Making the Most of Small Sample High Dimensional Micro-Array Data

1
Making the Most of Small Sample High Dimensional
Micro-Array Data

• Allan Tucker, Veronica Vinciotti, Xiaohui Liu
• Brunel University
• Paul Kellam Windeyer Institute

2
MicroArray Data

• High dimensional
• Small number of samples
• Need to identify predictive genes
• E.g. classification
• Rate confidence on genes based upon predictive
  ability / classification

3
Identifying Predictive Genes

• We use Naïve Bayes Classifier
• Well established
• Minimises parameters
• Feature selection using SA
• Repeated 10 times
• Apply cross validation

4
Identifying Predictive Genes

• Identify genes robustly
• Data perturbed during CV
• Repeats of stochastic SA search
• Assign confidence based upon the frequencies of
  genes being selected
• Limit maximum number of links

5
Effect of Model Complexity
6
Classification Accuracy

• Generally RSN performs best
• SA global search better than local
• Anomaly with B-Cell?
• Synthetic data supports global over local

7
Confidence Scores

• Relatively small number of genes
• Identified with high confidence
• Consistency between runs

8
Identified Genes
9
Conclusions

• When micro-array data only has small samples
• Simple models with small parameters best
• Global search for parameters better
• Proposed RSN successfully identifes genes of
  interest paving way for further biological
  analysis
• Need to explore different parameters