De Novo Design - Unistrainfochim.u-strasbg.fr/CS3/program/material/Schneider.pdf · De Novo Design...

De Novo Design

Adaptive Systems for Molecular Design

Gisbert Schneider

Beilstein Endowed Chair for Chem- & Bioinformatics

Goethe-University Frankfurt

23 June 2008, (c) G. Schneider

“[..] Thus between A & B immense

gap of relation. C & B the finest

Transmutation of Species

gap of relation. C & B the finest

gradation, B & D rather greater

distinction. Thus genera would be

formed. — bearing relation”

Charles R. Darwin (*1809)

Notebook B: Transmutation of Species (1837-1838), p.36

Source: www.darwin-online.org.uk


Voyages to the (un)known

Darwin and Henslow, letters 1831-1860.

The growth of an idea (N. Barlow, ed. 1967).

Voyages of the H.M.S. Beagle


Requirements

• Structure sampling

• Structure assessment

• Grow

• Link

• Lattice

• Stochastic

• Primary constraints

(receptor, ligand)

Implementation

De Novo Design Concepts

• Structure assessment

• Search / Optimization Method

(receptor, ligand)

• Secondary constraints

• Depth-first search

• Breadth-first search

• Random search

• Evolutionary Algorithm

• Monte Carlo / Metropolis

• Exhaustive enumeration

• (Free Energy Perturbation)23 June 2008, (c) G. SchneiderSchneider & Fechner (2005) Nat. Rev. Drug Discov. 4, 649.

Global optimum

Local optimum

Fit

ne

ss

Adaptive Walk

in a “Fitness Landscape”

Local optimum

X1X2


Schneider & Fechner (2005) Nat. Rev. Drug Discov. 4, 649.

Fechner & Schneider (2007) JCIM 47, 656.

Fit

ne

ssTypes of Fitness Landscapes

Fit

ne

ss

The Dream The Nightmare


à Chemical Similarity Principle!

O

O

O OH

Aspirin®

What is a Molecule?

Aspirin®

Increasing “Fuzziness”

• Pharmacophoric representation

on different levels of abstraction


Tanrikulu et al. (2007) ChemBioChem 8, 1932.

Renner & Schneider (2004) J. Med. Chem. 47, 4653.

Start

q C

BA

Local Neighborhood Search

Neighborhood N Local

optimumx

• fixed N• variable (“adaptive”) N


Local (neighborhood) search

Any-ascent/Stochastic

hill-climbing

Steepest/First-ascent

hill-climbing

Population-based

Local Neighborhood Strategies

Threshold-acceptingSimulated

annealing

Tabu search

Genetic

algorithms

Evolution

strategies

Evolutionary

programming


• Structures that undergo adaptation

• Initial structures (starting solutions)

• Fitness measure that evaluates the structures

• Operations to modify the structures

Elements of Artificial Adaptive Systems (Koza, 1992)

• Operations to modify the structures

• State (memory) of the system at each stage

• Method for designating a result

• Method for terminating the process

• Parameters that control the process


Start solution

Final

solution

Adaptive

neighborhood

Why “Adaptive”?

neighborhood

Adaptive optimization copes with …

• non-additive fitness functions

• multiple fitness functions

• very large search spaces

• nonlinear system behavior


• Genetic algorithms

• Evolution strategies

• Particle Swarm Optimization (PSO)

Learning from Nature

• Particle Swarm Optimization (PSO)

• Ant Colony Optimization (ACO)


Ant Colony Optimization

A Simple Combinatorial Case: Peptide Design

T-cell receptor

• Design of novel antigens presented

by MHC I (H-2Kb)

• Length: 8 residues (208 possibilities)

à Neural network ensemble + jury

(PDB: 1HSA, Madden et al. 1993)

à Neural network ensemble + jury

à Artificial ant system

Assay confirms the predictions:

89% correct for binding peptides

95% correct for non-binding peptides

Hiss et al. (2007) PEDS 20, 99.

Schneider et al. (1998) PNAS 95, 12179.

Ant System for Combinatorial Design


Which Search Strategy Should be Applied?

Global & local search control guided by

• Random parameter variation

• Stepwise systematic parameter variation

• Gradients in the fitness landscape ...

There is no one best method.

