# Homomorphisms of Multisource Trees into Networks with Applications to Metabolic Pathways

**ABSTRACT** Network mapping is a convenient tool for comparing and exploring biological networks; it can be used for predicting unknown pathways, fast and meaningful searching of databases, and potentially establishing evolutionary relations. Unfortunately, existing tools for mapping paths into general networks (PathBlast) or trees into tree networks allowing gaps (MetaPathwayHunter) cannot handle large query pathways or complex networks. In this paper we consider homomorphisms, i.e., mappings allowing to map different enzymes from the query pathway into the same enzyme from the networks. Homomorphisms are more general than homeomorphism (allowing gaps) and easier to handle algorithmically. Our dynamic programming algorithm efficiently finds the minimum cost homomorphism from a multisource tree to directed acyclic graphs as well as general networks. We have performed pairwise mapping of all pathways for four organisms (E. coli, S. cerevisiae, B. subtilis and T. thermophilus species) and found a reasonably large set of statistically significant pathway similarities. Further analysis of our mappings identifies conserved pathways across examined species and indicates potential pathway holes in existing pathway descriptions.

**0**Bookmarks

**·**

**97**Views

- [Show abstract] [Hide abstract]

**ABSTRACT:**In this article we address two pattern matching problems which have important applications to bioinformatics. First we address the topology-free network query problem : Given a set of labels L , a multiset P of labels from L , a graph H=(VH,EH)H=(VH,EH) and a function LabelH:VH→2LLabelH:VH→2L, we need to find a subtree S of H which is an occurrence of P . We provide a parameterized algorithm with parameter k=|P|k=|P| that runs in time O⁎(2k)O⁎(2k) and whose space complexity is polynomial. We also consider three variants of this problem. Then we address the alignment network query problem: Given two labeled graphs P and H, we need to find a subgraph S of H whose alignment with P is the best among all such subgraphs. We present two algorithms for cases in which P and H belong to certain families of DAGs. Their running times are polynomial and they are less restrictive than algorithms that are available today for alignment network queries. Topology-free and alignment network queries provide means to study the function and evolution of biological networks, and today, with the increasing amount of knowledge regarding biological networks, they are extremely relevant.Journal of Discrete Algorithms 07/2014; - SourceAvailable from: Robert Harrison[Show abstract] [Hide abstract]

**ABSTRACT:**Network mappings has been commonly applied on it to detect con- served subnetworks across species in biological network. Our previous work fo- cused on finding best mapping of pattern graph to text graph in which vertices in the image of pattern graph are allowed to be deleted (1),(2). Arising from the requirement of allowing deleting pattern vertices in mapping, we reformulate the graph comparison problem to be the one of a homo-home morphism mapping among arbitrary graphs and further analyze and compare our work with previous ones. - SourceAvailable from: digitalarchive.gsu.edu[Show abstract] [Hide abstract]

**ABSTRACT:**The network-mapping tool integrated with protein database search can be used for filling pathway holes. A metabolic pathway under consideration (pattern) is mapped into a known metabolic pathway (text), to find pathway holes. Enzymes that do not show up in the pattern may be a hole in the pattern pathway or an indication of alternative pattern pathway. We present a data-mining framework for filling holes in the pattern metabolic pathway based on protein function, prosite scan and protein sequence homology. Using this framework we suggest several fillings found with the same EC notation, with group neighbors (enzymes with same EC number in first three positions, different in the fourth position), and instances where the function of an enzyme has been taken up by the left or right neighboring enzyme in the pathway. The percentile scores are better when closely related organisms are mapped as compared to mapping distantly related organisms.

Page 1

Homomorphisms of Multisource Trees into Networks with Applications to

Metabolic Pathways

Qiong Cheng, Robert Harrison, Alexander Zelikovsky∗

Department of Computer Science, Georgia State University, Atlanta, Georgia 30303

Email: cscqxcx, rharrison, alexz@cs.gsu.edu

Abstract

Network mapping is a convenient tool for comparing and

exploring biological networks; it can be used for predict-

ing unknown pathways, fast and meaningful searching of

databases, and potentially establishing evolutionary rela-

tions. Unfortunately, existing tools for mapping paths into

general networks (PathBlast) or trees into tree networks

allowing gaps (MetaPathwayHunter) cannot handle large

query pathways or complex networks.

In this paper we consider homomorphisms, i.e., map-

pings allowing to map different enzymes from the query

pathway into the same enzyme from the networks.

momorphisms are more general than homeomorphism (al-

lowing gaps) and easier to handle algorithmically. Our

dynamic programming algorithm efficiently finds the min-

imum cost homomorphism from a multisource tree to di-

rected acyclic graphs as well as general networks.

