Latest / From First Principles / Can AI Help Wake Coma Patients? The Science of Consciousness (EP 35)

Transcript

Can AI Help Wake Coma Patients? The Science of Consciousness (EP 35) · From First Principles

0:00
There's a question that haunts them every time, which is, is my loved one still there?
0:05
And this question is one of the hardest problems in all of neuroscience, right?
0:10
What is consciousness?
0:11
He used AI to reverse engineer the biological mechanism of unconsciousness in coma patients.
0:19
AI has rediscovered known features of unconsciousness.
0:22
And then it's also predicted two entirely new biological mechanisms.
0:27
Hello, Internet.
0:27
This is your captain speaking, Lester Nare, joined as always by my co-host and our resident PhD, Krishna Chowdhury.
0:36
We are back for another deep dive episode where we are going to talk about a paper today that's trying to tackle one of the hardest questions in neuroscience about comas.
0:48
and how AI is helping us better understand the state of deep, prolonged unconsciousness.
0:54
This was a paper published in Nature Neuroscience by friend of the pod, Daniel Toker, et al.
1:00
on March 26th.
1:02
And as always, we are going to learn about the science from the ground up, because this is From First Principles.
1:24
Every year, tens of thousands of people fall into comas.
1:28
And it's either from traumatic brain injury or a stroke, cardiac arrest, drug overdoses.
1:34
And whenever this happens, some patients recover and some patients don't.
1:41
We don't really know why some recover and why some don't.
1:44
The other thing is for families that are sitting in that intensive care unit, there's a question that haunts them every time, which is, is my loved one still there?
1:53
And this question is one of the hardest problems in all of neuroscience.
1:58
What is consciousness?
1:59
It's not just philosophically hard, which that's above my pay grade and we're not going to get into, but even mechanistically and clinically, this is a very hard problem.
2:08
What is consciousness even physically speaking?
2:11
Obviously, it has something to do with the brain.
2:15
The human brain has 86 billion neurons.
2:19
It's a muscle of fat and neural tissue.
2:23
And it's quite incredible that that, like, sack of fat and tissue is where we get this profound phenomenon of consciousness, right?
2:32
Our ability to perceive, to feel, to think, and to have this subjective experience of ourselves and the world around us.
2:41
And it's all just from, like...
2:43
this sack of like blood and fat in our skull, right?
2:48
And one can ask, so it's somewhere in the brain, but what are the specific circuits that generate consciousness?
2:55
And if a part of that circuit breaks, how does...
3:00
it affect consciousness?
3:01
And then can I actually figure out how to repair it, right?
3:05
Now, doctors can observe that a patient is conscious or unconscious, right?
3:10
Because they're not responsive, and they can classify the depth of unconsciousness on behavioral scales.
3:16
But when
3:17
All we can do really is like wait for them to wake up.
3:20
There are some new therapies right now that we're going to get into that are maybe getting at trying to wake them up.
3:28
But at the end of the day, we still don't have a way to peer inside the machinery of awareness and start asking like, what is the thing that is broken and how do we fix it?
3:38
Okay.
3:38
Or how do we replace it?
3:47
and the actual mechanism that generates that it's kind of a little bit of a yeah dark cave yeah very much a black box in that sense right and that's why i really like this new paper it's a new study that i think represents a genuine step change in that field okay so it's a team that was led by daniel toker at ucla who is as you said a friend of the pod he went to college with us at princeton undergrad and
4:13
Then I think he went to Berkeley for a PhD and then came to UCLA and now has like a gig with UCLA and USC as a postdoc.
4:22
He's also the brain scientist on Instagram and TikTok.
4:26
He's got a bunch of followers, really into science communication.
4:28
You guys should check out his page.
4:30
It's really, really cool.
4:32
So...
4:34
The cool thing about this paper, the cool thing that Daniel did was he used AI to reverse engineer the biological mechanism of unconsciousness in coma patients.
4:45
And the way that he used AI is very different from, I think, when people think about using AI to do science.
4:53
They think like, oh, black box LLM model.
4:55
I like...
4:56
Train it on a bunch of data and then it's going to spit out like a classifier consciousness or not consciousness.
5:02
There's part of that here.
5:03
But what really I think is cool about it is there's the AI has rediscovered known features of unconsciousness.
5:11
And then it's also predicted to entirely new biological mechanisms.
5:16
Mm hmm.
5:17
Because the way that he has engineered this AI is it's not a black box.
5:22
There's actual mechanistic understanding about a model of the human brain that he's hard-coded in.
5:29
And that's what the AI is sort of acting on.
5:31
That's what the evolution of the AI is acting on.
5:34
And we're going to get into that.
5:35
But I think it's very, very cool.
5:37
And the implications are in every direction.
5:40
I just want to make a quick note here.
5:43
Yeah.
5:44
We use the term AI in a very general sense when we discuss on the pod.
5:48
It is a very expansive area of research with a lot of different subcategories and subboxes.
5:55
So a lot of times we will just say AI for shorthand because we have a lot to get to.
6:00
Yeah.
6:00
But we don't mean AI equals LLM when we do so.
6:04
Yes.
6:04
It is sort of talking about the entire field of research.
6:07
Yeah.
6:08
at a high level and we then get into the weeds of what we're talking about specifically.
6:13
But just as a small caveat note, that's important because it is a vast, vast space that is beyond just chatbots.
6:19
Exactly.
6:20
And this is very much not a chatbot.
6:22
Right.
6:22
It's very, very cool.
6:23
We're going to get into it, right?
6:24
And the implications, I think, radiate in all directions, right?
6:27
Something like 5,500 Americans currently are in a persistent vegetative state and –
6:34
There's a larger number that's in a minimally conscious state, and this research offers a genuine scientific pathway towards treatment.
6:41
From the philosophy of mind perspective, it provides a very mechanistic specific picture of what makes a brain conscious.
6:50
And then for an AI perspective, it's also very cool to think about how one can use brain.
6:56
underlying principles of AI architectures to understand something as complicated as consciousness, right?
7:02
So it's a deeply personal question that I think this is answering.
7:06
Like consciousness is the one thing I think that we can all be sure exists because at least we experience it.
7:12
It's the Descartes, I think, therefore I am,
7:15
Clearly consciousness is a thing.
7:17
Maybe the world isn't.
7:18
Right.
7:18
And quantum fields aren't.
7:20
But consciousness, like 100 percent, like I am me, right, sitting here talking in this microphone.
7:26
And you are you and everybody agrees that like that is a thing.
7:29
So it's a deeply personal question.
7:33
And it's still a phenomenon that we, I think, understand the least out of everything.
7:39
So any question.
7:41
Light that is shed on this mystery, I think, is certainly welcome.
7:45
All right.
7:45
So with that in mind, let's get into some of the history behind this type of research.
7:49
Yeah.
7:50
And we'll start all the way with the ancient Greeks.
7:53
Always.
7:53
OK, because the ancient Greeks, they had some crazy ideas.
7:57
Like Aristotle thought that the brain was just a cooling organ for the blood.
8:03
And consciousness actually lived in the heart.
8:05
We're going to let that slide.
8:08
Because other people like Hippocrates, the guy from the Hippocratic Oath, right?
8:13
He noted that patients who fell into deep, unresponsive sleep after a head trauma rarely ever recovered.
8:20
Okay?
8:21
And the word coma comes from deep sleep.
8:26
in Greek.
8:27
So they had already sort of classified normal sleep versus this pathological deep sleep coma.
8:33
They had already kind of figured that out.
8:35
But again, it was just sort of an observation.
8:38
Then in the Renaissance, there came the anatomical revolution with Visalius.
8:43
He published this massive manuscript called De Humani Corporis Fabrica in 1543.
8:51
He was basically really into dissecting human bodies.
8:56
and gave us incredible drawings of human anatomy.
8:59
So from there, now we've got an understanding of brain structure, right?
9:03
But it's kind of like the machine is getting cataloged, but we don't really know what any of the parts are doing.
9:08
So already by the early 1600s, we both had like a philosophical framework or concepts developing about classifying what these different parts of the body are doing and how that works.
9:19
Like how that translates behaviorally.
9:21
Yeah.
9:21
As well as sort of a very detailed cataloging of the physical structures.
9:27
Yeah.
9:27
At least as it related to the brain and other key parts of the human body.
9:31
You know, four or five hundred years ago.
