Korkin, Dmitry

professor of bioinformatics and computational biology, WPI

Dmitry Korkin: Evolution of Proteins, Viruses, Life, and AI

Lex Fridman Podcast · Jan 11, 2021

• Dmitry Korkin
• protein domains and modularity
• SARS-CoV-2 structural proteins
• alternative splicing
• Joshua Lederberg and DENDRAL
• AlphaFold and CASP
• protein folding mechanism
• Drake equation and origin of life
• dual-use virus research
• pathogen surveillance
• AI milestones ranking

Core ideas

Korkin argues that proteins are better understood not as single globular molecules but as modular assemblies of "protein domains" — stable structural and functional units strung together like beads, connected by flexible linkers — a picture that only emerged once structural biology moved beyond small, single-domain proteins (resolvable by X-ray crystallography and NMR) toward larger complexes newly accessible through cryo-electron microscopy. He illustrates this with the SARS-CoV-2 spike protein, a homotrimer whose receptor-binding domains open asynchronously to attach to the human ACE2 receptor, and notes that a documented mutation increases the frequency of two (rather than one) "arms" opening simultaneously, potentially altering infection dynamics. He surveys the coronavirus's four structural proteins — spike (trimer, receptor attachment), envelope (pentamer), membrane or "M" protein (dimers that self-assemble into a lattice making up most of the viral shell, which Korkin's lab has been studying via simulation), and nucleocapsid (which protects the viral RNA) — and argues that because the M protein is evolutionarily more stable (less prone to mutation) than the more variable spike protein, it may represent a more promising, durable drug target, since disrupting the viral outer shell could disable the whole particle. He is measured about mutation risk given vaccine rollout, noting no strong evidence yet of vaccine-resistance-driving mutation, while affirming that understanding viral evolutionary mechanisms — including cross-species jumps (citing human-to-dog transmission) — is critical for forecasting future viral behavior.

Zooming out to protein evolution generally, Korkin identifies "modularity" as the defining theme: protein domains function as both structural and evolutionary units that rarely split once established, while the flexible linkers between domains and even more so alternative splicing (where a single gene produces multiple distinct functional protein products by including or excluding different exons) add further, only partially understood layers of combinatorial complexity — often, but not always, with exon boundaries corresponding to domain boundaries. This connects to a discussion of self-replicating computer programs ("quines") as a toy parallel to biological self-replication, and Korkin's openness to the idea of AI agents that evolve, mutate, and reproduce with one another (citing nature-inspired algorithms like slime-mold-based optimization as an underexplored source of ideas beyond neural networks) as both fascinating and unsettling.

Discussing the origin of life in the context of the Drake equation, Korkin cites the 2018 discovery of glycine (a basic amino acid) in comet dust from Comet 67P/Churyumov-Gerasimenko as suggestive evidence for the cosmic availability of life's building blocks, while personally estimating the probability of life emerging on a habitable planet as lower than Lex's proposed 1%, invoking the "Rare Earth hypothesis" (that Earth's specific conditions — asteroid shielding, solar distance, water-land balance — may be unusually, perhaps uniquely, favorable) as a counterweight to naive assumptions of cosmic mediocrity, while ultimately favoring a middle position between "we are unique" and "life is everywhere." He recounts a formative personal anecdote: as a PhD student developing symbolic machine learning methods for chemistry, he posted a question to an online forum and received a direct reply from Joshua Lederberg — Nobel laureate and pioneer of the 1960s DENDRAL expert system, which used mass spectrometry data and chemical databases to automatically infer molecular structures of unknown (including potentially extraterrestrial) compounds — leading Korkin to later join a postdoc position partly hoping to meet Lederberg in person. He connects DENDRAL to a broader discussion of why symbolic expert systems declined (either failing to meet expectations or, alternately, succeeding so thoroughly they were absorbed into other paradigms), and traces an intellectual lineage from expert systems' domain-knowledge-encoding philosophy to modern approaches like AlphaFold.

On AlphaFold and protein structure prediction, Korkin explains the CASP competition's biannual, blind benchmarking format (researchers solve structures experimentally but withhold them, releasing only sequences for computational prediction), and situates the "protein folding problem" as fundamentally about understanding the still-mysterious mechanism by which a linear amino acid sequence reliably and rapidly folds into a stable 3D structure — noting folding begins co-translationally, before a protein is even fully synthesized. He describes AlphaFold's key technical innovation as predicting a "contact map" (spatial proximity between residue pairs) from which 3D structure can then be reconstructed, and situates AlphaFold2's apparent improvement as likely incorporating evolutionary/homology information (conserved structural patterns across species' sequence alignments) more deeply into the learning process itself, though he cautions that as of the conversation no peer-reviewed paper existed, only promotional material of uncertain fidelity to the actual method. He pushes back on unqualified claims that protein folding is "solved," noting CASP has focused mainly on compact, one-to-two-domain proteins, while a large share of real proteins (especially neurologically important ones) have five or more domains and multi-domain folding remains poorly understood — illustrating this with his favorite protein, PSD-95, a five-domain neurological scaffolding protein whose functional flexibility depends on some domains being stable and others intentionally disordered.

Asked to rank AI milestones, Korkin places Deep Blue's 1997 defeat of Garry Kasparov at or near the top for its shocking cultural impact (despite being search-based rather than learning-based, which he argues undersells search's continued centrality to intelligence broadly), alongside AlphaFold for scientific impact, and highlights an unusual favorite: the "Deep Rembrandt" project, in which a trained model generated an entirely new portrait convincingly mimicking Rembrandt's style, physically reproduced with period-appropriate materials — while noting no AI system has yet produced chart-topping music or a commercially dominant film, suggesting art and entertainment remain areas where human-generated work still commands clear preference. On dual-use risk, Korkin argues his lab's own work — training models to predict whether a coronavirus strain is pathogenic based on sequence data — could in principle be inverted to help design more dangerous strains, but voices more concern about natural pandemics than engineered bioweapons, reasoning that viruses are extremely difficult to target precisely (any engineered pathogen risks becoming an indiscriminate, self-defeating weapon against humanity as a whole) and expressing broader optimism that increasing global information transparency will make catastrophic secret military applications of biology less likely over time.

Reflecting on the scientific community's response to COVID-19 compared to SARS two decades earlier, Korkin is struck by how much faster structural characterization proceeded (months rather than years) thanks to advances in sequencing and imaging technology, crediting unprecedented viral genomic sequencing volume with enabling fine-grained evolutionary tracing tied to specific populations and geography, while expressing hope for future investment in dense, real-time "pathogen surveillance" infrastructure (echoing Michael Mina's proposed "weather map" of circulating viruses) to better anticipate future pandemics. He closes with personal reflections: three formative books — Bulgakov's The Master and Margarita, Solzhenitsyn's Cancer Ward (informed by Solzhenitsyn's own cancer diagnosis and gulag imprisonment), and John von Neumann's posthumously published essay The Computer and the Brain, written as von Neumann was dying — plus an "honorable mention" audiobook, Hope Jahren's Lab Girl; a 2020 realization that pandemic-driven remote work revealed how much family time he had previously sacrificed to academic distraction; and a recitation, in Russian, of a contemporary poem by his namesake Dmitry Kimelfeld about longing and love in winter.

Excerpts