Protein Tertiary Structures: Secondary article

Prediction from Amino Acid . Introduction

Article Contents

Sequences . Comparative Modelling

. Threading

Hongyu Zhang, Celera Genomics, Rockville, Maryland, USA . Ab Initio Prediction

. Discussion

Protein tertiary structures contain key information for the understanding of the
relationship between protein amino acid sequences and their biological functions. A large
collection of computational algorithms has been developed to predict protein tertiary
structures from their sequences in computers.

Introduction adopt similar structures. Therefore, the easiest and also

Proteins are polypeptide chains consisting of a large the most accurate way to predict the protein tertiary
number of amino acid residues that are covalently linked structure is to build the structure based on sequence
together via amide bonds. The order in which the 20 relatives that have high sequence similarities to the
different amino acids are arranged in a protein chain is also target protein according to the sequence alignment
called the primary structure of the protein. The polypep- results. Such an approach is called comparative modelling.
tide backbones of proteins exist in particular conforma- In most cases those sequence relatives and the target
tions known as the secondary structures. The secondary protein belong to the same functional family in
structures as well as their side-chains are then packed into biology, i.e. they are homologues of each other. Thus,
three-dimensional structures referred to as the tertiary traditionally, comparative modelling is also called homo-
structures. logy modelling.
The biological function of a protein is often intimately
dependent upon its tertiary structure. X-ray crystallogra-
phy and nuclear magnetic resonance are the two most
mature experimental methods used to provide detailed
information about protein structures. However, to date the Target protein
majority of the proteins still do not have experimentally
determined structures available. As at December 2000,
there were about 14 000 structures available in the protein
Database search
data bank (PDB, [Link] and there are
about 10 106 000 sequence records sequences in GenBank
([Link] Thus theoreti-
cal methods are very important tools to help biologists Homologue
obtain protein structure information. The goal of theore- detected? Yes
tical research is not only to predict the structures of
proteins but also to understand how protein molecules fold No Building alignment
into the native structures.
The current methods for protein structure prediction Threading
can be roughly divided into three major categories:
comparative modelling; threading; and ab initio prediction.
Yes
For a given target protein with unknown structure, the
general procedure for predicting its structure is described Fold Building model
in Figure 1. detected?

No

Comparative Modelling Ab initio prediction

From the available experimental data it has been observed Figure 1 Procedure for predicting a protein structure from its amino acid
that proteins with similar amino acid sequences usually sequence.

ENCYCLOPEDIA OF LIFE SCIENCES / & 2002 Macmillan Publishers Ltd, Nature Publishing Group / [Link] 1
 Protein Tertiary Structures: Prediction from Amino Acid Sequences

