a

r
X
i
v
:
0
7
0
5
.
1
8
4
5
v
1


[
p
h
y
s
i
c
s
.
b
i
o
-
p
h
]


1
3

M
a
y

2
0
0
7
Introduction to protein folding for physicists
Pablo Echenique

Theoretical Physics Department, University of Zaragoza,

Pedro Cerbuna 12, 50009, Zaragoza, Spain.
Institute for Biocomputation and Physics of Complex Systems (BIFI),
Edicio Cervantes, Corona de Aragon 42, 50009, Zaragoza, Spain.
February 1, 2008
Abstract
The prediction of the three-dimensional native structure of proteins
from the knowledge of their amino acid sequence, known as the protein
folding problem, is one of the most important yet unsolved issues of modern
science. Since the conformational behaviour of exible molecules is noth-
ing more than a complex physical problem, increasingly more physicists
are moving into the study of protein systems, bringing with them power-
ful mathematical and computational tools, as well as the sharp intuition
and deep images inherent to the physics discipline. This work attempts to
facilitate the rst steps of such a transition. In order to achieve this goal,
we provide an exhaustive account of the reasons underlying the protein
folding problem enormous relevance and summarize the present-day sta-
tus of the methods aimed to solving it. We also provide an introduction to
the particular structure of these biological heteropolymers, and we phys-
ically dene the problem stating the assumptions behind this (commonly
implicit) denition. Finally, we review the special avor of statistical
mechanics that is typically used to study the astronomically large phase
spaces of macromolecules. Throughout the whole work, much material
that is found scattered in the literature has been put together here to
improve comprehension and to serve as a handy reference.
1 Why study proteins?
Virtually every scientic book or article starts with a paragraph in which the
writer tries to persuade the readers that the topic discussed is very important
for the future of humankind. We shall stick to that tradition in this work; but
with the condence that, in the case of proteins, the persuasion process will
turn out to be rather easy and automatic.

E-mail address: pnique@unizar.es Web page: http://www.pabloechenique.com

1
Proteins are a particular type of biological molecules that can be found in
every single living being on Earth. The characteristic that renders them es-
sential for understanding life is simply their versatility. In contrast with the
relatively limited structural variations present in other types of important bio-
logical molecules, such as carbohydrates, lipids or nucleic acids, proteins display
a seemingly innite capability for assuming dierent shapes and for producing
very specic catalytic regions on their surface. As a result, proteins constitute
the working force of the chemistry of living beings, performing almost every
task that is complicated. Quoting the rst sentence of a section (which shares
this sections title) in Lesks book [1]:
In the drama of life on a molecular scale, proteins are where the
action is.
Just to state a few examples of what is meant by action, in living beings,
proteins
are passive building blocks of many biological structures, such as the coats
of viruses, the cellular cytoskeleton, the epidermal keratin or the collagen
in bones and cartilages;
transport and store other species, from electrons to macromolecules;
as hormones, transmit information and signals between cells and organs;
as antibodies, defend the organism against intruders;
are the essential components of muscles, converting chemical energy into
mechanical one, and allowing the animals to move and interact with the
environment;
control the passage of species through the membranes of cells and or-
ganelles;
control gene expression;
are the essential agents in the transcription of the genetic information into
more proteins;
together with some nucleic acids, form the ribosome, the large molecular
organelle where proteins themselves are synthesized;
as chaperones, protect other proteins to help them to acquire their func-
tional three-dimensional structure.
Due to this participation in almost every task that is essential for life, protein
science constitutes a support of increasing importance for the development of
modern Medicine. On one side, the lack or malfunction of particular proteins is
behind many pathologies; e.g., in most types of cancer, mutations are found in
2
Figure 1.1: Four molecular machines formed principally by proteins. Figures
taken from the Molecule of the month section of the RSCB Protein Data Bank
(http://www.pdb.org), we thank the RSCB PDB and David S. Goodsell, from the
Scripps Research Institute, for kind permission to use them. (a) ATP synthase: it
acts as an energy generator when it is traversed by protons that make its two cou-
pled engines rotate in reverse mode and the ATP molecule is produced. (b) RNA
polymerase: it slides along a thread of DNA reading the base pairs and synthesizing
a matching copy of RNA. (c) GroEL-GroES complex: it helps unfolded proteins to
fold by sheltering them from the overcrowded cellular cytoplasm. (d) Ribosome: it
polymerizes amino acids to form proteins following the instructions written in a thread
of mRNA.
3
the tumor suppressor p53 protein [2]. Also, abnormal protein aggregation char-
acterizes many neurodegenerative disorders, including Huntington, Alzheimer,
Creutzfeld-Jakob (mad cow), or motor neuron diseases [35]. Finally, to at-
tack the vital proteins of pathogens (HIV [6, 7], SARS [8], hepatitis [9], etc.),
or to block the synthesis of proteins at the bacterial ribosome [10], are common
strategies to battle infections in the frenetic eld of rational drug design [11].
Apart from Medicine, the rest of human technology may also benet from
the solutions that Nature, after thousands of millions
1
of years of research, has
found to the typical practical problems. And that solutions are often proteins:
New materials of extraordinary mechanical properties could be designed from
the basis of the spider silk [12, 13], elastin [14] or collagen proteins [15]. Also,
some attempts are being made to integrate these new biomaterials with living
organic tissues and make them respond to stimuli [16]. Even further away on
the road that goes from passive structural functions to active tasks, no engineer
who has ever tried to solve a dicult chemical problem can avoid to experience
a feeling of almost religious inferiority when faced to the speed, eciency and
specicity with which proteins cut, bend, repair, carry, link or modify other
chemical species. Hence, it is normal that we play with the idea of learning to
control that power and have, as a result, nanoengines, nanogenerators, nanoscis-
sors, nanomachines in general [17]. The author of this work, in particular, felt a
small sting of awe when he learnt about the pump and the two coupled engines of
the principal energy generator in the cell, the ATP synthase (gure 1.1a); about
the genetic Xerox machine, the RNA polymerase (gure 1.1b); about the hut
where the proteins fold under shelter, the GroEL-GroES complex (gure 1.1c);
or about the macromolecular factory where proteins are created, the ribosome
(gure 1.1d), to mention four specially impressive examples. Agreeing again
with Lesk [1]:
Proteins are fascinating molecular devices.
From a more academic standpoint, proteins are proving to be a powerful
centre of interdisciplinary research, making many diverse elds and people with
dierent formations come in contact
2
. Proteins force biologists, biochemists
and chemists to learn more physics, mathematics and computation and force
mathematicians, physicists and computer technicians to learn more biology, bio-
chemistry and chemistry. This, indeed, cannot be negative.
In 2005, in a special section of Science magazine entitled What dont we
know? [18], a selection of the hundred most interesting yet unanswered scientic
questions was presented. What indicates the role of proteins, and particularly of
the protein folding problem (treated in section 3), as focuses of interdisciplinary
collaboration is not the inclusion of the question Can we predict how proteins
will fold?, which was a must, but the large number of other questions which were
1
Herein, we shall use the British convention for naming large numbers; in which 10
9
=a
thousand million, 10
12
=a billion, 10
15
=a thousand billion, 10
18
=a trillion, and so on.
2
The Institute for Biocomputation and Physics of Complex Systems, which the author is
part of, constitutes an example of this rather new form of collaboration among scientists.
4
related to or even dependent on it, such as Why do humans have so few genes?,
How much can human life span be extended?, What is the structure of water?,
How does a single somatic cell become a whole plant?, How many proteins are
there in humans?, How do proteins nd their partners?, How do prion diseases
work?, How will big pictures emerge from a sea of biological data?, How far can
we push chemical self-assembly? or Is an eective HIV vaccine feasible?.
In this direction, probably the best example of the use that protein science
makes of the existing human expertise, and of the positive feedback that this
brings up in terms of new developments and resources, can be found in the ma-
chines that every one of us has on his/her desktops. In a rst step, the enormous
amount of biological data that emerges from the sequencing of the genomes of
dierent living organisms requires computerized databases for its proper l-
tering. The NCBI GenBank database
3
, which is one of the most exhaustive
repositories of sequenced genetic material, has doubled the number of deposited
DNA bases approximately every 18 months since 1982 (see gure 1.2a) and has
recently (in August 2005) exceeded the milestone of 100 Gigabases (10
11
) from
over 165,000 species.
Among them, and according to the Entrez Genome Project database
4
, the
sequencing of the complete genome of 366 organisms has been already achieved
and there are 791 more to come in next few years. In the group of the completed
ones, most are bacteria, and there are only two mammals: the poor laboratory
mouse, Mus Musculus, and, notably [19], the Homo Sapiens (with 3 10
9
bases and a mass-media-broadcast battle between the private rm Celera and
the public consortium IHGSC).
However, not all the DNA encodes proteins (not all the DNA is genes). Typ-
ically, more than 95% of the genetic material in living beings is junk DNA, also
called non-coding DNA (a more neutral term which seems recommendable in the
light of some recent discoveries [2022]). So, in a second step, the coding regions
must be identied and each gene translated into the amino acid sequence of a
particular protein
5
. The UniProt database
6
is, probably, the most comprehen-
sive repository of these translated protein sequences and also of others coming
from a variety of sources, including direct experimental determination [25, 26].
UniProt is comprised by two dierent sub-databases: the Swiss-Prot Protein
Knowledgebase, which contains extensively human-annotated protein sequences
with low redundancy; and TrEMBL, which contains computer-annotated se-
quences extracted directly from the underlying nucleotide entries at databases
such as GenBank and where only the most basic redundancies have been re-
moved.
The UniProt/Swiss-Prot database contains, at the moment (on 30 May
3
http://www.ncbi.nlm.nih.gov/Genbank/
4
http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=genomeprj
5
Note that many variations [23, 24] may occur before, during and after the process of
gene expression, so that the relation gene-to-protein is not one-to-one. The size of the human
proteome (the number of dierent proteins), for example, is estimated to be an order of
magnitude or two larger than the size of the genome.
6
http://www.uniprot.org
5
10
6
10
7
10
8
10
9
10
10
10
11
10
12
A
p
r 2
0
0
6
D
e
c
2
0
0
4
D
e
c
2
0
0
2
D
e
c
2
0
0
0
D
e
c
1
9
9
8
D
e
c
1
9
9
6
D
e
c
1
9
9
4
D
e
c
1
9
9
2
D
e
c
1
9
9
0
D
e
c
1
9
8
8
N
o
v
1
9
8
6
N
o
v
1
9
8
4
D
e
c
1
9
8
2
N
u
m
b
e
r o
f d
e
p
o
s
ite
d
D
N
A
b
a
s
e
s
a
t G
e
n
B
a
n
k
100 Gigabases
10
3
10
4
10
5
10
6
10
7
D
e
c
2
0
0
6
D
e
c
2
0
0
5
D
e
c
2
0
0
4
D
e
c
2
0
0
3
D
e
c
2
0
0
2
D
e
c
2
0
0
1
D
e
c
2
0
0
0
D
e
c
1
9
9
9
D
e
c
1
9
9
8
D
e
c
1
9
9
7
D
e
c
1
9
9
6
D
e
c
1
9
9
5
D
e
c
1
9
9
4
D
e
c
1
9
9
3
D
e
c
1
9
9
2
D
e
c
1
9
9
1
D
e
c
1
9
9
0
D
e
c
1
9
8
9
D
e
c
1
9
8
8
D
e
c
1
9
8
7
D
e
c
1
9
8
6
N
u
m
b
e
r o
f p
ro
te
in
s
e
q
u
e
n
c
e
s
a
t U
n
iP
ro
t
Swiss-Prot
TrEMBL
10
1
10
2
10
3
10
4
10
5
2005 2000 1995 1990 1985 1980 1975
N
u
m
b
e
r o
f p
ro
te
in
s
tru
c
tu
re
s
a
t P
D
B
10
3
10
4
10
5
10
6
10
7
10
8
10
9
2005 2000 1995 1990 1985 1980 1975 1970
N
u
m
b
e
r o
f tra
n
s
is
to
rs
o
n
a
n
in
te
g
ra
te
d
c
irc
u
it
4004
8008
8080
8086
286
386
486
Pentium
Pentium II
Pentium III
Pentium IV
Itanium
Itanium 2
Itanium 2
(9MB cache)
a b
d c
Figure 1.2: Recent exponential progress in genomics, proteomics and computer
technology. (a) Evolution of the number of DNA bases deposited at the GenBank
database. (b) Evolution of the number of protein sequences at the UniProt Swiss-Prot
and TrEMBL databases. (c) Evolution of the number of protein three-dimensional
structures at the Protein Data Bank. (d) Moores Law: evolution of the number of
transistors in the Intel CPUs.
2006), around 200,000 protein sequences from about 10,000 species, and it has
experienced an exponential growth (since 1986), doubling the number of records
approximately every 41 months (see gure 1.2b). In turn, the UniProt/TrEMBL
database contains almost 3 million protein sequences from more than 100,000
species, and its growth (from 1997) has also been exponential, doubling the
number of records approximately every 16 months (see gure 1.2b).
After knowing the sequence of a protein, the next step towards the under-
standing of biological processes is the characterization of its three-dimensional
structure. Most proteins perform their function under a very specic native
shape which involves many twists, loops and bends of the linear chain of amino
acids (see section 3). This spatial structure is much more important than the
sequence for biochemists to predict and understand the mechanisms of life and it
can be resolved, nowadays, by fundamentally two experimental techniques: for
small proteins, nuclear magnetic resonance (NMR) [27,28] and, more commonly,
for proteins of any size, x-ray crystallography [2931]. The three-dimensional
structures so obtained are deposited in a centralized public-access database
called Protein Data Bank (PDB)
7
[32]. From the 13 structures deposited in
1976 to the 33,782 (from more than a thousand species) stored in June 2006,
the growth of the PDB has been (guess?) exponential, doubling the number of
records approximately every 3 years (see gure 1.2c).
7
http://www.rcsb.org/pdb/
6
To summarize, in June 2006, we have sequenced partial segments of the ge-
netic material of around 160,000 species, having completed the genomes of only
366; we know the sequences of some of the proteins of around 100,000 species
and the three-dimensional structure of proteins in 1,103 species
8
. However, ac-
cording to the UN Millennium Ecosystem Assessment
9
, the number of species
formally identied is 1.7-2 million and the estimated total number of species
on Earth ranges from 5 million to 30 million [33]. Therefore, we should expect
that the exponential growth of genomic and proteomic data will continue to ll
the hard-disks, collapse the broadband connexions and heat the CPUs of our
computers at least for the next pair of decades.
Fortunately, the improvement of silicon technology behaves in the same way:
In fact, in 1965, Gordon Moore, co-founder of Intel, made the observation that
the number of transistors per square inch had doubled every year since the
integrated circuit was invented, and predicted that this exponential trend would
continue for the foreseeable future. This has certainly happened (although the
doubling time seems to be closer to 18 months) and this empirical law, which
is not expected to fail in the near future, has become to be known as Moores
Law (see gure 1.2d for an example involving Intel processors). So we do not
have to worry about running short of computational resources!
Of course, information produces more information, and public databases do
not end at the three-dimensional structures of proteins. In the last few years,
a number of more specic web-based repositories have been created in the eld
of molecular biology. There is the Protein Model Database (PMDB)
10
[34],
where theoretical three-dimensional protein models are stored (including all
models submitted to last four editions of the CASP
11
experiment [35]); the
ProTherm
12
and ProNIT
13
databases [36], where a wealth of thermodynamical
data is stored about protein stability and protein-nucleic acid interactions, re-
spectively; the dbPTM
14
database [37], that stores information on protein post-
translational modications; the PINT
15
database [38], with thermodynamical
data on protein-protein interactions; and so on and so forth.
In addition to the use of computers for storage and retrieval of enormous
quantities of data, the increasing numerical power of these machines is custom-
arily used for a wide variety of applications that range from molecular visual-
ization, to long simulations aimed to solve the equations governing biological
systems (the central topic discussed more in detail in the rest of this work).
Indeed, as Richard Dawkins has stated [39]:
What is truly revolutionary about molecular biology in the post-Watson-
Crick era is that it has become digital.
8
http://www.ebi.ac.uk/thornton-srv/databases/pdbsum/
9
http://www.millenniumassessment.org
10
http://www.caspur.it/PMDB/
11
http://predictioncenter.gc.ucdavis.edu
12
http://gibk26.bse.kyutech.ac.jp/jouhou/Protherm/protherm.html
13
http://gibk26.bse.kyutech.ac.jp/jouhou/pronit/pronit.html
14
http://dbPTM.mbc.nctu.edu.tw
15
http://www.bioinfodatabase.com/pint/
7
Finally, apart from all the convincing reasons and the appeals to authority
given above, what is crystal-clear is that proteins are an unsolved and dicult
enigma. And those are two irresistible qualities for any esh and blood scientist.
2 Summary of protein structure
In spite of their diverse biological functions, summarized in the previous section,
proteins are a rather homogeneous class of molecules from the chemical point
of view. They are linear heteropolymers, i.e., unbranched chains of dierent
identiable monomeric units.
Before they are assembled into proteins, these building units are called amino
acids and can exist as stand alone stable molecules. All amino acids are made up
of a central -carbon with four groups attached to it: an amino group (NH
2
),
a carboxyl group (COOH), a hydrogen atom and a fourth arbitrary group
(R) (see gure 2.2). In aqueous solvent and under physiological conditions,
both the amino and carboxyl groups are charged, the rst accepting one proton
and getting a positive charge, and the second giving one proton away and getting
a negative charge (compare gures 2.2a and 2.2c).
When the group R is not equal to one of the other three groups attached
to the -carbon, the amino acid is chiral, i.e., like our hands, it may exist in two
dierent forms, which are mirror images of one another and cannot be super-
imposed by rotating one of them in space (you cannot wear the left-hand glove
on your right hand). In chemical jargon, one says that the -carbon constitutes
an asymmetric centre and that the amino acid may exist as two dierent enan-
tiomers called L- (gure 2.2c) and D- (gure 2.2d) forms. It is common that,
when used as prexes, the L and D letters, which come from levorotatory and
dextrorotatory, are written in small capitals, as in L- and D-. This nomenclature
is based on the possibility of associating the amino acids to the optically active
L- and D- enantiomers of glyceraldehyde, and could be related to the +/- or to
the Cahn, Ingold and Prelogs R/S [41] notations. For us, it suces to say that
Figure 2.1: Color and size code for the atom types used in most of the gures in this
section. All the gures have been made with the Gaussview graphical front-end of
Gaussian03 [40] and then modied with standard graphical applications.
8
Figure 2.2: Amino acids. (a) Uncharged L-enantiomer. (b) CORN mnemotechnic
rule to remember which one is the L-form. (c) Charged L-enantiomer (the predomi-
nant form found in living beings). (d) Charged D-enantiomer.
the D/L nomenclature is, by far, the most used one in protein science and the
one that will be used in this work. For further details, take a look at the IUPAC
recommendations at http://www.chem.qmul.ac.uk/iupac/AminoAcid/.
In principle, amino acids may be L- or D-, and the group R may be any-
thing provided that the resultant molecule is stable. However, for reasons that
are still unclear [42], the vast majority of proteins in all living beings are made
up of L-amino acids (as a rare exception, we may point out the fact that D-amino
acids can be found in some proteins produced by exotic sea-dwelling organisms,
such as cone snails) and the groups R (called side chains) that are coded
in the genetic material comprise a set of only twenty possibilities (depicted in
gure 2.5).
A frequently quoted mnemotechnic rule for remembering which one is the
L-form of amino acids is the so-called CORN rule in gure 2.2b. According
to it, one must look from the hydrogen to the -carbon and, if the three re-
maining groups are labelled as in the gure, the word CORN must be read in
the clockwise sense of rotation. The author of this work does not nd this rule
very useful, since normally he cannot recall if the sense is clockwise or coun-
terclockwise. To know which form is the L- one, he draws the amino acid as
in gure 2.2a or 2.2c, with the -carbon in the centre, the amino group on the
left and the carboxyl group on the right, all of them in the plane of the paper
(which is very natural and easy to remember because it matches the normal
sense of writing with the fact that, conventionally, proteins start at NH
+
3
and
end at COO

