Predicting protein structures at any folding stage from sequence
| Source: arXiv AI
Tags: FoldScape, AlphaFold 3, Protein Folding, Rectified Flow, Computational Biology
FoldScape generates protein conformations at any specified folding stage from sequence, covering roughly ten times more of the folding coordinate than seed-varied AlphaFold 3 ensembles across 20 human proteins.
Details
Structure predictors find native conformations but do not describe the path from unfolded ensembles to folded states. FoldScape is a sequence-conditioned model that generates C-alpha ensembles at a chosen folding fraction Q. It rebuilds 3-D chains through hierarchical sliding-window regressions of inter-residue distances, then assembles and samples with a rectified-flow decoder. Training uses coarse-grained folding and all-atom mechanical-unfolding trajectories. Across 20 human proteins (36-857 residues), it covers about an order of magnitude more of the folding coordinate than AlphaFold 3 ensembles at matched sample size, with accuracy roughly constant along Q. AlphaFold 3 still anchors the near-native region better, so the authors position the two as complementary.