00:00:00 - 00:01:00 - Dr. Levin intro: PhD from Harvard, postdoc at Harvard Medical School, holds Vanevar Bush chair and directs Allen Discovery Center at Tufts, working to cracking the morphogenic code for applications of Regenerative medicine and basal cognition. Recent work: bioinformatics and ML tools for discovery of models and intervention in cancer, birth defects, organ regeneration. Talk title: “Biophysical and computational approaches to study decision making and basal cognition in cells, tissues, and synthetic living machines.”
00:01:00 - 00:01:20 - Interesting relationship between AI and biomedicine: to have regenerative medicine, need to understand the body’s natural intelligence; interplay between natural and artificial intelligence.
Websites:
Living systems are multi-scale (intelligence at different scales, aka “multi-scale competency architecture”):
Implications for biology:
so the way I see it and and what we do in in my Center is to bounce back and forth between machine learning and biology first of all to use machine learning to try to understand and control the the biological endpoints for example for medicine and conversely to use what we learned from the biology to create novel kinds of AI architectures that are based on much more fundamental and not necessarily neural principles so the main um points that I'm going to try to give you today are these that first I'm going to spend some time talking about the biology actually to try to paint in a different light what it is that we need to understand and these are very significant knowledge gaps about something that's important called anatomical homeostasis and um try to convince you that actually biology and and biom medicine is at a much um earlier stage than than one would think based simply on the progress in genetics and and stem cell biology I'm going to claim that fundamental advances and new medicine is only going to appear if we understand the decision- making by cells and tissues not only mechanisms but actually the algorithms of decision- making and I'm going to show you that a key medium for that kind of computation is non-neural bioelectricity this is this is it's it's it's basically a new kind of epigenetics but but it's it's a very important software layer of physiology that sits between the genome and the and the body and we need to understand this and we are working to crack the bioelectric code which will enable a huge class of electrocutica in heart defects regenerative medicine cancer and synthetic bi engineering so basically the whole the whole talk boiled down to two sentences is this I'm going to show you that like the brain your body tissues form electrical networks and these networks make decisions about Dynamic anatomy and new Tools in in Ai and machine learning are helping helping us to Target the system for some really amazing control that enable us enables us to to reap some some very serious benefits that are have advantages over competing competing Technologies so I show you this um this five-legged frog at the beginning just to say that I'm going to show you all kinds of weird creatures today these are not Photoshop these are these are real living organisms that are alive and well in our lab that serve as as tests of of the um the various theories that we develop okay so let's talk about the fundamental knowledge Gap what is it really that we want to do in biology and Medicine consider what the endgame might be of of all of our activities so at some point some someday in the future you should be able to sit down at a computer and draw right the anatomical layout of the animal or plant that you want so to really specify the structure the large scale structure of what you want to have and not not not at the genetic or or or um biochemical level but actually simply draw the thing you care about which is which is the shape of an organ or or or an appendage or whatever it is that we need for medicine whether for transplantation or directly in the patient you should just be able to drw and if we knew what we were doing we would have a piece of software which I call an anatomical compiler which would take that description and convert it into a set of stimuli that would have to be given to cells to build exactly that shape so here I've drawn a three diim a three-headed PL are and a three-headed flatworm in here is the three-headed flatworm that would arise if this thing was able to tell me what to do to all of the cells what what stimuli to give all the cells to get them to build this now why is this important well if we had something like this basically all of the problems of medicine except for infectious disease would disappear so if we knew how to tell groups of cells what it is that they should be building we would be able to resolve birth defects traumatic injury cancer aging degenerative disease all of these fundamental problems hang on one basic thing which is that we really need to be able to convince groups of cells to build whatever we want them to build healthy new um healthy new organs now um so let's let's ask this question how do our collectives of s know what to build so this is this is a group of embryonic blastam Mir shortly after fertilization and this is the cross-section through a human torso now look at this incredibly complex order right everything is in exactly the right position orientation the right shape the right size next to the right neighbors it's it's an incredible amount of order and so we ask how does this group of cells build exactly this every single time and so so where is this information of course immediately we're tempted to say DNA you know it's in the genome but of course we can read genomes now and we know now that it is in fact not in the genome there's nothing in the genome that talks directly about any of this the genome refers to proteins so it it specifies the micr level Hardware that every cell gets to have doesn't say anything directly about the shape the size the symmetry of the body so we have this this important question if we wanted to repair a piece that was missing or damaged we wanted the cells to to to rebuild it what would we have to do to these cells to get them to do it again and as Engineers we would like to know how far can we push this in other words could we get the same cells to build something completely different so let's think about what individual cells can do this is a single cell this is a creature called a lacr areia um you will notice that it has no brain it has no nervous system it has no stem cells no cellto cell communication it is incredibly competent in its own local single cell level goals so it handles all of its metabolic needs all of its transcriptional needs its um morpha genetic needs behavioral so you know here you can see it hunting for things in its environment of food how you know how how how can it do this without a brain so that's the topic of of a field an emerging field called basil cognition but what's what's interesting for us now is that these kinds of creatures these single cells that have local goals that they're very good at at pursuing can actually work together to for to pursue much bigger goals and I'm going to show you that momentarily but here's the basic framework we all start life as this single cell this fertilized egg and this is going to self assemble into one of these remarkably complex morphologies we know right away that stem cell biology is not sufficient this is a Teratoma this is a tumor it may have skin and hair and teeth and Bone and so the stem cell part has gone fine the stem cells have made their various mature tissue derivatives but what's missing of course is three-dimensional organization so not only do you need to have the building blocks of the body but they actually need to know where to go relative to each other so in in fundamental developmental biology the Paradigm looks something like this it's very much based on emergence and a kind of feed forward open loop system where you have Gene regulatory networks so genes turn each other on and off some of these genes make proteins that do things they diffuse or they're sticky or they exert force and then all of this happens in parallel and and then this this magical process of emergence where lots of simple things happening repeatedly in parallel give rise to a complex outcome such as this beautiful salamander of course all of that does happen so this is this is true but it's very limiting because it has a it has a fundamental difficulty which is if you wanted to make changes here let's say that I wanted this salamander to have three-fold Sy instead of bilateral symmetry or I wanted more legs or I wanted two tails or something this makes it look like all of the changes have to happen down here you have to make the changes at the genetic level that will then propagate and of course reversing this is a is a