Cells and Tissues Can Learn and Remember Without a Brain | Michael Levin | Episode 250

Yeah, it's good to be here. Good to meet you.

Yeah. Um, I'm Mike Levin. My background is first computer science and then cell and developmental biology. Uh, I run the Allen Discovery Center at Tufts University, and my lab works on many different issues around, um, how cognition manifests itself at all different scales and all different kinds of embodiment. And so cognition can manifest at different scales, meaning not just the organismal or whole brain level. We think it might manifest at smaller scales.

Correct. What we do is we take, uh, approaches both conceptual and practical bench tools from neuroscience and we apply them to all kinds of things that are evolved, engineered, and hybrids. So this means cells, this means trying to study memories in cells and tissues. It means trying to communicate, uh, goals to, uh, organs in biomedical settings. It means looking at materials and hybrids and cyborgs and various things in terms of what kind of minds do they have, what kind of processing can they do, and how many of the tools that we have in behavioral science can be applied far outside the brain.

Yeah. Um, I can give a rough definition. I don't get too caught up in definitions per se because I think especially as we apply some of our concepts outside their normal areas of application, the definitions can often hold you back more than they facilitate. And so much like with defining life or those kinds of things, I don't think that sharp definitions are doing us that much good. I tend to operate from a very kind of, um, pragmatic operationalist perspective. So in other words, I don't want to argue about the definition of cognition. But I do want to point out that isn't it interesting that, A, we can take tools that we've been using to study brainy behavior for, you know, many decades. We can take those tools, and I'm talking about everything from, um, optogenetics, neurotransmitters, perceptual multistability, active inference, you just across the spectrum of practical and conceptual frameworks that we study in brains. And we can show that those tools allow us to discover new science when you port them elsewhere. So, so my point is simply this: if the tools can't tell the difference, then you're not, uh, you're not able to say that, well, that's not real cognition. That's sort of as-if cognition. It's real if the tools are helping you find new science. That's how I see it. So, so by cognition I mean a very wide spectrum of properties, all the way, functional properties all the way from simple learning, habituation, sensitization, associative conditioning, predictions, uh, being able to make predictions, being able to hold counterfactuals, uh, planning, um, all of those things. I don't, you know, we haven't claimed, for example, language use in any of these things. So, so I don't, you know, obviously language is part of cognition. I don't, we've never seen an example of proper grammar at least yet in these things. So I don't say language, but most everything else that I've just mentioned exists far outside the brain.

Yeah, let me, let me, uh, take one step back before, before we get into that, just to, just to kind of set the stage a little bit. Uh, if we ask ourselves, and so, so, so what I think, what neuroscience is not the study of neurons. I think, I think neuroscience is the study of the cognitive glue that binds competent subunits into higher-level cognitive agents. So you and I are more than the sum of a bunch of neurons because there's a set of policies and mechanisms that allow a very particular kind of networking to form. And so that's the electrophysiology that allows our neurons to work together to have memories, goals, preferences, and to have computations that individual neurons are not doing. That's not to say they're not doing any. They're doing some, but they're not the same as what you and I are doing. And so we can ask then that question. Okay. So, so that's, that's a remarkable, um, that bioelectricity as a, or electrophysiology as a kind of cognitive glue in that sense. Where did that come from evolutionarily? Like, how old is it? And it turns out that evolution discovered that electrical networks are really good for this kind of thing. So, uh, um, integrating or synthesizing information across space and time, storing memories, uh, doing computations, that was discovered by evolution around the time of bacterial biofilms. It's extremely ancient. So all of the same components that, that we study in the brain, ion channels, electrical synapses or gap junctions, um, neurotransmitter machinery, all, all of that stuff is incredibly ancient. It predates multicellularity and it, um, it's, it's, you know, it's, it's homologous. The kinds of things I'm talking about are evolutionarily homologous with the things that people study in the brain. And so the next question you would, you would ask is, okay, so if these things were around before brain and muscle were here to move us in three-dimensional space, which is why the, the kind of, um, uh, digital coding is, is, is so prevalent, what, what, what were these networks thinking about before? And my claim is simply this: that everything we do in the brain is an evolutionary pivot of using these exact same mechanisms to traverse a different space. This is long before nerve and muscle came on the scene. That is a space of anatomical shapes. So here's where the analog and the analog computation is going to come in because you have a bunch of cells. The cells form a collective intelligence very much like we are a collective intelligence made of neurons. But cells were, were assembling into collective intelligences long before there were neurons or muscle. And what, what that set of electrical networks was doing is thinking about how to, how to move you from where you are at the beginning in anatomical space, your one cell, your fertilized, fertilized egg, how to move you to wherever you're going to be. And it's not a simple, it's very much a navigational task that requires all kinds of protocognitive competencies. It is not a simple, um, hardwired, well, follow these steps and here you go. And we can, I'll, I can discuss why, why that is, but it's, it's very much a navigational problem-solving process. It's actually what, what I call an intelligent process because it has the ability to solve novel problems. And so that process uses analog computation because the way that those electrical signals are interpreted are not the way that neurons in the brain are interpreting, uh, their activity using, using this digital coding. Because what, a couple of things happened in pivoting from, from developmental to, uh, behavioral, uh, controls is that first of all the rate, the speed had to, had to go up immensely. So we're talking about processes that take hours, sometimes days, as opposed to milliseconds. So the speed had to go up and then, then, then some, there's some other benefits of...

