We Shall Finally Map The Brain - EP 36 Andrew Payne

embed the brain in a swellable polymer, and then you add water, it swells, it makes the brain physically bigger. You can make it, you know, these days you can make it 20 times bigger in every dimension, and that physically moves apart the neurons. And so now two wires that were overlapping that you couldn't resolve in a light microscope now can be resolved.

So you're literally taking one of these samples and, and stretching it out. And I was just with Ed Boyden, I didn't know this until I, because my, immediately I was like, oh, well, you're going to get some crazy distortion when you start doing that. But I mean, he was telling me it's like still, I mean, they're finding it's like 99% holding, holding scale and everything.

It, it preserves the structure remarkably well. Even version 1.0, like preserves the structure remarkably well.

Welcome to the Core Memory Podcast, which I never usually say, but here we are. We're with Andrew Payne. You are the co-founder of E11 Bio.

Yes.

Yes. And, and I wrote a story about you guys 18 months ago.

It was back in December.

Okay.

Okay.

Okay. So it wasn't even that long ago. All right. Life is a blur.

But yeah, I mean, thanks for having me here.

Thank you so much for coming on. You— I mean, I'll do the, do the whole intro on you, but, you know, you, you work on a subject that is near and dear to my heart, which is imaging the brain. That's what E11 is doing.

Yeah, you used to cover that back in the day.

I did. And now I think, I think we got the brainiest podcast going. I mean, we go, we go deep.

You—

I learned as I was researching you, which I didn't know before, you went to the University of Toronto for undergrad. Are you Canadian?

I am.

You are? All right. I didn't know that. You were— you disguised it. Yeah.

I grew up in a suburb of Toronto, Mississauga.

Okay. Okay. And then you end up at MIT.

Well, eventually. Yeah.

Yes. Yes. Yeah. I mean, this is, this is the quick bit. We're going to unpack all this as we go. But, but you, you end up at MIT. You studied under Ed Boyden.

Yes.

And I was just with Ed Boyden.

Excellent.

Yeah. And he, he, we will have an episode with him. I mean, um, it's hard to be more famous as a brain scientist than Ed is, I think.

And it was, you know, really a wonderful adventure to be in his lab. So I would love to dig into that too.

I would like to talk about that because I feel like the most interesting people I tend to meet come from George Church's lab or Ed's lab and there's this, you know, like I guess I did do both. You did both. Okay, well, this disqualifies— I feel like in the sports world people do these like coaching trees where there's like some famous head coach and then it's everyone who studied under them. I think somebody should do like the George Church, Ed Boyden, you know, professorship tree at some point. But you're young and you're very distinguished. I mean, you like invented, you know, important things that led to other big things around, around how we might understand the brain and image it. And, and, um, like just super accomplished. I mean, I've, I knew you and I, we've talked, but as I was reading, I got, I got more impressed.

Well, I mean, there's a lot to do, so, you know, can't go too slowly right now, actually.

Okay. It's like a really exciting time for Neurotech, which is pretty much the heart of my first, my first question. Um, but I'm going to set the stage for people who are not total brain nerds a bit. I mean, in There's been this thing called the connectome, which is roughly, I mean, vaguely equivalent to like the genome of the brain. We've wanted to create this picture of a brain down to the neuron level, this very detailed map. I mean, I guess like conceptually it sounds easy enough, but it turns out to be quite, quite difficult to do because I think we, we created the first full brain of a worm, the C. elegans. I think it was like 20 years ago.

Longer. It was in the '70s.

In the '70s. Okay. So we did that and that's like an accomplishment, but then we kind of went all the way until last year or this year even, and did the fly.

There was a bit of a winter. Yes. It took some time for the technology to catch up. You know, I think. You know, in many ways, like the worm connectome was this incredible milestone along the way, but also it was like a cautionary tale. It's like, do not do this at home, kids. Like you need to do a lot of work. It was before we really had like good computer vision or like good industrial microscopes. And so, you know, actually, you know, getting into it just for a minute.

Yeah.

Oh, back in the day, people would like take pictures of the, of the C. elegans slice by slice. Using these electron microscopes, take pictures of them, line them up into like a 3D volume, and then do a color-by-numbers with colored pens to trace the neurons through these individual transparencies because we just didn't have the AI tools and the equipment that we now have that allows us to be done at scale industrially. And that's sort of like where the analogy between like connectomics and genomics come in. In both cases, you sort of have like industrial-scale microscopy that allows you to get lots of data. And then you also have like a bioinformatics or neuroinformatics backend that allows you to actually cope with this big stream of data, right?

In the, the amount of data is, is astonishing. How does one get a slice of a— before I do that, okay, so we're going to get to this point where people are, are taking these very thin slices of, of mice brains, of human brains, like like basically using a razor blade almost, or something even better than a—

like a diamond-tipped, uh, most often been a diamond knife.

Yeah. And, and so they create the slice, and then you have to stitch all the slices back to— you have to image them and then stitch all this back together. But how does— how do you even get a slice of a worm brain to do this in the '70s? What does that look like?

Oh, I mean, it's very tricky. I mean, you end up doing a bunch at once in a big block, but, you know, you find the right specimen, and then, and then you, uh, Uh, you know, there's a lot of trial and error involved. All right.

What do we do at Core Memory? We cover innovative, fast-moving, forward-thinking companies, which is why Core Memory is sponsored by Brex because Brex is the intelligent finance platform for many of these companies. 30,000 companies from startups to the world's largest corporations rely on Brex's technology for their finances. They've got smart corporate cards. High-yield business banking and expense automation tools that are fantastic. I hate doing my expenses and Brex's AI software run right through those expenses, figure out where we're spending money and take care of so much stuff for you so you don't have to waste your time on it yourself. Go to brex.com/corememory to learn more and just, you know, get with the program. Let's get going. Let's get out of this. Archaic finance software and move toward the future, core memory and Brax. Okay. Yes. Yeah, I would, I would assume so. So I've, I've been to Jeff Lichtman's lab at Harvard where they have these brain slicers, which I think they were using on, on mice and human, small human samples at the time. But, um, essentially with the Connectome, you want to image this whole brain. We did the worm. We just did the fly. There's been many millions, if not, I think billions at this point invested in trying to get a human connectome, as far as I'm aware, I mean, we're on, to try to image a human brain, we're on like the cubic centimeter of a brain type, type levels.

Well, certainly we haven't gotten billions. So, but so how about I back up? Yes. So, you know, there have been like a few milestones in this, in this field of connectomics, you know, at the very beginning, actually, at the like very beginning of the 20th century, there was this guy, Santiago Ramón y Cajal. And he, along with Golgi, who you might have heard of, who developed this Golgi stain, figured out how to stain these neurons with silver. And it was like, like one neuron in like thousands. And, you know, back then Cajal was like, you know, looking at this with a light microscope and, you know, drawing pictures of these neurons. And you can look up these like beautiful drawings from Cajal that show you, you know, the shapes and sizes of all of these different neurons. And That was like how like neuroanatomy got kicked off.

Sorry, what year, roughly what year was this?

Oh, this is like, like 1910-ish.

Okay. Okay. Wow. Okay.

