EXPLORING PROTEIN FOLDING TRAJECTORIES USING
D. RUSSEL and L. GUIBAS
Computer Science Department
353 Serra Mall
Stanford, CA 94305, USA
We describe the 3-D structure of a protein using geometric spanners — geometric
graphs with a sparse set of edges where paths approximate the n2inter-atom
distances. The edges in the spanner pick out important proximities in the structure,
labeling a small number of atom pairs or backbone region pairs as being of primary
interest. Such compact multiresolution views of proximities in the protein can
be quite valuable, allowing, for example, easy visualization of the conformation
over the entire folding trajectory of a protein and segmentation of the trajectory.
These visualizations allow one to easily detect formation of secondary and tertiary
structures as the protein folds.
There has been extensive work on visualizing the 3-D structure of proteins
in ways that attempt to make the certain aspects of the structure more ap-
parent. For example, commonly used software packages such as RasMol ,
, or SPV , among others, permit visualizations via
hard-sphere models, stick models, and ribbon models that emphasize differ-
ent aspects of the protein surface or secondary structure. Even more abstract
visualizations have been used as a tool for understanding intra-molecular prox-
imities, including contact maps and distance matrix images . None of these
approaches work very well, however, if the goal is to visualize proteins in mo-
tion and not just their static conformations.
Large corpora of molecular trajectories are becoming available through
efforts such as Folding@Home  where molecular simulations are carried
out on distributed networks of many thousands of computers. There is an
increasing need to compare, classify, summarize, and organize the space of such
protein trajectories with an eye toward advancing our understanding of protein
folding by studying their ensemble behaviors. Most currently used methods for
understanding such data revolve around computing a few summary statistics
for each conformation, such as radius of gyration or number of native contacts
and watching how these evolve during each trajectory. More similarly, the
chemical distance, a statistic of an adjacency graph of the amino acids, was
use to differentiate folded and unfolded states . In this paper we explore the
use of a more rich and abstract representation of the protein structure, based
on spanners, which makes the task of understanding and exploring the space
of protein motions easier.
Our basic idea is to take the continuous folding process and map it to
a more discrete combinatorial representation. This representation focuses on
higher-level geometric proximities that tend to form and be more stable over
time rather than atom coordinates or specific aspects of secondary/tertiary
structure. Specifically, we look at the formation of proximities between differ-
ent parts of the protein across a range of scales, and track the changes of such
proximities over time. Our more abstract description of the folding process is
in terms of ‘proximity events’ — when certain proximities are formed or de-
stroyed. Together, these characterize the folding process in a qualitative way
and capture the important aspects of the trajectory, the sequence of conforma-
tions adopted by a protein in a particular folding path. Just as an algebraic
topologist captures the essence of the connectivity of a continuous space in a
few discrete invariants (the homology groups), we aim to capture the signifi-
cant conformational changes during motion through a discrete representation
of proximities that form and break.
We use geometric spanners to accomplish this goal.
abstract graph with weights on its edges, a spanner is a sparse subgraph (in
the sense of having a number of edges roughly proportional to the number of
vertices), such that all edges in the full graph can be well approximated by
paths in the spanner (in the sense that the sum of the weights of edges of the
path in the spanner is very close to the weight of the original graph edge). In
the geometric setting the vertices in the original graph are points each pair
of which is connected by an edge with weight equal to the Euclidean distance
between the corresponding pair of points. The quality of the approximation
can be controlled by varying the number of edges in the spanner.
Note that spanners are at once generalizations of contact maps as well
as compressions of distance matrices. One can think of a spanner as a mul-
tiresolution contact map that allows an approximate reconstruction of the full
distance matrix (and therefore the full 3-D structure as well).
We propose to use these combinatorial structures as a tool for capturing
the important proximities of a protein conformation and, in this paper, for
comparing and visualizing sequences of protein conformations from molecular
trajectories. Key properties of the spanner that facilitate these goals include:
Starting from an
• Spanners are proximity based — this parallels proteins where local in-
teractions determine the behavior.
• Spanners are discrete — they have a combinatorial structure whose de-
scription does not include any geometric coordinates.
• Spanners are controllable — we can produce descriptors that more loosely
or more tightly capture the shape of the protein, converging to distance
matrices as the approximation gets tighter.
• Spanners are uniform — there is only one type of combinatorial element,
namely an edge. This makes comparison, processing and display simpler.
