Violeta Pérez-Nueno, David Ritchie · Violeta Pérez-Nueno, David Ritchie Orpailleur Team, INRIA...

11/31/3111ESCUELA TÉCNICA SUPERIOR

Using “Consensus clustering” to help understand how multiple ligands might bind

to multiple target sites. How this helps to choose the Right Query in Shape-Based Virtual Screening

Violeta Pérez-Nueno, David RitchieOrpailleur Team, INRIA Nancy - Grand Est

LORIA (Laboratoire Lorrain de Recherche en Informatique et ses Applications), INRIA Nancy – Grand Est, 615 rue du Jardin Botanique, 54506 Vandoeuvre-lès-Nancy, France

22/31/3122

1. Calculating SH Shapes

2. Calculating SH Consensus Shapes and their performance in VS

3. Consensus Clustering: Exploring CCR5 Multiple Binding Sites

4. All-against-all DUD Dataset Shape Clustering vs DUD Cross Docking

5. Choosing the Right Query in DUD Shape-Based VS

6. Examples of how Consensus Clustering can identify Multi-Site Pockets

Presentation Overview

33/31/3133

1. Calculating SH Shapes

44/31/3144

• Real SHs: • Coefficients:• Encode radial distances

from origin as SH series…• Solve coefficients by

numerical integration…

Surface shapes are represented as radial distance expansions of the molecular surface with respect to the center of the molecule.

Spherical Harmonic Surfaces

Ritchie, D.W. and Kemp, G.J.L. J. Comp. Chem. 1999, 20, 383–395.

55/31/3155

• Distance:

• Orthogonality:

• Rotation:

• Carbo:

• Hodgkin:

• Tanimoto:

• Multi-property:

ParaFit

Ritchie, D.W. and Kemp, G.J.L. J. Comp. Chem. 1999, 20, 383–395.

ParaFit calculates superpositions between pairs of molecules by exploiting the special rotational properties of the SH functions

66/31/3166

2. Calculating SH Consensus Shapes and their performance in VS

77/31/3177 Pérez-Nueno et al. J. Chem. Inf. Model. 2008, 48, 2146–2165.

1. Do all-v-all SH comparison2. Find best pair-wise match3. Calculate SH average of pair4. Treat average as new seed5. Superpose all onto seed6. Compute new average seed7. Rotate all onto new seed8. Iterate until convergence...9. Result = SH pseudo-molecule

Calculating Consensus Shapes

88/31/3188

CCR5 Antagonists (424):1) SCH-C derivatives 2) 1,3,5-trisubstituted pentacyclics3) Diketopiperazines4) 1,3,4-trisubstituted pyrrolidinepiperidines5) 5-oxopyrrolidine-3-carboxamides6) N,N’-Diphenylureas7) 4-aminopiperidine or tropanes8) 4-piperidines9) TAK derivatives10) Guanylhydrazone drivatives11) 4-hydroxypiperidine derivatives 12) Phenylcyclohexilamines13) Anilide piperidine N-oxides14) 1-phenyl-1,3-propanodiamines15) AMD derivatives16) Other

CXCR4 antagonists (248):1) AMD derivatives

2) Macrocycles

3) Tetrahydroquinolinamines

4) KRH derivatives5) Dipicolil amine zinc(II) complexes6) Other

PLUS…4696 inactive compounds from the

Maybridge Screening Collection with

similar 1D properties to the actives

Virtual Screening DatasetsThe consensus approach was first validated in a retrospective VS of two databases containing CCR5 and CXCR4 active compounds from the literature, and presumed inactive compounds from Maybridge.

99/31/3199

SH Consensus Shapes of theThree Most Active Inhibitors

Pseudo-molecules were obtained from the consensus shapes of the most active molecules for both CXCR4 and CCR5 targets, and used as VS queries against the database of known actives and decoys.

CCR5

CXCR4

Consensus shape KRH derivate superposition

Macrocycle derivate superposition

AMD derivate superposition

Consensus shape 1,3,4-trisubstituted pyrrolidinepiperidinederivate superposition

SCH derivate superposition

Piperidine derivate superposition

1010/31/311010

CXCR4

CCR5

Pérez-Nueno et al. J. Chem. Inf. Model. 2008, 48, 2146–2165.

