We study the problem of Protein Remote Homology Detection, which assesses the functional similarity of two proteins. We approach this as a problem of binary multiple-instance learning (MIL) that aims to distinguish between homologous and non-homologous proteins. The particular MIL approach employed is based on the dissimilarity representation in which various schemes of combining N-gram representations are considered. This approach allows us to cope with longer N-grams, capturing a richer biological context, and results in versatile framework offering competitive performance compared to state of the art.

A challenging Pattern Recognition problem in Bioinformatics concerns the detection of a functional relation between two proteins even when they show very low sequence similarity – this is the so-called Protein Remote Homology Detection (PRHD) problem. In this paper we propose a novel approach to PRHD, which casts the problem into a Multiple Instance Learning (MIL) framework, which seems very suitable for this context. Experiments on a standard benchmark show very competitive performances, also in comparison with alternative discriminative methods.