More driving factor in your life right

then your doctors and then

change you because the minute
the numbers are right General

we're in the horses for
the guy who's out first

really well with flowers or

no great mystery about

what you want one thing is for
good science very interactive books

they are good little things they
interact with right away or

another box was put in the parking
garage is a pointless exercise and

what Warchus fantasy reconstruction was

working on associated with

what is now called Iraq next door but
it will go on for.

Thank you so much.

Thank you for having me here very
very impressive campus here.

You know this but it's amazing to come and
see it from somewhere else.

Honestly the level of inter-disciplinary
that I just saw since last night over

dinner and this morning is I guess make
me jealous but honestly it's wonderful.

Anyway so I'll start with a quote and
thanks for the into.

Action by the way which is probably
better than what I will be able to

do for you guys.

I like this review by Paul Nurse and
if you're not.

I'm going to try to speak the language of
biologists to keep them interested this

morning but also try to reach out to
those who don't do biology every day

because we really need to work together.

Obviously if you go to biology every day.

This is a nice thing to read
Paul Nurse you Nobel Prize winner for

cell cycle discoveries.

Reminds us that there really
are four pillars in biology.

The cell right which is obviously
the basic structural and

functional unit of life.

Just imagine a time where you did
biology and you didn't know that right.

And that of course existed before we
knew about microscopes which again

was an engineering slash technology.

Discovery in a way.

The gene is the basis for hair ready.

We all know that now we think we have it
solved I think there are still questions

that are interesting though to be asked
as to what a gene is evolution by natural

selection is a card to stick of life and
life is based on chemistry and

many of you here and
I've seen examples of that this morning.

So those are four pillars that we all
accept it's basically under the pillow and

it's sort of by I want to one.

Now the gene is the basis of heredity.

Really started with this paper in
one thousand nine hundred one.

OK We always talk about double helix and
that sort of thing.

And obviously that's important.

The fifty three.

You know what's a quick but beetle and

got the Nobel Prize themselves
because they discovered this dry Klink

between a gene which at that time was
a very theoretical concept and enzymes.

Right.
So it's basically genes has you

need a very ready and
life is based on chemistry.

However the reason why I put this in
is look at the abstract of that paper

nine hundred forty one.

Just to remind us that
Systems Biology is nothing new.

What those guys want to do is this they
say an organism consists century of

an integrated system of chemical reactions
controlled in some manner by genes.

One hundred forty one since the components
of such a system are likely to be

interrelated in complex ways it would
appear that there must exist order of

the orders of directness of Gene control
ranging from simple one to one relations

which have been studied by biologists over
the last twenty years in great detail

to relations of great complexity
nine hundred forty one.

So based on that Paul Nurse
actually if you read the and

that's why I was interested in this review
is that he imagines that really the fifth

pillar the fifth key point of biology
with which we are much less of

course comfortable than the first for once
is summarized here as being biological

organization is based on logical and
informational processes and structures.

OK So you know the words were on
the stage now where the biologists

like to do their reasoning with those
four key points in mind the fifth one

is much less clear and how do we actually
go about this is the topic of this talk.

Now there are several ways that
you can think of organization.

First of all of you see how cells
are organized to form the brain.

For example or any other organ or
you can go inside the cell.

This is beautiful work actually we're just
talking about that earlier this morning.

Yes acquire electron tomography
by well my store where

this is actually a model that comes
from a set of pictures of the cell.

That are then being compared to
actual structures of enzymes that

people know about.

So for example in green here is the right.

So now you see in an intact living
cell the quote unquote organisation

of those structures.

Now obviously and
those are acting filaments here.

This is great I love to look at this right
and broadcaster gives amazing talks but

this is also if you think about it
from the point of view of the entire

proteome complete set of proteins and
also other micro molecules.

This is a very small number of
molecules right in terms of the type

of qualities of the genome
is capable of encoding.

Another way you can think
of organization and

that we know much more about because
of historical reasons life is

based on chemistry the work of biochemist
has been incredible over the years.

And so we have already here a notion
now of networks whereby basically

metabolize can be linked to each other
by virtue of how they are substrate and

products of common reactions and so
you of you've all seen this beautiful

metabolic network it turns
out this network is actually

in some organisation their properties
in this network that are interesting.

What we're more interested in
as you heard in the introduction

is what we call the intractable
networks for lack of a better word.

OK By interacting and we need we mean the
complete set of protein protein traction

in a cell or in an organ or
in an organism right.

So this is a beautiful structure right.

You have those two guys that homo demise
and then they interact with this other so

media over here.

What we would do if we
were looking at those

three interactions is we
would summarize it this way.

Beautiful structures three dimensional
structures as you just saw would be

summarised down to a node for
a single protein and

then the interactions that we detect
between them would be summarized by.

Links or alliance or sometimes called
edges as you will hear over and

over in the talk so in ninety
three I was working in a lab that

was very good at detecting interactions
and I started thinking you know how do you

actually start thinking of a way to get
out of the protein that we were staring at

a time in our B. And how would
the network look like if we're able to

map them all basically essentially so to
speak come out of our own neighborhood and

start zooming out like you could do on
Google Earth for example or something like

that and you know pretty soon you
go from the Georgia Tech campus or

the way to the United States of America or
something like that.

And so ninety three we didn't really know
the number of proteins will come back to

that in a second but we knew it was
a large number of things relative butt.

Tens of thousands and so the idea was
can we actually generate data that

would give us an idea of what this network
actually looks like so the hypothesis is

that global and local properties
are being tracked on networks relate to.

Biology.
That's that's the major assumption

of all this work.

OK that essentially that if you
were able to do map this right.

This huge beast that what you would find
there would not be random that there

would be some information there that
might relate back to to the biology.

Remember the fifth point
of Paul Nurse that

biological organization
might be the fifth pillar.

My main motivation was that even though
I trained as an engineer I then became

a geneticist and I realized very
quickly that they were very many and

still today unresolved aspect of genetics
which I'll tell you about the next slide

and also a second motivation for
me it was this idea that gene number.

Versus biological complexity
was very part Ochs ago.

I'll show you that in a second.

So those were my two main motivation
to tell you what I mean by and

resolve aspects of genetics.

