Applications

LOHMMs have been applied to several different problems. Publication VIII addresses the application to protein-fold recognition. The number of determined protein structures is growing rapidly and there are different classification schemes for them. There is a need for computer methods that can automatically extract structural signatures for classes of proteins. The secondary structure of a protein is represented as a sequence of structured symbols, so applying LOHMMs is very natural. The results on the database and classification scheme SCOP (Structural Classification Of Proteins from Murzin et al. (1995)) indicate that LOHMMs possess several advantages over other approaches.

Figure 6.4: A small protein fold represented emphasising the secondary structure with helices (blue) and strands (green).

Another application of LOHMMs in the biological domain is the mRNA signal structure detection presented in Publication VII. mRNA molecules fold to form a secondary structure which can be described with concepts such as stacking regions, hairpin loops, and interior loops. The secondary structure of an mRNA forms a tree which makes it more challenging than that of a protein. A LOHMM was used to parse a tree in in-order (the node itself between its children) while the tree structure is essentially stored in the arguments of the hidden state. Classification accuracy was higher than with the comparison method by Horváth et al. (2001).

UNIX command sequences have been studied with LOHMMs in Publication IX. Tasks that have been considered for UNIX command sequences include the prediction of the next command in the sequence by Davison and Hirsh (1998), the classification of a command sequence in a user category by Jacobs and Blockeel (2001); Korvemaker and Greiner (2000), and anomaly detection by Lane (1999). LOHMMs could be applied to all of these tasks and Publication IX reports experiments in the classification task with results comparable to other methods.

Landwehr et al. (2006) use a custom implementation of LOHMMs for haplotype reconstruction from genotype data. The proposed method offers a competitive trade-off between accuracy and computational complexity compared to other state-of-the-art systems developed for the task.