Dmitry Korkin: Evolution of Proteins, Viruses, Life, and AI
嘉宾：Dmitry Korkin
全文文稿 · 按对话顺序保留完整内容，深入理解上下文。
本篇目录13 章节
§01Transcript · 00:00–10:49§02Transcript · 10:52–20:54§03Transcript · 20:55–31:59§04Transcript · 32:03–42:21§05Transcript · 42:23–53:47§06Transcript · 53:50–1:04:03§07Transcript · 1:04:05–1:14:06§08Transcript · 1:14:11–1:25:04§09Transcript · 1:25:13–1:35:26§10Transcript · 1:35:27–1:46:27§11Transcript · 1:46:30–1:56:36§12Transcript · 1:56:39–2:08:23§13Transcript · 2:08:27–2:12:30§01Transcript · 00:00–10:49
The following is a conversation with Dmitry Korkin,
his second time in the podcast.
He's a professor of bioinformatics
and computational biology at WPI,
where he specializes in bioinformatics of complex disease,
computational genomics, systems biology,
and biomedical data analytics.
He loves biology, he loves computing,
plus he is Russian and recites a poem in Russian
at the end of the podcast.
What else could you possibly ask for in this world?
Quick mention of our sponsors,
Brave Browser, NetSuite Business Management Software,
Magic Spoon Low Carb Cereal,
and Eight Sleep Self-Cooling Mattress.
So the choice is browsing privacy, business success,
healthy diet, or comfortable sleep.
Choose wisely, my friends.
And if you wish, click the sponsor links below
to get a discount and to support this podcast.
As a side note, let me say that to me,
the scientists that did the best apolitical,
impactful, brilliant work of 2020
are the biologists who study viruses without an agenda,
without much sleep, to be honest,
just a pure passion for scientific discovery
and exploration of the mysteries within viruses.
Viruses are both terrifying and beautiful.
Terrifying because they can threaten
the fabric of human civilization,
both biological and psychological.
Beautiful because they give us insights
into the nature of life on Earth
and perhaps even extraterrestrial life
of the not-so-intelligent variety
that might meet us one day
as we explore the habitable planets
and moons in our universe.
If you enjoy this thing, subscribe on YouTube,
review it on Apple Podcasts, follow on Spotify,
support on Patreon, or connect with me
on Twitter @LexFriedman.
And now, here's my conversation with Dmitry Korkin.
It's often said that proteins
and the amino acid residues that make them up
are the building blocks of life.
Do you think of proteins in this way
as the basic building blocks of life?
- Yes and no.
So the proteins indeed is the basic unit,
biological unit, that carries out
important function of the cell.
However, through studying the proteins
and comparing the proteins across different species,
across different kingdoms,
you realize that proteins are actually
much more complicated.
So they have so-called modular complexity.
And so what I mean by that is an average protein
consists of several structural units.
So we call them protein domains.
And so you can imagine a protein as a string of beads,
where each bead is a protein domain.
And in the past 20 years,
scientists have been studying the nature
of the protein domains.
Because we realized that it's the unit.
Because if you look at the functions, right?
So many proteins have more than one function.
And those protein functions are often carried out
by those protein domains.
So we also see that in the evolution,
those proteins domains get shuffled.
So they act actually as a unit.
Also from the structural perspective, right?
So some people think of a protein
as a sort of a globular molecule.
But as a matter of fact,
is the globular part of this protein is a protein domain.
So we often have this, again,
the collection of this protein domains
align on a string as beads.
- And the protein domains are made up of amino acid residues.
- Yes.
- So we're talking--
- So it's--
- So this is the basic,
so you're saying the protein domain
is the basic building block of the function
that we think about proteins doing.
So of course, you can always talk about
different building blocks.
It's turtles all the way down.
But there's a point where there is,
at the point of the hierarchy,
where it's the most, the cleanest element block
based on which you can put them together
in different kinds of ways to form complex function.
And you're saying protein domains,
why is that not talked about as often in popular culture?
- Well, there are several perspectives on this.
And one, of course, is the historical perspective, right?
So historically, scientists have been able
to structurally resolve to obtain the 3D coordinates
of a protein for smaller proteins.
And smaller proteins tend to be a single domain protein.
So we have a protein equal to a protein domain.
And so because of that, the initial suspicion
was that the proteins, they have globular shapes,
and the more of smaller proteins you obtain structurally,
the more you became convinced that that's the case.
And only later when we started having
alternative approaches, so the traditional ones
are X-ray crystallography and NMR spectroscopy.
So these are sort of the two main techniques
that give us the 3D coordinates.
But nowadays, there is huge breakthrough
in cryo-electron microscopy.
So the more advanced methods that allow us
to get into the 3D shapes of much larger molecules,
molecular complexes, just to give you
one of the common examples for this year.
Right, so the first experimental structure
of a SARS-CoV-2 protein was the cryo-EM structure
of the S protein, so the spike protein.
And so it was solved very quickly.
And the reason for that is the advancement
of this technology is pretty spectacular.
- How many domains is the, is it more than one domain?
- Oh yes, oh yes, I mean, so it's a very complex structure.
- It's complex.
- We, you know, on top of the complexity
of a single protein, right, so this structure
is actually, is a complex, is a trimer.
So it needs to form a trimer in order to function properly.
- What's a complex?
- So a complex is agglomeration of multiple proteins.
And so we can have the same protein copied in multiple,
you know, made up in multiple copies
and forming something that we called a homo-oligomer.
Homo means the same, right?
So in this case, so the spike protein is an example
of a homotetrimer, homotrimer, sorry.
- So it needs three copies of a--
- Three copies, exactly. - In order to.
- Exactly.
- We have these three chains, the three molecular chains
coupled together and performing the function.
That's what, when you look at this protein from the top,
you see a perfect triangle.
- Yeah.
- So, but other, you know, so other complexes
are made up of, you know, different proteins.
Some of them are completely different,
some of them are similar.
The hemoglobin molecule, right, so it's actually,
it's a protein complex.
It's made of four basic subunits.
Two of them are identical to each other,
two other identical to each other,
but they are also similar to each other,
which sort of gives us some ideas
about the evolution of this, you know, of this molecule.
And perhaps, so one of the hypothesis is that, you know,
in the past, it was just a homotetrimer, right?
So four identical copies, and then it became, you know,
sort of modified, it became mutated over the time
and became more specialized.
- Can we linger on the spike protein for a little bit?
Is there something interesting
or like beautiful you find about it?
- I mean, first of all,
it's an incredibly challenging protein.
And so we, as a part of our sort of research
to understand the structural basis of this virus
to sort of decode, structurally decode
every single protein in its proteome,
we've been working on the spike protein.
And one of the main challenges was that
the cryo-EM data allows us to reconstruct
or to obtain the 3D coordinates
of roughly 2/3 of the protein.
The rest of the 1/3 of this protein,
it's a part that is buried into the membrane of the virus
and of the viral envelope.
And it also has a lot of unstable structures around it.
- So it's chemically interacting somehow
with whatever the heck it's connecting to.
- Yeah, so people are still trying to understand.
So the nature of and the role of this 1/3,
'cause the top part,
the primary function is to get attached
to the ACE2 receptor, human receptor.
There is also beautiful mechanics
of how this thing happens, right?
So because there are three different copies of this chains,
there are three different domains, right?
So we're talking about domains.
So this is the receptor binding domains, RBDs,
that gets untangled and get ready
§02Transcript · 10:52–20:54
to get attached to the receptor.
And now they are not necessarily going in a sync mode.
As a matter of fact--
- It's asynchronous?
- So yes, and this is where another level
of complexity comes into play,
because right now what we see is,
we typically see just one of the arms going out
and getting ready to be attached to the ACE2 receptors.
However, there was a recent mutation
that people studied in that spike protein.
And very recently, a group from Umayyad,
a group from Umayyad Medical School,
we happened to collaborate with groups.
So this is a group of Jeremy Luban
and a number of other faculty.
They actually solved the mutated structure of the spike.
And they showed that actually, because of these mutations,
you have more than one arms opening up.
And so now, so the frequency of two arms going up
increased quite drastically.
- Oh, interesting.
Does that change the dynamics somehow?
- It potentially can change the dynamics of,
because now you have two possible opportunities
to get attached to the ACE2 receptor.
It's a very complex molecular process, mechanistic process.
But the first step of this process
is the attachment of this spike protein,
of the spike trimer, to the human ACE2 receptor.
So this is a molecule that sits
on the surface of the human cell.
And that's essentially what initiates,
what triggers the whole process of encapsulation.
- If this was dating, this would be the first date.
So this is the--
(laughing)
- In a way, yes.
- So is it possible to have the spike protein
just floating about on its own?
Or does it need that interactability with the membrane?
- Yeah, so it needs to be attached,
at least as far as I know.
But when you get this thing attached on the surface,
there is also a lot of dynamics
on how it sits on the surface.
So for example, there was a recent work in,
again, where people use the cryo-lactone microscopy
to get the first glimpse of the overall structure.
It's a very low res,
but you still get some interesting details
about the surface, about what is happening inside,
because we have literally no clue until recent work
about how the capsid is organized.
- What's a capsid?
- So a capsid is essentially,
it's the inner core of the viral particle
where there is the RNA of the virus,
and it's protected by another protein, N-protein,
that essentially acts as a shield.
But now we are learning more and more,
so it's actually, it's not just this shield.
It potentially is used for the stability
of the outer shell of the virus.
So it's pretty complicated.
- And I mean, understanding all of this
is really useful for trying to figure out
like developing a vaccine or some kind of drug
to attack any aspects of this, right?
- So, I mean, there are many different implications to that.
I mean, first of all,
it's important to understand the virus itself, right?
So in order to understand how it acts,
what is the overall mechanistic process of this virus,
replication of this virus, proliferation to the cell, right?
So that's one aspect.
The other aspect is designing new treatments, right?
So one of the possible treatments
is designing nanoparticles.
And so some nanoparticles that will resemble the viral shape
that would have the spike integrated,
and essentially would act as a competitor to the real virus
by blocking the ACE2 receptors
and thus preventing the real virus entering the cell.
Now, there are also, you know,
there is a very interesting direction
in looking at the membrane,
at the envelope portion of the protein
and attacking its M protein.