• Spanners can be made smooth — small changes in the protein confor-
mation generally result in few large changes in the spanner, enabling
tracking of the spanner structure over time.
• Spanners are local — the combinatorial features, edges, are affected by a
small subset of the total point set. This means that changes in one part
of the protein do not generally affect the spanner edges in other parts.
As a result the edges can be assigned semantic meaning based on their
endpoints, rather than on larger regions of the protein.
We use our spanners to investigate the folding of the protein BBA5 
using simulation data produce produced by the Folding@Home project. The
spanners enable us to produce diagrams which show the formation (and some-
times dissolution) of secondary and tertiary structure during a whole folding
trajectory and allow us to segment these trajectories into logical parts. We
expect that the spanner approach will provide a valuable toolkit for the un-
derstanding and visualization of protein trajectories.
In the next sections we describe how we construct and smooth our spanner-
based representation and how we use it to visualize trajectories. Then we
discuss our how we have used spanners to try to understand the folding of
BBA5. Finally we mention other promising applications of our spanner based
2Representing Proteins Using Spanners
We first provide a more rigorous definition of a geometric spanner. Let P be a
set of points in R3, Euclidean three-space, and G be a Euclidean graph on P
(graph whose vertices are points from P and whose edge weights are Euclidean
distances between the endpoints of the edge). For a parameter s > 1, known
as the stretch factor, G is a spanner for P, if for all pairs of points i and j
in P with Euclidean coordinates piand pj, πG(i,j) ≤ s||pipj|| where πG(i,j)
denotes the shortest path distance between i and j in the graph G. Thus,
the spanner represents the quadratic number of interpoint distances in P by
the much sparser set of edges in G. There is a vast literature on spanners
that we will not attempt to review in here any detail; many different spanner
constructions are possible. It has been shown that for s arbitrarily close to 1,
there exist spanners whose number of edges is proportional to the size of P.
The reader is referred to a number of survey papers for background material
and additional references [1, 6].
For simplicity, we only use the backbone atoms of the protein. This allows
us to meaningfully identify each atom by its index along the backbone, so an
edge i,j connects the ith and jth atoms in the backbone. We can identify the
edge i,j with the point (i,j) where i < j. We define the distance, between two
edges i0,j0and i1,j1as the L1distance between the points (i0,j0) and (i1,j1),
namely |i0−i1|+|j0−j1|. We will write it d(i0,j0,i1,j1). Two edges are close
if they have a small L1distance between them the corresponding points. The
length as opposed to weight of an edge l(i,j) is defined as j − i. Throughout
the section s will designate the stretch factor. A s-spanner is a spanner with
stretch factor s.
We use what is known in the literature as the ‘greedy’ spanner. Its computation
is conceptually very simple: starting with graph G initially containing only the
points P, test each of the?|P|
πG(i,j). If so, add the edge i,j to the G. We call this test the inclusion test.
The algorithm runs in O(n3) time due to the quadratic number of edges and
the worst case linear time required to evaluate the inclusion test.
This greedy spanner construction has been shown to have asymptotically
optimal complexity (number of edges) and weight (the sum of the lengths of the
edges) as well as good practical complexity and weight . Having low weight is
important in our context since we want the spanner to consist of as many short
edges as possible in order to capture local interactions. Euclidean spanners can
be also be produced in O(nlog2n) time with the same asymptotic edge count
and weight bounds  although we have not implemented such methods.
If implemented naively, performing the inclusion test for long edges domi-
nates the running time as it requires a nearly linear time graph search for each
of these edges. However, such long edges are extremely unlikely to be in a
spanner of a packed protein. If we maintain an upper bound on the graph dis-
tance, dG(i,j) ≥ πG(i,j) between each pair of points, i,j, then we can quickly
eliminate any candidate edge for which s?ij? > dG(i,j). We can similarly
?interpoint candidate edges for inclusion, ordered
from shortest to longest. For each candidate edge i,j, check if s?pipj? <
prune many of the paths while searching.
