View Transcript

There are an estimated 86 billion neurons in the human brain. Neuroscientists are actively exploring the importance of single or small groups of neurons versus networks of billions of neurons in the processing of sensory information, storage of memories, generation of movement, and everything else the brain does. Until recently, it simply wasn't possible to study billions of neurons at once. In today's episode, we talk with U-M's Dr. Anne Draelos, assistant professor of biomedical engineering and assistant professor of computational medicine and bioinformatics, who is using A.I. and bioinformatics to reveal the hidden networks of the brain, and exploring how gaining this understanding could improve lives.

Dr. Anne Draelos

Draelos Lab

Learn more about The Fundamentals

Transcript

Kelly Malcom:

Welcome to The Fundamentals, a podcast where we explore biomedical research here at Michigan Medicine. Research is fundamental to University of Michigan's mission to improve the world. On each episode, we'll meet the people behind the research, learn more about their fields and the fundamental questions they are trying to answer. I'm Kelly Malcolm, a science writer and communication strategist for the University of Michigan Medical School. This season, we'll start by explaining a little bit of the history behind the questions our experts are asking and get a glimpse into the future of healthcare. In 1906, Camillo Golgi and Santiago Ramón y Cajal were awarded the first joint prize for medicine, despite the fact that they vehemently disagreed with each other. Golgi studied the brain and nervous system, working to reveal for the first time the branching beauty of neurons, using silver solution to make the neurons stand out in black.

Golgi believed neurons formed one continuous web, acting as a unit; however, Ramón y Cajal perfected Golgi's standing technique, and in the process revealed that neurons were not chained together, but individual cells with gaps we now know as synapses that are bridged, using electrical and chemical signals to communicate with each other. Today, this debate persists, albeit in a slightly different form. There are an estimated 86 billion neurons in the human brain. Neuroscientists are actively exploring the importance of single or small groups of neurons versus networks of billions of neurons in the processing of sensory information, storage of memories, generation of movement, and everything else the brain does. Until recently, it simply wasn't possible to study billions of neurons at once. U of M's Dr. Anne Draelos, assistant professor of biomedical engineering, assistant professor of computational medicine and bioinformatics, is using AI and bioinformatics to reveal these hidden networks. Welcome to the show, Dr. Draelos.

Dr. Anne Draelos:

Thank you for having me.

Kelly Malcom:

So, our first question for you is, where did the idea that individual neurons control brain functions come from?

Dr. Anne Draelos:

Well, I guess it was hard to say where it first came from because ideas are nebulous like that, but I think you just talked about Cajal and Golgi, their 1906 Nobel Prize, so I think the prevailing thought before Cajal's work was that there was a giant mesh of interconnected sort of continuous bits, and what Cajal showed was that neurons were individual units, separate from other units, and so that became the prevailing view, and that really, I think, sparked a lot of interest in how individual neurons can be the computational unit of the brain. A couple more Nobel Prizes resulted from a lot of neuroscience research. So, in the 30s, there was a Nobel Prize from Sherrington and Adrian. They worked on synapses, so those are the junctions in between two neurons, and it deals with a little bit about how neurons can talk to one another.

Then, the 50s, the 60s, the 1950s and 60s, to be clear, had a lot of very interesting neuroscience research. These were often done with what's called single unit or single neuron recordings, so scientists would stick some sort of electrical probe into the brain and target hopefully a single, possibly a few neurons, and study their electrical activity. So, how the neurons are firing is what that's called. By the 60s, there was another Nobel Prize, Hodgkin, Huxley, and Eccles, on how these neurons actually communicate via electrical signaling and, actually, also chemical signaling. And so, I think these are some of the seminal sort of work. There are many, many hundreds of other neuroscientists and other scientists who work on this kind of thing, and that's from the experimental point of view, I'd say. Some experiment and theory in neuroscience on how neurons in the brains of living animals function, why they're the essential unit.

Kelly Malcom:

So, then, what led to the idea that networks of neurons were also important?

Dr. Anne Draelos:

In science, if you say neurons are individual units, and they can do lots of things, but can they calculate something as difficult as moving your hand to pick up a bottle, right? One neuron, probably not. I think that's an obvious logic problem, and so ever since the early 1900s, theorists have been trying to put together neurons. What can one neuron do? What kind of thing can it compute? Well, okay, if it's limited to a certain thing, just like our computers are limited to bit operations, what can groups of neurons do, right? And so, this idea has some roots experimentally in the 50s, sorry, 40s. So Hebb's was a scientist who works on assemblies of cells, working together in a smaller network. I would say here's where a lot of more theory, early artificial networks start to be important, where we can model a neuron computationally and say how do groups of neurons, what can they combine to create, like what sort of computation?

