The following is a conversation with Manolis Kellis, his third time on the podcast.
He is a professor at MIT and head of the MIT Computational Biology Group.
This time we went deep on the science, biology, and genetics.
So this is a bit of an experiment.
Manolis went back and forth between the basics of biology to the latest state of the art
in the research.
He’s a master at this, so I just sat back and enjoyed the ride.
This conversation happened at 7am, so it’s yet another podcast episode after an all nighter
And once again, since the universe has a sense of humor, this one was a tough one for my
brain to keep up, but I did my best and I never shy away from a good challenge.
Quick mention of each sponsor, followed by some thoughts related to the episode.
First is SEMrush, the most advanced SEO optimization tool I’ve ever come across.
I don’t like looking at numbers, but someone probably should, it helps you make good decisions.
Second is Pessimist Archive, they’re back, one of my favorite history podcasts on why
people resist new things from recorded music to umbrellas to cars, chess, coffee, and the
Third is 8sleep, a mattress that cools itself, measures heart rate variability, has an app,
and has given me yet another reason to look forward to sleep, including the all important
And finally, BetterHelp, online therapy when you want to face your demons with a licensed
professional, not just by doing the David Goggins like physical challenges like I seem
to do on occasion.
Please check out these sponsors in the description to get a discount and to support this podcast.
As a side note, let me say that biology in the brain and in the various systems of the
body fill me with awe every time I think about how such a chaotic mess coming from its humble
origins in the ocean was able to achieve such incredibly complex and robust mechanisms of
life that survived despite all the forces of nature that want to destroy it.
It is so unlike the computing systems we humans have engineered that it makes me feel that
in order to create artificial general intelligence and artificial consciousness, we may have
to completely rethink how we engineer computational systems.
If you enjoy this thing, subscribe on YouTube, review it with 5 stars on Apple Podcast, follow
on Spotify, support on Patreon, or connect with me on Twitter at Lex Friedman.
And now, here’s my conversation with Manolis Callas.
So your group at MIT is trying to understand the molecular basis of human disease.
What are some of the biggest challenges in your view?
Don’t get me started.
I mean, understanding human disease is the most complex challenge in modern science.
So because human disease is as complex as the human genome, it is as complex as the
human brain, and it is in many ways, even more complex because the more we understand
disease complexity, the more we start understanding genome complexity and epigenome complexity
and brain circuitry complexity and immune system complexity and cancer complexity and
so on and so forth.
So traditionally, human disease was following basic biology.
You would basically understand basic biology in model organisms like, you know, mouse and
fly and yeast.
You would understand sort of mammalian biology and animal biology and eukaryotic biology
in sort of progressive layers of complexity, getting closer to human phylogenetically.
And you would do perturbation experiments in those species to see if I knock out a gene,
And based on the knocking out of these genes, you would basically then have a way to drive
human biology because you would sort of understand the functions of these genes.
And then if you find that a human gene locus, something that you’ve mapped from human genetics
to that gene is related to a particular human disease, you’d say, aha, now I know the function
of the gene from the model organisms.
I can now go and understand the function of that gene in human.
But this is all changing.
This is dramatically changed.
So that was the old way of doing basic biology.
You would start with the animal models, the eukaryotic models, the mammalian models, and
then you would go to human.
Human genetics has been so transformed in the last decade or two that human genetics
is now actually driving the basic biology.
There is more genetic mutation information in the human genome than there will ever be
in any other species.
What do you mean by mutation information?
So perturbations is how you understand systems.
So an engineer builds systems and then they know how they work from the inside out.
A scientist studies systems through perturbations.
You basically say, if I poke that balloon, what’s going to happen?
And I’m going to film it in super high resolution, understand, I don’t know, aerodynamics or
fluid dynamics if it’s filled with water, et cetera.
So you can then make experimentation by perturbation and then the scientific process is sort of
building models that best fit the data, designing new experiments that best test your models
and challenge your models and so on and so forth.
This is the same thing with science.
Basically if you’re trying to understand biological science, you basically want to do perturbations
that then drive the models.
So how do these perturbations allow you to understand disease?
So if you know that a gene is related to disease, you don’t want to just know that it’s related
to the disease.
You want to know what is the disease mechanism because you want to go and intervene.
So the way that I like to describe it is that traditionally epidemiology, which is basically
the study of disease, you know, sort of the observational study of disease has been about
correlating one thing with another thing.
So if you have a lot of people with liver disease who are also alcoholics, you might
say, well, maybe the alcoholism is driving the liver disease or maybe those who have
liver disease self medicate with alcohol.
So the connection could be either way.
With genetic epidemiology, it’s about correlating changes in genome with phenotypic differences
and then you know the direction of causality.
So if you know that a particular gene is related to the disease, you can basically say, okay,
perturbing that gene in mouse causes the mice to have X phenotype.
So perturbing that gene in human causes the humans to have the disease.
So I can now figure out what are the detailed molecular phenotypes in the human that are
related to that organismal phenotype in the disease.
So it’s all about understanding disease mechanism, understanding what are the pathways, what
are the tissues, what are the processes that are associated with the disease so that we
know how to intervene.
You can then prescribe particular medications that also alter these processes.
You can prescribe lifestyle changes that also affect these processes and so on and so forth.
That’s such a beautiful puzzle to try to solve.
Like what kind of perturbations eventually have this ripple effect that leads to disease
across the population.
And then you study that for animals or mice first and then see how that might possibly
connect to humans.
How hard is that puzzle of trying to figure out how little perturbations might lead to,
in a stable way, to a disease?
In animals, we make the puzzle simpler because we perturb one gene at a time.
That’s the beauty of this, the power of animal models.
You can basically decouple the perturbations.
You only do one perturbation and you only do strong perturbations at a time.
In human, the puzzle is incredibly complex because obviously you don’t do human experimentation.
You wait for natural selection and natural genetic variation to basically do its own
experiments, which it has been doing for hundreds and thousands of years in the human population
and for hundreds of thousands of years across the history leading to the human population.
So you basically take this natural genetic variation that we all carry within us.
Every one of us carries 6 million perturbations.
So I’ve done 6 million experiments on you, 6 million experiments on me, 6 million experiments
on every one of 7 billion people on the planet.
What’s the 6 million correspond to?
6 million unique genetic variants that are segregating in the human population.
Every one of us carries millions of polymorphic sites, poly, many, morph, forms.
Polymorphic means many forms, variants.
That basically means that every one of us has single nucleotide alterations that we
have inherited from mom and from dad that basically can be thought of as tiny little
Most of them don’t do anything, but some of them lead to all of the phenotypic differences
that we see between us.
The reason why two twins are identical is because these variants completely determine
the way that I’m going to look at exactly 93 years of age.
How happy are you with this kind of data set?
Is it large enough of the human population of Earth?
Is that too big, too small?
Yeah, so is it large enough is a power analysis question.
In every one of our grants, we do a power analysis based on what is the effect size
that I would like to detect and what is the natural variation in the two forms.
Every time you do a perturbation, you’re asking, I’m changing form A into form B. Form A has
some natural phenotypic variation around it and form B has some natural phenotypic variation
If those variances are large and the differences between the mean of A and the mean of B are
small, then you have very little power.
The further the means go apart, that’s the effect size, the more power you have, and
the smaller the standard deviation, the more power you have.
So basically when you’re asking, is that sufficiently large, certainly not for everything, but we
already have enough power for many of the stronger effects in the more tight distributions.
So that’s the hopeful message that there exists parts of the genome that have a strong effect
that has a small variance.
That’s exactly right.
Unfortunately, those perturbations are the basis of disease in many cases.
So it’s not a hopeful message.
Sometimes it’s a terrible message.
It’s basically, well, some people are sick, but if we can figure out what are these contributors
to sickness, we can then help make them better and help many other people better who don’t
carry that exact mutation, but who carry mutations on the same pathways.
And that’s what we like to call the allelic series of a gene.
You basically have many perturbations of the same gene in different people, each with a
different frequency in the human population and each with a different effect on the individual
that carries them.
So you said in the past there would be these small experiments on perturbations and animal
What does this puzzle solving process look like today?
So we basically have something like 7 billion people in the planet and every one of them
carries something like 6 million mutations.
You basically have an enormous matrix of genotype by phenotype by systematically measuring the
phenotype of these individuals.
And the traditional way of measuring this phenotype has been to look at one trait at
You would gather families and you would sort of paint the pedigrees of a strong effect,
what we like to call Mendelian mutation, so a mutation that gets transmitted in a dominant
or a recessive, but strong effect form where basically one locus plays a very big role
in that disease.
And you could then look at carriers versus non carriers in one family, carriers versus
non carriers in another family and do that for hundreds, sometimes thousands of families
and then trace these inheritance patterns and then figure out what is the gene that
plays that role.
Is this the matrix that you’re showing in talks or lectures?
So that matrix is the input to those stuff that I show in talks.
So basically that matrix has traditionally been strong effect genes.
What the matrix looks like now is instead of pedigrees, instead of families, you basically
have thousands and sometimes hundreds of thousands of unrelated individuals, each with all of
their genetic variants and each with their phenotype, for example, height or lipids or,
you know, whether they’re sick or not for a particular trait.
That has been the modern view instead of going to families, going to unrelated individuals
with one phenotype at a time.
And what we’re doing now as we’re maturing in all of these sciences is that we’re doing
this in the context of large medical systems or enormous cohorts that are very well phenotyped
across hundreds of phenotypes, sometimes with our complete electronic health record.
So you can now start relating not just one gene segregating one family, not just thousands
of variants segregating with one phenotype, but now you can do millions of variants versus
hundreds of phenotypes.
And as a computer scientist, I mean, deconvolving that matrix, partitioning it into the layers
of biology that are associated with every one of these elements is a dream come true.
It’s like the world’s greatest puzzle.
And you can now solve that puzzle by throwing in more and more knowledge about the function
of different genomic regions and how these functions are changed across tissues and in
the context of disease.
And that’s what my group and many other groups are doing.
We’re trying to systematically relate this genetic variation with molecular variation
at the expression level of the genes, at the epigenomic level of the gene regulatory circuitry,
and at the cellular level of what are the functions that are happening in those cells,
at the single cell level using single cell profiling, and then relate all that vast amount
of knowledge computationally with the thousands of traits that each of these of thousands
of variants are perturbing.
I mean, this is something we talked about, I think last time.
So there’s these effects at different levels that happen.
You said at a single cell level, you’re trying to see things that happen due to certain perturbations.
And then it’s not just like a puzzle of perturbation and disease.
It’s perturbation then effect at a cellular level, then at an organ level, a body, like,
how do you disassemble this into like what your group is working on?
You’re basically taking a bunch of the hard problems in the space.
How do you break apart a difficult disease and break it apart into problems that you,
into puzzles that you can now start solving?
So there’s a struggle here.
Super scientists love hard puzzles and they’re like, oh, I want to build a method that just
deconvolves the whole thing computationally.
And that’s very tempting and it’s very appealing, but biologists just like to decouple that
complexity experimentally, to just like peel off layers of complexity experimentally.
And that’s what many of these modern tools that my group and others have both developed
The fact that we can now figure out tricks for peeling off these layers of complexity
by testing one cell type at a time or by testing one cell at a time.
And you could basically say, what is the effect of these genetic variants associated with
Alzheimer’s on human brain?
Human brain sounds like, oh, it’s an organ, of course, just go one organ at a time.
But human brain has of course, dozens of different brain regions and within each of these brain
regions, dozens of different cell types and every single type of neuron, every single
type of glial cell between astrocytes, oligodendrocytes, microglia, between all of the neural cells
and the vascular cells and the immune cells that are co inhabiting the brain between the
different types of excitatory and inhibitory neurons that are sort of interacting with
each other between different layers of neurons in the cortical layers.
Every single one of these has a different type of function to play in cognition, in
interaction with the environment, in maintenance of the brain, in energetic needs, in feeding
the brain with blood, with oxygen, in clearing out the debris that are resulting from the
super high energy production of cognition in humans.
So all of these things are basically potentially deconvolvable computationally, but experimentally,
you can just do single cell profiling of dozens of regions of the brain across hundreds of
individuals across millions of cells.
And then now you have pieces of the puzzle that you can then put back together to understand
I mean, first of all, the cells in the human brain are the most, maybe I’m romanticizing
it, but cognition seems to be very complicated.
So separating into the function, breaking Alzheimer’s down to the cellular level seems
Is that basically you’re trying to find a way that some perturbation in the genome results
in some obvious major dysfunction in the cell.
You’re trying to find something like that.
So what does human genetics do?
Human genetics basically looks at the whole path from genetic variation all the way to
So human genetics has basically taken thousands of Alzheimer’s cases and thousands of controls
matched for age, for sex, for environmental backgrounds and so on and so forth.
And then looked at that map where you’re asking, what are the individual genetic perturbations
and how are they related to all the way to Alzheimer’s disease?
And that has actually been quite successful.
So we now have more than 27 different loci, these are genomic regions that are associated
with Alzheimer’s at these end to end level.
