How has AlphaFold changed protein structure prediction and what are its limitations?

AlphaFold2 achieves near-experimental accuracy for many protein structures using deep learning on sequence-structure relationships, multiple sequence alignments, and attention mechanisms. It revolutionized structural biology by providing predictions for most known proteins (AlphaFold DB). However, limitations include: uncertainty with proteins lacking homologs, difficulty with intrinsically disordered regions, inability to predict effects of post-translational modifications, challenges with protein-ligand complexes (though AlphaFold-Multimer addresses some protein-protein interactions), no direct dynamics or conformational flexibility information, and potential bias toward crystallographic conformations. Experimental validation remains essential for novel targets. Integration with molecular dynamics and cryo-EM fitting extends utility.