Transcript Slide 1

Chapter 14
Protein Secondary Structure Prediction
Refresher
Proteins have secondary structures
These structures are essential to maintain the 3D structure of the protein
Secondary structure can be either of
• -helix
• -strand
• Coil
-helix H-bond between C=O and N-H of every 4+ith residue
3.6 aa per turn
1.5 Å / aa (= 5.4 Å per turn)
(fully extended peptide backbone = 3.5 Å / aa)
-strand H-bond between C=O and N-H of distant regions
Parallel or anti-parallel
Coiled coil
Hydrophobic amino acids interact
Secondary Structure Predictions
Prediction of conformation of each amino acid:
• H: -helix
• E: -strand
• C: Coil (no defined 2° structure)
Used for classification of proteins
Defining domains and motifs
Intermediary step towards 3° structure prediction
Globular and trans-membrane proteins are structurally very different
Required different algorithms to predict these two classes of proteins
•
•
•
•
•
Problem is not trivial
-helix based on short distance (4+i interactions)
-strand based on long distance (5 – 50+ residues)
Long range interaction predictions less accurate
Accuracy about 75%
Ab initio based
Statistical calculation of residues in single query sequence
Homology-based
Common 2° structure patterns in homologous sequences
Ab initio Methods
Chou-Fasman
Intrinsic property of residue to be in helix, strand or turn
structure
A, E, M common in -helices
N: residues in all protein structures
M: residues in -helices
Y: Total Ala in protein structures
X: Ala in -helices
Propensity Ala in -helix: (X/Y)/(M/N)
Value = 1: same distribution as average
Value > 1: more often in -helix than average
Value < 1: less often in -helix than average
6 residue window of which 4 is H  -helix
Window extended bidirectionally until P < 1.0
5 residue window of which 3 is E  -strand
Helix
Sheet
A.A.
Desig
natio
n
P
Desig
natio
n
P
Ala
H
Cys
i
1.42
i
0.83
0.70
h
1.19
Asp
Glu
I
1.01
B
0.54
H
1.51
B
0.37
Phe
Gly
h
1.13
h
1.38
B
0.57
b
0.75
His
I
Ile
h
1.00
h
0.87
1.08
H
1.60
Lys
h
1.16
b
0.74
Leu
H
1.21
h
1.30
Met
H
1.45
h
1.05
Asn
b
0.67
b
0.89
Pro
B
0.57
B
0.55
Gln
h
1.11
h
1.10
Arg
i
0.98
i
0.93
Ser
i
0.77
b
0.75
Thr
i
0.83
h
1.19
Val
h
1.06
H
1.70
Trp
h
1.08
h
1.37
Tyr
b
0.69
H
1.47
http://fasta.bioch.virginia.edu/fasta_www2/fasta_www.cgi?rm=misc1
Example Chou-Fasman
10
20
30
40
50
60
SRRSASHPTY SEMIAAAIRA EKSRGGSSRQ SIQKYIKSHY KVGHNADLQI KLSIRRLLAA
70
80
90
GVLKQTKGVG ASGSFRLAKS DKAKRSPGKK
HELIX
HELIX
HELIX
SHEET
SHEET
SHEET
.
.
.
.
.
.
SRRSASHPTYSEMIAAAIRAEKSRGGSSRQSIQKYIKSHYKVGHNADLQIKLSIRRLLAA
helix
<-------->
sheet
<----->
EEEEEEEEE
EEEEEE
turns T T
T
.
T
.
.
GVLKQTKGVGASGSFRLAKSDKAKRSPGKK
helix ------->
<------->
sheet EEEEEEEEE
turns
T
T
<-----------------
TT
T
EEEEEEEEEEEEE
T
1 HA1 SER A
29 ALA A
2 HA2 ARG A
47 SER A
3 HA3 ALA A
64 ALA A
1 SA 3 SER A 45 SER A
2 SA 3 GLY A 91 ARG A
3 SA 3 LEU A 81 GLY A
38
56
78
46
94
86
Garnier-Osguthorpe-Robson (GOR)
•Makes use of distant influences on propensity
•Uses 17 residue window
•Adds propensity for four 2º structure states (H, E, T, C)
•Highest value defines 2º structure state of central residue in window
.
