Aligning Multiple Protein Structures using

Biochemical and Biophysical Properties

Paul Shealy and Homayoun Valafar
Department of Computer Science
University of South Carolina
Columbia, SC 29208
Abstract - Aligning multiple protein structures can they can be classified into a hierarchy by structure and
yield valuable information about structural similarities function, such as the manually curated database SCOP
among related proteins, as well as provide insight into [2]. This hierarchy provides a starting point for a
evolutionary relationships between proteins in a family. variety of investigations, such as determining the
We have developed an algorithm (msTALI) for aligning function of a new protein with a known structure or
multiple protein structures using biochemical and isolating the functional residues from a set of related
biophysical properties, including torsion angles, structures.
secondary structure, hydrophobicity, and surface
Comparing protein structures is an inherently
accessibility. The algorithm is a progressive alignment
difficult task. One problem is that is no single accepted
algorithm motivated by popular techniques from
definition of structural similarity. Many structure
multiple sequence alignment. It has demonstrated
alignment algorithms consider the protein to be a single
success in aligning the major structural regions of a set
rigid entity and use geometric properties to define
of protein from the s/r kinase family. The algorithm was
similarity. Some algorithms rely on interatomic distance
also successful at aligning functional residues of these
maps [3; 4]. Others use the distance and orientation of
proteins. In addition, the algorithm was also successful
secondary structure elements [5] or distance
in aligning seven members of the acyl carrier protein
comparisons between heptapeptide fragments [6],
family, including both experimentally derived as well as
Recently, algorithms on multiple structure alignment
computationally modeled structures.
have been developed, using heptapeptide fragments [7],
Keywords: structure alignment, multiple structure backbone RMSD of protein fragments of varying size
alignment, protein, active site, protein core, torsion [8], interatomic distance maps with a Monte Carlo
angles. search [9], or secondary structure elements [10].
Here we present an algorithm called msTALI
1 Introduction (multistructure torsion angle alignment) which utilizes
local structural information to create alignments of
Comparisons of protein structures can yield multiple protein structures. msTALI is inspired by and
valuable information about biologically relevant extends previous work on pairwise structure alignment
similarities between related structures. Identification of using torsion angles [11]. It computes alignments using
common structural motifs between two proteins can local structural motifs derived from torsion angles and
provide valuable information about their evolutionary biochemical properties. Proteins with very distant
relationship and yield insight into structural evolutionary relationships are more likely to exhibit
components required for the proteins to function. A local structural similarities. In addition, alignments
structure alignment can help identify the function for a using torsion angles are computationally efficient when
novel protein. Furthermore, structural comparison computing the optimal solution.
algorithms are an important validation step for
approaches to protein folding, such as ab initio methods
or threading algorithms. 2 Background and Method
While a pairwise comparison between two
structures is useful, a simultaneous analysis of multiple 2.1 Pairwise Structure Alignment
structures can be far more informative. With more than
50,000 protein structures in the PDB [1] as of 2009, Pairwise alignment of protein structures is a
core component of the multiple structure alignment
mode of msTALI. This algorithm is a straightforward S S  r a , r b =M  r a , r b  (4)
extension of sequence alignment algorithms to
structures. It treats each protein as a series of residues
and applies a generalized Needleman-Wunsch [12]
Where S H , S A , and S S are scores for the
algorithm to find the best global alignment. A
comparison between two residues is based on their hydrophobicity, surface accessibility, and sequence
biochemical and biophysical properties. One core scores, respectively; H r  is the hydrophobicity of
property is the residue's torsion angles. Treating a residue r , A r  is the surface acccessibility, and M
protein as a polymer of peptide planes, the relationship is the scoring matrix.
between two consecutive peptide planes can be
The final scoring function for two residues is
described by two torsion angles, φ and ψ. The atomic
positions of a protein's backbone can be almost entirely Sr a , r b = wT⋅S T r a ,r b
determined by a complete set of torsion angles.
Representing a protein in this form provides a compact, w H⋅S H r a ,r b
(5)
concise representation of the protein backbone. w A⋅S A r a , r b 
While all torsion angles are possible for a wS⋅S S r a ,r b
residue, many values yield an atomic structure that is
highly unfavorable due to steric hindrance and London Where w T , w H , w A , and w S are the weights for
dispersion forces and therefore are rarely observed.
Other values yield secondary structures, α-helices and the individual components.
β-sheets, which are locally stable. The space of all Gaps are implemented using a standard affine
possible torsion angles and their frequency of gap penalty. For a gap of n residues, with a gap open
observance is called a Ramachandran space [13]. penalty G O and gap extension penalty G E , the total
Ramachandran space has regions of high probability penalty is
that correspond to secondary structures, with much of
the remainder corresponding to angles of low Gn=GO n⋅G E (6)
probability. To compare two sets of torsion angles, the
difference in probabilities is used. Because separate
