OK I want to start with a story.

So I graduated ninety nine in so you
don't really know a whole lot about me so

I figured I'd add a little story and.

I had to make a decision
because several of you.

I know are in here maybe thinking
about an academic position and so for

me the decision was a post-doc.

Or an assistant professor position.

And you know I didn't know which

way to choose in I ended up
working a little bit with you.

I reckon Steen and I'm sure several of
the faculty it know or have heard of you.

I reckon Steen and

Dr reckon Steen everything that you
did he did about thirty years ago.

That's if his seminar if his presence
at seminars are any indication but

Dr reckon Steen.

So I worked with him and you know I
got to talk to me a little bit and

when I was the signing on schools.

He gave me advice and he basically said
well you can be a small fish in a big

sea these are his words a small fish in
a big sea or a big fish in a small city.

And then he followed up with discussing
about how you can effect change at

a smaller institution.

And that kind of stuck with me.

And so when I had the opportunity
to make a decision about.

Tennessee attack.

It was really a school that I had
an opportunity to effect change at and

that was really important for me because.

Maybe I'm the type of person
that sees something and

thinks well I know how to
do it better than that and.

So I kind of saw that in
the Tennessee Tech program so

when I started out at Tennessee Tech.

You know and
I just want you to to think about you

know how you can effect change so start
at Tennessee Tech my startup package.

Was a computer and a comfortable chair.

And literally.

And I am telling you that
because right now and I want.

I C H E meaning that year and
I was the only faculty member.

There was no graduate.

No graduate students no undergraduate
students at the meeting just myself.

From Tennessee Tacon everyone asked
me where is Tennessee Tech and

I say well it's halfway between Nashville
Knoxville along I forty to the point I

put a figure of Tennessee up and
show work for Tennessee Texas.

However you know last year all of our
faculty went to the meeting we had

thirty some odd contributions.

You know a big chunk of our graduate
students went there about thirty

undergraduate students so we really have
now a much larger presence at A I C H E.

And I think at some level I have it here
to Dr reckons Deans commentary about

you know kind of making an impact and I do
I have a lot invested in Tennessee Tech.

So I'm very very proud of that.

OK so my summer opportunities.

So without I don't have
a starter package and

so you know I want something
to do during the summer.

Hopefully that will catch a.

And so I started to send out.

Information to different national
labs in grand helpful thing or.

Sandia ended up contacting me and a Grant

said OK well you have this background in
molecular simulations we need someone.

How we have this larger algorithm.

To calculate to do some drug design we
need someone to calculate free energy of.

Lincoln's in binding pockets.

It's like OK you know
that's within my skill set.

I can do that.

So I'd be working with John
the fella who was a researcher.

At Sandia and so I show up in
Albuquerque and John Lewis and there

he actually just transferred to Livermore
about two weeks before that time.

And by the way they're not step or docking
for energies part of a larger algorithm.

They're actually one step one.

Which is chemical graph theory and
you know.

And so

I started out at step one in an area I
didn't really know anything about and

not work working with someone that wasn't
physically where I needed him to be so I.

You know this and
that's kind of a blessing.

Because oftentimes if you look at
something with unique perspective

you're not limited by what you already
know about a particular field so.

With that.
That's how this topic of computer aided

molecular design comes into play because
my background is in molecular simulations.

So this is you know working
on something totally new.

This seminar.

You know I never got a letter for
a seminar for

use in Taishan That's the first half or so
of the presentation should be an overview

or tutorial on your subject and I thought
there was very interesting and I wanted to

adhere to that because I've been in
seminars where I didn't really understand

the whole lot even though I'm the person
that sits where you are right there.

Trying to understand and
see the presentation clearly so

I'm going to here to this for
the first half of the Senate.

OK So what is Q Sorry so.

Q sort of quantitative structure to
relationships there quantitative in that

they're not trends things go up or things
go down you actually have values so.

That's why it's quantitative there's
a structure activity meaning

it's a structure of the molecule with
correlates to a certain activity change

and it hopefully model form so

this is quantitative structure
activity relationship that's what the.

That's what the definition of that is.

So the origin.

Any eighteen hundreds there is structure
activity relationship identified here.

And really the beginning of quantitative
structure activity relationships

can trace their.

Getting to about nine hundred sixty
two with Panshin coworkers and

really extensions on what they've done.

You know a little bit after
nine hundred sixty two.

So this is really the beginning

of quantitative structure
activity relationships.

So what about cues are right now so Q So

right now if you want to find a Q.
sorry if you want to publish a Q. so our.

You know if you look right now in
journalism in this no chemistry or

sarin Q. so our environmental
research there is a this.