Evolutionary algorithms are

• easily implemented

• robust

• able to cope with high-dimensional optimization tasks


Principle of Evolutionary Strategies

Step 1: Random searchStep 2 and following:

“Local hill climbing”

Best value

Optimum

Generate a diverse set of parameter

values and hope for a hit

Generate localized distributions of

parameter values

and improve steadily

Best value

à Adaptive neighborhood23 June 2008, (c) G. Schneider

Algorithm of the (1,λ),λ),λ),λ) Evolution Strategy

Initialize parent (xP,sP,FP)

For each generation:

Generate λ variants (xV,sV,FV) around the parent:

sV = abs(sP + G)

Fitness

Stepsize

Variables

(Rechenberg, 1973)

sV = abs(sP + G)

xV = xP + sV• G

Calculate fitness FV

Select best variant according to FV

Set (xP,sP,FP) = (xV,sV,FV)best

)2sin()ln(2),G( jiji π⋅−=

i,j: pseudo-random numbers in ]0,1]

in [-∞,∞]


De novo Fragment Assembly

DB

DB

Known drugs and leadsKnown drugs and leads

Set of reactionsSet of reactions

Generate µ initial

parent molecules

Generate λ new moleculesfrom µ parent(s)

Start

R10

R10

R3OHO

R1DB

Stock of building blocks

with reaction labels

Determine fitness

Select µ best molecules

End

No

Yes

Terminate?

R10

NR7

OR4

R1

• Template compound(s)

• Prediction tools• Biochemical assay

• Template compound(s)

• Prediction tools

• Biochemical assay

Gen

era

tio

n l

oo

p

23 June 2008, (c) G. SchneiderSchneider et al. (2000) Angew. Chem. Int. Ed. 39:4130

Operators of Evolutionary Optimization

• Mutation

• Crossover

VariationAdds new information to a population

Exploits information within a population• Crossover

• Selection

Exploits information within a population


HN

F

F

O

HN

F

HN

O

HN

Rx1

O

F

F HN

F

Rx1NH O

Rx1

Rx1

N

O

Rx1O

N

N Rx1

O

N

N

HN

HN

O

F

F

Breeding New Molecules

Virtual reaction

Mutation

HN

F

F

O

HN

F

HN

O

N

NHN

C

O

NHN

C

O

HN

Rx1

O

F

F

HN

F

Rx1NH O

Rx1

Rx1

C

O

Rx1NN

N

NHRx1

NH

Rx1C

O

Rx1

C

O

N HN OHN

F

HN

O

F

F

HNC

O N

NHN

Crossover


IC50 [µM]

40

50

60

70

80

90

100

Random Search

Evolution Strategy

How Many Iterations? How Many Compounds?

Limited resources:

300 compounds

à There is an optimum!

Trypsin

inhibition

Population size × Generations

1 x

300

2 x

150

3 x

100

5 x

60

10 x

30

15 x

20

20 x

15

30 x

10

37 x

8

50 x

6

60 x

5

75 x

4

100 x

3

150 x

2

0

10

20

30

preferred combinations

à There is an optimum!

23 June 2008, (c) G. SchneiderSchüller & Schneider (2008) JCIM, in press.

ONH

NH

HO

O O

NNH2

NH

ONH

NH

OH

OH

HONH2O OH

ONH

NH

HO

NH

NH2

O

OHO NH2

1IC50 = 22.5 µM

2IC50 = 13.0 µM

6IC50 = inactive

ONH

NH

O O

NNH2

NH

HN

ES PSO

“Molecular Evolution”

Trajectotries through

inactive regions of

chemical space are

possible O O ONH

NHN

O

ONH

NH NH2

NH

O

ONH

NH

O

NH2

NH

N

ONH

NH NH2

NH

N

3IC50 = 47.5 µM

4IC50 = 74.7 µM

5IC50 = 37.2 µM

7IC50 = inactive

8IC50 = 15.1 µM

9IC50 = 54.7 µM

ONH

NH

O

NH

NH2

N

possible


Particle Swarm Optimization (PSO)

• Biological motivated optimization paradigm

• Individuals move through a fitness landscape

• Particles can change their movement direction and velocities

• “Social” interactions between individuals of a population• “Social” interactions between individuals of a population

