A Hundred-Nanometer Machine That Is Not Quite Alive

Lex Fridman introduces Dmitry Korkin as a professor of bioinformatics and computational biology at Worcester Polytechnic Institute, working on complex diseases, computational genomics, systems biology, and biomedical data analytics. In February 2020, Korkin's group used the SARS-CoV-2 genome to reconstruct the three-dimensional structures of its major viral proteins and their contacts with human proteins, then released that structural genomics map for anyone to use. The conversation was recorded in the first spring of the pandemic, with the medical, psychological, and financial shock still unfolding.

Fridman asks whether viruses are terrifying or fascinating. Korkin pictures them as villains who do their work so well that fascination is almost unavoidable. He does not picture only a pandemic, or only a 100-nanometer particle. The useful idea, he says, is that a virus lets you move from the microscale to the macroscale. Isolated, it is not a living organism. "It's a machine," he says, limited to the functions it must perform, carrying just enough information to do them, plus the ability to modify itself.

Fridman worries that calling it a machine undersells it, especially once Korkin also calls it intelligent. Korkin draws a line: this is not human intelligence, and not the intelligence of an AI system. "The intelligence of a virus is in its simplicity," he says, the ability to do so much with so little material and information. He also wonders whether viruses are a basic case of swarm intelligence, inefficient as single particles and startlingly efficient when they act as a whole.

Fridman asks, looking a hundred years ahead, whether natural pandemics or engineered ones should worry us more, and how hard it is to build a virus. Korkin is more concerned with strains that keep emerging on their own. New influenza lineages appear, some of them pandemic; new coronaviruses appear too. That process is natural, he says, and that is why it is powerful. He has read papers on what modern biotechnology can do to a virus, and he hopes engineered pathogens will not be the main threat soon.

History's warning, for him, is smallpox: a double-stranded DNA virus, far more contagious than seasonal flu, SARS, or the new coronavirus, with remaining strains locked in a handful of controlled labs and a vaccine that made eradication possible. He walks through R0, the average number of new infections from one case. Estimates for the then-current coronavirus sat between 1.5 and 3; smallpox was 5 to 7; measles starts around 15. Those gaps matter because the growth is exponential.

What makes a virus more contagious is not one trick. Korkin lists biology and society together: how it spreads, how long it survives on surfaces, how fast it replicates once it is inside a host. The factor he thinks the field underweighted is the incubation window in which hosts are already infectious, and the share of infections that never produce symptoms. Those people still carry virus and can still pass it on.

An Iceland screening study, impressive to Korkin for its scale, put the asymptomatic share near 50 percent. Fridman adds swine flu: infection counts in the billions, on the order of 20 to 30 percent of the population, with a low fatality rate that still does not make a full-population takeover less frightening. Korkin calls it his favorite butterfly effect, except the butterfly is smaller. The virus used to sit in bat species; a couple of small genomic changes let it jump to humans, then from human to human. The change is the size of a few atoms, and the effect is a stopped world.

Scientists were still hunting patient zero and the exact mutations that made the jump. A few years earlier, influenza groups had shown experimentally that changing just a couple of residues could let a virus move between species. The work drew a moratorium, then was published. Fridman asks the obvious dual-use question: if you can search bats for the deadliest relatives and then test the jump, have you written a recipe? Korkin thinks the mechanism still has to be studied so it can be prevented. He also grants the nuclear-weapons analogy. You cannot control a virus once it works as a weapon, which is why he wants regulation and why he is "hopeful" rather than relaxed.

The payoff he wants from that knowledge is better vaccines, ideally universal ones. Seasonal flu shots exist because vaccines are in an arms race with a changing virus; the next influenza pandemic, he says, will likely not be stopped by the shot already in pharmacies. A vaccine that worked against influenza A no matter the strain, and no matter which species it jumped from, would be a huge scientific advance. Until then, he is still watching influenza, including the H7N9 avian strain that had emerged in China with a mortality rate he recalled as above 30 percent. Luckily it jumped from birds to people without becoming a sustained human-to-human pandemic.

That H7N9 example opens a question Korkin says scientists still cannot close: the balance between a virus that spreads efficiently and a virus that is highly pathogenic. Looking across outbreaks, the more pathogenic the virus, the less contagious it appears. He does not claim to know whether that is a law of biology. He does note that MERS, a deadlier coronavirus relative, never became a pandemic, even though its mortality sat far above 20 or 30 percent.

Fridman pushes the darker joke: whatever is running this show does not want us dead, or at least not yet, given how good viruses are at what they do. Korkin answers with the other side of the ledger. Hosts have immune systems. Vaccines exist against viruses that might have wiped populations out two centuries ago, and smallpox shows that eradication is sometimes possible. The virus is not negotiating. Hosts fight back, and recently they have better tools.

