An End-to-End Deep Learning Pipeline for Assigning Secondary Structure in Proteins

Abstract

Protein secondary structure assignment, a subdiscipline of computational chemistry, is yet to be explored using deep learning techniques. Protein secondary structure elements are assigned to support structural analysis and prediction. Algorithms like DSSP, generally regarded as the gold standard for assigning the secondary structure of proteins, need full atom information to label protein coordinates. The PDB database has been the major repository for data on the 3D structures of proteins, nucleic acids, and other complex assemblies since 1971. However, a significant fraction of protein structures contains missing atoms. As a result, new approaches to reliably and consistently assigning secondary structures based on coarse-grained atomic coordinates are needed. While deep learning architectures have an unparalleled track record in applications such as protein structure prediction, there are only a few known deep learning solutions for structure assignment problems. While the gold standard methods are based on bonding information and other geometric characteristics, deep learning methods extract features themselves without human intervention. The benefit that standardised datasets provide to the effectiveness of deep learning systems in multiple domains motivated us to create a labeled dataset for protein structure assignment tasks. Additionally, a deep learning model, named PSSADL, solely based on [Formula: see text] coordinates was trained on the generated dataset to validate its potency. The proposed method, which combines Convolutional Neural Networks (CNN) and Long Short Term Memory (LSTM)/Bidirectional LSTM networks, has been compared to the established standards and more recent techniques. The model achieved an accuracy of [Formula: see text] on the benchmark and individual test sets. The results show that deep learning techniques have a promising future in protein structure analysis, implying that the dataset developed as part of our work will be a valuable resource for further protein structure research.