By : Elham Khabiri Adviser : Dr. Hisham Al-Mubaid

Using Semantic Similarity Measures in the Using Semantic Similarity Measures in the Biomedical Domain for Computing Similarity Biomedical Domain for Computing Similarity

between Genes based on Gene Ontologybetween Genes based on Gene Ontology

By : Elham KhabiriBy : Elham Khabiri

Adviser : Dr. Hisham Al-MubaidAdviser : Dr. Hisham Al-Mubaid

University of Houston - Clear Lake2

Motivation

• Goal :– Measure functional similarity between genes and

Proteins

• Reason:– It is useful to measure the functional difference

between genes in different organisms

– Find the genes with unknown functions

HumanYeastDrug Target


MotivationMotivation

• To compute the similarity between two genes g1 and g2, we can use one of the following information sources:

– gene sequence information– gene functional annotations (GO terms)– biomedical literature and texts – gene expression profiles.

• In this work, we use Gene functional

annotations and the gene ontology GO to measure the similarity between genes.



• Given two genes Gp and Gq such that gene Gp is annotated with a set of n different GO terms, we call it the set GOp:

GOp = {tp1, tp

2, …., tpn}

• Similarly, the annotation set for gene Gq is:

GOq = {tq1, tq

2, …., tqn}

that is, gene Gq is annotated with m different GO terms • The terms tp

i or tqj are nodes in the GO

• If both genes Gp and Gq are annotated with only one term (n=m=1) and the same GO term ( tp

1 = tq1) then the

similarity between them is maximum.



• In general, if both genes Gp and Gq are annotated with the same set of GO terms (n=m≥1) (that is, tp

i = tqj) then

the similarity between them is maximum.


Motivation

• Many data resources in bioinformatics not only hold data in the form of sequences, but also as annotation– Scientific natural language– Suitable for human but not easy for

machine processing


Related Work:Semantic Measures in NLP

Resnik, 1995

Lin, 1998

Jiang and Conrath, 1997

Wu & Palmer, 1994

Leacock and Chodorow, 1998

Based on Information Based on Information

Content (IC) of Least Content (IC) of Least

Common AncestorCommon Ancestor (LCA)(LCA)

Based on Ontology Based on Ontology StructureStructure


1

151410

7

13

8

t2t1

9

54 6

3t

Related Work

• WordNet [Miller 1995] • Information Content Based Measures

– Resnik, 1995

)/(freq(t) log- ) t,(tsim) t,LCA(tt

21Resnik21

N

freq(t): Frequency of concept c in database.

N: the number of all the concepts in database.


Related Work

– Jiang and Conrath, 1997

– Lin, 1998

) t,LCA(tt2121JC

21

))p(t log)p(t (log-p(t) 2log ) t,(tdist

))(log)c(log

)(log.2(max),(

21),(21 21 cPP

cPccsim ccScLin


Related Work

• Ontology Structure Based Measures:– Wu & Palmer, 1994

• Based on the depths of the two concepts in the taxonomies, and the depth of the LCS

– Leacock and Chodorow, 1998: PL• Based on the PL(t1,t2) of the shortest path between two

concepts• Scale the measure by the overall depth D of the taxonomy

),(

2log),(

2121 cclen

Dccsimlch

)()(

)),((2),(

21

2121 cdepthcdepth

cclcsdepthccsimwup


Related Work:Measures in Biomedical Domain

• First semantic similarity measure in biomedical domain:– Rada et al., 1989 : Path Length between

biomedical terms in the MeSH ontology• Measure of semantic similarity in Gene

Ontology (GO)– Lord et al., 2003: Applied Resnik’s to GO– Validated the correlation between

sequence and semantic similarity


Related Work:Recent Works in Biomedical Domain

• Al-Mubaid and Nguyen, 2007– Investigated the effectiveness of using Medline

corpus as the information source for measuring the semantic similarity in the biomedical domain

• Al-Mubaid and Nguyen, 2007– A technique for computing the semantic similarity

between biomedical terms across multiple ontologies within a unified framework like UMLS

• Wang et. al , 2007– Functional similarity measure of GO terms based on

contributions of the term’s ancestors in GO Evaluation: Compare it with Resnik’s measure

• Found it was closer to human perception


Sequence Similarity

• Sequence Similarity– BLAST [Altschul 1990] :Finds regions of local

similarity between sequences of genes– WU-BLAST2

Output E-value Bit-score


Drawbacks of Sequence Similarity

• Sequence similarity holds for most genes with the same functionality

• Devos 2000: 30% of the functional similarity found by sequence similarity might be erroneous – Reason: Genes that are not evolved from a

common ancestors do not have a considerable sequence similarity

• One drawback for the sequence notation is that, it is not readable and understandable by human.