intractable inverse problem we have no idea in general what changes to make down here to give us um the desire changes at the system level right so so reversing this kind of recursive massively parallel system is is almost impossible now there there are other difficulties which are that we have really some some basic areas of of ignorance about genomes and Anatomy so I'm going to introduce you today to another creature called a planarian and plenaria and and so so the thing about plenaria is as I'll mentioned in a minute plenaria regenerate so if you cut off their heads and tails plenaria regenerate their heads and tails and they have a set of a set of stem cells called neoblasts that produce all of the new material so here's a species of plaria with a flathead here's a species with a round head and I can do a thought experiment I can say okay I'm going to take half of these stem cells from this flatheadedgorilla one of the shapes might be dominant to the other it might be an in between shape or maybe no it will never stop regenerating because neither set of cells is ever happy about the current state of the shape of the head so the interesting thing is that despite all the papers in in nature and science about the molecular biology of stem cells in this model system we have not a single model that will give a prediction on this okay because while we understand the pretty well the individual molecular events that guide individual cell decisions we we are have very poor understanding of what guides large scale anatomical level decisions and so how this Collective is going to decide what head shape it's going to make we actually have no models that that make predictions about that at all so overall where things stand are that we're very good at manipulating the the hardware so so cells and and individual Pathways proteins genes but we're really a long way from Form and Function and so control of Form and Function and so what I've argued before is that basically if we kind of think about um The Journey that computer science took that went from from this this is what programming looked like in the 40s and 50s where you literally had to rewire the system physically in order to get it to do something different and now of course why don't we do this right when you when you want to switch from from Photoshop to Microsoft Word on your laptop why do you not get out your soldering iron and start soldering away because we've realized that if the hardware is good enough then you can control it by experiences by stimuli right by by reprogramming not by editing the hardware and yet all the biology and biom medicine is down here right now all of the most exciting advances are all about you know crisper genome editing um changing promoters and protein engineering it's at the micr level Hardware so this whole journey is open to us in biology we are we are just at the beginning of this and I'm going to show you some of the ways in which the biological Hardware is absolutely good enough it's it's in fact amazing um one of the interesting things about about morphogenesis is that it is reliable but it's actually not hardwired it's quite flexible so here is this this Axel lottle it's a Mexican salamander these guys throughout their lifespan regenerate their limbs their eyes their jaws their spinal cords portions of the brain and heart so so um you know amazing everybody wants to be like like an aelole of course all the regenerative medicine is thinking about how to crank this this ability up in humans but but the most amazing thing about it is that if you for example amputate a limb here in fact you can amputate it anywhere you like it will grow very rapidly it will make exactly what's missing and then it stops that's the most amazing part how does it know when to stop well it stops when a correct salamander arm has been produced and if it's cut here it will grow this much if it's cut here it will only grow this much so it's context sensitive it can deal with novelty but it's reliable it can get to the same goal from different starting positions that by the way is William James is defin of intelligence is a system that can get to the same goal through different means um and the challenge is to ask how this collection Collective of cells can tell what a proper salamander arm looks like here are the plenaria I told you about so here are these um these these flatworms they have all kinds of internal organs they have a true brain lots of the same neurotransmitters that you and I have and if you cut them into pieces and the record is something like 275 pieces each piece becomes a tiny little worm it grows exactly what's missing no more no less and becomes a tiny little worm in fact um they're so good at regenerating they are Immortal so don't let anybody tell you that aging is some sort of consequence of um of entropy or or thermodynamics that it's you know we have to get old and die that's actually not true these plenaria don't have a lifespan limit they live forever um if and there's no such thing as an old planarian they simply keep regenerating whatever cells sesse and die now regeneration is not just for so-called lower animals so so the human the human liver um the human liver can regenerate and the the deer which are a large adult mammal can regrow their antlers every year they grow these antlers up to a centimeter and a half of new bone per day that they grow okay this is bone vasculature innervation um skin and then human children below a certain age can also regenerate their fingertips so so we also have some of this regenerative ability of course our body built all of these organs during embryonic development the question is how do we Kickstart them and how do we understand how the cells make decisions are they going to scar are they going to build an organ which organ are they going to build and I want to um kind of cement this idea of of of of of intelligence and go Direct in this in the following with the follow because it's going to be very important momentarily with with this example this is something we discovered a few years ago this is a tadpole of the Frog so here are the two eyes nostrils brain and gut um these tpes in order to become a frog they have to rearrange their face so the Jaws have to move forward the eyes have to move everything kind of has to has to remodel and it was thought that well every TPO looks the same and every frog looks the same so that all you would have to do is somehow encode a a distance and a direction for every piece of the head to move and and if everybody does that then you'll get from a normal frog to a from normal Tull to a normal frog well we decided to test that and we made these so-called Picasso tadpoles here's an example everything's in the wrong place the eyes on top of the head the Jaws are off to the side everything is just completely scrambled and what we found is that these animals become largely perfectly normal frogs because all of these pieces even though they start off in the wrong place um relative to the wrong thing they all will move in novel paths okay not not the paths that they normally do in in embryonic development they move through normal um through novel paths sometimes actually they go too far and they have to they have to track back a little bit but they keep doing it until they end up in this in this correct configuration through this through this interesting algorithm that we were able to um to to to infer so the amazing thing here is that genetics doesn't specify a hardwired set of movements what it specifies is a machine that can basically perform an error minimization task you can think of it as some sort of um means ends analysis where basically it just reduces the progressively reduces the error compared to the the correct pattern which again brings us to the this notion of what a correct pattern is so um so in my group what we focus on is is on top of this feedforward sort of emergent picture we put these these feedback loops that talk about what happens when the system is deviated from the correct state so now this could be injury it could be mutation it could be teratogens many different things there are these feedback loops that kick in both at the level of physics and at the level of genetics to get you back as as much as best as it can get get you back to where it's supposed to be this is basically homeostasis now on the one hand this should not be a surprise to anybody in biology we we know there are feedback loops of course we know homeostasis but um but there's a couple of things that are different here one is that whereas normal homeostasis usually um usually is is all about some sort of scaler you know pH hunger level something like that and and and so it's a single number here what you have is you have the the set point for this homeostatic process is actually a fairly complex descriptor it's some sort