Transcript PowerPoint Presentation - Secondary structure prediction

Protein structure
Predictive methods
Topics Covered
• Secondary structure prediction methods
• 3D fold prediction
– Ab initio protein structure prediction
– Homology-based methods of fold recognition
– Comparative model construction (aka homology model
construction)
• Community evaluation of protein structure prediction
– Critical Assessment of protein Fold Prediction (CASP)
http://predictioncenter.org/
– EVA (real-time continuous evaluation of protein fold prediction
methods) http://cubic.bioc.columbia.edu/eva/
– Astral datasets
• Structural Genomics Initiative
Why Protein Structure Prediction?
Y 2005
Sequences
2,300,000
Structures
29,000
We know the experimental 3D structure for
~1% of the protein sequences
Andras Fiser, Albert Einstein College of Medicine
Principles of Protein Structure
Anacystis nidulans
Anabaena 7120
Ab initio prediction
Condrus crispus
folding
Desulfovibrio vulgaris
GFCHIKAYTRLIMVG…
evolution
Fold Recognition
Comparative Modeling
Andras Fiser, Albert Einstein College of Medicine
Protein structure modeling
Ab initio prediction
Comparative Modeling
Applicable to any sequence
Applicable to those sequences only that
share recognizable similarity to a template
structure
Not very accurate (>4 Ang RMSD),
Fairly accurate ( <3 Ang RMSD), typically
comparable to a low resolution X-ray
experiment.
Attempted for proteins of <100 residues
Not limited by size
Accuracy and applicability are limited
by our understanding of the protein
folding problem
Accuracy and applicability are rather
limited by the number of known folds
Andras Fiser, Albert Einstein College of Medicine
Structural Genomics
Definition: The aim of structural genomics is to put every protein
sequence within a “modeling distance” of a known protein
structure.
Size of the problem:
There are a few thousand domain fold families.
There are ~20,000 sequence families (30% sequence id).
Solution:
Determine protein structures for as many different families as
possible.
Model the rest of the family members using comparative modeling
Andras Fiser, Albert Einstein College of Medicine
Structural Genomics
Characterize most protein sequences (red) based on related
known structures (green).
The number of “families” is
much smaller than the number
of proteins
Andras Fiser, Albert Einstein College of Medicine
The utility of a
comparative
model depends
on its accuracy
Accuracy is closely
linked to sequence
similarity
David Baker and Andrej Sali, Protein Structure
Prediction and Structural Genomics, Science 2001
Comparative Protein Structure Modeling
Ca RMSD Å (% EQV)
2 (50)
1 (80)
0 (100)
Flavodoxin
family
Anacystis nidulans
Anabaena 7120
Condrus crispus
COMPARATIVE
MODELING
KIGIFFSTSTGNTTEVA…
Desulfovibrio vulgaris
Clostridium mp.
20
50
100
% SEQUENCE IDENTITY
Andras Fiser, Albert Einstein College of Medicine
Steps in Comparative Protein Structure Modeling
START
TARGET
Template Search
ASILPKRLFGNCEQTSDEGLK
IERTPLVPHISAQNVCLKIDD
VPERLIPERASFQWMNDK
Target – Template
Alignment
TEMPLATE
ASILPKRLFGNCEQTSDEGLKIERTPLVPHISAQNVCLKIDDVPERLIPE
MSVIPKRLYGNCEQTSEEAIRIEDSPIV---TADLVCLKIDEIPERLVGE
Model Building
Model Evaluation
No
OK?
Yes
END
Andras Fiser, Albert Einstein College of Medicine
Typical Errors in Comparative Models
Incorrect template
Misalignment
MODEL
X-RAY
TEMPLATE
Region without a
template
Distortion in correctly
aligned regions
Side chain packing
Andras Fiser, Albert Einstein College of Medicine
Template identification
START
Template Search
Target – Template
Alignment
•
•
•
Fast but less sensitive: e.g. BLAST
Better: Intermediate sequence search
Even better: Profile/HMM and iterative search
methods (e.g. PSI-BLAST)
– Searching against libraries of HMMs and profiles for
solved structures
•
Profile-profile alignment (e.g., Hhalign, PHYRE)
– Including 2ary structure prediction
Model Building
Model Evaluation
No
OK?
