Dmitry Korkin on Protein Domains, Coronavirus Structure, AlphaFold 2 and the Origin of Life

High-Level Overview

In his second appearance on the Lex Fridman Podcast, Dmitry Korkin, a professor of bioinformatics and computational biology at WPI, walks Lex from the smallest working parts of life up to some of the largest open questions in science and artificial intelligence. They begin with proteins and the structural units called protein domains, move through the architecture of the SARS-CoV-2 particle and the way it mutates, then widen out to the evolution of proteins over billions of years, self-copying computer programs, the origin of life and the chance of finding life elsewhere in the solar system. The second half turns to AI: the story of Nobel laureate Joshua Lederberg and the 1960s Dendral expert system, why expert systems faded away, what DeepMind's AlphaFold 2 did and did not solve, whether AI will ever produce hit music or art, and whether machine learning could be used to engineer more dangerous viruses.

The tone is curious, warm and technical in places, with Lex repeatedly asking for the basics to be spelled out and Korkin answering with the enthusiasm of someone who plainly loves both biology and computing. It closes on a personal note, with book recommendations, a New Year's resolution about family, and a poem recited in Russian. The conversation swings between molecular biology, the history of AI, speculation about life in the universe, and personal reflection.

Table of contents

The Discussion Explores:

• Why protein domains, rather than whole proteins, are the real functional and evolutionary building blocks of life
• How the four structural proteins of SARS-CoV-2 assemble into a viral particle, and why the membrane protein may be a better drug target than the spike
• What virus mutations mean for vaccines, species jumps and the efficiency of the virus as a machine
• Modularity, linkers and alternative splicing as layers of complexity in protein evolution
• Self-replicating computer programs, nature-inspired algorithms and the idea of software that evolves
• The origin of life, the Drake equation, the rare earth hypothesis and hints of life on Venus, Mars and Europa
• Joshua Lederberg, the Dendral project and why expert systems disappeared into modern machine learning
• What AlphaFold 2 actually achieved, where protein folding remains unsolved, and whether AI could engineer pathogens

---

Deep Dive into Main Ideas

Proteins, and why domains are the real building block

Lex opens by asking whether proteins and their amino acid residues are the basic building blocks of life. Korkin answers yes and no. Proteins are the basic biological unit that carries out cellular function, but studying and comparing them across species and kingdoms reveals what he calls modular complexity: an average protein consists of several structural units known as protein domains. He offers a memorable image, describing a protein as a string of beads where each bead is a protein domain. Many proteins have more than one function, and those functions are often carried out by individual domains; in evolution, those domains get shuffled around as units.

Korkin explains why popular culture still talks about proteins rather than domains, and the reason is largely historical. Early structural biology could only resolve the 3D coordinates of small proteins, and small proteins tend to be single-domain, so protein and protein domain were effectively the same thing. That reinforced the idea that proteins are globular blobs; in fact, Korkin says, the globular part is the domain. The classic techniques were X-ray crystallography and NMR spectroscopy, and the recent breakthrough has been cryo-electron microscopy, which reaches much larger molecules and molecular complexes. As an example, the first experimental structure of a SARS-CoV-2 protein was the cryo-EM structure of the spike protein, solved very quickly.

He also describes complexes, where multiple protein chains assemble. Copies of the same protein forming a complex make a homo-oligomer; the spike is three chains coupled together, so that viewed from the top it looks like a perfect triangle. Hemoglobin is his example of a complex built from four subunits, two identical pairs that are also similar to each other, which hints that it may once have been four identical copies before mutating into something more specialized.

The spike protein and its unresolved third

Asked what is beautiful about the spike, Korkin says first of all that it is an incredibly challenging protein. As part of his lab's effort to structurally decode every protein in the virus's proteome, they worked on the spike and ran into the limits of the data: the cryo-EM structure covers roughly two thirds of the protein, while the remaining third is buried in the membrane of the viral envelope and surrounded by unstable structures. People are still trying to understand the nature and role of that third.

