1. Protein Subcellular Localization Prediction of Eukaryotes using a Knowledge-based Approach
Hsin-Nan Lin, Ching-Tai Chen, Ting-Yi Sung,
Shinn-Ying Ho, Wen-Lian Hsu
Bioinformatics Program, TIGP (Taiwan International Graduate Program), Academia Sinica, Taiwan

2. Outline Introduction Methods Dataset Prediction Results Conclusions
Construction of the knowledge base SPKB
KnowPredsite: a localization prediction method using SPKB
Dataset
Prediction Results
Conclusions

3. Protein Subcellular Localization (PSL)
Given a protein, determine its subcellular compartment
Mitochondria, cytoplasm, nuclear,….etc
It is important because
Help elucidate protein functions
Identify potential diagnostic and drug targets
Wet-lab experiment
Time consuming
Labor intensive
Computational prediction is needed.

4. Existing PSL Predictors
Using various features
Features extracted from literature or public databases
Phylogenetic profiling
Compartment-specific features
Main problem of many predictors
They only predict a limited number of locations
Limited to subsets of proteomes, e.g., those containing signal peptide sequences
Designed for specific species
Designed for single-localized protein sequences
Up to 35% of proteins move between different cellular compartments

5. Outline Introduction Methods Dataset Prediction Results Conclusions
Concept behind the method
Construction of the knowledge base SPKB
KnowPredsite: a localization prediction method using SPKB
Dataset
Prediction Results
Conclusions

6. Basic Concept Behind the Method
Transitivity Relations

7. A New Protein Feature – Similar Peptide
High Scoring Pair (HSP)
A significant local pairwise alignment of two proteins by PSI-BLAST
Interchangeable Amino Acid Pair
A positive score in the BLOSUM62.
Similarity Level
The number of interchangeable amino acid pairs within a sliding window
Similar Peptide
Represents possible sequence variation.
An n-gram peptide fragment from a similar protein

8. An Example of High Scoring Pairs (HSP)
MYKKILY
MY KIL
MYSKILL
Window size = 7
Pairwise similarity = 5

9. Construction of Similar Peptide Knowledge Base (SPKB)
A protein with annotated localization site CYT
HSP from PSI-BLAST
A protein from NCBI NR database
MENIKKE
ME +KK
MEAVKKS
If pairwise similarity ≧ similarity level ,
MEAVKKS is a similar peptide
Inherit CYT
Similarity level = 4
Pairwise similarity = 5

10. A similar peptide entry for protein subcellular localization
Peptide: MYSKILL
SPKB
We store the similar peptides into a knowledge base, which is called SPKB. Our knowledge base stores millions upon millions similar peptide records.
For example, this is a similar peptide record whose representative sequence is MYSKILL. And it has two protein members, one is 1ark, the other is 1ata.
It also inherits two different secondary structures. The similar peptide record is the basic unit we used for predicting protein secondary structures.

11. KnowPredsite: a localization prediction method using SPKB

12. Blast-hit Method Serves as a baseline approach
Use Blast to find the most similar sequence
Inherit the localization annotation
E-value = 10-3
If there is no hit, no annotation will be inherited

13. Evaluation

14. Outline Introduction Methods Dataset Prediction Results Conclusions
Concept behind the method
Construction of the knowledge base SPKB
KnowPredsite: a localization prediction method using SPKB
Dataset
Prediction Results
Conclusions

15. Dataset From ngLOC method 25,887 Single-localized proteins
2,169 multi-localized proteins
1923 different species
10 different subcellular locations
CYT (cytoplasm)
CSK (cytoskeleton)
END (endoplasmic reticulum)
EXC (extracellular)
GOL (Golgi apparatus)
LYS (lysosome)
MIT (mitochondria)
NUC (nuclear)
PLA (plasma membrane)
POX (perixosome)

16. Outline Introduction Methods Dataset Prediction Results Conclusions
Concept behind the method
Construction of the knowledge base SPKB
KnowPredsite: a localization prediction method using SPKB
Dataset
Prediction Results
Conclusions

17. Finding the Best Similarity Level and Window Size
Leave-one-out cross
validation is performed.
Similarity Level 1 2 3 4 5 6 7 8
Overall Accuracy
w=7
91.2
91.3
91.4
91.5
91.8
92.0
91.6 －
w=8
91.7
90.9

