design in 2d, model in 3d: live 3d pose generation from 2d ...€¦ · a simple flow chart will...

Design in 2D, model in 3D:live 3D pose generation from 2D sketches

Paolo Tosco

© Cresset

> Description of the grow3D workflow

Outline

> Introduction to the project

> Application to a case study

> Conclusions and outlook

© Cresset

How nice would it be…

…to design a molecule in the 2D sketcher and see it grow sensibly within the binding site in the 3D viewport?

Magenta:co-crystallizedcelecoxib in theCOX-2 active site

Grey: growing3D design

© Cresset

How hard can it be?

> Start from a blank 2D sketcher canvas

© Cresset

How hard can it be?

> Start from a blank 2D sketcher canvas> sketch something

© Cresset

How hard can it be?

> Start from a blank 2D sketcher canvas> sketch something> the largest 2D fragment is popped to a 3D

conformation…

© Cresset

How hard can it be?


conformation…> …and docked into the protein’s active site

© Cresset

How hard can it be?


conformation…> …and docked into the protein’s active site> …or aligned against a reference using a

combination of 3D fields and shape

© Cresset

How hard can it be?



combination of 3D fields and shape> using the protein as excluded volume (if

available)

© Cresset

How hard can it be?



combination of 3D fields and shape> using the protein as excluded volume (if

available)> then the grow3D process begins

© Cresset

Outline




> Description of the grow3D workflow

© Cresset

A simple flow chart will help: start from blank

We’ll keep it simple and start from a single fragment added to a blank 2D canvas

Look at disconnected fragments which

make up the sketch

YES

NO

Has the number of 2D

frags changed?

NO

YES Has the number of2D frags

increased?

No fragments

NH

N

© Cresset

A simple flow chart will help: generate 3D or grow3D?

Has the number of2D frags

increased?

Bonds/atomswere deleted

NO

N H 2

N

NH

N +

Has thetotal number

of atoms increased?

YES

No fragments

NH

N

Generate 3D structure for the

new 2D frag

YES

No atoms

NH

N

Frags were combined or deleted

NO

N H 2

N

NH

N+ or N

N H 2

Has the number of2D frags

changed?

Dock into binding siteor align to reference

grow3DNO

N

N

NH

N

© Cresset

Initial placement of the 3D design in the 3LN1 pocket

Co-crystallizedcelecoxib(magenta)

Aligned 3-methyl-pyrazole (grey)

© Cresset

A simple flow chart will help: growing the 2D design

Map current 2D frags to previous 2D frags

NO

No action

Loop over 2D frags: does this frag have 3D

coords?

N

N

NH

N

YES Pair current frag 2D indices

to previous

Compare# indices of current frag

with previous

Let’s add a methyl group in position 1 on the pyrazolesystem

>

Atoms were added

N

N

<

Atoms were deleted

NH

N

=

Check: • Element• Bonds• Bond order/

hybridization• Formal charge• Stereo

NH

N

© Cresset

A simple flow chart will help: growing the 3D design

A methyl groupwas added

NH

N

N

N

• Find bonds to added atoms

• Delete those bonds and cap with hydrogens

C H 4

NH

N

N

N

+

Generate 3D coords for newly added 2D frag

C H 4 Align frag Hs with parent’s heavy atoms and vice versa

Ready for the next grow3D iteration!

Attach to parent with appropriate bond order

Methyl was easy, can I have more of those, please?

© Cresset

Align frag Hs with parent’s heavy atoms and vice versa

1

23

A simple flow chart will help: scoring 3D designs

A phenyl ringwas added



N

N

N

N

+


N

N

N

N

© Cresset

A simple flow chart will help: scoring 3D designs

Attach to parent with appropriate bond order

Spin around bond and score against reference

1

23Rank models by score and do a local optimization

Align frag Hs with parent’s heavy atoms and vice versa



N

N

N

N

+


N

N

N

N

C C

A phenyl ringwas added

© Cresset

The devil is in the details (1): symmetries

When adding am-isopropyl substituent in 2D, there are actually two symmetry-equivalent positions it might fit in 3D

Let’s add it onone side first…

© Cresset


…and then onthe other side

When adding am-isopropyl substituent in 2D, there are actually two symmetry-equivalent positions it might fit in 3D

© Cresset


No matter which side the isopropyl substituent is drawn in 2D, the grow3Dalgorithm chooses the one which fits the active site best

© Cresset

The devil is in the details (2): chemically invalid states

When we draw a sulfone-containing functional group in 2D, we tend to go through a number of chemically invalid states

© Cresset



We draw a methyl…

© Cresset



…then we keep adding methyl groups until we get to a t-butyl…

We draw a methyl…

© Cresset



…then we add the double bonds…


We draw a methyl…

grow3D pauses,waiting for our next move

© Cresset



…then we turn terminal methylenes into oxygens…



We draw a methyl…

grow3D still patiently waiting for some decent chemistry

© Cresset



…and only in the end we turn the central carbon into a sulfur

…then we turn terminal methylenes into oxygens…



We draw a methyl…

grow3D applies the change

© Cresset

Outline


> Description of the workflow



© Cresset

A case study: CHK1 inhibitors

In 2018 AZ published an interesting scaffold morphing exercise on some Checkpoint 1 kinase (CHK1) inhibitors they had previously identified