We have performed pairwise mapping of all pathways

for four organisms (E. coli, S. cerevisiae, B. subtilis and T.

thermophilus species) and found a reasonably large set of

statistically significant pathway similarities. Further anal-

ysis of our mappings identifies conserved pathways across

examined species and indicates potential pathway holes in

existing pathway descriptions.

Ho-

Availability: The software is available from the authors

on request.

1. Introduction

The explosive growth of cellular network databases re-

quires novel analytical methods constituting a new interdis-

ciplinary area of computational systems biology. The main

problems in this area are finding conserved subnetworks,

integrating interacting gene networks, protein networks

and biochemical reactions, discovering critical elements or

∗Correspondence author.

modules and finding homologous pathways. With the im-

mense increase of good-quality data from high-throughput

genomic and proteomic technologies, studies of these ques-

tions are more and more challenging from analytical and

computational perspectives.

Network mapping is a convenient tool for comparing and

exploring biological networks. When mapping metabolic

pathways by matching similar enzymes and chemical reac-

tions chains we can match homologous pathways. Network

mapping can be used for predicting unknown pathways, fast

and meaningful searching of databases, and potentially es-

tablishing evolutionary relations.

Let the pattern be a pathway for which we are search-

ing for homologous pathways in the text, i.e., the known

metabolic network of a different species (see Figure 1).

This problem includes the Isomorphic Embedding problem,

therefore it is NP-hard (see [8]). Given a linear length-?

pathway as the pattern and a graph as the text, PathBlast

(see [9, 8, 14]) finds the image of the pattern in the text such

that no consecutive mismatches or gaps on the pattern and

the text are allowed. The path-to-path mapping algorithm

builds a global alignment graph and decomposes it to linear

pathways mapping.

A single enzyme in one pathway may replace a few se-

quential enzymes in homologous pathway and vise versa.

MetaPathwayHunter [12, 13] finds the optimal homeomor-

phic tree-to-tree mapping allowing an arbitrary number of

gaps. In contrast to the previous approaches, we allow for

the mapping of different enzymes from the pattern into the

same enzyme from the text while keeping the freedom to

mapasingleedgefromthepatterntoapathinthetext. Such

mappings (homomorphisms) are more general than homeo-

morphisms and easier to handle algorithmically.

Our contributions include: (1) efficient dynamic pro-

gramming based algorithm and its implementation finding

the minimum cost homomorphism from multisource trees

into arbitrary networks, (2) a new protein similarity score

scheme based on 4-digit EC enzyme hierarchy, (3) experi-

mental pairwise comparison of all pathways in four differ-

ent organisms (E. coli, S. cerevisiae, B. subtilis and T. ther-

Page 2

6.3.4.3

6.3.2.121.5.1.3 2.1.2.1

1.5.1.20

2.1.1.13

2.1.1.45 0.0.0.0

6.3.2.17

3.5.1.10

1.5.1.5 3.5.4.9

1.5.1.53.5.4.96.3.4.3

Pattern : Formaldehyde oxidation V pathway in B. subtilis

Text : Formy1THF biosynthesis pathway in E. coli

Figure 1. An example of network mapping to find an im-

age of pattern in text.

mophilus) resulting in a reasonably large set of statistically

significant pathway similarities, (4) identification of path-

ways conserved across examined species, potential holes in

existing pathway descriptions, and an estimation of the evo-

lutionary relationship between examined species.

The remainder of the paper is organized as follows. The

next section describes previous work. Section 3 presents the

proposed models for mapping with the definition of pro-

tein similarity score scheme. Section 4 introduces neces-

sary definitions and graph-theoretical problem formulation.

Section5presentsourdynamicprogrammingalgorithmand

analyzes its runtime. Section 6 describes our computational

study of metabolic pathways of four organisms. The analy-

sis and validation of experimental study is given in Section

7. Finally, we draw conclusions in Section 8.

2Previous Work

The earlier papers comparing pathways did not take into

account their topology and instead focused only on sim-

ilarity between proteins or genes. Biochemical pathway

similarity was defined in terms of sequence similarity of

involved genes (see [5]); a multiple pair-wise comparison

algorithm utilizing 4-digit EC enzyme hierarchy was pro-

posed in [15]. The alignment of linear pathways was re-

duced to the sequence alignment problem in [2].

A series of papers [9, 8, 14] has taken into account

the nonlinearity of protein network topology and formu-

lated the mapping problem as follows.

length-? pathway pattern T = (V T,ET) and text graph

G = (V G,EG), find an image of the pattern in the text

without consecutive gaps and minimizing mismatches be-

tween proteins. A global alignment graph in [9] was built

in which each vertex represents a pair of proteins and

each edge represents a conserved interaction, gap, or mis-

Given a linear

matches; their objective is to find the k-highest-scoring

path with limited length ? and no consecutive gaps or mis-

matchesbasedonthebuiltglobalgraph. Theapproachtakes

O(|V T|?+2|V G|2)).

