Protein contact distance and structure prediction driven by deep learning

Proteins, fundamental building blocks of living organisms, play a crucial role in various biological processes. Understanding protein structure is essential for unraveling their functions and designing therapeutics. However, experimentally determining protein structures is time-consuming and expensive, motivating the development of computational methods. The prediction of protein tertiary structure relies to a certain extent on the accurate prediction of protein secondary structure and protein contact/ distance map. A high-quality contact distance prediction is crucial in constructing an ideal protein tertiary structure. Similarly, accurate prediction of distances between protein chains aids in the construction of higher-quality protein complex structures, also known as quaternary structures. In recent years, the advancement of deep learning techniques and the continuous expansion of protein sequence databases has significantly improved the accuracy of protein contact distance prediction, consequently impacting the prediction of protein tertiary and quaternary structures. This dissertation presents four contributions. First, DNSS2, an innovative approach based on one-dimensional deep convolutional networks, is proposed for the accurate prediction of protein secondary structure. Secondly, DeepDist introduces a multi-task deep learning framework that facilitates the prediction of real-valued distances between residues. Thirdly, DeepDist2 represents an enhanced version of the deep learning-based protein distance prediction tool. Finally, CDPred, a 2D attention-based deep neural network is developed to predict inter-chain distances in protein complexes. All the methods are available as software tools or web servers which are freely available to the scientific community.