John Karanicolas Our computational toolbox Structure-based - - PowerPoint PPT Presentation

Computational chemical biology to address non-traditional drug targets

John Karanicolas

Our computational “toolbox”

Structure-based approaches Ligand-based approaches

Detailed MD simulations Docking (to predict pose) Virtual screening 2D fingerprints 3D matching (and pharmacophores) Fragment replacement

Ligand-based approaches for pose prediction and virtual screening

Ligand similarity metrics

Active Compound Neighborhood Region

• ther

compounds

Glen and Bender, Org Biomol Chem, 2004

The activity of a compound is most likely to be shared by “similar” compounds… but only IF we have an appropriate way to describe similarity.

2D “Fingerprints”

Bender, Mussa, and Glen, J Chem Inf Comput Sci, 2004

Define a string of bits based on the presence/absence of particular substructure features (from a “dictionary”).

Comparing 2D fingerprints

The Jaccard index is a statistic used for comparing the similarity and diversity of sample sets.

Image detection example from Wikipedia

Comparing 2D fingerprints

Bender, Mussa, and Glen, J Chem Inf Comput Sci, 2004

Given two fingerprints, their similarity is given by the Jaccard index (aka Tanimoto score).

A = 01010111110001 B = 01100101111001 T = 0.6

What about 3D ligand comparisons for pose prediction and virtual screening?

• Enzymes (including kinases) (47%)
• GPCR’s (30%)
• Ion channels (7%)
• Nuclear hormone receptors, transporters, other

receptors (12%)

• All are proteins evolved to bind small-molecules…

Targets of marketed drugs

Hopkins and Groom, Nat. Rev. Drug Discov. 2002

Non-traditional drug targets

• Protein-protein interactions
• Protein-RNA interactions
• Protein stabilizers

A detailed case study using a (traditional) kinase example…

Overview of an active kinase

“All active kinases are alike but an inactive kinase is inactive after its own fashion.” — Nobel et al, Science 2004

Jak1 + ADP

Type I kinase inhibitors

Jak1+Tofatinib Jak1+ADP

Schematic of an active site

Ghose et al., J Med Chem, 2008

Overview of kinase inhibitors

• Type I inhibitors bind the active conformation:

selectivity is a challenge!

• Type II inhibitors bind inactive conformations;

however, this hasn’t led to selectivity

• Type III and beyond: boutique inhibitors

Hinge-binding motifs

Jak1 + ADP Jak1 + Tofacitinib

Hinge-binding motifs

Ghose et al., J Med Chem, 2008

Three of the most common hinge-binding motifs in the PDB

Hinge-binding motifs

Ghose et al., J Med Chem, 2008

Introduction to today’s challenge

• A lab has screened for compounds that provide a

certain phenotype

• After optimizing activity they ultimately arrive at

a compound with an aminothiazole core, and they find it to be a CDK9 inhibitor

• How does this compound

engage CDK9?

• Why is it selective for CDK9?

Most CDK9 inhibitors also inhibit CDK2…

MC180295 IC50 (nM) CDK1-Cyclin B 138 CDK2-Cyclin A 233 CDK2-Cyclin E 367 CDK3-Cyclin E 399 CDK4-Cyclin D 112 CDK5-P35 159 CDK5-P25 186 CDK6-Cyclin D3 712 CDK7-CycH/MAT1 555 CDK9-Cyclin T1 5.1

Introduction to today’s challenge

Traditional docking

Sialic acid binding to Hemagglutinin (generated using Autodock software)

Modeling the MC180295/CDK9 complex

• We know what active kinases look like
• We know how Type I inhibitors bind
• Why would we possibly want to use docking?!

Modeling the MC180295/CDK9 complex

Comparative modeling for kinase inhibitors

• Start from an ATP-bound structure of CDK9
• Align all other inhibitor-bound CDK-kinases to this

(there are 389 in the PDB)

• Use each of these inhibitors to position MC180295
• Refine + re-score these 389 models

• Matching of features in 3D is more predictive of shared activity than 2D

shared features

• Template compound is represented by a series of Gaussians (“cloud”)
• Clouds from query (library) compounds are individually aligned with

template cloud to check their overlap (via Jaccard index)

Comparing molecules in 3D

Shape comparison matching

Rush et al., J Med Chem, 2005

ROCS: Rapid Overlay of Chemical Structures (very fast method for virtual screening!)