These upper bounds can be maintained lazily as graph searches are per-
formed. To tighten the upper bounds and further accelerate the process, it is
advantageous to periodically compute the all atoms shortest path distances in
the current spanner (an O(n2) process). In addition, we guide the search using
the Euclidean distance as a lower bound on the graph distance to bias our
search direction, as in the graph search algorithm A∗. Using these heuristics,
the 2-spanner of an 800 atom backbone of a protein can be computed in about
The kinetic spanner proposed in  is a possible alternative. It can be
cheaply maintained as the underlying points move around. However, it is non-
canonical, making comparison between trajectories tricky and it has more long
edges, which are hard to assign biological meaning.
2.2Spanners of Proteins
Figure 1 shows spanners computed using different expansion factors for single
protein and gives an estimate of the number of spanner edges per point for
typical proteins in their native state.
Figure 1: Example spanners and average edge per point for various expansion factors. We
mostly use 2-3 spanners for out computations and visualizations as spanners below 2 get
Secondary structure creates very well defined patterns in the spanner. If
each edge is visualized as a point (i,j), then α helices appear as a sequence of
points (i + kq,j + kq) where k is a counter variable and q is a stepsize which
depends on the expansion factor. For a expansion factors between 2 and 3 the
step size is 3, the edges are just longer when the expansion factor is larger.
For a expansion factor of 1.5, the stepsize is still 3 but there are several edges
leaving from each of the points. β hairpins appear as series of points heading in
an orthogonal direction to helices, namely, (i+kq,j−kq). k is 2 for 2 spanners
and rises to 4 or 6 for 3 spanners (depending on how the hairpin twists). Both
patterns are shown in Figure 2.
(a) Mostly α, 1AQ5(b) Mostly β, 1YTFD2
Figure 2: Helices and hairpins both give rise to a distinctive pattern of spanner edges: The
2-spanners and corresponding point patterns are shown for two proteins. The backbone
edges i,i + 1 are displayed in addition to the other spanner edges.
Some addition processing must be done before we can use the spanner for
detecting the creation and destruction of long-lasting structural patterns and
proximities. Simply computing the spanner for each frame independently re-
sults in a sequence of different structures for each frame as the atoms vibrate.
We need to be able to distinguish between edges which ‘move’ by a small
amount between frames and edges which are entirely new at a given frame.
Once we have done that we can filter the edges based on whether they repre-
sent long-lasting proximities.
We can set up the problem of associating edges from one frame with edges
from the next as a bipartite matching problem. Such problems can be solved
efficiently for a wide variety of similarity metrics. The similarity metric we
used with the most success is to make the score for the edges e0 and e1 be
min(l(e0),l(e1))/(1 + d(e0,e1)). This allows longer edges to move more easily
than shorter ones. In order to avoid spurious matches, we disallow matches of
edges which are far from one another, specifically edges which are more than
min(l(e0),l(e1))/3 + 2 apart. This threshold allows helix and strand edges to
be readily matched with one another over the expansion factors we use, but
cuts off long distance matches.
Using this technique we can associate edges from successive pairs of frames.
We call a sequence of paired edges a proximity. The lifetime of a proximity
is how many different consecutive frames contain edges that contributed to
that proximity. We can now talk about the creation and destruction of a
proximity—exactly the type of proximity events we mentioned in the intro-
Many of the proximities tracked using the previously mentioned technique will
have very short lifetimes, perhaps only one frame. These short-lived proximi-
ties form a sort of topological noise, which can obscure the real signal in the
In order to remove this noise we turn to the idea of persistence . A
proximity is persistent if its lifetime exceeds a certain threshold. Persistent
proximities represent stable aspects of the conformation of the protein back-
bone. We simply drop non-persistent proximities.
There are certain cases where persistence filtering removes too many edges.
For example, if the protein conformation is such that two candidate edges, both
of which independently pass the inclusion test, have nearly the same length,
then small perturbations can cause either edge to be included in the span-
ner. The spanner can rapidly alternate between the two edges in succeeding
frames, resulting in neither of them being persistent and both of the edges
being dropped. Dropping them both looses too much information about the
protein conformation. To solve this, in addition to matching edges from the
current frame against the edges from the previous frame, we add in an edge
from each proximity that was destroyed in a recent frame. As a result, in the
previously mentioned example of two edges flipping back and forth, both will
now be included.
