Title：Inside AlphaFold2: How and Why Does It Work?
Time：2023/03/02 (Thu.) 14:30
Place：R218, Life Science Building II, NTHU

Abstract：
AlphaFold2, published by DeepMind in July 2021 [Jumper et al., Nature 596, 583], enabled a structure prediction of proteins from mere amino acid sequence information. AlphaFold2 uses the multiple sequence alignment (MSA) of the target protein homologs and emits precisely predicted structures even when no structures of the homologs exist. AlphaFold2 consists of two components - Evoformer and Structure module. The former extracts the sequence and MSA information into the neural network (NN) representation, while the latter constructs the mainchain 3-dimensional structure from the NN representation. In this talk, I shed a light on the Structure module and its mathematical foundation that achieved the prediction. By combining the AlphaFold2 Structure module with the classical molecular structure generation, I will explain what enabled the prediction of the structure that was unprecedented in the previous structure prediction approaches.