, but it transcends all of life.
5:31 Lex Fridman? And can you elaborate maybe the inheritance part? What was the, what was the key thing that the ancients didn't understand?
5:39 Manolis Kellis? So the very theory of inheritance as discrete units, you know, throughout the life of Mendel and well after his writing, people thought that his pea experiments were just a little fluke, that they were just, you know, a little exception that would normally not even apply to humans. That basically what they saw is this continuum of eye color, this continuum of skin color, this continuum of hair color, this continuum of height. And all of these continuums did not fit with a discrete type of inheritance that Mendel was describing. But what's unique about genomics and what's unique about the genome is really that there are two copies and that you get a combination of these.
6:23 Manolis Kellis? But for every trait, there are dozens of contributing variables. And it was only Ronald Fisher in the 20th century that basically recognized that even five Mendelian traits would add up to a continuum-like inheritance pattern. And he, you know, wrote a series of papers that still are very relevant today about sort of this Mendelian inheritance of continuum-like traits. And I think that that was the missing step in inheritance. So well before the discovery of the structure of DNA, which is again another amazingly beautiful aspect, the double helix, what I like to call the most noble molecule of
7:04 Manolis Kellis? our time, is, you know, holds within it the secret of that discrete inheritance. But the conceptualization of discrete, you know, elements is something that precedes that.
7:16 Lex Fridman? So even though it's discrete, when it materializes itself into the actual traits that we see, it can be continuous, again, basically arbitrarily rich and complex.
7:27 Manolis Kellis? So if you have five genes that contribute to human height, and there aren't five, there is a thousand, if there's only five genes and you inherit some combination of them and everyone makes you two inches taller or two inches shorter, it'll look like a continuum trait, a continuous trait. But instead of five, there are thousands and every one of them contributes to less than one millimeter. We change in height more during the day than each of these genetic variants contributes. So by the evening, you're shorter than you were, you woke up with.
8:00 Lex Fridman? Isn't that weird then that we're not more different than we are? Why are we all so similar if there's so much possibility to be different?
8:09 Manolis Kellis? Yeah, so there are selective advantages to being medium. If you're extremely tall or extremely short, you run into selective disadvantages. So you have trouble breathing, you have trouble running, you have trouble sitting if you're too tall. If you're too short, you might, I don't know, have other selective pressures acting against that. If you look at natural history of human population, there's actually selection for height. In Northern Europe and selection against height in Southern Europe. So there might actually be advantages to actually being not not super tall. And if you look across the entire human population, you know, for many, many traits, there's a lot of push towards the middle.
8:49 Manolis Kellis? Balancing selection is, you know, the usual term for selection that sort of seeks to not be extreme and to sort of have a combination of alleles that sort of, you know, keep recombining. And if you look at, you know, mate selection, super, super tall people will not tend to sort of marry super, super tall people. Very often you see these couples that are kind of compensating for each other. And the best predictor of the kid's age is very often just take the average of the two parents and then adjust for sex and boom, you get it. It's extremely heritable.
9:25 Lex Fridman? Let me ask, you kind of took a step back to the genome outside of just humans. But is there something that you find beautiful about the human genome specifically?
9:35 Manolis Kellis? So I think the genome, if more people understood the beauty of the human genome, there would be so many fewer wars, so much less anger in the world. I mean, what's really beautiful about the human genome is really the variation that teaches us both about individuality and about similarity. So any two people on the planet are 99.9% identical. How can you fight with someone who's 99.9% identical to you? It's just counterintuitive. And yet, any two siblings of the same parent differ in millions of locations. So every one of them is basically two to the million unique from any pair of parents,
10:20 Manolis Kellis? let alone any two random parents on the planet. So that's, I think, something that teaches us about sort of the nature of humanity in many ways, that every one of us is as unique as any star and way more unique in actually many ways. And yet, we're all brothers and sisters.
10:38 Lex Fridman? Yeah, just like stars, most of it is just fusion reactions.
10:43 Manolis Kellis? Yeah, you only have a few parameters to describe stars.
10:45 Lex Fridman? Yeah, exactly.
10:46 Manolis Kellis? You know, mass size, initial size and, you know, stage of life. Whereas for humans, it's, you know, thousands of parameters scattered across our genome. So the other thing that makes humans unique, the other thing that makes inheritance unique in humans is that most species inherit things vertically. Basically, instinct is a huge part of their behavior. The way that, you know, I mean, with my kids, we've been watching this nest of birds with two little eggs, you know, outside our window for the last few months, for the last few weeks as they've been growing. And there's so much behavior that's hardcoded.
11:29 Manolis Kellis? Birds don't just learn as they grow. They don't, you know, there's no culture. Like, a bird that's born in Boston will be the same as a bird that's born in California. So there's not as much inheritance of ideas, of customs. A lot of it is hardcoded in their genome. What's really beautiful about the human genome is that if you take a person from today and you place them back in ancient Egypt, or if you take a person from ancient Egypt and you place them here today, they will grow up to be completely normal. That is not genetics. This is the other type of inheritance in humans. So on one hand, we have genetic inheritance, which is vertical from your parents down.
12:13 Manolis Kellis? On the other hand, we have horizontal inheritance, which is the ideas that are built up at every generation are horizontally transmitted. And the huge amount of time that we spend in educating ourselves, a concept known as neoteny, neo for newborn and then teny for holding. So if you look at humans, I mean, the little birds that were, you know, eggs two weeks ago and now one of them has already flown off, the other one's ready to fly off. In two weeks, they're ready to just fend for themselves. Humans? Sixteen years?
12:48 Manolis Kellis? Eighteen years, twenty-four, getting out of college.
12:50 Lex Fridman? I'm still learning. So that's so fascinating, this picture of a vertical and a horizontal. When you talk about the horizontal, is it in the realm of ideas?
12:58 Manolis Kellis? Exactly.
12:59 Lex Fridman? Okay, so it's the actual social interactions?
13:02 Manolis Kellis? That's exactly right. That's exactly right. So basically, the concept of neotony is that you spend acquiring characteristics from your environment in an extremely malleable state of your brain and the wiring of your brain for a long period of your life. Compared to primates, we are useless. You take any primate at seven weeks and a human at seven weeks, we lose the battle. But at eighteen years, you know, all better off. Like, basically, our brain continues to develop in an extremely malleable form till very late. And this is what allows education. This is what allows the person from Egypt to do extremely well now.
13:41 Manolis Kellis? And the reason for that is that the wiring of our brain and the development of that wiring is actually delayed. So, you know, the longer you delay that, the more opportunity you have to pass on knowledge, to pass on concepts, ideals, ideas, from the parents to the child. And what's really absolutely beautiful about humans today is that that lateral transfer of ideas and culture is not just from uncles and aunts and teachers at school, but it's from Wikipedia and review articles on the web and thousands of journals that are sort of putting out information for free and podcasts and videocasts and all of that stuff, where
14:26 Manolis Kellis? you can basically learn about any topic, pretty much everything that would be in any super advanced textbook in a matter of days, instead of having to go to the Library of Alexandria and sail there to read three books and then sail for another few days to get to Athens and et cetera, et cetera, et cetera. So the democratization of knowledge and the spread, the speed of spread of knowledge, is what defines, I think, the human inheritance pattern.
14:55 Lex Fridman? So you sound excited about it. Are you also a little bit afraid or are you more excited? By the power of this kind of distributed spread of information. So you put it very kindly that most people are kind of using the internet in, you know, looking Wikipedia, reading articles, reading papers and so on. But if we, if we're honest, most people online, especially when they're younger, probably looking at five second clips on TikTok or whatever the new social network is. Are you, given this power of horizontal inheritance, are you optimistic? Or a little bit pessimistic about the, this new effect of the internet?
15:39 Lex Fridman? And democratization of knowledge on our, on our, what would you call this? This genome, like, would you, would you use the term genome, by the way?
15:48 Manolis Kellis? Yeah, yeah, I think, you know, we use the genome to talk about DNA. But very often we say, you know, I mean, when I'm Greek, so people ask me, hey, what's in the Greek genome? And I'm like, well, yeah, what's in the Greek genome is both our genes and also our ideas and our ideals and our culture.
16:03 Lex Fridman? The poetic meaning of the word.
16:05 Manolis Kellis? Exactly, exactly.
16:06 Lex Fridman? Yeah.
16:06 Manolis Kellis? Yeah. So I think that there's a beauty to the democratization of knowledge. The fact that you can reach as many people as, you know, any other person on the planet, and it's not who you are, it's really your ideas that matter, is a beautiful aspect of the internet. The, I think there's of course a danger of my ignorance is as important as your expertise. The fact that with this democratization comes the abolishment of respecting expertise. Just because you've spent, you know, 10,000 hours of your life studying, I don't know, human brain circuitry, why should I trust you?
16:51 Manolis Kellis? I'm just going to make up my own theories and they'll be just as good as yours. Is an attitude that, that sort of counteracts the beauty of the democratization. And I think that within our educational system and within the upbringing of our children, we have to not only teach them knowledge, but we have to teach them the means to get to knowledge. And that, you know, it's very similar to sort of you fish, you catch a fish for a man for one day, you fed them for one day, you teach them how to fish, you fed them for the rest of their life. So instead of just gathering the knowledge they need for any one task, we can just tell them, all right. Here's how you Google it. Here's how you figure out what's real and what's not. Here's how you check the sources. Here's how you form a basic opinion for yourself.
17:36 Manolis Kellis? And I think that inquisitive nature is paramount to being able to sort through this huge wealth of knowledge. So you need a basic educational foundation based on which you can then add on the sort of domain-specific knowledge. But that basic educational foundation should not just be knowledge, but it should also be epistemology, the way to acquire knowledge.
18:04 Lex Fridman? I'm not sure any of us know how to do that in this modern day. We're actually learning. One of the big surprising things to me about the coronavirus, for example, is that Twitter has been one of the best sources of information. Basically, like building your own network of experts, as opposed to the traditional centralized expertise of the WHO and the CDC or maybe any one particular respectable person at the top of a department or some kind of institution. You instead look at 10, 20, hundreds of people, some of whom are young kids that are incredibly
18:49 Lex Fridman? good at aggregating data and plotting and visualizing that data. That's been really surprising to me. I don't know what to make of it. I don't know how that matures into something stable. I don't know if you have ideas. If you were to try to explain to your kids of where should you go to learn about coronavirus, what would you say?
19:13 Manolis Kellis? It's such a beautiful example. And I think the current pandemic and the speed at which the scientific community has moved in the current pandemic, I think, exemplifies this horizontal transfer and the speed of horizontal transfer of information. The fact that
19:30 Manolis Kellis? You know, the genome was first sequenced in early January. The first sample was obtained December 29, 2019. A week after the publication of the first genome sequence, Moderna had already finalized its vaccine design and was moving to production. I mean, this is phenomenal. The fact that we go from not knowing what the heck is killing people in Wuhan to wow, it's SARS-CoV-2 and here's the set of genes. Here's the genome. Here's the sequence. Here are the polymorphisms, et cetera. In the matter of weeks is phenomenal. In that incredible pace of transfer of knowledge, there have been many mistakes. So, you know, some of those mistakes may have been politically motivated.
20:14 Manolis Kellis? Our other mistakes may have just been innocuous errors. Others may have been misleading the public for the greater good, such as don't wear masks because we don't want the masks to run out. I mean, that was very silly in my view and a very big mistake. But the the spread of knowledge from the scientific community was phenomenal. And some people will point out to bogus articles that snuck in and made the front page. Yeah, they did, but within 24 hours, they were debunked and went out of the front page. And I think that's that's the beauty of science today. The fact that it's not, oh, knowledge is fixed. It's the ability to embrace that nothing is permanent when it comes to knowledge, that everything is the current best hypothesis and the current best model
20:58 Manolis Kellis? that best fits the current data and the willingness to be wrong. The expectation that we're going to be wrong. And the celebration of success based on how long was I not proven wrong for? Rather than, wow, I was exactly right. Because no one is going to be exactly right with partial knowledge. But the arc towards perfection, I think, is so much more important than how far you are on your first step. And I think that's what sort of the current pandemic has taught us. The fact that, yeah, no, of course, we're going to make mistakes, but at least we're going to learn from those mistakes and become better and learn better and spread information better. So if I were to answer the question of where would you go to learn about coronavirus?
21:44 Manolis Kellis? First textbook, it all starts with a textbook. Just open up a chapter on virology and how coronaviruses work. Then some basic epidemiology and sort of how pandemics have worked in the past. What are the basic principles surrounding these first wave, second wave? Why do they even exist? Then understanding about growth, understanding about the R naught and RT at, you know, various time points. And then understanding the means of spread, how it spreads from person to person. Then how does it get into your cells? From when it gets into the cells, what are the paths that it takes? What are the cell types that express the particular ACE2 receptor? How is your immune system interacting with the virus?
22:26 Manolis Kellis? And once your immune system launches its defense, how is that helping or actually hurting your health? What about the cytokine storm? What are most people dying from? Why are the comorbidities and these risk factors even applying? What makes obese people respond more or elderly people respond more to the virus? While kids are completely, you know, you know, very often not even aware that they're spreading it. So the, you know, I think there's some basic questions that you would start from. And then, I'm sorry to say, but Wikipedia is pretty awesome.
23:02 Lex Fridman? Yeah.
23:02 Manolis Kellis? Google is pretty awesome.
23:04 Lex Fridman? It used to be a time, it used to be a time, maybe five years ago, I forget, I forget when, but people kind of made fun of Wikipedia for being an unreliable source. I never quite understood it. I thought from the early days, it was pretty reliable or better than a lot of the alternatives. But at this point, it's kind of like a solid, accessible survey paper on every subject ever.
23:27 Manolis Kellis? There's an ascertainment bias and a writing bias. So, so I think this, this is related to sort of people saying, oh, so many nature papers are wrong. And they're like, why would you publish in Nature? So many Nature papers are wrong. And my answer is, no, no, no. So many Nature papers are scrutinized. And just because more of them are being proven wrong than in other articles is actually evidence that they're actually better papers overall because they're being scrutinized at a rate much higher than any other journal. So if you basically judge Wikipedia by not the initial content, but by the number of revisions,
24:09 Lex Fridman? Yeah.
24:09 Manolis Kellis? then of course it's going to be the best source of knowledge eventually. It's still very superficial. You then have to go into the review papers, et cetera, et cetera, et cetera. But I mean, for most scientific projects, topics, it's extremely superficial. But it is quite authoritative because it is the place that everybody likes to criticize as being wrong.
24:28 Lex Fridman? You say that it's superficial. On a lot of topics that I've studied a lot of, I find it, I don't know if superficial is the right word.
24:41 Lex Fridman? Because superficial kind of implies that it's not correct.
