Gerhard Ecker Lecture 1

Use your Brain – Artificial Neural Networks in Drug Design Gerhard F. Ecker Emerging Field Pharmacoinformatics Department of Medicinal Chemistry, University of Vienna Althanstrasse 14, A-1090 Wien, Austria e-mail: [email protected]; http://homepage.univie.ac.at/gerhard.f.ecker

G. Ecker

Use your Brain!

THM 1

G. Ecker

Drug Development - Failures

Kubinyi, Nature Rev Drug Discov 2, 665 (2003)

G. Ecker

Toxicity hERG Potassium Channel • Long QT syndrome • Highly promiscuous

Sanguinetti, Nature 443, 463 (2006) G. Ecker

Metabolism Cytochrome P450 Complex Problems: • Rapid and extensive metabolism • Toxification • Drug/Drug Interactions • Poor Metabolizers

de Graaf, J Med Chem 48, 2725 (2005) G. Ecker

Bioavailability and Transporter

Kim, Mol Pharm, 3, 26 (2006)

G. Ecker

ABC-Transporter • • • •

membrane-bound efflux pumps energy driven (ATP) ATP Binding Cassette (ABC-transporter) 48 ABC transporter in humans

– P-glycoprotein (P-gp) – Multidrug Resistance Related Protein (MRP) – Breast Cancer Resistance Protein (BCRP, MXR) • a lot of analogous transport proteins in bacteria, fungi, protozoes and plants very often multispecific in ligand recognition G. Ecker

P-Glycoprotein (ABCB1, MDR1) • • • • • • • •

170 kD 2 transmembrane domains 2 ATP-binding sites Xenotoxin transporter hydrophobic vacuum cleaner intestine, liver, kidney blood-brain barrier tumors G. Ecker

Linear „classical“ QSAR • Advantages – clear relationship between descriptors and biological activity (linear or e.g. bilinear) – influence and importance of given descriptors is visible (confidence intervals, t-test) – scaling gives relative importance of the descriptors G. Ecker

Linear „classical“ QSAR • Disadvantages: – type of relationship must be given (linear, bilinear, sigmoidal, ...) – difficult to treat complex interactions (receptor, membrane) – sensitive to noise in the data

G. Ecker

Artificial Neural Networks • Simulation of the human brain for recognition of complex relationships • examples: face recognition, stock market, ... • Currently more than 9000 references with the concept „Artificial Neural Networks“ in the Chemical Abstracts

G. Ecker

Artificial Neural Networks • Supervised Learning – Learning via presentation of cases – Train a kid to separate dogs from cats

• Unsupervised Learning – Classification of groups via pattern recognition – Separate cats and dogs without knowing the differences G. Ecker

Artificial Neural Networks Structure of the human brain

G. Ecker

Output

Transfer Function

Combining Function

Inputs

General Model for ANNs Weights

G. Ecker

Transfer Function

G. Ecker

Examples Feedforward Multilayer Network w1π+w2π+w3σ+...+w8I

σ MR

w1π w2π

Output

Inputs

π

TF

I G. Ecker

Training - Learning Training occurs via iterative adjustment of the weights till output signal corresponds to actual value (supervised learning): Delta Rule: ΔkWij = β(tj-aj)ai + αΔk-1Wij tj: aj: ai: β: α:

target output aktueller output aktueller input learning rate momentum factor G. Ecker

Stopping Conditions • • • • •

number of cycles maximum target error minimum improvment of target error maximum verification error minimum improvment of verification error

G. Ecker

Target Error - Verification Error

G. Ecker

Disadvantages of Artificial Neural Networks • Only little information on relationships between input data and output (positive or negative correlation?) • difficult to receive informations on relative importance of descriptors • most often global minimum not found; this leads to different results in different runs G. Ecker

Examples for use of Feedforward Networks • Regression analysis: caution with extrapolations, especially if you use min/max scaling of the output variable; (better: scale between 0.2 und 0.8). • Classification problems: in case of two classes you may use one or two out put neurons; examples: drug like/non drug like, active/non active, substrate or not,...; in case of more than two classes it is recommended to use one output neuron per class rather than 0-1-2 as output; G. Ecker

Examples • Classification: – exclusive and/or Problem – drug like/non drug like (H.Kubinyi)

• Regression – MDR-modulating Activity (G.Ecker)

G. Ecker

drug like/non drug like • ISIS Fingerprints as input • Training with 5000 compounds from ACD and 5000 from WDI • Test with the remaining compounds; 80% right predictions

G. Ecker

MDR-modulating Activity X1

N

F

X2

X8

(ortho) CH(OH)C2H4Ph

X9

(ortho) CH(OCH3)C2H4Ph

X10

(ortho) COCH3

X11

(ortho) COC2H4Naphth

X12

(ortho) COC2H5

X13

(ortho) COPh

X14

(ortho → para) COC2H4Ph

X15

5-OH

X16

5-OCH2Ph

X17

(ortho → para) COCH3

X18

(ortho → meta) COC2H4Ph

X19

(ortho → meta) COCH3

N N

H3C

X3

CH3 N

CH3 CH3

X4

OH O

H N

CH3

O

N O

N

X5

X6

N

N

OH

X7

H N

OH R2

O

R1

R1: X8-X19 R2: X1-X7 G. Ecker

MDR-modulating Activity log(1/EC50) predicted

Topology: 7:4:1

r = 0.92, s = 0.32, Q2cv = 0.85

log(1/EC50) beobachtet

G. Ecker

Virtual Library Screening OH

Predicted activity (μM)

