Pharmacophore Identification by Data Mining

1
Pharmacophore Identification by Data Mining

• Takashi Okada, Masumi Yamakawa, Satoshi Fujishima
  and Norito Ohmori
• Department of Informatics, Kwansei Gakuin
  University,
• 2-1 Gakuen, Sanda, Hyogo, 669-1337 Japan Email
  okada-office_at_ksc.kwansei.ac.jp

2
Contents

• Problem Definition Using an Example of Dopamine
  D1 Agonist
• Linear Fragments
• The Scheme of Analysis
• The Cascade Model
• Rules and Selection
• Refinement from a Rule Condition to the
  Supporting Structure Chart
• Basic Active Structures
• Concluding Remarks

3
Dopamine agonists antagonists

• Dopamine is a kind of neurotransmitter compound.
• There are 6 kinds of dopamine receptors D1 - D5
  and Dauto.
• If we have too much dopamine, then schizophrenic
  symptoms develop.
• If we have too little dopamine, then Parkinson's
  symptoms develop.
• Agonists and antagonists to the receptors are
  used as drugs.
• The problem is to extract characteristic
  substructures of agonists and antagonists for
  each receptor.

Ligand
4
Source Data and Basic Active Structures
Dopamine Agonists Data
List of Basic Active Structures

• MDDR database 2003.1 by MDL Prous
• 3 Dopamine agonists data D1, D2, Dauto
• Omission of stereo- optical isomers

DG-H
and 4 miscellaneous compounds
5
Contents

• Problem Definition Using an Example of Dopamine
  D1 Agonist
• Linear Fragments
• The Scheme of Analysis
• The Cascade Model
• Rules and Selection
• Refinement from a Rule Condition to the
  Supporting Structure Chart
• Pull-up of Basic Active Structures
• Concluding Remarks

6
Linear Fragments as Descriptors
Lowercase for aromatic atoms
CO treated as a united atom
2 neighbors Hydrogen gt 0
O2H-c3--CO2-N3H
2 terminal atoms
Atom symbols are omitted here
(colon) for aromatic bond
Friendly notation for chemists Avoids too
detailed expression
7
Selection of Descriptors

• Initial number of fragments generated 4626
• Selection of fragments by the appearnace
  probability 3 lt P(fragment) lt 97
  . 660 fragments selected.
• Omission of a fragment in a correlated
  pair Rij gt 0.9 306 fragments are
  selected.
• Meaningful fragments are added manually 29
  fragments are added
• Finally 335 fragments are used as descriptors.

8
Contents

• Problem Definition Using an Example of Dopamine
  D1 Agonist
• Linear Fragments
• The Scheme of Analysis
• The Cascade Model
• Rules and Selection
• Refinement from a Rule Condition to the
  Supporting Structure Chart
• Basic Active Structures
• Concluding Remarks

9
Scheme of Analysis (Dopamine Agonists)
Dopamine D1, D2, Dauto Agonists 407 compounds

• Final Result
• List of Basic Active Structures
• Supporting Structures Chart

Chemist Driven Refinement
Supporting Strcutures Chart
10
Contents

• Problem Definition Using an Example of Dopamine
  D1 Agonist
• Linear Fragments
• The Scheme of Analysis
• The Cascade Model
• Rules and Selection
• Refinement from a Rule Condition to the
  Supporting Structure Chart
• Basic Active Structures
• Concluding Remarks

11
Sum of Squares Decomposition for Categorical
DataUsing SS Definition by Gini
800/200 0.16 TSS 160
BSS(2)
BSS(1)
BSS(1) 18
BSS(2) 72
WSS(2)
760/40 0.0475 WSS(1) 38
40/160 0.16 WSS(2) 32
18 38 72 32 160
positives / negatives Sample variance Sum of
squares
TSS WSS(1) BSS(1) WSS(2) BSS(2)
12
Cascade Model Local Correlation and Rule
E the activity
A y
Main condition
IF B y added on A y Cases 100 ?
60 THEN E y 60 ? 90, BSS 5.40 D y 60 ?
93, BSS 6.67
Precondition
Collateral correlations
A y, B y
13
Contents

