Transcript Presentation File

Semantic Similarity over Gene
Ontology for Multi-label Protein
Subcellular Localization
Shibiao WAN and Man-Wai MAK
The Hong Kong Polytechnic University
Sun-Yuan KUNG
Princeton University
Outline
1.
2.
3.
4.
5.
6.
Introduction and Motivation
Retrieval of GO Terms
Semantic Similarity Measures
Multi-label Multi-Class Classification
Results
Conclusions
2
Proteins and Their Subcellular Locations
3
Subcellular Localization Prediction
• The subcellular locations of proteins help biologists
to elucidate the functions of proteins.
• Identifying the subcellular locations by entirely
experimental means is time-consuming and costly.
• Computational methods are necessary for subcellular
localization prediction.
• Previous research has found that gene ontology (GO)
based methods outperform methods based on other
protein features (e.g. AA composition).
4
Multi-label Problem
• Some proteins can simultaneously reside at, or move
between, two or more subcellular locations.
• Multi-label (Multi-location) proteins play important
roles in some metabolic processes taking place in
multiple subcellular locations.
• State-of-the-art multi-label predictors, such as PlantmPLoc, iLoc-Plant, and mGOASVM use frequency
counts of GO terms as features.
• In this work, we propose using semantic similarity of
GO terms as features for multi-label subcellular
localization prediction.
5
Method’s Flowchart
S
BLAST
Swiss-Prot
Database
SVM
GO Extraction by
searching GOA
database
Semantic
Similarity
Measure
SVM
GOA Database
GO of training
proteins
.
.
.
homolog AC
AC
SS: Semantic Similarity
M
.
.
.
Subcellular
Location(s)
SVM
Multi-label SVM
Semantic
Similarity
Vector
6
Gene Ontology
 Gene ontology is a set of standardized
vocabularies annotating the functions of genes
and gene products
 GO terms, e.g., GO:0000187
 A protein sequence may correspond to 0, 1 or
many GO terms.
7
Gene Ontology: Example
Search----GO:0000187 in http://www.geneontology.org/
8
GOA Database
• Gene Ontology Annotation database.
– Provide structured annotations to proteins in
UniProt Knowledgebase (UniProtKB) and other
protein databases using standardized GO
vocabularies.
– Include a series of cross-references to other
databases.
• Given an Accession Number, the GOA database
allows us to find a set of GO terms associated with
that accession number.
9
GOA Database
Accession Number
(AC)
GO term(s)
Search A0M8T9 in http://www.ebi.ac.uk/GOA/
1 AC maps to
many GO terms !
10
Finding GO Terms without an
Accession Number
S
BLAST
Swiss-Prot
Database
homolog AC
AC
GO Extraction by
searching GOA
database
GO Terms of Qi
GOA Database
11
Semantic Similarity Measure
GO term x
Find Common
Ancestors
A(x,y)
GO term y
Ancestors
Computing
Semantic
Similarity
sim(x,y)
SQL Query
GO
Database
12
Finding Common Ancestors, A(x,y)
13
Finding Common Ancestors, A(x,y)
GO:0000187
is_a
part_of
14
Semantic Similarity Measure
We use Lin’s measure to estimate the semantic similarity
between two GO terms (x and y):
p(c) =
#(proteins annotated to GO term c)
#(all proteins annotated to the GO taxonomy)
15
Semantic Similarity between 2 Proteins
Semantic similarity between 2 proteins (Gi, Gj):
where
Semantic Similarity Vector:
No. of training proteins
16
Multi-label SVM Scoring
GO of Qt
GO of
training
proteins
=
17
Benchmark Datasets
The Plant dataset
18
Performance Metrics
Overall locative accuracy:
Overall actual accuracy:
Actual accuracy is more
objective and stricter!
19
Performance Comparison
The Plant dataset
20
Conclusions
• Our Proposed predictor performs significantly better
than Plant-mPLoc and iLoc-Plant, and also better than
mGOASVM, in terms of locative and actual accuracies.
• As for individual locative accuracies, our proposed
predictor are significantly higher than the three
predictors for all of the 12 locations.
• In terms of GO information extraction, Plant-mPLoc,
iLoc-Plant and mGOASVM use the occurrences of GO
terms as features, whereas the proposed predictor
discovers the semantic relationship between GO terms,
from which the semantic similarity between proteins
can be obtained.
21
Web Servers
22
Thank you!
23
Multi-label SVM Classifier
Transformed labels for M-class problem:
24
Retrieving GO Terms with/without AC
AC known ?
Y
N
Retrieve
k max homologs by BLAST;k  1
k 0
Using the k - th homolog
N
k  kmax ?
Y
Retrieve a set of GO terms G i,ki
|G i ,k i | 0 ?
Using back-up methods
N
Multi-label SVM
classification
Y
k  k 1
25
Finding Common Ancestors
• The relationships between GO terms in the
GO hierarchy can be obtained from the SQL
database through the link:
http://archive.geneontology.org/latesttermdb/go_daily-termdb-tables.tar.gz.
• We only considered the ‘is-a’ relationship.
26