of coar grained description of what a correct anatomical pattern is so you have to think about how could biology store something like that that's the first thing and the second thing is that it suggests an important change of strategy instead of having to make changes down here which are very difficult to know what they should be for changes out here instead of solving this inverse problem what we might do is do the same thing you do with your thermostat when you change your thermostat to a different setting you don't rewire the thermostat you don't change the structure of the thermostat in fact you may not even know how the rest of it works you don't need to know the only thing you need to know is the fact that it is a homeostatic system and you need to know how to read and write the set point information so maybe what we can do biometic is leave the system intact it's very good at following different um set point information let's leave that alone let's exploit that and instead of trying to alter the Machine by gen editing or or changing Pathways let's try to rewrite the set point so here's what we've been trying to do we've been trying to identify the the mechanisms by which these set points are stored and learn to read and write them and I'm going to show you how that works now in this case the set point is quite large and one thing I want to point out is that as you know during evolution of course you have these individual cells that are operating on very small local single cell set points then they start to work together and there's this collective intelligence that can work towards very large goals you know this this this massive gold that's bigger than any individual cell in fact this is truly a collective intelligence because during this process many of the cells actually have to die the cells in between the fingers have to die and and so all the cells are kind of this this this massive Collective that's working on this giant thing not on their own individual goals but that process can break down and we'll talk about this momentarily this is cancer this is in fact a gleo gly blastoma cells crawling around and what happens is that individual cells can defect from this process they can disconnect from this network that binds them towards these large scale goals and they can start rolling back to their ancient amoeba level goals and they treat the rest of the body it's just external environment so so the the the size of the goals that sets that cells work on can can grow and shrink during the lifetime of the animal and there's some there's some very interesting aspects of this so let's so let's think about this now now what I want to talk about is exactly how does this happen how do how do cellular goals um scale up and what can we do about it and for that we have to understand bioelectricity and so I'm going to do is talk to you about some methods for addressing what I think is basically the the software of life and then let's talk about um what what can you do with this I'm going to show you some examples now we're going to spend the the next half an hour talking about bioelectricity and I want to be clear it's not because I think that bioelectricity does everything in the body it's clearly one layer of all kinds of interesting signaling modalities of course there's chemical gradients biophysical pressures tensions extracellular Matrix all kinds of things but bioelectricity actually is interesting in a in a very specific way it's not just another piece of physics that you need to know to track what cells are doing it's actually giving us access to a privileged access to a computational layer a layer where decisions are being made about what happens not just not just physics but actually um computation now we took as our in in in looking at this we took as our inspiration the one system that is uncontroversially known to be able to scale up from Individual cells to large scale goals and that's the brain so in the brain you have a you have a you have a collection of individual cells of neurons and they can keep memories and they can pursue goals and so on so how how does the brain do this trick well there's there's Hardware which is basically indiv each individual cell has little um proteins in the cell surface called ion channels they they pass ions such as potassium and sodium and so on in and out of the cell and as a result they acquire a voltage gradient so they acquire a potential a voltage potential across the meming and they can communicate that gradient across these controllable electrical synapses known as Gap Junctions that's the hardware the software and you can see it here this is this is Imaging a beautiful Imaging by this group of a zebra fish brain a living zebra fish brain thinking about whatever it is that zebra fish think about and you can see all of the electrical activity and it's a primary commitment of Neuroscience that these electrical Dynamics are holding the the mental content of this of this little mind that basically if we were able to decode what the what the patterns were were were doing we would know the animals Memories the you know the animals preferences um and and so on right so so this is where this is where all the software of cognition sits now it turns out that the brain did not invent this this kind of architecture from scratch it was based on Ancient ancient principles that evolution discovered way back around the time of bacterial biofilms and so every cell in your body does this every cell has ion channels most cells have have Gap Junctions to their neighbors and we can do a very parallel kind of effort to to to neural decoding in neuroscience and we can do sematic decoding we can say I mean here here's a here's a voltage sensitive fluorescent die showing you the electrical activity of an early frog embryo there are no neurons here there's no brain yet but already the cells are communicating electrically about who's going to be left right you know anterior posterior and so on and you can imagine that if we if we understood how to decode how to read and write this information we would have some amazing control over growth and form and so the first thing I want to do is um I want to show you um the tools that we use to kind of measure measure first first to measure and characterize these things so this is the voltage sensitive fluorescent die which allows us to see all of these things in in in in real time and then we do a lot of computational modeling and simulation so once we know what the channels and pumps are in in each cell we can ask where do these patterns come from I'm going to show you a couple of examples of these patterns so this is one called the electric face so here's a frog embryo trying to put its trying to put its face together it's in time-lapse and what you see is this is one frame from this video the the the depolarized region are are lighter here and what you can see is that already before all the genes come on that are um going to dictate where the eye is formed where the you know all the other organs already this tissue has decided where everything goes here's the here's where the animal's right eye is going to be here are the pla codes here's the mouth this is already and I'm showing you this as opposed to other patterns because this is the most the easiest one to decode I mean it literally looks like a face right we call this the electric face it literally looks like that um we know that this pattern is absolutely functionally instructive meaning it's it's required for normal development because if you interfere with it you create Picasso tadpoles so you can move all this stuff around move all the gene expression and all the anatomy by changing the electrical pattern so this is a normal um pre-pattern required and of course there are human channelopathies that have cranofacial defects that are that that are due to disruptions in this in this process now there are also pathological patterns such as this where we introduce a human oncogene into the embrio eventually it makes a tumor eventually the tumor starts to metastasize but before that happens you can already see the bioelectrics traces of these cells disconnecting from their neighbors in order to become individual individual amibas that are going to metastasize and and proliferate they have to disconnect electrically from their neighbors that were binding them into a collective that was able to store large scale goals such as form you know form a form nice chunk of muscle or or skin or something like that okay so these are these are the patterns now we have to be able to so we can read them now we have to be able to write them so we basically just stole all the techniques from from Neuroscience we do exactly the same thing it's very interesting the the tools cannot tell the difference between neurons and non-neural cells they work exactly the