Discovering Substructures in Chemical Toxicity Domain

1
Discovering Substructures in Chemical Toxicity
Domain
Masters Project Defense by Ravindra Nath
Chittimoori Committee DR. Lawrence B. Holder,
DR. Diane J. Cook , DR. Lynn Peterson Department
of Computer Science and Engineering University of
Texas at Arlington
2
Outline

• Chemical Toxicity Database
• Motivation and Goal
• Knowledge Discovery in Databases (KDD)
• SUBDUE Knowledge Discovery System
• Experiments with Unsupervised SUBDUE
• Experiments with Supervised SUBDUE
• Discussion of Results
• Conclusions
• Future Work

3
Chemical Toxicity Database

• Carcinogenesis Prediction Problem
• Toxicology Evaluation Challenge
• Domain
• Compounds - Total
• Training set 162 136 298
• Experimental set ? 27 ? 25 69

4
Motivation and Goal

• Ever-increasing number of chemical compounds
• Needs analysis to obtain the Structure-Activity
• relationships of a compound
• Determine SUBDUEs applicability to chemical
• toxicity domain

5
Knowledge Discovery in Databases (KDD)

• Process of identifying valid, novel, potentially
• useful and understandable patterns in data
• Goal of Knowledge Discovery
• Verification
• Discovery
• Data mining methods
• Model Representation, Evaluation and Search

6
Steps in KDD

• Identify the goal of the process
• Collect, create and prepare the dataset
• Select the data mining method
• Select the data mining algorithm
• Transform the data
• Execute the algorithm
• Interpret/evaluate the discovered patterns
• Consolidate the knowledge discovered

7
SUBDUE Knowledge Discovery System

• SUBDUE discovers patterns substructures in
  structural data sets

Vertices objects or attributes Edges
relationships
shape
triangle
object
shape
on
square
object
4 instances of
8
SUBDUE - Input Representation

• Each atom is represented as a vertex with
• directed edges to the name, type and the
  partial
• charge of the atom
• Bonds are represented as undirected edges
• Each group is represented as a vertex having a
• string label specifying the group name with
• directed edges to all participating atom
• vertices

9
SUBDUE - Input Representation

• Representation used in Unsupervised SUBDUE
• A vertex having a string label specifying
  the
• alert with directed edges to all the
  atoms in
• the compound
• Representation used in Supervised SUBDUE
• A vertex for all the compounds with string
  label
• compound
• The compound vertex has directed edges to
  all
• the vertices representing the activity of
  an
• alert on a compound

10
Unsupervised SUBDUE Input Representation Example
C
10
0.063
10
0.062
C
t
t
n
p
n
p
Atom
Atom
1
gr
n - Name t - Type p - Partial charge po -
Positive gr - group
po
gr
po
Ames
Methyl
11
Supervised SUBDUE Input Representation Example
C
10
0.063
10
0.062
C
t
t
n
p
n
p
Atom
Atom
1
gr
contains
n - Name t - Type p - Partial charge gr -
group Com - Compound
gr
contains
Com
Methyl
Positive
Ames
12
SUBDUE - Model Evaluation

• Minimum Description Length Principle
• Best theory to describe any graph
• Minimize I(S) I(G/S)
• Graph Compression

13
Other important Concepts of SUBDUE

• Inexact Graph Match Approach
• Concept - Learning
• Predefined Substructures

14
Unsupervised SUBDUE - Methodology

• Training set further divided
• 3 approaches to determine carcinogenicity of
  compounds in experimental set

-- Apply SUBDUE individually to the compounds --
Inclusion of pre-defined substructures -- Check
for matching of substructure in the compound
to be classified
15
Unsupervised SUBDUE - Results
10
3
0.062
0.057
c
br
t
p
t
p
n
n
atom
atom
1

• Third approach used to classify compounds in
  
• experimental set
• Accuracy Level -gt 0.322
• Cyanate ether groups are also discovered to
• be indicators of carcinogenic activity

16
Supervised SUBDUE - Methodology

• Create set of indicators of carcinogenic
  activity
• Create set of indicators of noncarcinogenic
• activity
• Calculate value of substructures discovered in
• carcinogenic and noncarcinogenic set
• Select a set of substructures to be used in
• classifying compounds in experimental set

17
Supervised SUBDUE - Methodology

• Check for the existence of these substructures
  in
• the compound to be classified
• Calculate the Carcinogenic Activity Value of the
• compound
• Calculate the NonCarcinogenic Activity Value of
  the
• compound
• Determine the activity of the compound

18
Supervised SUBDUE - Results

• A set of 12 substructures discovered by SUBDUE
  used to classify compounds in the experimental
  set
• 6 substructures from carcinogenic set include
  substructures which form part of groups like
  amino, di10, methyl, ether, halide10 and
  substructure which indicates compound testing
  positive on AMES, Salmonella, etc.
• 6 substructures from noncarcinogenic set include
  substructures which form part of groups like
  methoxy, Ar_Halide, di64, nitro and alkyl_halide
  and substructure which indicates compound testing
  negative on AMES, Salmonella, etc.

19
Supervised SUBDUE - Substructure Example -
Carcinogenic Set
positive
Ames
Salmonella
positive
Compound
Salmonella_n
positive
20
Supervised SUBDUE - Substructure Example -
Carcinogenic Set
Cl
93
-0.123
10
C
n
-0.024
t
t
p
n
p
Atom
Atom
n - Name t - Type p - Partial charge gr - group
gr
gr
Halide10
21
Supervised SUBDUE - Substructure Example -
NonCarcinogenic Set
negative
Ames
Salmonella
negative
Compound
Cytogen_ca
negative
22
Supervised SUBDUE - Substructure Example -
NonCarcinogenic Set
Cl
93
-0.124
10
0.477
C
n
t
t
p
p
n
Atom
Atom
n - Name t - Type p - Partial charge gr -
group A-H - Alkyl Halide
gr
gr
A-H
23
Supervised SUBDUE - Results

• PTE-1 Results
• Compounds
  - Total
• PTE-1
  20 19 39
• Correct Prediction 12 6
  18
• Incorrect Prediction
  8 13 22
• Accuracy 0.6 ( ), 0.315 (-) , 0.462 (total)

24
Supervised SUBDUE - Results

• PTE-2 Results
• Compounds
  - Total
• PTE-2
  7 6 13
• Correct Prediction 4 3 7
• Incorrect Prediction
  3 3 6
• of compounds whose activity is
  known
• Accuracy 0.572 ( ), 0.5 (-) , 0.538 (total)
  

25
Results - Discussion

• Unsupervised SUBDUE successful in discovering
• lead indicators of carcinogenic activity
• Supervised SUBDUE also successful in
• discovering lead indicators of carcinogenic
• activity
• ILP System PROGOL PTE-1 (0.72), PTE-2 (0.62)
• Ashby, TOPKAT are other toxicity prediction
• methods

26
Conclusions

• Consistent with results obtained by logic based
• systems like PROGOL
• Prefer to use Concept Learner when positive and
• negative examples of target concept available
  
• SUBDUE is capable of discovering lead
• indicators of carcinogenic/noncarcinogenic
• activity in chemical toxicity domain .

27
Future Work

• PTE-3 Evaluation Challenge
• Trimmed Data Sets (Partial Charge)
• Newer Version of Concept Learning SUBDUE being
• developed

28
Reference
http//cygnus.uta.edu/subdue