Regression based KNN for gene function prediction using heterogeneous data sources

1
Regression based KNN for gene function prediction
using heterogeneous data sources

• Zizhen Yao, Larry Ruzzo
• yzizhen, ruzzo _at_cs.washington.edu

2
Background

• E. Coli classification schemes
• KEGG , COG, MultiFun
• Common functional classes (10-19 classes)
• Metabolism, Translation, Transporter, Cell
  Motility
• Biological information used for inference
• Microarray expression, protein interaction,
  evolutionary history
• Methods
• Support vector machine, Bayesian, Rule-based

3
Introduction to KNN

• Idea for each query instance
• Choose k nearest neighbors
• Choose the class voted by majority of the
  neighbors.
• Design issues
• Similarity / Distance metric
• Voting schemes

4
Algorithm Flow Chart
Training
Testing
Training Data
Testing Data
For every pair of training genes, calculate the
predictors.
Calculate the predictors values using and
training data
Learn Similarity Metric
Choose k nearest neighbors
Voting
A list of predictions with confidence scores.
5
Predictors

• Microarray Expression Data
• Expression correlation
• Sequencing Data
• Chromosomal position
• Chromosomal distance
• Transcription direction
• Block indicator
• Protein sequence similarity
• Paralog indicator

6
Similarity (Distance) Metric

• Classical metrics are not appropriate because
  predictors are
• heterogeneous data type, scale
• different relevance
• correlated
• Goal estimate the likelihood that a pair of
  genes are in the same class based on predictors

7
Learning Similarity Metric

• Regression methods
• Response
• Find f
• Logistic regression
• Local regression

8
Probabilistic voting scheme

• Goal estimate the probability that the query
  gene belong to each class.
• Range 0 1
• Assigns higher confidence score to predictions
  voted by more neighbors, or neighbors with higher
  credibility.
• Report predictions that are above certain
  threshold value.

9
Performance comparison
10
Functional Classes ROC analysis (KEGG)
11
Confidence Score vs. Accuracy
12
Results Summary

• Combining all 4 predictors yields the best
  result.
• Using expression data only, regression based KNN
  methods outperforms SVM.
• Performance varies with different function
  classes
• Confidence scores are strongly correlated with
  accuracy.

13
Contribution

• KNN
• Simplicity, efficiency, flexibility
• Easy to interpret the results, useful to guide
  case studies
• Similarity metric
• integrate heterogeneous data sources
• voting scheme
• Statistic inference
• A general framework to incorporate other
  information.