Um, and you know, this was, there were like a couple of things that came out of that. There was like this neuron doctrine, which was, you know, this articulation that like the brain is made up of neurons, right? They're made up of like these individual discrete units that commute, that compute. They talk at synapses, which are the connections between neurons and that, you know, your brain has like 100 billion of these neurons. So that's like 1e11 in scientific notation. And we'll get to this, but that's why we called the company e11.bio. You know, one day we want to be able to build the tech that will let us map all 1e11 neurons of a human connectome. But we are, you know, in many ways still very early. So like you mentioned, you know, there have been a few milestones over time. First, you could look at the shapes of the neurons. But, you know, that didn't give you like the circuit map that, you know, these neurons are connected into circuits, but you don't have a wiring diagram. And so then Sydney Brenner and his colleagues in like the '70s are like, okay, let's like try to get the wiring diagram of an entire animal, the C. elegans organism, this nematode. And so, you know, there was like, that was like the proof of concept that it could be done. And we learned some things from that as a field. And then, like I mentioned, there was this like long winter where, you know, we're developing The tools, there are pioneers who are trying to figure out how to get the technology to do this in a scalable, automated way. So, you know, like, um, uh, Winfried Denk at Max Planck was developing this, you know, diamond knife slicing approach. Similarly, you mentioned Jeff Lichtman, you know, his team was also developing this diamond knife slicing approach, you know, and the fly was developed and mapped in very much the same way, just at like, you know, a larger scale. It was still sliced into, you know, these thousands of slices. Pictures were taken with electron microscopes of each one, and then it was stitched together into this 3D model. Um, and so, you know, the question is, like, if we want to go further than the fly, if we want to go to, like, the mouse or eventually the human— because we haven't actually gotten to a mouse connectome yet— yeah, you know, what do we need to do? And is it going to look like scaling up more electron microscopes and better computer vision Or is it going to look a little bit more like genomics? And so this is where, you know, this, you know, uh, training in George Church's lab and other, you know, genomics background is really interesting because things work very differently in genomics. The first big, you know, human genome project was completed with Sanger sequencing, this, you know, technology that was, you know, very industrialized. There was thousands of these Sanger sequencers running across the country day in, day out. To sequence the first human genome, and that cost billions of dollars. And connectomics has not had something equivalent to a billion-dollar Human Genome Project. We're not there yet. But around 2007, things start getting really interesting because new technology hits the scene. And these microscopes— a sequencer is actually a microscope in a box these days. It wasn't back then. But around 2007, Illumina and its competitors developed these next-generation sequencers, and the cost of sequencing falls like 4 orders of magnitude in like 5 years. It's this massive cost reduction, and suddenly people are able to do sequencing in their own labs. They're able to buy these instruments. It's just massively cheaper. And so the question that E11 and I have been sort of obsessing over is, is there new technology that will let us have that next-generation sequencing moment for connectomics where, you know, we can start to conceive of what it would, you know, what it would take to go from the fly where we are today to the mouse, which is 1,000 times bigger, or the human brain, which is 1,000 times bigger than that.

And that was fantastic context. And let's explain why it is so hard. The— you mentioned early on there's a, there's a technique part, there's the machines. Some of this is physical labor and how you decide to mass produce that. You, you have to do these slices. And then part of this is, is computing power. Um, I mean, I'm sure you'll tell us, but we're talking about ridiculous amounts of, of data. Uh, so there's, there's these different bottlenecks in this. And is it— is what is like the core If you hear that somebody's done a fly and they're slicing the brain, when I first started to learn about this, I was like, oh, well, why don't we just buy just tons of slicers then and, and, and just mass produce this thing and just do the labor? Why will you know, what are the reasons that this has been so difficult?

Yeah, I would say that there's two reasons why this is hard. Um, the first one is, and you know, people have been you know, running pilot projects to understand, you know, where the bottlenecks actually are. And so there have been, there have been these amazing efforts in the last few years. So for instance, you might have heard the Microns Project, which was a, you know, effort to map a cubic millimeter of mouse brain tissue. That's, you know, 1/500th of a mouse brain, but, you know, is the biggest volume of mouse brain that had been mapped to date. And that along with other efforts, you know, Harvard, you know, mapped a cubic millimeter of human tissue. Both of those efforts sort of illuminated where the current bottleneck is, and it's tracing neurons through the images. So when you get these images, you have to slice by slice, cross-section by cross-section, trace the neurons from one sequential section to another. And this is like, again, it's the ultimate color-by-numbers problem. Instead of colored pencils, now you have AI, but you still have to detect boundaries, trace cross-sections, and not make errors. Because when you make an error, then everything downstream of that, you know, the error propagates, the errors are correlated.

So, because it's like, you're, it's like you're, I don't think this is, I'm, I'm gendering wires here. I think this is mostly a male problem is the bucket of wires that you have in your garage where you've put all your, old cords and they're all tangled around each other and you're trying to either detangle them or find the end of one. It's sort of like that, but just exponentially worse.

Yeah. I mean, it's a rat's nest. Yeah. You can think of the rat's nest of wires as like a, as a good analogy here.

Yeah.

Um, and you know, unlike the rat's nest where you can sort of follow it with your fingertips here, all you have to work with is vision.

Okay.

Um, and so it turns out that this is, this is the bottleneck. Um, the Wellcome Trust, you know, released a report a couple of years ago where they sort of gamed out, you know, what would it take to scale up electron microscopes to the mouse connectome. And the conclusion is, you know, at least at that time, um, tracing the neurons would be 98% of the cost of mapping the mouse connectome. Even for the fly, and, you know, not to detract in any way from the fly as a massive accomplishment and milestone, you know, First, you use automatic computer vision to do a first draft of tracing, but there's then a bunch of errors. And so then what the Princeton team did was they got together human annotators to manually correct the outputs of the computer vision models. And that was like 33 person-years of manual annotation and correction to get to that first fly connecting. So tracing is the bottleneck, and that's the problem that the field is currently grappling with.

Were those poor grad students or mechanical work or?

You know, I mean, it was a collection. You know, there were citizen scientists, there were, you know, professional annotators. You know, there exists an occupation which is, you know, a professional, you know, neuroscientific annotator. So, you know, and it gets the job done, but it's, you know, expensive and it requires human input. And so, you know, there are a couple of different ways to attack the problem. Many people are attacking it with AI. As you get better AI, then you can bring these costs down. But the cost of tracing has to fall by at least 100-fold for it to be an equal cost driver with imaging and other preprocessing and then tracing. And it needs to fall by 1,000-fold if you want to have it be a negligible cost driver. So this is the bottleneck that needs to be solved.

Okay. And then before we get to different approaches to, to solving this problem. Let's explain why you would even want this in the first place. And I almost feel like there's a pre-AI answer and possibly a post-AI answer, because people were chasing the connectome before this recent takeoff in AI technology. And most of the interest was that, look, the brain defines everything about us, we should understand how it works, we know precious little about its mechanics, so let's build this wiring diagram. Let's get a whole— let's just map a whole thing so we can, we can at least understand one brain and how everything is interconnected. And hopefully we will come to like new insights about how this organ works by doing this. It's a slightly controversial thought, I think, because there's some people that would argue just having that map doesn't do a lot for you and you're almost definitely doing this on a on a brain in a moment of time that is dead. You don't get to see what's happening actively in this brain. And so, so there's been some debate about that, whether this is like a worthwhile exercise. I tend to fall— it seems obvious to me that we should map this thing and understand how it works. And then I feel like— and I know you're into this and I want to get into it, you know, I think there's this post-AI answer where it's, it's like, well, okay, now we're making these very powerful artificial minds that are loosely based on this neuronal structure that we have in our heads. And if we want these two things to one day maybe be aligned and not maybe mimicking each other, but borrowing from each other enough that we're not, not like destroying the AI, isn't destroying us, there's some kind of humanness in it, you know, you would, you would want to replicate that by having this model of the brain somehow injected into these AI systems. But what's your, like, why did you get into this? Why do you think it's important?

Yeah, I mean, you know, if you're asking me why I got into that, you know, maybe we can just like go back to the beginning for just a second.

I thought because this could go back to your childhood or I don't know where this would start.

