The obesity gene was never the obesity gene

Manolis Kellis, who leads the MIT Computational Biology Group, opens his third conversation with Lex Fridman by refusing a small answer. Every human disease, he says, is among the hardest problems in modern science, because disease is as complex as the genome, the brain, the immune system, and cancer — and each of those maps keeps getting more intricate as the others come into focus. For a century the standard workflow ran the other way: knock out a gene in yeast, fly, or mouse, learn the function, then climb toward mammals and finally to people. Once a human locus was mapped, you looked up the model-organism annotation and declared the problem understood.

That pipeline has inverted. Human genetics, transformed over the last decade or two, now drives basic biology. Scientists learn systems by perturbation; engineers already know how they built the machine. Natural selection has been running the experiment for hundreds of thousands of years. Each person carries about six million polymorphic sites inherited from both parents — six million tiny perturbations. Most do nothing. The ones that do, Kellis argues, explain why two twins still look alike at ninety-three, and why the rest of us do not.

Whether Earth's population is large enough is a power question: how far apart the means of allele A and allele B sit, and how tight the variance is around each. The sample is already enough for many strong effects with narrow distributions — which is only partly hopeful, because those same strong perturbations are often the basis of disease. What genetics really offers is an allelic series: many hits on the same gene, at different frequencies and effect sizes, so a pathway can be read even when a given patient does not carry the original mutation.

The old puzzle was pedigrees and Mendelian genes traced through families. The new one is a giant genotype-by-phenotype matrix: thousands or hundreds of thousands of unrelated people, millions of variants, and, increasingly, hundreds of traits from electronic health records. Computer scientists want to convolve the whole thing. Biologists want to peel the layers experimentally — one cell type, then one cell. Kellis's bet is to do both: single-cell profiles across dozens of brain regions, hundreds of donors, and millions of cells, then put the pieces back into a circuit.

Fridman asks which diseases are most important to understand. Kellis splits the question. One ruler is quality of life: Alzheimer's may take years to kill, but it hollows out patients and families; COVID-19 in 2020 mattered in the same register, stripping jobs, housing, and the social circle that keeps people intact. The other ruler is deaths, which is easier to quantify. In the United States he lists heart disease at about 650,000 a year, cancer at 600,000, accidents at 167,000, lower respiratory disease at 160,000, Alzheimer's at 120,000, then stroke, diabetes and metabolic disease at 85,000, flu at 60,000, and suicide at 50,000. COVID had already climbed to the number-three killer that year, past 100,000 and still rising.

The two rankings do not match, and neither is the whole intervention map. Car crashes make the news; suicide often does not. Societal levers — healthier cheap food priced for what it actually costs the health system, more empathy, education, and support — would cut suicides and homicides. Exercise and nutrition are still among the largest things a society can change. Every disease also mixes genetics and environment. Kellis puts Alzheimer's heritability at about 79 percent: genetics alone can explain most incidence. The remaining 21 percent is real — cognitive enrichment, social contact, books, a foreign language, running — but there is a ceiling those habits cannot lift.

Understanding the genetic component, he argues, is not a retreat from the environmental one. Genetics is access to mechanism. If you can move the mechanism, you can sometimes counteract what the environment is doing. That is the point of the work: not to declare people doomed by their genomes, but to find the valves.

Kellis's metaphor for obesity is a giant bathtub of fat. Public conversation stares at two openings: fat in from diet, fat out from exercise. That is the in-out equation everyone already knows. Any thermodynamic system, he reminds Fridman, has three terms, not two: energy in, energy out, and energy lost. The third term is heat. In biology it is thermogenesis, and it is regulated — it can be switched on or off. Metabolism is a fourth dimension on the same tub: it sets the rate at which energy is burned or stored, and it teaches you where the valves sit. Genetics is how you find those valves.

The same surplus calorie can be packed into white fat or burned as heat inside the fat cell itself. Those are two directions at one cellular fork, and which way a person goes is, to a large extent, genetic. Fat's public reputation, Kellis thinks, is undeserved. Fat sequesters excess lipids that would otherwise circulate and do damage. It can also burn those calories directly. When you are cold, brown fat is doing honest physiological work. The point is not that diet and exercise are fake. It is that they are two of the knobs, and the genome is pointing at the others.