The infection story starts with attachment. For this coronavirus, evidence already pointed to the same door SARS used: the spike protein on the virion binding the human ACE2 receptor, apparently more efficiently than SARS did. Spike does not work as a single copy. Three of them assemble into a trimer, the functional unit that docks ACE2, and that trimer is what puts the crown on the particle. Once anchored, the viral membrane fuses with the host membrane, RNA is released, and the cell's machinery is hijacked. Viruses as we know them have no ribosomes, Korkin notes. They cannot print proteins. They borrow the host's printer.

The envelope is built from three main surface proteins. Spike is the famous one. Membrane protein is the most abundant, shaping the envelope and its curvature. Envelope protein is scarce, and researchers were still arguing about what it actually does. Inside sits the nucleocapsid protein, wrapping and protecting the viral RNA. Influenza, Korkin says, encodes around eight or nine proteins. This coronavirus encodes at least 29. Coronaviruses have been studied for more than two decades, and the functions of those proteins are still not fully mapped.

This is where computational biology earns its keep. Decades of viral protein knowledge let researchers take a new genome, call potential proteins, and ask whether they resemble anything already understood. Polymerases and proteases often jump out. Truly novel proteins do not, and methods that work from scratch remain weak, including structure prediction without templates. Bioinformatics, Korkin stresses, produces predictions. Experiments have to validate them. The practical pipeline is a collaboration among virologists, molecular imaging specialists using electron microscopy, biochemists, pharmaceutical scientists, and computational groups. He sees his field as speeding discovery, not replacing the bench.

A protein is a chain of amino-acid residues that folds into a strange three-dimensional shape. Figuring out how that happens is one of the million-dollar challenges: heat can unfold a chain, and it often folds back into the same state, which means a lot is known about mechanisms and still too expensive to put together computationally. The usual picture is one canonical fold, though destabilized proteins, especially those built from more than one domain, can take other conformations. A domain is an evolutionarily preserved unit with its own function and fold. An average human protein, Korkin says, has two or three of them, each folding first and then packing into a larger structure.

The field is decent, he says, on proteins up to about 200 residues and still far from large ones. Transmembrane proteins are notoriously hard and also among the most important. The search space is combinatorial, but brute force is not the only attack. Machine learning has entered the problem. When Fridman asks about Folding@home, Korkin describes a protein-folding game he first saw as a postdoc, transparent enough that children get good at it. Players later appeared as co-authors on a top-journal paper, producing structures better than the scientists' own. That existence proof still matters: a problem that crushes computers is not always as hard for a human brain.

Asked whether folding is now a learned game in the AlphaGo sense or still Deep Blue-style search, Korkin points to CASP, the community's Olympics, where teams get the same sequences and compete on structure. In the most recent contest he had in mind, AlphaFold was among the top teams, if not the top team. Decades of structural data have reached the point where they can be summarized, generalized, and used to predict new folds. The prerequisite is the Protein Data Bank. Korkin says the earliest version lived on flash cards; every solved protein or complex is deposited, and there is no culture of hoarding. The pandemic, he adds, pushed collaboration to a new level of speed and willingness to help.

Journals were trying accelerated review, but the flood of preprints was changing what a paper even is. Korkin agrees that knowledge itself is becoming the core value, not the journal that wraps it. His group's first SARS-CoV-2 preprint never waited out bioRxiv's five- or six-day screening queue, which was overwhelmed; they posted a rougher version on their own site and it started getting read immediately. They tried to keep the write-up compact enough to explain the results without the usual ceremonial bulk.

The paper, *Structural Genomics of SARS-CoV-2 Reveals Evolutionary Conserved Functional Regions of Viral Proteins*, has a backstory that starts in September 2019. New England was dealing with Eastern equine encephalitis, Triple E, a mosquito-borne virus that attacks the brain, caused deaths, closed schools after 6 p.m., and kept children indoors because pediatric outcomes were especially bad. It hit the local economy from MetroWest into New Hampshire and Connecticut, and it was still a neglected disease: too small for companies to build a dedicated antiviral, too small for easy national grants. Korkin turned it into a problem-based bioinformatics course at WPI. Undergraduates modeled structures, interactions, and evolution. The projects wrapped in mid-December. A couple of weeks later he heard about a mysterious virus in one Chinese province and thought: we just did this. Can we do it again?

