Translational initiation site (TIS) prediction is a very important and actively studied topic in bioinformatics. In order to complete a comparative analysis, it is desirable to have several benchmark data sets which can be used to test the effectiveness of different algorithms. An ideal benchmark data set should be reliable, representative and readily available. Preferably, proteins encoded by members of the data set should also be representative of the protein population actually expressed in cellular specimens.
In this paper, we report a general algorithm for constructing a reliable sequence collection that only includes mRNA sequences whose corresponding protein products present an average profile of the general protein population of a given organism, with respect to three major structural parameters. Four representative transcript collections, each derived from a model organism, have been obtained following the algorithm we propose. Evaluation of these data sets shows that they are reasonable representations of the spectrum of proteins obtained from cellular proteomic studies. Six state-of-the-art predictors have been used to test the usefulness of the construction algorithm that we proposed. Comparative study which reports the predictors' performance on our data set as well as three other existing benchmark collections has demonstrated the actual merits of our data sets as benchmark testing collections.
The proposed data set construction algorithm has demonstrated its property of being a general and widely applicable scheme. Our comparison with published proteomic studies has shown that the expression of our data set of transcripts generates a polypeptide population that is representative of that obtained from evaluation of biological specimens. Our data set thus represents "real world" transcripts that will allow more accurate evaluation of algorithms dedicated to identification of TISs, as well as other translational regulatory motifs within mRNA sequences. The algorithm proposed by us aims at compiling a redundancy-free data set by removing redundant copies of homologous proteins. The existence of such data sets may be useful for conducting statistical analyses of protein sequence-structure relations. At the current stage, our approach's focus is to obtain an "average" protein data set for any particular organism without posing much selection bias. However, with the three major protein structural parameters deeply integrated into the scheme, it would be a trivial task to extend the current method for obtaining a more selective protein data set, which may facilitate the study of some particular protein structure.