Yeah, I mean, you know, I'm an optimist. Like I grew up like on Golden Age science fiction. I was really like drawn to like positive visions of the future. You know, one day when I was like, like a kid, like the school bus was late and I was like very upset because I was going to miss Star Trek. So like, you know, it's, it's like that was like what shaped my worldview. And then, you know, later on I, you know, had conversations with my uncle. I realized that, you know, maybe I could do something. He was an engineer and he convinced me to be an engineer. And, you know, like I mentioned, Went to the University of Toronto to study there. And, you know, the reason I got into this was that, like, I read this essay. I was studying physics engineering science, which was a rebrand of engineering physics. And I read this essay by Feynman, which was like, it's really easy to solve these biological problems. If you can just look at the thing, then you'll be able to answer it. You'll be able to see, like, the sequence of bases in the chain. And I'm like, oh, actually, this is a great approach to biology. This is what we should be doing. We should be building microscopes. And this was really important to me because when I was 12, I started experiencing chronic pain and arthritis for the first time. And I spent my fair share of time at SickKids Hospital. I spent a lot of time with my physical therapist. And in the process, I learned a lot about chronic pain and learned that actually there's lots of people in the world who have brain disorders. Actually, 1 in 3 people in the world at some point will experience a brain disorder. Um, and, you know, there are different ways to try to attack that problem and try to reduce, you know, that level of suffering. Um, for me, in, you know, the late 2000s, it was like genomics. But over time, I think a lot of people have realized that there's this potential in neurotech to like really take on some of these, you know, like serious like, you know, mental health, brain disorders, and also, you know, some of the other stuff that you just mentioned. We'll get to that. And that is like, you know, that realization is what got me into Connectomics.

Okay.

And so like, it's like there are— and, you know, I conceptualized chronic pain as like a connectopathy, like a disorder of the wiring. And that's like, I think, a good way to think about some of the utility of being able to map these brain circuits at the level of synaptic connections. And, you know, we haven't actually defined connectome for the audience, but like the connectome is just like the set, the wiring diagram of the entire brain. And you can, you know, you can map an entire brain, you can map, you know, smaller connectomes within brains. But, you know, the idea is that if you can actually look at the changes in the wiring between like a healthy brain and a disordered brain, that will give you a new target that you'd be able to intervene with, with a drug or a brain-computer interface in order to then restore function and health to the patient.

So in that sense, because sometimes I find the comparison with the genome unsatisfying, but in that sense, somewhat similar in that you could take a healthy person's genome, somebody with a specific type of cancer, and if you overlay those two sequences, Not that it usually works out this way, but you happen to see this one bit of DNA that's quite different, and then that's maybe a place to go looking for an underlying cause.

Yeah, I mean, there's— the analogy works in a couple of places. So, you know, the first thing you need in genomics is one genome, and that helps you get oriented, right? And so, you know, as we get these connectomes, these reference connectomes, that will really help us understand, you know, what normal or healthy wiring looks like. But the, the real value of genomics was when sequencing got cheap enough that you were able to map many genomes across many different conditions, you know, disease models and developmental models and so on. And then you could start looking at the variation between individuals. And as you do that, you then can like track down what are the differences. And that creates, you know, hypotheses that you can then pursue, you know, with, with drugs. And it's going to be similar for connectomics as we can, as we develop better tools that reduce the cost. And this is why, you know, E11 and I am like relentlessly focused on cost. If we can make it so cheap that, you know, we can do many connectomes, then we'll be able to look at these different conditions and start to diagnose, you know, what are the connectopathies, which, you know, psychiatric disorders are correctly understood as connectopathies and which, you know, might not be.

I guess it's very difficult because we don't know what worms suffer from, but were we able to do anything along those lines? How many, I don't know how many worm connectomes, did we just do the one and then we sort of stopped, or do we have multiple to compare against?

Well, we do now, but there was, there was definitely, you know, after even the first one was actually sort of a composite of several different worms because it was difficult to get the slicing technology working well enough. To be able to do an entire worm in one go. But, you know, this is why it's important to be able to get connectomes from, you know, disease models themselves or developmental models, because, you know, the worm is just not a good model for, you know, the mental health disorders or other disorders that, like, people, you know, live with. So, you know, the worm and the fly, maybe not. But if you can start talking about the mouse, where we actually do have some good models for diseases, then, then you get really, you know, interesting. And, you know, there are other things where connectomics might be useful as well. Um, yeah, not just mice, but, you know, there are organoid models which are good models for disease. You know, there are some cases where you might be able to look at postmortem tissue and be able to look at that. So I think there are, you know, many different cases where once we can get connectomics working in mammals, then it's going to be very transformative for therapeutics.

And for, you know, like for the brain and like for the naysayers, the people who have been skeptical that we learn that much just because we get a map. It's kind of funny because I think when the genome was first sequenced, every— a lot of people thought, oh, we've done this once. And then next day you wake up and I mean, it really was— I lived through this. I mean, I'm simplifying things a little bit, but it was sort of like, oh, you do it and then we're going to just cure all kinds of stuff kind of the next 2, 3, 4 years just by looking at this one genome. And then it turned out to be exactly like you said, we need a lot of this. It's been, what, 25 years now of— and we have made progress, but still, I think tip of the iceberg of where all this could go. But for the people who— I remember when I was doing a lot of connectomic stories, there's definitely very bright people who were just like, yeah, this is like a— this is not the best way we should spend our money. But are they— do you think they were so solely in the 'we're just getting one brain' camp? Is that the central argument? Does anybody debate that if we had lots of brain images that we would learn something from it? That seems like a hard thing to go against to me.

I mean, there are always debates over resource allocation, but, you know, I think attacking, you know, 'we're gonna get one brain and stop' is a bit of a straw man. You know, it's really about bending the cost curve so it becomes so cheap to get this kind of data that then you don't need to have the debate. And so I think that's what's really interesting around some of the new approaches where we're focusing on reducing the cost because then we can short-circuit the debate. But I think you're right. I mean, you have to set expectations in the right place. It took genomics many years to develop drugs because it takes a long time to make these drugs and to and, and to understand, you know, what are good targets. I think a good question is, you know, what are the alternatives? Like, some of these disorders are tractable using genomics tools, but, you know, many others are like very polygenic. It's like when there are many genes— there's something like 30 genes, for instance, implicated in schizophrenia. Many of them are synaptic, but, but there's no one target that is like an obviously good target for intervention. And so I think it is worth asking, can we get something new? And can we test the hypothesis that maybe if you look at the circuits as the origin of the connectopathy, if it is a connectopathy, then you might be able to make more progress in this area?

I see. So some people would, would argue we solve these issues just through, through DNA, and that's going to be the, the major way to get at this stuff, and, and the circuitry isn't as revealing about some underlying cause for something.

Yeah, I think there are some, like Huntington's, where it's like a very clearly monogenic, there's a single gene, you know what the gene is, that's less likely to be impacted by connectomics. But for places where we don't really have good hypotheses or targets that have panned out anyway, having new tools in the toolbox I think is going to be really important.

Well, and there's this— I want to get into what E11 does and how you guys are going about this. So I run the risk of sending us on a slight tangent really quick, but there's a frustrating part of this to me, which is whether no matter which way you want to debate it, it just seems so obvious to me that we should know as much as possible about the thing that defines our reality. And I know people have spent centuries studying the brain. Sometimes it seems frustrating to me that sometimes it feels like second fiddle. I honestly get I understand why we are spending, you know, what must be now trillions of dollars creating artificial brains. However, um, because of all the economic incentives behind that, I just— I find it baffling that the world hasn't pumped more money into brain research historically when it seems so important.

Well, you know, I think Uh, there are, there's a lot of exciting science to do. And so, you know, I can understand why, uh, many different projects are, are, you know, competitive and interesting. But I think something that has changed recently that is like, you know, worth, uh, observing and then reevaluating connectomics in its light is the fact that neurotech is changing, right? You know, in the last few years, our ability to interface with the brain using BCIs, and you've covered a bunch of them recently. There's, there's this proliferation of techniques to get into the brain, and I think we're getting to the point now where our ability to interface with the brain might actually be outstripping our ability to know what we're interfacing with. And so now there's a good argument, I think, or, you know, another good argument for like why we want to map out these circuits, because now we have the tools where if we can— I talked about drugs, but, you know, brain-computer interfaces are, are, you know, good treatments for brain disorders as well. If we know where we should put the electrodes or like which circuits to stimulate, that could be very transformative.

Okay. Since day one, the Core Memory podcast has been supported by the fine people at E1 Ventures. They are a young and ambitious VC firm in Silicon Valley investing in young and ambitious companies and people. Thank you so much to E1 Ventures for all your support. Right, okay. And then even like some basic stuff, I remember sitting with Jeff Lichtman and I think he was pointing me to some part of an axon and there was this turbine-like shape and he was like, we just, we didn't even know these things existed before we did this. So I mean, that just jumped out at me as well. If there's parts of the brain we didn't even, we didn't even know this thing lurked and he found them all over the place on this tiny sample.

