Afrikaans
Akan
Albanian
Amharic
Arabic
Armenian
Azerbaijani
Basque
Belarusian
Bemba
Bengali
Bihari
Bosnian
Breton
Bulgarian
Cambodian
Catalan
Cebuano
Cherokee
Chichewa
Chinese (Traditional)
Corsican
Croatian
Czech
Danish
Dutch
English
Esperanto
Estonian
Ewe
Faroese
Filipino
Finnish
French
Frisian
Ga
Galician
Georgian
German
Greek
Guarani
Gujarati
Haitian Creole
Hausa
Hawaiian
Hebrew
Hindi
Hmong
Hungarian
Icelandic
Igbo
Indonesian
Interlingua
Irish
Italian
Japanese
Javanese
Kannada
Kazakh
Kinyarwanda
Kirundi
Kongo
Korean
Krio (Sierra Leone)
Kurdish
Kurdish (Soranรฎ)
Kyrgyz
Laothian
Latin
Latvian
Lingala
Lithuanian
Lozi
Luganda
Luo
Luxembourgish
Macedonian
Malagasy
Malay
Malayalam
Maltese
Maori
Marathi
Mauritian Creole
Moldavian
Mongolian
Myanmar (Burmese)
Montenegrin
Nepali
Nigerian Pidgin
Northern Sotho
Norwegian
Norwegian (Nynorsk)
Occitan
Oriya
Oromo
Pashto
Persian
Polish
Portuguese (Brazil)
Portuguese (Portugal)
Punjabi
Quechua
Romanian
Romansh
Runyakitara
Russian
Samoan
Scots Gaelic
Serbian
Serbo-Croatian
Sesotho
Setswana
Seychellois Creole
Shona
Sindhi
Sinhalese
Slovak
Slovenian
Somali
Spanish
Spanish (Latin American)
Sundanese
Swahili
Swedish
Tajik
Tamil
Tatar
Telugu
Thai
Tigrinya
Tonga
Tshiluba
Tumbuka
Turkish
Turkmen
Twi
Uighur
Ukrainian
Urdu
Uzbek
Vietnamese
Welsh
Wolof
Xhosa
Yiddish
Yoruba
Zulu
I want you to picture something for a
I want you to picture something for a
I want you to picture something for a second.
second.
second. >> Okay. Um, imagine you're standing at the
>> Okay. Um, imagine you're standing at the
>> Okay. Um, imagine you're standing at the top of this really bumpy, uneven hill
top of this really bumpy, uneven hill
top of this really bumpy, uneven hill and you drop a marble. Okay. And as that
and you drop a marble. Okay. And as that
and you drop a marble. Okay. And as that marble rolls down, it starts hitting,
marble rolls down, it starts hitting,
marble rolls down, it starts hitting, you know, little ridges and it bounces
you know, little ridges and it bounces
you know, little ridges and it bounces into small valleys branching off into
into small valleys branching off into
into small valleys branching off into all these different paths until it
all these different paths until it
all these different paths until it finally comes to a complete stop at the
finally comes to a complete stop at the
finally comes to a complete stop at the bottom.
bottom.
bottom. >> Yeah, it is such a classic image. I
>> Yeah, it is such a classic image. I
>> Yeah, it is such a classic image. I mean, in developmental biology we refer
mean, in developmental biology we refer
mean, in developmental biology we refer to that as Waddington's landscape.
to that as Waddington's landscape.
to that as Waddington's landscape. >> Right, yeah.
>> Right, yeah.
>> Right, yeah. >> It was uh proposed decades ago to help
>> It was uh proposed decades ago to help
>> It was uh proposed decades ago to help visualize exactly how a cell makes
visualize exactly how a cell makes
visualize exactly how a cell makes decisions.
decisions.
decisions. >> Yeah, and instead of a marble, we're
>> Yeah, and instead of a marble, we're
>> Yeah, and instead of a marble, we're talking about a single generic stem
talking about a single generic stem
talking about a single generic stem cell.
cell.
cell. >> Exactly.
>> Exactly.
>> Exactly. >> As it rolls down this metaphorical
>> As it rolls down this metaphorical
>> As it rolls down this metaphorical landscape, it's um it's losing its
landscape, it's um it's losing its
landscape, it's um it's losing its limitless potential, right? It's
limitless potential, right? It's
limitless potential, right? It's figuring out its final identity. Like,
figuring out its final identity. Like,
figuring out its final identity. Like, will it slide into the valley that makes
will it slide into the valley that makes
will it slide into the valley that makes it a red blood cell?
it a red blood cell?
it a red blood cell? >> Mhm. Or maybe it bounces a different way
>> Mhm. Or maybe it bounces a different way
>> Mhm. Or maybe it bounces a different way and, you know, becomes a skin cell or a
and, you know, becomes a skin cell or a
and, you know, becomes a skin cell or a neuron. Right. And the holy grail for
neuron. Right. And the holy grail for
neuron. Right. And the holy grail for biologists has really always been trying
biologists has really always been trying
biologists has really always been trying to map out that exact landscape. We want
to map out that exact landscape. We want
to map out that exact landscape. We want to understand the invisible forces that
to understand the invisible forces that
to understand the invisible forces that push the marble down one specific valley
push the marble down one specific valley
push the marble down one specific valley instead of another because
instead of another because
instead of another because I mean, if we know that, we understand
I mean, if we know that, we understand
I mean, if we know that, we understand how life builds itself from scratch.
how life builds itself from scratch.
how life builds itself from scratch. >> Which brings us to the mission for
>> Which brings us to the mission for
>> Which brings us to the mission for today's deep dive.
today's deep dive.
today's deep dive. >> Mhm. We are looking at this incredible
>> Mhm. We are looking at this incredible
>> Mhm. We are looking at this incredible 2026 paper published in the journal
2026 paper published in the journal
2026 paper published in the journal Cell. Yeah, a massive paper.
Cell. Yeah, a massive paper.
Cell. Yeah, a massive paper. >> It really is. Uh, it's titled Reg-Velo,
>> It really is. Uh, it's titled Reg-Velo,
>> It really is. Uh, it's titled Reg-Velo, Gene Regulatory Informed Dynamics of
Gene Regulatory Informed Dynamics of
Gene Regulatory Informed Dynamics of Single Cells by Wang and a pretty large
Single Cells by Wang and a pretty large
Single Cells by Wang and a pretty large team of researchers.
team of researchers.
team of researchers. >> Yeah. And our goal today is to unpack
>> Yeah. And our goal today is to unpack
>> Yeah. And our goal today is to unpack how this team has finally built like a
how this team has finally built like a
how this team has finally built like a mathematical bridge between the speed at
mathematical bridge between the speed at
mathematical bridge between the speed at which a cell changes and the internal
which a cell changes and the internal
which a cell changes and the internal software that actually programs that
software that actually programs that
software that actually programs that change. To fully grasp why Reg-Velo is
change. To fully grasp why Reg-Velo is
change. To fully grasp why Reg-Velo is sending such shockwaves through
sending such shockwaves through
sending such shockwaves through computational biology, we really have to
computational biology, we really have to
computational biology, we really have to look at the massive blind spot that
look at the massive blind spot that
look at the massive blind spot that researchers were dealing with before
researchers were dealing with before
researchers were dealing with before this paper even came out.
this paper even came out.
this paper even came out. >> Right. Historically, we had two
>> Right. Historically, we had two
>> Right. Historically, we had two incredibly powerful, but um, completely
incredibly powerful, but um, completely
incredibly powerful, but um, completely isolated tools for studying cells. Let's
isolated tools for studying cells. Let's
isolated tools for studying cells. Let's lay those out for everyone listening.
lay those out for everyone listening.
lay those out for everyone listening. The first one is a concept called RNA
The first one is a concept called RNA
The first one is a concept called RNA velocity.
velocity.
velocity. >> Yes. And I think of RNA velocity as a
>> Yes. And I think of RNA velocity as a
>> Yes. And I think of RNA velocity as a well, like a cellular speedometer.
well, like a cellular speedometer.
well, like a cellular speedometer. That's a great way to put it.
That's a great way to put it.