secondary structure regions may have similar
probabilities, a separate penalty is imposed for 2.2 Multiple Structure Alignment
transitions across a secondary structure boundary. The
secondary structure scoring function is then The pairwise alignment algorithm can be
extended to multiple structure alignment in a
S T r a ,r b=∣Rr a −R r b∣Err r a , r b  (1) straightforward manner. This algorithm is a progressive
alignment algorithm inspired by ClustalW [16]. This
algorithm is based on the observation that the easiest
where S T is the torsion angle scoring function, alignments to compute are between structures that are
ra and r b are the residues to be scored, R is the the most similar. The algorithm to compute the
multiple structure alignment between a set S of protein
Ramachandran likelihood function, and Err is the
structures is as follows:
secondary structure penalty function.
1. Compute the pairwise distances between any
Additional properties used to compare two
two structures in S.
residues are hydrophobicity and surface accessibility.
The hydropobicity scale used is the Kyte-Doolittle scale 2. Compute a guide tree for S, using the pairwise
[14]. Hydrophobicities for two residues are compared distances from step 1.
by computing the difference between the
hydrophobicity values for the residues. Surface 3. Progressively align the structures of S according
accessibility is computed using DSSP [15], and the two to the guide tree, working from the leaves to the
values are compared as the difference in values. Finally, root.
sequence information can be incorporated if desired. This algorithm utilizes a structure profile,
The individual penalty functions are which is a set of aligned structures. In step 1, pairwise
distances are computed between all pairs of structures
S H  r a , r b =∣H  r a −H  r b ∣ (2) in S, using the pairwise alignment method described
earlier. These distances are used to compute a guide
S A r a , r b =∣Ar a − Ar b∣ (3) tree in step 2, using the neighbor-joining algorithm
 Scoring an affine gap requires two parameters:
a gap open penalty and a gap extension penalty, referred
to as GO and GE, respectively. These are specified by
the user. However, both parameters are modified by the
algorithm as follows:
1. If there are existing gaps at the new gap
location, the penalties are adjusted as follows,
where ng is the number of gaps at the location
and ns is the number of structures. In this case,
no other rules apply.
GO  0.3⋅GO−0.3⋅ng /ns⋅GO
GE  0.5⋅GE
2. If there is a loop at the new gap location, the
Figure 1: A guide tree for four structures. penalties are adjusted as follows, where nl is
the number of profiles with a loop at the
[17]. An example is shown in Figure 1. Once the guide location and ns is the number of structures.
tree is constructed, each leaf node is associated with an
alignment profile containing only that node's structure, GO GO−0.75⋅GO⋅nl/ns
so each structure belongs to a separate profile. Each GE  GE−0.5⋅GE⋅nl/ns
profile is weighted according to its distance from the
root in the guide tree [16]. Weighting reduces the Often, the initial alignment will produce
impact of several highly similar structures on the final reasonable results that can be improved by realignment.
alignment. Realigning a profile involves repeatedly extracting a
single structure from it, then aligning it to the remaining
In step 3, two leaf nodes that share a common profile. Here we realign each structure once.
parent node (i.e., siblings in the tree) are selected, and
their associated profiles are aligned. The aligned Using a guide tree to inform the order in which
profiles are combined into a single profile, and the to align the structures causes the most similar structures
parent node joining the leaves is replaced by a single to be aligned first, leading to subtle features in the
node associated with the combined profile. This is alignment being identified early. Structures are
repeated until only a single profile remains, which is the weighted to reduce the influence of multiple highly
full alignment. In Figure 1, the structures are aligned as related structures on the final alignment.
follows: 1PME versus 1O6L, 1UNL versus 1GZ8, and
1UNL/1GZ8 versus 1PME/1O6L.
3 Results and Discussion
To align two profiles, the same core alignment
algorithm in step 1 is used. To score a residue-to- We applied msTALI to two sets of proteins.
residue match between a position in profiles p1 and p2, One, s/r kinases, is a set of four highly conserved
each residue from p1 is compared to each residue from structures with a low sequence conservation of 17%
p2 using the scoring method described earlier for two across the set. The second set of proteins, acyl carrier
residues, the score is multiplied by the weights from the proteins, contains six structures of identical function but
two structures, and the average weighted score is with a number of small structural variations between
computed. The score for a residue versus a gap is zero. members.
Because all scoring metrics used are scaled to be non- For all experiments here, the weights used
negative, a residue versus gap score is the lowest were
possible value.
The structures in a profile remain fixed with w T =0.5
respect to one another during the rest of the algorithm. w H =0.2
Thus, the residues at each position do not shift, and
w A =0.3
once a gap is introduced in a profile, it remains fixed
for the remainder of the alignment. When introducing a wS =0.0
gap in a profile, the gap is introduced in all structures in
where the contributions to the total score are 50%
the profile.
torsion angles, 20% hydrophobicity, 30% surface
 1GZ8 -----enfqkvekigegtygvvykarnkl-tgevvalkkirv-------p---staireisllkelnhpniv---klld
1UNL -----qkyeklekigegtygtvfkaknre-theivalkrvrldd--ddeg-vpssalreicllkelkhkniv---rlhd
1O6L -vtm-ndfdylkllgkgtfgkvilvreka-tgryyamkilrkeviiakde--vahtvtesrvlqntrhpflt---alky
1BLX lcradqqyecvaeigegaygkvfkardlknggrfvalkrvrvqtgeegmplstirevavlrhletfehpnvvrlfdvct