But a lot of those theories were very limited by the amount of experimental data that we had, right? So, I mentioned single unit recordings, where scientists can stick an electrode in the brain and measure the electrical activity from a neuron, nearby neurons, but if you want to figure out how hundreds of neurons are acting together, you need the ability, perhaps, to record from hundreds of neurons. You can, absolutely, simulate from first principles from theories, but does that match reality, is one of the important questions? And so, I think it wasn't until the past couple of decades where we were really able to start recording from hundreds, thousands, hundreds of thousands, millions of neurons, that the strength of neural networks, so networks of neuron working together, really became prominent.

Kelly Malcom:

So, really simply, how do you record from hundreds of thousands of neurons?

Dr. Anne Draelos:

So, I mostly don't, but the experimentalists who do, there are a variety of methods. One of them is scaling up those single unit electrodes. So, instead of an electrode, a probe being put into the brain to record from one electrode, you put lots of different electrodes, so similar to the way that computers have gotten smaller, you can put lots of transistors on very small surfaces. You can do the same thing with electrical probe; put it in the brain, and now you can record from thousands of channels. There was a really big advance. It's this probe called neuro pixels. Couldn't quote you the exact number of channels, but at least 1000 channels to record from, and each of those channels is picking up on one or a couple of different neurons. So, simultaneously with transistor electrical research, there's been a lot of really great advances in genetics and imaging research, so genetics, so that we can genetically modify organisms so that, effectively, the neurons in our brain will light up when we image it.

And so, things like calcium fluorescence imaging, basically, if you can image an exposed part of the brain, then you can record an approximation of the electrical activity via these calcium fluorescence traces. What's, I think, really cool about that, so some of the work that I did during my postdoc at Duke was to do these kinds of calcium fluorescence recordings in the brains of larval zebrafish, so they've been genetically modified so that their brains are effectively translucent, so nothing invasive. You can simply just image the fish and see an estimate of their activity throughout almost the entire brain, which is really exciting for people who are studying networks of neurons so that we can record from all of them how they work together and how they produce behavior, right? Because imaging and living organisms, not asleep or anesthetized ones, is super important if we want to think about how to correlate brain activity with actual behavior.

Kelly Malcom:

Okay, so very simply, what is dimensionality reduction?

Dr. Anne Draelos:

Dimensionality reduction is a mathematical technique to help us get a handle on all of these neurons, so it's a very general technique used in lots of areas of science. In neuroscience, one of the ways that it's used is to think about how to model the data from tens of thousands of neurons, using only 10 or 20 different parameters. It's not exactly, but effectively like, instead of trying to track each person in a large crowd, which takes what you have to track each person and their position and their speed and where they're looking or something like this. Instead, you look at group movement. Think of fish shoaling, so instead of considering each neuron, you consider groups of neurons. It's not quite that simple, because it's not technically groups of neurons, it's groups of neural activity, so how the neurons are co-activated at different times. One neuron might be part of many groups, but it's a way of condensing the information.

What's particularly relevant about dimensionality reduction for neural data is that it appears that these dimension reduced or groups that we consider are the ones that are strongly correlated with the behavior of interest, so if you looked at a single neuron and you tried to say, "Oh. This neuron is active when the mouse sees red, or this neuron is active when the mouse sees blue," you might get a mixture of answers, because the neuron is active for lots of different things, but if instead you looked at the dimensionality reduced activity across large numbers of neurons, then you would find that group kind of activity is the one that strongly signals, "Oh. I saw something red," and a different group is something that signals, "Oh. I saw something blue." It not only gives us a way to track complicated pieces of information, because you're not going to really want to deal with tens of thousands of variables at once, but it also appears to be a potentially realistic representation in the brain of certain things.

Kelly Malcom:

So that makes me feel like if you saw a pattern, then you could figure out what the animal or even person would be doing, right? So, could we really figure that out, maybe using AI?

Dr. Anne Draelos:

I would say a lot of people who are using the sort of what's called population doctrine, as opposed to the neuron doctrine that we talked about earlier, where the neuron was the most important thing, now perhaps there's this population doctrine where a whole population of neurons is really relevant. So, a lot of people do use AI and machine learning to help find these patterns, right? If you're looking at flock of birds or group of fish and you're like, "Yes, I can see the group that's moving from left to right," but this is one group, and it's an easy pattern to find. Although our brains are really great pattern finders, they're maybe not as good dealing with ten-dimensional data sets. So, AI is really great at doing this. You can either use sort of deep learning to help you identify patterns, especially if you know sort of what you want to look for like, "I want to look for a pattern that correlates when the mouse is running to the right, and I want to find a different pattern, hopefully maybe when the mouse is running to the left," for instance.