A strong-effect mutation looks like an obvious drug target. Kellis wants the opposite reading. Every genetic association needs two things: the gene must affect the disease, and the population must still carry variation in it. Some genes are so important that a tiny change kills the person. Those genes have no standing variation, so they never appear in a genome-wide association study. That is not evidence they are irrelevant. It is evidence they tolerate no mutations. You can rank genes by how little variation they carry. Genes that are nearly invariant and still unassociated with any measurable adult trait are, he suspects, early embryonic lethals — phenotypes we are bad at measuring because the carriers never make it to the study.

Sperm, he notes, express thousands of proteins that have little to do with swimming. His reading is quality control: protein folding is an early failure signal. If this small thing is already wrong, something larger is likely wrong downstream. Mammals invest enormous resources in a newborn, so they evolved a fail-fast filter in the first months. At the personal level it is brutal. As an evolutionary resource strategy it is coherent. Humans with medicine can sometimes rescue carriers who would once have been lost; if a mutation does not survive the first trimester, we still never see it.

The intervention logic follows. What you observe is variant-on-gene times gene-on-disease. A huge GWAS hit may be a large shove on a gene that barely moves the disease, or a tiny shove on a gene that moves it a lot. The genes you want to drug are the second kind: causal for disease, but only weakly moved by the variants evolution left standing. Pharmaceutically you can push a gene 70 percent when nature only tolerated 2 percent. The dramatic Mendelian hits are often the ones you cannot shove that far without wrecking everything else.

Genetics keeps handing Kellis connections nobody had asked for. One of the worst things in his own genome, he says, is a predisposition to age-related macular degeneration, a progressive loss of central vision. His absolute risk is modest — about 8 percent against a 6 percent average — but the genes behind it were a shock. Until the associations were mapped, AMD was a mystery. The hits piled into the complement pathway, an immune module, which meant an eye disease had an immune basis that ophthalmology had not expected.

Complement showed up again in schizophrenia. Microglia, immune cells that patrol the brain, prune synapses on a use-it-or-lose-it rule: busy connections stay, idle ones go. If the timing of that pruning is off, schizophrenia risk rises. The same immune gene set now tied an eye disease, a psychiatric disease, and — soon — a neurodegenerative one. Kellis's immune system, he likes to say, has been co-opted for jobs all over the body.

The Roadmap Epigenomics project, which his group helped build, mapped active gene-regulatory elements across tissues: a first version with 27 tissues and samples and dozens of epigenetic marks, later EpiMap with 833 tissues and enrichments against 540 disorders. Overlay disease variants on those maps and most answers look like common sense. Immune traits such as allergy and type 1 diabetes enrich in T cells, B cells, and hematopoietic stem cells. Type 2 diabetes enriches in one sample: pancreatic islets. Blood pressure lands in left-ventricle enhancers, not in metabolism or stress circuits. Height enriches in embryonic stem-cell enhancers, which is where a developmental trait should act. Inflammatory bowel disease lights up both immune and digestive cells, matching the two words in its name.

Alzheimer's did not. They had seven different brain samples and found zero enrichment. "Our brains were literally hurting," Kellis says. Those samples are mostly neurons, oligodendrocytes, and astrocytes, so the genetics were apparently not acting in those three. The fourth major brain cell type is microglia, resident monocytes with surface markers such as CD14, cousins of the macrophages in blood and fat. Recalculate, and Alzheimer's is hugely enriched in microglia and not in the other brain cells. That is not a risk score. It is a place to start: screen molecules against the pathway in that cell type, then mice, then a trial. It also means immunologists have to walk into the Alzheimer's meeting. They used to work in another building.

Kellis's lab works on many disorders on purpose. MIT was early in folding zoology, bacteriology, and virology into a single biology department, because DNA, RNA, and protein — the central dogma — made life one subject. Genetics, he says, is doing the same thing to disease. Cardiovascular, neurodegeneration, immune, and cancer used to live in separate buildings because you had to understand the tissue first. Human genetics keeps revealing the wrong tissue, so the walls are crumbling. Immunologists belong in Alzheimer's. Neurologists belong in metabolic disease. His group goes where the genetics point.