). Finally, he must just remember that the side chain of the

L-amino acid goes out of the paper approaching the reader (which is also natural
9
Figure 2.3: Peptide bond formation reaction. The peptide plane is indicated in green.
because the side chain is the relevant piece of information and we want to look
at it closely).
The process through which amino acids are assembled into proteins (called
gene expression or protein biosynthesis) is typically divided in two steps. In the
rst one, the transcription, the enzyme ARN polymerase (see gure 1.1b) binds
to the DNA in the cellular nucleus and makes a copy of a section the gene
of the base sequence into a messenger RNA (mRNA) molecule. In the second
step, called translation, the mRNA enters the ribosome (see gure 1.1d) and is
read stopping at each base triplet (called codon). Now, a specic molecule of
transfer RNA (tRNA), which possesses the base triplet (called anticodon) that
is complementary to the codon, links to the mRNA bringing with her the amino
acid that is codied by the particular sequence of three bases. Each amino acid
that arrives to the ribosome in this way is covalently attached to the previous
one and so added to the nascent protein. In this reaction, the peptide bond is
formed and a water molecule is released (see gure 2.3). This process continues
until a stop codon is read and the transcription is complete.
The amino acid sequence of the resultant protein, read from the amino
terminus to the carboxyl terminus, is called primary structure; and the amino
acids included in such a polypeptide chain are normally termed amino acid
residues, or simply residues, in order to distinguish them from their isolated
form. The main chain formed by the repetition of -carbons and the C and N
atoms at the peptide bond is called backbone and the R groups branching out
from it are called side chains, as it has already been mentioned.
The specicity of each protein is provided by the dierent properties of the
twenty side chains in gure 2.5 and their particular positions in the sequence.
In textbooks, it is customary to group them in small sets according to dierent
10
criteria in order to facilitate their learning. Classications devised on the basis
of the physical properties of these side chains may be sometimes overlapping
(e.g., tryptophan contains polar regions as well as an aromatic ring, which, in
turn, could be considered hydrophobic but is also capable of participating in,
say, - interactions). Therefore, for a clearer presentation, we have chosen here
to classify the residues according to the chemical groups contained in each side
chain and discuss their physical properties individually.
Let us enumerate then the categories in gure 2.5 and point out any special
remark regarding the residues in them:
Special residues:
Glycine is the smallest of all the amino acids: its side chain contains only a
hydrogen atom. So, since its -carbon has two hydrogens attached, glycine
is the only achiral natural amino acid. Its anity for water is mainly
determined by the peptide groups in the backbone; therefore, glycine is
hydrophilic.
Proline is the only residue whose side chain is covalently linked to the
backbone (the backbone is indicated in purple in gure 2.5), giving proline
unique structural properties that will be discussed later. Since its side
chain is entirely aliphatic, proline is hydrophobic.
Sulfur-containing residues:
Cysteine is a very important structural residue because, in a reaction cat-
alyzed by protein disulde isomerases (PDIs), it may form, with another
cysteine, a very stable covalent bond called disulde bond (see gure 2.4).
Curiously, all the L-amino acids are S-enantiomers according to the Cahn,
Ingold and Prelog rules [41] except for cysteine, which is R-. This is proba-
bly the reason that makes the D/L nomenclature favourite among protein
scientists [24]. Cysteine is a polar residue.
Figure 2.4: Disulde bond between two cysteine residues.
11
Figure 2.5: Side chains of the twenty amino acid residues encoded in the genetic
material of living beings. They have been classied according to the chemical groups
they contain. The rotameric degrees of freedom i are indicated with small arrows over
the bonds. The name of the heavy atoms and the numbering of the branches comply
with the IUPAC rules http://www.chem.qmul.ac.uk/iupac/AminoAcid/). Below the
molecular structure, the one letter code (green), the three letter code (red) and the
complete name (blue) of each amino acid may be found. In the case of proline, the
N and the -carbon have been included in the scheme, and the backbone bonds have
been coloured in purple. The titratable residues Asp, Glu, Lys and Arg have been
represented in their charged forms, which is the most common one in aqueous solvent
under physiological conditions. Histidine is shown in its neutral 2-tautomeric form.
12
Methionine is mostly aliphatic and, henceforth, apolar.
Aliphatic residues:
Alanine is the smallest chiral residue. This is the fundamental reason for
using alanine models, more than any other ones, in the computationally
demanding ab initio studies of peptides that are customarily performed in
quantum chemistry [4353]. It is hydrophobic, like all the residues in this
group.
Valine is one of the three -branched residues (i.e., those that have more
than one heavy atom attached to the -carbon, apart from the -carbon),
together with isoleucine and threonine. It is hydrophobic.
Leucine is hydrophobic.
Isoleucines -carbon constitutes an asymmetric centre and the only enan-
tiomer that occurs naturally is the one depicted in the gure. Only
isoleucine and threonine contain an asymmetric centre in their side chain.
Isoleucine is -branched and hydrophobic.
Acid residues:
Aspartic acid is normally charged under physiological conditions. Hence,
it is very hydrophilic.
Glutamic acid is just one CH
2
larger than aspartic acid. Their properties
are very similar.
Amides:
Asparagine contains a chemical group similar to the peptide bond. It is
polar and can act as a hydrogen bond donor or acceptor.
Glutamine is just one CH
2
larger than asparagine. Their physical prop-
erties are very similar.
Figure 2.6: Three forms of histidine found in proteins. (a) Neutral 2-tautomer. (b)
neutral 1-tautomer. (c) Charged form.
13
Basic residues:
Histidine is a special amino acid: in its neutral form, it may exist as two
dierent tautomers, called
1
and
2
, depending on which nitrogen has
an hydrogen atom attached to it. The
2
-tautomer has been found to
be slightly more stable in model dipeptides [54], although both forms are
found in proteins. Histidine can readily accept a proton and get a positive
charge, in fact, it is the only side chain with a pKa in the physiological
range, so non-negligible proportions of both the charged and uncharged
forms are typically present. Of course, histidine is hydrophilic.
Lysines side chain is formed by a rather long chain of CH
2
with an amino
group at its end, which is nearly always positively charged. Therefore,
lysine is very polar and hydrophilic.
Arginines properties are similar to those of lysine, although its terminal
guanidinium group is a stronger basis than the amino group and it may
also participate in hydrogen bonds as a donor.
Alcohols:
Serine is one of the smallest residues. It is polar due to the hydroxyl
group.
Threonines -carbon constitutes an asymmetric centre; the enantiomer
that occurs in living beings is the one shown in the gure. The physical
properties of threonine are very similar to those of serine.
Aromatic residues:
Phenylalanine is the smallest aromatic residue. Its benzyl side chain is
largely apolar and interacts unfavourably with water. It may also partic-
ipate in specic -stacking interactions with other aromatic groups.
Tyrosines properties are similar to those of phenylalanine, being only
slightly more polar due to the presence of a hydroxyl group.
Tryptophan, with 17 atoms in her side chain, is the largest residue. It is
mainly hydrophobic, although it contains a small polar region and it can
also participate in - interactions, like all the residues in this category.
After having introduced the building blocks of proteins, some qualifying
remarks about them are worth to be done: On one side, why amino acids
encoded in DNA codons are the ones in the list or why there are exactly twenty of
them are questions that are still subjects of controversy [55,56]. In fact, although
the side chains in gure 2.5 seem to confer enough versatility to proteins in
most cases, there are also rare exceptions in which other groups are needed to
perform a particular function. For example, the amino acid selenocysteine may
be incorporated into some proteins at an UGA codon (which normally indicates
14
Figure 2.7: Typical denition of internal coordinates. r21 is the bond length between
atoms 2 and 1. 321 is the bond angle formed by the bonds (2,1) and (3,2), it ranges
from 0 to 180
o
. Finally 4321 is the dihedral angle describing the rotation around the
bond (3,2); it is dened as the angle formed by the plane containing atoms 1, 2 and
3 and the plane containing atoms 2, 3 and 4; it ranges either from 180
o
to 180
o
or from 0
o
to 360
o
, depending on the convention; the positive sense of rotation for
4321 is the one indicated in the gure. Also note that the denition is symmetric
under a complete change in the order of the atoms, in such a way that, quite trivially,
r21 = r12 and 123 = 321, but also, not so trivially, 4321 = 1234. (See reference [51]
for further information.)
a stop in the transcription), or the amino acid pyrrolysine at an UAG codon
(which is also a stop indication in typical cases). In addition, the arginine side
chain may be post-translationally converted into citrulline by the action of a
family of enzymes called peptidylarginine deiminases (PADs).
On the other hand, the chemical (covalent) structure of the protein chain
may suer from more complex modications than just the inclusion of non-
standard amino acid residues: A myriad of organic molecules may be covalently
linked to specic points, the chain may be cleaved (cut), chemical groups may
be added or removed from the N- or C-termini, disulde bonds may be formed
between cysteines, and the side chains of the residues may undergo chemical
modications just like any other molecule [54]. The vast majority of these
changes either depend on the existence of some chemical agent external to the
protein, or are catalyzed by an enzyme.
In this work, our interest is in the folding of proteins. This problem, which
will be discussed in detail in the next section, is so huge and so dicult that,
in the opinion of the author, there is no point in worrying about details, such
as the ones mentioned in the two preceding paragraphs, before the big picture
is at least preliminarily understood. Therefore, when we talk about the folding
15
of proteins in what follows, we will be thinking about single polypeptide chains,
made up of L-amino acids, in water and without any other reagent present, with
the side chains chosen from the set in gure 2.5, and having underwent no post-
translational modications nor any chemical change on their groups. Finally,
although some simple modications, such as the formation of disulde bonds
or the trans cis isomerization of Xaa-Pro peptide bonds (see what follows),
could be more easily included in the rst approach to the problem, we shall also
leave them for a later stage.
Now, with this considerations, we have xed the covalent structure of our
molecule as well as the enantiomerism of the asymmetric centres it may contain.
This information is enough to specify the three-dimensional arrangement of the
atoms of small rigid molecules. However, long polymers and, particularly, pro-
teins, possess degrees of freedom (termed soft ) that require small amounts of
energy to be changed while drastically altering the relative positions of groups
and atoms. In a rst approximation, all bond lengths, bond angles and dihe-
dral angles describing rotations around triple, double and partial double bonds
(see gure 2.7) may be considered to be determined by the covalent struc-
ture. Whereas dihedral angles describing rotations around single bonds may be
considered to be variable and soft. The non-superimposable three-dimensional
arrangements of the molecule that correspond to dierent values of the soft
degrees of freedom are called conformations.
In proteins, some of these soft dihedrals are located at the side chains; they
are the
i
in gure 2.5 and, although they are important in the later stages of
a
b
Figure 2.8: Trans and cis conformations of the peptide plane. The bonds dening
the peptide bond dihedral angle are indicated in purple. (a) Trans conformation
( 180
o
). The most common one in proteins. (b) Cis conformation ( 0
o
).
Signicantly found only in Xaa-Pro bonds.
16
Figure 2.9: Numeration of the heavy atoms and the dihedrals angles describing ro-
tations around backbone bonds. In agreement with IUPAC recommendations (see
http://www.chem.qmul.ac.uk/iupac/AminoAcid/). The peptide planes are indicated
as green rectangles.
the folding process and must be taken into account in any ambitious model of
the system, their variation only alters the conformation locally. On the contrary,
a small change in the dihedral angles located at the backbone of the polypeptide
chain may drastically modify the relative position of many pairs of atoms and
they must be given special attention.
That is why, the special properties of the peptide bond, which is the basic
building block of the backbone, are very important to understand the conforma-
tional behaviour of proteins. These properties arise from the fact that there is
an electron pair delocalized between the CN and CO bonds (using the com-
mon chemical image of resonance), which provokes that neither bond is single
nor double, but partial double bonds that have a mixed character. In particular,
the partial double bond character of the peptide bond is the cause that the
six atoms in the green plane depicted in gures 2.3, 2.8 and 2.9 have a strong
tendency to be coplanar, forming the so-called peptide plane. This coplanarity
allows for only two dierent conformations: the one called trans (corresponding
to 180
o
), in which the -carbons lie at dierent sides of the line contain-
ing the CN bond; and the one called cis (corresponding to 0
o
), in which
they lie at the same side of that line (see gure 2.8).
Although the quantitative details are not completely elucidated yet and the
very protocol of protein structure determination by x-ray crystallography could
introduce spurious eects in the structures deposited in the PDB [57], it seems
clear that a great majority of the peptide bonds in proteins are in the trans
conformation. Indeed, a supercial look at the two forms in gure 2.8 suggests
that the steric clashes between substituents of consecutive -carbons will be
more severe in the cis case. When the second residue is a Proline, however, the
special structure of its side chain makes the probability of nding the cis con-
17
Figure 2.10: Original Ramachandran plot drawn by Ramachandran and Ramakrish-
nan in 1963 [60]. In dark-green, the fully allowed regions, calculated by letting the
atoms approach to the average clashing distance; in light-green, the partially allowed
regions, calculated by letting the atoms approach to the minimum clashing distance;
in white, the disallowed regions. Some points representing secondary structure ele-
ments are shown as red circles at the ideal (, )-positions in table 1: () -helix.
() -helix. (310) 310-helix. (a) Antiparallel -sheet. (p) Parallel -sheet. (ppII)
Polyproline II.
former signicantly higher: For Xaa-nonPro peptide bonds in native structures,
the trans form is more common than the cis one with approximately a 3000:1
proportion; while this ratio decreases to just 15:1 if the bond is Xaa-Pro [57].
In any case, due to the aforementioned partial double bond character of the
CN bond, the rotation barrier connecting the two states is estimated to be of
the order of 20 kcal/mol [58], which is about 40 times larger than the ther-
mal energy at physiological conditions, thus rendering the spontaneous trans
cis isomerization painfully slow. However, mother Nature makes use of every
possibility that she has at hand and, sometimes, there are a few peptide bonds
that must be cis in order for the protein to fold correctly or to function prop-
erly. Since all peptide bonds are synthesized trans at the ribosome [59], the
trans cis isomerization must be catalyzed by enzymes (called peptidylprolyl
isomerases (PPIs)) and, in the same spirit of the post-translational modica-
tions discussed before, this step may be taken into account in a later renement
of the theoretical models.
Therefore, we shall assume in what follows that all peptide bonds (even the
Xaa-Pro ones) are in the trans state and, henceforth, the conformation of the
protein will be essentially determined by the values of the and angles, which
describe the rotation around the two single bonds next to each -carbon (see
gure 2.9 for a denition of the dihedral angles associated to the backbone).
This assumption was introduced, as early as 1963, by Ramachandran and
18

