On the Accuracy of Sequence Similarity Based Protein 3D Prediction

Abstract

In an article (Akcesme, and Can 2015), authors examined the relation between primary and secondary structure mismatches of the substrings of length seventeen residues from two different proteins. They have shown that the mismatches in the corresponding secondary structure sequence substrings of the same length mostly lag behind primary mismatches. In the PhD dissertation thesis (Akcesme 2016) author examined the possibility of secondary structure prediction by the use of smaller conserved segments and created a software AVISENNA that outperforms PSIPRED and all other available secondary structure prediction tools. In another article (Akcesme, et. al. 2017), the issue of how far secondary structure of proteins can be predicted based on hosts (larger proteins that contain the query protein as a subchain) of this protein in the set of solved structures currently deposited in PDB. It is seen that around 17% of proteins have hosts in PDB, and secondary structures of them can be predicted with a mean accuracy of 90.39 %. This accuracy of the host based secondary structure prediction set also an upper bound for the homology based tertiary structure predictions. In this article the impact of the mentioned inaccuracy on the homology based 3D structure predictions by the three predictors I-Tasser, Phyre2, and SwissModel are studied. Inaccuracies in predicted tertiary structures are seen in the visual comparison of the 3D structures of query proteins and their predicted 3D images by the three 3D predictors, and their counterparts in host proteins.