MicroArray Data Analysis Candice Quadros

1
MicroArray Data AnalysisCandice
QuadrosAmol Kothari
2
Neural Network for classification

• Harnessing the power of a neural network for
  classifying samples.

3
Neural Network for classification

• Reduce the no. of genes
• We have to reduce the data dimensionality, i.e.
  reduce the no. of genes to consider.
• PCA can be used to select most informative genes,
  but it is computationally expensive to obtain the
  Eigen vectors for high dimensional data.
• Use the method suggested by Golub et al. to
  obtain the informative genes.

4
Neural Network for classification

• Steps in classification
• Obtain the informative genes using Golubs
  method.
• Normalize the genes by shifting them to the mean
  dividing by the standard deviation.
• Train the neural network by using the training
  data targets, and get the weights.
• Classify the test data using the weights obtained
  above.

5
Neural Network for classification

• Results obtained

Inform. Genes No. of Hidden Units NN Accuracy Golub Accuracy
100 3 70.55 61.76
200 13 76.74 58.82
6
Hierarchical Merging When to stop?

• Question When to stop the merging?
• Suggested Solutions
• Diameter(C) ? MaxD
• Avg(sim(Oi,Oj)) ? (Oi,Oj ?C)
• Difficult to estimate the parameters in high
  dimensions.

7
Hierarchical Merging When to stop?

• Another solution When m clusters are present,
  stop merging.
• Problem The m clusters might contain single
  point clusters.
• Use the concept of MinPts (from DBScan). A set of
  points is a significant cluster only if the set
  has MinPts.
• When there are m significant clusters, then stop.

8
Hierarchical Merging When to stop?
No. of Significant Clusters
No. of iterations
9
Visualization of data Vizstruct
10
Visualization of data Vizstruct

• Equation used
• How do weigh each dimension, i.e. how do we
  select ?? Default value 0.5
• Use the Eigen Values of each dimension to obtain
  the value of ?.

11
Visualization of data Vizstruct

• Steps for visualization
• Project the data into Eigen space.
• The Eigen values of each dimension i ?i
• Now use the same formulae for calculating the 2D
  point

Where ?i Eigen value of the ith dimension
12
Visualization of data Vizstruct

• Results
• The visualization obtained by this method is more
  representative of the data, compared to
  Vizstruct.
• Demo

13

• Thank You