So there are, you know, to give you a, you know,
sort of a brief overview,
there are four structural proteins.
These are the proteins that made up
a structure of the virus.
So spike, S protein, that acts as a trimer,
so it needs three copies.
E, envelope protein, that acts as a pentamer,
so it needs five copies to act properly.
M is a membrane protein.
It forms dimers, and actually it forms beautiful lattice.
And this is something that we've been studying
and we are seeing it in simulations.
It actually forms a very nice grid or, you know,
threads, you know, of different dimers
attached next to each other.
- Just a bunch of copies of each other,
and they naturally, when you have a bunch of copies
of each other, they form an interesting lattice.
- Exactly.
And, you know, if you think about this, right,
so this complex, you know, the viral shape
needs to be organized somehow, self-organized somehow.
Right?
So, you know, if it was a completely random process,
you know, you probably wouldn't have the envelope shell
of the ellipsoid shape.
You know, you would have something, you know,
pre-random, right, shape.
So there is some, you know, regularity
in how this, you know, how these dimers
get attached to each other in a very specific, directed way.
- Is that understood at all?
- It's not understood.
We are now, we've been working in the past six months
since, you know, we met.
Actually, this is where we started working
on trying to understand the overall structure
of the envelope and the key components
that made up this, you know, structure.
- Wait, does the envelope also have
the lattice structure or no?
- So the envelope is essentially is the outer shell
of the viral particle.
The N, the nucleocapsid protein,
is something that is inside.
- Got it.
- But get that, the N is likely to interact with M.
- Does it go M and E?
Like, where's the E in--
- So E, those different proteins,
they occur in different copies on the viral particle.
So E, this pentamer complex,
we only have two or three maybe per each particle.
- Mm-hmm. - Okay?
We have thousand or so of M dimers
that essentially made up,
that makes up the entire, you know, outer shell.
- So most of the outer shell is the M--
- M dimer.
- It's the M protein. - And lipids.
- When you say particle, that's the viron,
the virates, the individual virus.
- The single, yes.
- Single element of the virus, single virus.
- Single virus, right.
And we have about, you know,
roughly 50 to 90 spike trimers, right?
So when you, you know, when you show a--
- Per virus particle.
- Per virus particle.
- Sorry, what did you say, 50 to 90?
- 50 to 90, right? - Cool.
- So this is how this thing is organized.
And so now, typically, right,
so you see these, the antibodies that target,
you know, spike protein, certain parts of the spike protein,
but there could be some, also some treatments, right?
So these are, you know, these are small molecules
that bind strategic parts of these proteins,
disrupting its functioning.
So one of the promising directions,
it's one of the newest directions,
is actually targeting the M-dimer of the protein,
targeting the proteins that make up this outer shell.
Because if you're able to destroy the outer shell,
you're essentially destroying the viral particle itself.
So preventing it from, you know, functioning at all.
- So that's, you think, is,
from a sort of cybersecurity perspective,
virus security perspective,
that's the best attack vector?
Is, or like, that's a promising attack vector?
- I would say, yeah.
So I mean, there's still tons of research needs to be,
you know, to be done.
But yes, I think, you know, so--
- There's more attack surface, I guess.
- More attack surface, but, you know,
from our analysis, from other evolutionary analysis,
this protein is evolutionarily more stable
compared to the, say, to the spike protein.
- Oh, and stable means a more static target?
- Well, yeah.
So it doesn't change, it doesn't evolve
from the evolutionary perspective so drastically
as, for example, the spike protein.
- There's a bunch of stuff in the news
about mutations of the virus in the United Kingdom.
I also saw in South Africa something,
maybe that was yesterday.
§03Transcript · 20:55–31:59
You just kind of mentioned about stability and so on.
Which aspects of this are mutatable
and which aspects, if mutated, become more dangerous?
And maybe even zooming out,
what are your thoughts and knowledge and ideas
about the way it's mutated,
all the news that we've been hearing?
Are you worried about it from a biological perspective?
Are you worried about it from a human perspective?
- So I mean, you know, mutations are sort of a general way
for these viruses to evolve, right?
So it's essentially, this is the way they evolve.
This is the way they were able to jump
from one species to another.
We also see some recent jumps.
There were some incidents of this virus
jumping from human to dogs.
So there is some danger in those jumps
because every time it jumps, it also mutates, right?
So when it jumps to the species and jumps back,
so it acquires some mutations that are sort of driven
by the environment of a new host, right?
And it's different from the human environment.
And so we don't know whether the mutations
that are acquired in the new species are neutral
with respect to the human host or maybe damaging.
- Yeah, change is always scary.
But so are you worried about, I mean, it seems like
because the spread is during winter now
seems to be exceptionally high,
and especially with a vaccine just around the corner
already being actually deployed,
is there some worry that this puts evolutionary pressure,
selective pressure on the virus for it to mutate?
Is that a source of worry? - Yes.
Well, I mean, there is always this thought
in the scientist's mind, what will happen, right?
So I know there've been discussions
about sort of the arms race between the ability
of the humanity to get vaccinated faster
than the virus essentially becomes resistant to the vaccine.
I mean, I don't worry that much
simply because there is not that much evidence to that.
- To aggressive mutation around a vaccine.
- Exactly.
Obviously there are mutations around the vaccine.
So the reason we get vaccinated every year
against the season of flu. - Because there's mutations.
- But I think it's important to study it, no doubts.
So I think one of the, to me, and again, I might be biased
because we've been trying to do that as well.
But one of the critical directions
in understanding the virus is to understand its evolution
in order to sort of understand the mechanisms,
the key mechanisms that lead the virus to jump,
the Nordic viruses to jump from species to another,
that the mechanisms that lead the virus
to become resistant to vaccines, also to treatments.
Right?
And hopefully that knowledge will enable us
to sort of forecast the evolutionary traces,
the future evolutionary traces of this virus.
- I mean, what, from a biological perspective,
this might be a dumb question,
but is there parts of the virus that if souped up
through mutation could make it more effective
at doing its job?
We're talking about this specific coronavirus.
'Cause we were talking about the different,
like the membrane, the M protein, the E protein,
the N and the S, the spike.
Is there some-- - And 20 or so more
in addition to that.
- But is that a dumb way to look at it?
Like which of these, if mutated, could,
have the greatest impact, potentially damaging impact
on the effectiveness of the virus?
- So it's actually, it's a very good question.
Because, and the short answer is we don't know yet.
But of course there is capacity of this virus
to become more efficient.
The reason for that is, you know,
so if you look at the virus, I mean, it's a machine, right?
So it's a machine that does a lot of different functions.
And many of these functions are sort of nearly perfect,
but they are not perfect.
And those mutations can make those functions more perfect.
For example, the attachment to ACE2 receptor, right?
Of the spike, right?
So, you know, is it,
has this virus reached the efficiency
in which the attachment is carried out?
Or there are some mutations that,
that still to be discovered, right?
That will make this attachment sort of stronger,
or, you know, something more, in a way more efficient
from the point of view of this virus functioning.
That's sort of the obvious example.
But if you look at each of these proteins,
I mean, it's there for a reason.
It performs certain function.
And it could be that certain mutations
will, you know, enhance this function.
It could be that some mutations
will make this function much less efficient.
So that's also the case.
- Let's, since we're talking about
the evolutionary history of a virus,
let's zoom back out and look at the evolution of proteins.
I glanced at this 2010 Nature paper
on the, quote, "Ongoing expansion of the protein universe."
And then, you know, it kind of implies
and talks about that proteins started
with a common ancestor, which is, you know,
kind of interesting.
It's interesting to think about, like,
even just like the first organic thing
that started life on Earth.
And from that, there's now, you know, what is it?
3.5 billion years later, there's now millions of proteins.
And they're still evolving.
And that's, you know, in part,
one of the things that you're researching.
Is there something interesting to you
about the evolution of proteins
from this initial ancestor to today?
Is there something beautiful,
insightful about this long story?
- So I think, you know, if I were to pick a single keyword
about protein evolution, I would pick modularity,
something that we talked about in the beginning.
And that's the fact that the proteins
are no longer considered as, you know,
as a sequence of letters.
There are hierarchical complexities
in the way these proteins are organized.
And these complexities are actually going
beyond the protein sequence.
It's actually going all the way back to the gene,
to the nucleotide sequence.
And so, you know, again, these protein domains,
they are not only functional building blocks.
They're also evolutionary building blocks.
And so what we see in the sort of,
in the later stages of evolution,
I mean, once this stable, structurally
and functionally building blocks were discovered,
they essentially, they stay, those domains stay as such.
So that's why if you start comparing different proteins,
you will see that many of them will have similar fragments.
And those fragments will correspond to something
that we call protein domain families.
And so they are still different
because you still have mutations and, you know,
different mutations are attributed to, you know,
diversification of the function of this, you know,
protein domains.
However, you don't, you very rarely see, you know,
the evolutionary events that would split
this domain into fragments because, and it's, you know,
once you have the domain split, you actually,
you know, you can completely cancel out its function
or at the very least you can reduce it.
And that's not, you know, efficient from the point of view
of the, you know, of the cell functioning.
So the protein domain level is a very important one.
Now, on top of that, right?
So if you look at the proteins, right?
So you have this structural units
and they carry out the function.
But then much less is known about things
that connect this protein domains.
Something that we call linkers.
And those linkers are completely flexible, you know,
parts of the protein that nevertheless
carry out a lot of function.
- It's like little tails, little heads.
- So we do have tails, so they're called termini,
C and N termini.
So these are things right on the,
on one and another ends of the protein sequence.
So they are also very important.
So they attributed to very specific interactions
between the proteins.
So-
- But you're referring to the links between domains.
- That connect the domains.
And, you know, apart from the,
just the simple perspective,
if you have, you know, a very short domain,
you have, sorry, a very short linker,
you have two domains next to each other.
They are forced to be next to each other.
If you have a very long one,
you have the domains that are extremely flexible
and they carry out a lot of sort of
spatial reorganization, right?
- That's awesome.
- But on top of that, right,
just this linker itself, because it's so flexible,
§04Transcript · 32:03–42:21
it actually can adapt to a lot of different shapes.
And therefore it's a very good interactor
when it comes to interaction between
this protein and other protein.