24:44 Manolis Kellis? No, no. I don't mean any implication of it not being correct. It's just superficial. It's basically only scratching the surface. For depth, you don't go to Wikipedia. You go to the review articles.
24:55 Lex Fridman? But it can be profound in the way that articles rarely. One of the frustrating things to me about like certain computer science, like in the machine learning world, articles, they don't as often take the bigger picture view. You know, there's a kind of data set and you show that it works and you kind of show that here's an architectural thing that creates an improvement and so on, so forth. But you don't say, well, like, what does this mean for the nature of intelligence for future data sets? We haven't even thought about. Or if you were trying to implement this, like if we took this data set of 100,000 examples and scaled it to 100 billion examples, would this method like, like look at the bigger picture,
25:38 Lex Fridman? which is what a Wikipedia article would actually try to do, which is like, what does this mean in the context of? Computer, the broad field of computer vision or something like that.
25:49 Manolis Kellis? Yeah, yeah, no, I, I agree with you completely, like, but it depends on the topic. I mean, for some topics, there's been a huge amount of work. For other topics, it's just a stub. So, you know.
25:58 Lex Fridman? I got it.
25:59 Manolis Kellis? Yeah.
25:59 Lex Fridman? Well, yeah, actually, the, which we'll talk on, genomics was not.
26:05 Manolis Kellis? Yeah, it's very shallow. Yeah, yeah. It's not wrong, it's just shallow.
26:08 Lex Fridman? It's shallow.
26:09 Manolis Kellis? Yeah. Every time I criticize something, I should feel partly responsible. Basically, if more people from my community went there and edited, it would not be shallow. It's just that there's different modes of communication in different fields. And in some fields, the experts have embraced Wikipedia. In other fields, it's relegated and perhaps the reason is that if it wasn't any better to start with, people would invest more time. But if it's not great to start with, then you need a few initial pioneers who will basically go in and say, ah, enough, we're just gonna fix that. And then I think it'll catch on much more.
26:46 Lex Fridman? So if it's okay, before we go on to genomics, can we linger a little bit longer on the the beauty of the human genome? You've given me a few notes. What else, what else do you find beautiful about the human genome?
26:58 Manolis Kellis? So the last aspect of what makes the human genome unique, in addition to the, you know, similarity and the differences and individuality, is that so very early on, people would basically say, oh, you don't do that experiment in human. You have to learn about that in fly. Or you have to learn about that in yeast first, or in mouse first, or in a primate first. And the human genome was in fact relegated to sort of, oh, the last place that you're gonna go to learn something new. That has dramatically changed. And the reason that changed is human genetics. We are the species in the planet that's the most studied right now.
27:41 Manolis Kellis? It's embarrassing to say that, but this was not the case a few years ago. It used to be, you know, first viruses, then bacteria, then yeast, then the fruit fly and the worm, then the mouse, and eventually, human was very far last.
27:59 Lex Fridman? So it's embarrassing that it took us this long to focus on it, or the...
28:03 Manolis Kellis? It's embarrassing that the model organisms have been taken over because of the power of human genetics. That right now, it's actually simpler to figure out the phenotype of something by mining this massive amount of human data than by going back to any of the other species. And the reason for that is that if you look at the natural variation that happens in a population of 7 billion, you basically have a mutation in almost every nucleotide. So every nucleotide you want to perturb, you can go find a living, breathing human being and go test the function of that nucleotide by sort of searching the database and finding that person.
28:39 Lex Fridman? Wait, why is that embarrassing? It's a beautiful data set.
28:41 Manolis Kellis? It's a beautiful data set. It's embarrassing for the for the model organism.
28:46 Lex Fridman? For the flies.
28:47 Manolis Kellis? Yeah, exactly.
28:47 Lex Fridman? I mean, do you do you feel, on a small tangent, is there something of of value in the in the genome of a fly? And are there these model organisms that you miss that we wish we would have would be looking at deeper?
29:04 Manolis Kellis? So directed perturbation, of course. So I think the place where the the place where humans are still lagging is the fact that in an animal model, you can go and say, well, let me knock out this gene completely.
29:15 Lex Fridman? Got it.
29:15 Manolis Kellis? And let me knock out these three genes completely. And as I said, the moment you get into combinatorics, it's something you can't do in the human because there just simply aren't enough humans on the planet. And again, let me be honest, we haven't sequenced all 7 billion people. It's not like we have every mutation, but we know that there's a carrier out there. So if you look at the trend with and the speed with which human genetics has progressed, we can now find thousands of genes involved in human cognition, in human psychology, in the emotions and the feelings that we used to think are uniquely learned. Turns out there's a genetic basis to a lot of that. So the, you know, the the the human genome has continued to elucidate
29:59 Manolis Kellis? through these studies of genetic variation so many different processes that we previously thought were, you know... Something that like free will. Free will is this beautiful concept that humans have had for a long time. You know, in the end, it's just a bunch of chemical reactions happen in your brain. And the particular abundance of receptors that you have this day, based on what you ate yesterday, or that you have been wired with based on, you know, your parents and your upbringing, et cetera, determines a lot of that quote unquote free will component to, you know, sort of narrow and narrow scale, you know, sort of slices.
30:38 Lex Fridman? So how much on that point, how much freedom do you think we have to escape the constraints of our genome? You're making it sound like more and more we're discovering that our genome is actually has the, a lot of the story already encoded into it. How much freedom do we have?
30:56 Manolis Kellis? I, so let me, let me describe what that freedom would look like. That freedom would be my saying, oh, I'm going to resist the urge to eat that apple because I choose not to. But there are chemical receptors that made me not resist the urge to prove my individuality and my free will by resisting the apple. So then the next question is, well, maybe now I'll resist the urge to resist the apple and I'll go for the chocolate instead to prove my individuality. But then what about those other receptors that, you know?
31:32 Lex Fridman? That, that might be all encoded in there.
31:34 Manolis Kellis? So it's kicking the bucket down the road and basically saying, well, your choice will may have actually been driven by other things that you actually are not choosing. So that's why it's very hard to answer that question.
31:47 Lex Fridman? It's hard to know what to do with that. I mean, if, if the genome has, if, if there's not much freedom, it's a.
31:55 Manolis Kellis? It's the butterfly effect. It's basically that in the short term, you can predict something extremely well by knowing the current state of the system. But a few steps down, it's very hard to predict based on the current knowledge. Is that because the system is truly free? When I look at weather patterns, I can predict the next 10 days. Is it because the weather has a lot of freedom? And after 10 days, it chooses to do something else? Or is it because in fact, the system is fully deterministic? And there's just a slightly different magnetic field of the earth, slightly more energy arriving from the sun, a slightly different spin of the gravitational pull of Jupiter that is now causing, you know, all kinds of tides and slight deviation of the moon, et cetera. Maybe all of that can be fully modeled.
32:39 Manolis Kellis? Maybe the fact that China is emitting a little more carbon today is actually going to affect the weather in, you know, Egypt in three weeks. And all of that could be fully modeled. In the same way, if you take a complete view of a human being now, you know, I model everything about you. The question is, can I predict your next step? Probably. But at how far? And if it's a little further, is that because of stochasticity and sort of chaos properties of unpredictability of beyond a certain level? Or was that actually true free will?
33:15 Lex Fridman? Yeah, so the number of variables might be so, you might need to build an entire universe to be able to model.
33:22 Manolis Kellis? To simulate a human. And then maybe that human will be fully simulatable. But maybe aspects of free will will exist. And where's that free will coming from? It's still coming from the same neurons, or maybe from a spirit inhabiting these neurons. But again, you know, it's very difficult empirically to sort of evaluate where this free will begin and sort of chemical reactions and electric signals and, you know.
33:43 Lex Fridman? So on that topic, let me ask the most absurd question that most MIT faculty roll their eyes on. But what do you think about the simulation hypothesis? And the idea that we live in a simulation?
33:57 Manolis Kellis? I think it's complete BS.
34:00 Lex Fridman? Okay.
34:01 Manolis Kellis? There's no empirical evidence, absolutely not.
34:04 Lex Fridman? Not in terms of empirical evidence or not, but in terms of a thought experiment, does it help you think about the universe? I mean, so if you look at the genome, it's encoding a lot of the information that is required to create some of the beautiful human complexity that we see around us. It's an interesting thought experiment. How much? You know, parameters do we need to have in order to model some, you know, this full human experience? Like, if we were to build a video game,
34:34 Manolis Kellis? Yeah.
34:34 Lex Fridman? how hard it would be to build a video game that's like convincing enough and fun enough? And, you know, it has consistent laws of physics, all that stuff. It's not interesting to use this thought experiment.
34:47 Manolis Kellis? I mean, it's cute, but, you know, it's Occam's razor. I mean, what's more realistic? The fact that you're actually a machine or that you're, you know, a person? What's, you know, the fact that all of my experiences exist inside the chemical molecules that I have? Or that somebody's actually, you know, simulating all that? I mean,
35:06 Lex Fridman? Well, you did refer to humans as a digital computer earlier, so.
35:09 Manolis Kellis? Of course, of course, but that does not.
35:11 Lex Fridman? It's a kind of a machine, right?
35:12 Manolis Kellis? I know, I know, but I think the probability of all that is nil. And let the machines wake me up and just terminate me now if it's not.
35:24 Manolis Kellis? I challenge you, machines.
35:25 Lex Fridman? They're gonna, they're gonna wait a little bit to see what you're gonna do next. It's fun. It's fun to watch, especially the clever humans. What's the difference to you between the way a computer stores information and the human genome stores information? So you also have roots and your work. Would you say you're, when you introduce yourself at a bar?
35:48 Manolis Kellis? It depends who I'm talking to.
35:51 Lex Fridman? Would you say it's computational biology? Do you, do you reveal your expertise in computers?
36:00 Manolis Kellis? It depends who I'm talking to, truly. I mean, basically, if I meet someone who's in computers, I'll say, oh, I'm a professor in computer science. If I meet someone who's in engineering, I say, computer science and electrical engineering. If I meet someone in biology, I'll say, hey, I work in genomics. If I meet someone in medicine, I'm like, hey, I work on, you know, genetics.
36:17 Lex Fridman? You're a fun person to meet at a bar. I got you, but, so.
36:20 Manolis Kellis? No, no, but what I'm trying to say is that I don't, I mean, there's no single attribute that I will define myself as. You know, there's a few things I know, there's a few things I study, there's a few things I have degrees on, and there's a few things that I grant degrees in. And, you know, I, I publish papers across the whole gamut, you know, the whole spectrum of computation to biology, et cetera. I mean, I, the complete answer is that I use computer science to understand biology.
36:51 Manolis Kellis? So I'm a, you know, I develop methods in AI and machine learning, statistics, in algorithms, et cetera, but the ultimate goal of my career is to really understand biology. If these things don't advance our understanding of biology, I'm not as fascinated by them. Although there are some beautiful computational problems by themselves, I've sort of made it my mission to apply the power of computer science to truly understand the human genome, health, disease, you know, and the whole gamut of how our brain works, our body works, and all of that, which is so fascinating.
37:30 Lex Fridman? So the dream, there's not an equivalent sort of complementary dream of understanding the human biology in order to create an artificial life, an artificial brain, artificial intelligence that supersedes the intelligence and the capabilities of us humans?
37:47 Manolis Kellis? It's an interesting question, it's a fascinating question. So, understanding the human brain is undoubtedly coupled to how do we make better AI? Because so much of AI has in fact been inspired by the brain. It may have taken 50 years since the early days of neural networks till we have, you know, all of these amazing progress that we've seen with, you know, deep belief networks and, you know, all of these advances in Go, in chess, in image synthesis, in deep fakes, in you name it.
38:28 Manolis Kellis? And, but, but the underlying architecture is very much inspired by the human brain, which actually posits a very, very interesting question.
38:39 Manolis Kellis? Why are neural networks performing so well? And they perform amazingly well. Is it because they can simulate any possible function? And the answer is no, no, they simulate a very small number of functions. Is it because they can simulate every possible function in the universe? And that's where it gets interesting. The answer is actually, yeah, a little closer to that. And here's where it gets really fun. If you look at human brain and human cognition, it didn't evolve in a vacuum. It evolved in a world with physical constraints.
39:15 Manolis Kellis? Like the world that inhabits us. It is the world that we inhabit. And if you look at our senses, what do they perceive? They perceive different, you know, parts of the electromagnetic spectrum. You know, the hearing is just different movements in air. The touch, et cetera. I mean, all of these things, we've built it intuitions for the physical world that we inhabit. And our brains, and the brains of all animals, evolved for that world. And the AI systems that we have built happen to work well with images of the type that we encounter in the physical world that we inhabit. Whereas if you just take noise and you add random signal that doesn't match anything in
40:00 Manolis Kellis? our world, neural networks will not do as well. And that actually, basically, has this whole loop around this, which is, this was designed by studying our own brain, which was evolved for our own world, and they happen to do well in our own world, and they happen to make the same types of mistakes that humans make many, many times. And of course, you can engineer images by adding just the right amount of, you know, sort of pixel deviations to make a zebra look like a bamboo and stuff like that, or like a table. But ultimately, the undocumented images at least are very often, you know, mistaken,
40:42 Manolis Kellis? I don't know, between muffins and dogs, for example, in the same way that humans make those mistakes. So it's, it's, you know, there's no doubt in my view that the more we understand about the tricks that our human brain has evolved to understand the physical world around us, the more we will be able to bring new computational primitives in our AI systems to, again, better understand not just the world around us, but maybe even the world inside us, and maybe even the computational problems that arise from new types of data that we haven't been exposed to, but are yet inhabiting the same universe that we live in. With a very tiny little subset of functions from all possible mathematical functions.
41:25 Lex Fridman? Yeah, and that small subset of functions is all that matters to us humans, really. That's what makes.
41:29 Manolis Kellis? It's all that has mattered so far. And even within our scientific realm, it's all that seems to continue to matter. But, I mean, I always like to think about our senses and how much of the physical world around us we perceive. And if you look at the LIGO experiment over the last, you know, year and a half has been all over the news. What did LIGO do? It created a new sense for human beings, a sense that has never been sensed in the history of our planet. Gravitational waves have been traversing the Earth since its creation a few billion years ago.
42:13 Manolis Kellis? Life has evolved senses to sense things that were never before sensed. Light was not perceived by early life. No one cared. And eventually photoreceptors evolved and, you know, the ability to sense colors by sort of catching different parts of that electromagnetic spectrum. And hearing evolved and touch evolved, et cetera. But no organism evolved a way to sense neutrinos floating through Earth or gravitational waves flowing through Earth, et cetera. And I find it so beautiful in the history of not just humanity, but life on the planet, that we are now able to capture additional signals
42:57 Manolis Kellis? from the physical world than we ever knew before. And axions, for example, have been all over the news in the last few weeks. The concept that we can capture and perceive more of that physical world is as exciting as the fact that we were blind to it is traumatizing before.