Database search There is no trivial solution for aligning two protein

sequences because of the vast number of combinations
An initial step for comparative modelling is to check between amino acid pairs. Fortunately, a classical algo-
whether there is any protein in the current PDB having the rithm, originally from the computer science field, called the
similar sequence or function to the target protein. A dynamic programming algorithm can guarantee to quickly
protein found will then serve as the structural template for find the optimal alignment given a score function (Needle-
modelling the target protein. In most situations, the man and Wunsch, 1970; Smith and Waterman, 1981). The
searching of the template has to proceed using a sequence basic philosophy of the algorithm is to build up an optimal
comparison algorithm that is able to identify the global alignment using previous solutions to smaller subse-
sequence similarity. In some cases, even when there is no quences.
global sequence similarity between two protein sequences, The central step in Needleman–Wunsch algorithm is the
a close match between some important sequence fragments construction of a score matrix. Each element in the score
or local sequence patterns (also called motifs) is still matrix, F(i,j), is the score of the best alignment between the
significant enough for us to identify the homologous initial segment x1,_i of sequence x and y1,_j of sequence y.
relationship between protein sequences. The ‘trick’ of the algorithm is that F(i,j) can be built
To start a database search, one first needs a score recursively according to eqn [1].
function that can evaluate the similarity between amino
acids. Various score functions are available. The simplest 8
one is the identity score function, which gives score 1 for an < Fði  1; j  1Þ þ sðxi ; yj Þ
amino acid matched to the same type of amino acid and Fði; jÞ ¼ max Fði  1; jÞ   ½1
:
score 0 for an amino acid mutated to a different type of Fði; j  1Þ  
amino acid. More advanced score functions are based on
the statistics of the amino acid substitution frequencies in In eqn [1], s(xi,yj) is the score of aligning a residue
known aligned homologous sequence families. Among pair (xi,yj), and d is the score of a residue aligned
them, most popular ones are Dayhoff (Dayhoff et al., 1978) to a gap. The principle is that the best alignment score
and Blosum (Henikoff and Henikoff, 1992) matrices. The F(i,j) can only come from the three possible ways
20  20 elements in the matrices represent the substitution shown in the above equation: either the last residues of
scores between 20 natural amino acids. the two sequences (xi and yj), aligned together, or any of
To search a large sequence database, the computer them aligned to a gap. At the beginning, F(0,0) is initialized
algorithms have to be able to find the close sequences to 0, and F(i,0), F(0,j) are initialized to 2 id and 2 jd
correctly and quickly. Some quite efficient algorithms have because they represent i or j residues that are aligned to
been developed to solve the database search time problem, gaps.
such as BLAST (Altschul et al., 1990, 1997) and FASTA The Needleman–Wunsch algorithm is used to look
(Pearson and Lipman, 1988). BLAST is currently the most for the best match between two sequences from one
popular database search protocol. Its central idea is to end to the other. A more common situation is looking
transform the whole sequence comparison problem into an for the best alignment between the subsequences of
easier problem of local fragment matching and extension. two sequences, which can locate the common regions
FASTA achieves much of its speed and selectivity by using shared between two proteins that could have little
a lookup table to locate all identities or groups of identities global similarity. A very similar dynamic algorithm called
between two sequences (Pearson, 1990). the Smith–Waterman algorithm was developed to solve
such local alignment problems (Smith and Waterman,
Sequence alignment 1981).
These algorithms have become the standard algorithms in
After finding the template sequences for the target this field after 20 years of improvement. Researchers can
sequence in the structure database, the second step in download their mature implementation programs from the
comparative modelling is to align the target sequence to the Internet, such as the popular CLUSTAL program (http://
template sequence. An alignment algorithm is used to find [Link]/BioInfo/ClustalW/) (Higgins et al.,
an optimal alignment for the two sequences. The result will 1989).
indicate the matching, insertion or deletion of the amino After automatically constructing the initial alignment
acids between the target sequence and the template using the dynamic programming technique, some human
sequence. Thus, from a sequence alignment one can decide intervention is helpful to adjust the errors in the
the structural features of each amino acid in the target computer-generated alignment; graphic tools with the
protein based on the structural features of its correspond- addition of human expertise can identify some possibly
ing template residue. If there are multiple templates, a inappropriate matches, such as a hydrophobic residue in
multiple sequence alignment can further improve the the target being matched to the surface region in the
accuracy of sequence–structure alignment. template structure.

2 ENCYCLOPEDIA OF LIFE SCIENCES / & 2002 Macmillan Publishers Ltd, Nature Publishing Group / [Link]
 Protein Tertiary Structures: Prediction from Amino Acid Sequences