All right, so these things also evolve, you know,
and they in a way have different sort of laws of,
the driving laws that underlie the evolution,
because they no longer need to preserve certain structure,
right, unlike protein domains.
And so on top of that,
you have something that is even less studied.
And this is something that attribute
to the concept of alternative splicing.
So alternative splicing, so it's a very cool concept.
It's something that we've been fascinated about
for over a decade in my lab
and trying to do research with that.
But so, you know, so typically, you know,
a simplistic perspective is that
one gene is equal one protein product.
Right, so you have a gene, you know,
you transcribe it and translate it,
and it becomes a protein.
In reality, when we talk about eukaryotes,
especially sort of more recent eukaryotes
that are very complex,
the gene is no longer equal to one protein.
It actually can produce multiple
functionally, you know, active protein products.
And each of them is, you know,
is called an alternatively spliced product.
The reason it happens is that if you look at the gene,
it actually has, it has also blocks.
And the blocks, some of which,
and it's essentially, it goes like this.
So we have a block that will later be translated,
we call it exon.
Then we'll have a block that is not translated, cut out.
We call it intron.
So we have exon, intron, exon, intron,
et cetera, et cetera, et cetera, right?
So sometimes you can have, you know,
dozens of these exons and introns.
So what happens is during the process
when the gene is converted to RNA,
we have things that are cut out,
the introns that cut out,
and exons that now get assembled together.
And sometimes we will throw out some of the exons.
And the remaining protein product will become--
- Still be the same.
- Different. - Oh, different.
- Right, so now you have fragments of the protein
that no longer there.
They were cut out with the introns.
Sometimes you will essentially take one exon
and replace it with another one, right?
- So there's some flexibility in this process.
- So that creates a whole new level of complexity.
'Cause now-- - Is this random though?
Is it random?
- It's not random.
We, and this is where I think now the appearance
of this modern single cell,
and before that tissue level sequencing,
next generation sequencing techniques such as RNA-Seq,
allows us to see that these are the events
that often happen in response.
It's a dynamic event that happens in response to disease,
or in response to certain developmental stage of a cell.
And this is an incredibly complex layer
that also undergoes, I mean,
because it's at the gene level, right?
So it undergoes certain evolution, right?
And now we have this interplay between what happening,
what is happening in the protein world,
and what is happening in the gene and RNA world.
And for example, it's often that we see
that the boundaries of these exons
coincide with the boundaries of the protein domains.
Right, so there is this close interplay to that.
It's not always, I mean,
otherwise it would be too simple, right?
But we do see the connection
between those sort of machineries.
And obviously the evolution will pick up this complexity
and-- - Select for whatever
is successful, whatever is functioning.
- We see that complexity in play.
And makes this question more complex, but more exciting.
- As a small detour, I don't know if you think about this
in into the world of computer science,
there's a Douglas Hostetter, I think,
came up with a name of Quine,
which are, I don't know if you're familiar with these things,
but it's computer programs that have,
I guess, exon and intron, and they copy,
the whole purpose of the program is to copy itself.
So it prints copies of itself,
but can also carry information inside of it.
That's a very kind of crude, fun exercise of,
can we sort of replicate these ideas from cells,
can we have a computer program that when you run it,
just prints itself, the entirety of itself,
and does it in different programming languages and so on.
I've been playing around and writing them.
It's a kind of fun little exercise.
- You know, when I was a kid,
so it was essentially one of the sort of main stages
in informatics Olympiads that you have to reach
in order to be any so good,
is you should be able to write a program
that replicates itself.
And so the task then becomes even sort of more complicated.
So what is the shortest program?
And of course it's a function of a programming language,
but yeah, I remember a long, long, long time ago
when we tried to make it short and short
and find the shortcuts.
- There's actually on Stack Exchange,
there's an entire site called Code Golf, I think,
where the entirety is just the competition.
People just come up with whatever task.
I don't know, like write code
that reports the weather today.
And the competition is about,
in whatever programming language,
what is the shortest program?
And it makes you actually, people should check it out
because it makes you realize
there's some weird programming languages out there.
But just to dig on that a little deeper,
do you think, in computer science,
we don't often think about programs.
Just like the machine learning world now,
that's still kind of basic programs.
And then there's humans that replicate themselves, right?
And there's these mutations and so on.
Do you think we'll ever have a world
where there's programs that kind of
have an evolutionary process?
So I'm not talking about evolutionary algorithms,
but I'm talking about programs
that kind of mate with each other and evolve
and on their own replicate themselves.
So this is kind of, the idea here is,
that's how you can have a runaway thing.
So we think about machine learning as a system
that gets smarter and smarter and smarter and smarter.
At least the machine learning systems of today are like,
it's a program that you can turn off,
as opposed to throwing a bunch of little programs out there
and letting them multiply and mate and evolve
and replicate.
Do you ever think about that kind of world
when we jump from the biological systems
that you're looking at to artificial ones?
- I mean, it's almost like you take
the sort of the area of intelligent agents, right?
Which are essentially the independent sort of codes
that run and interact and exchange the information, right?
So I don't see why not.
I mean, it could be sort of a natural evolution
in this area of computer science.
- I think it's kind of an interesting possibility.
It's terrifying too,
but I think it's a really powerful tool.
Like to have agents that,
we have social networks with millions of people
and they interact.
I think it's interesting to inject into that,
there's already injecting into that bots, right?
But those bots are pretty dumb.
They're probably pretty dumb algorithms.
It's interesting to think that there might be bots
that evolve together with humans.
And there's the sea of humans and robots
that are operating first in the digital space.
And then you can also think,
I love the idea, some people worked,
I think at Harvard, at Penn,
there's robotics labs that take as a fundamental task
to build a robot that given extra resources
can build another copy of itself,
like in the physical space,
which is super difficult to do,
but super interesting.
I remember there's like research on robots
that can build a bridge.
So they make a copy of themselves
and they connect themselves.
And so there's like self-building bridge
based on building blocks.
You can imagine like a building that self assembles.
So it's basically self-assembling structures
from robotic parts.
But it's interesting to, within that robot,
add the ability to mutate
and do all the interesting little things
that you're referring to in evolution
§05Transcript · 42:23–53:47
to go from a single origin protein building block
to like this weird complex--
- And if you think about this,
I mean, the bits and pieces are there.
So you mentioned the evolutionary algorithm, right?
So this is sort of,
and maybe sort of the goal is in a way different, right?
So the goal is to essentially to optimize your search.
So, but sort of the ideas are there.
So people recognize that the recombination events
lead to global changes in the search trajectories,
the mutations event is a more refined step in the search.
Then you have other sort of nature inspired algorithm, right?
So one of the reason that I think it's one of the funnest one
is the slime based algorithm, right?
So I think the first was introduced by the Japanese group
where it was able to solve some pretty complex problems.
So that's, and then I think there are still a lot of things
we've yet to borrow from the nature, right?
So there are a lot of sort of ideas
that nature gets to offer us
that it's up to us to grab it
and to get the best use of it.
- Including neural networks,
that we have a very crude inspiration
from nature on neural networks.
Maybe there's other inspirations to be discovered
in the brain or other aspects of the various systems,
even like the immune system, the way it interplays.
I recently started to understand that the immune system
has something to do with the way the brain operates.
Like there's multiple things going on in there,
which all of which are not modeled
in artificial neural networks.
And maybe if you throw a little bit
of that biological spice in there,
you'll come up with something cool.
I'm not sure if you're familiar with the Drake equation
that estimate, I just did a video on it yesterday
'cause I wanted to give my own estimate of it.
It's an equation that combines a bunch of factors
to estimate how many alien civilizations are.
- Oh yeah, I've heard about it, yes.
- So one of the interesting parameters,
you know, it's like how many stars are born every year,
how many planets are on average per star,
for this, how many habitable planets are there.
And then the one that starts being really interesting
is the probability that life emerges on a habitable planet.
So like, I don't know if you think about,
you certainly think a lot about evolution,
but do you think about the thing
which evolution doesn't describe,
which is like the beginning of evolution,
the origin of life?
I think I put the probability of life
developing in a habitable planet at 1%.
This is very scientifically rigorous.
Okay, first at a high level for the Drake equation,
what would you put that percent at on earth?
And in general, do you have something,
do you have thoughts about how life might have started?
You know, like the proteins being the first,
kind of one of the early jumping points?
- Yeah, so I think back in 2018,
there was a very exciting paper published in Nature
where they found one of the simplest amino acids,
glycine, in a comet dust.
So this is, and I apologize if I don't pronounce,
it's a Russian-named comet,
I think Chugryumov-Gerasimenko.
This is the comet where, and there was this mission
to get close to this comet
and get the star dust from its tail.
And when scientists analyzed it,
they actually found traces of glycine,
which makes up, it's one of the 20 basic amino acids
that makes up proteins.
So that was kind of very exciting.
But the question is very interesting.
So what, if there is some alien life,
is it gonna be made of proteins, right?
There may be RNAs, right?
So we see that the RNA viruses are certainly
very well-established sort of group of molecular machines.
Right, so yeah, it's a very interesting question.
- What probability would you put,
like how hard is this,
like how unlikely just on Earth do you think
this whole thing is that we got going?
Like are we really lucky or is it inevitable?
Like what's your sense when you sit back
and think about life on Earth?
Is it higher or lower than 1%?
Well, 'cause 1% is pretty low,
but it's still like, damn, that's a pretty good chance.
- Yes, it's a pretty good chance.
I mean, I would personally, but again,
I'm probably not the best person to do such estimations,
but I would, intuitively, I would probably put it lower.
But still, I mean--
- So we're really lucky here on Earth?
- I mean--
- Or the conditions are really good.
- It's, I think that everything was right in a way, right?
So still, it's not, the conditions were not like ideal
if you try to look at what was several billions years ago
when the life emerged.
- So there is something called the rare Earth hypothesis
that in counter to the Drake equation says that
the conditions of Earth,
if you actually were to describe Earth,
it's quite a special place, so special it might be unique
in our galaxy and potentially close to unique
in the entire universe.
Like it's very difficult to reconstruct
those same conditions.
And what the rare Earth hypothesis argues
is all those different conditions are essential for life.
