Alignment of Multiple Proteins with an Ensemble of Hidden Markov Models

The alignment of multiple protein sequences is a problem of fundamental importance in bioinformatics. In general, the optimal alignment can be obtained through the optimization of an objective function. However, such an optimization task is often computationally intractible, most of the existing alignment tools thus use statistical or machine learning based methods to avoid direct optimizations. In this paper, we develop a new method that can progressively construct and update a set of alignments by adding sequences in certain order to each of the existing alignments. In particular, each of the existing alignments is modeled with a profile hidden markov model (HMM) and an added sequence is aligned to each of these profile HMMs. The profile HMMs in the set are then updated based on the alignments with leading alignment scores. We performed experiments on BaliBASE benchmarks to compare the performance of this new approach with that of other alignment tools. Our experiments showed that, by introducing an integer parameter that controls the number of profile HMMs in the set, we are able to efficiently explore the alignment space and significantly improve the alignment accuracy on sequences with low similarity.