That's a great way to put it. >> It doesn't just tell you what a cell is
>> It doesn't just tell you what a cell is
>> It doesn't just tell you what a cell is doing right now. It measures the speed
doing right now. It measures the speed
doing right now. It measures the speed and the direction of the cell's
and the direction of the cell's
and the direction of the cell's development uh to predict what it's
development uh to predict what it's
development uh to predict what it's going to do next.
going to do next.
going to do next. >> Right. And the second tool is what we
>> Right. And the second tool is what we
>> Right. And the second tool is what we call a gene regulatory network or a GRN.
call a gene regulatory network or a GRN.
call a gene regulatory network or a GRN. Okay. Think of the GRN as the circuitry.
Okay. Think of the GRN as the circuitry.
Okay. Think of the GRN as the circuitry. It is the genetic software map that
It is the genetic software map that
It is the genetic software map that dictates, you know, which genes are
dictates, you know, which genes are
dictates, you know, which genes are turning on and which genes are turning
turning on and which genes are turning
turning on and which genes are turning off to actually drive the cell's
off to actually drive the cell's
off to actually drive the cell's development.
development.
development. >> So, we have a speedometer and we have a
>> So, we have a speedometer and we have a
>> So, we have a speedometer and we have a wiring diagram.
wiring diagram.
wiring diagram. >> Exactly.
>> Exactly.
>> Exactly. >> But here is the crazy part.
>> But here is the crazy part.
>> But here is the crazy part. >> Yeah. Before Reg-Velo, biologists had
>> Yeah. Before Reg-Velo, biologists had
>> Yeah. Before Reg-Velo, biologists had absolutely no way to connect the two.
absolutely no way to connect the two.
absolutely no way to connect the two. None. It's like imagine sitting in a car
None. It's like imagine sitting in a car
None. It's like imagine sitting in a car with a speedometer that tells you you're
with a speedometer that tells you you're
with a speedometer that tells you you're accelerating
accelerating
accelerating and you're holding a schematic of the
and you're holding a schematic of the
and you're holding a schematic of the engine's electrical wiring in your
engine's electrical wiring in your
engine's electrical wiring in your hands, but you have no mechanical way to
hands, but you have no mechanical way to
hands, but you have no mechanical way to see how pressing the gas pedal actually
see how pressing the gas pedal actually
see how pressing the gas pedal actually changes the speed.
changes the speed.
changes the speed. >> Yeah, it was a profound conceptual
>> Yeah, it was a profound conceptual
>> Yeah, it was a profound conceptual disconnect. I mean, we could watch the
disconnect. I mean, we could watch the
disconnect. I mean, we could watch the cells move along their developmental
cells move along their developmental
cells move along their developmental paths and we could draw these static
paths and we could draw these static
paths and we could draw these static maps of gene networks, but we just
maps of gene networks, but we just
maps of gene networks, but we just couldn't link them mechanically. We had
couldn't link them mechanically. We had
couldn't link them mechanically. We had literally no mathematical way to prove
literally no mathematical way to prove
literally no mathematical way to prove how the wiring was driving the motion.
how the wiring was driving the motion.
how the wiring was driving the motion. So, to understand how they finally
So, to understand how they finally
So, to understand how they finally solved this, I feel like we really need
solved this, I feel like we really need
solved this, I feel like we really need to lift the hood and look at how that
to lift the hood and look at how that
to lift the hood and look at how that speedometer, the RNA velocity, actually
speedometer, the RNA velocity, actually
speedometer, the RNA velocity, actually works. Good idea. Because from what the
works. Good idea. Because from what the
works. Good idea. Because from what the sources show, the old way we calculated
sources show, the old way we calculated
sources show, the old way we calculated it was frankly a little broken. Oh, it
it was frankly a little broken. Oh, it
it was frankly a little broken. Oh, it was heavily flawed. But, you know, the
was heavily flawed. But, you know, the
was heavily flawed. But, you know, the underlying biological mechanism of RNA
underlying biological mechanism of RNA
underlying biological mechanism of RNA velocity is actually quite elegant.
velocity is actually quite elegant.
velocity is actually quite elegant. Okay, break that down for us.
Okay, break that down for us.
Okay, break that down for us. >> It relies on the life cycle of RNA
>> It relies on the life cycle of RNA
>> It relies on the life cycle of RNA inside the cell. When a gene is
inside the cell. When a gene is
inside the cell. When a gene is activated, it doesn't instantly produce
activated, it doesn't instantly produce
activated, it doesn't instantly produce a perfect finished protein. Right.
a perfect finished protein. Right.
a perfect finished protein. Right. >> First, it transcribes this rough draft,
>> First, it transcribes this rough draft,
>> First, it transcribes this rough draft, which we call unspliced RNA. It's the
which we call unspliced RNA. It's the
which we call unspliced RNA. It's the raw, newly minted instruction manual.
raw, newly minted instruction manual.
raw, newly minted instruction manual. >> Exactly.
>> Exactly.
>> Exactly. >> has all the um the biological junk in it
>> has all the um the biological junk in it
>> has all the um the biological junk in it that hasn't been edited out yet.
that hasn't been edited out yet.
that hasn't been edited out yet. >> Yes. And then the sales machinery comes
>> Yes. And then the sales machinery comes
>> Yes. And then the sales machinery comes in, cuts out the unneeded parts, and
in, cuts out the unneeded parts, and
in, cuts out the unneeded parts, and processes it into what we call spliced
processes it into what we call spliced
processes it into what we call spliced RNA.
RNA.
RNA. >> Okay, that was okay.
>> Okay, that was okay.
>> Okay, that was okay. >> That is the mature, ready-to-use
>> That is the mature, ready-to-use
>> That is the mature, ready-to-use instruction. This makes perfect sense
instruction. This makes perfect sense
instruction. This makes perfect sense when you look at how scientists actually
when you look at how scientists actually
when you look at how scientists actually use it to predict the future.
use it to predict the future.
use it to predict the future. Like, by looking at the ratio of raw,
Like, by looking at the ratio of raw,
Like, by looking at the ratio of raw, unspliced RNA to mature, spliced RNA,
unspliced RNA to mature, spliced RNA,
unspliced RNA to mature, spliced RNA, you can tell if a gene is ramping up or
you can tell if a gene is ramping up or
you can tell if a gene is ramping up or powering down.
powering down.
powering down. >> Precisely. It reminds me of
>> Precisely. It reminds me of
>> Precisely. It reminds me of walking into a kitchen. Ooh, I like
walking into a kitchen. Ooh, I like
walking into a kitchen. Ooh, I like this.
this.
this. If you look on the counter and see raw
If you look on the counter and see raw
If you look on the counter and see raw ingredients, you know, flour, sugar,
ingredients, you know, flour, sugar,
ingredients, you know, flour, sugar, carton of eggs,
carton of eggs,
carton of eggs, and nothing is in the oven yet.
and nothing is in the oven yet.
and nothing is in the oven yet. >> Right, you know what's coming. Exactly.
>> Right, you know what's coming. Exactly.
>> Right, you know what's coming. Exactly. You can predict with near certainty that
You can predict with near certainty that
You can predict with near certainty that the chef is about to bake something. You
the chef is about to bake something. You
the chef is about to bake something. You are literally looking at the future
are literally looking at the future
are literally looking at the future state of the kitchen by just measuring
state of the kitchen by just measuring
state of the kitchen by just measuring the raw materials. That ratio of raw to
the raw materials. That ratio of raw to
the raw materials. That ratio of raw to mature RNA allowed previous mathematical
mature RNA allowed previous mathematical
mature RNA allowed previous mathematical models like Svelo to create these really
models like Svelo to create these really
models like Svelo to create these really beautiful vector fields. They could
beautiful vector fields. They could
beautiful vector fields. They could literally draw arrows showing exactly
literally draw arrows showing exactly
literally draw arrows showing exactly where cell was heading.
where cell was heading.
where cell was heading. >> Okay, but But, yeah, to make that math
>> Okay, but But, yeah, to make that math
>> Okay, but But, yeah, to make that math work, those older models had to rely on
work, those older models had to rely on
work, those older models had to rely on some incredibly restrictive and frankly
some incredibly restrictive and frankly
some incredibly restrictive and frankly unrealistic assumptions.
unrealistic assumptions.
unrealistic assumptions. >> This is where the old models hit a
>> This is where the old models hit a
>> This is where the old models hit a massive wall.
massive wall.
massive wall. What exactly were they assuming? Cuz I
What exactly were they assuming? Cuz I
What exactly were they assuming? Cuz I know the math gets a little dense here.
know the math gets a little dense here.
know the math gets a little dense here. Well, they assumed that every single
Well, they assumed that every single
Well, they assumed that every single gene in the cell operated completely
gene in the cell operated completely
gene in the cell operated completely independently of every other gene.
independently of every other gene.
independently of every other gene. Independently.