This is very cartoonish and it's designed
that way so I apologize for those who know

a lot about genetics but you know we keep
saying US Genesys this gene does that

this gene does that and if we knew
what all genes do that will solve it.

It's actually just not true.

You know for one thing.

Every time we do an experiment where
we compare wild type to mutants right

where we mutate a particular gene.

We have anyone sometimes
all four of those problems.

OK.

Many mutations gave rise to no fear
type whatsoever in which case what does

the gene do really.

We can't detect anything at least
under normal laboratory conditions.

Many mutations in one gene will give
rise to many different kinds of

phenotype depending on the conditions and

sometimes to classically depending on
the individuals that you're looking at and

more importantly in warms For example
we have beautiful examples where if you

compare a wild tribe to a several
mutant animals only a small

number of the early months even though
they're completely I said gently.

Tical genetically
identical eyes are Janet.

So they are genomes identical yet.

Their phenotypes can be
different between each other.

Called sometimes incomplete penetrance and

by the way the same sort of thing can be
explained with this weird word called

expressivity whereby different animals
with the same genotype can actually give

rise to different kinds of you know
quantitatively different kinds of type.

So what can explain the six right
sort of good old model one gene

one function from beetle into
two in the second slide.

It's not going to be correct.

There's just no way in the second
motivation that was mine and

also I'm sure you've been hearing.

That argument so many times is that how do
we explain that this is those organisms

with different could include I put quotes
because I get this question all the time.

What is complexity how do you measure it.

I don't know but I don't really
feel like a new cell myself and yet

we only have three thousand genes.

You know three fold difference
in number of genes between us.

So how can we explain that if we take
the simple model that D.N.A. makes R.N.A.

makes protein and that makes that
provides functions functions one time and

will solve it that way.

So perhaps we might retire
this sort of central dogma or

at least find a complimentary idea and

that would be the following is that
the cell is a system that's organized very

differently then then then this sort
of you know sort of focused view.

So the idea is that the interactive
network proteins but

also by the way I mean to include
everybody micro molecules are nice if

that's if that's if that's
when when that is relevant.

That this network actually impact of
the genome that you know of course

is important.

It contains all this information but it.

Genome is nothing without
the interactor network.

If you close your eyes one second.

If you isolate D.N.A.
D.N.A. will do nothing.

If you.
There are cases where we can

do a lot of beautiful biology with
interacting proteins without D.N.A.

that's really the motor of everything
that's the dynamic aspect and

obviously there's a transcript in between.

So you know one of the motivations
to do this now how do you start

building that network as I was trying to
show you there are several ways and it's

too short in fifteen minutes to really
cover other people's work I apologize for

those who might be in the room or.

And we can answer questions to
compare what we do to others.

Of course I'm here to tell you what what
or approach doesn't necessarily mean that

it's the best what it's meant to be
a complimentary one to what everybody does

in the field but where everybody
agrees I think is on this one and

that is that if you had a genome and
if it was well annotated and

if you knew what all proteins are of
an organism think about yeast C.

elegans whatever else and then you could
express all those proteins you could make

them your member in network theory
a lot of people map their networks and

the elements the components already
exist if you're sociologist

people exist already you just have to
map how they interact with each other.

In your field we first have
to make our own versions of

the components of the network before
we can actually map the network.

And obviously the goal is to test or
provide combinations for

possible possible physical interactions
and this is very important for the people

in the room who are more quantitative for
those who don't follow biology closely.

This is an old cartoon it's a very
famous paper in one thousand nine

hundred nine published by Stan field and.

Song where the idea is really for
us revolt revolutionary and

what it is is that you put two
proteins that I will call X. and Y.

you make marriage proteins make
you make basically hybrid proteins

one hundred between X. and the D.N.A.
binding domain that can bind D.N.A. and

another hybrid with Y. and an activation
of a second act of a transcription and

only when the two approaches
can interact with each other.

Do you get a transcription factor so
to speak and

that can be measured by the turning on
of the gene that's downstream here.

It's called a used to hybrid system again
trying to speak about those things for

people who have read different
kinds of literature.

I know some of you know this.

And the point of this
science to show that.

This idea where you can no hope to do this
in a high throughput way of systematic

well controlled unbiased that's very
important mapped a network as you'll

see a little bit mapped a network in
an unbiased matter is extremely important.

OK So that tool for us was really
what you know made a start thinking

about the making of this entire
enterprise of looking at the network

now before I go into five or
six sides of a little bit of history and

then we'll get into the current
stuff that's not published.

I want to do a couple of course notes
because I get that question all the time

and it's a good question and actually I'm
writing a little perspective on this.

We are mapping what we could call
the biophysical network meaning.

We're testing proteins two at a time and
the question we're asking can be interact.

So that's what I mean by possible
interactions in the reasonable reasonable

dynamic range.

OK this is to be contrast to what
biologists really would like to have and

that is networks depending on the
conditions the cell types the organs or

whatever the cell type is I said that that
you know different networks where do we

know that the interactions actually do
take place in vivo and are relevant for

whatever physiological reaction
that they're looking at and

I think that those two networks
don't need to miss that.

Excuse me that the addition of the union
of all those networks here doesn't

necessarily be need to be identical to
this one and I can discuss that during

the question session and yet reason this
sort of going like this with us also work.

I think there are physical
interactions that we can detect

that are absolutely not.

Artifactual you know.

Not at all they are happening
in their bio physical but

they are prevented from happening back in
vivo remember we're using an artificial

assay here I can discuss this more if you
want to but I wanted to put that into.

OK so the whole thing started
really with three papers one by.

One by Stanfield himself came up
with a used to hybrid system and

one from our lab in in two thousand and

what we have done is to basically started
this whole matrix by saying you know.

Do it all without any controls is
completely ridiculous we're not going

to be able to do anything with that let's
start with known proteins a small number

go deep into all the proteins predicted
from the genome sequence and take those

proteins as being well characterized from
a well can I trust biological process.

So we chose at that time something that
might you know it's just what it is it's

vocal development and they are proteins
that are really well known chemically

characterized genetically characterized
pass with you've seen those images.

