Prediction of subcellular localizations using amino acid composition and order.

Subcellular localization is important for proteins to function. For the prediction of subcellular localizations, we have developed a method, SortPred, using the amino acid composition and order. The composition represents the global features, e.g., the amino acid composition in the full or partial sequences, while the order represents the local features, e.g., the amino acid sequence order. The former was represented by neural networks and the latter was represented by a hidden Markov model. This method predicted the signal peptides (SP), the mitochondrial targeting peptides (mTP), the chloroplast transit peptides (cTP), and the nuclear or cytosolic sequences (other) comparing together the previous methods, this method achieved slightly higher prediction accuracy, 86% for plant and 91% for non-plant. We analyzed the trained neural networks and hidden Markov models and found out that these models well represent the biological features of the sequences.

Beta-sheet prediction using inter-strand residue pairs and refinement with Hopfield neural network.

Many secondary prediction methods have been studied, but the prediction accuracy is still unsatisfactory, since beta-sheet prediction is difficult. In this research, we gathered statistics of pairs of three residue sub-sequences in beta-sheets, calculated propensities for them. When a sequence is given, all possible three residue sub-sequences are examined whether they form beta-sheets. A short coming is that many false predictions are made. To exclude false predictions and improve the prediction, we employed a Hopfield neural network, in which the natural limitations on protein tertiary structure and preference of chemically stable long beta-sheet are expressed in a form of energy functions. To clarify the prediction for heads and tails of beta-sheets, special variables are introduced, which are similar to the line process proposed by Geman.

Prediction of Mitochondrial Targeting Signals Using Hidden Markov Model.

The mitochondrial targeting signal (MTS) is the presequence that directs nascent proteins bearing it to mitochondria. We have developed a hidden Markov model (HMM) that represents various known sequence characteristics of MTSs, such as the length variation, amino acid composition, amphiphilicity, and consensus pattern around the cleavage site. The topology and parameters of this model are automatically determined by the iterative duplication method, in which a small fully-connected HMM is gradually expanded by state splitting. The model can be used to predict the existence of MTSs for given amino acid sequences. Its prediction accuracy was estimated to be 86.9% using the cross validation test. Furthermore, a higher correlation was observed between the HMM score and the in vitro ATPase activity of MSF, which can be regarded as an experimental measure of signal strength, for various synthetic peptides than was observed with other methods.

Stochastic motif extraction using hidden Markov model.

In this paper, we study the application of an HMM (hidden Markov model) to the problem of representing protein sequences by a stochastic motif. A stochastic protein motif represents the small segments of protein sequences that have a certain function or structure. The stochastic motif, represented by an HMM, has conditional probabilities to deal with the stochastic nature of the motif. This HMM directly reflects the characteristics of the motif, such as a protein periodical structure or grouping. In order to obtain the optimal HMM, we developed the "ilerative duplication method" for HMM topology learning. It starts from a small fully-connected network and iterates the network generation and parameter optimization until it achieves sufficient discrimination accuracy. Using this method, we obtained an HMM for a leucine zipper motif. Compared to the accuracy of a symbolic pattern representation with accuracy of 14.8 percent, an HMM achieved 79.3 percent in prediction. Additionally, the method can obtain an HMM for various types of zinc finger motifs, and it might separate the mixed data. We demonstrated that this approach is applicable to the validation of the protein database; a constructed HMM has indicated that one protein sequence annotated as "leucine-zipper like sequence" in the database is quite different from other leucine-zipper sequences in terms of likelihood, and we found this discrimination is plausible.