Interpretable numerical descriptors of amino acid space.

Informative numerical representations of amino acid residues are essential for successful in silico modeling or establishing the structure-activity relationships of proteins. A straightforward approach is adopted here for representing more than 500 amino acid indices from the AAindex database by a set of uncorrelated scales, satisfying the varimax criterion. Different measures are considered in order to demonstrate the improved interpretability of the current scales as compared to previously published ones. Performance is also addressed in a classification problem of G-protein coupled receptors, and is found to be similar or higher than the performance achieved by six other scale sets. Finally, a unique correspondence between numerical indices and mutation matrices is derived and discussed in light of the evolutionary conservation of amino acid properties. Conclusions from this study highlight the discord between ease of interpretation of amino acid scales and their relevance to protein structure conservation, as well as general considerations for designing custom scale sets.