43:21 Lex Fridman? Right.
43:21 Manolis Kellis? Because that also tells us, you know, we're in 2020.
43:26 Manolis Kellis? Picture yourself in thirty twenty or in twenty, you know.
43:29 Lex Fridman? What new senses might we discover?
43:32 Manolis Kellis? Is it, you know, could it be that we're missing nine tenths of physics? That like, there's a lot of physics out there that we're just blind to, completely oblivious to it.
43:44 Lex Fridman? Yeah.
43:44 Manolis Kellis? And yet they're permeating us all the time.
43:46 Lex Fridman? Yeah, so it might be right in front of us.
43:48 Manolis Kellis? So, so when you're thinking about premonitions.
43:51 Lex Fridman? Yeah.
43:52 Manolis Kellis? Yeah, a lot of that is ascertainment bias, like, yeah, every, you know, every now and then you're like, oh, I remember my friend. And then my friend doesn't appear and I'll forget that I remembered my friend. But every now and then my friend will actually appear and I'm like, oh my God, I thought about you a minute ago. You just called me. That's amazing. So, you know, some of that is this, but some of that might be that there are, within our brain, sensors for waves that we emit that we're not even aware of. And this whole concept of when I hug my children, there's such an emotional transfer there that we don't comprehend. I mean, sure, yeah, of course, we're all like hardwired for all kinds of touchy feely things between parents and kids. It's beautiful between partners. It's beautiful, et cetera.
44:37 Manolis Kellis? But then there are intangible aspects of human communication that I don't think it's unfathomable that our brain has actually evolved ways and sensors for it that we just don't capture. We don't understand the function of the vast majority of our neurons. And maybe our brain is already sensing it, but even worse, maybe our brain is not sensing it at all. And we're oblivious to this until we build a machine that suddenly is able to sort of capture so much more of what's happening in the natural world.
45:07 Lex Fridman? So what you're saying is we're going, physics is going to discover a sensor for love.
45:14 Manolis Kellis? And maybe, maybe dogs are off scale for that. And we've been, you know, we've been oblivious to it the whole time because we didn't have the right sensor.
45:22 Lex Fridman? Yeah.
45:22 Manolis Kellis? And now you're going to have a little wrist that says, oh my God, I feel all this love in the house. I see, I sense a disturbance in the forest.
45:30 Lex Fridman? And dogs and cats will have zero.
45:32 Manolis Kellis? None.
45:33 Lex Fridman? None.
45:33 Manolis Kellis? None.
45:37 Lex Fridman? But let's take a step back to our unfortunate place.
45:41 Manolis Kellis? To one of the 400 topics that we had actually planned.
45:46 Lex Fridman? Yeah, but to our sad time in 2020, when we only have just a few sensors and very primitive early computers. So in your you have a foot in computer science and a foot in biology. In your sense, how do computers represent information differently than like the genome or biological systems?
46:09 Manolis Kellis? So first of all, let me let me correct that. No, we're in an amazing time in 2020. Computer science is totally awesome and physics is totally awesome. And we have understood so much of the natural world than ever before. So I am extremely grateful and feeling extremely lucky to be living in the time that we are. Because, you know, first of all, who knows when the asteroid will hit? And second, you know, of all times in humanity, this is probably the best time to be a human being. And this might actually be the best place to be a human being. So anyway, you know, for for anyone who loves science, this is this is it. This is awesome. It's a great time.
46:53 Lex Fridman? At the same time, just a sweet comment. All I meant is that if we look several hundred years from now, and we end up somehow not destroying ourselves.
47:04 Manolis Kellis? Yeah.
47:05 Lex Fridman? People will probably look back at this time in computer science and at your work of Manolis at MIT.
47:12 Manolis Kellis? As infantile.
47:13 Lex Fridman? As infantile and silly and how ignorant it all was.
47:16 Manolis Kellis? I like to joke very often with my students that, you know, we've written so many papers. We've published so much. We've been citing so much. And every single time I tell my students, you know, the best is ahead of us. What we're working on now is the most exciting thing I've ever worked on. So in a way, I do have this sense of, yeah, even the papers I wrote 10 years ago, they were awesome at the time. But I'm so much more excited about where we're heading now. And I don't mean to minimize any of the stuff we've done in the past, but You know, there's just this sense of excitement about what you're working on now that as soon as a paper is submitted, it's like, ugh, it's old. Like, you know, I can't talk about that anymore.
47:54 Lex Fridman? At the same time, you're not, you probably are not going to be able to predict what are the most impactful papers and ideas. When people look back 200 years from now at your work, what would be the most exciting papers? And it may very well be not the thing that you expected. Or
48:12 Manolis Kellis? Yeah.
48:13 Lex Fridman? The things you got awards for, or, you know.
48:15 Manolis Kellis? This might be true in some fields. I don't know, I feel slightly differently about it in our field. I feel that I kind of know what are the important ones. And there's a very big difference between what the press picks up on and what's actually fundamentally important for the field. And I think for the fundamentally important ones, we kind of have a pretty good idea what they are. And it's hard to sometimes get the press excited about the fundamental advances. But, you know, we take what we get and celebrate what we get. And sometimes, you know, one of our papers, which was in a minor journal, made the front page of Reddit and suddenly had like hundreds of thousands of views. Even though it was in a minor journal, because, you know, somebody pitched it the right way, that it suddenly caught everybody's attention. Whereas other papers that are sort of truly fundamental, you know, we have a hard time
49:00 Manolis Kellis? getting the editors even excited about them when so many hundreds of people are already using the results and building upon them. So I do appreciate that there's a discrepancy between the perception and the perceived success and the awards that you get for various papers. But I think that fundamentally, I know that, you know, some paper, I'm, so, so.
49:21 Lex Fridman? So is there a paper that you're most proud of? See now, you just, you trapped yourself.
49:26 Manolis Kellis? No, no, no, no.
49:27 Lex Fridman? Is there a line of work that you, you have a sense is really powerful that you've done to date? You've done so much work in so many directions, which is interesting. Is there something where you think is quite special?
49:42 Manolis Kellis? I I mean, it it's like asking me to say which of my three children I love best. I mean.
49:51 Manolis Kellis? Exactly. So I mean, it it and it's such a gimme question that is so so difficult not to brag about the awesome work that my team and my students have done. And I'll I'll just mention a few off the top of my head. I mean, basically, there's a few landmark papers that I think have shaped my scientific path. And, you know, I like to somehow describe it as a linear continuation of one thing led to another, led to another, led to another. And, you know, it kind of all started with.
50:28 Manolis Kellis? Skip, skip, skip, skip, skip. Let me try to start somewhere in the middle.
50:32 Manolis Kellis? So my my first PhD paper was the first comparative analysis of multiple species. So multiple complete genomes. So for the first time, we we basically developed the concept of genome-wide evolutionary signatures, the fact that you could look across the entire genome and understand how things evolve. And from these signatures of evolution, you could go back and study any one region and say, that's a protein coding gene, that's an RNA gene, that's a regulatory motif, that's a, you know, binding site and so on and so forth. So. I'm sorry, so comparing different. Different species. Species of the same. So. So take human, mouse, rat and dog. Yep.
51:15 Manolis Kellis? You know, they're all animals, they're all mammals, they're all performing similar functions with their heart, with their brain, with their lungs, et cetera, et cetera, et cetera. So there's many functional elements that make us uniquely mammalian. And those mammalian elements are actually conserved. Ninety-nine percent of our genome does not code for protein. 1% goes for protein. The other 99%, we frankly didn't know what it does until we started doing these comparative genomic studies. So basically, these series of papers in my career have basically first developed that concept of evolutionary signatures and then apply them to yeast, apply them to flies, apply them to four mammals, apply them to 17 fungi, apply them to 12 rhesus species,
52:00 Manolis Kellis? apply them to then 29 mammals and now 200 mammals.
52:03 Lex Fridman? So, so can we So the evolutionary signatures, it seems like such a fascinating idea. I'm probably going to linger on your early PhD work for two hours. But what is, how can you reveal something interesting about the genome by looking at the multiple, multiple species and looking at the evolutionary signatures?
52:28 Manolis Kellis? Yeah, so you basically align the matching regions. So everything evolved from a common ancestor way, way back. And mammals evolved from a common ancestor about 60 million years back. So after, you know, the meteor that killed off the dinosaurs landed alleged near Machu Picchu. We know the crater. It didn't allegedly land.
52:58 Lex Fridman? That was the aliens, okay.
52:59 Manolis Kellis? No, just slightly north of Machu Picchu in the Gulf of Mexico, there's a giant hole that meteor.
53:05 Lex Fridman? By the way, is that, sorry, is that definitive to people? Have people conclusively figured out what killed the dinosaurs?
53:15 Manolis Kellis? I think so.
53:16 Lex Fridman? So it was a meteor?
53:17 Manolis Kellis? Well, you know, volcanic activity, all kinds of other stuff is coinciding. But the meteor is pretty unique and we now have.
53:27 Lex Fridman? Terrifying. I wouldn't, if I, we still have a lot of 2020 left, so if anything.
53:32 Manolis Kellis? No, no, but think about it this way. So the dinosaurs ruled the Earth for 175 million years.
53:41 Manolis Kellis? We humans have been around for, what, less than 1 million years? If you're super generous about what you call humans. And you include chimps, basically. So we are just getting warmed up. And, you know, we've ruled the planet much more ruthlessly than Tyrannosaurus Rex. T-Rex had much less of an environmental impact than we did. And if you if you give us another 174 million years, you know, humans will look very different if we make it that far. So I think dinosaurs, basically, are much more of life history on Earth than we are in all respects. But look at the bright side.
54:24 Manolis Kellis? When they were killed off, another life form emerged: mammals.
54:27 Lex Fridman? And that's that whole evolutionary branching that's happened. So you kind of have, when you have these evolutionary signatures, is there basically a map of how the genome changed?
54:39 Manolis Kellis? Yeah, exactly. So now you can go back to this early mammal that was hiding in caves. And you can basically ask what happened after the dinosaurs were wiped out? A ton of evolutionary niches opened up. And the mammals started populating all of these niches. And in that diversification, there was room for expansion of new types of functions. So some of them populated the air with bats flying, a new evolution of flight. Some populated the oceans with dolphins and whales going off to swim, et cetera. But we all are fundamentally mammals. So you can take the genomes of all these species and align them on top of each other.
55:23 Manolis Kellis? And basically, create nucleotide resolution correspondences. What my PhD work showed is that when you do that, when you line up species on top of each other, you can see that within protein coding genes, there's a particular pattern of evolution that is dictated by the level at which evolutionary selection acts.
55:45 Manolis Kellis? If I'm coding for a protein and I change the third codon position of a triplet that codes for that amino acid, the same amino acid will be encoded. So that basically means that any kind of mutation that preserves that translation, that is invariant to that ultimate functional assessment that evolution will give, is tolerated. So for any function that you're trying to achieve, there's a set of sequences that encode it. You can now look at the mapping, the, you know, graph isomorphism, if you wish, between all of the possible DNA encodings of a particular function and that function.
56:27 Manolis Kellis? And instead of having just that exact sequence at the protein level, you can think of the set of protein sequences that all fulfill the same function. What's evolution doing? Evolution has two components. One component is random, blind, and stupid mutation. The other component is super smart, ruthless selection. That's my mom calling from Greece. Yes, I might be a fully grown man, but I am a Greek.
56:57 Lex Fridman? Did you just cancel the call?
56:59 Manolis Kellis? I know, I'm in trouble. No, she's gonna be calling the cops.
57:01 Lex Fridman? I'm gonna edit this clip out and send it to her. So, yeah, so there's a lot of encoding for the same kind of function.
57:11 Manolis Kellis? Yeah, so you now have this mapping between all of the set of functions that could all encode the same, all of the set of sequences that can all encode the same function. What evolutionary signatures does is that it basically looks at the shape of that distribution of sequences that all encode the same thing. And based on that shape, you can basically say, ooh, proteins have a very different shape than RNA structures, than regulatory motifs, et cetera. So just by scanning a sequence, ignoring the sequence, and just looking at the patterns of change, I'm like, wow, this thing is evolving like a protein, and that thing is evolving like a motif, and that thing is evolving. So that's exactly what we just did for COVID.
57:52 Manolis Kellis? So our paper that we posted in BioArchiv about coronavirus basically took this concept of evolutionary signatures.
57:58 Manolis Kellis? And applied it on the SARS-CoV-2 genome that is responsible of the COVID-19 pandemic.
58:05 Lex Fridman? And comparing it to?
58:07 Manolis Kellis? To 44 serbico virus species. So this is the beta.
58:11 Lex Fridman? What word did you just use?
58:13 Manolis Kellis? Serbico virus, the SARS-related beta coronavirus. It's a portmanteau.
58:18 Lex Fridman? So that whole family of viruses?
58:20 Manolis Kellis? Yeah.
58:20 Lex Fridman? How big is that family, by the way?
58:22 Manolis Kellis? We have 44 species that.
58:24 Lex Fridman? Forty-four species in the family?
58:26 Manolis Kellis? Yeah.
58:26 Lex Fridman? Virus is a clever bunch.
58:28 Manolis Kellis? No, no, but but there's just 44. And again, we don't call them species in in viruses. We call them strains. But anyway, there's 44 strains. And that's a tiny little subset of, you know, maybe another 50 strains that are just far too distantly related. Most of those only infect bats as the host, and a subset of only four or five have ever infected humans. And we basically took all of those and we aligned them in the same exact way that we've aligned mammals. And then we looked at what proteins are, you know, which of the currently hypothesized genes for the coronavirus genome are in fact evolving like proteins and which ones are not.
59:07 Manolis Kellis? And what we found is that ORF10, the last little open reading frame, the last little gene in the genome, is bogus. That's not a protein at all.
59:15 Lex Fridman? What is it?
59:16 Manolis Kellis? It's an RNA structure.
59:18 Lex Fridman? That doesn't have a?
59:20 Manolis Kellis? It doesn't get translated into amino acids.
59:22 Lex Fridman? And that's so it's important to narrow down to basically discover what's useful and what's not.
59:27 Manolis Kellis? Exactly. Basically, what are what is even the set of genes? The other thing that these evolutionary signatures showed is that within ORF3A lies a tiny little additional gene encoded within the other gene. So you can translate a DNA sequence in three different reading frames. If you start in the first one, it's, you know, ATG, et cetera. If you start on the second one, it's TGC, et cetera. And with there's a there's a gene within a gene. So there's a whole other protein that we didn't know about that might be super important. So we don't even know the building blocks of SARS-CoV-2. So if we want to understand coronavirus biology and eventually fight it successfully, we need to even have the set of genes. And and these evolutionary signatures that are developed in my PhD work.
