Cool.

So thanks for having me.

And it's not my first time in Atlanta but
it's it's I think my longest visit I

stayed here over the weekend and enjoyed
your beautiful city for some of its.

Better known features like the Atlanta
Braves and the aquarium and so forth so

that was really fun and it's my pleasure
to speak to you today on the subject

to motivate the presentation.

I'll tell you first that our.

The environment I'm in at the brode
where we do computational chemical

biology research is one that's
characterized by a tremendous

effort in synthetic organic chemistry.

So making molecules that will go into our
high throughput screening collections.

And we have probably forty.

Chemists real chemists not like
me who make such molecules and

who are constantly looking for
guidance on what to make.

We also have quite a number of different
high throughput measurement techniques

including sort of conventional high
throughput screening but others.

It's also some profiling techniques that
I'll mention in the middle portion that

provide an opportunity to
inform synthetic chemistry.

As to how to grow a more optimal
screening collection over time.

I think it's pretty common for

people thinking about a particular
biological problem protein or

pathway to to learn how to optimize
chemical matter around doing that job and

so that's sort of conventional Q.S.A.
are or essay our work.

What we've focused on is instead how to
how to improve the overall collection

performance and I'll go into the details
about what I mean by that in the first

part of the presentation I'm going to I'm
going to motivate why we need compound

synthetic compounds at all because there's
a fair community of chemists out there who

would argue that nature is a sufficient
source of starting points for small mom.

But all you need to really do is go out
to all the plants and the sponges and so

forth and squeeze them and
fraction of that stuff.

And now you have all the drugs you'll need

we undertook a study to ask whether or
not.

That was true.

That's based on biological
network connectivity.

Of proteins.

OK.

So as some of you may be
aware it's been known for

over a decade that biological
networks when considered by

protein protein functional connections
whether they be direct binding.

Enzyme substrate relationships or
other regulatory relationships can be

viewed as networks of interactions and
within such

networks there is in general small
number of highly connected proteins or

hubs and a much larger number of less
connected proteins which will call leaves.

And of course the continuum of
connectivity is in between.

Different biological network
graphs differ in their

connectivity distributions and
this can be quantified.

For example here we have a high
leaf to hub or a show and

this is a made up network for
cartoon illustration purposes and

a lower leaf to have ratio with the same
otherwise the same number of nodes and

edges and

the extent to which distributions
are different can be quantified using.

Cumulative distribution.

As as illustrated here with this cartoon
between the number of connections and

the log of the number of nodes that
is the number of nodes with each

number of connections.

So it was known since two thousand and

seven that in humans disease genes exhibit
intermediate network connectivity.

This is not surprising because
in in the human hitting us.

Hub with a small molecule or a gene or

some mutation is more likely to cause
acute death and not to be heritable.

Whereas these terminal terminal
nodes are more like or

less likely to make the organism sick.

So in between these are intermediate
disease genes that.

That because of their intermediate
position in the conduct of any

distribution are more likely to be
inherited by offspring and therefore

comprise the set of disease genes that
are present in the human population.

So this is a publisher result by others.

We hypothesized that natural product
targets would be hubs because

natural products very often are selected
as chemical warfare agents between

competitors and you don't want to
kill your neighboring later or

kill its children you
want to kill him now.

So you can eat his food right
that's that's the deal.

So the solution we surmised was that.

That hubs would be would be enriched
among the targets of natural

products so
to to look at this question we engaged to.

Existing databases one called strain

which is a publicly available database
of protein protein associations.

And it's available from the European
molecular biology laboratory and

we used it as a foundation for
analyzing network kind of activities.

Again some of these are direct

interactions some are and I'm substrate
some are indirect regulatory.

And string has a good scoring system for
how it weights the evidence for

these so we just took it at its word
we mapped natural product targets

on to string proteins using
a commercially available natural product

database from Bio they sell
a big product which has a ton of

compounds in it for all different
purposes and we negotiated with them.

Because we only we didn't want to
pay a bill for the whole thing.

So we only wanted cases where the compound
was known to be a natural product and

the protein target was no and
that turned out to be some

eight percent of their database and they
did in fact charitably sell it to us for

only ten percent of the cost of
the full database so from this.

From this database we identified
nine hundred forty six unique

proteins in from the human that
have no natural product targets.

We also looked at morbid map from N.C.B.

I which is a list of mostly
heritable human disease genes and

BILLION heritable human disease genes as
well as the emerging database of genome

wide association study genes so genes that
are implicated in human disease I.G.Y..

And together we identified
about twenty seven hundred

human proteins in the string database
that are implicated in human disease.

Our initial observation was
that our hypothesis that

natural product targets represent highly
connected nodes was indeed confirmed.

Shown here is just kind
of a been histogram.

Of number of connections per protein.

For small numbers of connections and

some intermediate ones and
then more than fifty.