same which tells you something interesting about our our Notions of of what neurons are and and how fundamental that distinction is exactly the same Technologies work here so so I can take a piece of skin or liver or anything else and I can control these electrical synapses by opening and closing them so this is the topology of the network who talks to whom or I can directly set their voltage by opening and closing these channels so I can use drugs I can mutate the channels I can use light in terms of optogenetics what we don't do is any electrical field application so there are no electrodes there are no electromagnetic fields or waves nothing like that what we what we do is use molecular physiology to Target the very um components that are used by the by this network to to hold electrical properties and when you do this some very interesting thing happens I'm going to show you some examples one example is at the Single Cell level is this I told you that when Electrical connection between cells and their neighbors are lost this is the first step towards tumorogenesis and so here what we do is we inject a red florescent red labeled enene into into this animal and of course it it makes a tumor and there's the there's the red you can see that but here here it is again blazingly strongly expressed in fact all over the place here not just here it's everywhere and there's no tumor and there's no tumor because we co-injected another Channel this happens to be a chloride channel that forces those cells to be hyperpolarized and to keep connected electrically connected with their neighbors so even though the enogen is is is is cranking the um the cells are not going to go off in form tumors because they are electrically coupled to a network that overrides their individual cellular goals to to proliferate dedifferentiate and and so and metastasize and keeps them harnessed towards a large scale a large scale goal now what might what might those large scale goals be well here's one example um I told you that in the electric face we were able to get an idea of what the patterns are that Kickstart the different types of organs so what we did was we took an ion Channel RNA and we injected it in a set of cells that are going to become gut so here's a side view of a tle here's the eye here's the m um here's the gut and what we did was we imposed on on a on a region of those cells the same voltage that normally would be associated with making an eye in the face and sure enough what they do is they make an eye okay so what we've done is we've convinced these sets of cells that they should become an eye now the so-called master I Gene the the transcription Factor pack six doesn't do this it only works up here in the anterior neum but but the bi electrics Works anywhere you can convince any cell in the body to to start making an eye these eyes have all the right layers retina optic nerve all of that stuff and so notice notice a couple of interesting things one is that what we provide is basically a sub routine call we don't provide enough information to tell those cells how to make an eye in fact we have no idea how to make an eye eyes are very complex and lots of cell types we provide a fairly simple bioelectrical state whose interpretation is build an eye here and then the cells do the rest it's basically a subroutine call we've identified a module that builds eyes and we've identified how to trigger it where we want and then and then it takes care of itself so that's very good for regenerative medicine you don't have to micromanage all the outcomes um there are also two layers of instruction here these this is this is a a section through a lens that's sitting out in the flank of a of a tail somewhere that we've induced the blue cells are the ones that we actually int that we actually um modified to induce this new ION channel it's a pottassium channel okay these cells start to make an eye but the whole system really izes that there's not enough of them to to make a proper lens and so what they do is recruit a bunch of their neighbors these all these Brown cells here they never got the channel they're completely wild type native cells so there's two levels of instruction we instruct these cells you you need to make an eye they instruct their neighbors saying come help us the eye is supposed to be of this size which again we did not have to micromanage so that's so that's very good you can make we can make the same way by in by inducing electrical States we can induce Autos cysts or inner ear balance or organs um we can induce ectopic Hearts all right so extra Hearts we can induce extra forbrain we can induce extra limbs as I showed you here and we can make fins now that's a little weird because tpes aren't supposed to have fins and so we'll get to that momentarily but you can see that that that by manipulating the large scale bioelectrical patterns you don't just change specific cells or cell differentiation into different cell types this is a language at the level of organs okay and it goes beyond organs it goes to whole body axis so you can see here this is a planarian um I take this middle I take this worm I chop off the head and the tail and I leave this middle fragment the middle fragment 100% of the time will regrow a head here and it will regrow a tail there how does it know where it go where everything goes well if you look there's this interesting voltage gradient that um says one head one tail okay now what we can do is we can manipulate the ion channels in these in these tissues to say actually I want two heads or I want no heads okay so you can do that um we'll get into this this shortly so you can control now now one thing here is that um these these heads are normal and they belong to this to this normal species but you can go further than this and you can ask a species to make heads that belong to a completely different species so here's an animal with a triangular head we chop off the head we prevent the rest of the body from electrically communicating amongst the the cells so basically block the Gap Junctions for about 48 hours then we withdraw the blocker the cells settle into a new electrical State and sometimes they settle into their correct attractor and you get your normal head shape but sometimes they get they make round heads like this s Mediterranean sometimes they make flatheads like this pina these animals are about 150 100 to 150 million years of genetic distance except there's no genetic difference here there's nothing we didn't do anything to the Genome of these animals this is purely temporarily physiologically block them from communicating with each other and they found states in the attract landscape of that electrical circuit that belong to other species okay without without any genetic change and in fact not only the head shapes change but the shapes of their brain changes change the shape of the the the distribution of number of stem cells changes you can go even further than this and you can explore regions of the morphospace that biology doesn't use at all and you can make planaria that aren't even flat so they can be spiky like this they can be cylindrical like that they can be kind of a combination form these cells are perfectly happy to build up other things if the Electrical pattern of what it is that they're building is altered now we can use this for biomedical application so so here we are working on leg regeneration frogs unlike salamanders do not regenerate their legs so you amputate 45 days later there's nothing you can um you can you can add introduce a cocktail the cocktail triggers a build-a leg here kind of signal immediately medely you see this Pro regenerative msx1 it starts to come on and within 45 days L later you've got some toes you've got a toenail the leg is touch sensitive and it's motile so the animal can use it and eventually it makes a it makes a very nice makes a very nice leg um the treatment here is only 24 hours so what we do is we treat with the Ion channel drugs for about 24 hours and we don't touch it again you from a 24-hour treatment in adult frogs these are froglets but we um we we've shown this in in complete adults in adult frogs after 24-hour treatment you get leg regeneration for a year and a half okay so it really is a very early trigger where you convince the cells that instead of scarring they're going to regrow whatever naturally is supposed to go there which is a leg and you don't get tumors you don't get eyes you don't get tails you get proper frog legs and and and they grow for about a year and a half without us having to do anything now you know for bi medicine we have to we have to remind ourselves that yes this is not some sort of frog specific or worm specific technology we've done these things with human meenal stem cells with with cardiomyocytes all the kinds of things you would expect to do in vitro