Yes
END
•
Structure-based threading
Target-template alignment
START
Template Search
Target – Template
Alignment
Model Building
Model Evaluation
No
OK?
Yes
END
• Note that the methods for identifying
candidate templates normally produce an
alignment
– but these alignments are unlikely to be
optimal
• The alignment method used must be
tuned to the level of evolutionary
divergence between the target and
template
• Manual refinement/editing of the
alignment is often used to improve the
comparative model
Constructing a comparative model
START
Template Search
Target – Template
Alignment
•
•
•
•
Rigid Body Assembly (COMPOSER)
Segment Matching (SEGMOD, 3DPSSM)
Satisfaction of Spatial Restraints (MODELLER)
Integrated (NEST)
Model Building
Model Evaluation
No
loop modeling, side chain modeling
OK?
Yes
END
Andras Fiser, Albert Einstein College of Medicine
Comparative model evaluation
START
Template Search
Target – Template
Alignment
• Stereochemistry (PROCHECK,
WHATCHECK)
• Environment (Profiles3D, Verify3d)
• Statistical potentials based methods
(PROSAII)
Model Building
Model Evaluation
No
OK?
Yes
Is the model reliable?
A model is reliable when it is based on a
correct template and on an approximately
correct alignment.
END
Andras Fiser, Albert Einstein College of Medicine
Primary,
Secondary,
Tertiary and
Quaternary
Structure
Hierarchical descriptions of proteins
(follows the folding process)
•
Primary structure: the amino acid sequence
•
Secondary structure: “regular local structure of linear segments of polypeptide chains”
(Creighton)
–
Helix (~35% of residues): subtypes: ,
–
–
Beta sheet (~25% of residues)
Both types predicted by Linus Pauling (Corey and Pauling, 1953;
 helix first described by Pauling in 1951)
Other less common structures:
–
•
•
•
–
•
Beta turns
3/10 helices
Ω loops
Remaining unclassifiable regions sometimes termed “random coil” or “unstructured regions”
Tertiary structure: “Overall topology of the folded polypeptide chain” (Creighton)
–
•
 and 310
Mediated by hydrophobic interactions between distant parts of protein
Quaternary structure: “Aggregation of the separate polypeptide chains of a protein”
(Creighton)
Baxevanis & Ouellette (Ch. 9, p.224, Wishart)
Information required for folding is (mostly)
contained in the primary sequence
• Early on, proteins were shown to fold into their native
structures in isolation
• This led to the belief that structure is determined by
sequence alone (Anfinsen, 1973)
• Over the last decade, a significant number of proteins
have been shown to not fold properly in the test tube
(e.g., requiring the assistance of chaperonins)
• Nevertheless, the native 3D structure is assumed to
be in some energetic minimum
• This led to the development of ab initio folding
methods
Baxevanis & Ouellette (Ch. 9, Wishart)
Folding pathways
• Evidence that local structure segments form first, and
then pack against each other to form 3D fold
– Exploited in protein fold prediction, Rosetta method
• Simons, Bonneau, Ruczinski & Baker (1999). Ab initio Protein
Structure Prediction of CASP III Targets Using ROSETTA. Proteins
• Semi-stable structural intermediates on folding pathway
to lowest-energy conformation
– Prof. Susan Marqusee, Berkeley
Baxevanis & Ouellette (Ch. 9, Wishart)
Secondary Structure Prediction
Why is secondary structure
prediction important?
• Secondary structure diverges less rapidly
than primary sequence
– Knowledge or prediction of 2ary structure
improves detection and alignment of remote
homologs
• 3d-pssm, PHYRE, SAM T02 (fold prediction servers)
Baxevanis & Ouellette (Ch. 9, Wishart)
Basic types of secondary structure
• Helices ( and others)
–  is most common; 3.6 residues/turn
– Side chains project outward
– Structure is stabilized between hydrogen bonds between the
carbonyl (CO) group of one amino acid and the amino (NH)
group of the amino acid that is 4 positions C-terminal to it
• -Strands (two or more strands interact to form a sheet)
• Other (sometimes called loop, coil, or non-regular)
• Most secondary structure prediction methods classify
residues to one of three states
Baxevanis & Ouellette (Ch. 9, Wishart)
Focusing on single residues
• Early structure prediction methods focused on the
structural characteristics of individual residues
• This enabled the larger problem to be decomposed
into smaller easier-to-solve problems (enabling the
combination of solutions to sub-problems to form a
global solution)
• This also enabled methods to focus on detecting
transmembrane regions, solvent-accessible residues,
and other important features of molecules
Baxevanis & Ouellette (Ch. 9, Wishart)
Secondary structure prediction
accuracy is boosted by using homologs
• Labeling residues in a sequence as -helix, -sheet or
turn/coil (3-state prediction).