Independently.
Independently. >> Yeah. In mathematical terms, they used
>> Yeah. In mathematical terms, they used
>> Yeah. In mathematical terms, they used what's called decoupled one-dimensional
what's called decoupled one-dimensional
what's called decoupled one-dimensional ordinary differential equations, or ODEs
ordinary differential equations, or ODEs
ordinary differential equations, or ODEs for short. Wait, wait, I need to stop
for short. Wait, wait, I need to stop
for short. Wait, wait, I need to stop you there. They assumed genes operate in
you there. They assumed genes operate in
you there. They assumed genes operate in a vacuum. Basically, yes. That's like
a vacuum. Basically, yes. That's like
a vacuum. Basically, yes. That's like trying to understand city traffic by
trying to understand city traffic by
trying to understand city traffic by focusing a camera on one single car,
focusing a camera on one single car,
focusing a camera on one single car, while completely ignoring the traffic
while completely ignoring the traffic
while completely ignoring the traffic lights, the stop signs, and the hundreds
lights, the stop signs, and the hundreds
lights, the stop signs, and the hundreds of other drivers swerving around it.
of other drivers swerving around it.
of other drivers swerving around it. >> It is a severe oversimplification, yes.
>> It is a severe oversimplification, yes.
>> It is a severe oversimplification, yes. But, um, it was the only way they could
But, um, it was the only way they could
But, um, it was the only way they could get the equations to resolve at the
get the equations to resolve at the
get the equations to resolve at the time. Decoupled just means the equation
time. Decoupled just means the equation
time. Decoupled just means the equation for gene A simply doesn't care what gene
for gene A simply doesn't care what gene
for gene A simply doesn't care what gene B is doing.
B is doing.
B is doing. And furthermore, they assumed the speed
And furthermore, they assumed the speed
And furthermore, they assumed the speed at which that raw RNA was being
at which that raw RNA was being
at which that raw RNA was being produced, the transcription rate, was
produced, the transcription rate, was
produced, the transcription rate, was just a constant unchanging background
just a constant unchanging background
just a constant unchanging background hum. The biology is basically the exact
hum. The biology is basically the exact
hum. The biology is basically the exact opposite of a vacuum. Exactly.
opposite of a vacuum. Exactly.
opposite of a vacuum. Exactly. >> It's a highly interactive, super messy
>> It's a highly interactive, super messy
>> It's a highly interactive, super messy system. You have upstream genes acting
system. You have upstream genes acting
system. You have upstream genes acting like, well, like traffic cops,
like, well, like traffic cops,
like, well, like traffic cops, constantly turning downstream genes on
constantly turning downstream genes on
constantly turning downstream genes on and off. And those upstream genes are
and off. And those upstream genes are
and off. And those upstream genes are called transcription factors. By
called transcription factors. By
called transcription factors. By excluding them, you know, by ignoring
excluding them, you know, by ignoring
excluding them, you know, by ignoring the regulatory traffic cops that
the regulatory traffic cops that
the regulatory traffic cops that actually drive a cell's differentiation,
actually drive a cell's differentiation,
actually drive a cell's differentiation, those old RNA velocity models were
those old RNA velocity models were
those old RNA velocity models were fundamentally missing the underlying
fundamentally missing the underlying
fundamentally missing the underlying mechanism of Waddington's landscape.
mechanism of Waddington's landscape.
mechanism of Waddington's landscape. They could see the marble rolling, but
They could see the marble rolling, but
They could see the marble rolling, but they had absolutely no idea why it was
they had absolutely no idea why it was
they had absolutely no idea why it was rolling.
rolling.
rolling. >> Precisely. Okay. So, let's do a quick
>> Precisely. Okay. So, let's do a quick
>> Precisely. Okay. So, let's do a quick recap for you listening, make sure we're
recap for you listening, make sure we're
recap for you listening, make sure we're all on the same page.
all on the same page.
all on the same page. >> Got you. Bullet one, we have
>> Got you. Bullet one, we have
>> Got you. Bullet one, we have Waddington's landscape, which is our
Waddington's landscape, which is our
Waddington's landscape, which is our mental image of how a cell decides its
mental image of how a cell decides its
mental image of how a cell decides its fate, rolling down a hill. Yep. Bullet
fate, rolling down a hill. Yep. Bullet
fate, rolling down a hill. Yep. Bullet two, we have RNA velocity, our
two, we have RNA velocity, our
two, we have RNA velocity, our speedometer that predicts future cell
speedometer that predicts future cell
speedometer that predicts future cell states by looking at raw versus mature
states by looking at raw versus mature
states by looking at raw versus mature RNA in the kitchen. Perfect. And bullet
RNA in the kitchen. Perfect. And bullet
RNA in the kitchen. Perfect. And bullet three, the old mathematical models
three, the old mathematical models
three, the old mathematical models completely ignored the genetic wiring,
completely ignored the genetic wiring,
completely ignored the genetic wiring, the GRNs, that connect all these
the GRNs, that connect all these
the GRNs, that connect all these changes.
changes.
changes. >> Because they use those decoupled
>> Because they use those decoupled
>> Because they use those decoupled one-dimensional equations.
one-dimensional equations.
one-dimensional equations. >> Right. So, the big question for you, the
>> Right. So, the big question for you, the
>> Right. So, the big question for you, the listener,
listener,
listener, is how do we actually fuse this wiring
is how do we actually fuse this wiring
is how do we actually fuse this wiring map with the velocity prediction? How do
map with the velocity prediction? How do
map with the velocity prediction? How do we build a better model? And that is
we build a better model? And that is
we build a better model? And that is exactly where the authors of this paper
exactly where the authors of this paper
exactly where the authors of this paper come in. They developed a framework
come in. They developed a framework
come in. They developed a framework called RegVelo. RegVelo. It stands for
called RegVelo. RegVelo. It stands for
called RegVelo. RegVelo. It stands for regulatory velocity.
regulatory velocity.
regulatory velocity. And rather than relying on those old
And rather than relying on those old
And rather than relying on those old decoupled equations we just talked
decoupled equations we just talked
decoupled equations we just talked about, RegVelo is built on deep
about, RegVelo is built on deep
about, RegVelo is built on deep generative modeling. Okay. It merges the
generative modeling. Okay. It merges the
generative modeling. Okay. It merges the RNA velocity equations and the gene
RNA velocity equations and the gene
RNA velocity equations and the gene regulatory networks into a single
regulatory networks into a single
regulatory networks into a single high-dimensional ordinary differential
high-dimensional ordinary differential
high-dimensional ordinary differential equation.
equation.
equation. >> A high-dimensional equation. Okay, let's
>> A high-dimensional equation. Okay, let's
>> A high-dimensional equation. Okay, let's translate that out of math speak for a
translate that out of math speak for a
translate that out of math speak for a second.
second.
second. >> Sure. If the old models were thousands
>> Sure. If the old models were thousands
>> Sure. If the old models were thousands of isolated single-lane dirt roads, Reg
of isolated single-lane dirt roads, Reg
of isolated single-lane dirt roads, Reg Velo is a mathematical superhighway
Velo is a mathematical superhighway
Velo is a mathematical superhighway system where every single road
system where every single road
system where every single road intersects. Oh, I see. Transcription is
intersects. Oh, I see. Transcription is
intersects. Oh, I see. Transcription is no longer modeled as a constant isolated
no longer modeled as a constant isolated
no longer modeled as a constant isolated event. It is now treated as a dynamic
event. It is now treated as a dynamic
event. It is now treated as a dynamic process that depends entirely on the
process that depends entirely on the
process that depends entirely on the regulations of other genes and the
regulations of other genes and the
regulations of other genes and the passage of time. I found the mechanics
passage of time. I found the mechanics
passage of time. I found the mechanics of how they actually built this
of how they actually built this
of how they actually built this superhighway fascinating because um
superhighway fascinating because um
superhighway fascinating because um Reg Velo doesn't just guess how these
Reg Velo doesn't just guess how these
Reg Velo doesn't just guess how these genes connect out of thin air, right?
genes connect out of thin air, right?
genes connect out of thin air, right? >> No, it can't.