A theory of everything for disease is not a romantic extra, in his view. It is inevitable. The genomic circuitry you need for schizophrenia is the same circuitry you need for Alzheimer's, metabolism, immune disease, and cancer. He teaches dynamic programming as the computer-science version of the same idea: do the shared subproblem once. Staffing a circuitry group in every disease building, each re-solving the wiring of the whole genome, is waste. One group should own the wiring of everything; the disease specialists should be its users — and the wiring group must be a user of its own maps. He calls that eating your own dog food. Microsoft Excel and Word worked, he says, because the people who built them used them every day. You cannot ship a circuit and walk away.

That is why a lab named Computational Biology is, in his telling, a biology lab. Computation is the method; the object is still biology, done genomically. Experiments this combinatorial, parallel, and noisy would never be designed unless the analysis pipeline was already in someone's head. Traditional one-gene biology is not dead as a follow-up. Once IRX3 and IRX5 are on the table, he wants every fact the single-gene literature has. You cannot use that literature as the way into polygenic disease.

The path from ignorance to a drug, Kellis says, is not one postdoc per gene. Diseases are too polygenic. The work now runs in parallel across thousands of genes, in stages. Step zero is the association itself: families for Mendelian effects, or the giant unrelated matrix for polygenic ones. What you get is not a gene. It is a genomic region. Most people miss that distinction.

They miss it because 93 percent of disease-associated variants do not affect a protein. The human genome has about 20,000 genes and 3.2 billion nucleotides; only 1.5 percent of it codes for protein. Genes are enriched, but only about threefold, which still leaves 93 percent of the hits outside coding sequence. A variant in an intron can be transcribed with a nearby gene, spliced out, and have nothing to do with it. The real target may sit a million nucleotides away. You cannot assign a locus to its nearest gene and call the circuitry done.

Step one is the tissue of action, via those epigenomic enrichments. The saving grace of the field is convergence: hundreds of Alzheimer's variants collapse onto a handful of processes — lipid metabolism, immune-cell function, neuronal energetics. Schizophrenia collapses onto excitatory neurons, inhibitory neurons, synaptic pruning, calcium signaling. Hits scattered across the genome turn out to be calcium-signaling proteins. A lab staring at one gene cannot see that picture. You have to look at the whole load.

Step two is the causal variant inside a linkage-disequilibrium block. Chromosomes recombine at hotspots guided by PRDM9, so variants are inherited in chunks — two dozen at a time, 89 common variants in the FTO obesity block alone. Step three is the motif the causal variant breaks, the regulatory "word." Step four is the factor that used to bind it. Step five is the target gene, from four independent clues: physical contacts (Hi-C-style folding), genetic links (eQTLs), activity links across 833 tissues, and causal perturbation with CRISPR. Step six is what that gene actually does.

In 2007 the strongest genetic association with obesity landed on FTO, the fat-mass and obesity-associated gene. Laboratories studied it for years. They showed that FTO encodes an RNA-modifying enzyme and built a story about epitranscriptomics and body weight. Kellis's group, with a team led by Melina Claussnitzer, turned the story over. The variants sit in an intron of FTO. They do not implicate FTO. They control IRX3 and IRX5, genes 600,000 and 1.2 million nucleotides away. At a conference a few years later, a colleague introduced him as the author of the most depressing paper of 2015. The pharmaceutical industry had been comfortable that one gene in the locus was the answer. The paper said the answer could be a random gene far, far away.

The 89 common variants in the block all associate with obesity. Global enrichment pointed at mesenchymal stem-cell enhancers — the progenitors that, in the first days after they commit, choose white fat that stores energy or brown and beige fat that burns it. About 42 percent of European chromosomes carry a predisposition to storing fat, likely selected on the way out of Africa and through ice ages, when extra stores helped people survive farther north. The predicted causal SNP was rs1421085, a T-to-C change. The T sits in an AT-rich motif bound by ARID5B, a repressor. ARID5B sits on a 12,000-nucleotide super-enhancer and keeps IRX3 and IRX5 off. Swap T for C, the motif vanishes, the repressor cannot bind, the enhancer wakes, and IRX3 and IRX5 rise.

Those two genes are a master switch between thermogenesis and lipid storage. When they go up, heat production shuts down and energy is stored. Across people, IRX3 and IRX5 correlate positively with lipid metabolism and negatively with mitochondrial biogenesis — mitochondria being the endosymbiotic organelles whose remaining 13 genes still run the electron-transport chain. Kellis himself is a homozygous carrier. He inherited the risk allele from both parents. In a food-scarce world that is an advantage: he can eat less and get by. In a world of abundant calories, he is the person who gets obese. The dissipation knob most people can turn, he cannot.