Correct. And what's playing out are sensing, perception, decision-making, attention, recognition, um, all, all of those same things, but, but they're, but they're stretched out in time. And, you know, we're not very good at visualizing behavior that happens at very slow or very fast time scales. And so people's, this is why we have separate departments of developmental biology and neuroscience. But we made a tool. I, I used to have my students do this by hand. And now, now there's an AI tool online that will do it for you. You can take pretty much any abstract of a neuroscience paper or a behavioral science paper. And all you do is you put it in, put it in a word processor and you just do a find-replace. And everywhere it says neuron, you just say cell. Everywhere it says millisecond, you say, you say hour. And you could make a couple of other tweaks, but basically that's it. You have yourself a developmental biology paper. I mean, it's amazing how much, uh, um, symmetry there is between these two fields. And I think it's no accident because evolution is telling us that this is, that this is what happened. So that's where you have a, um, an analog code because the interpretation of the bioelectric states that go on in the body as it's trying to figure out how many fingers you're supposed to have and all this kind of stuff, those patterns are, they're analog. They're, they're spatial distributions. Uh, sometimes they actually look like the thing that they encode. Sometimes they look nothing like the thing they encode and it's a lot of hard work to decode it. Much like neural decoding, people are trying to do that in the brain. Um, and yeah, and, and we pay a lot of attention to how the interpretation works and it has a little bit of an analog flavor, of a digital flavor in the sense that there seem to, there, there are ways to, so, so using this bioelectric code, we're trying to crack it exactly as neuroscience is trying to crack the, the code in the brain. And some of the messages in that code refer to specific organs. So we, we know a message that says build an eye here and we can, we can reproduce that message elsewhere and you can get eyes in the weird parts of the body and so on. So, so there is a kind of a digital aspect to it a little bit, but, but the patterns themselves fundamentally are very analog and they're very fluid.

Yeah. Um, there's, so, so, so the one thing that is basically, okay, so, so, so let me, let me list the things that are exactly the same and then the things that are different. So things that are exactly the same are that pretty much every cell in your body uses ion channels to set up a resting potential. So sodium, chloride, potassium, um, occasionally protons...

Every cell in your body is doing this. Okay, they're, they're using, they're using ion channels, ion pumps, ion transporters, cotransporters, all that stuff is being used to set up a particular voltage gradient, uh, in, of between the inside and the outside of the cell. So the resting membrane potential. Now there are gap junctions, which are electrical synapses, which sit in the membrane of cells. There are six on one side, six on the other side. They dock like a, you know, like a, like a, you know, two submarines trying to, trying to dock. Um, and they allow small molecules in particular, uh, like very small molecules. So in particular charge to, to go across, but also some other things that are not particularly ions as well. Um, and so that allows you to set up an electrical network of cells that just like, just like in the brain. So that, that part's the same. Um, some of the things that go through these gap junctions are neurotransmitters, things like serotonin. So that part's the same. The network itself is processing information that then it gets read out by other cells and other actions take place based on what the network is doing. That's just like, like neuroscience. Here are some things that are different. Um, first thing that's different is that neurons mostly like to sit at a particular highly polarized resting potential and then when they get activated they depolarize and they try to get back there as quickly as they can. Uh, the cardiac cells do it too. Most of the rest of the cells of the body are much more, um, a, much more flexible at what, uh, voltages they sit at, and b, those voltages change slowly. So they don't spike, they, they tend not to spike on a millisecond scale. They tend to change slowly as a function of computations that are happening in the network.

