Kein Folientitel

1
Artificial Neural Nets (ANN) An alternative way
of deriving QSAR models
2
J. Zupan, J. Gasteiger
Neural Networksin Chemistryand Drug Design
Second Edition
WILEY-VCH, Weinheim, 1999
3
Artificial Neural Nets
number of chemistry-related publications
927
855
743
441
498
290
105
30
3
5
4
Artificial Neural Nets
The "100 Steps Paradoxon"
human brain computer reaction time of firing
of neuron clock ratebasic unit 10-3 sec 10-9
sec (500 MHz) recognition of the faceof a
friend 10-1 sec 10-1 sec no. of processing
steps 100 100,000,000

• The human brain works highly parallel

5
Artificial Neural Nets
Visual Cortexof the Human Brain
6
Artificial Neural Nets
Biological and Artificial Neurons
7
Artificial Neural Nets
Input e.g. molecular descriptors Output e.g.
biological activity
8
Artificial Neural Nets - Supervised learning
predict new compounds
change weights
9
Artificial Neural Nets
Net architecture ("topology") in Capsaicin
example
10
Artificial Neural Nets
Capsaicin dataset Predicted vs Actual Log EC50
using a trained Neural Net(Corr. Coeff. R2
0.92)
11
Artificial Neural Nets

• Prediction of the test set (compound 6i)

• (Multiple) linear regression predictions
• Log EC50 0.76 - (0.82)? Pred. Log EC50 for 6i
  1.56
• Log EC50 1.14 - (0.07)MR Pred. Log EC50 for
  6i 0.42
• Log EC50 0.76 - (0.82)? (0.0003)MR Pred. Log
  EC50 for 6i 1.57

• Neural Net prediction Pred. Log EC50 for 6i
  1.05
• And the experimental affinity is......

• Log EC50 gt3!!!(totally inactive)

12
Artificial Neural Nets - Unsupervised learning
gtprojection with preservation of the topology of
the input space
13
Artificial Neural Nets - Unsupervised learning
Molecules exhibit their activity via complex
surface properties
14
Artificial Neural Nets - Unsupervised learning
112 Dopamine and 60 Benzodiazepine Agonists in
the bulk of 8,323 structures of unknown activity
Kohonen map (40x30)
15
Artificial Neural Nets - Unsupervised learning
benzodiazepine 50
benzodiazepine 39
benzodiazepine 43
benzodiazepine 22
benzodiazepine 49
16
Artificial Neural Nets - Unsupervised learning
Kohonen map unsupervised learning
multilayer network supervised learning
17
3D QSAR Linking QSAR with Molecular Modeling
18
3D QSAR

• Two standard methods
• CoMFA (Comparative Molecular Field Analysis,
  Cramer et al.)
• QuaSAR (Vedani et al.)

19
3D QSAR - The CoMFA Approach
20
3D QSAR - The CoMFA Approach
For the capsaicin example, CoMFA predicted Log
EC50-0.21!
21
3D QSAR - The CoMFA Approach