Views: Fundamental Building Blocks in the Process of Knowledge Discovery

1
Views Fundamental Building Blocks in the Process
of Knowledge Discovery

• Hideo Bannai1 , Yoshinori Tamada2
• Osamu Maruyama3, Kenta Nakai1, Satoru Miyano1
• 1 Human Genome Center, Inst. of
  Medical Science,
  Univ. of Tokyo
• 2 Dept. of Mathematical Sciences,
  Tokai Univ.
• 3 Faculty of Mathematics, Kyushu
  Univ.

2
Motivation

• Speed up the computational knowledge discovery
  process as a whole

3
The Knowledge Discovery Process
Evaluation/ Interpretation
Fayyad et al. 1996

• The keys to success
• Develop a good algorithm for the problem
• Find or generate appropriate attributes
• Allow human intervention

Data Mining
Transformation
Patterns
Preprocessing
Transformed Data
Selection
Preprocessed Data
Data
Target Data
4
Contents

• Basic Concepts
• Entity, View, View Operation, View Design
• Example discovery process
• Characterization of sub-cellular localization
  signals

5
Basic Concepts - Entity

• An entity is any object.
• Example
• The set E of S. cerevisiae gene names. Each gene
  name is an entity.
• YAL069W, YAL068C, YLR188W,

6
Basic Concepts - View

• A view is a function over entities
• v E ? R call R the range set
• Example
• View over the set of S. cerevisiae genes names

atgatcgtaaataa
atgcatcgtaaatc
atgattgtaagaat
7
Basic Concepts - View Operation

• A view operator is a function that creating new
  views from existing views.
• A function ? R?R over range sets induces a
  view operator ?v?v
• v E ? R, ?R?R
• ?v ? ?v v E ? R ? R
• ?v(e) ?(v(e))

8
Basic Concepts - View Operation

• Example codon translation

MIVNNTH
MHRKSLRR
MIVRMIRL
9
Basic Concepts - View Operation

• Example String Matching

Parameter patternatcgt
true
true
false
10
Basic Concepts - View Operation

• Hypothesis Generation is a view operation
  (Hypothesis view)
• Let E entity set, V views available
• Supervised Learning (v target view)
• SLearner E, V, v ? v where for most
  e?E, v(e) v(e)
• Unsupervised Learning
• ULearnerE, V ? v

11
Basic Concepts - View Operation

• Example ID3 algorithm and Decision Tree View

ID3
POS
POS
NEG
NEG
POS
POS
12
View Design

• Views and view operators can be combined to
  create new views.
• We call the structure of this combination, the
  view design
• Careful design of views by experts is an
  interface for human intervention

13
The Knowledge Discovery Process

• Data Preparation Transformation of entity set
  using views
• Data Mining Generation of new views from
  existing views and entities
• Interpretation/Evaluation Evaluation of the
  generated views
• Knowledge Consolidation Using newly created
  views, for the next discovery task

14
Sub-cellular Localization Signals

• Proteins are carried to specific locations after
  synthesized in the cytosol.
• A short amino acid sequence holds the information
  (localization signal)
• Signals at the N-terminal
• signal peptide (SP)
• mitochondrial targeting peptide (mTP)
• chloroplast transit peptide (cTP)

15
Previous Work

• PSORT 1992 Nakai, Kanehisa
• An expert system to predict localization sites
• Looks at the amino acid frequency of first 20
  residues -gt linear discrimination
• MitoProt 1996 Claros?Vincens
• Linear discrimination using 47 numerical
  attributes trying to model various knowledge
  about the signals (Mitochondrial only)
• TargetP 2000 Emanuelsson et al
• Neural network system (Best predictor so far)

16
Our Approach

• Find simple interpretable rules, which are still
  accurate
• Design views based on discussion with an expert
• Data
• 940 sequences of SP(269), mTP(368), cTP(141),
  Other(162)

17
View Design for Experiments
View Design 1 (binary classifier)

• Can we find structural characteristics?

MLRSAVRLAGKDVRFGEDA
(0, 15)
MLRSAVRLAGKDVRF
001000100000010
R?1 Other?0
True
0100010
18
View Design for Experiments
View Design 2 (binary classifier)

• What about overall composition?

MLRSAVRLAGKDVRFGEDA
(5, 10)
VRLAGKDVRF
0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 1.0, 0.0
Net Charge
0.1
False
0.8
19
View Design for Experiments
View Design 3 (binary classifier)
View Design 1
View Design 2
Logical And
20
View Design for Experiments

• Decision list of View Design 2 and 3
• Hypothesis generation
• DListE, view2, view3, v ?

P1
P2
P3
sequence
SP? View Design 2
Other? View Design 3
mTP? View Design 3
no
no
no
no
cTP
yes
yes
yes
SP
Other
mTP
21
Results

• 5-fold cross validation
• Accuracy was close to TargetP
• Our view
  TargetP
• SP 237/269 88.1 245/269 91.1
• mTP 304/368 82.6 300/368 81.5
• cTP 112/141 79.4 120/141 85.1
• Other 141/162 87.0 137/162 84.0

22
Results

• Other vs (cTP mTP)

View Design 1
View Design 2
Logical And
23
Results

• The rules discovered were interpreted by an
  expert
• We were able to extract interpretable knowledge
  about the signals.
• Hydropathy of SP
• A pattern capturing the amphiphilic a-helix of
  cTP and mTP
• The expert was surprised that such simple rules
  could predict so accurately

24
Hypothesis Creator

• Library implementing views, and view operators
  http//www.HypothesisCreator.net/

25
HC Core Team

• Yoshinori Tamada
• Osamu Maruyama

• Hideo Bannai
• Satoru Miyano

26
Collaborators

• Takao Goto
• Hirotaka Katou
• Kentaro Kawamoto
• Hideo Miyake
• Kaori Samuraki

27
Alphabet Indexing

• A mapping from one character to another.
• It is generally used to classify an alphabet, to
  a smaller alphabet.
• Example
• K,R,D,E ? 1
• ACFGHILMNPQSTVWY ? 2
• MALRAVRSVRA ? 22212212212

28
AAindex Database

• An entry in the AAindex database is
• a mapping from an amino acid to a numeric value,
  representing biochemical properties of the amino
  acid.
• Example