Estimating Genome Reversal Distance by Genetic Algorithm

Andy Auyeung Ajith Abraham

Oklahoma State University Oklahoma State University
Math Science 219 North Hall 328
Stillwater, OK 74078 Tulsa, OK 74106
1 (405) 744–5668 1 (918) 594–8188
[email protected] [email protected]

Abstract- Sorting by reversals is an important

problem in inferring the evolutionary relationship
between two genomes. The problem of sorting
unsigned permutation has been proven to be NP-hard.
The best guaranteed error bounded is the 3/2-
approximation algorithm. However, the problem of
sorting signed permutation can be solved easily. Fast
algorithms have been developed both for finding the
sorting sequence and finding the reversal distance of
signed permutation. In this paper, we present a way to
view the problem of sorting unsigned permutation as
signed permutation. And the problem can then be seen
n
as searching an optimal signed permutation in all 2
corresponding signed permutations. We use genetic
algorithm to conduct the search. Our experimental
result shows that the proposed method outperform the
3/2-approximation algorithm.

Figure 1. Transformation from Tabacco to Lobelia fervens by

1 Introduction reversals (Bafna and Pevzner; 1996).

Genome Rearrangement is a mechanism that happens in

mitochondrial genomes (Russell 2002). The genes order sort π (or π’) into identity permutation (i.e. the
in mitochondrial genome is constantly under permutation, 1 2 … n for unsigned permutation, and +1
rearrangement. Therefore, by estimating the +2 … +n for signed permutation). We called the
rearrangement distance between two genomes, the minimum number of reversal the reversal distance.
relationship between them can also be estimated (Pevzner The problem of sorting signed permutation can be
2001). Reversal is the most commonly seen mechanism 2
solved in O( n ) time (Kaplan et al. 1997). The problem
that genomes are rearranged. Figure 1 shows the
of finding the reversal distance of signed permutation can
estimated transformation from Tobacco to Lobelia fervens
be solved in O(n) time (Bader et al. 2001). However, both
by reversals (Bafna and Pevzner; 1996). There are two
sorting and finding the reversal distance of unsigned
variations of this problem, signed permutation and
permutation has been proven to be NP-hard (Caprara
unsigned permutation. For unsigned permutation, a
1997). So, error bounded heuristic solutions have been
genome is modeled as a permutation π with order n (i.e. a
proposed (Bafna and Pevzner 1996; Kececioglu and
permutation of {1, 2, …, n}), where n is the number of
Sankoff 1995). The lowest guaranteed error bound thus
gene blocks in the genome. Let the permutation π = π[1]
far is the 3/2-approximation algorithm (Christie 1998).
π[2] … π[n], the reversal operation ρ(i,j) rearrange π into
The 3/2-approximation algorithm uses the fact that any
π[1] … π[i-1] π[j-1] … π[i] π[j] … π[n]. For signed
cycle decomposition of the breakpoint graph that
permutation π’, each π’[k] has either a positive or a
maximize the number of 2-cycles exists a sorting
negative sign. Each reversal operation ρ(i,j) not only
sequence with length at most 3/2 of the optimal sorting
rearrange π’ but also negate the sign of π’[k] for i ≤ k < j.
sequence.
The problem of estimating reversal distance between two
In this paper, we propose a genetic algorithm for
genomes is formulated as sorting permutation by reversal
sorting unsigned permutation by reversal. Our method
operation. That is, given π (or π’), we want to find a
does not provide guaranteed error bound. However, our
sorting sequence that uses minimum number of reversal to
experiment shows that it finds better solution than the 3/2-
approximation algorithm. Also, so far, all heuristic
algorithms for this problem use a constructive manner to
find the solution. We would like to show an alternative
approach how this problem can be solved in inductive
manner.
The rest of the paper is organized as the following.
In Section 2, some background materials on sorting
permutation by reversal and the concept of genetic
Figure 2. Breakpoint graph for unsigned permutation 4 1 3 2.
algorithm are presented. In Section 3, the proposed
method is explained. In Section 4, the experimental setup
and results are shown. In Section 5, observations from the
experiment are discussed. Finally in Section 6, some
concluding remarks are made.

