Projects
![]()
No projects found
Investors
New
DAO Treasuries
DATs
Lend / Borrow APRs
Chains
Categories
Spaces
Introducing: Inside the Lab with @amelie_iska
Recorded: Dec. 4, 2024 Duration: 1:05:39Space Recording
Full Transcription
Thank you. Hello, hello. hello hello testing testing we hear you loud and clear stanley
sorry how do you say is it amelie amelie yeah amelie okay we can hear you we can hear you as
well amelie awesome that's great yeah we usually Amelie. Okay, we can hear you as well, Amelie. Awesome, that's great.
Yeah, we usually give like, you know,
two to three minutes for people to sort of filter in,
but, you know,
Stanley is the man in charge this week,
so I'll defer to y'all
actually. You are too
kind, sir. We are just here, humbly
excited to talk some science, and
thank you so much for having us on
the show. I do have to say the
D-Site mic is my favorite mic. Oh, I'm humbled and yes, let's leave it at humbled. I appreciate
the kind words. I'm sure Aaron and Crypto Shrimp do as well. Oh, and I cannot also tell you how
excited I am to see Crypto Shrimp in the audience
because he's actually a specialist in one of the diseases that our guest here is working
on right now.
And I think more than anything, I'm so excited for people to get to hear from Amelie.
Amelie, good morning.
How are you?
I'm good.
Yeah, things are good.
How are you guys?
Super excited to have you on. I was just chatting with Stanley earlier this morning about a lot of your background and
oh my gosh, it's amazing.
So excited to dive in today.
Yeah, it's going to be fun.
I think we're going to talk about a lot of cool stuff.
Absolutely. I guess to kick it off, so the D-Sci mic is an ongoing weekly space focused on different D-Sci topics. And this week, we're narrowing in on introducing this kind of new series called Inside the Lab with Stanley, who's up here right now, Science Stanley.
And we actually had connected back in 22, did a podcast with him back then. So super excited to
have him back up here and be collaborating to also bring you onto the stage right now and hopefully dive into a whole bunch of other cool topics in the future too.
I guess Stanley, to kick off the space today,
I would love for you to share some of the inspiration for Inside the Lab.
Yeah, so happy to share and my goodness Aaron I think it's been um maybe like two years since
since since our interview together yeah and you know I think that was in the first uh
desai summer and and man are we coming out of winter hot these past couple weeks.
And so very excited to be kind of rousing and sort of getting out to communicate here.
And I think this will be, yeah, just our kind of first toe in the water of a new space we're calling In the Lab.
So for Desai In the Lab, think inside the actor's studio. We want to, on a very personal level,
talk to some of the incredible researchers who are out there making a difference. And really,
because we got to make sure these people are supported, right? That's why we're here doing
Desai stuff. We've seen that the system as it exists doesn't support the right research all
the time. But also, I think we want to celebrate
these people, right? Because when you look at the world, and my goodness, when you look at the media,
I don't feel like the people who are being talked about are the people I want to be like,
or I want my kids to be like, but man, there are some scientists out there who are just heroic. And we got one of them with us here
today. And I think this space is just going to be the start of diving kind of deep with these people
who are telling really important stories. So, Amelie, let's kick off with just you and me.
How did we start working together? I mean, you kind of found me. I think you found me through Hugging Face for the
most part because I was doing a lot of open source stuff. I think you met me after I'd been
doing it for maybe like a year. And you were kind of like, okay, you're doing some really cool stuff.
Some of the people that I've been mentoring are doing projects based on your stuff. Maybe let's try to hook you up with some
compute and do some interesting projects or something. And then, yeah, it just kind of
went on from there. And we started trying to do some cool stuff with some proteins.
You know, that is very accurate.
And I have to say, when you work with Amelie,
the continual focus on cool proteins is infectious.
But I would even start the story a little earlier.
And 100%, Amelie is a bit of an open source superstar
on Hugging Face.
If you want to Google Amelie Hugging Face,
stuff comes up. And, you know, Amelie, maybe if you found to Google Amelie hugging face, uh, stuff comes up and,
you know, Amelie, maybe if you found a link of one of your favorite articles, we could post it
in the notes here so people could check it out. Um, Amelie's work is as sophisticated as it is
accessible. So, you know, if you're hearing about cool stuff like the Baker labs, Nobel prize,
and you want to see how actual, uh, scientists are using those tools, like Amelie is the person whose work you want to see.
Also, and Amelie, this is why I was just so excited when I met you.
I mentor for the Google Summer of Code for an open source project called DeepChem,
which is mostly postdocs from Stanford and other universities.
postdocs from Stanford and other universities. And this past summer, we had a cohort of 15
students, research fellows supported by Google, which is an unusual number for a single project,
reflective of the excitement about the space. And Amelie, literally more than half of our students
did projects based on your open source work. I think I've told you that before, right?
You have.
Sometimes it's a little hard to believe,
but yeah, you've said that a few times.
And I think it's really, I don't know,
it's really great that I can like plant some seeds
and maybe get a few good ideas out there
and then see somebody pick that up and run with it.
I, you know, sometimes, you know,
I can come up with a good idea and I can execute on it and
carry it through to the end. And sometimes I just kind of like, I can see where a thing might go,
but maybe just don't have the resources to fully follow through on a thing. And then to have other
people come in and be like, oh, you know what? This is a good idea. Let's pick this up and run with it.
And then have someone turn that into a fully-fledged project and see that really bloom.
It's a really fun thing to watch.
And it makes me feel very much part of the community.
Ah, well, I love to hear that. And then on this note, too, you say you talk about the resources that are needed to do this work.
Ah, well, I love to hear that.
