1. Protein Secondary Structures
Assignment and prediction
Pernille Haste Andersen

2. Outline What is protein secondary structure How can it be used?
Different prediction methods
Alignment to homologues
Propensity methods
Neural networks
Evaluation of prediction methods
Links to prediction servers

3. Secondary Structure Elements
ß-strand
Helix
Bend
Turn

4. Use of secondary structure
Classification of protein structures
Definition of loops (active sites)
Use in fold recognition methods
Improvements of alignments
Definition of domain boundaries

5. Classification of secondary structure
Defining features
Dihedral angles
Hydrogen bonds
Geometry
Assigned manually by crystallographers or
Automatic
DSSP (Kabsch & Sander,1983)
STRIDE (Frishman & Argos, 1995)
DSSPcont (Andersen et al., 2002)

6. Dihedral Angles phi - dihedral angle of the N-Calpha bond
From
phi dihedral angle of the N-Calpha bond
psi dihedral angle of the Calpha-C bond
omega dihedral angle of the C-N (peptide) bond

7. Helices phi(deg) psi(deg) H-bond pattern
alpha-helix i+4
pi-helix i+5
310 helix i+3
(omega = 180 deg )
From

8. Beta Strands Antiparallel Parallel phi(deg) psi(deg) omega (deg)
beta strand
Antiparallel
From
Parallel
From

9. Secondary Structure Elements
ß-strand
Helix
Bend
Turn

10. Secondary Structure Type Descriptions
* H = alpha helix
* G = helix
* I = 5 helix (pi helix)
* E = extended strand, participates in beta ladder
* B = residue in isolated beta-bridge
* T = hydrogen bonded turn
* S = bend
* C = coil

11. Automatic assignment programs
DSSP ( )
STRIDE ( )
DSSPcont ( )
The protein data bank visualizes DSSP assignments on structures in the data base
# RESIDUE AA STRUCTURE BP1 BP2 ACC N-H-->O O-->H-N N-H-->O O-->H-N TCO KAPPA ALPHA PHI PSI X-CA Y-CA Z-CA
A E , , , ,
A H , , , ,
A V , , , ,
A I E -A A , , , ,
A I E -A A , , , ,
A Q E -A A , , , ,
A A E +A A , , , ,
A E E +A A , , , ,
A F E -A A , , , ,
A Y E -A A , , , ,
A L E >> -A A , , , ,
A N T 45S , , , ,
A P T 45S , , , ,
A D T 45S , , , ,

12. Secondary Structure Prediction
What to predict?
All 8 types or pool types into groups
DSSP Q3
* H = alpha helix
* G = 310 -helix
* I = 5 helix (pi helix)
* E = extended strand
* B = beta-bridge
* T = hydrogen bonded turn
* S = bend
* C = coil H E C

13. Secondary Structure Prediction
Straight HEC
What to predict?
All 8 types or pool types into groups Q3
* H = alpha helix
* E = extended strand
* T = hydrogen bonded turn
* S = bend
* C = coil
* G = 310-helix
* I = 5 helix (pi helix)
* B = beta-bridge H E C

14. Secondary Structure Prediction
Simple alignments
Align to a close homolog for which the structure has been experimentally solved.
Heuristic Methods (e.g., Chou-Fasman, 1974)
Apply scores for each amino acid an sum up over a window.
Neural Networks
Raw Sequence (late 80’s)
Blosum matrix (e.g., PhD, early 90’s)
Position specific alignment profiles (e.g., PsiPred, late 90’s)
Multiple networks balloting, probability conversion, output expansion (Petersen et al., 2000).