Once the circuit is known, you can intervene at ARID5B, at the SNP, at IRX3 and IRX5, at thermogenesis, or at the developmental choice between storing and burning machines. In the paper they did the extreme version. They took primary adipocytes from risk and non-risk donors and used CRISPR to edit that one nucleotide out of 3.2 billion. The obese expression pattern and the lean pattern flipped like a switch. Kellis likes to say the paper is about one bit of information, C versus T, and that 3.2 billion nucleotides are still waiting.

Finding long-range targets still needs wet-lab geometry. Kellis describes Hi-C as a bowl of noodles: freeze the folds with glue, chop, let nearby ends religate, then sequence the chimeric junctions. Red glued to blue means those loci were touching in 3D even if they were far apart on the chromosome, which is how topological domains show up. Genetic links ask whether people with a T at a SNP express IRX3 differently from people with a C. Activity links ask whether, across 833 tissues, the enhancer and the gene turn on together. Perturbation is the causal version of the same question.

CRISPR, he reminds Fridman, is a bacterial adaptive immune system. After a bacterium survives a virus it stores a trophy snippet in a CRISPR array; next time the same genome appears, an RNA-guided enzyme cuts it. The yogurt industry found the system while trying to make starter cultures more virus-proof. Jennifer Doudna, Emmanuelle Charpentier, and others co-opted it as a programmable cutter. Cutting and homologous repair already existed with zinc fingers and TALENs. Those required a team of protein engineers for every address. CRISPR's revolution is the guide: you type an RNA on a computer. Dead Cas9 keeps the targeting and drops the cut, so you can dock an activator or repressor on an enhancer without breaking the chromosome. Kellis's classroom joke is genome vandalism followed by a stack of bandages that happen to carry the sequence you want.

Doing this one locus at a time will not finish 120,000 disease SNPs. The first industrial answer is robotics: thousands of plates, thousands of molecules a well. The second is massively parallel reporter assays, developed with people including Tarjei Mikkelsen and Eric Lander and used by many groups: DNA microarrays synthesize on the order of 150-base fragments across about 450,000 spots, drop them into a common plasmid, and read 10,000 enhancer hypotheses — risk versus non-risk — in one pot. Sharpr-MPRA tiles for higher resolution. HyDRA skips synthesis, cuts open chromatin from the genome, and tests about seven million fragments. STARR-seq, from former postdoc Alex Stark, lets the enhancer be part of the transcript it activates, so you sequence the element and the output together.

The third technology is single-cell RNA-seq. First came 384-well plates. Then droplets: one channel of cells, one channel of barcoded beads, oil-water chambers, smash, sequence. The next generation dropped the chip. Dissociate a million cells, throw them at random into 100 bottles, add a barcode; dump, reshuffle, barcode again; once more. Three choices of 1 in 100 are a million unique paths. Each cell is labeled by the route it walked. Kellis calls it a beautiful computer-science idea. For about $2,000 a sample you can now get 10,000 cells and thousands of genes each; the per-cell data are noisy, and computer scientists aggregate them.

From a single dissociated brain sample the lab now splits tens of thousands of cells: half for RNA, half for ATAC-seq of open DNA, both at single-cell resolution. Expression and regulation can be paired, then correlated across 2.3 million enhancers and 20,000 genes in millions of cells. Eight hundred thirty-three bulk tissues were already a lot. Ten million cells are another scale. In one year Kellis's team generated that many from human brain: about 1,500 individuals, a dozen regions, spanning schizophrenia, Alzheimer's, frontotemporal dementia, Lewy body dementia, ALS, Huntington's, PTSD, autism, bipolar disorder, and healthy aging.

The maps are already rewriting slogans. Microglia used to mean "the brain's immune cells," one sentence. They now resolve into about ten subtypes, each with its own expression program, playing different roles in Alzheimer's versus schizophrenia. Astrocytes are not just astrocytes: expression differs sharply by region. Recently evolved neocortex and older, more reptilian structures take different paths into the same diseases. A schizophrenia paper in the works uses the same single-cell and editing tools to argue that four master regulators tie together excitatory neurons, inhibitory neurons, synaptic pruning, and neural development — two processes that used to be studied apart. A metabolism collaboration with the Goodyear lab is watching cell-to-cell signals across six or seven tissues under exercise and diet, with immune cells and adipocyte stem cells reshaping the circuit.

