Received University of anyone

Medical University of
South Carolina to go to

war very happy to be here
very long to get here.

Of course and also like to inform
you that he has until five after

the Barclays.

Thank you.

And it's a great pleasure to meet you
as and he said I'm coming from far away

namely across the courtyard here but
I don't think I've ever been in this room.

So that's so it was a good experience.

No indie.

Asked me to talk about the biosystems
institute that we launched just

a few weeks ago and which I'm leading
right now and I said but I want to talk

about science and he said but you should
talk about it and asked So you get the.

Tool for the price of one.

And I try to stay within time anyway.

So here's a little preamble.

So when I was in college
in the Middle Ages.

I studied biology and mathematics and
people laughed at me and

they said you know.

Biology is too complicated
to use mathematics and

by now we know that biology
is too complicated.

Not to use mathematics.

So that it's sort of the idea.

So here's my overview I want to talk
a little bit about meant I want to

engine hearing and challenges and
needs for models and

then about a particular modeling
approach that we have been using for

a long time is called canonical modeling
or biochemical systems theory and

I want to show you some advantages
of that some examples and

then different issues of optimisation and

then in a second talk which
may start when the thing is.

Ten fifty four.

Or I will talk about a little bit.

So I probably don't have to tell you this
because you are in a chemical engineering

department and you know it meant I
want to engineering is but I put it in

from some working group and it's so
it's a new approach to understanding and

using metabolic processes so
most of this is done in microbes and

you want to of course the microbes
into producing something

that they really don't want to produce at
least not in significant quantities and

you want to coerce them to doing
that without killing them.

That's sort of the idea.

So in the past.

Most of metabolic engineering has been
done through motor Genesis and selection.

So you take a strain you select the one
that is doing the best according to your

criteria you improve the medium
you zap the strain and

you select new ones and the question is is
there a more rational way of doing this

and I claim that at least as an aid for

doing that mathematical modeling might be
a good way off approaching the problem.

So here's an example.

It's citric acid citric acid is produced
at a rate of about a million tons per year

million tons per year and you say why but
if you look at your marmalades preserves

all the things contain citric acid because
it's considered among carcinogenic

you can add it to anything with without
asking the F.D.A. or U.S.D.A. or anybody.

So there's a lot of that for
them it's also used as a start.

Material for kinds of chemical reactions
and as a cleaning agents for machinery.

So in one thousand nine hundred fifty
in this final Curry actually started

the large production off citric acid in
the United States and you say so what.

Well it was so important that
the United States became independent

of Europe with respect to such I guess it.

In fact they started exporting it.

So this is sort of a simplified scheme.

Of what's a trick as a metabolism
metabolism looks like in this

fungus as special as Niger and
the task is to reroute things so

that the organism produces more
citric acid than it usually does.

So the big challenge of that
is the complexity that we see

it's easier to show this way against
the complexity in biological systems.

So what does that mean complexity
organizational complexity means

you have a lot of components.

You have a lot of process sees
the processes are non-linear and

you have quantitative
changes in parameters

that can cause qualitative
changes in response so

that sounds obscure we're just think about
going out in the sun and getting a tan.

If you get a little bit more sun all
the sudden there's this invisible

threshold and you get a sunburn which
is a totally different reaction

totally different response of the body
to just a little bit more sun.

In terms of numbers probably
don't have to tell you those

about six thousand genes and
in the yeast thousands of proteins and you

call I we don't know how many proteins we
have it depends how you count but they're

sudden something like five octillion
atoms in the utan human body and I spell

it out because I thought you probably
haven't seen five octillion here it is.

And if we talk when we
talk about these numbers

we have to keep in mind what
a psychologist said in the fifty's.

He said that the human brain can manage
seven plus or minus two items at once.

So if you're young and smart.

Like most of you are I
guess then you can do.

Nine people like me can maybe do five.

I'm happy if I can do five.

But I cannot do five forty million and
I cannot do six thousand at once so

we need to have aids that help
us to deal with these numbers.

So if we look at the whole metabolic map
without even gene regulation without

them I now.

Makes without anything
is just the metabolism.

It's scary.

And so we need help.

And I claim that systems
analysis to the rescue.

OK Photoshop is a nice tool so

Systems Analysis always is based on
a model of the models at the center and

almost if you probably have an idea
of what you'd do with the model

once we have it so we want to understand
what what the system is doing.

We want to extrapolate to new conditions
then we want to manipulate it and

once we can manipulate we want
to optimize the manipulation.

So we want to find drug targets over
them to increase yield of some some

organic compound in these organisms.

So how do we get to that model
that's a different question.

So we need the model structure and
if you do much about it.

Modeling like I'm doing then you get
much from databases like A and E.

course like a metal psych from new
experiments or from the literature and

then you need data and in the olden
days which are still going on.

Most of the doubt.

I came from the literature on you know
enzyme kinetics correct touristic scale

and values of emacs and so on and then you
would put them together in some function

like you because many great law and
put them all together in one big model and

then hope for the best and usually
that didn't really work very well now

we have an alternative if the alternative
comes from what I would call global data.

And there is the Internet of course that
gives us all these wonderful things and

I'm talking about you know my courageous
espec N.M.R. methods and you can

do those as snapshots or what's much more
interesting as time series measurements.