Q So
uncommon Tauriel science which next year

is becoming molecular informatics and
I think that's.

Not only a change in the focus but

also kind of getting away from the work
you start and finally this journal of.

Chemical information and modeling
it used to be Journal of chemical.

Journal computer information.

Journal of chemical commission appears
science and it's split off in the Q.

journals one of which is this and

they're really well much
less focused on choose are.

To the point where this article came
out last year's Q star dead or alive.

And you know it's a perspective
from an industrial

person in the pharmaceutical industry.

About whether people are actually using
Q SARS in the way that they should be use.

So let's cue star one about quantitative
structure property relationships.

So quantitative structure
property relationships.

The way I like to think about
it is well it's just like a Q.

Star except it's generic so Q.S. P.R.

the property brick USA ours activity but
the property for Q.S.T. ours.

Whatever property you have me interested.

Now you know them.

You've likely used them but

maybe you haven't really
thought about them in that way.

And so if you look at say.

Plodding the normal boiling point for

normal alkane that's a function of carbon
number you can correlate that fairly well

linearly and so
this isn't a very simplistic example

of a quantitative structure activity
relationship and if you so desire.

You can pull off the properties
of gases and liquids and

really that's just a bunch of quantitative
structure property relationships.

This book and

sure I saw it on some bookshelves today
during my visit it's building the library.

So what about to espie aren't
chemical engineering it's not dead.

How much is it not dead.

So here's an article from this year in
a dusty engineer in chemistry research.

I'm predicting up refine ability
limits of organic compounds for

molecular structures and
you can see right there.

Q S.P.R. study.

Same year and
early point temperatures hewas P.R..

Another Q S.P.R. study same year.

And here talks about a molecular
descriptor which I'll mention momentarily.

But it's really another Q. S.P.R.

study so these are all four from two
thousand and nine from the same journal.

And this is just a handful
of several more so.

Q. S.P.R.
in chemical engineering is not dead.

However maybe people aren't
as familiar with with that as

as you might anticipate I've talked
to people when I've described this

technique to them and I'd mention
Well do you know what a Q. is P.R.

is and I get a head shake No What is
that and I have to explain it and

then they would understand it but
it's certainly not that in our field.

OK So molecular descriptors I kind
of mentioned little bit about

the previous one had the previous
paper talked about it.

I.Q.'s They are Q.S.

P.R.'s and by the way I'm going to use
this this framework interchangeably.

I might say to us they are not talking
about a property other than activity.

You can forgive me for
that these can be linear or

non-linear relationships the pen.

What you trying to do the property of
interest is your dependent variable and

the independent variables are related
to the structure of the molecules and.

Those independent variables
are called Molecular descriptors.

There's literally a cottage industry.

Of the finding and using molecular
descriptors and hewas P.R..

So there's thousands of these types
of descriptors that are available.

And a popular one that you might have
heard of would be the connectivity index

but you can use the lowest unoccupied
molecular orbital molecular weight shape

in the seas etc And
here's a couple handbooks to tell you

to kind of describe how many
descriptors there are available.

Ultimately their goal is to discriminate.

You wouldn't want to use molecular weight
if you're trying to discriminate for

example.

Amongst isomers So a structural changes
should give rise to differing values of

a particular descriptor so those are what
molecular descriptors are and what their.

What their goal is so I mentioned.

Connectivity index and

even though this is not a major
part of what I do research wise.

You might see this next week at the C.H.P.
meaning or if you've read papers on.

Quantitative structure
proper relationship or

anyone that's talked about connectivity
index I just want to introduce this here.

It's a popular top alogical index.

It's basically an operator on a two D.
Graph So you take a molecule and

the atoms are nodes and
the bonds are edges near to the graph and

so on top a logical index is an operator
on that graph and it gives you a value.

A connectivity index.

It's a counterfeit.

Given type of subgraphs you make subgraphs
within your larger chemical graph

of the molecule and it's weighted by
a function associated with the number of

skeletal neighbors and different subgroups
are associated with the path length.

Within that subgraph and

you denote that as order and so quick
example would be three methyl painting.

So three methyl ten pain if you look
at these Delta values the value in this

the hydrogen suppressed graph this and
tell you about the nearest that these

bonds are connections rather that are not
hydrogen so here this carbon atom has only

one connection here that's not hydrogen
this carbon has three this carbon has one.

It's set and so
you get these Delta values.

And so it's functionalized here.

And then your connectivity index is just
the sum of this functional zation and

just show a little example here.

And so they cause i and
it's pronounced the zeroth order chi and

extra methyl three methyl Pentium
is four point nine nine one.

And you can develop a similar
relationship for this path of one.