We make videos of these things. We use all the same tools. So, so we developed the first use of voltage-sensitive fluorescent dyes to track all these things. And so we, we make videos of what this actually looks like. And of course you have to, you have to, you have to time-lapse it because, and then speed it up because otherwise, you know, it takes too long. But you can, but you can see all the electrical conversations. You can see waves and you can see borders and boundaries. And here's where the, you know, this, this pattern we call the electric face that Danny Adams in my lab found years ago, which literally tells you where all the organs in the face are going to be. Here, here are the two eye spots. Here's the mouth spot. Here are the, the plots on the sides. So, um, yeah, and, and those, all of those changes are slow and they're gradual and the voltages can be anything. So a cell might be minus 30 and it might be minus 42 and it might be, you know, and the other important thing is much like in neuroscience is that it's generally not the voltage of any one cell that determines what it's going to do. This is not a single-cell-level code. It is the spatial pattern across a whole bunch of cells that determine what everybody's going to do. It's very much an emergent network, um, network property and it has a virtual governor, um, kind of aspect just like the, you know, just like the brain does.

Yeah. And, and, and the other thing to keep in mind is that when, you know, when I say the gap junctions allow, um, current to equalize between two cells, this is not just a giant syncytium where everything just ends up at the same voltage because both of the ion channels and the gap junctions themselves are voltage-sensitive. So what you have there is a voltage-sensitive current conductor, aka a transistor. So all of these things mediate very rich feedback loops, both positive, both positive and negative feedback loops, and that's why it can self-regionalize. So, you know, it's, it's not, I, I like to think it's, it's not an accident that, um, Alan Turing, who was, was very interested in intelligence and, and, uh, computation in other media and reprogrammability and all of that, he also wrote one of the first papers on a chemical model of morphogenesis, trying to understand how does order come spontaneously from a well-mixed set of chemicals. And I think this, this, I mean, he was, I think he was on to this idea that there's a profound symmetry between the self-assembly of bodies and the self-assembly of minds, and that, uh, you know, evolution happens to like bioelectricity for all of these things, but there's, there's a much more fundamental thing going on here, which is that you have to align and coordinate your subunits towards higher-order goals and spaces that they know nothing about. And, and this is, I think, works the same way.

Yeah. Um, I mean, the, the first thing that, just to, just to mention, and people say this all the time, is how could a clump of cells have goals, especially if they're not neural cells? And they say, well, how do your neural cells have goals? And then, yeah, I guess I don't know, although at some point that's not a great answer to this, although at some point say, well, it has something to do with the bioelectricity, the electrophysiology. I say, perfect, exactly, exactly the same way. And so also to be clear, when I say goal, I don't mean, um, some sort of human-level metacognitive, I know what my goal is, I'm happy about my goal. I don't, I don't mean any of that.

Absolutely. And so, and so I just, I just want to point this out that I'm not claiming second-order metacognition. What I'm, what I'm, I'm saying is as in a, in a basic, um, uh, homeostatic cybernetic type of system, there's a recorded set point which serves as the target of an error-minimization scheme. So that what's happening here is that there is a set point and you've got a system that's able to measure current state, it's able to access the memory of the target state, and as long as those two things are far enough apart, it's going to work hard to try to reduce the error to acceptable levels. Right? So this is, you can, you know, we, we have, it's no longer magical to say that machines have goals. We now have a mature science, you know, in cybernetics and, and, and so on, where we can actually cash this out in, in, in mechanisms. And so this is, this is of course not how people describe developmental biology normally. What developmental biology is typically, the story is described in a very open-loop, feed-forward kind of way, that you have a bunch of cells, they follow local rules, almost like a cellular automaton, they all follow local rules, and then they say, well, then there's emergence, meaning that if you do enough things in, in large numbers, eventually something complex will, will emerge. So that, that is, so I'm, I'm going to say here that that is a, that is true, emergence of complexity certainly does come from rules, but that is a very small part of the story because what we see in, and, and so now I'm going to tell why I think they have goals and, and, and how those goals are encoded, because, because what, what you have here typically, with the exception of certain things, maybe C. elegans, the, the nematode might work like that, maybe, but, but most things don't work like that. In embryonic development, in regeneration, in cancer suppression, you have a very active, continuous monitoring of large-scale properties and a repair to those properties if, if you introduce an error. And I'm just going to give you two, two examples. And by the way, I, I treat development and regeneration as exactly the same because I think development is just drastic regeneration from one cell. It's still, it's still an error-minimization scheme and it uses all the same components. Like, development is just, is just regeneration after you've been pruned down to, to, to one cell. But...