• Accuracy of prediction enhanced by ~6% when multiple
sequence alignments are used vs the use of a single
sequence (Cuff & Barton, 1999)
• Best methods for 2ary structure prediction -- PSIPRED
(Jones 1999) and JNET (Cuff & Barton, unpublished)
– Make use of homologs obtained using PSI-BLAST
– Have ~>76% accuracy for 3-state prediction
– Provide confidence values for each position
Baxevanis & Ouellette (Ch. 12, Barton)
Amino acid patterns indicative of
-strand structures
• Short runs of conserved hydrophobic
– Buried -strand
• An i, i+2, i+4 pattern of conserved
hydrophobic residues suggests a surface strand.
• Conserved residues sharing the same
physicochemical properties are likely to form
one face of a strand.
Baxevanis & Ouellette (Ch. 12, Barton)
Amino acid patterns indicative of
-helical structures
• Conservation patterns of i, i+3, i+4, i+7 and variations
(e.g., i, i+4, i+7) suggests an alpha helix
• Amphiphilic/amphipathic conservation patterns
(alternating hydrophobic and polar residues) following
an i, i+3, i+4, i+7 pattern (and variations, e.g., i, i+4,
i+7) are likely to represent surface helices
Baxevanis & Ouellette (Ch. 12, Barton)
Identifying loop regions
• Insertions and deletions are not well tolerated in
the hydrophobic core.
– Regions of an MSA that include many gap characters
are likely to indicate surface loops.
• Glycine and proline residues can be found in
any secondary structure.
– However, conserved glycine/proline residues are
strongly suggestive of loops.
Baxevanis & Ouellette (Ch. 12, Barton)
Amino acid preferences for different secondary structures
(and identifying loops/turns)
http://www.chembio.uoguelph.ca/educmat/phy456/456lec01.htm
Early schemes used observed preferences
• Various schemes give the amino acids numerical weights or
rankings for their preferences, and several computer programs
can predict the secondary structure from the given sequence.
• Preferences are weak, but provide some signal
• The simplest such scheme of Chou and Fasman, Ann. Rev
Biochem. (1978), examined the statistical distribution of amino
acids in alpha helix, beta sheet and turns or loops, using a set of
known protein structures from the protein databank.
• A novel sequence can then be scanned, and the tendency of
each portion of the sequence to form secondary structure is
assessed.
http://www.chembio.uoguelph.ca/educmat/phy456/456lec01.htm
Improving secondary structure prediction
•
Peer pressure (pressure from the neighbors): A minimum of 4
amino acids out of 6 should show alpha preference, or 3 out of 5 beta
preference, or clusters of 2-3 breakers in a sequence of 4 are needed
to set the secondary structure in any region, and individual misfits
adopt the secondary structure of their neighbours.
•
Learning secondary structure preferences from expanded data
sets: More recent prediction schemes take advantage of larger data
sets to examine amino acid preference for different regions in a helix or
different positions in a tight turn.
•
Up-weighting conserved residues: In addition, sequences of
homologous proteins may be compared. The rationale is that highly
conserved amino acids contribute more to the three dimensional
structure than unconserved, and different weightings can be introduced
to the statistical analysis.
