An analysis of the mechanism and prospect of AlphaFold2 model based on machine learning and neural network

Abstract

Protein is one of the most important biomolecules. For decades, the structure of proteins has been the main focus of structural biologists' research. It plays a decisive role in understanding the molecular mechanism of proteins in biological processes. Unfortunately, the number of protein structures determined by experiments is still minimal, and the task of accurate positioning is a difficult process, usually requiring a lot of time and effort. In addition, the accuracy and reliability of structure prediction of existing computational methods are not satisfactory. This paper comprehensively analyses and integrates the latest and most advanced computational method for protein structure prediction, AlphaFold2, its application on the human proteome, and a complex prediction architecture AlphaFold-Multimer adapted from AlphaFold2. We mainly explain the basic mechanism and performance of the models from three aspects: the fundamental framework of the models, the performance on CASP14 and public datasets, and case studies of the models. Firstly, we decipher the primary networks of AlphaFold2 and AlphaFold-Multimer by paraphrasing the typical constituents of general machine learning models. Then, the high accuracy and reliability on CASP14 and public data sets prove the excellent performance of this model. In addition, we also take several protein cases to illustrate the preeminent performance further. Finally, we discuss the limitations of the models such as the lack of capability to predict nonprotein components, while also exploiting the non-negligible potential application of the model with distinctive insight like the organic combination of experimental and computational methods or the tremendous effect on the field of medicine.