And I think there's there's something really interesting to talk about there, because because what I was doing right before we started this this sort of technical enablement work where, you know,
Amelie is one of the scientists we're supporting with compute.
Erin, did we ever talk about the Stanford Rare Disease AI hackathon?
We touched base kind of as it was going on, but I never heard the outcomes from it.
So please share.
Oh, well, it was just this fun thing.
I work largely in AI post-training,
so I configure models into sort of products
to be used for science and engineering.
And I had a really fun idea with the Snyder Labs,
where I was a research architect at Stanford Medicine
for a number of years, that we should use
like decentralized technology to do post-training and specifically for rare disease medicine.
Because, you know, for many patients, it's like five to six years before they can get
a diagnosis.
And that's if they're lucky.
And it really often is a case of just finding the right doctor because, you know, the doctors
their illness are rare, like the illness.
And so we actually ran this program.
And yeah, Aaron, I think I was telling you about it when we first started.
It went so well.
We ended up getting to work with AnyScale, Roche, Hugging Face, basically every partner
partner we talked to about the idea wanted to to work with us and then we we had you know over a
we talked to about the idea wanted to work with us.
couple hundred doctors actually involved in the post training talking to the models evaluating
them so it all went really well except you know the part where we were getting compute access for
the researchers like was consistently painful for for a number of different reasons you know
it was partially cultural. These are
people from academia who aren't used to how fast tech moves and the processes involved in getting
resources. Reciprocally, these are kind of technology people on the other side of the fence
who might not totally understand the motives of academic researchers. And so we really realized that like that was a core
bridge that needed to be built. And so that was when we started this project called the
Rare Compute Foundation. It's a nonprofit trying to bring compute resources to researchers. And
yeah, that's why we're so lucky to have Amelie to speak to today and to hear about her work,
because she was the first researcher we supported. And Amelie, what does that look like?
We got you like a pretty big machine, didn't we? Yeah. Yeah. So mostly I've been working on
a nice node with, I think, four A100s on it. So that's been pretty fun to work with.
I will say, though, like for some models, that's definitely necessary,
but for some other AI models for biochem,
you know, a lot of these can be run
on consumer grade hardware.
You don't always need like a huge giant node
to run some of them.
So like RF diffusion, for example,
like you can easily run that on like a 3090 or a 4090
and generate some new proteins for
yourself. And Erin, I think I was telling you a little bit about RF diffusion, but maybe that
would be a fun tool to double tap on if that's maybe interesting. Yes, let's dive into that.
So, Amelie, what is a diffusion?
Yeah.
So, I guess maybe the right way to think about it, if you guys are familiar with DALI or Stable Diffusion or Mid Journey,
any of these models where you type in text and you get out some kind of image or video,
those are diffusion models. They're basically taking, you train them on images. And what you
do is you add a little bit of noise to the image, and then you train
them to remove the noise. And you progressively add more and more noise. And they eventually learn,
given complete Gaussian noise, to turn it into an image, right? So you can do this with proteins.
So you can do this with proteins.
The idea is you split up the backbone into frames and add noise to the backbone.
And then have the model denoise the backbone and generate a new protein for you in a very similar way that you would type in text and generate a new image.
a new image. RF diffusion is a diffusion model. And in that way, it's very similar to like Dolly
RF diffusion is a diffusion model.
or mid-journey or stable diffusion or what have you. But it has a unique...
What I always say is, I'm so sorry, I just love this joke. It's a mid-journey for molecules.
It is. Yeah, it's exactly mid-journey for molecules. It has a unique architecture, though, that I was always really
fascinated by. And I was always really impressed by, too. Like, the mathematics behind it are just
really beautiful. My background originally was in physics, where we work with crazy stuff like
tensor fields and vector bundles. And yeah, you get to take some of the really cool tools off the shelf for diffusion.
For those of y'all who might not know some of those words, though, I always like to remember
that diffusion is kind of like trying to figure out how a Lego is built by smashing it with
a bowling ball and then playing the tape of it exploding in reverse, kind of watching
how the pieces come together.
And so the math that Amelie is talking about kind of relates to the kind of size of the bowling ball and where you're smashing it and how you're getting the tool.
Incredible tool.
And then also just for both of our hosts, like so that we can kind of ground this, what's y'all's familiarity with kind of proteins?
Would it be helpful if we said anything about, you know, the kind of nature of proteins and why they're so important?
I think for the context of this space, laying some of those foundations would be helpful for any other folks listening in later
and just kind of tying,
okay, how does this all connect to biology?
What does that end up looking like?
And starting to build up some of those blocks.
Oh, that sounds amazing.
And please, you guys too, thank you for having us
and jump in and any questions.
And it would be so fun to receive an answer.
But yeah, proteins, you guys, it's so simple.
They're the little self-building nanorobots that run all of life.
Just that.
I'm just kidding.
Yeah, right.
But it is kind of literally true. The proteins are sort of the
actors of biology. We hear about our genetic code, but the genetic code to some degree exists to
orchestrate a symphony of proteins that make up our cells, make up the processes of our biology.
make up our cells, make up the processes of our biology. So often in medical research,
when we're looking at a disease, particularly a rare genetic disease, but often, you know,
in just about any kind of disease state, there's some protein that's shaped wrong,
it's doing the wrong thing, it's interacting with another system in a problematic
way. So in many cases, like we might have a protein associated with an illness. For example,
helping me target a protein called PKA. And this is a protein that mediates metabolism. So when I
say that proteins kind of like are the,
the dancers or the actors in the dance of life, that's,
that's totally what I mean.