And so that's sort of the counter,
like all the things we,
thinking that Earth is pretty average,
I mean, I can't really,
I'm trying to remember to go through all of them,
but just the fact that it is shielded
from a lot of asteroids,
obviously the distance to the sun,
but also the fact that it's like a perfect balance
between the amount of water and land
and all those kinds of things.
I don't know, there's a bunch of different factors
that I don't remember, there's a long list.
But it's fascinating to think about
if in order for something like proteins
and then the DNA and RNA to emerge,
you need, and basic living organisms,
you need to be very close to an Earth-like planet,
which would be sad.
Or exciting, I don't know.
- If you ask me, in a way I put a parallel
between our own research
and from the intuitive perspective.
You have those two extremes,
and the reality is never, very rarely,
falls into the extremes.
It's always, the optimums always reach somewhere in between.
And that's what I tend to think.
I think that we're probably somewhere in between,
so they were not unique-unique,
but again, the chances are reasonably small.
The problem is we don't know the other extremes.
I tend to think that we don't actually understand
the basic mechanisms of what this is all originated from.
It seems like we think of life as this distinct thing,
maybe intelligence as a distinct thing,
maybe the physics from which planets and suns are born
is a distinct thing, but that could be a very,
it's like the Stephen Wolfram thing.
From simple rules emerges greater and greater complexity.
So I tend to believe that just life finds a way.
We don't know the extreme of how common life is,
'cause it could be life is like everywhere.
Like so everywhere that it's almost like laughable,
like that we're such idiots to think we're,
like it's ridiculous to even think,
it's like ants thinking that their little colony
is the unique thing and everything else doesn't exist.
I mean, it's also very possible that that's the extreme,
and we're just not able to maybe comprehend
the nature of that life.
Just to stick on alien life for just a brief moment more,
there is some signs of life on Venus in gaseous form.
There's hope for life on Mars, probably extinct.
We're not talking about intelligent life.
Although that has been in the news recently.
We're talking about basic like bacteria.
- Plural bacteria.
- Yeah, and then also I guess, there's a couple moons.
- Europa.
- Yeah, Europa, which is Jupiter's moon.
I think there's another one.
Are you, is that exciting or is it terrifying to you
that we might find life?
Do you hope we find life?
- I certainly do hope that we find life.
I mean, it was very exciting to hear about this news
about the possible life on Venus.
- It'd be nice to have hard evidence of something
which is what the hope is for Mars and Europa.
But do you think those organisms
would be similar biologically,
or would they even be sort of carbon-based
if we do find them?
- I would say they would be carbon-based.
How similar, it's a big question, right?
So it's, the moment we discover things outside Earth,
even if it's a tiny little single cell,
I mean, there is so much.
- Just imagine that, that would be so.
- I think that that would be another turning point
§06Transcript · 53:50–1:04:03
for the science.
And if, especially if it's different in some very new way,
that's exciting, 'cause that says,
that's a definitive statement, not a definitive,
but a pretty strong statement
that life is everywhere in the universe.
To me, at least, that's really exciting.
You brought up Joshua Lederberg in an offline conversation.
I think I'd love to talk to you about AlphaFold,
and this might be an interesting way
to enter that conversation, because,
so he won the 1958 Nobel Prize in Physiology and Medicine
for discovering that bacteria can mate and exchange genes,
but he also did a ton of other stuff,
like we mentioned, helping NASA find life on Mars,
and the-- - Dendro.
- Dendro, the chemical expert system,
expert systems, remember those?
Do you, what do you find interesting about this guy
and his ideas about artificial intelligence in general?
- So I have a kind of personal story to share.
So I started my PhD in Canada back in 2000,
and so essentially my PhD was,
so we were developing sort of a new language
for symbolic machine learning,
so it's different from the feature-based machine learning,
and one of the sort of cleanest applications
of this approach, of this formalism,
was to cheminformatics and computer-aided drug design.
Right, so essentially we were,
as a part of my research, I developed a system
that essentially looked at chemical compounds
of, say, the same therapeutic category,
male hormones, right, and tried to figure out
the structural fragments that are,
the structural building blocks that are important,
that define this class,
versus structural building blocks
that are there just because, you know,
to complete the structure,
but they are not essentially the ones
that make up the key chemical properties
of this therapeutic category.
And, you know, for me it was something new.
I was trained as an applied mathematician,
you know, with some machine learning background,
but, you know, computer-aided drug design
was a completely new territory.
So because of that, I often find myself
asking lots of questions on one of the sort of
central forums.
Back then there were, you know,
no Facebooks or stuff like that,
there was-- - What's a forum?
- It's a, you know, it's a forum,
it's essentially it's like a bulletin board.
- Yeah, on the internet. - Yeah, so you essentially,
you have a bunch of people and you post a question
and you get, you know, an answer from,
you know, different people.
And back then, one of the most popular forums was CCL.
Think Computational Chemistry Library,
not library, but something like that,
but CCL, that was the forum.
And there I, you know, I--
- Asked a lot of dumb questions.
- Yes, I asked questions, also shared some, you know,
some information about our formalism,
how we do and whether whatever we do makes sense.
And so, you know, and I remember that one of these posts,
I mean, I still remember, you know,
I would call it desperately looking for a chemist's advice,
something like that, right?
And so I post my question, I explained, you know,
how our formalism is, what it does,
and what kind of applications I'm planning to do.
And, you know, and it was, you know,
in the middle of the night and, you know,
I went back to bed and next morning
have a phone call from my advisor
who also looked at this forum.
He's like, "You won't believe who replied to you."
And it's like, "Who?"
He said, "Well, you know, there is a message to you
"from Joshua Lederberg."
And my reaction was like, "Who is Joshua Lederberg?"
(laughing)
- Your advisor hung up.
- So, and essentially, you know,
Joshua wrote me that we had conceptually similar ideas
in the Dendrel project.
You may wanna look it up.
And, you know--
- And we should also, sorry, and this is a side comment,
say that even though he won the Nobel Prize
at a really young age, in '58, but so--
- He was, I think, he was, what, 33.
- Yeah, it's just crazy.
So anyway, so that's, so hence in the '90s,
responding to young whippersnappers on the CCL forum.
Okay, so--
- And so back then he was already very senior.
I mean, he unfortunately passed away back in 2008.
But, you know, back in 2001, he was, I mean,
he was a professor emeritus at Rockefeller University.
And, you know, that was actually,
believe it or not, one of the reasons
I decided to join, you know, as a postdoc,
the group of Andrei Saleh,
who was at Rockefeller University,
with the hope that, you know, that I could actually,
you know, have a chance to meet Joshua in person.
And I met him very briefly, right?
Just because he was walking, you know,
there's a little bridge that connects
the sort of the research campus
with the sort of sky scrapper that Rockefeller owns,
the way, you know, postdocs and faculty
and graduate students live.
And so I met him, you know,
and had a very short conversation, you know.
But so I started, you know, reading about Dendral,
and I was amazed, you know, it's,
we're talking about 1960, right?
The ideas were so profound.
- Well, what's the fundamental ideas of it?
- The reason to make this is even crazier.
So Lederberg wanted to make a system
that would help him study
the extraterrestrial molecules, right?
So the idea was that, you know,
the way you study the extraterrestrial molecules
is you do the mass spec analysis, right?
And so the mass spec gives you sort of bits,
numbers about essentially,
gives you the ideas about the possible fragments,
or, you know, atoms, you know,
and maybe a little fragments,
pieces of this molecule that make up the molecule, right?
So now you need to sort of,
to decompose this information
and to figure out what was the whole
before, you know, became fragments, bits and pieces, right?
So in order to make this, you know,
to have this tool,
the idea of Lederberg was to connect chemistry,
computer science,
and to design this so-called expert system
that looks, that takes into account,
that takes as an input, the mass spec data,
the possible database of possible molecules,
and essentially try to sort of induce the molecule
that would correspond to this spectra.
Or, you know, essentially,
what this project ended up being
was that, you know, it would provide a list of candidates
that then a chemist would look at
and make final decision, so.
- But the original idea, I suppose,
is to solve the entirety of this problem automatically.
- Yes, yes.
So he, you know,
so he, back then, he--
- '60s. - Approached, yes.
Believe that.
You know, it's amazing.
I mean, it still blows my mind, you know,
that it's, that's,
and this was essentially the origin
of the modern bioinformatics, cheminformatics,
you know, back in '60s.
So that's, you know,
so every time you deal with projects like this,
with the, you know, research like this,
it just, you know,
so the power of the, you know,
intelligence of these people
is just, you know, overwhelming.
- Do you think about expert systems?
Is there, and why they kind of didn't become successful?
Especially in the space of bioinformatics,
where it does seem like there is a lot of expertise
in humans, and, you know,
is it possible to see that a system like this
could be made very useful?
- Right. - And be built up?
- So it's actually, it's a great question,
and this is something, so, you know,
so, you know, at my university,
I teach artificial intelligence,
and, you know, we start,
my first two lectures are on the history of AI.
And there we, you know, we try to, you know,
go through the main stages of AI,
and so, you know, the question of why expert systems
failed or became obsolete,
it's actually a very interesting one.
And there are, you know, if you try to read the,
you know, the historical perspectives,
there are actually two lines of thoughts.
§07Transcript · 1:04:05–1:14:06
One is that they were essentially
not up to the expectations,
and so therefore they were replaced, you know,
by other things, right?
The other one was that completely opposite one,
that they were too good.
And as a result, they essentially
became sort of a household name,
and then essentially they got transformed.
I mean, in both cases, sort of the outcome was the same.
They evolved into something.
- Yeah. - Right?
And that's what I, you know, if--
- That's interesting. - If I look at this, right?
So the modern machine learning, right?
So-- - So there's echoes
in the modern machine learning to that.
- I think so, I think so, because, you know,
if you think about this, you know,
and how we design, you know,
the most successful algorithms,
including AlphaFold, right?
You built in the knowledge about the domain
that you study, right?
So you built in your expertise.
- So speaking of AlphaFold, so DeepMind's AlphaFold2
recently was announced to have,
quote unquote, "solved protein folding."
How exciting is this to you?
It seems to be one of the exciting things
that have happened in 2020.
It's an incredible accomplishment from the looks of it.
What part of it is amazing to you?
What part would you say is overhyped
or maybe misunderstood?
