AlphaFold, small-angle X-ray scattering and ensemble modelling: a winning combination for intrinsically disordered proteins
Résumé
Artificial intelligence has revolutionized many societal and scientific domains, and structural biology benefits from one of its most spectacular breakthroughs. The deep learning algorithm AlphaFold can predict with unparalleled accuracy the threedimensional structure of folded proteins based solely on their sequence. In 1972, Christian Anfinsen was awarded the Nobel Prize in Chemistry for his groundbreaking work on ribonuclease which demonstrated that all the protein's folding information is encoded within its sequence. This ignited a surge of experimental and computational investigations into protein folding to unravel what was at the time referred to as the 'second half of the genetic code' and understand the mechanisms governing the protein's pathway towards its final three-dimensional conformation. Now, half a century after Anfinsen's seminal article (Anfinsen, 1973), it appears that a long-awaited milestone has been reached and that one can now correlate a onedimensional string of amino acids to an accurate three-dimensional structure. However, challenges still persist. The fundamental nature of artificial intelligence lies in its initial training using data that have accumulated over many years, the Protein Data Bank (PDB) in the case of AlphaFold. The PDB suffers from inherent biases: certain families of structures are under-represented due to the limitations of the methods used to solve the protein structures, in particular X-ray crystallization and cryo-electron microscopy (cryoEM). This is typically the case of proteins that exhibit a high conformational flexibility and cannot be described by a single three-dimensional structure. In particular, intrinsically disordered proteins (IDPs) or regions (IDRs) of proteins can adopt a tremendous number of conformations and this high dynamic is inherent to their function. Because of their flexibility, these polypeptides are often invisible in the structures solved by X-ray crystallization or cryoEM. Their flexibility may even preclude protein crystallization or lead to poorly diffracting crystals. In addition, a short disordered segment can undergo an induced folding upon binding to a partner, but may adopt different folds depending on the binding partner and the interface of interaction, further complicating accurate predictions. Therefore, AlphaFold could not train properly on disordered structures, which lack structural homologues in the PDB. Furthermore, AlphaFold uses multiple sequence alignments, while IDRs are very difficult to align because they are poorly conserved. For all these reasons AlphaFold cannot predict for now the 3D structure of long flexible disordered regions with accuracy and provides low confidence scores on these predictions. AlphaFold's creators even suggest that the low confidence scores can be used as a method to predict protein disorder. Therefore, the single static representation of the structure of proteins containing IDRs provided by AlphaFold cannot convey all their structural and functional properties. A very interesting strategy to circumvent this difficulty is to integrate AlphaFold predictions with ensemble molecular modelling and experimentally derived constraints, as proposed by Brookes et al. (2023) in their paper entitled AlphaFold-predicted protein structures and small-angle X-ray scattering: insights from an extended examination of selected data in the Small-Angle Scattering Biological Data Bank. Small-angle X-ray scattering (SAXS) is a particularly suitable experimental method for providing such constraints. SAXS data contain all the 3D information on the various conformations adopted by a macromolecule in solution. By calculating the SAXS profile from atomic coordinates and comparing it with experimental SAXS data, it becomes possible to select an ensemble of conformations that collectively contribute to the observed scattering
Origine : Fichiers éditeurs autorisés sur une archive ouverte
Licence : CC BY - Paternité
Licence : CC BY - Paternité