In practice, we discard proximities which exist for fewer than 20 frames
and allow edges to be matched against edges which disappeared up to 6 frames
3.3 Segmenting Trajectories
We can now define two spanners as being close if they have similar persis-
tent proximities. This gives us a means of segmenting trajectories into logical
components. To do this, assign each consecutive pair of trajectories a spanner
distance—the sum of the lengths of the persistent proximities which are cre-
ated or destroyed between this pair of trajectories. We can then smooth these
distances in time and use local maxima to segment the trajectory. We will
discuss the segmentations produced by this method more in the next sections
once we present our technique for displaying spanners.
4 Encoding Spanners in One Dimension
Simply displaying the spanner along with the backbone helps make interac-
tions easier to pick out, especially in less ordered states. However, by itself it
does make it easy to visualize the whole trajectory since each frame must be
displayed successively. To avoid that problem we want to encode the spanner
as a one dimensional object so that we can lay out a whole series of them for
Figure 3: Encoding a spanner into a strip spanner: a 2-spanner and corresponding strip
spanner are drawn and some edge and structure correspondences are indicated.
To encode a spanner for a frame, take each spanner edge which is part of a
persistent proximity and mark the first point of the edge with the length—i.e.
for edge i,j : i < j set S[i] = l(i,j). We call the resulting vector, S a strip
spanner. An example is shown in Figure 3.
One drawback of the strip spanner encoding is that it cannot directly
handle more than one spanner edge originating a single point. This can be
remedied in one of several manners. One solution is to expand each backbone
atom into a constant number of points, i.e.
This allows a constant number of edges per atom.
construction can enforce the one edge per point restriction. For expansion
factors 2 and above, this does not significantly distort the spanner, however,
it causes problems for factors much below 2 as there are too many edges. For
the purposes of generating strip spanners for this paper we simply take the
maximum length edge originating at each point. For a small protein such as
BBA5 where there are few tertiary edges, this is quite adequate.
Strip spanners from successive frames can then be stacked creating a view
of the trajectory as a whole. We call this stacking a strip history. An example
is shown in Figure 4.
The strip history gives us a good way of displaying the trajectories which
were segmented using the technique presented in Section 3.3. An example
segmentation is shown in Figure 5. The results of segmentation corresponds
atom i goes to i − ?,i,i + ?.
Edge length vs color
Figure 4: A strip history: In this
trajectory, the protein does not
fold completely, but forms sig-
nificant secondary and tertiary
structure.The patterns for α
helical secondary structure are
quite easily seen (lines 3 units
apart) on the right half of the
diagram. The β hairpin pattern
is slightly less easily picked out
on the left (lines 2 units apart).
Long tertiary edges show up as
the darker pixels on the left half
of the diagram.
quite well with manual segmentations we had previously performed using the
Figure 5: Subdivision of a trajec-
tory using the strip history:
non-folding trajectory for BBA5
was divided into 7 phases using
the strip history.
mation for the middle frame of
each phase is shown, always ori-
ented with the α-helical end to
the right. During the first seg-
ment the protein has little to no
ond the first part of the alpha he-
lix forms and stays fairly static
through the third. In the fourth
segment we see the hairpin form-
ing, but part of the helix disinte-
grates. The hairpin disintegrates
in the fifth. In the seventh the
hairpin reforms, and the end of the
helix forms for the first time.
In the sec-
5Understanding BBA5 Trajectories
We applied our spanner descriptor to the task of trying to understand the
folding process of the protein BBA5. It is a 23 residue protein which folds
comparatively quickly. The native state of BBA5 consists of a β hairpin,
involving the first 8 residues and an α helix involving the last 10 residues.
The hairpin is packed against the helix, although not very tightly due to the
presence of large sidechains between the two pieces of secondary structure.
Our data consists of 23 trajectories with frames every 200fs. Thirteen of
the trajectories come within 3.1˚ A cRMSDCαof the native structure.
Figures 4 and 5 show strip histories for two example trajectories. The times
of formation of the α helix are quite clearly visible. The β hairpin formation
is less clearly defined, but still can be seen on the left of the image, as can the
occasional large backbone distance tertiary interactions, which are the points
colored the darkest.
Perusal of the strip histories show that there are few constants among the
folding trajectories, as might be expected given its small size.
• α helical secondary structure rarely went away once created, although
there were a few cases where parts of the helix formed and then dissolved
(such an example can be seen in Figure 5). This may be an artifact of
how trajectories were selected for storage. Parts of the helix formed in
all possible orders.
• β hairpin structure was much less stable than the helix, however it did
stabilize without the presence of tertiary interactions in several of the
• tertiary interactions formed before secondary structure in some but after
strong secondary structure in others.