So I will come back to that later but
you have an organism like a bacterium you

let's say you started then you
give it something to eat and

then he can measure non-invasively
how the metabolites change over time.

So that's kind of.

For now obviously there's a lot
of information in these data and

the challenge is to get
that information out.

We can actually if we have good enough
data enough that I we can actually

infer what the structure of the model
should be and my lab is working very much

on this particular on those those
two green arrows right now.

And then of course you
have to do diagnostics and

testing validation of those things.

So there are two types of optimization I
want to talk about one is the optimization

once you have the model the other one
is optimization to get the model and

I would come back to those.

So the ways of setting up
the model could be in different

different forms the easiest
Just like your metric systems.

So you have a little Here's
a little toy sandbox problem.

And also you have metabolites A.B.C.
D. and

then your reactions are one through our
seven and you can formulate a connectivity

matrix that shows what's coming into
a poor Let's get going out of a poor book

and I don't have to explain that you
can probably infer what all that is and

then you can set up a differential
equation as it's here that says

the change in Suppes in the substrate
vector or is giving by the connectivity

matrix multiplied by a vector or
flex values.

So you're fluxes for all these reactions
and if you set that equal to zero.

It means that the system is in a steady
state and you can see if you ignore

the differential party when you
have an ultra brite system and

then you're home free because you can do a
lot with these types of old rock systems.

You can do what that little bit more
advanced like going to Paulson is

doing and I'm sure many of you
know often more have met him.

And he's doing Flex balance analysis
which is essentially the same thing but

it adds constraints and
adds other things to that system.

So the advantages are that you
don't need kinetic details.

So if you have K. M.
values here you don't need them.

You just into typology and
you need to identify flux valueless.

It's a linear system and you know that
any assistance can be sauf with sizes

up to the top thousands
optimisation is straightforward and

people like estimable us and poles and
other people are are using these

types of approaches to optimize
the production of valuable compounds.

The steady state solution is given
by a kernel of this equation and

also they are some tricks to make the
solution unique and in a good to go into

that now the limitations are if you have
kinetic information about the system.

There's no place where you can put it in.

So you get you cannot use your case you
don't you can't put in your inhibition

constants on those things you cannot
do your was known in the area T. S.

as soon as you put in
the mid eighty's in there.

The linear structure goes.

Goes go up is gone and so
you cannot do the analysis.

Because of that you cannot really
put regulatory signals and

they are not persons as you can but it's
cheating because he just puts major cities

in there that sort of change from time
to time and turn on and turn off things.

So that's not real regulation.

But if you have if you look for

optimal strategies for
altering the flux distribution.

Clearly you need to you need regulation.

Because if you change an organism if
you cut something out of an organism.

It's not going to sit there and
not do anything.

It's fighting against you it's trying to

circumvent the problems
you have create it.

So what would you want is
a non-linear formalism

that captures the essence of
biological systems and accounts for

physiological but not a large
trickle process sees us inside its

own lends itself to analysis and
simulation optimization.

And you need some computational
efficiency to make the whole thing work.

So if we work toward.

Such a nonlinear modeling framework
we can go back to the tenets

of our systems analysis so
you say that each component of the system.

Potentially depend on all other
components in the system and

if you want to understand
what the system is doing.

You need to understand what each of the
component is doing inside the system not

by taking it out.

And then the dynamic changes in
that system component driven

by inputs and outputs.

So if we formulate it the way we do it is
we have two types of quantities we have

exercised lots of metabolites in the
different contexts it could be a species

ecological ecological species or could be
genes and then we have processes which we

call a V. Plus the minus now to write
that as a model of easy here it is.

The issue is though.

What are the present I might as
well we know they are functions of

probably a lot of the inside variables and
probably some outside variables.

So what do we know about these functions
outside this this formulation here.

Well they're complicated and the big
problem is where do we get functions from.

Now if you're in physics then
the functions come from theory.

You know you know how Springs behave and

you know a lot about electricity
in optical features and

those things and so you can piece your
models together based on physics and

because that works so well we can
shoot little gadgets up to Mars and

actually land and
spit out a rover that's running around.

Because we have theory known
biology we don't have a theory.

We have educated guesses.

For instance for Growth functions
that we're pretty well but

at a very superficial level.

We have sort of partial theory Mz for

instance an enzyme kinetics and
then we have generic approximations.

So why don't we use true
functions like in physics.

So here's an example the simplest
of the simplest A plus B.

is going to go into a P. prosecuted
by substrate reaction was to products

like glue cause an A.T.P.
going to look or six phosphate and a D.P..

So that looks easy if
you do it right then.

You need to figure in how the enzyme
is getting into the action and

this is from the book by shorts and
shorts as well.

This is one mechanism to produce P. and Q.

of A and B. And if you're right
on every step with mass action.

This is what you get.

Now.

That is a mess.

Now you can say we have become pewters
you know toughen up computing is here.

So we can do this.

Yeah you can solve it but

the problem is not solving these equations
the problem is getting all these memes and

quaffs and constants that you have
to have to bring the thing to life.

And besides many other things
you don't really need.

So if this is for one single
reaction you can imagine what a big.

Metabolic network model would look like so
that not going to work.

Why don't we use linear functions.