Consensus Shape-Based VS

CXCR4 families very similar shapes -> results do not substantially improve

CCR5 families verydifferent shapes -> results considerably improve

1111/31/311111

3. Consensus Clustering: Exploring CCR5 Multiple Binding Sites

1212/31/311212

• Wards hierarchical clustering of chemical fingerprints

• We then used Kelley’s method to find the optimal number of clusters (16)

• These were manually merged to 10 groups based on known CCR5 families

•SH consensus shapes were calculated for the 10 groups

• These were then compared in ParaFit (all-vs-all)

Exploring CCR5 Multiple Binding Sites: Clustering the 424 CCR5 Ligands

1313/31/311313

From Consensus Shapes to Super-Consensus Clusters

•Another round of Ward’s clusteringproposed four super-consensusclusters

1414/31/311414

• Each SC pseudo-molecule was used as a VS query:

• NB. merging SC shapes significantly worsens the AUCs…• SC queries => CCR5 ligands form no less than FOUR groups

Using Super-Consensus Shapes as VS Queries

AUC=0.785 AUC=0.413

AUC=0.905 AUC=0.630

VS Super Consensus A VS Super Consensus B

VS Super Consensus C VS Super Consensus D

1515/31/311515

• SC-A docks to Site-1

(TMs 1, 2, 3, 7)

• SC-C docks to Site-2

(TMs 3, 5, 6)

• B and D dock to Site-3

(TMs 3, 6, 7)

• 3D pseudo-molecules were created as the union of all superposed ligands in each SC family for docking in Hex

Hex Blind Docking of SC Pseudo-Molecules to CCR5


1616/31/311616

• To confirm that the SC shapes were properly matched to their predicted target sites, the three proposed binding sites were treated as if they were separate targets for docking-based VS:• SC-As treated as actives for Site 1 (SCs B, C, D treated as inactives)• SC-Cs treated as actives for Site 2 (SCs A, B, D treated as inactives)• SC-B/Ds assumed active for Site 3 (SCs A and C treated as inactives)

• As before, merging SCs worsens the AUCs…• SC docking => no less than THREE CCR5 pocket sub-sites

A -> Site-1 C -> Site-2

B,D -> Site-3

Hex Docking VSw.r.t. Three CCR5 Sub-Sites


AUC=0.834 AUC=0.963

AUC=0.846

Docking VS onto Site 1 Docking VS onto Site 2

Docking VS onto Site 3

1717/31/311717

4. All-against-all DUD Dataset Shape Clustering vs DUD Cross Docking

1818/31/311818

All-against-all DUD Dataset Shape Clustering vs DUD Cross Docking

Cross-docking on DUDHuang et al. work

Huang et al. J. Med. Chem. 2006, 49, 6789-6801.

1919/31/311919


Cross-shape matching on DUD• All against all shape comparison for the ligands of the 40 DUD targets

• Clustering of ligands by shape into 40 groups

• Select a threshold that emphasizes the overall similarity to the cross-docking experiment

•Count the membership over the threshold for ligands belonging to each of the targets for all the 40 clusters and sum memberships.

•Normalize this sum by the total number of ligands for each target. This gives the cross-shape matching score.

1234...

.

.

.

40

When ace is over the threshold. Sum of the normalized number of ligands of each of the other targets over the threshold for all the clusters

2020/31/312020

Ligands for a given target are split into 2 or 3 sub-groups, which could suggest they might bind to different sub-sites of the same target

p38 ligands split in C36, C30, C29


2121/31/312121

Ligands from different targets are grouped together. It suggests that they could bind to the same targets.

Evidence of cox2 cross-docked with hsp90

Evidence of sahh cross-docked with tk and ada


2222/31/312222


8.45 h several months (from 20h to 12d/target)

Huang et al. J. Med. Chem. 2006, 49, 6789-6801.

2323/31/312323

5. Choosing the Right Query in DUD Shape-Based VS

2424/31/312424

Choosing the Right Query in DUD Shape-Based VS

Queries used:

• PARAFIT_B : PARAFIT bound conformation.

• PARAFIT_CM : PARAFIT consensus molecule.

• PARAFIT_C1_CM - PARAFIT_C3_CM: consensus molecules for SH clusters 1, 2, 3.

• PARAFIT_RCM: PARAFIT real center molecule closest to consensus.

• PARAFIT_C1_RCM - PARAFIT_C3_RCM: real center molecules for clusters 1, 2, 3.

• PARAFIT_SHEF_C: best SHEF molecule that fits pocket as PARAFIT query.

• ROCS_B: ROCS bound conformation.

• ROCS_RCM: real center molecule from PARAFIT as ROCS query.

• ROCS_C1_RCM - ROCS_C3_RCM: real PARAFIT center molecules as ROCS queries.

• ROCS_SHEF_C: best SHEF molecule that fits pocket as ROCS query.

• SHEF shape-based docking.

• GOLD conventional docking.

ROC plots Bar graphs

2525/31/312525


2626/31/312626


2727/31/312727

6. Examples of how Consensus Clustering can identify Multi-Site Pockets

2828/31/312828

Multi-Site Pockets: p38• 353 p38 DUD ligands clustered using Wards hierarchical clustering of chemical fingerprints

• Kelley’s method to find the optimal number of clusters (15)

• SH consensus shapes were calculated for the 15 groups

• These were then compared in ParaFit(all-vs-all)

• Another round of Ward’s clustering proposed three super-consensus clusters

2929/31/312929

Multi-Site Pockets: p38

Glu 71

Asp 168

Met 109

Glu 71

Asp 168Met 109

SC_A SC_B SC_ C

Glu 71

Asp 168

Met 109

3030/31/313030 Pargellis et al. Nature Structural Biology 2002, 9 (4), 268-272.