PathBlast with the same problem formulation as [9] was

presented in [8]. However, PathBlast’s solution is to ran-

domly decompose the text graph into linear pathways which

are then aligned against the pattern and then to obtain op-

timal mapping based on standard sequence alignment algo-

rithms. The algorithm requires O(?!) random decomposi-

tions to ensure that no significant alignment is missed, ef-

fectively limiting the size of the query to about six vertices.

In [13], metabolic pathways were modeled as outgo-

ing trees; the problem was reduced to the approximately

labeled tree homeomorphism problem.

bottom-up dynamic programming algorithm has runtime

O(m2n/logm+mnlogn) where m and n are the number

of vertices in pattern and text, respectively.

In [11] the problem was formulated as an integer

quadratic problem to obtain the global similarity score

based on the mapping of as many as node-to-node similar-

ity and as many as edge-to-edge similarity. An exhaustive

searching approach was employed in [17] to find the vertex-

to-vertexandpath-to-pathmappingswiththemaximalmap-

ping score under the condition of limited length of gaps or

mismatch. The algorithm has worst case time complexity

O(2m× m2). A label-diversity back-track algorithm was

proposed in [16] to align two networks with cycles based

on the mapping of as many as path-to-path similarity.

Finally, amoregeneralapproachtoaligningofmetabolic

pathways – homomorphisms allowing edge-to-path map-

ping – has been first proposed in [3].

The proposed

3Modeling Metabolic Pathway Mappings

A metabolic pathway is a series of chemical reactions

catalyzed by enzymes that occur within a cell. Metabolic

pathways are represented by directed networks in which

vertices correspond to enzymes and there is a directed edge

from one enzyme to another if the product of the reaction

catalyzed by the first enzyme is a substrate of the reaction

catalyzed by the second.

Mapping metabolic pathways should capture the similar-

ities of enzymes represented by proteins as well as topolog-

ical properties that cannot be always reduced to sequential

reactions represented by paths. Below we first describe our

approach to measure enzyme similarity and then discuss ad-

vantages and drawbacks of the homomorphism mappings of

metabolic networks.

Our implementation provides two alternative enzyme

similarity scores. One approach is to employ the lowest

common upper class distribution proposed in [15] and dis-

cussed in [13]. The corresponding penalty score for gap is

Page 3

2.0.

OurnewapproachmakesfulluseofECencodingandthe

tight reaction property classified by EC. The EC number is

expressed with a 4-level hierarchical scheme. The 4-digit

EC number, d1.d2.d3.d4represents a sub-sub-subclass in-

dication of biochemical reaction. If d1.d2of two enzymes

are different, their similarity score is infinite; if d3of two

enzymes are different, their similarity score is 10; if d4of

two enzymes are different, their similarity score is 1; or else

the similarity score is 0. The corresponding penalty score

for gap is 0.5. Our experimental study indicates that the

proposed similarity score scheme results in biochemically

more relevant pathway matches.

The topology of most metabolic pathways is a simple

path, but frequently pathways may branch or have several

incoming arcs – all such topologies are instances of a mul-

tisource tree, i.e., a directed graph which becomes an undi-

rected tree when edge directions are disregarded. The query

pathways are usually simple and can be represented as a

multisource tree but in some cases they can have a cycle or

alternative ways to reach the same vertex. Then we sug-

gest to follow the standard practice of breaking such cycle

or paths by removing edges.

The obvious way to preserve the pathway topology is

to use isomorphic embedding – one-to-one correspondence

between vertices and edges of the pattern and its image in

the text. The requirement on edges can be relaxed – an

edge in the pattern can be mapped to a path in the text

([12,13])andthecorrespondingmappingiscalledahomeo-

morphism. The computational drawback of isomorphic em-

bedding and homeomorphism is that the problem of finding

optimal mapping is NP-complete and, therefore, requires

severe constraints on the topology of the text to become ef-

ficient. In [12, 13], the text is supposed to be a tree, the pat-

tern should a directed tree while allowing multisource-tree

pattern complicates the algorithm. Their algorithm is com-

plex and slow because it repeatedly finds minimum weight

perfect matchings.

In this paper we propose to additionally relax one-to-one

correspondence between vertices – instead we allow differ-

ent pattern vertices to be mapped to a single text vertex. The

corresponding mapping is called a homomorphism. Such

relaxation may sometimes cause confusion – a path can be

mapped to a cycle. For instance, if two enzymes with sim-

ilar functions belong to the same path in the pattern and a

cycle with similar enzyme belongs to the text, then the path

can be mapped into a cycle (see Figure 2). However, if the

text graph is acyclic this cannot happen. Even if there are

cycles in the text, still one can expect that functionally sim-

ilar enzymes are very rare in the same path.

