TESTLoc: protein subcellular localization prediction from EST data

Abstract

Background: The eukaryotic cell has an intricate architecture with compartments and substructures dedicated to particular biological processes. Knowing the subcellular location of proteins not only indicates how bio-processes are organized in different cellular compartments, but also contributes to unravelling the function of individual proteins. Computational localization prediction is possible based on sequence information alone, and has been successfully applied to proteins from virtually all subcellular compartments and all domains of life. However, we realized that current prediction tools do not perform well on partial protein sequences such as those inferred from Expressed Sequence Tag (EST) data, limiting the exploitation of the large and taxonomically most comprehensive body of sequence information from eukaryotes.

Results: We developed a new predictor, TESTLoc, suited for subcellular localization prediction of proteins based on their partial sequence conceptually translated from ESTs (EST-peptides). Support Vector Machine (SVM) is used as computational method and EST-peptides are represented by different features such as amino acid composition and physicochemical properties. When TESTLoc was applied to the most challenging test case (plant data), it yielded high accuracy (~85%).

Conclusions: TESTLoc is a localization prediction tool tailored for EST data. It provides a variety of models for the users to choose from, and is available for download at http://megasun.bch.umontreal.ca/~shenyq/TESTLoc/TESTLoc.html.

Figure 1

Fragmentation procedure of plant protein sequences in order to expand the EST-peptide dataset. Open bars, full-length proteins; filled bars, fragmented protein sequences. Proteins shorter than 200 residues remained unchanged. Proteins ranging from 200 to 400 residues were fragmented into two pieces. Proteins longer than 400 residues were fragmented into three pieces. See text for details.

Figure 2

Training and evaluation of SVM predictors. The circle and pies indicate the dataset and portions thereof. The procedure in each dashed box was repeated ten times. The whole dataset was randomly divided into ten parts, with nine parts combined to construct the SVM model, and the remaining one to evaluate the model. The combined data for model construction were further divided randomly into ten subsets, in which nine subsets were combined to serve as training data, and the 10^th subset served as test data. See text for details.

Figure 3

Independent evaluation of SVM predictors based on different representations of amino acid composition. The performance was assessed by the Matthews Correlation Coefficient (MCC). For most classes, the best MCC was obtained with the 4^th order amino acid composition (the frequency of tetra-peptides). Amino acid group-C and group-D composition yielded similar results (see Additional file 5).

Figure 4

Integration of predictions from SVM models based on individual features. Each of the 41 SVM models built with single sequence features forms the first layer SVM and emits the probabilities for the query sequence to belong to the various classes. The probabilities are used as input for the second layer SVM.