Protein structure features inferred from sequence can change significantly with database growth over time

Protein structure predictions based on sequence normally use a reference database to carry out sequence similarity calculations. Typical references, such as nr or nt , essentially are continuously updated sequence sets and therefore their volume increase over time, as well as their annotations, fueled by the ever growing number of published genomes and proteomes in the last decade. In this work we measure to what extent the recent growth of a fairly standard sequence database, uniref90, affects the prediction of structural features of proteins, namely secondary structure, intrinsic disorder and the phenotype of point mutations. The results of widely used predictive software (PSIPRED, DISOPRED and SIFT, respectively), which relies on sequence searches for producing results, were compared with two carbon copies of uniref90 from 2010 and 2014. For this purpose four different sets of proteins from model organisms Escherichia coli, Arabidopsis thaliana and Homo sapiens, plus a subset of high quality structures from the Protein Data Bank (PDB), were analyzed. Our results indicate that the 2010 and 2014 predictions vary strongly (Q3 = 0.29 - 0.42) in the case of SIFT, suggesting that this algorithm probably overestimates the effect of point mutations that insert rare amino acids in poorly characterized protein families. Such variations were less pronounced, although significant, in the case of DISOPRED (Q3 = 0.83 - 0.89), and much lower with PSIPRED (Q3 = 0.92). Therefore, it can be concluded that out the three tested features, secondary structure predictions are the most stable over time. Furthermore, by comparing PSIPRED results with the corresponding secondary structure assignments of high resolution PDB structures, it was found that 2010 and 2014 predictions also conserve their quality, reporting Q3 values of 0.76 in both cases.Peer Reviewe