Building the homology model the misaligned residues or inappropriately generated side-
chains and repeat the modelling procedure again.
Once the alignment is completed, one can start to build the For some years there have been very good commercial
structure model for the target protein based on the packages available in this field that bundle the comparative
template structure. The major steps in building the modelling modules into one piece of software, plus some
homology model are conserved region modelling and loop other extra functions. They often run on powerful Unix
region modelling. Conserved regions refer to those regions workstations and provide very user-friendly graphic
with conserved amino acids in the sequence alignment, interfaces. Among the most popular ones are QUANTA
most often those regions having standard secondary and Insight-II produced by MSI and Sybyl produced by
structures (a helix and b strands) in the template structure. Tripos.
Those regions will very probably keep their conformations
unchanged from the template structure to the target
structure, and are therefore easy to build at the very
beginning. It is usually straightforward to copy the Threading
structure in those regions from the template to the target.
Loop regions are hard to model because they are less Of all the proteins in the current sequence database, only
conserved in structure. In most situations they are located about 10–20% of sequences can be modelled by compara-
on the protein surface exposed to the solvent and do not tive modelling methods. For all the other sequences, it is
have standard secondary structures. Traditionally, loop- difficult to find sequence relatives using plain sequence
modelling methods were categorized into two kinds of comparison methods.
approaches: knowledge-based approach and ab initio Threading improves the sequence alignment sensitivity
approaches (for details, see the review section in Zhang by introducing structural information into the alignment,
et al., 1997). The knowledge-based approach extracts the where the structural information refers to the secondary or
knowledge from the current protein structure database and tertiary structural features of proteins. This helps because
then applies it in the building of the new loops; the ab initio amino acids have different propensities for different
approach usually uses some kinds of theoretical conforma- secondary structures or tertiary structure environments.
tional search method such as the Monte Carlo or simulated For example, some amino acids are more often observed in
annealing methods (Leach, 1996) to build up the new a helices than in other secondary structure units, while
loops. Ab initio methods are more general methods because some amino acids appear more frequently in hydrophobic
they are not prohibited by the current size of the structure environments than do others.
database, but traditionally they are much slower than the The threading method is sometimes called the fold
knowledge-based methods and therefore are not suitable recognition method. Its basic assumption is that the
for modelling very long loops. Some improved ab initio number of protein folds existing in nature is limited, from
algorithms have achieved very high efficiency and can several hundreds to over 1000, according to different
successfully model long protein loops very quickly (Zhang theories (Wang, 1998). The goal of fold recognition is to
et al., 1997). Knowledge-based and ab initio algorithms can identify the correct fold for the target sequence.
be combined together to improve the modelling accuracy; Most of the threading algorithms are based on the
for example, one can apply both methods in the same loop dynamic algorithm, but the key difference is the scoring
region and, if they produce the similar result, have higher strategy: in most threading algorithms the score functions
confidence to one’s predictions. include the structure information in addition to the
After constructing the structures in both conserved sequence information. The earliest threading approach is
regions and loop regions, the last steps of comparative the ‘3D profiles’ method (Bowie et al., 1991; Luthy et al.,
modelling include side-chain modelling and model evalua- 1992), in which the structural environment in each residue
tion/refinement. Methods for side-chain modelling include position of the template is classified into 18 classes based on
Monte Carlo, genetic algorithm, side-chain rotamer the position’s burial status, local secondary structure and
library and others (Leach, 1996). They have already polarity. The threading score matrix is then deduced from
reached very high precision (Dunbrack, 1999). The model the probability of all amino acids present in those 18 classes
evaluation usually includes the checking of Ramanchan- of structure environment. For example, if a hydrophobic
dra graphs and atomic packing. Molecular mechanics and residue is aligned to a buried template position, the score
molecular dynamics are common tools (Leach, 1996) for matrix is supposed to give a high score to encourage such a
refining the final model. type of sequence–structure match. The threading methods
It should be pointed out that the above procedure is not a of Jones et al. (1992) and Godzik et al. (1992) are based on
simple one-way street; in most cases it is an iterative the protein residue pairwise interaction energy methods
procedure. For example, one can start with an initial such as the potential of mean force method of Sippl (1990).
sequence alignment and build the structure model; after The energy formulae are derived from statistical analyses
evaluating the structure model one can go back to correct of current protein structure database and reflect the