Computing minimum cost homomorphisms is much

simpler and faster than homeomorphisms. We will show

that a fast dynamic programming algorithm can find the

C

D

B

A

A’=f(A)=f(D)

B’=f(B)

C’=f(C)

PathEdge

Vertex-to-vertex Mapping

the pattern onto a cycle (A?,B?,...,C?,D?= A?) in the

text.

Path

Cycle

Figure 2. Homomorphism of a path (A,B,...,C,D) in

minimum cost homomorphism that allows edge-to-path

mapping for the multisource tree pattern and an arbitrary

text graph.

4Graph-Theoretical Problem Formulation

Wefirstgivenotationsanddefinitionsandthenformulate

the corresponding graph-theoretical problem.

A pattern T = (V T,ET) is a directed graph with vertex

set V T and edge set ET. We only consider the case of

T being a multisource tree. Following [13], a multi-source

tree is a directed graph, whose underlying undirected graph

is a tree. It is not necessarily a directed tree – each node can

have several incoming as well as several outgoing edges.

A text G = (V G,EG) is a directed graph with vertex set

V G and edge set EG. We further distinguish the case when

G is a general network and the case when G is a directed

acyclic graph.

A mapping f : T → G from pattern T = (V T,ET) to

text G = (V G,EG) is called a homomorphism if

(1) every vertex in V T is mapped to a vertex in V G ;

(2) every edge e = (u,v) ∈ ET is mapped to a directed

path f(e) = (u0= f(u),u1,u2,...,uk= f(v)) in G.

We will now introduce the cost of a homomorphism. Let

∆(u,v), u ∈ V T, v ∈ V G, be the cost of mapping an en-

zyme corresponding to the pattern vertex u into an enzyme

corresponding to the text vertex v.

Following [13], the rule (2) allows edge-to-path map-

ping, but edge-to-edge mapping is still preferable. There-

fore, the homomorphism cost should increase proportion-

ally to the number of extra hops in the images of edges, i.e.,

?

where |f(e)| = k is the number of hops in the path f(e) =

(u0= f(u),u1,u2,...,uk= f(v)).

e∈ET

(|f(e)| − 1)

Page 4

Following [13], the cost of a homomorphism f : T → G

takes in account cost of enzyme mapping and edge-to-path

mapping as follows

?

where λ is the cost of a single extra hop in an edge-to-

path mapping. This parameter balances enzyme mapping

and edge-to-path costs. If λ = 0, then only the enzyme

mapping cost is taken into account. If λ is very large, then

the enzyme mapping cost contribution is negligible. In our

computational experiments we use λ = 0.5.

Finally, the graph-theoretical problem formulation is as

follows.

cost(f) =

v∈V T

∆(v,f(v)) + λ

?

e∈ET

(|f(e)| − 1)

Minimum Cost Homomorphism Problem. Given a multi-

source pattern tree T and a text graph G, find the minimum

cost homomorphism f : T → G.

5Dynamic Programming Algorithm

We will first describe preprocessing of the text graph G

and ordering of vertices of the pattern graph T. Then we

define the dynamic programming table and show how to fill

that table in a bottom-up manner. We conclude with the

runtime analysis of the entire algorithm.

Text Graph Preprocessing. In order to compute the cost

of a homomorphism it is necessary to know the number of

hops for any shortest path in the text graph G. Although

finding single-source shortest paths in general graphs is

slow, in our case it is sufficient to run breadth-first-search

with runtime O(|EG| + |V G|). Assuming that G is con-

nected, i.e., |EG| ≥ |V G|, we conclude that the total run-

time of finding all shortest paths is O(|V G||EG|). In the

resulting transitive closure G?= (V G,EG?) of the graph

G, each edge e ∈ EG?is supplied with the number of hops

h(e) in the shortest path connecting its ends.

Pattern Graph Ordering. We will further need a cer-

tain fixed order of vertices in V T as follows. Let T?=

(V T,ET?) be the undirected tree obtained from T by dis-

regarding edge directions. Let us choose an arbitrary vertex

r ∈ V T as a root and run depth-first search (DFS) in T?

from r. Let {r = v1,...,v|V T|} be the order of the DFS

traversal of V T and let e?

to directed edge ei∈ ET) be the unique edge connecting

vito the set {v1,...,vi−1}. The vertex v ∈ {v1,...,vi−1}

is called a parent of viand viis called a child of v.

DP Table. Now we will describe our dynamic program-

ming table DT[1,...,|V T|][1,...,|V G|]. Each row and

column of this table corresponds to a vertex of T and

G, respectively. While the columns u1,...,u|V G|of DT

are in no particular order, the rows {r = v1,...,v|V T|}

i= (vi,v) ∈ ET?(corresponding

of DT are sorted according to the DFS traversal of T?.