Over here is the actual
cumulative distribution of all

the the whole database so
natural product targets are in fact.

Right.
Shifted in this plot indicating they're

more highly connected all all
three pairs of differences

between these are statistically.

So all human proteins show.

The most left shifted that's
because of what number

of lowly connected proteins disease
genes here in green are intermediate and

natural products
are the most highly active.

We next asked whether it's.

Whether proteins targeted by synthetic
small molecules and not natural products.

Have a different kind of committee
distribution the natural product targets

and we find that they do.

So what's shown here are the same.

The same.

Green and red as before human disease
genes and natural product targets and

now in gold.

Bioactive targets gotten from Campbell
that exclude natural product targets so

any protein for which we could
not find it naturally and but

we could find to synthetically and
is shown in gold.

These synthetic Lincolns
though are not are not drugs.

Some of them are probes and other tool
compounds so we look further using drug

bank at actual drug compounds
that are not synthetic.

That are that do not have natural product
targets but only synthetic targets and

we see a further leftward shift of
the approved drug targets from drug bank

over on top of the human disease gene so
this tells us that synthetic.

Compounds that are being made or that have
been made to target human disease genes.

Are our populating or antagonizing

if you will part of the proteome that's
not being hit by natural products.

So we take this as as evidence that.

Making molecules is a good idea.

That's the summary statement of this.

Of this first portion don't just go out
in nature and find starting points but

in fact dream up new molecules in
our chemists are quite good at this.

But what they always ask us as
computational biologists and

chemistry Maddux practitioners.

OK fine.

So which one should we be making which
compound I can dream up all these

molecules which one should we make.

To to articulate the challenge
of this problem.

I'm going to use some fun cartoon
slides to illustrate a concept

income informatics.

And I think in similarity measurement.

General similarity is a relative concept.

If I show you these pictures and
I ask you.

Are these two pictures left and
right similar you might

say well what do you mean by color or
by shape and that's a valid question.

Indeed with this pair of objects shape
looks kind of similar color not so much.

And now you're already
doing something subjective.

In addition to what shown on the screen
you're comparing two things in your mind

the taste of a mango versus
that of a capo and applying and

saying no way these are similar.

So it depends not only on on on
the descriptors you elect to use but

also on subject of interpretation.

And this can be further
illustrated by this picture where

almost all of your similarity difference.

A lot of arms that are going off
in your mind are subjective and

have nothing to do with the colors or
shapes that are represented on the.

Chemists do the same thing they
look at these molecules and

they make subjective assessments
both of these are drugs.

Both of them are effective in humans but

they're quite different to
even the untrained eye.

So.

I'm going to spend a couple minutes
talking about how in practice we can

try to dig into this a little bit more
using measures of of local structure

versus global structure of small
molecules so chemical fingerprints

are very popular way of characterizing
small molecule structure in a computer.

They are typically half coded bit strings
that are assigned by looking at all.

Substructures of a certain
size in the molecule and

assigning beds associated with
the substructure being present or absent.

In the molecule these measure look
because of their their their cutting them

the molecules up into small substructures
they measure local structure and

the familiar tandem OTOH
similarity measure which is used.

It is in fact a fingerprint based
measure it's it's tailor made for

problems of this kind where you have two.

Bit strings in this case
representing two compounds and

you want to know the fraction of on bits
that are shared between the two structures

so when somebody says
the tandem OTOH is blah.

This is what it means but.

OK.

But is that really what it means.

So different.

Fingerprints emphasize different groups.

So here's Tanna moto coefficients
calculated between this compound and

three other similar compounds
using a particular type

of fingerprint descriptor that's out there
in the public domain and you can tell

just from looking at the scores and
the differences between the compounds

that certain changes are considered
more important than other changes.

Based on this description.

So even within the world of
fingerprints it matters what you pick.

As far as what you get so
just just to be specific.

Here this methyl group is the same here.

The niter group is
different at this position.

Whereas here the final group is the same
whereas the Napco group is different.

This tells me that.

Because the drop in tandem
OTOH is much larger for the.

For the nature of group
versus the final group that.

These particular descriptors
consider methyl The nitrile more and

more important change than they do.

To that on the global structure side.

We can look at one popular
method of assigning.

SHAPE TO molecules This
is developed by Sauer and

Schwartz in the in the early part of the.

Last decade.

Called principal moments of inertia.

This is essential like finding
the best ellipsoid for

a small molecule some pose
of a small molecule but.

Weighted by math.

So it's a moment of inertia calculation
that gives you three amplitudes and

then by taking their ratios.

The short one over the long one.

The middle one over the long one.

You assign each compound to a position in

this triangular space that is
bounded at the corners by.

These canonical shapes Rod disc and
spear and

you can conceive a similarity
measure as a distance in the space

between two molecules that reflects their
difference in global shape properties as

opposed to local properties like
the fingerprints described.