10
.
20
.
30
.
40
.
50
.
60
SRRSASHPTYSEMIAAAIRAEKSRGGSSRQSIQKYIKSHYKVGHNADLQIKLSIRRLLAA
helix
HHHHHHHHHHH
HHHHHH
sheet
EEEEEEEE E
turns
coil
TTTT
C
TTTTT
CCCCC
.
70
.
80
.
C
90
GVLKQTKGVGASGSFRLAKSDKAKRSPGKK
helix HHHH
sheet
HHHHHHHHHHH
EEEEE
E
turns
coil
TTT
CCCC
Residue totals: H: 36
C
E: 21
EEEEEE
T TTTT
CCC
C
T: 17
C: 16
percent: H: 48.6 E: 28.4 T: 23.0 C: 21.6
HHHH
Expansion using larger crustal structure databases
Algorithms based on a larger database of crystal structure information:
•GOR II, III and IV
•SOPM
http://npsa-pbil.ibcp.fr/cgi-bin/npsa_automat.pl?page=/NPSA/npsa_server.html
SRRSASHPTYSEMIAAAIRAEKSRGGSSRQSIQKYIKSHYKVGHNADLQIKLSIRRLLAAGVLKQTKGVG
cccccccchhhhhhhhhhhhtccttcccchhhhhhhhhtcccccccthhhhhhhhhhhhhhhhhttttcc
ASGSFRLAKSDKAKRSPGKK
cccceeeecccccccccccc
Homology based methods
Neural Network programs
•
•
•
A neural net has an input layer, hidden layers composed of nodes
given different weights, and an output layer
Neural net trained with multiply aligned sequences
Accuracy >75%
PHD
1. BLASTP
2. MAXHOM (sequence alignment)
3. Neural Net
Layer one : 13 residue window
Layer two: 17 residue window
Layer three: “Jury layer” – removes very short stretches
PSIPRED
1. PSI-BLAST
2. Neural net
SSpro
PROTER
PROF
HMMSTR
Predictions with Multiple Methods
No single prediction program is correct, and it is generally good
practice to use the output from several programs
Some web servers do this:
JPred
•PHD, PREDATOR, DSC, NNSSP, Inet and ZPred
•First submitted to PSI-BLAST
•Multiple alignment
•Submitted to above 6 programs
•Consensus returned
•No consensus, uses PHD
SRRSASHPTYSEMIAAAIRAEKSRGGSSRQSIQKYIKSHYKVGHNADLQIKLSIRRLLAAGVLKQTKGVGASGSFRLAKSDKAKRSPGKK
---------HHHHHHHHHHH--------HHHHHHHHHH-------HHHHHHHHHHHHH---EEEEE------EEEE--------------
How accurate?