-helix 57 47
3
10
-helix 49 26
-helix 57 70
polyproline II 79 149
parallel -sheet 119 113
antiparallel -sheet 139 135
Table 1: Ramachandran angles (in degrees) of some important secondary structure
elements in polypeptides. Data taken from reference [1].
Ramakrishnan [60] and the and coordinates are commonly named Ra-
machandran angles after the rst one of them. In their famous paper [60], they
additionally suppose that the bond lengths, bond angles and dihedral angles on
double and partial double bonds are xed and independent of and , they
dene a typical distance up to which a specic pair of atoms may approach and
also a minimum one (taken from statistical studies of structures) and they draw
the rst Ramachandran plot (see gure 2.10): A depiction of the regions in the
(, )-space that are energetically allowed or disallowed on the basis of the local
sterical clashes between atoms that are close to the -carbon.
One of the main advantages of this type of diagrams as thinking tools
lies in the fact that (always in the approximation that the non-Ramachandran
variables are xed) some very common repetitive structures found in proteins
may be ideally depicted as a single point in the plot. In fact, these special
conformations, which are regarded as the next level of protein organization
after the primary structure and are said to be elements of secondary structure,
may be characterized exactly like that, i.e., by asking that a certain number
of consecutive residues present the same values of the and angles. In the
book by Lesk [1], for example, one may found a table with the most common of
these repetitive patterns, together with the corresponding (, )-values taken
from statistical investigations of experimentally resolved protein structures (see
table 1).
However, the non-Ramachandran variables are not really constant, and the
elements of secondary structure do possess a certain degree of exibility. More-
over, the side chains may interact and exert dierent strains at dierent points of
the chain, which provokes that, in the end, the secondary structure elements gain
some stability by slightly altering their ideal Ramachandran angles. Therefore,
it is more appropriate to characterize them according to their hydrogen-bonding
pattern, which, in fact, is the feature that makes these structures prevalent, pro-
viding them with more energetic stability than other repetitive conformations
which are close in the Ramachandran plot.
The rst element of secondary structure that was found is the -helix. It
is a coil-like
16
structure, with 3.6 residues per turn, in which the carbonyl
16
Here, we use the word coil to refer to the twisted shape of a telephone wire, a corkscrew
19
Figure 2.11: The three helices found in protein native structures. (a) 310-helix, (b)
-helix, and (c) -helix. In the three cases, the helices shown are 11-residues long. In
the standing views (above), the hydrogen bonds are depicted as green dotted lines and
the distance and number of turns spanned by 10 residues are indicated at the right of
the structures. Whereas in the standing views, the side chains and -hydrogens have
been removed for visual convenience, in the zenithal views (below), they are included.
group (C=O) of each i-th residue forms a hydrogen bond with the amino group
(NH) of the residue i +4 (see gure 2.11b). According to a common notation,
in which x
y
designates a helix with x residues per turn and y atoms in the ring
closed by the hydrogen bond [61], the -helix is also called 3.6
13
-helix.
She was theoretically proposed in 1951 by Pauling, Corey and Branson [62],
who used precise information about the geometry of the non-Ramachandran
variables, taken from crystallographic studies of small molecules, to nd the
structures compatible with the additional constraints that: (i) the peptide bond
is planar, and (ii) every carbonyl and amino group participates in a hydrogen
bond.
The experimental conrmation came from Max Perutz, who, together with
Kendrew and Bragg, had proposed in 1950 (one year before Paulings paper) a
series of helices with an integer number of residues per turn [61] that are not
so commonly found in native structures of proteins (see however, the discussion
or the solenoid of an electromagnet. Although this is common English usage, the same word
occurs frequently in protein science to designate dierent (and sometimes opposed) concepts.
For example, a much used ideal model of the denatured state of proteins is termed random
coil, and a popular statistical description of helix formation is called helix-coil theory.
20
about the 3
10
-helix below). Perutz read Pauling, Corey and Bransons paper
one Saturday morning [63] in spring 1951 and realized immediately that their
helix looked very well: free of strain and with all donor and acceptor groups
participating in hydrogen bonds. So he rushed to the laboratory and put a
sample of horse hair (rich in keratin, a protein that contains -helices) in the
x-ray beam, knowing that, according to diraction theory, the regular repeat of
the spiral staircase steps in Paulings structure should give rise to a strong x-
ray reection of 1.5

