CASP 5

1
CASP 5

• Fifth Meeting on the Critical Assessment of
  Techniques for Protein Structure Prediction
• Robert Langlois

2
Purpose

• Establish in structure prediction from sequence
• Capabilities
• Limitations
• Accomplished by analysis of a large number of
  blind predictions

3
Prediction Categories

• CASP 5
• Comparative Modeling
• Fold Recognition
• New Fold methods
• CAFASP 3
• Automated Servers

4
Category Target Overlap
Comparative Modeling
Fold Recognition
Ab Initio
5
Comparative Modeling

• Exploit evolutionary relationships to produce 3D
  structures
• Has not changed in a few decades
• General Protocol
• Start by identifying the template
• Align sequences of target and template
• Model conserved then diverged regions
• Assign side chain conformations, refine model

6
Scoring

• GDT-TS
• Global distance test
• Number of Ca in prediction not deviating more
  than di from Ca in the target
• Under the condition of optimal super-position
• RMSD of Ca
• Percent of Correctly Aligned Residues

7
Comparative Modeling Results

• Over 39 proteins, consisting of 51 domains
• Able to identify templates with 6 identity
• Improvement of CASP4, lt17 identity
• Top 5 groups Initial predictions
• Murzin Knowledge-based personal approach
• Bujnicki-Janusz Automatic servers
• VENCLOVAS Multiple sequence alignment
• Ginalski Automatic servers
• GeneSilico

8
Group Results cont.
448 Murzin 425 VENCLOVAS 020 Bujnicki 453
Ginalski 517 GeneSilico
9
Overall Results

• Servers better than most human predictors
• 3D Shotgun meta-predictor
• Bakers ROBETTA server
• However, no group was able to optimize
• I.e. create many good models but cannot pick the
  best
• No model comes significantly closer to target
  than the template

10
Overall Results cont.
11
CM Conclusion

• Successes
• Matching template to target
• Performance of different methods has leveled
• Failures
• Cannot produce model, closer than template
• Cannot model features not inherited
• Future select the best of several models

12
Fold Recognition

• Common approaches taxonometric (SVM, NN.),
  threading, homology modeling
• Combines fold recognition, comparative modeling,
  and de novo approaches

13
Fold Recognition Methods

• Template based combines
• Correct template, comparative modeling and fold
  recognition servers
• Refinement, available programs, and manual
  inspection
• Ginalski
• Fragment Assembly Ab initio
• Rosetta identify small fragments from a library
  of existing structures
• TOUCHSTONE conserved contacts threading

14
Scoring

• Livermore GDT_TS, SOV_O, and LGA_Q
• 3 structural superposition, 2 sequence depend
• Incorporate scores from
• Dali http//www.ebi.ac.uk/dali/
• CE http//cl.sdsc.edu/ce.html
• Mammoth http//icb.mssm.edu/services/
• Dali CE compare intra molecular Ca geometries
• Mammoth structural alignments independent of
  contact maps, depend on unit vector RMSD distances

15
Fold Recognition Results

• Top 3 Outperform the Rest
• Baker comparative modeling and ab initio
• Ginalski combine servers and manual inspection
• Rychlewski meta-server 3D Jury

16
Top 20 Predictors
17
Example Rossmann-like a/ß
18
Multi-domain Failures
19
Prediction Beats Template
20
Unrealistic Models Scoring Well
21
Conclusions

• While overall scores are higher than CASP4
• Does it reflect better predictions or larger
  database?
• Automatic servers nearly reach manual predictions

22
New Fold Techniques

• Ab Initio Folding Engine
• Metropolis Rule
• Potentials
• Successful Groups use
• Fragment Methods
• Contact Maps

23
Scoring

• RMSD, LCS, SOV, GDT_TS, Visual
• GDT_TS agreed best with visual inspection
• LCS CASP3 holdover
• Sequence dependent structural alignment

24
Example (NF) H. in?uenzae
25
Example (NF) E. coli
26
Visual Table Results
27
Summary of Methods
28
Discussion

• Fragment methods better at choosing fragments
  than assembling them.
• Automatic Servers similar performance
• PROTINFO-AB, I-site/Bystroff, BAKER-ROBETTA
• 20 out of 25 groups use PSI-PRED

29
Conclusions

• Coordinate Predictions
• Impressive accuracy
• Shows great progress in understanding folds
• Convergence of fold recognition
• Fragment template library construction
• Secondary Structure
• Have reached limits

30
Conclusions cont.

• Residue-Residue Contacts
• Very limited improvement
• Fragment templates, better accuracy
• Not accurate enough to build a model from scratch

31
References
Aloy, P., A. Stark, et al. (2003). "Predictions
Without Templates New Folds, Secondary
Structure, and Contacts in CASP5." PROTEINS
Structure, Function, and Genetics 53
436-456. Kinch, L. N., J. O. Wrabl, et al.
(2003). "CASP5 Assessment of Fold Recognition
Target Predictions." PROTEINS Structure,
Function, and Genetics 53 395-409. Tramontano,
A. and V. Morea (2003). "Assessment of
Homolgy-Based Predictions in CASP5." PROTEINS
Structure, Function, and Genetics 53 352-368.