Data Mining in DNA: Using the SUBDUE Knowledge Discovery System to Find Potential Gene Regulatory Sequences

1
Data Mining in DNA Using the SUBDUE Knowledge
Discovery System to Find Potential Gene
Regulatory Sequences

• by
• Ronald K. Maglothin

2
Committee Members

• Dr. Lawrence B. Holder, Supervisor
• Dr. Diane J. Cook
• Dr. Lynn L. Peterson

3
Outline

• DNA Sequence Domain
• SUBDUE Knowledge Discovery System
• Experiments with Unsupervised SUBDUE
• Experiments with Supervised SUBDUE
• Conclusion and Future Work

4
DNA Structure

• All cells use DNA to store their genetic
  information.
• A DNA molecule is composed of two linear strands
  coiled in a double helix.
• Each strand is made of the bases adenine (A),
  thymine (T), cytosine (C), and guanine (G),
  joined in a linear sequence.

5
DNA Sequence

• These four bases constitute a four-letter
  alphabet that cells use to store genetic
  information.
• Molecular biologists can break up a DNA molecule
  and determine its base sequence, which can be
  stored as a character string in a computer

TTCAGCCGATATCCTGGTCAGATTCTCT
AAGTCGGCTATAGGACCAGTCTAAGAGA
6
Genes

• A gene is a DNA sequence that encodes
  instructions for building a protein.
• Gene expression is the process of using a gene to
  make a protein

transcription
translation
DNA
RNA
Protein
gene
transcript
product
7
Gene Regulation

• Primary mechanism is to control the rate of DNA
  transcription
• Faster transcription more protein
• Slower transcription less protein
• Transcription rate is controlled by transcription
  factors, which are proteins which bind to
  specific DNA sequences.

8
Human Genome Project

• A U.S.-led, worldwide effort to determine the
  complete DNA sequence for humans, as well as
  several other organisms.
• These sequences will be used to study
• mechanisms of disease
• growth and development
• evolutionary relationships

9
A Genome is a LOT of Data

• Raw sequence (text)
• Human (2005) 3 x 10 9 base pairs
• Yeast (finished) 1.2 x 107 base pairs
• Annotated sequence (Relational DB)
• Links to 3D structures of protein products, other
  genes in family, known transcription factors,
  journal references, and other databases.

10
A Rich Domain for Knowledge Discovery

• Most of the sequences (and genes) have unknown
  function.
• Efficient algorithms are needed to
• identify important patterns
• identify and classify possible genes
• infer relationships between genes
• predict protein structure

11
The SUBDUE Knowledge Discovery System

• Input A graph G
• Output A list of substructures that compress G
  well
• Uses a computationally-constrained beam search
  and inexact graph match

12
What is a substructure?

• A definition subgraph and a list of subgraph
  instances

Input Graph
next
next
next
next
next
next
A
A
A
T
T
C
G
1
7
6
5
4
3
2
Substructure
Definition
Instances
next
A
T
next
A
T
1
2
next
A
T
6
5
13
MDL Heuristic

• SUBDUE uses the Minimum Description Length
  Principle to evaluate substructures.
• Description Length of a graph is the number of
  bits needed to send the graphs adjacency matrix
  to a remote computer.
• Goal is to minimize DL(S) DL(GS).

14
SUBDUE Parameters

• Iterations Graph is compressed using the best
  substructure, discovery is restarted
• Threshold Controls how much two subgraphs can
  differ to be considered similar
• Beam Width The number of substructures in the
  expansion list

15
Graph Representations

• Simple linear

next
next
next
next
A
A
T
C
G

• Downstream edges

4
3
3
2
2
2
1
1
1
1
A
A
T
C
G
16
Graph Representations

• Start vertex

5
4
3
2
next
next
next
next
1
A
A
T
C
G
Start

• Backbone

next
next
next
next
base
base
base
base
base
name
name
name
name
name
A
A
C
T
G
17
Graph Representations

• Backbone-star

star
star
star
star
star
next
next
next
next
base
base
base
base
base
name
name
name
name
name
A
A
C
T
G
18
Unsupervised SUBDUE

• Input An entire yeast chromosome

next
next
next
next
A
A
T
C
G

• Heuristic

• Results Not good patterns with two
• to three bases

19
Polynomial Heuristic
20
Unsupervised SUBDUE -Discussion

• Random noise is not a meaningful kind of pattern
  variation in DNA.
• Unsupervised SUBDUE finds DNA patterns that are
  hard to evaluate and that are not focused on any
  target concept.
• We need to give SUBDUE more targeted input data
  and to modify the system to use it effectively.

21
Supervised SUBDUE

• Give SUBDUE two graphs a graph of positive
  instances of a target concept, and a graph of
  negative instances.
• SUBDUE discovers substructures in the positive
  graph, finds instances in the negative graph, and
  bases the overall heuristic value on the values
  in both graphs.

22
New Data Sets

• Clusters of coexpressed yeast genes compiled by
  Brazma et al., from expression data generated by
  DeRisi et al.
• The expression level of each gene in a cluster
  changed at the same time and by a similar degree
  during the experiment perhaps some genes in a
  cluster are regulated by similar mechanisms?

23
New Data Sets

• Positive examples
• 300-bp upstream windows (both strands) for all
  genes in a given cluster
• Negative examples
• 300-bp upstream windows for genes not in the
  cluster, OR
• 300-bp windows randomly selected from the
  complete genome (probably not involved in gene
  regulation)

24
Supervised Heuristic

• Based on the substructures values in the
  positive and negative graphs

• Numerator set to 1 when no

negative instances
25
Compression Ratio

• Normalize the graph values by using the inverse
  of the graph compression

26
Negative Graph Value

• When there are no negative instances, setting
  numerator to 1 actually penalizes such
  substructures.
• Using 2 x DL(G-) in this situation gave better
  results.

27
Ratio Heuristic Results
28
Concept DL Heuristic

• Based on the size of a message containing the
  compressed positive graph, plus the errors
  (negative instances).

29
Concept DL Heuristic Results

• Relative graph size affected results

30
Backbone Representation
next
next
next
next
base
base
base
base
base
name
name
name
name
name
A
A
C
T
G

• Base vertices allowed dont-care positions, but
  heuristic had to be changed to accommodate them.
• Overlap became very important.

31
DL Equations
32
Negative Graph Value

• Using 2 x DL(G-) for no negative instances
  favored such substructures too strongly.

33
Compression Difference Heuristic

• Use subtraction with the compression values
  instead of division.

34
Results
Cluster cr4.111101.77
35
Results
Cluster c2_4.2222200.39
36
Results of Brazma et al.
Cluster c2_4.2222200.39
37
Brazma Heuristic

• Based on number of positive and negative instances

38
SUBDUE Using Brazma Heuristic
Cluster c2_4.2222200.39
39
Conclusion

• SUBDUE can be used to discover likely
  transcription factor binding sites.
• Patterns found by SUBDUE are different from those
  found by string-based algorithms, due to the
  graph representation, beam search, and different
  search heuristic.

40
Conclusion

• Patterns found by unsupervised SUBDUE in DNA are
  difficult to evaluate.
• Using supervised SUBDUE can greatly focus the
  search on the target concept.
• Choosing the right graph representation and
  heuristic are critical to success.

41
Future Work

• Further refinement of the supervised MDL
  heuristic.
• Application of graph grammar theory to SUBDUEs
  search.
• Close collaboration with molecular biologists to
  select data sets and evaluate results.