H2

(white) (orange)

Comparative modeling for kinase inhibitors

• Start from an ATP-bound structure of CDK9
• Align all other inhibitor-bound CDK-kinases to this

(there are 389 in the PDB)

• Use each of these inhibitors to position MC180295
• Refine + re-score these 389 models

Extending this case study

• Can we successfully model other inhibitors /

kinases, and ultimately use this to predict selectivity?

• Can we design new inhibitors that are more

potent / selective?

Modeling beyond CDK9

• Benchmark experiment: choose crystal structures of various

kinases, each bound to a different inhibitor

• Modeling pipeline: 1) start from ATP-bound kinase of interest,

2) align inhibitor-bound structure of a different kinase, 3) transfer template inhibitor into kinase of interest, 4) align designed inhibitor

• nto template inhibitor, 5) minimize and re-rank
• Remember: we don’t know active conformation of inhibitor…
• Whole PDB will serve as templates - all active kinases look the

same, so do all Type I complexes!

Inhibitor of Csk21 aka casein kinase II alpha aka CK2α (over-expressed in CRC and modulates EMT)

Csk21 example

Csk21 example

Csk21 inhibitor aligned to diverse PDB templates

Csk21 example

Top 10 models from diverse templates after minimization

Csk21 example

Models closely match crystal structure of this complex

Csk21 templates

B-Raf inhibitor EphB4 inhibitor CK2 inhibitor CDK2 inhibitor

Summary of benchmark expt

• Using diverse templates in the PDB helps model many

(Type I) kinase inhibitors

• Minimization is helpful for achieving refinement better

than the templates

• Pipeline is

VERY fast, ~100 CPU minutes per complex (with more degrees of freedom than typical docking)

• This opens the door to more virtual screening, and

especially to predicting selectivity

Designing better inhibitors

• Build kinase-focused libraries of inhibitors
• Given a synthetic route to arrive a known

inhibitor, enumerate analogs that can be built using commercially available building blocks

• Computationally screen these for potency and

selectivity

Generalized synthesis

O Br S N N S- O N S S H2N O N S S H2N R2 NH2 O N S H2N R2 HN

+ +

R1 R1 R1 R1

Step II Step I

2872 compounds

232k new compounds, 126k are “drug-like”

77 compounds

Designing new inhibitors

• This reaction scheme gives 126,000 drug-like compounds that could be

made using the Sigma catalog (including MC180295!)

• The vast majority of these are novel chemical matter
• Modeling this library against CDK9 gives many that score better than

MC180295, and are also predicted to be more selective

• Other members of this library are predicted to be potent and selective

for other kinases…

Pharmacophore mapping

39

• Cannabinoid receptor

antagonist (example from wikipedia showing rimonabant)

• Common software is

CATALYST, MOE, others

Pharmacophore mapping

Pharmacophore mapping

Cannabinoid receptor antagonists demonstrating shared pharmacophore

From pockets to exemplars

Protein surface pocket “Exemplar” (aka pharmacophore)

Exemplar screening

(using ROCS, in this case)

From traditional to non-traditional drug targets

• Well-validated (traditional) drug targets have the advantage of

ample extant knowledge, enabling new studies

• Biology doesn’t work this way though!
• The ability to access non-traditional drug targets may open new

avenues in drug discovery

• In both regimes, it’s important to think carefully about how to best

apply the tools at hand

Our computational “toolbox”

Structure-based approaches Ligand-based approaches

Detailed MD simulations Docking (to predict pose) Virtual screening 2D fingerprints 3D matching (and pharmacophores) Fragment replacement

Emerging themes in the field

• There is still lots of room for creative approaches, figuring out

new ways to use these tools

• Ligand-based methods (and hybrid methods) offer intriguing

advantages over purely structure-based methods

• At present, virtual screening often only identifies micromolar hits:

carrying out med chem optimization in silico may enable design of more potent compounds at the outset