There are other methods, more statistically-based methods, that machine learning has helped really scale up to these large numbers of neurons. Those work differently, but produce similar conclusions for a lot of these kinds of pattern matching. When we think about identifying patterns of activity, you can do so in what's called a supervised way, where you say, "I want to find this pattern," if you already know what the pattern is, or you say, "I want to find a pattern that best represents the total amount of variation in the data." The smallest pattern to reconstruct the whole data is a common one. Then, there's this sort of other kind of pattern finding in machine learning. It's called unsupervised, so it says, "We're not going to tell you what kind of pattern to look for, but we want you to find patterns effectively." Some of these approaches, some of which I work on, are intended to help us remove some of the human from the equation, some of our biases about what is relevant to a mouse, for instance, and look for features that might help explain parts of the mouse's behavior that we weren't necessarily noticing.

Kelly Malcom:

Okay. That makes sense, but it also begs the question, what would you do this for? Why would you do this? What information is this giving us?

Dr. Anne Draelos:

Some of these patterns that we might want to look for are relevant for discovery. We might want to take a video of a mouse's face while it's doing some sort of task and correlate that information with neural activity. It is actually really easy to find, depending on where you're recording in the brain, correlations of neural activity with how frequently the mouse twitches its nose or flicks its whiskers, right? But there's other neural activity that don't seem to have an obvious correlation with behavior, at least not one that the experimenters looked at, right? So, when people do experiments, they choose what to measure. They measure pupil diameter, the nose, maybe the speed, things like this. If you don't know what's relevant to measure, it's hard to choose what you should be measuring.

And so, a lot of video analysis has given us the capabilities of at least recording more things, recording more neurons, recording more of the behavior, but now, if we're not deciding ahead of time what's relevant, we need machine learning to help us figure this out. In sort of basic discovery neuroscience, these sorts of unsupervised learning or even supervised learning, pattern-finding algorithms can help us say, "Oh. This brain area where we see some information encoding about the nose or the paw is also apparently encoding really fine motor movements of the mouse's cheeks," and that would've been dismissed as noise decades previously, but now with the power of AI, we're starting to identify more of what is the neural activity actually relevant for? So, that's one thing, sort of neuroscience discovery. The other super important thing that these kinds of patterns are useful for is more along the lines of neural engineering, right?

So, how can we interpret or use these patterns in the brain to help us restore some sort of function, help us restore vision or motor control? So, that's actually been a really successful part of the population doctrine in neuroscience, using dimensionality-reduced features in larger populations of neurons, building a decoder from those features in order to predict how someone or an animal is desirous of moving their fingers, their wrist, or wanting to pick up an object. So, these kinds of feature identification from lots of complicated neurons have been super successful in identifying lower dimensional features, so we call them neural manifolds, lower dimensional spaces. So, instead of living in a 10-dimensional space, you live in a 3-dimensional space. That correlates really well with our 3-dimensional movements, and have been very successful for things like neural prosthetics.

Kelly Malcom:

If I were frozen in my body, you could look at my brain and maybe tell what it is that I want to do, but can't.

Dr. Anne Draelos:

That is the goal, yes. I would say that a lot of neural engineering has become sufficiently successful, that even companies are doing this now. So, you've probably heard of Neuralink. I don't know everything they do, but that's one example of a company that's tried to even start to commercialize some of this technology. Here at Michigan, there are a number of us who also work on this problem. How do we interpret, or how do we build better machine learning models to interpret the brain, so that someone who doesn't have the use of their right arm, for instance, could think about using it and have some sort of prosthetic or other way to restore function in the arm as well. There's been a lot more work, I think, on controlling a computer, of course, because that's very useful. I'm particularly interested in how can we control more naturalistic things, like a restorative hand kind of thing, but that comes with its own complexities, of course. Controlling a computer mouse to go in a two-dimensional space is a lower dimensional problem than how do you control your fingers and your wrist in an arbitrary space?

Kelly Malcom:

Wow, that's really cool. Okay, so who are your major collaborators in this?