But the moment you sort of break up that very long path into smaller levels, you can basically
say from genetics, what are the epigenomic alterations at the level of gene regulatory
elements where that genetic variant perturbs the control region nearby.
That effect is much larger.
You mean much larger in terms of this down the line impact or?
It’s much larger in terms of the measurable effect, this A versus B variance is actually
so much cleanly defined when you go to the shorter branches.
Because for one genetic variant to affect Alzheimer’s, that’s a very long path.
That basically means that in the context of millions of these 6 million variants that
every one of us carries, that one single nucleotide has a detectable effect all the way to the
I mean, it’s just mind boggling that that’s even possible, but indeed there are such effects.
So the hope is, or the most scientifically speaking, the most effective place where to
detect the alteration that results in disease is earlier on in the pipeline, as early as
It’s a trade off.
If you go very early on in the pipeline, now each of these epigenomic alterations, for
example, this enhancer control region is active maybe 50% less, which is a dramatic effect.
Now you can ask, well, how much does changing one regulatory region in the genome in one
cell type change disease?
Well, that path is now long.
So if you instead look at expression, the path between genetic variation and the expression
of one gene goes through many enhancer regions, and therefore it’s a subtler effect at the
But then now you’re closer because one gene is acting in the context of only 20,000 other
genes as opposed to one enhancer acting in the context of 2 million other enhancers.
So you basically now have genetic, epigenomic, the circuitry, transcriptomic, the gene expression
control, and then cellular, where you can basically say, I can measure various properties
of those cells.
What is the calcium influx rate when I have this genetic variation?
What is the synaptic density?
What is the electric impulse conductivity and so on and so forth?
So you can measure things along this path to disease, and you can also measure endophenotypes.
You can basically measure your brain activity.
You can do imaging in the brain.
You can basically measure, I don’t know, the heart rate, the pulse, the lipids, the amount
of blood secreted and so on and so forth.
And then through all of that, you can basically get at the path to causality, the path to
And is there something beyond cellular?
So you mentioned lifestyle interventions or changes as a way to, or like be able to prescribe
changes in lifestyle.
Like what about organs?
What about like the function of the body as a whole?
So basically when you go to your doctor, they always measure, you know, your pulse.
They always measure your height.
They always measure your weight, you know, your BMI.
So basically these are just very basic variables.
But with digital devices nowadays, you can start measuring hundreds of variables for
You can basically also phenotype cognitively through tests, Alzheimer’s patients.
There are cognitive tests that you can measure, that you typically do for cognitive decline,
these mini mental observations that you have specific questions to.
You can think of sort of enlarging the set of cognitive tests.
So in the mouse, for example, you do experiments for how do they get out of mazes?
How do they find food?
Whether they recall a fear, whether they shake in a new environment and so on and so forth.
In the human, you can have much, much richer phenotypes where you can basically say not
just imaging at the organ level and all kinds of other activities at the organ level, but
you can also do at the organism level, you can do behavioral tests.
And how did they do on empathy?
How did they do on memory?
How did they do on longterm memory versus short term memory?
And so on and so forth.
I love how you’re calling that phenotype.
I guess it is.
But like your behavior patterns that might change over a period of a life, your ability
to remember things, your ability to be empathetic or emotionally, your intelligence perhaps
Yeah, but intelligence has hundreds of variables.
You can be your math intelligence, your literary intelligence, your puzzle solving intelligence,
It could be like hundreds of things.
And all of that, we’re able to measure that better and better and all that could be connected
to the entire pipeline somehow.
We used to think of each of these as a single variable like intelligence.
I mean, that’s ridiculous.
It’s basically dozens of different genes that are controlling every single variable.
You can basically think of, imagine us in a video game where every one of us has measures
of strength, stamina, energy left and so on and so forth.
But you could click on each of those five bars that are just the main bars and each
of those will just give you then hundreds of bars and can basically say, okay, great
for my machine learning task, I want someone who, a human who has these particular forms
I require now these 20 different things.
And then you can combine those things and then relate them to of course performance
in a particular task, but you can also relate them to genetic variation that might be affecting
different parts of the brain.
For example, your frontal cortex versus your temporal cortex versus your visual cortex
and so on and so forth.
So genetic variation that affects expression of genes in different parts of your brain
can basically affect your music ability, your auditory ability, your smell, just dozens
of different phenotypes can be broken down into hundreds of cognitive variables and then
relate each of those to thousands of genes that are associated with them.
So somebody who loves RPGs or playing games, there’s too few variables that we can control.
So I’m excited if we’re in fact living in a simulation and this is a video game, I’m
excited by the quality of the video game.
The game designer did a hell of a good job.
So we’re impressed.
Oh, I don’t know.
The sunset last night was a little unrealistic.
Come on, NVIDIA.
To zoom back out, we’ve been talking about the genetic origins of diseases, but I think
it’s fascinating to talk about what are the most important diseases to understand and
especially as it connects to the things that you’re working on.
So it’s very difficult to think about important diseases to understand.
There’s many metrics of importance.
One is lifestyle impact.
I mean, if you look at COVID, the impact on lifestyle has been enormous.
So understanding COVID is important because it has impacted the wellbeing in terms of
ability to have a job, ability to have an apartment, ability to go to work, ability
to have a mental circle of support and all of that for millions of Americans, like huge,
So that’s one aspect of importance.
So basically mental disorders, Alzheimer’s has a huge importance in the wellbeing of
Whether or not it kills someone for many, many years, it has a huge impact.
So the first measure of importance is just wellbeing.
Impact on the quality of life.
Impact on the quality of life, absolutely.
The second metric, which is much easier to quantify is deaths.
What is the number one killer?
The number one killer is actually heart disease.
It is actually killing 650,000 Americans per year.
Number two is cancer with 600,000 Americans.
Number three, far, far down the list is accidents, every single accident combined.
So basically you read the news, accidents, like there was a huge car crash all over the
But the number of deaths, number three by far, 167,000.
Core respiratory disease.
So that’s asthma, not being able to breathe and so on and so forth, 160,000 Alzheimer’s
number five with 120,000 and then stroke, brain aneurysms and so on and so forth, that’s
147,000 diabetes and metabolic disorders, et cetera.
The flu is 60,000, suicide, 50,000 and then overdose, et cetera, you know, goes further
down the list.
So of course COVID has creeped up to be the number three killer this year with, you know,
more than 100,000 Americans and counting.
And you know, but if you think about sort of what do we use, what are the most important
diseases, you have to understand both the quality of life and the sheer number of deaths
and just numbers of years lost if you wish.
And each of these diseases you can think of as, and also including terrorist attacks and
school shootings, for example, things which lead to fatalities, you can look at as problems
that could be solved.
And some problems are harder to solve than others.
I mean, that’s part of the equation.
So maybe if you look at these diseases, if you look at heart disease or cancer or Alzheimer’s
or just like schizophrenia and obesity, Debbie, like not necessarily things that kill you,
but affect the quality of life, which problems are solvable, which aren’t, which are harder
to solve, which aren’t.
I love your question because he puts it in the context of a global effort rather than
just the local effort.
So basically if you look at the global aspect, exercise and nutrition are two interventions
that we can as a society make a much better job at.
So if you think about sort of the availability of cheap food, it’s extremely high in calories.
It’s extremely detrimental for you, like a lot of processed food, et cetera.
So if we change that equation and as a society, we made availability of healthy food much,
much easier and charged a burger at McDonald’s, the price that it costs on the health system,
then people would actually start buying more healthy foods.
So basically that’s sort of a societal intervention, if you wish.
In the same way, increasing empathy, increasing education, increasing the social framework
and support would basically lead to fewer suicides.
It would lead to fewer murders.
It would lead to fewer deaths overall.
So that’s something that we as a society can do.
You can also think about external factors versus internal factors.
So the external factors are basically communicable diseases like COVID, like the flu, et cetera.
And the internal factors are basically things like cancer and Alzheimer’s where basically
your genetics will eventually drive you there.
And then of course, with all of these factors, every single disease has both the genetic
component and environmental component.
So heart disease, huge genetic contribution, Alzheimer’s, it’s like 60% plus genetic.
So I think it’s like 79% heritability.
So that basically means that genetics alone explains 79% of Alzheimer’s incidents.
And yes, there’s a 21% environmental component where you could basically enrich your cognitive
environment, enrich your social interactions, read more books, learn a foreign language,
go running, you know, sort of have a more fulfilling life.
All of that will actually decrease Alzheimer’s, but there’s a limit to how much that can impact
because of the huge genetic footprint.
So this is fascinating.
So each one of these problems have a genetic component and an environment component.
And so like when there’s a genetic component, what can we do about some of these diseases?
And have you worked on what can you say that’s in terms of problems that are solvable here
So my group works on the genetic component, but I would argue that understanding the genetic
component can have a huge impact even on the environmental component.
Why is that?
Because genetics gives us access to mechanism.
And if we can alter the mechanism, if we can impact the mechanism, we can perhaps counteract
some of the environmental components.
So understanding the biological mechanisms leading to disease is extremely important
in being able to intervene.
But when you can intervene and what, you know, the analogy that I like to give is for example,
for obesity, you know, think of it as a giant bathtub of fat.
There’s basically fat coming in from your diet and there’s fat coming out from your
So that’s an in out equation and that’s the equation that everybody’s focusing on.
But your metabolism impacts that, you know, bathtub.
Basically your metabolism controls the rate at which you’re burning energy.
It controls the rate at which you’re storing energy.
And it also teaches you about the various valves that control the input and the output
So if we can learn from the genetics, the valves, we can then manipulate those valves.
And even if the environment is feeding you a lot of fat and getting a little that out,
you can just poke another hole at the bathtub and just get a lot of the fat out.
Yeah, that’s fascinating.
So we’re not just passive observers of our genetics.
The more we understand, the more we can come up with actual treatments.
And I think that’s an important aspect to realize when people are thinking about strong
effect versus weak effect variants.
So some variants have strong effects.
We talked about these Mendelian disorders where a single gene has a sufficiently large
effect, penetrance, expressivity, and so on and so forth, that basically you can trace
it in families with cases and not cases, cases, not cases, and so on and so forth.
But so these are the genes that everybody says, oh, that’s the genes we should go after
because that’s a strong effect gene.
I like to think about it slightly differently.
These are the genes where genetic impacts that have a strong effect were tolerated because
every single time we have a genetic association with disease, it depends on two things.
Number one, the obvious one, whether the gene has an impact on the disease.
Number two, the more subtle one is whether there is genetic variation standing and circulating
and segregating in the human population that impacts that gene.
Some genes are so darn important that if you mess with them, even a tiny little amount,
that person’s dead.
So those genes don’t have variation.
You’re not going to find a genetic association if you don’t have variation.
That doesn’t mean that the gene has no role.
It simply means that the gene tolerates no mutations.
So that’s actually a strong signal when there’s no variation.
That’s so fascinating.
Genes that have very little variation are hugely important.
You can actually rank the importance of genes based on how little variation they have.
And those genes that have very little variation but no association with disease, that’s a
very good metric to say, oh, that’s probably a developmental gene because we’re not good
at measuring those phenotypes.
So it’s genes that you can tell evolution has excluded mutations from, but yet we can’t
see them associated with anything that we can measure nowadays.
It’s probably early embryonic lethal.
What are all the words you just said?
Early embryonic what?
Meaning that that embryo will die.
There’s a bunch of stuff that is required for a stable functional organism across the
board for an entire species, I guess.
If you look at sperm, it expresses thousands of proteins.
Does sperm actually need thousands of proteins?
No, but it’s probably just testing them.
So my speculation is that misfolding of these proteins is an early test for failure.
So that out of the millions of sperm that are possible, you select the subset that are
just not grossly misfolding thousands of proteins.
So it’s kind of an assert that this is folded correctly.
This just because if this little thing about the folding of a protein isn’t correct, that
probably means somewhere down the line, there’s a bigger issue.
That’s exactly right.
So fail fast.
So basically if you look at the mammalian investment in a newborn, that investment is
enormous in terms of resources.
So mammals have basically evolved mechanisms for fail fast.
Where basically in those early months of development, I mean it’s horrendous of course at the personal
level when you lose your future child, but in some ways there’s so little hope for that
child to develop and sort of make it through the remaining months that sort of fail fast
is probably a good evolutionary principle for mammals.
And of course humans have a lot of medical resources that you can sort of give those
children a chance and we have so much more success in sort of giving folks who have these
strong carrier mutations a chance, but if they’re not even making it through the first
three months, we’re not going to see them.
So that’s why when we say what are the most important genes to focus on, the ones that
have a strong effect mutation or the ones that have a weak effect mutation, well the
jury might be out because the ones that have a strong effect mutation are basically not
mattering as much.
The ones that only have weak effect mutations by understanding through genetics that they
have a weak effect mutation and understanding that they have a causal role on the disease,
we can then say, okay, great, evolution has only tolerated a 2% change in that gene.
Pharmaceutically I can go in and induce a 70% change in that gene and maybe I will poke
another hole at the bathtub that was not easy to control in many of the other sort of strong
effect genetic variants.