Even if you don't reviled you every
day those names are proteins and

genes and that was the picture of the so
to speak regulatory

pathways that lead to normal
vocal development in C.

elegans as found out by a number
of generations by chemists.

When we went in and said let's be specific
be systematic and take each one of

those notes over here and try to establish
edges not only between them but with

the rest of the of the predicted proteome
this is the network that we found.

And now the southern those passwords
are completely collapsed into

basically a very highly connected
network much more highly connected and

obviously than than certainly I would
have been able to imagine at that time.

This is very small dose.

So a couple of years later we went in and
mapped what I would say is about

three to four percent of the entire
interact on network for this little beast.

So those networks of course
are descriptive right now

I'll show you in a second what we do with
them but I'm showing you this slide.

This is a paper that we published
in two thousand and four and

you can go have a look at it.

The point is that again.

It gives you this amazing
flavor of high connected.

In the cell.

You know sort of think about networks
think about it again sociological networks

how we are you know not more than six
degrees of separation away from any

any any person any other person
on earth and that sort of thing.

This is the sort of analysis this is the
sort of organisation that we can try to

know and understand about what
proteins are doing in the cell.

Now we were not only
the only lab in the game.

This is a map very early map
also a first draft map for east.

This is C.L. against this is just awful.

And it came out from
a company called Cura gen and

then we just recently and I'll say more
about that in a second we started to

we are tempted to go into the human
interactive networking try to map that.

OK So the bottom line about those maps

again is that they don't cover much
more depending on the organism

that you're looking at much more than one
two three percent of the problem there and

I'll give you a demonstration
of that in a little bit.

On the other hand there are sometimes
being criticised let me

put it this way in terms of quality and

I want to also to talk about the quality
of this data with you guys so

that people who are in the room
are interested in looking at this data and

do analysis on it finally start feeling
comfortable about about what's happening

here before I do that don't I just want
to show you a few more slides as to of

course to illustrate to you that looking
at Bio physical interactions between

a great number of persons is obviously not
enough to understand biological processes

right and that we proposed very early on
this is two thousand to two thousand and

one that we would need
to have some dynamic.

And also functional
perturbation based functional

assays with which we could then
figure out a little bit better.

How those networks actually relate to
biology remember that's my that's my

original question here so
I'll just show you again by.

Illustration And you know pointing to
some papers if you guys are interested

in finding more about this.

So for example.

Yeah sorry again while
I talk about sociology.

This is a very interesting case here.

I know for those so
this graph is really drawn by

Mark Newman who has incredible
craft Fareast And I think.

And so here what you're looking at is
a social network in a high school right.

And I want to make two points about about
this is a game to try to illustrate what

it is that we do.

And so basically one point is that

what you do here is you go to students
in a high school you say give me.

I forgot what the number is but
your father five best friends or

something like that.

And so for any student then an arrow
points out to their five best friends and

you do that with all students.

OK And that's all you do you don't know

how friends you know what they
do when they're together.

Why they're friends when they're friends
when they see each other that the dynamic

stuff is out another fee and yet you can
see from this graph that you know they're

interesting concepts are coming back
actually probably remind you of your own

time in high school that for example this
was already somewhat of a good memory for

me when I was in high school I always
felt I always felt I was on that side.

You know and also the interesting thing is
that some kids say that this one is their

best friend but this one doesn't say
the same on the other side you know that's

OK So I'm just trying to illustrate
with a funny example in a way this

group a work program that even though we
don't understand all the details once and

all the dynamics of all
those relationships yet

there are interesting features that are
coming back up this call is reflected in

the city of the city so and
I did put the code here

because I mean I don't even remember
what it is like but you can imagine that

this reflects the sort of society that
that we go through and that's the point.

What's the code.

So that's almost exactly the idea here
is if we could build networks where we

the edges are physical attractions but
we would give a color code to the nodes.

Relative to other kinds of aspects other
kinds of partners that can be measured for

genes then that would give
us images of this kind.

So one example that a graduate student who
he who is now she's a white had fellow.

She so she came up with this really
beautiful example where you basically take

an expression cluster matrix and you're
looking at my current data and you know

how you can do micro rate on all the genes
of yeast across many different conditions.

Now if you do the genes or
can be clustered with each other and

you can color them to simplify a supposed
to supposed to according to the type

of expression profiles that this that they
show across many different conditions.

So basically what we learned here is we
could take a very basic static almost

boring protein protein
physical interaction network.

And basically give it back some life

by just looking at how the dynamics of the
expression of the corresponding genes OK.

Another one I saw you had
five your piano coming here.

Maybe a few months ago or
something like that somebody was so

this is another example where we had
a network here of physical interactions

starting in bloom with proteins involved
in D.N.A. damage response in C. elegans.

And the white stuff was completely
new nodes unknown to that point and

they were again added to
the network based on the fact that

those proteins can interact
with the starting ones.

So what we did in this case is
called phenotypic profiling.

OK What we did here is take
the starting genes and

the genes that I could detractors and
then do R.N.A. I own all of them.

So I think that's pretty much covered for
anybody.

It's basically knocking down each one of
those genes using this tool called our AI

and that's important.

Conceptually is to then measure
systematically different

phenotypes different readouts So in other
words I'm going to that network here.

And I'm perturbing each node one of time
sort of basically making it go away.

And then I read what's at
the end in a systematic manner

looking at different things that
the worm is now not capable of doing

cell cycle arrest reduced up of ptosis and
the picture that Simon Bolton

when he was in the lab came up with
is basically a way to group genes.

According to the kind of field types
that they will show upon knock

knocking knocking them down or perturbing
them or knocking them out by the way

doing our AI OK So to not lose genes
are closer to each other because

the finity profile that they show upon
perturbation is very close relative to for

example another group here of genes
that seem to be doing a different

kind of biological function now and
probably most of showing you this

slide this is done in cooperation with
his lab work by Kristen consolidate and

many of the colleagues and now you get
to a point where in this particular case

we're studying cell against early embryo
Genesis is just to sell divisions

very early on in the against and
we're Genesis and this matrix is much more

complete Obviously this is six
hundred sixty one genes and

forty five phenotypes measured
again upon already I or

pro-traditional knock down of every
single one of them wanted time and

now obviously you start getting really
interesting profiles right groups

of genes that tend to show a very similar
profile of phenotypes out of this.