So to give you an idea of the size and
dependence of this.

If you consider this corner here both
the bucky ball big thing and Adam and

Cain would both be pegged
at this upper right corner.

Even though they're
very different in size.

So it's giving you global shape but
not size.

If you look at a random
collection of compounds

here are four hundred known bioactive
small molecules in in the broad collection

using the two different descriptors you
can kind of see what you're getting at

with two different sets of these
two different descriptions.

In the case of fingerprint descriptors you
get basically a low average similarity

between all the molecules a poor
mapping to overall shape or complexity.

But very good discrimination of
closely related structures which is

evidenced by these kind of these blue
little blocks right here on the diagonal.

With the P.M.I. descriptors.

In contrast you get essential a high
average similarity for all molecules and.

Despite the size of pendants
of independence a pretty

good mapping to shape and complexity.

But but as described a minute ago similar
molecules can have very very different.

Structures.

Similar similar as judged by this.

So what does this really mean it
means that chemists despite asking us

what should I make really need to decide
first what is it that they want they want

a shape to be a certain thing do they want
local structures to be a certain way.

What are the what are the properties
that they consider important and

I'll come back to how we can guide
that guide that decision making

in the last part is everybody.

Any questions here before I go.

OK cool.

So to start to get at.

What some of these products how some
of these properties are similar or

different between different
sources of compounds.

Natural Products compounds that are in
typical pharmaceutical screening

collections and those coming from the
academic synthetic chemistry community.

We characterize about fifteen common
fifteen thousand compounds that

at the time had no previously
known biological activities.

So again commercial C C diverse
compounds here we're calling.

D C and natural products and.

So we just want to know can we can we
quantify what chemists are telling us they

intuitively think are different
between these compounds.

Just a touch.

OK And so

I think this slide this is just five or
six randomly picked guys from each group.

That kind of give us a sense for
what the differences are these commercial

compounds in general are much
more flat aromatic compounds.

With.

Not a lot of side chain variability.

The natural products are pretty hairy
lots of stereo centers kind of bigger.

More.

Complex and these the synthetic
compounds from from academic

chemists are kind of intermediate
they have some features of each and

as we'll see that that's borne out.

With more quantitative studies.

So two things that chemists told us.

They were interested in
they they kept saying

we want to make things that
are more natural products like and

I don't I don't really know what
that means because natural selection

makes natural products and
there's no way to emulate that laboratory.

What you could do is make
compounds that have features

that are reminiscent of natural products
or similar to natural products but

it's not clear that that's going to
automatically confer the other properties

natural products have to wit
somewhere in the ecology of Earth.

There's a target for
every natural product.

Otherwise nature wouldn't
bother to make it.

It's too expensive to make
compounds even for nature.

So.

Two things that are chemists were
interested in as knowing whether

the stereo chemical complexity
of their compounds.

Or the shape complexity of their compounds
as judged by two metrics shown here.

Were approaching those
of natural products and

we quantified this in an AND
number of ways one of which is shown here.

So then the total fraction of carbons that
are stary genic centers chiral center.

And the total fraction of either S P two
or S B three carbon that are S.P. three so

both of these vary from zero to one and
provide a continuum in

in what we observe between the commercial
compounds which are much more flat and

less complex by both measures and natural
products which are much more complex.

By both measures and show that these
compounds from academics and that chemists

are in general sort of intermediate You
know they're getting where they are.

As judged by their phrase natural

product like they're making
what they think they want and.

We can take this further and

look at properties that say
medicinal chemists care about.

So here are.

Several properties that underlie love
Pinski rule of five and a related.

Rules for filtering compounds
molecular weight log and so

on and we see that and
in in four out of the six cases.

The diverse compounds are in fact
intermediate in their values and for

the two that that they're
outside of the C.C. N.P.

they're not far outside it and are still
for the most part within acceptable ranges

although these are a little
bit heavy Penske's rules.

That that chemical diversity is is
really relative like similarity is.

All illustrate with three
different ways of of Clare.

Characterizing these compounds sets.

So if you take the six properties
on the previous slide and

you principal component analysis and
then show as I've done here.

The top two principle components in this
case this accounts for about ninety three

percent of the diversity of this
particular compound collection.

You can see that those the compounds.

From Nature the green ones spread
out in a direction different than

the compounds from
the synthetic chemists and

this map of the factors on
to this space illustrates

how they're accomplishing that diversity
So nature is accomplishing its diversity

by varying the number of hydrogen bond
owners and accept or polar surface area.

Whereas.

It does not.

It does not do as much variation
of rotatable bonds and

size whereas most of the.

Most of the diversity of D.C.
is from changing in size.

If I look at simply atomic composition.

We can get another picture of what

chemists are doing versus
what nature is doing.