- It's definitely a very exciting achievement.
To give you a little bit of perspective, right?
So in bioinformatics, we have several competitions.
And so the way, you know, you often hear
how those competitions have been explained
to sort of to non-bioinformaticians is that,
you know, they call it bioinformatics Olympic games.
And there are several disciplines, right?
So the historical one of the first one
was the discipline in predicting the protein structure,
predicting the 3D coordinates of the protein.
But there are some others.
So the predicting protein functions,
predicting effects of mutations on protein functions,
then predicting protein-protein interactions.
So the original one was CASP
or a critical assessment of protein structure.
And the, you know, typically what happens
during these competitions is, you know, scientists,
experimental scientists solve the structures,
but don't put them into the protein databank,
which is the centralized database
that contains all the 3D coordinates.
Instead, they hold it and release protein sequences.
And now the challenge of the community
is to predict the 3D structures of these proteins,
and then use the experimentally solved structures
to assess which one is the closest one, right?
- And this competition, by the way,
just a bunch of different tangents.
And maybe you can also say, what is protein folding?
Then this competition, CASP competition,
has become the gold standard,
and that's what was used to say
that protein folding was solved.
So I just added a little, yeah, just a bunch.
- So if you can, whenever you say stuff,
maybe throw in some of the basics
for the folks that might be outside of the field.
Anyway, sorry.
- So, yeah, so, you know, so the reason it's, you know,
it's relevant to our understanding of protein folding
is because, you know, we've yet to learn
how the folding mechanistically works, right?
So there are different hypotheses
what happens to this fold.
For example, there is a hypothesis
that the folding happens by, you know,
also in the modular fashion, right?
So that, you know, we have protein domains
that get folded independently
because their structure is stable,
and then the whole protein structure gets formed.
But, you know, within those domains,
we also have so-called secondary structure,
the small alpha helices, beta sheets.
So these are, you know, elements that are structurally stable.
And so, and the question is, you know,
when do they get formed?
Because some of the secondary structure elements,
you have to have, you know, a fragment in the beginning
and say the fragment in the middle, right?
So you cannot potentially start
having the full fold from the get-go, right?
So it's still, you know, it's still a big enigma,
what happens.
We know that it's an extremely efficient
and stable process, right?
- So there's this long sequence,
and the fold happens really quickly.
- Exactly.
- So that's really weird, right?
And it happens like the same way almost every time.
- Exactly, exactly, right?
- That's really weird.
That's freaking weird.
- It's, yeah, that's why it's such an amazing thing.
But most importantly, right, so it's, you know,
so when you see the translation process, right,
so when you don't have the whole protein translated, right,
it's still being translated, you know,
getting out from the ribosome,
you already see some structural, you know, fragmentation.
So folding starts happening
before the whole protein gets produced, right?
And so this is obviously, you know,
one of the biggest questions in, you know,
in modern molecular biologies.
- Not like maybe what happens,
like that's not, that's bigger than the question of folding.
That's the question of like,
so like deeper fundamental idea of folding.
- Yes. - Behind folding.
- Exactly, exactly.
So, you know, so obviously if we are able to predict
the end product of protein folding,
we are one step closer to understanding
sort of the mechanistics of the protein folding.
Because we can then potentially look and start probing
what are the critical parts of this process
and what are not so critical parts of this process.
So we can start decomposing this, you know,
so in a way, this protein structure prediction algorithm
can be used as a tool, right?
So you change the, you know, you modify the protein,
you get back to this tool, it predicts,
okay, it's completely unstable.
- Yeah, which aspects of the input
will have a big impact on the output.
- Exactly, exactly.
So what happens is, you know,
we typically have some sort of incremental advancement.
You know, each stage of this CASP competition,
you have groups with incremental advancement.
And, you know, historically, the top performing groups
were, you know, they were not using machine learning.
They were using very advanced biophysics
combined with bioinformatics,
combined with, you know, the data mining.
And that was, you know, that would enable them
to obtain protein structures of those proteins
that don't have any structurally solved relatives.
Because, you know, if we have another protein,
say the same protein, but coming from a different species,
we could potentially derive some ideas,
and that's so-called homology or comparative modeling,
where we'll derive some ideas
from the previously known structures.
And that would help us tremendously in, you know,
in reconstructing the 3D structure overall.
But what happens when we don't have these relatives?
This is when it becomes really, really hard, right?
So that's so-called de novo, you know,
de novo protein structure prediction.
And in this case,
those methods were traditionally very good.
But what happened in the last year,
the original alpha fold came into,
and all of a sudden, it's much better than everyone else.
- This is 2018.
- Yeah.
- Oh, and the competition is only every two years, I think.
- And then, so, you know, it was sort of
kind of of a shockwave to the bioinformatics community
that we have like a state-of-the-art machine learning system
that does structure prediction.
And essentially what it does, you know,
so if you look at this, it actually predicts the context.
So, you know, so the process of reconstructing
the 3D structure starts by predicting the context
between the different parts of the protein.
And the context is essentially the part of the proteins
that are in the close proximity to each other.
- Right, so it's actually,
the machine learning part seems to be estimating,
you can correct me if I'm wrong here,
but it seems to be estimating the distance matrix,
which is like the distance between the different parts.
- Yeah, so we call it a contact map.
- Contact map.
- Right, so once you have the contact map,
the reconstruction is becoming more straightforward.
- Yeah.
- Right, but so the contact map is the key.
And so, you know, so that what happened.
§08Transcript · 1:14:11–1:25:04
And now we started seeing in this current stage, right,
where in the most recent one,
we started seeing the emergence of these ideas
in other people works, right?
But yet here's, you know, AlphaFold2,
that again outperforms everyone else.
And also by introducing yet another wave
of the machine learning ideas.
- Yeah, there does seem to be also an incorporation.
First of all, the paper's not out yet,
but there's a bunch of ideas already out.
There does seem to be an incorporation of this other thing.
I don't know if it's something that you could speak to,
which is like the incorporation of like other structures,
like evolutionary similar structures
that are used to kind of give you hints.
- Yes, so evolutionary similarity is something
that we can detect at different levels, right?
So we know, for example, that the structure of proteins
is more conserved than the sequence.
The sequence could be very different,
but the structural shape is actually still very conserved.
So that's sort of the intrinsic property that, you know,
in a way related to protein folds, you know,
to the evolution of the protein,
of proteins and protein domains, et cetera.
But we know that, I mean, there've been multiple studies.
And, you know, ideally if you have structures,
you know, you should use that information.
However, sometimes we don't have this information.
Instead, we have a bunch of sequences.
Sequences we have a lot, right?
So we have, you know, hundreds, thousands of, you know,
different organisms sequenced, right?
And by taking the same protein,
but in different organisms and aligning it,
so making it, you know,
making the corresponding positions aligned,
we can actually say a lot about sort of what is conserved
in this protein and therefore, you know,
structure more stable, what is diverse in this protein.
So on top of that, we could provide sort of the information
about the sort of the secondary structure of this protein,
et cetera, et cetera.
So this information is extremely useful
and it's already there.
So while it's tempting to, you know,
to do a complete ab initio,
so you just have a protein sequence and nothing else,
the reality is such that we are overwhelmed with this data.
So why not use it?
- Mm-hmm.
- And so, yeah, so I'm looking forward
to reading this paper.
- It does seem to, like they've,
in the previous version of AlphaFold, they didn't,
for this evolutionary similarity thing,
they didn't use machine learning for that.
Or rather, they used it as like the input
to the entirety of the neural net,
like the features derived from the similarity.
It seems like there's some kind of, quote unquote,
iterative thing where it seems to be part of the learning
process is the incorporation of this evolutionary similarity.
- Yeah, I don't think there is a bioarchive paper, right?
- There's nothing. - No, nothing.
- There's a blog post that's written
by a marketing team, essentially.
- Yeah. - Which, you know,
it has some scientific similarity probably
to the actual methodology used,
but it could be, it's like interpreting scripture.
It could be just poetic interpretations
of the actual work as opposed
to direct connection to the work.
- So now, speaking about protein folding, right?
So, you know, in order to answer the question
whether or not we have solved this, right?
So we need to go back to the beginning of our conversation,
you know, with the realization that, you know,
an average protein is that typically what the cusp
has been focusing on is, you know,
this competition has been focusing on the single,
maybe two domain proteins that are still very compact.
And even those ones are extremely challenging to solve,
right, but now we talk about, you know,
an average protein that has two, three protein domains.
If you look at the proteins that are in charge of the,
you know, of the process with the neural system, right,
perhaps one of the most recently evolved sort of systems
in the organism, right?
All of them, well, the majority of them
are highly multi-domain proteins.
So they are, you know, some of them have five, six, seven,
you know, and more domains, right?
And, you know, we are very far away from understanding
how these proteins are folded.
- So the complexity of the protein matters here,
the complexity of the protein modules
or the protein domains.
So you're saying solved, so the definition of solved here
is particularly the cast competition,
achieving human level, not human level,
achieving experimental level performance
on these particular sets of proteins
that have been used in these competitions.
- Well, I mean, you know, I do think that, you know,
especially with regards to the alpha fold, you know,
it is able to, you know, to solve, you know,
at the near experimental level,
pretty big majority of the more compact proteins,
like, or protein domains, because, again,
in order to understand how the overall protein,
you know, multi-domain protein fold,
we do need to understand the structure
of its individual domains.
- I mean, unlike if you look at alpha zero
or like mu zero, if you look at that work,
you know, it's nice, reinforcement learning,
self-playing mechanisms are nice
'cause it's all in simulation,
so you can learn from just huge amounts.
Like, you don't need data.
It was like the problem with proteins,
like the size, I forget how many 3D structures
have been mapped, but the training data is very small,
no matter what, it's like millions,
maybe a one or two million, something like that,
but it's some very small number,
but like, it doesn't seem like that's scalable.
There has to be, I don't know,
it feels like you want to somehow 10X the data
or 100X the data somehow.
- Yes, but we also can take advantage
of homology models, right?
So the models that are of very good quality
because they are essentially obtained
based on the evolutionary information, right?
So you can, there is a potential to enhance this information
and use it again to empower the training set.
And it's, I think, I am actually very optimistic.