>> No, it can't.
>> No, it can't. >> a starting point.
>> a starting point.
>> a starting point. >> Yeah. So, the scientists feed the
>> Yeah. So, the scientists feed the
>> Yeah. So, the scientists feed the algorithm a starter map of the gene
algorithm a starter map of the gene
algorithm a starter map of the gene regulatory network. Right. And they
regulatory network. Right. And they
regulatory network. Right. And they derive that starter map from other types
derive that starter map from other types
derive that starter map from other types of single-cell data. Often, they use
of single-cell data. Often, they use
of single-cell data. Often, they use something called multiomic data.
something called multiomic data.
something called multiomic data. >> Multiomic?
>> Multiomic?
>> Multiomic? >> Which involves looking at the physical
>> Which involves looking at the physical
>> Which involves looking at the physical accessibility of the DNA itself to guess
accessibility of the DNA itself to guess
accessibility of the DNA itself to guess which genes might be able to talk to
which genes might be able to talk to
which genes might be able to talk to each other.
each other.
each other. We refer to these starting maps as
We refer to these starting maps as
We refer to these starting maps as epigenetic priors.
epigenetic priors.
epigenetic priors. >> Epigenetic priors.
>> Epigenetic priors.
>> Epigenetic priors. >> Yeah. I love that concept. It's like
>> Yeah. I love that concept. It's like
>> Yeah. I love that concept. It's like looking at the physical shape of the
looking at the physical shape of the
looking at the physical shape of the DNA, like seeing which chapters of the
DNA, like seeing which chapters of the
DNA, like seeing which chapters of the instruction manual are physically open
instruction manual are physically open
instruction manual are physically open and readable to guess the wiring
and readable to guess the wiring
and readable to guess the wiring diagram.
diagram.
diagram. So,
So,
So, >> take that rough starter map and they
>> take that rough starter map and they
>> take that rough starter map and they feed it into a deep neural network. And
feed it into a deep neural network. And
feed it into a deep neural network. And that neural network has to calculate
that neural network has to calculate
that neural network has to calculate something absolutely critical called
something absolutely critical called
something absolutely critical called latent time.
latent time.
latent time. >> I'm really hoping we'd get to latent
>> I'm really hoping we'd get to latent
>> I'm really hoping we'd get to latent time.
time.
time. >> Oh, it's so cool.
>> Oh, it's so cool.
>> Oh, it's so cool. >> Because when you scoop up a sample of
>> Because when you scoop up a sample of
>> Because when you scoop up a sample of cells to study them, you aren't
cells to study them, you aren't
cells to study them, you aren't capturing them all at the exact same
capturing them all at the exact same
capturing them all at the exact same moment in their life cycle.
moment in their life cycle.
moment in their life cycle. >> No, it's a mixed bag.
>> No, it's a mixed bag.
>> No, it's a mixed bag. >> Right. So, latent time is like the
>> Right. So, latent time is like the
>> Right. So, latent time is like the algorithm calculating an internal
algorithm calculating an internal
algorithm calculating an internal biological clock for each individual
biological clock for each individual
biological clock for each individual cell, placing them in perfect
cell, placing them in perfect
cell, placing them in perfect chronological order along a timeline of
chronological order along a timeline of
chronological order along a timeline of development. It is a crucial step. And
development. It is a crucial step. And
development. It is a crucial step. And alongside that internal clock, the
alongside that internal clock, the
alongside that internal clock, the neural network also generates a massive
neural network also generates a massive
neural network also generates a massive matrix of weights to represent the
matrix of weights to represent the
matrix of weights to represent the regulatory network.
regulatory network.
regulatory network. >> of weights.
>> of weights.
>> of weights. >> Yes. It is assigning a mathematical
>> Yes. It is assigning a mathematical
>> Yes. It is assigning a mathematical score to how different genes interact
score to how different genes interact
score to how different genes interact with each other. This weight matrix is
with each other. This weight matrix is
with each other. This weight matrix is where this concept really clicked for me
where this concept really clicked for me
where this concept really clicked for me when I was reading the paper.
when I was reading the paper.
when I was reading the paper. >> Yeah. Yeah. If the neural network
>> Yeah. Yeah. If the neural network
>> Yeah. Yeah. If the neural network observes the data and gives a connection
observes the data and gives a connection
observes the data and gives a connection a positive number, it means gene A
a positive number, it means gene A
a positive number, it means gene A actively turns on gene B. Yes. And if it
actively turns on gene B. Yes. And if it
actively turns on gene B. Yes. And if it gives it a negative number, gene A
gives it a negative number, gene A
gives it a negative number, gene A represses or shuts down gene B. Exactly.
represses or shuts down gene B. Exactly.
represses or shuts down gene B. Exactly. And a zero means those two genes simply
And a zero means those two genes simply
And a zero means those two genes simply don't interact at all.
don't interact at all.
don't interact at all. Okay. But here is the true stroke of
Okay. But here is the true stroke of
Okay. But here is the true stroke of genius in the Reg-Velo framework. What's
genius in the Reg-Velo framework. What's
genius in the Reg-Velo framework. What's that? It knows that biological starter
that? It knows that biological starter
that? It knows that biological starter maps, those epigenetic priors we talked
maps, those epigenetic priors we talked
maps, those epigenetic priors we talked about, are notoriously noisy and
about, are notoriously noisy and
about, are notoriously noisy and imperfect. It does not blindly trust the
imperfect. It does not blindly trust the
imperfect. It does not blindly trust the prior information it was given. Because
prior information it was given. Because
prior information it was given. Because in reality, a lot of those initial
in reality, a lot of those initial
in reality, a lot of those initial guesses about which genes talk to each
guesses about which genes talk to each
guesses about which genes talk to each other are just flat out wrong.
other are just flat out wrong.
other are just flat out wrong. >> Exactly. Biological networks are
>> Exactly. Biological networks are
>> Exactly. Biological networks are governed by a principle called sparsity.
governed by a principle called sparsity.
governed by a principle called sparsity. >> Sparsity. Let's define that. It just
>> Sparsity. Let's define that. It just
>> Sparsity. Let's define that. It just means not every gene interacts with
means not every gene interacts with
means not every gene interacts with every other gene. Most only communicate
every other gene. Most only communicate
every other gene. Most only communicate with a highly specific select few.
with a highly specific select few.
with a highly specific select few. Right. So, Reg-Velo's deep learning
Right. So, Reg-Velo's deep learning
Right. So, Reg-Velo's deep learning model uses the actual RNA velocity data
model uses the actual RNA velocity data
model uses the actual RNA velocity data to learn new edge weights.
to learn new edge weights.
to learn new edge weights. It essentially self-corrects.
It essentially self-corrects.
It essentially self-corrects. >> That's wild.
>> That's wild.
>> That's wild. >> Yeah. It fixes erroneous assumptions in
>> Yeah. It fixes erroneous assumptions in
>> Yeah. It fixes erroneous assumptions in the original map and discovers hidden
the original map and discovers hidden
the original map and discovers hidden connections that the researchers hadn't
connections that the researchers hadn't
connections that the researchers hadn't even mapped yet.
even mapped yet.
even mapped yet. >> It's rebuilding the car's wiring diagram
>> It's rebuilding the car's wiring diagram
>> It's rebuilding the car's wiring diagram while the car is actively driving down
while the car is actively driving down
while the car is actively driving down the highway. Yes. That's incredible. So,
the highway. Yes. That's incredible. So,
the highway. Yes. That's incredible. So, time for another quick recap.
time for another quick recap.
time for another quick recap. >> Let's do it. Bullet one, Reg-Velo merges
>> Let's do it. Bullet one, Reg-Velo merges
>> Let's do it. Bullet one, Reg-Velo merges RNA velocity with those gene regulatory
RNA velocity with those gene regulatory
RNA velocity with those gene regulatory networks.
networks.
networks. >> Yeah. Bullet two, it uses complex
>> Yeah. Bullet two, it uses complex
>> Yeah. Bullet two, it uses complex coupled math, those high-dimensional
coupled math, those high-dimensional
coupled math, those high-dimensional ODEs, instead of isolated equations.