So there’s this beautiful map of across the population of things that you’re saying strong
and weak effects, so stuff with a lot of mutations and stuff with little mutations with no mutations
and you have this map and it lays out the puzzle.
So when I say strong effect, I mean at the level of individual mutations.
So basically genes where, so you have to think of first the effect of the gene on the disease.
Remember how I was sort of painting that map earlier from genetics all the way to phenotype.
That gene can have a strong effect on the disease, but the genetic variant might have
a weak effect on the gene.
So basically when you ask what is the effect of that genetic variant on the disease, it
could be that that genetic variant impacts the gene by a lot and then the gene impacts
the disease by a little, or it could be that the genetic variants impacts the gene by a
little and then the gene impacts the disease by a lot.
So what we care about is genes that impact the disease a lot, but genetics gives us the
full equation and what I would argue is if we couple the genetics with expression variation
to basically ask what genes change by a lot and which genes correlate with disease by
a lot, even if the genetic variants change them by a little, then those are the best
places to intervene.
Those are the best places where pharmaceutical, if I have even a modest effect, I will have
a strong effect on the disease, whereas those genetic variants that have a huge effect on
the disease, I might not be able to change that gene by this much without affecting all
kinds of other things.
So that’s what we’re looking at.
What have we been able to find in terms of which disease could be helped?
Again, don’t get me started.
We have found so much.
Our understanding of disease has changed so dramatically with genetics.
I mean places that we had no idea would be involved.
So one of the worst things about my genome is that I have a genetic predisposition to
age related macular degeneration, AMD.
So it’s a form of blindness that causes you to lose the central part of your vision progressively
as you grow older.
My increased risk is fairly small.
I have an 8% chance.
You only have a 6% chance.
I’m an average.
By the way, when you say my, you mean literally yours.
You know this about you.
I know this about me.
Which is kind of, I mean philosophically speaking is a pretty powerful thing to live with.
Maybe that’s, so we agreed to talk again by the way for the listeners to where we’re going
to try to focus on science today and a little bit of philosophy next time.
But it’s interesting to think about the more you’re able to know about yourself from the
genetic information in terms of the diseases, how that changes your own view of life.
So there’s a lot of impact there and there’s something called genetics exceptionalism,
which basically thinks of genetics as something very, very different than everything else
as a type of determinism.
And you know, let’s talk about that next time.
That’s a good preview.
So let’s go back to AMD.
So basically with AMD, we have no idea what causes AMD.
You know, it was, it was a mystery until the genetics were worked out.
And now the fact that I know that I have a predisposition allows me to sort of make some
life choices, number one, but number two, the genes that lead to that predisposition
give us insights as to how does it actually work.
And that’s a place where genetics gave us something totally unexpected.
So there’s a complement pathway, which is an immune function pathway that was in, you
know, most of the loci associated with AMD.
And that basically told us that, wow, there’s an immune basis to this eye disorder that
people had just not expected before.
If you look at complement, it was recently also implicated in schizophrenia.
And there’s a type of microglia that is involved in synaptic pruning.
So synapses are the connections between neurons.
And in this whole use it or lose it view of mental cognition and other capabilities, you
basically have microglia, which are immune cells that are sort of constantly traversing
your brain and then pruning neuronal connections, pruning synaptic connections that are not
So in schizophrenia, there’s thought to be a change in the pruning that basically if
you don’t prune your synapses the right way, you will actually have an increased role of
This is something that was completely unexpected for schizophrenia.
Of course, we knew it has to do with neurons, but the role of the complement complex, which
is also implicated in AMD, which is now also implicated in schizophrenia, was a huge surprise.
What’s the complement complex?
So it’s basically a set of genes, the complement genes that are basically having various immune
And as I was saying earlier, our immune system has been coopted for many different roles
across the body.
So they actually play many diverse roles.
And somehow the immune system is connected to the synaptic pruning process, the process.
So the prune cells were coopted to prune synapses.
How did you figure this out?
How does one go about figuring this intricate connection, like pipeline of connections out?
Let me give you another example.
So Alzheimer’s disease, the first place that you would expect it to act is obviously the
So we had basically this roadmap epigenomics consortium view of the human epigenome, the
largest map of the human epigenome that has ever been built across 127 different tissues
and samples with dozens of epigenomic marks measured in hundreds of donors.
So what we’ve basically learned through that is that you basically can map what are the
active gene regulatory elements for every one of the tissues in the body.
And then we connected these gene regulatory active maps of basically what regions of the
human genome are turning on in every one of different tissues.
We then can go back and say, where are all of the genetic loci that are associated with
This is something that my group, I think was the first to do back in 2010 in this Ernst
Nature Biotech paper, but basically we were for the first time able to show that specific
chromatin states, specific epigenomic states, in that case enhancers, were in fact enriched
in disease associated variants.
We pushed that further in the Ernst Nature paper a year later.
And then in this roadmap epigenomics paper a few years after that, but basically that
matrix that you mentioned earlier was in fact the first time that we could see what genetic
traits have genetic variants that are enriched in what tissues in the body.
And a lot of that map made complete sense.
If you looked at a diversity of immune traits like allergies and type one diabetes and so
on and so forth, you basically could see that they were enriching, that the genetic variants
associated with those traits were enriched in enhancers in these gene regulatory elements
active in T cells and B cells and hematopoietic stem cells and so on and so forth.
So that basically gave us a confirmation in many ways that those immune traits were indeed
enriching immune cells.
If you looked at type two diabetes, you basically saw an enrichment in only one type of sample
and it was pancreatic islets.
And we know that type two diabetes sort of stems from the dysregulation of insulin in
the beta cells of pancreatic islets.
And that sort of was spot on, super precise.
If you looked at blood pressure, where would you expect blood pressure to occur?
You know, I don’t know, maybe in your metabolism and ways that you process coffee or something
Maybe in your brain, the way that you stress out and increases your blood pressure, et
So the blood pressure localized specifically in the left ventricle of the heart.
So the enhancers of the left ventricle in the heart contained a lot of genetic variants
associated with blood pressure.
If you look at height, we found an enrichment specifically in embryonic stem cell enhancers.
So the genetic variants predisposing you to be taller or shorter are in fact acting in
developmental stem cells, makes complete sense.
If you looked at inflammatory bowel disease, you basically found inflammatory, which is
immune, and also bowel disease, which is digestive.
And indeed we saw a double enrichment both in the immune cells and in the digestive cells.
So that basically told us that this is acting in both components.
There’s an immune component to inflammatory bowel disease and there’s a digestive component.
And the big surprise was for Alzheimer’s.
We had seven different brain samples.
We found zero enrichment in the brain samples for genetic variants associated with Alzheimer’s.
And this is mind boggling.
Our brains were literally hurting.
What is going on?
And what is going on is that the brain samples are primarily neurons, oligodendrocytes, and
astrocytes in terms of the cell types that make them up.
So that basically indicated that genetic variants associated with Alzheimer’s were probably
not acting in oligodendrocytes, astrocytes, or neurons.
So what could they be acting in?
Well, the fourth major cell type is actually microglia.
Microglia are resident immune cells in your brain.
They are CD14 plus, which is this sort of cell surface markers of those cells.
So they’re CD14 plus cells, just like macrophages that are circulating in your blood.
The microglia are resident monocytes that are basically sitting in your brain.
They’re tissue specific monocytes.
And every one of your tissues, like your fat, for example, has a lot of macrophages that
And the M1 versus M2 macrophage ratio has a huge role to play in obesity.
And so basically, again, these immune cells are everywhere, but basically what we found
through this completely unbiased view of what are the tissues that likely underlie different
disorders, we found that Alzheimer’s was humongously enriched in microglia, but not at all in the
other cell types.
So what are we supposed to make that if you look at the tissues involved, is that simply
useful for indication of propensity for disease, or does it give us somehow a pathway of treatment?
It’s very much the second.
If you look at the way to therapeutics, you have to start somewhere.
What are you going to do?
You’re going to basically make assays that manipulate those genes and those pathways
in those cell types.
So before we know the tissue of action, we don’t even know where to start.
We basically are at a loss.
But if you know the tissue of action, and even better, if you know the pathway of action,
then you can basically screen your small molecules, not for the gene, you can screen them directly
for the pathway in that cell type.
So you can basically develop a high throughput multiplexed robotic system for testing the
impact of your favorite molecules that you know are safe, efficacious, and sort of hit
that particular gene and so on and so forth.
You can basically screen those molecules against either a set of genes that act in that pathway
or on the pathway directly by having a cellular assay.
And then you can basically go into mice and do experiments and basically sort of figure
out ways to manipulate these processes that allow you to then go back to humans and do
a clinical trial that basically says, okay, I was able indeed to reverse these processes
Can I do the same thing in humans?
So the knowledge of the tissues gives you the pathway to treatment, but that’s not the
There are many additional steps to figuring out the mechanism of disease.
So that’s really promising.
Maybe to take a small step back, you’ve mentioned all these puzzles that were figured out with
the Nature paper for, I mean, you’ve mentioned a ton of diseases from obesity to Alzheimer’s,
even schizophrenia, I think you mentioned.
What is the actual methodology of figuring this out?
So indeed, I mentioned a lot of diseases and my lab works on a lot of different disorders.
And the reason for that is that if you look at biology, it used to be zoology departments
and botanology departments and virology departments and so on and so forth.
And MIT was one of the first schools to basically create a biology department, like, oh, we’re
going to study all of life suddenly.
Why was that even a case?
Because the advent of DNA and the genome and the central dogma of DNA makes RNA makes protein
in many ways, unified biology.
You could suddenly study the process of transcription in viruses or in bacteria and have a huge
impact on yeast and fly and maybe even mammals because of this realization of these common
And in the same way that DNA unified biology, genetics is unifying disease studies.
So you used to have, I don’t know, cardiovascular disease department and neurological disease
department and neurodegeneration department and basically immune and cancer and so on
and so forth.
And all of these were studied in different labs because it made sense, because basically
the first step was understanding how the tissue functions and we kind of knew the tissues
involved in cardiovascular disease and so on and so forth.
But what’s happening with human genetics is that all of these walls and edifices that
we had built are crumbling.
And the reason for that is that genetics is in many ways revealing unexpected connections.
So suddenly we now have to bring the immunologists to work on Alzheimer’s.
They were never in the room.
They were in another building altogether.
The same way for schizophrenia, we now have to sort of worry about all these interconnected
For metabolic disorders, we’re finding contributions from brain.
So suddenly we have to call the neurologist from the other building and so on and so forth.
So in my view, it makes no sense anymore to basically say, oh, I’m a geneticist studying
I mean, that’s ridiculous because, I mean, of course in many ways you still need to sort
But what we’re doing is that we’re basically saying we’ll go wherever the genetics takes
And by building these massive resources, by working on our latest map is now 833 tissues,
sort of the next generation of the epigenomics roadmap, which we’re now called epimap, is
833 different tissues.
And using those, we’ve basically found enrichments in 540 different disorders.
Those enrichments are not like, oh great, you guys work on that and we’ll work on this.
They’re intertwined amazingly.
So of course there’s a lot of modularity, but there’s these enhancers that are sort
of broadly active and these disorders that are broadly active.
So basically some enhancers are active in all tissues and some disorders are enriching
in all tissues.
So basically there’s these multifactorial and this other class, which I like to call
polyfactorial diseases, which are basically lighting up everywhere.
And in many ways it’s, you know, sort of cutting across these walls that were previously built
across these departments.
And the polyfactorial ones were probably the previous structural departments wasn’t equipped
to deal with those.
I mean, again, maybe it’s a romanticized question, but you know, there’s in physics, there’s
a theory of everything.
Do you think it’s possible to move towards an almost theory of everything of disease
from a genetic perspective?
So if this unification continues, is it possible that, like, do you think in those terms, like
trying to arrive at a fundamental understanding of how disease emerges, period?
That unification is not just foreseeable, it’s inevitable.
I see it as inevitable.
We have to go there.
You cannot be a specialist anymore.
If you’re a genomicist, you have to be a specialist in every single disorder.
And the reason for that is that the fundamental understanding of the circuitry of the human
genome that you need to solve schizophrenia, that fundamental circuitry is hugely important
to solve Alzheimer’s.
And that same circuitry is hugely important to solve metabolic disorders.
And that same exact circuitry is hugely important for solving immune disorders and cancer and,
you know, every single disease.
So all of them have the same sub task.
And I teach dynamic programming in my class.
Dynamic programming is all about sort of not redoing the work.
It’s reusing the work that you do once.
So basically for us to say, oh, great, you know, you guys in the immune building go solve
the fundamental circuitry of everything.
And then you guys in the schizophrenia building go solve the fundamental circuitry of everything
separately, is crazy.
So what we need to do is come together and sort of have a circuitry group, the circuitry
building that sort of tries to solve the circuitry of everything.
And then the immune folks who will apply this knowledge to all of the disorders that are
associated with immune dysfunction and the schizophrenia folks will basically interacting
with both the immune folks and with the neuronal folks.
And all of them will be interacting with the circuitry folks and so on and so forth.
So that’s sort of the current structure of my group, if you wish.
So basically what we’re doing is focusing on the fundamental circuitry.
