Loading…
Diffusion model that generates an ordered protein folding trajectory from sequence, predicting intermediates and transition states, not one fold.
No providers recorded yet. Browse all providers
A single predicted structure says what a protein ends up as, not the order in which it got there. For a biochemist asking which hairpin nucleates first or which residues are already native-like in the transition state, the route is the object of study — and it has historically been reachable only through Φ-value analysis, hydrogen exchange, single-molecule pulling, or molecular dynamics runs too expensive to do at scale.
PathFold, from Daisuke Kihara's group at Purdue University with a collaborator at the University of Edinburgh, predicts the whole route. Given a sequence it emits an ordered series of Cα structures running from a fully extended chain to the native fold. The generator is a denoising diffusion model conditioned on the single and pair representations produced by the AlphaFold 2 Evoformer, which encode the native structure that serves as the endpoint of folding, together with one or more preceding intermediate states supplied as contact maps. Each iteration predicts the next intermediate, and the selected structure becomes the input for the one after it.
That ordering is what separates PathFold from the sequence-to-conformation generators it sits beside. AlphaFlow and BioEmu sample equilibrium ensembles — sets of plausible conformations with no intrinsic sequence. The authors are direct about the cost of chasing ordered trajectories instead: large-scale atomistic folding trajectories do not exist, so PathFold was trained on reversed steered molecular dynamics (SMD) unfolding runs, a surrogate that supplies ordered conformational transitions but does not reproduce genuine folding kinetics or thermodynamics.
PathFold is a denoising diffusion probabilistic model built on the Genie backbone and
extended with conditioning inputs. Cα coordinates are converted to Frenet–Serret frames
and passed through an SE(3)-invariant encoder producing single and pair encodings, which
are fused with linearly projected AF2 representations (single L × 384, pair
L × L × 128) and with the previous state's contact map encoded by two ResNet blocks. An
SE(3)-equivariant decoder using Invariant Point Attention predicts the next state's Cα
coordinates against a per-residue L2 loss. Training used AdamW at a learning rate of
1e-3 and batch size 4, and took roughly eight weeks on one 48 GB NVIDIA RTX 6000.
The data came from a January 2019 PISCES set filtered to below 25% sequence identity and 2.5 Å resolution, from which 1,290 chains of 400 residues or fewer were pulled by harmonic-spring SMD in NAMD with CHARMM36. Training used 940 proteins and 728,597 frames, with 175 proteins held out for validation and 175 more for testing. Mean PathScorer alignment scores on the test set were 0.804, 0.815 and 0.733 for PathFold-1, -3 and -6; for scale, five independent SMD replicas of the same protein agree with each other at 0.894. Inference takes under 45 minutes for a 400-residue protein. On the FBP28 WW domain, cold shock protein B and ubiquitin, residue folding scores derived from the predicted trajectories tracked the experimentally measured Φ-value profiles — WW domain residues 8–23 averaged 0.79 against 0.43 elsewhere.
PathFold suits researchers who want a testable hypothesis about folding order before committing to experiments or simulation: which secondary-structure elements form early, which domains assemble first in a multi-domain protein, and which residues are native-like in the transition state and therefore worth mutating for a Φ-value measurement. Running it requires AF2 single and pair embeddings plus a folded reference structure, so it slots after a structure-prediction step rather than replacing one.
PathFold is the first deep learning method to generate complete, ordered folding pathways, and PathScorer gives the field a metric for comparing trajectories that differ in rate and duration. The limits are stated plainly by the authors: supervision comes from reversed mechanical unfolding rather than real folding, the model operates at Cα resolution with no side chains, and the native structure enters as conditioning, so this is pathway inference downstream of structure prediction rather than independent of it. The work is a preprint and has not been peer reviewed; code and all three checkpoints are released under GPL-3.0.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.