The neural network that doesn't just predict proteins: it holds their secrets express
Imagine a vast library where every book is a precise instruction manual for assembling one protein molecule. AlphaFold2 read millions of such books and learned to instantly produce a ready-made 3D model. But inside this network, in its very 'brain,' something more remained — not just assembly rules, but an entire atlas of possible protein motions, its hidden life.
Researchers devised a way to peek into that atlas. They took the weight tensors (matrices of numbers storing the network's knowledge) from the Evoformer block — a key part of AlphaFold2—and applied a Gaussian blur convolution to them. In simpler terms, they slightly 'defocused' the network's vision to see not one sharp image but a whole fan of possibilities. They call this method neural spectroscopy.
The result was astonishing. For the protein ubiquitin, such a 'perturbation' showed that its parts come apart in exactly the order that decades of laboratory folding experiments had established.
For the protein KaiB, which can adopt two completely different shapes, all five independently trained versions of the network unanimously showed that the second shape cannot be reached through perturbation. And for alpha-synuclein, linked to Parkinson's disease, the five versions produced five different but internally consistent maps. This indicated where the training data gave a clear signal and where the network remained ignorant. A control check, injecting random noise of the same magnitude, produced only formless garbage, not meaningful motions. The network's 93 million parameters turned out to be not just a mechanism but a frozen imprint of the physical reality of proteins.
Original abstract
as written by the authors arXiv:2607.16087
AlphaFold2's 93 million parameters, shaped by the evolutionary record of protein structure encoded in the Protein Data Bank and in sequence alignments, are conventionally treated only as machinery for converting sequence to structure. We propose they are also a scientific object that can be analyzed directly: a learned encoding of protein conformational organization that can be probed and characterized. By smoothing the Evoformer's weight tensors with a Gaussian convolution and scaling the result, we show that the trained model produces physically structured conformational landscapes. Under perturbation, ubiquitin's native contacts break in the order established by decades of folding experiments. For KaiB, five independently trained models agree that the alternative fold is not recovered under perturbation. For alpha-synuclein, five models produce five different but coherent landscapes, mapping where the training signal has determined the representation and where it has not. Matched-power noise controls confirm that random corruption of equal magnitude produces debris, not conformations. The model learned to predict static structures; the conformational organization visible under perturbation was not an explicit training target, suggesting it emerged as a byproduct of that objective. AlphaFold2's weights appear to encode structural constraints, shaped by evolutionary and structural training data, that extend beyond what unperturbed inference reveals. We call the approach of reading them neural spectroscopy, and Scaled Gaussian Convolution one such protocol.
The protein ubiquitin, named for its ubiquitous presence (from the English 'ubiquitous'), consists of just 76 amino acids, yet its folding pathway was studied for decades, and the neural network reproduced it simply by 'pondering' the static shape.
This echoes the idea from 'Jurassic Park,' where scientists extracted dinosaur DNA from amber — except here, hidden molecular dynamics are extracted from a 'frozen' neural network.