AbstractPrediction of Proteins’ three dimensional structure and their contact maps from their amino-acid sequences is a fundamental problem in structural computational biology. The structure and contacts shed light on protein function, enhance our basic understanding of their molecular biology and may potentially aid in drug design. In recent years we have seen significant progress in protein contact map prediction from Multiple Sequence Alignments (MSA) of the target protein and its homologous, using signals of co-evolution and applying deep learning methods.Homology modelling is a popular and successful approach, where the structure of a protein is determined using information from known template structures of similar proteins, and has been shown to improve prediction even in cases of low sequence identity. Motivated by these observations, we developed Periscope, a method for homology-assisted contact map prediction using a deep convolutional network. Our method automatically integrates the co-evolutionary information from the MSA, and the physical contact information from the template structures.We apply our method to families of CAMEO and membrane proteins, and show improved prediction accuracy compared to the MSA-only based method RaptorX. Finally, we use our method to improve the subsequent task of predicting the proteins’ three dimensional structure based on the (improved) predicted contact map, and show initial promising results in this task too - our overall accuracy is comparable to the template-based Modeller software, yet the two methods are complementary and succeed on different targets.
Magnesium-dependent Serratia marcescens endonuclease: three-dimensional structure at 1.1 Å resolution and molecular mechanism of enzyme action