ENCYCLOPEDIA OF LIFE SCIENCES / & 2002 Macmillan Publishers Ltd, Nature Publishing Group / [Link] 3
 Protein Tertiary Structures: Prediction from Amino Acid Sequences

residue–residue distance distribution probabilities in Ab Initio Prediction

known protein structures. In each step of the threading
procedure, the alignment score is calculated by adding Despite the great effort previously spent on comparative
up all the pairwise interaction energies between each modelling and threading, there remains a large proportion
target residue and the template residues surrounding of protein sequences with neither homologues nor clear
them. folds detected. From the early 1970s, people began starting
In addition to the above methods using the sequence– to look for ambitious ab initio algorithms that could
structure match scores, some other threading methods also directly attack the protein folding problem, that is to use
use the structure–structure match scores to evaluate the supercomputers to explore the huge conformational space
alignment between the target and the template. In those of protein molecules and find the pathways that lead
methods, although the target structure is unknown, proteins to their native conformations. The methods are
one can still characterize it using some predicted based on the assumption that a protein molecule’s native
structure properties, such as the predicted secondary structure is the lowest free energy state among all its
structures or the predicted residue burial status (Rost possible alternative conformations. This assumption has
and Sander, 1994). been demonstrated to be true by much experimental data,
Another important threading method is the Profile most famously the pioneering experiment of Christian
Hidden Markov Model method (HMM, see review of Anfinsen. The attraction of the ab initio approach is that it
Durbin et al., 1998). This is a very sensitive tool in not only promises to solve the protein structure prediction
searching for remote homologues because of its strong problem without being limited by the current protein
statistics background. A HMM is basically a probability structure database but it can also provide theoretical
distribution model. To build the profile HMM, first all the explanations of how proteins fold into their native
sequences in the database need to be clustered into a structures – in other words the answer to the famous
handful of families. Each family is then used to train a protein folding enigma.
HMM. Finally, the target sequence is aligned to From the 1970s, scientists from various fields, including
those HMMs to identify the family to which it belongs. biology, chemistry, physics, computer science and mathe-
Although the structural information usually is not matics, have collaborated to develop all sorts of ab initio
explicitly characterized in HMMs, it is implied in the structure prediction methods and have published numer-
corresponding statistical models. A HMM algorithm ous papers. However, no significant progress was made
developed by Di Francesco et al. (1997a,b) used over a very long period. In recent years, because of the
the structure information directly, in which the target rapid expansion of experimental data and the rapid
structure is characterized by the predicted secondary increase in computer speeds, deeper insight has been
structure while the template structures are represented by gained to the protein folding problem and new algorithms
profile HMMs trained on the template’s secondary have been developed that are beginning to show encoura-
structure patterns. ging results in the blind protein structure prediction tests
Some advanced sequence search methods such as PSI- (Moult et al., 1999).
BLAST (Altschul, 1997) utilize more sensitive position- Figure 2 gives a schematic view of ab initio prediction
dependent score matrices, which are very good at detecting algorithms (after Lin, 1996). The figure indicates that three
remote homologues. Some people also consider them to components are essential for designing an ab initio
belong to threading methods because of their high algorithm, shown as the three dimensions in the figure.
searching sensitivity compared to basic database searching All the ab initio folding algorithms can be considered
algorithms. different combinations of the three components.
Although threading methods are good at detecting The first dimension in Figure 2 is the protein model, which
remote homologues, they are often not able to give good is used to characterize the protein molecules in the
sequence–structure alignment. The main reason is that the computer. This can be as complicated as the explicit
structure information is included in threading with many atomic model in the classic molecular dynamics programs,
approximations, and thus can introduce significant noise in which all protein atoms and their related physical
into the final alignment. For example, most threading chemical properties (bond, order, length and angle,
methods use the so-called ‘frozen’ approximation, that is electronic charge, etc.) are explicitly described; or it can
they assume that the target residues are in the same be a simple model like the simplified residues model, in
environments as the template residues if they belong which each residue is represented as a single particle in
to the same structural fold. In reality, even two closely space. The lattice model represents the protein atoms or
homologous structures can have slightly different residues using discrete integer points in three-dimensional
residue environments, especially in loop regions. This is space, so the program is faster. Generally, the more
one reason why Bryant’s group use only conserved regions complicated a model, the better it can describe the physical
in threading (Bryant and Lawrence, 1993; Madej et al., chemical properties of proteins, but also the slower the
1995). algorithm will be.

