Transcript Slide 1

Combining Predictors for Short
and Long Protein Disorder
Zoran Obradovic, Slobodan Vucetic and Kang Peng
Information Science and Technology Center, Temple University, PA 19122
A. Keith Dunker and Predrag Radivojac
Center for Computational Biology and Bioinformatics, Indiana University, IN 46202
NIH grant R01 LM007688-01A1 to A.K. Dunker and Z. Obradovic is gratefully acknowledged
Introduction
Protein Structure - under physiological condition, the amino acid
sequence of a protein folds spontaneously into specific (native) three
dimensional (3-D) structure or conformation
hydrogen
bond
-strand
4 levels of
protein structure
hydrogen
bond
Importance of Protein Structure
The “central dogma” – amino acid sequence determine protein
structure, and protein structure determine its biological function
Amino Acid Sequence
> 1NLG:_ NADP-LINKED GLYCERALDEHYDE-3-PHOSPHATE
EKKIRVAINGFGRIGRNFLRCWHGRQNTLLDVVAINDSGGVKQASHLLKYDSTLGTFAAD
VKIVDDSHISVDGKQIKIVSSRDPLQLPWKEMNIDLVIEGTGVFIDKVGAGKHIQAGASK
VLITAPAKDKDIPTFVVGVNEGDYKHEYPIISNASCTTNCLAPFVKVLEQKFGIVKGTMT
TTHSYTGDQRLLDASHRDLRRARAAALNIVPTTTGAAKAVSLVLPSLKGKLNGIALRVPT
PTVSVVDLVVQVEKKTFAEEVNAAFREAANGPMKGVLHVEDAPLVSIDFKCTDQSTSIDA
SLTMVMGDDMVKVVAWYDNEWGYSQRVVDLAEVTAKKWVA
3-D Structure
Biological Function
Function: Gene Transfer
Thus, it is important to know a protein’s structure to understand its
function and other biological properties
Protein Structure Prediction

The sequence-structure gap




Current experimental structure determination techniques, e.g. X-ray diffraction
and NMR spectroscopy, are still slow, expensive and have their limitations
As a result, there are less than 30,000 experimental protein structures, compared
to more than 1.6 million known protein sequences
Protein structure prediction – predicting protein structures from
amino acid sequences using computational methods
Aspects of protein structure prediction



1D – secondary structures, solvent accessibility, transmembrane helices, signal
peptides/cleavage sites, coiled coils, disordered regions
2D – inter-residue contacts, inter-strand contacts
3D – individual atom coordinates in the tertiary structure (the ultimate goal)
The CASP Experiments

Critical Assessment of Techniques for Protein Structure Prediction

The primary goal


To obtain an in-depth and objective assessment of current methods for predicting protein
structure from amino acid sequence
The procedure



Proteins with “soon to be solved” structures are selected as prediction targets, and their amino
acid sequences are made available
Prediction teams submit their prediction models before the experimental structures are released
Prediction models are compared to experimental structures for detailed evaluation by
independent assessors
# of targets
# of participating groups
# of submitted models
CASP6 (2004)
76
208
41283
CASP5 (2002)
67
215
28728
CASP4 (2000)
43
163
11136
CASP3 (1998)
43
98
3807
CASP2 (1996)
42
72
947
CASP1 (1994)
33
35
135
CASP Website: http://predictioncenter.llnl.gov/
Prediction Categories in CASP6

Tertiary structure (3-D coordinates for individual atoms) prediction







Comparative/Homology modeling
Fold recognition
New fold modeling
Disordered region prediction (since CASP5)
Domain boundary prediction (new)
Residue-residue contact prediction (new)
Secondary structure prediction was excluded in CASP6
In CASP6 there were 20 groups participated in Disordered
Region prediction, while only 6 groups in CASP5
Disordered Region (DR)

Part of a protein or a whole protein that does NOT have stable 3D
structure in its native state




Perform important biological functions
Have distinct sequence properties
Evolve faster than ordered regions
Common in nature
Other definitions of disordered region



Kissinger et al, 1995

Missing coordinates (used by CASP)
High B-factors
Random coils
NOn-Regular Secondary Structure (NORS)
Prediction of Disordered Regions
K Q L L W C Y L A A M A H Q F G A G K L K C T S A T T W Q G
Input Window
of length Win
Class label 0/1:
disordered / ordered
Amino Acid
Sequence
Attributes derived from the local
window
• 20 AA frequencies
• K2-entropy (sequence complexity)
• Flexibility
• Hydropathy
• more …
One example for each sequence position (residue)
Long DR Predictors on Short DR