The top part's primary function is to attach to the human ACE2 receptor, and Korkin describes the mechanics with evident pleasure. Because the trimer has three chains, there are three receptor binding domains, or RBDs, which untangle and get ready to bind, and they do not move in sync. Typically only one arm goes out at a time. He cites recent work by a group from UMass Medical School, including Jeremy Luban and other faculty, who solved a mutated spike structure and showed that with most of these mutations more than one arm opens up, so the frequency of two arms being up increases drastically. That potentially changes the dynamics, since the particle now has two chances to attach to ACE2. The attachment of the spike trimer to ACE2, a molecule sitting on the surface of the human cell, is the first step that triggers encapsulation. Lex jokes that if this were dating, it would be the first date.

How the viral particle is built, and where to attack it

Korkin gives a compact inventory of the four structural proteins that make up the virus. S, the spike, acts as a trimer and needs three copies. E, the envelope protein, acts as a pentamer and needs five copies. M is the membrane protein, forms dimers, and assembles into what he calls beautiful lattices that appear in their simulations as a neat grid of dimers attached to one another. N is the nucleocapsid protein, which sits inside, protects the viral RNA and likely interacts with M; it acts as a shield and may also contribute to the stability of the outer shell.

The numbers matter. Per viral particle there are only two or three E pentamers, on the order of a thousand M dimers making up most of the outer shell, and roughly 50 to 90 spike trimers. Korkin stresses that the envelope's assembly is not random: a completely random process would not produce the ellipsoid envelope shell that is observed, so there must be regularity in how the dimers attach in a specific, directed way. That regularity is not yet understood, and it is what his lab has been working on.

This leads to the most practical part of the discussion. Antibodies target parts of the spike, and small molecules can bind strategic parts of proteins to disrupt their function, but Korkin points to a newer direction: targeting the M dimer that makes up the outer shell, because destroying the shell destroys the particle itself. Lex frames this as choosing an attack vector, as in cybersecurity, and Korkin agrees, noting that evolutionary analyses show the M protein is more stable than the spike, meaning it does not evolve as drastically and is therefore a more static target. He also mentions another possible treatment: nanoparticles resembling the viral shape, with spike integrated, that would compete with the real virus by blocking ACE2 receptors.

Mutations, species jumps and the arms race with vaccines

Lex raises the mutation news out of the United Kingdom and South Africa. Korkin frames mutation as the general way viruses evolve, and the way they were able to jump from one species to another. He notes recorded incidents of the virus jumping from humans to dogs, and explains the danger of such jumps: every jump comes with mutations driven by the environment of the new host, which differs from the human environment, and there is no way to know in advance whether mutations acquired in another species are neutral or damaging for humans.

On the worry that vaccination puts selective pressure on the virus, Korkin is measured. He acknowledges the discussion of an arms race between humanity vaccinating faster and the virus becoming resistant, and notes that mutation is the reason we get vaccinated every year against seasonal flu, but says he does not worry that much because there is not much evidence of aggressive mutation around the vaccine. What he does argue for is study: understanding the virus's evolution is, to him, one of the critical directions, because it could reveal the mechanisms that let viruses jump between species and become resistant to vaccines and treatments, and might eventually let scientists forecast future evolutionary traces.

When Lex asks which parts of the virus, if souped up by mutation, would make it more effective, Korkin says the short answer is that we do not know yet, but the capacity is there. The virus is a machine performing many functions, and many of those functions are nearly perfect but not perfect. Attachment to ACE2 is the obvious example: there may be undiscovered mutations that make the attachment stronger or more efficient from the virus's point of view. Equally, some mutations could make a function much less efficient.

The evolution of proteins: modularity, linkers and splicing

Lex references a 2010 Nature paper on the ongoing expansion of the protein universe and the idea of a common protein ancestor, and asks what is beautiful about the story from that first ancestor to the diversity of today. If he had to pick a single keyword for protein evolution, Korkin says, it would be modularity. Proteins are no longer treated as a sequence of letters; there are hierarchical complexities in how they are organized, reaching back to the gene and the nucleotide sequence. Domains are not only functional building blocks but evolutionary ones. Once structurally and functionally stable building blocks were discovered, they stayed, which is why comparing different proteins reveals similar fragments belonging to what are called protein domain families. Mutations within them drive diversification of function, but evolutionary events that split a domain into fragments are very rare, because splitting a domain cancels or reduces its function.