vihten--klylvfeflhq-dlkkfmdasaltgiplpliksylfqllqglafchshrvlhrdlkpqnllintegaikladfgla
vlhsdk--kltlvfefcdq-dlkkyfdscn-gdldpeivksflfqllkglgfchsrnvlhrdlkpqnllinrngelklanfgla
afqthd--rlcfvmeyanggelffhlsrer--vfteerarfygaeivsaleylhsrdvvyrdiklenlmldkdghikitdfglc
vsrtdretkltlvfehvdq-dlttyldkvpepgvptetikdmmfqllrgldflhshrvvhrdlkpqnilvtssgqikladfgla

1GZ8 -----llBBBBBBBBBllllBBBBBBBll-lllBBBBBBBll-------l---HHHHHHHHHHlllllllBl---lBBB
1UNL -----llBBBBBBBBBllllBBBBBBBll-lllBBBBBBBBlll--llll-HHHHHHHHHHHHlllllllBl---lBBB
1O6L -llH-HHBBBBBBBBBlllBBBBBBBBll-lllBBBBBBBBHHHHHHlll--HHHHHHHHHHHHllllllBl---lBBB
1BLX lllHHHlBBBBBBBBBBllBBBBBBBBlllllBBBBBBBBBBBBlllllBllHHHHHHHHHHHHHlllllBllBBBBBB

BBBBll--BBBBBBBlllB-BHHHHHHHlllllllHHHHHHHHHHHHHHHHHHHHlllllllllHHHBBBlllllBBBllllHH
BBBlll--BBBBBBBlllB-BHHHHHHHll-llllHHHHHHHHHHHHHHHHHHHHllBBlllllHHHBBBlllllBBBllllll
BBBlll--BBBBBBBlllllBHHHHHHHHl--lllHHHHHHHHHHHHHHHHHHHHlllBlllllHHHBBBlllllBBBllllll
BBBlllBBBBBBBBBllll-BHHHHHHHlllllllHHHHHHHHHHHHHHHHHHHHlllllllllHHHBBBlllllBBBllllll