When he needs to communicate that growth should happen or that metabolism
should happen,
there's a protein that actually is sort of emitted and then regulates that
growth. And then a certain case,
this protein can have a mutation where like you know like when you have
a piece of printer paper going through the printer and it kind of sticks and then it double prints
over one section that's like almost literally what happens with this protein it's what's called a
fusion mutation and so basically like one piece gets overwritten and so this one protein that
is supposed to regulate metabolism just kind of
puts the pedal to the metal on metabolism and it leads to an oncological process. So this particular
single protein mutation causes a very nasty liver cancer. And so Amelie, what does it look like for
you? Because we took a call recently where you sat down with an oncologist and we were talking about these proteins and how to
target them. Right. So, so when, when you get a protein, how, how do you kind of follow it from
there? Yeah. So I think that's like, that's hard to answer because it, it really depends on the
target. Every target has its own personality and its own quirks and it,
you know, its own special considerations and everything.
And so I think like the first step is always to just sit down and study your
target and really get to know the target and,
and figure out like how to approach the problem,
what kind of molecules to design things like that.
Like sometimes it might be as simple as just,
you know, figuring out what region to target and designing a binder. And in other cases,
it might be a little more complicated. You may need to do something a little different.
So I don't know. I think it just depends. For some cancer targets, like a basic inhibitor,
you're basically, you're just picking a region on the
protein, usually something that differentiates it from other proteins. And you're trying to
block some kind of interaction by designing a binder, which is something that RF diffusion,
for example, is really good at. But you can also do like more complicated things, like try to
design like
allosteric modulators to, to like change the confirmation of the protein.
Um, and that's a little trickier.
Uh, you have to, you have to find sites on the protein that are, that are allosteric
sites or that are likely to, uh, trigger that conformational change.
Um, and, and you have to kind of think about like a little bit about the dynamics of the protein to
some extent. But there are different techniques that you can use. So there's a technique that
Baker Lab uses called two-sided diffusion, where they can kind of induce certain confirmations
of the target to some extent by sort of diffusing it and then letting RF diffusion predict its
structure while it's designing the binder. So yeah, I guess the short answer is it really
depends on the target and you have to just get to know your target before you start.
And I get this vibe too, working with Amelie, that she comes to know these proteins very personally.
It's pretty remarkable, the intuition she has.
I see your hand up, but please just jump in.
Like, we're so excited to just kind of get input.
I didn't want to interrupt or, like, ruin the vibe or anything like that.
But just, like, you know, the more I learn about, you know, biology, right, the more I think it's, like, a the more i learn about you know biology right the
more i think it's like a wonder that we're alive at all right um you know learning all these details
about proteins um but so here's my sort of you know i mean yeah i mean i sort of my my research
in my past life as a scientist was funded by the by uh you know uh pharmaceutical companies and so
forth so i have a smattering of, even though it's an engineering background.
But like, I guess my question, perhaps like a layman's type question here would be,
does this work relate to prions at all?
Because like, you know, I've learned about prions.
I'm like, wow, that's kind of scary stuff.
Ooh, ooh, very scary stuff.
I apologize for everyone that I'm going to be talking about prions this early. And first answer is it might. And that's something to be concerned about. Okay, I should I should talk about prions. Amelie, you know prions, right? I don't know if we've ever spoken about prions.
Yeah, we, we've mentioned them to each other a little bit. And I do know a little bit about them. I wouldn't say I'm an expert or anything, but I understand sort of how that works. Yeah.
Okay. So first off, proteins are so freaking cool. It's just one of those things where you're
like, how did anyone ever think of that? And no one did, right? It was nature. Nature is real smart.
nature. Nature is real smart. Proteins start as amino acid sequences. And there are actual real
scientists in the audience, so forgive me for kind of butchering things, but you can almost
picture an amino acid chain as a string of pearls. And the pearls on the string are kind of like
molecules, and they have electromagnetic properties.
So you can almost imagine that each of these strings of pearls in a line all kind of attract each other differently with magnetic and electrical force.
And then they kind of use that to program their own creation. So certain clusters of amino acids group together,
other ones group together, the groups sort of start to interact and fold over each other. And
it allows these very complex shapes to be created. Amelie, I don't know if it may be possible we
could add the gap junction protein to the thread just so that people can see sort of like
how beautiful these things can be.
But this protein Amelie is gonna share, for example,
is the protein that goes between our cells
and allows our cells to communicate and execute metabolism.
And yeah, so amazing that something like that
can just build itself from a string of pearls.
All of the pearls that exist, all of the folded pearls that exist in our sort of like biological
reality, let's just say they have a certain shape. They're kind of folded in a certain kind of way.
And prions are these little protein fragments that are folded differently. And not just that, but in a very complicated process, when prions interact with normal proteins,
they will slowly cause those proteins to fold in the incorrect way that is not compatible with our biology.
And so this is the cause of something you guys might have heard of called
mad cow disease. Mad cow disease happens when you eat a burger that has prions in it,
and it may cause you to get a neurodegenerative illness in 30 years. There's no test for it.
Prions require temperatures, I believe, over thousand degrees to, um, to break them
down.
So there are even concerns, for example, and again, sorry to the black pill you guys with
this, um, that there, there may be prion disease rampant in deer populations and that, uh,
deers may be kind of like depositing prions in nature.
So, um, yeah, this is a big, a big scary topic.
In the data sets that we have to train these models, we don't have an enormous amount of
prions. And so while I haven't worked in this area specifically, I think we would struggle to
generate new prions. Prions are also not super compatible with our biology.
So yeah, I would say of the 10,000 things I'm very afraid of with AI-driven biology,
that's lower down on the list.
And actually, I almost have optimism because of prions,
not because of obviously their negative interactions with the human proteome, but more like their physical properties, right?
