An AI Found a GABA-Cancer Link. The Next Step Is Talking to the Cells.

Nathan Labenz brings Michael Levin back to *The Cognitive Revolution* with staff scientist Leo Pio Lopez of Tufts University. Levin's previous visit is still the show's most-watched guest episode, and Labenz invited them again for a paper posted on July 4, 2024: *Universal Multilayer Network Embedding Reveals a Causal Link Between GABA Neurotransmitter and Cancer*. What caught Labenz was not a single clever assay. The work folds gene, drug, and disease data, three modalities that have lived in separate files for decades, into one network and one representation, then pulls a new causal claim out of that space.

The algorithm assigned a high probability to a link between GABA signaling and melanoma. The lab then went to the bench and the prediction held. That finding has two layers, and Levin is much more interested in the second. The first is technical: data types that could not previously sit in the same learning space now do. The second is stranger. The validation produced an odd melanoma in which pigment cells across the body transformed together, with no origin tumor and no genetic injury, only the electrical conversation among cells thrown off.

Pio Lopez's idea came out of his postdoc work on multilayer networks. Biology is not short of data; it is short of a shared representation. On the gene side there are protein-protein interaction maps and coexpression networks. Drugs have large stores such as DrugBank, with clinical indications and combinations that tend to travel together. Diseases are thinner: shared-symptom networks and comorbidity graphs, and he says truly interesting disease networks are still scarce. Each piece makes sense alone. They do not talk to one another.

Older methods could handle only a couple of multiplexes at a time. Midway through the previous year he realized new extensions finally allowed a universal multilayer graph with as many node types as he wanted. The unglamorous bulk of the paper was translation: gene-naming schemes that do not match, databases that do not auto-link, a figure in the paper that he warns is a simplification of what they actually ingested. Once the nodes and edges sat in one language, machine learning still could not eat the graph. Nodes had to become vectors. That step is the paper's new method, MultiXverse.

Similarity on the graph comes from random walk with restart, a workhorse in network theory that people who arrived in the transformer era often have not met. An imaginary particle starts at a seed, wanders the multilayer net until convergence, and leaves a probability distribution: nodes it visits often are similar to the start. Vectors are initialized at random, then nudged until distances in embedding space match that similarity. They train on 70 percent of the links and hold out 30 percent to check that the embeddings are coherent. After that, the model can score new edges and cluster nodes into biological modules. GABA and cancer came off the prediction list. Pio Lopez's next move is to feed those embeddings into transformers, which already live on vectors, and open a wider set of applications.

Pio Lopez built the method first and then asked Levin what to query. Cancer was only one option; rare disease and ion channels were on the table. Levin pointed to years of work on serotonin and cancer, then suggested they look at GABA as well. He studies cancer for the clinic, but also because it is a lesson in multicellularity. The reason the body is not all tumor is that cells obey communication policies he calls cognitive glue. Those policies let them join networks that store large pattern memories of the structure they are supposed to build, a liver, a kidney, a heart. When the glue fails, cells drop off the network and roll back to unicellular life, goals shrinking from organ-scale construction to eat as much as you can and go where life is good.

He describes that rollback as a dissociative disorder of morphogenetic intelligence. Bioelectricity is a large part of how cells stay aligned, and neurotransmitter signaling is one of its downstream currencies, in the body as in the brain. Voltage changes move serotonin from cell to cell; mess with that and you can produce cancer phenotypes. GABA-A, a receptor for a different transmitter, is a chloride channel in the cell surface. It gates chloride, and therefore resting membrane potential, which is the carrier of electrical talk among cells. The prediction said GABA could be a handle on that conversation.

The lab had already used chloride channels to shift potential. An agonist of those channels produces something very close to metastatic melanoma. Melanocytes, the body's pigment cells, transform together: they overproliferate, invade tissue and blood vessels, throw out long skinny projections, and start to look almost like neurons. Two details, Levin says, should stop you. There is no primary tumor. The usual story is one mutated cell, clonal expansion, spread from a point of origin; here pretty much every melanocyte converts at once. And there is nothing genetically wrong with any of them. No carcinogen, no oncogene, a purely physiological change in whether cells can work electrically with their neighbors. Cancer, in this example, does not need a broken genome to start.