By the way.

So if you look at the relationship
between being close to each other

as measured here by the phenotypic
correlation on this matrix and

having proteins that
interact with each other.

You have this very nice positive
correlation between the two so

you know where it's at that scale
trying to look at the organization of

a cell making two and
two cells making four.

If you look at two very different
things for that to be profiling and

physical interactions between proteins
the genes that tend to be doing the same

thing here in this matrix then to
encode interacting proteins and

from that you start getting those costers
of those clicks another like this.

So logical network I was showing
you as a as an illustration.

OK.

Now remember we said global properties
of interactive networks relate to.

If you will function.

Every solve that.

We know for sure.

Not quite I showed you a number of
correlations I showed you a number

of maps as an attempt to
try to get there to do what

I would like to do is to ask you know
three basic questions in the context OK.

How am I doing on time.

That's that's OK.

So what is the size of this beast.

So if you think the human and

since we know I have that on my next slide
I think we know the number of genes.

Even though actually the number of genes
has been this is a slide from one thousand

nine hundred four.

I think OK at that time the estimates
range between fourteen thousand and

three hundred thousand genes for
example this ninety four.

It's actually not bad More fact
that I love the paper because and

that's the point of the next part of my
talk is it provides a framework to try to

estimate the number of genes without
the full genome sequence at that time.

OK so now we think it's about you know
twenty twenty two thousand genes each and

putting obviously at least one
protein sometimes many more than one.

You know that's the phenomenon of eyes
of forms coming from splice variants.

So basically what you can try to do is

here is to look at how you want to do this
how you want to estimate the number of

proteins protein protein
interactions in the cell.

If you could say OK I'm going to go try
to read everything that's known about

protein protein interactions.

OK So there are tens of
thousands of papers if not more.

And so their databases that have
attempted to cure that information.

Five of them are represented here.

And so in other words it's one
way to get into the network by

saying it's sort of a bottom up where
you're looking at single labs having

worked a couple of interactions at a time
you really tens of thousands of papers and

you try to extract a.

Network from that.

And then like I tried to explain to you a
little bit earlier is the sort of you know

let me show you this one is sort of
the converse of that if it's basically

taking a totally systematic unbiased
way tried to express all the proteins.

You know map the network out of that and
so this is what of course what we've been

trying to do and I was referring to that
paper a little bit earlier on remember

that's the first attempt of trying to get
to the human interact on network map and

I just want to make a point here about the
fact that this other team in Germany in

Berlin team of any Quaker only work
has also published about the same time

as we did making descent some
sort of attempt so you know what.

How do you do it.

What's the best way to proceed.

How good is this how bad is that etc

Those are the ideas that I would
like to develop in the next.

You know five or six minutes.

Another way to put the question is
this is it if we close our eyes and

try to imagine the real.

Binary biophysical interactive network
that's encoded by the genome basically by

MIT by virtue of making proteins that
can interact with each other and

if we look at current interact on maps
would they be bottom up or top down right.

How close are we.

And how well can we use this to try to
estimate what the network looks like and

just simply today.

How many interactions they are you
know how can we size this beast.

OK.

I have the feeling
the colors are going up.

Is it me the.

All right so we just set up a framework
basically recently where we realised that

you know one of the things that
when people read those papers and

try to compare data sets for
example the shells on the reaal

data says that we just published
in a couple of sites ago.

You know a couple of number of people
forget that those networks are very

limited by what we call the or fail.

So for those who don't do biology every
day we do have a genome sequence.

Agree.
Yes those attractions not only

are by physical but
there are a physiologically relevant etc.

And then we took one hundred and

eighty eight pairs completely randomly in
this huge matrix that we have right now.

So if you have ten thousand genes
that you've cloned you have fifty

million combinations
of pairs between them.

Right.
And in those fifty million we pick two

hundred of them to some extent control
ingrained negative control in red.

So the next slide shows a number
of used to hybrid our say

is different versions on
the system remember the fields and

song thing with the reconstitution
of the transcription factor.

Well it turns out that we detect about
fifteen to twenty percent depending on

the versions depending on how to titrate
depending on the stringency fifteen

to twenty percent of the green that
is the gold standard procedure.

So the detectability of those
acids is about one fifth.

So for for one hundred interactions
are we know are right we

see about twenty percent of them and
that's under conditions where most

of the versions although not all some
of them are a little bit more sloppy.

But most of the version do not detect any
of the two hundred randomly generated

only selected pairs out of
fifty million Mexicans.

So that's a number that we have not
measured before and then lastly and

this is an engineering place so
I hope that I can actually make a good job

at this one but you know if you
go in the space I dislike or

first reality paper right where we
map seven thousand by seven thousand.

Again that's forty nine million
divided by two combinations right.

You do that you do the assaye once and it
turns out you're not successful at every

test the first time around
we call this coverage.

So we decided to go and use a subspace
of that and then in that subspace now.

Repeating our say four times which is
a lot of work I can tell you that.

And so now what we see here is the first
time around the number of interactions we

found and then essentially the number
of interactions from twice on.

Three times found four times now.

Each one of those by the way as
are as reproducible as each other.

It's not like those guys
are of lesser quality.

They just artistically have been found.

Once after four times relative to of
course a lower number that has been found

four times and if we take this
data model it it turns out that

to reach about eighty five
ninety percent coverage here.

We will need to do this test
about between eight and

ten times depending on the sort of thing.

And finally and this you know I'm sure
many of you who read this literature have

in mind is that if you take the current
interaction interactive maps and try to

imagine the real interactive network you
have to take specific city or should call

precision probably into account in other
words how many of your edges are wrong.

Artefactual due to the kind of a say
that you are actually using and so

we wanted to measure this.

And so to do that we said OK all the ass
is that we've been using that we've

done so far all the tests have used this
as a let's summarize two proteins common

interact with each other from a hybrid
with the D.B. from a hybrid with A.D.T.

were nice cells were in the nucleus and
that's where we testing the interactions.

Now we're going to do is we
basically getting out of all this.

