Item

Abstract

pi-helices are short and unstable protein secondary structure motifs. They are present in 15% of all known protein structures, often in functionally important regions such as ligand-binding sites. Thus, the correct prediction of pi-helices is essential for the detection of potential functional sites and can aid design tasks aiming at the creation of ligand-binding pockets. Due to the similarity of pi-helices to much more abundant alpha-helices, it is a challenging task to predict them based on the sequence data and there are no methods devoted to this problem. We present a deep learning neural network trained with sequences where pi-helical residues were assigned based on high-quality X-ray structures [1]. The model achieved 48% precision and 46% sensitivity in per-residue prediction on a test set. Moreover, in the benchmark on commonly used datasets like CB6133, CB513, CASP10, and CASP11 the model outperforms other state-of-the-art methods used for secondary structure prediction.