The building block of life is a bead on a string, not the whole protein

Lex Fridman opens with the high-school slogan: proteins, and the amino acids that make them, are the building blocks of life. Dmitry Korkin, a professor of bioinformatics and computational biology at Worcester Polytechnic Institute, answers yes and no. Proteins are the biological units that carry out the important functions of the cell. Compare them across species and kingdoms, though, and they stop looking like single globules. They have what he calls modular complexity.

An average protein is several structural units strung together. He asks you to picture a string of beads. Each bead is a protein domain. Over the past twenty years, he says, scientists have become convinced that domains are the unit that actually does the work. Many proteins have more than one function, and those functions are often carried out by those beads. Evolution shuffles the beads. The necklace is just how they happen to be strung.

Popular culture almost never talks this way, and Korkin thinks the reason is historical inertia. Early structures came from small proteins, and small proteins tend to be a single domain, so protein equaled globular molecule. X-ray crystallography and NMR spectroscopy locked that picture in. Only when electron microscopy could resolve much larger molecular machines did the assembled truth become obvious. Fridman tries the turtles-all-the-way-down joke; Korkin still wants the cleanest hierarchical block you can recombine into complex function, and that block is the domain.

One of the first experimental SARS-CoV-2 structures was a cryo-EM map of the spike, solved unusually fast because the imaging had leapt forward. Complexity does not stop at one chain. The spike has to form a trimer — three copies of the same protein coupled together — before it functions. Look at it from the top, Korkin says, and you see a perfect triangle. Hemoglobin, by contrast, is a more mixed complex: four subunits, two pairs of similar chains, a hint that it may once have been four identical copies that specialized.

His lab has been trying to decode every protein in the viral proteome, and the spike is "an incredibly challenging protein." Cryo-EM reconstructs roughly two thirds of it. The rest is buried in the viral envelope, surrounded by unstable structure, chemically tangled with whatever it sits on. The business end is the receptor-binding domain, the RBD, which has to unlatch and attach to the human ACE2 receptor. The three copies do not move in sync. Typically one arm goes up and gets ready.

A group at UMass Medical School led by Jeremy Luban, collaborators of Korkin's, solved a mutated spike and showed that more than one arm opens. The frequency of two arms going up rose quite drastically. If this were dating, Fridman says, it would be the first date. Two arms mean two chances to attach. That first grip on ACE2 is the trigger for the whole entry process, which is why the mechanics of a single raised domain sit at the start of vaccine and drug design.

Antibodies and headlines go after spike. Korkin keeps pointing at the envelope. A particle is built from four structural proteins. Spike is a trimer, roughly 50 to 90 copies per particle. E, the envelope protein, is a pentamer, maybe two or three complexes per particle. M, the membrane protein, forms dimers and a lattice; there are on the order of a thousand of those dimers, and they make up most of the outer shell. N, the nucleocapsid, sits inside as a shield around the RNA. Until recent low-resolution electron microscopy, he says, people had almost no clue how that inner core was organized.

His group had spent the previous six months simulating how M dimers self-organize into a grid. An ellipsoid shell does not grow from a fully random stack. There has to be regularity, a directed way the dimers attach, and that rule is not fully understood. "If you're able to destroy the outer shell, you're essentially destroying the viral particle itself." Small molecules that bind the M dimer and disrupt that lattice are, in his view, a promising direction, not because they are fashionable but because they hit the particle rather than one of its doors.

Fridman reaches for a cybersecurity analogy: is the shell the best attack vector? Promising, Korkin says, with tons of research still to do, and more attack surface. The evolutionary argument is cleaner. From their analysis, M is more stable than spike. Stable means a more static target. It does not evolve as drastically. Spike is the obvious door, heavily guarded and quick to change. M is closer to a load-bearing wall. A target that evolution conserves is one the virus has a harder time mutating away from.

News from the United Kingdom and South Africa is, to Korkin, the ordinary business of viral evolution. Mutation is how these viruses jump species. There had already been incidents of this coronavirus moving from humans to dogs and back. Each jump accumulates mutations driven by a new host environment, and nobody knows whether those changes are neutral, or worse, once the virus returns to people. Change is always a little scary. He is not especially worried that vaccination is already forcing an arms race. There is not much evidence, in early 2021, of aggressive mutation around the vaccine, even though seasonal flu is the reminder that such races exist.