9:33
Yes.
9:33
Yeah, exactly.
9:34
It was a catalog.
9:35
We didn't know what...
9:36
Each thing did, but there was a pretty good catalog already, right?
9:40
And then in the 1800s, we had some pretty revolutionary ideas.
9:45
Paul Broca in 1861, he discovered this idea of cortical localization.
9:50
There's an area of the brain now called Broca's area, and now we know that that damage to that area causes language loss.
9:58
We also know that electrical stimulation in the motor cortex causes muscle movements.
10:06
Thank you.
10:07
a machine that has specialized parts, and each of these parts is doing something, right?
10:12
The other invention in the 1800s was electrophysiology.
10:15
In the 1870s, Richard Caton placed electrodes on exposed skulls of rabbits and found that there was electrical activity.
10:22
And then we see Hans Berger in 1929.
10:25
He develops the EEG, the electroencephalogram, which is you put leads on the skull and you can actually read electrical activity.
10:33
And he found that the brain electrical activity
10:36
dramatically changes with the state of consciousness.
10:40
Okay?
10:41
So different patterns mean waking, sleep, anesthesia.
10:46
So already now it's like, okay, the brain is doing different things during unconsciousness, so perhaps the brain is really the seat of consciousness.
10:54
That kind of makes sense.
10:55
Yes.
10:55
Yeah, yeah, yeah.
10:58
Finally, 1949, Giuseppe Morozzi and Horace Maguan, they published this landmark paper...
11:04
And this demonstrates something called the reticular activating system.
11:09
It's a diffuse system that runs in the brainstem.
11:12
The brainstem is where your brain meets the rest of the body with the spinal cord.
11:15
So it's kind of like the connecting part of the brain.
11:19
And they found that that part of the brain, the part that connects the brain to the spinal cord, that is essential for maintaining wakefulness.
11:28
So now we've honed in on this is the sort of general area.
11:32
where consciousness is probably happening.
11:35
So to make a crude analogy, we figured out the brain is the factory, and we now know that different parts of the factory have different workers.
11:45
The engineers are downstairs.
11:47
The sort of manufacturing line is upstairs.
11:50
The finance department is over here.
11:53
But we didn't really know where the boss's office was.
11:56
Yeah.
11:56
Where's the boss's office?
11:57
Yeah.
11:58
Yeah.
11:58
It seems like that's where it's the brainstem.
12:00
Right.
12:00
It's that part.
12:01
Right.
12:02
And over the following decades, we've mapped out that brainstem in more and more detail.
12:07
So we figured out there's something called the locus coriolis.
12:10
There's something called a basal forebrain.
12:13
There's something that's part of the brainstem.
12:15
There's the hypothalamus.
12:17
There's the raphe nuclei.
12:18
And so all of these like sort of parts, you see where the spinal cord is going in, that brainstem.
12:24
junction is where all of this is happening which it's it's almost like even just from a crude structural like just looking at it you would imagine if the things that are most sensitive have the most protection just locationally yes locationally that makes sense from one perspective of evolution and the other perspective of evolution i think is that like the brain sort of grew out from that part right yeah
12:47
So it's like the oldest part of the brain is the part that connects it to the rest of the body.
12:51
Because, I mean, let's face it, like the brain evolved to control movement.
12:56
Right.
12:56
And so the brain controlling the rest of the spinal cord is sort of probably how it started.
13:01
And then like higher and higher cognition formed around that.
13:03
Right.
13:04
Like, oh, OK, I need a sense of place.
13:06
I need a sense of time.
13:09
I need a sense of problem solving.
13:11
All of that starts forming around this like nucleus.
13:14
Right.
13:15
And then the prefrontal cortex being the most advanced.
13:17
The most advanced.
13:18
That's the one that's like...
13:19
Right in the front.
13:20
Which, you know, it's very cool to think about it that way.
13:23
Just like in terms of stages of development over the millions of years of evolution.
13:27
Okay.
13:28
Okay.
13:28
So now we know that it's like kind of the brainstem, but we still don't have really an idea of like...
13:34
Okay, this thing connects to this, this thing connects to this, so this is powering that, that, that, that, and so on and so forth.
13:40
2007, there's a pivotal moment in consciousness research because Nicholas Schiff and his colleagues published this paper in Nature.
13:48
And it showed that continuous deep brain stimulation, this is the idea of we take an electrode...
13:53
and we pulse it with electricity, that is going to excite the neurons near that electrode.
13:59
It's just you're pumping electricity deep into the brain.
14:02
That's why it's deep brain stimulation.
14:04
And deep brain stimulation of the central thalamus helped a minimally conscious patient
14:10
regain the ability to communicate and feed himself okay this was a very big deal because this was an ability that he had lost about six years earlier after a traumatic brain injury it's a preliminary result single patient notoriously in human consciousness studies and things like that the n is like very low single digits is you're already getting a lot because human i mean you're like experimenting on human beings right there's tons of ethical concerns yep there's not a lot of like
14:37
You know, people who are willing to do this.
14:39
And so even though the data is low, the fact that it worked means that there's something going on.
14:44
So that sort of rejuvenated this research.
14:47
And in 2010, that same author, Nicholas Schiff, he came up with something called the Mesocircuit Hypothesis.
14:54
Okay?
14:55
Effectively, what he's saying is the following.
14:57
The consciousness relies on a network.
15:00
of interactions between the cortex and the central thalamus.
15:04
Okay.
15:04
And that interaction is regulated by a part of the brain called the basal ganglia.
15:09
We're going to focus on effectively three...
15:14
regions of the brain.
15:15
Okay.
15:16
And I'm going to use an analogy of an electrical grid.
15:20
Okay.
15:20
Okay.
15:21
A city's electrical grid.
15:22
Yes.
15:22
So the cerebral cortex, this is the part that's on the outside, the most sort of recently developed in humans.
15:29
Um, this is the, when we think about the brain and all the wrinkles, that's the cerebral cortex that we're seeing on the outside.
15:35
And that outer layer, it's responsible for higher thought, memory, sensory processing,
15:40
And in this analogy of a city's electrical grid, this is your network of homes, podcast studios, businesses that consume the electricity, right?
15:50
Now, the central thalamus is the deep brain relay station.
15:54
That's the part that's sort of on the right over there in the pink.
15:58
Just above the brainstem?
16:00
Just above the brainstem.
16:01
Okay, so that's getting input from the cortex, and it's also giving input to the cortex.
16:05
Okay, so it's a two-way street.
16:07
It's kind of like the primary power substation.
16:09
Yes.
16:10
It's pumping excitatory input to the cortex, but it's also getting excitatory input back from the cortex.
16:16
So it's relaying a bunch of electricity.
16:18
One way or the other.
16:19
So one of the examples being if you have solar on your house, you're sending electricity back to the grid.
16:25
It's not just you're receiving it as a practical example.
16:27
That's very good.
16:28
Yeah, yeah, yeah.
16:28
No, that's exactly right.
16:30
And this feedback loop kind of helps both run.
16:33
Right.
16:34
Okay?
16:34
In some sense.
16:35
And then finally there's the basal ganglia, and that's the striatum and the globus pallidus.
16:40
Mm-hmm.
16:40
That is a kind of regulatory system, and it's kind of like a voltage controller that manages the flow of power in this entire network.
16:49
So in the summer when it gets hot, they have to decide where to send energy, where to not send energy in order to not overflow.
16:55
Like there has to be – somebody decides, okay, we're going to shut down this neighborhood because we need to power the nuclear.
17:01
Yeah, exactly.
17:03
As a crude analogy, it's kind of routing signals, and it's controlling both the power stations and the homes to make sure there's balance and nothing's going out of the way.
17:13
Okay, so that's how sort of the consciousness mechanism is working in healthy patients.
17:18
What happens during a coma?
17:19
Well, according to the mesocircuit hypothesis, here's what happens.
17:23
When there's widespread brain injury, like from a traumatic brain injury or a stroke, so a bunch of neurons die because they don't get oxygen and things like that.
17:32
your power lines are going to be down on the cortex, right?
17:34
So you're going to get localized areas of power cuts, like we do in LA sometimes when somebody steals our copper from underground.
17:43
But anyways, the cortex is now going to reduce its excitatory input to the thalamus.
17:49
That connection is broken, as you can see, on the left.
17:51
The thalamus and the cortex, that's now a dotted line.