In one of the strip histories a helix like pattern can be seen forming in
the loop end of the hairpin. It persists for 30 frames. Direct inspection of the
backbone structure confirmed that there is indeed a one and a half turn helix
like structure there. This trajectory is not shown here. We did not observe
similar structure formation in any of the other trajectories.
Sidechains play an important role in the protein folding process. For example,
in BBA5, the rings from a phenylalanine and a tyrosine occupy much of the
space between the helix and the hairpin in the native state, making the helix
and hairpin pack much less tightly than they would in the absence of such
rings. The position of the sidechains is currently ignored in our calculations,
although it may be important to the folding process and ignoring it makes
the tertiary structure show up much less clearly. Simply adding the whole
sidechains adds too many new edges to the spanner making the relevant data
hard to extract and disrupts the linear order which we depend on for matching
and visualization. A better approach may be to add a single point per sidechain
which will capture its location without complicating the structure too much.
The strip spanner is a one dimensional descriptor which captures key as-
pects of the proteins conformation. Searching and matching of one dimensional
structures is a much easier problem than matching three dimensional curves,
suggesting that the strip spanner might have applications in protein structure
motif searching and structure alignment. However, there are a number of issues
with incorporating gaps which need to be resolved.
We are trying to apply spanners to the problem of understanding the parts
of protein conformation space relevant to folding. The strip history based seg-
mentation provides one way of dividing trajectories into chunks which could
be matched against one another to find common paths through conformation
space. There are a number of problems with measuring the distances between
spanners which need to be resolved first. In addition, we suspect simple pro-
teins such as BBA5 fold too quickly and have too small an energy barrier for
its fold space to have significant structure. As a result we plan to apply the
techniques to unfolding data.
This work has been supported by NSF grants CARGO 0310661, CCR-0204486,
ITR-0086013, ITR-0205671, ARO grant DAAD19-03-1-0331, as well as by the
Bio-X consortium at Stanford.
The authors would like to thank Rachel Kolodny for many valuable sug-
gestions regarding the project and Vijay Pande for providing the data.
 S. Arya, G. Das, D. Mount, J. Salowe, and M. Smid. Euclidean spanners:
short, thin, and lanky. In Proceedings of the twenty-seventh annual ACM
symposium on Theory of computing, pages 489–498. ACM Press, 1995.
 G. Das, P. Heffernan, and G. Narasimhan. Optimally sparse spanners in 3-
dimensional Euclidean space. In Symposium on Computational Geometry,
pages 53–62. ACM Press, 1993.
 G. Das and G. Narasimhan. A fast algorithm for constructing sparse
Euclidean spanners. In Symposium on Computational geometry, pages
132–139. ACM Press, 1994.
 Nikolay Dokholyan, Lewyn Li, Feng Ding, and Eugene Shakhnovich.
Topolical determinants of protein folding. Proceedings of the National
Academy of Science, pages 8637–8641, 2002.
 H. Edelsbrunner, D. Letscher, and A. Zomorodian. Topological persistence
and simplification. Discrete Compututational Geometry, 28:511–533, 2002.
 D. Eppstein. Spanning trees and spanners. Handbook of Computational
Geometry, pages 451–461, 2000.
 J. Gao, L. J. Guibas, and A. Nguyen. Deformable spanners and applica-
tions. In Symposium on Computational Geometry, pages 179–199, June
 N. Guex and M. C. Peitsch. Swiss-model and the swiss-pdbviewer: An
environment for comparative protein modeling. Electrophoresis, 18:2714–
2723, 1997. URL http://www.expasy.org/spdbv/.
 E. Martz.
Protein explorer: Easy yet powerful macromolecular visu-
Trends in Biochemical Sciences, 27:107–109, 2002.URL
 Rasmol. URL http://www.umass.edu/microbio/rasmol/.
 Y.M. Rhee, E. J. Sorin, G. Jayachandran, E. Lindahl, and V. Pande.
Simulations of the role of water in the protein-folding mechanism. In
Proceedings of the National Academy of Science, volume 101, pages 6456–
 M. Shirts and V. Pande. Screen savers of the world, unite! Science, 2000.
 M. J. Sippl. On the problem of comparing protein structures. Journal
Molecular Biology, 156:359–388, 1982.