Like they are strong, resilient molecules.
And, you know, that's my positive take on prions is they may lead to some really interesting synthetic biology.
synthetic biology. Yeah. I mean, with, sorry, sorry to jump in again, Stanley, you know, with
this, uh, you know, uh, blind watchmaker kind of idea, you know, of, of why, why we exist sort of
thing. We'd be better off being made of prions. I guess they're like indestructible sounds like,
but, um, looking, thank you for, you know, I learned a lot on this space. Um, you know, uh,
if I'm quiet or if I have to drop off, I actually have to take a call in a couple of minutes. But I already feel like I learned a ton.
So thank you.
Thank you, Stanley and Emily for sharing.
Oh, dude.
No, thank you for your questions.
And jump in with more while you're here.
And Aaron, you as well.
And this is really like our pilot.
Like I was just having a convo with Aaron and we were talking about how there really are these cool stories to share, you know? And I think that kind of growing this
into a format where we can be talking to cool scientists is going to be a lot of fun, man. So
please jump in with more questions and we'll hope we'll have you next time we're chatting too.
And on that note, we did bring up the gap the gap junctions and um um aaron i think i was
telling you before that our actual first um uh compound that we we sort of helped amelie uh
develop in the lab will will actually be going into wet lab production in in the new year and
it's a a molecule that's related to those gap junction proteins. Would
you like to tell us a little bit about that, Amelie? Yeah, sure. It's pretty exciting for me.
I think it's a really fun project. And I think the implications could be really big. So that's also really exciting. So basically, I guess, like, it's kind of a funny story. I
randomly reached out to Michael Levin. And I was like, you know, like, if I could design some
molecules for you, what kind of molecules would you need? And we got to talking and he was talking
to me about gap junctions, which is one of his big things.
And he was like, we don't really have any good gap junction openers. And if we had some gap junction openers, that would be really useful. And so I was like, okay, let's see if we can get
some gap junction openers. So I jumped right on that and started trying to figure out how to design some gap junction openers. And so the current approach is designing a small peptide binder,
basically to lock connections into a specific position
so that the gap junctions allow ions to travel through.
And so for those of you who aren't familiar with gap junctions,
they're these channels between
our cells that connect our cells up to each other.
And they allow our cells to communicate with each other.
And they pass, ions pass through them.
And sometimes other small molecules and things can pass through them too.
But they have electrical current flowing, right?
Because the ions are passing through them.
And so it's like, it's a way for your cells to communicate with each other and sort of
orchestrate larger scale behavior among themselves and have sort of like a collective intelligence
to them.
And this is actually such a cool topic I want to tap on, double tap on.
Hey, Erin, are you able to see?
I think Amelie did post a picture of this protein we're talking about, this gap junction protein.
Are you able to see?
Isn't it kind of cool looking?
I just pinned it up above.
And when I first saw it, I'm like, oh, my gosh, this is so pretty.
Everyone has to see this.
Yeah, so it should be pinned up above.
100%. There's some real art. There's some real art here, isn't it? And I have to say,
there's some literal art. Amelie, for my birthday, I think, gave me my favorite present I've ever
received, a beautiful print of this protein that we designed a binder for. So honored to receive
that, Amelie. But yeah, these gap junctions are so interesting. And then also, if anyone's not
familiar with Dr. Michael Levin, he is worth checking out. He is, to my mind, like a top 10
research biologist in the world right now. His episode of the Lex Friedman podcast is really
a good listen. Beyond just his sort of like traditional bona fides in biology,
he is a person with a pretty radical philosophy about biology. And for that reason, even if he's quite, you know, web to side, I think of him as a real
radical and a trailblazer and one of the most incredible people to be trying to work with and
support in the whole world right now. He is a person who has a pilling. He has ideas that are
so radical, you can be leaven pilled. And basically, like, Amelie, you'll have to correct me if I'm wrong, but the
leaven pill has to do with what you were, I think, alluding to, the intercellular intelligence of our
bodies and our tissues. If you do check out the beginning of the Lex Friedman podcast,
Dr. Leaven is talking to Lex about an ocean organism called a planarian, which is a little wormy dude.
And when you cut a planarian in half, both halves regrow into a full planarian, which both fully regrown planarians somehow seem to share the memories of the original planarian.
And it kind of begs the question, right? How do the
memories get from the old brain to the new brain? Right, Amelie? Yeah, it's pretty wild. I think
it's a really good indication that there is actually a lot of information contained outside
of the brain that's encoded in these electrical patterns. So it's really wild.
So the way that our cells communicate with each other with these gap junctions with the ion flow,
they sort of implement these electrical patterns that remind me a lot of Turing patterns.
And if you can like figure out what the right electrical pattern is, you can
give a top-down signal to the tissue to do a certain thing. And so he does these really wild
experiments where he goes in and he like gives this like top-down signal by modulating the gap
junction activity. And then it implements an electrical pattern. And then the electrical
pattern kind of orchestrates the large scale behavior of the tissue. And then you can do like
really crazy things like regrow amputated limbs in animals that don't normally do that,
or like turn off cancer without editing the genome or anything, or, like, correcting birth defects, like, just all kinds of really
cool stuff. It's pretty wild.
You know, I think that it is, like, we've discovered there's, like, a firmware layer,
you know, that's kind of operating the body intercellularly.
And yeah, it's going to be so interesting to see how that picture develops.
I personally like not qualified to comment, but it seems like the gentleman is on to something.
But also, and Aaron, this is just like so fun, though.
But also, and Aaron, this is just like so fun, though.
These tools that Amelie is talking about, it's kind of like, you know, like you could think about biology as a place where previously we just had a sledgehammer.