17:54
Instead of on the left, it's a solid line.
17:56
Okay, the other thing is,
17:58
There are these special neurons in the striatum called medium spiny neurons.
18:04
And what they do, they're super prone to metabolic stress death.
18:09
Meaning you, like, remove a little bit of oxygen and they'll just die.
18:13
Okay?
18:13
They're not very resilient.
18:14
They're very finicky.
18:16
And if those guys die, then the striatum is going to not inhibit the part that is inhibiting the thalamus.
18:24
It's a double negative now.
18:25
Okay.
18:25
Okay?
18:26
So...
18:28
Because of that, so what it's doing is usually the striatum is like inhibiting the part that is putting brakes on the thalamus.
18:36
Okay?
18:36
But now that that brake is gone, the first brake is gone, right?
18:41
Right.
18:42
The second negative is going to go even more negative.
18:46
Okay?
18:46
And it's going to start inhibiting the thalamus even more.
18:48
That's why that red line just got fatter.
18:50
That makes sense.
18:51
Right?
18:51
So the thalamus is, one, not getting any input from the cortex because the power lines are down.
18:55
Yes.
18:56
And second, the part that was putting brakes on the thalamus has now gone hyperactive and putting even more brakes on the thalamus.
19:02
So the thalamus is getting inhibited.
19:04
It's getting shut down more and more in a coma because of these very tiny...
19:10
Defects that have happened in exactly the wrong areas.
19:13
Yeah.
19:14
So this is very interesting.
19:15
And what we're sort of saying here is the traumatic brain injuries are disrupting the normal flow between these three fundamental component parts.
19:26
Yeah.
19:28
The cortex, which is like the end destination, the Holmes example, the thalamus, which is sort of sending and receiving, and then the striatum and palliatum that are...
19:37
kind of controlling the amount of what is or is not being sent and where it is or is not being sent.
19:42
Exactly.
19:42
And it's exactly at the right part where the thalamus, which is the relay station, is just getting signaled to shut down.
19:49
Yeah, yeah, yeah.
19:50
Okay?
19:50
And when that shuts down, now this entire circuit of consciousness is getting shut down.
19:55
It's almost like the fail-safe system that happens at a substation if there's some overload, just gets stuck in the locked position.
20:00
Exactly.
20:00
Exactly.
20:01
Exactly.
20:02
And there's some evidence that this mesocircuit hypothesis is probably correct.
20:07
There's certain stimulant drugs, for example, amantadine, which enhances the inhibition to that break, like the part that's inhibiting the thalamus.
20:18
amantadine inhibits that part.
20:20
So it's no longer going to inhibit the thalamus kind of works.
20:24
And it kind of wakes up patients from turning it to overclock.
20:26
Yes, yes, exactly.
20:28
So, so there, there's some mechanism that,
20:32
where these drugs are kind of working with the mesocircuit hypothesis.
20:35
But at the end of the day, still a hypothesis because it's not very granular, right?
20:39
It's just like, oh, giant brain area.
20:43
But the giant brain area has millions of neurons that are all different types.
20:47
What type?
20:48
What type is doing what, right?
20:50
And also, you can't really test it directly because the bottleneck is standard animal models are really bad at mimicking neurons.
20:58
human comas, this prolonged unconsciousness.
21:01
And that just has to do with the fact that animal models like rodents, the brain is structurally a very different thing, right?
21:09
Even though it's still a mammal...
21:11
One of the things that we think happens during traumatic brain injury that causes this kind of coma is the idea of a diffuse axonal injury.
21:20
Here's the idea.
21:20
You've got the cerebral cortex on the outside.
21:23
There's gray matter, which is like sort of the cell bodies.
21:26
And then there's white matter.
21:27
You might have heard of that gray matter and white matter.
21:29
Gray matter is kind of like the cell bodies.
21:31
The white matter is the axons.
21:32
These are these like cables.
21:35
Like, you know, the high voltage cables that come from, you know, Hoover Dam to L.A.?
21:39
The axons are kind of like that.
21:40
They're the high voltage cables.
21:42
Now, if we go to a photo of the 14, what you'll see is the white matter is kind of in the middle and it's like connecting these areas of gray matter.
21:51
But if you have a traumatic brain injury, let's say you got in a car crash or like, boom, like your brain had severe G-forces on it.
22:00
Those axons, which are the cables, those are going to rupture first.
22:03
Yeah.
22:04
Right?
22:04
Yeah.
22:05
Now, with a small rodent, the brain is not big enough to feel that kind of G-force stress.
22:13
Right?
22:13
Ours is big enough.
22:14
We're like, and we're like doing crazy things as humans.
22:17
We're like, we're driving around at 70 miles an hour.
22:20
We're like, we're now subject to G-forces if something goes wrong.
22:24
Right?
22:24
And so our brain is just not evolved.
22:28
When we were hunter-gatherers in Africa, like evolving, the evolution didn't think, hey, I should probably make this brain, you know, withstand getting hit by another 350-pound man in an NFL game.
22:41
Yeah, exactly.
22:42
Exactly.
22:44
So there's a mismatch between what the brain has designed, the environment that the brain is designed in, and then what the brain was capable of creating, which is like human beings in cars and all this other crazy stuff, skiing and then hitting a tree and things like that.
23:00
Right.
23:01
Right.
23:01
So – and we simply can't test –
23:05
In rodent models.
23:06
Right.
23:07
Okay.
23:07
And this is something we talked about in a recent previous episode too about just, you know, there are a lot of opportunities to use animal models, which the purpose of which, yes, for folks who are, you know, PETA adjacent, et cetera.
23:22
The ethical concerns about testing directly on humans are so prevalent that it's perceived as a lower ethically bad option.
23:32
100%.
23:33
And for an outcome that's going to save millions of lives at some point in the future.
23:36
Yeah.
23:37
Yeah, but here, like...
23:38
It doesn't.
23:39
It doesn't.
23:39
The ethics, like, is even more problematic because, well, you're not going to get anything out of it.
23:44
Right, right.
23:45
Right.
23:45
And so you're just... Yeah, you're just, like, wasting animals at that point, right?
23:50
Right.
23:50
And that's not something that we want to do in science research.
23:53
Right.
23:53
So that's one bottleneck.
23:55
Okay.
23:55
Okay.
23:55
which is just animal models are not a thing.
23:58
The second bottleneck is the diagnostic crisis.
24:01
Okay.
24:01
This is in general with consciousness in medicine.
24:05
Okay.
24:05
Consciousness has two fundamental components.
24:07
There's wakefulness, which is like arousal, like when you're awake.
24:10
And then there's also awareness.
24:11
That's the qualitative experience of like me being me and seeing you and everyone else.
24:17
Right.
24:17
This is some qualitative subjective experience that only I have about what it's like to be me.
24:22
Right.
24:23
So,
24:24
disorders of consciousness that comes from physical trauma and oxygen deprivation, there's several different types.
24:29
So in a coma, you have neither arousal nor awareness, right?
24:33
Because you can't be awoken.
24:35
Yes.
24:35
There's no sleep-wake cycle.
24:37
Yes.
24:38
And they fail to respond to stimuli.
24:40
The lights are off.
24:41
Yeah, lights are off and maybe nobody's home.
24:43
Right.
24:43
Okay?
24:44
Vegetative state means it's also called unresponsive wakefulness syndrome.
24:49
Okay.
24:49
You have wakefulness.
24:52
And you have a sleep-wake cycle, but perhaps you have no awareness.
24:55
So the lights are on, but maybe nobody's home.
24:58
Okay.
24:59
But now it gets kind of weird.
25:01
Like, what if they are aware, but they just can't...
25:06
have any agency over their motor controls right so how can you tell so someone's home but they can't answer the door when you ring the doorbell exactly yeah lights are on lights are on and someone is home yeah you ring the doorbell but they can't yeah that's very good yeah exactly how do you how can you tell right this is a yeah that's subtle that's very subtle and for the
25:30
Science paper 2006 by Adrian Owen.
25:33
It's a landmark paper.
25:34
It's quite incredible.
25:35
This was a wake-up call, so to speak.
25:38
Okay.
25:39
He used functional MRI to ask a patient that was diagnosed in a vegetative state to imagine playing tennis and imagine walking around her home.
25:50
And the brain activity from that fMRI study was identical to someone who was fully aware.
25:57
So she's there in this vegetative state.
