Data Mining For Bioinformatics: Tools and Applications

1
Data Mining For Bioinformatics Tools and
Applications

• Craig A. Struble
• Department of Mathematics, Statistics, and
  Computer Science
• craig.struble_at_marquette.edu

2
Overview

• Clustering
• Hierarchical Clustering, SOMs, Model-based
  clustering
• Classification
• SVMs, neural networks
• Tool building

3
Clustering

• Basic idea
• Group similar things together
• d(x,y) Distance function between x and y
• Euclidean, mismatches, etc.
• Bioinformatics context
• Similar expression profiles imply similar
  function
• This is under some scrutiny
• Unsupervised
• Useful when no other information is available
• Just to see what happens

4
Example Data

• Genes x Experiments
• 6000 genes x 16 experiments
• Could use ratios or other values for data

5
Hierarchical Clustering

• Bottom up (agglomerative)
• Top down (divisive)
• Linkage
• How groups are combined (or split)

6
Hierarchical Clustering (Example)

• Analyzing yeast data (different experiment)

7
Self Organizing Maps

• Also called Kohonen maps
• Example of neural networks

8
Self Organizing Maps

• Yeast data set
• Typical values
• Error bars
• Using GeneSOM in R
• Many other visualizations

9
Clustering With Models

• Create/select representative points (i.e. models)
• Perform cluster analysis (K-means, K-medoids,
  etc.)
• Classify/identify real data items by finding
  which representative they cluster with

10
Example Yeast SporulationChu et.al. Science 282
11
Example Yeast Sporulation
12
After Clustering

• Try multiple sequence alignment of genes closely
  clustered together
• Include upstream/downstream sequence to look for
  promoter regions, etc.
• Search in KEGG for metabolic pathways genes may
  be involved with
• Look at functional classification of genes in the
  same cluster
• May be able to assign putative function to genes
  with unknown function (again, this is under
  scrutiny)

13
Classification

• Use data to create a classifier
• A predictive model for labeling new data items
• Supervised learning
• Generate data associated with known labels
• Train the specific technique with labeled data

14
Support Vector Machines

• Find a hyperplane to separate data points

15
Feature Selection

• Identify subset of attributes as most important
• Identify group of genes that play most important
  role in distinguishing classes
• Use information gain or other statistical
  measures to determine the importance of a data
  item
• In many cases, feature selection is the true goal

16
Example Leukemia Classification

• Ovarian Cancer (Furey et al, 2000)
• Also tested on Leukemia data (Golub et al, 1999)
• 97,802 DNA clones used
• 31 tissue samples
• Cancerous and normal ovarian tissue
• Non-ovarian tissue

17
Example (cont.)

• Feature selection
• High scoring genes differ most on average and
  have small deviations in value
• 50 relevant clones identified
• Leave one out testing
• 80 accuracy
• Really testing if selected features are good

18
Neural Networks

• Network of connected neurons
• Trained with data with known output values
• Errors propogated through network for learning

Input Layer
Output Layer
Hidden Layer
19
Example Cleavage Site Prediction (Nielsen et al,
1997)

• Predict where cleavage sites are in protein
  precursors
• Training data from SWISS-PROT database

20
Example (cont.)

• Input layer
• Groups of 20 neurons per location
• 20 amino acids
• Sliding window of 5-39 amino acids
• Hidden layer
• 0-10 neurons tried
• Output layer
• 2 neurons per location, P(c),P(s)
• Trained with backpropogation

a
P(c)

P(s)
v
a



v
a

P(c)
v
P(s)
21
Tool Building

• Many commercial packages have lots of tools
• Not always integrated
• May not provide enough flexibility
• Sometimes youve just gotta do it yourself

22
Tool Building

• Identify problem to work on
• E.g. predict where miRNAs are on the genome
• Determine where to get data
• E.g. NCBI, KEGG, literature
• Determine the final format of your data
• E.g. Oracle, PostgreSQL, CSV, XML, etc.
• Select data mining techniques to use
• Literature and experience

23
Tool Building

• Select user interface style
• Web-based vs. applet vs. application
• Upload data/download data?
• Select visualizations
• Decide how to present the data
• Communicate with your users

24
Tool Building

• Select software to use
• Does it contain the data mining techniques to
  use?
• Is the source code available?
• Library vs. application vs. interpreter
• What do you know and what are you willing to
  learn?
• Eventually, youll build up a collection of tools
  to build on

25
Typical architecture for small project
Neural Network
Output
Perl Scripts
Perl Scripts
Clustering
26
Where to Find Examples

• Search Google (http//www.google.com)
• E.g. self organizing maps bioinformatics
• Citeseer (http//citeseer.nj.nec.com/)
• PubMed (http//www.ncbi.nih.gov)
• Web pages and papers usually contain links to
  software used

27
Be Prepared

• Programming for bioinformatics/data mining often
  requires knowing many languages (Perl, C, Java
  at a minimum)
• Practice on supplied sample data sets if any
• Read, read, read!

28
Useful free tools

• R (http//www.r-project.org)
• Statistics, plotting, clustering, etc.
• Octave (http//www.octave.org)
• Matlab clone (Matlab itself is EXTREMELY useful
  and popular)
• PostgreSQL (http//www.postgresql.org)
• Powerful ORDBMS (when you cant afford Oracle)
• http//bistro.mscs.mu.edu/cstruble/biolinks.html