The research program he wants is evolutionary: the mechanisms of species jumps, of vaccine resistance, of drug resistance, in the hope that the traces can be forecast. Fridman asks the blunt engineering question. If you souped up one part of this coronavirus through mutation, which part would make it better at its job? The short answer, Korkin says, is that we do not know yet. The capacity to become more efficient is real.

He treats the virus as a machine that performs many functions, most of them nearly perfect and none of them perfect. Spike-receptor attachment is the obvious example. Has this virus already reached the efficiency of that handshake, or is there an undiscovered mutation that would make the grip stronger? Each protein is there for a reason. Some mutations will enhance a function; some will wreck it. The open optimization space is what a nearly-good machine still has left.

Fridman zooms out to a 2010 *Nature* paper on "the ongoing expansion of the protein universe": a common ancestor, three and a half billion years, millions of proteins still evolving. If Korkin had to pick one keyword for that story, it would be modularity. Proteins are no longer a sequence of letters. They have hierarchical complexity that runs back through the gene to nucleotide sequence. Domains are not only functional building blocks. They are evolutionary ones. Once a structurally stable, functionally clear domain is discovered, it tends to stay.

That is why comparisons across proteins light up similar fragments, the families of domains. Mutations still diversify function inside a family. What you very rarely see is evolution splitting a domain in half. Split the domain and you cancel or at least wreck the function, a bad bargain for the cell. Between domains sit linkers, flexible stretches that are much less studied. A short linker nails two domains together. A long one lets them reorganize in space. Because a linker can adapt to many shapes, it is a good interface for protein-protein interaction, and it evolves under different rules: it does not have to preserve a fold.

Above that sits something his lab has been fascinated by for a decade: alternative splicing. The textbook equation, one gene, one protein, fails in complex eukaryotes. A gene is exons interrupted by introns; the introns are cut out, the exons assembled, and sometimes an exon is dropped or swapped. The products are different, functionally active proteins. The process is not random. RNA-seq and single-cell sequencing show splicing as a dynamic response to disease or developmental stage. Exon boundaries often coincide with domain boundaries. The gene machinery and the protein machinery are coupled. It would be strange, Korkin says, if they were not.

Fridman detours into computer science and the Quine, a program whose only job is to print its own source, named by Douglas Hofstadter. Korkin laughs because it was a gate in informatics olympiads: you were not serious until you could write one. Then the contest became shortness, which is how you end up on a site like Code Golf, watching people solve the weather in bizarre languages. Fridman wants the larger version. Not evolutionary algorithms as search. Actual programs that meet, mutate, copy, and form an ecology you cannot simply power off the way you power off today's machine-learning job.

Korkin does not dodge. He sees no reason it could not happen. It would be a natural evolution of work on intelligent agents, independent pieces of code that run, interact, and exchange information. Fridman finds the possibility both powerful and terrifying: bots that evolve with humans in a sea of social networks, first in digital space. He also likes the physical version, robotics labs at Harvard and Penn whose fundamental task is a robot that, given extra material, builds a copy of itself, or robots that assemble a bridge from copies of themselves.

Add mutation to that self-assembly and you have, in spirit, the same walk from a single origin block to weird complexity that proteins already took. The bits and pieces of the algorithms are already around, Korkin says, even if the usual goal is still to optimize a search: recombination for global jumps, mutation for refined steps. He recommends the slime-mold algorithm, introduced by a Japanese group, which can solve some fairly hard problems. Neural nets are a very crude borrowing from nature. The immune system, and its interplay with the brain, are spices we have barely put into the models.

Fridman had just put a one-percent number on the Drake-equation term for life arising on a habitable planet, and he wants Korkin's instinct. Korkin points to a 2018 *Nature* paper: glycine, one of the twenty amino acids that make proteins, found in comet dust. The ingredients of life are not scarce. That is exciting. It does not tell you whether alien life would even be made of proteins. RNA viruses are already a well-established class of molecular machines. Ingredients are not a self-replicating system.

Intuitively he would put the probability lower than one percent, with the caveat that he is not the person to be making such estimates. Earth got a lot of things right without being ideal. He raises the Rare Earth hypothesis as the counter to Drake: Earth's conditions — shielding from asteroids, distance to the sun, the balance of water and land — may be special enough to be unique in the galaxy, even close to unique in the universe, and those conditions may all be essential. Reconstructing them elsewhere would be hard.