ODEs, instead of isolated equations.
ODEs, instead of isolated equations. >> Right. And bullet three, it refines
>> Right. And bullet three, it refines
>> Right. And bullet three, it refines imperfect prior knowledge using deep
imperfect prior knowledge using deep
imperfect prior knowledge using deep learning, keeping sparsity in mind.
learning, keeping sparsity in mind.
learning, keeping sparsity in mind. Perfectly summarized. So, for you
Perfectly summarized. So, for you
Perfectly summarized. So, for you listening, think about this question. If
listening, think about this question. If
listening, think about this question. If we have this highly accurate,
we have this highly accurate,
we have this highly accurate, mathematically coupled virtual model,
mathematically coupled virtual model,
mathematically coupled virtual model, what happens if we start tinkering with
what happens if we start tinkering with
what happens if we start tinkering with it? And this is where Reg Velo
it? And this is where Reg Velo
it? And this is where Reg Velo transitions from being just an
transitions from being just an
transitions from being just an observational tool to being a full-on
observational tool to being a full-on
observational tool to being a full-on experimental playground.
experimental playground.
experimental playground. >> Oh, this part is so cool. The
>> Oh, this part is so cool. The
>> Oh, this part is so cool. The researchers connected Reg Velo to an
researchers connected Reg Velo to an
researchers connected Reg Velo to an analytical tool called Cell Rank 2. Cell
analytical tool called Cell Rank 2. Cell
analytical tool called Cell Rank 2. Cell Rank 2? And this combination allows
Rank 2? And this combination allows
Rank 2? And this combination allows scientists to perform what are known as
scientists to perform what are known as
scientists to perform what are known as in silico perturbations.
in silico perturbations.
in silico perturbations. >> In silico meaning entirely inside the
>> In silico meaning entirely inside the
>> In silico meaning entirely inside the computer. Exactly. They have essentially
computer. Exactly. They have essentially
computer. Exactly. They have essentially constructed a fully actionable virtual
constructed a fully actionable virtual
constructed a fully actionable virtual cell.
cell.
cell. I'm trying to picture what this looks
I'm trying to picture what this looks
I'm trying to picture what this looks like in practice because if I'm a
like in practice because if I'm a
like in practice because if I'm a researcher studying a specific organ,
researcher studying a specific organ,
researcher studying a specific organ, normally I'd spend what, a year breeding
normally I'd spend what, a year breeding
normally I'd spend what, a year breeding mice? Oh, easily.
mice? Oh, easily.
mice? Oh, easily. >> Chemically altering their DNA and
>> Chemically altering their DNA and
>> Chemically altering their DNA and waiting to see what a specific gene
waiting to see what a specific gene
waiting to see what a specific gene mutation does.
mutation does.
mutation does. >> Yeah. But with this, am I literally just
>> Yeah. But with this, am I literally just
>> Yeah. But with this, am I literally just typing code to delete a gene and
typing code to delete a gene and
typing code to delete a gene and watching the virtual organ break?
watching the virtual organ break?
watching the virtual organ break? >> That's exactly what you were doing. You
>> That's exactly what you were doing. You
>> That's exactly what you were doing. You go into the matrix, you digitally erase
go into the matrix, you digitally erase
go into the matrix, you digitally erase a gene and all of its downstream
a gene and all of its downstream
a gene and all of its downstream targets, you run a forward simulation,
targets, you run a forward simulation,
targets, you run a forward simulation, and you watch the mathematical
and you watch the mathematical
and you watch the mathematical consequences ripple through the cell's
consequences ripple through the cell's
consequences ripple through the cell's fate. Wow. It allows for counterfactual
fate. Wow. It allows for counterfactual
fate. Wow. It allows for counterfactual inference. You know, asking the model,
inference. You know, asking the model,
inference. You know, asking the model, "What would happen if this specific
"What would happen if this specific
"What would happen if this specific traffic cop disappeared?" So, to prove
traffic cop disappeared?" So, to prove
traffic cop disappeared?" So, to prove this wasn't just a fancy math trick, the
this wasn't just a fancy math trick, the
this wasn't just a fancy math trick, the team obviously had to test it on some
team obviously had to test it on some
team obviously had to test it on some highly researched biological systems.
highly researched biological systems.
highly researched biological systems. Naturally. And I know they started with
Naturally. And I know they started with
Naturally. And I know they started with mouse pancreatic development.
mouse pancreatic development.
mouse pancreatic development. >> Yeah, the pancreas is a great testing
>> Yeah, the pancreas is a great testing
>> Yeah, the pancreas is a great testing ground because we already know a lot
ground because we already know a lot
ground because we already know a lot about how its different cells develop,
about how its different cells develop,
about how its different cells develop, like the insulin-producing beta cells
like the insulin-producing beta cells
like the insulin-producing beta cells and the ductal cells.
and the ductal cells.
and the ductal cells. >> Right. When they fed the raw pancreatic
>> Right. When they fed the raw pancreatic
>> Right. When they fed the raw pancreatic data into Reg Velo, the model
data into Reg Velo, the model
data into Reg Velo, the model independently identified the known
independently identified the known
independently identified the known driver genes.
driver genes.
driver genes. >> Which is a huge first step. And then,
>> Which is a huge first step. And then,
>> Which is a huge first step. And then, when they did a virtual knockout of a
when they did a virtual knockout of a
when they did a virtual knockout of a gene called E2F1,
gene called E2F1,
gene called E2F1, >> Yes. the simulation instantly showed a
>> Yes. the simulation instantly showed a
>> Yes. the simulation instantly showed a massive disruption in the cell cycle of
massive disruption in the cell cycle of
massive disruption in the cell cycle of the ductal cells.
the ductal cells.
the ductal cells. >> Exactly. And then when they digitally
>> Exactly. And then when they digitally
>> Exactly. And then when they digitally knocked out a gene called RFF6, the
knocked out a gene called RFF6, the
knocked out a gene called RFF6, the virtual cell suddenly failed to
virtual cell suddenly failed to
virtual cell suddenly failed to differentiate into beta cells. It
differentiate into beta cells. It
differentiate into beta cells. It perfectly mirrored decades of hard
perfectly mirrored decades of hard
perfectly mirrored decades of hard thought, real-world biological
thought, real-world biological
thought, real-world biological experiments just by crunching the data.
experiments just by crunching the data.
experiments just by crunching the data. But an even more striking example from
But an even more striking example from
But an even more striking example from the paper involves human hematopoiesis.
the paper involves human hematopoiesis.
the paper involves human hematopoiesis. The formation of blood cells, yes. This
The formation of blood cells, yes. This
The formation of blood cells, yes. This system relies on a classic textbook
system relies on a classic textbook
system relies on a classic textbook example of gene regulation known as a
example of gene regulation known as a
example of gene regulation known as a toggle switch. This is one of my
toggle switch. This is one of my
toggle switch. This is one of my favorite parts of the paper.
favorite parts of the paper.
favorite parts of the paper. >> Let's break it down. It's the tug-of-war
>> Let's break it down. It's the tug-of-war
>> Let's break it down. It's the tug-of-war between GATA1 and SPI1. Yes. And SPI1 is
between GATA1 and SPI1. Yes. And SPI1 is
between GATA1 and SPI1. Yes. And SPI1 is also commonly known as PU.1 just for
also commonly known as PU.1 just for
also commonly known as PU.1 just for anyone who might know it by that name.
anyone who might know it by that name.