A spacing from planes perpendicular to the ber axis. The
result of the experiment was positive
17
and, in the last years of the 50s, Perutz
and Kendrew saw again the same signal in myoglobin and hemoglobin, when
they resolved, for the rst time in history, the structure of these proteins [64,65].
However, despite its being, by far, the most common, the -helix is not
the only coil-like structure that can be found in native proteins [6668]. If
the hydrogen bonds are formed between the carbonyl group (C=O) of each
i-th residue with the amino group (NH) of the residue i + 3, one obtains a
3
10
-helix, which is more tightly wound and, therefore, longer than an -helix
of the same chain length (see gure 2.11a). The 3
10
-helix is the fourth most
common conformation for a single residue after the -helix, -sheet and reverse
turn
18
[67] but, remarkably, due to its having an integer number of residues
per turn, it seemed more natural to scientists with crystallographic background
and was theoretically proposed before the -helix [61, 69]. On the other hand,
if the hydrogen bonds are formed between the carbonyl group (C=O) of each
i-th residue with the amino group (NH) of the residue i + 5, one obtains a
-helix (or 4.4
16
-helix), which is wider and shorter than an -helix of the same
length (see gure 2.11c). It was originally proposed by Low and Baybutt in
1952 [70], and, although the exact fraction of each type of helix in protein
native structures depends up to a considerable extent on their denition (in
terms of Ramachandran angles, interatomic distances, energy of the hydrogen
bonds, etc.), it seems clear that the -helix is the less common of the three [66].
Now, it is true that, in addition to these helices that have been experimentally
conrmed, some others have been proposed. For example, in the same work in
which Pauling, Corey and Branson introduce the -helix [62], they also describe
another candidate: the -helix (or 5.1
17
-helix). Finally, Donohue performed, in
1953, a systematic study of all possible helices and, in addition to the ones
already mentioned, he proposed a 2.2
7
- and a 4.3
14
-helix [71]. None of them
has been detected in resolved native proteins.
Among the secondary structure elements of proteins, not all regular local
patterns are helices: there exist also a variety of repetitive conformations that
do not contain strong intra-chain hydrogen bonds and that are less curled than
the structures in gure 2.11. For example, the polyproline II [7274], which is
17
Linus Pauling was awarded the Nobel prize in chemistry in 1954 for his research into the
nature of the chemical bond and its application to the elucidation of the structure of complex
substances, and Max Perutz shared it with John Kendrew in 1962 for their studies of the
structures of globular proteins.
18
A conformation that some residues in proteins adopt when an acute turn in the chain is
needed.
21
Figure 2.12: -sheets in the pure (a) antiparallel, and (b) parallel versions. On
the left, the top view is shown, with the side chains and the -hydrogens omitted for
visual convenience and the directions of the strands indicated as yellow arrows. The
hydrogen bonds are represented as green dotted lines. On the right, the side view of
the sheets is depicted. In this case, the side chains and the -hydrogens are included.
thought to be important in the unfolded state of proteins, and, principally, the
family of the -sheets, which are, together with the -helices, the most recog-
nizable secondary structure elements in native states of polypeptide chains
19
.
The -sheets are rather plane structures that are typically formed by sev-
eral individual -strands, which align themselves to form stabilizing inter-chain
hydrogen bonds with their neighbours. Two pure arrangements of these single
threads may be found: the antiparallel -sheets (see gure 2.12a), in which the
strands run in opposite directions (read from the amino to the carboxyl termi-
nus); and the parallel -sheets (see gure 2.12b), in which the strands run in
the same direction. In both cases, the side chains of neighbouring residues in
contiguous strands branch out to the same side of the sheet and may interact.
Of course, mixed parallel-antiparallel sheets can also be found.
The next level of protein organization, produced by the assembly of the ele-
ments of secondary structure, and also of the chain segments that are devoid of
regularity, into a well dened three-dimensional shape, is called tertiary struc-
ture. The protein folding problem (omitting relevant qualications that have
been partially made and that will be recalled and made more explicit in what
follows) may be said to be the attempt to predict the secondary and the tertiary
19
It is probably more correct to dene the secondary structure as the conformational
repetition in consecutive residues and, from this point of view, to consider the -strand as the
proper element of secondary structure. In this sense, the assembly of -strands, the -sheet,
together with some other simple motifs such as the coiled coils made up of two helices, the silk
broin (made up of stacked -sheets) or collagen (three coiled threads of a repetitive structure
similar to polyproline II), may be said to be elements of super-secondary structure, somewhat
in between the local secondary structure and the global and more complex tertiary structure
(see below).
22
structure from the primary structure, and it will be discussed in the next section.
The quaternary structure, which refers to the way in which protein monomers
associate to form more complex systems made up of more than one individual
chain (such as the ones in gure 1.1), will not be explored in this work.
3 The protein folding problem
As we have seen in the previous section and can visually check in gure 1.1, the
biologically functional native structure of a protein
20
is highly complex. What
Kendrew saw in one of the rst proteins ever resolved is essentially true for most
of them [75]:
The most striking features of the molecule were its irregularity and
its total lack of symmetry.
Now, since these polypeptide chains are synthesized linearly in the ribosome
(i.e., they are not manufactured in the folded conformation), in principle, one
may imagine that some specic cellular machinery could be the responsible of
the complicated process of folding and, in such a case, the prediction of the
native structure could be a daunting task. However, in a series of experiments
in the 50s, Christian B. Annsen ruled out this scenario and was awarded the
Nobel prize for it [76].
The most famous and illuminating experiment that he and his group per-
formed is the refolding of bovine pancreatic ribonuclease (see the scheme in
gure 3.1 for reference). They took this protein, which is 124 residues long and
has all her eight cysteines forming four disulde bonds, and added, in a rst
step, some reducing agent to cleave them. Then, they added urea up to a con-
centration of 8 M. This substance is known for being a strong denaturing agent
(an unfolder) and produced a scrambled form of the protein which is much
less compact than the native structure and has no enzymatic activity. From
this scrambled state, they took two dierent experimental paths: in the positive
one, they removed the urea rst and then added some oxidizing agent to reform
the disulde bonds; whereas, in the negative path, they poured the oxidizing
agent rst and removed the urea in a second step.
The resultant species in the two paths are very dierent. If one removes
the urea rst and then promotes the formation of disulde bonds, an homoge-
neous sample is obtained that is practically indistinguishable from the starting
native protein and that keeps full biological activity. The ribonuclease has been
unscrambled ! However, if one takes the negative path and let the cysteines
form disulde bonds before removing the denaturing agent, a mixture of prod-
ucts is obtained containing many or all of the possible 105 isomeric disulde
20
Most native states of proteins are exible and are comprised not of only one conformation
but of a set of closely related structures. This exibility is essential if they need to perform
any biological function. However, to economize words, we will use in what follows the terms
native state, native conformation and native structure as interchangeable.
23
Figure 3.1: Scheme of the refolding of the bovine pancreatic ribonuclease by Annsen.
The black arrows indicate fundamental steps of the experiment and the red labels next
to them designate: (R) addition of reducing agent (cleavage of the disulde bonds),
(O) addition of oxidizing agent (reformation of the disulde bonds), (U) and (U)
increase of the urea concentration up to 8 M and decrease to 0 M respectively. The
conformation of the backbone of the protein is schematically depicted by a black line,
the cysteines are shown as small yellow circles and the disulde bonds as line segments
connecting them. The dierent states are labelled: (a) starting native enzyme with
full activity, (b) non-disulde bonded, folded form, (c) representant of the ensemble
of inactive scrambled ribonuclease, (d.1) non-disulde bonded, folded form, (e.1)
refolded ribonuclease indistinguishable from (a), (d.2) representant of the ensemble of
the scrambled, disulde bonded form, and, nally, (e.2) representant of the mixture
of the 105 isomeric disulde bonded forms.
bonded forms
21
. This mixture is essentially inactive, having approximately 1%
the activity of the native enzyme.
One of the most clear conclusions that are commonly drawn from this ex-
periment is that all the information needed to reach the native state is encoded
in the sequence of amino acids. This important statement, which has stood the
test of time [24, 77], allows to isolate the system under study (both theoreti-
cally and experimentally) and sharply denes the protein folding problem, i.e.,
the prediction of the three-dimensional native structure of proteins from their
amino acid sequence (and the laws of physics).
It is true that we nowadays know of the existence of the so-called molecular
chaperones (see, for example, the GroEL-GroES complex in gure 1.1c), which
help the proteins fold in the cellular milieu [7882]. However, according to the
most accepted view [24,77], these molecular assistants do not add any structural
information to the process. Some of them simply prevent accidents related to
the cellular crowding from happening. Indeed, in the cytoplasm there is not
much room: inside a typical bacterium, for example, the total macromolecular
21
Take an arbitrary cysteine: she can bond to any one of the other seven. From the
remaining six, take another one at random: she can bond to ve dierent partners. Take the
reasoning to its nal and we have 7 5 3 = 105 dierent possibilities.
24
concentration is approximately 350 mg/ml, whereas a typical protein crystal
may contain about 600 mg/ml [77]. This crowding may hinder the correct fold-
ing of proteins, since partially folded states (of chains that are either free in
the cytoplasm or being synthesized in proximate ribosomes) have more sticky
hydrophobic surface exposed than the native state, opening the door to aggre-
gation. In order to avoid it, some chaperonins
22
are in charge of providing a
shelter in which the proteins can fold alone. Yet another pitfall is that, when
the polypeptide chain is being synthesized in the ribosome, it may start to fold
incorrectly and get trapped in a non-functional conformation separated by a
high energetic barrier from the native state. Again, there exist some chaper-
ones that bind to the nascent chain to prevent this from happening. As we have
already pointed out, all this assistance to fold is seen as lacking new structural
information and meant only to avoid traps which are not present in vitro. It
seems as if molecular chaperones aim is to make proteins believe that they are
not in a messy cell but in Annsens test tube!
The possibility that this state of aairs opens, the prediction of the three-
dimensional native structure of proteins from the only knowledge of the amino
acid sequence, is often referred to as the second half of the genetic code [83,84].
The reason for such a vehement statement lays in the fact that not all proteins
are accessible to the experimental methods of structure resolution (mostly x-
ray crystallography and NMR [54, 85]) and, for those that can be studied, the
process is long and expensive, thus making the databases of known structures
grow much more slowly than the databases of known sequences (see gure 1.2
and the related discussion in section 1). To solve the second half of the genetic
code and bridge this gap is the main objective of the hot scientic eld of
protein structure prediction [24, 86, 87].
The path that takes to this goal may be walked in two dierent ways [88,89]:
Either at a fast pragmatic pace, using whatever information we have avail-
able, increasingly rening the everything-goes prediction procedures by exten-
sive trial-and-error tests and without any need of knowing the details of the
physical processes that take place; or at a slow thoughtful pace, starting from
rst principles and seeking to arrive to the native structure using the same
means that Nature uses: the laws of physics.
The dierent protocols belonging to the fast pragmatic way are commonly
termed knowledge-based, since they take prot from the already resolved struc-
tures that are deposited in the PDB [32] or any other empirical information that
may be statistically extracted from databases of experimental data. There are
basically three pure forms of knowledge-based strategies [90]:
Homology modeling (also called comparative modeling) [85, 91] is based
on the observation that proteins with similar sequences frequently share
similar structures [92]. Following this approach, either the whole sequence
of the protein that we want to model (the target ) or some segments of
it are aligned to a sequence of known structure (the template). Then, if
some reasonable measure of the sequence similarity [93,94] is high enough,
22
A particular subset of the set of molecular chaperones.
25
the structure of the template is proposed to be the one adopted by the
target in the region analysed. Using this strategy, one typically needs
more than 50% sequence identity between target and template to achieve
high accuracy, and the errors increase rapidly below 30% [87]. Therefore,
comparative modeling cannot be used with all sequences, since some recent
estimates indicate that 40% of genes in newly sequenced genomes do not
have signicant sequence homology to proteins of known structure [95].
Fold recognition (or threading) [86, 96] is based on the fact that, increas-
ingly, new structures deposited in the PDB turn out to fold in shapes that
have been seen before, even though conventional sequence searches fail to
detect the relationship [97]. Hence, when faced to a sequence that shares
low identity with the ones in the PDB, the threading user tries to t it
in each one of the structures in the databases of known folds, selecting
the best choices with the help of some scoring function (which may be
physics-based or not). Again, fold recognition methods are not awless
and, according to various benchmarks, they fail to select the correct fold
from the databases for 50% of the cases [86]. Moreover, the fold space
is not completely known so, if faced with a novel fold, threading strate-
gies are useless and they may even give false positives. Modern studies
estimate that approximately one third of known protein sequences must
present folds that have never been seen [98].
New fold (or de novo prediction) methods [99, 100] must be used when
the protein under study has low sequence identity with known structures
and fold recognition strategies fail to t it in a known fold (because of
any of the two reasons discussed in the previous point). The specic
strategies used in new fold methods are very heterogeneous, ranging from
well-established secondary structure prediction tools or sequence-based
identication of sets of possible conformations for short fragments of chains
to numerical search methods, such as molecular dynamics, Monte Carlo
or genetic algorithms [97].
These knowledge-based strategies may be arbitrarily combined into mixed
protocols, and, although the frontiers between them may be sometimes blurry
[35], it is clear that the more information available the easier to predict the
native structure (see gure 3.2). So that the three types of methods described
above turn out to be written in increasing order of diculty and they essentially
coincide with the competing categories of the CASP experiment
23
[35, 97]. In
this important meeting, held every two years and whose initials loosely stand
for Critical Assessment of techniques for protein Structure Prediction, exper-
imental structural biologists are asked to release the amino acid sequences of
proteins (the CASP targets) whose structures are likely to be resolved before
the contest starts. Then, the prediction community gets on stage and their
23
Since CASP1, people has drifted towards knowledge-based methods and, nowadays, very
few groups use pure ab initio approaches [101].
26
Figure 3.2: Schematic classication of protein structure prediction methods.
members submit the proposed structures (the models), which may be found
using any chosen method. Finally, a committee of assessors, critically evaluate
the predictions, and the results are published, together with some contributions
by the best predictors, in a special issue of the journal Proteins.
Precisely, in the latest CASP meetings, the expected ordering (based on the
available experimental information) of the three aforementioned categories of
protein structure prediction has been observed to translate into dierent quali-
ties of the proposed models (see gure 3.2). Hence, while comparative modeling
with high sequence similarity has proved to be the most reliable method to
predict the native conformation of proteins (with an accuracy comparable to
low-resolution, experimentally determined structures) [87, 102], de novo model-
ing has been shown to remain still unreliable [35, 88] (although a special remark
should be made about the increasingly good results that David Baker and his
group are achieving in this eld with their program Rosetta [103, 104]).
Opposed to these knowledge-based approaches, the computer simulation of
the real physical process of protein folding
24
without using any empirical in-
formation and starting from rst principles could be termed ab initio protein
folding or ab initio protein structure prediction depending whether the emphasis
is laid on the process or on the goal.
Again, the frontier between de novo modeling and ab initio protein folding is
not sharply dened and some confusion might arise between the two terms. For
example, the potential energy functions included in most empirical force elds
such as CHARMM [106, 107] contain parameters extracted from experimental
data, while molecular dynamics attempts to fold proteins using these force elds
will be considered by most people (including the author) to belong to the ab
initio category. As always, the limit cases are clearer, and Bakers Rosetta
[103, 104], which uses statistical data taken from the PDB to bias the secondary
24
Not a new idea [105].
27
structure conformational search, may be classied, without any doubt, as a
de novo protocol; while, say, a (nowadays unfeasible) simulation of the folding
process using quantum mechanics, would be deep in the ab initio region. The
situation is further complicated due to the fact that score functions which are
based (up to dierent degrees) on physical principles, are commonly used in
conjunction with knowledge-based strategies to prune or rene the candidate
models [87, 108]. In the end, the classication of the strategies for nding the
native structure of proteins is rather continuous with wriggly, blurry frontiers
(see gure 3.2).
It is clear that, despite their obvious practical advantages and the superior
results when compared to pure ab initio approaches [86], any knowledge-based
features included in the prediction protocols render the assembly mechanisms
physically meaningless [109]. If we want to know the real details of protein
folding as it happens, for example, to properly study and attack diseases that
are related to protein misfolding and aggregation [3], we must resort to pure ab
initio strategies. In addition, ab initio folding does not require any experimental
information about the protein, apart from its amino acid sequence. Therefore, as
new fold strategies, it has a wider range of applicability than homology modeling
and fold recognition, and, in contrast with the largely system-oriented protocols
developed in the context of knowledge-based methods, most theoretical and
computational improvements made while trying to ab initio fold proteins will
be perfectly applicable to other macromolecules.
The feedback between strategies is also an important point to stress. Apart
from the obvious fact that the knowledge of the whole folding process includes
the capability of predicting the native conformation, and the problem of pro-
tein structure prediction would be automatically solved if ab initio folding were
achieved, the design of accurate energy functions, which is a central part of ab
initio strategies (see the next section), would also be very helpful to improve
knowledge-based methods that make use of them (such as Rosetta [108]) or to
prune and rene the candidate models on a second stage [87]. Additionally,
to assign the correct conformation to those chain segments that are devoid of
secondary structure (the problem known as loop modeling), may be considered
as a mini protein folding problem [87], and the understanding of the physical
behaviour of polypeptide chains would also include a solution to this issue. In
the light of all these sweet promises, the long ab initio path to study protein
folding constitutes an exciting eld of present research.
Before we delve deeper in the details, let us dene clearly the playeld in
which the match shall take place: Although some details of the protein folding
process in vivo are under discussion [110] and many cellular processes are in-
volved in helping and checking the arrival to the correct native structure [81];
although some proteins have been shown to fold cotranslationally [111] (i.e.,
during their synthesis in the ribosome) and many of them are known to be as-
sisted by molecular chaperones (see the discussion above and references therein);
although some proteins contain cis proline peptide bonds or disulde bonded
cysteines in their native structure, and must be in the presence of the respec-
tive isomerases in order to fold in a reasonable time (see section 2); although
28
some residues may be post-translationally changed into side chains that are not
included in the standard twenty that are depicted in gure 2.5; and, although
some non-peptide molecules may be covalently attached to the protein chain or
some cofactor or ion may be needed to reach the native structure, we agree with
the words by Alan Fersht [112]:
We can assume that what we learn about the mechanism of folding
of small, fast-folding proteins in vitro will apply to their folding in
vivo and, to a large extent, to the folding of individual domains in
larger proteins.
and decided to study those processes that do not include any of the afore-
mentioned complications but that may be rightfully considered as intimately
related to the process of folding in the cellular milieu and regarded as a rst
step on top of which to build a more detailed theory.
Henceforth, we dene the restricted protein folding problem, as the full de-
scription of the physical behaviour, in aqueous solvent and physiological condi-
tions, and (consequently) the prediction of the native structure, of completely
synthesized proteins, made up just of the twenty genetically encoded amino acids
in gure 2.5, without any molecule covalently attached to them, and needless
of molecular chaperones, cofactors, ions, disulde bonds or cis proline peptide
bonds in order to fold properly.
Explicitly mentioned or tacitly assumed, it is this restricted version of the
problem the one that is most amenable to physics-based methods and the one
that is more commonly tackled in the literature.
4 Folding mechanisms and energy functions
After having drawn the boundaries of the problem, we should ask the million-
dollar question associated to it: How does a protein fold into its functional
native structure? In fact, since this feat is typically achieved in a very short
time, we must add: How does a protein fold so fast? This is the question about
the mechanisms of protein folding, and, ever since Annsens experiments, it
has been asked once and again and only partially answered [76, 89, 109, 113, 114].
In order to dene the theoretical framework that is relevant for the descrip-
tion of the folding process and also to introduce the language that is typically
used in the discussions about its mechanisms, let us start with a brief reminder
of some important statistical mechanics relations. To do this, we will follow the
main ideas in reference [115], although the notation and the assumptions regard-
ing the form of the potential energy, as well as some other minor details, will
be dierent. The presentation will be axiomatic and we will restrict ourselves
to the situation in which the macroscopic parameters, such as the temperature
T or, say, the number of water molecules N
w
, do not change. In these condi-
tions, that allow us to drop any multiplicative terms in the partition functions
or the probabilities, and also to forget any additive terms to the energies, we
29
can only focus on the conformational preferences of the system (if, for exam-
ple, the temperature changed, the neglected terms would be relevant and the
expressions that one would need to use would be dierent). For further details
or for the more typical point of view in physics, in which the stress is placed in
the variation of the macroscopic thermodynamical parameters, see, for example,
reference [116].
The system which we will talk about is the one dened by the restricted
protein folding problem in the previous section, i.e., one protein surrounded by
N
w
water molecules
25
; however, one must have in mind that all the subsequent
reasoning and the derived expressions are exactly the same for a dilute aqueous
solution of a macroscopic number of non-interacting proteins.
Now, if classical mechanics is assumed to be obeyed by our system
26
, then
each microscopic state is completely specied by the Euclidean
27
coordinates
and momenta of the atoms that belong to the protein (denoted by x

and

,
respectively, with = 1, . . . , N) and those belonging to the water molecules
(denoted by X
m
and
m
, with m = N + 1, . . . , N + N
w
). The whole set of
microscopic states shall be called phase space and denoted by
w
, explicitly
indicating that it is formed as the direct product of the protein phase space
and the water molecules one
w
.
The central physical object that determines the time behaviour of the system
is the Hamiltonian (or energy) function
H(x

, X
m
,

,
m
) =

2M

2
m
2M
m
+V (x

, X
m
) , (4.1)
where M

and M
m
denote the atomic masses and V (x

, X
m
) is the potential
energy.
After equilibrium has been attained at temperature T, the microscopic de-
tails about the time trajectories can be forgot and the average behaviour can be
described by the laws of statistical mechanics. In the canonical ensemble, the
partition function [116] of the system, which is the basic object from which the
rest of relevant thermodynamical quantities may be extracted, is given by
Z =
1
h
N+Nw
N
w
!