Well here's something that's
not metabolic engineering but

let's say you want to model the heartbeat
and you start was a sine function.

So it's it's going up and down pumping and
now here are the sudden you.

You're in a horror movie or
you're running up the stairs and

warm you go into this bigger amplitude and

if you do there was a sine function
of this amplitude States forever.

So you never recover from from
your from your shock in that

horror movie unless you go to a chick
flick and maybe it goes back.

On the other hand if you have a nominee
a system like event up or oscillator.

But see the first you know there's this
baseline here looks almost exactly

the same the portal Basin looks
the same but there's a vent up or

non-linear oscillator comes back almost
immediately to the original oscillation So

it's a stable sustained oscillation.

You cannot do these stable sustain
oscillations linear systems.

So the challenge is the linear process
missions unsuited if we look for

money a function of the infinitely many
out of them and somebody says if you

divide the function into a linear
function and you have functions is like.

Dividing the animal kingdom into
elephants and none elephants.

So we have all kinds of choices in
the Norris of peace no guidelines for

what you're to pick.

So here's a solution was potential Mike so
I'm sure who was in Michigan for

a long time as Noah Davis and I worked
with them some while long time ago

he said well we can take the functions
we have plus the minus and

approximate them and lower US My
coordinate system using Taylor theory and

Taylor of course this has been dead for
a long time.

So what he does must be right and
so that leads to

something that we call canonical
modeling or biochemical system theory.

The result this is this and only a mother
can love it but it is very useful.

So you have the incoming prophecies
you have the X. come out going

process sees all the variables that
affecting the influx are given here

with with exponents that we call kinetic
orders and their the rate constant.

The kinetic or it can be positive or
negative integer and

on into Cheney thing real.

The F. was and made us on the negative.

Not important is that each
term contains only those

variables exactly those variables that
have a direct effect for instance of X.

two has no effect on this particular
term the G.I. two would be zero.

The thing would be you want
to drops out of this product.

OK.

So it's an unusual way of looking at
things but it is a very useful way and

I hope I can convince you
that it really is useful.

So there are twenty formulations
within this by chemical system theory.

One is what I just showed you.

So you take the sum of all these incoming
processes to aggregate them all and

then you do the approximation you do the
same with everything that's going out of

this poor and you do the approximation.

Or you can do each separately and

then you get a sum of all these products
of par functions mathematically able.

So different.

Mathematically they are exactly
the same at some point if your

choices go to the operating point.

And if you go away from it.

You know the two differ a little bit you
know in most cases it doesn't really make

much difference which one you are so
both are very similar in accuracy but

there are differences between
those mathematical formats.

This gives you sort of the doable
size that we can manage these days.

Something like twenty five
metabolites it has to do with

single digit metabolism
the details are not so important.

Something like thirty enzymes
many many parameters and

the values of those parameters
this poor guy of our US year.

You spend three or four or five years
to getting those parameters to gather.

It's a hard job.

It's a hard job.

So there are many applications
of that in pathways fury in

you know dopamine LICHTMAN sense of
this now for biofuel There are papers on

Gene circuitry genome expression doubt
on metalwork engineering and so on.

They're very interesting mathematical
features of these equations but

I don't have time to talk about them.

So one of the features that is
very interesting of these S.

is the models is if you look
at the city state equations.

So he has a differential
equation in the ice of them and

we have internal variables with external
variables that treat exactly the same

except that these external variables
don't have their own equations.

So we said that equal to zero.

And throw it to dot and then we take logs
and you can see log here and you have

log there and the whole thing becomes a
linear system in logarithmic coordinates.

Now this is very amazing because the X.

system format itself is highly non-linear
in fact it is sort of any of that anything

you can write as a differential equation
you can write as an assistant Exactly.

So I don't have time to
talk about that either but

it's if you're curious
come to me afterwards.

So the systems are very very non-linear
but the steady state equation.

A linear.

Now that is very cool because now
you can compute the city states and

you compute features like stability and
sensitivity and houses

is one application you can very easily do
past weight optimization with this system.

So you say if I want to run my
optimisation and I batch conditions for

the cultures and steady state.

I can write a linear program that says.

So even though the money in the kinetics
is everything is taken into account

the only thing is it's happening at
steady state close to steady state.

So maximize the log of the flux for
the log of the variable.

So for some of the vapors citric acid then
I maximize the log of that or I can look

at the flux with which the organ spits
out citric acid into the medium and takes

a lot of that I can form a steady state
conditions it's all in the unlock space.

I could afford me the constraints on the
variables and mock space constraints and

lots of flux.

So everything is a linear system linear so
we can use methods of

operations research and you know that also
applicable to thousands of equations.

That's what a T. and T.

uses to optimize your connection when you
reach out and try to talk to somebody.

And very well understood.

It's very robust and of course it's income
probably faster than the new methods.

So we have done some
applications out there and

then people that you
probably know by name.

Bailey is dead now but Flo DOS was here
some while ago had some money cut this so

they use these types of
optimization schemes for

optimisation of past where structure and
we have done a lot of work for

G M A Models which I told you
alternatives to those things.

So example coming back to
us to trick acid here.

So here's the past weigh in these arrows
sort of indicate home much you have

to change at least theoretically have
to change the endemic to the TS or

gene expression to to optimize
the flux the coming out of the.