So here's one two three four five
there's five paths we get five here.

And the connectivity index that
order one is given Heron So

this values two point eight zero eight etc
You could see how you might do this for

order to order three etc So this is
a sample of a type of top logical index.

OK so how are Q.S.A.
are skewers P.R. is used so

you use the model to relate dependent
variable to some independent variables and

so the value or the sign of a regression
coefficient that happens to make it into

the model potentially can provide insights
for researchers so you might have a.

A large number of
descriptors to choose from.

You do a statistical analysis
to determine which one goes or

which ones end up into your model and

the values of those regression
coefficients for example provide insights

for researchers and they might allow them
to test conjectures that are made or

really give new design direction so
you might run into star and

and depending on what goes into your model
might give you an idea about how to modify

your training set based on the cue star
in order to see improved activity.

However you can do this in a little bit
in really an approved rational approach.

So a more rational approach
to molecular design is

called Computer Aided micro design and so
that's what I'm going to talk about here.

So the community computer aided most
of design has some general steps and

these are certainly General So the first
part is this selection of fragments or

groups and it's either defined by the set.

You're working with or you select
a pre the find number of groups.

The second step is really making
molecules and now you put those

fragments together in a certain way based
on balance arguments or other types of.

Constraints you put them together in
a certain way to make a molecule and

then ultimately you have
to evaluate the fitness.

And that molecule in some way
that you've made and then

the pending on the algorithm you might
have a structure generation step here.

So that's really the general steps.

On the Origin is from one
nine hundred eighty three.

Really.

This comes from a gun in his lab and
Danny is still.

Working on this.

So this is one nine hundred
eighty three and his work.

Looked at using functional groups.

From your a fact in a generate and

test approach would generate
meaning that you generate.

Structures and then you test
the fitness of those structures so

it's called the generating test approach.

And he used this for solvent selection.

Really suffered greatly from
carbon authorial explosion so

you could consider or you can think about.

Well if I have a lot of groups and

I want to put them together in a wide
variety of ways this could quickly get out

of hand the number of combinations
that that you could come up with.

And so more rational strategies were
needed to limit this type of carbon

the toilet flush.

And really I mean he

has a technique that he still uses
today that's basically has its

framework in the same general approach
it's still generating test technique.

It has many it's more complex.

It has more steps to it but

it's still similar to this generation
test technique there was a paper.

On polymer design on this.

That was published this year so

it's really a technique that's still being
used in computer and molecular design.

And so what he does is he uses
group contribution techniques to

score the fitness of the molecules.

And I want to take an aside
a little bit to describe for

those that that are in for
me with contribution techniques so

groups are fragments of a molecule and
they're looked at in two dimensions.

And so.

His method Petone and
you have four groups.

Each group will contribute
something towards a property of

the molecule and
three the groups are unique.

One group is repeated group
contribution technique and I kind of.

Give you a little example there and
by way of comparison.

The formula gives you three hundred forty
five in the experiment gives you three

hundred fifty three.

So there's a simple example of your
contribution technique that's used to rank

the fitness in the generating test
approach that there's other approaches.

You can envision computer
aided molecular design being

approached as an optimisation way and so
you would formulate really entire problem

as an optimization
problem with constraints.

And this was introduced really
by prominent around us.

In ninety eight and
they had the objective function is for

whatever property are interested in and

it's right in terms of those connectivity
and the seas that I spoke about earlier.

Their contribution was rewriting these
connectivity indices in terms of variables

related to a partitioned adjacency
matrix partition adjacency matrix

tells how each of those
groups are connected.

Which ones are connected to which
what's the binding is single binding

double binding etc So they saw
the saw as an optimization problem.

Because then they can use optimization
techniques techniques from optimization

theory in the solution of these problems.

However you really needed templating
with called templating to make

efficient progress.

And so what is template so

templating is when part of
the structure is fixed up for you.

Ari.

And so you have an idea of what you
want the structure to look like.

And so then you would design
around a certain scaffold or

template it promotes certainly
method efficiency or

allows you to introduce some sort
of expert knowledge as well.

It's less likely to design not intuitive

compounds however because not really
could target the space very much from.

Your training set compounds if
you've already this scaffold so

really you're not going to design
any new scuffles this way.

Bo.

The generate and test technique of Danny.

And this.

This after musician approach
both of us templating and

I give some examples of this.

So this generate text technique here's
a case study to design active herbicides

with a certain backbone.

And they kept the structure would have
the highest activity and so this published

in two thousand and seven and so
here you see the template and

so they modified at the R one are two or
three positions.

And so here they put.

Method her appear here they've put to hide
and so that was the optimal structure.