© Cresset

A (very) quick introduction to CHK1

> CHK1 is a promising target to improve the therapeutic index of DNA-damaging anti-cancer agents

> DNA damage triggers CHK1 activation, which in turn arrests the cell cycle, thus allowing DNA repair to take place

> Inhibiting CHK1 abrogates the cell cycle arrest, causing apoptosis

> CHK1 inhibitors would thus sensitize tumour cells to the action of DNA-damaging drugs

© Cresset

TCU and TZQ leads

ONH 2

NH

N HNH

O

F

S

1 (AZD7762)Thiophene carboxamide

urea (TCU)Clinical candidate

2Triazoloquinolone

(TZQ)

The carbonyl group interacts with the backbone N-H of Cys-87

The amino group interacts with the backbone C=O of Glu-85

3Thiophene carboxamide

urea (TCU)Matched pair with 1

Both 1 and 3 feature an intramolecular hydrogen bondwhich stabilizes the bioactive conformation

NO

NH

N N

N HNH

O

NH

ONH 2

S

© Cresset

Ring closure, ring opening

ONH 2

NH

N HNH

O

F

S

1 (AZD7762)Thiophene carboxamide

urea (TCU)Clinical candidate

2Triazoloquinolone

(TZQ)

S

FNH 2 O

N

NHN H

44-Carboxamide

thienopyridine (4-CTP)

Ringclosure

Ringopening

NO

NH

N N

57-Carboxamide

thienopyridine (7-CTP)

Matched pairsN H

NH

N

ONH 2

SF

© Cresset

Ring closure in grow3D

ONH 2

NH

N HNH

O

F

S

1 (AZD7762)

S

FNH 2 O

N

NHN H

4 (4-CTP)

Ringclosure

Let’s see what grow3D can do about the ring closure of 1 to 4

I’ll use the 2YDJ PDB structure (CHK1 co-crystallized with 1) as a reference

© Cresset

Positioning the initial 3D structure

ONH 2

NH

N HNH

O

F

S

1 (AZD7762)

S

FNH 2 O

N

NHN H

4 (4-CTP)

Ringclosure

I’ll start by generating a 3D conformation of 1 and aligning it to the X-ray reference

In the background, multiple 3D conformations of 1 are generated, and the one with the highest field/shape similarity score to the reference is chosen

© Cresset

Scaffold morphing begins

ONH 2

NH

N HNH

O

F

S

1 (AZD7762)

S

FNH 2 O

N

NHN H

4 (4-CTP)

Ringclosure

Then I turn the carbonyl group into an imine…

© Cresset


ONH 2

NH

N HNH

O

F

S

1 (AZD7762)

S

FNH 2 O

N

NHN H

4 (4-CTP)

Ringclosure

…now the difficult bit: ring closure! (holding my breath)…


…the urea into an N-methylurea…

© Cresset


ONH 2

NH

N HNH

O

F

S

1 (AZD7762)

S

FNH 2 O

N

NHN H

4 (4-CTP)

Ringclosure



Whoa, that worked!

© Cresset


ONH 2

NH

N HNH

O

F

S

1 (AZD7762)

S

FNH 2 O

N

NHN H

4 (4-CTP)

Ringclosure



…let’s clean it up a little bit (both 2D and 3D)…

© Cresset


ONH 2

NH

N HNH

O

F

S

1 (AZD7762)

S

FNH 2 O

N

NHN H

4 (4-CTP)

Ringclosure




…now I turn the dihydropyrimidineinto a dihydropyridine…

© Cresset

Scaffold morphing: done!

ONH 2

NH

N HNH

O

F

S

1 (AZD7762)

S

FNH 2 O

N

NHN H

4 (4-CTP)

Ringclosure




…now I turn the dihydropyrimidineinto a dihydropyridine…

…and aromatize to pyridine: done!

© Cresset

More transformations at constant #atoms

S

FNH 2 O

N

NHN H

4 (4-CTP)

Ringflip

I will now flip the thiophene ringto turn 4 into 5 in 4 moves:

N HNH

N

ONH 2

SF

5 (7-CTP)

© Cresset


S

FNH 2 O

N

NHN H

4 (4-CTP)

Ringflip


1. Hydrogenate the thiophene…

5 (7-CTP)

N HNH

N

ONH 2

SF

© Cresset


S

FNH 2 O

N

NHN H

4 (4-CTP)

Ringflip



2. Turn it into a cyclopentene…

5 (7-CTP)

N HNH

N

ONH 2

SF

© Cresset


S

FNH 2 O

N

NHN H

4 (4-CTP)

Ringflip


5 (7-CTP)



3. Dehydrogenate cyclopentene…

N HNH

N

ONH 2

SF

© Cresset


S

FNH 2 O

N

NHN H

4 (4-CTP)

Ringflip


5 (7-CTP)




4. Turn cyclopentadiene intothiophene…

N HNH

N

ONH 2

SF

© Cresset


S

FNH 2 O

N

NHN H

4 (4-CTP)

Ringflip


5 (7-CTP)




4. Turn cyclopentadiene intothiophene…and clean it up

N HNH

N

ONH 2

SF

© Cresset

More transformations at constant #atoms: how did we do?