I think it's been one of these sort of, you know,
churning events where you have a system
that is, you know, a machine learning system
that is truly better than the sort of the more conventional
biophysics-based methods.
- That's a huge leap.
This is one of those fun questions,
but where would you put it in the ranking
of the greatest breakthroughs
in artificial intelligence history?
So like, okay, so let's see who's in the running.
Maybe you can correct me.
So you got like AlphaZero and AlphaGo
beating the world champion at the game of Go.
Thought to be impossible like 20 years ago,
or at least the AI community was highly skeptical.
Then you got like also Deep Blue original Kasparov.
You have deep learning itself, like the,
maybe what would you say, the AlexNet image in that moment.
So the first, you know,
network achieving human level performance,
super not, that's not true.
Achieving like a big leap in performance
on the computer vision problem.
There is OpenAI, the whole like GPT-3,
that whole space of transformers and language models,
just achieving this incredible performance
of application of neural networks to language models.
Boston Dynamics, pretty cool.
Like robotics, even though people are like,
there's no AI, no, no, no.
There's no machine learning currently,
but AI is much bigger than machine learning.
So that just the engineering aspect,
I would say it's one of the greatest accomplishments
in engineering side, engineering meaning
like mechanical engineering of robotics ever.
Then of course, autonomous vehicles,
you can argue for Waymo,
which is like the Google self-driving car,
or you can argue for Tesla,
which is like actually being used
by hundreds of thousands of people
on the road today, machine learning system.
And I don't know if you can, what else is there?
But I think that's it.
So, and then AlphaFold, many people are saying
is up there, potentially number one.
Would you put them at number one?
- Well, in terms of the impact on the science
and on the society beyond, it's definitely,
to me would be one of the--
- Top three?
- I mean, I'm probably not the best person to answer that.
But I do have, I remember my,
back in, I think 1997, when Deep Blue,
but Kasparov, it was, I mean, it was a shock.
I mean, it was, and I think for the,
for the pretty substantial part of the world,
§09Transcript · 1:25:13–1:35:26
that especially people who have some experience
with chess, right, and realizing how incredibly human
this game, how much of a brain power you need
to reach those levels of grandmasters, right, level.
- Yeah, and it's probably one of the first time,
and how good Kasparov was.
- And again, yeah, so Kasparov's arguably
one of the best ever, right?
And you get a machine that beats him, right?
So it's--
- First time a machine probably beat a human
at that scale of a thing, of anything.
- Yes, yes, so that was, to me, that was like,
you know, one of the groundbreaking events
in the history of AI.
- Yeah, it's probably number one.
- As probably, like, we don't, it's hard to remember.
It's like Muhammad Ali versus, I don't know,
any other, Mike Tyson, something like that.
It's like, nah, you gotta put Muhammad Ali at number one.
Same with Deep Blue, even though
it's not machine learning based.
Still, it uses advanced search,
and search is the integral part of AI, right?
So as you said, it's--
- People don't think of it that way at this moment.
In Vogue, currently, search is not seen
as a fundamental aspect of intelligence,
but it very well, it very likely is.
In fact, I mean, that's what neural networks are,
is they're just performing search
on the space of parameters.
And it's all search. (laughs)
All of intelligence is some form of search,
and you just have to become clever and clever
at that search problem.
- And I also have another one that you didn't mention
that's one of my favorite ones.
So you probably heard of this.
I think it's called Deep Rembrandt.
It's the project where they trained,
I think there was a collaboration between
the experts in Rembrandt painting in Netherlands,
and a group, an artificial intelligence group,
where they train an algorithm
to replicate the style of the Rembrandt,
and they actually printed a portrait
that never existed before in the style of Rembrandt.
I think they printed it on a sort of,
on the canvas that, using pretty much
same types of paints and stuff.
To me, it was mind-blowing.
- Yeah.
- It's--
- In the space of art, that's interesting.
There hasn't been, maybe that's it,
but I think there hasn't been an image
in that moment yet in the space of art.
You haven't been able to achieve superhuman level
performance in the space of art,
even though there was this big famous thing
where a piece of art was purchased,
I guess, for a lot of money.
- Yes.
- Yeah, but it's still, people are like,
in the space of music at least,
that's, it's clear that human-created pieces
are much more popular.
So there hasn't been a moment where it's like,
oh, this is, we're now, I would say in the space of music,
what makes a lot of money, we're talking about serious money,
it's music and movies, or like shows and so on,
and entertainment.
There hasn't been a moment where AI created,
AI was able to create a piece of music
or a piece of cinema, like Netflix show,
that is, that's sufficiently popular to make a ton of money.
- Yeah.
- And that moment would be very, very powerful,
'cause that's like, that's an AI system
being used to make a lot of money.
And like direct, of course, AI tools,
like even Premiere, audio editing, all the editing,
everything I do, to edit this podcast,
there's a lot of AI involved.
Actually, there's this program,
I wanna talk to those folks just 'cause I wanna nerd out,
it's called iZotope, I don't know if you're familiar with it.
They have a bunch of tools of audio processing,
and they have, I think they're Boston-based.
Just, it's so exciting to me to use it,
like on the audio here, 'cause it's all machine learning.
It's not, 'cause most audio production stuff
is like any kind of processing you do,
it's very basic signal processing.
And you're tuning knobs and so on.
They have all of that, of course,
but they also have all of this machine learning stuff,
like where you actually give it training data.
You select parts of the audio you train on,
you train on it, and it figures stuff out.
It's great, it's able to detect,
like the ability of it to be able to separate voice
and music, for example, or voice in anything, is incredible.
Like it just, it's clearly exceptionally good
at applying these different neural networks models
to separate the different kinds of signals from the audio.
Okay, so that's really exciting.
Photoshop, Adobe people also use it.
But to generate a piece of music
that will sell millions, a piece of art, yeah.
- No, I agree, and that's,
as I mentioned, I offer my AI class,
and an integral part of this is the project, right?
So it's my favorite, ultimate favorite part,
because typically we have these project presentations
the last two weeks of the class,
it's right before the Christmas break,
and it's sort of, it adds this cool excitement.
And every time, I'm amazed with some projects
that people come up with.
And so, and quite a few of them are actually,
they have some link to arts.
I mean, I think last year, we had a group
who designed an AI producing hokus, Japanese poems.
- Oh, wow.
- And some of them, so it got trained
on the English base. - Hokus?
- Hokus, hokus, right?
So, and some of them, they get to present
like the top selection.
They were pretty good.
I mean, of course, I'm not a specialist,
but you read them, and you see--
- It seems profound.
- Yes, yeah, it seems reason, so it's kind of cool.
We also had a couple of projects
where people tried to teach AI how to play like rock music,
classical music, I think, and popular music.
Interestingly enough, classical music
was among the most difficult ones.
Of course, if you look at the grandmasters of music,
like Bach, right?
So, there is a lot of almost math.
- Yeah, well, he's very mathematical.
- Exactly, so this is, I would imagine
that at least some style of this music could be picked up,
but then you have this completely different spectrum
of classical composers, and so it's almost like,
you don't have to sort of look at the data.
You just listen to it, and say, "Nah, that's not it,
"not yet." - That's not it.
Yeah, that's how I feel, too.
There's OpenAI has, I think, OpenMuse
or something like that, the system.
It's cool, but it's like, eh,
it's not compelling for some reason.
It could be a psychological reason, too.
Maybe we need to have a human being,
a tortured soul behind the music, I don't know.
- Yeah, no, absolutely, I completely agree.
But yeah, whether or not we'll have,
one day we'll have a song written by an AI engine
to be in top charts, musical charts,
I wouldn't be surprised.
I wouldn't be surprised.
- I wonder if we already have one,
and it just hasn't been announced.
(both laugh)
We wouldn't know.
How hard is the multi-protein folding problem?
Is that something you've already mentioned,
which is baked into this idea
of greater and greater complexity of proteins?
Like multi-domain proteins,
is that basically become multi-protein complexes?
- Yes, you got it right.
So it has the components of both,
of protein folding and protein-protein interactions.
Because in order for these domains,
I mean, many of these proteins,
actually, they never form a stable structure.
One of my favorite proteins,
and pretty much everyone who works in the,
I know, whom I know who works with proteins,
they always have their favorite proteins.
Right, so one of my favorite proteins,
probably my favorite protein,
the one that I worked when I was a postdoc,
is so-called post-synaptic density 95, PSD95 protein.
So it's one of the key actors
in the majority of neurological processes
at the molecular level.
And essentially, it's a key player
in the post-synaptic density.
So this is the crucial part of the synapse,
where a lot of these chemological processes
are happening.
So it has five domains, right?
So five protein domains.
§10Transcript · 1:35:27–1:46:27
It's a pretty large proteins,
I think 600 something, I mean,
but the way it's organized itself, it's flexible, right?
So it acts as a scaffold.
So it is used to bring in other proteins.
So they start acting in the orchestrated manner, right?
So, and the type of the shape of this protein,
it's in a way, there are some stable parts of this protein,
but there are some flexible.
And this flexibility is built in,
into the protein in order to become
sort of this multifunctional machine.
- So do you think that kind of thing is also learnable
through the AlphaFold2 kind of approach?
- I mean, the time will tell.
- Is it another level of complexity?
Is it, like how big of a jump in complexity
is that whole thing?
- To me, it's yet another level of complexity,
because when we talk about protein-protein interactions,
and there is actually a different challenge for this,
called CAPRI.
And so this, that is focused specifically
on macromolecular interactions.
Protein-protein, protein-DNA, et cetera.
So, but it's, you know, there are different mechanisms
that govern molecular interactions,
and that need to be picked up,
say by a machine learning algorithm.
Interestingly enough, we actually,
we participated for a few years in this competition.
We typically don't participate in competitions,
I don't know, don't have enough time,
you know, 'cause it's very intensive.
It's a very intensive process.
But we participated back in, you know,
about 10 years ago or so.
And the way we entered this competition,
so we design a scoring function, right?
So the function that evaluates whether or not
your protein-protein interaction is supposed to look like
experimentally solved, right?
So the scoring function is very critical part
of the model prediction.
So we design it to be a machine learning one.
And so it was one of the first
machine learning-based scoring function used in CAPRI.
And, you know, we essentially, you know,
learned what should contribute,
what are the critical components contributing
into the protein-protein interaction.
