Ubiquitination Sites Prediction

1
Ubiquitination Sites Prediction

• Dah Mee Ko
• Advisor Dr.Predrag Radivojac
• School of Informatics
• Indiana University
• May 22 2009

2
Outline

• Ubiquitination
• Machine Learning
• Decision Tree
• Support Vector Machines
• Prediction of ubiquitination sites
• Influence of sequence
• Influence of structure
• Influence of evolutionary consideration

3
Ubiquitin

• A small protein that occurs in all eukaryotic
  cells.
• Highly conserved among eukaryotic species.
• Consists of 76 amino acids and has a molecular
  mass of 8.5 kDa.
• Key features
• its C-terminal tail and Lys residues
• Human ubiquitin sequence
• MQIFVKTLTGKTITLEVEPSDTIENVKAKIQDKEGIPPDQQRLIFAGKQL
  EDGRTLSDYNIQKESTLHLVLRLRGG

4
Ubiquitination

• Post-translational modification of a protein
• Covalent attachment of one or more ubiquitin
  monomers to Lys residues
• Reversible
• Target proteins for degradation by the proteasome

5
Functions of Ubiquitination

• Monoubiquitination
• Histone regulation
• DNA repair
• Endocytosis
• Budding of retroviruses from the plasma membrane
• Polyubiquitination
• Protein kinase activation

6
Machine Learning

• Machine learning is programming computers to
  optimize a performance criterion using data and
  past experience.
• Learn general models from a data set of
  particular examples.
• Build a model that is a good and useful
  approximation to the data.

7
Machine Learning

• Supervised learning
• Learn input/output patterns from given correct
  output.
• Split data into training and test set.
• Train element on training data.
• Evaluate performance on test data.
• Unsupervised learning
• Learn input/output patterns without known output.

8
Machine Learning Decision Tree

• One of classification algorithms
• Each internal node tests the value of a feature
  and branches according to the results of the
  test.
• Each leaf node assigns a classification.

9
Machine Learning Random Forest

• A machine learning ensemble classifier
• Consists of many decision trees
• Each tree is constructed using a bootstrap sample
  of training data.
• After a large number of trees are generated each
  tree casts a unit vote for the most popular class.

10
Machine Learning Support Vector Machines

• Viewing input data as two sets of vectors in an
  n-dimensional space an support vector machine
  will construct a separating hyperplane in that
  space.
• The hyperplane maximizes the margin between the
  two data sets.

11
Machine Learning Support Vector Machines

• H3 does not separate the classes.
• H1 separates with a small margin.
• H2 separates with the maximum margin.

• If a data set is not linearly separable map into
  a higher-dimensional space using kernel approach.

12
Data Sets for Prediction

• 334 protein sequences from yeast
• Positive and negative sites with 25 amino acid
  residues centered at lysine
• Remove all positive and negative sites that have
  more than 40 identity inside the data sets.

13
Features Sequence Information

• Relative amino acid frequencies Entropy Net
  charge Total charge Aromatics
  Charge-hydrophobicity ratio Protein disorder
  probability Vihinens flexibility Hydrophobic
  moments B-factors
• 64 X 4 256 features
• Relative amino acid frequencies
• Window size 11
• A 1/11 G 0/11 M 0/11 S 0/11
• C 0/11 H 0/11 N 1/11 T 1/11
• D 2/11 I 0/11 P 3/11 V 1/11
• E 0/11 K 1/11 Q 0/11 W 0/11
• F 0/11 L 0/11 R 0/11 Y 1/11

14
Features Evolutionary Information

• Position Specific Scoring Matrix
• 21 X 4 84 features
• Window size 11
• 256(Seq) 84(Evol) 340 features.

15
Features Structure Information

• BLAST each sequence against PDB database.
• Select alignments with greater than 30 identity.
• For each mapped site five shells with 1.5 3
  4.5 6 75Å radial boundaries are constructed
  from the residues alpha-carbon atom using X Y
  Z coordinates from PDB.
• Amino acid at the center site 20 features
• e.g. K A C D E F G H I K L M N P Q R S T V W Y
• 0 0 0 0 0 0 0 0 1 0 0 0 0 0
  0 0 0 0 0 0
• Each shell contains 24 features.
• 4 for counts of C N O S and 20 for counts
  of amino acids
• 20 24 x 5 140 features
• 60 sites among 245 positive sites 24
• 3239 sites among 12906 negative sites 25
• 1X140 zero vector for the other sites
• 256(Seq) 84(Evol) 140(Str) 480 features

16
Prediction Results Random Forest
17
Prediction Results Random Forest
18
Prediction Results SVM
19
Prediction Results SVM
20
Feature Selection

• Rank features using correlation coefficients (r).

21
Conclusions

• Ubiquitination sites are predictable.
• The accuracy is modest.
• Long range interactions
• Flexibility of structure
• Noise in positive sites
• Small data set
• The sequence features are the most important.

22
Acknowledgements

• Prof. Predrag Radivojac
• Wyatt Clark
• Arunima Ram
• Nils Schimmelmann
• Prof. Sun Kim
• Linda Hostetter
• School of Informatics

23
Thank you!