Protein Secondary Structure Prediction C606 Paper presentation

1
Protein Secondary Structure PredictionC606 Paper
presentation
2
The Papers

• Protein Secondary Structure Prediction based on
  Position Specific Scoring Matrices
• David T. Jones
• J. Mol. Biol. (1999)
• (PSIPRED)
• A Two-stage Classifier for Protein ?-turn
  Prediction using Support Vector Machines -
• Chiu et al.
• Unpublished (2006)

3
Protein Structure Review

• What you should already know
• 8 protein states
• HBEGITSC
• Reduce to 3 states
• G, H ? H
• E, B ? E
• All others ? C

4
Why Predict Structure?

• Structure implies function
• Functional sites are often
• Drug targets
• Protein interaction sites
• Thus, learning about structure helps us to
  understand life itself

5
PSIPRED Method

• Create a PSIBLAST database
• Extract non-redundant proteins from DB
• Filter to remove
• Repetitive regions
• Transmembrane regions

6
PSIPRED Method

• Run PSIBLAST using our DB
• 3 iterations
• Extract PSSM
• l x 20 where l is sequence length
• Each matrix entry Mij represents the likelihood
  of an amino type j appearing at i
• Scale to 0,1

7
PSIPRED Method

• Entries of PSSM becomes features
• Take a window of 15 aa

8
PSIPRED Method

• Neural network
• Inspired by biology
• Highly interconnected elements
• Learn by example to associate certain inputs with
  certain outputs
• Hidden layers encode this relation

9
PSIPRED Method

• 2 Neural networks
• 1st network
• 315 inputs (1521) (why 21?)
• 75 hidden units
• 3 outputs
• 2nd network
• 60 inputs (153)
• 60 hidden units
• 3 outputs

10
PSIPRED Evaluation

• Cross validation for SS is tricky
• if proteins in test and train fold are too
  similar results will be artificially high
• 3 partitions, 187 proteins
• No similarity across partitions
• Sequence
• Fold
• CATH (structure classification DB)

11
PSIPRED Results

• Cross validation
• Q3 76
• SOV 73.5
• CASP3 Q3
• PSIPRED 76.3
• JPRED 72.4 (includes PHD)

12
Predicting ?-turns

• ?-turns are a type of coil
• Frequent
• Important in ?-hairpins and ?-sheets
• Often in exposed regions
• Most likely to interact with other proteins

13
SVM

• Support Vector Machines
• Again, learn to associate input with output
• Represent input as points in n-dimensional space
• Define a maximum margin hyperplane to separate
  two kinds of input in space

14
?-turn Methods

• Input
• Primary sequence
• PSSM
• HYPROSPII output
• Combines PSIPRED and PROSP
• PROSP uses DB of small structure segments
• Sliding window
• w1 15
• w2 9

15
?-turn Methods

• 2 levels of SVMs

16
?-turn Methods

• 1st level

H/E
SVM_TREE1
SVM_TREE3
SVM_TREE2
17
?-turn Methods

• 2nd level
• Combine results from 1st level
• SVM_MAX_D
• pick most sure answer
• SVM_VOTE
• Majority rules vote
• If classified as Coil proceed to ?/? SVM

18
?-turn Methods

• ?/? SVM
• Takes
• output of SVM_VOTE/SVM_MAX_D
• PSSM
• Primary sequence
• HYPROSPII (SSE)

?/?
?
?

19
?-turn Results

• 10 fold cross validation, 426 proteins
• SVM_VOTE gt SVM_MAX_D
• SVM_TREE3 performs poorly, but has best precision
  
• reluctant to classify a coil, but usually right
  when it does

20
?-turn Results

• Q3 79.25 (?1.1)
• Current best is BetaTurn 1.1
• Q3 77.3
• No standard deviation reported

21
Questions?