Figure 2: Multiple structure alignment of protein kinases. The top half displays the residues; the bottom half
displays the secondary structure elements – B for β-strand, H for α-helix, l for loop. The nucleotide binding site is
highlighted in blue, the ATP binding site is in yellow, and the proton acceptor site in green.

accessibility. The sequence component is unused. The majority of the secondary structure regions are
Because some key residues are conserved across all conserved across all structures; in all cases, msTALI
proteins being aligned, this provides a method for has aligned these regions together. Furthermore, for
validation of the algorithm. many of these conserved secondary structures of equal
length, msTALI has aligned the ends of the structures.
Where there is ambiguity about the precise alignment,
3.1 Protein S/R Kinases such as the additional α-helices present in 1O6L, the
We aligned four members of the protein kinase hydrophobicity and surface accessibility components
catalytic subunit family as indicated in SCOP: 1GZ8, provide valuable information in determining the exact
1UNL, 1O6L, and 1BLX. All are serine/theronine local alignment.
kinases from homo sapiens. The entire chain indicated The functional regions of the protein s/r
in SCOP was aligned, although the s/r domain is the kinases are annotated in UniProtKB [18]. These sites
first ~50% of each chain. Here we only display the are highlighted in Figure 2. These sites were all aligned
alignment of the s/r kinase domain for brevity. correctly by msTALI. Furthermore, these sites are all
These structures were simultaneously aligned scored highly by msTALI, indicating well-conserved
using msTALI. No sequence information was used to sites.
create the alignment. Because many of the structures
have a high sequence identity to other structures in the
3.2 Acyl Carrier Proteins
alignment, sequence information provides a valuable,
independent method for judging the quality of the The acyl carrier protein (ACP) family is
alignment. involved in fatty acid synthesis, linking intermediates
during synthesis via a thioester linkage. Here we have
The full alignment is displayed in Figure 2.
aligned a set of acyl carrier proteins from varying
The score is an indication of the degree of conservation
sources. Three are crystal structures: 2FAC, 1L0I, and
at a position, ranging from a low of 0 to a high of 9.
2EHS. Two are NMR structures: 2JQ4 and 1ACP. One,
2FAC tieervkkiigeqlgv--kqeevtnnasfvedlgadsldtvelvmaleeefdteipdeeaek--ittvqaaidyin---g-hq-
1L0I tieervkkiigeqlgv--kqeevtnnasfvedlgadsldtvelvmaleeefdteipdeeaek--mttvqaaidyin---g-hq-
2EHS -leervkeiiaeqlgv--ekekitpeakfvedlgadsldvvelimafeeefgieipdedaek--iqtvgdvinylk---e-k--
2JQ4 --natireilakfgqlptpvdtiadeadl-yaaglssfasvqlmlgieeafdiefpdnllnrksfasikaiedtvklildgkea
1ACP tieervkkiigeqlgv—kqeevtnnasfvedlgadsldtvelvmaleeefdteipdeeaek--ittvqaaidyin---g-hq-
AcpXL atfdkvadiiaetsei--dratitpeshtiddlgidsldfldivfaidkefgikiplekwtq---e-----vn-----------

2FAC lHHHHHHHHHHHHHll--lHHHlllllBllllllllHHHHHHHHHHHHHHllllllHHHHHl--llBHHHHHHHHH---H-Hl-
1L0I lHHHHHHHHHHHHHll--lHHHlllllBllllllllHHHHHHHHHHHHHHHlllllHHHHll--llBHHHHHHHHH---H-ll-
2EHS -HHHHHHHHHHHHHll--lHHHlllllBllllllllHHHHHHHHHHHHHHHlllllHHHHHl--llBHHHHHHHHH---H-H--
2JQ4 --HHHHHHHHHHlllllllHHHllllllH-HHHlllHHHHHHHHHHHHHHHlllllHHHHllHHHHlHHHHHHHHHHHHHlHHH
1ACP llHHHHHHHHHHHlll--llllllllllllllllllHHHHHHHHHHHHHHHlllllHHHHll--llllHHHHHHHH---H-Hl-
AcpXL lHHHHHHHHHHHHHll—lHHHlllllBllllllllHHHHHHHHHHHHHHHlllllHHHHll---l-----lB-----------