anyone who might know it by that name. >> Okay, so GATA1 versus SPI1. When a human
>> Okay, so GATA1 versus SPI1. When a human
>> Okay, so GATA1 versus SPI1. When a human stem cell is preparing to become a blood
stem cell is preparing to become a blood
stem cell is preparing to become a blood cell, it faces a critical fork in
cell, it faces a critical fork in
cell, it faces a critical fork in Waddington's landscape. Right. It can
Waddington's landscape. Right. It can
Waddington's landscape. Right. It can slide down one valley to become an
slide down one valley to become an
slide down one valley to become an erythroid cell, which is a red blood
erythroid cell, which is a red blood
erythroid cell, which is a red blood cell that carries oxygen. Okay. Or it
cell that carries oxygen. Okay. Or it
cell that carries oxygen. Okay. Or it can slide down a completely different
can slide down a completely different
can slide down a completely different valley to become a monocyte, which is a
valley to become a monocyte, which is a
valley to become a monocyte, which is a white blood cell that fights off
white blood cell that fights off
white blood cell that fights off infections. And the thing deciding which
infections. And the thing deciding which
infections. And the thing deciding which valley the cell falls into is a literal
valley the cell falls into is a literal
valley the cell falls into is a literal biological tug-of-war between those two
biological tug-of-war between those two
biological tug-of-war between those two transcription factors. Exactly. GATA1
transcription factors. Exactly. GATA1
transcription factors. Exactly. GATA1 and SPI1 mutually repress each other. So
and SPI1 mutually repress each other. So
and SPI1 mutually repress each other. So if GATA1 gets the upper hand, it
if GATA1 gets the upper hand, it
if GATA1 gets the upper hand, it physically blocks SPI1 from functioning
physically blocks SPI1 from functioning
physically blocks SPI1 from functioning and the cell becomes a red blood cell.
and the cell becomes a red blood cell.
and the cell becomes a red blood cell. Yes. And if SPI1 wins the tug-of-war, it
Yes. And if SPI1 wins the tug-of-war, it
Yes. And if SPI1 wins the tug-of-war, it shuts down GATA1 and the cell becomes a
shuts down GATA1 and the cell becomes a
shuts down GATA1 and the cell becomes a monocyte. What is so impressive is that
monocyte. What is so impressive is that
monocyte. What is so impressive is that Rigvel captured this toggle switch
Rigvel captured this toggle switch
Rigvel captured this toggle switch perfectly. The researchers went into the
perfectly. The researchers went into the
perfectly. The researchers went into the virtual model, computationally deleted
virtual model, computationally deleted
virtual model, computationally deleted the GATA1 network, and they watched the
the GATA1 network, and they watched the
the GATA1 network, and they watched the mathematical fate probabilities
mathematical fate probabilities
mathematical fate probabilities instantly shift entirely toward the
instantly shift entirely toward the
instantly shift entirely toward the monocyte lineage. It's a perfect digital
monocyte lineage. It's a perfect digital
monocyte lineage. It's a perfect digital mirror of biological reality. It really
mirror of biological reality. It really
mirror of biological reality. It really is. And it's worth noting that the older
is. And it's worth noting that the older
is. And it's worth noting that the older decoupled RNA velocity models totally
decoupled RNA velocity models totally
decoupled RNA velocity models totally failed at this. Miserably. Because they
failed at this. Miserably. Because they
failed at this. Miserably. Because they couldn't see the regulatory traffic
couldn't see the regulatory traffic
couldn't see the regulatory traffic cops, they couldn't predict the
cops, they couldn't predict the
cops, they couldn't predict the tug-of-war. Even when previous models
tug-of-war. Even when previous models
tug-of-war. Even when previous models were fed extra highly detailed metabolic
were fed extra highly detailed metabolic
were fed extra highly detailed metabolic data, they still couldn't accurately
data, they still couldn't accurately
data, they still couldn't accurately predict these driver genes because they
predict these driver genes because they
predict these driver genes because they fundamentally lacked the regulatory
fundamentally lacked the regulatory
fundamentally lacked the regulatory context.
context.
context. >> That's the power of coupling the math.
>> That's the power of coupling the math.
>> That's the power of coupling the math. Okay, time for recap number three. Go
Okay, time for recap number three. Go
Okay, time for recap number three. Go for it. Bullet one, RegVelo acts as a
for it. Bullet one, RegVelo acts as a
for it. Bullet one, RegVelo acts as a virtual cell allowing in silico or
virtual cell allowing in silico or
virtual cell allowing in silico or computer simulated gene knockouts.
computer simulated gene knockouts.
computer simulated gene knockouts. Bullet two, it successfully mapped
Bullet two, it successfully mapped
Bullet two, it successfully mapped pancreatic cell drivers like E2F1 and
pancreatic cell drivers like E2F1 and
pancreatic cell drivers like E2F1 and RFY6.
RFY6.
RFY6. >> Correct.
>> Correct.
>> Correct. >> And bullet three, it perfectly simulated
>> And bullet three, it perfectly simulated
>> And bullet three, it perfectly simulated the GATA1 versus SP1 toggle switch in
the GATA1 versus SP1 toggle switch in
the GATA1 versus SP1 toggle switch in human blood cells. Awesome summary. But
human blood cells. Awesome summary. But
human blood cells. Awesome summary. But you know, if you're hearing this and
you know, if you're hearing this and
you know, if you're hearing this and thinking, "Okay, that's really cool. The
thinking, "Okay, that's really cool. The
thinking, "Okay, that's really cool. The computer built a nice virtual cell that
computer built a nice virtual cell that
computer built a nice virtual cell that confirms things we already discovered
confirms things we already discovered
confirms things we already discovered years ago in textbooks."
years ago in textbooks."
years ago in textbooks." >> Right, confirming the knowns.
>> Right, confirming the knowns.
>> Right, confirming the knowns. >> Yeah. But does this algorithm actually
>> Yeah. But does this algorithm actually
>> Yeah. But does this algorithm actually have the power to discover something
have the power to discover something
have the power to discover something completely new? And that is exactly what
completely new? And that is exactly what
completely new? And that is exactly what the researchers wanted to prove next.
the researchers wanted to prove next.
the researchers wanted to prove next. Because validating known biology is one
Because validating known biology is one
Because validating known biology is one thing. Predicting unknown biology is the
thing. Predicting unknown biology is the
thing. Predicting unknown biology is the ultimate test of any computational
ultimate test of any computational
ultimate test of any computational model. So to do this, the researchers
model. So to do this, the researchers
model. So to do this, the researchers chose one of the most complex chaotic
chose one of the most complex chaotic
chose one of the most complex chaotic systems in developmental biology. The
systems in developmental biology. The
systems in developmental biology. The zebrafish cranial neural crest.
zebrafish cranial neural crest.
zebrafish cranial neural crest. >> The neural crest. I mean, this is
>> The neural crest. I mean, this is
>> The neural crest. I mean, this is Waddington's landscape on extreme fast
Waddington's landscape on extreme fast
Waddington's landscape on extreme fast forward, isn't it?
forward, isn't it?
forward, isn't it? >> Oh, it truly is. These are a temporary
>> Oh, it truly is. These are a temporary
>> Oh, it truly is. These are a temporary group of cells in a developing embryo
group of cells in a developing embryo
group of cells in a developing embryo that migrate incredibly fast and
that migrate incredibly fast and
that migrate incredibly fast and differentiate into a wildly diverse
differentiate into a wildly diverse
differentiate into a wildly diverse array of tissues. Like what? They turn
array of tissues. Like what? They turn
array of tissues. Like what? They turn into the pigment cells that give the
into the pigment cells that give the
into the pigment cells that give the fish its stripes, the cartilage in its
fish its stripes, the cartilage in its
fish its stripes, the cartilage in its face, peripheral neurons, glia. It is a
face, peripheral neurons, glia. It is a
face, peripheral neurons, glia. It is a very noisy, highly dynamic system. So to
very noisy, highly dynamic system. So to
very noisy, highly dynamic system. So to give the algorithm the best possible
give the algorithm the best possible
give the algorithm the best possible chance at mapping this chaos, the
chance at mapping this chaos, the
chance at mapping this chaos, the researchers didn't just use standard
researchers didn't just use standard
researchers didn't just use standard sequencing. No, they couldn't.
sequencing. No, they couldn't.
sequencing. No, they couldn't. >> They threw the absolute state of the art
>> They threw the absolute state of the art
>> They threw the absolute state of the art at it, generating a massive multiomic
at it, generating a massive multiomic
at it, generating a massive multiomic data stack. They needed extremely
data stack. They needed extremely
data stack. They needed extremely high-fidelity input. So they used a
high-fidelity input. So they used a
high-fidelity input. So they used a technique called full-length Smart-seq3
technique called full-length Smart-seq3
technique called full-length Smart-seq3 to get incredibly high-resolution RNA
to get incredibly high-resolution RNA
to get incredibly high-resolution RNA data. Smart-seq3? Yes. But more
data. Smart-seq3? Yes. But more
data. Smart-seq3? Yes. But more importantly, they measured that RNA
importantly, they measured that RNA
importantly, they measured that RNA expression while simultaneously
expression while simultaneously
expression while simultaneously measuring the chromatin accessibility.
measuring the chromatin accessibility.
measuring the chromatin accessibility. Meaning how physically open the DNA was.