- So this could be converted into a learning problem
and thereby it could be learned.
- I believe so, yes.
- Do you think AlphaFold2 or something similar to it
from DeepMind or somebody else will be,
will result in a Nobel Prize or multiple Nobel Prizes?
So like, you know, obviously, maybe not so obviously,
you can't give a Nobel Prize to a computer program.
You, at least for now, give it to the designers
of that program.
But do you see one or multiple Nobel Prizes
where AlphaFold2 is like a large percentage
of what that prize is given for?
Would it lead to discoveries at the level of Nobel Prizes?
- I mean, I think we are definitely destined
to see the Nobel Prize becoming sort of,
to be evolving with the evolution of science.
And the evolution of science is such
that it now becomes like really multifaceted, right?
So where you don't really have like a unique discipline,
you have sort of the, a lot of cross-disciplinary talks
in order to achieve sort of, you know,
really big advancements, you know.
So I think, you know, the computational methods
will be acknowledged in one way or another.
And as a matter of fact, you know,
they were first acknowledged back in 2013, right?
Where, you know, the first three people were, you know,
awarded the Nobel Prize for the,
for study of the protein folding, right, the principle.
And, you know, I think all three of them
are computational biophysicists, right?
So, you know, that I think is unavoidable, you know.
It will come with the time.
The fact that, you know, alpha fold and, you know,
similar approaches, 'cause again, it's a matter of time
that people will embrace this, you know, principle.
And we'll see more and more such, you know,
such tools coming into play.
But, you know, these methods will be critical
in a scientific discovery, no doubts about it.
- On the engineering side, maybe a dark question,
but do you think it's possible
to use these machine learning methods
to start to engineer proteins?
And the next question is something quite a few biologists
are against, some are for, for study purposes,
is to engineer viruses.
Do you think machine learning, like something
like alpha fold could be used to engineer viruses?
- So to answering the first question, you know,
it has been, you know, a part of the research
in the protein science.
The protein design is, you know,
is a very prominent areas of research.
Of course, you know, one of the pioneers is David Baker
and Rosetta algorithm that, you know,
essentially was doing the de novo design
and was used to design new proteins, you know.
- And design of proteins means design of function.
So like when you design a protein, you can control,
I mean, the whole point of a protein,
with the protein structure comes a function,
like it's doing something.
- Correct, correct.
- Or design different things.
- So you can, yeah, so you can, well,
you can look at the proteins from the functional perspective,
you can also look at the proteins
from the structural perspective, right?
So the structural building blocks.
So if you want to have a building block of a certain shape,
you can try to achieve it by, you know,
introducing a new protein sequence
and predicting, you know, how it will fold.
So with that, I mean, it's a natural,
one of the natural applications of these algorithms.
Now, talking about engineering a virus.
- With machine learning.
- With machine learning, right?
So, well, you know, so luckily for us,
I mean, we don't have that much data, right?
We actually, right now, one of the projects
that we are carrying on in the lab
is we're trying to develop a machine learning algorithm
that determines whether or not
the current strain is pathogenic.
And-- - The current strain
of the coronavirus.
- Of the virus, I mean, so there are applications
to coronaviruses because we have strains of SARS-CoV-2,
also SARS-CoV, MERS, that are pathogenic,
but we also have strains of other coronaviruses
that are not pathogenic, I mean,
the common cold viruses and some other ones, right?
So-- - Pathogenic meaning spreading.
- Pathogenic meaning it's actually inflicting damage.
Correct.
There are also some seasonal versus pandemic strains
of influenza, right?
And determining what are the molecular determinant, right,
so that are built in into the protein sequence,
into the gene sequence, right?
So, and whether or not the machine learning
can determine those components, right?
- Oh, interesting, so like using machine learning,
that's really interesting, to given,
the input is like, what, the entire--
- Protein sequence.
- The protein sequence, and then determine
if this thing is gonna be able to do damage
to a biological system.
- Yeah.
So, I mean-- - It's good machine learning,
you're saying we don't have enough data for that?
- I mean, for this specific one, we do.
We might actually have to back up on this,
'cause we're still in the process.
There was one work that appeared in bio-archive
by Eugene Kunin, who is one of these pioneers
in evolutionary genomics, and they tried to look at this,
but the methods were sort of standard,
supervised learning methods, and now the question is,
can you advance it further by using not so standard methods?
So, there's obviously a lot of hope in transfer learning,
where you can actually try to transfer the information
that the machine learning learns
about the proper protein sequences, right?
And so, there is some promise in going this direction,
but if we have this, it would be extremely useful,
because then we could essentially forecast
the potential mutations that would make a current strain
more or less pathogenic.
- Anticipate them from a vaccine development,
for the treatment, antiviral drug development.
- That would be a very crucial task.
- But you could also use that system to then say,
how would we potentially modify this virus
to make it more pathogenic?
- That's true, that's true.
I mean, you know, again, the hope is, well, several things.
So, one is that, even if you design a sequence, right?
So, to carry out the actual experimental biology,
to ensure that all the components working,
you know, is a completely different matter.
- Difficult process.
- Yes, then we've seen in the past,
there could be some regulation of the moment
the scientific community recognizes
§11Transcript · 1:46:30–1:56:36
that it's now becoming no longer a sort of a fun puzzle
for machine learning.
- Could be a weapon.
- Yeah, so then there might be some regulation.
So, I think back in, what, 2015,
there was an issue on regulating the research
on influenza strains, right?
So, there were several groups use sort of mutation analysis
to determine whether or not this strain will jump
from one species to another.
And I think there was like a half a year moratorium
on the research, on the paper published,
until scientists analyzed it
and decided that it's actually safe.
- I forgot what that's called,
something of function, test of function.
- Gain of function, loss of function.
- Gain of function, yeah, gain of function,
loss of function, that's right, sorry.
It's like, let's watch this thing mutate for a while
to see what kind of things we can observe.
I guess, I'm not so much worried about that kind of research
if there's a lot of regulation
and if it's done very well
and with competence and seriously.
I am more worried about kind of this,
the underlying aspect of this question
is more like 50 years from now.
Speaking to the Drake equation,
one of the parameters in the Drake equation
is how long civilizations last.
And that seems to be the most important value, actually,
for calculating if there's other alien
intelligent civilizations out there.
That's where there's most variability.
Assuming, like if life, if that percentage
that life can emerge is like not zero,
like if we're super unique,
then it's the how long we last
is basically the most important thing.
So from a selfish perspective,
but also from a Drake equation perspective,
I'm worried about our civilization lasting.
And you kind of think about all the ways
in which machine learning can be used
to design greater weapons of destruction.
And I mean, one way to ask that,
if you look sort of 50 years from now,
100 years from now,
would you be more worried about natural pandemics
or engineered pandemics?
Like who is the better designer of viruses,
nature or humans, if we look down the line?
- I think, in my view,
I would still be worried about the natural pandemics,
simply because, I mean, the capacity
of the nature producing this.
- It does a pretty good job, right?
- Yes.
- The motivation for using virus,
engineering viruses as a weapon is a weird one,
because maybe you can correct me on this,
but it seems very difficult to target a virus, right?
The whole point of a weapon, the way a rocket works,
if a starting point, you have an end point,
and you're trying to hit a target,
to hit a target with a virus is very difficult.
It's basically just, right?
It's the target would be the human species.
Oh man.
- Yeah, I have a hope in us,
I'm forever optimistic that we will not,
there's insufficient evil in the world
to lead to that kind of destruction.
- Well, I also hope that, I mean, that's what we see.
I mean, with the way we are getting connected,
the world is getting connected,
I think it helps for the world to become more transparent.
- Yeah.
- So the information spread is,
I think it's one of the key things for the society
to become more balanced.
- Yeah. - One way or another.
- This is something that people disagree with me on,
but I do think that the kind of secrecy
that governments have,
so you're kind of speaking more to the other aspects,
like research community being more open,
companies are being more open,
government is still like,
we're talking about like military secrets.
I think military secrets of the kind
that could destroy the world
will become also a thing of the 20th century.
It'll become more and more open.
- Yeah.
- I think nations will lose power in the 21st century,
like lose sufficient power towards secrecies.
Transparency is more beneficial than secrecy,
but of course it's not obvious.
Let's hope so, let's hope so,
that the governments will become more transparent.
- So we last talked, I think, in March or April.
What have you learned?
How has your philosophical, psychological,
biological worldview changed since then?
Or you've been studying it nonstop
from a computational biology perspective.
How has your understanding and thoughts
about this virus changed over those months,
from the beginning to today?
- One thing that I was really amazed
at how efficient the scientific community was.
I mean, and even just judging on this very narrow domain
of protein structure,
understanding the structural characterization
of this virus from the components point of view,
whole virus point of view.
If you look at SARS, right,
the something that happened, you know,
less than 20, but close enough, 20 years ago.
And you see what, when it happened,
what was sort of the response by the scientific community.
You see that the structural characterizations did occur,
but it took several years, right?
Now, the things that took several years,
it's a matter of months, right?
So we see that, you know, the research pop up.
We are at the unprecedented level
in terms of the sequencing, right?
Never before we had a single virus sequenced so many times.
You know, so which allows us to actually,
to trace very precisely the sort of the evolutionary nature
of this virus, what happens.
And it's not just the, you know,
this virus independently of everything.
It's, you know, it's the, you know,
the sequence of this virus linked,
anchored to the specific geographic place,
to specific people, because, you know,
our genotype influences also, you know,
the evolution of this, you know,
it's always a host pathogen co-evolution that, you know,
occurs.
- It'd be cool if we also had a lot more data about,
sort of, the spread of this virus.
Not maybe, well, it'd be nice if we had it
for like contact tracing purposes for this virus,
but it'd be also nice if we had it
for the study for future viruses,
to be able to respond and so on.
But it's already nice that we have geographical data
and the basic data from individual humans.
- Exactly.
No, I think contact tracing is obviously a key component
in understanding the spread of this virus.
There is also, there is a number of challenges, right?
So XPRICE is one of them.
We, you know, just recently, you know,
took a part of this competition.
It's the prediction of the number of infections
in different regions.
So, and, you know, obviously the AI is the main topic
in those predictions.
- Yeah, but it's still the data.
I mean, that's a competition,
but the data is weak on the training.