Disordered regions can be divided into 2 groups according to their
lengths



Our previous disorder predictors were specific to long DRs



Predictors – VL-XT, VL2, VL3, VL3H, VL3P, VL3B
Accuracies – 70% (VL-XT) ~ 85% (VL3P)
They were less successful on short DRs, as shown in CASP5



short DRs – 30 consecutive residues or shorter
long DRs – longer than 30 consecutive residues
25~66% per-residue accuracy on short DRs
75~95% per-residue accuracy on long DRs
Possible reasons



The window lengths for attribute construction and post-filtering were optimized
for long DRs
Training data did NOT include any short DRs
Short DRs are different from long DRs in terms of amino acid compositions,
flexibility index, hydropathy and net charge
Amino Acid Compositions of Short DRs
Amino acid frequency
from Globular-3D
difference
Dataset-Globular
3D
0.05
Rigid Order
Flexible Order
Short Disorder
Long Disorder
0.03
0.01
-0.01
-0.03
Radivojac et al., Protein Science, 2004
-0.05
W
C
F
I
Y
V
L
H
M
A
T
R
G
Q
S
N
P
D
E
K
Residues
Consequence – a predictor specialized for short disordered
regions is necessary
Our Approach in CASP6
Idea – two specialized predictors for long and short disordered
regions, and a meta predictor to estimate which specialized
predictor is more suitable for current input
Meta
Predictor
wL
Input
Long Disorder
Predictor (>30aa)
wS
OL

Short Disorder
Predictor (30aa)
Final
Prediction
OS
In CASP5, we used only Long Disorder Predictor component
The Training Dataset
Dataset
Number of
Chains
Number of
long DRs
Number of short
DRs
LONGa
153
163
24
SHORTc
511
43
630
ORDERa,b
290
0
0
XRAYd
381
24
329
TOTALe
1335
230
983
a) LONG and ORDER – training data for VL3 predictors (Z. Obradovic, K. Peng, S. Vucetic, P. Radivojac, C. J. Brown, A. K.
Dunker, Proteins, 53 (S6): 566-572, 2003; K. Peng, S. Vucetic, P. Radivojac, C. J. Brown, A. K. Dunker, Z. Obradovic,
Journal of Bioinformatics and Computational Biology, in press)
b) ORDER – training data for a B-factor predictor and used in a study of flexibility index (P. Radivojac, Z. Obradovic, D. K.
Smith, G. Zhu, S. Vucetic, C. J. Brown, J. D. Lawson, A. K. Dunker, Protein Science, 13 (1):71-80, 2004; D. K. Smith, P.
Radivojac, Z. Obradovic, A. K. Dunker, G. Zhu, Protein Science, 12 (5):1060-1072, 2003)
c) SHORT – training data for a short disorder predictor (Radivojac et al., Protein Science, 13 (1):71-80, 2004)
d) XRAY – a non-redundant set of PDB chains released between June 2003 and May 2004
e) TOTAL - the merged sequences are non-redundant with less than 50% identity
Specialized Disorder Predictors
Optimized for long and short disordered regions, respectively
Window Length
Predictor
Attributes
Wina
Woutb
Accuracyc (%)
short DR long DR
order
Long Disorder
(>30aa)
• Amino acid frequencies
• K2-Entropy
• Flexibility index
• Hydropathy/net charge ratio
41
31
50.13.6
76.54.2
85.10.9
Short Disorder
(30aa)
(In addition to the attributes above)
• PSI-BLAST profile
• Secondary structure prediction
(PSIPred)
• An indicator of terminal regions
15
5
81.52.1
66.73.5
82.40.5
a) Length of input window for attribute construction
b) Length of output window for post-filtering
c) Out-of-sample per-chain accuracies were estimated by 1) randomly split the 1335 sequences into 75%:25%, 2) the first part
for training and the second for testing, 3) repeat steps 1 and 2 for 30 times and average the accuracies
The Prediction Process

For each sequence position (residue)



The three predictors construct attributes and output OL, OS and OG
The final output is calculated as O = OL * OG + OS * (1 – OG)
If O > 0.5, predict disorder
Otherwise, predict order
Meta Predictor
OG
Input
Long Disorder
Predictor (>30aa)
1-OG
OL

OS
Short Disorder
Predictor (30aa)
The final output
O = OL* OG + OS * (1 - OG)
Training the Meta Predictor


The meta predictor was then trained as a 2-class classifier (short
disorder vs. long disorder)
Constructing labeled dataset for training of meta predictor