Score 458899888988888900878898898786888888998899998898988789988988880067666667866600060530

Figure 3: Secondary structure of the residues of acyl carrier proteins. H denotes an α-helix (of any type), B denotes
a β-bridge or β-sheet, and l denotes a turn, bend, or loop.
AcpXL, is computationally modeled using I-TASSER so its alignment is more difficult to evaluate. However,
[19]. These structures were simultaneously aligned the conserved secondary structure elements are aligned
using msTALI. well with respect to the crystal structure 2FAC. The
conserved secondary structure elements – the first,
The full alignment of all acyl carrier proteins
second, fourth, and fifth α-helices of 2JQ4, are aligned
structures is shown in Figure 3. The three crystal
to their corresponding helices from 2FAC. The
structures are aligned perfectly with respect to one
alignment of the third α-helix of 2JQ4 to loop regions
another. One NMR structure, 1ACP, has high sequence
from other structures is reasonable, given that the
similarity to the crystal structures. It is clearly aligned
surrounding helices are precisely aligned and only a
well from its sequence identity and secondary structure
single gap was inserted in this region. Finally, 2JQ4 has
similarity to the crystal structures. The second NMR
three α-helices at its C-terminus, whereas the other
structure, 2JQ4, has much lower sequence identity, and
structures only have one. TALI has aligned the second