SCH3

ON

N

HN

I

ATP

1

2

2

BIRB

1

1

3131/31/313131


Glu 71

Asp 168

Met 109

Glu 71

Asp 168Met 109

SC_A SC_B SC_ C

Glu 71

Asp 168

Met 109

ATP

1

21

SITE 1 SITE 1SITE 2

SITE 2SITE 1

SITE 2

SITE 1

(Allosteric site)(ATP site) (ATP site)

(ATP site)(ATP site) (Allosteric site)

(Allostericsite)

diaryl urea inhibitors

pyridinyl-imidazole inhibitors

SC_A

SC_B

3232/31/313232

Multi-Site Pockets: Alr2• 26 alr2 DUD ligands clustered using Wards hierarchical clustering of chemical fingerprints

• Kelley’s method to find the optimal number of clusters (5)

• SH consensus shapes were calculated for the 5 groups

• These were then compared in ParaFit (all-vs-all)

• Another round of Ward’s clustering proposed three super-consensus clusters

3333/31/313333

Multi-Site Pockets: Alr2

SC_A SC_B SC_ C

Tyr 48His 110

Trp 111

Thr 113

Leu 300 Cys 298

His 110

Trp 111

Thr 113

Leu 300

Tyr 48

Cys 298

His 110

Trp 111Thr 113

Leu 300

Tyr 48

Cys 298

3434/31/313434


3

2

(d) (e)

Steuber et al. J. Mol. Biol. 2007, 369, 186-197.

3535/31/313535


SC_A SC_B SC_ C

Tyr 48His 110

Trp 111

Thr 113

Leu 300 Cys 298

His 110

Trp 111

Thr 113

Leu 300

Tyr 48

Cys 298

His 110

Trp 111Thr 113

Leu 300

Tyr 48

Cys 298

Small ligand Sub-site 1 (catalytic cleft)

Super ligand Sub-site 2 (whole pocket)



Small ligand Sub-site 1 (catalytic cleft)

3636/31/313636

• Consensus shape provides good queries. It is also a good

way to select a center molecule.

• Cross-shape matching is much faster than cross-docking in

identifying possible cross-docked targets.

• Consensus clustering allows to detect multi-site targets.

• Multi-site targets are more common than previously assumed.

We recommend clustering ligands if poor performance is

observed with the single query.

Conclusions

3737/31/313737

Acknowledgements

• Dave Ritchie• Vishwesh Venkatraman• Lazaros Mavridis

• INRIA Nancy - Grand Est• IQS, Universitat Ramon-Llull

ParaSurf + ParaFit: http://www.ceposinsilico.de/

Papers: http://www.loria.fr/~pereznue/http://www.loria.fr/~ritchied/

• Agence Nationale de la Recherche (ANR-08-CEXC-017-01)• Generalitat de Catalunya-DURSI (FI2008 and BE-DGR2009)

3838/31/313838

Thank you!

3939/31/313939

• From MOPAC or VAMP, calculate:– Density contours of 2x10-4e/A3 (i.e. approx = SAS)– MEP – electrostatic potential– IEL – ionization energy– EAL – electron affinity– aL – polarizability

• Encode as Spherical Harmonic expansions to order L=15…

Lin, J.-H. and Clark, T. J. Chem. Inf. Model. 2005, 45, 1010–1016.

Shapes/Properties From Semi-Empirical QM

ParaSurf

4040/31/314040


0.8920.9040.6890.9600.908RXR_ALPHA

0.8810.9470.9350.9450.919COX2

0.9230.9310.8710.9440.859SAHH

0.6420.4520.8040.7690.572ACE

0.5100.7690.6610.7520.524CDK2

0.5810.7160.6520.7580.574COX1

0.6760.7410.7520.7970.591PR

0.7750.8080.4260.8060.623THROMBIN

0.6290.5560.4720.6090.598VEGFR2

0.7530.8090.5520.8250.498p380.5790.6780.6670.7420.448PDGFRB0.6390.6390.7400.7280.207TRYPSIN0.3780.6740.5080.3510.371COMT

0.4190.6170.5200.6090.508ALR2

0.630(B) 0.413(D)

0.505 (B+D)

0.9050.7850.8660.594CCR5

PFT_C3_CMPFT_C2_CMPFT_C1_CMPFT_CMPF_BTarget

POOR

POOR

POOR

POOR

POOR

RANDOM

RANDOM

RANDOM

RANDOM

RANDOM

RANDOM

SUCCESSFUL

SUCCESSFUL

SUCCESSFUL

4141/31/314141


4242/31/314242


Violeta Pérez-Nueno, David Ritchie · Violeta Pérez-Nueno, David Ritchie Orpailleur Team, INRIA...

Documents

Transcript of Violeta Pérez-Nueno, David Ritchie · Violeta Pérez-Nueno, David Ritchie Orpailleur Team, INRIA...