Trans-membrane proteins
Two types of trans-membrane proteins
•-helix
•-barrel
•Many consists solely of -helix and are
found in the cytoplasmic membrane
•-barrel normally found in outermembrane of gram negative bacteria
•Difficult to get X-ray or NMR structure
•-helix perpendicular to membrane 17-25 residues
•Hydrophobic residues separated by hydrophilic loops (<60 residues)
•Residues bordering hydrophobic module is generally charged
•Inner cytosolic region most often highly charged (orientation info)
•Positive inside rule
•Scan window 17-25 residues calculate hydrophobicity score
•Many false positives
•Signal peptide sequences confuse algorithm
TMHMM
•Trained with 160 known TM sequences
•Probability of having an -helix is given
•Orientation of -helix based on positive inside rule
Phobius
•Incorporates distinct HMM models for signal peptides and TM helices
•Signal peptide sequence ignored
•Can use sequence homologs and multiply aligned sequences
Prediction of -barrel proteins
•-strand forming trans-membrane section is amphipatic
•10-22 residues
•Alternating hydrophobic and hydrophilic sequence arrangement
•-helix TM prediction programs thus not applicable to -barrel proteins
TBBpred
•Neural net trained with -barrel protein sequences
Coiled coil prediction
Two or more -helices winding around each other
For every 7 residues, 1 and 4 are hydrophobic, facing central core
Coils
•Scan window of 14, 21 or 28 residues
•Compares residues to probability matrix based on known coiled coils
•Accurate for left-handed coil, but not right-handed coil
Multicoil
•Scoring matrix based on 2-strand and 3-strand coils
•Used in several genome-wide studies
Leucine zippers
•sub-class of coiled coils
•L-X6-L-X6-L•Found in transcription factors
•Anti-parallel -helices stabilized by leucine core
Chapter 13
Protein Tertiary Structure Prediction
The need for predicting 3D structures
•
•
•
•
X-ray crystallography is extremely tedious
DNA sequences and therefore protein sequences are rapidly
generated
A gap between sequence and structure is widening
Protein structure often provides insight info function
Thee main methods for 3D prediction
1. Homology modeling
2. Threading
3. Ab initio
Homology Modeling
Template Selection
•Search PDB for homologous sequences with BLAST or FASTA
•Should have >30% sequence identity (20% at a stretch)
•In case of multiple hits, choose
•Highest identity
•Highest resolution
•Most appropriate co-factors
Sequence Alignment
Critical
Incorrectly aligned residues will give an incorrect model
Use Praline or T-Coffee for alignment
Inspect visually to confirm alignment of key residues
Backbone Model Building
•Copy the backbone atoms of the query sequence to that of the
corresponding aligned residue
•If the residues are identical, the coordinates of the whole residue can
be copied
•If the residues are different, only the C are copied
•The remaining atoms of the residue are modeled later
Loop Modeling
It often happens that there are “gaps” in the aligned sequences
Two techniques to connect the protein on either side of the gap:
Database
•Search database for fragments that fit the gap
•Measure coordinates and orientation of backbone on either side of gap
•Search for fragments that can fit
•Best loop gives no steric clash with structure
Ab Initio
•Generate random loop No clash with nearby side-chains
• And  angles in acceptable region of Ramachandran plot
Side Chain Refinement
•Need to model side-chains where these differ from aligned template
sequence
•Search database for all occurrences of given side-chain in backbone
conformation and minimal clash with neighbouring residues
•Computationally prohibitive
•Library of rotamers
•Collection of conformations for each residue that is most often
observed in structure database
•Select rotamer with conformation that best fits backbone
•Minimal interference with neighbouring side-chains
•SCWRL
Model Refinement using Energy Function
•After loop modeling and side-chain refinement the follwing remain
•Unfavourable torsion angles
•Unacceptable proximity of atoms
•Use energy minimization to alleviate such problems
•Limit number of iteration (<100) to ensure that the entire model does
not change form the template
•Molecular Dynamic can be used to search for a global minimum
Model Evaluation
•Check consistency in - angles
•Bond lengths
•Close contacts
•Flag regions below acceptability threshold
•Procheck
•WHATIF
•ANOLEA
•Verify3D
Comprehensive Modeling Programs
•Modeler
•Swiss-Model
•3D-Jigsaw
Threading and Fold Recognition
Pairwise Energy Method
•Fit sequence to each fold in database
•Use local alignment to improve fit
•Calculate energies
•Pairwise residue interaction
•Solvation Hydrophobic
Profile Method
•Fit sequence to fold
•Calculate propensity of each amino acid to be present at each
profile position
•Secondary structure types
•Solvent exposure
•Hydrophobicity
•Use structure fold that best fits profile of parameters
Ab Initio Prediction
Protein fold into a native, low-energy native state
The mechanism driving this process is poorly understood
Computationally untenable to explore all possible states and calculate
energies
A 40 residue peptide will require 1020 years to calculate all states using
a 1×1012 FLOPS computer
Not realistic approach currently