Each element DT[i,j] is equal to the best cost of a ho-

momorphism from the subgraph of T induced1by vertices

{v|V T|,v|V T|−1,...,vi} into G?which maps viinto uj.

Filling DP Table. The table DT is filled bottom-up for

i = |V T|,|V T| − 1,...,1 as follows. If viis not a parent

for any vertex in T, then viis a leaf and

DT[i,j] = ∆(vi,uj)

In general, let vibe a parent for the vertices vi1,...,vik.

In order to compute DT[i,j], we should find the cheapest

mapping of each of the children vi1,...,viksubject to vi

being mapped to uj. The mappings of the children do not

depend on each other since the only connection between

them in the tree T is through vi. Therefore, each child vil,

l = 1,...,k should be mapped intoujlminimizing the con-

tribution of vilto the total cost

C[il,jl] = DT[il,jl] + λ(h(j,jl) − 1)

where h(j,jl) depends on direction of eil, i.e., h(j,jl) =

h(uj,ujl) if eil= (vi,vil) and h(j,jl) = h(ujl,uj) if

eil= (vil,vi). Finally,

DT[i,j] = ∆(vi,uj) +

k

?

l=1

min

j?=1,...,|V G|C[il,j?]

Runtime Analysis. As we mentioned earlier, the runtime

for constructing the transitive closure G?= (V G,EG?) is

O(|V G||EG|). The runtime to fill a cell DT[i,j] is propor-

tional to

tij= degT(vi)degG?(uj)

where degT(vi) and degG?(uj) are degrees of viand ujin

graphs T and G?, respectively. Indeed, the number of chil-

dren of viis degT(vi) − 1 and for each child vilof vithere

are at most degG?(uj) feasible positions in G?since f(vi)

and f(vil) should be adjacent. The runtime to fill the entire

table DT is proportional to

|V G|

?

Thus the total runtime is O(|V G||EG|+|EG?||V T|). Even

thoughGissparse, |EG?|maybeaslargeasO(|V G|2), i.e.,

the runtime is O(|V G|(|EG| + |V G||V T|)).

j=1

|V T|

?

i=1

tij=

|V G|

?

j=1

degG?(uj)

|V T|

?

i=1

degT(vi) = 2|EG?||ET|

6Mapping Metabolic Pathways

In this section we first describe the metabolic pathway

data, then explain how we measure statistical significance

1A subgraph induced by the subset of vertices S includes only edges

that have both ends in S.

Page 5

pattern network

(tree pathways)

T. thermophilus

text network (number of pthways)

B. subtilis(226)

21

14

20

217

143

153

5

3

5

12

7

12

T. thermophilus(208)

38

28

35

162

80

106

9

2

9

24

9

21

E. coli(113)

18

12

18

121

85

92

38

3

14

12

6

12

S. cerevisiae(151)

18

13

17

58

39

40

3

2

5

14

13

14

# of mapping pairs

# of mapped pattern pathways

# of mapped text pathways

# of mapping pairs

# of mapped pattern pathways

# of mapped text pathways

# of mapping pairs

# of mapped pattern pathways

# of mapped text pathways

# of mapping pairs

# of mapped pattern pathways

# of mapped text pathways

B. subtilis

E. coli

S. cerevisiae

Table 1. Pairwise statistical homomorphisms among T. thermophilus, B. subtilis, E. coli and S. cerevisiae.

of homomorphisms and report the results of pairwise map-

pings between four species.

Data. The genome-scale metabolic network data in our

studies were drawn from BioCyc [1, 7, 10], the collection of

260 Pathway/Genome Databases, each of which describes

metabolic pathways and enzymes of a single organism. We

havechosenmetabolicnetworksofE.coli, theyeastS.cere-

visiae, the eubacterium B. subtilis and the archeabacterium

T. thermophilus so that they cover major lineages Archaea,

Eukaryotes, and Eubacteria. The bacterium E. coli with

113 pathways is the most extensively studied prokaryotic

organism. T. thermophilus with 208 pathways belongs to

Archaea. B. subtilis with 226 pathways is one of the best

understood Eubacteria in terms of molecular biology and

cell biology. S. cerevisiae with 151 pathways is the most

thoroughly researched eukaryotic microorganism.

Statistical Significance of Mapping. Although the cost

of a homomorphism reflects the similarity of pathways, it

alone cannot assure us that such cost is not obtained by

chance. Only statistically significant cost values can be

taken in account. Statistical significance is measured by p-

value, i.e., the probability of the null hypothesis that the cost

value is obtained by pure chance. Following a standard ran-

domization procedure, we randomly permute pairs of edges

(u,v) and (u?,v?) if no other edges exist between these 4

vertices u,u?,v,v?in the text graph by reconnecting them

as (u,v?) and (u?,v). This allows us to keep the incoming

and outgoing degree of each vertex intact. We find the min-

imum cost homomorphism from the pattern graph into the

fully randomization of the text graph and check if its cost is

at least as big as the minimum cost before randomization of

the text graph. We say that the homomorphism is statisti-

cally significant with p < 0.01 if we found at most 9 better

costs in 1000 randomization of the text graph.