But at the same time, we’re the users of our own tools by collaborating with many other
labs in every one of these disorders that we mentioned.
We basically have a heart focus on cardiovascular disease, coronary artery disease, heart failure
and so on and so forth.
We have an immune focus on several immune disorders.
We have a cancer focus on metastatic melanoma and immunotherapy response.
We have a psychiatric disease focus on schizophrenia, autism, PTSD, and other psychiatric disorders.
We have an Alzheimer’s and neurodegeneration focus on Huntington’s disease, ALS and, you
know, AD related disorders like frontotemporal dementia and Lewy body dementia.
And of course, a huge focus on Alzheimer’s.
We have a metabolic focus on the role of exercise and diets and sort of how they’re impacting
metabolic organs across the body and across many different tissues.
And all of them are interfacing with the circuitry.
And the reason for that is another computer science principle of eat your own dog food.
If everybody ate their own dog food, dog food would taste a lot better.
The reason why Microsoft Excel and Word and PowerPoint was so important and so successful
is because the employees that were working on them, were using them for their day to
You can’t just simply build a circuitry and say, here it is guys, take the circuitry,
we’re done without being the users of that circuitry because you then go back.
And because we span the whole spectrum from profiling the epigenome, using comparative
genomics, finding the important nucleotides in the genome, building the basic functional
map of what are the genes in the human genome, what are the gene regulatory elements of the
I mean, over the years we’ve written a series of papers on how do you find human genes in
the first place using comparative genomics?
How do you find the motifs that are the building blocks of gene regulation using comparative
And how do you then find how these motifs come together and act in specific tissues
How do you link regulators to enhancers and enhancers to their target genes using epigenomics
and regulatory genomics?
So through the years we’ve basically built all this infrastructure for understanding
what I like to say, every single nucleotide of the human genome and how it acts in every
one of the major cell types and tissues of the human body.
I mean, this is no small task.
This is an enormous task that takes the entire field.
And that’s something that my group has taken on along with many other groups.
And we have also, and that sort of a thing sets my group perhaps apart, we have also
worked with specialists in every one of these disorders to basically further our understanding
all the way down to disease and in some cases collaborating with pharma to go all the way
down to therapeutics because of our deep, deep understanding of that basic circuitry
and how it allows us to now improve the circuitry.
Not just treat it as a black box, but basically go and say, okay, we need a better cell type
specific wiring that we now have at the tissue specific level.
So we’re focusing on that because we’re understanding the needs from the disease front.
So you have a sense of the entire pipeline, I mean, one, maybe you can indulge me.
One nice question to ask would be, how do you, from the scientific perspective, go from
knowing nothing about the disease to going, you said, to go into the entire pipeline and
actually have a drug or a treatment that cures that disease?
So that’s an enormously long path and an enormously great challenge.
And what I’m trying to argue is that it progresses in stages of understanding rather than one
gene at a time.
The traditional view of biology was you have one postdoc working on this gene and another
postdoc working on that gene, and they’ll just figure out everything about that gene
and that’s their job.
But we’ve realized how polygenic the diseases are, so we can’t have one postdoc per gene
We now have to have these cross cutting needs.
And I’m going to describe the path to circuitry along those needs.
And every single one of these paths, we are now doing in parallel across thousands of
So the first step is you have a genetic association, and we talked a little bit about sort of the
Mendelian path and the polygenic path to that association.
So the Mendelian path was looking through families to basically find gene regions and
ultimately genes that are underlying particular disorders.
The polygenic path is basically looking at unrelated individuals in this giant matrix
of genotype by phenotype, and then finding hits where a particular variant impacts disease
all the way to the end.
And then we now have a connection, not between a gene and a disease, but between a genetic
region and a disease.
And that distinction is not understood by most people.
So I’m going to explain it a little bit more.
Why do we not have a connection between a gene and a disease, but we have a connection
between a genetic region and a disease?
The reason for that is that 93% of genetic variants that are associated with disease
don’t impact the protein at all.
So if you look at the human genome, there’s 20,000 genes, there’s 3.2 billion nucleotides.
Only 1.5% of the genome codes for proteins.
The other 98.5% does not code for proteins.
If you now look at where are the disease variants located, 93% of them fall in that outside
the genes portion.
Of course, genes are enriched, but they’re only enriched by a factor of three.
That means that still 93% of genetic variants fall outside the proteins.
Why is that difficult?
Why is that a problem?
The problem is that when a variant falls outside the gene, you don’t know what gene is impacted
by that variant.
You can’t just say, oh, it’s near this gene, let’s just connect that variant to the gene.
And the reason for that is that the genome circuitry is very often long range.
So you basically have that genetic variant that could sit in the intron of one gene.
An intron is sort of the place between the exons that code for proteins.
So proteins are split up into exons and introns and every exon codes for a particular subset
of amino acids and together they’re spliced together and then make the final protein.
So that genetic variant might be sitting in an intron of a gene.
It’s transcribed with the gene, it’s processed and then excised, but it might not impact
this gene at all.
It might actually impact another gene that’s a million nucleotides away.
So it’s just riding along even though it has nothing to do with this nearby neighborhood.
That’s exactly right.
Let me give you an example.
The strongest genetic association with obesity was discovered in this FTO gene, fat and obesity
So this FTO gene was studied ad nauseum.
People did tons of experiments on it.
They figured out that FTO is in fact RNA methylation transferase.
It basically impacts something that we call the epitranscriptome.
Just like the genome can be modified, the transcriptome, the transcript of the genes
can be modified.
And we basically said, oh great, that means that epitranscriptomics is hugely involved
in obesity because that gene FTO is clearly where the genetic locus is at.
My group studied FTO in collaboration with a wonderful team led by Melina Klausnitzer.
And what we found is that this FTO locus, even though it is as associated with obesity,
does not implicate the FTO gene.
The genetic variance, it’s in the first intron of the FTO gene, but it controls two genes
IRX3 and IRX5 that are sitting 1.2 million nucleotides away, several genes away.
What am I supposed to feel about that because isn’t that like super complicated then?
So the way that I was introduced at a conference a few years ago was, and here’s Manolis Kellis
who wrote the most depressing paper of 2015.
And the reason for that is that the entire pharmaceutical industry was so comfortable
that there was a single gene in that locus.
Because in some loci, you basically have three dozen genes that are all sitting in the same
region of association and you’re like, oh gosh, which ones of those is it?
But even that question of which ones of those is it is making the assumption that it is
one of those as opposed to some random gene just far, far away, which is what our paper
So basically what our paper showed is that you can’t ignore the circuitry.
You have to first figure out the circuitry, all of those long range interactions, how
every genetic variant impacts the expression of every gene in every tissue imaginable across
hundreds of individuals.
And then you now have one of the building blocks, not even all of the building blocks
for then going and understanding disease.
So embrace the wholeness of the circuitry.
So back to the question of starting knowing nothing to the disease and going to the treatment.
So what are the next steps?
So you basically have to first figure out the tissue and then describe how you figure
out the tissue.
You figure out the tissue by taking all of these non coding variants that are sitting
outside proteins and then figuring out what are the epigenomic enrichments.
And the reason for that, you know, thankfully is that there is convergence, that the same
processes are impacted in different ways by different loci.
And that’s a saving grace for our field.
The fact that if I look at hundreds of genetic variants associated with Alzheimer’s, they
localize in a small number of processes.
Can you clarify why that’s hopeful?
So like they show up in the same exact way in the, in the specific set of processes.
So basically there’s a small number of biological processes that underlie, or at least that
play the biggest role in every disorder.
So in Alzheimer’s you basically have, you know, maybe 10 different types of processes.
One of them is lipid metabolism.
One of them is immune cell function.
One of them is neuronal energetics.
So these are just a small number of processes, but you have multiple lesions, multiple genetic
perturbations that are associated with those processes.
So if you look at schizophrenia, it’s excitatory neuron function, it’s inhibitory neuron function,
it’s synaptic pruning, it’s calcium signaling and so on and so forth.
So when you look at disease genetics, you have one hit here and one hit there and one
hit there and one hit there, completely different parts of the genome.
But it turns out all of those hits are calcium signaling proteins.
You’re like, aha.
That means that calcium signaling is important.
So those people who are focusing on one doctor at a time cannot possibly see that picture.
You have to become a genomicist.
You have to look at the omics, the om, the holistic picture to understand these enrichments.
But you mentioned the convergence thing.
The whatever the thing associated with the disease shows up.
So let me explain convergence.
Convergence is such a beautiful concept.
So you basically have these four genes that are converging on calcium signaling.
So that basically means that they are acting each in their own way, but together in the
But now in every one of these loci, you have many enhancers controlling each of those genes.
That’s another type of convergence where dysregulation of seven different enhancers might all converge
on dysregulation of that one gene, which then converges on calcium signaling.
And in each one of those enhancers, you might have multiple genetic variants distributed
across many different people.
Everyone has their own different mutation.
But all of these mutations are impacting that enhancer.
And all of these enhancers are impacting that gene.
And all of these genes are impacting this pathway.
And all these pathways are acting in the same tissue.
And all of these tissues are converging together on the same biological process of schizophrenia.
And you’re saying the saving grace is that that conversion seems to happen for a lot
of these diseases.
For all of them.
Basically that for every single disease that we’ve looked at, we have found an epigenomic
How do you do that?
You basically have all of the genetic variants associated with the disorder.
And then you’re asking for all of the enhancers active in a particular tissue.
For 540 disorders, we’ve basically found that indeed there is an enrichment.
That basically means that there is commonality.
And from the commonality, we can just get insights.
So to explain in mathematical terms, we’re basically building an empirical prior.
We’re using a Bayesian approach to basically say, great, all of these variants are equally
likely in a particular locus to be important.
So in a genetic locus, you basically have a dozen variants that are coinherited.
Because the way that inheritance works in the human genome is through all of these recombination
events during meiosis, you basically have, you know, you inherit maybe three, chromosome
three, for example, in your body is inherited from four different parts.
One part comes from your dad, another part comes from your mom, another part comes from
your dad, another part comes from your mom.
So basically, the way that it, sorry, from your mom’s mom.
So you basically have one copy that comes from your dad and one copy that comes from
But that copy that you got from your mom is a mixture of her maternal and her paternal
And the copy that you got from your dad is a mixture of his maternal and his paternal
So these breakpoints that happen when chromosomes are lining up are basically ensuring through
these crossover events, they’re ensuring that every child cell during the process of meiosis,
where you basically have, you know, one spermatozoid that basically couples with one ovule to basically
create one egg to basically create the zygote.
You basically have half of your genome that comes from dad and half your genome that comes
But in order to line them up, you basically have these crossover events.
These crossover events are basically leading to coinheritance of that entire block coming
from your maternal grandmother and that entire block coming from your maternal grandfather.
Over many generations, these crossover events don’t happen randomly.
There’s a protein called PRDM9 that basically guides the double stranded breaks and then
leads to these crossovers.
And that protein has a particular preference to only a small number of hotspots of recombination,
which then lead to a small number of breaks between these coinheritance patterns.
So even though there are 6 million variants, there are 6 million loci, this variation is
inherited in blocks and every one of these blocks has like two dozen genetic variants
that are all associated.
So in the case of FTO, it wasn’t just one variant, it was 89 common variants that were
all humongously associated with obesity.
Which one of those is the important one?
Well, if you look at only one locus, you have no idea.
But if you look at many loci, you basically say, aha, all of them are enriching in the
same epigenomic map.
In that particular case, it was mesenchymal stem cells.
So these are the progenitor cells that give rise to your brown fat and your white fat.
Progenitor is like the early on developmental stem cells?
So you start from one zygote and that’s a totipotent cell type.
It can do anything.
You then, you know, that cell divides, divides, divides, and then every cell division is leading
to specialization where you now have a mesodermal lineage and ectodermal lineage and endodermal
lineage that basically leads to different parts of your body.
The ectoderm will basically give rise to your skin, ecto means outside, derm is skin.
So ectoderm, but it also gives rise to your neurons and your whole brain.
So that’s a lot of ectoderm.
Mesoderm gives rise to your internal organs, including the vasculature and you know, your
muscle and stuff like that.
So you basically have this progressive differentiation and then if you look further, further down
that lineage, you basically have one lineage that will give rise to both your muscle and
your bone, but also your fat.
And if you go further down the lineage of your fat, you basically have your white fat
These are the cells that store energy.
So when you eat a lot, but you don’t exercise too much, there’s an excess set of calories,
What do you do with those?
You basically create, you spend a lot of that energy to create these high energy molecules,
lipids, which you can then burn when you need them on a rainy day.
So that leads to obesity if you don’t exercise and if you overeat because your body’s like,
oh great, I have all these calories.
I’m going to store them.
Ooh, more calories.
I’m going to store them too.
Ooh, more calories.
So basically the 42% of European chromosomes have a predisposition to storing fat, which
was selected probably in the food scarcity periods, like basically as we were exiting
Africa before and during the ice ages, there was probably a selection to those individuals
who made it North to basically be able to store energy, a lot more energy.
So you basically now have this lineage that is deciding whether you want to store energy
in your white fat or burn energy in your beige fat.