26:01
There's no way that she can express her awareness.
26:04
Mm-hmm.
26:05
But if we look into the brain MRI, it's the same as someone who's aware.
26:10
The idea being they may not be able to move their hands to touch a button or verbalize.
26:17
Or even move their eyes.
26:18
Or move their eyes, which is techniques that have been used before to communicate.
26:23
Now we're just saying it's because the awareness is in the brain.
26:27
And so you shouldn't actually need any of these secondary things.
26:30
uh meth avenues yes to communicate and the only way to make the differentiation between the uh is someone home versus someone not home reference point you have to go directly to the source yeah if all of their physical capacities yeah are are not available to them which does not necessarily mean yeah that no one is home
26:51
Exactly.
26:52
It's like no one's answering the door, but here it's kind of like we did an X-ray of the home.
26:55
Yeah, yeah.
26:55
And the person is moving around.
26:57
Right, right.
26:57
We have the thermal and you can see.
26:58
Yeah, exactly.
26:59
You know?
26:59
Okay.
27:00
I thought that was a very cool paper.
27:01
That is very clever.
27:02
It's very clever.
27:03
And now it's a real crisis, right?
27:07
Right.
27:07
Because now there's a third.
27:08
It's called minimally conscious state.
27:10
Mm-hmm.
27:10
Okay?
27:10
And it turns out that 15% to 20% of patients that are classified as vegetative... Mm-hmm.
27:17
actually might have covert conscious awareness.
27:20
That's a lot.
27:21
That's got to be terrifying.
27:22
And that's got to be so terrifying.
27:24
Right?
27:24
Because you can hear and see everything happening around you, and you cannot engage in that environment.
27:28
That's just unbelievable.
27:29
And this was only 20 years ago, 2006, right?
27:32
After this 500-year cycle we've just talked about.
27:34
Yeah.
27:34
So, I mean, this field is just one of those where you can really sit back and be like, wow, we really don't know much about the brain.
27:42
Right.
27:42
Right.
27:43
Right, right.
27:44
Because this is like kind of, you know, people's, this goes to the materialist, non-materialist debate.
27:51
You know, is it that consciousness arises out of matter or is consciousness primary, which gets a little, that's, again, that was above our pay grade and not the focus of this podcast.
28:03
Yeah.
28:03
But it is one of those.
28:05
It's the first question.
28:07
Yes.
28:07
Because once you get, you know what I mean?
28:10
You can't, everything else downstream is really impacted by which of the two.
28:14
Which of those two, right?
28:15
What are we talking about?
28:16
Right.
28:17
Who am I?
28:18
Who are you?
28:19
And what the hell is going on?
28:21
Right.
28:22
That's how I like to say it.
28:23
Right.
28:24
Right.
28:25
You know?
28:25
And so with these kinds of studies, it's really like underscoring how much we don't know about this, right?
28:33
Right.
28:34
And now 15% to 20% of patients, we're saying, are classified incorrectly.
28:39
Which is fascinating.
28:40
Because, I mean, I'm sure after that study, a variety of... Because the fMRIs are not hugely...
28:47
negative for a patient no so at least as a way especially for families or anybody to try um but in any event exactly so now we finally get to um daniel's paper daniel toker in nature neuroscience adversarial ai reveals mechanisms and treatments for disorders of consciousness it's a really innovative way of using machine learning and the vast amount of data that's now available
29:12
On comatose patients, vegetative patients.
29:16
He's made an in-silico model of coma.
29:19
So a computer model of the brain that circumvents the lack of mouse models.
29:26
So now I have a model in the brain that I can play with.
29:30
A model of the brain in my computer that I'm fairly confident is quite good.
29:34
That I can play around with.
29:35
Because the idea is it's based off of this real-world data.
29:38
Anonymized.
29:38
Yeah.
29:39
Yeah.
29:41
data.
29:42
Exactly.
29:42
And then I can figure out a kind of mechanistic model about how consciousness works, what gets disrupted to make a coma, and then it becomes a kind of discovery engine.
29:52
From that, we can mess around with it and propose new treatments.
29:55
And he is doing all of these.
29:57
Very well done, Daniel, by the way.
29:59
Quite good.
30:00
Quite good.
30:01
Again, thebrainscientist on Instagram.
30:04
And before we deep dive into it, let's do a little bit of housekeeping.
30:08
Yes.
30:09
So as always, we are so grateful for all of you to join us for these research deep dives.
30:15
It's just an incredible, not only an incredible time to be alive, but...
30:19
But the opportunity to really understand not only what we're learning today, but how it works and what the history of how we got here, you're not going to get it anywhere else.
30:29
And this is why we love doing this show at From First Principles.
30:33
And it is hugely helpful to us to continue to bring you the best and greatest breaking science research.
30:40
As we, as I always say, fight the billionaire algorithms, a like, a share, a comment, a DM, sending it to a friend, bringing it into Journal Club is hugely, hugely helpful for us where it's, again, still just the two of us running the show to be able to give this to you multiple times every week.
30:57
If you would like to become a patron, you can donate to the pod at ffppod.com backslash donate.
31:04
Any and all support monetarily or non-monetarily is hugely, hugely valuable, and we really greatly appreciate it.
31:12
I have no other show notes this week, and I'm very excited to dive back into this paper.
31:17
Let's do it.
31:17
Okay.
31:18
So first thing I want to do, right?
31:21
I want to use artificial intelligence to make a discovery engine for my brain, figure out why comas happen, and then propose treatment.
31:29
That's the end goal of this paper, okay?
31:31
Okay.
31:31
First thing I want to do is train an AI detective that is capable of diagnosing consciousness from raw brainwaves.
31:38
Okay.
31:38
Okay?
31:39
Just can I identify consciousness if you give me the, you know, sort of brain activity, the electrical activity of a brain?
31:48
So he made something called a deep convolutional neural network.
31:51
These are just our normal, you know, CNNs that we hear about.
31:55
Giant network of artificial neurons that is trained through backpropagation.
32:00
You know, supervised learning, which is just we've got about 680,000 10-second neurons.
32:07
electrophysiological recordings.
32:09
So 680,000 recordings of brains, electrical activity.
32:14
These recordings come from humans, monkeys, rats, and bats under anesthesia, in a coma, and wakeful.
32:20
So a swath all across mammals.
32:24
So that means it's not going to really hone in on some human feature.
32:27
It's going to really start thinking about what is the nature of consciousness itself.
32:32
Right?
32:33
And he made three discriminators.
32:37
neural network that was going to get trained on the cortical data.
32:40
There's a thalamic and then there's a palladial.
32:43
That's the three sort of big brain regions that we were talking about in that mesocircuit model, right?
32:49
The core, the cortex is your grid with all the homes and the podcast studios.
32:56
The thalamus is the relay station and the palladial station is kind of like governing what's going on between these two.
33:05
Right?
33:05
Yes.
33:05
Okay.
33:06
Okay.
33:08
Once he trains it, he validates it.
33:10
It's pretty good.
33:11
The AI score for consciousness, which is plotted on the y-axis, is...
33:17
is correlated with the ground truth score of consciousness that it was given to each sample in the test sample.
33:24
Okay?
33:24
Which means that, like, the AI, this deep convolutional neural network, is given a sample that it's never seen before, and it's told to guess how conscious it is based on a number between 0 and 1.
33:36
That number is correlated with the ground truth.
33:38
Okay, so that means it's working.
33:39
Yes.
33:40
The other big one that I quite liked is from figure 1J, if you see over there.
33:44
What that shows is that the AI was able to classify fully paralyzed ALS patients as conscious.
33:53
Remember in ALS, we did the story where it's a motor neuron disease, right?
33:57
All of your motor neurons go away.
33:59
So you can still be conscious, but you won't be able to...
34:03
fully express that.
34:05
Here, they looked at EFIS recordings from ALS patients, and there's no significant difference between them and healthy cortical activity.
34:15
This is actually a fascinating follow-up to our ALS story because we did a great deep dive on understanding what it actually is mechanistically.
34:22
Yes.
34:23
Yes.
34:24
And it dovetails with what we said earlier, which is just because your motor functions go away.
34:29
ALS is a perfect example of what we were just talking about.
34:32
It doesn't necessarily mean that there's no one home.
34:35
Yeah, yeah.
34:36
That's very – Exactly.
34:37
And this was able to find that, right?