• Problem Definition Using an Example of Dopamine
  D1 Agonist
• Linear Fragments
• The Scheme of Analysis
• The Cascade Model
• Rules and Selection
• Refinement from a Rule Condition to the
  Supporting Structure Chart
• Basic Active Structures
• Concluding Remarks

14
Computation of rules for D1 agonists

• Lattice generation
• 5928 nodes using thres 0.125
• 1242 candidate links (BSS gt 4070.007)
• Optimization of links resulted in 323 rules with
  BSS gt 3690.015
• Organization of rules resulted in 12 principal
  rules and 41 relative rules.
• Select rules with D1 ratio gt 0.8 and compounds
  gt 10,
• Finally, 2 principal and 19 relative rules
  resulted.

15
Self Organizing Map of 21 D1 Agonist Rules

• SOM procedure placed 21 rules in the right figure
  using supporting compounds (y/n) as variables.
• The coverage-based algorithm selected 3 rules
  (shown by red color), which covers 69 compounds
  out of 74 actives.

16
The Strongest Rule for Dopamine D1 agonists

• Rule 1 Cases 407 -gt 60 BSS 35.41
• IF O2H-c3c3-O2H y
• added on
• THEN D1Ag 0.82 0.18 ? 0.05 0.95 (off on)

Main condition catechol
THEN DAuAg 0.52 0.48 ? 1.00 0.00 (off
on) THEN C4H-C4H-c3-O2H 0.79 0.21 ? 0.07
0.93 (n y) THEN C4H-C4H-c3-O2H 0.81 0.19 ?
0.02 0.98 (n y) THEN N3H-C4H--c3-O2H 0.86
0.14 ? 0.33 0.67 (n y) THEN N3H-C4H--c3-O2H 0
.90 0.10 ? 0.32 0.68 (n y) THEN
C4H-N3H--C4H-c3 0.85 0.15 ? 0.38 0.62 (n y)

• Catechol is the key substructure.

• Real pharmacophore needs an amino group,which
  appears in collateral correlations.

• Interpretation is a hard process.

17
Contents

• Problem Definition Using an Example of Dopamine
  D1 Agonist
• Linear Fragments
• The Scheme of Analysis
• The Cascade Model
• Rules and Selection
• Refinement from a Rule Condition to the
  Supporting Structure Chart
• Basic Active Structures
• Concluding Remarks

18
Refinement Process Starting from OccO
0
1
2
3
4

• Insertion of bond-atom pair to every position.
• Greedy search for the highest BSS structure.
• If a seed captures the essential
  substructure,the refinement reaches a reasonable
  result.

8
19
Structure Refinement System
All data Active 74 Inactive 334
a seed SMARTS string
a seed fragment
Select by seed Active 57 Inactive 3
Target D1 agonist
Reference D1, D2, Dauto agonists
At step 7 Active 57 Inactive 2
20
Part of Supporting Structures chart at Step 7
21
Contents

• Problem Definition Using an Example of Dopamine
  D1 Agonist
• Linear Fragments
• The Scheme of Analysis
• The Cascade Model
• Rules and Selection
• Refinement from a Rule Condition to the
  Supporting Structure Chart
• Basic Active Structures
• Concluding Remarks

22
Necessity of Listing Basic Active Structures

• Given a common substructure
  supporting structure chart
• There are still dissimilar structure types.
• A computer scientist thinks OK, if supplemented
  by a comment to frequently appearing skeletons.
• Synthetic organic chemists thinklisting basic
  active structures is essential
• because the function as a drug necessitates the
  presence of surrounding substructures.
• They said they can extract these structures
  easily viewing the supporting structures chart..

