NLM DIR Seminar Schedule

Abstract:

AlphaFold has transformed structural biology but remains unreliable at predicting multiple conformations, a limitation whose mechanistic causes are unknown. Here we show that AlphaFold selects between structural outcomes using sparse subsets of residues. Using fold-switching proteins that adopt distinct conformations depending on sequence context as adversarial examples, we identify a hierarchical selection mechanism we term gating. Gating occurs when small networks of interacting residues with fold-specific sequence and geometric features determine whether mutations elsewhere influence the predicted conformation. Gating generalizes to diverse fold classes and to single-fold proteins without requiring explicit coevolutionary inference. Across proteins, AlphaFold concentrates mutational sensitivity more strongly than experimental measurements, and gating residues need not be thermodynamically significant. Thus, AlphaFold encodes a sparse, hierarchical decision process in which a small subset of residues governs prediction of structural states, providing a framework for understanding when and why its predictions diverge from physical reality.