Application of the multi-modal relevance vector machine to the problem of protein secondary structure prediction

The aim of the paper is to experimentally examine the plausibility of Relevance Vector Machines (RVM) for protein secondary structure prediction. We restrict our attention to detecting strands which represent an especially problematic element of the secondary structure. The commonly adopted local principle of secondary structure prediction is applied, which implies comparison of a sliding window in the given polypeptide chain with a number of reference amino-acid sequences cut out of the training proteins as benchmarks representing the classes of secondary structure. As distinct from the classical RVM, the novel version applied in this paper allows for selective combination of several tentative window comparison modalities. Experiments on the RS126 data set have shown its ability to essentially decrease the number of reference fragments in the resulting decision rule and to select a subset of the most appropriate comparison modalities within the given set of the tentative ones.