23
Extraction of Basic Active StructuresDriven by a
Chemist

• A chemist recognizes a candidate basic structure,
  and gives it to the refinement system.
• The process repeats until it gives a satisfactory
  yes/no hits.
• The extraction of an active structure repeats
  until the list of basic active structures cover
  most compounds.
• The final knowledge base consists of
• A list of basic active structures with SMARTS
• Reference to the supporting compounds chart for
  each active structure
• Miscellaneous compounds not covered by the active
  structures

24
Basic Active Structures from D1 Agonists
actives / inactives
25
Miscellaneous Compound Structures Not Covered by
the 8 Basic Active Structures
26
Basic Active Structure

• The basic active structure is a subjective
  concept.
• The structure possesses the discriminating power
  for the target activity tested by the refinement
  system.
• In this sense, the structure is different from
  the frequent appearing skeleton in active
  compounds.
• This listing stimulates chemists' recognition of
  the pharmacophore, and it is very useful for the
  design of new drugs.

Non DG
DG-A
DG-C
DG-unknown1
Non DG
DG-B
DG-unknown2
DG-D
Miscellaneous Active compounds
Nobody knows the exact boundary.
Non DG
27
Contents

• Problem Definition Using an Example of Dopamine
  D1 Agonist
• Linear Fragments
• The Scheme of Analysis
• The Cascade Model
• Rules and Selection
• Refinement from a Rule Condition to the
  Supporting Structure Chart
• Basic Active Structures
• Concluding thought

28
Discrimination vs. Understanding

• Discrimination can be done by just HO-cc-OH.
• An amino group is definitely necessary as a D1
  agonist. Success of a rule as a discrimination
  model depends on the biased character of the
  dataset.
• Incorporation of Surrounding Information
• Inclusion of collateral correlations into a rule.
• Aggregation of the surrounding components by the
  refinement system.
• Essential components in the surroundings are
  added and catalogued as the basic active
  structure by medicinal chemists using the
  refinement system.

29
Conclusion and Future Plan

• We could pull up basic active structures of
  dopamine D1 agonists, which are effective in the
  design of new drugs.
• Project of Pharmacophore Knowledge Base
• Current stage (some entry is ongoing)
• Dopamine D1, D2, Dauto Agonist
• Dopamine D1-D4 Antagonist
• 5HT 1A, 1B, 1C, 1D, 1F, 3, 4 Agonist
• 5HT 1A, 1D, 2A, 2B, 2C, 3, 4 Antagonist
• Dopamine, 5HT Reuptake Inhibitor
• Adrenergic a1, a2, ß1 Blocker
• Adrenergic a2 Agonist
• Adenosine A1, A2 Agonist
• Adenosine A1 - A3 Anatagonist
• Website will open by the end of this year.
• Analysis results of 100 activities by March 2009

30
Acknowledgements

• Thanks go to
• Chemists Dr. Mori and Dr. Horikawa and Ms.
  Kamiguchi for their efforts in the analysis and
  useful comments to the system.
• Students Mr. Nakano and Mr. Kitajima for the
  development of supporting software.
• Granted by Ministry of Education, Culture,
  Sports, Science and Technology, Japan.
• Thank you for your attention.

31
Structure Refinement Using Unconnected Structure

• Seed fragment C-NH0CCCcc-OH1.N!H0
  active/inactive 16 / 1
• Refined fragment C(-C(-C(-C(-C(-C)))))-NH0(-C
  (-C(-C)))C(-C(-c(c(c(c)))))CCcc-OH1.N!H0
  (-C(-C)) active/inactive 16 / 0contains 2
  structures

32
Cascade model 1

• Itemset Lattice
• itemset as node
• items inclusion as link
• class distribution as node property

33
Cascade model 2

• Nodes as Lakes with Potential
• Links as Waterfalls with Power
• High Power Waterfalls as Rules

Mixed

• Questions
• Potential definition
• Power definition
• Cascade construction
• Selection of waterfalls

Pure
Potential Class purity