And now we kind of have a scalpel.
We can start interacting with these systems kind of on their turf.
You know, we can kind of look at particular molecules
and figure out, you know, what we need to do to them.
It's really a pretty exciting time to be working.
Yeah, I think it's like one of the most exciting times
to be working.
Like just the applications of AI to biochemistry
are blowing my mind the last couple of years, for sure.
I feel like that is a really interesting transition into kind of a larger, maybe observation that the rest of the world is starting to be able to see as well of just even
at the Nobel Prize level, AI being heavily involved there. And Emily, I know you work with
some of that work as well, possibly out of the Baker lab. I would love for you to comment on kind of that whole progress happening now at
that larger scale and some of your work there. Yeah, I mean, I would say Baker Lab is a huge
inspiration for me. I definitely love following their work and taking inspiration from them and also just like trying to get really deep into their techniques and learning some of their really specialized techniques.
So I think like the Nobel Prize, I think, was, I don't know, not that much of a surprise to some people, especially for like David Baker
and DeepMind. People have been saying for a while, this work is going to get a Nobel Prize.
And I think it was spot on. I think the way that it changed what we can do with these molecules is just very paradigm changing.
Like we, before just a year or two ago, we didn't have good tools to design new proteins to do
specific things. And the fact that that's come onto the stage now really changes the game. Because we, I mean, like Stanley was saying, we have a scalpel now.
And we can go in and design these very customized molecules to do these very specific things.
And accomplish things that we just couldn't do before.
Because we didn't have the ability to design these kinds of proteins.
And I think RF diffusion was one of the first indications of this.
RF diffusion has been out for a couple of years now,
and it spurred a whole flurry of newer, similar models
that tried to build on that work and do similar things.
And I think it's a game changer. It really is. And I think also, like, it showed
us, like, what AI can really do for biochem. And you see, like, newer diffusion models or flow
matching models that are coming out that are doing, like, even crazier stuff now. Like there's a new model out of MIT
that I was really impressed by
that I think could be taken a lot further called MDGen.
And it does really wild stuff.
Like it will, it's basically Sora, but for molecules.
And so it'll generate like molecular dynamics trajectories
for you, like Sora would generate a video for you.
And so you can get these like basically like videos of molecules from the,
from this AI model.
And it lets you do really interesting things.
Like you can like interpolate between two different states,
which,
which could be used for like calculating like binding free energy,
which is like a huge, a a huge add, a huge unlock. You can do inpainting
with it, where you go in and you mask out part of the molecule, and then it generates some new
stuff for you where you masked out part of it. It's really cool. I'm very impressed by not just what Baker Lab did, but what everybody else did after to build on top of that, because people really took it and ran with it. And I think it's opening up a lot that we just didn't have before.
I couldn't agree more. And it is obviously a big milestone for the Nobel to go both to
the DeepMind team and Baker's Lab. It's kind of funny, though, because the writing has been on
the wall there for quite a while. If you Google like AlphaFold citation graph, you can see these graphs of like how many papers cited alpha
fold after it was released and exponential growth has completely changed the field. You know,
previously, like we needed systems like folding at home to just get like 80% of the way to predicting
to just get like 80% of the way to predicting how a protein folds.
And then now, yeah, just a single GPU gets you like 99% accuracy.
It's a pretty remarkable thing.
And I think too, like if it had been like a human who figured that out
or, you know, a traditional method,
I think the Nobel would have come quite a bit faster
because, yeah, there's definitely some, you know, interesting tension between kind of traditional biology and AI tools.
Something people might be familiar with is there are certain companies that are now being called tech bio companies instead of biotech companies.
tech companies, which is kind of some funny wordplay, but it is reflective of the fact
that, you know, a biology project that used technology might have previously been, you
know, 10 biologists and one or two engineers.
And now that could be flipped.
It could be, you know, one or two biologists telling a large team of engineers the relevant
science and then sort of like a lot of the complexity of the project being on the engineering or the computer science side.
So this is just one thing I really wanted to shout out the Baker Lab for my perspective, being the most Web3, Web2 research group out there.
most web three, web two research group out there.
Amelie, maybe you can say a little bit more about this,
but Dr. Baker just has a reputation
as like the kind of guy you want to work for, right?
Like he wants his researchers to be successful.
He sort of, I mean, there's a lot of discussion
about how he fosters really unorthodox collaboration, right?
Like you can kind of find the right people in his group to work with, something like that? Yeah, I've heard so many good
things about him, coming from like, you know, former students and collaborators and stuff.
Like the vibes are just good all around, I would say.
And the vibes, you know, vibes,
effermal, but all too concrete when it comes to the outcome of projects,
because, you know, it's kind of funny, like on these tech bio projects, you have,
you know, people with really deep expertise in different areas who might have very different
language, but also who sort of like might have
not ever been a neophyte in their field for many, many years. And, you know, that's one of the
things that's challenging here is sometimes I have to ask incredibly stupid questions on a
biological level. And then reciprocally, sometimes I'm explaining stuff to people who,
you know, I feel just incredibly humbled to talk to that, you know, you would be learning as a freshman in a machine learning program.
So, yeah, I just think this is a really cool Web3 DSi part of the story is it's a new field that's kind of growing at the moment when DSi is growing.
that's kind of growing at the moment when D-Sci is growing.
And it's a field that really needs decentralized collaboration
without the kind of centralized power structures.
And case in point, you know, it was the DeepMind team
that created the first protein folding system.
But, you know, Baker Labs at University of Washington,
where they really have a healthy pattern of collaboration
between biologists and computer scientists, University of Washington, where they really have a healthy pattern of collaboration between
biologists and computer scientists, they kind of regularly outperform AlphaFold on the benchmarks.