New approach

• Ontology structure based – Path Length (PL) between the two terms– Number of minimum paths between terms– Depth of LCA of two terms

• Ontology used: Gene Ontology– A comprehensive resource for gene functional

information

• Validation – Correlation with sequence similarity– Correlation with two other semantic measures


Gene Ontology

• One of the greatest project in bioinformatics• Created in 2000 by GO Consortium [Ashburner et. al]

• Consists of a set of controlled vocabularies for– Biological Process– Molecular Functions– Cellular Components

• Shows the functional and biological terms related to genes in a hierarchical and structured way


Gene Ontology


Gene Ontology

• Directed Acyclic Graph

• Each term may have more than one parent

• There may be more than one path between two nodes (terms)

• Each two node have at least one LCA (Least Common Ancestor)


3 Proposed Measures

1. Plain Path Length (PL)– Number of edges between the two terms

2. Path Length with Variation (PLm)

– Number of common terms– Number of minimum paths

3. Path Length with Depth (SimPLD)

– Path Length between two terms– Depth of LCA of the two terms


Plain Path Length

1

151410

7

13

8

1211

9

54 6

32

11 2584712 6

Parents of 11

Parents of 12 Parent of 4

Parents of 8

Considers the first level ancestor of each node in the list

Parent of 5


PL between two Genes

• genep is annotated with terms {t1,..., tn}

• geneq is annotated with terms {t1,..., tm }

1..m}:j 1..n,:i | dij avg{

Facl6

Annotated with 3 MF

dij: Shortest PL between ti of

gene1 and tj of gene2


PL Evaluation

• Based on Correlation with Sequence Similarity

• Genome Used: – SGD (Saccharomyces cerevisiae): 3 datasets– FlyBase (Drosophila Melanogaster): 1 dataset

• Divide datasets Based On E-Value:– High Sequence Similarity (HSS): E-value ≤ 10-5 – Low Sequence Similarity (LSS): 10-5 < E-value <1– No Sequence Similarity (NSS): E-value = 1


Evaluation: Compare PL with Sequence Similarity

Dataset1 From SGD

0

10

20

30

40

50

60

70

PL<=2 2<PL<=7 PL>7

Path Length

Per

cen

tag

e HSS

LSS

NSS

DataSet2 From SGD

0102030405060708090

PL<=2 2<PL<=7 PL>7

Path Length

Per

cen

tag

e HSS

LSS

NSS


Evaluation: Compare PL with Sequence Similarity

Dataset3 From SGD

0

10

20

30

40

50

60

70

80

PL<=2 2<PL<=7 PL>7

Path Length

Per

cen

tag

e

HSS

LSS

NSS

Distribution of Path Length in FlyBase dataset

0

20

40

60

80

100

PL<=2 3 <PL<=7 PL>7

Path LengthP

erce

ntag

e

HSS

NSS

70% of HSS have PL<=2

7% of HSS have PL>7

7% of NSS have PL<=2

80% of HSS have PL<=2

4% of HSS have PL>7

17% of NSS have PL<=2


3 Proposed Measures







Path Length with Variation

• More than one LCA

• Two minimum Paths– “6-10-7-5-1” – “6-10-11-5-1”

• More functional similarity that those who have only one minimum path between them

5

2 41 3

10

6

9

7 8

11

12


PL with Variation

PL(gox, goy) if nmp = 1

PL(gox, goy)/w1.nmp, otherwise

PLm (gox, goy)

PL(gox, goy) = the minimum path length in

the GO graph between the two GO terms

gox and goy

mn

n

i

m

j

q

j

p

i gogom

PL

1 1

qpm

),(

)G ,(G PL


Path Length with Variation

• genep is annotated with terms {t1,..., tn}

• geneq is annotated with terms {t1,..., tm }

Max go_pl = 15

mnnct

n

i

m

j

q

j

p

i gogom

PL

1 1

qpm

),(

2

1 )G ,(G PL

nct = number of common GO terms between Gp, Gq.