Used same attributes as for the short disorder predictor
Residues from long DRs and their flanking regions were labeled as class 1
Residues from short DRs (3aa) and their flanking regions were labeled as class 0
The remaining residues were discarded (u)
Example:
Ordered
Region
A Short Disordered
Region (8aa)
Ordered
Region
Sequence:
GKKGAVAEDGDELRTEPEAKKSKTAAKKNDKEAAGEGPALYEDPPDHKTS
Disorder labels:
ooooooooooooooooooooDDDDDDDDoooooooooooooooooooooo
Class labels:
uuuuuuuuuuuuuu00000000000000000000uuuuuuuuuuuuuuuu
Input Window
(Length Win)
Current
Residue
The input window (of length Win =61) centered at current residue must overlap with more than half of a
disordered region
CASP6 Targets

63 targets with 3-D coordinates information available, with 90
disordered regions and 90 ordered regions
Length range
Disordered
regions
Ordered regions
Number of regions
Number of residues
1-3
35
58
4-15
41
304
16-30
9
201
31-100
4
266
>100
1
102
Total
90
931
90
12,520
Prediction Accuracy
VL2
VL3E
NEW
NEW/short
NEW/long
100
100
80
80
60
Accuracy (%)
Accuracy (%)
VL2
VL3E
NEW
NEW/short
NEW/long
40
20
60
40
20
0
0
1-3
4-15
16-30
31-100
>100
order
1-3
4-15
16-30
31-100
>100
Length range
Length range
(a) per-region accuracy
(b) per-residue accuracy
order
• VL2 (CASP6 model-3) – a previously developed long disorder predictor (S. Vucetic, C.J. Brown, A.K. Dunker and Z. Obradovic, Proteins: Structure, Function and
Genetics, 52:573-584, 2003)
• VL3E(CASP6 model-2) – a previously developed long disorder predictor (Z. Obradovic, K. Peng, S. Vucetic, P. Radivojac, C. J. Brown, A. K. Dunker, Proteins, 53
(S6): 566-572, 2003; K. Peng, S. Vucetic, P. Radivojac, C. J. Brown, A. K. Dunker, Z. Obradovic, Journal of Bioinformatics and Computational Biology, in press )
• NEW (CASP6 model-1) – the combined predictor
• NEW/short – the specialized predictor for short disordered regions (30aa)
• NEW/long – the specialized predictor for long disordered regions (>30aa)
Prediction on Long Disordered Regions
T0206 (1-78)
T0206 (1-78)
1
0.9
long (OL)
0.8
meta (OG)
short (OS)
0.7
0.6
0.6
0.4
0.5
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
1
20
40
60
80
100 120 140 160 180 200 220
residue
(a) Prediction by component predictors
VL3E
VL2
0.8
0.7
0.5
NEW
0.9
prediction
prediction
1
0
1
20
40
60
80
100 120 140 160 180 200 220
residue
(b) Comparison to previous predictors
Notes: (1) red segments indicate disordered regions (of missing coordinates), (2) The threshold for predicting
disorder is 0.5
Prediction on Short Disordered Regions
Notes: (1) red segments indicate disordered regions (of missing coordinates), (2) The threshold for predicting
disorder is 0.5
In both targets, all short DRs were identified, but with considerable amount of false positives.
More detailed analysis shows that the new predictor tend to over-predict at N- and C- termini
Correlation with High B-factor Regions
T0203 (1-4, 105-111, 377-382)
T0233 (1-13, 81-92, 106-108, 137-138)
1
1
0
50
100
150
200
residue
250
300
350
0
0.5
50
0
50
100
150
200
250
300
350
residue
Notes: (1) red segments indicate disordered regions (of missing coordinates), (2) The threshold for predicting
disorder is 0.5, (3) no B-factor data for disordered regions
B-factor
0.5
disorder predictions
B-factor
disorder predictions
50
Conclusion by CASP6 Assessor
“Group 193 is best on all measures, on both no-density segments
and B-factors, and is significantly better than next 3 groups, 096,
003, 347 on no-density segments, who are about the same as each
other. Groups 3, 347, and 472 are good at B-factors”
Group IDs:

193 ISTZORAN (Zoran Obradovic, Temple University)
096 CaspIta (Tosatto et al., Univ. of Padova)
003 Jones UCL (David Jones, University College London)
347 DRIP PRED (server from Bob MacCallum, Stockholm)

472 Softberry (good at B-factor correlation)



Assessor’s report is available at CASP6 website:
http://predictioncenter.llnl.gov/casp6/meeting/presentations/DR_assessment_RD.pdf
Future Directions

The length threshold 30 for dividing DRs into long and short is
artificial and may not be the best choice


The new predictor produced considerable amount of false positives,
especially at the N- and C- terminals.


A better method for partitioning the DRs into more homogenous length groups
(maybe more than 2)
Build predictors specific to terminal and internal regions, and combine them (a
similar approach to VL-XT)
The dataset contains noises, i.e. mislabeling, since not all missing
coordinate regions may not necessarily be due to disorder
The End
Thank You!!