Above domains sit the linkers that connect them, which are much less studied. These are completely flexible parts of the protein that nevertheless carry out a great deal of function, along with the C and N termini at each end of the sequence, which are involved in very specific protein-protein interactions. Linker length has direct consequences: a very short linker forces two domains to sit next to each other, while a long one allows extensive spatial reorganization. Because linkers are so flexible, they can adapt to many shapes, which makes them good interactors with other proteins. They also evolve under different rules than domains, since they do not need to preserve a particular structure.

The least studied layer is alternative splicing, something Korkin's lab has been fascinated by for over a decade. The simplistic picture is one gene, one protein product, but in eukaryotes, especially complex recent ones, a gene can produce multiple functionally active protein products. He explains the mechanism: genes alternate blocks that are translated, called exons, and blocks that are cut out, called introns, sometimes dozens of each. When the gene is converted to RNA the introns are cut out and the exons assembled, but sometimes exons are thrown out or swapped, producing a different protein product. Lex asks whether it is random, and Korkin says it is not. Single-cell and tissue-level next-generation sequencing techniques such as RNA-seq show that these are dynamic events happening in response to disease or to a particular developmental stage of a cell. He adds a striking connection: the boundaries of exons often coincide with the boundaries of protein domains, though not always, so there is an interplay between the gene and RNA world and the protein world that evolution picks up on.

Self-replicating programs and nature-inspired computing

Lex brings up quines, self-printing computer programs he attributes to Douglas Hofstadter, which he compares to exons and introns and enjoys writing in different languages. Korkin recalls that as a kid, writing a self-replicating program was one of the stages you had to reach in mathematics olympiad circles to be considered any good, and that the harder version was finding the shortest such program, which depends on the language. Lex mentions Code Golf, the competition site where people write the shortest program for an arbitrary task and where you discover how strange some programming languages are.

The deeper question Lex asks is whether we will ever have programs that evolve on their own: not evolutionary algorithms, but programs that mate, mutate and replicate, so that the system can run away rather than being something you turn off. Korkin sees no reason why not, treating it as a natural evolution of the intelligent agents area of computer science, where independent codes run, interact and exchange information. Lex speculates about bots evolving alongside humans in social networks, and about robotics labs, he mentions Harvard and Penn, whose task is to build a robot that can build another copy of itself given extra resources, along with self-assembling bridges and buildings made of robotic parts.

Korkin connects this back to algorithms, noting that recombination events cause global changes in a search trajectory while mutation events are a more refined search step, and points to other nature-inspired methods. His favorite example is a slime-based algorithm, first introduced, he believes, by a Japanese group, which solved some complex problems. He argues there is much more to borrow from nature, noting that neural networks take only a very crude inspiration from it. Lex adds that the immune system seems to have something to do with how the brain operates, and that none of that is modeled in artificial neural networks.

The origin of life and the odds of finding it elsewhere

Lex introduces the Drake equation, which combines factors such as how many stars are born each year and how many habitable planets there are to estimate the number of alien civilizations in the galaxy, and says he puts the probability of life emerging on a habitable planet at one percent. Korkin's entry point is a 2018 Nature paper reporting that glycine, one of the simplest of the twenty basic amino acids that make up proteins, was found in comet dust collected from a comet's tail, which he identifies as Churyumov-Gerasimenko. That raises the question of whether alien life would be made of proteins at all, or of RNA, given that RNA viruses are a well established group of molecular machines.

Pressed for his own number, Korkin says he would intuitively put it lower than one percent while stressing he is not the best person to make such estimates, and adds that Earth was really lucky, since conditions several billion years ago were far from ideal. Lex describes the rare earth hypothesis as the counterpoint to the Drake equation: Earth may be a special, perhaps near-unique place, with its shielding from asteroids, its distance from the sun, and its balance of water and land all essential for life. Korkin's instinct is that reality rarely falls at either extreme, so his guess is somewhere in between: not unique, but with reasonable if small chances. Lex offers the other extreme, a Stephen Wolfram style view in which great complexity emerges from simple rules and life simply finds a way, so that life could be so common it is laughable, like ants believing their colony is the only thing that exists.