1:00:12 Lex Fridman? Are you really useful here?
1:00:13 Manolis Kellis? Recently used.
1:00:14 Lex Fridman? You know what? Let's let's run with that tangent for a little bit, if it's okay. Is?
1:00:20 Lex Fridman? Can we talk about the COVID-19 a little bit more? Like, how, what's your sense about the genome, the proteins, the functions that we understand about COVID-19? Where do we stand in your sense? What are the big open problems? And also, you know, you kind of said it's important to understand what are the, like, the important proteins? And like, why is that important?
1:00:51 Manolis Kellis? So what else does the comparison of these species tell us? What it tells us is how fast are things evolving? It tells us about at what level is the acceleration or deceleration pedal set for every one of these proteins? So the genome has, you know, 30-some genes. Some genes evolve super, super fast. Others evolve super, super slow. If you look at the polymerase gene that basically replicates the genome, that's a super slow evolving one. If you look at the nucleocapsid protein, that's also super slow evolving. If you look at the spike one protein, this is the part of the spike protein that actually touches the H2 receptor
1:01:32 Manolis Kellis? and then enables the virus to attach to your cells.
1:01:38 Lex Fridman? That's the thing that gives it that, that visual.
1:01:40 Manolis Kellis? Yeah, the corona look, basically.
1:01:41 Lex Fridman? The corona look, yeah.
1:01:42 Manolis Kellis? Yeah, so basically the spike protein sticks out of the virus. And there's a first part of the protein, S1, which basically attaches to the H2 receptor. And then S2 is the latch that sort of pushes and channels the fusion of the membranes and then the incorporation of the viral RNA inside our cells, which then gets translated into all of these 30 proteins. So the S1 protein is evolving ridiculously fast. So if you look at the stop versus gas pedal, the gas pedal is all the way down.
1:02:20 Manolis Kellis? Orf 8 is also evolving super fast and Orf 6 is evolving super fast. We have no idea what they do. We have some idea, but nowhere near what S1 is. So what the-
1:02:30 Lex Fridman? Isn't that terrifying that S1 is evolving? That means that's a really useful function. And if it's evolving fast, doesn't that mean new strains could be created or it does something?
1:02:39 Manolis Kellis? That means that it's searching for how to match, how to best match the host. So basically, anything in general in evolution, if you look at genomes, anything that's contacting the environment is evolving much faster than anything that's internal. And the reason is that the environment changes. So if you look at the evolution of these serbico viruses, the S1 protein has evolved very rapidly because it's attaching to different hosts each time. We think of them as bats, but there's thousands of species of bats. And to go from one species of bat to another species of bat, you have to adjust S1 to the new S2 receptor that you're going to be facing in that new species.
1:03:15 Lex Fridman? Sorry, quick tangent.
1:03:16 Manolis Kellis? Yeah.
1:03:16 Lex Fridman? Is it fascinating to you that viruses are doing this? I mean, it feels like they're this intelligent organism. I mean, is it like, does it give you pause how incredible it is that they're, that the evolutionary dynamics that you're describing is actually happening? And they're figuring out, figuring out how to jump from bats to humans, all in this distributed fashion? And then most of us don't even say they're alive or intelligent, whatever?
1:03:44 Manolis Kellis? So intelligence is in the eye of the beholder. You know, stupid is as stupid does, as Forrest Gump would say.
1:03:51 Lex Fridman? Yes.
1:03:52 Manolis Kellis? And intelligent is as intelligent does. So basically, if the virus is finding solutions that we think of as intelligent, yeah, it's probably intelligent. But that's again in the eye of the beholder.
1:04:01 Lex Fridman? Do you think viruses are intelligent?
1:04:02 Manolis Kellis? Oh, of course not.
1:04:04 Lex Fridman? Really?
1:04:04 Manolis Kellis? No.
1:04:06 Lex Fridman? It's so incredible.
1:04:07 Manolis Kellis? So remember, remember when I was talking about the two components of evolution? One is the stupid mutation.
1:04:12 Lex Fridman? Yeah.
1:04:12 Manolis Kellis? Which is completely blind. And the other one is the super smart selection, which is ruthless. So- It's not viruses who are smart. It's this component of evolution that's smart. So it's evolution that that sort of appears smart. And how is that happening? By huge, parallel search across thousands of, you know, parallel infections throughout the world right now.
1:04:38 Lex Fridman? Yes, but so to push back on that. So, yes, so then the the intelligence is in the mechanism. But then by that argument, viruses would be more intelligent because there's just more of them. So the search, they're basically the the brute force search that's happening with viruses because there's so many more of them than humans, then they're, taken as a whole, are more intelligent. I mean, so you don't think it's possible that, I mean, who runs? Would we even be here if viruses weren't? I mean, who runs this thing?
1:05:15 Manolis Kellis? So, so. So let me answer, yeah, let me answer your your question. So we would not be here if it wasn't for viruses.
1:05:27 Lex Fridman? Yes.
1:05:27 Manolis Kellis? And part of the reason is that if you look at mammalian evolution, early on in this mammalian radiation that basically happened after the death of the dinosaurs, is that some of the viruses that we had in our genome spread throughout our genome and created binding sites for new classes of regulatory proteins. And these binding sites that landed all over our genome are now control elements that basically control our genes and sort of help the complexity of the circuitry of mammalian genomes. So, you know, everything's co-evolution.
1:06:02 Lex Fridman? That's fascinating. We're working together.
1:06:03 Manolis Kellis? Yeah.
1:06:04 Lex Fridman? And yet you say they're dumb.
1:06:06 Manolis Kellis? No, I never said they're dumb. They just don't care. They don't care. Another thing, oh, is the virus trying to kill us? No, it's not. The virus is not trying to kill you. It's not, it's actually actively trying to not kill you. So when you get infected, if you die, pff, bummer, I killed him, is what the reaction of the virus will be. Why? Because that virus won't spread. Many people have a misconception of, oh, viruses are smart, or, ooh, viruses are mean. They don't care. It like you have to clean yourself of any kind of anthropomorphism out there.
1:06:40 Lex Fridman? I don't know.
1:06:41 Manolis Kellis? Oh, yes.
1:06:42 Lex Fridman? So there's a, there's a sense, when taken as a whole, that there's a.
1:06:48 Manolis Kellis? It's in the eye of the beholder. Stupid is as stupid does, intelligent is as intelligent does. So if you want to call them intelligent, that's fine. Because the end result is that they're finding amazing solutions.
1:06:59 Lex Fridman? Right.
1:06:59 Manolis Kellis? I mean, I am in awe.
1:07:01 Lex Fridman? But they're so dumb about it. They're just doing dumb.
1:07:03 Manolis Kellis? They don't care. They're not dumb and they're not instinctive. They just don't care.
1:07:05 Lex Fridman? They don't care. The care word is really interesting.
1:07:08 Manolis Kellis? Exactly.
1:07:08 Lex Fridman? I mean, there could be an argument that they're conscious.
1:07:11 Manolis Kellis? They're just dividing. They're not. They're just dividing. They're just a little entity which happens to be dividing and spreading. It doesn't want to kill us. In fact, it prefers not to kill us. It just wants to spread. And when I say wants, again, I'm anthropomorphizing. But it's just that if you have two versions of a virus, one acquires a mutation that spreads more, that's going to spread more. One acquires a mutation that spreads less, that's going to be lost.
1:07:38 Lex Fridman? Yes.
1:07:38 Manolis Kellis? One acquires a mutation that enters faster, that's going to be kept. One acquires a mutation that kills you right away, it's going to be lost. So over evolutionary time, the viruses that spread super well but don't kill the host are the ones that are going to survive.
1:07:53 Lex Fridman? Yeah, but so you brilliantly described the basic mechanisms of how it all happens. But when you zoom out and you see the entirety of viruses, maybe across different strains of viruses, it seems like a living organism.
1:08:09 Manolis Kellis? I am in awe of biology. I find biology amazingly beautiful. I find the design of the current coronavirus, however lethal it is, amazingly beautiful. The way that it is encoded, the way that it tricks your cells into making 30 proteins from a single RNA. Human cells don't do that. Human cells make one protein from each RNA molecule. They don't make two, they make one. We are hardwired to make only one protein from every RNA molecule. And yet this virus goes in, throws in a single messenger RNA. Just like any messenger RNA, we have tens of thousands of messenger RNAs in our cells in any one time, in every one of our cells.
1:08:52 Manolis Kellis? It throws in one RNA and that RNA is so, I'm gonna use your word here, not my word, intelligent. Yeah. That it hijacks the entire machinery of your human cell. Yeah. It basically has at the beginning, a giant open reading frame. That's a giant protein that gets translated. Two thirds of that RNA make a single giant protein. That single protein is basically what a human cell would make. It's like, oh, here's a start codon, I'm gonna start translating here. Human cells are kinda dumb, I'm sorry. Again, this is not the word that we normally use. But the human cell basically is, oh, this is an RNA, must be mine, let me translate. And it starts translating it, and then you're in trouble. Why?
1:09:36 Manolis Kellis? Because that one protein, as it's growing, gets cleaved into about 20 different peptides. The first peptide and the second peptide start interacting, and the third one and the fourth one, and they shut off the ribosome of the whole cell to not translate human RNAs anymore. So the virus basically hijacks your cells and it cuts, it cleaves every one of your human RNAs to basically say to the ribosome, don't translate this one, junk, don't look at this one, junk. And it only spares its own RNAs because they have a particular mark that it spares.
1:10:18 Manolis Kellis? Then all of the ribosomes that normally make protein in your human cells are now only able to translate viral RNAs. And they have more and more and more and more of them. That's the first 20 proteins. In fact, halfway down, about protein 11, between, you know, 11 and 12, you basically have a translational slippage where the ribosome skips reading frame. And it translates from one reading frame to another reading frame. That means that about half of them are gonna be translated from 1 to 11, and the other half are gonna be translated from 12 to 16. Wow. It's gorgeous. And then...
1:10:53 Manolis Kellis? Then you're done. Then that mRNA will never translate the last day in proteins. But spike is the one right after that one. So how does spike even get translated? This positive strand RNA virus has a reverse transcriptus, which is an RNA-based reverse transcriptus. So from the RNA on the positive strand, it makes an RNA on the negative strand. And in between every single one of these genes, these open reading frames, there's a little signal, AAC, GCA, or something like that, that basically loops over to the beginning of the RNA. And basically, instead of sort of having a single full negative strand RNA, it basically has a partial negative strand RNA that ends right before the beginning of that gene. And another one that ends right before the beginning of that gene.
1:11:38 Manolis Kellis? These negative strand RNAs now make positive strand RNAs that then look to the human wholesale, just like any other human mRNA. It's like, oh, great, I'm going to translate that one because it doesn't have the cleaving that the virus has now put on all your human genes. And then you've lost the battle. That cell is now only making proteins for the virus that will then create the spike protein, the envelope protein, the membrane protein, the nucleocapsid protein that will package up the RNA and then sort of create new viral envelopes. And these will then be secreted out of that cell in new little packages that will then infect the rest of the cells. Repeat the whole process again. It's beautiful, right? It's mind-blowing. It's hard not to anthropomorphize it.
1:12:23 Manolis Kellis? I know, but it's so gorgeous. So there is a beauty to it. Of course. Is there... Is it terrifying to you? So this is something that has happened throughout history. Humans have been nearly wiped out over and over and over again, and yet never fully wiped out. So I'm not concerned about the human race. I'm not even concerned about, you know, about the impact on sort of our survival as a species. This is absolutely something... I mean, you know, human life is so invaluable and every one of us is so invaluable. But if you think of it as sort of, is this the end of our species? By no means, basically. So let me explain. The Black Death killed what, 30% of Europe?
1:13:09 Manolis Kellis? That has left a tremendous imprint, you know, a huge hole, a horrendous hole in the genetic makeup of humans. There's been series of wiping out of huge fractions of entire species or just entire species altogether. And that has a consequence on the human immune repertoire. If you look at how Europe was shaped and how Africa was shaped by malaria, for example, all the individuals that carry a mutation that protects you from malaria were able to survive much more. And if you look at the frequency of sickle cell disease and the frequency of malaria,
1:13:55 Manolis Kellis? the maps are actually showing the same pattern, the same imprint on Africa. And that basically led people to hypothesize that the reason why sickle cell disease is so much more frequent in Americans of African descent is because there was selection in Africa against malaria,
1:14:12 Manolis Kellis? leading to sickle cell. Because when the cell's sickle, malaria is not able to replicate inside your cells as well. And therefore, you protect against that. So if you look at human disease, all of the genetic associations that we do with human disease, you basically see the imprint of these waves of selection killing off gazillions of humans. And there's so many immune processes that are coming up as associated with so many different diseases. The reason for that is similar to what I was describing earlier, where the outward-facing proteins evolve much more rapidly because the environment is always changing. But what's really interesting in the human genome is that we have co-opted many of these immune genes
1:14:57 Manolis Kellis? to carry out non-immune functions. For example, in our brain, we use immune cells to cleave off neuronal connections that don't get used. This whole use it or lose it, we know the mechanism. It's microglia that cleave off. Neuronal synaptic connections that are just not utilized. When you utilize them, you mark them in a particular way that basically when the macroglia come tell it, don't kill this one, it's it's used now. And the macroglia will go off and kill it once you don't use. This is an immune function which is co-opted to do non-immune things. If you look at our adipocytes, M1 versus M2 macrophages inside our fat will basically determine whether you're obese or not.
1:15:39 Manolis Kellis? And these are again, immune cells that are resident and living within these tissues. So many disease associations.
1:15:47 Lex Fridman? That's anybody that we co-opt these kinds of things for incredibly complicated functions.
1:15:53 Manolis Kellis? Exactly. Evolution works in so many different ways, which are all beautiful and mysterious.
1:15:59 Lex Fridman? But not intelligent.
1:16:00 Manolis Kellis? Not intelligent. It's in the eye of the beholder. But but but the the point that I'm trying to make is that if you look at the imprint that COVID will have, hopefully it will not be big. Hopefully, the US will get it packed together and stop the virus from spreading further. But if it doesn't, it's having an imprint on individuals who have particular genetic repertoires. So if you look at now the genetic associations of blood type and immune function cells, et cetera, there's actually association genetic variation that basically says, how much more likely am I or you to die if we contact the virus? And it's it's through these rounds of shaping the human genome that humans have basically made it so far.