And going into the medium.

Is So if you.

If you look at the system it has a barge

twenty enzymes that one could probably
muck around for us twenty genes.

If we allow an optimisation where
you can change every single

of these twenty enzymes or
twenty two actually then you can.

According to a theory according to these

optimizations you can increase the yield
about twelve thought another twelve for

does depend on how tightly
you set your boundaries.

For us.
You want to say well the flux from

here to there cannot be more than that or

a variable cannot be lower in
concentration that a certain level.

Otherwise the organism isn't viable.

So you can increase a twelve volt no
question if you know that's a lot of

work to change twenty enzymes.

If I can only change one enzyme
which one should it be and

if I change it you can pick
any one of those twenty and

you can change it any way you want to up
or down doesn't matter how much can you

improve you know the answer is you
can pick whichever you want to.

It's not going to do anything good for
you.

It's not going to change.

OK.

Now that's a that's a damper So
OK let's do it.

Let's change two or three or you know
enzymes how much you improvement or

we get well if you do two
nothing is happening.

You yield will not increase if you strange
three nothing is going to happen for

just six almost nothing is
happening going up a little bit.

If you want to have at least a three four
year old you need seven that you need to

change now.

Philosophically that's a very interesting
observation because what it means is

the standard techniques
of frenemy to Genesis and

optimizing the medium up to a strain and
so on.

You do that.

Interrupt of really four hundred
years like curry since Curry I should

say if you do that you find
all the simple solutions.

So combinations of one or two or three.

There.

Just give you the right set up.

You can find by you know random.

But if you imagine rolling seven dice at
once you don't find that by coincidence.

So here we can predict what you could do
potentially and what is not going to work.

So that's going to call.

So now I want to talk about the other
optimization of my lab is working in that

quite intensely and

the other up to me is parameter
estimation from time series data.

No you go with your problem to a computer
scientist and you feel like you're with

a bicycle at a four way stop and
the other guy has an F three fifty and

he's coming in a minute later and
he's just going to go off.

I mean what are you going to do so

that's the impression you get when
you talk with computer scientists.

So they say well.

We have solved this problem
in the forty's and fifty's.

You know there are many
algorithms out there.

So that many methods.

Now I have not yet

seen a computer scientist come back to
me with a solution to our problems.

So they say good.

Here my data find me a solution and
then there is this black hole for

computer scientists who lie and
they never come back out.

OK which may be good.

So there are many methods out there.

If you look at B.S.T. estimation of B.S.D.

there are about hundred
papers since two thousand.

Trying to optimize parameter values from
from time series data hundred papers.

All of them say they work all of them
do work for the case that in the paper.

Almost all of them don't
work if you change the data

in the example a little bit.

None of them works if you
change the data a lot.

That's a situation.

Now if you do a brute force and
this fellow Kikuchi

believe it was the Japanese to do
what they did it was great force.

Fyfe able system.

Try to estimate parameters.

It took seventy hours on
a cluster of one thousand P.C.'s.

Now there are seventy
thousand hours of P.C.

time now that is not a happy situation.

OK so
the time to conversions is a problem.

Problems with current colinear data
are noisy doubt our problems with models

permitting redundancies and
compensation of error among terms.

And since I got so many slides that.

Let me tell you what we
are doing now one thing we do

is that we replace
the differential equations for

the differentials in the differential
equations with estimated slopes.

So we take our time series that we
estimate slopes at all the different time

points and we substitute the slopes for
the differentials what that does is

each differential equation gets
substituted by K. algebraic equations.

So there were no you have K.
times an equation that's even worse.

No it isn't we have clocked how much
time algorithms like today goggles and

so regression algorithm
spend on optimizing

these types of equations and
which part of the process.

And I don't we don't have time to play
a guessing game but they spent something

like ninety five percent on
integrating integrating the equations.

If the equations are nasty
stiff in the in in jargon.

If they are stiff The other
than spend ninety nine.

If not cloth almost one hundred
percent on integrating equations.

And if you have a system that's not stiff.

That's not nasty and
you do these random searches for

parameters you are bound to run into
combinations that make the system stiff.

So even if you have a good
system well behaved system.

The other is most fail very very very
often because they become stiff and

then the algorithms keep
on keep on doing that.

So we do this.

Estimation and
slopes of situation and that

cuts down the time from seventy thousand
hours to about one hour on a P.C..

So that's a good improvement and
we have done some other things.

No one one issue I want to talk about
is compensation of error among terms.

So what does that mean for
you as an example.

It doesn't really matter.

It's a made up example.

So it's in G.M.A.
form here's here's my little pass way so

there's one one branch point in
this a feedback and a feedback and

if you feed forward activation of some
activation you know this process here and

so I created a model off that and
used that model as data.

And then I look for the solutions that
would give me a decent solution cities and

fits and if you look at this.

And you imagine that would be
experimental that I would say you know

this this is a good fit.

You know this is a good fit.

But it has two were wrong terms
not just one wrong term or

you know we start it was one long term and

then it's compensating all
with all the other terms.

So everything is a little bit wrong.

Now if you are just one to look at it.

Look at this or put it in a picture
frame you can say well this is OK but

if you want to extrapolate this
situation out to something

like doubling the input exit row here.