OH N

O

ANN CoMFA CoMSIA HQSAR

O

O

0.036 0.011

0.004

0.003

0.040 0.054

0.042

0.018

OH OH O O

O

CH3

G. Ecker

Classification Problems • Feedforward Networks – classification of compounds according to their odour impression

• Self Organizing Maps – identification of new lead compounds

G. Ecker

Aroma Quality of 1,4-Pyrazines • 98 Pyrazine derivatives with green, bell-pepper and nutty aroma • feedforward network with 5:3:3 topology: sum of electrotopological indices, number of carbon atoms of R2, charge on the first atom of R4, molecular surface of R1 and R3 • nominal output variable with accept and R3 R2 N reject threshold levels for each aroma • 86% right classifications for the test set R4 N R1 B. Wailzer, J. Med. Chem. 44, 2805 (2001)

G. Ecker

Support Vector Machines

linear

SVM G. Ecker

Support Vector Machines

Higher Dimensions G. Ecker

Self Organising Maps

G. Ecker

SONNIA Self-Organizing Neural Network for Information Analysis

C3

© Gasteiger et al.

G. Ecker

The Data Set O OH

R1

O R1 O

R2

R1 O

R2

Propafenon Type

Benzopyranone Type

OH

Benzofurane Type

OH

O

O R1

R1

R1

O R2 O R2

O

R2

Indanone Type G. Ecker

Connectivity - Autocorrelation • An autocorrelation vector gives informations about the distribution of atom properties within a molecule:

• AC(d) = Σ p(i) • p(j) i,j ∈ M(d) p(i), p(j): Properties of atoms i,j d: distance of atoms i and j • 2D-Autocorrelation: d = number of bonds • 3D-Autocorrelation: d = distance in angström

G. Ecker

Autocorrelation • Example: 4-Hydroxy-2-butanon – Property: atom weight – Distance: 2 bons

O 1

H3C

C

6 2 3

C

C

4 5

OH

AC(2) = p(1)p(3) + p(2)p(4) + p(3)p(5) + p(1)p(6) + p(3)p(6) = 12•12 + 12•12 + 12•16 + 12•16 + 12•16 = 864

G. Ecker

Kohonen Maps • Data set: 131 Propafenon-analogs • autocorrelations vectors (PETRA) lead to separation of highly active from inactive compounds • increase network size to 250 x 250 • projection of the propafenons together with the SPECS database (150.000 compounds) • analyse co-localisations

G. Ecker

Results

G. Ecker

Kohonen Maps - Hits O

O

N

N

O

N

N

N

N

O

N

O

N N N

N N

S

AG-690/11972772

N

O

AG-690/12887361

S

Cl

AJ-131/15197008

NH2

AJ-292/13162028

O

N

O

N

O

S

N N

S

N

N O

N

S

S N

AJ-292/15089034

AN-989/14669159

AO-364/14480185

G. Ecker

Kohonen Maps - Results • out of 7 compounds colocalizing with highly active propafenones are – 2 compounds with EC50 < 1 μM – 4 compounds with EC50 < 10 μM – 1 compound inactive

• out of 8 compounds colocalizing with inactive propafenones are – 1 compound with EC50 = 17 μM – 5 compounds with 100 μM < EC50 < 500 μM – 2 compounds with EC50 > 500 μM D. Kaiser, J. Med. Chem., 2007.

G. Ecker

Hit Follow-up • Select compounds similar to the two hits • Order and test them

G. Ecker

Software Packages • SONNIA: Molecular Networks • SNNS: University of Stuttgart • TRAJAN: implemented in Statistica

G. Ecker

References • J. Zupan & J. Gasteiger: Neural Networks in Drug Design, Wiley-VCH, 1999 • List of Web-Resources: – – – –

www.dsi.unifi.it/neural/w3-sites.html www.ncst.ernet.in/education/apgdst/aifac/resources/Neural psychology.about.com/od/neuralnetworks/ www2.chemie.uni-erlangen.de/publications/ANN-book/publications/ G. Ecker

Neural Networks - Conclusions • ANNs are good – for noisy data (HTS) – for complex relationships (ADMET) – for classification problems (yes/no)

• ANNs are bad – if you have a distinct drug/receptor interaction – if you want to know what‘s going on G. Ecker

Thank you! Gerhard Ecker

Michael Gottesman

Silke Schindler Barbara Zdrazil Karin Pleban Michael Demel Dominik Kaiser Claudia Hoffer

Gergely Szarkasz

Peter Chiba Stephan Kopp Manuela Hitzler

Edina Csaszar G. Ecker