Applying the SUBDUE Substructure Discovery System to the Chemical Toxicity Domain

1
Applying the SUBDUE Substructure Discovery System
to the Chemical Toxicity Domain

• Ravindra N. Chittimoori, Diane J. Cook,
• Lawrence B. Holder
• Department of Computer Science and Engineering
• University of Texas at Arlington
• http//cygnus.uta.edu/subdue/

2
Motivation and Goal

• Ever-increasing number of chemical compounds in
  use today (100,000).
• Needs to identify relationships between the
  molecular structure and the toxicity of a
  chemical compound.
• Apply knowledge discovery to the U.S. National
  Toxicology Program (NTP) to identify such
  relationships.

3
Knowledge Discovery in SUBDUE

• Structural discovery system
• Graph-based input representation
• Beam search through substructure (subgraph) space
• Graph compression heuristic based on minimum
  description length
• Inexact, polynomial graph match

4
SUBDUE Example
5
Chemical Toxicity Domain

• Database of 367 chemicals
• Levels of evidence assigned by NTP
• CE clear evidence of cancerous activity
• SE some evidence
• E equivocal evidence
• NE no evidence

6
Predictive Toxicology Evaluation

• Predictive Toxicology Evaluation (PTE) challenge
• PTE-2 ended November 1998
• http//dir.niehs.nih.gov/dirlecm/pte2.htm
• PTE-3 scheduled for July 1999 - July 2000

7
Chemical Toxicity Data

• Atoms (name, type, partial charge)
• Bonds (type)
• Chemical groups
• Alcohol, amine, amino, benzene, ester, ether,
  ketone, methanol, methyl, nitro, phenol and
  sulfide

8
Chemical Toxicity Data

• Carcinogenicity-related tests
• Ames
• Chromex
• Chromaberr
• Drosophilia
• Mouse-Lymph
• Salmonella Assay

9
Chemical Compound Representation
10
Input Representation

• Sample Atomic Structure
• SUDBUE graph input

C
H
1
v 1 atom v 2 C v 3 atom v 4 H d 1 2
name d 3 4 name u 1 3 1
11
Methodology

• Training set further divided into learning and
  testing sets
• Find best substructures in learning-set positives
  not prevalent in negatives
• Find occurrences of substructure in testing

12
Results
10
3
0.062
0.057
c
br
t
p
t
p
n
n
atom
atom
1

• Learning set 268
• Positive compounds 134/143
• Negative compounds 24/125
• Testing set 30
• Positive compounds 15/19
• Negative compounds 4/11

13
Results
1
10
32
0.34
0.211
h
?0.778
t
p
c
n
t
p
t
p
1
n
atom
1
atom
n
atom
n
1
0.36
h
1
t
p
n
atom

• Learning set 268
• Positive compounds 60/143
• Negative compounds 0/125
• Testing set 30
• Positive compounds 8/19
• Negative compounds 0/11

14
Discussion

• Consistent with results obtained by ILP system
  PROGOL (Srinivasan et al., ILP-97).
• Groups discovered by SUBDUE (e.g., Amino) are
  unique substructures found only in compounds
  which test positive on carcinogenicity.

15
Conclusion

• SUBDUE has the ability to discover interesting
  patterns (substructures) that might be helpful in
  predicting carcinogenicity.
• SUBDUE is suitable for knowledge discovery in the
  chemical toxicity domain.

16
Future Research

• Applying concept-learning SUBDUE to the chemical
  toxicity database
• Find substructures compressing positive graph,
  but not negative graph
• Incorporate more domain knowledge
• PTE-3 challenge (July 1999)