4 ENCYCLOPEDIA OF LIFE SCIENCES / & 2002 Macmillan Publishers Ltd, Nature Publishing Group / [Link]
 Protein Tertiary Structures: Prediction from Amino Acid Sequences

Conformational function, in which b is the bond length and the values

search method of the minimum energy bond length b0 and force
constant kb are dependent on the specific bond type.
The second term is the bond angle bending potential,
Fragment assembling which is a three-body interaction; y, y0 and ky are the
bond angle, minimum-energy bond angle and force
Genetic algorithm constant. The four-body interactions fall into two cate-
Molecular mechanics potential

Simulated annealing gories: one is a harmonic potential to constrain the

dihedral angle x, the other is a cosine potential that
Monte Carlo allows the dihedral angle j to rotate 3608; kx, kj, x0, d
Molecular dynamics and n are the corresponding constants. The last summation
term is the sum of two terms representing nonbonding
Potential of mean force

Graph theory interactions, which consist of the van der Waals potential
Systematic search and the electrostatic potential between atoms i and j. C12
3D profile

and C6 are the Lennard-Jones constants, rij is the distance

between atoms i and j, and e0 and er are the dielectric
Protein constant in vacuum and the relative dielectric constant in a
s
ue
s
om

model
el

sid
s

medium.
m

od

re
at
to

ed
d
ta

The advantage of the molecular mechanics potentials is

ite

ce

ifi
ici

tti
Un

pl

that they can explicitly characterize the physical chemical

pl

La

m
Ex

Si

interactions in proteins at detailed atomic scale; but they

are very slow to compute and also are not good for
Potential evaluating the solvent interactions, especially the impor-
function
tant solvent entropy effect in protein folding. Thus, many
Figure 2 Schematic view of ab initio prediction methods (revised from of the latest ab initio folding algorithms prefer to use simple
Lin, 1996). threading potentials as described earlier. The threading
potentials, which are also called knowledge-based poten-
Potential function is the second dimension in Figure 2; this tials, are derived from the current protein structure
describes the physical chemical interactions both within database and reflect either residue–residue distance
protein molecules and between protein molecules and their distribution probabilities or residue-to-environment and
environments. The ideal potential function is expected to residue-to-structure propensities.
rank the native conformation as the lowest free energy The last dimension in Figure 2 is the conformational
conformation among all possible alternatives. One of the search method, which is how the conformational space of
most popular potential functions used in ab initio algo- proteins is explored to look for the lowest free energy
rithms is the molecular mechanics potential widely adopted conformation. Since proteins are long-chain biopolymers,
in molecular dynamics and molecular mechanics simula- they have a large number of internal degrees of freedom
tions, such as CHARMM (Brooks et al., 1993), AMBER originating from both main-chain and side-chain dihedral
(Pearlman et al., 1995) and GROMOS (van Gunsteren and angles. The simplest conformational search method is the
Berendsen, 1990). Its general form is shown in eqn [2]. systematic search. This divides each dihedral angle into a
few discrete states approximately representing the local
X1 X1 energy minima of that angle. One can then generate
Vðr1 ; r1 ; :::; rN Þ ¼ kb ðb  b0 Þ2 þ k ð  0 Þ2 approximately all the possible conformations of the whole
2 2
bonds angles molecule by combining all the states of each dihedral
X 1 angle. Because of the exponential increase in the number
þ k ð  0 Þ2 of combinations as the molecular size increases, it is
improper
2
dihedrals actually impossible to use this method in any real protein
X systems.
þ k’½1 þ cosðn’  Þ The problem of exploring the conformational space of
dihedrals proteins is a typical combinatorial problem in computer
" # science, which has been demonstrated to be NP-complete
X C12 ði; jÞ C6 ði; jÞ qi qj in complexity (Ngo and Marks, 1992). This means that no
þ 12
 6
þ efficient algorithm is guaranteed to find the answer to the
pairsði; jÞ
rij rij 4"0 "r rij
problem in a time bounded by a polynomial function of the
[2] protein size.
The first term in eqn [2] is the bond stretch interaction along Present ab initio prediction algorithms use virtually
the covalent bond direction. It is represented by a harmonic every kind of advanced algorithm that has been used in