Oh, there's no doubt we're going to learn like a lot of new biology, just like, you know, along the way. Like if you look at this Microns project, you know, some of the most active users of, of that mouse cubic millimeter are cell biologists who are interested in glial cells. They're not, they're not actually there for the connectomics. They're there for, you know, the, the cell biology. So, you know, but this is, this happened in genomics as well. You know, you build DNA sequencers because you want to sequence DNA and then like fast forward 10 years and people are using them to sequence RNA for cell typing. They're using them to, like as diagnostic tools. They're even using them for data storage for like IT, right? Like how do you want to archive your data for the long run? You turn it into DNA and then bury it, right? So, um, I think, you know, these general purpose tools like close to the root of the tech tree, like genomics or genomics, I think, you know, have many unexpected, um, you know, applications as you make them accessible and cheap.

Okay, well, I will not embarrass myself fully trying to explain what E11 does, and thankfully I don't have to, but loosely, just to get us going, all the stuff we've been talking about, you guys have a new technique to try to image the brain that gets around some of these limitations that we've talked about. And like you mentioned, you're very focused on cost. And so this is just a major, major issue. And so, yeah, I mean, you guys are taking a swing that that other people, um, have not. Give us the, like, 30,000-foot view of, of how this technology works.

Yeah, so, you know, E11 Bio is a nonprofit research org. It's a focused research organization, and we can talk about that. And the purpose of E11 is to, you know, develop these new tools that will make it possible, tractable, and inexpensive to map mammalian connectomes. So, you know, we're really focused on this problem of how do we, how do we go to the next level? How do we create this like next generation sequencing moment for the mammalian brain as we go from the fly where we are today to where we want to go, which is the mouse and then ultimately, you know, human brains. And, you know, the idea that we are, our thesis is not a new idea. It's been around um, since like the mid-2000s. Um, the idea is to make every neuron in the brain a different color. And if you go back to your rat's nest of wires in the garage, you know, if they're all the same color, then it's going to be really hard to untangle them. But, you know, if they're different colors, then it's suddenly much easier to, you know, uh, untangle that, that big ball of wire.

So when we usually image these brains, I mean, you just— it's just tissue, it's all coming back. It's the Wires look the same.

Yeah, I mean, the state of the art is you use electron microscopy, which is typically black and white, um, and you know, you get these grayscale images out and then you have to trace neurons in, in grayscale. And when you use light microscopes and you use genetic engineering, you can make every neuron a different color by delivering combinations of fluorescent proteins to the neurons in the brain. And, you know, this idea was first implemented as Brainbow. This was actually, you know, another technology out of Jeff Lichtman's lab.

Great name.

And, you know, Google it. It's beautiful. I mean, it creates, you know, these beautiful images.

The images are— people should Google it. They're— it's awesome.

But, you know, the challenge with— and so this was one of the angles of attack on connectomics. But the challenge, and this is the same challenge that, you know, Cajal had back at like the dawn of neuroanatomy, is that you can't generally see the connections between neurons, or at least not in a, in a dense, you know, efficient way in light microscopes. And therefore, you know, you can't get the connectome out. And so you can make these neurons different colors, but, you know, the, the wiring, the wires themselves overlap because, you know, a wavelength of light is big. It's like 500 nanometers or more depending on where you are in the spectrum. And, you know, the cleft between neurons is tens of nanometers. So at a synapse. So, you know, it all blurred together and the state of the art ended up becoming electron microscopy because that was the only method with the resolution necessary to see these connections between cells. But then something changes. And, you know, I mentioned I did my PhD in Ed Boyden's lab and around 2015, Fei Chen and Paul Tilburg and Ed Boyden develop expansion microscopy. And that is, you know, one of the game changers. So I'm going to call it eXm, expansion microscopy. And you essentially embed the brain in a swellable polymer, and then you add water. It swells. It makes the brain physically bigger. You can make it, you know, these days you can make it 20 times bigger in every dimension. And that physically moves apart the neurons. And so now two wires that were overlapping and that you couldn't resolve in a light microscope now can be resolved.

So you're literally taking one of these samples and stretching it out. And I was just with Ed Boyden. I didn't know this until— because my— immediately I was like, oh, well, you're going to get some crazy distortion when you start doing that. But I mean, he was telling me it's like still— I mean, they're finding it's like 99% holding, holding scale and everything.

It, it preserves the structure remarkably well, even version 1.0, like preserves the structure remarkably well. You're doing a little bit of a balancing act. You're digesting some of the structural proteins enough so that instead of stretching it, you're just moving it apart. You're moving it apart isotropically, which just means everything, all the biomolecules move the same distance apart from each other. And, you know, over the last 10 years, this has been very, you know, effectively optimized. And so now it really sort of preserves this 3D structure, which is very important.

It's crazy, right?

It's, it's—

so you were in the lab when this was— were you in the lab when this was taking place or being discovered?

Just after.

Okay, okay. I mean, people must have been freaking out.

Uh, I mean, it, it launched a field.

Yeah, yeah, yeah. I mean, just, it boggles my mind that this works. But yeah, I mean, Ed described it as like the same— I don't know if he was saying it was the same material as what goes in a baby diaper, or—

yeah, or like, you know, those dinosaurs that you would put in the bathtub that would expand, right? I actually, I'm not sure that that's okay, but it's the same concept. It's like, you know, you add water and then, you know, the the material expands.

Okay. So you're studying at MIT, you're seeing the development of this technology take hold, and I assume this starts to get you thinking about—

Yeah. Yeah. So, yeah. So I'm at, you know, I started in the lab. I'm very interested in, you know, how do we, how do we apply this new tool expansion to connectomics? And I'm also inspired by this essay that, you know, Tony Zador wrote in like 2012 called Sequencing the Connectome. And it was, you know, an articulation of this idea that, you know, if you can make neurons different colors, or even better, give them different, you know, molecular ID tags, you can distinguish them from each other. And you can use many of the tools that we have in sequencing and then apply them in order to, you know, map the brain more effectively by using these— they're called cell barcodes. Barcode is just a term of art. It just means, you know, the cell now has a molecular fingerprint.

Like a tag.

Yeah. And you use the tag to correct errors in the tracing. And so I spent a couple years working on that. And, you know, there are some challenges. You know, it's difficult to get the barcodes to express throughout the cell uniformly. You know, like when you implement them in RNA, sometimes they don't make it down the axon. Other times they don't make it up the dendrites. These are the two different branches of the neuron. And so this is a challenge. So the idea is nice, but the implementation is difficult. This brainbow technique is another thing that I look at, but the challenge with brainbow, even when combined with expansion microscopy, is that there's only about 100 or so distinguishable hues in brainbow. You have red, green, and blue fluorescent proteins, and they, you know, can be expressed at different levels. But, you know, 100 is much, much less than the 100 million neurons in a mouse brain, right? So it's, it's not enough. And After a couple of years, I'm like, man, this is hard. This is going to need a team of 10 or 15 people and 5 years. I'm not going to be able to do this as a graduate student on my own. And I spent some time working on comparatively simple stuff, sequencing genomes inside intact cells with microscopes, looking at the folds of the 3D structure of the genome. This is actually useful tech that we're now implementing at E11, but I had to take a break from the problem. And around like 2019, like Adam Marblestone, Sam Rodriguez, and I all overlapped in Ed's lab. And, you know, we asked this question, like, wait a second, like, what if we did have, you know, an organization where, you know, you had 10 or 15 people working on this? Like, would that move the needle? And actually, that was how we developed this, uh, initial notion of a focused research organization, which is, you know, in a nutshell, it's just if you took a vertically integrated, like, Series A biotech startup and a really important nonprofit public benefit science goal, and you put them in a blender, what would you get out? You would get out something like E11.