You know, uh, to, two, two, so, so, so the one that most people have, would have heard of is, uh, axolotl limbs. So, you've got this, you got this amphibian and, and you, you know, you chop off the limb or they bite each other's legs off and the cells very quickly start to, to, to proliferate. They grow. They make a new limb and then they stop. That's the most amazing part of this whole process. Not the growth, it's the stoppage because when do they stop? They stop when a perfect axolotl limb has formed. That's when they stop. So what you're looking at there is, uh, you know, basically an open-loop kind of emergence thing misses the whole point that this is, this thing is a, is a, is a homeostatic error-reduction scheme in anatomical space. It's at a particular point in morphospace that describes a limb. If you try to deviate it, it will, it will, it will try to get back there. Another thing that you might not have heard of that's, I think, even more instructive is what happens when you take a tail and surgically graft it to the flank of a, to the side of a, of a, of a newt or an axolotl. Over time that tail remodels into a limb.

You can put the wrong piece of anatomy on another part of the body and it will turn into what it should be.

And if you look at, and the most instructive thing is to look at the tip of that tail. There's a bunch of cells there that turn into fingers. Now these are tail tip cells sitting at the end of a tail. There's nothing locally wrong. Why are they turning into fingers? Because like in, just like in cognitive systems, there's an element of top-down causation that what happens is in the service of a large-scale goal, no individual cell knows where a finger is or how many fingers you're supposed to have, but the collective absolutely knows. And the, and how you know it knows this is not philosophy or linguistics. The reason that you know it knows is because you do the experiment and if you don't understand that this thing is going to, uh, reduce error from this large-scale pattern, you're going to make the wrong prediction. The, the, the reality is it's going to try to get back to the correct state either by growing something new or remodeling something in place. There is a very strong influence of this, of this, uh, set point on the lower-level subunit. So all of the cell behaviors, all of the molecular, you know, the gene transcriptions and the methylations and God knows what else has to happen in order to turn, um, the tip of a tail into some fingers. All of that gets triggered not by any kind of local error. There's nothing wrong with those cells, by this large-scale pattern minimization scheme to a specific goal. Okay? And so, and so this is what I mean. This is very much an experimentally based, uh, cybernetic kind of model. So, uh, we've looked, and so now that, okay, so now your question, okay, well, how the heck does it know what it's supposed to grow? And so I'm going to give you, there's a few examples we can talk about.

Yeah, exactly. Where's the set point encoded? Because, because, uh, the, the, and it's really important because when you change the thermostat setting at your house, it's, it's very critical that you don't need to know how the entire heating and cooling plant works. You're not going to micromanage that thing. All you need to know is A, that it is a homeostat, and B, how to change the set point, and after that everything goes well. The reason that's important is because in regenerative medicine, if we want to be able to regrow limbs, to normalize tumors, to repair birth defects, we could try micromanaging it from the bottom up, which is what molecular medicine is trying to do today. I think, I think that's a losing proposition in many ways, but also we can take advantage of the native intelligence of the material and find the trigger points and the set points that we can rewrite and let the thing do what it does best, which is to execute to those set points. So this is what we, this is the basis of our regenerative medicine, um, approach in our lab. So, so I'll give you one example that's kind of one, one of my favorites. So, so imagine a flatworm. So, this is a planarian. It's got a head and a tail, central nervous system, true centralized brain, uh, two eyes in the head, and a very particular head shape. If you cut that thing into pieces, what you'll find is that every piece regrows a perfect little worm. Okay? So, so every piece has the information needed to remake, remake an entire worm. So, they're highly, highly regenerative. You can ask the question, uh, how does it know how many heads it's supposed to have? It's an interesting question because, uh, you know, just saying, well, it's in the DNA, that doesn't help. We know the DNA doesn't contain things like head number or head shape directly. The DNA codes for proteins. So if you look at a wound, when, when you cut a planarian in half, just imagine, you know, head, head here, tail here. You cut it in half like this. The cells on this side of the cut need to make a tail. The cells on that side of the cut need to make a head. They were near neighbors before you cut them apart. Why do they have different anatomical fates? It's not a local decision. You need to know where, like, what's the rest of the, the fragment doing? Right?

