NLM DIR Seminar Schedule

Abstract:

In recent years, advances in artificial intelligence have significantly transformed the field of structural biology, particularly in protein structure prediction. While AlphaFold2 often predicts structures that resemble known conformations in the Protein Data Bank (PDB), the question remains: has the protein structure prediction problem truly been solved? AlphaFold2 typically produces models representing a single conformational state with limited structural diversity. To address this, various enhanced sampling techniques have been developed to broaden the range of predicted structures or to steer predictions toward specific conformers or folds. However, subsampling of multiple sequence alignments (MSAs) and other methods aimed at increasing structural diversity can sometimes result in lower-confidence models, often favoring familiar structures from its training data over novel folds. We explored both the strengths and potential limitations of AlphaFold (AF2 and AF3), particularly in predicting the alternative conformations of fold-switching proteins. These features point to approaches in predicting alternative conformations more reliably.