It's the same idea you dealing with two
proteins that only they interact they

reconstitute a function but

now that function is a receptor it's
in a million cells and it's a membrane.

And so the problems of this one
are totally eliminated by using this

sort of speak or
orthogonal independent assay of that one.

So in other words the specific potential
problems of of using your yeast are say on

this particular site.

So now we're going to do is we're
going to go back to the gold

standard positive right.

The two controls the positive
control green negative control

red and now those are the actually
attractions that we're testing here real

ones here completely random
over here very small bag.

This is this new assaye called mapping.

It doesn't completely overlap

doesn't completely overlap with different
versions of these two hybrid system but

you can see there is some clustering
here where we have all for us.

Is that the union of all for

us is detecting about forty five to fifty
percent of what's known up there now.

This is the important site at
least from my perspective.

OK.

Remember the two high throughput papers
I showed you were all the nature and

styles all from any rank or burning group.

This is remember high throughput
close your eyes doing systematically

doing well do it well do
it well controlled but

do we do it do it unbiased systematic
This is literature curators.

Remember the five databases that
have read those tens of thousands

of papers where interactions
are described one interaction at a time.

If you use Map it.

Here's the experiment.

This is the rate of detectability
of the gold standard.

Plus other words it's it's this right.

It's your positive control.

This is the gold standard negative.

So the negative control over here
other those conditions no background.

This is the retest rate the precision
again if you want to do is too high to put

datasets relative to literature
curated OK So remember those are those

databases that go into a time to read
literature one paper at a time and

what we are starting to find here and
we by that we were so surprised by this

that we also requested it just by use to
Highbridge is the way we've actually done

generated the first two
the first two datasets and even

there the literature curated stuff doesn't
look as great to the point is this is that

there's a lot of literature out there and
many of those papers are great but

it's very hard to read them and
extract the information out of that and

rebuild your network bottom
up is my point here.

And so I really advertise for
you know a way to do this

in a very systematic way
the way I just tried to this.

To you considering the know
the number of nodes the eyes of forms

doing this very cleanly and
or in an organized fashion.

You know in a systematic way.

Considering the limitations of what I say
we need more I say is that's why I was

excited about talking to some of
the nanotechnology people here we are only

covering twenty percent of
the covered stuff that's money and

also reducing the size of
the experiments the volumes but

I think we can get there and I think
we'll be able to one day get a image

of the interactive map that resembles that
what's what's actually possible out there.

OK At this stage.

If you take the framework has we
have it and you really compute

back the number of interactions we
find in this given given space and

then introduce completeness
detectability and coverage.

Depending on the way we do it and
it's again it's a range.

That's why I showed you
the the gene number slide.

You know ten.

I don't know we think they might
be anywhere between one hundred

twenty one hundred thirty to a billion two
hundred million three hundred thousand

interactions possible
between human proteins.

OK.

And at this stage.

If you go to the databases
the bottom up ones which are not so

good as I showed you the top
down which are getting better.

You'll be in the range
of five to ten thousand.

So we really only have
the tip of the iceberg.

OK that's that I think it's one of the.

Take all messages in the next slide but
first I want to thank everything catus and

Jeffery Vasquez the first
author of the other paper and

my limits and this is done with
the labs of young and my own lab.

So the digital message of this
is to summarize the binary

file by physical I'm trying to
network might contain in a range

of a billion interactions
something like that.

According to the framework we now have
high throughput data can be higher quality

than loop with data that's you know
a barrel just at least are very surprised

by this when I show this data and
I'm ready to discuss it with you guys.

During the question session and

the current challenges with those maps
is not so much false positives but

it's really false negatives so
again we only have the tip of the iceberg.

And I think we have provided you
with that sort of a new framework

to map this network.

OK.

So this teach like to put this
more into perspective I like

to compare the same to
the genome sequence project

because a lot of people know at
least a little bit about that.

And so you might say well.

So you only have the tip of the iceberg
you still worried about your

quality What are you doing
exactly what those networks mean

I just want to put into
perspective this is.

This is a very famous
paper about this one.

There are people who can see this

right the first the first
sequence one of the first.

Sequencing reaction fifteen years
later you have one chromosome of use

that sequenced and
fifteen years later now you know what

the genome sequences are like It's amazing
it's great quality it's out there.

We all use it every day.

The fields in song Paper which I think
is almost as important as this one.

Well it is not much more than
fifteen years ago it was

one thousand nine hundred nine.

So when we generated those first maps it's
as if we had like a small comes out of

a small organism.

And now we need to get organized.

We need a good framework we need to
think Nano and we need to increase

the the Thora of assets to try to
get there but I think it's possible.

Again how much time I always
lose track of time and

I don't want to be rude
about them is that.

So another vignette that I would like to
discuss is despite all those limitations

when we start using
interactive networks and

put that back into a framework
of of human disease.

OK So because at the end of the day and
I think I have this over here.

Yes.

If you have taken any
class even in high school.

I'm sure you understand
this biology class.

That is in the genome or or differences
into genome might make us more susceptible

to diseases of this kind or that kind but
really the relationship between.

In them hasn't been that clear.

I mean the except for

a few examples right what really
probably happens is we didn't is here.

I think the nodes are and or

the edges of the network the network
properties are changing and

that gives rise to the phenotypic
consequences that we see in the clinic.

So we're trying to get at this idea and
try to relate this

interactive network models to what's
seen in diseases in general and

more particular in genetic diseases.

So who does Zogby who's at
Baylor She's a specialist of a taxi.

I was telling me you know you were
very good at looking at patients and

trying to help them doing their pathology
figuring out the genetics right.

Map it cloning to genes and

then looking at their product we
were pretty good at that right.

And so far I think she was telling me that
I showed that in two slides that there

are about three dozen genes so
between thirty and forty genes.

Mutations in which can give rise to
those very bad diseases right when she

was telling me is as I was giving a talk
and I met with her in her office right

after that she said I really want to see
where those genes or the products of those

genes are in your intractable networks are
there relationships that again would not

be expected by that would be far from
what you would expect by chance.

So this is just a one example of
those you know thirty or forty.