So it was such a different situation or
if you say what happens if there is

no activation or what happens if I take
out some of the X. to that type of X.

or palatial if you do that
then you are in big trouble.

And here's an example.

So scale is different and
gives you the true system is

behaving like the dots the model which
which smoked very very promising.

Here read it as a bad job
especially in the beginning.

And it comes from compensation of heiress.

So if you look at these hundred page.

Both of us cited zero of them
including ours are looking

into these compensation issues so
that is bad so because it's bad.

We did something about it.

We come came up with something
it's called Dynamic flux

estimation just got published
in bioinformatics and

it was inspired by stoichiometric analysis
but it was extended to time courses.

So we study the flux
bands at each time point.

And of course of balances
not in equals out but

in equals Plus what's
the change in that variable S.

So it's a different system
than we have before.

So the changing of A but at time T. is all
in flux us at T. minus all F.X. as a T..

So if you if you say this
various not changing it.

T. then you go back to a strike in
metric and Nastase it in houses.

Now even if you do this if you set it up
like that you get a linear system and

we saw that as far as possible and
the result is that

we get numerical values of
Fluxus at all time points.

Now we don't get functions out of it.

We just get what the flux
values individual

flux values look like at each time point.

And then we can go along to go ahead and

represent these large fluxes with
appropriate functions appropriate models.

So this if thing consists of two phases
the first is totally model free.

The second one is model based so
the model for

a restart was a time series we do some
technical stuff like smoothing and stuff

like that we estimate the slopes from the
data and that can be done automatically.

If we nor the system topology
we can set up the system

of fluxes which flux is going to this
pull which ones come out of this poor and

we list them all and then we do so
many are brought in and

put you in conditions and we are working
on an opportunity conditions right now but

I'm not put your conditions
we get a dynamic flow.

It's profile that says he darts

where you have fluxes versus time of
Fluxus versus the contributing variables.

Now you leave the model free
estimation face and you can make

assumptions of what the function should
look like that is crybabies Fluxus So

if you like because men who can do that if
you like Palm lost by coming system theory

you do use that if you like not models or
something else you can do that.

You fit that into the symbolic flux
representation in the you can do some

parameter estimates can you get numerical
flux representations that are in

a functional form and from that you get
a fully parameterized he medic model.

So let me show you an example.

So this is an example that has to do with
black cause as in luck to caucus luck this

is the little blue dots year luck
to caucus as I'm have been told for

many years I showed a wrong picture
that was an Arctic office but

I'm not doing that anymore so.

When I did I got it from some
websites as like to caucus and

somebody wrote to me says you
are using my picture and it's wrong.

So OK OK.

He says for a fee I will give you
a better picture I said no thank you.

And I could then make I don't have money.

Anyway this picture is involved in in your
words Jesus dairy products of wine bread

pickle all those the things you like to
eat and we were interested in regulation.

After a car service and we're working
with people in Portugal who are producing

really cool dad up and the goal is to
understand how this past was regulated and

even though it's played causing It's
a yeah you know I've been in the book for

fifty years one hundred years
just a question about regulation.

We want to extrapolate to new
situations for instance not also or

expression studies or other things
they want to max may yield and so

what's the thing is doing is taking Group
course and converting it into lactate and

by doing that the medium becomes sour and
become so sour that the tape.

If you're over it becomes so
solid that other critters like equal and

struck the carcass don't like it so
they don't go into your bird and

because of that you can keep the your
work longer than if you didn't have

the lactate and they're OK so
that's that sort of the idea.

But they're secondary products
that improve the creaminess so

that your grower and
the texture and all that eventually

they would like to have a solar wind
model but we're far from that I think.

So your typical back up these
are produced was in vivo N.M.R.

measurements and
if you're really interested in that I

have a postmark right now from the lab in
my lab so you can tell you all the ins and

outs of how that is done it's really
cool so here the group was in

the medium that's of course easy
Local six phosphate for Plus this.

Phosphate A.T.P. other things you

can see in some cases the cursor various
most of the board every thirty seconds.

In other cases you have a lot of
fluctuation here and that's probably

because these two are in very very fast
it with every of these two reactions so

we have some problems
with calling near eighty.

Anyway so we've been playing with
these data for quite a while.

And we had actually fitted
these data before and

first we didn't find any solutions and
then we were happy for

a moment and then we found all kinds of
solution there were really different.

And he said How can that be.

Well it's because of compensation
compensation of error between terms.

So that was our problem so
every time we want to extrapolate to

a different set of conditions the whole
thing blew up and said This is not good.

So here's our dynamic fax estimation
Here's the flux years sort of the model.

So in contrast to high organism that use
A.T.P. they actually use this paper.

S. as a fast food donor and

that produces glucose X. phosphate and
it has a side effect that you produce

pirated immediately and that goes into
lactate you start souring the meeting.

Medium very very quickly.

So here's our first attempt with
with idea that smoothed out

there were generated artificially and
you can see the fit so good.

The the individual flux
profiles are done very well.

So that's all good.

You can do that.

It's sort of tolerant to noise so
we did that in that we're working.

But what's really
important is this one here.

So here we knew that
the glue causes this estate

uptake function but we pretend that
that it was something like this

it's something like an exponential decay
functional style and we did the analysis.

