Leo Molokov

1
Predicting protein-protein interaction using SVM

• This presentation will probably involve audience
  discussion, which will create action items. Use
  PowerPoint to keep track of these action items
  during your presentation
• In Slide Show, click on the right mouse button
• Select Meeting Minder
• Select the Action Items tab
• Type in action items as they come up
• Click OK to dismiss this box
• This will automatically create an Action Item
  slide at the end of your presentation with your
  points entered.

• Leo Molokov
• Xidan Li
• Parhati Aibibula

2
Basic information

• Project is based on paper 1
• Different data set had been used.
• Tweaks in results.
• Comparison will be provided.
• Tests on parameter value were performed
• Data summarization procedure was changed.

3
Model built

• Feature vector
• Symmetric pairs introduced
• x p1p2, pi represents protein,
• p dwf, d domain information,
• wf amino acid frequencies.
• Kernel function
• Gaussian RBF with various parameter values
• (0.01, 0.02, 0.04, 0.06 and 0.08 were checked)

4
Data sets

• SGD
• S.Cerevisiae Genome Database
• http//www.yeastgenome.org/
• DIP
• Database of Interacting Proteins
• http//dip.doe-mbi.ucla.edu/
• pFAM-A
• Protein family classification
• http//www.sanger.ac.uk/Software/Pfam/

5
Data extraction
6
(No Transcript)
7
Training and test samples.

• Training
• Negatives to Positives ratio of 1.0, 2.0 and 4.0
  were considered.
• N/P1.0, 2.0, 4.0
• Positives number is constant and equal to 1500.
• P1500
• Symmetric pairs are generated
• Test
• P400, N/P180 (compared to used before N/P60)

8
Multiple SVM

• Dependency between training sample size and
  working time is polynomial.
• Training data set is split into subsets upon
  which several SVMs are trained.
• Data is summarized by
• Taking weighted average -OR-
• Taking minimum
• The first approach seems to work better on gt3
  SVMs
• Precision/sensitivity are actually increased

9
Comparison tables.

• Different data sets were used, although overall
  parameters dont fluctuate much.
• N/P60 on test set gives the similar values
• N/P180 is actually used.

10
Old data different N/P ratio
11
New data different N/P ratio
12
(No Transcript)
13
Old data various number of SVMs
14
New data various number of SVMs
15
Different summary approach
16
G-parameter of RBF
17
Conclusions

• Method may be used to yield a number of unknown
  interactions by chance.
• FPR is reasonably low.
• Sensitivity is high
• Precision exceeds random assignment analogue.
• Training on high N/P ratio sets is required
• Multiple SVMs increase the performance in both
  senses

18
Some Contras

• Negatives are assigned to some possibly positive
  pairs.
• Positive might not mean interacting. Old
  problem of statistics.
• Low sensitivity
• Studies performed only on one specie.

19
Acknowledgements

• Shinsuke Dohkan, Asako Koike and Toshihisa Takagi
• for Improving the performance of an SVM-based
  method for predicting protein-protein
  interactions and for keeping in touch and some
  valueable advices!
• Dia project team
• For the tool Dia!
• You all
• For listening