SOLeNNoID: A Deep Learning Pipeline For Solenoid Residue Detection in Protein Structures

Abstract

Solenoid proteins are a subset of tandem repeat proteins, which are structurally distinct from globular proteins. Solenoid proteins are defined by their modular, elongated structures, dependent on interactions between adjacent repeats. These proteins are found across all domains of life and have many important functions such as protein binding, enzymatic catalysis, ice binding and nucleic acid binding. Furthermore, engineered variants of solenoid proteins such as DARPins and designed PPR proteins have therapeutic commercial applications. In order to advance the study of natural solenoid proteins and the design of novel solenoid proteins, accurate tools for solenoid detection and annotation are required. As solenoid structures are more conserved than solenoid sequences and owing to recent developments in protein structure prediction, structure-based solenoid detection is preferred. Here we propose SOLeNNoID - a deep learning pipeline for solenoid residue prediction in protein structures. We cover all three solenoid sub-classes: alpha-, alpha/beta- and beta-solenoids. We use a CNN architecture to reason over protein distance matrices and compare our method to existing structure-based methods. Finally, we produce predictions on the entire PDB and demonstrate a 71 percent increase in solenoid-containing entries over the gold-standard RepeatsDB database using our method.GitHubhttps://github.com/gnik2018/SOLeNNoIDZenodoTBC