Both of the synthetic classes of
compounds that the more flat like.

Typical commercial and

the academic synthetic chemistry compounds
are very nitrogen composition and

to a lesser extent carbon composition
to get most of their variability.

Whereas the variation in oxygen
composition is much more suppressed than

that nature on the other
hand does a lot of it's very

easy compounds by changing
the composition of option.

So the number of options and
molecules much more variable on this that.

You can do the same trick with.

Any descriptors you like I show here
aromatic rings versus side chains because

it sort of maps onto these again so
changing the number of of rings and

aromatic Rings is a big driver for
the diversity of the synthetic compounds

changing the number of side chains for
natural products.

So this gives us some clue of
what it really might mean to be

natural product like.

In in their parlance.

Something that chemists have become
increasingly interested in the past two or

three years is this notion of a scaping
Flatland earlier studies by us and

others have indicated this
theory of chemical complexity.

As I'll show in a few minutes is in fact.

Better for outcomes in biology
either as drugs or probes.

I'll come to our evidence for
that in a few minutes and

this paper by Lovering was
entitled escaping Flatland and

took this to clinical compounds and
addressed the extent to which more

storage and compounds were survive longer
during the process of clinical trials.

What we see by assessing using
the principle moment of inertia

is that the efforts chemists
are making to make more globular

like compounds that is those more.

At the top corner.

Are in fact working they're shifting the
distribution of the blue on up that way.

Relative to natural products and
commercial compounds that the data

here are are shown as
cumulative distributions from

either this corner or from the middle
of the side in these two plots.

And you can see that the D.C. are.

More like spherical than either and
P. or C C.

So is this is what chemists have
decided over the last couple of

years they really want to do.

I don't know whether
it's the right thing but

we know that we can guide them
with that goal toward doing it.

I'll just give you two quick
examples where we published a study

with a chemist retrospectively that.

Considered the relative role of Styria
chemical diversity within any of these

skeletons or skeletal to versity within
the branched reaction pathway on shape.

And were able to quantify
the cases in which.

The the difference between
a precursor skeleton and

its antecedents Gallatin
were more important or

less important than the difference in
dietary Merce at each of those positions.

So again this was done.

Retrospectively the chemists in
this case did not have screening

data at the time to support
which of these was better.

I'll kind of come to that in
the last part of the topic.

But it shows that we can quantify
these choices that they're making in

terms of relative importance of
different types of diversity.

A more a more global library
decision making process can also

be measured with respect to
its effect on shape yours and

other chemistry collaboration we engaged
in in which two different libraries

were made one published in two thousand
and eight one in two thousand and ten.

Where.

But essentially only one intermediate part
of the library synthesis was changed and

we wanted to look at the consequences and
shape of the old versus the new library

and were able to quantify that again in
terms of even in this case distance from

the the sphere like corner and
make the make the statistical

statement that indeed the blue library is
more spiritual in nature than the read

that may be true but no one has shown
me that more spiritual is better to

my satisfaction and
I'm in the business of showing that.

So this gets at the last last.

Sort of half of the talk to our chemists
really picking on their own the right.

Desired property.

To go after it's clear that if
they have an idea what they

should go after we can
help them get there.

That's that's kind of a straight
up computational exercise but

how can we help them pick the good ones.

So I'm going to I'm going to
take a slight diversion into.

And what is sort of another
area of research in my group

to tell you a story there.

But that I think importantly sets up.

The last part of the talk the last
five minutes I'll come back to this.

So what inspires me is the following
all of this coming from magic

stuff that you've seen is about taking
compound structures that you may not have

made before and computing something about
them in advance of their synthesis.

That's really important.

But it doesn't really trump what happens
when you take and make a molecule and

go and measure something or some things
as we'll see about it in the laboratory.

So to me and
a completely alternative representation of

chemical similarity is
performance similarity and

that's where that's what I'll
talk about in this third section.

Anyone have any yet and they told me to
wait till the end but I can't I like it.

I like questions too much during.

No.

No we didn't consider that.

As all show in this section we actually we
do by having experiments with one hundred

protein.

And we treat all the proteins equally.

In that study.

But you could do that.

I can even imagine clever ways to
use some measure of complexity or

binding site complexity to weight.

Some of the things that you'll see need
present in the context of each protein is

considered as good as any other
that we haven't done that.

I mean it's that's a useful thing
to think about for future work and

collaboration particularly with people
that think about protein structure

complexity a lot which we don't.

But that that's definitely
the subject of this section so.

We make a measurement we take a small
molecule and we put it on some biology.

And we make a measurement and ideally on
the same day at the same time we take

D.M.S.O. vehicle and
put it on the same biology.

Ideally a bunch of times.

And then we ask how likely is
the warm that had the compound then.

To be a member of the MT
treatment distribution.

And that's kind of a score for

whether the compounds presence was
required to observe the effect.

