Kaustav Mehta
AlphaFold2 is normally treated as a machine: sequence goes in, structure comes out, and the 93 million parameters in between are plumbing. The proposal here is to treat those parameters as a scientific object in their own right.
The reasoning is that the weights were shaped by the entire evolutionary record of protein structure, everything in the Protein Data Bank and the sequence alignments behind it. Whatever the network learned about how proteins organise themselves is sitting in there, encoded, whether or not anyone reads it out. The paper probes it by smoothing the Evoformer's weight tensors with a Gaussian convolution, borrowing the instinct of spectroscopy: perturb something systematically and see what structure the response reveals.
It is an unusual inversion. Instead of asking what the model can predict, it asks what the model has come to know, and treats a trained network as a fossil record of the data that formed it.
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…
DAPGNet: Dynamic Adaptive Physics-Guided Graph Diffusion Network for Hyperspectral Image Classification
arXiv (cs.AI) · July 16, 2026NIFA: Nonlinear IMC enhanced FPGA for efficient ML inference
arXiv (cs.LG) · July 16, 2026Learning in Infinitesimal Non-Compositional Sketches
arXiv (cs.AI) · July 16, 2026Long-Context Fine-Tuning with Limited VRAM