Rapid Discrimination of Decoy Protein Structures Using Multiple Energy Functions

1
Rapid Discrimination of Decoy Protein Structures
Using Multiple Energy Functions

• Beating back bogus biomolecules since 2004.

2
Machine Learning vs. Bioinformatics vs.
Structural Biology in a battle to the finish!
3
Designing a Potential

• Preserve hydrophobic character
• Use standard threading f (Bryant Lawrence)
• Preserve hydrogen bonding
• Calculate Hbond geometry and strength
• Preserve van der Waals character
• Count bad contacts/distances
• Preserve good stereochemistry
• Ramachandran allowed/disallowed regions
• Preserve globular shape/size
• Limit radius of gyration
• Preserve secondary structure
• Assess predicted 2o to observed 2o

4
Energy Functions

• Threading Energy
• Bryant Lawrence
• Ramachandran Score
• Secondary Structure Score

Si score for residue i according to the
Ramachandran plot core 0, allowed 1,
generally allowed 2, disallowed 3
5
Energy Functions

• Radius of Gyration Score

Dij C? distance between residue i and
centroid Pradgyr n0.33 2.4
6
Energy Functions

• Hydrogen Bond Score

Ei Hydrogen bond energy for residue i with
hydrogen bond Nh Number of residues involved
in hydrogen and disulfide bonds
7
Energy Functions

• Bump Score

8
Other Scores

• Average Hbond Energy
• Hbonds (and various functions thereof)
• Hbonds/SS
• (Actual - Theoretical) Radius of Gyration (and
  absolute value)
• Secondary Structure Errors
• Beta, Helix, Coil
• Statistical chi functions

9
Other Inputs

• Secondary structure as calculated by PsiPred
  based only on the protein sequence
• Actual secondary structure as computed by VADAR

10
Development Data Sources

• Real structures Richardson lab Top500 set, as
  assessed by their All-Atom Contact method
• http//kinemage.biochem.duke.edu/databases/top500.
  php
• Decoy Structures 76,200 decoys of 41 structures
  derived from the Rosetta process
• (Baker Lab, from Tsai et al, 2001)

11
Upcoming Data Sources

• PDB Select 25 set
• Trents NMR structures
• Structures generated by Homodeller for BacMap
• Homodeller forced to do bad threading

12
Initial Idea

• Try to predict RMSD from real structure based on
  some combination of the scoring functions (using
  MINER and other methods)
• Modest success R0.7, Spearman 0.7

13
Simpler Task

• Just distinguish real from decoy structures.
• Unclear which scores are important, so use data
  driven method.

14
C4.5 (Quinlan)

• Induces a decision tree from data
• Splits the data set at branch based on
  information-theoretic criteria
• Very general
• Handles discrete and numeric attributes equally
  well.

15
Sample Decision Tree

• chi4n lt 0.502986
• beta_perc lt 0.1875
• hbond_erg lt -1.301656
• helix_perc lt 0.675
• rofg lt 0.455691 no
  (10.0)
• rofg gt 0.455691 yes
  (16.0/1.0)
• helix_perc gt 0.675
• bump lt 0.000773
• hbond_erg lt
  -1.448668
• ss lt 0.740741
  yes (13.0/2.0)
• ss gt 0.740741 no
  (2.0)

16
Should I Go To This Talk?
Should I Go To This Talk?
Information S(prob)-log(prob) 0.811 bits
17
Split on Interesting
Info 0.721 bits
Info 0.918 bits
Total Info 6/8 0.721 2/8 0.918 0.771
18
Split on Food
Info 0 bits
Info 0.918 bits
Total Info 3/8 0.918 0.344
19
Real/Decoy Decision Tree

• 127 branches, 64 leaves (equivalently 64 rules)
• Real structures (446) 384 right, 62 wrong
• Decoys (63546) 63537 right, 9 wrong
• Given a real structure, the tree will identify it
  correctly 86 of the time.
• Given a decoy, theres an extremely good chance
  (99.986) it will be identified as such.

20
The Most Effective Rule

• If abs(actual radius of gyration theoretical
  radius of gyration) lt 2.698442, AND
• (Fraction of Residues Predicted As Beta and
  Evaluated as Helix) 0, AND
• Chi Score4 gt 0.502986, AND
• (Fraction of Residues Predicted As Beta and
  Evaluated as Beta) lt 0.468750, THEN
• Its a decoy
• Correct in 49862 cases, wrong in 36.
• A real finding, or a bias in how Rosetta
  generates decoys?

21
Chi Scoring Functions

• Another decision tree
• Inputs
• amino acid type
• Phi
• Psi
• Next AA
• Next chi
• Preceding AA
• Preceding Chi
• Coil or Not
• (Chi2 angle)

22
Chi Results

• 1296 branches, 917 leaves
• w/ chi2, correctly predicts 80 of chi angles
• w/o chi2, correctly predicts 73 of chi angles

23
Going Forward

• Use a lot more datasets and different
  combinations thereof.
• Try predicting RMSD within classes (0 2 A, 2
  4 A, etc.)
• Examine incorrectly classified structures more
  carefully
• Test with contrived structures (i.e. real
  structure with polar residue in core)
• Different machine learning methods.

24
Acknowledgements

• Joey Cruz
• The Proteome Analysts (Roman Eisner, Alona Fyshe,
  Brandon Pearcey, Brett Poulin)
• Warren Ward
• Dr. Wishart
• Haiyan Zhang