So it's like a probabilistic
measurement of success.

And we do that in a bunch of contacts
which I'm about to go through that.

All right so
here's here's the idea right and

this is the this is this is the sort of
generic slide and I'll show you several.

Samples.

OK though.

So this is what I just
said we take a compound.

We take some biology and we test whether
the compound did anything and it.

If we do that a lot of different times
with a lot of different biologies we can

get a vector of measurements
that come from all these

different assays whatever they are so
this is not a new idea.

OK The N.C.I. pioneered this idea.

Twenty five years ago.

With the N.C.I. cell line
sensitivity profiling experiments.

In which they took compounds and exposed
them to sixty different cancer cell

lines and took a second
compound that did the same and

articulated a level of similarity
between the compounds that had to

do with the differential
sensitivity with these lines.

So.

Other work that has been it and these
kinds of sensitivity profiles they used

them to do things like connect
compounds and targets so

they had some of their original work that
we reviewed in in two thousand and nine.

Like sixty cell lines by four
thousand small molecules and

they'd also measured protein levels.

So these are just so killing experiments
and these are protein levels for

those same sixty lines against
some number of protein targets and

they did a nice matrix algebra thing to.

Connect small molecules with
targets through the common

intermediary of the sixty's cell.

So this is good.

This is kind of the this is
the grandfather of all profiling as far as

I'm concerned.

You'll hear the people like that.

That I'm about to talk
about forget about this but

this is the this is the beginning.

Profiling has been done
using gene expression data.

This is probably the the kind of profiling
that's most familiar to most people.

The two studies that were most relevant
to small molecule perturbation pro.

Filing are the compendium in
yeast that was done by Rosetta

in two thousand and the work at
the brode of the Connectivity Map or

small molecules in Penge on a cell system
you measure gene expression profile and

use that to make connections
between different compounds between

induced states of cells say by our
own AI or whatever and different

cell lines whatever it may be and so the
conic to the map some of you may have read

about is is the sort of web tool that
enables this functional annotation.

Profiling can be done with image features
so high for high content screening

assayers where you impinge on it on
a cellular system with small molecules and

measure a bunch of different features
by microscopy with different

fluorescent dies or whatever can be
used as the basis to form profiles.

And.

This can connect small
molecules sensitivity to

different phenotypes with with a compound.

Profiler is an example of this that
was also developed at the brode So

these are these last two are examples
of datasets to which we have access and

think about and then.

The one thing my group has developed
over the last couple years is assembling

profiles over many diverse assaye So
the example I told you about for

the investigators set
out to build a profile.

By taking sixty cell lines and doing
something or running gene expression.

We built chem bank in part to enable
assaye performance profiling across

many different assayers that were not
done with the idea of building profiles.

They were done because each biologist
wanted to discover a pinch of magic dust

for their own biology.

And so we just needed a way to
reconcile the scoring between these

to put them onto a common footing.

And this kind of idea.

We've used it as what the other
profile when you've seen to enable

lead hopping and target identify.

Studies I'll come back to one of those in
a minute I'm going to spend three slides

on how how we do it.

Which gets back to your question.

So we score.

So the one thing you can be sure to
have in any small molecule perturbation

experiment is the corresponding
vehicle control.

More of them is better and replicate so
the compound treatment is also better.

So we handle replicates of this which
we try to do a lot of times and

of this we try to do as many times
we can afford usually two or three.

In a principled way.

So that the evidence is accumulated over

the appropriate number of mosques
that share a plate with the compound.

We do transform the data then to
emphasise the decision boundaries.

So that for a particular confidence
of where the compound scoring

we suppress differences down here
where it's all in the noise anyway and

we also suppress very these are kind
of like Z. scores on their own.

So after a certain point it doesn't
matter anymore it did it right.

Point nine nine nine nine nine nine
nine nine we don't care anymore.

So we transform into using this double
sigmoid into a regime where we're

emphasizing the decision
boundaries were really matters.

And then we normalized to account for
missing data using Fisher's the prime.

To take.

Pairs of compounds that have
shared different numbers of path.

And normalize their similarity
scores between each other so

that we can rank for
any given compound we can say what asses

has it seen in common
with any other compound.

And did it perform similarly And
the answer doesn't depend on.

The number of assayers that
that share being different.

So this is now a fourth
novel profiling type

that we developed and can give rise

like the others to pairwise similarity
between any collection of common.

Right.

So.

One so.

So again here's the summary going back to
the through these various types that no

matter what you're doing you end up
with a measurement vector that you can

use as the basis for similarity and
diversity between compounds.

So this is an instead of a cam
informatic profile we now have

a measurement vector and ask
a performance profile for each compound.

And this can support target ID I'll
give you an anecdote of how that.

Organ.

Of how that works for some we published
last year is just one example.