•
Improved accuracy: The accuracy of prediction has risen from about
55% using the simple Chou-Fasman method, where the tendency is to
overpredict, to almost 80% using current methods.
http://www.chembio.uoguelph.ca/educmat/phy456/456lec01.htm
Amino acid propensities for
different structural environments
• Propensities are weak but contribute to prediction
accuracy
– E.g., Glu (E) occurs in alpha helices only 59% more frequently
than random
• Helical propensities
– Partial charge of helix dipole favors
• Acidic Asp (D) and Glu (E) residues at N-terminus of helices
• Basic Lys (K), Arg (R ) and His (H) residues at C-terminus
– Pro (P) residues are more common at the N-terminal first turn of
helix
– Asn (N), Asp (D), Ser (S) and Thr (T) residues often occur at first
turn of helix (side chain hydrogen bonding to backbone of third
residue)
Creighton, Proteins
The new generation of secondary
structure prediction
• Based on machine learning concepts
– Training set: learn implicit rules, principles and
model parameters from labelled data (sequences
whose secondary structures are known for each
position)
– Test set: sequences of unknown structure
– Used machine learning method called artificial
neural networks (designed to simulate biological
neural networks in the brain)
– PHDsec (Rost et al 1994, Rost et al 1996)
Baxevanis & Ouellette Ch 8 (Ofran and Rost)
Neural Network for Protein Structure
Prediction
Key to success in machine learning
algorithms
• “The success of machine learning algorithms
depends on the careful choice of the biologically
based features used for training… and a sufficiently
large and accurate training set”
• To enhance prediction accuracy on novel data,
training data diversity is also critical
• Exploit knowledge that local environment is
important: to predict 2ary structure of residue ‘i’,
consider all residues in a window around i: i-n, … i,
… i+n.
Baxevanis & Ouellette Ch 8 (Ofran and Rost)
PHDsec
• Employs homology detection and a feed-forward
artificial neural network
• Step 1: homolog search and MSA construction
• Step 2: label each position with conservation signal
(across MSA) and observed substitutions
• Step 3: submit representative annotated “sequence”
to a system of neural networks.
• Output is a prediction of the most likely secondary
structure at each position, with the estimated
confidence in that prediction
Baxevanis & Ouellette Ch 8 (Ofran and Rost)
Assessing performance evaluations
• “Overall, the correct evaluation of performance for
prediction methods is an art in itself; only a handful of
methods turned out over time to not have been
overestimated by their developers.”
– Evaluation must be performed on a standard dataset
– Training and test data should be rigorously kept separate
– Standard deviations of estimates should be provided
Baxevanis & Ouellette Ch 8 (Ofran and Rost)
Other problems with comparing
different methods
• Performance reported in literature can take different forms
–
–
–
–
–
Accuracy and coverage
Positive (or negative) predictive power
Sensitivity and specificity
Machine learning terms (e.g., Matthews coefficients)
Wilcoxon paired score signed rank tests
• Or might be based on different criteria for success
– per residue
– per secondary structure element
– per protein
• Others measure performance only in cases where a prediction has
high confidence (with a likelihood of a lower FP rate)
Baxevanis & Ouellette Ch 8 (Ofran and Rost)
How do the methods compare?
• Best methods now reach 76% accuracy at 3-state
prediction (helix, strand, random coil)
– Rost 2001
– See EVA website for detailed comparisons
• Metaservers:
– Consensus approaches combining weighted predictions
from different servers
– These almost always outperform individual methods
– Shown in both CASP and EVA
Baxevanis & Ouellette Ch 8 (Ofran and Rost)
Caveats
• Even when an experimental structure is available, it
is sometimes unclear where one secondary structure
element ends and another begins
• Low-confidence predictions (and regions of
disagreement across servers) can correspond to
structurally ambiguous regions
• Real-life example: Prion protein (involved in bovine
spongiform encephalopathy, Creutzfeld-Jakob
disease, etc).
– Region assumed to be responsible for aggregation believed
to flip from experimentally determined helical structure to
(predicted) strand in diseased individuals
– All the best secondary structure prediction methods predict
this region to be beta (“incorrect”)
Baxevanis & Ouellette Ch 8 (Ofran and Rost)
Secondary structure prediction
programs
• PSI-PRED (David Jones; makes use of distant
homologs detected using PSI-BLAST - most popular)
• JNET (Cuff & Barton)
• PHD (Rost & Sander)
Baxevanis & Ouellette Ch 8 (Ofran and Rost)
PSIPRED
Consensus and jury approaches
produce best results
• Primary conclusion of CASP experiments is that
structure prediction meta-servers (which combine
results from several independent prediction methods)
have the highest accuracy
• This kind of consensus approach can be applied to
both the template selection and the pairwise
alignment between the target and template
• We have also shown in class that a consensus
approach can be applied to predicted structures for
numerous homologs in a family related to the target
3D-structure prediction
Basic premise: The function and structure of
a protein are encoded in its primary sequence
The amino acid sequence determines
• a protein’s 3D structure,
• subcellular localization,
• intermolecular interactions,
• biochemical physiological tasks, and
• (eventually) how and when it will be broken down into
its component building blocks.