That's super cool. Are you able to kind of tie this into some of the bigger
trends happening across open science, pairing that in with DSI or like open
source versus closed source. And I don't know, I feel like that's just coming to mind as you're
talking about vibes and collaboration and any trends or things you're kind of seeing across
each of these domains. Oh, my goodness.
I sure can, Aaron.
Thank you so much.
And let me start by sharing something really fun that happened for me this year.
So my birthday is October 20th.
And I had always thought I was just the luckiest person in the world to share a birthday with
Snoop Dogg, if you can believe that.
That's amazing.
But I found out.
Fun fact.
I know, right?
I know, right?
Come on, come on.
I'm a California guy.
What could be better?
But in all seriousness, though, and I think Mr. Doggy Dogg would also love this,
And I think Mr. Doggy Dog would also love this because we share a birthday with something potentially even more special, more important, something called the PDB Data Bank.
Amelie, what is the PDB Data Bank?
Yeah, so the Protein Data Bank, it's a structural database for proteins.
I mean, there are other kinds of molecules in there too.
Like they have DNA and RNA and small molecules and so on.
But it was started like 50 years ago.
Is that right?
And then it's been going and they've just been slowly producing more and more data each year of structures,
of protein and molecular structures.
And this is actually what AlphaFold was trained on.
And there aren't that many in there because it's actually really hard to get experimental structures, right?
Like cryo-EM, X-ray crystallography, and NMR are all like pretty hard to do and they're time
consuming. And so it takes a while to get an experimental structure out of these and actually
uploaded into the protein databank. And so over the years, we've accumulated like something like
200,000 experimental structures
or something like that.
And, you know, it just it reached a point and it came came at just the right time that
the data set was just big enough and the AI was just far enough along that they converged
and an alpha fold was able to actually happen.
And I think like if that data hadn't been there,
AlphaFold would never have been able to have been trained in the first place. And so the PDB is like, you know, in large part is why AlphaFold even exists in the first place.
And I think that brings us back to a really good point that, you know, I'm sure most AI people will know. The devil is in the details,
and your data set really matters. And so if you have a good data set that's been well curated
that you can train on, you can usually solve the problem. You can design some clever architecture
architecture and train on that data and figure out a way to leverage it to solve your problem.
and train on that data and figure out a way to leverage it to solve your problem.
And the Protein Data Bank really enabled that. And it also enabled, you know, things like RF
diffusion and subsequent models that came after it for designing proteins, because that's what
those were trained on as well. You know, you alluded to a joke I was going to make, Amelie.
I was going to make a joke like, and hey, imaging a protein is really easy, isn't it?
Yeah, no, not really.
Because it's really hard.
No, I mean, listen, we didn't know they existed for thousands of years, right?
They're good at being hidden.
They're so small. But yeah, so it's a pretty remarkable thing that's been built by the collective effort of scientists all over the world.
This library of Alexandria of protein structures built protein by protein over decades, again, by thousands of scientists like probably mostly funded with public money,
creating this public good in the PDB data bank. And quite literally, that's the key piece of
these tools. And that's how we have them. So there's some interesting like questions here,
right? Like, how is it that a company like DeepMind trains a model on a public good that hundreds of thousands,
millions of hours of work and public money went into, and then they own the model?
I mean, I'm a red-blooded capitalist just like the next guy, but I don't know if the
incentives are well aligned there, are they?
Yeah, I mean, I think it's a good point. just like the next guy, but I don't know if the incentives are well aligned there, are they?
Yeah, I mean, I think that's, I think it's a good point. I also will say, like, on this point,
Baker Lab is really good about this. Like, all of their models thus far have been open sourced and released with a very, like, permissible license, like an MIT license or something similar.
And so we have a lot of these AI models out in the open source available to use to just anybody
for anything that they want to use it for, in large part because of Baker Lab. And I would say like AlphaFold, DeepMind people,
like they definitely open source
their first two versions, right?
Like AlphaFold 2 was open source.
And I think there was like a little bit
of like friendly competition going on
between DeepMind and Baker Lab, you know?
Cause you can kind of see like one of them
will release a model and then the other one will release a similar model or a better model or something. And they'll kind of
go back and forth and really like push each other to like grow and expand and do new interesting
things. But yeah, I think like not having an open source model that's trained on open source data,
on open source data, there's definitely a bit of a bit of cognitive dissonance.
there's definitely a bit of a bit of cognitive dissonance.
Oh, my God, I'm so sorry to realize I was host and I have a mute everyone button. I'm so sorry,
Amelie, I would never want to mute you. I totally agree. You know, forgive me for
you were but then I unmuted you just the last last word or two, we lost and it was,
was my fault. So forgive me everybody. I hang on every word when Amelie's speaking,
so I wouldn't want anyone to miss one. But I was going to take a second to, to shill for Desai as,
as the Desai guy and say Desai could kind of deal with this, right? Like in, in a, in a fully
imagined Desai world, you could attach these data to Web3
primitives and you could kind of like, you know, track and understand exactly what data and what
contributions went into allowing a model to be trained. And, you know, we certainly have many
problems to solve to get to that point. But, you know, it feels like a lot of stuff that used to
be uncompensated because it was too hard to kind of track will be beautifully
tracked on the blockchain so yeah this is very exciting oh you know looks like we have a request
for someone to come up and we're in our last 10 minutes so anyway you guys thank you so much for
being here for this kind of experimental space this week and maybe what we'll do is we'll um
you know open it up to some questions to kind of wrap us up.
Does that sound good?
That sounds great.