It turns out that your fat is, you know, like we have such a bad view of fat.
Fat is your best friend.
Fat can both store all these excess lipids that would be otherwise circulating through
your body and causing damage, but it can also burn calories directly.
If you have too much energy, you can just choose to just burn some of that as heat.
So basically when you’re cold, you’re burning energy to basically warm your body up and
you’re burning all these lipids and you’re burning all these calories.
So what we basically found is that across the board, genetic variants associated with
obesity across many of these regions were all enriched repeatedly in mesenchymal stem
So that gave us a hint as to which of these genetic variants was likely driving this whole
And we ended up with this one genetic variant called RS1421085.
And that genetic variant out of the 89 was the one that we predicted to be causal for
So going back to those steps, first step is figure out the relevant tissue based on the
Second step is figure out the causal variant among many variants in this linkage disequilibrium
in this coinherited block between these recombination hotspots, these boundaries of these inherited
That’s the second step.
The third step is once you know that causal variant, try to figure out what is the motif
that is disrupted by that causal variant.
Basically how does it act?
Variants don’t just disrupt elements, they disrupt the binding of specific regulators.
So basically the third step there was how do you find the motif that is responsible
like the gene regulatory word, the building block of gene regulation that is responsible
for that dysregulatory event.
And the fourth step is finding out what regulator normally binds that motif and is now no longer
able to bind.
And then once you have the regulator, can you then try to figure out how to, what after
it developed, how to fix it?
That’s exactly right.
You now know how to intervene.
You have basically a regulator, you have a gene that you can then perturb and you say,
well, maybe that regulator has a global role in obesity.
I can perturb the regulator.
Just to clarify, when we say perturb, like on the scale of a human life, can a human
being be helped?
I guess understanding is the first step.
No, no, but perturbed basically means you now develop therapeutics, pharmaceutical therapeutics
Or you develop other types of intervention that affect the expression of that gene.
What do pharmaceutical therapeutics look like when your understanding is on a genetic level?
Sorry if it’s a dumb question.
No, no, no.
It’s a brilliant question, but I want to save it for a little bit later when we start talking
So let’s talk about the first four steps.
There’s two more.
So basically the first step is figure out, I mean, the zero step, the starting point
is the genetics.
The first step after that is figure out the tissue of action.
The second step is figuring out the nucleotide that is responsible or set of nucleotides.
The third step is figuring out the motif and the upstream regulator, number four.
Number five and six is what are the targets?
So number five is great.
Now I know the regulator.
I know the motif.
I know the tissue and I know the variant.
What does it actually do?
So you have to now trace it to the biological process and the genes that mediate that biological
So knowing all of this can now allow you to find the target genes.
By basically doing perturbation experiments or by looking at the folding of the epigenome
or by looking at the genetic impact of that genetic variant on the expression of genes.
And we use all three.
So let me go through them.
Basically one of them is physical links.
This is the folding of the genome onto itself.
How do you even figure out the folding?
It’s a little bit of a tangent, but it’s a super awesome technology.
Think of the genome as again, this massive packaging that we talked about of taking two
meters worth of DNA and putting it in something that’s a million times smaller than two meters
worth of DNA.
That’s a single cell.
You basically have this massive packaging and this packaging basically leads to the
chromosome being wrapped around in sort of tight, tight ways in ways, however, that are
functionally capable of being reopened and reclosed.
So I can then go in and figure out that folding by sort of chopping up the spaghetti soup,
putting glue and ligating the segments that were chopped up but nearby each other, and
then sequencing through these ligation events to figure out that this region of this chromosome,
that region of the chromosome were near each other.
That means they were interacting even though they were far away on the genome itself.
So that chopping up, sequencing and reglueing is basically giving you folds of the genome
that we call.
Sorry, can you backtrack?
How does cutting it help you figure out which ones were close in the original folding?
So you have a bowl of noodles.
And in that bowl of noodles, some noodles are near each other.
So you throw in a bunch of glue, you basically freeze the noodles in place, throw in a cutter
that chops up the noodles into little pieces.
Now throw in some ligation enzyme that lets those pieces that were free religate near
In some cases, they religate what you had just cut, but that’s very rare.
Most of the time they will religate in whatever was proximal.
You now have glued the red noodle that was crossing the blue noodle to each other.
You then reverse the glue, the glue goes away and you just sequence the heck out of it.
Most of the time you’ll find red segment with, you know, red segment, but you can specifically
select for ligation events that have happened that were not from the same segment by sort
of marking them in a particular way and then selecting those and then you sequence and
you look for red with blue matches of sort of things that were glued that were not immediate
proximal to each other.
And that reveals the linking of the blue noodle and the red noodle.
You’re with me so far?
So we’ve done these experiments.
That’s the physical.
That’s step one of the physical.
And what the physical revealed is topologically associated domains, basically big blocks of
the genome that are topologically connected together.
That’s the physical.
The second one is the genetic links.
It basically says across individuals that have different genetic variants, how are their
genes expressed differently?
Remember before I was saying that the path between genetics and disease is enormous,
but we can break it up to look at the path between genetics and gene expression.
So instead of using Alzheimer’s as a phenotype, I can now use expression of IRX3 as the phenotype,
expression of gene A. And I can look at all of the humans who contain a G at that location
and all the humans that contain a T at that location and basically say, wow, it turns
out that the expression of each gene is higher for the T humans than for the G humans at
So that basically gives me a genetic link between a genetic variant, a locus, a region,
and the expression of nearby genes.
Good on the genetic link?
I think so.
The third genetic link is the activity link.
What’s an activity link?
It basically says if I look across 833 different epigenomes, whenever this enhancer is active,
this gene is active.
That gives me an activity link between this region of the DNA and that gene.
And then the fourth one is perturbations where I can go in and blow up that region and see
what are the genes that change in expression, or I can go in and over activate that region
and see what genes change in expression.
So I guess that’s similar to activity?
So that’s basically similar to activity.
I agree, but it’s causal rather than correlational.
Again, I’m a little weird.
No, no, you’re 100% on.
It’s exactly the same as the perturbation where I go in and intervene.
I basically take a bunch of cells.
So you know CRISPR, right?
CRISPR is this genome guidance and cutting mechanism.
That’s what George Church likes to call genome vandalism.
So you basically are able to, you can basically take a guide RNA that you put into the CRISPR
system, and the CRISPR system will basically use this guide RNA, scan the genome, find
wherever there’s a match, and then cut the genome.
So I digress, but it’s a bacterial immune defense system.
So basically bacteria are constantly attacked by viruses, but sometimes they win against
the viruses and they chop up these viruses.
And remember as a trophy inside their genome, they have these loci, these CRISPR loci that
basically stands for clustered repeats, interspersed, et cetera.
So basically it’s an interspersed repeats structure where basically you have a set of
repetitive regions and then interspersed where these variable segments that were basically
So when this was first discovered, it was basically hypothesized that this is probably
a bacterial immune system that remembers the trophies of the viruses that managed to kill.
And then the bacteria pass on, you know, they sort of do lateral transfer of DNA and they
pass on these memories so that the next bacterium says, Ooh, you killed that guy.
When that guy shows up again, I will recognize him.
And the CRISPR system was basically evolved as a bacterial adaptive immune response to
sense foreigners that should not belong and to just go and cut their genome.
So it’s an RNA guided RNA cutting enzyme or an RNA guided DNA cutting enzyme.
So there’s different systems.
Some of them cut DNA, some of them cut RNA, but all of them remember this sort of viral
So what we have done now as a field is, you know, through the work of, you know, Jennifer
Donne, Manuel Carpentier, Feng Zhang and many others is coopted that system of bacterial
immune defense as a way to cut genomes.
You basically have this guiding system that allows you to use an RNA guide to bring enzymes
to cut DNA at a particular locus.
That’s so fascinating.
So this is like already a natural mechanism, a natural tool for cutting those useful as
And we’re like, well, we can use that thing to actually, it’s a nice tool that’s already
in the body.
It’s not in our body.
It’s in the bacterial body.
It was discovered by the yogurt industry.
They were trying to make better yogurts and they were trying to make their bacteria in
their yogurt cultures more resilient to viruses.
And they were studying bacteria and they found that, wow, this CRISPR system is awesome.
It allows you to defend against that.
And then it was coopted in mammalian systems that don’t use anything like that as a targeting
way to basically bring these DNA cutting enzymes to any locus in the genome.
Why would you want to cut DNA to do anything?
The reason is that our DNA has a DNA repair mechanism where if a region of the genome
gets randomly cut, you will basically scan the genome for anything that matches and sort
of use it by homology.
So the reason why we’re deployed is because we now have a spare copy.
As soon as my mom’s copy is deactivated, I can use my dad’s copy.
And somewhere else, if my dad’s copy is deactivated, I can use my mom’s copy to repair it.
So this is called homologous based repair.
So all you have to do is the cutting and you don’t have to do the fixing.
That’s exactly right.
You don’t have to do the fixing.
Because it’s already built in.
That’s exactly right.
But the fixing can be coopted by throwing in a bunch of homologous segments that instead
of having your dad’s version, have whatever other version you’d like to use.
So you then control the fixing by throwing in a bunch of other stuff.
That’s exactly right.
And that’s how you do genome editing.
So that’s what CRISPR is.
That’s what CRISPR is.
In popular culture, people use the term.
I’ve never, wow, that’s brilliant.
So CRISPR is genome vandalism followed by a bunch of band aids that have the sequence
that you’d like.
And you could control the choices of band aids.
And of course there’s new generations of CRISPR.
There’s something that’s called prime editing that was sort of very, very much in the press
recently that basically instead of sort of making a double stranded break, which again
is genome vandalism, you basically make a single stranded break.
You basically just nick one of the two strands, enabling you to sort of peel off without sort
of completely breaking it up and then repair it locally using a guide that is coupled to
your initial RNA that took you to that location.
Dumb question, but is CRISPR as awesome and cool as it sounds?
I mean, technically speaking, in terms of like as a tool for manipulating our genetics
in the positive meaning of the word manipulating, or is there downsides, drawbacks in this whole
context of therapeutics that we’re talking about or understanding and so on?
So when I teach my students about CRISPR, I show them articles with the headline, genome
editing tool revolutionizes biology.
And then I show them the date of these articles and they’re 2004, like five years before CRISPR
And the reason is that they’re not talking about CRISPR.
They’re talking about zinc finger enzymes that are another way to bring these cutters
to the genome.
It’s a very difficult way of sort of designing the right set of zinc finger proteins, the
right set of amino acids that will now target a particular long stretch of DNA because for
every location that you want to target, you need to design a particular regulator, a particular
protein that will match that region well.
There’s another technology called talons, which are basically just a different way of
using proteins to sort of guide these cutters to a particular location of the genome.
These require a massive team of engineers, of biological engineers to basically design
a set of amino acids that will target a particular sequence of your genome.
The reason why CRISPR is amazingly, awesomely revolutionary is because instead of having
this team of engineers design a new set of proteins for every locus that you want to
target, you just type it in your computer and you just synthesize an RNA guide.
The beauty of CRISPR is not the cutting, it’s not the fixing.
All of that was there before.
It’s the guiding, and the only thing that changes is that it makes the guiding easier
by sort of just typing in the RNA sequence, which then allows the system to sort of scan
the DNA to find that.
So the coding, the engineering of the cutter is easier in terms of SP.
That’s kind of similar to the story of deep learning versus old school machine learning.
Some of the challenging parts are automated.
But CRISPR is just one cutting technology, and then that’s part of the challenges and
exciting opportunities of the field is to design different cutting technologies.
So now this was a big parenthesis on CRISPR, but now when we were talking about perturbations,
you basically now have the ability to not just look at correlation between enhancers
and genes, but actually go and either destroy that enhancer and see if the gene changes
in expression, or you can use the CRISPR targeting system to bring in not vandalism and cutting,
but you can couple the CRISPR system with, and the CRISPR system is called usually CRISPR
Cas9 because Cas9 is the protein that will then come and cut.
But there’s a version of that protein called dead Cas9 where the cutting part is deactivated.
So you basically use the dead Cas9 to bring in an activator or to bring in a repressor.
So you can now ask, is this enhancer changing that gene by taking this modified CRISPR,
which is already modified from the bacteria to be used in humans, that you can now modify
the Cas9 to be dead Cas9, and you can now further modify to bring in a regulator, and
you can basically turn on or turn off that enhancer and then see what is the impact on
So these are the four ways of linking the locus to the target gene, and that’s step
Step number five is find the target gene, and step number six is what the heck does
that gene do?
You basically now go and manipulate that gene to basically see what are the processes that
change, and you can basically ask, well, in this particular case, in the FTO locus, we
found mesenchymal stem cells that are the progenitors of white fat and brown fat or
We found the RS1421085 nucleotide variant as the causal variant.
We found this large enhancer, this master regulator.
I like to call it OB1 for obesity one, like the strongest enhancer associated with it,
and OB1 was kind of chubby as the actor.
I don’t know if you remember him.