On our own solar system, they touch on possible signs of life on Venus in gaseous form, the hope for extinct life on Mars, and moons such as Europa. Korkin hopes we find life, says the Venus news was exciting, and expects any such life to be carbon-based, though how similar it would be is the big question. Discovering even a tiny single cell outside Earth would, he says, be another turning point for science; Lex adds that if its specifics were different in some new way, that would be a strong statement that life is everywhere in the universe.

Joshua Lederberg, Dendral, and the fate of expert systems

Korkin tells a personal story about Joshua Lederberg, who won the 1958 Nobel Prize in Physiology and Medicine for discovering that bacteria can mate and exchange genes. Korkin began his PhD in Canada in 2000, developing a new language for symbolic machine learning as an alternative to feature-based approaches, applied to chemoinformatics and computer-aided drug design; his system looked at chemical compounds in the same therapeutic category, such as male hormones, and tried to identify which structural fragments defined the class rather than merely completing the structure. Trained as an applied mathematician, he asked many questions on CCL, a computational chemistry bulletin board. One night he posted a question he still remembers titled along the lines of desperately looking for a chemist's advice, and the next morning his advisor phoned to say Lederberg had replied. Korkin's reaction was to ask who Joshua Lederberg was. Lederberg wrote that they had conceptually similar ideas in the Dendral project and that Korkin might want to look it up.

Lederberg was by then professor emeritus at Rockefeller University, and Korkin says the hope of meeting him was one reason he joined Andrej Sali's group there as a postdoc. He did meet him briefly, on the little bridge connecting the research campus to the Rockefeller skyscraper where postdocs, faculty and graduate students live, for a very short conversation. Lederberg died in 2008.

Reading about Dendral amazed him, because the ideas from around 1960 were so profound. Lederberg wanted a system to help study extraterrestrial molecules via mass spectrometry: the spectrum gives information about possible fragments and atoms, and the task is to work backwards to the whole molecule. His idea was to connect chemistry and computer science in an expert system that took the mass spec data plus a database of possible molecules and induced the molecule corresponding to the spectrum. What the project ended up doing was providing a list of candidates for a chemist to choose from, though Lex notes the original ambition was full automation. Korkin calls this essentially the origin of modern bioinformatics and cheminformatics, in the 1960s.

Asked why expert systems did not succeed, Korkin says he teaches artificial intelligence at his university and devotes his first two lectures to the history of AI, where this is a genuinely interesting question. He describes two opposing lines of thought in the historical literature: that expert systems simply were not up to expectations and were replaced, or the opposite, that they were so good they became a household name and were transformed. Either way the outcome was the same, they evolved into something else, and Korkin thinks that something is modern machine learning, because the most successful algorithms, including AlphaFold, build in knowledge about the domain being studied.

AlphaFold 2: what was solved, and what was not

On AlphaFold 2 and the claim that protein folding is solved, Korkin calls it definitely a very exciting achievement and then supplies context. Bioinformatics has several competitions, sometimes described to outsiders as the bioinformatics Olympic games, covering protein structure prediction, protein function prediction, the effects of mutations on function, and protein-protein interactions. The oldest is CASP, the critical assessment of protein structure prediction, where experimentalists solve structures but withhold them from the Protein Data Bank, releasing only sequences so that the community can predict the 3D structures, which are then compared against the experimental results.

Folding itself remains mechanistically mysterious. One hypothesis is that folding happens modularly, with domains folding independently before the whole structure forms, but within domains there is secondary structure, small alpha helices and beta sheets, which are structurally stable elements, and it is unclear when these form, since some require a fragment from the beginning of the sequence and one from the middle. Korkin notes that folding is extremely efficient and stable, happening very quickly and almost the same way every time, which both he and Lex find remarkable. Folding also begins before translation finishes, with structural fragmentation visible as the protein emerges from the ribosome. Being able to predict the end product brings us a step closer to the mechanism, because the predictor can be used as a tool: modify a protein, run the prediction, and see whether the result is unstable.