what we're doing now is we're trying for limb regeneration in mammals so the idea would be can we can we use a bioreactor a wearable bioreactor produced by our collaborators in the K lab they make these bioreactors inside the bioreactor is a gel with ION channel drug payLo and again only for the first day it's on and then basically it it protects the the the cells as they make decisions about what they're going to do and I have to do a disclosure here because David and I have started a company called morical Inc which is which is dedicated to using these kind of approaches for regeneration of all kinds of organs you know appendages and organs now um I should say that it is it is known at the Single Cell level how individ how voltage changes talk to the genome so the transduction Machinery by which voltage changes and individual cells control gene expression that's known we know about at least half a dozen mechanisms by which this happens some very common ones like calcium channels some exotic ones like voltage sensitive phosphatases so so we know this we know how the transduction works at a single cell level and in fact we know what's Downstream so we've studied using NextGen sequencing and microarray we've shown that all your favorite BMP pathway Sonic Hedgehog pathway fgf um all the wi all of these kinds of things are Downstream of the bi Electric decisions that the circuit makes then you have to implement them by rearranging morphogens and so on but all of this is even though we know these molecular details at the Single Cell level all of this has been highly unsatisfying because that isn't what we're trying to understand we're not trying to understand individual cell decisions we are trying to understand um large scale where large scale shape comes from and how to control it and so what we really need to do is this and so this is what we're working on a kind of full stack approach a multiscale approach where we start with transcriptional circuits that that result in specific channels being expressed then we model computationally all the bioelectric Dynamics in the tissue that result from these channels operating then we can ask how does this relate to large scale axial you know sort of whole body patterning and then from this we can derive algorithmic kinds of descriptions of what's going on that are really suitable for manipulations you know once you have this kind of white box model where you can see what it's measuring what what decisions it's making what are the control points you can start to intervene intervening down here is is is really really hard and so this is this is a um a simulator called Betsy that our collaborator Alexis pek wrote that lets us take individual cells in a virtual tissue and model every cell has these these networks both both chemical and bi electrical and then we can ask tissue level questions like hey when we cut it into pieces How come every piece rescales the electrical gradient how does that work and so here's a you know there's a model that could explain that and so here's where the AI comes in there's a couple things that that that that we've done and and if people are interested here here's all the software you can you can download it and described in various papers and you can play with it two two things that are really important on the machine learning side one is to help us um infer models so so knowing the the pieces that are in there the various channels and and transcription factors and everything else what is a what is a minimal model that explains the functionality that we see that explains how it grows how it regenerates what shape it makes so trying to discover models and in fact um we've done this so so so here in um a few years ago we showed how you can use evolutionary computation so basically just simulate Evolution over the space of possible models to try to match functional data of regeneration so basically we had a system we code it into it every known functional piece of data in the planarian literature so hundreds and hundreds of papers not not large data in the sense of um like genomic sequencing but actually functional data like we cut it in a particular way we treated it with particular drugs and then look it made you know a head and two tails or something and so that kind of that kind of information goes into the database and then there's a there's a there's an evolutionary computation across a simulator that tries to identify models whose behavior in the experiments matches what the database has in it from the Publications okay so it tries to infer act and so this is this is an amazing um I'm I'm still kind of amazed about the fact that this actually came up with a human understandable model of what's going on in planaria you know human scientists have been trying to do this for for probably 120 years and and the system actually guessed a very nice very nice model of how heads different from Tails and that's the most creative part of the scientist job right is is is trying to come up with these models and then here we have ai doing that so I thought thought that was that was pretty cool but we're going to have to have we're going to have to have this because um the amount of information functional information is is flooding us at a rate at which no no human scientist can keep all of these things in their in their head so so not only do we do this for inferring inferring models the next step is to infer interventions so now that you have a decent model you can ask this question um if I want the system to do something different how would I intervene where you know which of these nodes and and in what way do I need to intervene and you can look for very rare kinds of interventions where where most of them don't do what you want and in this case we were able to find some nice trigger points that let us control um melanoma basically the conversion of normal melanocytes into melanoma so so the AI is is important both on the on the on the model inference and on the intervention inference and um to come back to the to the bi electrics a little bit um because because I want to get back to the to the medical aspects um one one thing that you hear a lot I think it's important to understand the biology in in a new way in order to enable all these things and and one one thing I I'd like to talk about is is this metaphor of the code so you will often hear that the DNA is the software of the of the body and that actually the cells are the hardware that interpret the software and I think I think that's um that's okay for certain kinds of things but I think it's profoundly incomplete and limiting and I I'll give you a different metaphor to to consider what I actually think is is more more valuable is the idea that what the genome actually does is it nails down the hardware of every cell the gunome is just a description of what what proteins including ion channels are going to be present in the cell but now once you make um a tissue out of these cells even though they all have the same ion channels you basically have made an excitable medium in which symmetry breaking and and amplification can take place and this is this is now a a simulation a model of that of that scenario where where all the cells are proteomically identical they all have the same channel and yet look at all this Rich structure at the bioelectrical level if you've seen touring patterns um that's kind of a chem the chemical version of of that same thing same thing happening in the brain you can have all kinds of activity without having to wait for new new gene expression to come on and so basically you can build circuits that have this interesting property like a flip-flop where they can store different electrical patterns in the same set of cells they without changing out the cells or without changing the the proteins in the cells much like this flip-flop that can store a zero or a one via the the pattern of energy flow through the system and just by taking stock of what the parts are you can't tell what's in it this is why bioelectricity work is so hard you can't kill the cell and then fix it and then analyze it all all the information is gone at that point you have to study the living system and you have to do the physic profiling it's all in the in the real-time electrical activity it's not in the in the proteins so if this is true it makes an interesting prediction that we should be able to edit the software while keeping the hardware constant and I've kind of shown you all that already in the examples with the eye and and the heads and all that but I want to drill down to one specific example so here's a planarian head and tail has normal gene expression so the anterior genes are in the head not in the tail and when I cut it like this it makes a perfectly normal one-headed worm as it's supposed to and it's very reliable now here we have the same