In all the fluxes came
out OK like this one and

you see the target flux in the fifth
century the same except for this one.

So here you see clearly a very very big
difference of Holder's flux depends on

Fed and glucose Well you say so
what where this is a big deal because

it allows us to diagnose
assumptions that we had

an other people had automatically put into
these models and had no way of checking.

Now we can check them we can
diagnose wrong assumptions.

So it's really cool.

Weapon problems of course this will come.

Convergence issues with doing mask
on the vacation didn't have time

to go into that in many cases of course
these fluxes are under determined just so

we have fewer time courses
that we have measured and

then we have fluxes that are in the system
so we have an under determined system and

we are working right now on methods for
combining this with other

matter of the laws to true and
lies those data.

There are big big issues
with data that are colinear.

So they behave sort of the same.

Or they are very very.

A quick Librium and
it will take some really good stead of

stations who can get out of their
little silo of linear analysis and

major season regression analysis into
something that's dynamic and the many or

so if anybody if you're a statistician you
can come and I'll give you a job or so.

And then there are models and
then just mathematically allow different

solutions and we have worked on that this
is a source for sleep it's not a lie.

It's so physically from Norway one of
the two Norwegians in mathematics who have

become famous able being the other.

So those are problems.

So off here the summary
of my technical talk.

I'm doing OK on time.

OK.

So biology has become too complicated.

Not to use math and
the reasons I hope were clear to you.

It's just the number of things we're
talking about twenty five thousand

genes in the human body and something
like probably two hundred fifty thousand

proteins and nobody has any clue how many
minutes more metabolites we have it's

just a lot of stuff to keep track off and
that alone is not sufficient because

now all the things are interacting in
the into act in non linear fashion.

Sometimes you have competing effects this
wants to increases this wants to decrease

is which one is going to win to what
degree is the one going to win and so on.

So the non-linear interactions
are really our brain is not made for

non-linear interactions at
least on in this context.

So really the only solution.

In the long run is going to
be a computational model.

And how you get it.

It's a matter of taste.

What you do with it.

It's a matter of strategy and
technique but

there is a model that
needs to be involved.

Otherwise you cannot do it.

The nature of nature does not
provide us with guidelines for

what model we should use and that.

Gives us some.

Some freedom and it gets us contests and

scientific fights and
other things as part of the deal.

So the Infinity infinitely
many choices and

he's canonical models that I showed you
are not the solution to every problem on

earth even though we would
like to think so but

I admit that it's not the case but they
are good pro they're good defaults and

if you don't know how to get started
this really helps you to get started.

Because you can set up your diagram
of a biological diagram and

it's made it's an automated procedure.

So as to set up the equations.

We have a computer program
that does it for you.

So the adventures are that you
know exactly what parameters mean.

So effect of this on that or
turnover rates.

There's not nothing else for computation

associate with the steady state refute
the system form become very simple and

indicate that with optimization where
you can really solve a big systems.

Now there are some problems with it
to what in principle it's a great

great way of doing it.

So optimization is facilitated by
the model structure and optimization means

both something like year old and
the identification of parameters.

So I don't enjoy it because now
the second part is going to start.

OK So this is my current crew and they're
all over there in that building and on.

And I'll be happy to send you information
or you can Google me and stuff.

All right now integrated by a system
in situ and this is much shorter.

So it's bear with me.

So we call it C..

Integrative Biosystems Institute.

I don't know who a few looks at
these research arise and but.

I think in the last issue actually summer
two thousand and eight they had a long

article about different activities in
the center and the institute and so on.

So having the problems your little
puzzle pieces and the computers and

the drugs of the D.N.A.
and all those things play nicely together.

We also have a website.

And I just.

Dot edu so that should be easy to do.

So there's a lot of information out there.

So here's a history in the history
starts with two thousand and

four because that's when I came and
everything before and I'm not.

So in two thousand and four.

I came in they put me on a committee
with all these esteemed people here.

And they said sure.

Georgia Tech do something
in computational biology or

structural biology or
systems biology or integrative.

And didn't you read enough.

And of course they were said yes.

So there was good and
then in two thousand and

five the three deans sure store and

get in and de Mello at the time
said OK here's your committee.

You should tell us what such
an entity should look like.

And they asked us to
write a white paper and

they had expected something like fifteen
pages maybe but with German thornless

we wrote a white paper that was
one hundred twenty three pages and

nobody read it but it made an impression
because it has a lot of weight.

OK.

So if you see doorstopper
us all over campus.

That's what it is.

So anyway we were so

impressed by the quantity if not
the quality that they say good.

Now you have to write an implementation
plan so we got a new committee.

And there was Richard Fujimoto who
was a Computer Science years from

biology and myself.

All right so
in these activities we did surveys and

we had many meetings we
talk with a lot of people.

Nobody's really concrete ideas but
he says you know there's good.

We should do that.

And so the vision was and

is that Georgia Tech would be nationally
recognized internationally recognized

as a leading institution for research and
study of integrative biological systems.

And clearly we need
a situation wide participation.

It's not just biology and
it's not just computer sciences.

It really needs to have
cross disciplinary linking.

We clearly need to leverage the strengths
in the lots of strengths at Georgia Tech

that are relevant.