Figure 4: Alignment of the crystal structure 2EHS, Figure 5: Alignment of the crystal structure 2EHS,
in green, the NMR structures 2JQ4 , in magenta, in green, to the computational structure AcpXL, in
and 1ACP, in silver. purple.
helix from 2JQ4 to the single helix from the other structural components required to perform particular
proteins. functions.
The computational structure, AcpXL, has a msTALI is currently implemented in
high sequence identity to the crystal structures, so the MATLAB®, a flexible development environment
alignment is easy to assess. In addition, the secondary centered around matrices. We plan to re-implement it in
structure elements are identical. From these items, we C++ for speed and general-purpose use, and also to
see that these structures are precisely aligned to their make a web version available. Furthermore, msTALI
crystal structure counterparts. The only questionable could be beneficial in a number of other tools from our
part is the tail of AcpXL, which appears to be shifted to lab, such as PDPA.
the right by six residues.
The (rigid) protein structures were aligned in
Acknowledgements
MolMol [20] using the major conserved region from
each protein, as indicated by msTALI. This region is Thanks to the Rothberg Fellows program in the
annotated in Figure 3. Department of Computer Science and Engineering at
the University of South Carolina for their support to
A full alignment of this set of structures
PGS.
provides valuable insight into this protein family. When
considering only the three crystal structures, the
structures are highly similar, with pairwise backbone
RMSDs ranging from 0.39 to 0.67. Incorporating the
NMR structures as well provides insight into the 5 Bibliography
regions of similarity and difference between the crystal [1] Berman HM, Westbrook J, Feng Z, Gilliland G,
and NMR structures. Figure 4 displays a partial region Bhat TN, Weissig H, Shindyalov IN & Bourne PE.
from this alignment, isolating the two NMR structures The Protein Data Bank. Nucleic Acids Res (2000)
and a single crystal structure. The α-helices on the right 28: pp. 235-242.
are aligned well, while the helical regions in the top left [2] Murzin AG, Brenner SE, Hubbard T & Chothia C.
show a clear divergence. These regions of divergence SCOP - A Structural Classification Of Proteins
could be highly dynamical regions or areas of varying Database For The Investigation Of Sequences And
functionality. Structures. Journal of Molecular Biology (1995)
Finally, displaying the multiple structure 247: pp. 536-540.
alignment between the crystal structure 2EHS and the [3] Holm L & Sander C. Protein structure comparison
computationally modeled structure AcpXL provides by alignment of distance matrices. J Mol Biol
valuable information on the conservation of core (1993) 233: pp. 123-138.
structural and functional residues in the modeled [4] Shindyalov IN & Bourne PE. Protein structure
structure. This is illustrated in Figure 5. alignment by incremental combinatorial extension
(CE) of the optimal path. Protein Engineering
(1998) 11: pp. 739-747.
4 Conclusion [5] Krissinel E & Henrick K. Secondary-structure
matching (SSM), a new tool for fast protein
We have presented an algorithm for aligning structure alignment in three dimensions. Acta
multiple protein structures. It is computationally Crystallographica Section D-Biological
efficient, with a computational complexity on the same Crystallography (2004) 60: pp. 2256-2268.
order as multiple sequence alignment with ClustalW. It [6] Ortiz AR, Strauss CEM & Olmea O. MAMMOTH
provides both pairwise and multiple structure (Matching molecular models obtained from
alignments using several germane biochemical and theory): An automated method for model
biophysical properties. comparison. Protein Science (2002) 11: pp. 2606-
An alignment between multiple protein 2621.
structures provides a wealth of information about a [7] Lupyan D, Leo-Macias A & Ortiz AR. A new
protein family. For example, aligning proteins that progressive-iterative algorithm for multiple
contain a common active site could provide a method to structure alignment. Bioinformatics (2005) 21: pp.
identify the key residues in the active site. It could also 3255-3263.
identify the structural elements required to position the [8] Shatsky M, Nussinov R & Wolfson HJ. A method
atoms in the active site correctly. Aligning structurally for simultaneous alignment of multiple protein
related proteins can also elucidate their structural structures. Proteins (2004) 56: pp. 143-156.
differences, which may provide insight into the [9] Guda C, Scheeff ED, Bourne PE & Shindyalov
 IN. A new algorithm for the alignment of multiple
protein structures using Monte Carlo optimization.
Pac Symp Biocomput (2001) : pp. 275-286.
[10] Krissinel E & Henrick K. Multiple Alignment of
Protein Structures in Three Dimensions. Lecture
Notes in Computer Science (2005) 3695: pp. 67-
78.
[11] Miao X, Waddell PJ & Valafar H. TALI: local
alignment of protein structures using backbone
torsion angles. J Bioinform Comput Biol (2008) 6:
pp. 163-181.
[12] Needleman SB & Wunsch CD. A general method
applicable to the search for similarities in the
amino acid sequence of two proteins. J Mol Biol
(1970) 48: pp. 443-453.
[13] Ramachandran GN, Ramakrishnan C &
Sasisekharan V. Stereochemistry Of Polypeptide
Chain Configurations. Journal Of Molecular
Biology (1963) 7: p. 95-&.
[14] Kyte J & Doolittle RF. A simple method for
displaying the hydropathic character of a protein.
J Mol Biol (1982) 157: pp. 105-132.
[15] Kabsch W & Sander C. Dictionary of protein
secondary structure: pattern recognition of
hydrogen-bonded and geometrical features.
Biopolymers (1983) 22: pp. 2577-2637.
[16] Thompson JD, Higgins DG & Gibson TJ.
CLUSTAL W: improving the sensitivity of
progressive multiple sequence alignment through
sequence weighting, position-specific gap
penalties and weight matrix choice. Nucleic Acids
Res (1994) 22: pp. 4673-4680.
[17] Saitou N & Nei M. The neighbor-joining method:
a new method for reconstructing phylogenetic
trees. Mol Biol Evol (1987) 4: pp. 406-425.
[18] The UniProt Consortium. The universal protein
resource (UniProt). Nucleic Acids Res (2008) 36:
p. D190-5.
[19] Zhang Y. I-TASSER server for protein 3D
structure prediction. BMC Bioinformatics (2008)
9: p. 40.
[20] Koradi R, Billeter M & Wuthrich K.
MOLMOL: A program for display and analysis of
macromolecular structures. Journal of Molecular
Graphics (1996) 14: p. 51-55.