34:38
Without motor output, you can now classify.
34:41
That's it.
34:42
So already big win.
34:43
Figure one.
34:43
Yes.
34:43
Yes, that is huge.
34:45
Right?
34:45
That's good.
34:45
That is huge.
34:46
Okay.
34:47
Now next, we've made a detector.
34:49
We've made a detector of consciousness.
34:51
Yes.
34:51
Now that we have a consciousness detector, let's make a brain simulator that can make my own brain signals.
34:58
Okay.
34:59
Okay?
34:59
And this is where we get into the real why there's AI in the title.
35:04
Okay?
35:04
Okay?
35:05
Because just making a deep convolutional neural network that will classify, that's just an autonomous classifier.
35:11
Right, right.
35:12
That's like 10 years old.
35:13
What are you doing?
35:13
Right?
35:14
You're not going to get into nature neuroscience like that.
35:16
And Daniel knows that.
35:18
And this is for the social clips.
35:19
If we end up clipping the first half without the second half...
35:23
Please watch the pod so you get the full story.
35:25
Yeah, so they're not like, that's not what I mean.
35:27
I can't believe that's a nation you're not saying.
35:29
Yes, we know.
35:30
We can't fit an hour-long thing onto Instagram.
35:33
Yeah.
35:34
Okay, so now what he's going to do is make a brain simulator.
35:37
Mm-hmm.
35:38
He's going to combine that discriminator that he got, the AI, the consciousness detector, with something called a generative adversarial network.
35:46
These are GANs.
35:47
Traditionally with GANs, I mean, they're kind of used to make, like, generative AI, like, photographs and things like that.
35:53
Here's how it works.
35:54
So this is very different from, like, a diffusion model, which is usually kind of used in the zeitgeist.
35:58
GANs kind of didn't have popularity.
36:00
They had popularity for a while.
36:02
And then diffusion kind of went up and became the generative network.
36:06
sort of paradigm.
36:07
But then now GANs are kind of back up.
36:10
We don't really know which architecture is going to win in the end.
36:12
But here's the idea.
36:13
Okay?
36:14
So you've got something called a forger, which is trying to forge...
36:21
real life data.
36:22
For example, in this case, what we're trying to do is create real faces.
36:26
So we're going to have a data set of real faces, and then we're going to have a forger, that's the generator over there, and that's going to generate faces.
36:34
And then we're going to have a discriminator,
36:37
On the other side, that is going to take the real data and going to take the forged data and try to figure out which one is forged and which one is real.
36:44
Now, in the beginning, it's going to be obvious.
36:46
In the Beninging.
36:47
Yeah.
36:47
In the beginning, because the forger doesn't have any training.
36:50
He's just going to be making up random nonsense.
36:52
Right.
36:52
So the discriminator is going to be able to look and be like, oh, yeah, this is real.
36:55
This is fake.
36:56
But as the forger gets better and better, because...
36:59
That output of, hey, this is fake, you keep making fake stuff, I want the real stuff, that output is getting back propagated through to the forger.
37:07
And the forger is thinking, okay, how do I get better?
37:09
How do I get better?
37:10
Pretty soon after multiple rounds of training, the forger is going to get good at making the real data such that it fools...
37:19
You're a discriminator.
37:20
Yes.
37:20
Okay?
37:21
That's the generative part, and the adversarial part is you've got two networks that are adversaries.
37:26
There's a detective that's trying to detect the fakes, and then there's a forger that's trying to create even better fakes.
37:32
That's the adversarial part, right?
37:34
For anyone who's ever used image generation, particularly mid-journey, this structure is exactly the reason why you'll see a blurry, weird thing happening.
37:44
when you first initiate the image generation, and then it's going through the cycle, you just talked about it, the fidelity gets closer and closer and closer over until it gives you the final, like produced like image.
37:56
This is like the whole stable diffusion kind of came
37:59
Yeah, and that one's slightly different.
38:02
Slightly different because – so what you're talking about is very, very close to generative adversarial networks because at the end of the day, how the forger creates the fake stuff is it starts with some kind of random noise, which is what you see there.
38:13
And then it starts manipulating that noise to create the real thing.
38:18
So what you're seeing there is the act of the forger, in some sense, going from noise to the real thing.
38:25
But the training part is something that's already been done by mid-journey on the backend.
38:29
100%.
38:29
And so that's the generative part.
38:34
Yes.
38:34
Okay, great.
38:35
So now...
38:37
Great.
38:38
That's how you make a fake Picasso and a real Picasso and things like that.
38:41
How are we going to use that to figure out coma?
38:43
Yes.
38:44
All right.
38:45
This is where Dan comes in hot with an interpretable biophysically grounded mean field model.
38:54
Okay.
38:54
And that is what is cool.
38:55
So on the upper left, that's A, figure A.
38:58
Yes.
38:58
Okay.
38:58
On the left-hand side, we've got our DCNN, the convolutional neural network.
39:02
That's our detector.
39:04
that we had trained previously in figure one, right?
39:06
There's other detectors in here that'll actually figure out, is this real or fake data?
39:12
That one is saying, is this unconscious or conscious data?
39:15
There's other ones that are not in this backend.
39:17
You have to go to the supplement to see it, where it's like, is this real data or fake data?
39:21
Because first we want to just create...
39:23
Real looking data.
39:24
Then we'll worry about, okay, am I creating conscious data or unconscious data, right?
39:28
So there's multiple steps to this process.
39:30
But the key is that the forger is not a black box, okay?
39:35
The forger is actually a three component within which there's multiple components.
39:41
Yes.
39:42
But you know that mesocircuit hypothesis of like there's a cortex, there's the basal ganglia, there's the thalamus, and these guys are talking to one another?
39:49
Each of those boxes is its own little neural network, right?
39:53
And it's a biophysically realistic neural network in the sense that it's not just a bunch of artificial neurons that start at noise.
40:01
What it's doing is...
40:03
It's saying, okay, how does a biophysical neural network work?
40:06
Well, some input comes in, then there's going to be like channels, like, you know, some neurotransmitter is going to have some characteristic timescale.
40:15
Other neurotransmitters are going to have other characteristic timescales.
40:18
There's going to be excitatory neurons, which do positive feedback.
40:22
There's going to be negative inhibitory neurons that do negative feedback.
40:26
These I can hook up mathematically using differential equations, and I can create a biophysically realistic model.
40:34
It is no longer a black box.
40:36
And instead of using backpropagation to change the weights between each neuron, what I'm going to do is use something called a genetic algorithm to...
40:46
to change the parameters of my biophysically realistic model, meaning how many excitatory neurons are there?
40:53
How many inhibitory neurons are there?
40:55
How is the connectivity, right?
40:57
What is the timescale?
40:59
of the interaction things like that things that are actually relevant when we think about what is a brain doing and is in some sense the reason why it made it's like the that algorithm is necessarily bounded by what the realistic value ranges are exactly and that's why it's not a black box because you know that oh it's either going to be excitatory neurons are either going to go between x and y yeah and that's a realistic normalized range
41:25
Exactly.
41:26
That's one of them.
41:27
And the other one is we can literally point to components of the model and be like, oh, it's the excitation and the thalamus.
41:32
Ah, yeah, yeah.
41:33
That's right.
41:35
I can literally be like, oh, it's this part.
41:37
Yeah.
41:37
With again, it's like with normal neural networks, I don't have no idea.
41:41
Yeah.
41:42
Like, you know, I type in make me a cat on a horse.
41:46
I have no idea in the trillions of parameters where it decided.
41:51
cat, where it decided horse, how it figured out to put the cat on top of the horse.
41:56
Here, literally, I could look into the parameters.
41:59
I can see how they evolve with training and figure out what is actually going on.
42:03
And then make that extrapolation from there to the brain.
42:07
That's very good, I think.
42:08
That's very good.
42:09
Right?
42:09
Yeah.
42:10
And that's one of the things I like a lot.
42:12
Yeah.
42:12
Right?
42:13
This loss function, he had a loss function that sort of, like, figures out, right, how to change the parameters.
42:19
And this loss function incorporates the outputs from the networks that's trained to classify real versus synthetic data.
42:26
It also has outputs from the consciousness detector that we had done earlier.
42:30
It's also got a way to identify seizures.
42:32
Yeah.
42:33
And as you said, it's got these empirical constraints on the firing rates of these neurons, as you said, like, right?