)G ,(GPL - max )G ,Sim(G qpmgo_plqp


Validate PLm

• We measured the similarity of gene pairs in SGD pathways

• Pathway is a series of chemical reactions occurring within a cell – Pathway #5 (allantoin degradation): 4 genes – pathway #6 (arginine biosynthesis): 7 genes– pathway #141 (tryptophan degradation): 12 genes

• Compare with – Resnik measure – Wang et. al measure


Validate PLm: Compare with Resnik

• Pathway 5: allantoin degradation – 4 genes, 6 pairs

gene1 gene2 Res Ours

DAL1 DAL2 2.4 11

DAL1 DAL3 2.4 11

DAL1 DUR1,2 1.7 9.5

DAL2 DAL3 5.2 13

DAL2 DUR1,2 1.7 9.5

DAL3 DUR1,2 1.7 9.5

They Correlate well with

each other

Minimum

Maximum


Validate PLm : Compare with Resnik

Pathway 6: 7 genes, 21 pairs gene1 gene2 Res Ours

ARG1 ARG3 0.28 8

ARG1 ARG4 0.28 8

ARG2 ARG3 1.38 7.5

ARG3 ARG5,6 1.01 8.5

ARG4 ARG8 0.28 7

ARG1 ARG8 0.28 8

PL(ARG2, ARG3) > PL(ARG3, ARG5,6)

PL(ARG4, ARG8) > PL(ARG1, ARG8)


Evaluation: Clusters of Genes Wang et. al vs. Our Method


3 Proposed Measures







Similarity between GO terms

• PL(gox, goy ) = minimum path length between the two GO terms gox and goy

)2

),(log()

)),((log(),(

MaxDepth

gogoPL

MaxDepth

gogolcadepthgogoSim yxyx

yxPLD

1

151410

7

13

8

1211

9

54 6

32


SimPLD between two Genes

• gp is annotated with terms {go1,..., gon}

• gq is annotated with terms {go1,..., gom }

1..m}:y 1..n,:x|)go ,(go{sim avg )g ,(gsim yxPLDqpPLD


Evaluation: SimPLD

• Correlation between SimPLD and sequence similarity

• Dataset:– SGD – FlyBase – Human-Yeast

• Ontology Used: – Molecular function (MF)


Compare SimPLD with Sequence Similarity

Distribution of sim for Human-Yeast dataset

0

20

40

60

80

100

-2<sim<0 0=<sim<2.8

Perc

enta

ge HSS

LSS

NSS

Distribution of sim for FlyBase

0

20

40

60

80

100

-2<sim<0 0=<sim<2.8

Perc

enta

ge HSS

NSS

Distribution of sim for SGD

0

20

40

60

80

100

-2<sim<0 0=<sim<2.8

Perc

enct

age

HSS

LSS

NSS

Based On BLAST E-Value:

High Sequence SimilarityLow Sequence SimilarityNo Sequence Similarity


Conclusion

• Gene Ontology is a reliable source to be used for functional similarity

• Our semantic measures – Can be used as an automated tool to

determine the genes with the similar functionalities

– Has a fairly well agreement with Blast sequence similarity and results of other famous semantic measures


Resulted Publications

• Khabiri E., Al-Mubaid H. (2007) “A path length method for gene functional similarity using GO annotations.” 16th International Conference on Software Engineering and Data Engineering SEDE 2007. Las Vegas, Nevada USA, 2007

• Khabiri E. (2007) “A Preliminary study of Correlation between depth and Path Length of GO nodes with Gene Sequence Similarity.” IEEE 7 International Conference on BioInformatics and BioEngineering BIBE07, Boston, Massachusetts USA, 2007

• Al-Mubaid H., Khabiri E., “A New Path Length Based Measure for Functional Similarity of Genes with Evaluation Using SGD Pathways.” Computational Structural Bioinformatics Workshop (CSBW), San Jose, CA (Accepted, not finalized)


Future Work

• Apply path length-based measures to more datasets from different model organisms

• More accurate evaluation – Biomedical literature – Microarray data analysis

• Consider the number of distinct paths

• Prediction of functionally unknown genes



By : Elham Khabiri Adviser : Dr. Hisham Al-Mubaid

Documents

biomedical terms

semantic similarity

measure functional similarity

functional similarity

computing similarity

gene gq

gene gp

genes gp