Because that's generating test
here is computed like design

using mixed unit in
the linear programming.

They wanted to design a fuel additive.

And so this is two thousand and
five with these following properties.

And so you can see the template
molecule that they started with.

And then the additions that
they've made over here so.

The C H two O. H.
Group on top of that was the compound

that had the properties that best
matched what they were looking for.

OK so now we get into.

Kind of an area where I'd
normally start my talk.

This is what we do this is computed
like a design with the signature

molecular descriptor.

And we use this technique since signature
creates a useful quantitative structure

property relationship.

I'm going to show in a little bit
of compares favorably to other

popular molecular descriptors.

I hazard to nibble degeneracy.

Really we can control the amount
of solutions that come.

I'm out of our computer aided
molecular design algorithm and

a number of structures personal Lucian
based on the signature height we use and

it's it's tunable and we go from high
degeneracy to non degenerate depending on

what height we go to so it's monotonic Lee
the degeneracy monotonically decreases.

And really the best thing about this
technique I think is that it can develop

novel non-intuitive structures
you don't have to template.

And so you're not.

Predisposed to choosing
a certain scaffold.

And we also have an exhaustive and
effect a fish in structure

generation algorithm designed specifically
for signature and oftentimes this is

a bottleneck in some other techniques
the structure generation system.

Right.
So what a signature it started in one

thousand nine hundred four.

As part of a computer assisted
structural elucidation code.

And it really ended up identifying
the structure of weakness.

So this was one thousand nine hundred
forty really similar to the smiles

notation if you happen to
know anything about that.

The signature of an atom.

Is a root it three that
spans all of the atoms

bonds that are a certain
distance from the root.

So that's the atomic signature in each
atom has its own atomic signature.

And then the molecular signature is just
a vector of the occurrence numbers of

those atomic signatures.

So this molecular signature that
describes the molecular signature of

the molecule is an atomic signature
of that or fragment all parts.

So here's a signature example and
I show method all up here.

And I'm going to focus on the carbon atom.

That's right here and
we can do this for all of the.

Atoms in the system.

But here's the carbon atom.

What you do is you draw turrets so
this carbon atom is bonded to four things

three hydrogen is an oxygen
you drop like that and

then this oxygen is bonded to the hydrogen
that's not you know backtrack.

So this is your tree.

For carbon.

At Highgate zero.

It's just the carbon atom.

That's the atomic signature and
one it's the carbon and

what I've responded to and it had two is
the carbon whatever it's been a two and

whatever those atoms are positive and
so you see the signature it had zero and

I want to take two for that carbon atom.

You would do this then for
each of the six atoms in your system.

And Adam up and you would have
a molecular signature for methanol.

And you see those
are the occurrence numbers and

you might see well these hydrogen's
are all equivalent and that's

demonstrated here but this hides in this
difference in this and that's shown there.

And so I showed this is a very simple
example we can do this for six structures

for aromatics for double bind IT
systems for triple bonded systems etc.

OK So signatures in in these
types of quantitative structure.

Relationships.

We have value weighted
the biological activity

of HIV you want protease
inhibitors as well as

close to thirteen thousand optimal
water partitioning coefficients.

And we compared signature to
those found in commercial

descriptors and there was a whole you
know bunch of commercial descriptors

we could look at here's just a sample
there were hundreds of them available.

And we found that we had
results that were similar

our correlate of results were similar
to that of commercial descriptors and

really the fragment like nature signature
allows them to be more a thug and all and

it creates a still more stable Q.

So are.

Then what.
You would

have from those molecular descriptors.

All right so say you want to do computer
aided molecular design the signature.

What's the steps for the first steps and

I think I've mentioned that to a few
faculty already today when they were

talking about this technique is
we have to start with a data set.

So I set of compounds that have
certain property or properties so

we need to be given a set of molecules and
then we break them down into atomic

signatures and so after we do that we
generate a quantitative structure property

relationship for
our property of properties of interest and

then we must generate equations
which describe how those

atomic signatures merge together to form
a new molecule and these are based on.

The valance arguments in some other hand
shaking arguments that are shown a moment

and then we solve these equations and
those are our inverse solutions.

Then we scored a solutions based on
the U.S. P.R.'s we generated in step one.

And then from that

mathematical description of the molecule
from these inverse solutions you generate

an actual structure from the molecular
signature then we go backwards and

we generate structures and

the best going structure has
become your focus database.

It's really easier to see in an algorithm
and so we start with a database of N.

compounds and
we generate two dimensional structure so

we draw these are we
have them from somewhere.

Then you calculate your unique height age
atomic signatures you might say well I'm

going to start going to do this problem
at how you want to do this problem.