So we take sparse data in can
bank bunch assayers were done

up to a thousand on a bunch of compounds
but not every compound sorry as a.

We had in a separate assaye we had

identified a compound that
induces insulin expression and

pancreatic Alpha cell so this part
of our disease biology groups work.

And they came to us and
said we just found a suit compound.

We want to know what it does so
can you tell us what compounds in

the in the whole historical data
set perform similarly to it.

Several compounds had similar profiles.

According to our.

Threshold of similarity metric.

In fact eleven and

they were similar enough that we were
willing to tell them about them.

And two of those nine
were an annotated but

two were previously known to be
kindness inhibitors so that immediately

gives us the hypothesis that our new
compound is a kindness and had better.

So we sent it out for
kindness inhibitory profiling to her and

indeed confirmed that it is an inhibitor
it's not super selective but

it is a kind of inhibitor and it's it's
inhibitory characteristics or enrich for

particular parts of the kind.

So it's not a broad spectrum and
have better but it's within a fan.

Only it's not so good but
it kind of goes to two ish families and

they're still investigating what
this compound really does and

in out in Alpha cells
that makes this work.

So this is an example where a profile
similarity can support target ID So

that's sort of the similarity
version of performance profiling.

Another thing that we're that we're doing
this is more like just a heads up for

the future because this dataset I think
is going to be valuable to the community.

Later this year will be releasing
a dataset where we did.

All three of.

Gene expression.

So you are morphology and cross essay
profiles for three sets of compounds.

Twenty thousand from our own internal
model synthetic chemistry collection.

Ten thousand from the National Screening
collection them up the C.N.

collection and three thousand
known drug and tool compounds.

And so by relate relating similarity
relationships among these people will be

able to go and find novel chemical
matter that does something similar

to a compound that they've studied or to a
chemically similar compound the something

they've studied so we for this middle
set we pick them so that they be.

Most representative of the.

The active subset of
the national collection.

So we're going to talk more
about this today but you can.

Look for a dataset to come out
later this year or early next year.

OK So that that's performance.

As an alternative small
molecule similarity.

So we've talked about
chemical similarity and

diversity and now we've talked
about small molecule similarity.

The last bit.

I want to talk about small
molecule performance diversity and

how we can use it to guide
synthetic chemistry so

coming back to the idea of can
we actually tell chemists what.

They should be making.

So remember this slide this is where
we left off that little segment.

Chemists may or
may not be picking the right properties so

how can we get at this.

So remember we looked at
structure properties for

these three classes of compounds.

We also did a.

Performance Analysis of these compounds
using small molecule microarray So as a as

an ass a platform so we print compounds on
glass lights come in with labeled protein.

Wash.

Then look at what secondary
antibody it what's value.

So we have a matrix of fifteen thousand
compounds by one hundred proteins.

That that represent primary
punitive binding interactions.

This is what those look like.

This is just a binary.

Heat map of.

Showing the the composition of the groups
and then the one hundred proteins.

So one of the things we looked at is
whether that be the hit rates overall

were different.

Among these.

And we did find that that
the diverse synthetic

chemistry compounds had a slightly higher
hit rate and less disperse hit rate than.

Commercial compounds and
both were much higher.

Maybe ten fold for this one.

The natural products.

Which makes sense because natural
products are exquisitely selective for

something in the ecology and we only
picked one hundred proteins from humans

and so we shouldn't really expect that
we found the right target for that.

We found that when we look
across the hundred proteins and

ask whether compounds hit exactly one or
two to five or six or

more of the protein individual
compounds did this that the diverse and

natural compounds were in rich over
represented for highly selective binders

or in the case of natural products both
the highly selective and the media.

Selectivity binders whereas the commercial
compounds were enriched for.

Four nonselective binders and under and
under represented for selected binders.

This was just using compounds
that bound at least one protein.

We look at all the compounds again so
hit and not hit compounds and

now regardless of what source.

They came from natural or otherwise.

But instead mapping onto our stereo genic.

Ratio.

So compounds with no stir genic centers up
to a quarter of the carbon tax enters and

more than that.

We see that the most complex compounds
again also in rich in natural products or

in rich for non binders.

The simplest compounds are enriched for
permissiveness Viner's and

the intermediate complexity favors.

Favor selected binders.

More recently and we also analyze the same
data set with respect to the P.M.I.

plots that I showed you a couple times and

we do find that there is this
significant enrichment for.

The specific behaviors among sphere
like the significant enrichment for

severely compounds among.

Selective binary than.

Rotten dislike compounds among
the more promiscuous by.

OK So this this is all great right.

Or chemists love this.

They were like this means our
compounds are the greatest thing

because this happened.

Medium hit rate low selectively for

these typical guys commercial
compounds that rates are too low.

Even though they have high selectively and
are compounds are good at both but

this doesn't address an important
distinction that we're new to thought

experiments real quick.