Experiments. For each pair of four species (B. subtilis, E.

coli, T. thermophilus and S. cerevisiae), using our algorithm

we find the best homomorphism from each pathway of one

species to each pathway of the other and check if this homo-

morphism is statistically significant, i.e., if p < 0.01. We

have run our experiments on a Pentium 4 processor, 2.99

GHz clock with 1.00 GB RAM. The total runtime was 1.5h

for the input/output of pathways and computing the opti-

mal pattern-to-text mapping and its p-value for every pair

of pathways (there are in total 516052 pattern-text pathway

pairs).

Results. The results of our experiments are reported in Ta-

ble 1. The first column contains the name of the species

from whose metabolic network the pattern pathways have

been chosen. Note that if a pathway is not a multisource

tree or degenerate (i.e., has less than 3 nodes), then it is

omitted. We did not omit any pathway from the text species

since our algorithm supports any network as a text. For

every species-to-species mapping, we compute the number

of mapped pairs with p < 0.01, the number of the pattern

pathways that have at least one statistically significant ho-

momorphic image and the number of the text pathways that

have at least one statistically significant homomorphic pre-

image.

For example, for homomorphism from T. thermophilus

to B. subtilis, there are in total 21 statistically significant

mapped pairs, 14 non-degenerate tree T. thermophilus path-

ways have statistically significant homomorphic images in

B. subtilis and 20 out of 226 B. subtilis pathways have sta-

tistically significant homomorphic pre-images.

7 Implications of Pathway Mappings

In this section we identify pathways conserved across

multiple species, show how one can resolve enzyme am-

biguity and identify potential holes in pathways, and phylo-

genetically validate our pathway mappings.

Identifying Conserved Pathways. We first have identified

the pathways that are conserved across all 4 species under

consideration. Table 2 contains a list of all 20 pathways in

Page 6

Pathway name

alanine biosynthesis I

biotin biosynthesis I

coenzyme A biosynthesis

fatty acid beta

fatty acid elongation saturated

formaldehyde oxidation V (tetrahydrofolate pathway)

glyceraldehyde 3 phosphate degradation

histidine biosynthesis I

homoserine biosynthesis

lysine biosynthesis I

ornithine biosynthesis

phenylalanine biosynthesis I

phenylalanine biosynthesis II

polyisoprenoid biosynthesis

proline biosynthesis I

quinate degradation

serine biosynthesis

superpathway of gluconate degradation

tyrosine biosynthesis I

UDP galactose biosynthesis

alanine biosynthesis

biotin biosynthesis

fatty acid oxidation pathway

fructoselysine and psicoselysine degradation

Table 2. The list of all 20 pathways in B. subtilis that

have statistically significant homomorphic images simulta-

neously in all 3 other species E. coli, T. thermophilus and S.

cerevisiae. The lower part contains 4 more different path-

ways with statistically significant images in all 4 species.

Pathway name

triple: B. subtilis, E. coli, and T. thermophilus

4 aminobutyrate degradation I

de novo biosynthesis of pyrimidine deoxyribonucleotides

de novo biosynthesis of pyrimidine ribonucleotides

enterobacterial common antigen biosynthesis

phospholipid biosynthesis I

PRPP biosynthesis II

salvage pathways of pyrimidine deoxyribonucleotides

ubiquinone biosynthesis

flavin biosynthesis

glycogen biosynthesis I (from ADP D Glucose)

L idonate degradation

lipoate biosynthesis and incorporation I

menaquinone biosynthesis

NAD biosynthesis I (from aspartate)

triple: B. subtilis, E. coli, and S. cerevisiae

oxidative branch of the pentose phosphate pathway

S adenosylmethionine biosynthesis

triple: B. subtilis, T. thermophilus, and S. cerevisiae

tyrosine biosynthesis I

fatty acid elongation unsaturated I

Table 3. The list of 14 pathways conserved across B. sub-

tilis, E. coli, and T. thermophilus; 2 more pathways con-

served across B. subtilis, E. coli, and S. cerevisiae; 2 more

pathways conserved across B. subtilis, T. thermophilus, and

S. cerevisiae.

Page 7

B. subtilis that have statistically significant homomorphic

images simultaneously in all species. The lower part of Ta-

ble 2 contains 4 more pathways with different names in E.

coli, T. thermophilus and S. cerevisiae, which have simulta-

neous statistically significant images in all species.