I hate to you kept it all of
your unique atomic signatures.

And from those atomic signatures.

You've you generate your
constraint equations and

I will show an example of
a constrained equation.

In a moment.

So you generate these contrasts constrain
equations and then you have to solve them

and they constrain equations
are solved using a brute force.

Technique.

Originally we used.

A form background arisen these equations
are Diophantine equations which means they

have energy coefficients
an interesting Lucian's Fort Barker

algorithm was coded and
it's an efficient algorithm for

some types of problems for our problems
that would work either really well.

Or they wouldn't work at all and
we never had an idea about which when it

when the problem when the bottleneck would
be and so we redesigned our own algorithm.

That gives us a much
better handle on how long

the solution to this problem
is going to take and

then here's the demi check because you
should always get the compounds back.

From the ones that you're
originated you create your model or

models to score your candidates
that you've generated.

In this box here is
interesting this is the box of.

This is the box that really
refines the problem.

So no where do we impose whether or

not our aromatic signatures
have to obey huckle rule and so

this is the step where we say OK
huckle rule has to be enforced.

This is the step where we introduce
some stability calculation so

we might calculate driving force field so
in intermolecular.

Energy and eliminate some of those that
have higher energies we might impose a C.

rule of five if you're looking for a drug
like ness cycle of can rule of five.

We might limit the number of rings
you can have in your system.

You could also put in a sort of analysis
of synthetic feasibility at this point.

So this step.

Really refines what comes
out of this part and

you end up ultimately with a focused
database and your hope is that

this focused database are things that
you want to explore in more detail.

Maybe involving us instead of chemist and
then evaluating the predictions.

For this for these substances.

OK so

a simple in first design example you would
never do something like this with just

one compound so but I'm just going to show
this with with one cop and his pen Tain.

Normal pen thing and

it has seventeen atoms the twelve
hydrogen atoms at height one.

Are all equivalent they're all hydrogen
is fine that the carbon was called this

variable X. one then there's these
two terminal carbon atoms here.

They all upon it to another
carbon in three hydrogen's.

And so you get two of those and there's
these three internal carbon say see here.

So we represent each of
the unique atomic signatures with

a variable the graphic
ality equation first.

Determines whether or not.

The you can create a connected graph and
this is based on.

Valid arguments associated with
a particular atom type and

so this star indicates a modulus equation
here so this is the first equation.

That's the graph calculation
you get one persist.

And then you have handshaking or
bonding arguments.

So you have two types of bonding in
the system carbon bonding to hydrogen and

hydrogen bonding to hydrogen and so these
are questions in a relate these variables.

About what type of plan
you have in your system so

depending on the types of bonding
you have you will have that many.

That many consistency equations and those
that represent those sets of Diophantine

is a quick Diophantine equations that
I talked about and you solve those.

OK so I would like to move
on to an application here.

University of New Mexico is in
Albuquerque this was done in

collaboration with
Sandia National Labs which is also in

Albuquerque New interested
enough designing inhibitors.

For I cam one stronger than
what's currently available.

And so.

We looked at designing.

Rather they gave us
sixteen seqlock pap tides.

There were nine amino acids long and

they ranged from strong to
weak to night inhibiting.

And so here is a R Q Some results which.

Span.

You know a few orders of magnitude aren't
great and the non binding you know you get

data and it wasn't a number just said
you know greater than a thousand.

And so we lumped all of those
to one thousand and so for

such a small set of sixteen and
for the quality of the Q. so

we were not confident in these predictions
but we want to head we solve the problem.

So.

We generated height one
signatures at the amino acid

level we didn't have to go to
an atomic level to do this because.

You know you don't have to go that finer
grained because you know how what type of

bond you have an amino acid so we did
the high one signatures relative to me.

No assets.

In those of the constrain equations
associated with how the new acids can

combine for our system we distribute we

create a distribution of solutions
based on the quantitative structure and

activity relationship that we had created.

And you can see the predicted I.C.
fifty values are down here and

these are the most active
compounds ones with the lower.

I.C. fifty values.

And so we gave University of
New Mexico twenty compounds.

I'm sorry twenty peptide sequences.

That we had predicted to be.

More active than the most active
in there in the original sixteen.

They picked two here and they synthesize
those and you could see how well.

The actual values compare with
the predicted values and so this was a.

You know a success of this particular
technique and right now these

are the two most active compounds
relative to that particular receptor and

I can one known to date.

So I've done some a variety
of different work with this.

Technique in a variety of projects.

This is a sample of this list.

This is a sample of this list I'm going to
show here we designed hydro floor ethers

as replacements for foam blowing agents.

