The role of independent test set in modeling of protein folding kinetics

The testing of a bioinformatics algorithm on the training set is not the best indicator of its future performance because of the misleadingly optimistic results. The optimal method of testing is the calculation of error rate on an independent dataset (test set). We have tested the validity of the FOLD-RATE method for the prediction of protein folding rate constants [ln(k ( f ))] using sequences, structural class information and experimentally verified folding rate constants of the Protein Folding Database (PFD). PFD is a publicly accessible repository of thermodynamic and kinetic data of interest for the researchers of different profiles, standardized by the International Foldeomics Consortium. Our results show that when the standardized PFD dataset is used to test a protein fold rate prediction method, the estimation of validity may differ significantly.