Dr. Anne Draelos:

Yeah, so Cindy Chestek, a professor here in biomedical engineering, and she and I have started working together on both different kinds of neural decoders. There's a lot of very interesting mathematical dynamics in these decoders that we study, where there's sort of an open problem in this field, where the best decoder for a person can change day to day, because your brain is quite dynamic, right? You're a dynamic person. Your brain is also a dynamic thing, and it can be hard to build stable decoders that continue to accurately translate what's happening in the brain, to what did we want to happen or what's also happening with behavior? And there are a couple other people in sort of biomedical engineering. We have a pretty strong neural engineering program here, and they work on a variety of these problems. Something I didn't talk about at all yet is how do you tell the difference between different models?

If you recorded from lots of neurons, you can fit a model to that and be like, "That is your neural network. The end," and then the next day you can do the same thing, and you get a different network. Which one of those networks is true? It's probably neither of them, but how do you tell? So, one of the ways that you tell is through causal interventions. You poke the network, right? So, you stimulate things. Stimulation, actually, has a long history in neuroscience for ascertaining function, and it is one of the ways, in addition to dimensionality reduction, that we deal with large numbers of variables. How do you figure out which neuron is most relevant? Well, you can use things like electrical stimulation or what's called optogenetic photo stimulation to cause a neuron to fire, cause it to be active, and look at what happens, right? Look at what happens in the brain.

And so, part of my lab works on how do we choose what stimulations to do? How do we choose what interventions to do, to learn something quickly, either in a neural network or in a couple of other kinds of problems. And so, some of the other work that the Chestek lab and my lab are doing are not just how do we build better decoders, more sort of dynamic decoders that might adapt themselves to an individual in real time, but also a kind of therapy called functional electrical stimulation, so stimulate the muscles to move. If we could interpret brain signals and then stimulate muscles to move, we could better restore function, but figuring out how to combine different stimulation patterns is a similar kind of high-dimensional problem that we also work on.

Some of my other collaborators at Michigan, so I also work with Catherine Kaczorowski. 

We have a project together where she has a lot of data. So, in addition to neural data and behavioral data, there's more going on in your brain than just electrical signaling, and there's more going on in your body than just your brain and your movement. And so, she and I worked together to combine genetic data, proteomic data, still electrophysiological data, maybe behavioral data, cognitive data, right? How do we think about doing dimensionality reduction across data types? How do we find patterns that don't just involve groups of neurons, but groups of neurons plus maybe your gene, plus maybe your gender? How do those work together? And so, she studies what's called cognitive resilience to things like Alzheimer's disease, and so we're working on some very large scale, probabilistic models. How do we find these features in that kind of data?

Then, Elise Saviera and Christian Burgess and I, all at Michigan, have a Rresearch Scout award from Michigan Medicine to tackle another kind of high dimensional problem where, instead of thinking about how do we find features and neurons, we look at sort of the opposite problem or inverse problem, which is the complicated kinds of visual inputs. Instead of thinking about how lots of neurons produce a simple behavior like movement, simple or not, we think about how do many different kinds of visual input affect a neuron or groups of neurons. This was actually very relevant for determining the utility of single neurons. The firing rate of a single neuron can change brain state during sleep, so single neurons are still important. The typical way that people used to do visual neuroscience is you might show a mouse some dots moving from left to right.

Then, you show the same mouse, dots moving from up to down or right to left. You would measure the same neuron each time, and you'd get a picture of how that neuron responds to movement from different directions. Then, later, you might record the same neuron, and show the mouse a red dot, a blue dot, a green dot, or something like this. I'm being generalistic, but rarely would you record a blue dot moving to the right and a green dot moving to the left, right? So, the combination of features, so color, direction, speed, size, contrast, shape, all of these things combine, and the neuron might respond to any one of these things, but how does it respond to all combinations of these things? And so, you can see how this could be a really high dimensional problem, because if you have 10 different speeds, 10 different directions, 10 different colors, you get billions of possibilities.

You are hopefully not just recording from one neuron. You want to get a picture of how do hundreds of neurons represent combinations of visual input, and so this is a problem where you cannot measure all of these things, especially not in the same mouse and not in the same day. What we do collaboratively is put a model of what the neurons are currently responding to that day for that mouse into the experimental system, and have that model pick what sort of thing should I show the mouse next? This is what's known as an adaptive design. The common analogy that I like to give people is that game, Wordle, where I think you have six guesses to make a five-letter word.