In two thousand and five.

Brown and coworkers use the signature
technique in two thousand and

six to design.

Up.

Repeat units that had similar properties
to nylon six if I remember correctly.

We have current work on fire retardant
polymers and soap and example of this.

If you look at this data set of
fire retardant polymers here.

These are three properties he
released capacity mass loss rate and

heat of combustion.

And so this is the dataset that.

We utilized and so
we used our technique and we generated.

Several compounds and
I'm going to focus on these two and

you can see the predicted values are very
low and you want low values here and

what I think what's most interesting
is this unit that's right here.

It doesn't really appear by itself at all.

In this set.

However it does appear here so
certainly it's part of this comp.

However this.

Hydra fear and ring.

That's right here is not on any
of the training set compounds.

So what's happening is it's borrowing from
this ring here it's taking this oxygen and

putting it in this ring to make this
structure here that ends up over here and.

Forms this particular compound.

And so this kind of shows
you I think some of the.

Not intuitive solutions that you can hear
it through this particular approach.

We've designed Gluco quarter cord
receptor link instead of nonsteroidal.

We've just published
a paper it's in press.

Basically it extends
the Glaxo Smith Kline solvent selection.

Database they have this green database and
we've utilized that they're in.

First design an extended it and
in the during our extending we actually

identified things that are already green
solvent in there so the fact that we've

identified green solvent that weren't
originally part of the softened selection

got kind of gives us some confidence
potentially that some of our compounds.

Would be interesting things to look at for
further study.

We're talking about this on
Wednesday afternoon at A I C H D.

There's also an Auburn group that's using.

Signature and
product a combined product process design.

And interesting thing about this technique
is you can kind of take in the point

in a in a direction and so we pointed
this in the direction of our strength and

Hansard and set set retards for
the concrete industry and

certainly so W.R.
Grace is very interested in this and so

we have funding for this through N.S.F.
to design admixtures.

For the concrete industry using
this technique and it's a combined.

Design with experiment.

Approach which is similar.

What we're doing with the design
of also timers disease.

Aggregation hitters with
a colleague at South Carolina.

And finally there's a static
liquids as pretreatment for

sale is like ethanol and we are working
with the joint Bio Energy Institute.

On this and I want to talk
a little bit about this one.

In more detail so Tennessee has
committed to switch grass as the say

last source to make what
they call grass only.

They're paying farmers to grow this

is fairly robust You can
see a picture of it there.

So pretreatment of the biomass is required
to expose the say Los there are different

techniques that have advantages and
disadvantages.

I act liquids have been
of interest in this area.

Because it's identified
as a green solvent and

it has a basically a negligible
vapor pressure and

so the energy constraints or
the energy input associated with

recovering the solvent
is is is much less so.

Salts I think liquids or
salts that melt below one hundred.

On environmentally friendly solvent.

Their design are solvent because you can
kind of pick between cabins and eyes.

And here's just a sample of some.

Projected to be a very large combination
rather large number of can I.

And I develop one so
you'd say well this is.

Amenable to a computer aided
molecular design approach.

So for sailors pretreatment.

We are looking at melting point
viscosity and sailor site delivery.

OK so we're going to play or
technique here and

we need to score the hand first Aleutians.

However.

Previous studies.

We've had all wrapping data sets
if we wanted multiple properties.

So if we wanted to do things.

Only point and vapor face there
are a kind of clarity like we did for

a hydro floor of foam blowing agents.

The data sets are overlapping
that doesn't happen here.

I like liquid there's very limited
overlapping data at the melting point set

is large and diverse viscosity
there's data available but

only minor structural diversity and
a very limited data on site Realty.

So we had to redesign or
we had a modifier design strategy.

Because there are only few ironically
quids that had all of the data for

the desired properties so
we focused on the cattle in the side.

And there's more diversity
among the kadai And

then there are hands and so
while this is a more qualitative approach.

What we've done is we've kept
the constant for each property so.

For the melting point we were
keeping this and I'm constant for

the discussing this and I stand for
the Cybill of this and I asked him and

he said we have a decreasing number
of compounds available to us.

So we've selected the melting point.

And by the way there's the.

Cuter statistics for
the melting point viscosity and the.

Cycle the work.

OK so
the inverse design process Here's a sample

of the five hundred ninety nine
training said compounds for

the melting point set so what we did
is we took our set of five nine and

we solve the inverse problem the way to
really spoken about we get about a hundred

thousand solutions we keep those that
have a Melting Point less than fifty and

we end up with close to three hundred
thousand solutions we don't generate

structures at this point what we
do is we take those solutions and

we score them relative to our frisk us the
set and we keep those that have a predict.