Um, yeah, which we could talk about. Froze, I'm going to borrow you to talk about brains, um, but I am going to direct people to an episode that we did with, uh, Adam Marblestone and Anastasia Gomick, who run Convergent Research and It's episode 22 and you can learn all about focused research organizations and sort of the reasoning behind this. But yeah, so yeah, I mean, it fits into this category of, of somebody maybe too expensive for a university to go after, might not have totally clear kind of like product on the other side for a venture capitalist to definitely fund, but is extremely important to do and somebody should put money behind it.

Yeah, I mean, you know, to be, you know, put a little bit of a spin on it, uh, my spin on it, it's like the project is pre-commercial. It doesn't make sense to do as a startup yet, but it does require more coordination and more vertical integration than you can typically find in academia. And we're not the first people to do this, right? Like the Allen Institute has, you know, a long track record of doing team science projects. Janelia, the Broad, they all have team science projects. But I think the one thing that we really noticed is that startup culture is a little bit different than, you know, nonprofit academic institute culture. And, you know, the FROs are typically, you know, focused research organizations are typically time-bound. And so that creates a real sense of urgency to get the job done.

Like 5 years-ish. And yeah, and there's this pressure to keep it small, tight, focused, hardworking, some of that like startup buzz to charge after this goal. And then, and then you mentioned, I mean, other people have done it, but then Some folks like Eric Schmidt and James Fickle come to mind as people who are willing to put money behind a few of these efforts.

Yeah, so E11 Bio was anchored by Eric and Wendy Schmidt very generously. And since then, many others have understood what we're doing and have wanted to support us. And so James Fickle and Protocol Labs and the Bechtel Foundation and the Hearst Foundations and others are now very interested in the, the project that we are pursuing, which is to really try to solve some of these fundamental problems around how do we make neurons different colors, enough different colors that you can actually trace them, and then apply that notion to make the cost of connectomics much cheaper. And then there's one other thing that is, you know, fairly exciting, which is when you're using light microscopes, you can use these different colors to, you know, identify different neurons and distinguish them from each other, but you can also use different wavelengths and different colors to look at other things in the sample. And so for the first time, we can actually bring molecular information, you know, biomarkers or synaptic markers or things that tell you cell type into Connectomics assays. And so that's, that's like a new capability that hasn't existed before now.

So if there's people who are keeping up, I'm sure recently they're scratching their head a little bit, which is that you You said we were trying to do 100 million neurons in a mouse brain. You were running out of colors on these other techniques after like 100, I guess. Um, so how on earth do you possibly even begin to mark up 100 million things differently?

Yeah, so for people who are interested, we just dropped our first preprint, um, so E11 uh, you can find a link on our website or you can check it out on BioRxiv.

Yeah, yeah, open it up too. Yep.

Describes, you know, the methods that we're using in order to make this work. But in short, you know, we've extended the number of possible colors by splitting up the assay into sequential cycles. And this is the same idea that, you know, how DNA sequencers work actually in, in many ways. So When you're sequencing DNA, you know, ideally you would want like one color per base or something like that, but there's no way you can split up the visible spectrum in that way. And so what you do is you have four colors, A, C, T, or G, and you look at, you know, the next base in the chain and it'll light up as one of those four colors. And then you'll extinguish that and then you'll take another step down the chain and you'll look, is it red, green, blue, or yellow? And you'll, and you'll do that sequentially. And we're doing something similar. So, you know, we have these different proteins that we can deliver to the brain that are like fluorescent proteins that can be read out in sets, say 5 at a time or 4 at a time. And so then we can, for the first set of those proteins, say, you know, is it red, green, blue, or yellow? And then we can extinguish that and then we can read out the second set. And so we can do that in a sequential way. And each time you do this, you exponentially increase the number of possible combinations. And so it's the difference between colors and color codes. And so we're reading out color codes, and what we've found is that we can make, at this point in time, 18 of these different, uh, combination— or like, like building blocks that can be combined in combinatorial ways. So if there's 18 different proteins, there's 2 to the 18 different possible combinations, which is, you 260-something thousand combinations. So this is, you know, um, like hundreds of times or, or even thousands of times more than what people have been able to do previously.

And how do you— when you talk about the, the barcoding, the tagging, um, and putting these proteins in the brain, if I remember right, you're using a virus more or less, right? I mean, you're injecting brain tissue with the virus which is carrying the protein to an individual neuron. Is that the right way to think about it?

Yeah, we as a field have gotten remarkably good at using viruses to deliver genes to the mouse brain. And, you know, these are called adeno-associated viruses, AAVs. They're, they're very safe. They're just a workhorse in biology these days. People use them, um, to deliver genes to the brain. They don't replicate. They just sort of hang out after they deliver the gene of interest. And there has been a lot of work to, you know, develop these viruses so they can be administered systemically, so they can cross the blood-brain barrier, so they can, you know, get into many different neurons in the brain. And so this is really, you know, a, a frontier in how we can genetically engineer organisms.

And so, okay, so you take— you've got a mouse brain You put in the virus. Are you putting in many viruses or one?

It's a pool of viruses.

A pool of viruses.

You have, you have, you know, I mentioned there's 18 different variants in this publication that you can check out. And, you know, you make a pool where, you know, each of them is present, you know, like at an equal ratio. And then when you deliver them to the brain, the barcodes are delivered essentially randomly. Each cell will just get some subset of that total. Set of different viruses. So, you know, cell 1 might get, you know, variants 1, 2, and 4, and, you know, now the barcode is 10110000 all the way up to 18, and, and so on. Um, and so you, you basically use like random— call it stochastic co-infection, where it just means that each cell gets infected multiple times to deliver the barcode randomly.

And so some neuron is absorbing a different combination of the proteins from its neighbor. I mean, this is what's helping you distinguish because each neuron is going to pick up something different in this process. And so then because they fluoresce and give off this different color, instead of getting this black and white map where everything looks the same, now you're starting to see differentiation. Between, um, between the different neurons in the brain. And, and the, the, and you can like follow a blue wire and a green wire and essentially.

And so we, we apply this, uh, this is our first pilot study. We apply it in, in mouse hippocampus, which is, you know, the hippocampus is very important for learning and for memory. Um, and, you know, I should say that this is a very collaborative study. One that actually, we call it PRISM. One of the most fun things about, you know, developing Prism was that it was this long-term collaboration that started in Ed Boyden's lab. Sam Rodriguez, Jurgen Kornfeld, who's now at Max Planck, Sam went to the CRIC and is now at Futurehouse in San Francisco. And Bobayanne in Ed's lab and I all worked together on the initial instantiation of this. And then many others have brought their intellectual lineage to it. So Kathleen Leeper and Johan Winnebst and Sven Truckenbrod and Arla Sheridan are all new members of B11 who have brought different aspects of barcoding or expansion microscopy or of machine learning to this problem. But we all got together to perform this initial pilot study where we delivered these barcodes to the mouse hippocampus and then applied and optimized the expansion microscopy protocols in combination with these barcodes because you have to have both of them working together. In order to actually be able to, you know, look at these neurons in a way that accommodates the fact that light is big and neurons are small. So, um, so that required a lot of co-optimization. And then we needed new machine learning methods and AI methods that would actually take this multicolor information and use it automatically to correct errors. And we can do this in, you know, two different ways. One is when you're actually, you know, drawing the boundaries around the shapes and identifying the boundaries. And then, you know, there's a— I mentioned earlier that when you are correcting the outputs of like normal computer vision models for Connectomics, you have humans who are proofreading the outputs and correcting them. And so we actually can use this color information again in an automatic proofreading step where there's, you know, a model that is able to, um, or an algorithm that is able to identify, uh, when errors have occurred and then use the barcodes to correct those errors.

Okay. So many questions. Okay. So you mentioned, okay, in this, in this, in this paper, you're talking about part of a mouse brain. You're talking about the levels you've been able to get to, 2 to the 18th, I think. Was that right?

Yeah, we can encode 2 to the 18th combinations.

Right. Okay. And then if you wanted to do an entire mouse brain, it had the numbers have to go up, right?