1:16:44 Manolis Kellis? And selection is ruthless and it's brutal and it only comes with a lot of killing. But this is the way that viruses and environments have shaped the human genome. Basically, when you go through periods of famine, you select for particular genes. And what's left is not necessarily better, it's just whatever survived. And it may have been the surviving one back then, not because it was better, maybe the ones that ran slower survived. I mean, you know, again, not necessarily better. But the surviving ones are basically the ones that then are shaped for any kind of subsequent evolutionary condition and environmental condition.
1:17:24 Manolis Kellis? But if you look at, for example, obesity. Obesity was selected for, basically the genes that now predisposes to obesity were at 2% frequency in Africa. They rose to 44% frequency in Europe.
1:17:35 Lex Fridman? Wow, that's fascinating.
1:17:37 Manolis Kellis? Because you basically went through the ice ages and there was a scarcity of food. So you know, there was a selection to being able to store every single calorie you consume. Eventually, environment changes. So the better allele, which was the fat storing allele, became the worse allele. Because it's the fat storing allele. It still has the same function. So if you look at my genome, speaking of mom calling. Mom gave me a bad copy of that gene, this FTO locus. Basically.
1:18:07 Lex Fridman? The one that has to do with.
1:18:09 Manolis Kellis? Obesity.
1:18:09 Lex Fridman? With obesity.
1:18:10 Manolis Kellis? Yeah, I basically now have a bad copy from mom that makes me more likely to be obese. And I also have a bad copy from dad that makes me more likely to be obese. I'm homozygous. And that's the allele. It's still the minor allele, but it's at 44% frequency in Southeast Asia, 42% frequency in Europe. Even though it started at 2%. It was an awesome allele to have 100 years ago. Right now, it's a pretty terrible allele. So the other concept is that diversity matters. If we had 100 million nuclear physicists living on the earth right now, we'd be in trouble. You need diversity, you need artists, and you need musicians, and you need mathematicians, and you need, you know, politicians, yes, even those. And you need like.
1:18:55 Lex Fridman? Well, let's not, let's not get crazy enough. But so because then if a virus comes along or whatever.
1:19:01 Manolis Kellis? Exactly, exactly. So, so no, there's two reasons. Number one, you want diversity in the immune repertoire, and we have built in diversity. So basically, there are the most diverse, basically if you look at our immune system, there's layers and layers of diversity. Like the way that you create your cells generates diversity because of the selection for the VDJ recombination that basically eventually leads to a huge number of repertoires. But that's only one small component of diversity. The blood type is another one, the major histocompatibility complex, the HLA alleles are, you know, another source of diversity. So the immune system of humans is by nature incredibly diverse. And that basically leads to resilience. So basically what I'm saying that I don't worry for the human species.
1:19:47 Manolis Kellis? Because we are so diverse immunologically, we are likely to be very resilient against so many different attacks like this current virus.
1:19:56 Lex Fridman? So you're saying natural pandemics may not be something that you're really afraid of because of the diversity in our genetic makeup. What about engineered pandemics? Do you have fears of us messing with the makeup of viruses? Or, well, yeah, let's say with the makeup of viruses to create something that we can't control and would be much more destructive than would come about naturally?
1:20:23 Manolis Kellis? Remember how we were talking about how smart evolution is? Humans are much dumber.
1:20:28 Lex Fridman? You mean like human scientists?
1:20:29 Manolis Kellis? Yeah, humans, humans.
1:20:30 Lex Fridman? Humans overall?
1:20:31 Manolis Kellis? Yeah, humans overall. But I mean, even, you know, the sort of synthetic biologists, you know, basically, if you were to create, you know, virus like SARS that will kill a lot of people, you would probably start with SARS. So whoever, you know, would like to design such a thing would basically start with a SARS tree or at least some relative of SARS. The source genome for the current virus was something completely different. It was something that has never infected humans. No one in their right mind would have started there.
1:21:09 Lex Fridman? But when you say source, it's like the nearest.
1:21:11 Manolis Kellis? The nearest relative is in a whole other branch. No species of which has ever infected humans in that branch. So, you know, let's put this to rest. This was not designed by someone to kill off the human race.
1:21:25 Lex Fridman? So you don't, you don't believe it was engineered?
1:21:28 Manolis Kellis? The, yeah, the, the path to engineering a deadly virus would not come from this strain that, that was used. Moreover, there's been various claims of, haha, this was mixed and matched in a lab because the S1 protein has three different components, each of which has a different evolutionary tree. So, you know, a lot of popular press basically said, aha, this came from pangolin and this came from, you know, all kinds of other species. This is what. Has been happening throughout the coronavirus tree. So basically, the S1 protein has been recombining across species all the time. Remember when I was talking about the positive strand, the negative strand, subgenomic RNAs? These can actually recombine.
1:22:12 Manolis Kellis? And if you have two different viruses infecting the same cell, they can actually mix and match between the positive strand and the negative strand and basically create a new hybrid virus with recombination that now has the S1 from one and the rest of the genome from another. And this is something that happens a lot in S1, in orfate, et cetera. And that's something that's true of the whole tree.
1:22:30 Lex Fridman? For the whole family of coronavirus.
1:22:32 Manolis Kellis? So it's not like someone has been messing with this for millions of years and, you know, changing.
1:22:37 Lex Fridman? This happens naturally. That's again, beautiful that that somehow happens, that they recombine. So two different strands can infect the body and then recombine.
1:22:47 Lex Fridman? So all of this actually magic happens inside hosts. Like all like.
1:22:53 Manolis Kellis? Yeah, that's why that's why classification wise, virus is not thought to be alive because it doesn't self replicate. It's not autonomous. It's something that enters a living cell and then co-opt it to basically make it its own. But by itself, people ask me, how do we kill this bastard? I'm like, you stop it from replicating. It's not like a bacterium that will just live in a, you know, puddle or something. It's a virus. Viruses don't live without their host and they only live within their host for very little time. So if you stop it from replicating, it'll stop from spreading. I mean, it's not like HIV, which can stay dormant for a long time. Basically, coronaviruses just don't do that. They're not integrating genomes. They're RNA genomes. So if it's not expressed, it degrades. RNA degrades.
1:23:38 Manolis Kellis? It doesn't just stick around.
1:23:40 Lex Fridman? Well, let me ask also about the immune system you mentioned. A lot of people kind of ask, you know, how can we strengthen the immune system to respond to this particular virus, but the viruses in general? Do you have from a biological perspective, thoughts on what we can do as humans?
1:24:00 Manolis Kellis? If you look at the death rates across different countries, people with less vaccination have been dying more. If you look at North Italy, the vaccination rates are abysmal there, and a lot of people have been dying. If you look at Greece, very good vaccination rates. Almost no one has been dying. So, yes, there's a policy component. So Italy reacted very slowly. Greece reacted very fast. So, yeah, many fewer people died in Greece. But there might actually be a component of genetic immune repertoire. Basically, how did people die off, you know, in the history of the Greek population versus the Italian population?
1:24:37 Lex Fridman? Wow.
1:24:38 Manolis Kellis? There's a...
1:24:39 Lex Fridman? That's interesting to think about.
1:24:41 Manolis Kellis? And then there's a component of what vaccinations did you have as a kid? And what are the off-target effects of those vaccinations? So basically, a vaccination can have two components. One is training your immune system against that specific insult. The second one is boosting up your immune system for all kinds of other things. If you look at allergies. Northern Europe, super clean environments, tons of allergies. Southern Europe, my kids grew up eating dirt. No allergies. So, growing up, I never had even heard of what allergies are. Like, was like, really, allergies? And the reason is that I was playing in the garden. I was putting all kinds of stuff in my mouth, from, you know, all kinds of dirt and stuff. Tons of viruses there, tons of bacteria there.
1:25:26 Manolis Kellis? You know, my immune system was built up. So, the more you protect your immune system from exposure, the less opportunity it has to learn about non-self repertoire in a way that prepares it for the next insult.
1:25:41 Lex Fridman? So it's a horizontal thing too? Like, so it's throughout your lifetime and the lifetime of the people that are your ancestors?
1:25:50 Manolis Kellis? Yeah.
1:25:50 Lex Fridman? That kind of thing?
1:25:50 Manolis Kellis? Yeah.
1:25:51 Lex Fridman? What about the... So, again, it returns against free will. On the free will side of things, is there something we could do to strengthen our immune system in 2020? Is there, you know, exercise, diet, all that kind of stuff?
1:26:07 Manolis Kellis? So, it's kind of funny. There's a cartoon that basically shows two windows with a teller in each window. One has a humongous line, and the other one has no one. The one that has no one above says health. No, it says exercise and diet. And the other one says pill. And there's a huge line for pill. So we're looking basically for magic bullets for sort of ways that we can, you know, beat cancer and beat coronavirus and beat this and beat that. And it turns out that the window with like just diet and exercise is the best way to boost every aspect of your health. If you look at Alzheimer's, exercise and nutrition. I mean, you're like, really? For my brain? Neurodegeneration?
1:26:51 Manolis Kellis? Absolutely. If you look at cancer, exercise and nutrition. If you look at coronavirus, exercise and nutrition. Every single aspect of human health gets improved. And one of the studies we're doing now is basically looking at what are the effects of diet and exercise? How similar are they to each other? We're basically taking diet intervention and exercise intervention in human and in mice. And we're basically doing single cell profiling of a bunch of different tissues to basically understand how are the cells, both the stromal cells and the immune cells of each of these tissues responding to the effect of exercise? What are the communication networks between different cells?
1:27:35 Manolis Kellis? Where the muscle that exercises sends signals through the bloodstream, through the lymphatic system, through all kinds of other systems, that give signals to other cells that I have exercised and you should change in this particular way, which basically reconfigure those receptor cells with the effect of exercise.
1:27:56 Lex Fridman? How well understood is those reconfigurations?
1:28:00 Manolis Kellis? Very little. We're just starting now, basically.
1:28:03 Lex Fridman? Is the hope there to understand the effect on, so like the effect on the immune system?
1:28:11 Manolis Kellis? On the immune system, the effect on brain, the effect on your liver, on your digestive system, on your adipocytes. Adipose, you know, the most misunderstood organ. Everybody thinks, oh, fat, terrible. No, fat is awesome. Your fat cells is what's keeping you alive. Because if you didn't have your fat cells, all those lipids and all those calories would be floating around in your blood and you'd be dead by now. Your adipocytes are your best friend. They're basically storing all these excess calories so that they don't hurt all of the rest of the body. And they're also fat burning in many ways. So, you know, again, when you don't have the homozygous version that I have, your cells are able to burn calories much more easily by sort of flipping a master metabolic
1:28:55 Manolis Kellis? switch that involves this FTO locus that I mentioned earlier, and it's target genes, RX3 and RX5, that basically switch your adipocytes during their three first days of differentiation as they're becoming mature adipocytes to basically become either fat burning or fat-storing fat cells. And the fat-burning fat cells are your best friends. They're much closer to muscle than they are to white adipocytes.
1:29:19 Lex Fridman? Is there a lot of difference between people like that you could give, science could eventually give advice that is very generalizable? Or is our differences in our genetic makeup, like you mentioned, is that going to be basically something we have to be very specialized individuals? Any advice we give in terms of diet, like we were just talking about?
1:29:42 Manolis Kellis? Believe it or not, the most personalized advice that you give for nutrition don't have to do with your genome. They have to do with your gut microbiome, with the bacteria that live inside you. So most of your digestion is actually happening by species that are not human inside you. You have more non-human cells than you have human cells. You're basically a giant bag of bacteria with a few human cells along.
1:30:05 Lex Fridman? And so how, and those do not necessarily have to do with your genetic makeup?
1:30:11 Manolis Kellis? They interact with your genetic makeup. They interact with your epigenome. They interact with your nutrition. They interact with your environment. They're, you know, basically an additional source of variation. So when you're thinking about sort of personalized nutritional advice, part of that is actually how do you match your microbiome? And part of that is how do we match your genetics? But again, you know, this is a very diverse set of, you know, contributors, and the effect sizes are not enormous. So I think the science for that is not fully developed yet.
1:30:44 Lex Fridman? Speaking of diets, because I've wrestled in combat sports, sports my whole life where weight matters. So you have to cut and all that stuff. One thing I've learned a lot about my body, and it seems to be, I think, true about other people's bodies, is that you can adjust to a lot of things. That's the miraculous thing about this biological system, is like I fast often. I used to eat like five, six times a day and thought that was absolutely necessary. How could you not eat that often? And then when I started fasting, your body adjusts to that and you learn how to not eat.
1:31:20 Manolis Kellis? I know.
1:31:21 Lex Fridman? And it's, it was, if you just give it a chance for a few weeks, actually, over a period of a few weeks, your body can adjust to anything.
1:31:28 Manolis Kellis? Yeah.
1:31:29 Lex Fridman? And that's a miracle. That's such a beautiful thing.
1:31:31 Manolis Kellis? So I'm a computer scientist and I've basically gone through periods of 24 hours without eating or stopping. And, you know, then I'm like, oh, must eat. And I eat a ton. I used to order two pizzas just with my brother and, you know.
1:31:44 Manolis Kellis? Like, so I've gone through these extremes as well. And I've gone the whole intermittent fasting thing. So I can sympathize with you both on the seven meals a day to the zero meals a day. So I think when I say everything with moderation, I actually think your body responds interestingly to these different changes in diet. I think part of the reason why we lose weight with pretty much every kind of change in behavior is because our epigenome and the set of proteins and enzymes that are expressed in our microbiome are not well suited to that nutritional source. And therefore, they will not be able to sort of catch everything that you give them. And then, you know, a lot of that will go undigested. And that basically means that your body can then, you know, lose weight in the short
1:32:29 Manolis Kellis? term, but very quickly will adjust to that new normal. And then we'll be able to sort of perhaps gain a lot of weight
1:32:35 Lex Fridman? Yeah.
1:32:36 Manolis Kellis? from the diet. So anyway, I mean, there's also studies in factories where basically people, you know, dim the lights and then suddenly everybody started working better. It was like, wow, that's amazing. Three weeks later, they made the lights a little brighter. Everybody started working better. So any kind of intervention has a placebo effect of wow, now I'm healthier and I'm going to be running more often, etc. So it's very hard to uncouple the placebo effect of wow, I'm doing something to intervene on my diet from the wow, this is actually the right thing for me. So, you know.
1:33:07 Lex Fridman? Yeah, from the perspective from nutrition science, psychology, both things I'm interested in, especially psychology, it seems that it's extremely difficult to do good science because there's so many variables involved, it's so difficult to control the variables, so difficult to do sufficiently large scale experiments, both sort of in terms of number of subjects and temporal, like how long you do the study for, that it just seems like it's not even a real science for now, like nutrition science.