So, we started studying this, and what we found, long, so this compresses like 12 years of work. Um, long story short is that we found that there is a bioelectrical pattern across the entire worm that, uh, that we partially decoded that says how many heads you're supposed to have. And what we found is that if we go in and rewrite that pattern, and the way that we rewrite that pattern, in the case of the planarian, we use drugs that open and close certain ion channels. In the case of the frog, we've used optogenetics and...

No, it's more complex. It's more complex than that. I mean, in the, in the planarian it's like actually almost that simple, but, but in many of these cases it's a much richer pattern. So, so, and in fact it's often quite complex and you need, we, we're working on software to try to deconvolve it the way that, you know, when you do neural decoding, somebody's looking at a picture of a rose, you don't look in their brain and see a rose, right? You see a bunch of crazy signals that you have to somehow interpret. It's a little like that. There are complex spatial patterns that you can see and you have to try to figure out which of those, what aspects of that pattern encode what organ shape, orientation, size.

Correct. And, and, and when I say time-evolving, it's, yes, it can evolve with time, especially in development, but overall it's quite, especially in adults, planaria or adults, it's quite stable. So you could look and it just sits there and then this is, you know, it's quite a rich, I mean, I can't, you know, if you go to my website you can see videos of and, and of images of all this stuff. So, so what we figured out how to do was to reset that pattern so that by opening and closing certain ion channels, so that instead of reading, instead of saying, instead of being the one-head pattern, it becomes a two-head pattern.

Yeah. But the, the key here is that we're not just disturbing them. This is not, this is not, uh, you know, sort of whack the television set until, until, you know, it does something. This is a very specific, we have a computational platform that given a particular circuit with known ion channels and whatnot, it can help you derive if you want the pattern to, to be this instead of this, which ion channel should you open and close. That's really a non-trivial problem. But for example, this is how we do repair of birth defects because oftentimes there's a, there'll be a birth defect where you can see the pattern is incorrect and, and we know that if, if that goes on, there's going to be a very severe craniofacial, heart, gut defect. We can actually go in by using that computational platform and say, okay, I'm going to fix that pattern to be, to be the normal pattern, and the, and the, the model will say, well, to do that you need to close this channel and open that channel and, and, and tweak some gap junctions or whatever it is. And if you do that, then the pattern changes shape and then, and then you get the repair. So in the case of the, so, so it's, it's very targeted stuff. Uh, in the case of the, the planarian, yeah, it turns out that if you change that voltage, you get two-headed worms. And the amazing thing about it, so, so here are some amazing things about it. First of all, we, in it, it's extremely modular in the sense that I, I didn't have to tell it what genes to turn on or what cells belong in a head or how to make a head. We don't have any idea how to do any of that. All we said is at the very highest level, you, you should have two heads, and that propagates down to the, to the cells and the molecules, just like when you and I are talking. I'm not worried about tweaking the synaptic proteins in your brain to make you remember what I'm saying. You'll, you'll do that yourself. I'm giving you a very high-level signal and, and you, and with our cognitive systems, just like that, are multi-, multiscale kind of organizational architecture. So, so incredibly modular. We didn't have to sweat the details. Even things like size control, we didn't have to make sure that the head was matched the right size. It already does that. By the way, if you try to cheat and you do this molecularly with, um, uh, uh, for example, a Wnt pathway modulation, as other people have done in the field traditionally, you can get extra heads that way, but those heads are not, they're not scaled correctly because, because yeah, you turn them into a head, but you didn't handle the size control. The bioelectric switch is way on top of that. Once the thing's committed to a second head, it will handle all the...

Exactly. Exactly. And I'm not, and I'm not saying they couldn't. I mean, eventually, eventually, if you wanted to be in charge of what all the other knobs were that you needed to do in parallel, but you're saying you can go, you can go above that.