So this is a great question. So, um, there's a couple of different levers. So one is to increase the number of possible combinations. And with each new protein variant that we discover— and we're running a pipeline at E11 to discover these variants— with each new variant, we double the number of possible barcodes. So this grows very exponentially. So with, you know, 18, you have 262,000. With 30, you have about a billion We've been finding these at a rate of like, you know, one per couple months. So, you know, uh, this has the potential to grow quite quickly. But I think there's another question, which is how many do you need? Because we're not throwing out the shape information as well. We're still using the shapes of the neurons and like other methods from, uh, electron microscopy, connectomics, computer vision in combination with these barcodes. And so you can still trace the shapes as well. So I think we, You know, how many barcodes do you need? How much shape information do you need is like a research problem that we're pursuing right now. But what we have found is that when you, you know, in this particular pilot study, when you apply these barcodes, we improve the accuracy of tracing the neurons by almost an order of magnitude. And we like see a clear path to extending that further to that 100x that I'm talking about. So, you know, we're, We're, you know, one log there. We need to get a second log there. Um, and when we, when we do, and, uh, when we combine that with other methods that are also, you know, being developed right now for EM, for electron microscopy, I think, you know, we'll be really making a dent in this, you know, big bottleneck, which is the cost of tracing neurons.

Yeah. So, I mean, there's, as far as I understand it, I mean, and you mentioned this one study that have been done, it's just people are like, oh, well, we're not getting to a human connectome like anytime soon unless somebody figures something out or has infinite money. What, what is like the best, um, what's the most public date or, or goal for the, the mouse brain, a whole mouse brain at the moment?

I, I think that's an interesting topic of discussion that, you know, different people in the field have different reference points for. I think most people would be surprised if it showed up sooner than 5 years from now.

Okay.

And also surprised if it took more than 15. But, you know, I think people were also very surprised in genomics when, you know, the cost of sequencing fell 10,000-fold in 5 years. And so when new technologies come along, it's sometimes difficult to predict exactly how quickly they'll be able to, you know, reshape the landscape. Yeah.

And then, and then granted Your company's dedicated to this and you've dedicated this portion of your life to this. So possibly some bias, but you know, given what you've seen with this research and the breakthroughs that you guys have had, you know, how has that altered your thinking? What is your like most optimistic case for your technology being able to do a whole mouse brain? Or do we still not know if that's feasible?

I think there's a lot of uncertainty here. Part of the fun of running a focused research organization is we get to test these hypotheses very empirically. We get to go and find out how good is the technology going to be. I would say a lot of the hard problem has been just figuring out how to do this at all. I would be less I would say that like 5 to 15 years seems right for any technology. But if we can make that happen in more like 5 years rather than in 15, that would be really transformative for all of the people who would benefit from having Connectome sooner. And there's another thing. I mean, we're focused, we've been talking a lot about what would it take to do an entire brain, but the medical value of being able to do parts of the brain or, you know, many subvolumes of brain is really important. For example, you know, people are working on cell therapies where, you know, in cases of, you know, Parkinson's or traumatic brain injury or other— or stroke, you need to be able to put stem cells into the brain that then, you know, integrate and regrow and connect, you know, back up in many of the same ways that you would want the healthy cells. And the nice thing about the cell therapies is you have a lot of control over how you engineer those cells, you can actually conceive of putting things like barcodes into those neurons or into those stem cells and then, you know, deliver them to the brain. And so there's, there's a lot to do when you have multiple colors. You can see how those cell therapies are integrating and connecting and rewiring using those color— that color information. You can use that to distinguish it from, you know, the, like, the, the, the host tissue. Um, and so this could be, you know, a, a medical application where it could happen a lot sooner than being able to map the entire brain. And so I think we, we should think about, you know, different ways that, you know, connectomics technology can be a little unbundled from, you know, the, the main push, which is to get the whole brain.

Don't cop out on me, Andrew. I want the whole human brain.

Oh, I mean, you know, the mission of the company is to make that happen.

All right, I'm just— I'm looking up a message at the moment from Sebastian Sung, who worked, um, I think it's fair to say he was like a central figure on the Fly Connectome. He wrote the book Connectome.

It's a great book.

And then, but the Fly Project came after, came after the book and, and was a, was a pretty central figure. I don't want to get this wrong. So I'm just looking it up because I, if I remember correctly, and I will, I'll launch my question and then I will, I'll like dig in and see if I'm, if I'm right or not. But I think coming out of the fly, his argument to me was that the AI software had, had been surprisingly good, um, and incredibly helpful and that he thought we, we get over a lot of these cost and technical issues just through the, the steady sometimes exponential improvement in AI software.

Yeah, I would say, you know, it was surprising in many ways how little human proofreading was needed to get, you know, that FlyBrain only 33 person-years versus, you know, the many thousands that, you know, would have been required before advances in AI. So that sort of brings us to the other challenge, which is Attacking the cost of tracing is the biggest bottleneck. We need just many different methods. We need to be able to combine improvements in AI, like the kind that Sebastian is talking about, with other modalities, like when we're literally injecting more information into the brain that will just make it easier and more effective. Because we want to do many brains, right? Why would you say no to another factor of 10 reduction in cost, for example? For example. But the other really challenging thing, which we haven't touched on, is that slicing the brain into these thin sections or ablating it is actually a very precision engineering process that is challenging to do in a scalable way thousands and thousands of times in a row. And the Microns Project, which was this cubic millimeter of mouse brain tissue, You know, they, they took a subvolume of the mouse brain, a cubic millimeter, and then sliced it into, you know, 28,000 individual thin sections. And this was like an incredible engineering achievement. And also the diamond knife jammed about, you know, 40% of the way through. And so what we actually have is a subvolume, which is, you know, like 400 microns thick and a subvolume, which is like 600 microns thick. And we don't actually have the wires across. That gap of 6 slices in the middle. It's, you know, you just can't trace it if there's nothing there. And so I think, you know, one of the additional challenges with this approach to slicing is that, you know, if you're going to do one mouse brain, you need to do that 28,000-slice problem 500 times for all of the 1 cubic millimeter subvolumes of the mouse brain. And the chances of making an error are, you know, still high. And so one of them—

yeah, well, because these aren't machines you just like go, go buy, right? I mean, they're sort of like hand-built, more or less.

I mean, took months to do the slicing. Yeah. So one of the interesting things about this expansion microscopy approach is that you can work with much thicker sections. Instead of working with 50-nanometer sections, you can work with 50-micron sections. So the sections are 1,000 times bigger. And so now when you're talking about a cubic millimeter, You're talking about maybe 20 sections rather than 28,000. So, you know, that's, that's the sort of, uh, difference where, um, I think new technology is still going to be really helpful. You know, we, because we're asking this question not necessarily like, what is the, the method that will get the first connectome, but like, what is the technology that will give us that next generation sequencing, like bend the cost curve moment? That will make it so cheap that, you know, anyone can do it. And I think, you know, that's, that's one of the really interesting questions in connectomics.

Okay. Well, you get off the hook a little bit here. In my messages with Sebastian from— this is, this is from months ago, I think, when the Fly came out. He says, I have a somewhat different take regarding the value proposition of E11 and its competitors, but I've got to run now. And we'll answer your questions later. And he, he, he didn't. Do you have competitors? He lists, he lists, um, he lists Penluminate.

Oh, I love Penluminate.

Okay, well, I mean, he just sort of says they're impressive. I, I, I don't know if competitor is fully fair. And then he links me to a paper by a researcher named Tavakoli.

Yeah.

Okay. Yeah. I don't know. Are there, are there different techniques?