ENCYCLOPEDIA OF LIFE SCIENCES / & 2002 Macmillan Publishers Ltd, Nature Publishing Group / [Link] 5
 Protein Tertiary Structures: Prediction from Amino Acid Sequences

solving combinatorial problems, such as molecular dy- Discussion

namics (Duan and Kollman, 1998), Monte Carlo
(Simons et al., 1999; Ortiz et al., 1999), genetic algorithms From the review in the previous sections, it can be seen that
(Pederson and Moult, 1997), simulated annealing the comparative modelling method has become a very
and graph theory methods. The molecular dynamics mature approach for protein structure prediction, while
algorithm simulates the movement of the atoms of proteins more recent advances in threading methods effectively
and solvents based on classical Newtonian laws, and thus extend the structure prediction scale to remote homo-
has a strong physics background. However, most of the logues. Finally, cutting-edge developments in software and
latest ab initio prediction algorithms tend to use Monte hardware have brought ab initio algorithms very close to
Carlo algorithms or genetic algorithms because the most real application.
effective potential functions nowadays for ab initio One of the latest developments related to protein
prediction are knowledge-based threading functions, structure prediction is the emergence of the structural
which in most cases are discrete and unable to calculate genomics project in the post-human genomics era. After
molecular forces for molecular dynamics simulations. Celera Genomics and the public effort headed by NIH
Some workers have also tried to combine the molecular dramatically finished the human genome project (HGP)
dynamics method with the Monte Carlo method in one ahead of the expected timetable, scientists around the
algorithm, as well as combining different potentials world started to collaborate on the structural genomics
(Zhang, 1999). project (Sanchez et al., 2000). The idea is to classify all the
Fragment-assembling algorithms increase the confor- proteins in the genome into homologous families and then
mational search efficiency by enumerating the limited to pick a representative sequence for each family to make
number of possible structures for any given protein experimental structures. Subsequently, the structures of all
fragment. The possible candidate structures are selected the sequences in the genome can be modelled using plain
on the basis of statistical analysis of the current protein comparative modelling methods. In other words, all future
structure database. Using these algorithms, it is not protein structure prediction work would be comparative
necessary to spend a great deal of time exploring the modelling. On the other hand, other structure prediction
conformational space of every fragment; instead, whole methods still can be useful in the future; for example, ab
protein conformations can be obtained by assembling the initio algorithms can still be used to study the theoretical
limited number of fragment conformations. As a result, the basis of the protein folding problem.
program can be fast enough to search the conformational
space of small to medium-sized proteins currently using
Monte Carlo or genetic algorithms. In addition to the References
speed advantage, the fragment-assembling algorithms can Altschul SF, Gish W, Miller W, Myers EW and Lipman DJ (1990) Basic
guarantee to give reasonable local structures, at least for local alignment search tool. Journal of Molecular Biology 215:
the fragment structures selected. 403–410.
In the history of protein structure prediction, the authors Altschul SF, Madden TL, Schaffer AA et al. (1997) Gapped BLAST and
of ab initio algorithms have tended to overestimate the PSI-BLAST: a new generation of protein database search programs.
performance of their algorithms because of the lack of Nucleic Acids Research 25: 3389–3402.
Bowie JU, Luthy R and Eisenberg D (1991) A method to identify protein
objective assessment methods. Starting from 1994, John
sequences that fold into a known three-dimensional structure. Science
Moult and his co-workers organized a series of conferences 253: 164–170.
named CASP (Critical Assessment of techniques for Bryant SH and Lawrence CE (1993) An empirical energy function for
protein Structure Prediction). The procedure of CASP is threading protein-sequence through the folding motif. Proteins 16:
to first collect a number of protein targets whose structures 92–112.
are soon to be solved by X-ray crystallography, those Brooks BR, Bruccoleri RE, Olafson BD, States DJ, Swaminathan S and
targets are posted on the Internet, inviting predictors Karplus M (1993) CHARMM: a program for macromolecular energy
minimization, and dynamics calculations. Journal of Computational
around the world to submit their predictions before the
Chemistry 4: 187–217.
experimental structures become public. After the experi- Dayhoff MO, Schwartz RM and Orcutt BC (1978) A model of
mental structures are solved, the committee of CASP uses evolutionary change in protein matrices for detecting distant relation-
objective criteria such as the root mean square deviation ships. In: Dayhoff MO (ed.) Atlas of Protein Sequence and Structure,
between the predicted structure and the real structure to vol. 5, supplement 3, pp. 345–352. Washington, DC: National
evaluate the success of all predictions. Biomedical Research Foundation.
The CASP3 results showed that several ab initio Di Francesco V, Garnier J and Munson PJ (1997a) Protein topology
recognition from secondary structure sequences: application of the
prediction groups have produced reasonably accurate
hidden Markov models to the alpha class proteins. Journal of
models of protein fragments of up to 60 residues or so Molecular Biology 267: 446–463.
(Orengo et al., 1999; Simons et al., 1999; Ortiz et al., 1999), Di Francesco V, Geetha V, Garnier J and Munson PJ (1997b) Fold
especially the fragment assembling algorithm (Simons recognition using predicted secondary structure sequences and hidden
et al., 1999). Markov models of protein folds. Proteins (supplement 1): 123–128.

