Last Updated $Date: 2010/04/30 08:16:37 $

About WoLF PSORT

WoLF PSORT predicts the subcellular localization sites of proteins based on their amino acid sequences. The method, which is a major extension to the venerable PSORTII program, makes predictions based on both known sorting signal motifs and some correlative sequence features such as amino acid content. Like PSORT and PSORTII, WoLF PSORT displays some information about detected sorting signals which is useful in helping users determine the reliability of the prediction in specific cases. Our experiments (presented at APBC06) show that the overall prediction accuracy of WoLF PSORT is over 80%. For common localization sites (e.g. cytosol, nucleus, mitochondria, etc) WoLF PSORT makes better than majority classifier predictions even for queries that do not have strong sequence similarity to any sequence in the dataset. Thus WoLF PSORT is a useful complement to tools such as BLAST. The current dataset used to train WoLF PSORT contains over 12,000 animal sequences and more than 2,000 plant and fungi sequences respectively. It was gathered mainly from Uniprot but several hundred Arabidopsis thaliana sequences from the Gene Ontology database were also included.

Citation

Server:

Paul Horton, Keun-Joon Park, Takeshi Obayashi, Naoya Fujita, Hajime Harada, C.J. Adams-Collier, & Kenta Nakai,
"WoLF PSORT: Protein Localization Predictor",
Nucleic Acids Research, doi:10.1093/nar/gkm259, 2007.
[Abstract] [Paper]

Prediction Method:

Paul Horton, Keun-Joon Park, Takeshi Obayashi & Kenta Nakai,
"Protein Subcellular Localization Prediction with WoLF PSORT",
Proceedings of the 4th Annual Asia Pacific Bioinformatics Conference APBC06, Taipei, Taiwan. pp. 39-48, 2006.
[Abstract] [Paper]

Developers

WoLF PSORT is being developed by

Paul HORTON and others at CBRC

These people have contributed to WoLF PSORT in the past

Keun-Joon PARK at CBRC, now at Korea Center for Disease Control & Prevention, Initial dataset preparation
Takeshi OBAYASHI at Tokyo Institute of Technology, now at Tokyo University Human Genome Center, Initial plant dataset preparation
C.J. Adams-Collier of Collier Technologies, Initial server design

Dataset

The dataset is based mainly on annotation from Uniprot and Gene Ontology. The table below gives a correspondence between our localization site definitions and Gene Ontology. However, many of our entries are based solely on Uniprot "Subcellular Localization" field keywords and in some of these cases the site assignment may not be completely consistent with the GO cellular component annotation.

*Localization Sites and corresponding GO cellular components.*
Abbrev	Localization Site	GO Cellular Component
chlo	chloroplast	0009507, 0009543
cyto	cytosol	0005829
cysk	cytoskeleton	0005856(2)
E.R.	endoplasmic reticulum	0005783
extr	extracellular	0005576, 0005618
golg	Golgi apparatus	0005794(1)
lyso	lysosome	0005764
mito	mitochondria	0005739
nucl	nuclear	0005634
pero	peroxisome	0005777(2)
plas	plasma membrane	0005886
vacu	vacuolar membrane	0005774(2)

Abbreviation, Localization Site, and corresponding GO Cellular Component(s) are given for each localization site. Numbers in parentheses, such as "0005856(2)" indicate that descendant "part_of" cellular components were also included, up to the specified depth (2 in this case). For example, all of the children and grandchildren of "GO:0005856" were included as "cysk".

Stand alone Package

WoLF PSORT package version 0.2 has been released September 2006. It is academic free and also relatively easy for industrial users to use as well. Please see the package documentation for details.

Prediction Accuracy by Localization Site

The accuracy varies greatly between different localization sites -- the general trend being that sites with few uniprot annotated proteins are seldom correctly predicted. In a separate localization accuracy by utility page, we have compiled some statistics to help answer this question quantitatively.

What's in a name

"WoLF" does not necessarily stand for anything. A rather dramatic mnemonic would be "Where Life Functions". Originally it was going to be "Learned Weight Features" but I wanted the acronym to be a pronouncable English word. Women only Love Fools.

Presentations

[Slides] from a presentation introducing the issues involved in protein localization prediction.

Acknowledgements

WoLF PSORT Relies heavily on features inherited from PSORT(Nakai & Kanehisa)
WoLF PSORT also uses some sequence features from iPSORT (Bannai et. al).
The original server design was done by C.J.Collier. (But he is not to blame for subsequent hacking...)

Independent Evaluations

Several evaluations have been done on the prediction accuracy of various protein subcellular localization methods, including WoLF PSORT. In some WoLF PSORT has performed very poorly, but in some very well. To bias the reader, I list two here in which WoLF PSORT performed well:

W. Qian & J. Zhang, Genome Biology and Evolution, 198-204, 2009. [PMID:20333190]
Comparing methods which do not rely strongly on sequence similarity on predicting the subcellular localization of yeast proteins, they found WoLF PSORT to perform the best (even so it only attained an accuracy of 45%, Multiloc was second at 36%).
E.W. Klee & C.P. Sosa, Drug Discovery Today, 12:234-40, 2007. [PMID:17331888]
Evaluated several programs on predicting human secreted proteins and concluded that WoLF PSORT was the best overall