So like I said, you know, expansion microscopy started a field, right? So, you know, many, like it's been around for 10 years now. And in the recent past, there has been some pretty exciting advances where you can make these expansion microscopy samples look more and more like electron microscopy. So you can, you know, you can expand it more, you know, in our PRISM study, which we just released, you know, we expand it 5-fold. If you expand it, you know, 16 or 20-fold, you can actually see the, you know, it looks like EM essentially. And so, you know, this was pioneered by a couple of people. Hans M. Saad, who founded PanLuminate, developed this Pan-EXT. I think that's what it's called, Pan-Stain called PAN because it stains all of the proteins, which stains the cell morphology in this way and it looks like EM. Hugo Damstra at E11 did something similar with T-REx. It's another expansion technique. I'm getting into the weeds here. But one of the most interesting things was out of IST Austria, Mo Takavoli last year developed this technique called LICCON, which took it a step further and he worked with Viren Jain's group, um, and this was in the Donzel lab. So Motakaboli and Johann Donzel worked with Viren Jain at Google to apply the EM tools to these morphology stains in order to actually get the connectome out. And actually, that was one of the missing pieces, you know, before that people hadn't managed to actually get connectomes out of expansion microscopy. So that's actually one of the why now moments. And, you know, this is one of the things that we have integrated into our pipeline. And our thesis is that all of the problems, the challenges that EM faces with tracing these neurons through black and white images are also challenges that these pan stains have. And you're going to have similar error rates and similar challenges tracing neurons. And so if you can bring in this additional color information, it's only going to help. And so it's like a yes and.

Yeah. There's something— Okay. Well, I feel like we've— Sebastian, ordered me to, uh, you know, at least, at least talk about some of these, the other folks working on some parts of this. So we've done that. Um, the— when I was at— I don't know if I want to get Jeff Lichtman in trouble or not. I don't think I will. I mean, I, I found a strange part of this was that the amount of compute is so high that Harvard was like basically in line for to get space on some Google cluster somewhere to do all the calculations that they needed and the imaging, you know, to stitch all these, these things together. And yeah, I guess I was sort of surprised that an institution of even Harvard had to— yeah, I was like, well, how do you, how do you move up and down in the line? It was like, well, you got to kind of like call in favors. You got to know somebody. And, and I found it interesting that this is such a big computing problem that you're like horse trading to get space on these supercomputers to do this. Work, and that also seemed like a massive bottleneck to me and just a strange situation in some ways.

I mean, I think it's not that strange in the sense that when you are, you know, pioneering for the, you know, and trying a lot of the stuff for the first time, you haven't had the time to optimize the methods. And so, you know, a lot of the more recent approaches to machine learning for neuron tracing were pioneered by, you know, uh, people who had access to Google resources because just getting it to work at all is step 1 and then getting it to work cheaply is, you know, steps 2 through n. Yeah. And, you know, uh, folks like, uh, like Arla Sheridan at E11 and, and Jan Funke at Genelia have, you know, been working on, uh, you know, different versions that, you know, maybe aren't quite as accurate, but, you know, are much more computationally efficient. I think it's just in general, there's this give and take where people will come up with a new method that breaks new ground and then you have additional work to optimize it and make it cheaper and make it more effective. In that sense, I think it's not too surprising, but it just underscores this point that we need to make connectomics much cheaper and develop new tools and orient our research towards making it really cost-effective.

Yeah. Okay. All right. For this last section, let's go philosophical for a second. Or I just want to ask you my big picture questions. Sometimes I run across people like I talked about on the AI front that think if we, the more we know about the human brain, the better we can inform our AI models. Like I always like get the gist of it, but I sometimes stumble upon the mechanics of how that works. I know another group that is trying to stick a ton of probes in a brain to create like a digital, a very complete digital model of the brain. And again, the hope is some kind of alignment type thing. Yeah, I stumble when I, I mean, again, I get the gist, right? It's like, okay, if we know more about what, if we find consciousness in here somewhere, or If we know more about what makes the human brain special and we can somehow inject that into these AI models that are very, at least for me, like, although over time I think I'm changing my opinion, but roughly based on the structure of a human brain in the sense of just connecting lots of computing elements. Um, yeah, but I, you know, I saw you talking about the subject and you have thoughts on this. Can you, I mean, can you explain? Where you might see this going and if you think this knowledge does have some application in AI.

Yeah, I mean, I think there's a lot of excitement around this, you know, like nascent field of neuro-AI, which is, you know, how can AI learn from neuroscience? How can neuroscience learn from AI? We know from history that, you know, there has been an interplay between neuroscience and AI. The original artificial neural networks that we use day in, day out these days were named because they resembled the brain. The convolutional neural networks that drove the deep learning revolution were very similar to, or organized in very similar ways to the visual cortex and the visual system. So there's definitely these parallels that exist, but it's a little hard to say that, you know, those advances in AI were caused by neuroscience or, you know, it directly led to the other, but it could have, you know, there's definitely, you know, worlds in which that could have happened. And so I think one of the most interesting questions is if you have a field which can very systematically map brain circuits, mine them, look for patterns and motifs and, you know, like changes between different organisms. Can you drive cross-pollination to, you know, speed up the development of new or safer AI tools that are based on the brain? And can you actually, you know, do that in an intentional and systematic way rather than in a, you know, serendipitous or loosely inspired way? And so I think like that's a, that's a thesis that needs to be tested out. But, you know, one of the challenges for a long time has been we just haven't been able to have the data. And so, you know, it's been hard to drive that cross-pollination. So what if you could do that in a systematic way? That's really interesting to me.

Like, what have you made— going back to that original thing, I mean, yes, they called the original neural networks neural networks and always talked about, you know, I don't know, for a long time, like I said, I found it unsatisfying because 'cause it was just like, ah, you just stitched a bunch of computing elements together and you, you know, it's like, okay, it's like mimics the brain, but in, in the most basic possible way. And then, and then as we moved along, I was like, oh God, well, I mean, apparently if you put enough complexity in this thing, it starts to do pretty interesting stuff that we can't quite explain. And it does seem to, whether you want to call it a parrot or not, I mean, it's, it's, it's clear, it's like, thinking something on some level, you know, and so that makes me feel like, okay, this seems to be some kind of precursor or related to the brain.

Yeah.

I mean, where do you, what have you made of AI models over the last 5 years, just given your background in neuroscience and, and, um, I don't know. I mean, it's a giant question, but I'm just, I'm always curious how you read these things.

Yeah, I mean, they've been making phenomenal progress. I think anybody who says they knew that it was coming would be either directly working on it or making it up after the fact. I did not expect that the AI models would advance as quickly as they did, and I updated on that. I think there's a lot of things that they do well. There are some things that are still really hard, right? Continuous learning is not really a capability that these models have. They have to be trained and then you have a snapshot. And so, you know, I think there's still a lot to be learned from what biological systems and living systems do that current models can't. And indeed, one of those things is, you know, we generally are are pretty well aligned with each other as humans. You know, we like cooperate. We're generally prosocial. We generally are altruistic. And, you know, there exists like other organisms which that is not true for. And so, you know, what is special about, you know, the wiring of humans or other, you know, prosocial animals that is different than, you know, antisocial animals. And if we can figure that out, if we can figure out like What is special about those connectomes that could, in an intentional way rather than a serendipitous or loosely, a loose analogy way, inspire a new set of AI safety work?

Why do you— you mentioned brain-computer interfaces before and how exciting they are, and obviously tons of money has been pumped into whatever it is, 12 to 20 startups all approaching BCIs in different ways. Why do you choose to work on the problem you're working on as opposed to getting into that field where you can be quite directly interfacing with and manipulating and learning from brains?

Well, like I said, I think there is a— I mean, there's a sense of urgency here around like our ability to interface with the brain is outstripping our ability to know what we're interfacing with. And so, you know, there's a lot of opportunity cost, I think, in not having the maps of the brain that would be useful to inform, you know, where we should be plugging in our BCIs or how we should be, you know, stimulating particular circuits to restore function. And, you know, this is, this is a kind of a public good. It's a little hard to, to commercialize that, but it is possible to, you know, create these tools and make that available, um, to many different, you know, BCI companies, um, and therefore lift all of the boats, as opposed to, you know, throwing my hat into the ring for one more competing BCI company. And so I think, I think this is really like an important thing for humanity, you know.