6 ENCYCLOPEDIA OF LIFE SCIENCES / & 2002 Macmillan Publishers Ltd, Nature Publishing Group / [Link]
 Protein Tertiary Structures: Prediction from Amino Acid Sequences

Duan Y and Kollman PA (1998) Pathways to a protein folding simulate the structural and energetic properties of molecules.
intermediate observed in a 1-microsecond simulation in aqueous Computational Physics Communications 91: 1–41.
solution. Science 282: 740–744. Pearson WR and Lipman DJ (1988) Improved tools for biological
Dunbrack RL Jr (1999) Comparative modeling of CASP3 targets using sequence comparison. Proceedings of the National Academy of
PSI-BLAST and SCWRL. Proteins (supplement 3): 81–87. Sciences of the USA 85: 2444–2448.
Durbin R, Eddy S, Krogh A and Mitchison G (1998) Biological Sequence Pearson WR (1990) Rapid and sensitive sequence comparison with
Analysis: Probabilistic Models of Proteins and Nucleic Acids. Cam- PASTP and FASTA. Methods in Enzymology 183: 63–98.
bridge: Cambridge University Press. Pederson JT and Moult J (1997) Ab initio protein folding simulations
Godzik A and Skolnick J (1992) Sequence–structure matching in with genetic algorithms: simulations on the complete sequence of small
globular proteins: application to supersecondary and tertiary struc- proteins. Proteins (supplement 1): 179–184.
ture determination. Proceedings of the National Academy of Sciences of Rost B and Sander C (1994) Combining evolutionary information and
the USA 89: 12098–12102. neural networks to predict protein secondary structure. Proteins 19:
Henikoff S and Henikoff JG (1992) Amino acid substitution matrices 55–72.
from protein blocks. Proceedings of the National Academy of Sciences Sanchez R, Pieper U, Melo F et al. (2000) Protein structure modeling for
of the USA 89: 10915–10919. structural genomics. Nature Structural Biology 7 (supplement):
Higgins DG and Sharp PM (1989) CLUSTAL: a package for performing 986–990.
multiple sequence alignments on a microcomputer. Gene 73: 237–244. Simons KT, Bonneau R, Ruczinski I and Baker D (1999) Ab initio
Jones DT, Taylor WR and Thornton JM (1992) A new approach to protein structure prediction of CASP III targets using Rosetta.
protein fold recognition. Nature 358: 86–89. Proteins (supplement 3): 171–176.
Leach AR (1996) Molecular Modelling: Principles and Applications. Sippl MJ (1990) Calculation of conformational ensembles from
Essex: Addison Wesley Longman. potentials of mean force. An approach to the knowledge-based
Lin D (1996) Knowledge-based Protein Fold and Folding Study. PhD prediction of local structures in globular proteins. Journal of