Vega, go for it.
Hey, Aaron.
Yeah, this is such a cool space.
Hey, Emily, Merrick, Stanley, what's up?
It's been so long since to hear your voice.
And obviously, when I come in and hear you, I'm like, wow, dropping some awesome wisdom.
And I'm always so grateful and honored.
So thank you for everyone's time and stuff.
I guess my question.
Well, first, I had a bit of insight and was like, wow, I can't.
The whole like string of pearls with these with the way these things are built.
I was like thinking blockchain the entire time.
Like they're just connecting and gravitating towards that.
So I was like, wow, that's an interesting way to think about it.
So that was pretty cool.
And then I guess, yeah, and I guess my question is a couple.
But the first one is I've been kind of messing around and I'm a meditation teacher on the side of certain things.
And I've been kind of practicing a lot and coming back to impermanence and chaos and whatnot.
And I sometimes think about, well, I was doing research on inflammation and overthinking.
And I was like, okay, how can I, one, develop a way, and I guess in this conversation, I was thinking about this molecule, interleukin-6 and cytokines
and was kind of thinking, okay, how can I connect those
or how can I think of it in a way where we can figure out
an analogy of like a thermal paste?
Like is it a way, because it has to do with like inflammation
and I was like, how can I upgrade or look into figuring that out
with these models or these open source tech to kind of develop
a way to upgrade that sort of molecule for people who are in the ground level of sort of
for lack of a better word suffering or processing their their experience on this on this plane
so i'm kind of thinking like how can i look into that as a person on my end, since I could get one of those GPUs and work on that.
But yeah, I kind of like wanted to see if you guys have any, like, ideas or thoughts.
Oh, man.
Yeah.
You know, I've done a bit of research data science.
Have you ever heard of the organization MAPS?
organization maps no tell me more oh it's um the multidisciplinary association of psychedelic
No, tell me more.
science and so they're a big group that's developing oh yeah molecules that people feel
oh and such a cool project worth worth reading more about too um, kind of like a DSI project before it was DSI,
the founder of MAPS, Rick Doblin. Same way, you know, we're seeing these incredible founders
going like, there needs to be an Athena DAO for female health and there needs to be a cryo DAO,
you know, we need to advocate for this unaddressed field. Mr. Doblin kind of like did that back in the 70s or 80s. Like he
had some experiences where psychedelics helped him and people around him with, with kind of trauma
and processing and was kind of like, man, we got to get these things mainstream. We got to get the
science done so people can be helped. And, you know, MAPS is the result of that. It's like a
fully grown pre-Dao-Dao kind of thing. And yeah,
I would just say, like, I think for people interested in consciousness, interested in
mindfulness, like what I think MAPS is doing is pretty profound. And, you know, not just because
of the medications, but the way that they're using them, you know, they're actually developing
data-driven clinical methodologies that, you know, use low-dose psychedelics to minimize risk and then sort of
interface that treatment with talk therapy and other modalities that already have proven efficacy.
And yeah, so I would just say like that's one area I'm really passionate about. On the more
biochemical front, Amelie, have we ever talked about like why mushrooms invented psilocybin?
No. Tell me more.
Oh, it's an interesting story.
But then I also see we got my brother Ken up here.
And Ken, you weren't here for the earlier part of the show where I was celebrating the, you know, quickly thawing Desai spring we're wandering
into. But yeah, Ken, one of my dearest and oldest bros in the Desai space, how you doing?
Hey, how's it going, Stanley?
Can you hear me?
So good, man. And yeah, I can hear you so good. And man, you would have loved the work we did for MAPS
because we were actually working with patients
who were doing talk therapy
and we were doing sentiment analysis on their journals.
Yeah, super cool.
I think you mentioned this before.
Something with trees, right?
Well, that was one of the interesting learnings is we found that it was unusually common for
patients experiencing healing, as demonstrated through our sentiment analysis, to be
expressing how they felt as if they were a tree.
Right. Go on, go on.
How do you mean by tree?
Yeah, kind of an interesting one.
Like rooted into the ground, spreading out branches,
the sense between it. Tell me more.
But yes, yes, that and more.
And again, like just,
it seemed to be that those poetic metaphors were what grew out
of the experience for these patients. And very timely. I see that we also have another one of
my DSI heroes, Ed, in the audience. And Ed, really one of the most incredible leaders on the subject
of regenerative agriculture, you know, which is a field that tells us that the
ecosystem of the soil and of everything that cohabits the soil is much more interconnected
than we ever would have thought. So on some level, it kind of makes sense, right? Like these patients
are recognizing something profound, you know, that in the same way that a tree has these deep roots
to connect to its environment, You know, we as humans
who are social animals, like our relationships are kind of like our roots. And when you experience
trauma, like to me, it makes so much sense to see that as your roots being damaged or the nutrients
you need not flowing through your roots correctly. And that was just kind of what we found is that
over and over again, that particular metaphor occurred to these patients. And that was just kind of what we found is that over and over again, that particular
metaphor occurred to these patients. And, you know, it's sort of surprising on one level,
but then on another level, like when we take psychedelics, we kind of all see the same patterns,
right? And so that's, you know, at least once one sort of indication, there is some really deep
shared experience going on. Absolutely. I want to bring it back to the ground level,
you know, in a day-to-day person.
Just like the person, right,
take the train conductor
or the mother who can't afford to have time
to sit with shrooms.
You know, those people,
I kind of think about them a lot.
