Selforganizing maps in microarray data clustering

1
Self-organizing maps in microarray data clustering

• Tomas Singliar
• CS3750 Final Project Report

2
Self-organizing maps

• Neural network architecture
• One unit active winner
• Lateral inhibition is equivalent concept
• Found in biological systems

• Model
• Neural units arranged in grid (various
  topologies)
• Weights associated with units

3
Self-organizing maps - computation

• Initialize the weights
• Picture in 2D input space!
• Repeat
• Choose a random datapoint
• Find winning neuron and its neigborhood
• Update neuron and neighborhood
• Pull them closer to datapoint
• Winner-more, neighbors - less
• Result after one iteration
• After many iterations
• The units disentangle
• Positions correspond toagglomerations of data
• Sensitive to initial settings

4
Microarray data

• Gene expression data -DNA array
• thousands of assaysat one time
• High throughput
• One drop one gene
• Fluorescent light applied
• Colors expression level
• Computational nightmareLots of data, little
  structure
• thousands-dimensionalinput space

5
Figure of Merit

• Does the clustering make sense?
• A squared distance based measure (smallgood)
• measures predictive power wrt e
• aggregate
• adjustment for number of classes

6
FOM, observation clustering
7
Visualization with SOMs

• You know the number of clusters
• Need to see their spatial distribution

8
Clustering pictures - nice
9
Clustering pictures - real

10
Human lymphoma dataset

11
Human lymphoma dataset

12
Take-home messages

• Self Organizing Maps are used for
• clustering and visualization
• Computational resources an issue
• How many clusters are there?
• Figure of Merit measure worked
• Microarray data are notoriously hard
• dont have clearly defined clusters
• preprocess them right!
• Semester is ending, festivities begin

13
Thank you!

• Questions welcome for 3 minutes