S

FNH 2 O

N

NHN H

4 (4-CTP)

Ringflip

This is how our designed 5compares against its experimental X-ray structure 6FC8

5 (7-CTP)

N HNH

N

ONH 2

SF

© Cresset

Ring opening in grow3D

2 (TZQ)

S

FNH 2 O

N

NHN H

4 (4-CTP)

Ringopening

Let’s try and get to 4 through a ring opening approach starting from 2

I’ll use the 2X8E PDB structure (CHK1 co-crystallized with 2) as a reference

NO

NH

N N

This morphing involves far more changes to the scaffold than the previous one

© Cresset

A more complicated design case

2 (TZQ)

S

FNH 2 O

N

NHN H

4 (4-CTP)

Ringopening

As previously, I start by aligning my 2D design to the X-ray reference

NO

NH

N N

© Cresset


2 (TZQ)

S

FNH 2 O

N

NHN H

4 (4-CTP)

Ringopening


NO

NH

N N




Turn the naphthalene into an isoquinoline

© Cresset


2 (TZQ)

S

FNH 2 O

N

NHN H

4 (4-CTP)

Ringopening

NO

NH

N N





Break the phenyl ring as I need to make it 5-term

© Cresset


2 (TZQ)

S

FNH 2 O

N

NHN H

4 (4-CTP)

Ringopening

NO

NH

N N





Turn cyclohexyl into 3-piperidyl

Final clean-up, and we’re done!

© Cresset


2 (TZQ)

S

FNH 2 O

N

NHN H

4 (4-CTP)

Ringopening

NO

NH

N N

This is how our designed 4compares against the experimental X-ray structure of 1 (2YDJ)…

ONH 2

NH

N HNH

O

F

S

1 (AZD7762)

© Cresset


2 (TZQ)

S

FNH 2 O

N

NHN H

4 (4-CTP)

Ringopening

NO

NH

N N

…and against the experimental X-ray structure of its close analogue 5 (6FC8)

5 (7-CTP)

N HNH

N

ONH 2

SF

© Cresset

Conclusions

> grow3D is an algorithm which aims at generating sensible 3D poses in response to edits to a 2D sketch

> Poses are scored against their ability to fit a binding site and/or mimic electrostatics and shape of a reference ligand

> The algorithm is enough quick and accurate to enable real time assessment of design ideas

© Cresset

The simple flow chart (to be continued)

Atoms were added• Find bonds to added atoms, note their

indices and delete them, re-adding H’s• Fragment curr frag• Loop over originated frags, find those with

identical size and 2D coordsto prev(parent)

Break the input mol into frags

Has the number of frags changed?

NO

Pair prev frag indices with identical 2D coords to currand copy across 3D coords (if any)

Does curr frag have more indices

than prev?

NO

Atoms were deleted• Delete bonds to deleted atoms

from 3D coords• Fix hybridization of previously

bonded atom if needed• Delete atoms from 3D coords

Does curr frag have less indices

than prev?

YES

Loop over curr frags Is frag 3D?

NO

No action on currfrag

YES

NO YES

Check: • Element• Bond order• Hybridization• Formal chargeApply change to 3D coords

Atoms were added• Find bonds to added

atoms, note their indices and delete them, re-adding H’s

• Fragment curr frag• Loop over originated

frags, find the one with identical size and 2D coords to prev(parent)

Loop over children frags

• Generate 3D coordsfor the child frag

• Merge child frag into parent, aligning child H’s with parent’s heavies and vice versa (underdetermined if only one bond was broken: rotational scan)

Is the child frag a single atom?

YES NOReplace parent H with child atom

Regenerate bonds to parent with correct bond order

Render 3D coords

Are there more frags?

Frags have been combined or deleted

Has the total number of atoms

changed?

NO Frags were combined

Loop over prev frags which were combined

Is any frag 3D?

No action on combined frag

NO

YES

Minimize grown frag 3D coords

YES

Has the total number of atoms

increased?

NO

YES

Does any of the parent frags have

3D coords?

NOYESGenerate 3D coordsfor parent frags if needed

YES

NO

YES Has the total number of atoms

increased?

YES

No action on newly added frag

NO

Bonds (and possibly atoms) were deleted

© Cresset

cressetgroup

[email protected]

Thank you foryour attention

Acknowledgments

@Cresset: James Foley, Harry Jubb, Mark Mackey, Tim Cheeseright

The RDKit

design in 2d, model in 3d: live 3d pose generation from 2d ...€¦ · a simple flow chart will...

Documents