Did you discuss the Lower than
one hundred seven points.

At this point though we
introduce an overlap metric.

So what you overlap metric does is it
looks to see the number of signatures

that are available to the model relative
to the number of signatures available to

the compound you're evaluating and if
that overlap in signature is high enough.

We keep that and the conjecture is that

the higher the overlap the more confidence
she would have in those predictions.

So we kept.

At the seventy percent level the viscosity
set we do the same thing for the side or

the set.

You might think five percent side ability
is in small bits of very large number for

cellulose.

And if we keep those that had
a greater than fifty percent overlap.

We end up with about twenty
two hundred solutions.

Those solutions we
generate structures for.

And each solution doesn't necessarily
have to generate a structure.

However one solution can
generate multiple structures so.

You don't get two thousand one hundred
thirty four solutions structures here.

We end up with a focused database
of close to four hundred compounds

and here's a sample of some of them.

And obviously they're going to have.

Predictive properties that are within
the range that we're interested in.

And so what we did and by the way here is.

Hey here's just a.

More conceptual level.

Presentation of actually how this works.

So really we take these two compounds here
that are part of our training set and what

you're doing is you're cutting parts of
them and making new molecules out of him.

Hopefully you know it's the parts that
are really hopefully you're designing

it such that.

It gives you the optimal properties.

And so here's a little example of one
of the compounds that we've generated.

So experimental verification so

I have the student that they
had want to see fellowship.

They were sending to Australia with

Doug McFarlane to actually synthesize
some of these and so he gets there.

And we give him the west.

And he's like Well in a couple of months
we're not going to be able to synthesize

and testes you know our top performers.

So they identified those that
were feasible to synthesize

within the limited time frame
that they had during the summer.

So here's an example of some training set.

I had a quick wit So these are among
the five hundred ninety nine.

And so these are some of the ones
that are inverse designs identified.

And so they were synthesize and

experimentally tested their
melting point and you could see.

You know how these parts come together.

To form the different parts
of these new liquids.

And so.

You can see the experimental values and
then the predicted values and

you can see that they're you know
they're fairly close and so.

On This is very encouraging for
us because the next step then.

Is to set up a set of system lab and
so right now this is happening.

We're looking through a set up
a synthesis that have a Tennessee text to

the design some of these more.

Additionally and
evaluate their melting point we already

have protocols established in our lab for
sale side ability that we had

worked on in the previous semester
I didn't discuss this at all.

But we have work that looks that

data mining pub cam Cam is a bio
asked a database for me and I.

Across multiple bio Athanase.

We have a paper on this from last year.

It's really not computed
Michael design but it.

It's really an exciting technique and

we're trying to do this on
multiple bio Athanase and

we have an experimental verification
step that we're involved in right now to

validate the or verify some of the
predictions that this approach has made.

You know fundamentally going
to thermodynamics this.

So I have projects involving modeling and

experimental work of poly also
flowing agency ability and

ultimately a question of
state creditor's ation.

Using quantum mechanical approaches to
aid in the equation of state primitives

ation So those are some future directions
and some additional work them involved in.

And I know I've exceeded my time limit.

But at this point I'd like to acknowledge
the Department of Energy and.

An S.F..

As well as collaborative and so

here's a group of students that
helped with this and these are all.

Actually the greener
the undergraduate students.

So I had several in graduate students.

Involved in this.

So thank you for your attention and
I'll take any questions at this time.

We would we want to do that once we.

Once we.

Experimentally determine
the properties of the models.

So as part of our Also
timers disease work.

It's a much larger algorithm where
we're going to then refine the mom.

Predictions based on the the predictions
that our models end up making.

Alternately we hoped that
we wouldn't have to do.

That's that.
But practically I suspect we will.

But yeah there is a certainly you
know that aspect helps us really

refine the models that we develop.

You mean the inverse design.

Well

I mean it really depends and
so let me give an example.

You know I don't want to
bore you with the fact

that we have to draw the compounds
are going to take a longer right.

So but I think practically you're asking.

I how long this computer have to
work on the system and it's really.

If it doesn't finish.

You know in like a few days then carbon
authorial explosion will kill us and

we won't be able to store all
the information we have so

we have an idea of how long things take.

And so you know we're not talking about
weeks if that if it takes a few days.

That's at the high end.

So it wouldn't take longer than that.

Yeah

I mean I haven't used it for
anything like that.

So I don't know if I'd really
be qualified to speak to

that I do have an idea as to why something
like this wouldn't work in an area.

Well.

Yeah well so
you know depending on the states right so

you can have different atoms that
have different states and so

we would identify that in the atom
type if that was the situation.