They waited for the genome, a string of fewer than 30,000 nucleotides encoding those 29 genes. With SARS, MERS, and bat coronavirus as templates, finding gene boundaries was comparatively straightforward. Homology, or template-based, modeling then turns similar letter sequences into similar structures, locally and across a whole protein, using sophisticated tools and the PDB. After single-protein models come complexes: spike needs three copies, envelope protein five, and several other SARS-CoV-2 proteins oligomerize. Interaction, Korkin says, is structural, but the structure is there to carry out a function. On a figure of cyan protein blobs, pink patches mark differences from SARS and bat relatives. In one dimension those mutations can look scattered; in three dimensions they often cluster, sometimes right on a functional surface. Evolution, in this genome, is not uniform. Some proteins are almost intact relative to SARS. Others are heavily mutated.

One of the most useful findings, Korkin says, is that binding sites on viral surfaces targeted by known small molecules were pretty much unaffected. The same drug-like compounds that worked conceptually against earlier coronaviruses might therefore work against this one. The map also flags what has changed, which is equally useful for new chemistry. All of it still needs wet-lab tests. The next modeling target is the whole virion: a lipid-bilayer envelope with proteins stuck in it, something like 50 to 100 spikes on an average particle, 200 to 400 membrane-protein dimers in a lattice that looks like a carpet, and occasional envelope-protein pentamer rings. Internet renderings are perfect spheres bristling with spikes. Real micrographs are messier. Collaborator Benjamin Newman at Texas A&M had just argued, with some evidence, that the particle is an elongated ellipsoid, and Korkin's group was folding that shape into the model.

A faithful virion model is not only descriptive. It opens nanoparticle designs that mimic the particle and might even compete with real virions. Vaccine logic is more established: train the immune system so antibodies bind spike and block ACE2. Classic routes include passaging a virus through cell culture until it prefers embryonic cells and replicates poorly in people, or stripping genetic material so the particle cannot copy itself. Asked whether the bottleneck is molecular understanding or experimental validation and approval, Korkin says both. Deeper mechanism makes design more efficient; a designed vaccine still has to be tested. Eighteen-month projections already look exceptionally fast. He remembers reading that some earlier vaccines took up to ten years to design, test properly, and reach mass production.

Vaccines prevent infection. Antivirals try to stop a virus that is already inside, by blocking a critical protein function. Korkin names remdesivir as perhaps the most promising candidate then: developed for Ebola and Marburg, later shown effective against SARS, it mimics an RNA nucleotide and stalls replication. Other candidates inhibit proteases that clip protein sequences; one began as a malaria drug. The computational question is brutally specific. A compound that is perfect against SARS or MERS may fail against SARS-CoV-2 if the ligand-binding site mutated. For the ligands his group inspected, those sites were largely intact.

Fridman asks which short-term path is real: antivirals, the pandemic burning out under distancing and masks, or something worse. Korkin picks antivirals as the short-term solution. He was seeing more hospital use of new candidates, and he thinks reopening an economy that is already deep into a worldwide outbreak will need them. He will not forecast the reopening itself. A second wave is possible, which is why precautions still matter. Whether the virus can mutate around a new antiviral is an open question. It already mutated enough to jump species and become human-transmissible.

"This is such a beautiful and terrifying process," he says: some viruses mutate around whatever is put in front of them. Nothing is imprinted into the particle. It simply evolves with its host. That is one of the cleanest demonstrations of evolution, and the simplicity is not a miracle. The virus does not have to store every necessary function; it hijacks most of them from the host, which slashes the complexity of the machine. Recent discoveries complicate the origin story: viruses as large as bacteria, mimiviruses and mamaviruses, have pushed scientists to rethink how viruses appeared.

Fridman returns to the living-or-not question. Korkin repeats that viruses are not living organisms because most of their functions depend on the host. Fridman plays devil's advocate: humans need a host planet too, so dependence cannot be the whole definition of life. Korkin admits the question is hard if you press it that way. His own cut is autonomy. A virus lacks critical tools, including the printer for its proteins, so it is not autonomous, and that is how he separates a living organism from a machine at a high level.

Korkin also runs epidemiological simulations, specifically agent-based models of individual people and the spread between them. The timing was grimly neat. The group had worked on the project for five years; on New Year's Eve a student emailed that the last experiments were done; three weeks later Diamond Princess happened. The undergraduate code name had been a zombie apocalypse on a cruise ship. The serious prompt was norovirus, the stomach flu that has long been one of the biggest threats to the cruise economy, sometimes forcing ships to cancel and turn around. Unlike classical epidemiology at national scale, they wanted a confined setting: a ship, a school, a large company, anywhere people interact densely, with the whole course of host-pathogen contact visible in something like real time.