thesis, Peking University, p. 76. Molecular Biology 213: 859–883.
Luthy R, Bowie JU and Eisenberg D (1992) Assessment of protein Smith TF and Waterman MS (1981) Identification of common
models with three-dimensional profiles. Nature 356: 83–85. molecular subsequences. Journal of Molecular Biology 147: 195–197.
Madej T, Gibrat JF, Bryant SH (1995) Threading a database of protein van Gunsteren WF and Berendsen HJC (1990) Computer simulation of
cores. Proteins 23: 356–369. molecular dynamics: methodology, applications and perspectives in
Moult J, Hubbard T, Fidelis K and Pedersen JT (1999) Critical chemistry. Angewandte Chemie. International Edition in English 29:
assessment of methods of protein structure prediction (CASP): round 992–1023.
III. Proteins (supplement 3): 2–6. Wang ZX (1998) A re-estimation for the total numbers of protein folds
Needleman SB and Wunsch CD (1970) A general method applicable to and superfamilies. Protein Engineering 11: 621–626.
the search for similarities in the amino acid sequences of two proteins. Zhang H (1999) A new hybrid Monte Carlo algorithm for protein
Journal of Molecular Biology 48: 443–453. potential function test and structure refinement. Proteins 34: 464–471.
Ngo JT and Marks J (1992) Computational complexity of a problem in Zhang H, Lai L, Wang L, Han Y and Tang Y (1997) A fast and efficient
molecular structure prediction. Protein Engineering 5: 313–321. program for modeling protein loops. Biopolymers 41: 61–72.
Orengo CA, Bray JE, Hubbard T, LoConte L and Sillitoe I (1999)
Analysis and assessment of ab initio three-dimensional prediction,
secondary structure, and contacts prediction. Proteins (supplement 3): Further Reading
149–170.
Ortiz AR, Kolinski A, Rotkiewicz P, Ilkowski B and Skolnik J (1999) Ab Leach AR (1996) Molecular Modelling: Principles and Applications.
initio folding of proteins using restraints derived from evolutionary Essex: Addison Wesley Longman.
information. Proteins (supplement 3): 177–185. Eisenhaber F, Persson B and Argos P (1995) Protein structure
Pearlman DA, Case DA, Caldwell JW et al. (1995) AMBER, a package prediction: recognition of primary, secondary, and tertiary structural
of computer programs for applying molecular mechanics, normal features from amino acid sequence. Critical Reviews in Biochemistry
mode analysis, molecular dynamics and free energy calculations to and Molecular Biology 30: 1–94.

ENCYCLOPEDIA OF LIFE SCIENCES / & 2002 Macmillan Publishers Ltd, Nature Publishing Group / [Link] 7