The understanding of the factors governing protein stability is of outmost importance for a variety of research fields in biochemistry. In addition to a detailed understanding of protein folding it allows the design of prote ins with increased thermostability. Whereas a number of successful em pirical strategies for the design of stabilized proteins can be found in the literature, most of the designs were done by "manual design" compared to an automatic calculation of stabilizing protein sequence variations that should be theoretically possible as far as only small changes in the 3D structure occur.
In order to establish an automatic procedure, two different approaches were taken: first we developed a "distance dependent aminoacidatom potential", second we used a "torsionanglepotential" to predict the ther mostability of proteins. The aminoacidatompotential considers the spa tial distribution of atoms within a given 3D structure, whereas the torsion anglepotentials is derived from the sequential preference between a given amino acid and its neighbors. Both potentials are optimized to obtain a good correlation between predicted and experimental stabilization values.
The quality of the predictions is judged by the degree of linear de pendency, expressed as the correlation coefficient. During the develop ment of the potentials and algorithms we used a "development dataset" which consists of about 700 thermodynamical characterized point muta tions from 8 structures. After each step of development the performance was checked within a "evaluation dataset", which contains about 2000 sta bilization data points from about 25 structures.