thing one-headed animal the anterior genes are in the front so that's good I cut that and I get a two-headed flat worm why why is that because I just told you that it was reliable how come the two-headed flatworm and it's because what we've done in the meantime is edit the electrical pattern of this body to say not one head like here but actually two and when we do this sure enough it makes a two-headed worm now here's the most important part of this slide this electrical pattern is not a readout of this animal this electrical pattern is the readout of this animal in other words this is not the electrical pattern isn't what the anatomy is doing now electrical pattern is a stored memory in this case a latent memory because it doesn't do anything until you injure the the worm it's a latent memory of what a correct planarian is supposed to look like which gets consulted after injury it doesn't do anything until then and and and the bioelectric pattern does not have to match the current Anatomy the currently it's a one-headed animal and it can have this this you know so so one a normal body can store at least two different ideas of what a correct planarian looks like so I promised you at the beginning that we were going to find the set point for anatomical homeostasis we're going to be able to read and write the memory of what the anatomy is supposed to be this is it and there are many more I'm just showing you a nice simple example this is we can now literally see how this tissue is encoding how many heads it's going to form if it gets injured and that information can live in the same in the same kind of body here now one interesting question you could ask is well what happens if you take one of these two-headed animals and recut it in plain water no more manipulations leave the I Chan alone you just going to recut it so now the standard Paradigm would tell you that well you you remove the primary head you remove this ectopic secondary head um the all all you've got left is a nice normal gut fragment we didn't do any genomic editing so or mutation so surely all you're going to get is a normal worm now right in the absence of this crazy reprogrammed tissue you're just going to have a normal worm and that's of course what um what you would think except that if you analyze the electric circuit you see that it actually has two stable points it has one stable point at the single head state but it has another stable point at the double head State and so when you do this when you actually cut them they don't go back to normal they one they continue to re to regenerate as two-headed animals so this system again remember nothing has been changed about the genetic sequence here and and and and of and we can change it back to normal we can take the two heads and and convert them back to normal by um blocking a particular ion pump for for a couple of days so so this has all the properties of memory it's long-term stable it's rewritable it's got conditional recall and it has you know two possible outcomes so think about what this means for for for biology um the question of how many heads this thing is going to have is not something that's nailed down by the DNA it's not in the genome what the genome gives you is a machine that reliably defaults to a pattern memory of one head but you can rewrite that and not only does the system keep it once you rewrite it apparently forever but it will obey the new pattern until you set it back to normal so that's very that's very important that means you have Hardware that has a default Behavior but you also have the software lever layer that you can interact with if you know how to read and write that information that does not require you to do any genome editing and so what we're trying to do right now is to really understand um really understand how the state space of these electrical circuits correspond to the different shapes and especially to merge it with with um paradigms out of connectionist machine learning where it's very understandable how networks how electrical networks store patterns how they store memories and in particular how they do pattern completion when part of that information is gone you know hopfield Nets and other things will restore it and so there's a very nice um mer kind of a merging of dynamical systems and energy minimization and learning and information and so so I think by biology has has tons to learn from um from the kinds of things that that that machine learning has been has been working on okay so so in the last couple minutes um I want to tell you about where where I think this is all going and U and and first I just want to show you something about the rational design of so-called electrocutica frog brain the electrical pre pattern of a frog brain and you can see the important thing to see here is that there's a particular pattern here that tells the brain exactly how wide it should be so it sets the size and shape of the brain out out of this whole you know out of this whole entire tissue and it has a very particular bi electrical pattern and the thing is that this bi electrical pattern is critical because if that bi electrical pattern goes wrong and it can be affected by nicotine by alcohol by various teratogens by mutations all kinds of things if that pattern goes wrong what you get is birth defects of the brain so so here are the eyes here's the forbrain here's the midbrain the hindbrain um you can have a hindbrain that kind of looks okay although not not not great and then the midbrain is is just completely screwed up the forebrain is just gone the eyes end up connecting directly to the midbrain it is just terrible so we asked a simple question if we know what the correct pattern is and we know that the pattern has become Incorrect and th led to this this defective the crania facial structure what would we have to do to change that so we made a model um and this is again work with our collaborator Alexis pyac she and VAV Pai in my group made this made this model of a um of the electrical circuit of the early brain and what goes wrong with it and then they asked the simple question they they interrogated the model to say okay if if that's what's wrong what channel would we need to open or close to get back to the correct State and notice that what you can't do so so here's this bell curve pattern is what you want you can't just raise everybody up to to to be at this level because those those that gives you birth brain defects and you can't drop everybody down to this level because that also gives you brain defects you have to try to recover this this particular pattern that's very complicated but the model was was was was amazing in that it told us that there's one particular Channel called hcn2 that would actually do this it would leave these cells down low it would raise these cells up high and and would it would predict that it would recover this amazingly enough um and I'm always shocked when when when these models actually give you something that that that works to to fix these very complex defects it actually works whether you introduce new hcn2 channels or you use drugs to open existing hcn2 channels here here's a normal brain here's a defective brain in this case it's even worse than a teratogen these animals have been um have been a mutant form of a of a notch protein has been introduced Notch is a very important neurogenesis Gene and without that the forbrain is gone the midbrain and hindbrain are a bubble these animals have no Behavior they just lay there doing nothing it's a it's a really strong defect despite the notch mutation look the the the structure of the brain is back and in fact their IQs come back so they um they they their learning rates you you can test test their learning rates and so on in in behavioral assays the learning rates are indistinguishable from normal so you can rescue even even not only from teratogens but actually from really nasty genetic defects if you manage the bi electrical pattern correctly and the way so this is important this this was not a screen we didn't try 100 different or a thousand different drugs the model told us very specifically which channel we would have to open to get the right bioelectrical pattern so we've moved you know when I started this work in 99 or so 1999 we were at the point of doing screens just asking what are what are bi electrical States for what do they do and now we're to the point where in certain cases we can make a model that specifically suggests Therapeutics it tells you which channel drug to use to fix a very complex defect and so the the the the vision I see going forward for for biom medicine looks something like this um there's a there's a a simulator that takes as input expression data which already exist from from RNA SE studies of what channels are present in what cells those are your control knobs those are the things