A list of thirty or forty genes very weird
names you know attacks in Calcium channel

dependent something top box binding
protein very little to do with each

other not not really a pathway you know
not not some sort of really nice coherent

biological story and imagine a you
know a list that goes down to thirty.

I'm sorry did I forget the the actual
numbers between thirty and forty G.'s.

So it's very simple question just to give
you a sense of where this is going you

know well what if we measured the shortest
path length between those notes.

You know about this in
that work theory right.

So.

You can actually not only measure
the shortest path as between two notes but

also measure the average shortest
path as between a series of notes and

that's what we did here if we look at
those proteins involved in the facts

in this otherwise you know quote
unquote static and boring network.

Well what we see is the shortest path
the average the mean shortest path and

sometimes between them is extremely
different from what you would expect

from a distribution of randomly generated
either networks or attacks on nodes.

So somehow they're closer to each other
in the network models as we have them

today than what would be expected by just
now remember this is important because

we take turns in those nodes that
are relatively close to each other

we need to do is absolutely amazing
diseases out there and doing.

They look very much the same like those
kids behave the same way in a way even

though the molecular defects that affects
them can be summed up some very different.

So let's about a bash the many
of you know him who's worked at

a number of fundamental aspects of
any network with a biological or

not came on sabbatical to my lab and.

I was telling him about this and he said
Well can we do if you can do this with

thirty to forty genes can you
do it with all of the genes.

We didn't see which are known to
cause one or more of these orders and

you could you do that in a bipartite graph
matter where you would consider not only

all genes involved in disorders but
also all disorders that those

genes are known to be involved with OK so
if you don't do biology every day.

This might be surprised if you but

they are today in those databases
me show you one in fifteen.

This is the growth of the number
of genes known to be involved with

a number of disorders.

It's absolutely phenomenal.

You know some fifteen hundred genes
are known to be involved in one or

more disorders there which I think
it's an incredible information so.

Go have post-op in decided to look at
this problem from a from a bipartite.

Right.
So now you have the diseases right.

And you want to link them to each other
if they have at least one gene in common.

And then you have a gene network where
you're going to link them to each other if

they are known to be involved in
at least one common disease and

that's because there's not again
remember my very very first light.

We tend to hear in the news this
gene is involved in this disease.

Right.
And from what we remember from that

is that they sort of are one
to one relationship but

that's very far from the truth.

Many You know the majority of
diseases can be obtained by

mutations in more than one gene and
vice versa.

OK anyway so here's a network of diseases.

So the nodes here are this is causing
cancer leukemia obesity right.

And they are linked to each other by
virtue of the fact that they share common

genes mutations in which are so shared
with perceptibility to get those diseases.

I won't give you all the details obviously
because of time but this was generated

completely blindly the way to draw those
graphs is that what we did is we colored

You know it's again somebody asked about
the color code in one of the networks.

Well this relates to the kind of this
is this is in blue here or green.

This is cancer for example.

Now we didn't put those notes here.

OK the algorithm by virtue of
trying to draw this network

in the most I guess comfortable
way of looking at it and it up

with this particular region including
most of the cancer diseases here.

We didn't do that.

OK so what I'm trying to show you with
this image is that diseases are related to

each other in a way that most of them
are actually part of a main component.

There are exceptions to the other one here
which I could discuss until tonight if you

want to remember there's the counterpart
of that this is the gene network.

So now the nodes are genes and
the links between them indicate

that they're involved in at least one
common disease that makes sense and

I want to stop with that
to just show you one.

I think.

Very interesting observation
to just leave both worlds

the sort of all out world of bio physical
interactions that are possible between

proteins of the human genome and that.

I guess more practical of genes can cause
such and such diseases it turns out that

if you look at the two graphs and
put them on top of each other.

The relationship of genes
causing common diseases and

protein protein interactions is
again far from what you would expect

from a set of random randomly
generated negative control networks.

OK So in other words proteins that include
that in trying to have each other.

Tend to be included by genes that are
involved with the same sort of diseases.

Again the same sort of
separation I was showing you

a little bit earlier with with C.L.
you get.

To finish one very early.

I mean a very recent observation I don't
know really what to do with that you

probably have if you know anything about
network you probably know that there's

this idea that some of the notes on
those networks are called helps because

they're highly connected.

I've heard that Atlanta is a hub at least
three times since I got here yesterday and

and so you know so if you actually look at
the distribution of number of interactions

and the proportion of nodes that have
those number of interactions you get

some sort of a parallel.

I want this because
the details of that but

it turns out that in the cell the proteins
that tend to be highly connected like this

have been shown to tend to be more
essential for example for life then

the proteins are much less connected again
using the current truck term models.

So I think that's Atlanta right.

Probably.
So you've seen this before.

So what we wanted to do is well.

Those global properties
that we're starting to.

Observe in those networks it is
still a lot of work to be done for

one thing we need better maps I tried to
tell you that two or three times today or

more but still given the limitations
of what we have today.

Can we start looking at
a global problem like this.

When a virus infects the host cell.

The virus has itself twenty's right.

And those proteins of the virus
are used to basically.

In the worst scenario kill the cell but
in many scenarios to rewire the network so

to speak to change
the phenotype of the cell and

make sure that the cell is now doing what
the virus needs the cell to do right.

So we said well can we actually
use a bio physical interaction

strategy to try to look at where
it is that the viral proteins

are going to grow in the interactive
network what do they tend to interact with

so we generated very easy dataset
to think of the yellow nodes

here are virus viral proteins and the
green ones are the host proteins right so

we're looking at the viral proteins
interacting with the host proteins and

we've made a couple of I think relatively
interesting observations we can't explain

them right now quite yet but
remember the Hub's Well it turns out

that relative to all other
nodes in the host network.

The viral proteins tend to
target the hubs more so

than you would expect by chance
as well and not only that

remember the idea of shortest path length
between nodes and example of attacks.

The target proteins in the host
tend to be more closer to

all the other nodes of the network then
again then you would expect by chance so

the occupy a position that's what that's

significantly different than
the entire set of nodes in the network.

Again we don't know what that means but
the hypothesis here and

it's really only a hypothesis is that
is there an evolution when you know

when there is a relationship of
host pathogen whereby the pathogen

needs to quote include circumvent
some of the global properties

of the interactive network of the host
in order to do what it's supposed to do.

