Prediction Of Protein Stability Changes Upon Amino Acid Mutations

During recent years, many methods have been reported for predicting protein stabilitychanges upon amino acid mutations. However, the problem of too many parameters andoverfit still exists in these methods. In this paper, we bulit two new models for predictingprotein stability in mutations by analysising the protein primary structure, protein tertiarystructure and physicochemical properties. We also bulit an integrated prediction methodbased on the meta approach by comparing these methods.(1) By analyzing protein primary structure, tertiary structure and physicochemicalproperties, we bulit M8and M47models. The number of input vectors of M8is8,which isthe least among current methods, including only8physicochemical properties, while thatof M47is47, including sequence, structure, physicochemical information. Theperformance of M8is quite good at small datasets and is an effective solution to overfitproblem. The performance of M47is significantly better than other methods and itssuperiority is more evident for big dataset.(2) Each method has both advantage and disadvantage, so we integrated the currentmethods by using support vector machine(SVM). We proposed three ways forintegration,which are Oneclass, Wildresidue and Structure. The performance of theseintegrated methods is much better than the single method, among them of Wildresidue isthe best. With the twenty cross-validation in S2760, the accuracy of Wildresidue is0.864,MCC is0.676, r is0.837, and RMSE is0.974kcal/mol.(3) Based on the model M8and M47indicated in (1), we developed software PPSC;and based on the model indicated in (2), we developed software IPSPS.