Historically, top CASP groups advanced incrementally and did not use machine learning, relying instead on advanced biophysics combined with bioinformatics and data mining, which allowed them to predict structures for proteins without structurally solved relatives. Where relatives exist, homology or comparative modeling borrows heavily from known structures; de novo prediction without relatives is where it gets very hard. Then the original AlphaFold arrived in 2018 and was suddenly much better than everyone else, which Korkin describes as a shockwave through the community. What it does, he explains, is predict contacts between different parts of the protein, the parts in close proximity to each other, producing what is called a contact map; Lex describes it as estimating a distance matrix. Once the contact map exists, reconstruction becomes much more straightforward, so the contact map is the key. In the most recent round, those ideas had started appearing in other groups' work, and yet AlphaFold 2 again outperformed everyone with another wave of machine learning ideas.

On the use of evolutionary information, Korkin explains that protein structure is more conserved than sequence: sequences can differ greatly while the shape stays conserved. Structures would be ideal input but are often unavailable, whereas sequences are abundant, with hundreds or thousands of organisms sequenced, so aligning the same protein across organisms reveals what is conserved and therefore structurally stable, and what is diverse. He argues that while a purely ab initio approach using only the sequence is tempting, we are overwhelmed with this data, so why not use it. Both note that no paper was out at the time, only a blog post, and Lex jokes that interpreting it is like interpreting scripture.

The key caveat is scope. CASP has focused on single or maybe two domain proteins that are still very compact, and even those are extremely challenging. An average protein has two or three domains, and proteins involved in the nervous system, one of the most recently evolved systems, are mostly highly multi-domain, some with five, six, seven or more domains. We are very far from understanding how those fold. Korkin does think AlphaFold solves a large majority of the more compact proteins and domains at near experimental level, and that this matters because understanding multi-domain folding requires understanding individual domains first. Lex raises the data problem: unlike AlphaZero or MuZero, which learn in simulation without data, solved structures number only in the low millions at most. Korkin suggests good-quality homology models, derived from evolutionary information, could enhance the training set, and says he is very optimistic, calling this one of those turning events where a machine learning system is truly better than conventional biophysics-based methods.

When Lex asks where AlphaFold ranks among AI breakthroughs, listing AlphaGo and AlphaZero, Deep Blue beating Kasparov, AlexNet, GPT-3 and transformers, Boston Dynamics, and autonomous vehicles from Waymo and Tesla, Korkin says that in terms of impact on science and society it would be in his top three, while placing Deep Blue's 1997 victory first as a genuine shock, given how human a game chess is and how good Kasparov was. Lex notes that Deep Blue was not machine learning but used advanced search, and argues search is not currently seen as fundamental to intelligence but very likely is, since neural networks are themselves performing search over parameters.

AI in art and music

Korkin offers his own favorite addition to the list: Deep Rembrandt, a collaboration between Rembrandt experts in the Netherlands and an AI group that trained an algorithm to replicate Rembrandt's style and printed a portrait that never existed before, on canvas and with much the same kinds of paint. He found it mind-blowing. Lex argues there has not yet been an AlexNet moment for art, and that in music human-created pieces remain far more popular; the moment that would matter, he says, is an AI-created piece of music or a show popular enough to make serious money. He praises iZotope's audio tools, which he uses on the podcast and which let you select parts of the audio to train on, and he singles out their ability to separate voice from music as incredible.

Korkin adds examples from the project component of his AI class, presented in the last two weeks before the Christmas break, which is his favorite part of teaching. One group built an AI producing haiku, trained on English text, and the best selections seemed profound and reasonable. Other groups tried teaching AI to play rock, classical and popular music, and classical turned out to be the most difficult, even though composers like Bach are very mathematical, because classical composers span such a different spectrum that you can simply listen and tell it is not there yet. Lex feels the same about OpenAI's music work: the system is cool but not compelling, and he wonders half-jokingly whether we need a tortured human being behind the music. Korkin says he would not be surprised if one day an AI-written song reaches the top charts.