1:33:39 Manolis Kellis? I want to jump into the whole placebo effect for a little bit here and basically talk about the implications of that. If I give you a sugar pill and I tell you it's a sugar pill, you won't get any better. But if I tell you a sugar pill and I tell you, wow, this is an amazing drug, it actually will stop your cancer, your cancer will actually stop with much higher probability. What does that mean?
1:34:04 Lex Fridman? That's so amazing.
1:34:04 Manolis Kellis? That means that if I can trick your brain into thinking that I'm healing you, your brain will basically figure out a way to heal itself, to heal the body. And that tells us that there's so much that we don't understand in the interplay between our cognition and our biology, that if we were able to better harvest the power of our brain to sort of, you know, impact the body through the placebo effect, we would be so much better in so many different things. Just by tricking yourself into thinking that you're doing better, you're actually doing better. So there's something to be said about sort of positive thinking, about optimism, about sort of, you know, just.
1:34:45 Manolis Kellis? Getting your brain and your mind into the right mindset, that helps your body and helps your entire biology.
1:34:54 Lex Fridman? Yeah, from a science perspective, that's just fascinating. Obviously, most things about the brain is a total mystery for now. But that's a fascinating interplay that the brain
1:35:04 Manolis Kellis? Yeah.
1:35:05 Lex Fridman? That the brain can reduce.
1:35:08 Lex Fridman? The brain can help cure cancer is a. I don't even know what to do with that.
1:35:13 Manolis Kellis? I mean, the way to think about that is the following. The converse of the equation is something that we are much more comfortable with. Like, oh, if you're stressed, then your heart might rise and all kinds of sort of toxins might be released. And that can have a detrimental effect in your body, et cetera, et cetera, et cetera. So maybe it's easier to understand your body healing from your mind by your mind is not killing your body. Or at least it's killing it less. So I think the, you know, that aspect of the stress equation is a little easier for most of us to conceptualize. But then the healing part is, you know, perhaps the same pathways, perhaps different pathways. But again, something that is totally untapped scientifically.
1:35:56 Lex Fridman? I think we tried to bring this question up a couple of times. But let's return to it again. Is what's do you think is the difference between the way a computer represents information, the human genome represents and stores information? Like what and maybe broadly, what is the difference between how you think about computers and how you think about biological systems?
1:36:17 Manolis Kellis? So I made a very provocative claim earlier that we are a digital computer. Like that at the core lies a digital code. And that's true in many ways, but surrounding that digital core, there's a huge amount of analog. If you look at our brain, it's not really digital. If you look at our sort of RNA and all of that stuff inside our cells, it's not really digital. It's really analog in many ways. But let's start with the code and then we'll expand to the rest. So the code itself is digital. So there's genes, you can think of the genes as, I don't know, the procedures, the functions inside your language. And then somehow you have to turn these functions on. How do you call a gene? How do you call that function? The way that you would do it in old programming languages is go to address whatever in your
1:37:01 Manolis Kellis? memory and then you'd start running from there. And, you know, modern programming languages have encapsulated this into functions and objects and all of that. And it's nice and cute, but in the end, deep down, there's still an assembly code that says, go to that instruction and it runs that instruction. If you look at the human genome and, you know, the genome of pretty much most species out there, it's, there's no go-to function. You just don't start in, you know, transcribing in position thirteen hundred five, you know, thirteen thousand five hundred twenty-seven in chromosome twelve. You instead have content-based indexing. So at every location in the genome, in front of the genes that need to be turned on, I
1:37:45 Manolis Kellis? don't know, when you drink coffee, there's a little coffee marker in front of all of them. And whenever your cells that metabolize coffee need to metabolize coffee, they basically see coffee and they're like, ooh, let's go turn on all the coffee-marked genes. So there's basically these small motifs, these small sequences that we call regulatory motifs. They're like patterns of DNA. They're only eight characters long or so, like GAT, GCA, et cetera. And these motifs work in combinations and every one of them has some recruitment affinity for a different protein that will then come and bind it.
1:38:26 Manolis Kellis? And together, collections of these motifs create regions that we call regulatory regions that can be either promoters near the beginning of the gene. And that basically tells you where the function actually starts, where you call it. And then enhancers that are looping around of the DNA that basically bring the machinery that binds those enhancers and then bring it on to the promoter, which then recruits the right sort of the ribosome and the polymerase and all of that thing, which will first transcribe and then export and then eventually translate in the cytoplasm, you know, whatever RNA molecule. So the...
1:39:06 Manolis Kellis? Beauty of the way that the digital computer, that's the genome, works is that it's extremely fault tolerant. If I took your hard drive and I messed with 20% of the letters in it, of those zeros and ones, and I flipped them, you'd be in trouble. If I take the genome and I flipped 20% of the letters, you probably won't even notice. And that resilience
1:39:32 Lex Fridman? That's fascinating.
1:39:34 Manolis Kellis? is a key design principle, and again, I'm turboamorphizing here, but it's a key driving principle of how biological systems work. They're first resilient and then anything else. And when you look at this incredible beauty of life, from the most, I don't know, beautiful, I don't know, human genome, maybe, of humanity and all of the ideals that should come with it, to the most terrifying genome, like, I don't know, COVID-19, SARS-CoV-2, and the current pandemic, you basically see this elegance as the epitome of clean design. But it's dirty. It's a mess. It's, you know, the way to get there is hugely
1:40:19 Manolis Kellis? messy. And that's something that we as computer scientists don't embrace. We like to have clean code. You know, like in engineering, they teach you about compartmentalization, about sort of separating functions, about modularity, about hierarchical design. None of that applies in biology.
1:40:36 Lex Fridman? Testing.
1:40:38 Manolis Kellis? Testing, sure. Yeah, biology does plenty of that. But, I mean, through evolutionary exploration. But if you look at biological systems, first, they are robust, and then they specialize to become anything else. And if you look at viruses, the reason why they're so elegant, when you look at the design of this, you know, genome, it seems so elegant. And the reason for that is that it's been stripped down from something much larger because of the pressure to keep it compact. So many compact genomes out there have ancestors that were much larger. You don't start small and become big.
1:41:17 Manolis Kellis? You go through a loop of add a bunch of stuff, increase complexity, and then, you know, slim it down. And one of my early papers was in fact on genome duplication. One of the things we found is that baker's yeast, which is the, you know, yeast that you use to make bread, but also the yeast that you use to make wine, which is basically the dominant species when you go in the fields of Tuscany and you say, you know, what's out there? It's basically Saccharomyces cerevisiae. Or the way my Italian friends say Saccharomyces cerevisiae. So. So. Uh, which means what? Oh, Saccharomy-- okay, I'm sorry, I'm Greek, so yeah. Zaharo mikis, zaharo is sugar, mikis is fungus. Yes. Cerevisiae, cerveza, beer.
1:42:01 Manolis Kellis? So it means the sugar fungus of beer. Yeah. You know, let's, let's, let's keep sounding to the. Still poetic, yep. Uh, so anyway, Saccharomyces cerevisiae, basically the major baker's yeast out there, is the descendant of a whole genome duplication. Why would a whole genome duplication even happen? When it happened is coinciding with about 100 million years ago and the emergence of fruit-bearing plants.
1:42:31 Manolis Kellis? Why fruit-bearing plants? Because animals would eat the fruit, would walk around and poop huge amounts of nutrients along with a seed for the plants to spread. Before that, plants were not spreading through animals. They were spreading through wind and all kinds of other ways. But basically, the moment you have fruit-bearing plants, the, the, the, these plants are basically creating this abundance of sugar in the environment. So there's an evolutionary niche that gets created. And in that evolutionary niche, you basically have enough sugar that a whole genome duplication, which initially is a very messy event, allows you to then, you know, relieve some of that complexity. So I had to pause. What does genome duplication mean?
1:43:16 Manolis Kellis? That basically means that instead of having eight chromosomes, you're gonna now have 16 chromosomes.
1:43:23 Lex Fridman? So but the duplication, at first, when you have sixteen, when you go to sixteen, you're not using that.
1:43:30 Manolis Kellis? Oh, yeah, you are. Yeah. So basically, from one day to the next, you went from having eight chromosomes to having sixteen chromosomes. Probably a non-disjunction event during a duplication, during a division. So you basically divide the cell. Instead of half the genome going this way and half the genome going the other way after duplication of the genome, you basically have all of it going to one cell. And then there's sufficient messiness there that you end up with slight differences that make most of these chromosomes be actually preserved. It's a long story short to basically-
1:44:00 Lex Fridman? But that's a big upgrade, right? So that's-
1:44:02 Manolis Kellis? Not necessarily. Because what happens immediately thereafter is that you start massively losing tons of those duplicated genes. So ninety percent of those genes were actually lost very rapidly after whole genome duplication. And the reason for that is that biology is not intelligent. It's just ruthless selection, random mutation. So the ruthless selection basically means that as soon as one of the random mutations hit one gene, ruthless selection just kills off that gene. It's just, you know, you, you know, if you have a pressure to maintain a small, compact genome, you will very rapidly lose the second copy of most of your genes. And a small number, ten percent, were kept into copies. And those had to do a lot with environment adaptation,
1:44:45 Manolis Kellis? with the speed of replication, with the speed of translation, and with sugar processing. So I'm making a long story short to basically say that evolution is messy. The only way, like, so, so, you know, the example that I was giving of messing with twenty percent of your bits in your computer, totally bogus. Duplicating all your functions and just throwing them out there in the same, you know, function, just totally bogus, like, this would never work in an engineer system. But biological systems, because of this content-based indexing and because of this modularity that comes from the fact that the gene is controlled by a series of tags. And now if you need this gene in another setting, you just add some more tags that will basically turn it on also in those settings.
1:45:29 Manolis Kellis? So this gene is now pressured to do two different functions. And it builds up complexity. I see hydrogen duplication and gene duplication in general as a way to relieve that complexity. So you have this gradual buildup of complexity as functions get sort of added on to the existing genes. And then boom, you duplicate your workforce. And you now have two copies of this gene. One will probably specialize to do one and the other one will specialize to do the other. Or one will maintain the ancestral function. The other one will sort of be free to evolve and specialize while losing the ancestral function and so on and so forth. So that's how genomes evolve.
1:46:07 Manolis Kellis? They're just messy things, but they're extremely fault tolerant and they're extremely able to deal with mutations. Because that's the very way that you generate new functions. So new functionalization comes from the very thing that breaks it. So even in the current pandemic, many people are asking me which mutations matter the most. And what I tell them is, well, we can study the evolutionary dynamics of the current genome to then understand which mutations have previously happened or not. And which mutations happen in genes that evolved rapidly or not. And one of the things we found, for example, is that the genes that evolved rapidly in the past
1:46:51 Manolis Kellis? are still evolving rapidly now in the current pandemic. The genes that evolved slowly in the past are still evolving slowly.
1:46:58 Lex Fridman? Which means that they're useful?
1:46:59 Manolis Kellis? Which means that they're under the same evolutionary pressures. But then the question is, what happens in specific mutations? So if you look at the D614G mutation that's been all over the news. So in position 614, in the amino acid 614 of the S protein, there's a D2G mutation that sort of has creeped over the population. That mutation, we found out through my work, disrupts a perfectly conserved nucleotide position that has never been changed in the history of millions of years of equivalent mammalian evolution of these. Viruses, that basically means that it's a completely new adaptation to human.
1:47:44 Manolis Kellis? And that mutation has now gone from 1% frequency to 90% frequency in almost all outbreaks.
1:47:50 Lex Fridman? So there's a mutation, I like how you say the 416, what was it?
1:47:56 Manolis Kellis? Yeah, 614, sorry.
1:47:57 Lex Fridman? 614.
1:47:59 Manolis Kellis? D614G.
1:48:00 Lex Fridman? D6, so, so literally, so what you're saying is this is like a chess move.
1:48:05 Manolis Kellis? Yeah.
1:48:05 Lex Fridman? So it just mutated one letter to another.
1:48:07 Manolis Kellis? Exactly.
1:48:08 Lex Fridman? And that hasn't happened before.
1:48:09 Manolis Kellis? Yeah.
1:48:10 Lex Fridman? And, and this somehow, this mutation is really useful.
1:48:14 Manolis Kellis? It's really useful in the current environment of the genome, which is moving from human to human. When it was moving from bat to bat, it couldn't care less for that mutation. But it's environment specific, so now that it's moving from human to human, whoo, it's moving way better, like by orders of magnitude.
1:48:32 Lex Fridman? What do you, okay, so you're like tracking this evolutionary dynamics, which is fascinating. But what do you do with that? So what does that mean? What does this, what do you make, what do you make of this mutation in trying to anticipate, I guess? Is, is the, is one of the things you're trying to do is anticipate where, how this unrolls into the future? This evolutionary dynamics?
1:48:56 Manolis Kellis? Such a great question. So, so there's, there's two things. Remember when I was saying earlier, mutation is the path to new things, but also the path to break old things. So what we know is that this position was extremely preserved through gazillions of mutations. That mutation was never tolerated when it was moving from bats to bats. So that basically means that that mutation, that position is extremely important in the function of that protein. That's the first thing it tells. The second one is that that position was very well suited to bat transmission, but now is not well suited to human transmission. So it got rid of it and it now has a new version of that amino acid. That basically makes it much easier to transmit from human to human. So.
1:49:41 Manolis Kellis? In terms of the evolutionary history teaching us about the future, it basically tells us here's the regions that are currently mutating. Here's the regions that are most likely to mutate going forward. As you're building a vaccine, here's what you should be focusing on in terms of the most stable regions that are the least likely to mutate. Or here's the newly evolved functions that are most likely to be important because they've overcome this local maximum that it had reached in the bat transmission. So anyway, it's a tangent to basically say that evolution works in messy ways.
1:50:21 Manolis Kellis? And the thing that you would break is the thing that actually allows you to first go through a lull and then reaching new local maximum. And I often like to say that if engineers had basically designed evolution, we would still be perfectly replicating bacteria. Because it's by making the bacterium worse that you allow evolution to reach a new optimum.
1:50:49 Lex Fridman? That's just a pause on that. That's so profound. That's so profound for the entirety of this scientific and engineering disciplines.
1:51:00 Manolis Kellis? Exactly. We as engineers need to embrace breaking things. We as engineers need to embrace robustness as the first principle beyond perfection. Because nothing's going to ever be perfect. And when you're sending a satellite to Mars, when something goes wrong, it'll break down. As opposed to building systems that tolerate failure and are resilient to that and in fact, get better through that.
1:51:27 Lex Fridman? So the SpaceX approach versus NASA for the...
1:51:31 Manolis Kellis? For example.
1:51:33 Lex Fridman? Is there something we can learn about the incredible... Take lessons from the incredible biological systems in their resilience, in their... in the mushiness, the messiness to our computing systems, to our computers?
