Neural spectroscopy of AlphaFold2 reveals encoded protein conformational landscapes
The paper treats AlphaFold2’s trained weights as something to probe directly, not just as machinery for structure prediction.
It proposes “neural spectroscopy,” using Scaled Gaussian Convolution on Evoformer weight tensors to expose conformational landscapes. In tests, ubiquitin contacts break in the experimentally known folding order, while KaiB’s alternative fold does not reappear under perturbation across five models. For alpha-synuclein, separate models yield different but coherent landscapes, which the author reads as a map of where training fixed the representation and where it did not. Equal-power noise controls produce debris rather than conformations. ArXiv · AI/CL/LG's note
It proposes “neural spectroscopy,” using Scaled Gaussian Convolution on Evoformer weight tensors to expose conformational landscapes. In tests, ubiquitin contacts break in the experimentally known folding order, while KaiB’s alternative fold does not reappear under perturbation across five models. For alpha-synuclein, separate models yield different but coherent landscapes, which the author reads as a map of where training fixed the representation and where it did not. Equal-power noise controls produce debris rather than conformations. ArXiv · AI/CL/LG's note
score 5