Or you can consider it one if you like.

So one hundred protein
binding ass ace right.

Suppose you find one hundred specific
compounds each one binds to one protein.

There's two extreme cases.

All the one hundred compounds that
bind exactly one protein bind

the same one protein versus
each compound that binds that

exactly one protein binds
a different protein.

So from the standpoint of
a screening collection.

If you if you like this protein and
you care about it you're happy as a clam.

Right but from the standpoint of
a building a better screening collection

this is clearly a preferred outcome.

And so we wanted a way to get at this.

This distinction.

So what's the mechanism that we're going
to use to do that and we turn to not and

probably not surprisingly to some of you.

The Shannon information entropy as
a measure of outcome diversity.

Since it's going to doing us in in
the single assaye case you can imagine for

a binary ass a non hit you
would get the best answer when.

Half your compounds were heads and
we'll come back to that in a minute

because that's probably
not a good situation.

We're using this as in more of
a joint entropy sense for profiles.

So.

You.

Just for those of you who may be
less familiar quickly the intuitive

idea here is that when you have different
states that are symmetrically disposed

the entropy of the thing calling a die or
whatever is is simply this

is the log in whatever base you're using
I'm using two of the number of different.

Of the number of different states.

It's a log of fifty two logs and so on.

Right.

In the unfair case there's a distribution
of outcomes depending on how unfair it

is and it's correspondingly more complex
when multiple states are possible and

joint entropy is defined here for
performance profiles

as OK now we've got a bunch
of basically unfair.

Points that are separate from each
other and what we're going to

be computing is the probability
of getting a particular pattern

across all one hundred proteins
versus other proteins other patterns.

And so this joint entropy measures how
uniformly the states are distributed.

And so
we can kind of walk through this and

context of screening experiments
in this toy example.

So consider three binary assaye readouts
there's eight things that can happen.

If you've got one compound that did
each of these things you'd maximise.

The the joint entropy.

Indeed any uniform distribution of
compounds across the years gives

that answer.

If you're missing States
you you pay a cost.

Certain patterns didn't happen if
you dilute with an active which

is now getting closer to
our real situation or

we have most compounds that don't
do anything you get another.

Penalty.

So this performs well in our in our
thought experiment in a sense in the sense

you know if you have one hundred
proteins they all bind of

one hundred compounds at all by the
protein you get you have no information.

Whereas And in the case where you
bind a different protein you.

OK so.

There's two issues with this one.

That are in the reality of applying
into our data and the first one is this

entropy doesn't care how much you
prefer a given state over another.

So this these two examples
have the same joint entropy.

But this one is reflecting a situation
where you had these specific states

one zero zero zero one zero and
zero zero one having outcomes in them.

If you put the same distribution of
outcomes on states that are less specific

like this permissive us compound
that you get the same answer.

So we need some way to deal with that and
we actually.

The so far have.

Addressed this by scoring separately
sets of profiles with the same number of

non-zero entries and
then computing a weighted sum that

emphasizes more contributions
from the more specific states.

The other one has to do with missing data.

Of course entropy is perfectly happy
to reward you for having missing data

which it should do because you don't
you haven't measured it yet so

you actually don't know yet whether.

Something's going to happen or

not to your level of uncertainty
is higher in the advanced

advance of making a measurement and
we haven't addressed this yet.

We simply show both that both missing
data sensor to zero to inactive and

missing data not sensor to an active
which gives a higher answer.

So again coming back from two to you know
one hundred protein binding assaye is

our observation is that in fact while D.C.
did have.

More enrichment for specificity it had
lower joint entropy than C C it which

means that if you see in these asses with
this particular hundred proteins was

concentrating its effects
on a subset of proteins or

patterns of protein relative to C C which
was accessing more different patterns

to see how general this was we looked at
the exact same compounds across all data

known for these compounds and can bang.

So this is functional assaye data.

It's a sparse matrix because
different compounds have

seen different numbers of assayers each
compound here and seen at least fifty one

ass A's no more than one hundred
fifty four and here's the case where.

Untested is considered different than
inactive so the sort of inflated.

Caister issue to that I described and
again D.C.

is is slightly lower and
the result doesn't change really

the relative result from these three
classes doesn't really change when you.

Untested equal to an active.

OK.

So this means that despite
the excitement about

the diverse compounds
being more specific and.

Hit rate higher.

It's not fair to say that they're
a better collection as a collection.

Necessarily than C.C..

In these in these particular assay.

So last piece.

How can we get this back to what
the chemist should be making.

So here.

What we've done is to take the joint
entropy notion on subsets of compound.

Of equal size centered around different
areas in a molecular weight distribution.

So we can ask does it matter
whether compounds are lighter or

heavier with respect to
the question of whether they

perform better as a group
as judged by joint entropy.