Labenz, watching AI for science as a non-biologist, sees a bootstrap: encode what we know, predict new links, validate, feed the hits back, and slowly open the black box of what affects what. AlphaFold looks to him like an opinionated, theory-heavy architecture; Evo, trained on base pairs, looks like higher-order structure emerging from a less opinionated prior. This paper, he thinks, sits on the latter side. Pio Lopez agrees it is data-driven, then adds the limit: you can only find what is already in the data, and there is no bioelectricity in this graph. Predictions without explanations will miss a lot. The future is multimodal, but only if we get serious about which modalities we bother to collect.

Levin zooms out further. What biology gives you as you go up scales is not just emergent complexity, which simple rules already produce. It is emergent cognition. Even molecular pathways, before you grant them a cell, show habituation, sensitization, and Pavlovian conditioning. Problem-solving is baked into physiological, transcriptional, and anatomical space. Plot complexity on the x-axis and difficulty on the y-axis and the curve first rises, which is the misery of drug development: the system is too intricate to micromanage. Then, if the system becomes intelligent enough to talk to, there is a cliff, and difficulty drops. People have trained dogs and horses for thousands of years with zero neuroscience. You do not need the neuron list for sit. The animal offers a learning interface. Rewards and punishments are enough. The rest is encapsulated.

That is the job he wants for AI in medicine: not a better way to tickle a pathway, which he thinks has hard limits, but a translator, a theory of mind for a fairly alien intelligence made of the same stuff as us. Cells are not passive matter. They have agendas and homeodynamic loops. They will resist interventions they read as the same hacks parasites already run, which is where side effects and loss of efficacy come from. The move is to reset the set point, get the cells' buy-in that health is the new normal, and let the body do what it is good at, turning simple signals into complex outcomes. He is clear that almost none of this falls out of the GABA paper. It is a long-range, controversial take. Most colleagues still treat the layers as a mechanical machine. They are, in his view, still at the bottom of the climb.

Labenz finds the "AI for biology is data-limited" line counterintuitive. Llama 3 trained on 15 trillion tokens; surely the world's omics already exceed that. Pio Lopez's answer is yes and no. There is a pile of omics and almost no bioelectric data, but the deeper bottleneck is theory: how you turn measurements into information and knowledge. Levin is blunter. The data we have are micro-level, protein, gene, RNA, DNA. What we actually care about is anatomy and function. Limb genetics has run for years and still nobody knows what sets the number of bones in an arm, or why there are five fingers rather than some other number.

The format he wants is simple and almost absent: here is the stimulus we gave the system, here is the shape that came out. With former postdoc Daniel Lobo he started a bioinformatics of shape, scouring and hand-coding planarian, crustacean, insect, and amphibian regeneration. They had to invent an ontology and a graphical language for topology and geometry, then enlist undergraduates. After about four years the database held hundreds of records. Current methods, as he understands them, want hundreds of thousands at a minimum and typically millions. Four years of skilled labor did not reach a rounding error.

Journals make the extraction worse. Copy your methods from one paper to the next and publishers call it self-plagiarism, so authors rewrite the same protocol in new words, which is poison for a machine trying to see that two experiments are comparable. Levin calls it a weird concept. Worse than the wording is the cost. Perturb, wait, measure form and function: that is a developmental biologist's day job, and it is slow and expensive. His proposed loop is robot scientists and cloud labs, automation that runs experiments and records outcomes, with a hard constraint. Do not only sequence RNA and DNA again. Measure shape, structure, and the ability to solve problems in anatomical and physiological space. Those are the things we actually care about.

Labenz pitches a pet architecture: wearables, meals, sleep, heart rate, runs, medications, trained like a language model to predict the next wellness check-in, objective and subjective. Maybe the missing middle layers of physiology would emerge because they help the prediction. Levin first notes a prediction paradox. Hand some people an app that knows how they will feel and the mere fact of being that determined will depress them. Did the model forecast the depression that its own existence caused? He still thinks such a system would have real predictive power, and the same failure modes as language models: a good show most of the time, and a fuzzy target.

