Prediction of protein functional specificity without an alignment

We propose a new approach to predict functional specificity of proteins from their amino acid sequences. Our approach is based on two things: structural Multilevel Neighborhoods of Atom (MNA) descriptors and an original Bayesian algorithm. Usually, a protein sequence is presented as a string of amino acid symbols. Here we introduce a new description of an amino acid sequence: a set of structural MNA descriptors. The MNA descriptor is a string describing an atom and its neighbor atoms according to the selected level. In this work, we also use description of a protein sequence as a set of peptides (strings of amino acid symbols). We performed a case study on two subsubclasses of enzyme nomenclature (EC). It is shown that B-statistics give a sufficient predictive power of enzyme specificity prediction for both MNA descriptors and peptides. We also showed that MNA descriptors give higher accuracy values in comparison with peptides and also provide a choice of MNA descriptor levels for best accuracy prediction. The highest average accuracy prediction that was achieved was 0.98.