Protein structure quality assessment by deep learning

Proteins, considered fundamental building blocks of life, play an important role across various scientific domains, including novel protein design, protein structure prediction, and new drug discovery. Each of these fields necessitates the detailed analysis of a large number of protein structures. Traditionally, the determination of protein structures through biological and physical methods has been prohibitively expensive. Consequently, computational methods for predicting protein structures have gained significant traction. Such methods often generate multiple structural approximations, or decoys, which are then evaluated using protein quality assessment (QA) techniques to identify the most promising candidates for further research. This evaluation process, also known as estimation of model accuracy (EMA), is crucial for advancing our understanding of protein. The task of protein structure QA can be viewed as mapping a protein's structure to its evaluative score, a process increasingly addressed through deep learning technologies. These technologies, recognized for their powerful pattern learning capabilities, are at the forefront of tackling significant research challenges within this domain. This dissertation describes five major contributions to the field. Firstly, DeepRank2, a multimodel method, utilizes a 1D multi-layer perceptron to predict the GDT-TS score of tertiary protein structures. Secondly, the performance of DeepRank2 and its variants was assessed in the CASP14 category for model accuracy estimation. Thirdly, DISTEMA employs an attentive 2D convolutional neural network (CNN) that relies solely on distance-based features to predict the GDT-TS score of tertiary structures. Fourthly, DProQA introduces an innovative gated-graph transformer approach for evaluating protein quaternary structures. Lastly, this work includes a comprehensive survey on deep learning-based methods for quaternary protein structure QA, covering relevant datasets, evaluation metrics, protein representations, and advancements in deep learning techniques over the past five years. Both DISTEMA and DProQA have been released as open-source software and are available to the scientific community.