42:39
Excitation can't be that high.
42:41
Inhibitory neurons are usually higher firing rate than excitation.
42:44
Things like that, right?
42:45
And the interregional communication patterns.
42:47
Right.
42:48
Like excitation is long range.
42:50
Inhibition is usually short range.
42:52
This is stuff you can bake into the model.
42:54
This is quite nice.
42:55
This is quite nice I love the fact that it's not black box Because that's one of the things that I hate about Like just large neural networks I have no idea where anything is happening And it's also so different from the culture of software Since its beginning Which has been that everything is explicit And you can literally point to Where exactly, like the trace callback Of where something is coming from This is the going back to that type Which is great This is like, yeah, fantastic Fantastic
43:23
And so once this genetic algorithm has trained my brain maker, it's an artificial brain, now we can ask, okay, how about we reprogram the objective now?
43:35
Okay.
43:35
It was just making real world data.
43:37
Now I reprogram the objective to make comatose data.
43:41
So I can use that first neural network to be like, give me coma.
43:45
Right.
43:46
Right.
43:46
Yes.
43:47
Right.
43:47
And then now that genetic algorithm is going to tweak the parameters such that I get coma type data.
43:52
Yes.
43:53
And just I just want to bring this back in because we have done the training and done the validation based off of a ground truth, real source of actual information.
44:03
Yeah.
44:03
We have all of those 10 second long clips.
44:05
Right.
44:06
That's like I just it's not making this from whole cloth.
44:11
No.
44:12
That's a really important concept.
44:13
Yeah, and there's multiple stages of training.
44:16
Right.
44:16
He's figured out a way to, you know, you can't just tell a biophysical model, make me coma.
44:22
Right.
44:22
Right.
44:23
You got to first be like, no, no, no, let's make just normal data.
44:27
Give me data that can fool even me to thinking, I don't know if that's from a real patient or from my brain simulation.
44:34
Right?
44:34
Right.
44:35
I'm sure he did like checks just on his own.
44:37
Yes.
44:37
Where like after the whole thing, I mean, you look at it, it's like, okay, that looks pretty good.
44:40
Yes.
44:40
Right.
44:40
And then you go, okay, now give me a coma.
44:43
And he was successfully able to recreate known phenotypes.
44:46
Like on the right-hand side, we've got high voltage delta oscillations.
44:49
This is something that happens in coma.
44:50
Like the slow sort of oscillations that are super high voltage.
44:54
We've got burst suppression that's on the bottom side.
44:56
Yeah.
44:57
Big bursts and then just seconds of inactivity.
45:00
Again, something that's there in coma patients.
45:03
What's cool is,
45:04
There's no explicit programming of prior neuroanatomical knowledge on what is the difference between a coma versus a normal brain.
45:15
Okay?
45:16
Just the genetic algorithm and that detector that I had of coma versus conscious.
45:22
Yes.
45:23
Just that and the genetic algorithm has now independently discovered some of the core tenets of that mesocircuit hypothesis, right?
45:32
What did the mesocircuit hypothesis said?
45:34
It said that if I weaken cortical drive coming in from the cortex, that's going to cause a bunch of random crap.
45:41
That's what my model is doing.
45:45
So as a first pass, it's already kind of on the right track.
45:48
You can smell...
45:50
it in the water like there's i'm on the right track because right the normal stuff that i kind of know about the mesocircuit that i hadn't baked in explicitly it's already figuring that out it is in it it is being it is able to intuit uh what we've already done over this last five six hundred years that we talked about um out of the box without this is sort of what people say about benchmarks and ai models it's like
46:16
well, if you train it on the benchmarks, then it's going to do good on the benchmarks.
46:19
And so the point here, it's not, the answers to the test were not provided as a part of training.
46:26
And so it's getting answers that are not tainted by sort of giving them the cheat sheet beforehand.
46:33
Yeah, yeah.
46:33
Yeah, it's just looking at how do I make a coma and already it's recreating parts of the mesocircuit hypothesis that we know are probably true.
46:41
Right.
46:42
Okay.
46:42
Right.
46:42
So that's already impressive.
46:44
The next part is the impressive part, which is what is the new stuff?
46:48
I just want to say this is all already very impressive.
46:50
And just as another caveat, we're talking about Daniel specifically.
46:53
This was obviously a team and this is incredible work by everybody on the team.
46:56
yeah yeah yeah everybody is our friend so that's why we're talking about daniel yeah all right all right so but clearly yeah this is this is very i mean there takes a lot of data it takes a lot of like it takes a lot of brain it's always a team effort guys every time we're talking that's always a team yeah yeah so let's get back to daniel yeah okay um here's here's what he's gonna do next okay okay so
47:22
The one part of the mesocircuit hypothesis, which is like cortical drive goes down, so the normal power grid is down, or there's elevated firing in the palladial population, so that's going to inhibit the thalamus even more and things like that.
47:38
That stuff is already corroborated.
47:40
Stuff we already knew, this thing is kind of confirming.
47:42
Okay.
47:43
Now—
47:44
Does it do anything new?
47:45
Well, yes.
47:46
There's something called the selective disruption of the indirect pathway.
47:50
This is a new prediction from the model.
47:54
Here's what's happening.
47:55
So it turns out the prediction is that the coma is actually driven by the degradation of the...
48:04
Certain medium spiny neurons.
48:06
Remember those neurons that I was telling you about?
48:08
They just die if there's any stress whatsoever.
48:10
Or a little bit of loxia deprivation immediately.
48:14
They're just like, okay, I'm done.
48:17
There's selective degradation of a certain type of medium spiny neuron in the striatum that is projecting to the part that's putting on brakes to the thalamus.
48:28
And this very specific pathway is
48:33
is new.
48:34
Okay?
48:34
The classical mesocircuit hypothesis is just like, oh, there's like striatal dysfunction.
48:40
Right?
48:40
It's like the whole brain region.
48:42
Yeah, yeah.
48:42
Right?
48:43
This thing is saying that the indirect pathway, which is specifically this D2 receptor expressing medium spiny neuron, this very specific subpopulation in the striatum, that's the part
48:56
That is getting weakened.
48:57
And when that gets weakened, that's going to ultimately suppress the thalamus at the end of the day.
49:02
Like that thing that I was telling you about.
49:04
But now it's identified a little part.
49:06
It's not just all of L.A. County.
49:08
We've got it down to a street block.
49:10
Yeah, and a specific type of house type thing.
49:12
Right.
49:12
Right?
49:13
Okay, so how do you... Fine, that's your prediction.
49:16
Predictions are only as good as the data that...
49:19
Corroborate.
49:20
Right.
49:20
So how do we corroborate it?
49:21
We use something called diffusion tensor imaging.
49:24
So this is an advanced form of, like, MRI.
49:26
Oh, this is beautiful.
49:28
What we can look at is trace.
49:30
This is very cool.
49:30
What you can do with MRI is magnetic resonance imaging, right, where you look at, like, hydrogen atoms in whatever biological tissue, and you can actually...
49:43
See that using this magnetic resonance imaging technique.
49:45
I'm not going to get into how MRI works, but effectively we can trace hydrogen atoms.
49:50
Okay.
49:51
Water has a bunch of hydrogen atoms.
49:53
Okay.
49:53
So we can measure the directional diffusion rate of water molecules along the axons, along those fibers.
50:01
Mm-hmm.
50:02
Those neurons right and we can measure in white matter the water is going to move faster along the length of the axon Then across and so we have a directional idea of like how Diffusion of water is going and diffusion of water is a proxy for kind of how neurons talk to one another Where is it?
50:19
How's the traffic and flowing right?
50:20
Yeah, exactly.
50:22
So what they did was
50:23
was use this DTI, diffusion tensor imaging, of 51 patients with disorders of consciousness.
50:30
And you can show that there's significantly lower striatum to GPE streamlines.
50:37
That's the break on the thalamus.
50:39
That prediction is lower in vegetative state versus minimal consciousness patients.
50:45
No way!
50:47
Yeah.
50:47
So just to take a step back, right?
50:50
From this model...
50:52
it made a prediction about where specifically in this region of the brain that is controlling the flow between the thalamus and the cortex is the point where the degradation of this specific area is what's actually driving the problem.
51:07
And this connection is where we would find it.
51:11
And so it said this is the road and the house where the problems are.
51:16
Arise out of.