• Each particle knows about its own best position

• All particles know about the overall best position


Particle Swarm Optimization (PSO)

T. Krink, University of Aarhus

PSO: Update Rule 1 (Standard Type)

Individual

memory

Social

memory

))((2))((2)(v)1(v 21ii txprtxprtt ibii −⋅⋅+−⋅⋅+=+

v: Velocity vector

x: Coordinates of a particle

i: Dimension

t: Time (epochs)

r1 , r2: Random numbers in ]0,1]

pi , pb: Best positions (individual and social)


Global optimum

PSO convergence after 100 epochs (Griewangk‘s function)

PSO: Individual and Social Memory

n1 = 1

n2 = 2

n1 = 2

n2 = 1

Higher weight on

the social memory


Visualization of Particle Trajectories

• The particle visited several local optima bevor finding

the global optimum

à PSO Demo23 June 2008, (c) G. Schneider

COLIBREE®

Combinatorial Library BreedingCombinatorial Library Breeding


Disassembly Rules

Ar Ar Ar

Retain building blocks with

MW < 200 Da

à 7,184 Building Blocks

avg MW = 141 ± 35 Da

23 June 2008, (c) G. SchneiderHartenfeller et al. (2008) Chem. Biol. Drug Des., in press

Linker Library


Molecule Dissection


PSO Progress Over Time

Library size:

40 molecules40 molecules

Fitness:

Distance to Reference

(minimization!)


Positive and Negative Design

PPARα(negative weight)

PPARγ(positive weight)


Nonlinear

mapping

1

2

3

4

5

SOM projection

Self-Organizing Map (SOM)Kohonen , 1982

5

6

1 2 3 4 5 6 7 8

Original

data space

à SOMMER Demo

• Nonlinear projection of high-dimensional space

• Topology-preserving projection

• Data clustering

• Prediction with probability estimation


Pharmacophore Road Map of Chemical Space

Proteases MMPsKinases

Ion channelsGPCRs Unidentified (orphan)

• Topological pharmacophores (CATS descriptor)

• COBRA collection (10,500 drugs and lead compounds)


Thrombin inhibitors PPARα modulators

Land Ho! Exploration of Chemical Space

SO H

N

O

O

NH

O

N

NH2H2N

NAPAP

HO

O

O

HN

O

S

N

CF3

GW590735


Let’s Be Positive!

De Novo Design of Thrombin Inhibitors

CATS descriptor no. 10:

PP0 =

Top-ranking designed compounds O O

NH

OHN

S

O

O

HN NH2

O

NH

O N

NHH2N

HN

S

O

OSCl

O O

NH

OHN

S

O

O

S

Cl

HN NH2

O O

NH

OHN

S

O

O

HN NH2

O O

NH

OHN

SO

OF S

NHH2N


Climbing “Acid hill”:

De Novo Design of PPARαααα Agonists

CATS descriptor no. 19:

DN1 =

HO

O

Top-ranking designed compounds

O

HOO

HN

O

O

Cl

O

HO O

HN O

CF3

O

HO

OHN

O

NS

O

HO

O

HN

O

NS

Cl

CF3

O

CF3

O

HOO

HN

O

O

O

HO O

HN

O

N

O

HOO

HN

O

O

I

HO

O

HO

O

HN

O

NS

CF3

O

HOO

HN

O

NS

S

O

HO

OHN

O

NS

Cl

Cl


Conclusions

• Fragment-based design often results in bioactive compounds

• “Novelty” of the designs depends on local convergence

• Adaptive optimization enables “scaffold-hopping”• Adaptive optimization enables “scaffold-hopping”

• Visual inspection / expert knowledge is crucial


Acknowledgement

Dr. Karl-Heinz Baringhaus

Prof. Manfred Schubert-Zsilavecz

Prof. Holger Stark

Prof. Dieter Steinhilber

Dr. Tanja WeilDr. Tanja Weil

modlab® Team

www.modlab.de

De Novo Design - Unistrainfochim.u-strasbg.fr/CS3/program/material/Schneider.pdf · De Novo Design...

Documents

Transcript of De Novo Design - Unistrainfochim.u-strasbg.fr/CS3/program/material/Schneider.pdf · De Novo Design...