w
exp

H(x

, X
m
,

,
m
)

dx

dX
m
d

d
m
,
(4.2)
25
At this point of the discussion, the possible presence of non-zero ionic strength is consid-
ered to be a secondary issue.
26
Although non-relativistic quantum mechanics may be considered to be a much more
precise theory to study the problem, the computer simulation of the dynamics of a system with
so many particles using a quantum mechanical description lies far in the future. Nevertheless,
this more fundamental theory can be used to design better classical potential energy functions
(which is one of the main long-term goals of the research performed in our group).
27
Sometimes, the term Cartesian is used instead of Euclidean. Here, we prefer to use the
latter since it additionally implies the existence of a mass metric tensor that is proportional to
the identity matrix, whereas the Cartesian label only asks the n-tuples in the set of coordinates
to be bijective with the abstract points of the space [117].
30
where h is Plancks constant, we adhere to the standard notation := 1/RT
(per-mole energy units are used all throughout this work, so R is preferred over
k
B
) and N
w
! is a combinatorial number that accounts for the quantum indistin-
guishability of the N
w
water molecules. Additionally, as we have anticipated,
the multiplicative factor outside the integral sign is a constant that divides out
for any observable averages and represents just a change of reference in the
Helmholtz free energy. Therefore, we will drop it from the previous expression
and the notation Z will be kept for convenience.
Next, since the principal interest lies on the conformational behaviour of
the polypeptide chain, seeking to develop clearer images and, if possible, re-
duce the computational demands, water coordinates and momenta are custom-
arily averaged (or integrated) out [115, 118], leaving an eective Hamiltonian
H
e
(x

; T) that depends only on the protein degrees of freedom and on the

temperature T, and whose potential energy (denoted by W(x

; T)) is called
potential of mean force or eective potential energy.
This eective Hamiltonian may be either empirically designed from scratch
(which is the common practice in the classical force elds typically used to per-
form molecular dynamics simulations [106, 107, 119128]) or obtained from the
more fundamental, original Hamiltonian H(x

, X
m
,

,
m
) actually perform-
ing the averaging out process. In statistical mechanics, the theoretical steps
that must be followed if one chooses this second option are very straightforward
(at least formally):
The integration over the water momenta
m
in equation (4.2) yields a T-
dependent factor that includes the masses M
m
and that shall be dropped by
the same considerations stated above. On the other hand, the integration of the
water coordinates X
m
is not so trivial, and, except in the case of very simple
potentials, it can only be performed formally. To do this, we dene the potential
of mean force or eective potential energy by
W(x

; T) := RT ln

exp

V (x

, X
m
)

dX
m

, (4.3)
and simply rewrite Z as
Z =

exp

H
e
(x

; T)

dx

, (4.4)
with the eective Hamiltonian being
H
e
(x

; T) =

2M

+W(x

; T) . (4.5)
At this point, the protein momenta

may also be averaged out from the

expressions. This choice, which is very commonly taken in the literature, largely
simplies the discussion about the mechanisms of protein folding and the im-
ages and metaphors typically used in the eld. However, to perform this av-
erage is not completely harmless, since it brings up a number of technical and
31
interpretation-related diculties mostly due to the fact that the marginal prob-
ability density in the x

-space in equation (4.7) is not invariant under a change

of coordinates
28
(see appendix A and reference [52] for further details).
Bearing this in mind, the integration over

produces a new T-dependent

factor, which is dropped as usual, and yields a new form of the partition function,
which is the one that will be used from now on in this section:
Z =

exp

W(x

; T)

dx

, (4.6)
where now denotes the positions part of the protein phase space .
Some remarks may be done at this point: On the one hand, if one further
assumes that the original potential energy V (x

, X
m
) separates as a sum of
intra-protein, intra-water and water-protein interaction terms, the eective po-
tential energy W(x

; T) in the equations above may be written as a sum of two

parts: a vacuum intra-protein energy and an eective solvation energy [115].
Nevertheless, this simplication is neither justied a priori, nor necessary for
the subsequent reasoning about the mechanisms of protein folding; so it will not
be assumed herein.
On the other hand, the (in general, non-trivial) dependence of W(x

; T) on
the temperature T (see equation (4.3)) and the associated fact that it contains
the entropy of the water molecules, justies its alternative denomination of
internal or eective free energy, and also the suggestive notation F(x

) :=
W(x

; T) used in some works [130]. Here, however, we prefer to save the name
free energy for the one that contains some amount of protein conformational
entropy and that may be assigned to nite subsets (states) of the conformational
space of the chain (see equation (4.10) and the discussion below).
Finally, we will stick to the notational practice of dropping (but remember-
ing) the temperature T from W and H
e
. This is consistent with the situation of
constant T that we wish to investigate and also very natural and common in the
literature. In fact, most Hamiltonian functions (and their respective potentials)
that are considered to be fundamental actually come from the averaging out of
degrees of freedom more microscopical than the ones regarded as relevant, and,
as a result, the coupling constants contained in them are not really constant,
but dependent on the temperature T.
Now, from the probability density function (PDF) in the protein conforma-
tional space , given by,
p(x

) =
exp

W(x

Z
, (4.7)
we can tell that W(x

) completely determines the conformational preferences

of the polypeptide chain in the thermodynamic equilibrium as a function of
each point of . On the opposite extreme of the details scale, we may choose
to describe the macroscopic state of the system as a whole (like it is normally
28
Note that, if the momenta are kept in the integration measure, any canonical transfor-
mation leaves the probability density invariant, since its Jacobian determinant is unity [129].
32
done in physics [116]) and dene, for example, the Helmholtz free energy as
F := RT lnZ, where no trace of the microscopic details of the system remains.
In protein science, it is also common practice to take a point of view some-
what in the middle of these two limit descriptions, and dene states that are
neither single points of nor the whole set, but nite subsets
i
compris-
ing many dierent conformations that are related in some sense. These states
must be precisely specied in order to be of any use, and they must fulll some
reasonable conditions, the most important of which is that they must be mutu-
ally exclusive, so that
i

j
= , i ,= j (i.e., no point can lie in two dierent
states at the same time).
Since the two most relevant conceptual constructions used to think about
protein folding, the native (^) and the unfolded (|) states, as well as a great
part of the language used to talk about protein stability, t in this formalism,
we will now introduce the basic equations associated to it.
To begin with, one can dene the partition function of a certain state
i
as
Z
i
:=

i
exp

W(x

; T)

dx

, (4.8)
so that the probability of
i
be given by
P
i
:=
Z
i
Z
. (4.9)
The Helmholtz free energy F
i
of this state is
F
i
:= RT lnZ
i
, (4.10)
and the following relation for the free energy dierences is satised:
F
ij
= F
j
F
i
= RT ln
Z
j
Z
i
= RT ln
P
j
P
i
= RT ln
[j]
[i]
= RT lnK
ij
,
(4.11)
where [i] denotes the concentration (in chemical jargon) of the species i, and
K
ij
is the reaction constant (using again images borrowed from chemistry) of
the i j equilibrium. It is precisely this dependence on the concentrations,
together with the approximate equivalence between F and G at physio-
logical conditions (where the term PV is negligible [115]), that renders equa-
tion (4.11) very useful and ultimately justies this point of view based on states,
since it relates the quantity that describes protein stability and may be estimated
theoretically (the folding free energy at constant temperature and constant pres-
sure G
fold
:= G
N
G
U
) with the observables that are commonly measured
in the laboratory (the concentrations [^] and [|] of the native and unfolded
states) [24, 54, 131].
The next step to develop this state-centred formalism is to dene the mi-
croscopic PDF in
i
as the original one in equation (4.7) conditioned to the
knowledge that the conformation x

lies in
i
:
33
p
i
(x

) := p(x

[ x

i
) =
p(x

)
P
i
=
exp

W(x

Z
i
. (4.12)
Now, using this probability measure in
i
, we may calculate the internal
energy U
i
as the average potential energy in this state:
U
i
:= W)
i
=

i
W(x

)p
i
(x

) dx

, (4.13)
and also dene the entropy of
i
as
S
i
:= R

i
p
i
(x

) lnp
i
(x

) dx

. (4.14)
Finally, ending our statistical mechanics reminder, one can show that the
natural thermodynamic relation among the dierent state functions is recovered:
F
ij
= U
ij
TS
ij
G
ij
= H
ij
TS
ij
, (4.15)
where H is the enthalpy, whose dierences H
ij
may be approximated by U
ij
neglecting the term PV again.
Retaking the discussion about the mechanisms of protein folding, we see
(again) in equation (4.7) that the potential of mean force W(x

) completely
determines the conformational preferences of the polypeptide chain in the ther-
modynamic equilibrium. Nevertheless, it is often useful to investigate also
the underlying microscopic dynamics. The eective potential energy W(x

)
in equation (4.3) has been simply obtained in the previous paragraphs using
the tools of statistical mechanics; the dynamical averaging out of the solvent
degrees of freedom in order to describe the time evolution of the protein subsys-
tem, on the other hand, is a much more complicated (and certainly dierent)
task [132136]. However, if the relaxation of the solvent is fast compared to the
motion of the polypeptide chain, the function W(x

) turns out to be precisely

the eective dynamical potential energy that determines the microscopic time
evolution of the protein degrees of freedom [133]. Although this condition could
be very dicult to check for real cases and it has only been studied in simpli-
ed model systems [132,134,136], molecular dynamics simulations with classical
force elds and explicit water molecules suggest that it may be approximately
fullled [133, 137, 138]. For the sake of brevity, in the discussion that follows, we
will assume that this fast-relaxation actually occurs, so that, when reasoning
about the graphical representations (commonly termed energy landscapes) of
the eective potential energy W(x

), we are entitled to switch back and forth

from dynamical to statistical concepts.
Now, just after noting that F(x

) is the central physical object needed to

tackle the elucidation of the folding mechanisms, we realize that the number of
degrees of freedom N in an average-length polypeptide chain is large enough
for the size of the conformational space (which is exponential on N) to be
astronomically astronomical. This fact was, for years, regarded as a problem,
and is normally called Levinthals paradox [139]. Although it belongs to the set of
34
paradoxes that (like Zenos or Epimenides) are called so without actually being
problematic
29
, thinking about it and using the language and the images related
to it have dominated the views on folding mechanisms for a long time [89]. The
paradox itself was rst stated in a talk entitled How to fold graciously given
by Cyrus Levinthal in 1969 [140] and it essentially says that, if, in the course of
folding, a protein is required to sample all possible conformations (a hypothesis
that ignores completely the laws of dynamics and statistical mechanics) and the
conformation of a given residue is independent of the conformations of the rest
(which is also false), then the protein will never fold to its native structure.
For example, let us assume that each one of the 124 residues in Annsens
ribonuclease (see section 3) can take up any of the six dierent discrete backbone
conformations in table 1 (side chain degrees of freedom are not relevant in
this qualitative discussion, since they only aect the structure locally). This
makes a total of 6
124
10
96
dierent conformations for the chain. If they
were visited in the shortest possible time (say, 10
12
s, approximately the
time required for a single molecular vibration [141]), the protein would need
about 10
76
years to sample the whole conformational space. Of course, this
argument is just a reductio ad absurdum proof (since proteins do fold!) of the
a priori evident statement that protein folding cannot be a completely random
trial-and-error process (i.e., a random walk in conformational space). The golf-
course energy landscape in gure 4.1a represents this non-realistic, paradoxical
situation: the point describing the conformation of the chain wanders aimlessly
on the enormous denatured plateau until it suddenly nds the native well by
pure chance.
Levinthal himself argued that a solution to his paradox could be that the
folding process occurs along well-dened pathways that take every protein, like
an ordered column of ants, from the unfolded state to the native structure,
visiting partially folded intermediates en route [114, 142]. The ant-trail energy
landscape in gure 4.1b is a graphical depiction of the pathway image.
This view, which is typically referred to as the old view of folding [133, 139,
143], is largely inuenced by the situation in simple chemical reactions, where
the barriers surrounding the minimum energy paths that connect the dierent
local minima are very steep compared to RT, and the dynamical trajectories
are, consequently, well dened. In protein folding, however, due to the fact
that the principal driving forces are much weaker than those relevant for chem-
ical reactions and comparable to RT, short-lived transient interactions may
form randomly among dierent residues in the chain and the system describes
stochastic trajectories that are never the same. Henceforth, since the native
state may be reached in many ways, it is unlikely that a single minimum energy
path dominates over the rest of them [133].
In the late 80s, a new view of folding mechanisms began to emerge based on
these facts and inspired on the statistical mechanics of spin glasses [133,139,141,
144146]. According to it, when a large number of identical proteins (from 10
15
29
In fact, Levinthal did not use the word paradox and, just after stating the problem, he
proposed a possible solution to it.
35
Figure 4.1: Possible energy landscapes of a protein. The conformational space is
assumed to be two-dimensional, the degrees of freedom being q1 and q2. The degrees
of freedom of the solvent have been integrated out (see the text), and the eective
potential energy W(q1, q2) is a function of these two variables, which are internal
degrees of freedom of the molecule. N stands for native state and it is assumed here
to be the global minimum. (a) Flat golf course: the energy landscape as it would
be if Levinthals paradox were a real problem. (b) Ant trail : the old-view pathway
solution to Levinthals paradox. (c) New-view smooth funnel. (d) More realistic
partially rugged funnel. (Figures taken from reference [130] with kind permission and
somewhat modied.)
to 10
18
[147]) are introduced in a test tube in the conditions of the restricted
protein folding problem dened in section 3, a conformational equilibrium is
attained between the native ensemble of states ^ and the ensemble made up of
the rest of (non-functional) conformations (the unfolded state |). At the same
time, what is happening at the microscopic level is that each single molecule is
following a partially stochastic trajectory determined by the intrinsic energetics
of the system (given by W(x

)) and subject to random uctuations due to the

thermal noise. Of course, all trajectories are dierent, some towards the native
state and some towards the unfolded state, but, if we focus on a single molecule
at an arbitrary time, the probability that she is wandering in the native basin
is very high (typically more than 99%) and, in the rare case that we happen to
choose a protein that is presently unfolded, we will most certainly watch a very
fast race towards the native state.
In order for this to happen, we need that the energy landscape be funneled
towards the native state, like in gures 4.1c and 4.1d, so that any microscopic
trajectory has more probability to evolve in the native direction than in the
opposite one at every point of the conformational space (the ruggedness of the
funnel must also be small in order to avoid getting trapped in deep local minima
during the course of folding). In this way, the solution to Levinthals paradox
could be said to be funnels, not tunnels [148], and the deterministic pathway
image is changed by a statistical treatment in which folding is a heterogeneous
reaction involving broad ensembles of structures [149], the kinetic intermediates
36
that are sometimes observed experimentally being simply more or less deep wells
in the walls of the funnel. Anyway, although this new view has been validated
both experimentally [150] and theoretically [147], and it is widely accepted as
correct by the scientic community, one must note that it is not contradictory
with the old view, since the latter is only a particular case of the former in
which the funnel presents a deep canyon through which most of the individual
proteins roll downwards. In fact, in some studied cases, one may nd a single
pathway that dominates statistically [138, 147].
A marginal issue that arises both in the old and new views, is whether the
native state is the global minimum of the eective potential energy W(x

) of
the protein (in which case the folding process is said to be thermodynamically
controlled) or it is just the lowest-lying kinetically-accessible local minimum (in
which case we talk about kinetic control ) [115]. This question was raised by
Annsen [76], who assumed the rst case to be the correct answer and called
the assumption the thermodynamic hypothesis. Although Levinthal pointed out
a few years later that this was not necessary and that kinetic control was per-
fectly possible [140], and also despite some indications against it [151, 152], it is
now widely accepted that the thermodynamic hypothesis is fullled most of the
times, and almost always for small single-domain proteins [24, 77, 81, 115]. Of
course, nothing fundamental changes in the overall picture if the energy land-
scape is funneled towards a local minimum of W(x