2 Reviews

2.1 Breakpoint Graph

An unsigned permutation can be modeled by a breakpoint
graph. For each gene (a number in the permutation), we
will create a node for it. The idea of breakpoint graph is to
mark the desired and realistic relationship between these
nodes in the permutation. For each pair of the nodes we
draw a black edge between them if they are adjacent in
the permutation, and we draw a red edge between them if Figure 3. Breakpoint graph for signed permutation +4 -1 +3 -2.
they are adjacent in the identity permutation. In order to
model the orientation, we expand the unsigned
permutation to have a zero at the front and a n+1 at the is to use evolutionary model to determine the next
end. An example is shown in Figure 2. It has been shown searching area from that is the next population, is
that given a cycle decomposition of the breakpoint graph, generated by selection, crossover and mutation operation
any reversal can at most change the number of cycles by according to the fitness of solutions in the current
one. Besides, it has also been shown that given a cycle population.
decomposition of the breakpoint graph, the corresponding The norm of the genetic algorithm can be describe
shortest sorting sequence can then be found. Thus, the key as the following. A population of possible solutions is
problem is to find a cycle decomposition that provides the initially generated. The algorithm is divided in
shortest sorting sequence for the unsigned permutation. generations. In each generation, if the termination
However, the problem of finding an optimal cycle condition has not met, next population will be first
decomposition is NP-hard. determined by selection and crossover. In selection,
On the other hand, sorting signed permutation individuals in the current population are probabilistically
can be solved easily. We first create the breakpoint graph selected according to their fitness to move to the next
as above. Then for each gene node, we split it into two population. In crossover, pairs of individuals are
nodes according to its sign (and we shift to number of the probabilistically selected according to their fitness to
node accordingly). An example is shown in Figure 3. The generate a new pair of individuals (offspring) by the
advantage is that now each node has exactly one red edge crossover operator. Once the next population has been
and one black edge associated with it. Thus, there is only determined, then mutation operation will be
one cycle decomposition of this breakpoint graph. Many probabilistically applied to individuals in the next
algorithms have been proposed to find the sorting population. Finally, we re-label the next population to be
2 the current population that is to symbolize the old
sequence. The best known time bound is O( n ) time.
population has died out. And the fitness of the new
And the best known time bound time for finding the
individuals will be evaluated. In each generation, only fit
reversal distance is O(n).
individuals can produce offspring and survive. Thus, a
2.2 Genetic Algorithm population of fit solutions would be expected when the
Genetic algorithm is a searching technique. The idea is algorithm terminates. And the best individual would be
inspired by natural selection happens in evolution. A used as the solution to the problem.
genetic algorithm works with a population of individuals,
each representing a solution to a given problem. The idea
3 Proposed Method π is the sorting sequence of π*. However, it is not feasible
n
to go through all 2 π’. Thus, we use genetic algorithm
The idea of the proposed method is to view all the to find π*. There is no guarantee that the genetic
possible cycle decomposition of the unsigned permutation algorithm would find π*, however, we could expect the
as the signed permutations that have the same gene order. genetic algorithm would find a π’ that has low reversal
Except node 0 and n+1, every node in the breakpoint distance.
graph has degree 2 for red edges and also for black edges.
So, a cycle decomposition is used to define the (color)
alternating paths. However, there is a corresponding
4 Experiment
signed permutation that actually defines the same
alternating paths. Therefore, we can now turn our focus
The genetic algorithm is the following. We allow the
on signed permutation instead of cycle decomposition. 2
Define the set Signed(π) be the set of signed permutations population size to be n . The initial population is
that have the same gene order as π. For example, when π randomly generated binary strings representing the signs
is 2 1, then Signed(π) is { -2 -1, -2 +1, +2 –1, +2 +1}. of the genes in the permutation. However, we apply a
Thus, the size of Signed(π) is 2 . The following two
n heuristic on taking all trivial cycles (i.e. cycles that
observations are required for our method. compose of exactly one red edge and one black edge).
Thus, sorted substrings would be assigned to the same
sign (positive for ascending sorted substring; negative for
Observation 1: descending sorted substring). The fitness is evaluated by
Each π’ ∈ Signed(π) can deduces a valid the reversal distance of this signed permutation. Single
sorting sequence for π. point crossover and mutation are used with rate 0.5 and
0.3 respectively. And the genetic algorithm is terminated
Proof: when the best reversal distance in the population remains
Let sorting sequence ρ sort π’ into identity unchanged in three generations.
(i.e. π ' ρ = id). Because |π’[i]| = π[i], then π ' ρ [i] We conduct the experiment by randomly generated
permutations, where permutations are generated by
= π ρ [i], for all i. Thus, ρ can also sort π (i.e. performing n random swap operations on the identity
π ρ = id). permutation. Figure 4 shows the comparison between the
3/2-approximation algorithm and the proposed method.
(The actual data is shown in Appendix A.) The figure
shows the number of reversals required in the sorting
sequences found by the two methods. The comparison is
Observation 2: on the average solution of ten runs with n is between 10
There exist π* ∈ Signed(π) that deduces to 150. We can see that the proposed method produce
an optimal sorting sequence for π. better solution than the 3/2-approximation algorithm.

