Neural Networks to predict protein stability changes upon mutation

Black box modelling is used here to improve the performances of the PoPMuSiC program that predicts protein stability changes caused by single-site mutations. For that purpose previously developed statistical energy functions are exploited, which are based on a formalism that highlights the coupling between 4 different protein descriptors (sequence, distance, torsion angles and solvent-accessibility), as well as the volume variation of the mutated amino acid. As the importance of the different types of interactions may depend on the protein region, the stability change is expressed as a linear combination of these energetic functions, whose proportionality coefficients vary with the solvent-accessibility of the mutated residue. Two alternative structures are considered for these coefficients: a Radial Basis Function network, and a MultiLayer Perceptron with sigmoid nodes. These two structures are identified, leading to an improvement of the predictive capabilities of PoPMuSiC, and are discussed in terms of their biophysical interpretation.