Multi-protein folding and protein-protein interactions

Asked how hard multi-protein folding is, Korkin says it has components of both protein folding and protein-protein interactions. Many proteins never form a stable structure. His favorite, from his postdoc years, is postsynaptic density 95, or PSD-95, a key actor in the majority of neurological processes at the molecular level and a crucial part of the postsynaptic density where many molecular processes happen. It has five domains and is a large protein, and it is organized flexibly so that it acts as a scaffold, bringing in other proteins so they act in an orchestrated manner. Some parts are stable and some flexible, and that flexibility is built in precisely so the protein can be a multifunctional machine.

Whether AlphaFold-style approaches can learn this is, he says, something time will tell, but to him it is yet another level of complexity, because molecular interactions are governed by different mechanisms that a machine learning algorithm would have to pick up. There is a separate challenge for this, called CAPRI, focused on macromolecular interactions including protein-protein and protein-DNA. Korkin's lab does not usually enter competitions because they are so intensive, but it participated about ten years ago, entering with a scoring function that evaluates whether a predicted interaction looks like the experimentally solved one. They made it a machine learning function, one of the first machine learning based scoring functions used in CAPRI, and learned which components contribute critically to protein-protein interaction, which he takes as evidence that the problem can be converted into a learning problem.

Nobel prizes, protein design and engineered pathogens

On whether AlphaFold 2 will lead to Nobel prizes, Korkin argues the prize itself is evolving with science, which has become multi-faceted and cross-disciplinary, and that computational methods will be acknowledged one way or another. He notes they already were in 2013, when three people, all of them computational biophysicists in his description, were awarded the prize for work on protein folding principles. He regards further recognition as unavoidable with time and has no doubt these methods will be critical in scientific discovery.

On engineering, Korkin notes protein design is already a prominent area of research, and names David Baker and the Rosetta algorithm as pioneers of de novo design used to create new proteins. Designing a protein means designing a function, and one can approach it structurally too: if you want a building block of a certain shape, you can introduce a new sequence and predict how it will fold.

Asked whether machine learning could engineer viruses, Korkin says that luckily we do not have that much data. His lab is currently developing an algorithm to determine whether a given strain is pathogenic, meaning inflicting damage rather than merely spreading. The comparison is possible because there are pathogenic coronavirus strains, SARS-CoV-2 and MERS among them, alongside non-pathogenic coronaviruses such as common cold viruses, plus seasonal versus pandemic strains of influenza. The goal is to find the molecular determinants built into the protein and gene sequence. He mentions a bioRxiv paper by Eugene Koonin, a pioneer of evolutionary genomics, that attempted this with standard supervised learning, and says the open question is whether less standard methods, including transfer learning from what models learn about protein sequences, can go further. Success would let researchers forecast which mutations make a strain more or less pathogenic, which would be crucial for vaccine and antiviral development, though Lex points out the same system could suggest how to make a virus more pathogenic.

Korkin's reassurance is twofold. Designing a sequence is one thing; carrying out the experimental biology to make all the components work is a completely different and difficult process. And regulation has precedent: he recalls an issue around 2015 over research on influenza strains, where groups used mutation analysis to determine whether a strain would jump between species, and there was something like a half-year moratorium on publication until scientists analyzed it and decided it was safe. Lex supplies the term they are reaching for, gain of function and loss of function research, and says he is less worried about well-regulated research than about the longer view, fifty years out. He ties this to the Drake equation parameter for how long civilizations last, which he believes carries the most variability and therefore matters most. Asked whether he would fear natural or engineered pandemics more in fifty or a hundred years, Korkin says he would still worry about natural pandemics, because nature has a pretty good capacity for producing them. Lex adds that viruses are strange weapons because they cannot be targeted, and says he remains optimistic that there is insufficient evil in the world for such destruction. Korkin hopes connectivity makes the world more transparent, and that information spread helps society become more balanced; Lex predicts nations will lose power and that world-destroying military secrets will become a thing of the twentieth century.