Besides 24 pathways conserved across all 4 species we

have also found 18 pathways only common for triples of

these species. Table 3 gives the pathway names for each

possible triple of species (the triple E. coli, T. thermophilus

and S. cerevisiae does not have extra conserved pathways).

2.6.1.1

2.6.1.1

1.2.4.- 2.3.1.-

2.3.1.61

6.2.1.5

6.2.1.5

1.3.99.1

1.3.99.1

4.2.1.2

4.2.1.2

1.1.1.1.82

1.1.1.82

1.2.4.2

Figure 3. Mapping of glutamate degradation VII path-

ways from B. subtilis to T. thermophilus (p < 0.01). The

node with upper part and lower part represents a vertex-to-

vertex mapping. The upper part represents the query en-

zyme and the lower part represents the text enzyme. The

shaded node reflects enzyme homology.

3.5.3.6

3.5.3.6

1.5.1.21.5.1.

1.5.1.2

2.1.3.3

2.1.3.3

2.6.1.13

2.6.1.13

0.0.0.0

0.0.0.0

5.1.1.4

5.1.1.4

0.0.0.0

0.0.0.0

0.0.0.0

0.0.0.0

1.4.1.12

1.4.1.12

6.1.1.12

6.1.1.12

5.4.3.5

5.4.3.-

Figure 4. Mapping of interconversion of arginine, or-

nithine and proline pathway from T. thermophilus to B. sub-

tilis (p < 0.01). The node with upper part and lower part

represents a vertex-to-vertex mapping. The upper part rep-

resents the query enzyme and the lower part represents the

text enzyme. The shaded node reflects enzyme homology.

Resolving Ambiguity. Currently multiple pathways con-

tain unresolved enzymes. Completely unresolved enzymes

have EC notation 0.0.0.0/-.-.-.- and partially unresolved en-

zymes have less ”-”’s, e.g., EC notation 1.2.4.-. We can

use our mapping tool to suggest possible resolution of these

ambiguities as follows.

Let us consider two examples of homomorphism –

the mapping of glutamate degradation VII pathway in B.

subtilis to glutamate degradation VII pathway in T. ther-

mophilus (shown in Figure 3), and the mapping of inter-

conversion of arginine, ornithine and proline pathway in T.

thermophilus to interconversion of arginine, ornithine and

proline pathway in B. subtilis (shown in Figure 4). When

some enzymes in pathway are labeled with the end of ”.-”, it

denotes their exact reactions are not explicit. The mapping

results indicate that a potential similar enzyme with similar

functions of the unclear enzyme can be found in some other

species.

6.3.2.12

1.5.1.51.5.1.5

1.5.1.53.5.4.9

3.5.4.9

6.3.4.3

6.3.4.3

1.5.1.32.1.2.1

1.5.1.20

2.1.1.13

2.1.1.450.0.0.0

6.3.2.17

3.5.1.10

Figure 5. Mapping of formaldehyde oxidation V path-

way in B. subtilis to formy1THF biosynthesis pathway in

E. coli (p < 0.01). The node with upper part and lower part

represents a vertex-to-vertex mapping. The upper part rep-

resents the query enzyme and the lower part represents the

text enzyme. The node with dashed box represents a gap.

Holes in Pathways. Pathway holes happen when a genome

appears to lack the enzymes needed to catalyze reactions

in a pathway [6]. We can use our mapping tool to identify

potential pathway holes as shown in the following example

(see Figure 5).

There is a statistically significant mapping from

formaldehyde oxidation V (tetrahydrofolate pathway) in

B. subtilis to formyITHF biosynthesis in E. coli.

correspondence between enzymes shows a gap – enzyme

3.5.1.10 is missing. We found that the enzyme 3.5.1.10

exists in B. subtilis according to Enzyme and Swiss-prot

databases.

The

Phylogenetic Validation. One can measure similarity be-

tween species based on the number of conserved pathways.

Thelargestamountofconservedpathwaysisfoundbetween

B. subtilis and T. thermophilus – two species-to-species

mappings have in total 183 statistically significant pairs of

pathways. The next closest two species are E. coli and

B. subtilis which have 126 statistically significant pairs of

Page 8

pathways. This agrees with fact that B. subtilis, T. ther-

mophilus, and E. coli are prokaryote and S. cerevisiae is a

eucaryote.

8 Conclusions

In this paper we have introduced a new method of map-

ping metabolic pathways based on homomorphism. The

proposed mapping approach allows to map different en-

zymes of the pattern pathway into a single enzyme of a text

network. We have also define a novel scoring scheme for

computing similarity between enzymes based on their EC

notation.

We have formulated the graph-theoretical problem cor-

responding to finding optimal homomorphism from the

pattern network to a text network.

cient dynamic-programming method for exactly solving

this problem when the pattern is multisource tree an the text

is an arbitrary network.