So you're working on like a foundational technology. I mean, yeah, and, and that's exciting. Um, what do you like on the The more I cover BCI, the more I get interested and fascinated. And then sometimes I feel I see its limitations because I hang out with a lot of the patients that get these. And so, you know, I've covered tech long enough not to judge things in just a small period of time. I mean, all these developments tend— the best ones take decades, but But I get a little frustrated with the limitations. I think most people have no idea that these implants do go into this tiny, tiny, tiny, tiny part of one part of the brain. And it's very limited functionality that, you know, we're like kind of good at reading out data. But as far as I'm aware, barely not so good at reading in, although, you know, we can, Ed Boyden has developed some techniques to stimulate the brain and, and things like that, but still quite limited. Yeah. So, and then when I hear Elon or someone talking about, you know, elective consumer brain implants and, and going all Matrix and downloading kung fu, I don't know, where, where are you on this? I mean, do you— is that like obvious to you that that's gonna happen? Do you think that's many years away?

Well, I mean you know, like I mentioned, like, I, I grew up with chronic pain. I have chronic pain. I see this from, you know, I don't know if a BCI is gonna, is gonna help me particularly, but I, it's like 1 in 3 people are gonna have a brain disorder in some time, some time in their lives. And so like that, that medical need is huge. Like it's really right there, right in front of us. And so like maybe we will get to that point as these technologies get good. Um, but like the medical need is right there.

Okay. So you're not going to sci-fi with me. Would you, you have pain from arthritis? Um, I can ask.

I do have arthritis. Um, you know, chronic pain is a little bit hard to track down. There are medicines that help with arthritis, but don't necessarily help, you know, when, when there are changes in brain circuitry, right? You know, there's this expression, pain is in the brain. And yeah, like that's not, that doesn't mean it's in your head. It means it's like literally changes in wiring. Yeah.

I get cluster headaches. Have you ever heard of those?

Yeah.

Yeah. It's like, not fun, dude. I'm in the middle of a spell right now. It is like so debilitating and like the pain is just unlike anything I think most people feel in life. It's, yeah, I would give anything to turn them off. And there's, there's like pretty much zero medicine. You have to, you have to like immediately get access to some high-grade oxygen if you want to have any hope of this. So it's quite difficult. So I'm with you on that. But you, come on, you watch Star Trek sci-fi with me. When you're, when you're, when you're just letting your, your mind go, you don't have any, you don't have any inklings, any thoughts on where the implant stuff is heading in like 10, 20 years?

You know, I think it's, I think it's a little hard to predict. Yeah. I wouldn't bet against it, right? You know, whenever there are new technologies, like, they will be used for medical purposes. But, you know, people don't just use, you know, GLP-1 agonists for medical purposes, right? Like, they use it because they want to. And so I think there are many people who, you know, would want to take advantage of BCIs that are very powerful. And we will see what happens and we'll see, you know, what the regulatory regime looks like and so on. But You know, many of these technologies are dual use in that sense.

You said you're an optimist. I mean, are you, are you trans— are you kind of like pro-transhumanist?

I think that there are a couple, you know, notions from transhumanism in that movement which are, which are interesting and, you know, make for really good science fiction. Um, I'm an optimist, but I also very practically, you know, build technologies, and I'm an engineer. And so You know, I think there's, there's, you know, interesting like sources of inspiration from, from those stories.

Yeah.

But we, we need to build technologies that help people right now. Yeah.

Yeah. I mean, half the time I spent so much time in AI, BCI land, all this other stuff, you know, half the days I'm so excited and then half the days I'm like, oh my God, the world is getting so, so insane so quickly.

And you know, sometimes it, it's like you're living in a sci-fi future already. Right. So. Yeah.

Okay. Um, okay. I'm just keeping— I'm being cognizant of time. Um, so the— what— taking everything you said, I mean, the thought that we will get a human connectome, just— it's just too TBD. It's like you said, we— this would be— realistically, it's dependent on a breakthrough of some sort along the lines of a sequencing breakthrough?

Yeah, I think the breakthroughs are, are going to be very important. You know, there was, there was this next-generation sequencing breakthrough that got us to the point where, you know, genomes were cheap enough that, you know, anyone could get their genome sequenced. The technology is still going. There's now a third generation of sequencers that are based on nanopores that are, you know, potentially going to be even cheaper. And so I don't know if the technology that gets us the mouse brain is going to be the same technology that then takes us all the way to the human brain, like whether you ride that cost curve all the way there or whether there needs to be, you know, another breakthrough in the meantime. But what I do think is we're going to learn a lot about that as we start applying these new tools that are based on, you know, barcoding and expansion microscopy and new AI methods to the problem. We'll be able to actually very quantitatively say How much better are we getting? And, you know, then you'll be able to plot some data points on a graph and extrapolate and, you know, draw some conclusions around that.

Yeah.

Um, and so I just want to say, you know, I'm the, the PRISM study was led by Rosa Park and Arla Sheridan. There's some amazing scientists, you know, definitely if you take a look at that, you'll be able to, you know, draw some of your own conclusions and, and also, you know, maybe email them if you have questions.

Okay. Okay. And then by the time this episode comes out, your, the paper will be Prizit, I will put a link in. You can go to E11 Bio. Um, like you mentioned one other spot where it will—

it'll be on Bio Archive.

On Bio Archive, right? Of course. And then, um, I've been to your lab in Alameda. It's very cool. You have a cool collection of people, kind of like, um, it's an interesting collection of people, you know, like you mentioned a little bit earlier, um, different talents in, in quite—

it's a really interdisciplinary team, you know, and it's a fun project in the sense that we are bringing together, you know, expertise from genomics and from, you know, imaging and material science and from AI and vertically integrating them and getting them to, you know, work really tightly together on this problem. So, you know, I think that that is really like a wonderful group of people.

And how many years are you into this?

So we've been doing this for about 3 years.

3. And so are you on the clock? Do you have 2 left or is there flexibility?

You know, there's a little bit of flexibility, but I think the sense of urgency is really strong. You know, we've brought our Prism platform from zero to one. So, you know, when we started, like, these ideas were out there, but nobody had, like, integrated them into a single package where, you know, you can use barcodes and expansion microscopy to trace neurons much more accurately than you can using, you know, single-color fluorescent proteins alone. So like, that's, that's a big step. And then the question now is like, how well does that scale as we go from, you know, 2 to n?

So in an ideal world, over the next 2 years, you would come out of that and, and you would show that it can scale and you would have the techniques to do that. And you guys would, it's fairly, it'd be open source to some degree, right?

So yeah, I mean, we want to make the methods available for people to build on. Especially for non-commercial research use. And so, you know, plasmids will be on Addgene, the methods will be on BioRxiv. We have software libraries that, you know, people can use, and those are open source.

Okay. And since FROs do tend to have this fixed timescale, I don't think I'm out of line asking this question. I mean, do you have any idea what you would like to do next?

Well, I mean, I think everyone who is part of E11 is, is very mission-driven. And that's like, you know, one of why you, why you start a new org, right? Because you want to get like-minded people together who share a sense of purpose and a sense of mission. And so, you know, we're building this technology because we want to apply this technology, right? Like if we have found an angle of attack that can reduce the cost of connectomics by 100x, then, you know, we want to be able to use that to get connectomes. And to continue to push that number further, you know, we don't want to stop there. So, um, you know, we're, we're connectomics people.

And if you, if you get to that point, the commercial end of that would be finding cures to illnesses.

It might be. I mean, I think, I think there's a question that still needs to be answered about, you know, is, is what are the commercial applications of connectomics for therapeutics, for for BCI, for neuro AI. All of this is sort of an open research question and we want to help find out.

Okay. All right. Well, Andrew Payne, thank you so much for coming in. It's fascinating work, like among the most fascinating, I think, around. And it's been, it's really cool to have gotten to know you and to hang out a little. We were at a dinner, we got to shoot the shit. It was good.

Yeah.

Yeah. Thank you so much.

Thanks for having me on the show, Ashlee. This was fun.

The Core Memory Podcast is hosted by me, Ashlee Vance. It is produced by David Nicholson and me. Our theme song is by James Mercer and Jon Sortland, and the show is edited by the Jon Sortland. Thanks as always to Brex and A1 Ventures for making this possible. Please visit our Substack, YouTube, and podcast channels to get more of what Core Memory makes. Thanks, y'all.