Mark Mackey, Cresset, 'Meet Molecular Architect, A new product for understanding SAR and gaining better activity predictions'

Meet Molecular Architect

Dr Mark MackeyChief Scientifc Officer

Outline

> Fields, Field points and the good things you can do with them

> The alignment problem

> 3D-QSAR using Fields

> Examples> SARS PLpro – small data set, known xtal structure

> NK3 – large data set, unknown xtal structure

NN

Br

F FF

SH2NO

O

Field Points

Condensed representation of electrostatic, hydrophobic and shape properties (“protein’s view”)

> Molecular Field Extrema (“Field Points”)

3D Molecular Electrostatic

Potential (MEP)

Field Points= Positive = Negative= Shape= Hydrophobic

2D

Field Points have lots of applications

> Virtual screening

> Alignment

> Pharmacophore elucidation

> Bioisosteres

> etc

Field Points have lots of applications

> Virtual screening

> Alignment

> Pharmacophore elucidation

> Bioisosteres

> etc

> What about 3D QSAR?

The Alignment Problem

> Historically very difficult

> Early approaches template-based> Issues with side chain orientations

> Some success with docked data sets

> Easy to fool yourself> Correlation/causation

Alignment issues

> Ligand-centric view vs protein-centric

Cramer, JCAMD, 2010, DOI 10.1007/s10822-010-9403-z

NN

Br

F FF

SH2NO

O

NN

Br

F FF

SH2NO

O

Which is better?

> “The superior statistical qualities of 3D-QSAR models based on poses that superimpose presumably critical ligand features, rather than docked conformations.” Clark R., JCAMD 2007, p587

Doweyko, J. Comp-Aided Mol. Des., 2004, p 587

Free alignment adds signal, but also noise. Worse statistics, better predictability?

N-methyl acetamideImidazole

Field Alignment

H3N

N

OO

HN

O

NH

NH2H2N

O

NH

NH

HN

SOO

N-methyl acetamideImidazole

Field Scoring

Cheeseright et al, J. Chem Inf. Mod., 2006, 665

To score a particular alignment, we use the field points of molecule 1 to sample the actual field of molecule 2

Field Scoring

N-methyl acetamideImidazole

Cheeseright et al, J. Chem Inf. Mod., 2006, 665

To score a particular alignment, we use the field points of molecule 1 to sample the actual field of molecule 2 and vice-versa

Field Sampling

Field-point based QSAR descriptors

Field Sampling

Field-point based QSAR descriptors

Advantages

> Many fewer sample points than grid-based methods

> E.g. Vegfr2 data set

Field Sampling Grid Sampling Filtered Grid Sampling

q2 0.62 0.42 (0.48) 0.40

Number of descriptors

466 3940 1243

Du et al., J Mol Graph Model. 27 (2009) 642-652

Advantages

> Many fewer sample points than grid-based methods

> Sample points physically rather than statistically chosen

> Gauge invariant

> Consistent framework for alignment and QSAR

Initial validation

> Tested against literature CoMFA datasets> 15 datasets with alignments available

> CoMFA average cross-validated RMSE is 0.72

> Field QSAR using CoMFA alignments is 0.74

> Simple model (volume indicator variable) is 0.83

> Data sets re-aligned using field alignment> RMSE 1.00

Interpretability

Electrostatic Steric Variance

SARS PLpro

The target

> PLpro (Papain-like protease) is a DUB target which is critical for the replication of the coronavirus responsible for SARS

> Crystal structures available with bound ligands from 2 series of compounds: structurally related (PDB entries 3E9S and 3MJ5)

> Small number of analogues – challenge to see if we can use 3D-QSAR for small data sets

Alignment

Sampling points

Model

3.50 4.00 4.50 5.00 5.50 6.00 6.50 7.003.50

4.00

4.50

5.00

5.50

6.00

6.50

7.00

R² = 0.838410392100944

R² = 0.987121600034644Training SetLinear (Training Set)Test SetLinear (Test Set)

PLS Components = 5 RMSE = 0.09 RMSEP = 0.38

Summary

> Able to build a predictive 3D-QSAR model based on small number of analogues

> Guided (by volume of Xtal structure) alignment worked best. Free alignment was OK, but noisier.

NK3 antagonists

NK3 example

> GPCR target (Tachykinin receptor 3) – selectively binds Neurokinin B – target for treatment of neurological disorders such as schizophrenia

> Three series of inhibitors from Euroscreen> Scaffold-1 – 81 compounds with pIC50 (radioligand

binding) in range 4.6-8.7

> Scaffold-2 – 80 compounds with pIC50 in range 4.8-7.7

> Errors in radioligand binding data c. ± 0.4

NK3 binding mode

> For a 3D method you need a 3D alignment

> FieldAlign can align to a reference

> FieldTemplater generates the reference

O

N+

NH

N

HO

O

O

OHO

H

N

N+

N

O

N

N

O

H

F

F F

F

F

O

HN N+

H

H

N N

NH

FieldTemplater

NK3 binding mode prediction

> FieldTemplater> Selection of 3 highly active scaffold-1 compounds

plus 2 structurally dissimilar literature NK3 actives (Talnetant and SB-218795).

> Generated Templates filtered and candidate selected

> Conformation of most active scaff-1 structure then used as alignment target for other structures

3D-QSAR details

> Alignment> Free alignment to template conformation

> Field selection> Generated Field points for both steric and electrostatic

fields, with both sets at independent locations.

> 80/20 training/test split > Most active and least active training set

> 2nd most active, 2nd least active test set

> Random distribution of remaining compounds

Initial models problematic

> When all else fails, talk to the chemists

> “Are you using the right tautomer?”

N N

N

N

H

N N

N

N

NK3 Series 1

RMSE 0.19, RMSEpred 0.64

4.5 5 5.5 6 6.5 7 7.5 8 8.5 94.5

5

5.5

6

6.5

7

7.5

8

8.5

9

Training Set

Test Set

NK3 Series 1

Electrostatics Sterics

Extend to scaff-2?

4 5 6 7 8 9 10 114

5

6

7

8

9

10

11

Actual vs Predicted - scaff2 compounds on scaff1 model

> Complete lack of predictivity

> Visual analysis suggests a shift in binding mode for scaff-2

> Cross-series QSAR difficult

> Requires consistent binding modes!

NK1 Scaffold 2

4 4.5 5 5.5 6 6.5 7 7.5 8 8.54

4.5

5

5.5

6

6.5

7

7.5

8

8.5

TrainingTest

Activity

Pred

icted

Acti

vity

RMSEpred 0.60

Summary

> Able to generate models based on alignment to predicted active conformation by templating

> Independent models within each of two series show reasonable predictivity and can be used to guide further work

> Cross-series analysis suggests different binding modes for the two series

Molecular Architect

Molecular Architect

> Initially FieldAlign + QSAR

> Align your molecules

> Build models

> Test models

> Fit new compounds to models

> Interactive feedback

> Add additional alignment options

Molecular Architect

> One tool for molecule designers> Align

> QSAR

> Pharmacophore elucidation

> Bioisosteres

> What do I make next?

> Beta Q4 2011

Acknowledgements

> Cresset> Andy Vinter

> Tim Cheeseright

> James Melville

> Chris Earnshaw

> Euroscreen> Hamid Hoveyda

> Julien Parcq

[email protected]