Using neural networks expert system to predict protein thermostability

Some biological or chemical reactions need to be performed at high temperatures to decrease reaction time. However, many proteins are not very stable when heated. Research is needed that helps proteins to remain active and stable at high temperatures overcoming many limitations to their industrial applications. Recently we have shown that some structural features of proteins are related to the thermostability. However, a general model of how these features are related to thermostability remained a scientific challenge. Our goal was to develop a predictive model for thermostability based on sequence and structural features by using neural networks modeling. Amino acid sequences of 4636 proteins, including 3069 mesophilic and 1294 thermophilic, were extracted from databases, 74 protein features calculated and feature selection algorithm applied to determine the most important features contributing to thermal stability. Six different neural network modeling (Quick, Dynamic, Multiple, Prune, Exhaustive Prune and RBFN) applied on important features, and the best method with high accuracy selected. The model applied on some mesoand thermophilic xylanase proteins with known temperatures. The results showed 53 features (p value 0.993 to 1.0) such as frequency of glutamine and frequency of hydrophilic residues were important in thermal stability of proteins, and this model can be applied to classify other proteins as either mesophilic or thermophilic with more than 0.95% confidence. The findings show a suitable tool to classify and predict enzymes thermal stability using automated expert systems.