So you basically are using this Jedi mind trick to basically find out the location of
the genome that is responsible, the enhancer that harbors it, the motif, the upstream regulator,
which is ARID5B for AT rich interacting domain 5B.
That’s a protein that sort of comes and binds normally.
That protein is normally a repressor.
It represses this super enhancer, this massive 12,000 nucleotide master regulatory control
gene, and it turns off IRX3, which is a gene that’s 600,000 nucleotides away, and IRX5,
which is 1.2 million nucleotides away.
So those things.
And what’s the effect of turning them off?
That’s exactly the next question.
So step six is what do these genes actually do?
So we then ask, what does RX3 and RX5 do?
The first thing we did is look across individuals for individuals that had higher expression
of RX3 or lower expression RX3.
And then we looked at the expression of all of the other genes in the genome.
And we looked for simply correlation.
And we found that RX3 and RX5 were both correlated positively with lipid metabolism and negatively
with mitochondrial biogenesis.
You’re like, what the heck does that mean?
Does this sound related to obesity?
Not at all superficially, but lipid metabolism should, because lipids is these high and
energy molecules that basically store fat.
So RX3 and RX5 are negatively correlated with lipid metabolism.
So that basically means that when they turn on, positively, when they turn on, they turn
on lipid metabolism.
And they’re negatively correlated with mitochondrial biogenesis.
What do mitochondria do in this whole process?
Again, small parenthesis, what are mitochondria?
Mitochondria are little organelles.
They arose, they only are found in eukaryotes.
U means good, karyote means nucleus.
So truly like a true nucleus.
So eukaryotes have a nucleus.
Prokaryotes are before the nucleus.
They don’t have a nucleus.
So eukaryotes have a nucleus, compartmentalization.
Eukaryotes have also organelles.
Some eukaryotes have chloroplasts.
These are the plants, they photosynthesize.
Some other eukaryotes like us have another type of organelle called mitochondria.
These arose from an ancient species that we engulfed.
This is an endosymbiosis event.
Symbiosis bio means life, sim means together.
So symbiotes are things that live together.
Symbiosis endo means inside, so endosymbiosis means you live together holding the other
one inside you.
So the pre eukaryotes engulfed an organism that was very good at energy production and
that organism eventually shed most of its genome to now have only 13 genes in the mitochondrial
genome and those 13 genes are all involved in energy production, the electron transport
So basically electrons are these massive super energy rich molecules.
We basically have these organelles that produce energy and when your muscle exercises, you
basically multiply your mitochondria.
You basically sort of, you know, use more and more mitochondria and that’s how you get
So basically the muscle sort of learns how to generate more energy.
So basically every single time your muscles will, you know, overnight regenerate and sort
of become stronger and amplify their mitochondria and so forth.
So what does mitochondria do?
The mitochondria use energy to sort of do any kind of task.
When you’re thinking, you’re using energy.
This energy comes from mitochondria.
Your neurons have mitochondria all over the place.
Basically this mitochondria can multiply as organelles and they can be spread along the
body of your muscle.
Some of your muscle cells have actually multiple nuclei, they’re polynucleated, but they also
have multiple mitochondria to basically deal with the fact that your muscle is enormous.
You can sort of span these super, super long length and you need energy throughout the
length of your muscle.
So that’s why you have mitochondria throughout the length and you also need transcription
through the length so you have multiple nuclei as well.
So these two processes, lipids store energy, what do mitochondria do?
So there’s a process known as thermogenesis.
Thermal heat, genesis generation.
Thermogenesis is the generation of heat.
Remember that bathtub with the in and out?
That’s the equation that everybody’s focused on.
So how much energy do you consume?
How much energy do you burn?
But in every thermodynamic system, there’s three parts to the equation.
There’s energy in, energy out, and energy lost.
Any machine has loss of energy.
How do you lose energy?
You emanate heat.
So heat is energy loss.
Which is where the thermogenesis comes in.
Thermogenesis is actually a regulatory process that modulates the third component of the
You can basically control thermogenesis explicitly.
You can turn on and turn off thermogenesis.
And that’s where the mitochondria comes into play.
So Irix3 and RX5 turn out to be the master regulators of a process of thermogenesis versus
lipogenesis generation of fat.
So Irix3 and RX5 in most people burn heat, burn calories as heat.
So when you eat too much, just burn it off in your fat cells.
So that bathtub has basically a sort of dissipation knob that most people are able to turn on.
I am unable to turn that on because I am a homozygous carrier for the mutation that changes
a T into a C in the RS1421085 allele and locus, a SNP.
I have the risk allele twice from my mom and from my dad.
So I’m unable to thermogenize.
I’m unable to turn on thermogenesis through Irix3 and RX5 because the regulator that normally
binds here, Irix5b, can no longer bind because it’s an AT rich interacting domain.
And as soon as I change the T into a C, it can no longer bind because it’s no longer
But doesn’t that mean that you’re able to use the energy more efficiently?
You’re not generating heat or is that?
That means I can eat less and get around just fine.
So that’s a feature actually.
It’s a feature in a food scarce environment.
But if we’re all starving, I’m doing great.
If we all have access to massive amounts of food, I’m obese basically.
That’s taken us to the entire process of then understanding that why mitochondria and then
the lipids are both, even though distant, are somehow involved.
Different sides of the same coin.
And you basically choose to store energy or you can choose to burn energy.
And then all of that is involved in the puzzle of obesity.
And that’s what’s fascinating, right?
Here we are in 2007, discovering the strongest genetic association with obesity and knowing
nothing about how it works for almost 10 years.
For 10 years, everybody focused on this FTO gene and they were like, oh, it must have
to do something with RNA modification.
And it’s like, no, it has nothing to do with the function of FTO.
It has everything to do with all of these other processes.
And suddenly the moment you solve that puzzle, which is a multiyear effort by the way, a
tremendous effort by Melina and many, many others.
So this tremendous effort basically led us to recognize this circuitry.
You went from having some 89 common variants associated in that region of the DNA sitting
on top of this gene to knowing the whole circuitry.
When you know the circuitry, you can now go crazy.
You can now start intervening at every level.
You can start intervening at the arid 5B level.
You can start intervening with CRISPR Cas9 at the single SNP level.
You can start intervening at iRx3 and iRx5 directly there.
You can start intervening at the thermogenesis level because you know the pathway.
You can start intervening at the differentiation level where the decision to make either white
fat or beige fat, the energy burning beige fat is made developmentally in the first three
days of differentiation of your adipocytes.
So as they’re differentiating, you basically can choose to make fat burning machines or
fat storing machines.
And sort of that’s how you populate your fat.
You basically can now go in pharmaceutical and do all of that.
And in our paper, we actually did all of that.
We went in and manipulated every single aspect.
At the nucleotide level, we use CRISPR Cas9 genome editing to basically take primary adipocytes
from risk and non risk individuals and show that by editing that one nucleotide out of
3.2 billion nucleotides in the human genome, you could then flip between an obese phenotype
and a lean phenotype like a switch.
You can basically take my cells that are non thermogenizing and just flip into thermogenizing
cells by changing one nucleotide.
It’s mind boggling.
It’s so inspiring that this puzzle could be solved in this way and it feels within reach
to then be able to crack the problem of some of these diseases.
What are the technologies, the tools that came along that made this possible?
What are you excited about?
Maybe if we just look at the buffet of things that you’ve kind of mentioned, what’s involved?
What should we be excited about?
What are you excited about?
I love that question because there’s so much ahead of us.
There’s so, so much.
So basically solving that one locus required massive amounts of knowledge that we have
been building across the years through the epigenome, through the comparative genomics
to find out the causal variant and the controller regulatory motif through the conserved circuitry.
It required knowing these regulatory genomic wiring.
It required high C of these sort of topologically associated domains to basically find these
long range interaction.
It required EQTLs of these sort of genetic perturbation of these intermediate gene phenotypes.
It required all of the arsenal of tools that I’ve been describing was put together for
And this was a massive team effort, huge investment in time, energy, money, effort, intellectual,
You’re referring to, I’m sorry, just for the obesity one.
Yeah, this one paper.
This one single paper.
This one single locus.
I would like to say that this is a paper about one nucleotide in the human genome, about
one bit of information, C versus T in the human genome.
That’s one bit of information and we have 3.2 billion nucleotides to go through.
So how do you do that systematically?
I am so excited about the next phase of research because the technologies that my group and
many other groups have developed allows us to now do this systematically, not just one
locus at a time, but thousands of loci at a time.
So let me describe some of these technologies.
The first one is automation and robotics.
So basically, you know, we talked about how you can take all of these molecules and see
which of these molecules are targeting each of these genes and what do they do?
So you can basically now screen through millions of molecules through thousands and thousands
and thousands of plates, each of which has thousands and thousands and thousands of molecules,
every single time testing, you know, all of these genes and asking which of these molecules
perturb these genes.
So that’s technology number one, automation and robotics.
Technology number two is parallel readouts.
So instead of perturbing one locus and then asking if I use CRISPR Cas9 on this enhancer
to basically use dCas9 to turn on or turn off the enhancer, or if I use CRISPR Cas9
on the SNP to basically change that one SNP at a time, then what happens?
But we have 120,000 disease associated SNPs that we want to test.
We don’t want to spend 120,000 years doing it.
So what do we do?
We’ve basically developed this technology for massively parallel reporter assays, MPRA.
So in collaboration with Tarsha Mikkelsen, Eric Lander, I mean, Jason Durie’s group has
done a lot of that.
So there’s a lot of groups that basically have developed technologies for testing 10,000
genetic variants at a time.
How do you do that?
You know, we talked about microarray technology, the ability to synthesize these huge microarrays
that allow you to do all kinds of things like measure gene expression by hybridization,
by measuring the genotype of a person, by looking at hybridization with one version
with a T versus the other version with a C, and then sort of figuring out that I am a
risk carrier for obesity based on these differential hybridization in my genome that says, oh,
you seem to only have this allele or you seem to have that allele.
These can also be used to systematically synthesize small fragments of DNA.
So you can basically synthesize these 150 nucleotide long fragments across 450,000 spots
at a time.
You can now take the result of that synthesis, which basically works through all of these
sort of layers of adding one nucleotide at a time.
You can basically just type it into your computer and order it, and you can basically order
10,000 or 100,000 of these small DNA segments at a time.
And that’s where awesome molecular biology comes in.
You can basically take all these segments, have a common start and end barcode or sort
of like Gator, just like pieces of a puzzle.
You can make the same end piece and the same start piece for all of them.
And you can now use plasmids, which are these extra chromosomal small DNA circular segments
that are basically inhabiting all our, all our genomes.
We basically have, you know, plasmids from floating around and bacteria use plasmids
for transferring DNA.
And that’s where they put a lot of antibiotic resistance genes.
So they can easily transfer them from one bacterium to the other.
After one bacterium evolves a gene to be resistant to a particular antibiotic, it basically says
to all its friends, Hey, here’s that sort of DNA piece.
We can now coopt these plasmids into human cells.
You can basically make a human cell culture and add plasmids to that human cell culture
that contain the things that you want to test.
You now have this library of 450,000 elements.
You can insert them each into the common plasmid and then test them in millions of cells in
And the common plasmid is all the same before you add it.
The rest of the plasmid is the same.
So it’s, it’s called an epizomal reporter assay.
Epizome means not inside the genome.
It’s sort of outside the chromosomes.
So it’s an epizomal assay that allows you to have a variable region where you basically
test 10,000 different enhancers and you have a common region which basically has the same
You now can do some very cool molecular biology.
You can basically take the 450,000 elements that you’ve generated and you have a piece
of the puzzle here, piece of the puzzle here, which is identical.
So they’re compatible with that plasmid.
You can chop them up in the middle to separate a barcode reporter from the enhancer and in
the middle put the same gene again using the same piece of the puzzle.
You now can have a barcode readout of what is the impact of 10,000 different versions
of an enhancer on gene expression.
So we’re not doing one experiment, we’re doing 10,000 experiments.
And those 10,000 can be 5,000 of different loci and each of them in two versions, risk
or non risk.
I can now test tens of thousands.
Just a little hypothesis.
And then you can do 10,000 and we can test 10,000 hypothesis at once.
How hard is it to generate those 10,000?
But it’s biology.
Generating the 10,000 is trivial because you basically add, it’s biotechnology.
You basically have these arrays that add one nucleotide at a time at every spot.
So it’s printing and so you’re able to, you’re able to control.
Is it super costly?
So this isn’t millions.
10,000 bucks for 10,000 experiments sounds like the right, you know.
I mean, so that’s super, that’s exciting because you don’t have to do one thing at a time.
You can now use that technology, these massively parallel reporter assays to test 10,000 locations
at a time.
We’ve made multiple modifications to that technology.
One was sharper MPRA, which stands for, you know, basically getting a higher resolution
view by tiling these, these elements so you can see where along the region of control
are they acting.
And we made another modification called Hydra for high, you know, definition regulatory
annotation or something like that, which basically allows you to test 7 million of these at a
time by sort of cutting them directly from the DNA.
So instead of synthesizing, which basically has the limit of 450,000 that you can synthesize
at a time, we basically said, Hey, if we want to test all accessible regions of the genome,
let’s just do an experiment that cuts accessible regions.