His scientific habit is to distrust both extremes. Reality almost never falls at the poles. "We're probably somewhere in between. So not unique, unique, but again, the chances are reasonably small." Fridman offers the other extreme: life so common that we are ants convinced our colony is the universe, unable to recognize forms that do not look like us. Korkin hopes we find something anyway — possible signs in the Venusian atmosphere, extinct life on Mars, oceans under Europa's ice. Even a single cell, if it differed in some fundamental way, would be a turning point. It would be a strong statement that life is everywhere.

The personal story starts in 2000, when Korkin was a Ph.D. student in Canada building a language for symbolic machine learning and aiming it at cheminformatics and computer-aided drug design. He had been trained as an applied mathematician. Chemistry was new land, so he lived on CCL, the Computational Chemistry List, a bulletin board from before Facebook, asking questions that included a lot of dumb ones. One night he posted something like a plea for a chemist's advice, explained the formalism and the planned applications, and went to bed.

In the morning his adviser called: you will not believe who replied. Joshua Lederberg. Korkin's first reaction was, who? Lederberg won the 1958 Nobel Prize in Physiology or Medicine at 33 for showing that bacteria can mate and exchange genes. He helped NASA think about life on Mars. In the 1960s he led DENDRAL, an expert system that took mass-spec fragments and a library of possible molecules and tried to reconstruct the whole, ending as a candidate list for a chemist to judge. The original ambition was to automate the entire loop. Lederberg's note said the ideas were conceptually similar; go look it up. Korkin later took a postdoc at Rockefeller partly for a chance to meet him, and did, briefly, on a little bridge between the research campus and the towers. Lederberg died in 2008.

Why did expert systems fade, especially in a field so full of human expertise? Korkin teaches the history of AI in his first two lectures and offers the two standard lines. One: they did not meet expectations and were replaced. The other, which he prefers: they were too good, became household, and transformed until the name disappeared and the thing remained. Modern machine learning still echoes them. Successful algorithms, including AlphaFold, work by building in knowledge of the domain you study. Encoding expert experience in the method is the same idea with a new label.

DeepMind had just announced that AlphaFold 2 had, in quotes, solved protein folding. Korkin calls it a very exciting achievement and then puts it on the right scoreboard. Bioinformatics has its Olympic Games. The original discipline is CASP, the Critical Assessment of protein Structure Prediction: experimentalists solve structures, withhold them from the Protein Data Bank, release only sequences, and the community predicts coordinates against the hidden answers. Other events cover function, mutational effects, and interactions. Historically the leaders mixed advanced biophysics, bioinformatics, and data mining, not machine learning. Homology, or comparative modeling, is the easier case, when a relative's structure already exists. De novo prediction, with no close relative, is the hard one.

AlphaFold's first appearance sent a shock wave through the community by being much better than everyone else. The machine-learning core predicted a contact map, which parts of the chain sit near which, a distance matrix from which reconstruction is more straightforward. AlphaFold 2 added another wave, including evolutionary information that used to live in feature engineering: align the same protein across hundreds of thousands of sequenced organisms, see what is conserved and therefore structurally important, and fold that signal into learning. At the time of the conversation there was still no biology paper, only a blog post from a marketing team. Reading it, Korkin says, is a bit like interpreting scripture.

Predicting the endpoint is not the same as knowing the mechanism. Folding is extremely fast, almost always reaches the same result, and begins while the chain is still coming off the ribosome. One hypothesis is modular: domains fold independently, then assemble, with secondary-structure elements — alpha helices, beta sheets — forming inside. Some of those elements need a fragment from the start of the chain and a fragment from the middle, so you cannot simply zip from one end. Nobody really knows. The practical gift of a good predictor is as a probe: change a residue, watch the predicted structure hold or collapse, and decompose which parts of the process are critical.

The word "solved" has to go back to the beads. CASP has mostly focused on compact single-domain, maybe two-domain, proteins, which are already extremely hard. An average protein has two or three domains. Proteins that run signaling in the nervous system often have five, six, seven, or more, and science is far from understanding how those fold. Solved, here, means near-experimental accuracy on the particular set of compact proteins in the contest. It is a huge leap over conventional biophysics. It is not the necklace.