Alzheimer's itself has a geography. It starts in a small patch, the entorhinal cortex, spreads through the hippocampus, and eventually blankets neocortex. Each region that falls takes a cognitive function with it: orientation, then short-term memory, then long-term memory. The patient's decline is that anatomical march made visible. The group is ordering cells, regions, and people by how well they predict pathology, so every gene can be placed on a timeline. The practical stake is the lag: the disease is already moving for about ten years before the first cognitive loss. See that path early and you get diagnostics, prognostics, and biomarkers while there is still something to save.

There is a third tree of variation that post-mortem genetics used to miss. Common variants are shared across humanity; rare variants sit on your branch of the species; somatic variants arise after the zygote, so the body — especially the brain — is a mosaic. Well-based single-cell RNA-seq is expensive but transcript-rich enough to infer mutations in the thousands of genes a cell expresses. Genotype and phenotype then meet inside one cell. Common, rare, and somatic hits light up almost three different pathway sets. Somatic mutations point at neuronal energetics and oligodendrocyte function, which barely appear among common variants — probably, Kellis says, because those effects are too strong for evolution to tolerate them as standing polymorphisms. They can exist only as private, lifetime accidents.

Those energetic pathways connect to what is strange about our species. Relative to close relatives, each human brain cell is more energy-efficient, there are far more of them, and a new diet of energy-rich food feeds the factory. The bill is oxidative damage: a high-power plant of ideas, running for decades, whose debris-clearing cannot always keep up — one road into late-life neurodegeneration. The immune system is equally unusual. Dense societies, the Black Death and other die-offs, famines, and ice ages put selection on human immunity that no other species has lived through. Metabolism is unusual too. Humans can run marathons; in the endurance contest Kellis recalls, the horse tires first.

Lifespan is the fourth variable. Recent human longevity has no precedent among nearby species, and late disorders arrive on the back of it. His car analogy is blunt. Re-engineering a car from 70 miles an hour to 120 is a parts problem. Pushing it to 400 means redesigning the vehicle. The human body, he guesses, evolved to live to about 120; minor changes might reach 150. Push well-being into the eighties, nineties, and hundreds, and you meet engineering problems — cancer from extra cell divisions, Alzheimer's, metabolic disease, regeneration — that evolution never signed up to solve.

Kellis wants a distinction drug developers already know in one direction and underuse in the other. An off-target side effect is a molecule aimed at gene A that hits gene B. An on-target side effect is a molecule that does exactly what it was designed to do. The gene is pleiotropic — many ways of acting — so a 70 percent knockdown that helps the disease pathway harms the others. That is why the strongest GWAS genes can be the worst drugs, and why modest, clearly causal nodes can be the better ones: a small push is enough, and the rest of the network stays standing.

The next interventions, as he sees them, sit at the network: multi-input, multi-target modulation that admits the wiring is dense and nudges a system from one state to another. Delivery has to match. Collaborator Ron Weiss's synthetic-biology constructs can sense whether they are in a tumor cell, an immune cell, or a stromal cell and fire only there; other cassettes can be tuned to liver, heart, or neuron. Coupled with cell-type-specific enhancers, a drug stops being a whole-body flood. Logic gates in the tumor microenvironment are already running in the lab.

Put the pieces together and you get what he calls systems medicine: diagnosis from your particular mix of common, rare, and somatic marks; intervention from your exposures, biomarkers, and current state; tools that modulate specific pathways in specific cells. The aim is not only to drag someone from disease back to normal, but sometimes from normal to better. The unfinished board includes generative models for chemistry, molecular tricks at DNA and RNA, electronic health records for prediction, polygenic scores weighted across the three variant trees, and cell-type delivery of the kind Fridman has discussed with Bob Langer. The historical lag from discovery to therapy, Kellis says, ran around 50 years, then 40, 30, 20, and now about 10. The density of discovery tools now in play will, he thinks, produce the most dramatic manipulation of human biology in the next few years. Asked whether some of the diseases they started with can be cured, he does not hedge. "Absolutely. Absolutely. It's only a matter of time."