– Paraphrased from class text (Ofran and Rost), p 198
How many unique protein folds are there?
• Many structural biologists believe that all protein domains will
eventually be classified into only 1,000 different fold classes
(Koonin et al 2002)
• Number of unique SCOP folds already close to 1,000
• Structural Genomics Initiative is designed to populate that fold
space
– However:
even with attempts to solve novel structures,
many new structures are clearly members of existing structural
classes
Baxevanis & Ouellette (Ch. 9, Wishart)
3D structure classification schemes
• Structure classification databases:
– SCOP (http://scop.berkeley.edu/
– CATH (http://www.cathdb.info/)
• Three main classes for folds
– All alpha (>50% helix; <10% beta sheet)
– All beta (>30% beta sheet; <5% helix)
– Mixed or alpha/beta (everything else)
Baxevanis & Ouellette (Ch. 9, Wishart)
3D structure prediction
• Decompose into two subtasks
– Fold recognition “Protein X is related by evolution to structure Y”
• Assumed evolutionary relationship is used to infer a similarity in 3D
fold (but no comparative model construction)
• Can be achieved by pairwise sequence comparison, scoring a
sequence against a library of profiles or HMMs, and by other
methods
• Newer “threading” methods can enable correct fold recognition in
the Twilight Zone
– Comparative model construction
• May be restricted to higher sequence identity (e.g., above 30%) due
to the likelihood of serious alignment error below this range.
– Some servers do both
• 3d-pssm/PHYRE, Superfamily, etc.
Baxevanis & Ouellette Ch 8 (Ofran and Rost)
Threading
• Limited to generating approximate models or
suggesting approximate folds
– >5 Angstroms for 3D threading
– >3Angstroms for 2D threading
• Name based on “threading” a tube (called a snake)
through a plumbing system.
• Each unique threading of a sequence through the 3D
model can be evaluated using empirically derived
energy function or measure of packing efficiency
• Sequences can be scored based on how well they fit
the model (i.e., the best score achievable)
Baxevanis & Ouellette Ch 9 (Wishart)
Three-dimensional threading
• First described by Novotny et al (1984)
• Rediscovered in early 1990s
– Jones et al 2992; Sippl & Weitckus 1992; Bryant & Lawrence 1993
– Based largely on heuristic contact potentials (interactions
between pairs of residues)
– 3D coordinates of theoretical structure (based on threading of
sequence through PDB structure model) used to evaluate predicted
contacts and derive a fitness score based on a pseudoenergy
function
• Powerful for predicting 3D structure of unknown proteins, and
for evaluating structure of known proteins
• Limitations found in this method:
– interactions are not always conserved between distant homologs
– Computational complexity (very slow)
– Modest accuracy (early methods ignored amino acid
information; model accuracy >5Angstroms)
Baxevanis & Ouellette Ch 9 (Wishart)
Contact maps
• 2D plots of
distances between
C-alpha atoms of
all pairs of residues
– Observed
interactions
between amino
acids used to form
“contact potentials”
for 3D threading
methods
Figure 6.14
Creighton, Proteins Ch. 6
Two-dimensional threading
• Sequence-profile methods; combines predictions of secondary
structure prediction (and possibly solvent accessibility) with
standard profile methods to score and align proteins
• Improved accuracy through combined use of 2ary structure
prediction and amino acid similarity
• Much faster than standard 3D threading
• Model accuracy good but not excellent (RMSD >3 Angstroms)
– However, for model construction for proteins with no close
homologs with solved structure, these methods are among the best
• Examples:
– UCSC SAMT99 (two-track HMMs), PHYRE, FUGUE
Baxevanis & Ouellette Ch 9 (Wishart)
Assessing method performance
• Astral benchmark datasets
– Park et al
• CASP experiments
• EVA and Livebench
– Continuous evaluation of webservers
Park et al experimental design and
conclusions
• discussed in class
The EVA server
• Continuous assessment of the predictions of automatic servers
using the same measurements, the same standards, and the
same sequences to all methods
• New structures (pre-release to PDB) given to EVA by
participating structural biologists. EVA submits the amino acid
sequences to online servers.