15. Improvement of accuracy
1974 Chou & Fasman ~50-53%
1978 Garnier 63%
1987 Zvelebil 66%
1988 Quian & Sejnowski 64.3%
1993 Rost & Sander %
1997 Frishman & Argos <75%
1999 Cuff & Barton 72.9%
1999 Jones %
2000 Petersen et al %

16. Simple Alignments Solved structure of a homolog to query is needed
Homologous proteins have ~88% identical (3 state) secondary structure
If no close homologue can be identified alignments will give almost random results

17. Propensities: Amino acid preferences in -Helix
Capping

18. Propensities: Amino acid preferences in -Strand

19. Propensities: Amino acid preferences in coil

20. Chou-Fasman propensities
Name P(a) P(b) P(turn) f(i) f(i+1) f(i+2) f(i+3)
Ala
Arg
Asp
Asn
Cys
Glu
Gln
Gly
His
Ile
Leu
Lys
Met
Phe
Pro
Ser
Thr
Trp
Tyr
Val

21. Chou-Fasman Generally applicable
Works for sequences with no solved homologs
But the accuracy is low!
The problem is that the method does not use enough information about the structural context of a residue

22. Neural Networks Benefits Drawbacks Generally applicable
Can capture higher order correlations
Inputs other than sequence information
Drawbacks
Needs a high amount of data (different solved structures). However, today nearly 2500 structures with low sequence identity/high resolution are solved
Complex method with several pitfalls

23. Architecture Weights Input Layer I K H E Output Layer E E C H V I I Q
Hidden Layer
Window
IKEEHVIIQAEFYLNPDQSGEF…..

24. Sparse encoding
Inp Neuron
AAcid A R N D C Q E

25. Input Layer I K E 1 E H V I I Q A E

26. BLOSUM 62 A R N D C Q E G H I L K M F P S T W Y V B Z X *

27. Input Layer I K E E H V I I Q A E -1 2 -4 2 5 -2 -3 -3 1 -2 -3 -1 -1I 2 K -4 E 2 E 5 H -2 V I -3 I -3 Q 1 A -2 E -3 -1 -1 -3 -2 -2

28. Secondary networks (Structure-to-Structure)
Weights
Input Layer H E H C
Output Layer E H C E C H E C
Window
Hidden Layer
IKEEHVIIQAEFYLNPDQSGEF…..

29. PHD method (Rost and Sander)
Combine neural networks with sequence profiles
6-8 Percentage points increase in prediction accuracy over standard neural networks
Use second layer “Structure to structure” network to filter predictions
Jury of predictors
Set up as mail server

30. PSI-Pred (Jones)
Use alignments from iterative sequence searches (PSI-Blast) as input to a neural network
Better predictions due to better sequence profiles
Available as stand alone program and via the web

31. Position specific scoring matrices (PSI-BLAST profiles)
A R N D C Q E G H I L K M F P S T W Y V
1 I
2 K
3 E
4 E
5 H
6 V
7 I
8 I
9 Q
10 A
11 E
12 F
13 Y
14 L
15 N
16 P
17 D

32. Several different architectures
Sequence-to-structure
Window sizes 15,17,19 and 21
Hidden units 50 and 75
10-fold cross validation => 80 predictions
Structure-to-structure
Window size 17
Hidden units 40
10-fold cross validation => 800 predictions
Output:
C C H H C C C
Output:
C C C C C C C

33. The majority rules
Combining predictions from several networks improves the prediction
Combinations of 800 different networks were used in the method described by
Petersen TN et al. 2000, Prediction of protein secondary structure at 80 % accuracy. Proteins

34. Activities to probabilities
Helix activities (output)
Strand activities (output)
Coil probabilities! (calculated)
Coil conversion
… 1.0
0.10 .
1.0

35. Benchmarking secondary structure predictions
EVA
Newly solved structures are send to prediction servers.
Every week

36. EVA results (Rost et al., 2001)
PROFphd 77.0%
PSIPRED 76.8%
SAM-T99sec 76.1%
SSpro %
Jpred %
PHD %
Cubic.columbia.edu/eva

37. Links to servers Several links: ProfPHD PSIPRED JPred
ProfPHD
PSIPRED
JPred

38. Practical Conclusions
If you need a secondary structure prediction use the newer methods based on advanced machine learning methods such as :
ProfPHD
PSIPRED
JPred
And not one of the older ones such as :
Chou-Fasman
Garnier