The Importance of Window Length in Splice Site Prediction

The performance of gene prediction programs strongly depends on the methods that they use to locate splice sites. Different pattern recognition techniques are available to assess the quality of candidate splice sites, see [1] for an overview and further references. All of these techniques proceed by computing a score derived from the distribution of the nucleotides in the neighbourhood of a splice site consensus sequence. These scores are normally obtained with splice sites models that have been estimated from large training sets of exemplary neighbourhoods. The training sets may also include negative examples, i.e. sequences that contain the consensus sequence, but that are actually no splice sites. Unfortunately, the concept of ‘neighbourhood’ is rather ambiguous, and there is no general recommendation about the positions of the nucleotides that should be included in the calculation, i.e. the analysis window that should be employed. In principle, the window length is an important parameter, because it determines the amount of information that has to be evaluated. On the one hand, the window should be long enough to provide as many details as possible about the patterns. On the other hand, the window should be short enough to take only the relevant information, in order to improve generalization. In the present study, we investigate how splice-site prediction accuracy depends on the window size and shape, using support vector machines (SVM) [2]. Our results show that the choice of the window is crucial for splice site prediction, and therefore we suggest that the window length should be considered as an essential parameter of the model.