Multi-domain folding is already mixed up with protein-protein interaction. Korkin's favorite example from his postdoc is PSD-95, a scaffold in the postsynaptic density with five domains and some 600 amino acids, flexible on purpose so it can bring other proteins into an orchestra. CAPRI is the separate contest for those macromolecular interactions. His lab entered about ten years earlier with one of the first machine-learning scoring functions in that challenge, learning which components make a predicted interface look like an experimental one. He thinks that class of problem is learnable. On a Nobel Prize for AlphaFold, he notes that computational biophysicists already won chemistry in 2013 for protein-folding simulation. Cross-disciplinary work will keep getting acknowledged. The program itself cannot win; the designers might.

Protein design is not new. David Baker and the Rosetta algorithm were pioneers of de novo design: a new sequence, a predicted fold, a chosen function. Using the same tools to engineer viruses is the darker question. Korkin's lab was building a model to tell whether a coronavirus strain is pathogenic, training across SARS-CoV-2, SARS, MERS, and the common-cold coronaviruses, looking for molecular determinants in sequence. Eugene Koonin had tried something similar with standard supervised learning. The hope is transfer learning from protein sequence more generally. Forecast which mutations make a strain more dangerous and you can prepare vaccines and antivirals. The same system, he agrees, could be asked how to make a virus worse.

Even a designed sequence is not a working particle. Experimental biology that checks every component is a different process. Around 2015, gain-of-function work on influenza jumps sat under a moratorium of half a year or more while scientists argued safety. Fridman is less worried about carefully regulated academic work than about fifty or a hundred years of machine-learned weapons, and he asks who is the better virus designer, nature or humans. Korkin still worries more about the natural pandemics. Nature's production capacity is enormous. A virus is also a strange weapon: rockets have a start and an endpoint; a virus hits the human species. He is "forever optimistic that we will not" — that there is insufficient evil to take that step — and he hopes connection and transparency keep making secrecy less useful.

Fridman asks what has changed since they last spoke in March or April. Korkin had been living in the virus computationally, and the thing that amazed him was the efficiency of the scientific community. Judge it even on the narrow domain of protein structure. After SARS in 2003, structural characterization of the virus took several years. In 2020 the same class of work was a matter of months. Sequencing is at an unprecedented level. Never before had a single virus been sequenced so many times, which lets researchers trace evolutionary trajectories with geographic and even personal resolution, because the host genotype also shapes how the virus changes.

Host-pathogen coevolution is the deepest lesson he takes from the pandemic. Contact tracing is obviously central and still incomplete. His group had just joined an XPRIZE-style contest to predict infection counts by region; the bottleneck was data, and the data were still thin. Fridman cites Michael Mina's dream of a weather map for viruses: high-resolution sensors of how every kind of virus moves from person to person, day to day. That, both of them agree, is what it would take to watch evolution in something like real time and to be ready for the next pandemic.

Asked for books, Korkin offers three plus an audiobook. Mikhail Bulgakov's *The Master and Margarita* hit him earliest: parallel lives centuries apart, braided into one story, romance powered by magic, Soviet irony, humor, and pain. The ending still gives him goosebumps. For anyone trying to understand Russian culture, he says, this is the book. Aleksandr Solzhenitsyn's *Cancer Ward*, first read in graduate school, follows a man sentenced for life, sent to a clinic, surrounded by people who die. Doctors later called the patient's inner life uncannily accurate; some said every doctor should read it. Under the tumor that grows and recedes sits an allegory of the Soviet system, written by a man who had been part of the regime, a Gulag prisoner, and a cancer patient at once.

The third is a thin English essay, John von Neumann's *The Computer and the Brain*, written while he knew he was dying of cancer and published in 1958 after his death. Every paragraph is packed. In 1957 he was already trying to draw parallels between computers and the nervous system. The honorable mention is Hope Jahren's *Lab Girl*, narrated by the author. In some passages you can hear her cry. Korkin had never felt that kind of connection while reading, and he calls it a must-listen for anyone who wants to know what academic science actually feels like.

For 2021 he misses travel, especially a European summer school in theoretical molecular biology for gifted students from around the world, the highlight of his year, which 2020 had forced online. The more private resolution came from working at home. Research, teaching, and academic noise had been a distraction so complete he did not notice he was away from his family. This time he noticed. The resolution is to spend as much time as possible with his kids even after the world opens. He ends by reciting, from memory, a winter poem by another Dmitry, a namesake, that he associates with Margarita. Something far away, he says, that is still very close.