So this is just one example where

I've taken a window size of
a little over a thousand compounds.

And computed either the unweighted
OK I work to three syphon and

so I take the science stuff as go.

So and computed either the unweighted or
the specificity weighted.

Joint entropy for that group of
compounds in these same half days.

Relative to random samples of the same
size to give you a little bit of an idea

to the on the samples
it's about point one.

So you can consider like.

Heights that are inside of that
maybe aren't as meaningful.

But.

But what this shows is that.

For an waited for
an weighted by specificity getting

actually heavier compounds helps
you with performance diversity.

But if you wait by specificity that is the
more specific cases you're actually doing.

Worse.

And that was true for the for the.

That's And the.

I'm sorry.
This should

be functional data not the specificity
weighted That's a copy paste mistake.

That's true.

More for the cellular assayers
than our protein binding Athanase

which kind of makes sense
in terms of compounds.

Either getting in or having getting into
cells more easily when they're bigger

increase your but
then having nonspecific effects when you.

I don't know if that's what's
happening but it's consistent.

But if you imagine chemists now making
molecules and wanting to know as

a collection should I be aiming down here
in the molecular weight range or up here.

We can give them guidance now on what
prior collections have done when we viewed

molecular weight or viewed performance
diversity as a function of like a way.

And we can look at that also for other
properties like calculated log P. your.

Which shows that until you
get to a certain point.

For for at least for

the unweighted case you're really
performing relatively poorly.

But that in the weighted case.

That that's a muted.

But similar trend.

We can do this for
our own hybridization ratio.

We can do it for distance from the P.M.I.
sphere so any descriptor that we can

quantify and
group compounds relative to positions

in a chemical property access
we can ask whether it helped or

not yet to be at a particular position X.
So there's this is how we'll be engaging

chemists in the future when we talk about
prior performance of compound collections.

And how they might tuner libraries to
have properties that fall into sir.

Ranges based on these
these kinds of analysis.

OK So that's it.

So I think I've covered that new
sources of compounds are needed for

disease biology.

That.

Are chemists want to learn which
new compounds to make that.

Small molecule performance can be used

as not only an alternative measure
of small molecule similarity.

But also as a means to guide
synthetic chemistry activities

when viewing performance diversity for
collections.

I believe that's yes.

And to end all thank my team

who are comprised of a very
talented set of about half and

half postdocs and staff scientists
at the brode we have and have had

some very nice visiting scholars and some
are students who contributed to this work.

As well as our collaborators in the rest
of the chemical biology program a platform

at the brode cancer program and
imaging platform and.

Obviously and I H. and other disease
foundations for funding you for

your attention and hopefully we
still have time for questions.

Thank you.

So for the it's a great question.

So for.

These ones where we were
working with the lips.

With a chemist directly in
a really small library.

We made.

So these once and
these ones we made several and

actually made Gianni and Daniela stare
at them and say this one I believe.

So that that was very chemistry
have been small enough to do it.

No we show them several of each like we
get we gave them like eight minimum energy

ones from a piece of software and let them
decide yeah I really believe this one or

whatever based on you know they had
been doing exploratory and a Mars an X.

ray during their synthesis they didn't
have that data for every compound so

we didn't want to use X.
ray for some and not others.

But they had a good sense for
how these guys were folding up so.

We did that in this case in the case
of the large collection where

we're looking at.

Fifteen thousand compounds.

We took like these we took thirty.

We asked for up to thirty two low
energy come from Earth from the cam X.

on plug in a pipeline pilot.

And then kept those within
three of the lowest and

then took that actually then took
the median values of the three.

P.M.I. coordinates and used the ratio
of medians to do these plots because

we really had no there's no way we're
going to inspect these yet this many.

Lose.

Same thing.

Yet.

The other is everybody.

And the above were so wrong about he

says one of them was this.

You know.

Rocky plays this by now.

Are you close to the spot the White House.

Is certainly possible but I'm sure shaped.

Chemical areas.

As you like.

You know what spots on yours yours.

Yeah.

You know and
you could actually live rocks or

civil arteries was you know
is what you will doctors.

You know.

It's really not guys yours or
couple who should.

Yes So there's a couple of avenues
here I think one that's been explored

in the literature and
a little bit by us playing around is the.

Charge density moments
instead of mass moments and

so that gets you part of the way there.

I think what a lot of people have.

In the Kevin for manic field been
thinking about is for is three or

four point pharmaco for.

Objects with pseudo atomic features
that are ranged in space that.

May.

So when you're dealing with a protein
that's pretty easy to think about how

you're going to orient them.

But when you don't with a library
then of the orientation problem where

you've got compounds that have one or
more of these and

you have to decide before you can
start computing a similarity on that

across all compounds you have to
decide how they should be oriented.

If you like.

Like.

Who pays orders.

So it has been through space and
through bond.

So yes sure.