) instead of being funneled

to a global one, however, from the computational point of view there is a dif-
ference: In the latter case, the prediction of the native state may be tackled
both dynamically and by simple minimization
30
of the function W(x

) (for
example, using simulated annealing [153, 154] or similar schemes), whereas, if
the thermodynamic hypothesis is broken, the native structure may still be found
performing molecular dynamics simulations, but minimization procedures could
be misleading and technically problematic. This is so because, although local
minima may also be found and described, the knowledge about towards which
one of them the protein trajectories converge depends on kinetic information,
which is absent from the typical minimization algorithms.
Now, even though a funneled energy function provides the only consistent
image that accounts for all the experimental facts about protein folding, one
must still explain the fact that the landscape is just like that. If one looks
at a protein as if it were the rst time, one sees that it is a heteropolymer
made up of twenty dierent types of amino acid monomers (see section 2).
Such a system, due to its many degrees of freedom, the constraints imposed
by chain connectivity and the dierent anities that the monomers show for
their neighbours and for the environment, presents a large degree of frustration,
that is, there is not a single conformation of the chain which optimizes all the
interactions at the same time
31
. For the vast majority of the sequences, this
30
See appendix A for some technical but relevant remarks about the minimization of the
eective potential energy function.
31
In order to be entitled to give such a simple denition, we need that the eective potential
energy of the system separates as a sum of terms with the minima at dierent points (either
because it is split in few-body terms, or because it is split in dierent types of interactions,
37
would lead to a rugged energy landscape with many low-energy states, high
barriers, strong traps, etc.; up to a certain degree, a landscape similar to that
of spin glasses. A landscape in which fast-folding to a unique three-dimensional
structure is impossible!
However, a protein is not a random heteropolymer. Its sequence has been
selected and improved along thousands of millions of years by natural selec-
tion
32
, and the score function that decided the contest, the tness that drove
the process, is just its ability to fold into a well-dened native structure in a
biologically reasonable time
33
. Henceforth, the energy landscape of a protein is
not like the majority of them, proteins are a selected minority of heteropoly-
mers for which there exists a privileged structure (the native one) so that, in
every point of the conformational space, it is more stabilizing, on average, to
form native contacts than to form non-native ones (an image radically imple-
mented by Go-type models [156]). Bryngelson and Wolynes [146] have termed
this fewer conicting interactions than typically expected the principle of min-
imal frustration, and this takes us to a natural denition of a protein (opposed
to a general polypeptide): a protein is a polypeptide chain whose sequence has
been naturally selected to satisfy the principle of minimal frustration.
Now, we should note that this funneled shape emerges from a very delicate
balance. Proteins are only marginally stable in solution, with an unfolding free
energy G
unfold
typically in the 5 15 kcal/mol range. However, if we split
this relatively small value into its enthalphic and entropic contributions, using
equation (4.15) and the already mentioned fact that the term PV is negligible
at physiological conditions [115],
G
unfold
= H
unfold
TS
unfold
, (4.16)
we nd that it is made up of the dierence between two quantities (H
unfold
and TS
unfold
) that are typically an order of magnitude larger than G
unfold
itself [115,157], i.e., the native state is enthalpically favoured by hundreds of kilo-
calories per mole and entropically penalized by approximately the same amount.
In addition, both quantities are strongly dependent on the details of the ef-
fective potential energy W(x

) (see equations (4.13) and (4.14)), which could be

imagined to be made up of the sum of thousands of non-covalent terms each one
of a size comparable to G
unfold
. This very ne tuning that has been achieved
after thousands of millions of years of natural selection is easily destroyed by a
such as van der Waals, Coulomb, hydrogen-bonds, etc.). This is a classical image which is
rigorously wrong but approximately true (and very useful to think). If one does not want to
assume the existence of interactions or few-body terms that may conict with one another,
one may jump directly to the conclusion, noting that the energy landscape of a random
heteropolymer is glassy but without introducing the concept of frustration.
32
The problem of nding the protein needle in the astronomical haystack of all possible
sequences and its solution are presented as another paradox, the blind watchmaker paradox,
and inspiringly discussed by Richard Dawkins in reference [155].
33
One may argue that the ability to perform a catalytic function also enters the tness
criterium. While this is true, it is probably a less important factor than the folding skill, since
the active site of enzymes is generally localized in a small region of the surface of the protein
and it could be, in principle, assembled on top of many dierent folds.
38
single-residue mutation or by slightly altering the temperature, the pH or the
concentration of certain substances in the environment (parameters on which
W(x

) implicitly depends).
For the same reasons, if the folding process is intended to be simulated
theoretically, the chances of missing the native state and (what is even worse)
of producing a non-funneled landscape, which is very dicult to explore us-
ing conventional molecular dynamics or minimization algorithms, are very high
if poor energy functions are used [144, 158, 159]. Therefore, it is not surpris-
ing that current force elds [106, 107, 119128], which include a number of
strong assumptions (additivity of the interactions, mostly pairwise terms, sim-
ple functional forms, etc.), are widely recognized to be incapable of folding
proteins [24, 86, 100, 102, 160163].
The improvement of the eective potential energy functions describing poly-
peptides, with the long-term goal of reliable ab initio folding, is one of the main
objectives pursued in our group, and probably one of the central issues that
must be solved before the wider framework of the protein folding problem can
be tackled. The enormous mathematical and computational complexity that
the study of these topics entails, renders the incorporation of the physicists
community essential for the future advances in molecular biology. That the
boundaries of what is normally considered physics are expanding is obvious,
and so it is that the investigation of the behaviour of biological macromolecules
is a very appealing part of the new territory to explore.
Acknowledgments
I wish to thank J. L. Alonso, J. Sancho and I. Calvo for illuminating discussions
and for the invaluable help to perform the transition mentioned in the title of
this work.
This work has been supported by the research projects E24/3 and PM048
(Arag on Government), MEC (Spain) FIS2006-12781-C02-01 and MCyT (Spain)
FIS2004-05073-C04-01. P. Echenique and is supported by a BIFI research con-
tract.
A Probability density functions
Let us dene a stochastic or random variable
34
as a pair (X, p), with X a subset
of R
n
for some n and p a function that takes n-tuples x (x
1
, . . . , x
n
) X to
positive real numbers,
p : X [0, )
x p(x)
Then, X is called range, sample space or phase space, and p is termed proba-
bility distribution or probability density function (PDF). The phase space can be
34
See Van Kampen [164] for a more complete introduction to probability theory.
39
discrete, a case with which we shall not deal here, or continuous, so that p(x) dx
(with dx := dx
1
dx
n
) represents the probability of occurrence of some n-
tuple in the set dened by (x, x + dx) := (x
1
, x
1
+ dx
1
) (x
n
, x
n
+ dx
n
),
and the following normalization condition is satised:

X
p(x) dx = 1 . (A.1)
It is precisely in the continuous case where the interpretation of the function
p(x) alone is a bit problematic, and playing intuitively with the concepts derived
from it becomes dangerous. On one side, it is obvious that p(x) is not the
probability of the value x happening, since the probability of any specic point
in a continuous space must be zero (what is the probability of selecting a random
number between 3 and 4 and obtaining exactly ?). In fact, the correct way
of using p(x) to assign probabilities to the n-tuples in X is to multiply it by
dierentials and say that it is the probability that any point in a dierentially
small interval occurs (as we have done in the paragraph above equation (A.1)).
The reason for this may be expressed in many ways: one may say that p(x) is an
object that only makes sense under an integral sign (like a Dirac delta), or one
may realize that only probabilities of nite subsets of X can have any meaning.
In fact, it is this last statement the one that focuses the attention on the fact
that, if we decide to reparameterize X and perform a change of variables x

(x),
what should not change are the integrals over nite subsets of X, and, therefore,
p(x) cannot transform as a scalar quantity (i.e., satisfying p

(x

) = p(x(x

))),
but according to a dierent rule.
If we denote the Jacobian matrix of the change of variables by x/x

, we
must have that
p

(x

) =

det

x
x

p(x(x

)) , (A.2)
so that, for any nite set Y X (with its image by the transformation denoted
by Y

), and indicating the probability of a set with a capital P, we have the
necessary property
P(Y ) :=

Y
p(x) dx =

Y

p

(x

) dx

=: P

(Y

) . (A.3)
All in all, the object that has meaning content is P and not p. If one needs
to talk about things such as the most probable regions, or the most probable
states, or the most probable points, or if one needs to compare in any other way
the relative probabilities of dierent parts of the phase space X, an arbitrary
partition of X into nite subsets (X
1
, . . . , X
i
, . . .) must be dened
35
. These
X
i
should be considered more useful states than the individual points x X
and their probabilities P(X
i
), which, contrarily to p(x), do not depend on the
35
Two additional reasonable properties should be asked to such a partition: (i) the sets
in it must be exclusive, i.e., X
i
X
j
= , i = j, and (ii) they must ll the phase space,
S
i
X
i
= X
40
5a/2
2a
3a/2
a
a/2
0
a 3a/5 a/2 0
p(x)
p(x(x))
Figure A.1: Probability density functions p(x) and p

(x

(x)) in equations (A.4) and

(A.5) respectively. In the axes, the quantities x and p(x) are shown for convenience.
Note that the area enclosed by the two curves is dierent; this is because p

(x

(x)) is
normalized with the measure dx

and not with dx, which is the one implicitly assumed

in this representation.
coordinates chosen, should be used as the meaningful quantities about which to
make well-dened probabilistic statements.
To illustrate this, let us see an example: suppose we have a 1-dimensional
PDF
p(x) =
6
a
3
x(a x) . (A.4)
The maximum of p(x) is at x = a/2, however, it would not be very clever
to declare that x = a/2 is the most probable value of x, since one may choose
to describe the problem with a dierent but perfectly legitimate variable x

,
whose relation to x is, say, x = x

2
, and nd the PDF in terms of x

using
equation (A.2):
p

(x

) =
12
a
3
x

3
(a x

2
) . (A.5)
Now, insisting on the mistake, we may nd the maximum of p

(x

), which
lies at x

= (3a/5)
1/2
(see gure A.1), and declare it the most probable value
of x

. But, according to the change of variables given by x = x

2
, the point
x

= (3a/5)
1/2
corresponds to x = 3a/5 and, certainly, it is not possible that
x = a/2 and x = 3a/5 are the most probable values of x at the same time!
To sum up, only nite regions of continuous phase spaces can be termed
states and meaningfully assigned a probability that do not depend on the co-
ordinates chosen. In order to do that, an arbitrary partition of the phase space
must be dened.
41
Far for being an academic remark, this is relevant in the study of the equilib-
rium of proteins, where, very commonly, Annsens thermodynamic hypothesis
is invoked (see section 4). Loosely speaking, it says that the functional native
state of proteins lies at the minimum of the eective potential energy (i.e., the
maximum of the associated Boltzmann PDF, proportional to e
W
, in equa-
tion (4.7)), but, according to the properties of PDFs described in the previous
paragraphs, much more qualifying is needed.
First, one must note that all complications arise from the choice of integrat-
ing out the momenta (for example, in equation (4.6)) to describe the equilibrium
distribution of the system with a PDF dependent only on the potential energy.
If the momenta were kept and the PDF expressed in terms of the complete
Hamiltonian as p(q

) = e
H
/Z, then, it would be invariant under canon-
ical changes of coordinates (which are the physically allowed ones), since the
Jacobian determinant that appears in equation (A.2) equals unity in such a
case. If we now look, using this complete description in terms of H, for the
most probable point (q

) in the whole dynamical phase space, the answer

does not depend on the coordinates chosen: It is the point with all momenta

set to zero (since the kinetic energy is a positive dened quadratic form on
the

), and the positions q

set to those that minimize the potential energy

V (q

), denoted by q

min
. If we now perform a point transformation, which is a
particular case of the larger group of canonical transformations [165],
q

(q

) and

=
q

, (A.6)
the most probable point in the new coordinates turns out to be the same one,
i.e., the point (q

) = (q

(q

min
), 0), and all the insights about the problem
are consistent.
However, if one decides to integrate out the momenta, the marginal PDF
on the positions that remains has a more complicated meaning than the joint
one on the whole phase space and lacks the reasonable properties discussed
above. The central issue is that the marginal p(q

) (for example, the one in

equation (4.7)) quanties the probability that the positions of the system be
in the interval (q

, q

+ dq

) without any knowledge about the momenta, or,

otherwise stated, for any value of the momenta.
In Euclidean coordinates, the volume in momenta space does not depend
on the positions, however, in general curvilinear coordinates, the accessible mo-
menta volume is dierent from point to point, and one can say the same about
the kinetic entropy (see reference [52]) associated with the removed

, which,
apart from the potential energy, also enters the coordinate PDF.
If, despite these inconveniences, the description in terms of only the positions
q