1:51:48 Manolis Kellis? It would basically be starting from scratch in many ways.
1:51:52 Manolis Kellis? It would basically be building new paradigms that don't try to get the right answer all the time, but try to get the right answer most of the time, or a lot of the time.
1:52:04 Manolis Kellis? Do you see deep learning systems in the whole world of machine learning as kind of taking a step in that direction?
1:52:08 Manolis Kellis? Absolutely, absolutely. Basically, by allowing this much more natural evolution of these parameters, you basically and if you look at sort of deep learning systems, again, they're not inspired by the genome aspect of biology, they're inspired by the brain aspect of biology. And again, I want you to pause for a second and realize the complexity of the entire human brain with trillions of connections within our, you know, neurons with millions of cells talking to each other is still encoded within that same genome.
1:52:49 Manolis Kellis? That same genome encodes every single freaking cell type of the entire body. Every single cell is encoded by the same code. And yet, specialization allows you to have this single viral-like genome that self-replicates, this single module, modular automaton, work with other copies of itself. It's mind-boggling. Create complex organs through which blood flows. And what is that blood? The same freaking genome. Create organs that communicate with each other. And what are these organs? The exact same genome.
1:53:31 Manolis Kellis? Create a brain that is innervated by massive amounts of blood pumping energy to it, 20% of our energetic needs, to the brain.
1:53:44 Manolis Kellis? From the same genome and all of the neuronal connections, all of the axillary cells, all of the immune cells, the astrocytes, the ligandocytes, the neurons, the excitatory, the inhibitory neurons, all of the different classes of parasites, the blood-brain barrier, all of that, same genome.
1:53:59 Lex Fridman? One way to see that in a sad, this one is beautiful. The sad thing is thinking about the trillions of organisms that died to create that.
1:54:12 Manolis Kellis? You mean on the evolutionary path to?
1:54:13 Lex Fridman? On the evolutionary path to humans. It's crazy. There's two descendants of apes just talking on a podcast. Okay. It's so mind-boggling.
1:54:25 Manolis Kellis? Just to boggle our minds a little bit more.
1:54:27 Lex Fridman? Yeah.
1:54:28 Manolis Kellis? Us talking to each other. We are basically generating a series of vocal utterances through our pulsating of vocal cords. Received through this, the people who listen to this are taking a completely different path to that information transfer, yet through language. But imagine if we could connect these brains directly to each other. The amount of information that I'm condensing into a small number of words is a huge funnel, which then you receive and you expand into a huge number of thoughts from that small funnel.
1:55:12 Manolis Kellis? In many ways, engineers would love to have the whole information transfer, just take the whole set of neurons and throw them away. I mean, throw them to the other person.
1:55:22 Manolis Kellis? This might actually not be better because in your misinterpretation of every word that I'm saying, you are creating new interpretation that might actually be way better than what I meant in the first place. The ambiguity of language perhaps might be the secret to creativity.
1:55:42 Manolis Kellis? Every single time you work on a project by yourself, you only bounce ideas with one person. And your neurons are basically fully cognizant of what these ideas are. But the moment you interact with another person, the misinterpretations that happen might be the most creative part of the process. With my students, every time we have a research meeting, I very often pause and say, "Let me repeat what you just said in a different way." And I sort of go on and brainstorm with what they were saying. But by the third time, it's not what they were saying at all. And when they pick up what I'm saying, they're like, "Oh, well, da-da-da." Now they've sort of learned something very different from what I was saying. And that is the same kind of messiness that I'm describing in the genome itself.
1:56:27 Manolis Kellis? It's sort of embracing the messiness.
1:56:30 Lex Fridman? And that's a feature, not a book.
1:56:32 Manolis Kellis? Exactly. And in the same way, when you're thinking about sort of these deep learning systems that will allow us to sort of be more creative perhaps, or learn better approximations of these complex functions, again, tuned to the universe that we inhabit, you have to embrace the breaking. You have to embrace the, you know, "How do we get out of this local optima?" And a lot of the design paradigms that have made deep learning so successful are ways to get away from that. Ways to get better training by sort of sending long-range messages. These LSTM models and the, you know, sort of feed-forward loops that, you know, sort of jump through layers of a convolutional neural network.
1:57:17 Manolis Kellis? All of these things are basically ways to push you out of this local maxima. And that's sort of what evolution does. That's what language does. That's what conversation and brainstorming does. That's what our brain does. So, you know, this design paradigm is something that's pervasive and yet not taught in schools, not taught in engineering schools, where everything's minutely modularized to make sure that we never deviate from, you know, whatever signal we're trying to emit, as opposed to "Let all hell breaks loose," because that's the path to paradise.
1:57:50 Lex Fridman? The path to paradise. Yeah, I mean, it's difficult to know how to teach that and what to do with it. I mean, it's difficult to know how to build up the scientific method around messiness.
1:58:05 Manolis Kellis? I mean, it's not all messiness. We need, we need some cleanness. And going back to the example with Mars, that's probably the place where I want to sort of moderate error as much as possible and sort of control the environment as much as possible. But if you're trying to repopulate Mars, well, maybe messiness is a good thing then.
1:58:23 Lex Fridman? Annette, you quickly mentioned this in terms of us using our vocal cords to speak on a podcast. So Elon Musk and Neuralink are working on trying to plug, as per a discussion with computers and biological systems, to connect the two. He's trying to connect our brain to a computer to create a brain-computer interface where they can communicate back and forth. On this line of thinking, do you think this is possible to bridge the gap between our engineered computing systems and the messy biological systems?
1:59:06 Manolis Kellis? My answer would be absolutely. You know, there's no doubt that we can understand more and more about what goes on in the brain, and we can sort of train the brain. I don't know if you remember the Palm Pilot?
1:59:20 Lex Fridman? Yeah, Palm Pilot, yeah.
1:59:21 Manolis Kellis? Remember this whole sort of alphabet that they had created? Am I thinking of the same thing? It's basically, you had, you had a little pen, and for every character, you had a little scribble that was unique, that the machine could understand, and that instead of trying the machine, trying to teach the machine to recognize human characters, you had basically, they figured out that it's better and easier to train humans to create human-like characters that the machine is better at recognizing. So, in the same way, I think what will happen is that humans will be trained to be able to create the mind pattern that the machine will respond to before the machine truly comprehends our thoughts.
2:00:04 Manolis Kellis? So the first human-brain interfaces will be tricking humans to speak the machine language, where with the right set of electrodes, I can sort of trick my brain into doing this. And this is the same way that many people teach, like, learn to control artificial limbs. You basically try a bunch of stuff, and eventually you figure out how your limbs work. That might not be very different from how humans learn to use their natural limbs when they first grow up. Basically, you have this, you know, neotony period of, you know, this puddle of soup inside your brain, trying to figure out how to even make neural connections before you're born.
2:00:42 Manolis Kellis? And then learning sounds in utero of, you know, all kinds of echoes and, you know, eventually getting out in the real world. And I don't know if you've seen newborns, but they just stare around a lot. You know, one way to think about this as a machine learning person is, oh, they're just training their edge detectors.
2:01:01 Lex Fridman? Yeah.
2:01:02 Manolis Kellis? And eventually, they figure out how to train their edge detectors. They work through the second layer of the visual cortex and the third layer and so on and so forth. And you basically have this, um, learning how to control your limbs that probably comes at the same time. You're sort of, you know, throwing random things there and you realize that, oh, wow, when I do this thing, my limb moves. Let's do the following experiment. Take a breath. What muscles did you flex? Now, take another breath and think, what muscles do I flex? The first thing that you're thinking when you're taking a breath is the impact that it has on your lungs. You're like, oh, I'm now going to increase my lungs, or I'm now going to bring air in. But what you're actually doing is just changing your diaphragm.
2:01:44 Lex Fridman? Yeah.
2:01:45 Manolis Kellis? That's not conscious, of course. You never think of the diaphragm as a thing.
2:01:51 Lex Fridman? Yeah.
2:01:52 Manolis Kellis? And why is that? That's probably the same reason why I think of moving my finger when I actually move my finger. I think of the effect instead of actually thinking of whatever muscle is twitching that actually causes my finger to move. So we basically, in our first years of life, build up this massive lookup table between whatever neuronal firing we do and whatever action happens in our body that we control. If you have a kid grow up with a third limb...
2:02:21 Manolis Kellis? I'm sure they'll figure out how to control them probably at the same rate as their natural limbs.
2:02:25 Lex Fridman? And a lot of the work would be done by the, if a third limb is a computer, you kind of have a, not a faith, but a thought that the brain might be able to figure out, like the plasticity would come from the brain.
2:02:43 Manolis Kellis? Yeah.
2:02:44 Lex Fridman? Like the brain would be cleverer than the machine at first.
2:02:46 Manolis Kellis? When I talk about a third limb, that's exactly what I'm saying, an artificial limb that basically just controls your mouse while you're typing, you know, perfectly natural thing. I mean, again, you know, in a few hundred years.
2:02:55 Lex Fridman? Maybe sooner than that.
2:02:58 Manolis Kellis? But, but basically, there's, as long as the machine is consistent in the way that it will respond to your brain impulses, you'll figure out how to control that and you could play tennis with your third limb. And let me go back to consistency. People who have dramatic accidents that basically take out a whole chunk of their brain can be taught to co-opt other parts of the brain to then control that part. You can basically build up that tissue again and eventually train your body how to walk again and how to read again and how to play again and how to think again, how to speak a language again, et cetera. So there's a massive amount of malleability that happens, you know, naturally in our
2:03:41 Manolis Kellis? way of controlling our body, our brain, our thoughts, our vocal cords, or limbs, et cetera. And human-machine interfaces are inevitable if we sort of figure out how to read these electric impulses, but the resolution at which we can understand human thought right now is nil, is ridiculous. So how are human thoughts encoded? It's basically combinations of neurons that co-fire. And these create these things called n-grams that eventually form memories and so on and so forth. We know nothing of all that stuff. So before we can actually read into your brain that you want to build a program that does this and this and this and that.
2:04:25 Manolis Kellis? We need a lot of neuroscience.
2:04:27 Lex Fridman? Well, so to push back on that, do you think it's possible that without understanding the functionality about the brain or the from the neuroscience or the cognitive science or psychology, whichever level of the brain will look at, do you think if we just connect, connect them, just like per your previous point, if we just have a high enough resolution between connection between Wikipedia and your brain, the brain will just figure it out with less understanding, because that's one of the innovations of Neuralink is they're increasing the number of connections to the brain to like several thousand, which before was, you know, in the dozens or whatever.
2:05:05 Manolis Kellis? You're still off by a few orders of magnitude, only order of seven.
2:05:11 Lex Fridman? Right, but the thing is, the hope is if you increase that number more and more and more, maybe you don't need to understand anything about the actual how human thought is represented in the brain. You could just let it, let it figure it out by itself.
2:05:25 Manolis Kellis? Yeah, it's all Keanu Reeves waking up and saying, I know Kung Fu.
2:05:27 Lex Fridman? Yeah, exactly.
2:05:30 Manolis Kellis? So, yeah, sure.
2:05:31 Lex Fridman? You don't have faith in the plasticity of the brain to that degree?
2:05:35 Manolis Kellis? It's not about brain plasticity. It's about the input aspect. Basically, I think on the output aspect, being able to control a machine is something that you can probably train your neural impulses that you're sending out to sort of match whatever response you see in the environment. If this thing moved every single time I thought a particular thought, then I could figure out, I could hack my way into moving this thing with just a series of thoughts. I could think guitar, piano, tennis ball, and then this thing would be moving. And then, you know, I would just have the series of thoughts that would sort of result in the impulses that will move this thing the way that I want. And then eventually it'll become natural because I won't even think about it. I mean, in the same way that we control our limbs in a very natural way. But babies don't do that. Babies have to
2:06:19 Manolis Kellis? figure it out. And, you know, some of it is hardcoded, but some of that is actually learned based on the whatever soup of neurons you ended up with, whatever connections you pruned them to, and eventually you were born with. You know, a lot of that is coded in the genome, but a huge chunk of that is stochastic. Instead of the way that you sort of create all these neurons, they migrate, they form connections, they sort of, you know, spread out, they have particular branching patterns, but then the connectivity itself, unique in every single new person.
2:06:48 Manolis Kellis? All this to say that on the output side, absolutely, I'm very, very, you know, hopeful that we can have machines that read thousands of these neuronal connections on the output side, but on the input side, oh, boy. I don't expect any time in the near future, we'll be able to sort of send a series of impulses that will tell me, oh, earth to sun distance, seven point five million, et cetera, like nowhere. I mean, I think language will still be the input way, rather than sort of any kind of more complex.
2:07:24 Lex Fridman? It's a really interesting notion that the ambiguity of language is a feature.
2:07:28 Manolis Kellis? Yeah.
2:07:29 Lex Fridman? And we evolved for millions of years to take advantage of that ambiguity.
2:07:36 Manolis Kellis? Exactly. And yet, no one teaches us the subtle differences between words that are near cognates, and yet evoke so much more than, you know, one from the other.
2:07:47 Lex Fridman? Wow.
2:07:48 Manolis Kellis? And yet, you know, when you're choosing words from a list of 20 synonyms, you know exactly the connotation of every single one of them. And that's something that, you know, is there. So, yes, there's ambiguity, but there's all kinds of connotations. And in the way that we select our words, we have so much baggage that we're sending along, the way that we're emoting, the way that we're moving our hands every single time we speak, the pauses, the eye contact, et cetera. So much higher baud rate than just a vocal, you know, string of characters.
2:08:21 Lex Fridman? Well, let me just take a small tangent on that.
2:08:24 Manolis Kellis? Oh, tangent, we haven't done that yet.
2:08:25 Lex Fridman? Haven't done that.
2:08:26 Manolis Kellis? It's a good idea, let's just tangent.
2:08:27 Lex Fridman? We'll return to the origin of life after. So, I mean, you're Greek, but I'm going on this personal journey. I'm going to Paris for the explicit purpose of talking to one of the most famous, a couple who's a famous translators of Russian literature, Dostoevsky, Tolstoy.
2:08:51 Manolis Kellis? Yeah.
2:08:52 Lex Fridman? And they go, that's their art, is the translation.
2:08:55 Manolis Kellis? Yeah.
2:08:56 Lex Fridman? And everything I've learned about the translation art, it makes me feel, it's so profound in a way that's... So much more profound than the natural language processing papers I read in the machine learning community, that there's such depth to language that I don't know what to do with. I don't know if you've experienced that in your own life with knowing multiple languages. I don't know what to, I don't know how to make sense of it, but there's so much loss in translation between Russian and English and getting a sense of that. Like, for example, there's like, just taking a single sentence from Dostoevsky,
2:09:40 Lex Fridman? and like, there's a lot of them. You could, you could talk for hours about how to translate that sentence properly. That captures the meaning, the, the, the period, the culture, the humor, the wit, the suffering that was in the context of the time, all of that. It could be a single sentence. You could, you could talk forever about what it takes to translate that correctly. I don't know what to do with that.