you get to to tweak and of course what bioelectric pattern is desired for your tissue now this this is largely missing this this needs to be acquired these um physiological data sets need to be acquired for for most in most cases once you have those two pieces of information you can ask the model what channel and pump would I need to open and close to to to get the right pattern and then you can simply look for existing channels something like 20% of all human drugs are human used drugs are are ION channel drugs and then you can simply look look for either develop new ones or look for them and so this is again we've already um you can go you can go play with this a little bit we've already started on the basic framework here where it's called the electrocutica design environment where you would sit down you would pick the cells and tissues that you want you would say my pattern is wrong what do I do and this thing would actually then tell you to tell you which drugs you actually need to use okay so that's the that's kind of the vision and all all of this part is the machine learning part because it interrogating those models is is is not is not easy you can't do it just in your head you have to do the simulation and you have to be able to work the simulation backwards through a sech of some sort to know what the intervention should be so um just to close here here here are my conclusions that um there's a very important physiological layer of software that sits between the genotype and the anatomy and it's a really tractable Target for new biomedicine evolution discovered very early on that electrical signaling is a great way to process information to integrate decisions across distance so so the fact that we have organs and tissues that scale correctly and that connect up to each other other is is a consequence of that fact um we now have the ability to read and write some of these pattern memories and to thus to reprogram large scale shape heads eyes appendages not by micromanaging not by controlling stem cell properties but by telling the collective what it is that it's supposed to be building right taking advantage of the fact that it is a homeostatic system and that new machine learning tools are coming online as a kind of robot scientist platform to help humans scientists infer bioelectric and and biochemical circuits from functional data and then to infer interventions right so so how do we manipulate those circuits to give us the outcome that we want and um I think two things will come from all this not only um fantastic um regenerative medicine approaches that that will do way more than I think is actually doable with classical let's say crisper and things like that but will also actually feed backwards and help us design new AI platforms that are not based on brain architecture but on much more ancient principles of of computation and and and scaling of of of cognition in early cells so that's it um I want to thank the the students and the postto who who did all this who did all this work um thank our funders thank the model systems who actually do all the all the heavy lifting for all of this research and um again a disclosure this this there's a company called morphos euticals zinc that David and I are part of um yeah and I thank you for your time and I'll take question questions hello thank you Dr 11 for that talk I know a lot of our attendees signed up for our Symposium just to hear it and I'm certain that it lived up to their expectations it was truly revolutionary and I had no idea that was there was such a complex bioelectrical l that controlled our physiological outcomes now we will have a 10-minute question period so if anyone in person has questions I will ask them to please line up here at the podium and any online attendees as before may put their questions in the chat and one of our execs will relay them to Dr Len here the microphone please hello doctor how you do sorry I was just curious that when you mentioned that about inducing bioelectric signals with ions to make changes is more or like is additional energy required from Ela to fuel to to like fuel the changes in development like say addition nutrients and minerals as they're like an order of complexity to it yeah interesting interesting question um I I don't think so now nobody's measured this so it's possible that there are subtle differences but I don't think there are major differences because what you're doing is you're not you're not necessarily introducing new proteins you're not doing anything other than opening and closing the existing the existing channels that are there um and so I you know I I think well what it does require is a baseline level of resting potential which is kept up usually by the sodium pottassium atpa so that has to kept be kept going but I think I think not much energy is actually used to open and close these channels so um as far as I know I don't think there's an energy limitation here should any nobody's talking I can't hear anyone I think you guys are muted oh there we go just just press that because it's just on at this time no okay can I okay um Professor leine thank you so much for the talk it was really fascinating thank you um so I don't have much of a biomedical background so all of this looks like magic to me much of it but I do have computer science background and I'm trying to figure out how AI could help in this endeavor and so from what I understand the models that you're speaking of involve could involve like looking at a bunch of data that you collected from all of these different animals and then finding out some patterns which can then extrapolate to new kinds of data make new predictions that sort of thing so is that is that the kind of models that you're implying or do you have something else in mind could you maybe say yeah so there's been there's been a couple of different types of of of uses of AI here and then and then we can talk about what's still missing the thing that we did in planaria was to basically build a database basically build an expert system that knows every published result in plaria research so every experiment M of the form I did this and then I saw that I did this and then I saw that right so these you know experimental setup result right so those P you know those kind of pairs um we have hundreds I think like 800 maybe in the in the database so that is a knowledge database that's everything we know about planarian and then what the machine learning has to do is using a s using a simulator so a virtual planarian it has to try to come up with a a model of the regulatory circuit within each cell such that when you pull them together into a tissue and you start doing virtual experiments on them give the same answers as what's in the database right most models of course don't match that don't match that and so this was an evolutionary algorithm that eventually found a model that acts exactly like all not not only did it act like the experiments in the database of course we held back a bunch of experiments that it never got to see and then it correctly predicted those experiments and then it actually gave us some that had never been done before that we did and we saw the the right outcome so so so one thing you want to do with with with machine learning is to in infer models of what's going on not just predict numbers and and you know and um and and that kind of thing but actually infer human understandable descriptions of what the algorithm is that it's operating inside the cells and also across the whole tissues that's one that's one kind of thing another another thing sounds more like what I think you were describing which is that if if we had a large number of bioelectric measurements across different organs different ages different types of animals then then you could ask the system to try to infer patterns in that in that in that data and try to help us guess the bi Electric Code that would be super cool and I think we should do that the problem is that typically machine learning requires large numbers of data these experiments are very difficult to do they're very timec consuming we do not have anything like the volume of data that's required for typical approaches in this field so I think until the technology C up that kind of thing is going to be is going to be pretty pretty hard um but nevertheless what we what we have also done is apply apply machine learning to to to models of the code that we do have for example the brain this is how we discovered the intervention that fixes the brain and and ask ask the the the machine learning to pick interventions given an existing model so you know we still lots of technology to be developed but I think already we can see how how it's going to be used okay now we will take a question from the online chat it says what slow does The evolutionary process work for the cells