I don't know.

So if I had five more minutes.

I would do a last vignette but
I can stop here.

I always do this and
I always get five more minutes.

I'm just trying to be nice and.

Is that OK.

So.

Which is essentially to introduce
back to the unit of those networks.

You know.

So basically here is the question
that I want to finish with so

I told you about interest.

That's what models.

I told you about fear
types like in the worm.

But also in people right.

So but one thing is that we never
really investigate in this field

is whether the perturbations that we're
talking about a network related to nodes.

You know you remove a node and then that
changes the properties and that's working.

You see if you type.

Well or and
actually they are not mutually exclusive.

Obviously Or
do you sometimes actually affect and that.

And so and the way the network is
rewired is by keeping both nodes

here completely normal at the exception
that they cannot see each other anymore.

You know words to talk about proteins
that a protein would have retained it

would be a mutant would be abnormal.

But it would have retained the ability to
interact with all its products except one.

And so we wanted to ask in human disease.

Considering the interactive network as we
have it today for the for humans can we

distinguish between the two so and
again this will be a little quick for

those who don't know biology every day but
I don't have much time.

If you have a stop going
on in a gene right.

Essentially very early on.

For example you don't have the protein
at all and if you have if you don't have

the protein the node is gone in the
network on the other hand many patients.

I'll show you the second are due to single
I mean ASA change as we call them so

the D.N.A.

is changed so that you have a new I mean
a lawsuit here but if to the normal one.

And that potentially not always but
potentially we thought could

be associated with the loss of an edge and
not the loss of those necessarily right.

So for

example if you look at those Remember all
those genes in the only database that

are known to be associated with all the
disorders Well it turns out that about how

of the mutations are variable so
far that have been found in patients.

Our of this kind right here.

Single very I guess.

You know not not destructive
right type of change.

OK so having that in mind we then
went ahead and said Well we know

of genes that sometimes are associated
with more than one disease right.

When we have that case and
we know what mutations what

changes in those genes
are socially to those diseases.

Can we actually find relationships
that would indicate that

this model is actually not completely
ridiculous and we find that for

dozens and dozens of genes so
let me just restate this.

This is a genes mutations in this portion
of the gene encoding this domain in

the protein will tend to give rise to this
disease mutations in that part of the gene

will give rise to another type of
disease that in the clinic you

really see the difference between them
supporting this model of a specific

patient I'll skip this one.

So just to finish.

I want to tell you that we're now
experimentally going after this idea.

And so we basically cloning hundreds of

we take that are found in patients
that are known to cause disease.

OK And we basically are looking at their
interaction capabilities relative to

the normal protein that is present
in unaffected individuals and

I'll just show you one example
with out up in the area.

It's a very famous gene mutations
in which gave rise to this disease.

We picked those are the illegals
that are known to

you know single I mean I
was a change on the other.

No it is protein and
essentially Here's the analysis.

Here's five examples all of which
are known to cause a disease right.

But again when you look at the physical
interactions of the corresponding proteins

with three of its interactors
the behaviors are very different and so

in this case for example the protein
is not affected at all in its ability

to interact with its partners those two
only of the protein seems completely.

Dead.

This would be more of
a node perturbation and

then those guys are somewhere in between
we call them as a joke right now

a jet it because it's sort of genetics
of edges and so those two proteins

cause an amazing disease out there and
yet retain the ability to interact

with some of their partners let me
show you that in a graphic idea so.

So if you don't look at those
two hundred stuff every day.

This is a normal guy.

And what we're finding is so most of
mutations that cause disease really make

the protein dead quote unquote.

But many and we have more examples
than what I have time to show you here

many actually affect the protein in
a very subtle way and only affect one of

sometimes two edges that that protein is
capable of mediating so I'll summarize and

remind you that we're in the business
of trying to tackle this question.

It's probably a very
important pillar of biology.

Now that I showed you the last answer
is obviously I think you might have

a feeling that by systematically mapping
interactive networks we're now going into

finding some global properties I didn't
show you and your local stuff but

I'm sure you know the work
of many other people and

more particularly I tried to tell you that
who are still very early on in this field.

If you look at this estimated size from
the first part of the experimental

side of the talk.

It's scary.

It's in the range of a million.

We think it's worth it because it looks
like it might help us understanding

human disease and
my point here was more to confuse you.

It's very early on it's
not ready to go but

just to say that you know we have
this one gene one phenotype idea.

It's very possible that we could
start thinking about this but

in terms of one edge one function one
phenotype and I'll put the side back I

think I thank the people as I was going
through and I'm running out of time.

So thank you very much.

I'm sorry I.

How.

Yes So.

Yes Well those are even more likely
to be agenda grade in a way right.

So so obviously as you all know

there are mutations that make
the protein apparently at least.

Make the protein unable to mediate some
functions and there those are usually

called also function you're right and
they usually are recessive noways and then

you have this petition that make proteins
do stuff that the wild type doesn't do.

And those are called gain a function
then they're usually dominant but

that's that's also not always the case.

So we've actually looked at that
one surprise that I had you know is

when you look at the human population.

And you look at all those diseases
some are called dominant and

others are called recessive obviously so
take Huntington's Disease for example.

Right.
That's a clear dominant one and

it's clearly again a function there is
this sort of probably going to mean

stretch that appears in the mutant and

those proteins do things that
the wall type doesn't do.

However I was surprised by the proportion
maybe because of my own lack of knowledge

but I was surprised by the number of
mutations that actually are dominant.

In the human population right.

In other words you only
need to inherit one version

of one of the offer from one
of your parents and that's it.

But which are actually lots of functions
when you look at the biochemistry.

OK And that actually is much more
than I than I would have thought and

it's you know Gene dollars each for
example or

it's usually explained by this I had
a slide that I skipped it not probably as

an answer to your question
were mine here right.

So you have the dominant negative also.

Mode of action right.

And so basically what happens is that is
that what I guess what I'm trying to say

is the dominant phenotypes back there in
the population do not always correspond.

If you look at the literature if you
look at what's known about those earlier

do not always correspond
to get a function.

I see.