2:10:05 Manolis Kellis? So, being Greek, it's very hard for me to think of a sentence or even a word without going into the full etymology of that word, breaking up every single atom of that, that sentence, and every single atom of these words, and rebuilding it back up. I have three kids, and the way that I teach them Greek is the same way that, you know, the documentary I was mentioning earlier about sort of understanding the deep roots of all of these, you know, words. And it's very, it's very interesting that every single
2:10:47 Manolis Kellis? time I hear a new word that I've never heard before, I go and figure out the etymology of that word, because I will never appreciate that word without understanding how it was initially formed.
2:10:56 Lex Fridman? Interesting. But how does that help? Because that's, that's not the full picture.
2:11:00 Manolis Kellis? No, no, of course, of course. But what I'm trying to say is that knowing the components teaches you about the context of the formation of that word, and sort of the original usage of that word. And then, of course, the word takes new meaning as you create it, you know, from its parts. And that meaning then gets augmented, and two synonyms that, that sort of have different roots will actually have implications that carry a lot of that baggage of the historical provenance of these words. So before working on genome evolution, my passion was evolution of language.
2:11:36 Manolis Kellis? And sort of tracing cognates across different languages through their etymologies.
2:11:44 Lex Fridman? And that's fascinating that there's parallels between, I mean,
2:11:47 Manolis Kellis? Of course.
2:11:48 Lex Fridman? The idea that there's evolutionary dynamics to our language.
2:11:53 Manolis Kellis? Yeah. Every single word that you utter, parallels, parallels. What does parallels mean? Para means side by side. Alleles from alleles, which means identical twins. Parallels. I mean, name any word and there's so much baggage, so much beauty in how that word came to be and how this word took a new meaning than the sum of its parts.
2:12:20 Lex Fridman? Yeah, and they're just words. They don't have any physical grounding.
2:12:25 Manolis Kellis? Exactly, and now you take these words and you weave them into a sentence. The emotional invocations of that weaving are fathomless.
2:12:36 Lex Fridman? And they're all, all of those emotions all live in our, in the brains of humans.
2:12:42 Manolis Kellis? In the eye of the beholder. In the eye, no, seriously, you have to embrace this concept of the eye of the beholder. It's, it's the conceptualization that nothing takes meaning with one person creating it. Everything takes meaning in the receiving end. And the emergent properties of these communication networks where every single, you know, if you look at the network of our cells and how they're communicating with each other, every cell has its own code. This code is modulated by the epigenome. This creates a bunch of different cell types. Each cell type now has its own identity, yet they all have the common root of the stem cells that sort of led to them. Each of these identities is now communicating with each other.
2:13:25 Manolis Kellis? They take meaning in their interaction. There's an emergent property that comes from a bunch of cells being together that is not in any one of the parts. If you look at neurons communicating, again, these n-grams don't exist in any one neuron. They exist in the connection, in the combination of neurons. And the meaning of the words that I'm telling you.
2:13:47 Manolis Kellis? Is empty until it reaches you and it affects you in a very different way than it affects whoever's listening to this conversation now. Because of the emotional baggage that I've grown up with, that you've grown up with and that they've grown up with.
2:14:00 Lex Fridman? Yeah.
2:14:01 Manolis Kellis? And that's, I think, the magic of translation. If you start thinking of translation as just simply capturing that emotional set of reactions that you evoke, you need a different set of words to evoke that same set of reactions to a French person than to a Russian person because of the baggage of the culture that we grew up in.
2:14:25 Lex Fridman? Yeah, I mean, there's.
2:14:26 Manolis Kellis? So, so, so basically, you shouldn't find the best word. Sometimes it's a completely different sentence structure that you will need matched to the cultural context of the target audience that you have.
2:14:40 Lex Fridman? Yeah, the, it's, it's, I mean, you're just, I usually don't think about this, but right now there's this feeling, as a reminder, that it's just you and I talking, but there's several hundred thousand people will listen to this. There's some guy in Russia right now, running, like in Moscow, listening to us. And there's somebody in India, I guarantee you, there's somebody in China and South America, there's somebody in Texas, and, and they all have different.
2:15:09 Manolis Kellis? Emotional baggage.
2:15:10 Lex Fridman? They probably got angry earlier on about the whole discussion about coronavirus and about some aspect of it. Uh, yeah, it's, and there's that network effect.
2:15:21 Manolis Kellis? Yeah, yeah, yeah.
2:15:22 Lex Fridman? That's, uh.
2:15:23 Manolis Kellis? It's a beautiful thing, and, and this lateral transfer of information, that's what makes the collective quote unquote genome of humanity so unique from any other species.
2:15:36 Lex Fridman? So you somehow miraculously wrapped it back to the very beginning of when we're talking about the human, the beauty of the human genome. So I think this is the right time, unless we want to go for a six to eight hour conversation, we're gonna have to talk again, but I think for now, to wrap it up, this is the right time to talk about the, uh, the biggest, most ridiculous question of all, meaning of life. Off mic, you mentioned to me that you, um. You had your 42nd birthday, 42nd being a very special, absurdly special number. And you had a kind of a get together with friends to discuss the meaning of life.
2:16:21 Lex Fridman? So let me ask you, in your, as a biologist, as a computer scientist, and as a human, what is the meaning of life?
2:16:31 Manolis Kellis? I've been asking this question for a long time, ever since my 42nd birthday, but well before that, in even planning the Meaning of Life Symposium. And symposium, sim means together, posse actually means to drink together. So symposium is actually a drinking party.
2:16:50 Lex Fridman? Can you actually elaborate about this Meaning of Life Symposium that you put together? It's like the most genius idea I've ever heard.
2:16:58 Manolis Kellis? So 42 is obviously the answer to life, the universe and everything, from the Hitchhiker's Guide to the Galaxy. And as I was turning 42, I've had the theme for every one of my birthdays. When I was turning 32, it's 1 0 0 0 0 in binary. So I celebrated my 100,000th binary birthday, and I had a theme of going back 100,000 years. You know, let's dress something in the last 100,000 years. Anyway, it was, I've always had these.
2:17:26 Lex Fridman? That's, you're such an interesting human being. Okay, that's awesome.
2:17:30 Manolis Kellis? I've always had these sort of numerology-related announcements for my birthday parties. So what came out of that Meaning of Life Symposium is that I basically asked 42 of my colleagues, 42 of my friends, 42 of my collaborators, to basically give seven minute speeches on the meaning of life, each from their perspective. And I really encourage you to go there, because it's mind-boggling that every single person said a different answer. Every single person started with, "I don't know what the meaning of life is, but..." and then gave this beautifully, eloquently answer, eloquent answer.
2:18:12 Manolis Kellis? And they were all different, but they all were consistent with each other, and mutually synergistic, and together forming a beautiful view of what it means to be human in many ways.
2:18:25 Manolis Kellis? Some people talked about the loss of their loved one, their life partner for many, many years, and how their life changed through that. Some people talked about the origin of life. Some people talked about the difference between purpose and meaning. I'll, you know, maybe quote one of the answers, which is this linguistics professor, friend of mine at Harvard, who basically said that she was gonna... She's Greek as well, and she said, "I will give a very Pythian answer." So Pythia was the oracle of Delphi, who would basically give these very cryptic answers, very short, but interpretable in many different ways. There was this whole set of priests who were tasked with interpreting what Pythia had said.
2:19:10 Manolis Kellis? And very often, you would not get a clean interpretation. But she said, "I will be like Pythia and give you a very short and multiply interpretable answer. But unlike her, I will actually also give you three interpretations." And she said, "The answer to the meaning of life is become one." And the first interpretation is, "Like a child, become one-year-old with the excitement of discovering everything about the world." Second interpretation, "In whatever you take on, become one, the first, the best, excel, drive yourself to perfection for every one of your tasks. And become one when people are separate.
2:19:55 Manolis Kellis? Become one, come together, learn to understand each other." Damn, that's an answer. And one way to summarize this whole Meaning of Life symposium is that the very symposium was illustrating the quest for meaning, which might itself be the meaning of life. This constant quest for something sublime, something human, something intangible, some, you know, aspect of what defines us as a species and as an individual. Both the quest of me as a person through my own life, but the meaning of life could also be the meaning of all of life. What is the whole point of life? Why life?
2:20:39 Manolis Kellis? Why life itself? Because we've been talking about the history and evolution of life, but we haven't talked about why life in the first place? Is life inevitable? Is life...
2:20:51 Manolis Kellis? Part of physics. Does life transcend physics? By fighting against entropy? By compartmentalizing and increasing concentrations rather than diluting away? Is life a distinct entity in the universe? Beyond the traditional, very simple, physical rules that govern gravity and electromagnetism and all of these forces? Is life another force? Is there a life force? Is there a unique kind of set of principles that emerge? Of course, built on top of the hardware of physics. But is it sort of a new layer of software? Or a new layer of a computer system? So that's at the level of, you know, big questions.
2:21:35 Manolis Kellis? There's another aspect of gratitude of basically... What I like to say is, during this pandemic, I've basically worked from 6:00 a.m. until 7:00 p.m. every single day, nonstop, including Saturday and Sunday. I've basically broken all boundaries of where life, personal life begins and work life, you know, ends. And that has been exhilarating for me. Just the intellectual pleasure that I get from a day of exhaustion, where at the end of the day, my brain is hurting, and I'm telling my wife, "Wow, I was useful today."
2:22:17 Manolis Kellis? And there's a certain pleasure that comes from feeling useful. And there's a certain pleasure that comes from feeling grateful. So I've written this little sort of prayer for my kids to say at bedtime every night, where they basically say, "Thank you, God, for all you have given me, and give me the strength to give on to others with the same love that you have given on to me."
2:22:50 Manolis Kellis? We as a species are so special. The only ones who worry about the meaning of life. And maybe that's what makes us human. And what I like to say to my wife and to my students during this pandemic work extravaganza is every now and then they ask me, but how do you do this? And I'm like, I'm a workaholic. I love this. This is me in the most unfiltered way. The ability to do something useful, to feel that my brain is being used, to interact with the smartest people on the planet, day in, day out, and to help them discover aspects of the human genome, of the human brain,
2:23:35 Manolis Kellis? of human disease and the human condition that no one has seen before with data that we're capturing that has never been observed. And there's another aspect, which is on the personal life. Many people say, oh, I'm not going to have kids. Why bother? I can tell you, as a father, they're missing half the picture, if not the whole picture. Teaching my kids about my view of the world and watching through their eyes the naivete with which they start and the sophistication with which they end up. The understanding that they have of not just the natural world around them,
2:24:19 Manolis Kellis? but of me too. The unfiltered criticism that you get from your own children that knows no bounds of honesty. And I've grown components of my heart that I didn't know I had until you sense that fragility, that vulnerability of the children.
2:24:51 Manolis Kellis? That immense love and passion, the unfiltered egoism that we as adults learn how to hide so much better. It's just this back of emotions that tell me about the raw materials that make a human being and how these raw materials can be arranged with more sophistication that we learn through life to become truly human adults.
2:25:19 Manolis Kellis? But there's something so beautiful about seeing that progression between them. The complexity of the language growing as more neural connections are formed to realize that the hardware is getting rearranged as their software is getting implemented on that hardware. That their frontal cortex continues to grow for another 10 years. There's neural connections that are continuing to form, new neurons that actually get replicated and formed. And it's just incredible that we have this, not just you grow the hardware for 30 years and then you feed it all of the knowledge. No, no, the knowledge is fed throughout and is shaping these neural connections as they're forming.
2:26:05 Manolis Kellis? So seeing that transformation from either your own blood or from an adopted child is the most beautiful thing you can do as a human being. And it completes you, it completes that path, that journey. The create life? Oh, sure, that's a conception, that's easy. But create human life to add the human part, that takes decades of compassion, of sharing, of love and of anger and of impatience and patience. And as a parent, I think I've become a very different kind of teacher. Because again, I'm a professor.
2:26:44 Manolis Kellis? My first role is to bring adult human beings into a more mature level of adulthood, where they learn not just to do science, but they learn the process of discovery and the process of collaboration, the process of sharing, the process of conveying the knowledge, of encapsulating something incredibly complex and sort of giving it up in sort of bite-sized chunks that the rest of humanity can appreciate. I tell my students all the time, if you, you know, like when an apple falls, when a tree falls in the forest and no one's there to listen, has it really fallen? The same way, you do this awesome research. If you write an impenetrable paper that no one will understand, it's as if you never did the awesome research.
2:27:28 Manolis Kellis? So conveying of knowledge, conveying this lateral transfer that I was talking about at the very beginning, of sort of humanity and sort of the sharing of information, all of that has gotten so much more rich by seeing human beings grow in my own home. Because that makes me a better parent, and that makes me a better teacher and a better mentor to the nurturing of my adult children, which are my research group.
2:28:00 Lex Fridman? First of all, beautifully put, connects beautifully to the vertical and the horizontal inheritance of ideas that we talked about at the very beginning. I don't think there's a better way to end it on this poetic and powerful note. Manaus, thank you so much for talking to us. It's a huge honor. We'll have to talk again about the origin of life, about epigenetics, epigenomics, and some of the incredible research you're doing. Truly an honor. Thanks so much for talking today.
2:28:32 Manolis Kellis? Thank you. Such a pleasure. It's such a pleasure. I mean, your questions are outstanding. I've had such a blast here. I can't wait to be back.
2:28:38 Lex Fridman? Awesome.
2:28:40 Lex Fridman? Thanks for listening to this conversation with Manolis Kellis, and thank you to our sponsors, Blinkist, Eight Sleep, and MasterClass. Please consider supporting this podcast by going to blinkist.com/lex, eightsleep.com/lex, and masterclass.com/lex. Click the links, buy the stuff, get the discount. It's the best way to support this podcast. If you enjoy this thing, subscribe on YouTube, review it with five stars on Apple Podcasts, support it on Patreon, or connect with me on Twitter at Lex Fridman. And now, let me leave you with some words from Charles Darwin that I think Manolis represents quite beautifully. "If I had my life to live over again, I would have made a rule to read some poetry
2:29:24 Lex Fridman? and listen to some music at least once every week." Thank you for listening and hope to see you next time.
Previous #112 – Ian Hutchinson: Nuclear Fusion, Plasma Physics, and Religion Next #114 – Russ Tedrake: Underactuated Robotics, Control, Dynamics and Touch Page updated October 26, 2026 · transcript and summary are machine-generated