Let’s take those accessible regions, put them all with the same end joints of the puzzles,
and then now use those to create a much, much larger array of things that you can test.
And then tiling all of these regions, you can then pinpoint what are the driver nucleotides,
what are the elements, how are they acting across 7 million experiments at a time.
So basically this is all the same family of technology where you’re basically using these
parallel readouts of the barcodes.
And then to do this, we used a technology called StarSeq for self transcribing reporter
assays, a technology developed by Alex Stark, my former postdoc, who’s now API over in Vienna.
So we basically coupled the StarSeq, the self transcribing reporters where the enhancer
can be part of the gene itself.
So instead of having a separate barcode, that enhancer basically acts to turn on the gene
and it’s transcribed as part of the gene.
So you don’t have to have the two separate parts.
So you can just read them directly.
So there’s a constant improvements in this whole process.
By the way, generating all these options, is it basically brute force?
How much human intuition is?
Oh gosh, of course it’s human intuition and human creativity and incorporating all of
the input data sets.
Because again, the genome is enormous.
3.2 billion, you don’t want to test that.
You basically use all of these tools that I’ve talked about already.
You generate your top favorite 10,000 hypothesis, and then you go and test all 10,000.
And then from what comes out, you can then go to the next step.
So that’s technology number two.
So technology number one is robotics, automation, where you have thousands of wells and you
constantly test them.
The second technology is instead of having wells, you have these massively parallel readouts
in sort of these pooled assays.
The third technology is coupling CRISPR perturbations with these single cell RNA readouts.
So let me make another parenthesis here to describe now single cell RNA sequencing.
So what does single cell RNA sequencing mean?
So RNA sequencing is what has been traditionally used, well, traditionally the last 20 years,
ever since the advent of next generation sequencing.
So basically before RNA expression profiling was based on these microarrays.
The next technology after that was based on sequencing.
So you chop up your RNA and you just sequence small molecules, just like you would sequence
a genome, basically reverse transcribe the small RNAs into DNA, and you sequence that
DNA in order to get the number of sequencing reads corresponding to the expression level
of every gene in the genome.
You now have RNA sequencing.
How do you go to single cell RNA sequencing?
That technology also went through stages of evolution.
The first was microfluidics.
You basically had these, or even chambers, you basically had these ways of isolating
individual cells, putting them into a well for every one of these cells.
So you have 384 well plates and you now do 384 parallel reactions to measure the expression
of 384 cells.
That sounds amazing and it was amazing, but we want to do a million cells.
How do you go from these wells to a million cells?
So what the next technology was after that is instead of using a well for every reaction,
you now use a lipid droplet for every reaction.
So you use micro droplets as reaction chambers to basically amplify RNA.
So here’s the idea.
You basically have microfluidics where you basically have every single cell coming down
one tube in your microfluidics and you have little bubbles getting created in the other
way with specific primers that mark every cell with its own barcode.
You basically couple the two and you end up with little bubbles that have a cell and tons
of markers for that cell.
You now mark up all of the RNA for that one cell with the same exact barcode and you then
lyse all of the droplets and you sequence the heck out of that and you have for every
RNA molecule, a unique identifier that tells you what cell was it on.
That is such good engineering, microfluidics and using some kind of primer to put a label
on the thing.
I mean, you’re making it sound easy.
I assume it’s beautiful, but it’s gorgeous.
So there’s the next generation.
So that’s the second generation.
Next generation is forget the microfluidics altogether.
Just use big bottles.
How can you possibly do that with big bottles?
So here’s the idea.
You dissociate all of your cells or all of your nuclei from complex cells like brain
cells that are very long and sticky so you can’t do that.
If you have blood cells or if you have neuronal nuclei or brain nuclei, you can basically
dissociate let’s say a million cells.
You now want to add a unique barcode, a unique barcode in each one of a million cells using
only big bottles.
How can you possibly do that?
Sounds crazy, but here’s the idea.
You use a hundred of these bottles, you randomly shuffle all your million cells and you throw
them into those hundred bottles randomly, completely randomly.
You add one barcode out of a hundred to every one of the cells.
You then you now take them all out.
You shuffle them again and you throw them again into the same hundred bottles.
But now in a different randomization and you add a second barcode.
So every cell now has two barcodes.
You take them out again, you shuffle them and you throw them back in.
Another third barcode is adding randomly from the same hundred barcodes.
You’ve now labeled every cell probabilistically based on the unique path that he took of which
of a hundred bottles did he go for the first time, which of a hundred bottles the second
time and which of a hundred bottles the third time.
A hundred times a hundred times a hundred is a million unique barcodes in every single
one of these cells without ever using microfluidics.
It’s beautiful, right?
From a computer science perspective, that’s very clever.
So you now have the single cell sequence technology.
You can use the wells, you can use the bubbles or you can use the bottles and you have way
The bubbles still sound pretty damn cool.
The bubbles are awesome.
And that’s basically the main technology that we’re using.
So the bubbles is the main technology.
So there are kits now that companies just sell to basically carry out single cell RNA
sequencing that you can basically for $2,000, you can basically get 10,000 cells from one
And for every one of those cells, you basically have the transcription of thousands of genes.
And you know, of course the data for any one cell is noisy, but being computer scientists,
we can aggregate the data from all of the cells together across thousands of individuals
together to basically make very robust inferences.
So the third technology is basically single cell RNA sequencing that allows you to now
start asking not just what is the brain expression level difference of that genetic variant,
but what is the expression difference of that one genetic variant across every single subtype
of brain cell?
How is the variance changing?
You can’t just, you know, with a brain sample, you can just ask about the mean, what is the
If I instead have 3000 cells that are neurons, I can ask not just what is the neuronal expression.
I can say for layer five excitatory neurons of which I have, I don’t know, 300 cells,
what is the variance that this genetic variant has?
So suddenly it’s amazingly more powerful.
I can basically start asking about this middle layer of gene expression at unprecedented
So when you look at the average, it washes out some potentially important signal that
corresponds to ultimately the disease.
So that, I can do that at the RNA level, but I can also do that at the DNA level for the
So remember how before I was telling you about all this technology that we’re using to probe
the epigenome, one of them is DNA accessibility.
So what we’re doing in my lab is that from the same dissociation of say a brain sample
where you now have all these tens of thousands of cells floating around, you basically take
half of them to do RNA profiling and the other half to do epigenome profiling, both at the
single cell level.
So that allows you to now figure out what are the millions of DNA enhancers that are
accessible in every one of tens of thousands of cells.
And computationally, we can now take the RNA and the DNA readouts and group them together
to basically figure out how is every enhancer related to every gene.
And remember these sort of enhancer gene linking that we were doing across 833 samples?
833 is awesome, don’t get me wrong, but 10 million is way more awesome.
So we can now look at correlated activity across 2.3 million enhancers and 20,000 genes
in each of millions of cells to basically start piecing together the regulatory circuitry
of every single type of neuron, every single type of astrocytes, oligodendrocytes, microglial
cell inside the brains of 1,500 individuals that we sample across multiple different brain
regions across both DNA and RNA.
So that’s the data set that my team generated last year alone.
So in one year, we basically generated 10 million cells from human brain across a dozen
different disorders, across schizophrenia, Alzheimer’s, frontotemporal dementia, Lewy
body dementia, ALS, Huntington’s disease, post traumatic stress disorder, autism, bipolar
disorder, healthy aging, et cetera.
So it’s possible that even just within that data set lie a lot of keys to understanding
these diseases and then be able to like directly leads to then treatment.
So basically we are now motivating.
So our computational team is in heaven right now and we’re looking for people.
I mean, if you have super smart.
So this is a very interesting kind of side question.
How much of this is biology?
How much of this is computation?
So you’re the head of the computational biology group, but how much of, should you be comfortable
with biology to be able to solve some of these problems?
If you just find, if you put several of the hats you were on fundamentally, are you thinking
like a computer scientist here?
You have to.
This is the only way.
As I said, we are the descendants of the first digital computer.
We’re trying to understand the digital computer.
We’re trying to understand the circuitry, the logic of this digital core computer and
all of these analog layers surrounding it.
So the case that I’ve been making is that you cannot think one gene at a time.
The traditional biology is dead.
There’s no way you cannot solve disease with traditional biology.
You need it as a component.
Once you figured out RX3 and RX5, you now can then say, Hey, have you guys worked on
those genes with your single gene approach?
We’d love to know everything you know.
And if you haven’t, we now know how important these genes are.
Let’s now launch a single gene program to dissect them and understand them.
But you cannot use that as a way to dissect disease.
You have to think genomically.
You have to think from the global perspective and you have to build these circuits systematically.
So we need numbers of computer scientists who are interested and willing to dive into
these data fully, fully in and extract meaning.
We need computer science people who can understand machine learning and inference and decouple
these matrices, come up with super smart ways of dissecting them.
But we also need computer scientists who understand biology, who are able to design the next generation
Because many of these experiments, no one in their right mind would design them without
thinking of the analytical approach that you would use to deconvolve the data afterwards.
Because it’s massive amounts of ridiculously noisy data.
And if you don’t have the computational pipeline in your head before you even design the experiment,
you would never design the experiment that way.
So in designing the experiment, you have to see the entirety of the computational pipeline.
That drives the design.
That even drives the necessity for that design.
Basically, you know, if you didn’t have a computer scientist way of thinking, you would
never design these hugely combinatorial, massively parallel experiments.
So that’s why you need interdisciplinary teams, you need teams.
And I want to sort of clarify that what do we mean by computational biology group?
The focus is not on computational, the focus is on the biology.
So we are a biology group.
What type of biology?
That’s the type of biology that uses the whole genome.
That’s the type of biology that designs experiments, genomic experiments, that can only be interpreted
in the context of the whole genome.
So it’s philosophically looking at biology as a computer.
So which is in the context of the history of biology is a big transformation.
You can think of the name as what do we do?
That’s not true.
How do we study it?
That is true.
So all of these single cell sequencing can now be coupled with the technology that we
talked about earlier for perturbation.
So here’s the crazy thing.
Instead of using these wells and these robotic systems for doing one drug at a time or for
perturbing one gene at a time in thousands of wells, you can now do this using a pool
of cells and single cell RNA sequencing.
You basically can take these perturbations using CRISPR and instead of using a single
guide RNA, you can use a library of guide RNAs generated exactly the same way using
this array technology.
So you synthesize a thousand different guide RNAs.
You now take each of these guide RNAs and you insert them in a pool of cells where every
cell gets one perturbation.
And you use CRISPR editing or CRISPR, so with either CRISPR Cas9 to edit a genome with these
thousand perturbations or with the activation or with the repression.
And you now can have a single cell readout where every single cell has received one of
And you can now in massively parallel ways, couple the perturbation and the readout in
a single experiment.
How are you tracking which perturbations each cell received?
So there’s ways of doing that, but basically one way is to make that perturbation an expressible
vector so that part of your RNA reading is actually that perturbation itself.
So you can basically put it in an expressible part so you can self drive it.
So the point that I want to get across is that the sky’s the limit.
You basically have these tools, these building blocks of molecular biology.
We have these massive data sets of computational biology.
We have this huge ability to sort of use machine learning and statistical methods and, you
know, linear algebra to sort of reduce the dimensionality of all these massive data sets.
And then you end up with a series of actionable targets that you can then couple with pharma
and just go after systematically.
So the ability to sort of bring genetics to the epigenomics, to the transcriptomics, to
the cellular readouts using these sort of high throughput perturbation technologies
that I’m talking about and ultimately to the organismal through the electronic health record
endophenotypes and ultimately the disease battery of assays at the cognitive level,
at the physiological level and, you know, every other level.
There is no better or more exciting field, in my view, to be a computer scientist then
or to be a scientist in period.
Basically this confluence of technologies, of computation, of data, of insight and of
tools for manipulation is unprecedented in human history.
And I think this is what’s shaping the next century to really be a transformative century
for our species and for our planet.
Do you think the 21st century will be remembered for the big leaps in understanding and alleviation
If you look at the path between discovery and therapeutics, it’s been on the order of
50 years, it’s been shortened to 40, 30, 20, and now it’s on the order of 10 years.
But the huge number of technologies that are going on right now for discovery will result
undoubtedly in the most dramatic manipulation of human biology that we’ve ever seen in the
history of humanity in the next few years.
Do you think we might be able to cure some of the diseases we started this conversation
It’s only a matter of time.
Basically the complexity is enormous and I don’t want to underestimate the complexity
but the number of insights is unprecedented and the ability to manipulate is unprecedented
and the ability to deliver these small molecules and other non traditional medicine perturbations,
there’s a new generation of perturbations that you can use at the DNA level, at the
RNA level, at the micro RNA level, at the epigenomic level, there’s a battery of new
generations of perturbations.
If you couple that with cell type identifiers that can basically sense when you are in the
right cell based on the specific combination and then turn on that intervention for that
cell, you can now think of combinatorial interventions where you can basically sort of feed a synthetic
biology construct to someone that will basically do different things in different cells.
