AlphaFold2 Reveals Encoded Protein Conformational Landscapes Through Neural Spectroscopy.

AlphaFold2, with its massive parameter count, is typically viewed as a tool for predicting protein structures from sequences. However, new research proposes treating the model itself as a scientific object, capable of revealing deeper insights into protein behavior. By applying a technique called "neural spectroscopy," specifically smoothing AlphaFold2's Evoformer weight tensors with a Gaussian convolution and scaling, researchers found that the model generates physically structured conformational landscapes. These landscapes reflect how proteins might behave under perturbation. For example, when ubiquitin was perturbed, its native contacts broke in an order consistent with decades of experimental folding studies. This suggests that AlphaFold2's weights encode structural constraints and conformational organization that emerged as a byproduct of its training, extending beyond its explicit task of predicting static structures.