is chosen to be kept (which is typically recommendable from the computa-

tional point of view), two dierent approaches may be followed to assure the
meaningfulness of the statements made: Either some partition of the conforma-
tional space into nite subsets must be dened, as it is described in the beginning
of this appendix and as it is done in reference [118], or the position-dependent
kinetic entropies that appear when curvilinear coordinates are used and that are
42
introduced in reference [52] must be included in the eective potential energy
function.
References
[1] A. M. Lesk, Introduction to Protein Architecture, Oxford University Press,
Oxford, 2001.
[2] Y. Cho, S. Gorina, P. D. Jeffrey, and N. P. Pavletich, Crystal structure of
a p53 tumor suppressor-DNA complex: Understanding tumorigenic mutations,
Science 265 (1994) 346355.
[3] C. M. Dobson, Protein-misfolding diseases: Getting out of shape, Nature 729
(2002) 729.
[4] J. W. Kelly, Towards and understanding of amyloidgenesis, Nat. Struct. Biol.
9 (2002) 323.
[5] E. H. Koo, P. T. Lansbury Jr., and J. W. Kelly, Amyloid diseases: Ab-
normal protein aggregation in neurodegeneration, Proc. Natl. Acad. Sci. USA
96 (1999) 99899990.
[6] M. H. Nielsen, F. S. Pedersen, and J. Kjems, Molecular strategies to inhibit
HIV-1 replication, Retrovirology 2 (2005) 10.
[7] H. Ohtaka and E. Freire, Adaptive inhibitors of the HIV-1 protease, Prog.
Biophys. Mol. Biol. 88 (2005) 193208.
[8] U. Bacha, J. Barrila, A. Vel azquez-Campoy, S. Leavitt, and E. Freire,
Identication of novel inhibitors of the SARS associated coronavirus main pro-
teinase 3CLpro, Biochemistry 43 (2004) 49064912.
[9] S. Venkatraman, F. G. Njoroge, V. M. Girijavallabhan, V. S. Madison,
N. H. Yao, A. J. Prongay, N. Butkiewicz, and J. Pichardo, Design and
synthesis of depeptidized macrocyclic inhibitors of Hepatitis C NS3-4A protease
using structure-based drug design, J. Med. Chem. 48 (2005) 50885091.
[10] J. Poehlsgaard and S. Douthwaite, The bacterial ribosome as a target for
antibiotics, Nat. Rev. Microbiol. 3 (2005) 870881.
[11] C. Smith, Drug target validation: Hitting the target, Nature 422 (2003) 341
347.
[12] F. Vollrath and D. Porter, Spider silk as archetypal protein elastomer, Soft
Matter 2 (2006) 377-385.
[13] J. Scheller, K.-H. G uhrs, F. Grosse, and U. Conrad, Production of spider
silk proteins in tobacco and potato, Nat. Biotech. 19 (2001) 573577.
[14] C. M. Bellingham and F. W. Keeley, Self-ordered polymerization of elastin-
based biomaterials, Curr. Opin. Sol. State Mat. Sci. 8 (2004) 135139.
43
[15] A. Y. Wang, X. Mo, C. S. Chen, and S. M. Yu, Facile modication of
collagen directed by collagen mimetic peptides, J. Am. Chem. Soc. 127 (2005)
41304131.
[16] S. A. Maskarinec and D. A. Tirrell, Protein engineering approaches to
biomaterials design, Curr. Opin. Biotech. 16 (2005) 422426.
[17] M. Strong, Protein nanomachines, PLoS Biology 2 (2004) 0305.
[18] Many authors, What dont we know?, Science 309 (2005) 78102.
[19] L. D. Stein, Human genome: End of the beginning, Nature 431 (2004) 915916.
[20] A. Woolfe, M. Goodson, D. K. Goode, P. Snell, G. K. McEwen,
T. Vavouri, S. F. Smith, P. North, H. Callaway, K. Kelly, K. Walter,
I. Abnizova, W. Gilks, Y. J. K. Edwards, J. E. Cooke, and G. Elgar,
Highly conserved non-coding sequences are associated with vertebrate develop-
ment, PLoS Biology 3 (2005) 0116.
[21] M. A. Nobrega, I. Ovcharenko, V. Afzal, and E. M. Rubin, Scanning
human gene deserts for long-range enhancers, Science 302 (2003) 413.
[22] W. W. Gibbs, The unseen genome: Gems among the junk, Scientic American
289 (2003) 4653.
[23] G. Gibson and S. V. Muse, A Primer of Genome Science, Sinauer, Sunderland,
2nd edition, 2004.
[24] C. G omez-Moreno Calera and J. Sancho Sanz, editors, Estructura de
Protenas, Ariel ciencia, Barcelona, 2003.
[25] B. Boeckmann, A. Bairoch, R. Apweiler, M.-C. Blatter, A. Estreicher,
E. Gasteiger, M. J. Martin, K. Michoud, C. 0Donovan, I. Phan, S. Pil-
bout, and M. Schneider, The SWISS-PROT protein knowledgebase and its
supplement TrEMBL in 2003, Nucleic Acids Research 31 (2003) 365370.
[26] S. Orchard, H. Hermjakob, and R. Apweiler, Annotating the human pro-
teome, Mol. Cell. Proteomics 4 (2005) 435440.
[27] J. T. Pelton and L. R. McLean, Spectroscopic methods for analysis of protein
secondary structure, Anal. Biochem. 277 (2000) 167176.
[28] D. A. Case, H. J. Dyson, and P. E. Wright, Use of chemical shifts and
coupling constants in nuclear magnetic resonance structural studies on peptides
and proteins, Methods Enzymol. 239 (1994) 392416.
[29] G. J. Kleywegt, Validation of protein crystal structures, Acta Crystallogr. D
56 (2000) 249265.
[30] J. Drenth, Principles of Protein X-Ray Crystallography, Springer-Verlag, New
York, 1999.
[31] J. P. Glusker, X-ray crystallography of proteins, Methods Biochem. Anal. 37
(1994) 172.
44
[32] H. M. Berman, J. Westbrook, Z. Feng, G. Gilliland, T. N. Bhat,
H. Weissig, I. N. Shindyalov, and P. E. Bourne, The Protein Data Bank,
Nucleic Acids Research 28 (2000) 235242.
[33] Millennium Ecosystem Assessment. Ecosystems and Human Well-being: Biodi-
versity Synthesis, Washington, DC, 2005, http://www.millenniumassessment.
org/en/Products.Synthesis.aspx.
[34] T. Castrigian` o, P. DOnorio De Meo, D. Cozzetto, I. G. Talamo, and
A. Tramontano, The PMDB Protein Model Database, Nucleic Acids Research
34 (2006) 306309.
[35] A. Tramontano, Of men and machines, Nat. Struct. Biol. 10 (2003) 8790.
[36] M. D. S. Kumar, K. A. Bava, M. M. Gromiha, P. Prabakaran, K. Ki-
tajima, H. Uedaira, and A. Sarai, ProTherm and ProNIT: Thermodynamic
databases for proteins and protein-nucleic acid interactions, Nucleic Acids Re-
search 34 (2006) 204206.
[37] T.-Y. Lee, H.-D. Huang, J.-H. Hung, H.-Y. Huang, Y.-S. Yang, and T.-H.
Wang, dbPTM: An information repository of protein post-translational modi-
cation, Nucleic Acids Research 34 (2006) 622627.
[38] M. D. S. Kumar and M. M. Gromiha, PINT: Protein-protein Interactions
Thermodynamic Database, Nucleic Acids Research 34 (2006) 195-198.
[39] R. Dawkins, River Out of Eden: A Darwinian View of Life, BasicBooks, New
York, 1995.
[40] M. J. Frisch, G. W. Trucks, H. B. Schlegel, G. E. Scuseria, M. A. Robb,
J. R. Cheeseman, J. A. Montgomery, Jr., T. Vreven, K. N. Kudin, J. C.
Burant, J. M. Millam, S. S. Iyengar, J. Tomasi, V. Barone, B. Men-
nucci, M. Cossi, G. Scalmani, N. Rega, G. A. Petersson, H. Nakatsuji,
M. Hada, M. Ehara, K. Toyota, R. Fukuda, J. Hasegawa, M. Ishida,
T. Nakajima, Y. Honda, O. Kitao, H. Nakai, M. Klene, X. Li, J. E.
Knox, H. P. Hratchian, J. B. Cross, V. Bakken, C. Adamo, J. Jaramillo,
R. Gomperts, R. E. Stratmann, O. Yazyev, A. J. Austin, R. Cammi,
C. Pomelli, J. W. Ochterski, P. Y. Ayala, K. Morokuma, G. A.
Voth, P. Salvador, J. J. Dannenberg, V. G. Zakrzewski, S. Dapprich,
A. D. Daniels, M. C. Strain, O. Farkas, D. K. Malick, A. D. Rabuck,
K. Raghavachari, J. B. Foresman, J. V. Ortiz, Q. Cui, A. G. Baboul,
S. Clifford, J. Cioslowski, B. B. Stefanov, G. Liu, A. Liashenko,
P. Piskorz, I. Komaromi, R. L. Martin, D. J. Fox, T. Keith, M. A.
Al-Laham, C. Y. Peng, A. Nanayakkara, M. Challacombe, P. M. W.
Gill, B. Johnson, W. Chen, M. W. Wong, C. Gonzalez, and J. A. Pople,
Gaussian 03, Revision C.02, 2004, Gaussian, Inc., Wallingford, CT.
[41] R. S. Cahn, S. C. Ingold, and V. Prelog, Specication of molecular chirality,
Angew. Chem. Int. Ed. 5 (1966) 385415.
[42] M. Klussmann, H. Iwamura, S. P. Mathew, D. H. Wells Jr., U. Pandya,
A. Armstrong, and D. G. Blackmond, Thermodynamic control of asymmet-
ric amplication in amino acid catalysis, Nature 441 (2006) 621623.
45
[43] T. Head-Gordon, M. Head-Gordon, M. J. Frisch, C. Brooks III, and
J. Pople, A theoretical study of alanine dipeptide and analogs, Intl. J. Quant.
Chem. 16 (1989) 311-322.
[44] R. F. Frey, J. Coffin, S. Q. Newton, M. Ramek, V. K. W. Cheng, F. A.
Momany, and L. Sch afer, Importance of correlation-gradient geometry opti-
mization for molecular conformational analyses, J. Am. Chem. Soc. 114 (1992)
53695377.
[45] I. R. Gould, W. D. Cornell, and I. H. Hillier, A quantum mechanical
investigation of the conformational energetics of the alanine and glycine dipep-
tides in the gas phase and in aqueous solution, J. Am. Chem. Soc. 116 (1994)
92509256.
[46] M. Beachy, D. Chasman, R. Murphy, T. Halgren, and R. Friesner, Ac-
curate ab initio quantum chemical determination of the relative energetics of
peptide conformations and assessment of empirical force elds, J. Am. Chem.
Soc. 119 (1997) 59085920.
[47] C.-H. Yu, M. A. Norman, L. Sch afer, M. Ramek, A. Peeters, and C. van
Alsenoy, Ab initio conformational analysis of N-formyl L-alanine amide in-
cluding electron correlation, J. Mol. Struct. 567568 (2001) 361374.
[48] R. Vargas, J. Garza, B. P. Hay, and D. A. Dixon, Conformational study of
the alanine dipeptide at the MP2 and DFT levels, J. Phys. Chem. A 106 (2002)
32133218.
[49] A. Perczel,

O. Farkas, I. J akli, I. A. Topol, and I. G. Csizmadia, Pep-
tide models. XXXIII. Extrapolation of low-level Hartree-Fock data of peptide
conformation to large basis set SCF, MP2, DFT and CCSD(T) results. The Ra-
machandran surface of alanine dipeptide computed at various levels of theory,
J. Comp. Chem. 24 (2003) 10261042.
[50] Z.-X. Wang and Y. Duan, Solvation eects on alanine dipeptide: A MP2/cc-
pVTZ//MP2/6-31G** study of (, ) energy maps and conformers in the gas
phase, ether and water, J. Comp. Chem. 25 (2004) 16991716.
[51] P. Echenique and J. L. Alonso, Denition of Systematic, Approximately Sep-
arable and Modular Internal Coordinates (SASMIC) for macromolecular simu-
lation, J. Comp. Chem. 27 (2006) 10761087.
[52] P. Echenique, I. Calvo, and J. L. Alonso, Quantum mechanical calculation
of the eects of sti and rigid constraints in the conformational equilibrium of
the Alanine dipeptide, J. Comp. Chem. 27 (2006) 17481755.
[53] P. Echenique and J. L. Alonso, Ecient model chemistries for peptides. I.
Split-valence Gaussian basis sets and the heterolevel approximation in RHF and
MP2, Submitted, 2007.
[54] T. E. Creighton, Proteins: Structures and Molecular Properties, Freeman,
W. H., New York, 2nd edition, 1992.
[55] M. Di Giulio, The origin of the genetic code: Theories and their relationships,
a review, Biosystems 80 (2005) 175184.
46
[56] R. D. Knight, S. J. Freeland, and L. F. Landweber, Selection, history and
chemistry: The three faces of the genetic code, Trends Biochem. Sci. 24 (1999)
241247.
[57] M. S. Weiss, A. Jabs, and R. Hilgenfeld, Peptide bonds revisited, Nat.
Struct. Biol. 5 (1998) 676.
[58] J. F. Swain and L. M. Gierash, A new twist for an Hsp70 chaperone, Nat.
Struct. Biol. 9 (2002) 406408.
[59] V. I. Lim and A. S. Spirin, Stereochemical analysis of ribosomal transpeptida-
tion conformation of nascent peptide, J. Mol. Biol. 188 (1986) 565574.
[60] G. N. Ramachandran and C. Ramakrishnan, Stereochemistry of polypeptide
chain congurations, J. Mol. Biol. 7 (1963) 9599.
[61] L. Bragg, J. C. Kendrew, and M. F. Perutz, Polypeptide chain cong-
urations in crystalline proteins, Proc. Roy. Soc. London Ser. A 203 (1950)
321357.
[62] L. Pauling, R. B. Corey, and H. R. Branson, The structure of proteins:
Two hydrogen-bonded helical congurations of the polypeptide chain, Proc. Natl.
Acad. Sci. USA 37 (1951) 205211.
[63] D. Eisenberg, The discovery of the -helix and -sheet, the principal structural
features of proteins, Proc. Natl. Acad. Sci. USA 100 (2003) 1120711210.
[64] J. C. Kendrew, G. Bodo, H. M. Dintzis, R. G. Parrish, H. Wyckoff, and
D. C. Phillips, A three-dimensional model of the myoglobin molecule obtained
by x-ray analysis, Nature 181 (1958) 662666.
[65] M. F. Perutz, M. G. Rossman, A. F. Cullis, H. Muirhead, G. Will,
and A. C. T. North, Structure of haemoglobin: A three-dimensional Fourier
synthesis at 5.5