We need to create a physical not just a
virtual institute and we've been promised

a significant part of the building that's
to be built on Atlantica tenth Street and

on whatever the recent unpleasantness
of the economy is letting up.

I guess we will be back in plant and

we are encouraged to do strategic
hiring in relevant areas.

So the courts are to create such a
learning environment research environment

consolidate all the strings.

And then to develop a graduate program
interior of interview with BAE Systems.

Why do we do it.

Why do we do it now.

Well if you if you look at the paradigm
of biology over the last fifty or

one hundred years.

It's what's called reductionist So

the idea is if you want to
understand how a heart works.

You need to take it apart and

then you need to study the valves and the
tissue us and if you want to understand

how they work you need to take the tissues
apart and you look at the cells.

If you want to understand how the cell
works you have to take them apart.

So you're taking apart a part a part a
part until you down at the building blocks

and the idea was well once we know the
building blocks we can put together and

we understand what's going on.

Well this last part.

Which is conveniently ignored in many
cases is the hard part because we still

don't know how we can put all these pieces
back together into functioning entities.

So this reductionist
reductionist approach has to be

complemented with something like
a Reconstructionist approach where you put

pieces back together and
that requires modeling.

So the opportunities there because many of
the things you know Georgia Tech is really

strong in you know simulation
pattern recognition visualization.

We have some really cool people who
are tagging molecular tagging the so on.

So if we look at biology in general
there's sort of data collection.

And that's the good old fashioned
biology that's still going on that

should be going on.

I'm not minimizing this at all that goes
from observation to quantitative data.

And then a bioinformatics came along
about fifteen or twenty years and

then organizes these data are especially
the high throughput data.

And so from quantitative data you
go to information what does it

mean what's clustered together.

What's causing what next one is my
chemical biological systems analysis

that's where my lab is and many other
labs here and that's where the focus of

this is it was going to be and that
goes from information to understanding.

So we have all these pieces of an
astounding how can we put them together so

we really know what.

What is happening in this organism and
then this is just starting a synthetic

biology where you want to reconstruct
biological systems from the bottom up.

So you want to go into innovation and

create little biological systems that do
a job that is difficult to do otherwise.

So if you take these pieces
it goes on observation

to quantitative information and
the standing to innovation.

So that for me is integrative biosystems.

Or systems biology.

So all the the big question that has.

Come up all the time is what's in it for
G.T. for take what's in it for you.

Steak or.

Because I could I mix I like cats and you
know you can hardly tell them what to do

if you don't believe it as your chair he
tries to tell his people what to do and

they really do it unless there's
something in it for them right.

And that's human nature.

So what's in it for everybody is
that the prospect of solving gram

challenge problems in biology in
medicine that nobody by himself or

herself is able to solve because
the solution require complex and

flexible teams that are to
disciplinary interdisciplinary and

cross disciplinary if
you want to be picky.

So the guiding themes
are multi level assessments.

You know at this point.

Probably ninety nine percent
of all analyses are one level.

Maybe the genome never maybe physiological
level or maybe the metabolic level.

We need to combine those levels because
they are carbide they work together.

We need to develop molecular inventories.

So we need to know what all the little
blocks are that nature has as at

its disposal to solve these
complicated tests and then we need and

able to technologies that are suited for
systemic approaches and

so again low to understanding to creating
a pathway systems in some sense.

So when we have decided to begin
with three streams are sure like

one of the development from
the normal SOS to work cancer.

One has to do with sensory microbial
system the environment and

there's a lot of interest in that
because infectious diseases and

sustainable environments cleaning up
the environment those things in sort of

support it is supporting the both
high performance computing and

that's where computing is in this mix.

So we have these multi-level level
a system of course in a general.

If you like his rate systems
we need modeling simulation

molecular inventories and
then devices at the engineering part

points of the creation
of tagging devices or

methods sensors visualization two worlds
in those all the enabling technologies.

So the implementation has begun.

We haven't executive committee and
there consists of

these three folks here I'm sharing that
right now we have an advisory board

with high ranking Georgia Tech faculty
like your own Professor Bill Maher Yes.

We have a rotating directorship So
we've done by the end of next year.

And the hiring would be through
schools but other entities and

if there would be involved in that.

And if anybody has good ideas for

hiring superstars from outside the process
come to me and I will find money for you.

In terms of research we have funds for C.
projects and if you're a member and

you can work with somebody from
a different unit that's a you are here and

you find someone in biology or in physics
to work on some interesting project

then you can apply for that as a faculty
there funds for shared graduate students.

So the graduate student has at least
two mentors into different units

that may be a blessing and maybe a curse.

It's probably both.

If you are a student.

But that is supposed to trigger
a sort of collaboration because

students usually don't go away.

So you have to deal with them.

Then we have a distinguished
seminar serious and

we had a speaker today Dr Ryan it's us or
there is every So

we have that eventually we will have
this Biosystems building the name

is not quite decided yet but we made a big

advance by getting into
the capital campaign.

And the building into
the Capitol campaign.

So for those of you who don't know what
that means is if somebody comes and

says thirty thousand or

thirty million dollars to spend what
should I do with it then we are there.

OK.

If you have thirty
million dollars to spend.

I'm here.

OK come forward.

We have a Chalk Talk series that
happens every Wednesday at noon.