So basically for cancer, this is one of the therapeutics that our collaborator Ron Weiss
is using to basically start sort of engineering the circuits that will use micro RNA sensors
of the environment to sort of know if you’re in a tumor cell or if you’re in an immune
cell or if you’re in a stromal cell and so forth and basically turn on particular interventions
You can sort of create constructs that are tuned to only the liver cells or only the
heart cells or only the brain cells and then have these new generations of therapeutics
coupled with this immense amount of knowledge on the sort of which targets to choose and
what biological processes to measure and how to intervene.
My view is that disease is going to be fundamentally altered and alleviated as we go forward.
Next time we talk, we’ll talk about the philosophical implications of that and the effect of life,
but let’s stick to biology for just a little longer.
We did pretty good today.
We stuck to the science.
What are you excited in terms of the future of this field, the technologies in your own
group, in your own mind, you’re leading the world at MIT in the science and the engineering
of this work.
So what are you excited about here?
I could not be more excited.
We are one of many, many teams who are working on this.
In my team, the most exciting parts are, you know, many folds.
So basically we’ve now assembled these battery of technologies.
We’ve assembled these massive, massive data sets and now we’re really sort of in the stage
of our team’s path of generating disease insights.
So we are simultaneously working on a paper on schizophrenia right now that is basically
using the single cell profiling technologies, using this editing and manipulation technologies
to basically show how the master regulators underlying changes in the brain that are sort
of found in schizophrenia are in fact affecting excitatory neurons and inhibitory neurons
in pathways that are active both in synaptic pruning, but also in early development.
We’ve basically found this set of four regulators that are connecting these two processes that
were previously separate in schizophrenia in sort of having a sort of more unified view
across those two sides.
The second one is in the area of metabolism.
We basically now have a beautiful collaboration with the Goodyear lab that’s basically looking
at multi tissue perturbations in six or seven different tissues across the body in the context
of exercise and in the context of nutritional interventions using both mouse and human,
where we can basically see what are the cell to cell communications that are changing across
And what we’re finding is this immense role of both immune cells as well as adipocyte
stem cells in sort of reshaping that circuitry of all of these different tissues and that’s
sort of painting to a new path for therapeutical intervention there.
In Alzheimer’s, it’s this huge focus on microglia and now we’re discovering different classes
of microglial cells that are basically either synaptic or immune.
And these are playing vastly different roles in Alzheimer’s versus in schizophrenia.
And what we’re finding is this immense complexity as you go further and further down of how
in fact there’s 10 different types of microglia, each with their own sort of expression programs.
We used to think of them as, oh yeah, they’re microglia, but in fact now we’re realizing
just even in that sort of least abundant of cell types, there’s this incredible diversity
The differences between brain regions is another sort of major, major insight.
Often one would think that, oh, astrocytes are astrocytes no matter where they are.
But no, there’s incredible region specific differences in the expression patterns of
all of the major brain cell types across different brain regions.
So basically there’s the neocortical regions that are sort of the recent innovation that
makes us so different from all other species.
There’s the sort of reptilian brain sort of regions that are sort of much more very extremely
There’s the cerebellum.
Each of those basically is associated in a different way with disease.
And what we’re doing now is looking into pseudo temporal models for how disease progresses
across different regions of the brain.
If you look at Alzheimer’s, it basically starts in this small region called the entorhinal
cortex and then it spreads through the brain and through the hippocampus and ultimately
affecting the neocortex.
And with every brain region that it hits, it basically has a different impact on the
cognitive and memory aspects, orientation, short term memory, long term memory, et cetera,
which is dramatically affecting the cognitive path that the individuals go through.
So what we’re doing now is creating these computational models for ordering the cells
and the regions and the individuals according to their ability to predict Alzheimer’s disease.
So we can have a cell level predictor of pathology that allows us to now create a temporal time
course that tells us when every gene turns on along this pathology progression and then
trace that across regions and pathological measures that are region specific, but also
cognitive measures and so on and so forth.
So that allows us to now sort of for the first time, look at can we actually do early intervention
for Alzheimer’s where we know that the disease starts manifesting for 10 years before you
actually have your first cognitive loss.
Can we start seeing that path to build new diagnostics, new prognostics, new biomarkers
for this sort of early intervention in Alzheimer’s?
The other aspect that we’re looking at is mosaicism.
We talked about the common variants and the rare variants, but in addition to those rare
variants as your initial cell that forms the zygote divides and divides and divides, with
every cell division there are additional mutations that are happening.
So what you end up with is your brain being a mosaic of multiple different types of genetic
Some cells contain a mutation that other cells don’t have.
So every human has the common variants that all of us carry to some degree, the rare variants
that your immediate tree of the human species carries, and then there’s the somatic variant,
which is the tree that happened after the zygote that sort of forms your own body.
So these somatic alterations is something that has been previously inaccessible to study
in human postmortem samples.
But right now with the advent of single cell RNA sequencing, in this particular case, we’re
using the well based sequencing, which is much more expensive, but gives you a lot richer
information about each of those transcripts.
So we’re using now that richer information to infer mutations that have happened in each
of the thousands of genes that sort of are active in these cells, and then understand
how the genome relates to the function, this genotype phenotype relationship that we usually
build in GWAS between in genome wide association studies between genetic variation and disease.
We’re now building that at the cell level, where for every cell, we can relate the unique
specific genome of that cell with the expression patterns of that cell, and the predicted function
using these predictive models that I mentioned before on this regulation for cognition for
pathology in Alzheimer’s at the cell level.
And what we’re finding is that the genes that are altered and the genetic regions that are
altered in common variants versus rare variants versus somatic variants are actually very
different from each other.
The somatic variants are pointing to neuronal energetics and oligodendrocyte functions that
are not visible in the genetic legions that you find for the common variants, probably
because they have too strong of an effect that evolution is just not tolerating them
on the common side of the allele frequency spectrum.
So the somatic one, that’s the variation that happens after the zygote, after you individual.
I mean, this is a dumb question, but there’s mutation and variation, I guess that happens
And you’re saying that they’re through this, if we focus in on individual cells, we’re
able to detect the story that’s interesting there, and that might be a very unique kind
of important variability that arises for, you said neuronal or something that would
Energetics, sounds like a cool term.
So, I mean, the metabolism of humans is dramatically altered from that of nearby species.
We talked about that last time that basically we are able to consume meat that is incredibly
energy rich, and that allows us to sort of have functions that are meeting this humongous
brain that we have.
So basically on one hand, every one of our brain cells is much more energy efficient
than our neighbors, than our relatives.
Number two, we have way more of these cells.
And number three, we have this new diet that allows us to now feed all these needs.
That basically creates a massive amount of damage, oxidative damage from this huge super
powered factory of ideas and thoughts that we carry in our skull.
And that factory has energetic needs, and there’s a lot of sort of biological processes
underlying that, that we are finding are altered in the context of Alzheimer’s disease.
So you have to consider all of these systems if you want to understand even something like
diseases that you would maybe traditionally associate with just the particular cells of
The immune system, the metabolic system, the metabolic system.
And these are all the things that makes us uniquely human.
So our immune system is dramatically different from that of our neighbors.
Our societies are so much more clustered.
The history of infection that have plagued the human population is dramatically different
from every other species.
The way that our society and our population has sort of exploded has basically put unique
pressures on our immune system.
And our immune system has both coped with that density and also been shaped by, as I
mentioned, the vast amount of death that has happened in the Black Plague and other sort
of selective events in human history, famines, ice ages, and so forth.
So that’s number one on the sort of immune side.
On the metabolic side, again, we are able to sort of run marathons.
I don’t know if you remember the sort of human versus horse experiment where the horse actually
tires out faster than the human and the human actually wins.
So on the metabolic side, we’re dramatically different.
On the immune side, we’re dramatically different.
On the brain side, again, you know, no need to sort of, you know, it’s a no brainer of
how our brain is like just enormously more capable.
And then, you know, in the side of cancer, so basically the cancers that humans are having,
the exposures, the environmental exposures is again, dramatically different.
And the lifespan, the expansion of human lifespan is unseen in any other species in, you know,
recent evolutionary history.
And that now leads to a lot of new disorders that are starting to, you know, manifest late
So you know, Alzheimer’s is one example where basically, you know, these vast energetic
needs over a lifetime of thinking can basically lead to all of these debris and eventually
saturate the system and lead to, you know, Alzheimer’s in the late life.
But there’s, you know, there’s just such a dramatic set of frontiers when it comes to
aging research that, you know, so what I often like to say is that if you want to engineer
a car to go from 70 miles an hour to 120 miles an hour, that’s fine.
You can basically, you know, fix a few components.
If you wanted to now go at 400 miles an hour, you have to completely redesign the entire
car because the system has just not evolved to go that far.
Basically our human body has only evolved to live to, I don’t know, 120, maybe we can
get to 150 with minor changes.
But if, you know, as we start pushing these frontiers for not just living, but well living,
the Fzine that we talked about last time.
So to basically push Fzine into the 80s and 90s and a hundreds and, you know, much further
than that, we will face new challenges that have, you know, never been faced before in
terms of cancer, the number of divisions, in terms of Alzheimer’s and brain related
disorders, in terms of metabolic disorders, in terms of regeneration, there’s just so
many different frontiers ahead of us.
So I am thrilled about where we’re heading.
So basically I see this confluence in my lab and many other labs of AI, of, you know, sort
of, you know, the next frontier of AI for drug design.
So basically these sort of graph neural networks on specific chemical designs that allow you
to create new generations of therapeutics.
These molecular biology tricks for intervening at the system at every level, these personalized
medicine prediction, diagnosis, and prognosis using the electronic health records and using
these polygenic risk scores weighted by the burden, the number of mutations that are accumulating
across common rare and somatic variants, the burden converging across all of these different
molecular pathways, the delivery of specific drugs and specific interventions into specific
And again, you’ve talked with Bob Langer about this, there’s, you know, many giants in that
And then the last concept is not intervening at the single gene level.
I want you to sort of conceptualize the concept of an on target side effect.
What is an on target side effect?
An off target side effect is when you design a molecule to target one gene and instead
it targets another gene and you have side effects because of that.
And on target side effect is when your molecule does exactly what you were expecting, but
that gene is plyotropic.
Plyo means many, tropos means ways, many ways, it acts in many ways.
It’s a multifunctional gene.
So you find that this gene plays a role in this, but as we talked about the wiring of
genes to phenotypes is extremely dense and extremely complex.
So the next stage of intervention will be intervening not at the gene level, but at
the network level.
Intervening at the set of pathways and the set of genes with multi input perturbations
to the system, multi input modulations, pharmaceutical or other interventional, and that basically
allow you to now work at the sort of full level of understanding, not just in your brain,
but across your body, not just in one gene, but across the set of pathways and so on and
so forth for every one of these disorders.
So I think that we’re finally at the level of systems medicine of basically instead of
sort of medicine being at the single gene level, medicine being at the systems level
where it can be personalized based on the specific set of genetic markers and genetic
perturbations that you are either born with or that you have developed during your lifetime.
Your unique set of exposures, your unique set of biomarkers, and your unique set of
current set of conditions through your EHR and other ways.
And the precision component of intervening extremely precisely in the specific pathways
and the specific combinations of genes that should be modulated to sort of bring you from
the disease state to the physiologically normal state or even to physiologically improved
state through this combination of interventions.
So that’s in my view, the field where basically computer science comes together with artificial
intelligence statistics, all of these other tools, molecular biology technologies and
biotechnology and pharmaceutical technologies that are sort of revolutionary in the way
And of course, this massive amount of molecular biology and data gathering and generation
and perturbation in massively parallel ways.
So there’s no better way.
There’s no better time.
There’s no better place to be sort of looking at this whole confluence of ideas.
And I’m just so thrilled to be a small part of this amazing, enormous ecosystem.
It’s exciting to imagine what humans of 100, 200 years from now, what their life experience
is like, because these ideas seem to have potential to transform the quality of life
that, when they look back at us, they probably wonder how we were put up with all the suffering
in the world.
Manolis, it’s a huge honor.
Thank you for spending this early Sunday morning with me.
I deeply appreciate it.
See you next time.
Sounds like a plan.
Thank you, Lex.
Thanks for listening to this conversation with Manolis Kellis.
And thank you to our sponsors, SEMrush, which is an SEO optimization tool.
Pessimist Archive, which is one of my favorite history podcasts.
8Sleep, which is a self cooling mattress with smart sensors and an app.
And finally, BetterHelp, which is an online therapy service.
Please check out these sponsors in the description to get a discount and to support this podcast.
If you enjoy this thing, subscribe on YouTube, review it with 5 Stars and Apple Podcasts,
follow on Spotify, support on Patreon, or connect with me on Twitter at Lex Friedman.
And now, let me leave you with some words from Haruki Murakami.
Human beings are ultimately nothing but carriers, passageways for genes.
They ride us into the ground like racehorses from generation to generation.
Genes don’t think about what constitutes good or evil.
They don’t care whether we’re happy or unhappy.
We’re just means to an end for them.
The only thing they think about is what is most efficient for them.
Thank you for listening, and hope to see you next time.