Like, it's great.
It's much more than probably before,
but like, it would be nice if it was like really rich.
Like I talked to Michael Mina from Harvard.
I mean, he dreams that the community comes together
with like a weather map to where a viruses, right?
Like really high resolution sensors on like how,
from person to person, the viruses that travel,
all the different kinds of viruses, right?
Because there's a ton of them.
And then you'd be able to tell the story
that you've spoken about of the evolution of these viruses,
like day-to-day mutations that are occurring.
I mean, that'd be fascinating,
just from a perspective of study
and from the perspective of being able to respond
to future pandemics.
That's ultimately what I'm worried about.
People love books.
Is there some three or whatever number of books,
technical fiction, philosophical,
that brought you joy in life,
had an impact on your life,
and maybe some that you would recommend others?
- So I'll give you three very different books,
and I also have a special runner-up.
- Honorable mention.
- I mean, it's an audio book,
and there's some specific reason behind it.
So the first book is something
that sort of impacted my earlier stage of life,
and I'm probably not gonna be very original here.
It's "Bulgakov's Master and Margarita."
§12Transcript · 1:56:39–2:08:23
So that's probably, you know.
- Well, not for a Russian, maybe it's not super original,
but it's a really powerful book for even in English.
So I read it in English, so.
- It is incredibly powerful,
and I mean, it's the way it ends, right?
So I still have goosebumps
when I read the very last sort of,
it's called prologue,
where it's just so powerful.
- What impact did it have on you?
What ideas, what insights did you get from it?
- I was just taken by, you know,
by the fact that you have those parallel lives
apart from many centuries, right?
And somehow they got sort of intertwined
into one story.
And that, to me, was fascinating.
And, you know, of course, the romantic part of this book
is like, it's not just romance,
it's like the romance empowered by sort of magic, right?
And maybe on top of that, you have some irony,
which is unavoidable, right?
Because it was that, you know, the Soviet time.
- But it's very, it's deeply Russian.
So that's the wit, the humor, the pain, the love,
all of that is one of the books
that kind of captures something about Russian culture
that people outside of Russia should probably read.
- I agree.
- What's the second one?
- So the second one is, again, another one that,
it happened, I read it later in my life.
I think I read it first time when I was a,
a graduate student.
And that's the Solzhenitsyn's "Cancer Ward."
That is amazingly powerful book.
It's-- - What is it about?
- It's about, I mean, essentially based on,
you know, Solzhenitsyn was diagnosed with cancer
when he was reasonably young and he made a full recovery.
But, you know, so this is about a person
who was sentenced for life in one of these, you know, camps.
And he had some cancer.
So he was, you know, transported back
to one of these Soviet republics,
I think, you know, South Asian republics.
And the book is about, you know,
his experience being a prisoner,
being a, you know, a patient in the cancer clinic,
in a cancer ward, surrounded by people,
many of which die, right?
But in the way, you know, the way it reads,
I mean, first of all, later on,
I read the accounts of the doctors
who described these, you know, the experiences,
in the book, by the patient as incredibly accurate, right?
So, you know, I read that there was some doctors saying
that, you know, every single doctor should read this book
to understand what the patient feels.
But, you know, again, as many of the Solzhenitsyn's books,
it has multiple levels of complexity.
And obviously, you know, if you look above the cancer
and the patient, I mean, the tumor that was growing
and then disappeared in his body with some consequences,
I mean, this is, you know, allegorically the Soviet,
and, you know, and he actually, he agreed,
you know, when he was asked, he said that this is
what made him think about this, you know,
how to combine these experiences.
Him being a part of the, you know, of the Soviet regime,
also being a part of the, you know,
of someone sent to the Gulag camp, right?
And also someone who experienced cancer in his life.
You know, the Gulag Archipelago and this book,
these are the works that actually made him,
you know, receive a Nobel Prize.
But, you know, to me, I've, you know,
I've read other, you know, books by Solzhenitsyn.
This one is, to me, is the most powerful one.
- And by the way, both this one and the previous one,
you read in Russian?
- Yes, yes.
So now there is, the third book is an English book,
and it's completely different.
So, you know, we're switching the gears completely.
So this is the book, which it's not even a book,
it's an essay by Jonathan Neumann.
- Oh, wow.
- Called "The Computer and the Brain."
And that was the book he was writing,
knowing that he was dying of cancer.
So the book was released back,
it's a very thin book, right?
But the power, the intellectual power
in this book, in this essay is incredible.
I mean, you probably know that von Neumann
is considered to be one of the biggest thinkers, right?
So his intellectual power was incredible, right?
And you can actually feel this power in this book
where, you know, the person is writing
knowing that he will be, you know, he will die.
The book actually got published only after his death,
back in 1958, he died in 1957.
And, but, so he tried to put as many ideas
that, you know, he still, you know, hadn't realized.
And, you know, so this book is very difficult to read
because, you know, every single paragraph is just compact.
You know, it's filled with these ideas
and, you know, the ideas are incredible.
You know, nowadays, you know, so he tried
to put the parallels between the brain computing power,
the neural system, and the computers, you know,
as they were understood.
- Do you remember what year he was working on this?
Like approximately?
- '57. - '57.
- So that was right during his, you know,
when he was diagnosed with cancer and he was essentially.
- Yeah, he's one of those, there's a few folks
people mention, I think Ed Witten is another,
that like, everyone that meets them,
they say he's just an intellectual powerhouse.
- Yes.
- Okay, so who's the honorable mention?
- So, so, so, and this is, I mean,
the reason I put it sort of in this separate section
because this is a book that I recently listened to.
So it's an audio book, and this is a book
called Lab Girl by Hope Jarren.
So Hope Jarren, she is a scientist,
she's a geochemist that essentially studies
the fossil plants and so she uses the fossil plants,
the chemical analysis to understand what was the climate
back in, you know, in thousand years,
hundreds of thousands of years ago.
And so something that incredibly touched me by this book,
it was narrated by the author.
- Nice, excellent. - And it's an incredibly
personal story, incredibly.
So certain parts of the book you could actually
hear the author crying.
And that, to me, I mean, I never experienced
anything like this, you know, reading the book.
But it was like, you know, the connection
between you and the author.
And I think this is, you know, this is really a must read,
but even better, a must listen to audio book
for anyone who wants to learn about sort of, you know,
academia, science, research in general.
Because it's a very personal account
about her becoming a scientist.
- So we're just before New Year's, you know,
we talked a lot about some difficult topics,
the viruses and so on.
Do you have some exciting things you're looking forward
to in 2021, some New Year's resolutions,
maybe silly or fun, or something very important
and fundamental to the world of science
or something completely unimportant?
- Well, I'm definitely looking forward
towards, you know, things becoming normal, right?
So yes, so I really miss traveling.
Every summer I go to a international city,
international summer school,
it's called the School for Molecular and Theoretical Biology.
It's held in Europe,
it's organized by very good friends of mine.
And this is the school for gifted kids
from all over the world, and they're incredibly bright.
It's like, every time I go there, it's like, you know,
it's the highlight of the year.
And we couldn't make it this August,
so we did this school remotely, but it's different.
So I am definitely looking forward
to next August coming there.
I also, I mean, you know, one of my personal resolutions,
I realized that being in-house and working from home,
you know, I realized that actually,
I apparently missed a lot, you know,
spending time with my family, believe it or not.
So you typically, you know, with all the research
and teaching and everything related to the academic life,
I mean, you get distracted.
And so, you know, you don't feel that, you know,
the fact that you are away from your family
doesn't affect you because you're, you know,
naturally distracted by other things.
And, you know, this time I realized that, you know,
that that's so important, right?
Spending your time with the family, with your kids.
And so that would be my new year resolution
in actually trying to spend as much time as possible.
- Even when the world opens up.
Yeah, that's a beautiful message.
That's a beautiful reminder.
I asked you if there's a Russian poem you could read
that I could force you to read.
And you said, okay, fine, sure.
Do you mind reading?
§13Transcript · 2:08:27–2:12:30
And you said that no paper needed, so.
- Nope.
So yeah, so this poem was written by my namesake,
another Dmitry, Dmitry Kemmerfeld.
And it's a recent poem,
and it's called "Sorceress", "Vedma" in Russian.
Or actually, "Kaldunya".
So that's sort of another connotation of sorceress or witch.
And I really like it.
And it's one of just a handful poems
I actually can recall by heart.
I also have a very strong association
when I read this poem with Master Margarita,
the main female character, Margarita.
And also it's happening about the same time
we are talking now, so around New Year, around Christmas.
- Do you mind reading it in Russian?
- I'll give it a try.
(speaking in foreign language)
(speaking in foreign language)
(speaking in foreign language)
(speaking in foreign language)
(speaking in foreign language)
(speaking in foreign language)
(speaking in foreign language)
(speaking in foreign language)
(speaking in foreign language)
(speaking in foreign language)
- That's beautiful.
I love how it captures a moment of longing
and maybe love even.
- Yes.
To me, it has a lot of meaning about this,
something that is happening, something that is far away,
but still very close to you.
And yes, it's the winter.
- There's something magical about winter, isn't it?
What is the, well, I don't know.
I don't know how to translate it,
but a kiss in winter is interesting.
Lips in winter and all that kind of stuff.
It's beautifully, I mean, Russian has a way.
As a reason, Russian poetry is just,
I'm a fan of poetry in both languages,
but English doesn't capture some of the magic
that Russian seems to, so thank you for doing that.
That was awesome.
Dmitry, it's great to talk to you again.
It's contagious how much you love what you do,
how much you love life.
So I really appreciate you taking the time to talk today.
- And thank you for having me.
- Thanks for listening to this conversation
with Dmitry Korkin, and thank you to our sponsors,
Brave Browser, NetSuite Business Management Software,
Magic Spoon Low Carb Cereal,
and Asleep Self-Cooling Mattress.
So the choice is browsing privacy, business success,
healthy diet, a comfortable sleep.
Choose wisely, my friends.
And if you wish, click the sponsor links below
to get a discount and to support this podcast.
Now, let me leave you with some words
from Jeffrey Eugenides.
Biology gives you a brain.
Life turns it into a mind.
Thank you for listening, and hope to see you next time.
(upbeat music)
(upbeat music)