• Predictions stored until release of 3D coordinates to PDB. Then
the predicted (2D or 3D) structures can be compared against the
solved structures, and given various scores.
• Approach enables the community to compare methods, and
gives developers concrete feedback that is critical for method
improvement.
Baxevanis & Ouellette Ch 8 (Ofran and Rost)
Critical Assessment of Protein
Structure Prediction (CASP)
Kryshtafovych et al, “Progress over the First
Decade of CASP Experiments” Proteins:
Structure, Function and Genetics 2005
Rosetta/Robetta
Red=first model
Yellow=models 2-5
Black=other groups
Red=first model
Yellow=models 2-5
Black=other groups
Selected protein structure
prediction servers
• Superfamily (Sequence-profile alignment; UCSD,
MRC/Cambridge, U. Bristol UK)
– http://supfam.org/SUPERFAMILY/index.html
• PHYRE (Profile-profile alignment; Imperial College of London)
– http://www.sbg.bio.ic.ac.uk/phyre/
• SwissModel (Swiss Institute of Bioinformatics)
– http://swissmodel.expasy.org//SWISS-MODEL.html
• MODBASE (precomputed models; Sali lab at UCSF)
– http://modbase.compbio.ucsf.edu/modbase-cgi/index.cgi
• PhyloFacts
– http://phylogenomics.berkeley.edu/phylofacts/
Summary
• Experimental determination of protein structure is expensive and
not always straightforward
• Predictive methods are relied upon to obtain clues to protein fold
(and function)
• Knowing what (which parts of a protein structure) you can
believe and what you can’t is critical for both experimental and
predicted structures
• Consensus and jury methods produce the best results
– E.g., protein structure prediction meta-servers
Summary (cont’d)
• Ab initio methods of protein fold prediction use physics-based
energy minimization to simulate the process of protein folding
– These methods are generally less successful than homology-based
fold prediction (limited to short peptides/small proteins)
– Exception: Rosetta/I-sites methods (Baker group) which employ
both types of approach
• Threading methods fall into the homology-based class of
approaches.
– 2D profiles use 2ary structure (prediction/knowledge) as well as
sequence information (and perhaps additional information).
– 3D profiles use 3D models and assign scores to proteins based on
inter-residue contacts based on the observed contacts in the
original structure template and derived contact potentials from other
structures
– It is possible to use threading approaches to predict structure for
non-homologous molecules (but this is rarely very successful)
Summary (cont’d)
•
Community assessment of 2D and 3D structure prediction uses various
approaches
–
EVA and LiveBench (continuous real-time assessment of methods)
–
CASP (Critical Assessment of Protein Structure Prediction)
–
Benchmark datasets (e.g., Astral PDB40 for fold recognition)
•
Reported accuracy of 2D structure prediction between 75-77% (for best
methods)
•
Reported accuracy of comparative models derived by 3D structure prediction
servers is harder to assess.
•
Fold prediction (ignoring the comparative model construction) is fairly accurate
for the best servers provided
•
–
A homologous structure has already been deposited in the PDB
–
That structure can be detected with a significant E-value using sequence information
alone, e.g., by PSI-BLAST)
The inclusion of 2ary structure prediction (e.g., in 2D profiles) can improve the
alignment and give a modest boost to fold recognition accuracy when %ID is
very low, but can also yield errors in prediction
Questions on the reading
•
•
David Baker and
Andrej Sali, “Protein
Structure Prediction •
and Structural
Genomics” Science
•
2001
What is the single most significant source of error in a comparative
model construction, if based on a template with <30% identity with
the target?
What is an additional probable source of error if the percent identity
drops below 20%?
What is the reason cited by Baker and Sali for why errors in a
comparative model tend to not lie in functionally important sites?
•
What example do Baker and Sali give to demonstrate the utility of a
low-accuracy comparative model?
•
How would protein-protein interaction interfaces be predicted with a
comparative model?
•
What are the possible applications of a comparative model?
•
What fraction (approximately) of comparative models produced by
Rosetta for proteins <150 residues in length are considered
accurate?
•
Does model refinement improve models or not? Under what
conditions?