51:17
And then we said, okay, well, we have all this actual MRI data from real patients.
51:23
So can we look at this road, this house and real patients that the model predicted you would see degradation in an outsized way as compared to surrounding area?
51:35
Yeah.
51:36
I mean, no, they compared vegetative versus minimally cognitive.
51:39
Right, right.
51:40
Because we've got those two sets.
51:41
Yes.
51:41
And if there's a difference, there should be some statistically...
51:44
Some delta between those two.
51:46
The vegetative will have more of the degradation than the minimally... Oh, my God.
51:51
Yeah, yeah.
51:52
And so on the left-hand side, we see that significance.
51:57
There is a significant difference.
51:58
On the right-hand side, he did sneak the sin.
52:00
The p-value is only 0.07.
52:01
So it's not significant.
52:04
Yeah.
52:05
But it's there.
52:06
And perhaps if you were to pull the two data, it would become even more significant.
52:09
But there's at least something.
52:13
And the N is low.
52:15
So obviously significance is not... We're not going to get Higgs boson 10 to the minus 5 significance here.
52:21
Where literally to do that, they had billions of...
52:26
particles colliding with each other, right?
52:28
So you'll never get the significance level of traditional physics, but the effect, it seems, is there, right?
52:36
And with more and more data, perhaps it's going to be an even bigger effect.
52:39
Makes total sense.
52:40
It was really kind of a proof of, like, minimally viable proof that this hypothesis of this little...
52:49
Prediction that I have, it's probably true.
52:51
It was good enough for nature neuroscience.
52:53
Yes.
52:53
Not bad.
52:54
Exactly.
52:55
The second thing that they did was predict, this whole AI architecture predicted, that there's increased synaptic coupling between inhibitory interneurons, meaning those negative feedback neurons that I was telling you about, these are sometimes called fast spiking PV plus neurons in the cerebral cortex.
53:13
Those negative feedback neurons in our cerebral cortex have a lot of...
53:18
Coupling in between them.
53:20
So it's like the negative feedback is coupled to the negative feedback, which is coupled to the negative feedback.
53:24
Right?
53:24
That is what the AI model is predicting.
53:28
So to test this, you go to transcriptomics.
53:32
Transcriptomics means I'm going to now read...
53:35
the mRNA that is in my cerebral cortex, okay?
53:39
And I'm going to see what genes are being expressed in patients that have vegetative state and patients that don't.
53:47
Okay, healthy versus those in coma.
53:49
Yes.
53:50
What they find is there's an upregulation of two genes, the VGF and the SCG2.
53:57
These are genes that, when expressed in these interneurons,
54:02
They drive synaptogenesis, meaning they drive this feedback mechanism, this coupling.
54:08
Yeah, yeah.
54:09
That.
54:09
Right?
54:10
This is really...
54:11
This one's pretty significant.
54:12
Yeah.
54:13
This one's way more than...
54:15
In both those genes.
54:16
Yeah, yeah, yeah.
54:16
Right?
54:17
And this is, again, data that's out there.
54:20
Right, right, right.
54:21
Right?
54:22
Exactly.
54:22
And I think the...
54:24
This is kind of, again, in terms of like a frame of reference for...
54:28
I've seen the conversation around quote-unquote AI be very different within some corners of the science community versus the general public because a lot of researchers view it in this way where it's like, okay, this can be a sort of intermediary layer.
54:44
where I can rapidly prototype and generate predictions and rapidly be able to test that against real data in a way where I don't have to abuse my postdocs.
54:57
Yeah, yeah, yeah.
54:58
Just abuse the undergrads.
55:03
here's a thousand images labeled right no right sorry undergrads are always gonna that's that's tough yeah it's a tough life we've all been there we all want those letters of recommendation you know you gotta you gotta work the sweat this is really interesting though because now there's the predictions are also on two very different planes
55:25
Yes, that's a very good point.
55:26
You know what I mean?
55:27
Like in terms of what's... Yeah, because when we think about those three components of our city, right?
55:33
The cerebral cortex is the homes and the businesses.
55:38
That PV plus interneuron thing has to do with that part.
55:41
Right.
55:42
And then the...
55:45
The brain imaging part, that one had to do with the thalamus and the relay stations.
55:50
Correct.
55:50
Right, right, right.
55:51
So all three parts of the mesocircuit hypothesis are working together.
55:54
And this thing is giving predictions, as you said, on all of them.
55:57
Right.
55:58
Right.
55:58
Right.
55:59
In very interesting ways.
56:01
And again, this continues to go back to the fact that we have all these existing tools like transcriptomics.
56:06
Mm-hmm.
56:10
enables the ability to know what genes are being expressed yeah and one thing i just want to say is like to the team that that made this paper happen right i admire the resourcefulness right because it's one thing to have an ai model and make predictions right and it's another thing to think okay here are the predictions how can i
56:33
make the argument that this is real.
56:34
Right.
56:35
How can I check?
56:36
It's always about checking.
56:37
Yes.
56:37
And they were so resourceful that they found these kinds of data sets.
56:41
You know, maybe they talked, I don't know exactly the details.
56:44
I don't know if these are open or not, but you, you go talk to somebody who has that data set.
56:47
Right.
56:47
You meet someone at a conference and they're like, Hey, actually you can look into mine.
56:51
Put me as an author You know so It's very clever I think that part is nice The whole from the ideation Of like let's give it a try To then how are we going to be able to Prove to reviewer two That this is not all nonsense Yeah yeah exactly I got lunch with Daniel about three weeks ago And he was telling me about this paper And he told me like The thing predicted this And the first thing I asked well yeah but like
57:21
you know,
57:22
How do you know the model isn't just bullshitting?
57:25
And then he told me about these two techniques.
57:27
I was like, okay, that's actually pretty dope.
57:29
Yeah.
57:29
Well, as soon as it's out, we'll cover it.
57:32
You know?
57:34
So, yeah, this was pretty cool.
57:36
Okay, the final thing we're going to talk about is a proposed strategy for awakening patients with coma, right?
57:43
Right.
57:43
Because at the end of the day, that's what really matters.
57:46
Clinically, that's how it started, and that's how we're going to end it.
57:50
So...
57:51
The intervention that they looked for and how to test was deep brain stimulation.
57:56
This is, again, the idea of you test, you put in an electrode deep into the brain, and then you give it electrical activity.
58:04
And you try to wake up the neurons, okay?
58:07
Now, the researchers here, what they did was test a bunch of targets in their biophysical model and say, what if I provide stimulation here?
58:16
What if I provide stimulation here?
58:17
What's the best target in my three-component model, right?
58:22
Because there's a bunch.
58:23
There's cortex.
58:24
There's thalamus.
58:24
There's subthalamic nuclei.
58:26
There's the palladium.
58:28
And each of those has their own little subpopulations.
58:32
So we can get really granular now, right?
58:35
We can get really granular and try to think what is the best target.
58:39
This is so good.
58:40
They settled on the subthalamic nucleus.
58:43
It had an overwhelmingly standout result for consciousness recovery in their model.
58:49
Okay.
58:49
Okay?
58:50
Then again, they tested this with humans.
58:53
They tested 130 hertz stimulation in the subthalamic nucleus in six awake human patients.
59:00
And the cortical CNN, the cortical consciousness detector that they had previously trained, detected a significant shift towards optimal consciousness.
59:14
This is crazy.
59:15
With only stimulation in the STN.
59:18
If they went anywhere else, only the sub-dilemic nucleus showed this effect.
59:21
If they did it in the cortex, if they did it in the palladium, didn't show any significant increase in consciousness.
59:27
Yes.
59:28
And that's what the model predicted.
59:29
So are you saying we found the boss's office?
59:32
Yeah, it seems we found... We may have found the boss's office.
59:35
Where if you turn the lights on and off, the boss will wake up.
59:38
Yeah, right, right, right.
59:41
This is...
59:42
I'm so mad because it's so clever.
59:45
Each of the three layers, also in how we went through it, of like detection, characterization, and evaluation almost is like, is so good.
59:55
Because I think...
59:57
The part that's getting me about this piece is because you've set up like the genomic algorithm and the model, the brain simulator, the way they thought to set it up that way means that you almost have this like alpha fold adjacent kind of thing where you can in silica like test stuff.
1:00:18
Yeah.
1:00:20
To try to narrow the sandbox of where to look.
1:00:24
Exactly.