We have applied our mapping tool pairwise mapping of

all pathways for four organisms (E. coli, S. cerevisiae, B.

subtilis and T. thermophilus species) representing main dif-

ferent lineages and found a reasonably large set of statisti-

cally significant pathway similarities.

We report 24 pathways that are conserved across all 4

species as well 18 more pathways that are conserved across

at least three of these species. We show that our mapping

toolcanbeusedforidentificationofpotentialpathwayholes

as well resolving enzyme notation ambiguities in existing

pathway descriptions.

We give an effi-

Acknowledgments

We thank Professor Pinter and Oleg Rokhlenko for giv-

ing access to their software package MetaPathwayHunter,

Professor Karp and Dr. Kaipa for providing the pathway

databases for T. thermophilus, B. subtilis, E. coli and S.

cerevisiae. We also thank Amit Sabnis, Dipendra Kaur,

Kelly Westbrooks for helpful discussions.

References

[1] http://www.biocyc.org/.

[2] M. Chen and R. Hofest. An algorithm for linear metabolic

pathway alignment. In silico biology (In silico biol.) ISSN,

1386-6338: 111-128, 2005.

[3] Q. Cheng and A. Zelikovsky. Optimal mapping of multi

sourcetreesintodaginbiologicalnetwork. ISBRA’07Poster,

May 2007.

[4] T. Dandekar, S. Schuster, B. Snel, M. Huynen, and P. Bork.

Pathway alignment: application to the comparative analysis

of glycolytic enzymes. Biochem. J., 1: 115-124, 1999.

[5] C. V. Forst and K. Schulten. Evolution of metabolism: a

newmethodforthecomparisonofmetabolicpathwaysusing

genomics information. J. Comput. Biol., 6: 343-360, 1999.

[6] M. L. Green and P. D. Karp. A bayesian method for iden-

tifying missing enzymes in predicted metabolic pathway

databases. BMC Bioinformatics, Sep. 2004.

[7] I. M. Keeler, V. J. Collard, C. S. Gama, J. Ingrafts, S. Palely,

I. T. Paulson, M. Peralta-Gil, and P. D. Karp. Ecocyc: a

comprehensive database resource for escherichia coli. Nu-

cleic Acids Research, 33(1):D334-337, 2006.

[8] B. P. Kelly, R. Sharan, R. M. Karp, T. Sittler, D. E. Root,

and B. R. Stockwell. Pathblast: a tool for alignment of pro-

tein interaction networks. Nucleic Acids Research, Vol.32 :

W83-W88, 2004.

[9] B. P. Kelly, R. Sharan, R. M. Karp, T. Sittler, D. E. Root,

and B. R. Stockwell. Conserved pathways within bacteria

and yeast as revealed by global protein network alignment.

PNAS, 11394-11399, Sep. 30 2003.

[10] C. J. Krieger, P. Zhang, L. A. Mueller, A. Wang, S. Paley,

M. Arnaud, J. Pick, S. Rheme, and P. Karp. Metacyc: a mi-

croorganism database of metabolic pathways and enzymes.

Nucleic Acids Research, 32(1):D438-42, 2006.

[11] Z. Li, Y. Wang, S. Zhang, X.-S. Zhang, and L. Chen. Align-

ment of protein interaction networks by integer quadratic

programming. EMBS ’06. 28th Annual International Con-

ference of the IEEE, 5527-5530, Aug. 2006.

[12] R. Pinter, O. Rokhlenko, D. Tsur, and M. Ziv-Ukelson. Ap-

proximate labeled subtree homeomorphism. In Proceedings

of 15th Annual Symposium of Combinatorial Pattern Match-

ing.

[13] R. Y. Pinter, O. Rokhlenko, E. Yeger-Lotem, and M. Ziv-

Ukelson. Alignmentof metabolic pathways. Bioinformatics.

[14] R. Sharan, S. Suthram, R. M. Kelley, T. Kuhn, S. McCuine,

P. Uetz, T. Sittler, R. M. Karp, and T. Ideker. Conserved

patterns of protein interaction in multiple species. PNAS,

Vol.102 : 1974-1979, 2005.

[15] Y. Tohsato, H. Matsuda, and A. Hashimoto.

alignment algorithm for metabolic pathway analysis using

enzyme hierarchy. Proc. 8th International Conference on

Intelligent Systems for Molecular Biology, 376-383, ISMB

2000.

[16] S. Wernicke. Combinatorial algorithms to cope with the

complexity of biological networks. Dissertation, December

2006.

[17] Q. Yang and S.-H. Sze. Path matching and graph matching

in biological networks. Journal of Computational Biology,

Vol. 14, No. 1: 56-67 : 5527-5530, 2007.

A multiple

#### View other sources

#### Hide other sources

- Available from A. Zelikovsky · May 31, 2014
- Available from gsu.edu
- Available from psu.edu