Meaning how physically open the DNA was.
Meaning how physically open the DNA was. >> Exactly. In the exact same individual
>> Exactly. In the exact same individual
>> Exactly. In the exact same individual cells. So, they were essentially reading
cells. So, they were essentially reading
cells. So, they were essentially reading the text of the instruction manual while
the text of the instruction manual while
the text of the instruction manual while simultaneously checking the spine of the
simultaneously checking the spine of the
simultaneously checking the spine of the book to see which pages naturally fell
book to see which pages naturally fell
book to see which pages naturally fell open. That is a brilliant way to picture
open. That is a brilliant way to picture
open. That is a brilliant way to picture it, and it gave RegVelo an unbelievably
it, and it gave RegVelo an unbelievably
it, and it gave RegVelo an unbelievably rich starting map.
rich starting map.
rich starting map. >> So, they fed all of this high-resolution
>> So, they fed all of this high-resolution
>> So, they fed all of this high-resolution data into the framework. Right. And
data into the framework. Right. And
data into the framework. Right. And RegVelo's algorithms crunched those
RegVelo's algorithms crunched those
RegVelo's algorithms crunched those high-dimensional equations, calculated
high-dimensional equations, calculated
high-dimensional equations, calculated the latent times, self-corrected the
the latent times, self-corrected the
the latent times, self-corrected the network weights. Using sparsity, like we
network weights. Using sparsity, like we
network weights. Using sparsity, like we talked about.
talked about.
talked about. >> Yes. And it output a prediction. It
>> Yes. And it output a prediction. It
>> Yes. And it output a prediction. It pointed to specific very early
pointed to specific very early
pointed to specific very early transcription factors as the master
transcription factors as the master
transcription factors as the master switches for pigment cell development.
switches for pigment cell development.
switches for pigment cell development. >> Specifically, it flagged two genes, Tfec
>> Specifically, it flagged two genes, Tfec
>> Specifically, it flagged two genes, Tfec and Elf1. Now, scientists knew these
and Elf1. Now, scientists knew these
and Elf1. Now, scientists knew these genes were involved in development
genes were involved in development
genes were involved in development somewhere. Right. But their role as the
somewhere. Right. But their role as the
somewhere. Right. But their role as the primary earliest drivers dictating the
primary earliest drivers dictating the
primary earliest drivers dictating the pigment lineage was heavily
pigment lineage was heavily
pigment lineage was heavily underappreciated. RegVelo basically
underappreciated. RegVelo basically
underappreciated. RegVelo basically looked at the math and said, "These two
looked at the math and said, "These two
looked at the math and said, "These two genes are the most critical traffic cops
genes are the most critical traffic cops
genes are the most critical traffic cops at the very top of the hill." And this
at the very top of the hill." And this
at the very top of the hill." And this is honestly my favorite part of the
is honestly my favorite part of the
is honestly my favorite part of the entire paper.
entire paper.
entire paper. >> Oh, the validation. Yes.
>> Oh, the validation. Yes.
>> Oh, the validation. Yes. The researchers didn't just publish
The researchers didn't just publish
The researchers didn't just publish their neat little computer simulation
their neat little computer simulation
their neat little computer simulation and walk away. No, they didn't.
and walk away. No, they didn't.
and walk away. No, they didn't. >> They took RegVelo's mathematical
>> They took RegVelo's mathematical
>> They took RegVelo's mathematical prediction and marched straight into the
prediction and marched straight into the
prediction and marched straight into the physical wet lab. They performed what we
physical wet lab. They performed what we
physical wet lab. They performed what we call in vivo knockout validation. They
call in vivo knockout validation. They
call in vivo knockout validation. They tested it in living animals.
tested it in living animals.
tested it in living animals. >> They took actual living zebrafish
>> They took actual living zebrafish
>> They took actual living zebrafish embryos. Yes. They used CRISPR Cas9
embryos. Yes. They used CRISPR Cas9
embryos. Yes. They used CRISPR Cas9 genetic scissors to physically snip the
genetic scissors to physically snip the
genetic scissors to physically snip the Tfec and Elf1 genes right out of the
Tfec and Elf1 genes right out of the
Tfec and Elf1 genes right out of the DNA. And then they let the embryos grow
DNA. And then they let the embryos grow
DNA. And then they let the embryos grow and used advanced single-cell
and used advanced single-cell
and used advanced single-cell sequencing, specifically Perturb-seq, to
sequencing, specifically Perturb-seq, to
sequencing, specifically Perturb-seq, to measure exactly what happened to those
measure exactly what happened to those
measure exactly what happened to those physical cells. And the results were
physical cells. And the results were
physical cells. And the results were undeniable. The physical biological
undeniable. The physical biological
undeniable. The physical biological knockout perfectly validated the digital
knockout perfectly validated the digital
knockout perfectly validated the digital mathematical simulation. When Tfec and
mathematical simulation. When Tfec and
mathematical simulation. When Tfec and Elf1 were removed from in living fish,
Elf1 were removed from in living fish,
Elf1 were removed from in living fish, the pigment lineage specification
the pigment lineage specification
the pigment lineage specification completely broke down. The cells
completely broke down. The cells
completely broke down. The cells entirely failed to differentiate into
entirely failed to differentiate into
entirely failed to differentiate into pigment cells, matching RegVelo's
pigment cells, matching RegVelo's
pigment cells, matching RegVelo's virtual prediction flawlessly. The
virtual prediction flawlessly. The
virtual prediction flawlessly. The algorithm nailed it. The virtual model
algorithm nailed it. The virtual model
algorithm nailed it. The virtual model was so tightly coupled to reality that
was so tightly coupled to reality that
was so tightly coupled to reality that it predicted a physical outcome in a
it predicted a physical outcome in a
it predicted a physical outcome in a living organism. And the statistics back
living organism. And the statistics back
living organism. And the statistics back up just how significant a leap this is.
up just how significant a leap this is.
up just how significant a leap this is. When the team measured how accurately
When the team measured how accurately
When the team measured how accurately RegVelo predicted the downstream effects
RegVelo predicted the downstream effects
RegVelo predicted the downstream effects of knocking out these individual
of knocking out these individual
of knocking out these individual transcription factors, RegVelo achieved
transcription factors, RegVelo achieved
transcription factors, RegVelo achieved a mean Spearman correlation of 0.52.
a mean Spearman correlation of 0.52.
a mean Spearman correlation of 0.52. Okay, I know a Spearman correlation of
Okay, I know a Spearman correlation of
Okay, I know a Spearman correlation of 0.52 sounds like deep statistical
0.52 sounds like deep statistical
0.52 sounds like deep statistical jargon. It is a bit dense, yes.
jargon. It is a bit dense, yes.
jargon. It is a bit dense, yes. >> But let's put that number in context for
>> But let's put that number in context for
>> But let's put that number in context for the listener. Compared to the older
the listener. Compared to the older
the listener. Compared to the older decoupled RNA velocity methods we talked
decoupled RNA velocity methods we talked
decoupled RNA velocity methods we talked about at the beginning of the show, how
about at the beginning of the show, how
about at the beginning of the show, how much better is that 0.52? It is nearly
much better is that 0.52? It is nearly
much better is that 0.52? It is nearly double the accuracy.
double the accuracy.
double the accuracy. >> Double? Yes.
>> Double? Yes.
