Since proteins play vital roles in cell and life functioning, researchers have spent lots of time and resources in determining protein structures using traditional experimental approaches such as X-ray diffraction, solution NMR, and cryo-EM. Thanks to the rapid growth of information technology, in silico methods grow fast in predicting protein structures these years. AlphaFold2 is a model published by Deepmind presenting outstanding performance in CASP14, 2021, raising our interest to conduct some case analyses to tell its pros and cons. The network architecture, mechanism, and performance of AlphaFold2 are systematically reviewed. Then taking T-cell surface glycoprotein CD4 (Human CD4) for instance, experimental and predicted structures of the protein are obtained, aligned, and analyzed to evaluate the predicted Human CD4 protein structure generated by AlphaFold2. By aligning experimental structures, the structural changes of Human CD4 tend to be caused by the combination between Human CD4 and other tertiary structures and predicted structures. By aligning predicted structures with experimental ones, the predicted structure generated by AlphaFold2 is proved to be more similar to that in combination situations. Since proteins function when they interact with other small molecules or macromolecules, hotspot residues are detected and their prediction confidence is reviewed. The similarity between predicted structures and experimental structures in their combination situations is beneficial for structure-based drug discovery. The work suggests that applying deep-learning methods in protein structure prediction, i.e., AlphaFold2 will bring about further potential and development in novel drug discovery.
|