What the pandemic taught him, books, and family

Asked what changed since their previous conversation in the spring, Korkin says he was amazed at how efficient the scientific community was. Judging by the narrow domain of protein structure, the structural characterization of SARS in the early 2000s took several years; the equivalent work for SARS-CoV-2 took months. Sequencing is at an unprecedented level, with a single virus sequenced so many times that its evolution can be traced precisely, and sequences anchored to specific geographic places and people, which matters because host genotype influences what is always a host-pathogen coevolution. Lex wishes for richer spread data, mentions that they recently took part in an XPRIZE competition on predicting infection numbers in different regions where AI is central but the training data is weak, and describes Michael Mina's dream of a high-resolution weather map for viruses that would let you tell the day-to-day story of viral mutation and respond to future pandemics.

For books, Korkin offers three plus an honorable mention. The first is Bulgakov's The Master and Margarita, which affected him early in life and still gives him goose bumps at the very end; he was taken by parallel lives centuries apart becoming interwoven into one story, by romance empowered by magic, and by the irony that was unavoidable in Soviet times. The second is Solzhenitsyn's Cancer Ward, which grew out of Solzhenitsyn's own cancer diagnosis and recovery and follows a man sentenced to the camps who develops cancer and is treated in a Soviet republic; Korkin says doctors have called its account of what a patient feels incredibly accurate, and that above the literal story the tumor is an allegory for the Soviet regime. He read both in Russian and notes that while The Gulag Archipelago and other works won Solzhenitsyn the Nobel Prize, Cancer Ward is to him the most powerful. The third is John von Neumann's essay The Computer and the Brain, written while he was dying of cancer and published after his death in 1957. It is thin and very difficult to read because every paragraph is compact and filled with ideas, drawing parallels between the brain, the nervous system and the computers of the time. The honorable mention is an audiobook, Lab Girl by Hope Jahren, a geochemist who studies fossil plants and their chemistry to reconstruct ancient climate. Narrated by the author, it is so personal that in places you can hear her crying, and Korkin calls it a must-listen for anyone who wants to understand academia, science and research, because it is a personal account of becoming a scientist.

His hopes for 2021 are for things to become normal again, especially travel: every summer he goes to an international summer school called the School of Molecular and Theoretical Biology, held in Europe and organized by good friends, for gifted kids from all over the world, and it is the highlight of his year. It had to be held remotely, which was different. His personal resolution comes from working at home and realizing he had been missing time with his family: with research and teaching you are naturally distracted, so being away does not register, and he now intends to spend as much time as possible with his wife and children even once the world reopens. The episode closes with Korkin reciting from memory a recent Russian poem by his namesake Dmitry, titled Sorceress, which he associates with Margarita and which is set around the new year, and with Lex reflecting that Russian poetry captures a magic English does not.

---

Conclusion

The thread running through this conversation is modularity: the idea that life's complexity is built by shuffling and recombining stable units, from protein domains and exons up to the thousand membrane dimers that form a viral shell, and that the same instinct shows up in computing, in self-replicating programs, agent systems and nature-inspired algorithms. Korkin uses that lens to make the pandemic legible, explaining why the membrane protein may be a better drug target than the fast-evolving spike, why species jumps are the mutations to watch, and why the field's speed since SARS is itself a scientific story. He is generous about AlphaFold 2 and precise about its limits, crediting it with near-experimental accuracy on compact proteins and domains while insisting that multi-domain proteins, flexible scaffolds like PSD-95, and protein-protein interaction remain open problems, and arguing that modern machine learning is what the old expert systems became rather than their opposite. The wider speculation, on the origin of life, the Drake equation, whether nature or humans are the better designer of viruses, is held loosely and honestly, with Korkin repeatedly declining to overclaim. What lingers is the personal register of the close: the graduate student who did not know who Joshua Lederberg was, the books read in Russian, and the realization, arrived at while stuck at home, that family time is the thing the busy years quietly take.

---

This is a summary of "#153 – Dmitry Korkin: Evolution of Proteins, Viruses, Life, and AI" from Lex Fridman Podcast, giving you the GIST so you can decide if you want the REST. Listen to the full episode.