18. Prediction Performance for Single-localized Proteins
*KnowPredsite : leave-one-out cross validation
#KnowPredsite : ten-fold cross validation
Overall Accuracy (%)
Methods
Top 1
Top 2
Top 3
Top 4
Single-localized
*KnowPredsite
92.0
95.7
96.8
98.1
#KnowPredsite
91.7
95.4
96.6
97.9
ngLOC
88.8
92.2
94.5
96.3
Blast-hit
86.0 －

19. Prediction Performance for Multi-localized Proteins
Overall Accuracy (%)
Methods
Top 1
Top 2
Top 3
Top 4
Multi-localized
(at least 1 correct)
*KnowPredsite
90.8
96.4
98.2
98.9
#KnowPredsite
90.1
96.1
98.1
ngLOC
81.9
92.0
97.4
Blast-hit
78.8 －
(both correct)
74.3
83.3
88.7
72.1
82.2
87.5
59.7
73.8
83.2
45.7

20. Site-Specific Prediction Performance
Site i
Occurrence in the dataset (%)
Precision (%)
Accuracyi (%)
MCCi
CYT
11.1
75.7
84.4
0.774
CSK
1.0
81.1
52.0
0.645
END
3.6
92.9
84.1
0.88
EXC
29.1
98.5
93.9
0.946
GOL
1.1
79.1
70.9
0.746
LYS
0.6
87.2
81.9
0.844
MIT
9.4
96.7
86.9
0.907
NUC
18.0
87.3
93.8
0.884
PLA
25.2
94.4
96.4
0.938
POX
0.8
85.1
0.861

21. Multi-localized Confidence Score (MLCS)
We follow King and Guda’s method, for a protein t
MLCS(t) = (CS1 + CS2) － (CS12 －CS22)/100
CS1: highest confidence score among all the sites
CS2: 2nd highest confidence score among all the sites
Best MLCS threshold of KnowPredsite = 20
86.3% of multi-localized proteins have MLCS > 20
82.8% of single-localized proteins have MLCS < 20

22. Case Study: EF1A2_RABIT Swiss-Prot : NUC Gene Ontology: CYT & NUC
Query
CYT
CSK
END
EXC
GOL
LYS
MIT
NUC*
PLA
POX
MLCS
EF1A2_RABIT
95.45
1.45
0.04
2.97
0.05
7.40
Template
NUC SI
EF1A2_RAT
2.94
99.78
EF1A_CHICK
2.77
92.22
EF1A1_HUMAN
2.75
EF1A1_RAT
EF1A0_XENLA
2.69
90.06
EF1A_BRARE
2.64
EF1A2_XENLA
88.79
EF1A3_XENLA
2.60
88.55
Swiss-Prot : NUC
Gene Ontology: CYT & NUC

23. Case Study: MCA3_MOUSE Query CYT* CSK END EXC GOL LYS MIT NUC* PLA POX
MLCS
MCA3_MOUSE
95.46
0.3
0.27
0.36
0.2
0.01
1.13
93.59
1.82
0.22
100
Template
CYT
NUC SI
MCA3_HUMAN
89.16
88.51
EF1G1_YEAST
2.74
2.47
8.67
EF1G2_YEAST
0.49
8.50
GSTA_PLEPL
0.35
15.86
SYEC_YEAST
0.16
3.86
CCNA1_MOUSE
0.15
7.36
NU155_RAT
0.14
3.17
GCYB2_HUMAN
4.86

24. Outline Introduction Methods Dataset Prediction Results Conclusions
Concept behind the method
Construction of the knowledge base SPKB
KnowPredsite: a localization prediction method using SPKB
Dataset
Prediction Results
Conclusions

25. Conclusions
We proposed a sequence based prediction method KnowPredsite which can
Predicts single-localized proteins with 92% accuracy
Predicts multi-localized proteins with 74.3% accuracy
Suitable for proteome-wide prediction
25887 single-localized proteins
2169 multi-localized proteins
Provides interpretable prediction results through template proteins
KnowPredsite can be easily improved by incrementally expanding the SPKB

26. Thank You.

27. Multi-localized Confidence Score
TP: a multi-localized protein with MLCS > 20
TN: a single-localized protein with MLCS < 20

28. Site-Specific Comparison
ngLOC performs better in CYT & LYS