In this paper, we adapt a statistical learning approach, inspired by automated topic segmentation techniques in speech-recognized documents to the challenging protein segmentation problem in the context of G-protein coupled receptors (GPCR). Each GPCR consists of 7 transmembrane helices separated by alternating extracellular and intracellular loops. Viewing the helices and extracellular and intracellular loops as 3 different topics, the problem of segmenting the protein amino acid sequence according to its secondary structure is analogous to the problem of topic segmentation. The method presented involves building an n-gram language model for each 'topic' and comparing their performance in predicting the current amino acid, to determine whether a boundary occurs at the current position. This presents a distinctly different approach to protein segmentation from the Markov models that have been used previously and its commendable results is evidence of the benefit of applying machine learning and language technologies to bioinformatics.
展开▼
