Modelling thermal stability changes upon mutations in proteins with artificial neural networks

The change in protein thermal stability upon mutation (monitored by changes in melting temperature Tm) is usually predicted from the computed folding free energy change at room temperature. Although the correlation between these two quantities is as high as 0.88, predicting the former from the latter yield poor performances, with correlation coefficients between measured and predicted Tm changes as low as 0.67 for 90% of 829 mutations in cross validation. Predictive algorithms that are specifically focused on thermal stability changes upon mutation are therefore needed. For this purpose, we use a set of previously developed statistical energy functions, describing the coupling between four protein descriptors (sequence, distance, torsion angles and solvent accessibility), and accounting for the volume variation of the mutated amino acid. The change in melting temperature is expressed as a linear combination of these energy functions, whose weights are sigmoid functions of the solvent accessibility of the mutated residue. These weights are identified with the help of an artificial neural network, and their physical meaning is discussed. In particular, the importance of local interactions in predicting thermal stability changes is higher in the protein core than on the surface, although the opposite trend is observed for the prediction of thermodynamic stability changes. The performance of the prediction is strongly improved, as witnessed by a correlation coefficient of 0.73 in cross validation for 90% of the set of 829 mutations.