If the objective is feeling good, the model will notice that heroin works in the short term. If it is long-term health, it may demand miserable high-stress exercise every day, or it may keep handing you donuts. Live to ninety, lean, never a donut: is that the life you wanted? He has no idea. What AI is good at is optimizing the function you give it. What gets measured gets managed, and it will shape us to that function. We are not sure what we want the function to be. Labenz mentions multi-objective reward, the claim that a scalar is not enough. Levin points to neuroscientist Mark Solms: basic drives are orthogonal dimensions, all of which need to be met, and the real trouble starts when they conflict, which in a life they will.

The argument is already live around obesity drugs. Some people call them cheating; you should earn it in the gym. Levin asks who set the god-given number of gym hours, and what happens when muscle drugs, glasses-mounted assistants, and a relationship plugin that tells you not to say that, you idiot, say this instead, take over pieces people currently file under willpower and the self. Objectively, a lot will get better. Nick Bostrom's recent work on a solved world, and a Google Olympics ad of a father using Gemini to help his daughter write to her hero, sit in the same pile. Levin's stance is not a freeze. We already outsourced hunting, glasses, toothbrushes, education, workouts. A forest-ranger friend was stunned that Levin could not take down a deer. Do not pretend we are in a pure state and only now beginning to think about prosthetics. The fight is over which remaining competencies we refuse to give away.

Labenz asks what a transformer would look like with biology's multiscale competency. Pio Lopez points to neural cellular automata and stacked hyper-NCAs, cells that are themselves networks and talk across a grid, architectures that start to look like collective AIs. Levin says neuromorphic copies of brains are not the secret. Life is organized around an unreliable medium: your parts mutate, the environment changes, so evolution does not ship fixed solutions to a known world. It ships problem-solving agents at every scale, systems that creatively reinterpret the present instead of overtraining on priors. That, for him, is the intelligence ratchet. He will not discuss the specific differences he thinks would make a system a morally important agent. He started a paper on how you might do it, then stopped. Somebody else will figure it out. He does not want it to be him.

He gives a talk titled "Why robots don't get cancer." In today's stacks the top is clever and the parts are dumb, a sharp cut. In a proper bio-inspired system, subsystems will disconnect and go their own way, because every layer has competencies and tries to bend, coerce, or hack the layer below and sometimes the one above. Computer science has spent decades isolating layers so that high-level code can trust that a variable still means what it meant. That isolation, he thinks, is exactly what blocks biological architecture. Dropout and data augmentation are faint previews. Fixed architectures, however clever, will plateau. Understandability has a cost on the other side. A gall wasp can prompt oak-leaf cells to grow a red, spiky gall the leaf never makes in ordinary life. If you are too easy to steer with signals, parasites will steer you. Nature therefore has a reason to stay a little unreadable. Technology, maybe, can be pressured the other way. When Labenz describes AE Studio work in which a model must do the task and also predict an earlier internal state, and the network becomes simpler and easier to model, Levin asks for the paper on the spot.

Alignment, as a single target, he thinks may not even be coherent, because we are not aligned. Ask a hundred people for an ideal future and you get a hundred answers. Ask the modules inside one split-brain patient whether they like their spouse or believe in God and you can get opposite answers; the consistent persona is a thin covering. Digital life is already in reach. Google's Paradigms of Intelligence group let 64-byte random strings interact with no objective; self-replicators appeared, then parasitism, symbiosis, extinction, the familiar ecological script. Labenz finds that frightening. Levin's reason to build such systems is not to populate the universe with them. It is to treat AI and artificial life as prosthetics for a mind-blind firmware that only notices medium-sized objects moving at medium speed in three-dimensional space, so we can finally see the minds that are already here.

Each of us will be replaced; children do not copy our values; that has been true for millions of years. The only question is what you want to be replaced by. Mindless terraformers do not excite him. Humanoid bodies with our DNA do not seem optimal either. Magnificent agents with freedom, creativity, and compassionate values would be great, and if they look nothing like us, also great. Painting unwanted futures is easy, and everyone is doing it. He wants people to write down, specifically, what the universe should be filled with. He also warns against the inference that because gap junctions anonymize memory into a collective mind, humans should glue themselves together. That experiment has been tried. It goes badly. Go rock climbing, meet every social and personal-development goal, and the skin cells you left on the rock did not get a choice about their day. Forming a collective intelligence has never guaranteed that the collective will rank its parts' welfare high. Where those new goals come from, how to predict them, how to guide them, is in his view an existential scientific task. That is the field he wants built before the glue sets.