1:00:24
Right?
1:00:25
And, like, that's so valuable.
1:00:27
Yeah.
1:00:27
Yeah, exactly.
1:00:28
It doesn't necessarily have to be the exact answer.
1:00:30
No.
1:00:30
But if it's even somewhat directionally correct, that's unbelievably valuable.
1:00:35
Yeah, dude, it's great.
1:00:36
And one thing that one thing that's kind of funny is so Daniel, he is himself a science communicator.
1:00:43
Right.
1:00:43
So he's got an Instagram and a TikTok.
1:00:45
And he did like a short three minute video being like, hey, my paper came out and he kind of explained it at a very high level.
1:00:51
I mean, here we've gone really in depth.
1:00:54
So.
1:00:55
At that high level, he was describing his generative AI framework, and there was someone in the comment who was like, that's not really AI, right?
1:01:01
Because you're using a biophysical model.
1:01:03
And it's like, dude, that's the point.
1:01:06
That's the point.
1:01:06
It's like without a biophysical model, I wouldn't be able to make these predictions of, oh, it's the PV plus neurons in the cortex.
1:01:13
It's this highway in between the striatum and the palladium.
1:01:17
It's the subthalamic nucleus that we need to probe, right?
1:01:21
It's just going to be subthalamic.
1:01:22
random neuron in some random layer of a trillion parameter model what good is that the the a hundred percent i still come back to one of the there's so many interesting insights in this entire papers architecture like the experimental design architecture but i think a key piece was uh
1:01:42
You can get traceability and then subsequently reproducibility because of that brain simulator.
1:01:50
Yeah.
1:01:52
And the way in which it is explicit, it is...
1:01:59
What's the, there's the terminology, and this may not be correct.
1:02:02
There's this terminology where probabilistic versus deterministic, and it's more on the deterministic side and less on the probabilistic side, which in science, it's a good thing.
1:02:12
Yeah, it's very much a good thing.
1:02:15
And then finally we now have an idea that it should be the sub-dalamic nucleus that we target, which is this tiny, like, lentil-sized structure
1:02:24
Right?
1:02:25
It's actually also quite incredible if you think about it.
1:02:27
Like, that is the boss's office.
1:02:29
Right.
1:02:30
Right?
1:02:30
For at least stimulation of the brain.
1:02:32
Consciousness could be a distributive thing.
1:02:34
But just stimulating that tiny structure of maybe, let's say, hundreds of thousands of neurons at the max is going to wake up a brain that is 86 billion neurons.
1:02:47
Right?
1:02:48
Consciousness is a crazy thing.
1:02:50
And we're getting closer to...
1:02:52
So would it actually a correct analysis?
1:02:54
Because there's going to be probably some thoughts about, oh, like consciousness is, like you discussed at the beginning, a very expansive thing.
1:03:03
And really all we're identifying in this last piece is, I keep using the grill lighter example.
1:03:10
This might just be the igniter, the little button you press on the grill to turn consciousness on.
1:03:16
Yeah.
1:03:16
But that's all we're saying right now.
1:03:19
I don't know what the grill being on means.
1:03:21
Right.
1:03:22
We don't know how big the grill is.
1:03:23
Is it a black stone?
1:03:24
Yeah.
1:03:25
Is it whatever?
1:03:27
Exactly.
1:03:27
We're just saying we may have found where we can wake an unconscious from this vegetative or unconscious state.
1:03:35
Yes.
1:03:36
And the interesting dovetail of how.
1:03:42
The 2007 paper we talked about that identified this idea of minimally waking consciousness was actually a key step in order for this to actually also work because that was where we started to zoom in and be able to validate the mesocentric model.
1:03:59
It's just like everything has to work.
1:04:03
Yeah, one after the other.
1:04:04
It has to stack.
1:04:05
Right, yeah.
1:04:06
In order for this to even be possible.
1:04:08
Yeah.
1:04:08
This was really quite nice.
1:04:09
It was a good paper.
1:04:11
Daniel does a lot of really cool work with Koma.
1:04:15
He's actually also in an organoid lab at UCLA.
1:04:18
So one of these days, we'll come by UCLA and see your lab, Daniel.
1:04:22
But until then, nicely done.
1:04:24
Very, very nicely done.
1:04:26
Obviously, the...
1:04:28
This is early, but there's a huge...
1:04:32
The unlocks that are there, you know, you could talk about it forever, just in terms of not only the clinical or medical context, but even for the philosophy of mind kind of stuff, having some mechanistic details.
1:04:47
Yes, that's big.
1:04:48
And I think...
1:04:51
The part that is going to be most influential for this paper is actually the way that he went about using this biophysical model.
1:05:02
Yes, I totally agree.
1:05:03
And tweaking.
1:05:04
I totally agree.
1:05:04
That's a very interesting way to do things.
1:05:07
I totally agree.
1:05:08
And I can already imagine...
1:05:11
it being used for all sorts of stuff, material science, like physical things, you know, like we've got really good models about how atoms work with each other.
1:05:21
Exactly.
1:05:22
Make a GAN genetic network type thing, you know, like that is the part that is really like, I think mechanistically very interesting to me.
1:05:30
The methods is very nice.
1:05:31
I mean, obviously the results are incredible, but to me, the, the one I'm most impressed about is the way that he used this biophysical model.
1:05:39
Um,
1:05:40
As someone who's coming more from the technology and software world, it is naturally what I'm able to better engage in at a deep level.
1:05:52
But I agree with you.
1:05:53
I think they've basically created a conceptual framework around how to design things.
1:06:04
ML and AI architecture as a, as an end to end system and process that is subject matter agnostic.
1:06:15
Mm hmm.
1:06:16
Yeah, yeah.
1:06:17
Like it could work.
1:06:17
Yeah, yeah, yeah.
1:06:18
That thing can work.
1:06:19
And a whole variety of – the brain was just the first target.
1:06:23
But the exact same conceptual framework could work in a variety of different areas.
1:06:27
Yeah.
1:06:27
Particularly the data feedback loop and validation process, et cetera.
1:06:31
Really, really – I just – this was great.
1:06:33
Yeah, good stuff.
1:06:33
Yeah.
1:06:34
Well done, Daniel Toker et al.
1:06:36
Again, this was in Nature Neuroscience on March 24th.
1:06:40
Fresh hot off the presses.
1:06:42
Again, there is nowhere else on the internet, not even on the direct pages of the authors themselves, where you're going to get this level of excitement, passion, and deep dive.
1:06:54
They need a break.
1:06:54
Daniel needs a break anyway.
1:06:55
So it's our job to do this and step in.
1:06:58
Yeah, we'll see what he thinks about it.
1:07:01
You actually guys did pretty well.
1:07:02
You did pretty well.
1:07:03
No, just another fantastic paper.
1:07:05
And we just really want to thank those of you who have stayed this long and listened to the episode for staying with us and enjoying this journey of curiosity and discovery.
1:07:18
I will do a brief pause to ask if we would like to do a comment.
1:07:23
for the audience.
1:07:26
Gan.
1:07:27
Gan.
1:07:28
This has become our shtick.
1:07:29
We can't think of anything.
1:07:32
What's another alternate...
1:07:36
full form of GAN.
1:07:37
Yeah, acronym definition for what?
1:07:39
Generative Adversarial Networks.
1:07:40
Come on, you guys can think of something better.
1:07:42
Something great.
1:07:43
We really appreciate you all.
1:07:44
My name is Lester Nare, joined as always by my co-host and our resident PhD, Krishna Chowdhury.
1:07:51
This is the last reminder I'll make on this.
1:07:53
We are moving to single-story episodes multiple times a week.
1:07:58
We have gotten confirmation from our deep listeners that...
1:08:01
You do appreciate it and you like it.
1:08:02
And there was a good idea that we will think through.
1:08:06
I don't know if actually, I don't even know if you saw this, but the rundown, when we do the rundown at the end of the week, we don't really go in depth.
1:08:14
And there's a lot of folks who would like us to go in depth or any number of those particular stories.
1:08:19
So,
1:08:20
We will look into having basically a patron system to vote on what rundown story should we follow up on for a deep dive.
1:08:27
It's a great idea.
1:08:28
Excuse me for not remembering the commenter who said it, but you can comment again.
1:08:32
We do see it.
1:08:32
Thank you so much.
1:08:34
We will see you all later this week.