The novelty was to make the pathogen its own agent, not only the passengers. Ordinary agent-based epidemiology treats each person as a program on a contact graph. Adding an explicit pathogen agent let them encode how the virus transmits, how long it survives on surfaces, and how much a host sheds while asymptomatic versus symptomatic. They started with norovirus and zombies, then modeled Ebola, flu, and SARS. For the undergraduates they also modeled the virus from the film *Contagion*. Scientific consultant Ian Lipkin of Columbia had based that fictional agent on Nipah, with airborne ideas from SARS and flu, and the movie contained enough detail to extract a model. One classroom joke: a week after detection, the film shows scientists looking at a surface-protein structure. Experimental biologists would call that impossible, months or years of crystallography. A bioinformatician would say, sure, homology-model it.

Fridman asks whether large-scale simulations predict the future or only compare interventions. Korkin says both. For the new coronavirus, existing interventions may not be efficient enough. The overlooked piece from past pandemics is the asymptomatic period. This virus's long silent window, he says, creates a completely new containment game. Shedding versus time is the parameter that decides how contagious a person is before symptoms. Work already suggested people are infectious in that window and shed enough to infect another host. Fridman mentions a paper, possibly in *Nature*, in which the first week of a 14-day observation window looked like peak contagiousness; Korkin wants larger samples. A separate recent result he liked: after recovery, there is a period with no detectable shedding, meaning not contagious.

Fridman has been reading the messy literature on masks and on whether the virus is airborne over long distances. Korkin tracks that debate because it feeds the epidemiology models, and his view is simple: masks work both ways. They cut incoming particles, and they stop a potentially contagious person from spreading them, including people who do not know they are infected. Fridman notes the evidence that surgical and homemade masks, which is what a shortage leaves you, protect others better than they protect the wearer. Korkin agrees. At least 30 percent of asymptomatic cases, as he remembers the figure, are fully asymptomatic: no cough, no obvious signs, still shedding.

Fridman had worn a mask in a grocery store a week earlier and gotten looks. Korkin hopes the country will not struggle with it; CDC guidance was starting to shift. He had been looking at Spanish flu photographs in which almost everyone is masked, including an iconic San Francisco streetcar image of a conductor refusing a rider without one. How scared we are, and how seriously we treat the problem, is the real variable. His take is that we should treat it as very serious. Fridman worries about the psychology: masks erase facial expression and create distance, yet they also signal that the wearer cares about other people's health. Korkin notes they are already physically distant. When the economy reopens and people come closer, the long road will be rebuilding trust, not turning everyone into a permanent germophobe.

Korkin has a thick Russian accent he does not pretend otherwise. He moved first to Canada and then to the United States in 1999, at 22, and the accent from that year stuck. The Soviet Union collapsed while he was a child old enough to feel the change. For about ten years he wanted to be a passenger-jet pilot; he was also hooked on math, biology, chemistry, and physics. Cybernetics, closer to computational robotics than to computer science, pulled him, then mathematics did, and at Moscow State University he realized he wanted that math mixed with real problems. Fridman asks whether open, collaborative science is hard in Russia. Korkin says he has been lucky: strong bioinformatics in Moscow, Novosibirsk, St. Petersburg, and Kazan, people with ideas who are open to collaboration. He cannot recall a major AI conference in Russia, and as a U.S. citizen he finds going back bureaucratically messy. He hopes the pandemic at least normalizes remote work for computational scientists, even while experimental colleagues are stuck.

In his office, spaced out for social distancing, sits a collection of scientist bobbleheads he has built for 12 or 13 years. It started with Watson and Crick. The newest, and his favorite, is Rosalind Franklin. Newton, Einstein, Marie Curie, Tesla, and Darwin are already on the shelf. He wants Kolmogorov, both for the science and for the school Kolmogorov founded in Moscow, the high school Korkin attended that now bears the name. He also wants Turing and von Neumann, and he is still amazed no one has made an Alan Turing bobblehead. Linus Pauling is on the wish list too: the chemist who discovered protein secondary structure and came close to DNA. Some people, Korkin notes, think that without Photograph 51, the X-ray image of Franklin's that Watson and Crick saw, Pauling might have solved it. Science, he says, is a funny race.

Fridman asks the large question: does studying the body's defenses and these enemies change how Korkin sees a human life? It makes him realize how fragile that life is. A tiny thing can affect the whole of humankind, which is something to appreciate and to remember. We are fragile, he says, and we have to bond together as a society. It also gives him hope that what scientists do is useful. Fridman closes with E.O. Wilson: the variety of genes in viruses is likely to exceed that in all the rest of life combined. A particle that may not even count as alive still outguns the living world in genetic diversity. It does not need to be complicated. It needs to be enough.