Yes.

OK OK I see your question.

I'm sorry.

So so yes that's right.

We've done a small number of genes so
far and but

we really want to go
large scale with this and

we are starting to now include known
dominant mutations probably so

many of them gain a function it's going to
be a little bit harder of course because.

Lots of interaction we can measure
right I mean if we know of

interactions that the wild Had can mediate
and we can look at the illegals and

see whether or
not they can mediate if you if you this.

Again of interaction that's much harder
we have to actually go to a screen.

Basically yes that's right.

So it's a harder problem practically but
a fascinating one.

Yeah.

OK anyway.

We're just talking about that I
get that question all the time so

we do we can estimate it we can
give us sort of a ballpark.

But honestly the assay that we use is so
interact.

OK what you can measure the very end.

Right.
How something to do.

Obviously with the ability of two proteins
to interact but it also is influenced

tremendously by the stability of those two
proteins by how easily they go to the use

nucleus by how easily the U.S. will make
them will degrade them so there's so

many parameters that we'll end up giving
you what we measure here that yes there is

a correlation overall for the few actually
these are people just published where.

So e-mail me I can I can point you to it
where overall the story of the new and

strong interaction score very highly
in the south in the week a one ten

ten ten to two.

To score much more likely but

I wouldn't give a number to it now does
that mean we don't we should do that.

Yes we should and actually really
excited about this but we don't.

That's why again no technology or
right now measuring the Cady's

of thousands of edges from those
networks is very very hard.

Right.
I mean with the current technology that we

have you know by a quarter
would be one of them that I

think if we could do this
in much smaller volumes.

Yes it's absolutely critical
that we do that but

honestly I don't think we have the tools
of course I mean many people perhaps

in this room to have measured Cady's for
you know one or two or three tractions and

have done it well but
to do it at the network.

You know level is another is
another problem and you know so

let me tell you why I say this is not just
to be smiley and say yeah you're right but

a fundamental question I have is if you
look at the distribution of Cady's for

physical interactions between proteins and
compared to species.

You can we you know basic question Do
we expect the same distribution or

from nice to ones do you see.

Do you see different distributions
do you do you go through lower or

higher kiddie is that you know are there
changes there that evolution has

given to the net to do all those
proteins so that the network can have

the properties is supposed to have so
they are really fundamental questions.

I love them but honestly practically
you know we need Current T.V.

to hear from people like you guys
to do to start at that scale.

OK

so you're pointing back to
the slide right where I was making

the point that we're looking at by
a physical and biological interactions.

Well.

So you know where we're measuring this
right now we're actually I'm trying

to remember I tried to say that
I'm glad you asked the question.

So imagine yeast.

For example and
all the conditions possible or

at least a good number of conditions
that the yeast cells can live under.

And so you have a bunch of
different tract of networks right.

And I was trying to make the argument
that if you take the union of that

that might be different than what we
measuring here which is the network of

possible interactions.

OK.

So that means there is if it's not if
the two are not exactly identical.

Obviously I'm thinking that
the biophysical interactive networks heart

has more interactions.

OK.

So the number the delta
is totally unknown.

But here is here's why I think
it's an interesting problem so

I tend to put words on things you know.

And in this case I'm not I'm not

successful at all because I've tried
to convince people in my life to work

on those things I call those
interactions pseudo interactions.

Like pseudogenes.

And you might say that's crazy.

You know it's just you just saying that
false positives have a cute name now but

that's not exactly true.

So if you sequence a genome for
example right.

You use a sangar reaction for
example and you go through things and

you see a beautiful gene or another.

Gene and then at some point you see
something else and at first you think

it's a gene then somebody else comes
in it's a no that's because of some

things they've seen that you hadn't seen
before the sequence is not wrong but

the interpretation of the sequence can go
back and forth as my fact the list of two

genes and real genes have been changing
over time because you know people

are just trying to do the best job
they can and trying to predict them.

Nobody is wrong.

It's just that it's hard to do it.

So if you had a biophysical network right.

It's like that.

It's like you know how would you universal
I mean I wouldn't be surprised at all of

those into that soldiers interactions
would never take place back if you go.

They would be to go right now.

How do you find them.

How do you measure them how do you measure
their proportion their percentages is

a very hard work and I would like to
finish this by telling you the reason why

I started thinking about this is it's
actually going the other way around.

And if you think of the problem.

Take you straight six thousand by said.

Isn't right.

That's eighteen million
combinations right combinations.

So of course no one knows
what the size of the U.S.

interactive network is I just showed
you an attempt at estimating the size

of the human interactive I'm sure that you
know we have a large range and it's just

sort of more methodology call that's that
was the point of that of telling you that

but still we can try to do it right so
many biologists by a physicist

biochemist will tell you I don't know I
think you know on average a protein trucks

with five other words for example have
done this exercise many many times and

it's sort of the number you
get from what's known today.

In other words the brain the the number of
expected interactions there is probably in

the range of tens of thousands something
like that so that means that from eighteen

million combinations you could down to
tens of thousands something like that.

So you have three orders of
magnitude of pairs that apparently

in vivo would not see each other or
could not see each other or

it would be really much better for
the cell if they didn't.

Now the problem becomes from
an evolution point of view.

How do you do that.

How do you actually keep
proteins away from each other.

And it might not always be the best
solution to do it from there and

I mean the acid.

You know structural point of view it
might be that in some cases you keep two

proteins away from each other by giving
them two different signal peptides and

sending one to the nucleus and
the other one to the membrane some extent.

Now when I do my path that aspect is now

confusing because I'm actually
taking the two genes and

I'm putting them in one assay where I'm
sending everybody to the nucleus and

I'm testing whether or not they
can reform a transcription factor.

You see so you know that's the point is is
OK so and lastly what we're trying to do

right now is to look at by physical I mean
the by physical interaction that works and

try to look for those pairs that
are very clean technically but

for which there is absolutely no other
times of information available to.

Suggest that they might be real back
in vivo you see where I'm trying to go.

So their genes are not core expressed.

If you knew something about their field.

You don't see any similarity you know so
you go on the other side.

I guess of the distribution of.

Of similarity of parameters long
answer I'm sorry but thank you.