Proof:
Let ρ be a sorting sequence for π that uses
minimum number of reversals. For each π[k], let
count[k] be the number of times that π[k] is
included in reversals of ρ, i.e. ρ(i,j) i ≤ k < j. Then
π* is the following, π*[k] has a positive sign if
count[k] is even, otherwise it has a negative sign.
Because |π’[i]| = π[i], π ρ [i] =| π ' ρ [i]|, for all i.
However, all π ' ρ [i], must be positive by our
construct. Thus ρ can also π* and the reversal
distance for π* is equal to the reversal distance for Figure 4. Performance comparison of 3/2-approximation and
π. genetic algorithm.

From the two observations, we can see that the 5 Discussion

n
problem can be solved in O( 2 n) time. That is to find the
n Sorting unsigned permutation has a trivial n-1 upper
reversal distance for all 2 π’. Let π* be the π’ that has
bound. That is, simply using one reversal to put one gene
minimum reversal distance. Then the sorting sequence of
block in place. When n-1 gene blocks are in place, the n-
th block is already in place. In fact, this is the reversal 6 Conclusion
distance for the “hard-to-sort” Gollan permutation
γ n (and γ n−1 ). From Figure 4, we can see that the Sorting unsigned permutation by reversals served an
solutions found by both the proposed method and the 3/2- important role in inferring evolutionary history. A 3/2-
approximation are not very far from the n-1 upper bound. approximation has been proposed and is the lowest
That observation is consistent with the analysis done by guaranteed error bound thus far. All previous methods use
Bafna and Pevzner that says the reversal distance between constructive approach to produce the solution. This paper
2 random permutations is very close to the n-1 upper introduces a new inductive approach that uses genetic
bound. algorithm to find the solution. Experimental result shows
that the proposed method outperforms the 3/2-
Figure 4 shows that the solutions produced by
approximation algorithm, although it does not
both the proposed method and the 3/2-approximation are
mathematically guarantee the quality of the solution. Due
growing linearly with n. Although the experiment shows
to the inductive perspective, different searching methods
that the proposed method only outperform the 3/2-
can then be applied to solve the problem. Also, this
approximation by a small distance, we can see that the
perspective creates possibilities to estimate the reversal
3/2-approximation has a steeper slope than the proposed
distance by sampling the solution space. Such estimation
method. In another word, the improvement from the
allows the estimated answer on both higher and lower
proposed method scales up with n.
sides, while the previous methods always produce over
During our implementation, we found our
estimated answer.
heuristically initialized population enhances the average
performance of the genetic algorithm. In fact, sorted
substrings do not always stay together. Notice when we Acknowledgements
optimally sort the permutation 3 4 1 2, one of the sorted
substrings (i.e. 3 4 or 1 2) has to be broken. However We would like to thank David Christie for all his help
interestingly, it has been proven that there exists a sorting during our implementation of the 3/2-approximation
sequence that does not break sorted substring with length algorithm. We also would like to thank Itsik Mantin and
longer than 2 (Hannenhalli and Pevzner, 1996). Ron Shamir make available to public the Java applet and
Our genetic operators and parameters to the the source code for the sorting signed permutation
genetic algorithm are not carefully chosen. The main goal algorithm.
of this paper is to introduce the representation that the
sorting unsigned permutation problem can be solved by
genetic algorithm. When selecting the genetic operators Bibliography
and parameters of the genetic algorithm, we believe it
should be domain specific. For example, a mathematician Bader, D. A., Moret, B. M. E. and Yan, M. (2001) “A
may be interested in purely random permutations; while a Linear-Time Algorithm for Computing Inversion
biologist may be interested in biological shuffled Distance between Signed Permutations with an
permutations. And it is natural to expect that different Experimental Study,” Journal of Computational Biology,
permutation groups have different characteristics in their 8(5):483-491.
breakpoint graphs. So, different heuristic would be
required in the genetic algorithm. Bafna, V. and Pevzner, P. A. (1996) “Genome
The computational time was not our major Rearrangements and Sorting by Reversals,” SIAM Journal
concern during the implementation of the proposed on Computing, 25(2):273-289.
2
method. So, we used the O( n ) time Java implementation Bergeron, A. (2001) “A Very Elementary Presentation of
created by Itsik Mantin to find the reversal distance of the Hannenhalli-Pevzner Theory,” The Twelfth Annual
signed permutations for the fitness evaluation (Mantin Symposium on Combinatorial Pattern Matching, 106-117.
and Shamir, 1999). However, O(n) time algorithm is
known and can be used. Here we provide a time Berman, P. and Hannenhalli, S. (1996) “Fast Sorting by
complexity comparison. The 3/2-approximation requires Reversal,” The Seventh Annual Symposium on
2
O( n ) time. On the other hand, the proposed method Combinatorial Pattern Matching, 1075:168-185.
2
requires O( n ) size of population; each requires O(n)
time to evaluate its fitness in each generation; and there Caprara, A. (1997) “Sorting by Reversals is Difficult,”
can be at most O(n) generations (because the upper bound Proceedings of the First International Conference on
for reversal distance is n-1 and the genetic algorithm Computational Molecular Biology, 75-83.
enforces improvement on the best reversal distance in
every 3 generations otherwise terminates). Therefore, the Christie, D. A. (1998) “A 3/2-Approximation Algorithm
4 for Sorting by Reversals,” Proceedings of the Ninth
proposed method requires O( n ) time.
Annual ACM-SIAM Symposium on Discrete Algorithms,
244-252. Sankoff, D., Leduc, G., Antoine, N., Paquin, B., Lang, B.
and Cedergren, R. (1992) “Gene Order Comparisons for
Diestel, R. (2000) “Graph Theory,” Springer Verlag. Phylogenetic Inference: Evolution of the Mitochondrial
Genome,” Proceedings of the National Academy of
Hannenhalli, S. and Pevzner, P. A. (1995) “Transforming Science USA, 89:6575-6579.
Cabbage into Turnip,” Proceedings of the Twenty-seventh
Annual ACM Symposium on the Theory of Computing, Slinko, A. (2002) “Combinatorics. Tutorial: 1
178-189. Permutations,” http://www.matholymp.com/ .