So you know nitrogen can have.

I think it's you know in five or
something like that.

So there are ways up front if the if
if you have a different animal types or

if you have the same atom types that have
different number of patterns you would

denote that differently by
the atom type in the signature.

So you can handle something like that.

Which.

I'm curious that you have in mind.

Just people want precipitate.

Just.

This kind of.

Well.

You know we.

You know we haven't.

We've done that.

But we've never you know it's
not something quantitative So

yeah it's exactly the issue so

we've used water to as a an anti
stop in gotten the sailors out but

we were really most interested at that
point in measuring the site delivery and

you know it was cool for the student
to foil the kind of quick Woodward.

Dr the water off but
certainly we have water in that.

Yeah and.

Maybe you know that the issue with the
Arctic liquids is that when you do that.

Another time.

You know you might end up.

Degrading the Anik liquid and that's
one of the issues in this particular

technique or rather in this in using and

cliquishness area that people
are are evaluating it.

Well I

mean fundamentally you know if it's not.

Provided the groups that
you've identified upfront.

Are able to come together to do that
our technique isn't limited by that.

So in say the you know
the general in test technique.

They start with like eight groups and
only those eight groups can.

You know what will merge together and
what might be a group of benzene ring or

like a fluorine and so
really there's kind of it's like a mixed.

It's a mixed approach.

So.

You know if you can generate
those molecules potentially.

In the same way that we have as long as
you have those predefined groups upfront.

Now the problem that they
have is that they wouldn't

have to SARS are curious P.R.
is developed for those groups.

And so that's where they're
you know they would have to.

End up doing that for a wide variety of
systems in order to work in a book and

find like a drawback technique.

Well ultimately.

I mean ultimately the if you're using it.

In a predictive manner.

Right.
Ultimately what will generate

the compounds since that you know
you're synthesize the compounds.

And then evaluate them experimentally and

so that's really I think the only
way to evaluate their their.

You know their predictive
ability outside of the set.

They're using it.

I mean we you know there
are techniques to evaluate the models.

There's leave one out techniques where
you leave home a molecule out then.

Run a prayer of write
have different models and

calculate how well they predict
the one that you've left out.

You can leave out just a whole host
of them and train on this model and

see how well they evaluate
this set that you've excluded.

So we done that but ultimately it's
it's how well it predicts the set.

You've created.

It's not melting quite yet.

Correct.

Yeah.

Now we are we have well.

You mean for
the stuff that we've generated.

Yeah we're not even I mean this is so for

example the melting point data we had
just maybe gotten a couple weeks ago.

So we're really this is this is kind of.

Yeah absolutely.

I mean I think that we do.

And one of the problems that you run
into is that even a little bit of

cellulose in an act liquid
makes it very viscous and so.

One of the things that you.

You know one of the things that's
potentially on tap is how can we.

Maybe not use a pure ironically can
we mix of the something else so

that when you add the sailors to it
it will retain the site will be but

you won't have it won't
won't cost you as much.

Money because the goods are expensive and
maybe you can modify some of

the properties in that way but
yeah I would certainly expect the quality

of the predictions to the grid
because the set sizes in this large.

OK.
Yes So that's really high one and two so

what happens is that if you
use a higher signature height.

You become too specific because now
you're looking at the atom and its

nearest neighbors and it's next nearest
neighbors it's next nearest neighbors.

So you get really specific signatures.

When you go to a higher height you also
because degeneracy is a function of

the height you end up with a reduced
number of inverse solutions at that point.

The other way.

If you end up at zero Now we're talking
about just the molecular formula.

So really there's there is some sort of
optimal height between one and two and

we've done some similarity studies
that have been shown here that also

kind of call out that one of the two
are really where you need to be so

that's where we've had it up.

We've done things where we've combined
Heights one in two we think you

SARS with multiple heights.

So there's opportunities to kind
of mix mix and match that level.

Yet.

Yes but we haven't we haven't.

I want to we haven't needed to do it yet.

I mean.

But there you know we thought about
ways to add in Cairo receptors through

modifying.

You know what the signature would be or
what the how we would denote the atom

type so yeah there are there are ways
that we could do that and I suspect this.

This matures.

We might have to do something like
that because you know what hat what we

would do now.

For example.

If you have a stereo isomers and
the sis is active in the trances and

active and you make it to the graph and
it's the same molecule.

What do you do we just
get rid of both of them.

I mean.

And so we move on.

But yeah I mean if we were.

You know normally it doesn't come up but
in situations we've just got rid of it.

When it's going to come up more and
more in the future.

You know we can certainly
consider that absolutely.