to know the correct bioelectric patterning to produce the ultimate Anatomy yeah yeah great question I mean where do the where does the pre- pattern come from in the first place so um we we've tracked this down the best the best example we have of this is in the the early U is in the of of of left right patterning in the Frog embryo because we can track it all the way down to the Single Cell the fertilized egg and and so so we have all that information from the from the beginning the best way to explain that is I'll tell you I'll tell you the physics side of it and then I'll tell you what I think is the kind of computational side of it the physics side has to do with symmetry breaking and self-organization basically much like touring patterns where you can have very specific kinds of patterns appear in a WellMed medium if you have ion channels expressed across a sheet of cells and they're all identical so no patterning has taken place yet there's no pattern everybody's the same pretty soon you're going to have very characteristic patterns you know and the touring patterns would be St spots and stripes and bi electrics it's slightly different but but pretty soon you're going to have patterns spontaneously arising and what determines the shape of those patterns are fundamentally the property two two things the properties of the ion channels and the laws of physics that's that's what determines the shape of the patterns that spontaneously arise in tissue and so what evolution does is optimally tweak the properties of the of the electrical of the ion channels and of specific signals that get kicked off at different time points by by various cells that optimally shape those patterns such that they end up provi producing a useful um producing a useful organism okay so so that's kind of the physics end of it you can see that in its you know kind of using conventional Concepts there's there's another way to think about this which is you know my students often well when they first get introduced to this they'll say but where is the pattern recorded like I see the electric face where was that pattern recorded and I have to remind them that patterns don't have to be recorded so for example I can take um I can take a couple of transistors and I can make an and gate and it's it's it's a piece of physics but it also does logic and it has a very nice truth table associated with that or maybe I build a calculator and it can do signs and cosiness and then you can ask the question where is that truth table encoded right how does it know how to do that and again we have to come back to the fact that all of these things take advantage of the laws of physics and computation if you if you produce a specific kind of physical machine it's going to have behaviors this is the software part it's going to have behaviors that are not necessarily directly encoded anywhere by anybody it's a it's interest various interes laws of physics working through this machine and so that's the other way to think about this much like with with computers you end up with all kinds of calculation you know interesting calculations and computations that don't have to be explicitly pre sort of provided in the genome or anywhere else it's not where they are they're they're like kind of like the the properties of mathematics so um yeah so those are the those are the two two ways of thinking about it thank you Dr Lan we still have many people lined up for questions but we want to be respectful of your time if you have five more minutes we could take maybe two more questions and then we'll finish sure let's do two more questions and anybody that I don't get to hear feel free to drop me an email and later you know I'll get back to you offline okay thank you very much another plenaria related question here um so you mentioned that when you cut the original head off that another um it could potentially grow the head of another species um would that mean that the head part would be genetically different from the body part no no the genetics are are completely intact what it tells you is that the genetics don't determine don't determine the shape of the head and and and we already know that right we shouldn't be you know anybody who uses computers should not be so um really shocked that the same piece of Hardware can do multiple different things so there's a default right so any kind of like think about a um a programmable calculator you turn it on and by default very reliably the first thing you see is zero that circuit on the on a little LCD screen so that circuit has an electrical state that it reliably starts off in but it's a programmable calculator which means that you can now give it stimuli or inputs that alter the way that the pieces talk to each other internally and then you can cause it to remember various things and to do things that it wasn't doing when you first turned it on same thing here the genetics produce cells that by default reliably end up making triangular heads in that particular species but the cool thing about that Hardware is that's not the only thing it knows how to do that's just what it typically does you can by by by judicious interventions you can you can basically push it into different regions of the state space of the electrical circuit that that genome knows how to make so it it it's it's all of this stuff is forcing us to rethink what what we what is the relationship between the genome and the anatomy you know this idea that the genome somehow codes for the anatomy is that you know that's just not the right way to think about it the genome codes for a machine that by default knows how to make that anatomy And by the way if you want to see a really um kind of egregious example of this kind of thing I can go go to my website and under the presentations there's a bunch of talks about xenobots and we show that cells with skin cells with a perfectly normal frog genome can build a whole other type of creature it's it's kind of a synthetic Proto organism has different behaviors and different shapes and does all kinds of things it's the exact same genome so so we got to be just keep that in mind that the genome is does not directly set the shape at all thank you all right I had I had a question about the methods you're using to to change the electrical patterns why are you not using electrical fields and are you solely using drugs or genetic manipulations of changing how much ion channels are expressed right we use we use drugs we use ion channels modifications and we use optogenetics so light so so why are we not using electrical Fields um for the following reason there's nothing fundamentally wrong with electrical fields in fact back in the day starting in I think the earliest experiment was like 190 three when people didn't have these other Technologies they that's what they used they used electrical Fields the problem is that electrodes and applied electrical fields are really good at two things they're good at spiking excitable cells so if you want to Trigger action potentials that neuroscientists do all the time they're great at that so you can you know put a put in a like kilohertz um kind of a thing and and and Spike spike neurons that's fine and they're good for providing vectors for cell migration so if I put if I put two electrodes across a piece of tissue here um the cells will feel the electric field most cells love to crawl in electrical Fields they're just going to start crawling okay though that you can do but but notice what we have here what we have what what's important here are the stable steady state long-term distributions in resting potential that means I need cells over so in the electric face I need cells over here that are minus 30 and they need to sit that way for hours I need cells over here that are minus 80 and then some more minus 30 and then some minus 20 so this complicated pattern of wrting potentials that you you can't set that up with electrodes if I put electrodes on either side of a of a of a cell what happens is this side of the cell gets depolarized that side of the cell gets hyperpolarized the net voltage of the cell does not change at all I there's there's no way that I know of to use electrodes to set up a stable steady long-term complex image or you know a pattern of electrical potential it's just not it's just not not doable for now maybe someday somebody will will invent the way to do it but but for now it can't be done that way okay thank you Dr 11 so lastly for anyone whose questions are still remaining could I ask you to put your email in the chat so that people can see it yes absolutely yeah great questions thank you so much um thank you for the opportunity to share these ideas that is my email address please feel free to drop me an email and we'll chat thank you very much byebye everybody thank you thank you so much • Generated with https://kome.ai