>> Double? Yes. By connecting the gene regulatory
By connecting the gene regulatory
By connecting the gene regulatory network to the velocity dynamics, they
network to the velocity dynamics, they
network to the velocity dynamics, they doubled their ability to predict complex
doubled their ability to predict complex
doubled their ability to predict complex biological reality. That is huge. So,
biological reality. That is huge. So,
biological reality. That is huge. So, let's do our final mini recap. Bullet
let's do our final mini recap. Bullet
let's do our final mini recap. Bullet one, RegVelo was tested on the extremely
one, RegVelo was tested on the extremely
one, RegVelo was tested on the extremely complex zebrafish cranial neural crest.
complex zebrafish cranial neural crest.
complex zebrafish cranial neural crest. Yep. Bullet two, it predicted TFEB and
Yep. Bullet two, it predicted TFEB and
Yep. Bullet two, it predicted TFEB and Elf1 as the crucial drivers for pigment
Elf1 as the crucial drivers for pigment
Elf1 as the crucial drivers for pigment cells.
cells.
cells. >> Right. And bullet three, CRISPR-Cas9
>> Right. And bullet three, CRISPR-Cas9
>> Right. And bullet three, CRISPR-Cas9 knockouts in live embryos proved the
knockouts in live embryos proved the
knockouts in live embryos proved the computer model was right. Spot on. So,
computer model was right. Spot on. So,
computer model was right. Spot on. So, the question to ask yourself now is,
the question to ask yourself now is,
the question to ask yourself now is, what does this actually mean for the
what does this actually mean for the
what does this actually mean for the future of how we study diseases and
future of how we study diseases and
future of how we study diseases and developmental biology? Well, it means
developmental biology? Well, it means
developmental biology? Well, it means we're moving from simply drawing lines
we're moving from simply drawing lines
we're moving from simply drawing lines on a graph and observing a car drive by
on a graph and observing a car drive by
on a graph and observing a car drive by to actually understanding how the engine
to actually understanding how the engine
to actually understanding how the engine turns the wheels. It represents a
turns the wheels. It represents a
turns the wheels. It represents a fundamental paradigm shift in
fundamental paradigm shift in
fundamental paradigm shift in computational biology. It does. We are
computational biology. It does. We are
computational biology. It does. We are transitioning from an observational
transitioning from an observational
transitioning from an observational phase into a causal phase. By honoring
phase into a causal phase. By honoring
phase into a causal phase. By honoring the fact that biology is highly coupled
the fact that biology is highly coupled
the fact that biology is highly coupled and interactive, RegVelo provides an
and interactive, RegVelo provides an
and interactive, RegVelo provides an interpretable, actionable framework. We
interpretable, actionable framework. We
interpretable, actionable framework. We can now test hypotheses in a computer
can now test hypotheses in a computer
can now test hypotheses in a computer simulation with a level of mechanical
simulation with a level of mechanical
simulation with a level of mechanical confidence that simply did not exist
confidence that simply did not exist
confidence that simply did not exist before this paper. And that leads to a
before this paper. And that leads to a
before this paper. And that leads to a thought I really want to leave you with
thought I really want to leave you with
thought I really want to leave you with as we wrapped up today's deep dive.
as we wrapped up today's deep dive.
as we wrapped up today's deep dive. Think about the implications of this for
Think about the implications of this for
Think about the implications of this for medicine.
medicine.
medicine. We just spent a lot of time talking
We just spent a lot of time talking
We just spent a lot of time talking about predicting how a healthy stem cell
about predicting how a healthy stem cell
about predicting how a healthy stem cell becomes a blood cell or how it becomes a
becomes a blood cell or how it becomes a
becomes a blood cell or how it becomes a pigment cell in a fish. Mhm. But
pigment cell in a fish. Mhm. But
pigment cell in a fish. Mhm. But Waddington's landscape isn't just about
Waddington's landscape isn't just about
Waddington's landscape isn't just about healthy development. Cancer is
healthy development. Cancer is
healthy development. Cancer is essentially a problem of cell fate and
essentially a problem of cell fate and
essentially a problem of cell fate and gene regulation gone horribly tragically
gene regulation gone horribly tragically
gene regulation gone horribly tragically wrong. Exactly. It is a mutated
wrong. Exactly. It is a mutated
wrong. Exactly. It is a mutated landscape where the marble gets trapped
landscape where the marble gets trapped
landscape where the marble gets trapped in a valley it was never supposed to
in a valley it was never supposed to
in a valley it was never supposed to enter and it just keeps dividing. So if
enter and it just keeps dividing. So if
enter and it just keeps dividing. So if we now have the computational power to
we now have the computational power to
we now have the computational power to accurately simulate a virtual cell and
accurately simulate a virtual cell and
accurately simulate a virtual cell and predict exactly how altering its wiring
predict exactly how altering its wiring
predict exactly how altering its wiring changes its fate.
changes its fate.
changes its fate. >> Where does that leave us? Right. How far
>> Where does that leave us? Right. How far
>> Where does that leave us? Right. How far are we from a world where a doctor
are we from a world where a doctor
are we from a world where a doctor biopsies your specific tumor, they
biopsies your specific tumor, they
biopsies your specific tumor, they sequence the RNA, they measure the DNA
sequence the RNA, they measure the DNA
sequence the RNA, they measure the DNA accessibility and they upload your exact
accessibility and they upload your exact
accessibility and they upload your exact cancer cells into a framework like
cancer cells into a framework like
cancer cells into a framework like RigVelo.
RigVelo.
RigVelo. >> They could create a completely
>> They could create a completely
>> They could create a completely personalized digital twin of your tumor.
personalized digital twin of your tumor.
personalized digital twin of your tumor. Yes. And instead of testing highly toxic
Yes. And instead of testing highly toxic
Yes. And instead of testing highly toxic chemotherapy drugs on your actual body
chemotherapy drugs on your actual body
chemotherapy drugs on your actual body through months of agonizing trial and
through months of agonizing trial and
through months of agonizing trial and error, the doctor performs in silico
error, the doctor performs in silico
error, the doctor performs in silico perturbations. Exactly. They digitally
perturbations. Exactly. They digitally
perturbations. Exactly. They digitally simulate thousands of drug combinations
simulate thousands of drug combinations
simulate thousands of drug combinations in the computer knocking out different
in the computer knocking out different
in the computer knocking out different transcription factors until they find
transcription factors until they find
transcription factors until they find the exact molecular mechanism that
the exact molecular mechanism that
the exact molecular mechanism that forces your specific cancer cells to
forces your specific cancer cells to
forces your specific cancer cells to stop growing. Or forces them to
stop growing. Or forces them to
stop growing. Or forces them to differentiate into harmless tissue.
differentiate into harmless tissue.
differentiate into harmless tissue. >> Yes. They solve the puzzle virtually
>> Yes. They solve the puzzle virtually
>> Yes. They solve the puzzle virtually before you ever have to swallow a single
before you ever have to swallow a single
before you ever have to swallow a single pill. That is the ultimate promise of
pill. That is the ultimate promise of
pill. That is the ultimate promise of this technology.
this technology.
this technology. By finally linking the genetic software
By finally linking the genetic software
By finally linking the genetic software to the cellular dynamics, we aren't just
to the cellular dynamics, we aren't just
to the cellular dynamics, we aren't just observing biology anymore. We are
observing biology anymore. We are
observing biology anymore. We are learning how to actively rewrite it. So
learning how to actively rewrite it. So
learning how to actively rewrite it. So the next time you picture that single
the next time you picture that single
the next time you picture that single marble rolling down the bumpy slopes of
marble rolling down the bumpy slopes of
marble rolling down the bumpy slopes of Waddington's landscape bouncing between
Waddington's landscape bouncing between
Waddington's landscape bouncing between the ridges and valleys, remember that we
the ridges and valleys, remember that we
the ridges and valleys, remember that we are no longer just watching it roll. No,
are no longer just watching it roll. No,
are no longer just watching it roll. No, we're not. We finally have the tools to
we're not. We finally have the tools to
we're not. We finally have the tools to read the map and maybe very soon we'll
read the map and maybe very soon we'll
read the map and maybe very soon we'll be the ones drawing the valleys.
Can't find what you're looking for?
Get subtitles in any language from opensubtitles.com, and translate them here.