Hannenhalli, S. and Pevzner, P. A. (1995) “Transforming Tao, J., Ying, X. and Zhang, M. Q. (2002) “Current
Men into Mice,” Proceedings of the Thirty-sixth Annual Topics in Computational Molecular Biology,” MIT Press.
IEEE Symposium on Foundations of Computer Science,
581-592.
Appendix A
Hannenhalli, S. and Pevzner, P. A. (1996) “To Cut … or
not to Cut,” The Seventh Annual ACM-SIAM
n 3/2-approximation Genetic algorithm
Symposium on Discrete Algorithms, 304-313.
10 7.6 5.9
Kaplan, H., Shamir, R. and Tarjan, R. E. (1997) “Faster 20 15.8 13.3
and Simpler Algorithm for Sorting Signed Permutations 30 25.8 22.2
by Reversals,” Proceedings of the Eighth Annual ACM- 40 36.1 32.1
SIAM Symposium on Discrete Algorithms, 344-351. 50 43.9 40.4
60 54.9 50.6
Kececioglu, J. and Ravi, R. (1995) “Of Mice and Men: 70 65.1 60.4
Evolutionary Distance between Genomes under 80 73.9 69.3
Translocation,” Proceedings of the Sixth Annual ACM- 90 82.8 78.7
SIAM Symposium on Discrete Algorithms, 604-613. 100 93.5 89.2
110 104.5 99.0
Kececioglu, J. and Sankoff, D. (1995) “Exact and
120 112.8 107.8
Approximation Algorithms for Sorting by Reversals, with
130 124.1 118.0
Application to Genome Rearrangement,” Algorithmica,
140 132.9 128.1
13:180-210.
150 140.7 136.0
Lovasz, L. and Plummer, M. D. (1986) “Matching
Table 1. Permformance comparison of 3/2-approximation and
Theory,” North-Holland.
genetic algorithm.

Mantin, I. and Shamir, R. (1999) “An Algorithm for

Sorting Signed Permutations by Reversals,”
http://www.math.tau.ac.il/~rshamir/GR/ .

Pevzner, P. A. (2001) “Computational Molecular Biology

An Algorithmic Approach,” MIT Press.

Russell, P. J. (2002) “iGenetics,” Benjamin Cummings.

Russell, S. J. and Norvig, P. (2002) “Artificial

Intelligence: A Modern Approach,” 2nd ed. Prentice Hall.

Sankoff, D. (1992) “Edit Distance for Genome

Comparison based on Non-local Operations,” The Third
Annual Symposium on Combinatorial Pattern Matching,
644:121-135.

Sankoff, D., Cedergren, R. and Abel, Y. (1990) “Genomic

Divergence through Gene Rearrangement,” Molecular
Evolution: Computer Analysis of Protein and Nucleic
Acid Sequences, 26:428-438.