Protein classification by stochastic modeling and optimal filtering of amino-acid sequences

Abstract

The prediction of a protein's tertiary structural class from its amino-acid sequence is formulated as a signal-processing problem. The amino-acid sequence is treated as a “time series” of symbols containing signals that determine the protein's structural class. A methodology is described for building detailed stochastic signal models for recognized structural classes of single-domain proteins. We solve the problem of determining that model, from a set of candidates, which is the most probable generator of a protein's entire amino-acid sequence. The solution employs a nonlinear, optimal filtering algorithm, which is suited for implementation on parallel computer architectures. Previous approaches have only been able to classify correctly 80% of single-domain proteins within three very broad structural types, while our approach achieves this level across twelve much more detailed classes.