It's an informal setting

where people usually from inside
Georgia Tech talk about their research.

And in particular about open research
questions we encourage speakers not

to use Powerpoint which changes
the dynamics dramatically.

So you can ask they scribble
on the whiteboard and

thousands you can't read it.

They have to explain it.

And so
there is there's a lot of interaction.

We sort of food capital F O O D.

So if you're a student
on Wednesday free food.

It's in the clouds building.

OK.

Catholic I mean I talked about that
we have some come committees we have

a website of course.

We had some events where we start was
a brainstorming workshop or shot or lawn.

We like to do things in style.

We had a poster workshop where we had
something like hundred twenty posters and

two hundred people showing interesting and
we had featured speakers including your

professor Lou and we had an announcement
celebration in February this year and

then we had our international launch
conference just three weeks ago in there.

We had really really good speakers.

So you missed it.

It was great at about two hundred fifty
in ten days we had a good poster session

it all happened.

So tell we had dinner in
the ballroom of the aquarium.

And it was it was announced plenty
fully so she didn't come to bed.

So work in progress.

So we are working on
gradual post-doc program.

We would like to have an Dolman to where
we have some money that we can we can have

posts that are not associated
with one faculty member but

this sort of floating post box for
you to work on training runs.

Building this building a city G.
partnerships and

all then of course we need to work
towards sustainability so sustainability.

On this or money stuff but you know your
thirty million will be put to good use.

If you have them.

So it's a great opportunity.

There's a lot of buzz.

All the way from the president of the
provost to the international faculty all

across campus and we want to
integrate that buzz of our goal.

So all almost made it.

In that time so I thank you very much.

And

that's a great question and

my students always mock me because
my standard answer is It depends.

And that's my answer here.

It depends on for instance how good the
data are if the data are very noisy you're

not going to you're not
going to diagnose anything.

If the data are very clean and you see
systematic deviations from those data.

You may be able to figure out
that there's some time delay and

you say well as time delay coming from.

Well the most likely one is that there's
something in between that is stealing

this reaction.

So I'm sure that to some degree.

You can diagnose other things that
are not just drugs or presentations and

I'm sure that them many things
you would not be able to find.

So this is all at the beginning and
you know maybe in three or

four years I can give
you a better answer but

this is sort of I think partially yes and
partially No At this point.

Yeah.

We're doing that right now right now.

If things work.

You don't know if things
don't work you read on.

Now it's so if you go to fit a sort of OK.

You don't know whether it's wrong
assumptions for the for the functions or

whether it happened suddenly something
that was screwed up the experiment.

Or it's a combination of they're off
especially if you have noisy data.

It's very difficult to find out
what the reason is no the solution

of course is that you have replica data
and eventually that will be coming and

so we're getting a big data
set of metabolic profiles.

I'm almost almost scared.

How big.
There's going to be from the group in

Japan who have a lot of machinery and
a lot of students who are working there

twenty four hours a day and so
that would be a very good test bed because

most of the things we have done on this
very very small system where we know a lot

about the topology and some about the
regulation and we have really good data.

So as soon as any of these
ideal assumptions deteriorate.

You know problems are lurking.

And we know that if.

Yeah.
So in the ideal case you get a specific

number you have to increase this and
them to you by three point five four and

where you have to decrease the security
by you know to eighty percent

that sort of the solution that comes
out of that and else a good question is

what happens if you cannot get three point
five four and you have to do a four fold.

And then we know that that the theoretical
maximum will not be reached.

How much it deteriorates
depends on the special case.

Now there are also assumptions
about how much you can allow

Fluxus to deviate from normal and
very very well from normal and

most of these boundaries
are set by your own.

Little feel guessing educated

biologists who say well you know
if it's more than twenty fold.

There's not going to work itself.

It may be able to.

It may be possible to get a better
solution in the theoretical maximum but

than you violate some of these
constraints that you put on.

In the case of optimization or
in case of Premier estimation.

With the optimization.

If everything is working correctly and

ideally and everything is good life is
good and you set it up as an assistant.

The size doesn't mean a difference
because a linear system so you can do it.

Now the trouble comes in with
getting the right primitive ideas.

And so if you don't have
the right primitive values then

you know you on the on and the question is
how do you get those so it comes through

comes back to the primer estimation issue
but with the with this as just a model.

It's becomes a linear system and
then you only have

feeling your objective function that you
can muck around with and you constraints.

And that's under your control.

So you can play around with it.

It's very fast.

It depends it depends on what you
have to cross validate your data.

So if one doubts that
with fifty doubt up ones

when you take out twenty five of them.

And you put them in a drawer for
later use.

It's not going to help you.

Because the number of dead up points
in these small time profiles as.

Really you need enough so
that you know what the profile looks like.

So if you take those are using for
for validation it's not going to work.

If you have a data set that is
produced on the different conditions

that's a different substrate.

That would help to some degree it
will help by telling you all wrong.

What you've got with your
model is not working.

Period.

It does not tell you what's not working.

Really.

So all then you can go and
say well let me fit both together and

see what they can find something
like that but we have done that and

it's really not a not a pretty picture.

Because you just them all of the arrows
are with a you have compensation

was composition with this with an equation
of between all these equations.

So this if this really works and
in a stable fashion.

It's much.

It's a big step forward.

I think.