A resolution, obtained by x-ray analysis, Nature 185 (1960)
416422.
[66] M. N. Fodje and S. Al-Karadaghi, Occurrence, conformational features and
amino acid propensities for the -helix, Protein Eng. 15 (2002) 353358.
[67] M. E. Karpen, P. L. De-Haseth, and K. E. Neet, Dierences in the amino
acid distributions of 310-helices and -helices, Prot. Sci. 1 (1992) 13331342.
[68] E. N. Baker and R. E. Hubbard, Hydrogen bonding in globular proteins,
Prog. Biophys. Mol. Biol. 44 (1984) 97179.
[69] M. L. Higgins, The structure of brous proteins, Chem. Rev. 32 (1943) 195
218.
[70] B. W. Low and R. B. Baybutt, The helix A hydrogen bonded conguration
of the polypeptide chain, J. Am. Chem. Soc. 74 (1952) 58065807.
[71] J. Donohue, Hydrogen bonded helical congurations of the polypeptide chain,
Proc. Natl. Acad. Sci. USA 39 (1953) 470478.
47
[72] B. Zagrovic, J. Lipfert, E. J. Sorin, I. S. Millett, W. F. van Gunsteren,
S. Doniach, and V. S. Pande, Unusual compactness of a polyproline type II
structure, Proc. Natl. Acad. Sci. USA 102 (2005) 1169811703.
[73] R. V. Pappu and G. D. Rose, A simple model for polyproline II structure in
unfolded states of alanine-based peptides, Prot. Sci. 11 (2002) 24372455.
[74] B. J. Stapley and T. P. Creamer, A survey of left-handed polyproline II
helices, Prot. Sci. 8 (1999) 587595.
[75] J. C. Kendrew, Myoglobin and the structure of proteins, Nobel Lecture, 1962,
http://nobelprize.org/nobel prizes/chemistry/laureates/1962/.
[76] C. B. Anfinsen, Principles that govern the folding of protein chains, Science
181 (1973) 223230.
[77] C. M. Dobson, The nature and signicance of protein folding, in Mechanism
of Protein Folding, edited by R. H. Pain, pp. 133, Oxford University Press,
New York, 2000.
[78] F. U. Hartl and M. Hayer-Hartl, Molecular chaperones in the cytosol: from
nascent chain to folded protein, Science 295 (2002) 18521858.
[79] F. U. Hartl, Molecular chaperones in cellular protein folding, Nature 381
(2002) 571580.
[80] J. Ellis, Proteins as molecular chaperones, Nature 328 (1987) 378379.
[81] C. M. Dobson, Protein folding and misfolding, Nature 426 (2003) 884890.
[82] A. L. Horwich, E. U. Weber-Ban, and D. Finley, Chaperone rings in protein
folding and degradation, Proc. Natl. Acad. Sci. USA 96 (1999) 1103311040.
[83] C. Hardin, T. V. Pogorelov, and Z. Luthey-Schulten, Ab initio protein
structure prediction, Curr. Opin. Struct. Biol. 12 (2002) 176181.
[84] Y. Duan and P. A. Kollman, Computational protein folding: From lattice to
all-atom, IBM Systems Journal 40 (2001) 297309.
[85] E. Krieger, S. B. Nabuurs, and G. Vriend, Homology modeling, in Structural
Bioinformatics, edited by P. E. Bourne and H. Weissig, pp. 507521, Wiley-
Liss, 2003.
[86] K. Ginalski, N. V. Grishin, A. Godzik, and L. Rychlewski, Practical
lessons from protein structure prediction, Nucleic Acids Research 33 (2005)
18741891.
[87] M. Jacobson and A. Sali, Comparative protein structure modeling and its
applications to drug discovery, Ann. Rep. Med. Chem. 39 (2004) 259276.
[88] V. Daggett and A. Fersht, The present view of the mechanism of protein
folding, Nat. Rev. Mol. Cell Biol. 4 (2003) 497.
[89] B. Honig, Protein folding: From the Levinthal paradox to structure prediction,
J. Mol. Biol. 293 (1999) 283293.
48
[90] D. Baker and A. Sali, Protein structure prediction and structural genomics,
Science 294 (2001) 9396.
[91] M. A. Marti-Renom, A. C. Stuart, A. Fiser, R. S anchez, F. Melo, and
A. Sali, Comparative protein structure modeling of genes and genomes, Annu.
Rev. Biophys. Biomol. Struct. 29 (2000) 291325.
[92] C. Chothia and A. M. Lesk, The relation between the divergence of sequence
and structure in proteins, EMBO J. 5 (1986) 823826.
[93] S. F. Altschul, T. L. Madden, A. A. Sch affer, J. Zhang, Z. Zhang,
W. Miller, and L. D. J., Gapped BLAST and PSI-BLAST: A new generation
of protein database search programs, Nucleic Acids Research 25 (1997) 3389
33402.
[94] S. Henikoff, Scores for sequence searches and alignments, Curr. Opin. Struct.
Biol. 6 (1996) 352360.
[95] U. Pieper, N. Eswar, F. P. Davis, H. Braberg, M. S. Madhusudhan,
A. Rossi, M. Marti-Renom, R. Karchin, B. M. Webb, D. Eramian, M.-Y.
Shen, L. Kelly, F. Melo, and A. Sali, MODBASE: A database of annotated
comparative protein structure models and associated resources, Nucleic Acids
Research 34 (2006) 291295.
[96] J. U. Bowie, R. Luthy, and D. Eisenberg, A method to identify protein
sequences that fold into a known three-dimensional structure, Science 253 (1991)
164170.
[97] J. Moult, K. Fidelis, B. Rost, T. Hubbard, and A. Tramontano, Criti-
cal Assessment of Methods of Protein Structure Prediction (CASP)Round 6,
PROTEINS: Struct. Funct. Bioinf. 7 (2005) 37.
[98] C. A. Orengo and J. M. Thornton, Protein families and their evolutionA
structural perspective, Annu. Rev. Biochem. 74 (2005) 867900.
[99] P. Bradley, K. M. S. Misura, and D. Baker, Toward high-resolution de
novo structure prediction for small proteins, Science 309 (2005) 18681871.
[100] O. Schueler-Furman, C. Wang, P. Bradley, K. Misura, and D. Baker,
Progress in modeling of protein structures and interactions, Science 310 (2005)
638642.
[101] J. Moult, A decade of CASP: Progress, bottlenecks and prognosis in protein
structure prediction, Curr. Opin. Struct. Biol. 15 (2005) 285289.
[102] R. Bonneau and D. Baker, Ab initio protein structure prediction: Progress
and prospects, Annu. Rev. Biophys. Biomol. Struct. 30 (2001) 173189.
[103] C. A. Rohl, C. E. Strauss, K. M. Misura, and D. Baker, Protein structure
prediction using Rossetta, Methods Enzymol. 383 (2004) 6693.
[104] C. A. Rohl, C. E. Strauss, D. Chivian, and D. Baker, Modeling structurally
variable regions in homologous proteins with Rossetta, PROTEINS: Struct.
Funct. Bioinf. 55 (2004) 656677.
49
[105] M. Levitt and A. Warshel, Computer simulation of protein folding, Nature
253 (1975) 694698.
[106] A. D. MacKerell Jr., B. Brooks, C. L. Brooks III, L. Nilsson, B. Roux,
Y. Won, and M. Karplus, CHARMM: The energy function and its parameter-
ization with an overview of the program, in The Encyclopedia of Computational
Chemistry, edited by P. v. R. Schleyer et al., pp. 217277, John Wiley &
Sons, Chichester, 1998.
[107] B. R. Brooks, R. E. Bruccoleri, B. D. Olafson, D. J. States, S. Swami-
nathan, and M. Karplus, CHARMM: A program for macromolecular energy,
minimization, and dynamics calculations, J. Comp. Chem. 4 (1983) 187217.
[108] D. Baker, Prediction and design of macromolecular structures and interactions,
Phil. Trans. R. Soc. London B Biol. Sci. 361 (2006) 459463.
[109] J. Skolnick, Putting the pathway back into protein folding, Proc. Natl. Acad.
Sci. USA 102 (2005) 22652266.
[110] M. A. Basharov, Protein folding, J. Cell. Mol. Med. 7 (2003) 223237.
[111] B. Hardesty and G. Kramer, Folding of a nascent peptide on the ribosome,
Prog. Nucleic Acid Res. Mol. Biol. 66 (2001) 4166.
[112] A. R. Fersht and V. Daggett, Protein folding and unfolding at atomic
resolution, Cell 108 (2002) 573582.
[113] V. Daggett and A. R. Fersht, Is there a unifying mechanism for protein
folding?, Trends Biochem. Sci. 28 (2003) 1825.
[114] C. Levinthal, Are there pathways for protein folding?, J. Chim. Phys. 65
(1968) 4445.
[115] T. Lazaridis and M. Karplus, Thermodynamics of protein folding: a micro-
scopic view, Biophys. Chem. 100 (2003) 367395.
[116] W. Greiner, H. Stocker, and L. Neise, Thermodynamics and Statistical
Mechanics, Classical Theoretical Physics, Springer, New York, 2004.
[117] B. A. Dubrovin, A. T. Fomenko, and S. P. Novikov, Modern Geometry
Methods and Applications, volume I. The Geometry of Surfaces, Transformation
Groups and Fields, Springer, New York, 1984.
[118] T. Lazaridis and M. Karplus, Eective energy function for proteins in solu-
tion, PROTEINS: Struct. Funct. Gen. 35 (1999) 133152.
[119] W. F. Van Gunsteren and M. Karplus, Eects of constraints on the dy-
namics of macromolecules, Macromolecules 15 (1982) 15281544.
[120] W. D. Cornell, P. Cieplak, C. I. Bayly, I. R. Gould, J. Merz, K. M.,
D. M. Ferguson, D. C. Spellmeyer, T. Fox, J. W. Caldwell, and P. A.
Kollman, A second generation force eld for the simulation of proteins, nucleic
acids, and organic molecules, J. Am. Chem. Soc. 117 (1995) 51795197.
50
[121] D. A. Pearlman, D. A. Case, J. W. Caldwell, W. R. Ross, T. E.
Cheatham III, S. DeBolt, D. Ferguson, G. Seibel, and P. Kollman, AM-
BER, a computer program for applying molecular mechanics, normal mode anal-
ysis, molecular dynamics and free energy calculations to elucidate the structures
and energies of molecules, Comp. Phys. Commun. 91 (1995) 141.
[122] W. L. Jorgensen and J. Tirado-Rives, The OPLS potential functions for
proteins. Energy minimization for crystals of cyclic peptides and Crambin, J.
Am. Chem. Soc. 110 (1988) 16571666.
[123] W. L. Jorgensen, D. S. Maxwell, and J. Tirado-Rives, Development
and testing of the OPLS all-atom force eld on conformational energetics and
properties of organic liquids, J. Am. Chem. Soc. 118 (1996) 1122511236.
[124] T. A. Halgren, Merck Molecular Force Field. I. Basis, form, scope,
parametrization, and performance of MMFF94, J. Comp. Chem. 17 (1996)
490519.
[125] T. A. Halgren, Merck Molecular Force Field. II. MMFF94 van der Waals and
electrostatica parameters for intermolecular interactions, J. Comp. Chem. 17
(1996) 520552.
[126] T. A. Halgren, Merck Molecular Force Field. III. Molecular geometrics and
vibrational frequencies for MMFF94, J. Comp. Chem. 17 (1996) 553586.
[127] T. A. Halgren, Merck Molecular Force Field. IV. Conformational Energies
and Geometries for MMFF94, J. Comp. Chem. 17 (1996) 587615.
[128] T. A. Halgren, Merck Molecular Force Field. V. Extension of MMFF94 using
experimental data, additional computational data, and empirical rules, J. Comp.
Chem. 17 (1996) 616641.
[129] V. I. Arnold, Mathematical Methods of Classical Mechanics, Graduate Texts
in Mathematics, Springer, New York, 2nd edition, 1989.
[130] K. A. Dill and H. S. Chan, From Levinthal to pathways to funnels: The new
view of protein folding kinetics, Nat. Struct. Biol. 4 (1997) 1019.
[131] A. Fersht, Structure and Mechanism in Protein Science: A Guide to Enzyme
Catalysis and Protein Folding, Freeman, W. H., New York, 1998.
[132] S. Reich, Smoothed Langevin dynamics of highly oscillatory systems, Physica
D 118 (2000) 210224.
[133] C. M. Dobson, A.

Sali, and M. Karplus, Protein folding: A perspective from
theory and experiment, Angew. Chem. Int. Ed. 37 (1998) 868893.
[134] D. Perchak, J. Skolnick, and R. Yaris, Dynamics of rigid and exible
constraints for polymers. Eect of the Fixman potential, Macromolecules 18
(1985) 519525.
[135] M. R. Pear and J. H. Weiner, Brownian dynamics study of a polymer chain
of linked rigid bodies, J. Chem. Phys. 71 (1979) 212.
51
[136] E. Helfand, Flexible vs. rigid constraints in Statistical Mechanics, J. Chem.
Phys. 71 (1979) 5000.
[137] T. Lazaridis and M. Karplus, Discrimination of the native from misfolded
protein models with an energy function including implicit solvation, J. Mol. Biol.
288 (1999) 477487.
[138] T. Lazaridis and M. Karplus, New view of protein folding reconciled with
the old through multiple unfolding simulations, Science 278 (1997) 19281931.
[139] K. A. Dill, Polymer principles and protein folding, Prot. Sci. 8 (1999) 1166
1180.
[140] C. Levinthal, How to fold graciously, in Mossbauer Spectroscopy in Biological
Systems, edited by J. T. P. DeBrunner and E. Munck, pp. 2224, Allerton
House, Monticello, Illinois, 1969, University of Illinois Press.
[141] J. D. Bryngelson, J. N. Onuchic, N. D. Socci, and P. G. Wolynes, Fun-
nels, pathways, and the energy landscape of protein folding: A synthesis, Pro-
teins 21 (1995) 167195.
[142] H. S. Chan and K. A. Dill, Protein folding in the landscape perspective:
Chevron plots and non-Arrhenius kinetics, PROTEINS: Struct. Funct. Gen. 30
(1998) 233.
[143] R. L. Baldwin, The nature of protein folding pathways: The classical versus
the new view, J. Biomol. NMR 5 (1995) 103109.
[144] J. N. Onuchic and P. G. Wolynes, Theory of protein folding, Curr. Opin.
Struct. Biol. 14 (2004) 7075.
[145] S. S. Plotkin and J. Onuchic, Understanding protein folding with energy
landscape theory. Part I: Basic concepts, Quart. Rev. Biophys. 35 (2002) 111
167.
[146] J. D. Bryngelson and P. G. Wolynes, Spin-glasses and the statistical-
mechanics of protein folding, Proc. Natl. Acad. Sci. USA 84 (1987) 75247528.
[147] R. Day and V. Daggett, Ensemble versus single-molecule protein unfolding,
Proc. Natl. Acad. Sci. USA 102 (2005) 134451450.
[148] S. E. Radford and C. M. Dobson, Insights into protein folding using physical
techniques: Studies of lysozyme and alpha-lactalbumin, Phil. Trans. R. Soc.
London B Biol. Sci. 348 (1995) 1725.
[149] C. D. Snow, H. Nguyen, V. S. Pande, and M. Gruebele, Absolute compari-
son of simulated and experimental protein-folding dynamics, Nature 420 (2002)
102106.
[150] A. A. Deniz, T. A. Laurence, G. S. Beligere, M. Dahan, A. B. Martin,
D. S. Chemla, P. E. Dawson, P. G. Schultz, and S. Weiss, Single-molecule
protein folding: Diusion uorescence resonance energy transfer studies of the
denaturation of chymotrypsin inhibitor 2, Proc. Natl. Acad. Sci. USA 97 (2000)
51795184.
52
[151] D. Baker, Metastable states and folding free energy barriers, Nat. Struct. Biol.
5 (1998) 10211034.
[152] J. L. Sohl, S. S. Jaswal, and D. A. Agard, Unfolded conformations of -lytic
protease are more stable than its native state, Nature 395 (1998) 817819.
[153] S. Kirkpatrick, C. D. Gelatt, and M. P. Vecchi, Optimization by simulated
annealing, Science 220 (1983) 671680.
[154] V. Cerny, A thermodynamical approach to the travelling salesman problem:
An ecient simulation algorithm, J. Optimiz. Theory App. 45 (1985) 4151.
[155] R. Dawkins, The Blind Watchmaker: Why the Evidence of Evolution Reveals
a Universe without Design, W. W. Norton & Company, New York, 1987.
[156] N. G o and H. Taketomi, Respective roles of short- and long-range interactions
in protein folding, Proc. Natl. Acad. Sci. USA 75 (1978) 559563.
[157] G. I. Makhatadze and P. L. Privalov, Energetics of protein structure, Adv.
Prot. Chem. 47 (1995) 307425.
[158] P. Derreumaux, Ab initio polypeptide structure prediction, Theo. Chem. Acc.
104 (2000) 16.
[159] R. A. Abagyan, Protein structure prediction by global energy optimization, in
Computer Simulations of Biomolecular Systems, edited by W. F. Van Gun-
steren, volume 3, Kluwer academic publishing, Dordrecht, 1997.
[160] C. D. Snow, E. J. Sorin, Y. M. Rhee, and V. S. Pande, How well can
simulation predict protein folding kinetics and thermodynamics?, Annu. Rev.
Biophys. Biomol. Struct. 34 (2005) 4369.
[161] A. R. MacKerell Jr., M. Feig, and C. L. Brooks III, Extending the
treatment of backbone energetics in protein force elds: Limitations of gas-
phase quantum mechanics in reproducing protein conformational distributions
in molecular dynamics simulations, J. Comp. Chem. 25 (2004) 14001415.
[162] A. V. Morozov, T. Kortemme, K. Tsemekhman, and D. Baker, Close
agreement between the orientation dependence of hydrogen bonds observed in
protein structures and quantum mechanical calculations, Proc. Natl. Acad. Sci.
USA 101 (2004) 69466951.
[163] M. Karplus and J. A. McCammon, Molecular dynamics simulations of
biomolecules, Nat. Struct. Biol. 9 (2002) 646652.
[164] N. G. van Kampen, Stochastic processes in Physics and Chemistry, North-
Holland, Amsterdam, 1981.
[165] H. Goldstein, C. Poole, and J. Safko, Classical Mechanics, Addison-Wesley,
3rd edition, 2002.
53