摘要：AbstractIn this paper, we show how the adequate use of the intrinsic symmetry of a system when setting up its model structure can avoid unwanted biases in the parameter optimization phase. The playground of our analysis is the prediction of protein thermodynamic stability changes upon single amino acid substitutions (point mutations). Using a simple artificial neural network (ANN), sixteen different energy-like contributions are combined to predict the change in folding free energy (ΔΔG). We show that the presence of terms violating the symmetry under inverse mutations induces a bias towards the dataset on which the ANN is trained, even if a strict n-fold cross-validation procedure is performed. A completely symmetric free energy functional is then introduced, which gives predictions that are slightly less efficient in terms of root mean square error with respect to the experimental ΔΔG's, but appear to be basically independent of the training dataset and are thus more satisfactory.