And I'm like, okay,
well, what are the things that are interesting
that can be prescribed in a way
or look at interleukin-6
or these cytokines, these proteins
to interact and
kind of bring it to a better story. And like for most people, it's pretty funny because you can
mention this to a person who can't afford health insurance in the States and they go to a doctor
who's not like an upgraded doctor, right? Who doesn't, like you said earlier, you have to find
the right doctor and that takes like time to get your diagnosis right because they have their
biases, right? And that's kind of one of the things right because they have their biases. Right.
And that's kind of one of the things I kind of I sit with. And I sit there and I guess in some experiences, you'll have someone who'll say, oh, what do you think about supplements?
And they'll be like, well, they're a load of whatever.
And they'll associate like fish pills.
And you're like, well, that's not what I meant by supplements.
What about ashaglanda?
You know, those are easy supplements.
And that's connected to like
anti-inflammatory stuff and i guess people are now kind of realizing that like in meta psychology at
least in the behind the scenes of it you're looking um at like over inflamed experiences in your mind
so you're constantly thinking and causing chronic sort of inflammations of stress and survival
so in a way you're breaking yourself down.
So I'm trying to figure out how can I create like a better thermal paste for
people who are over processed right now, a small upgrade, tiny,
so that they can kind of go about their day a little bit better.
And I guess smoking weed might dampen those things, but I guess,
I don't know, I'm trying to thinking other things in interesting ways.
Yeah. So that's an interesting one.
And myself, too, like, you know, just did some data analysis for this kind of research.
And what's really exciting is this research is happening.
There are significant clinical trials going on with low-dose MDMA. And it kind
of seems like MDMA is a little more effective in this context than cannabis. Cannabis does
alleviate some of the stress of a traumatic experience, but it seems to be slightly through
kind of like distracting you from the processing of that experience, not allowing that process to happen with better thermal processing.
But the low-dose MDMA does seem to allow the patients to exist in these areas that are very uncomfortable, but not have the sort of heat that leads to that cycle of rumination. Early moments, and early moments
where maybe we're going to find some even more incredible molecules. Anyone here know the books
Tycal and Phycal? Oh, wow. You guys, what's going on here go go find ty cal and
no one said anything no listen i mean i'm kidding too i was just maybe looking for some emojis but
but yeah these are kind of weird books but they're they're also really fun
amelie have you heard of these these books oh we lost her that's right she had a meeting at 10
well anyway I'll do the Tychal Fychal thing real quick and I also just thank
everyone for coming um Erin I'm feeling excited I think this pilot
we hit into some good energy don't you think
definitely uh yeah good kind of crowd listening in and
covered a lot of really interesting topics. And we didn't even get through kind of the a regular thing, and we hope everyone will come back.
And quite frankly, some of y'all want y'all to come up on the stage,
like Crypto Shrimp, Ed, Ken.
Like, we would love to do an episode about you guys
and, you know, what you guys are working on.
And, yeah, just to close,
because I really think this is such an inspiring Desai story.
There was a woman named Alexander Shuglin.
And I believe in the, I always get the data this wrong.
I forget if it was the post-World War I era,
or I think it was the post-World War II era.
So it would have been maybe the 50s or something.
Yeah, that makes sense to me.
And he was, I think he was the dean of the organic chemistry department at Harvard,
something like that. Like he was a very prestigious high-level academic and his specialty was
phenethylamines and related organic compounds. And these are
psychedelic compounds at Harvard. And this was before psychedelic compounds were outlawed.
So when they were outlawed, Dr. Shuglin had a choice. Did he want to continue pursuing this
research area he really believed in and was passionate about? Or did he want to switch to a
more centrally approved topic? And he decided to follow his passion and his love. And the story
in these two books, Tycal and FICAL, FICAL stands for phenethylamines I have known and loved.
And what Dr. Shuglin did is he actually moved to the Bay Area, set up a psychedelic research lab in his basement and synthesized hundreds and hundreds
of potential novel psychedelics based on the phenethylamine core chain. So he sort of took
this core Lego block that seems to have a lot of psychedelic activation and then, you know,
tried to stick lots of other molecules on it in different ways to kind of see what they would do.
And then he would actually host parties in the Bay Area where people would come and they would all dose a psychedelic
no one had ever tried before. To me, that's the craziest thing I've ever heard. Not something I
would recommend, but a pretty crazy story and a pretty cool story of certainly decentralized
science. And yeah, so this book is pretty fun.
And the first half of the book contains the story
of Dr. Shuglin's life and information
about all the different kind of stuff he built.
And the second part contains recipes
for all the compounds and reports
from the people who he gave them to try.
So it's a pretty unusual work
and definitely worth checking out.
That's such a great share.
I hadn't heard of that before.
So definitely a lot of things from this space
to go check out afterwards.
Some of those things are replied down below
in the comments.
So go check out those episodes,
those models,
all of it.
Yeah.
So much good content.
Yay.
So fun, you guys.
Thanks for having us.
And we'll be doing this again soon.
So I hope to catch you guys again soon.
And Erin,
thank you so much for helping us make this happen.
I really had a blast.
Yes.
And for everyone listening in,
this was the kickoff of a new series called Inside the Lab,
where we're really focusing on talking with scientists,
people doing kind of lab work day-to-day,
and hearing from kind of that more insider perspective,
different updates in their field and their work
and being able to showcase some of that to the outside world
beyond just kind of the small circles of people
that might be in these a bit niche down fields.
So if you have any recommendations of someone else
we should have on in the upcoming episodes,
definitely tag them.
Feel free to DM Stanley or myself, Aaron, and we'd love to get them scheduled out.
Amazing.
In the meantime, have an amazing week, and we'll see you back here same time next week.
Thanks so much, everyone. Later next week. Thanks so much everyone.
Later everyone, thanks so much. I can't wait to catch everyone back in the lab. Love it, thanks, bye.