The important thing is that each time that you have a guess, you get feedback on your answer, right? And so, then you can really narrow your choices down. You probably don't have billions of options for five-letter words, but you have lots of options, and if you're ever going to get it right in six guesses, you need to be adaptive in your strategy for how you choose, so we use the same kind of logic for being adaptive in what do we show the mouse, to quickly figure out what do these neurons in this brain area, in this mouse, care about? How do they combine different kinds of visual inputs? So, that's a project I'm really excited about, so those are sort of my major collaborators here at Michigan.

Kelly Malcom:

Really cool. So, what drew you to U of M originally?

Dr. Anne Draelos:

So, I'm not from here, and I really wasn't expecting to like it as much as I did when I first interviewed here, so it was sort of a surprise, but people here were just amazing. I came here and I met with maybe dozens of other faculty and some of their students, and I was blown away by the excitement for collaborative opportunity, and really just the sense that it wasn't just about our individual sciences, but how we could combine them in way more powerful ways. So, maybe some of the Midwestern niceness, I'm not sure, but it's fantastic.

Then, ultimately decided to come to Michigan because of that, but also because I think there are some unique advantages. So, I'm in biomedical engineering and computational medicine and bioinformatics, so I'm part of the med school, but I'm also sort of in and around the college of engineering, right? And so, having those two things come together to be able to work on applying my work to basic discovery neuroscience, and applying it to very applied neural engineering problems and moving towards more human neuroscience kind of applications was just a fantastic opportunity, and I'm really glad that I came here.

Kelly Malcom:

So, I know you also teach, what do you tell your students, or what are some messages that you might have for people who are interested in studying neuroscience?

Dr. Anne Draelos:

So, I think, actually, a pretty common question I get from the students that I teach are more like, "Is it too late for me to go into neuroscience?" I think if neuroscience as an undergraduate major is still relatively less common, or maybe they knew they wanted to be a biomedical engineer, which has, again, tons of different sub areas, and they didn't really hear about neuroscience until maybe my class or a different class. Neuroscience is a huge field, so one of the things I tell is that you're not going to learn everything and you're not going to be an expert on everything. People from different backgrounds who are adept at communicating with one another and valuing each other's skills, it's a key pathway to getting involved in making an impact in neuroscience. My background was computer science and physics.

So, I did my PhD in experimental physics and did not think that I would end up in neuroscience, but as I was transitioning to a postdoc and wanting to do something new, I was reached out to by a professor who wanted some of my math and computer skills applied to neuroscience. I felt like I wasn't allowed to do neuroscience, because I didn't have the background, the biology, the chemistry, but he assured me that my skills were useful. And as long as I was interested in neuroscience, in the brain, and in learning how to talk to neuroscientists with very different backgrounds from mine, that would be okay. I'd say he's right. My own lab has people from different backgrounds, neuroscience, computer science, bioinformatics. I think that's what you want in neuroscience. I think some of what we talked about today in advances in recording, imaging, a lot of that honestly came from physics. Things like the genetic manipulations came from more of the field of genetics, potentially even bioinformatics. I think that's a really powerful way to move the field forward.

Kelly Malcom:

So, be like a neuron and get connected.

Dr. Anne Draelos:

Be like a neuron and get connected. That's very cute. Yeah, focus in on what you're good at, but then also talk to the neuroscientists who perhaps are coming from a more biological background and have those biological questions to help you figure out how you should be applying those different skills to address a really important question.

Kelly Malcom:

Okay. So you've hinted at potential applications for this work. What do you see is the future for this particular line of inquiry?

Dr. Anne Draelos:

Well, I will say one of the reasons that I did get into neuroscience is because I thought I would be challenged for the next 40 or 50 years, but I think a lot of the work that my lab is currently focusing on in hopefully maybe 10 years, give it 10 years, really focusing on this idea on using AI and machine learning to help custom tune and adapt different kinds of decoders to an individual and across time. I don't think we're going to solve all of this, not my lab, but I think it's entirely possible that the field will solve more of this with greater access to data, better AI combining strategies, sharing more of the data. We will be able to build models that can continually evaluate how well they're doing. How can they be better? That'll be per person, so that you can really have that individualized, personalized kind of therapies, and I think that would be a huge blessing to a large number of people.

Kelly Malcom:

All right. This has been super fascinating. I've always loved neuroscience, so all the work you're doing sounds so cool, and it's been great having you on the show. Thank you so much.

Dr. Anne Draelos:

Yeah, no. Thank you. I love any excuse to talk about research and neuroscience.

Kelly Malcom:

The Fundamentals is produced by the Michigan Medicine Department of Communication in partnership with the University of Michigan Medical School. Find us and subscribe wherever you listen to podcasts.